Thesis

Graph Conne
tivity: Relaxations and Algorithms
by
Adrian Roshan Vetta
B.S . E onomi s, London S hool of E onomi s 1994
M.S . Mathemati s, London S hool of E onomi s 1995
Submitted to the Department of Mathemati s
in partial fulllment of the requirements for the degree of
Do tor of Philosophy
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
June 2002
Adrian Roshan Vetta, MMII. All rights reserved.
The author hereby grants to MIT permission to reprodu e and
distribute publi ly paper and ele troni opies of this thesis do ument
in whole or in part.
Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Department of Mathemati s
May 1, 2002
Certied by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Santosh Vempala
Assistant Professor
Thesis Supervisor
A epted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chair, Applied Mathemati s Committee
A epted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Tomasz S. Mrowka
Chair, Department Committee on Graduate Students
2
Graph Conne tivity: Relaxations and Algorithms
by
Adrian Roshan Vetta
Submitted to the Department of Mathemati s

on May 1, 2002, in partial fulllment of the
requirements for the degree of
Do tor of Philosophy
Abstra t
The task of nding minimum ost subgraphs with pres ribed onne tivity require-
ments is a fundamental problem in network optimisation. In general, though, this
task is NP-hard. As a onsequen e, there has been mu h interest in the development
of approximation algorithms for su h network problems. This thesis fo uses predomi-
nantly upon the design of approximation algorithms for the two lassi al problems of
this genre: the k-edge onne tivity problem and the k-vertex onne tivity problem.
For the k-vertex onne tivity problem in undire ted graphs we present two dierent
types of algorithm. The rst is a k-phase augmentation algorithm whi h in reases the
vertex onne tivity of a solution subgraph by one in ea h phase. The se ond algorithm
is based upon the te hnique of iteratively rounding a linear program. The iterative
rounding method also works for the k-vertex onne tivity problem in dire ted graphs,
and for the element onne tivity problem. For the spe ial ase of bi- onne tivity prob-
lems in unweighted graphs, we present ombinatorial algorithms that give improved
approximation guarantees. These algorithms have two broad phases. In the initial
phase de omposition methods are applied whi h allow us to fo us upon subgraphs
with a simplied stru ture. In the nal phase a feasible solution is obtained by om-
bining te hniques from mat hing theory with lo al sear h methods.
Thesis Supervisor: Santosh Vempala
Title: Assistant Professor
3
4
A knowledgments
I am very grateful to my advisor Santosh Vempala. I ouldn't have hoped for a better
advisor than Santosh; he is also a very ni e person. I would also like to thank the
other members of my thesis ommittee, Dimitris Bertsimas and Mi hel Goemans,
for their omments and insights regarding this thesis. I am also grateful to Bru e
Shepherd, Madhav Marathe and Joseph Cheriyan for an assortment of reasons.
Thanks go to my old atmates Pete and Fran is for the odd pint, and to Pablo
for the footy mat hes. The visits from home of the following people served as a
sanity he k: Amber, Amita, Anne, Billy, dad, Dave, Elysia, Justin, Konni, Lambert,
Messer, mum, Paul, Ralphie, Rohit, Simmy and Sohit. Last, but not least, thanks to
Juli.
5
6
Contents
1 Introdu tion 9
2 Vertex Conne tivity in Weighted Graphs 13
2.1 Setpair Relaxations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.1.1 The Cut Covering Model . . . . . . . . . . . . . . . . . . . . . 14
2.1.2 The Setpair Model . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Conne tivity Augmentation . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 The k-Out onne ted Subgraph Problem . . . . . . . . . . . . 17
2.2.2 Augmentation Algorithm for k-Vertex Conne tivity . . . . . . 19
2.3 Iterative Rounding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.1 Setpairs and Skew Bisupermodularity . . . . . . . . . . . . . . 26
2.3.2 Chara terising a Basi Solution via a Non-Overlapping Family 28
2.3.3 Iterative Rounding and k-Vertex Conne tivity . . . . . . . . . 31
2.3.4 Iterative Rounding and Element Conne tivity . . . . . . . . . 35
3 Bi-Conne tivity in Unweighted Graphs 45
3.1 Two Simple Approximation Algorithms . . . . . . . . . . . . . . . . . 47

3.1.1 The DFS Tree Method . . . . . . . . . . . . . . . . . . . . . . 47
3.1.2 The Cy le Contra tion Method . . . . . . . . . . . . . . . . . 47
3.2 The Lower Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.2.1 The Lower Bound in Undire ted Graphs . . . . . . . . . . . . 48
3.2.2 The Lower Bound in Dire ted Graphs . . . . . . . . . . . . . . 49
3.3 The De omposition Theorems . . . . . . . . . . . . . . . . . . . . . . 49
7
3.3.1 2-Edge Conne tivity . . . . . . . . . . . . . . . . . . . . . . . 50
3.3.2 2-Vertex Conne tivity . . . . . . . . . . . . . . . . . . . . . . 56
3.3.3 Strong Conne tivity . . . . . . . . . . . . . . . . . . . . . . . 57
3.4 The Component Graph . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.5 In orporating a Depth First Sear h Tree . . . . . . . . . . . . . . . . 62
3.6 Edge Conne tivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.6.1 The Lo al Improvement Method . . . . . . . . . . . . . . . . . 65
3.6.2 Comparison with the Subtour Relaxation . . . . . . . . . . . . 78
3.7 Vertex Conne tivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.8 Strong Conne tivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.8.1 Tools and Operations . . . . . . . . . . . . . . . . . . . . . . . 93
3.8.2 The Lo al Improvement Method . . . . . . . . . . . . . . . . . 95
8
Chapter 1
Introdu tion
Let G = (V; E ) be a graph (undire ted or dire ted), with n verti es and m edges.
Asso iated with ea h edge e is a ost e. Given this framework, the task of nding
low ost spanning subgraphs of a pres ribed onne tivity is a fundamental problem
in network optimisation. This thesis fo uses predominantly upon the two lassi al
problems of this genre: the k-edge onne tivity problem (kEC ) and the k-vertex on-
ne tivity problem (kV C ). In undire ted graphs, the k-edge onne tivity problem is
to nd the minimum ost subgraph in whi h there are at least k edge disjoint paths
between every pair of verti es. In the k-vertex onne tivity problem we require at
least k vertex disjoint paths between every pair of verti es. For k = 1 we have the
minimum spanning tree problem, whi h an be solved exa tly in polynomial time.
In general, however, these problems are hard. For example, even the problems of
nding the minimum ardinality 2-edge onne ted spanning subgraph (2EC ) or 2-
vertex onne ted spanning subgraph (2V C ) are NP-hard. This an be seen via a
simple redu tion from the Hamiltonian y le problem. Moreover, it is known that
both these problem are maxSNP-hard ([9℄, [3℄). In dire ted graphs, the k-edge on-
ne tivity problem is to nd the minimum ost subgraph in whi h, for ea h pair of
verti es u and v, there are at least k edge disjoint dire ted paths from u to v, and at
least k edge disjoint dire ted paths from v to u. The dire ted version of the k-vertex
onne tivity problem is dened analogously.
Given the hardness results, this thesis will on entrate on approximation algo-
9
rithms for onne tivity problems. Hen e, we desire subgraphs whose ost is guaran-
teed to be within a ertain fa tor of that of the optimal solution. An overview of
this thesis is as follows. In Chapter 2 we onsider onne tivity problems in weighted
graphs. We begin by presenting a linear programming formulation (with onstraints
upon the edges indu ed by two disjoint subsets of verti es) that proves useful in
the design and analysis of onne tivity algorithms. Sin e there exist very good ap-
proximation algorithms for edge onne tivity problems in weighted graphs (see, for
example, Jain [25℄), our fo us will be on vertex onne tivity problems. In parti u-
lar, we will on entrate on the k-vertex onne tivity problem. We give two dierent
types of algorithm. The rst algorithm, given in Se tion 2.2, is a k phase augmen-
tation algorithm whi h in reases the vertex onne tivity of a solution subgraph by
one in ea h phase. The key stru tural result, here, is that in phase r, 1 r k, a
suitable r-vertex onne ted subgraph an be obtained from the union of three easily
onstru ted subgraphs. The performan e of the algorithm is analysed via the linear
programming relaxation. This algorithm gives an O(log k) approximation guarantee
in undire ted graphs ontaining at least 6k2 verti es.
The se ond algorithm, given in Se tion 2.3, is based upon the iterative rounding of
the setpair linear program (LP). At ea h step in the algorithm an edge is judi iously
sele ted (by onsideration of the weight given to the edge by the LP solution) until a k-
vertex onne ted subgraph is obtained. The existen e, at ea h step, of a high weight
edge in the LP solution is suÆ ient to produ e a good approximation guarantee.
It is shown that for a large lass of fun tions a basi LP solution is hara terised
by a \non-overlapping family" of setpairs. Su h families, in turn, guarantee the
existen e of a high weight edge. In parti ular, for the k-vertex onne tivity problem
p
iterative rounding gives an O( n" ) approximation guarantee in both undire ted and
dire ted graphs, where k (1 ")n. The previous best approximation guarantee for
the undire ted k-vertex onne tivity problem, due to Kortsarz and Nutov [31℄, was
O(k). For dire ted graphs the only known non-trivial guarantee was O( m), due
p
to Cheriyan and Vempala [6℄. We also show how the iterative rounding method an
be applied to give a fa tor 2 approximation guarantee for the element onne tivity
10
problem (to be dened). Chapter 2 is based upon joint work with Joseph Cheriyan
and Santosh Vempala.
In Chapter 3 we onsider unweighted graphs, that is, graphs for whi h e = 1
for ea h edge e. Here we present improved approximation guarantees for various bi-
onne tivity problems. The basi te hnique is as follows. We apply an assortment of
de omposition theorems that allow us to fo us upon a lass of graphs with a simplied
stru ture. Then, mat hing theory te hniques are used to nd subgraphs from whi h
bi- onne ted subgraphs an be onstru ted via lo al improvement methods. These
methods are shown to produ e fa tor 43 -approximation algorithms for the 2-edge on-
ne tivity and 2-vertex onne tivity problems in undire ted graphs. These results
improve upon the previous best guarantees of 1217 and 23 , due to Cheriyan et al [5℄ and
Garg et al [18℄, respe tively. These methods also produ e a fa tor 23 -approximation
algorithm for the strong onne tivity problem (this is the dire ted 1-edge (vertex) on-
ne tivity problem) in dire ted graphs. This approximation guarantee improves upon
the previous best guarantee of 1:617 due to Khuller et al [29℄. The results in Chapter
3 regarding undire ted graphs are based upon work with Santosh Vempala.
11
12
Chapter 2
Vertex Conne tivity in Weighted
Graphs
In this hapter, we onsider onne tivity problems in weighted graphs; the results
presented here are drawn from the papers [7℄ and [8℄. As stated, our fo us will be on
approximations algorithms. Good approximation algorithms exist for weighted edge
onne tivity problems in undire ted graphs. In parti ular, Jain [25℄ has shown that
the te hnique of iterative rounding may be applied to the ut overing LP formula-
tion, des ribed in Se tion 2.1, of the generalised Steiner network problem to give a
fa tor 2-approximation algorithm. Consequently, we will fo us on vertex onne tivity
problems in this hapter.
In parti ular, we present two approximation algorithms for the k-vertex onne -
tivity problem. The rst algorithm works by su essively augmenting by one the
onne tivity of an initial subgraph until we obtain a feasible solution. The se ond
algorithm is based upon iterative rounding of the setpair LP relaxation. The augmen-
tation algorithm only works on undire ted graphs; the iterative rounding algorithm
works for both undire ted and dire ted graphs. We brie y dis uss how our results
ompare with to the previous best approximation guarantee of more than k=2 due to
Kortsarz and Nutov [31℄. The onne tivity augmentation algorithm has an approxi-
mation guarantee of O(log k). This algorithm, however, only works on graphs with at
least 6k2 verti es. Thus the guarantee is a stri t improvement on the previous guar-
13
p
antee for 56 k n=6, and is omparable (within a onstant fa tor) for k < 56.
This algorithm also provides, for graphs with at least 6k2 verti es, an O(log k) upper
bound on the integrality ratio for the setpair LP relaxation. The iterative round-
p
ing algorithm provides an approximation guarantee of O( n=), for any > 0 and
p
k (1 )n. This is stri t improvement for
( n) k n
(1); it is omparable
for k = (pn) and k n (1).
2.1 Setpair Relaxations
In this se tion we present two integer program formulations that are useful in network
design problems. The rst formulation has been widely used in edge onne tivity
problems. The se ond formulation generalises the rst formulation and an be used
for vertex onne tivity problems.
2.1.1 The Cut Covering Model
Our rst formulation is known as the ut- overing model. Here there is a requirement
fun tion f that pla es a onstraint upon the edges Æ(S ) indu ed by ea h ut (S; V n S )
in the graph.
X
min e xe
X
2
e E
s:t: xe f (S ) 8S V
2( )
e Æ S
xe 2 f0; 1g 8e 2 E
Now onsider the following fundamental theorem of Menger.

Theorem 2.1.1. The maximum number of pairwise edge-disjoint (i; j )-paths equals
the minimum number of edges whose deletion separates i and j .
By Menger's Theorem, the k-edge onne tivity problem may be given using the
ut overing formulation. Simply set f (S ) = k, for all S , ; S V . In fa t,
14
the generalised Steiner network problem may also be formulated via the ut overing
model. Here, given two verti es i and j , we require rij edge disjoint paths between i
and j . Setting f (S ) = maxi2S;j2V nS rij allows us to use the ut overing model.
2.1.2 The Setpair Model
A more general integer programming formulation, developed by Frank and Jordan

[12℄, is the setpair model. A setpair W = (Wt; Wh) is an ordered pair of disjoint vertex
sets. We say that Wt is the tail of W , and Wh is the head of W . Let S denote the set of
all setpairs. We abuse our notation slightly and denote by Æ(W ) = Æ(Wt ; Wh) the set
of edges with one end-vertex in Wt and the other in Wh. We set W 0 = V n (Wt [ Wh).
Then, given a requirement fun tion f that pla es a onstraint upon ea h edge set
indu ed by a setpair, we obtain the following integer program.
X
min e xe
2
X
e E
s:t: xe f (W ) 8W 2 S
2( )
e Æ W
xe 2 f0; 1g 8e 2 E
Note that a ut (Wt; V n Wt) is indu ed when the head and tail of W form a om-
plementary pair of vertex sets. Thus, the ut- overing model is just a spe ial ase of
the setpair model in whi h requirements are only pla ed upon setpairs W for whi h
jW 0j = 0. We remark that Menger's Theorem an be re-formulated in terms of
vertex-disjoint paths.
Theorem 2.1.2. For non-adja ent verti es i and j , the maximum number of pairwise
(internally) vertex-disjoint (i; j )-paths equals the minimum number of verti es from
V n fi; j g whose deletion separates i and j .
Then, by Menger's Theorem, the k-vertex onne tivity problem may be given
using the ut- overing formulation. Simply set f (W ) = max(k jW 0j; 0), for ea h
setpair W .
15
Given the omplexity of our onne tivity problems, we do not expe t to be able to
solve these integer programs exa tly in polynomial time. Thus, the fo us of our atten-
tion will be their linear program relaxations. Observe that there are an exponential
number of onstraints in these relaxations. However, for many of the spe i prob-
lems in network design that are aptured by these models, the linear programming
relaxation an be written as a ompa t linear program via a \ ow formulation". As
we shall see, this is true for the k-edge and k-vertex onne tivity problems. Moreover,
optimal solutions an be omputed in strongly polynomial time [38℄.
16
2.2 Conne tivity Augmentation
In this se tion, we present an approximation algorithm for the k-vertex onne ted
subgraph problem. The algorithm onsists of k phases; during ea h phase the vertex
onne tivity of the urrent solution is augmented by one. The algorithm works under
the restri tion that n 6k2; the method fails for graphs of smaller order. The ap-
proximation guarantee is 6 times the kth harmoni number, H (k), whi h is O(log k).
This is also an upper bound on the integrality ratio of the setpair LP for the k-vertex
onne tivity problem. The previous best upper bound on the integrality ratio was
(k). (An O(log k) approximation guarantee was laimed earlier in [37℄, but sub-
sequently an error has been found and that laim is withdrawn.) Our algorithm is
based on two results:
(i) A polynomial-time algorithm of Frank and Tardos [13℄ for nding a minimum ost
k-out onne ted subgraph in a dire ted graph.
(ii) An upper bound on the order of 3- riti al graphs by Mader [32℄.
The Frank-Tardos algorithm has been applied earlier to the minimum ost k-
vertex onne ted subgraph problem by several authors, beginning with Khuller and
Raghava hari [27℄; see also [1℄ and [2℄. The s aling tri k used in Lemma 2.2.4 has
been used earlier by Goemans et al [20℄, and also by Jain et al [26℄.
2.2.1 The k-Out onne ted Subgraph Problem
Consider again the setpair formulation of the k-vertex onne tivity problem. It will
be onvenient to keep a parameter `, where ` is a positive integer, and denote by P (`)
the LP relaxation of the `-vertex onne tivity problem.
X
P (`) : z (`) = min e xe
X
2
e E
s:t: xe ` jW 0 j 8W 2 S
2( )
e Æ W
0 xe 1 8e 2 E
17
Let opt(`) denote the ost of a minimum ost `-vertex onne ted subgraph. Then
P (`) gives a lower bound z (`) on the ost opt(`). In parti ular, P (k) gives a lower
bound z(k) on the ost opt = opt(k) of an optimal solution.
A graph is said to be k-out onne ted from a root vertex r if there exist k openly
disjoint paths from r to v, for ea h vertex v, v 6= r. The minimum ost k-out onne ted
subgraph problem is as follows: given an undire ted graph G = (V; E ) and a root vertex
r 2 V , nd a minimum ost subgraph that is k-out onne ted from r. This problem
is NP-hard for k 2. A fa tor 2 approximation algorithm has, though, been given
by Frank and Tardos [13℄ (see also Khuller and Raghava hari [27℄).
Theorem 2.2.1. There is an algorithm for the minimum ost `-out onne ted subgraph
problem that produ es a solution with ost at most 2z (`).
Proof. In the dire ted version Gb of G, ea h edge e of G is repla ed by two oppositely

oriented ar s, and ea h of these two ar s has ost e. The LP relaxation of the dire ted
minimum ost `-vertex onne ted subgraph problem on Gb has optimal value at most
2z(`); the dire ted subgraph orresponding to the undire ted solution of P (`) has
double the ost and is feasible for the dire ted LP.
An optimal solution to this LP is also a feasible solution to the following LP
relaxation Pb(`) of the dire ted minimum ost `-out onne ted subgraph problem on
Gb (with any vertex as the root). Denote by R the set of all setpairs W = (Wt ; Wh )
su h that the root r is in Wt and Wh 6= ;; let Æ^(W ) denote the set of ar s (u; v) in Gb
with u 2 Wt ; v 2 Wh.
X
Pb(`) : min a xa
2 ( )
Xb
a E G
subje t to xa ` jW 0 j 8W 2 R
2^( )
a Æ W
0 xa 1 8 a 2 E (Gb)
Now Pb has an integer optimal solution (see [11℄). The Frank-Tardos algorithm nds
a minimum ost subgraph Hb that is `-out onne ted from r, whose ost (Hb ) equals
18
the optimal value of Pb. The undire ted version of Hb is then a subgraph of G that is
`-out onne ted from r, with ost at most 2z (`).
3-Criti al Graphs
Let (G) denote the vertex onne tivity of G, that is, the minimum number of verti es
whose removal results in a dis onne ted graph. A -separator of a onne ted graph
is a set of verti es whose removal results in a dis onne ted graph.
A graph G = (V; E ) is alled r- riti al if the vertex onne tivity de reases by jS j
on removing the verti es in any set S of at most r verti es, that is, if
(G n S ) = (G) jS j; 8S V; jS j r
If G is not r- riti al, then there exists a set S of r verti es su h that no (G)-separator
ontains all the verti es in S . The algorithm whi h we present will work in polynomial
time, if r is a onstant, for any graph that is not r- riti al. Due to the following result
of Mader [32℄, we will fo us our attention on the ase r = 3.
Theorem 2.2.2. A 3- riti al graph of vertex onne tivity k has less than 6k2 verti es.
It is not know if this result is tight. It is known, however, that 3- riti al graphs
ontaining
(k ) verti es do exist [34℄.
3
2
Remark. Mader's proof [32℄ is written in German. He dis usses this result in two
survey papers written in English [33, 34℄.
2.2.2 Augmentation Algorithm for k-Vertex Conne tivity
The Algorithm
The algorithm starts begins with a minimum- ost spanning tree H1. At iteration i,
1 i k 1, the algorithm augments Hi to Hi+1 by adding edges from E (G) n E (Hi)
su h that the vertex onne tivity in reases by at least one. In addition, the augment-
ing ost, (Hi+1 ) (Hi ), is approximately minimum. A detailed des ription of an
19
iteration follows. Let ` = (Hi). If ` = k, then we stop and output Hi as the desired
solution. Now, suppose ` < k. For ea h edge in Hi, we hange the edge ost to zero;
the other edges keep the original osts. Hen e we use the augmenting ost of the edge.
By Mader's theorem, and the fa t that n is at least 6k2, there exist a set S of at most
three verti es su h that no `-separator of Hi ontains ea h vertex of S ; we all su h a
set non- riti al. We nd a non- riti al set, say r1; r2 and r3 , by exhaustively he king
for ea h vertex set S of ardinality three whether (Hi n S ) > ` 3. For ea h vertex
in our non- riti al set, we apply the Frank-Tardos algorithm with root rj (j = 1, 2,
or 3) and the modied edge osts, to nd a subgraph Hij that is (` + 1)-out onne ted
from rj . We take Hi+1 to be the union of (the edge sets of) Hi1, Hi2, Hi3 and Hi.
An Example
To illustrate how the algorithm works, we present a small example. Suppose we wish
to nd the minimum 2-vertex onne ted subgraph in the graph shown in Figure 2-1.
In the rst phase, see Figure 2-2, we begin with H0 = ;. As the vertex onne tivity
v1
1 1
1 1
v5 v2
3
3 3
2 2
v4 v3
2
Figure 2-1: An Example
of H0 is zero, the removal of any vertex does not de rease the onne tivity. Thus any
vertex indu es a non- riti al set S . So take v1 to be our non- riti al set. We then nd
a minimum ost 1-out onne ted subgraph H01 (in the dire ted version of the graph)
rooted at v1 using the augmenting osts. We then set H1 = H0 [ H01.
The se ond phase of the algorithm is shown in Figure 2-3. Sin e H1 has vertex
onne tivity one, the removal of any two verti es redu es the onne tivity by at
20
v1
1 1
1 1
3 3
2 2
H0 H10
Figure 2-2: Phase I
most one. Thus, any pair of verti es indu es a non- riti al set S . So take v1 and
v2 to be our non- riti al set. Using the augmenting osts we nd minimum ost 2-
out onne ted subgraphs H11 and H12 rooted at v1 and v2, respe tively. We then set
H2 = H1 [ H11 [ H12 . Observe that, as required, H2 is 2-vertex onne ted.
v1
0 0
0 0
3 3
2 2
2
1
H1
H1
v2
H1
2 H2
Figure 2-3: Phase II
We also remark that the algorithm works on this example despite the fa t that
n < 6k2 . In general, we will require that n 6k2 in order to guarantee the existan e
21
of non- riti al sets in ea h phase of the algorithm.
An Analysis
First we show that the algorithm does indeed work. That is, we show that the vertex
onne tivity in reases by one in ea h phase of the algorithm.
Lemma 2.2.3. We have (Hi+1 ) (Hi ) + 1, for ea h i 1.
Proof. Let ` = (Hi). Note that ` < k. Suppose that (Hi+1 ) = `. Then Hi+1 has
an `-separator C V . By Mader's theorem, Hi is not 3- riti al sin e n 6k2 > 6`2.
Hen e, there exist a set S of at most three verti es in Hi su h that for ea h `-separator
of Hi, at least one of verti es of S is absent from the `-separator. The algorithm nds
su h a non- riti al set, say r1; r2 and r3. Assume, without loss of generality, that r1 is
absent from C . The graph Hi1, whi h is a subgraph of Hi+1, is (` + 1)-out onne ted
from r1 . Therefore Hi+1 has (` + 1) openly disjoint paths between r1 and any other
vertex v. Observe that at least one of these paths survives in Hi+1 n C . Thus Hi+1 n C
is onne ted, a ontradi tion.
We now measure the quality of the solution produ ed by the algorithm.
Lemma 2.2.4. For ea h i 1, we have (H ) (H )
6z(k) , where ` = (H ).
i +1 i
k ` i
Proof. Note that ` < k. We will prove that for ea h of the three supergraphs Hij
(j = 1, 2, or 3) of Hi, the augmenting ost (Hij ) (Hi) is at most 2z(k)=(k `).
The lemma then follows immediately. Re all that, during the onstru tion of Hij , we
use the augmenting edge osts. Let x : E ! R + be an optimal solution to the linear
program P (k); note that the ost of x is z(k). Let x0 : E ! R + be given by
8
>
<1 if e is in Hi
x0e = >
: xe
k `
otherwise
Clearly, x0 has modied ost at most z(k)=(k `). We laim that x0 is a feasible
solution to the linear program P (` + 1). Then, by Theorem 2.2.1, the Frank-Tardos
22
algorithm nds an (` + 1)-out onne ted supergraph of Hi with augmenting ost at
most 2z(k)=(k `). To see the laim, onsider any setpair W 2 S and its onstraint
in P (` + 1),
X 0
xe (` + 1) jW 0j
2( )
e Æ W
Suppose that Hi has p edges in Æ(W ). Then

X X X
x0e = x0e + x0e
2( ) 2 ( )\Hi 2 ( )nHi
X
e Æ W e Æ W e Æ W
= p + k1 ` xe
e2Æ (W )nH i
Thus if p ` + 1 jW 0j we are done. Moreover, sin e Hi is `-vertex onne ted,

we have that p ` jW 0j. Hen e, we may assume that p = ` jW 0j. Given
that x is an optimal solution to P (k), it follows that Pe2Æ(W ) xe k jW 0j. So,
P jW 0j p = k `. Thus Pe2Æ(W ) x0e p + 1 = (` + 1) jW 0j.
e2Æ (W )nH xe k
i
Theorem 2.2.5. The algorithm produ es a k-vertex onne ted subgraph with ost at
most 6 H (k) z (k) 6 H (k) opt.
Proof. The vertex onne tivity of Hi in reases by at least one in every iteration,
starting from one, so the algorithm terminates with a k-vertex onne ted subgraph
in at most k 1 iterations. Note that a minimum spanning tree H1 in G indu es
a dire ted minimum 1-out onne ted subgraph in Gb for any root vertex r. Thus the
ost of H1 is at most 2z(1). So the ost of the k-vertex onne ted subgraph obtained
is at most
X1
k
2 z (k ) Xk 1
6z ( k )
(H1 ) + ( (Hi+1 ) (Hi )) +
i=1
k `=1
k `
6 (1 + 21 + 13 + + k1 ) z(k)
6 (1 + 21 + 13 + + k1 ) opt
23
Observe that the algorithm does run in polynomial time. Ea h iteration runs the
Frank-Tardos algorithm at most three times, and tests (Hi n S ) for at most n3 sets
S of three verti es. Gabow's algorithm [16℄ tests the vertex onne tivity in time
O((n + 2:5 ) n), and there is a version of the Frank-Tardos algorithm, due to Gabow
[15℄, that runs in time O(k2mn2 ).
Lower Bounds on the Integrality Gap
So the integrality gap of the setpair LP for the k-vertex onne tivity problem is at
most O(log k). It is easy to see that the integrality gap is at least 2. Take the 1-vertex
onne tivity problem on a y le; the optimal LP solution has value n2 , whereas the
optimal solution ontains n 1 edges. An important open problem is to determine
whether the integrality gap in reases with k.
24
2.3 Iterative Rounding
In this se tion we show how the iterative rounding method, due to Jain [25℄, may be
applied to the setpair model. As a result, we develop approximation algorithms for
network problems with various vertex onne tivity requirements. In parti ular, we
give approximation algorithms for the (dire ted and undire ted) k-vertex onne tivity
problem and the element onne tivity problem.
The iterative rounding method works as follows. Formulate the problem as an
integer program, and then solve the LP relaxation to nd a basi (extreme point)
optimum solution x. Pi k an edge e of highest weight (i.e., xe xe; 8e 2 E ) and
add it to the solution subgraph H . Then update the LP and the integer program
(sin e the variable xe is now impli itly xed at value 1). Spe i ally, de rease by
1 the right hand side of every onstraint in whi h the variable xe o urs, and then
remove this variable from the LP. The resulting LP is the same as the original LP for
the \redu ed" problem where the edge e is pre-sele ted for H . Under appropriate
onditions on the requirement fun tion f , the problem turns out to be \self redu ible",
that is, the essential properties of the original problem are preserved in the redu ed
problem. Then iteratively solve the redu ed problem.
Jain [25℄ applied this method to the ut overing model, and proved that it a hieves
an approximation guarantee of 2 provided that the requirement fun tion f is weakly
supermodular. (Su h requirement fun tions apture several interesting problems, e.g.,
the Steiner network problem.) The proof follows from a key property of the LP: every
non-zero extreme point has an edge of weight at least 12 . This result is based on an
extension of a lassi result that, under appropriate onditions on the requirement
fun tion f , every extreme point of the LP is hara terised by a laminar family of
\tight sets".
In Se tion 2.3.1, we introdu e a spe ial lass of requirement fun tions f , namely,
skew bisupermodular fun tions. These fun tions are a ommon generalisation of
weakly supermodular (symmetri ) fun tions ([11℄, [20℄) and rossing bisupermodu-
lar fun tions [12℄. Skew bisupermodular requirement fun tions are useful be ause
25
the extreme points of the LP relaxation of the setpair model are hara terised by
\non-overlapping families" of setpairs (see Se tion 2.3.2). Note the orresponden e
with the ut overing model, where the extreme points of the LP relaxation may be
hara terised by a laminar family of sets. Based on this, we obtain lower bounds on
the maximum weight of an edge for the basi solutions of the LP relaxation. Also,
this LP relaxation has the self-redu ibility property needed for the iterative rounding
method. Hen e, these stru tural results immediately give approximation algorithms.
2.3.1 Setpairs and Skew Bisupermodularity
Crossing Setpairs.
Two setpairs W and Y are said to be omparable if

(i) Wt Yt and Wh Yh, or (ii) Wt Yt and Wh Yh.
The former ase is denoted by W Y and the latter ase by W Y . An example of
the former ase is given in Figure 2-4. For larity, the example has Wt and Yh being
disjoint; in general, Wt and Yh may interse t.
Yh
Wt
Wh
Yt
Figure 2-4: Comparable setpairs.
Two setpairs W and Y are said to be non- rossing if they are

(i) omparable, or (ii) tail-disjoint, or (iii) head-disjoint.
Otherwise, W and Y are said to ross. Note that W and Y ross if and only if their
tails interse t, their heads interse t and they are not omparable. An example of
rossing setpairs is given in Figure 2-5. In general, Wt and Yh may interse t and/or
Wh and Yt may interse t. Again, for larity, this is not the ase in the example shown.
26
W t W h
Yh
Yt
Figure 2-5: Crossing setpairs.
For a setpair W = (Wt ; Wh), let W = (Wh; Wt) be the reverse setpair. Note that
Æ (W ) = Æ (W ) for an undire ted graph. Two setpairs W and Y are overlapping if
(i) W and Y ross, or (ii) W and Y ross (equivalently, W and Y ross).
Otherwise, the setpairs are alled non-overlapping. A family of setpairs L S is
alled non- rossing (non-overlapping) if no two setpairs in L ross (overlap).
Bisubmodular Fun tions.
Given two setpairs W = (Wt ; Wh) and Y = (Yt; Yh), let W

Y denote the setpair
(Wt [ Yt; Wh \ Yh); and let W Y denote the setpair (Wt \ Yt ; Wh [ Yh): A real-valued
fun tion f on S is alled bisubmodular if for any two setpairs W and Y we have
f (W ) + f (Y ) f (W
Y ) + f (W Y )
A non-negative, integer-valued fun tion f on S is alled rossing bisupermodular if

for any two rossing setpairs W and Y with f (W ) > 0 and f (Y ) > 0, we have
f (W ) + f (Y ) f (W
Y ) + f (W Y )
Let W denote the edge in iden e ve tor of Æ(W ). For any two setpairs W and Y ,
note that if an edge is present in Æ(W Y ) or Æ(W
Y ), then it is present in Æ(W )
or Æ(Y ), and if an edge is present in both Æ(W Y ) and Æ(W
Y ), then it is present
in both Æ(W ) and Æ(Y ). Hen e, we have W
Y + W Y W + Y . Consequently,
27
for any non-negative ve tor x : E ! R + on the edges, the orresponding fun tion on
setpairs, x(Æ(W )) = Pe2Æ(W ) xe, is bisubmodular.
Skew Bisupermodular Fun tions.
A non-negative, integer-valued fun tion f on S is alled skew bisupermodular if for

any two overlapping setpairs W and Y with f (W ) > 0 and f (Y ) > 0, one of the
following holds
(i) f (W ) + f (Y ) f (W
Y ) + f (W Y )
(ii) f (W ) + f (Y ) f (W
Y ) + f (W Y )
This is motivated by problems on undire ted graphs. For these problems, it turns
out that the relevant fun tions f are symmetri , i.e., f (W ) = f (W ); 8W 2 S .
Note that every rossing bisupermodular fun tion is skew bisupermodular. Also,
every weakly supermodular symmetri fun tion is skew bisupermodular (see the dis-
ussion at the start of Se tion 2.3.4). Thus the skew bisupermodular property is a
generalisation of these earlier notions.
2.3.2 Chara terising a Basi Solution via a Non-Overlapping
Family
The main result of this se tion, Theorem 2.3.3, hara terises a basi solution of the
setpair LP via a non-overlapping family of setpairs, assuming that the requirement
fun tion f is skew bisupermodular. This result is an extension of lassi results to
our setting. Our proof is similar to earlier proofs in [25℄ and [6℄; however, when we
add a new setpair W to our non-overlapping family, we have to ensure that neither
W nor W rosses a setpair in the family.
Given a feasible solution x to the LP, a setpair W is alled tight if x(Æ(W )) = f (W ).
Now take a skew bisupermodular requirement fun tion f , and let x be a feasible
solution su h that xe > 0 for all edges e 2 E . Then we obtain the following lemma.
28
Lemma 2.3.1. Let W and Y be overlapping, tight setpairs (with f (W ) > 0; f (Y ) >
0). Then one of the following holds:
(i) The setpairs W
Y and W Y are tight, and W + Y = W
Y + W Y .
(ii) The setpairs W
Y and W Y are tight, and W + Y = W
Y + W Y .
Proof. Assume that f (W )+ f (Y ) f (W

Y )+ f (W Y ). The other ase is similar
(note that W is tight, overlaps Y , and f (W ) > 0). Then
f (W
Y ) + f (W Y ) x(Æ (W
Y )) + x(Æ (W Y ))
x(Æ (W )) + x(Æ (Y ))
= f (W ) + f (Y )
f (W
Y ) + f (W Y )
Hen e, all the inequalities hold as equations, therefore, W

Y and W Y are both
tight. This proves the rst part of (i). The se ond part of (i) follows from the following
fa ts: W + Y W
Y + W Y ; x(Æ(W
Y ))+ x(Æ(W Y )) = x(Æ(W ))+ x(Æ(Y ));
xe > 0 for all edges e.
The next result is in [6℄ and is used to prove the main result of this se tion.
Lemma 2.3.2. Let W , Y and Z be setpairs. If Z rosses W
Y (or W Y ) then
either Z rosses W or Z rosses Y .
Re all we may also assume that 1 > xe > 0, otherwise we either x the edge or
remove it and solve the indu ed problem.
Theorem 2.3.3. There exists a non-overlapping family L of tight, linearly indepen-
dent setpairs that span the support of x.
Proof. We will show that there exists a family L of tight, non-overlapping setpairs
su h that:
(a) Every setpair W 2 L has f (W ) 1.
(b) jLj = jE j.
( ) The ve tors W , W 2 L, are linearly independent.
29
(d) x is the unique solution to fx(Æ(W )) = f (W ); 8W 2 Lg.
Sin e x is a basi solution and none of the onstraints 0 xe 1 (e 2 E ) holds
with equality, there exist jE j tight setpair onstraints x(Æ(W )) = f (W ) > 0 su h
that the edge in iden e ve tors W orresponding to these onstraints are linearly
independent. Let L be a maximal, non-overlapping family of tight setpairs and let
span(L) denote the ve tor spa e spanned by the ve tors W , where W 2 L. The
theorem holds be ause span(L) equals span(F ), where F is the family of all tight
setpairs. We prove this by ontradi tion. Suppose that there is a tight setpair Y su h
that Y 2= span(L). Take su h a Y that overlaps the minimum number of setpairs in
L (this is a key step). Choose any setpair W 2 L that overlaps Y ; there exists su h
a W by the maximality of L.
Either part (i) or part (ii) in the statement of Lemma 2.3.1 holds for W and Y .
Assume that part (ii) holds. The other ase is similar. Thus, the setpairs W
Y
and W Y are tight. In addition, W + Y = W
Y + W Y . It follows that either
W
Y 2= span(L) or W Y 2= span(L). Consider the former ase. Again, the other
ase is similar. Let Q = W
Y and take any setpair Z 2 L. Suppose that Z overlaps
Q. Then either Z rosses Q or Z rosses Q = W
Y = W Y . Suppose that Z
rosses Q. We apply Lemma 2.3.2 to Z and Q, noting that Z does not ross W , to
see that Z rosses Y . Similarly, if Z rosses Q, then by Lemma 2.3.2, Z rosses Y .
Hen e, Z overlaps Y . So every setpair in L that overlaps Q also overlaps Y .
It follows that Q overlaps fewer setpairs in L than Y . To see this, note that (1) W
overlaps Y , by our hoi e of W , and (2) W does not overlap Q, be ause Q W (so
W does not ross Q) and W and Q are head disjoint (so W does not ross Q). Thus
we get a ontradi tion to the hoi e of Y , sin e Q is a tight setpair that overlaps fewer
setpairs in L than Y , and Q 2= span(L).
Remarks. The results in this se tion apply to undire ted graphs but not to dire ted
graphs, sin e we use the fa t that a setpair W with f (W ) > 0 is tight i the reverse
setpair W has f (W ) > 0 and is tight. In the theorem, if we repla e a setpair W 2 L
by its reverse W , then this preserves the properties (a)-(d).
30
2.3.3 Iterative Rounding and k-Vertex Conne tivity
The main result of this se tion is as follows.

Theorem 2.3.4. Let k (1 )n, for some < 1. Then iterative rounding gives an
p
O( n=)-approximation algorithm for the k-vertex onne tivity problem.
Now re all that the k-vertex onne tivity problem an be formulated by the setpair
model using the following requirement fun tion.
8
>
< max f0; k jW 0jg if Wt 6= ; and Wh 6= ;
f (W ) =
>
: 0 otherwise
The k-vertex onne tivity requirement fun tion f is rossing bisupermodular [12℄,
hen e, f is also skew bisupermodular. This an be seen as follows. For ea h setpair
W , let p(W ) denote jW 0 j. For any two setpairs W = (Wt ; Wh ) and Y = (Yt ; Yh), we
have p(W ) + p(Y ) = p(W
Y ) + p(W Y ), be ause
j(Wt [ Yt) [ (Wh \ Yh)j + j(Wt \ Yt) [ (Wh [ Yh)j = jWtj + jWhj + jYtj + jYhj:
Hen e, the fun tion f is rossing bisupermodular (if either W

Y or W Y has
an empty head or tail then note that W and Y do not ross). The results in [12℄
imply that a basi optimal solution to setpair linear programming relaxation for
k-vertex onne tivity an be found in polynomial time. Moreover, the requirement
fun tion f has the following \self-redu ibility" property: for any edge e, the \redu ed"
requirement fun tion f 0 is also rossing bisupermodular (and skew bisupermodular),
where f 0 is given by f 0(W ) = f (W ) 1 if the setpair W has e 2 Æ(W ) and f 0(W ) =
f (W ) otherwise (sin e the zero-one in iden e fun tion of e on S is bisubmodular).
Theorem 2.3.4 then follows dire tly from the next result.
Theorem 2.3.5. Let k (1 )n, for some < 1. Then any nonzero basi solution
of the setpair LP for k-vertex onne tivity has an edge of weight
(
p ).

n
We give the proof for undire ted graphs, using results on skew bisupermodular
31
fun tions. After that, we sket h the proof for dire ted graphs via rossing bisuper-
modular fun tions. Consider the setpair model LP with a rossing bisupermodular
requirement fun tion f . Then we have the following result from [6℄ (whi h applies
also to dire ted graphs).
Theorem 2.3.6. Let L be a non- rossing family of setpairs hara terising a nonzero
basi solution x. Then there exists an edge e with xe
(p1 jLj) .
Fo us on a nonzero basi solution x. By Theorem 2.3.3, x orresponds to a

non rossing family of setpairs L su h that ea h setpair has a nonempty head and a
nonempty tail. We may assume that ea h setpair W 2 L has jWhj jWt j, else we
may repla e W by W and this preserves all the properties of L in Theorem 2.3.3.
Let P denote the poset (and the Hasse diagram) representing L, and re all that P
has a node W for ea h setpair W in L and it has an ar (W; Y ), where Y 2 L, if
W Y and there is no other Z 2 L su h that W Z Y (we omit ar s implied by
transitivity from P ); we all W a hild of Y , and all Y a parent of W . (Note that
the terms \node" and \ar " refer to the poset P , and the terms \vertex" and \edge"
refer to G.) A diamond of a poset is a set of four nodes W; X; Y; Z su h that X and
Y are in omparable, W is a ommon des endant of X and Y , and Z is a ommon
an estor of X and Y . A poset is alled diamond-free if it ontains no diamond. It is
easy to show that poset P representing L is diamond-free [6℄.
Lemma 2.3.7. Suppose that k (1 )n. Then there are at most 2 pairwise in om-
parable setpairs in L su h that all their tails ontain a ommon vertex.
Proof. Let the setpairs Y 1 ; Y 2; : : : ; Y q 2 L be pairwise in omparable, and let v be a
vertex that o urs in the tail of ea h of these setpairs. Sin e the Y i are in omparable
and their tails interse t, their heads are disjoint. Ea h head Yhi has ardinality at
least 12 (n k) 21 n; this follows from the properties of L, and the fa t that ea h
requirement f (Y i) is at least one. Hen e, there are at most 2 su h setpairs.
We partition the setpairs in L into several sets.
(i) W is a type I setpair if it has no parents in P .
32
(ii) W is a type II setpair if it has at least two parents in P .
(iii) W is a type III setpair if it has exa tly one parent Y in P , and Wt = Yt.
(iv) W is a type IV setpair if it has exa tly one parent Y in P , and Wt 6= Yt .
Lemma 2.3.8. The family of setpairs L ontains at most 2n setpairs of type IV. In
addition, ea h node (setpair) W 2 P has at most 2n hildren.
Proof. To ea h type IV setpair W with parent Y we assign a vertex v in Wt n Yt. We
laim that a vertex an be assigned to at most 2 setpairs of type IV. This will imply
the laim. Take a vertex v that is assigned to the type IV setpairs W 1; W 2; : : : ; W q .
Let their parents be Y 1; Y 2; : : : ; Y q , respe tively. Now v 2 Wt1 \ Wt2 \ \ Wtq ,
but v 62 Yt1 [ Yt2 [ [ Ytq . The setpairs W 1; W 2; : : : ; W q asso iated with v must
be pairwise in omparable (sin e v is in the tail of every W i, but for every proper
an estor of a W i note that v is not in the tail). Then by Lemma 2.3.7 applied to
W 1 ; W 2 ; : : : ; W q , there are at most 2 su h setpairs whose tails ontain v .
For the se ond part of the lemma, onsider all the hildren Y 1 ; Y 2; : : : ; Y q of a
node W . Then Y 1 ; Y 2; : : : ; Y q are pairwise in omparable (otherwise, one of the ar s
(Y i; W ) in P is redundant), and moreover, ea h of the tails Yti, i = 1; : : : ; q, ontains
Wt whi h is nonempty. Then, applying Lemma 2.3.7 to Y 1 ; Y 2 ; : : : ; Y q , we have
q 2 .
Proof of Theorem 2.3.5.

qLet m denote jLj. If m < 4n=, then by Theorem 2.3.6 there
is an edge e with xe (n) . Hen e, in this ase, Theorem 2.3.5 holds. The rest of
the proof fo uses on the other ase, m 4n=. We show that there exists an edge e
with either xe (1) 1 or xe
()
(p ) (note that k n 1, hen e 1 , and
p n
so n ). We have jLj = m , hen e, by by Lemma 2.3.8, L has at least m6

4 n
n
setpairs of one of the types I, II, or III.

(Type I) Suppose that there are at least m6 type I setpairs. The type I setpairs are
pairwise in omparable. Hen e the edge sets Æ(W ), where W is a type I setpair, are
quasi-disjoint, i.e., for ea h edge (u; v ), there are at most two setpairs W; Y (among the
type I setpairs) su h that (u; v) 2 Æ(W ) (say u 2 Wt and v 2 Wh) and (u; v) 2 Æ(Y )
33
(say v 2 Yt and u 2 Yh). Hen e, there is a type I setpair W 2 L su h that Æ(W ) has
size at most 12, and so there is an edge e 2 Æ(W ) with xe 121 .
(Type II) Suppose that there are at least m6 type II setpairs. Let Y 1 ; : : : ; Y q be
the minimal nodes in P (those with no hildren); learly, these nodes are pairwise
in omparable. Ea h Y i and its an estors forms a (dire ted) tree (sin e P is diamond-
free). Ea h node of P is in at most 2 of these trees, by Lemma 2.3.7. Fo us on one
of the trees and note that the number of leaf nodes (maximal nodes of P ) of the tree
is greater than or equal to the number of type II nodes in the tree. Summing over all
the trees, we see that the total number of leaf nodes (maximal nodes of P ) is at least
2 times the number of type II nodes in P , hen e, P has at least 12 maximal nodes;
m
learly, these nodes are pairwise in omparable. Then the edge sets Æ(W ), where W
is a maximal node of P , are quasi-disjoint, so there is an edge e with xe 24 .
(Type III) Finally, suppose that there are at least m6 type III setpairs. Consider all of
the type III setpairs W 1; W 2; : : : W q and their unique parents Y 1 ; Y 2; : : : Y q . Observe
that f (W i)+1 f (Y i), sin e jWt [ Wh j < jYt [ Yhj. Moreover, if W i and W j (where
1 i < j q) have distin t parents Y i and Y j , then the edge sets Æ(Y i ) n Æ(W i) and
Æ (Y j ) n Æ (W j ) are quasi-disjoint. To see this, onsider an edge (u; v ) 2 Æ (Y i ) n Æ (W i )
with u 2 Yti = Wti and v 2 Yhi n Whi . Now, every proper an estor Z of W i has
v 2 Zh , so if W j is an an estor of W i , then we annot have (u; v ) 2 Æ (Y j ) n Æ (W j ),
u 2 Ytj = Wtj , and v 2 Yhj n Whj . Consequently, by Lemma 2.3.8, any edge o urs in
at most 4 of the edge sets Æ(Y i) n Æ(W i), and moreover, x(Æ(Y i) n Æ(W i)) 1, for
ea h i = 1; : : : ; q. Hen e, there is an edge e with xe 24 .
This on ludes the proof of Theorem 2.3.5 for undire ted graphs: any nonzero
basi solution of the setpair LP for the k- onne tivity problem has an edge of weight
p
( =n).
The proof extends to dire ted graphs via Theorem 2.3.6. We partition the non-
rossing family of setpairs L into two subfamilies L1 and L2 , depending on whether
or not a setpair W 2 L has jWhj jWt j. Suppose that jL1j jL2j; otherwise, we
use a symmetri argument. Then we apply the arguments for undire ted graphs to
L1 (partitioning it into types I, II, III, IV, et .). We obtain similar lower bounds on
34
maxfxeg, the main dieren e being that we lose a fa tor of two in the lower bound
(sin e we have jL1j m2 rather than jLj = m). Hen e, Theorem 2.3.5 holds for
dire ted graphs.
Theorem 2.3.4 follows from Theorem 2.3.5: for k (1 )n, iterative rounding
p
gives an O( n=) approximation algorithm for the (dire ted or undire ted) k-vertex
onne tivity problem.
Remark. There is an example of the undire ted problem that has an extreme point
x for whi h ea h edge e has xe

(p1 k) . The example is obtained from the example
in Figure 1 of the Erratum of [37℄ with appropriate modi ations. Hen e, in Theo-
rem 2.3.5, the lower bound on the maximum weight of an edge annot be improved
beyond
(p1 k) . Note, however, this does not imply a similar lower bound on the in-
tegrality gap of the linear program. As we have seen, the intergrality gap is at most
O(log k).
2.3.4 Iterative Rounding and Element Conne tivity
In the element onne tivity problem [26℄, there is a set of terminal verti es T V and
between ea h pair of distin t terminals i and j there is a onne tivity requirement
rij (ea h rij is a non-negative integer). Terminals are reliable and do not fail. Edges
and non-terminal verti es are unreliable and are subje t to failure. The edges and
non-terminals are alled elements. The problem is to nd a minimum- ost subgraph
H that ontains rij element disjoint paths between ea h pair of terminals i and j .
Alternatively, for every pair of terminals i and j , H should ontain a path between
i and j despite the failure of up to rij 1 elements. The problem is NP-hard. Re-
ently, Fleis her, Jain, and Williamson [14℄ gave a 2-approximation algorithm for this
problem, improving on the previous best approximation guarantee of 2 log(maxfrij g)
due to Jain et al [26℄. The work in this se tion is independent of [14℄ and gives a dif-
ferent proof of the same fa tor 2 approximation guarantee. The element onne tivity
problem is a generalisation of the Steiner network problem, see Goemans et al [20℄
and Jain [25℄. Even for rather spe ial ases of this problem, the best approximation
guarantee known is 2, and moreover, the LP relaxation has an integrality ratio of at
35
least 2 n2 .
Now, for any W 2 S , let g(W ) = i2Wmax
; j 2W
t
rij . Dene a requirement fun tion f
h
on the setpairs W 2 S as follows:
8
>
< max f 0; g(W ) jW 0j g if T Wt [ Wh
f (W ) =
>
: 0 otherwise
Note that if a setpair W has a positive requirement, then all the terminals are in Wt [
Wh . It is easily seen that the setpair model with the requirement fun tion f models the
element onne tivity problem (for more details see [26℄). For the element onne tivity
problem, we obtain the following lemma on erning the setpair requirement fun tion.
Lemma 2.3.9. The element onne tivity requirement fun tion f is skew bisupermod-
ular.
Proof. We need to show that the fun tion f = g p is skew bisupermodular. We

have already seen that p(W ) + p(Y ) = p(W
Y ) + p(W Y ). Thus, it suÆ es to
show that g is skew bisupermodular. Sin e f (W ) and f (Y ) are positive it follows that
T (W1 \ Y1 ) [ (W1 \ Y2 ) [ (W2 \ Y1 ) [ (W2 \ Y2 ). Let 1 be the value of the maximum
requirement between a node in (W1 \ Y1) and a node in (W2 \ Y2). Similarly let 1
be the value of the maximum requirement between a node in (W1 \ Y1) and a node in
(W2 \ Y1). We dene 2 ; 2; 3 and 4 in a similar fashion, as shown in Figure 2-6a).
Here the edges are also labelled by their values. It follows that
g (W ) + g (Y )
= max(1; 2; 1 ; 3) + max(1; 2; 2; 4 )
g (W
Y ) + g (W Y ) = max(1 ; 3 ; 4 ) + max(1 ; 1 ; 2 )

Y ) + g(W Y ) = max(2; 2; 3 ) + max(2 ; 1; 4)
g (W
Observe that ea h of the terms g(W ), g(Y ), g(W

Y ) et . may thus be represented
by a single edge. Note that the edge representing g(W
Y ) is one of the three edges
in ident to (W2 \ Y2), whilst the edge representing g(W
Y ) is one of the three edges
in ident to (W1 \ Y1 ). These subsets are shown as triangles in Figure 2-6b). Similarly
36
a) W1 W2 b) α1
α2 β α4
α1 1
Y1 β 1
β 2
α2 α4
α3
Y2 β 2
α3
Figure 2-6: An illustration for skew bisupermodularity.
the edge representing g(W

Y ) is one of the three edges in ident to (W1 \ Y2 ), whilst
the edge representing g(W Y ) is one of the three edges in ident to (W2 \ Y1). These
subsets are shown as squares in Figure 2-6b).
Thus g(W ) + g(Y ) an be represented by a pair of (possibly identi al) edges. We
need to show that this pairing is also a feasible pairing for either g(W
Y )+ g(W Y )
or g(W
Y ) + g(W Y ). It suÆ es to show that either the two triangles are
in ident to one edge ea h in the optimal edge pairing for g(W ) + g(Y ) or the two
squares are. In the former ase g(W ) + g(Y ) g(W
Y ) + g(W Y ), whilst
in the latter ase g(W ) + g(Y ) g(W
Y ) + g(W Y ). Note that ea h edge
i ; 1 i 4; is in ident to both a square and a triangle. Thus we are learly done
if the optimal pairing is (i ; j ), where i 6= j . The pairing (i; i) is not a feasible
pairing for g(W ) + g(Y ). So suppose the optimal pairing ontains one of the i . By
symmetry we may assume it ontains 1 . Thus the optimal pairing is either (1 ; 1)
or (1 ; i); 1 i 4. In these ases both triangles are in ident to dierent edges in
the pairing and thus g(W ) + g(Y ) g(W
Y ) + g(W Y ). Hen e g, and thus f , is
skew bisupermodular.
A basi optimal solution to LP relaxation for the element onne tivity problem
an be found in polynomial time. This follows be ause a strong separation ora le
may be obtained using standard network ow te hniques (for more details, see [26℄).
37
To prove the key result for the 2-approximation guarantee for element onne tivity,
namely, there exists an edge of value at least 12 in any non-zero basi solution of the
setpair LP relaxation. we need to develop some preliminaries and a key lemma. Based
on this, the rest of the proof is an extension of the ounting arguments used in [25℄.
Let x be a non-zero basi solution to the LP. We may assume that ea h edge e 2 E
has xe > 0, sin e any other edges may be dis arded. In addition, we may assume
that ea h edge e 2 E has xe < 1, otherwise we are done. Let L be the orresponding
non-overlapping family of setpairs of Theorem 2.3.3. Sin e E is the support of a basi
solution, jE j = jLj. Let m denote jE j. Let be an arbitrarily hosen terminal vertex
in T . For ea h setpair in L, we assume that the head ontains . Otherwise, if is in
the tail of W 2 L, then we repla e W by W and this preserves all the properties of
L in Theorem 2.3.3. Then the tails of the setpairs in L form a laminar family. This
holds be ause the setpairs in L are pairwise non-overlapping, and the heads of any
two setpairs in L interse t (sin e is in every head). We may also view L as a forest
of rooted trees. Call W 2 L the parent of Y 2 L (and Y a hild of W ) if Wt Yt
and any other setpair Z 2 L with Zt Yt either has Zt Wt or has Zt = Wt and
Zh Wh . We may refer to the setpairs in L as nodes of the forest. We say that a
setpair Y 2 L is smaller than a setpair W 2 L if Y W , i.e., Yt Wt and Y is a
des endant of W in the forest.
In the rest of this se tion, we take the edges in E to be bidire ted. That is, we
repla e every undire ted edge (p; q) in E by a pair of dire ted ar s (p; q) and (q; p).
For a setpair W = (Wt; Wh), let Æ(W ) denote the set of ar s (p; q) with p 2 Wt and
q 2 Wh . (Though we use the same notation Æ () for undire ted edges and dire ted
ar s, the ontext will resolve any ambiguity.) The other notation, x, W , et . remains
the same. An ar (p; q) is alled good if there is a setpair W 2 L su h that (p; q) is in
Æ (W ). Note that, in the undire ted setting, every edge e is in Æ (W ) for some W 2 L,
but in the dire ted setting, there may exist ar s that are not good.
The proof of the main result hinges on assigning token ar s to the setpairs in L.
Here is a brief sket h; the details are given in the proof of Theorem 2.3.12. We have
jLj = jE j = m and thus a total of 2m ar s. If a basi solution x of has xe < 12 for
38
ea h edge e, then we will be able to distribute the token ar s among the setpairs in
L su h that ea h gets at least two token ar s and some setpair gets at least three
token ar s. Thus, we end up with at least 2jLj + 1 = 2m + 1 ar s. This ontradi tion
shows that our assumption on maxe fxeg is false. We use the following two rules to
distribute the token ar s su h that no ar is assigned to two dierent setpairs.
rule 1. If (p; q ) is a good ar , then we assign it to the smallest setpair W 2 L su h
that (p; q) 2 Æ(W ).

rule 2. If (p; q ) is not a good ar , then we assign it to the smallest setpair W 2 L
su h that q 2 Wt and p 62 Wh (note that neither (p; q) nor (q; p) is in Æ(W )).
We note that this se ond rule is essential in the sense that the analysis fails if we
insist on assigning ar s (p; q) only to those setpairs W su h that (p; q) 2 Æ(W ).
Given two setpairs W and Y , let Æ(W ) O Æ(Y ) denote the symmetri dieren e
of Æ(W ) and Æ(Y ). Given S 1; S 2; : : : ; S `, a olle tion of mutually disjoint vertex sets,
we denote by (S 1; S 2; : : : ; S `) the set of ar s (p; q) su h that p and q are in dierent
sets S i and S j , where 1 i; j `.
Lemma 2.3.10. Let W be a setpair in the forest L, and let its hildren be the setpairs
Y 1 ; Y 2 ; : : : ; Y ` . Then for every ar (p; q ) in
(Æ(W ) O (Æ(Y 1) [ [ Æ(Y `))) n (Yt1; : : : ; Yt`)
either (p; q ) or (q; p) is assigned to W by Rule 1 or Rule 2.
Proof. Note that the lemma applies to all ar s in the symmetri dieren e of Æ(W )
and S i=1 Æ (Y ) ex ept for ar s whose two end-verti es are in the tails of dierent
` i
hildren. Figure 2-7 illustrates the arguments that follow, for the parti ular ase of
` = 2. [Note, for example, that the ar a4 is in (Yt1 ; Yt2 ).℄
We have two ases to deal with.
(a) (p; q) 2 Æ(W ) n S`i=1 Æ(Y i).
Now q 2 Wh Yhi, for i = 1; : : : ; `. However, by assumption, p 2 Wt n [`i=1Yti and
so W is the smallest setpair in L su h that p is in the tail. Thus, W is the smallest
setpair in L with (p; q) 2 Æ(W ), and hen e (p; q) is assigned to W by Rule 1. [The
39
Y1h Y2h
Wh
a1
Wt a3
a2
Y1t Y2t
a4
Figure 2-7: Ar assignment rules.
ar a1 in Figure 2-7 is an example.℄

(b) (p; q) 2 Æ(Y i) n Æ(W ), for some i 2 f1; : : : ; `g, and (p; q) 2= (Yt1 ; : : : ; Yt`).
Note that p 2 Yti Wt and q 62 Wh. We have two sub ases.
(i) q 2 Wt .
First, we show that (q; p) is not a good ar . Clearly, q 62 Yti as (p; q) 2 Æ(Y i). For any
other hild Y j , observe that q 62 Ytj as (p; q) 2= (Yt1; : : : ; Yt`). Hen e, q 2 Wt n[`j=1Ytj .
Then W is the smallest setpair in L whose tail ontains q, and there is no setpair
Z 2 L su h that (q; p) 2 Æ (Z ) (otherwise, p 2 Zh \ Wt and q 2 Zt Wt , whi h is
impossible). Thus (q; p) is not a good ar . Moreover, p 62 Wh (sin e p 2 Wt), and so
Rule 2 assigns (q; p) to W . [The ar a2 in Figure 2-7 is an example.℄
(ii) q 62 Wt .
We rst show that (q; p) is not a good ar . Suppose Z 2 L is a setpair su h that
(q; p) 2 Æ(Z ). We derive a ontradi tion by showing that W and Z overlap; in
parti ular, W and Z ross. Note that Wt and Zt are disjoint, be ause p 2 Wt n Zt and
q 2 Zt n Wt . Now Zt and Wh have a terminal in ommon, be ause Zt has at least one
terminal and Wt [ Wh ontains all the terminals. Then Zt and Wh interse t properly,
be ause
40
(1) There is a terminal in Zt \ Wh
(2) The terminal is in Wh n Zt
(3) We have q 2 Zt n Wh
Finally, note that Zh and Wt interse t sin e both ontain the vertex p. Hen e, Z
and W ross. Therefore, (q; p) is not a good ar and Rule 2 applies to (q; p). Rule
2 assigns (q; p) to W sin e p 2 Wt , q 2= Wh, and for any smaller setpair X 2 L with
p 2 Xt we have q 2 Xh (sin e p 2 Yti and q 2 Yhi ). [The ar a3 in Figure 2-7 is an
example.℄
We need the following lemma in order to apply the parity arguments required in
the proof of Theorem 2.3.12.
Lemma 2.3.11. If Y and Z are tail disjoint setpairs in L then Yt Zh and Zt Yh .
Proof. Let Y = (Yt ; Yh) and Z = (Zt ; Zh) be setpairs in L. To see the lemma, note
that Yt and Zh have a terminal in ommon, sin e Yt has at least one terminal, and
Zt [ Zh ontains all the terminals. Similarly, Zt and Yh have a terminal in ommon.
Suppose that Yt is not a subset of Zh. Then it an be seen that Yt and Zh interse t
properly. Hen e, Z rosses Y , and so Y and Z overlap. This ontradi tion shows
that Yt Zh. Similarly, we have Zt Yh.
Theorem 2.3.12. Any non-zero basi solution of the setpair LP for element onne -
tivity has an edge of weight at least 12 .
Proof. Let x 6= 0 be a basi solution su h that xe < 12 for ea h edge e. Then

the following laim implies that we end up with at least 2jLj + 1 ar s. This is a
ontradi tion. Hen e, maxe fxeg 12 . Dene the orequirement of an edge e, denoted
(e), to be 21 xe > 0, and the orequirement of a setpair W , denoted (W ), to be
P (e). Note that a tight setpair W has (W ) = 1 jÆ(W )j f (W ) > 0. We
e2Æ (W ) 2
will use the fa t that if a setpair W has (W ) = 21 , then jÆ(W )j is odd.
Claim. Suppose that xe < 21 ; 8e 2 E . Then the ar s may be redistributed to the
nodes in the forest L su h that for ea h rooted subtree of L, ea h node gets at least
two token ar s and the root gets at least three token ar s. Moreover, a root with
orequirement 6= 12 gets at least four token ar s.
41
Again, we prove the laim by indu tion. For the base ase take a leaf node W 2 L.
Then Æ(W ) has at least 3 ar s. By Rule 1, all these ar s are assigned to W . If W gets
exa tly 3 token ar s, then note that f (W ) = 1 and jÆ(W )j = 3, so we have (W ) = 21 .
Thus W satises the indu tion hypothesis.
For the indu tion step, onsider a subtree rooted at a node W 2 L. We have four
ases.
(a) W has at least four hildren.
By the indu tion hypothesis, ea h hild has at least 3 token ar s. We reassign one
token ar from ea h hild to W . Hen e, W gets at least 4 token ar s.
(b) W has three hildren.
Call the hildren X , Y and Z . If one of the hildren has at least 4 token ar s,
then we an reassign the token ar s from X , Y and Z to W so that W has at least 4
token ar s and ea h of X , Y and Z is left with at least 2 token ar s.
Otherwise, by the indu tion hypothesis, ea h of X , Y and Z has 3 token ar s, so
ea h has orequirement 21 and ea h of jÆ(X )j, jÆ(Y )j and jÆ(Z )j is odd. We reassign
one token ar from ea h of X , Y and Z to W . Now fo us on the symmetri dieren e
of Æ(W ) and Æ(X ) [ Æ(Y ) [ Æ(Z ). By the linear independen e of W , X , Y and
Z , there is at least one ar (p; q ) in the symmetri dieren e. If possible, hoose
(p; q) 62 (Xt ; Yt; Zt). Then by Lemma 2.3.10 either (p; q) or (q; p) is assigned to W as
a token ar . Hen e, W ends up with at least 4 token ar s. Otherwise, every ar in the
symmetri dieren e is in (Xt ; Yt; Zt). Therefore Æ(W ) Æ(X ) [ Æ(Y ) [ Æ(Z ), and for
ea h ar (p; q) in (Æ(X ) [ Æ(Y ) [ Æ(Z )) n Æ(W ) the ar (q; p) is also in the same set, by
Lemma 2.3.11. Hen e, jÆ(W )j is odd, and so the orequirement (W ) is a semi integer
(i.e., an odd integer multiple of 21 ). Moreover, (W ) < (X ) + (Y ) + (Z ) = 23 .
Hen e, (W ) = 12 , and so W satises the indu tion hypothesis.
( ) W has two hildren.
Call the hildren Y and Z . If we an assign 4 token ar s to W by reassigning token
ar s from Y and Z , and by applying Lemma 2.3.10 to (Æ(W )O(Æ(Y )[Æ(Z )))n (Yt ; Zt),
then we are done. Otherwise, we are left with two sub ases.
(i) Ea h of Y and Z has 3 token ar s and a orequirement of 12 , and there is one ar
42
(p; q) in (Æ(W ) O (Æ(Y ) [ Æ(Z ))) n (Yt; Zt). Then W gets 3 token ar s, one ea h by
reassigning ar s from Y and Z and one by Lemma 2.3.10. Both jÆ(Y )j and jÆ(Z )j are
odd (sin e (Y ) = (Z ) = 12 ) and the ar s (if any) in (Yt; Zt) o ur in pairs (u; v)
and (v; u), by Lemma 2.3.11. Therefore, jÆ(W )j is odd and (W ) is a semi integer.
Moreover, 0 < (W ) (Y ) + (Z ) + (fp; qg) < 32 . Hen e, (W ) = 12 , and so W
satises the indu tion hypothesis.
(ii) At least one of the hildren, say Z , has 3 token ar s and a orequirement of 21 , and
there are no ar s in (Æ(W ) O (Æ(Y ) [ Æ(Z ))) n (Yt; Zt). We will prove that this ase
annot o ur. First, note that jÆ(W )j and jÆ(Y )j have opposite parity (one is odd
and the other is even), be ause jÆ(Z )j is odd and the ar s in the symmetri dieren e
of Æ(W ) and Æ(Y ) [ Æ(Z ) o ur in pairs (u; v) and (v; u). Furthermore, note that
W , Y and Z are linearly independent. It follows that there is an ar in (Yt ; Zt )
and there is an ar in Æ(Z ) \ Æ(W ). Hen e, (W ) < (Y ) + (Z ) (due to the ar
in (Yt; Zt )) and (Y ) < (W ) + (Z ) (due to the ar in Æ(Z ) \ Æ(W )). Therefore,
(Y ) 12 < (W ) < (Y )+ 12 , implying that (Y ) = (W ) (sin e (W ) is an integer
multiple of 21 ). Then jÆ(W )j and jÆ(Y )j have the same parity. This ontradi tion
shows that this ase annot o ur.
(d) W has one hild.
Call the hild Y . If the symmetri dieren e of Æ(W ) and Æ(Y ) has exa tly 2 ar s
and (Y ) = 12 , then W gets 3 token ar s (two by Lemma 2.3.10 and one reassigned
from Y ). In this ase, note that f (W ) = f (Y ) and jÆ(W )j = jÆ(Y )j. As a result,
(W ) = (Y ) = 21 . Otherwise, W gets at least 4 token ar s. In either ase, W
satises the indu tion hypothesis. This ompletes the proof of the laim.
The theorem follows.
Thus, we get a fa tor 2 approximation for the element onne tivity mat hing the
result of Fleis her et al [14℄.
Theorem 2.3.13. There is a polynomial-time algorithm for the element onne tivity
problem that nds a solution of ost at most twi e the optimal ost.
43
44
Chapter 3
Bi-Conne tivity in Unweighted
Graphs
In this hapter, we onsider bi- onne tivity problems in unweighted graphs. For
undire ted graphs we ta kle the minimum 2-edge onne ted spanning subgraph prob-
lem (2EC ) and the minimum 2-vertex onne ted spanning subgraph problem (2V C ).
The results given in this hapter are from the papers [39℄ and [40℄. Sin e these bi-
onne tivity problems are maxSNP-hard ([9℄, [3℄) our fo us will remain approxima-
tion algorithms. Khuller and Vishkin [30℄ gave a 32 -approximation algorithm for 2EC .
Cheriyan et al [5℄ have re ently improved upon this, with a 1217 -approximation algo-
rithm. In [30℄, Khuller and Vishkin also gave a 35 -approximation algorithm for 2V C .
This was improved to 23 by Garg et al [18℄. This hapter presents 34 -approximation al-
gorithms for both 2EC and 2V C . The ratio 34 has a spe ial signi an e; a elebrated
onje ture in ombinatorial optimisation says that the travelling salesman problem
on metri s is approximable to within 43 via a linear programming relaxation alled
the subtour relaxation. A (previously unveried) impli ation of this onje ture is that
2EC is also approximable to within 34 .
Bi- onne tivity in dire ted graphs orresponds to the property of strong onne tiv-
ity. A dire ted graph is strongly onne ted if, for every pair of verti es u and v , there
is a dire ted path from u to v and vi e versa. Thus, for dire ted graphs we ta kle the
minimum strongly onne ted spanning subgraph problem (SCS ). A slightly more
45
general problem is the minimum equivalent digraph problem, denoted MED. Here
we are given a dire ted graph G and we wish to nd the smallest subgraph that
maintains the rea hability relations between all pairs of verti es in G. This is a
long studied problem, with Moyles and Thompson [36℄ showing that MED de om-
poses into two parts; a solution to the strongly onne ted subgraph problem on ea h
strongly onne ted omponent of G and a solution of MED on an indu ed a y li
graph. The latter problem an be solved exa tly in polynomial time [24℄. It follows
that an -approximation algorithm for the strongly onne ted subgraph problem an
easily be onverted into an -approximation algorithm for the minimum equivalent
digraph problem. The urrent best approximation algorithm for SCS was given by
Khuller, Raghava hari and Young ([28℄,[29℄). Their algorithm has an approximation
fa tor of just over 1:617. In this hapter, we improve this fa tor to 32 .
The basi algorithmi approa h we take is the same for all three problems, that
is 2EC , 2V C and SCS . An overview of our approa h is as follows. In Se tion 3.2 we
present the basi lower bound against whi h the performan e of our algorithm will
be measured. To obtain this bound we nd the minimum sized spanning subgraph
in whi h every vertex has degree at least two (or in-degree and out-degree at least
one in dire ted graphs). This an be done in polynomial-time via a redu tion to the
maximum mat hing problem. Se tion 3.3 ontains de omposition theorems whi h are
the basis of our algorithm. These de ompositions allow us to fo us on graphs whi h do
not ontain ertain problemati stru tures. In su h graphs the lower bound solution
an be modied to obtain a solutions to 2EC , 2V C and SCS without in reasing
their size signi antly. The bulk of the analysis lies in showing that su h a low-
ost modi ation is possible. This involves detailed arguments for several tri ky
ongurations; these are given in Se tion 3.6, Se tion 3.7 and Se tion 3.8 for the edge
onne tivity, vertex onne tivity, and strong onne tivity ases, respe tively.
46
3.1 Two Simple Approximation Algorithms
In this se tion we give two simple approximation algorithms for the 2EC problem.
Both algorithms give fa tor 2 guarantees for the problem. We present these algorithms
as the basi ideas they rely on will be in orporated into our algorithms for 2EC , 2V C
and SCS .
3.1.1 The DFS Tree Method
The rst algorithm is a \depth rst sear h tree based" algorithm. First we grow a
depth rst sear h (DFS) spanning tree T rooted at any vertex r in G. A property
of DFS trees is that any non-tree edge is in ident to two verti es that lie on a path
from the root to a leaf in the tree. We say that an edge (u; w) is a ba k edge with
respe t to v if w, v, and u lie (in that order) on the same path from the root to a leaf
node l. Observe that for ea h vertex v there is at least one asso iated ba k edge. If
not then the removal of the tree edge above v in T dis onne ts the subtree rooted at
v from the rest of the graph; thus G is not 2-edge (vertex) onne ted. Our solution
is the edges of T along with a ba k edge with respe t to ea h vertex. It is easy to
see that su h a solution is 2-edge onne ted. In an n vertex graph this produ es a
solution with at most 2n 2 edges. Sin e the optimum has at least n edges, this is a
2-approximation.
3.1.2 The Cy le Contra tion Method
The se ond algorithm is a \ y le- ontra tion" algorithm. Find a y le C in G that

ontains a spe ied vertex r, alled the root vertex. We will take C to form part of
our nal solution. Next onsider G0 = G=C , the graph formed by the ontra tion of
the y le C . Let the super-vertex formed by the ontra tion of C be our root vertex
r0 in G0 . Re urse until the graph ontains a single vertex. At this point it is lear that
the y les found in ea h iteration together form a strongly onne ted subgraph in G.
We now show that this y le ontra tion algorithm has a performan e guarantee of
2. Given a ontra ted y le C , harge the edges in the y le to the verti es of the
47
y le as follows. The root vertex r is not harged. One of the verti es neighbouring r
in C , say v2 , is harged the two edges adja ent to it in the y le. The other verti es
in the y le (if C ontains more than two verti es) are harged for one adja ent edge
ea h. Noti e that a vertex re eives a harge only when it is ontra ted into the root
vertex. Sin e every edge in our solution is harged to some vertex and every vertex
is harged at most 2 edges, we see that our solution ontains at most 2n 2 edges.
Hen e this approa h does indeed give a 2-approximation algorithm.
Observe that the y le- ontra tion algorithm does not ne essarily give a 2-vertex
onne ted subgraph; this, though, will not be of on ern to us. Note, however, that
dire ted versions of both these algorithms work for the SCS problem.
3.2 The Lower Bound
3.2.1 The Lower Bound in Undire ted Graphs
The idea behind the algorithms is to take a minimum sized spanning subgraph with
minimum degree 2 and, at a small ost, make it 2-edge onne ted. We will denote
by U 2 the problem of nding, in an undire ted graph, a minimum sized spanning
subgraph in whi h ea h vertex has degree at least two. Noti e that the solution to
U 2 (whi h may not be onne ted) provides a lower bound for both 2EC and 2V C .
Lemma 3.2.1. The size of the optimal solution to U 2 is a lower bound on the size
of the optimal solutions to 2EC and 2V C .
Proof. Any 2- onne ted subgraph must have minimum degree at least two.
This gives a possible method for approximating 2EC and 2V C . Find an optimal
solution to U 2. Alter the solution in some way to get a 2- onne ted graph. The rest
of this hapter is devoted to spe ifying the \some way" and proving that the method
gives a solution whose size is at most 34 the size of the optimum solution for U 2.
Hen e it is at most 34 the size of the optimum solution for 2EC or 2V C .
The onversion of a solution of U 2 to a solution of 2EC or 2V C will only take
O(m) time and, thus, the overall running time is dominated by the time to solve U 2.
48
The latter problem an be solved exa tly in polynomial time. One fast way to do
this is to nd a maximum sized subgraph in whi h ea h vertex has degree at most
two (this is the maximum y le-path over problem whi h asks for the maximum
sized subgraph onsisting of disjoint y les and paths) and add an ar to the end
verti es of ea h path. This will give an optimal solution to U 2. This follows from the
fa t that if the minimum U 2 solution ontains n + p edges then there are 2p edges
whose removal indu es a olle tion of p disjoint paths plus some disjoint y les. The
maximum y le-path over problem has a simple redu tion to maximum mat hing so
the running time is O(mpn).
3.2.2 The Lower Bound in Dire ted Graphs
For the SCS problem we use a dire ted version of the lower bound for the 2EC
and 2V C problems. Consider the problem of nding, in a dire ted graph, the mini-
mum sized subgraph in whi h every node has both in-degree and out-degree at least
one. Call this problem D2. Again, an optimal solution to D2 an be omputed in
polynomial time via a redu tion to the undire ted maximum mat hing problem.
Lemma 3.2.2. D2 gives a lower bound on the size of the optimal solutions to SCS .
Proof. Every node must be in ident to both an in-ar and an out-ar in any strongly
onne ted subgraph.
We will take an optimal solution to D2 and alter it in order to obtain a strongly
onne ted subgraph. We will show that this transformation pro ess gives a strongly
onne ted subgraph that is at most 32 times the size of the optimum solution for SCS .
3.3 The De omposition Theorems
How good are our lower bounds though? Take the lower bound U 2 for our undire ted
problems. The following examples show that, on its own, it is not suÆ ient for our
purposes. Consider nding a minimum 2-edge onne ted subgraph in the graph shown
in Figure 3-1i). The optimal solution is 23 times larger than the optimal solution to
49
U 2. To see this, assume that the number of verti es is odd. The minimum 2-edge
onne ted subgraph is the whole graph and, thus, ontains 32 (n 1) edges. However,
the graph ontains a Hamiltonian path. Therefore, the lower bound ontains (n 1)+2
edges (we need to add an extra edge at ea h end-vertex of the path to give a feasible
U 2 solution). Similarly if we wish to nd a minimum 2-vertex onne ted subgraph
in the graph in Figure 3-1ii), it is easy to see that this is also 32 times larger than the
optimal solution to U 2.
ii)
i)
Figure 3-1: Troublesome examples.
In this se tion we will show how to strengthen our lower bounds via the use of
de ompositions.
3.3.1 2-Edge Conne tivity
Cut Verti es
Dealing with the diÆ ulty in the example given in Figure 3-1i) is not diÆ ult. We
assume that for the problem 2EC our graph ontains no ut verti es. This ondition
entails no loss of generality. An -approximation algorithm for 2-vertex onne ted
graphs an be used to give -approximate solutions for general graphs. Simply apply
the algorithm separately to ea h maximal 2-vertex onne ted blo k, then ombine
the solutions to ea h blo k.
We may also assume that G has no adja ent degree 2 verti es. This also entails
no loss of generality. Observe that both edges in ident to a degree two vertex must
be ontained in any 2-edge onne ted spanning subgraph. Hen e, for a path whose
internal verti es are all degree 2, every edge in the path must be hosen. This allows
us to transform an instan e of the general ase into an instan e with non-adja ent
50
degree 2 verti es. Given G, ontra t, to just two edges, any su h path. Call the
resulting graph G0. Given the trivial bije tion between solutions to G and solutions
to G0, it follows that any -approximate solution to G0 indu es an -approximate
solution to G.
Beta Stru tures
Beta Stru tures The example of Figure 3-1ii) is more troublesome. To ounter it we
need to eliminate a ertain graph stru ture, whi h we will all a beta-stru ture. A
vertex u is a beta-vertex if the removal of two verti es v1 and v2 indu es at least three
omponents, one of whi h onsists of just u. This situation is illustrated in Figure
3-2, with the other two omponents labelled C1 and C2. Without loss of generality,
we may assume that the omponent C1 is smaller (i.e. ontains fewer verti es) than
C2 .
v1
C1 C2
v2
Figure 3-2: A beta-vertex.
We say that two adja ent verti es u1 and u2 are a beta-pair if there exist two
nodes, v1 and v2 , whose removal leaves at least 3 omponents, one of whi h ontains
just u1 and u2.
This situation, with the other two omponents labelled C1 and C2, is shown in
Figure 3-3. Note, the gure shows all four edges (u1; v1 ); (u1; v2 ); (u2; v1) and (u2; v2)
present. It may be the ase that, for a beta-pair, only three of these edges are present.
This will not ae t our argument in any way. In addition, for both beta-verti es and
beta-verti es we may assume that the edge (v1; v2 ) is not present in the graph G. To
51
v1
C1 C2
u1 u2
v2
Figure 3-3: A beta-pair.
see this note that there is some optimal subgraph in G that does not ontain the edge
(v1 ; v2); thus it suÆ es to nd an -approximate solution in G (v1; v2 ). We say that
a graph without a beta-vertex or beta-pair is beta-free.
Here we will show how to re ursively obtain, in polynomial time, a lower bound
l(G) on the optimal solution. In the following se tions we will show how to augment
the solution indu ed by the lower bound to give a feasible 2-edge onne ted subgraph
whose size is at most 43 that of the lower bound. The lower bound is obtain by
de omposing the graph around beta-stru tures.
First we show that the ase of beta-pairs an be redu ed to that of beta-verti es.
Observe that both edges in ident to a beta-vertex u must be in any solution. These
two edges indu e a path of length two from v1 to v2 . Similarly given a beta-pair u1
and u2 it is easy to show that there is some optimal solution whi h ontains a path
of length three from v1 to v2 using the edge (u1; u2). It follows that if we nd a
beta-pair u1 and u2 in G we may ontra t the edge (u1; u2) and solve our problem on
a smaller graph G0 that ontains a beta-vertex. Note, also, that this edge ontra tion
redu es the size of the optimal solution in the graph by exa tly one; the lower bound
that we present will also be redu ed by exa tly one. Any approximation guarantee
we may obtain, via the lower bound, for the ontra ted graph G0 will then apply
to the original graph G. Thus, it suÆ es to assume that all our beta-stru tures are
beta-verti es.
Our basi te hnique to deal with a beta-vertex is as follows. We will de ompose the
52
graph into two pie es around the beta-vertex and iterate the pro edure. Eventually
we will have a de omposition into beta-free graphs. We will then work on these pie es
and show that the sub-solutions obtained an be ombined to give a good solution to
the whole graph. Given a beta-vertex, u, with omponents C1 and C2 , let G1 be the
subgraph formed by removing C2 and its in ident edges. Let G01 be the graph formed
by ontra ting V C1. These situations are shown in Figure 3-4i) and ii), where the
super-node formed by ontra ting V C1 is labelled v0. The graphs G2 and G02 are
dened analogously. Let k1; k10 ; k2 and k20 be the sizes of the optimal 2-edge onne ted
v1
i) ii)
C1 C1
u v’
v2 G’1
G1
Figure 3-4: A split.
spanning subgraphs of G1; G01; G2 and G02, respe tively. Let opt(G) be the size of the
optimal solution to the whole graph.
Lemma 3.3.1. Given a 2-edge onne ted graph, opt (G) = min(k1 + k20 ; k10 + k2).
Proof. It is lear that the ombination of a solution to G1 and G02 , or vi e versa, gives
a 2-edge onne ted subgraph for G. Conversely any solution to G an be de omposed
to give a solution to G1 and G02, or to give a solution to G01 and G2.
In addition, there is the following relationship between ki and ki0 .
Lemma 3.3.2. For i = 1; 2, we have ki0 + 2 ki ki0 + 3.
Proof. First note that to onne t up u we must use both its in ident edges. It follows
that ki0 +2 ki ki0 +3, i = 1; 2. This is be ause, given a 2-edge onne ted subgraph
for G0i, we are for ed to use 2 extra edges in onne ting up u and we may need to
53
use an extra edge from Ci in ertain instan es in whi h the solution to G0i had all its
edges leaving Ci going to the same vertex (either v1 or v2).
From Lemma 3.3.1 and Lemma 3.3.2, we obtain the following simple orollary.
Corollary 3.3.3. If G has a beta-vertex indu ing the omponents C1 and C2 then
(a) k10 + k20 + 2 opt(G) k10 + k20 + 3

(b) k1 + k2 3 opt(G) k1 + k2 2
(Note that if the omponents C1 and C2 were indu ed by a beta-pair not a beta-
vertex then we would obtain, for example, the inequalities ki0 + 3 ki ki0 + 4,
i = 1; 2. The lower bound dened below may then be onstru ted with regards to
both beta-pairs and beta-verti es. However, it is leaner to use the fore-mentioned
redu tion of beta-pairs to beta-verti es.)
To ompute optimally the values k1 ; k10 ; k2 and k20 is as hard as the original problem.
So instead we will need to nd approximate solutions. First, though, we present a
lower bound on opt(G) that is dened re ursively. If G is beta-free then let l(G)
be given by the lower bound provided by D2. If G ontains a beta-vertex then we
evaluate l(G1 ) and l(G01 ) and, based upon their relative magnitudes, we either set
l(G) = l(G1 ) + l(G02 ) or set l(G) = l(G01 ) + l(G2 ). The exa t pro edure is:
LowerBound(G)
f
If G has no beta-vertex
Then let l(G) = jD2(G)j
Else if G has a beta-vertex
Then if l(G1 ) l(G01 ) + 2 let l(G) = l(G1 ) + l(G02 )
Else if l(G1 ) l(G01 ) + 3 let l(G) = l(G01 ) + l(G2 )
g
This re ursive lower bound pro edure indu es the following de omposition into
beta-free subgraphs.
54
De ompose(G)
f
If G is beta-free
Then output G
Else if G has a beta-vertex
If l(G1 ) l(G01 ) + 2 then output De ompose(G1 ) and De ompose(G02 )
If l(G1 ) l(G01 ) + 3 then output De ompose(G01 ) and De ompose(G2 )
g
Sin e we have that jC1j jC2j, this re ursive pro edure an be applied in poly-
nomial time.
Theorem 3.3.4. The de omposition pro edure an be implemented in polynomial

time.
Proof. Let f (n) be the running time for our algorithm on a graph of n verti es. The
time required to sear h for possible de ompositions is O(mn) in G. In addition, if
p
G is beta-free then we require O(m n) to nd a solution to D2. For graphs with
beta-stru tures, the running time is then at most f (n) = f (n a) + f (a + 1) +
f (a + 3) + O(mn), where 1 a 12 n. Hen e, f (n) max(f (n a) + 2f (a + 3) +
p
O(mn); O(m n)). Solving the re urren e we get that f (n) = O(mn2 ).
We now prove that l(G) is indeed a lower bound on the size of the optimal 2-edge
onne ted subgraph.
Theorem 3.3.5. Given a 2-edge onne ted graph, l(G) opt(G).
Proof. We pro eed by indu tion. If G is beta-free then l(G) is given by the solution to
D2, a lower bound on the optimal solution. So assume that G ontains a beta-vertex.
By onstru tion of l(G), we have two possibilities:
55
(i) l(G1 ) l(G01) + 2: thus we have
l(G) = l(G1 ) + l(G02) by LowerBound(G)

l(G01 ) + 2 + l(G02)
k10 + 2 + k20 by indu tion
opt(G) by Corollary 3.3.3a)
(ii) l(G1) l(G01) + 3: thus we have
l(G) = l(G01 ) + l(G2) by LowerBound(G)

l(G1 ) 3 + l(G2 )
k1 3 + k2 by indu tion
opt(G) by Corollary 3.3.3b)
Corollary 3.3.6. Given an algorithm that nds a solution with at most jD2(G)j
edges for any beta-free graph, we get an -approximation algorithm for any graph.
Proof.Consider the de omposition tree indu ed by the de omposition pro edure.

Let H1; H2; : : : Hp be its leaves. Then opt(G) l(G) = jD2(H1)j + jD2(H2)j + +
jD2(Hp)j, whilst the number of edges used by the algorithm is at most (jD2(H1)j +
jD2(H2)j + + jD2(Hp)j).
Thus, it suÆ es to onsider only beta-free graphs. It is in these graphs that we will
nd 43 -approximate solution based upon the subgraph indu ed by the D2 solution.
3.3.2 2-Vertex Conne tivity
Beta Stru tures
Consider the ase of a beta-vertex u for the 2-vertex onne tivity problem. We deal
with this situation as follows. Given G, let G0 = G n fug. See Figure 3-5i) and ii).
Given that the two edges in ident to u are for ed in any solution to G, we have the
following simple lemma.
56
i) G ii) G’
v1 v1
C1 C2 C1 C2
v2 v2
Figure 3-5: Beta-de ompositions in vertex onne tivity.
Lemma 3.3.7. Consider the problem 2V C on a graph G that ontains a beta-vertex

u. An -approximation for G an be obtained from an -approximation for G0 by
adding the ar s in ident to u.
So for 2V C our de omposition is straightforward. Remove the beta-vertex and

onsider the indu ed subproblem.
3.3.3 Strong Conne tivity
Cut Verti es
In dire ted graphs, a vertex v is a ut vertex if there are two omponents C1 and C2
su h that for every pair of verti es, u1 2 C1 and u2 2 C2, ea h dire ted path from u1
to u2, and vi e versa, passes through v.
Lemma 3.3.8. It is suÆ ient to onsider only graphs that ontain no ut verti es.
Proof. Suppose G ontains a ut vertex v. Then every strongly onne ted subgraph
of G is the ar -disjoint union of a strongly onne ted subgraph on C1 [ fvg and a
strongly onne ted subgraph on C2 [fvg. Hen e we an de ompose the problem into
two and work on the pie es. Combining -approximations for the pie es gives an
-approximation for G.
Re all that we will present a fa tor 32 -approximation algorithm for SCS . The
graph in Figure 3-6i) shows that this approximation guarantee is the best we will be
able to obtain using the D2 lower bound in the absen e of ut verti es. The solution
57
to D2, shown in Figure 3-6ii), is a perfe t y le over and hen e ontains n ar s.
There is a 23 gap, though, between this solution and the optimal solution to SCS ,
whi h is illustrated in Figure 3-6iii).
i)
ii)
iii)
Figure 3-6: A tight example.
3.4 The Component Graph
Our optimal solution to U 2 an be divided into onne ted omponents of minimum

degree two. Our rst step is to partition these onne ted omponents into sub-
omponents. These sub- omponents will be either 2-edge onne ted subgraphs or
paths. This we an a hieve in linear time in the following manner. Take a omponent
of the solution to U 2. Find and ontra t a y le in the omponent. Re urse on
the resultant ontra ted graph. This pro ess terminates with a tree, T , onsisting
of verti es from the original graph and mega-verti es (representing 2-edge onne ted
omponents). A diagram will help larify this pro edure, see Figure 3-7a). The 2-edge
onne ted omponents ( ir led in the diagram) onstitute nodes in T .
Given T , we nd our sub- omponents as follows. All the mega-verti es will be
hosen as sub- omponents. In addition, we partition the other verti es of the tree into
path sub- omponents. To do this, we partition the tree into paths via, say, depth rst
58
a) b)
Figure 3-7: Mega-verti es.
sear h. We add the ondition that any path, derived in this manner, whi h ontains
an internal mega-vertex be split into two disjoint paths. The key observation here is
that asso iated with ea h path sub- omponent are two \free" edges; these are edges
that are part of the lower bound but are not being used by any of the sub- omponents,
This implies that we will be able to add extra edges later to make up for these free
edges. Again a diagram illuminates this pro ess, see Figure 3-7b). Here the sub-
omponents are shown in bold, whilst the \free edges" are dashed. We denote by G
the graph obtained by ontra ting these omponents so that they form nodes. This
indu ed graph is alled the omponent graph. Note that, for larity, we refer to nodes
in G and verti es in G.
For motivation, assume for a moment that ea h of the omponents is 2-edge on-
ne ted and ontains at least six edges. Let there be p su h omponents and hen e
p nodes in the omponent graph G . Performing the depth rst sear h tree based
algorithm gives a 2-edge onne ted subgraph on G ontaining at most 2p 2 edges.
Hen e, at most two edges are asso iated with ea h node or omponent. These edges
together with the edges in the original 2-edge onne ted omponents form a 2-edge
onne ted subgraph on G. Sin e two additional edges are \ harged" to ea h ompo-
nent and, in turn, ea h omponent is of size at least 6 this represents a 43 -optimal
59
solution. This idea forms the basis of our algorithm: apply the depth rst sear h tree
based algorithm to the omponent graph G . We have seen that this approa h is ap-
propriate when the omponents are 2-edge onne ted and ontain at least 6 ar s. The
omponents, though, may also represent small y les or paths. In su h ir umstan es
we need to be more areful. For example, harging two ar s to a 3- y le results in
a lo al blow-up fa tor of 35 , whi h is una eptable. We also need to ensure that we
indu e a 2-edge onne ted subgraph when path omponents are onsidered. These
issues will be dealt with in subsequent se tions.
We nd a slightly dierent set of omponents for the vertex onne tivity ase. Re all
that we observed that we ould nd an optimal solution to U 2 by adding ar s to the
end verti es of ea h path in an optimal solution to the maximum y le-path over
problem. As su h, for the vertex onne tivity ase, our omponents will be the paths
and y les given by the solution to the maximum y le-path over problem. Note
that we again have two \free" edges asso iated with ea h path.
We use a similar method for the strong onne tivity problem. Our lower bound is a
minimum sized subgraph in whi h every vertex is the tail of some ar and the head
of another. This subgraph onsists of disjoint omponents that are either strongly
onne ted or they onsist of strongly onne ted sub- omponents onne ted together
by dire ted paths. An example of this latter type is shown in Figure 3-8i).
We break apart these omponents as before. We separate the strongly onne ted
sub- omponents. We also remove the two end ar s of ea h path. These ar s, whi h we
all the \free ar s" asso iated with a path, will not be forgotten, though, as they are
in luded in our lower bound. This method is illustrated in Figure 3-8ii). After this
pro ess is omplete we have a olle tion of disjoint sub- omponents. For larity we
will subsequently refer to these sub- omponents as omponents. These omponents
60
are either paths or are strongly onne ted. Note that the paths may onsist of just
one vertex.
i) ii)
Figure 3-8: Finding omponents.
For motivation, assume for a moment that ea h of the omponents is strongly

onne ted and ontains at least four ar s. As with our undire ted problems, the
graph, G , indu ed by ontra ting these omponents is alled the omponent graph.
Let there be p su h omponents and hen e p nodes in our indu ed graph. Performing
the y le ontra tion algorithm gives a strongly onne ted subgraph on G ontaining
at most 2p 2 ar s. Hen e, at most two ar s are harged to ea h node or omponent.
These ar s together with the ar s in the original strongly onne ted omponents form
a strongly onne ted subgraph on G. Sin e two additional ar s are asso iated with
ea h omponent and, in turn, ea h omponent is of size at least 4 this represents a
2 -optimal solution.
3
This idea forms the basis of our algorithm: apply the y le ontra tion algorithm
to the omponent graph G . We have seen that this approa h is appropriate when
the omponents are strongly onne ted and ontain at least 4 ar s. The omponents,
though, may also represent 2- y les (digons), 3- y les or paths. In su h ir umstan es
we need to be more areful. For example, harging two ar s to a digon results in a
lo al blow-up fa tor of 2, whi h is una eptable. We also need to ensure that we
indu e a strongly onne ted subgraph when path omponents are ontra ted via the
y le- ontra tion algorithm. These issues will be dealt with in subsequent se tions.
61
3.5 In orporating a Depth First Sear h Tree
We now have a omponent graph G . Our next step is to add a olle tion of edges so
that these omponents be ome onne ted. In parti ular, we will grow a depth rst
sear h spanning tree, T , on G . Note that the tree T is not the same as the tree T ,
previously des ribed. For example, the tree formed may use other edges in the graph,
not just those that were in our U 2 solution.
Sin e the nodes in G may ontain several verti es we need to larify exa tly how to
build a DFS tree on the omponent graph. As we grow the tree we order the verti es
within ea h omponent. We then examine the verti es for new tree edges a ording
to that ordering. We use the notation x y when a vertex x is given priority over a
vertex y. We have three ases.
1. The node represents a y le. First sear h for edges in ident to verti es adja ent
to v in the y le, where v is the vertex in ident to the tree edge. Next onsider
the verti es at distan e 2 from v along the y le, et . Finally onsider v itself.
(We add one extra te hni al ondition. Note there are two verti es on dierent
sides of the y le that are the same distan e from v. Should we en ounter a
vertex u at distan e i that produ es no hildren in the tree, then of the two
verti es at distan e i + 1 we give higher priority to the vertex on the same side
as u.)
2. The node represents a path v0 ; v1; : : : ; vk . Use the ordering v0 v1 vk .
3. The node is 2-edge onne ted: in this ase, where the node is not a y le, sear h
the verti es in any order.
Building the tree in a depth rst sear h manner gives the ni e property that every
edge in the graph is between a des endant node and one of its an estor nodes (ex ept
for edges between verti es in the same node). We will all su h edges \ba k edges"
signifying that they go ba k towards the root.
62
For the vertex onne tivity problem we apply exa tly the same method as for the
edge onne tivity problem.
We wish to apply the y le ontra tion algorithm to the general ase where the
omponents may be digons, 3- y les or paths. In order to apply this method eÆ iently
we rst grow a depth rst sear h tree T , as with the undire ted graph problems, on
the omponent graph G . The tree will help us dis over appropriate y les in G . In
addition, the spe ial stru ture provided by depth rst sear h trees will help simplify
the proofs on erning the performan e of the algorithm.
We in orporate the tree as follows. We grow our tree from a node ( omponent) r,
hen eforth termed the root node or root omponent. The tree, T , will be an out-tree,
a spanning tree with the property that there is a dire ted path from r to every other
node. The tree is built via a DFS approa h with root node r. Re all the important
property of DFS trees that for any non-tree ar s (u; v), either v was added to the
DFS tree before u or u lies on the dire ted path from r to v in the tree. Again, as the
nodes in G may ontain many verti es, we need to give a priority s heme for growing
the tree.
1. Suppose that in growing the DFS tree we rea h a omponent that represents
a dire ted path. Let the path onsist of the verti es vk ; vk 1; : : : v0 with ar s
(vi+1; vi). Use the ordering v0 v1 vk .
2. If the omponent represents a digon on the verti es v0 and v1 , where v0 is the
vertex in ident to the tree ar , use the ordering v1 v0 .
3. If the omponent represents a 3- y le, f(v0 ; v1); (v1; v2); (v2 ; v0)g, where v0 is the
vertex in ident to the tree ar , use the ordering v0 v2 v1.
4. For a omponent representing a strongly onne ted omponent of size at least
4, we may order the verti es in any manner.
63
Given T and a omponent C 2 G , let TC (or just T if there is no onfusion) be
the subtree of T rooted at C . If C ontains the verti es v1 ; v2; : : : ; vk then Ti TC
is the subtree with root vertex vi 2 C . Denote by Ci the hild omponents of vi in
Ti . There may be several su h Ci for ea h vi . In su h ir umstan es, the ontext will
make it apparent to whi h Ci we are referring.
We will use the following labels. We all an ar a down ar with respe t to a
omponent C , if it is dire ted from G n TC to TC . We all an ar a ba k ar with
respe t to a omponent C , if it is dire ted in the opposite dire tion.
We all the rst non-root omponent C , dis overed in growing T , the target om-
ponent. Our aim will be to apply a version of the y le ontra tion algorithm in whi h
we nd and ontra t a y le through the root and the target omponent, and then
re urse. This method will be des ribed in detail in the Se tion 3.8.2.
64
3.6 Edge Conne tivity
3.6.1 The Lo al Improvement Method
Given our urrent subgraph, these omponents within a tree stru ture, how do we
reate a good 2-edge onne ted subgraph? We attempt to 2- onne t up the tree,
working up from the leaves, adding extra ba k edges as we go. Ea h edge that we
add will be harged to a omponent. Our proof will follow from the fa t that no
omponent is harged more than one third its own weight.
First some terminology. Ea h omponent has one tree edge going up towards
the root node. This edge we will all an upper tree edge w.r.t. the omponent. The
omponent may have several tree edges going down towards the leaf nodes. These will
be alled lower tree edges w.r.t. the omponent. Initially ea h non-root omponent
is harged one. This harge is for its in ident upper tree edge.
Lemma 3.6.1. For nodes representing 2-edge onne ted omponents the harge is at
most 2.
Proof. If we examine su h a omponent then we may add any ba k edge from the node,
in addition to the upper tree edge. This gives a harge of two to the omponent.
Re all that if the omponent ontains at least 6 edges then the asso iated harge
is at most one third the omponents weight. It remains to deal with 2-edge onne ted
omponents of at most 5 edges, and with paths.
Lemma 3.6.2. For path nodes the harge is at most d k3 e + 2, where k is the number
of edges in the path.
Proof. Consider a path, P , of k edges. Let the node representing P be alled the
path node, N0. Let the upper tree edge, e0 , harged to the path node be in ident to
the vertex v0. Let the subtree rooted at the path node be alled T0. Let the hildren
of the path node, N1 ; : : : ; Nr , have asso iated upper tree edges e1 ; : : : ; er . We may
assume that these edges are in ident to the verti es v1 ; : : : ; vr on the path P . These
65
e0
T0
P
N
V-T 0
0
e1 er
e2
N1 N2 Tr Nr
T1 T2
Figure 3-9: Subtrees.
hild nodes are the root nodes of r indu ed subtrees, T1 ; : : : ; Tr . This situation is
shown in Figure 3-9.
Re all we work our way up from the leaf nodes. Hen e, we may assume that all
of the subtrees, T1; : : : ; Tr , are already 2- onne ted up when we onsider N0 . This
allows us to arry out some prepro essing. First ontra t V n T0 into v0 . Consider the
subtree T1 . There are at least two sele ted edges going beyond its root node N1. One
of these edges is e1 , let the other be e01. Let e01 be in ident to the path P at vertex v10
(note that if e01 goes beyond the path, P , then, sin e we ontra ted V n T0 into v0 , it
is now in ident to v10 = v0 ). Then the se tion of the path between v1 and v10 is already
2- onne ted. Hen e we may ontra t into one super vertex that se tion of the path
and T1. We repeat this ontra tion pro ess, in turn, for ea h subtree T2 ; : : : ; Tr .
After this ontra tion pro ess we are left with a single path, P 0 of at most k
verti es. Two observations, on erning edges between verti es in P 0, follow from
the DFS priority s heme used to hoose tree edges. First, there are no edges that
orrespond to edges in ident to a vertex in Ti and another in Tj , i 6= j . Su h edges
annot exist as they would have been hosen instead of the tree edge ej , where we
66
assume that ei was hosen before ej . Se ond, onsider an edge in P 0 orresponding to
an edge between a vertex in Ti and a vertex in P . This edge is between vi and some
vertex vj to the right of vi in the path P 0, (possibly vi = vj ), otherwise it would have
been sele ted before ei.
Suppose we 2- onne t P 0. We know the tree edge e0 goes up to V n T0. If another
su h edge is added in 2- onne ting up P 0, or if su h an edge was already hosen for
one of the subtrees T1 ; : : : ; Tr , then we have su essfully 2- onne ted up the subtree
T0 . Otherwise, in addition to the edges hosen for P 0 we need one extra edge (e0 has
been hosen and we need one more). So the ost of onne ting up T0 is, at most, the
ost of 2- onne ting P 0 plus two edges.
So what is the ost of 2- onne ting P 0? In the worst ase P 0 also ontains k edges,
and it turns out that the ost of 2- onne ting su h a path is at most d k3 e. To show
this, we break up the path into blo ks of at least 3 verti es su h that ea h blo k is
harged only one extra edge in the onne tion pro ess.
Start with the leftmost vertex in the path. We add rightward edges wherever a
path edge threatens to be a bridge. On adding an edge it may be possible to remove
some urrent edge. If not, then all of the verti es to the left of the urrent bridge, not
ontained in any other blo k, form a new blo k. The blo ks formed have the property
that, ex ept for the blo ks ontaining the path endpoints, they ontain at least three
verti es. To see this, let (u0; u1) be the urrent bridge, with the last blo k losed at
u0 . Let u2 and u3 be the two verti es to the right of u1 . Choose the edge, e, from
the last blo k rea hing furthest to the right. Now e is not in ident to u1, otherwise
u1 is a ut vertex. If e is in ident to u2 then there is a rightward edge e0 out of u1 ,
otherwise u2 is a ut vertex. We an hoose e0 for free by dropping the edge (u1; u2),
see Figure 3-10. If e goes to u3 or beyond then we are already done. Note that in
our pre-pro essing we may have merged subtrees from below, i.e. some ontra ted
verti es might onsist of two or more verti es. In this ase the arguments with regard
to ut verti es may not apply; however, sin e these ontra ted verti es ontain free
edges, the proof is easier.
Sin e the blo k ontaining the right endpoint does not ontribute an extra edge,
67
e e’
u0 u1 u2 u3
Figure 3-10: Paths.
we have added an extra d k3 e edges. So to deal with T0 we add, at most, an extra

d k3 e +2 edges. Re all, that asso iated with P we have two free edges, (i.e. k +2 edges
in total), so these extra two edges are a ounted for. Note that d k3 e 13 (k + 2) so the
asso iated blow-up fa tor for ea h path is 34 .
Now onsider the ase of small 2- onne ted omponents. First note that the only
2-edge onne ted omponents of at most 5 edges are y les of at most 5 edges (for
example 4- y les with hords ould not have been hosen in the solution to U 2). Our
aim is to show that in pi king an extra ba k edge to 2- onne t up su h y les we an
remove some edge. In ee t, we get the ba k edge for free and the harge to the node
is just one.
Lemma 3.6.3. For y le nodes of size 3, the harge is one.
Proof. We will label a ba k edge from the node going up the DFS tree by b. A ba k
edge from below that is in ident to the node will be labelled f . The deepest ba k
edge from the des endants of the node will be labelled d. In all of our diagrams we use
the following s heme. Solid edges were already sele ted at the point of examination
of the omponent; dashed edges were not. Bold edges are sele ted by the algorithm
after the examination; plain edges are not.
(I) Lower tree edges are all adja ent at the same vertex as the upper tree edge.
Assume this vertex is v1. Note that neither v2 nor v3 have in ident edges going down
the tree, otherwise the DFS algorithm would have hosen them as tree edges. Now
v1 is not a ut vertex, so there must be a ba k edge b from v2 or v3 . Suppose it is
in ident to v3; the other ase is similar. We may add b and remove the edge (v1 ; v3)
to 2- onne t upwards the node. This is shown in Figure 3-11.I).
68
(II) The upper tree edge is in ident to a vertex whi h has an adja ent vertex that is
in ident to a lower tree edge.
1. There is a ba k edge b from v2 or v3. We may assume it is from v3 . The other ase
is similar. Add b and remove the edge (v1; v3). This is shown in Figure 3-11.II)1).
I) II)1) II)2)
b
b d v1
v1
v1 v3 v3
v3
v2 v2
v2
Figure 3-11: 3 y les.
2. Otherwise, sin e v1 is not a ut vertex, there is an edge d going beyond v1. Add d
and remove (v1 ; v2). See Figure 3-11.II)2).
Proof.
1. There is a ba k edge b from v2 or v4 . We may assume it is from v4 ; the other ase
is similar. Add b and remove the edge (v1; v4). This is shown in Figure 3-12.I)1).
2. Otherwise, sin e v1 is not a ut vertex, there is a ba k edge b in ident to v3 . In
addition, v4 is not a beta-vertex. Therefore, as there is no edge from v4 going down
the tree (otherwise it would be a tree edge) there must be an edge between v2 and
v4 . Add b and (v2 ; v4 ). Remove (v1 ; v4 ) and (v2 ; v3 ). See Figure 3-12.I)2).
1. There is an edge d going beyond v1 . Add d and remove (v1 ; v2). See Figure
3-12.II)1).
is similar. Add b and remove the edge (v1; v4). This is shown in Figure 3-12.II)2)).
69
I)1) b I)2)
v4 b
v4
v1 v3 v1
v3
v2 v2
II)1) II)2)
b
v4
d v1 v4 v1
v2 v3 v2 v3
II)3)a) II)3)b)
b
v4
v1 v4 b v1
v2 v2
v3 v3 f
III)2)a) III)2)b)
b
d v1 v1
v2 v4 v2 v4
v3 v3
3. Otherwise, sin e v1 is not a ut vertex, there is a ba k edge b in ident to v3 . In

addition, v4 is not a beta-vertex. Therefore, either there is an edge between v2 and v4
or there must be an edge f from v4 going down the tree. In the former ase, Figure
3-12.II)3)a), add b and (v2 ; v4). Remove (v1; v4 ) and (v2; v3 ). In the latter ase, gure
3-12b)iii)B), we have two possibilities. First suppose that the lower tree edge and
f are in ident to the same subtree below the 4- y le. Then add b and f , remove
70
also (v1 ; v4) and (v2; v3 ). Otherwise the lower tree edge and f originate in separate
subtrees with root nodes below the 4- y le. It this ase observe that the edge f is an
upper tree edge for a lower omponent. As su h we re eive the edge for \free" and
onne ting up the 4- y le is then easy. Similar situations to this se ond ase, where
there is su h a type f edge, may arise again. In these situations we an always show
that we obtain su h a \free" edge. Thus, in the ases that follow, we will assume that
all su h type f edges belong to the same subtree as the lower tree edge.
(III) The upper tree edge is in ident to a vertex whi h has an opposite vertex that
is in ident to a lower tree edge, but adja ent verti es are not in ident to a lower tree
edge.
1. There is a ba k edge b from v2 or v4. We may assume it is from v4. Add b and
remove the edge (v1 ; v4).
2. Otherwise, note that v4 is not a beta-vertex. Therefore, as there is no edge from
v4 going down the tree (otherwise it would be a tree edge) there must be an edge
between v2 and v4. In addition, sin e v1 is not a ut vertex, there is either an edge d
going beyond v1 or there is a ba k edge b in ident to v3 . We deal with both ases in
a similar fashion, see Figures 3-12.III)2)a) and 3-12.III)2)b). Add (v2 ; v4) and either
d or b. Remove (v1 ; v4 ) and (v2 ; v3 ).
So by dealing with ea h node individually we were able to obtain a 57 approximation
fa tor. We will begin by ontinuing this lo alised approa h for the situation in whi h
nodes represent 5- y les. It will turn out that su h an approa h is not quite suÆ ient
in all instan es. We will deal with these instan es by pairing nodes and harging
them together.
The following observations will also be useful. Suppose there is a y le in G
onsisting of ve edges, (v0; v1 ); (v1; v2); (v2; v3 ); (v3; v4) and (v4 ; v0). Note that the
verti es v0; : : : ; v4 may indu e up to ten edges in G. Suppose that every optimal
minimum 2- onne ted subgraph ontains at least 4 of these ten possible ar s. Then
we may hoose the edges in the 5- y le to be part of our solution, ontra t the y le,
and nd an -optimal subgraph on the resultant graph. The following lemma shows
that this approa h will be suÆ ient to produ e an -approximation fa tor, provided
71
that 54 . Re all, we use the notation G=C to denote the graph resulting from
ontra ting a set of verti es C V .
Lemma 3.6.5. If 5 verti es indu e 4 edges in every optimal solution then we may
hoose any 5- y le through the verti es and onsider the resultant sub-problem.
Proof. Let the optimal solution have opt edges, 4 of whi h are indu ed by v0 ; : : : ; v4.
So k = opt 4 ar s are used on the edge set indu ed by G0 = G=fv0; : : : ; v4g. Let
opt0 be the size of the optimal solution to D 0 . Clearly opt0 k . So if we an nd an
-optimal solution to D0 and then add the 5 ar s in the y le we will have a solution
to G. This solution will use at most opt0 + 5 k + 5 (k + 4) = opt edges,
provided that 54 . Hen e we will have a -optimal solution to G.
This lemma will be useful in bounding the harge on a 5- y le. As mentioned
earlier, we annot bound the harge on ea h 5- y le, but we an bound the amortized
harge.
Lemma 3.6.6. The amortized harge on a 5- y le is at most 35 .
Proof. First, let's begin by attempting to deal with 5- y les using our lo alised ap-
proa h.
1. There is a ba k edge b from v2 or v5. We may assume it is from v5. Add b and
remove the edge (v1 ; v5).
2. Otherwise, sin e v1 is not a ut vertex there is a ba k edge b from either v3 or
v4 . These ases are symmetri so assume it is from v4 . In addition, for the same
reason, there must be an edge d. Also, v5 is not a beta-vertex so one of the following
situations must arise.
(a) There is an edge (v2 ; v5). Add b and (v2 ; v5). Remove (v1; v2) and (v4 ; v5). See
Figure 3-13.2)a).
(b) Otherwise every optimal solution indu es four edges within the 5- y le, sin e
two must be in ident to v2 and two in ident to v5. These four edges are learly
distin t so we may apply Lemma 3.6.5 and solve on the smaller subproblem.
72
2)a)
b
v5
v4
v1
d v2 v3
Figure 3-13: A 5 y le
1. There is an edge d going beyond v1 . Add d and remove (v1 ; v2). See Figure 3-14.1).
is similar. Add b and remove the edge (v1; v5).
3. There is a ba k edge b in ident to v4.
Figure 3-14.3)a).
(b) There is an edge (v3 ; v5). Add b and (v3 ; v5). Remove (v1; v5) and (v3 ; v4). See
Figure 3-14.3)b).
( ) There is a down edge f from v5. Add b and f , remove (v1; v2 ) and (v4 ; v5). See
Figure 3-14.3) ).
(d) Otherwise, sin e v5 is not a beta-vertex, there must be a ba k edge b0 from v3 .

Add b and b0 , remove (v3; v4 ) and the upper tree edge in ident to v1. See Figure
3-14.3)d). Note that we have removed a tree edge. We will return to this ase
subsequently.
73
1) 3)a) b 3)b)
b
v5 v5 v5
v4 v4
d v4
v1 v1 v1
v2 v2 v2
v3 v3 v3
3)c) b 3)d) b
v5 v5
v4 v4
v1 v1 b’
f
v2 v2
v3 v3
C1
4)a) 4)b) 4)c)
b v5 b
v5
v4 v4 C3
v1 v1
v1
v2 C2
v2 v3 v3
v3
4)d) 4)e)
v5 v5 b
v4 v4
v1 v1
f b
v2 v2
v3 v3
f
Figure 3-14: More 5 y les.
4. There is a ba k edge b in ident to v3. Now, at least one of v4 and v5 has degree
greater than two, otherwise we would have ontra ted the edge (v4; v5).
Figure 3-14.4)a).
(b) There is an edge (v2 ; v4). Add b and (v2 ; v4). Remove (v1; v2) and (v3 ; v4). See
Figure 3-14.4)b).
74
( ) There are edges (v1 ; v4) and/or (v3; v5). This situation is prohibited as we have
a beta-pair. The removal of verti es v1 and v3 leaves three omponents, one of
whi h C3 is of size two. The three omponents are shown in Figure 3-14.4) ).
(d) There is a down edge f from v5. Add b and f , remove (v1; v5 ) and (v2 ; v3). See
Figure 3-14.4)d).
(e) There is a down edge f from v4. Add b and f , remove (v1; v2 ) and (v3 ; v4). See
Figure 3-14.4)e).
(III) The upper tree edge is in ident to a vertex whi h has an opposite vertex that
is in ident to a lower tree edge, but adja ent verti es are not in ident to a lower tree
edge.
is similar. Add b and remove the edge (v1; v5).
2)a)i) 2)b)
v1 v1 b
v2 v5 v2 v5
v3 v3
v4 v4
d
Figure 3-15: Even more 5 y les.
2. There is an edge (v2; v5 ). Note that v1 is not a ut vertex.

(a) There is an edge d from below the node that goes beyond v1 . Add d and Remove
(v1; v5 ) and (v2; v3 ). See Figure 3-15.2)a).
(b) There is a ba k edge b from v4 (the ase on erning v3 is similar). Add b and
(v2; v5 ), remove (v1; v2) and (v4; v5). See Figure 3-15.2)b).
3. Otherwise the 5- y le indu es four ar s in any optimal solution and we may ontra t
the y le.
So we're not done yet. We have one troubling situation: when an upper tree
edge was removed. See Figure 3-14.3)d). This tree edge is a lower tree edge for the
75
omponent above the 5- y le. Sin e our arguments for this omponent relied upon
the positioning of its lower tree edges, our previous arguments may be ome invalid
for that omponent. We will thus attempt to pair our troublesome omponent 5- y le
with the above omponent. We will then add edges to onne t up both omponents
at the same time. The omponents will be harged the extra edges as a pair.
So we need to re onsider the ase of omponents that lie above these types of
5- y le. We attempt to pair o these 5- y les with the omponent above them. If
the node above is a 2- onne ted omponent of size 4 or more then we are done as
we have two spare edges to play with, sin e the 5- y le is harged only one edge. So
suppose the node represents a 3- y le. Let e be the tree edge removed, and let b and
b0 be the two ba k edges hosen instead.
(I) The 3- y le had lower tree edges all adja ent at the same vertex, y1, as the upper
tree edge. If either b and b0 is in ident to y1 then our pre ious arguments for three
y les apply, sin e swapping e for b, say, does not hange the stru ture of the lower
tree edges for the 3- y le. Neither b nor b0 are in ident to y2 or y3 as otherwise they
would have been hosen as lower tree edges for the 3- y le, not edge e. So both b and
b0 go beyond the 3- y le. In this ase, see Figure 3-16.I), sin e y1 is not a ut vertex
there is a ba k edge from either y2 or y3. Adding this edge allows us to remove one
from the 3- y le, and hen e the net harge to the 3- y le and the 5- y le is just two
edges.
I)
y1 y3
y2
b
e b’
v5
v4
v1
v2 v3
Figure 3-16: A pair of y les.
(II) The 3- y le had a lower tree edge and the upper tree edge in ident at adja ent
verti es.
76
1. Some edge from the 5- y le goes beyond the 3- y le. The ase in whi h it is edge
b is shown in Figure 3-17.II)1). Add ba k in edge e and (v4 ; v5 ), remove edges b0 and
(y1; y2). The net harge to the two omponents is then just two edges.
II)1) II)2)
y1 y1
y2 y3 y2 y3
b b
e e
v5 v5
v4 b’ v4 b’
v1 v1
v2 v3 v2 v3
Figure 3-17: More pairs of y les.
2. Both b and b0 are in ident to y1. If either is in ident to y2 or y3 then it has the
same priority as e; hen e the loss of e for b would have no ee t on our previous proof
method. Sin e y1 is not a ut vertex either y2 or y3 has a ba k edge. If there is a
ba k edge then we an drop an edge from the 3- y le, Figure 3-17.II)2).
e’
e b
v1 b’
v5
v2
v3 v4
Figure 3-18: A y le and path pair.
Finally suppose the node above the 5- y le represents a path, P . Let e be the
original tree edge harged to the 5- y le. Suppose that in 2- onne ting up the path
another edge (in addition to e) from the 5- y le or below is used. Then this edge
also two onne ts up the 5- y le whi h gets the edge for \free", giving a net harge of
one to the 5- y le. So no su h edges are used. In whi h ase adding b and b0, whilst
removing (v4; v5) has no ee t on the argument on erning the path. Hen e they may
77
be onsidered separately. See Figure 3-18.
Our result follows.
Theorem 3.6.7. The algorithm is a 34 -approximation algorithm for 2EC .
3.6.2 Comparison with the Subtour Relaxation
Re all the ut- overing LP relaxation for 2EC .
min P xe
X 2
e E
xe 2 8S V; S 6= ;
2( )
e Æ S
xe 0 8e 2 E
It may be noted that this relaxation is similar to the familiar subtour relaxation
for the Travelling Salesman Problem (T SP ). The subtour relaxation imposes the
additional onstraints that Pe2Æ(fvg) xe = 2; 8v 2 V . For metri osts, though, there
is an optimal solution to the LP relaxation whi h is also optimal for the subtour
relaxation. A famous onje ture is that, for a graph with metri osts, opt(T SP )
3 opt(LP). Whilst this onje ture remains unresolved, some progress has been made.
4
For example, Goemans [19℄ onsidered the ee ts of adding lasses of valid TSP
inequalities to the subtour onstraints. He showed that, for a range of lasses, the
optimal solution produ ed was always within a fa tor 43 of the solution produ ed by
just the subtour onstraints. In addition, Monma et al [35℄ have shown a weaker
result, that, given metri osts, opt(T SP ) 34 opt(2EC ). One impli ation of the
onje ture is that for a 2-edge onne ted graph, opt(2EC ) 43 opt(LP). Carr
and Ravi [4℄ provide a proof of the impli ation for the spe ial ase of half-integral
solutions of the relaxation. Unfortunately, the impli ation (in its entirety) does not
follow simply from our result. This is due to the fa t that our lower bound, based
upon integral degree 2 subgraphs, may be stronger in some situations than the bound
oered by the linear program. The graph in Figure 3-19a) is an example. Figure
3-19b) shows the half integral optimal solution given by the LP. Here, edges given
78
weight 1 are solid whilst edges of weight 12 are dashed. This has total value 9, whereas
the minimum integral degree 2 subgraph, Figure 3-19 ), has value 10. A weaker
impli ation of the TSP onje ture is that 2EC is approximable to within 43 , and our
result does indeed verify this impli ation.
a) b) c)
Figure 3-19: Subtour vs integral degree 2 subgraph.
79
3.7 Vertex Conne tivity
Our approa h for the problem 2V C is similar to that of 2EC . Re all that our lower
bound is a solution to the maximum y le-path over problem rather than a solution
to U 2. This ensures our initial omponents are either paths or 2-vertex onne ted
omponents. Then, given our tree of omponents we attempt to 2-vertex onne t up
the graph using a limited number of extra edges. In doing so, though, we maintain
the following property. Again, given the tree of omponents T , denote by TC the
subtree in T rooted at a omponent C
Property 1. After dealing with omponent C we require that at least two edges
ross from TC to G n TC . In addition these two edges must be vertex-disjoint.

As before we harge additional edges to the omponents.
Lemma 3.7.1. For path nodes the harge is at most d k3 e + 2, where k is the number
of edges in the path.
Proof. We take a similar approa h to the method used for edge onne tivity. Starting
at the leftmost vertex in the path we add rightward edges (these edges may originate
from omponents below the path as before). We attempt to split the path into blo ks
of at least three verti es with the property that we add only one edge per blo k.
Suppose the last edge added, a, is in ident to u0. We lose the previous blo k at
the vertex in the path before u0. We show that the next blo k onsists of at least three
verti es. Take the edge, e, from the previous blo k that goes furthest rightwards. If
e rea hes u3 then we are done. Suppose e rea hes u2 . If there is an e0 from u1 going
beyond u2 then we get this edge for free by removing (u1; u2). See Figure 3-20a).
Otherwise there is an e0 originating at u0 going beyond u2. If there is an edge e00
from the previous blo k entering u1 then we get e0 for free. This is shown in Figure
3-20a). Suppose there is no su h e00. Sin e u1 is not a Beta-vertex we then have
the possibilities shown in Figure 3-20 ). In ases i) and ii) by onsideration of the
subtrees rooted at u1 and u0, respe tively, we an remove a path edge as shown. We
avoid ase iii), by for ing an edge going down from u1 to be a tree edge.
80
a) a e e’ b) a e’’ e’
u0 u1 u2 u3 u0 u1 u2 u3
c)i) a e ii) a e
u0
u0 u1 u2 u3 u1 u2 u3
iii) a e
u0 u1 u2 u3
Figure 3-20: Paths and vertex onne tivity.
Proof. We will use the following notation. Call the two ba k edges from the subtree
below the 3- y le a and a0. Let them originate at the verti es z1 and z2. We will
presume that a is the lower tree edge for the 3- y le. In ontrast to the 2EC ase
our arguments for 2V C are based upon both of the edges from below, not just one of
them. Let the upper tree edge from the 3- y le be e = (v1 ; y1). In the ourse of these
proofs we will take y2 to be a generi vertex in T n (TC [ fy1g). We will assume that
any ba k edge leaving TC , that is not in ident to y1, is in ident to the same vertex
y2 . This is the worst ase s enario. The proofs are simpler if this does not happen to
be the ase. See Figure 3-21.
1) y2 y 2)a) y2 2)b) y2
1
e’ y1 e’ y1
e’
e e e
v1 v3 v1 v3 v1 v3
a’ a v2 a’ a v2 a’ a v2
z2 z1 z2 z1 z2 z1
81
(I) Lower tree edges are all adja ent at the same vertex, v1, as the upper tree edge.
Noti e that a0 is not in ident to the 3- y le. It an not be in ident to v1 as a is. It is
not in ident to v2 or v3 as otherwise it would have been hosen as a lower tree edge
not a.
1. Edge a0 is in ident to a vertex y2 6= y1. Sin e v1 is not a ut vertex there is a ba k
edge e0 from either v2 or v3 . These ases are symmetri , the latter is shown in Figure
3-21.1). We may add e0 and remove (v1 ; v3) for a net harge of one to the y le.
2. Edge a0 is in ident to y1. Again there must be a ba k edge e0 from v2 or v3. If e0 is
in ident to y2, add e0 and remove e and (v1; v3 ) for zero net harge to the y le. See
Figure 3-21.2)a). If e0 is in ident to y1, add e0 and remove e and (v1 ; v3). We still have
one spare edge available. Choose any ba k edge that is in ident to some y2 6= y1. See
Figure 3-21.2)b).
(II) The upper tree edge e is in ident to v1 and the lower tree edge a is in ident to
an adja ent vertex v2 .
1. If a0 is in ident to y2 then we are already done. In fa t, we may also remove the
edge (v1 ; v2). See Figure 3-22.1)a). If a0 is in ident to y1 then remove (v1; v2 ). We
then have one extra edge to play with. This we add to onne t the 3- y le upwards
to some vertex y2. This is shown in Figure 3-22.1)b).
1)a) y2 1)b) y2 2) y2
y1 y1 y1
e e’ e e’ e e’
v1 v1 v1
v3 v3 v3
v2 v2 v2
a’
a’ a a’ a a
z2 z1 z2 z1 z1 z2
2. Otherwise a0 is in ident to v1 or v3. The ases are similar and we deal with the
latter. We may remove (v2 ; v3) and so have one extra edge available. See Figure
3-22.2).
82
Lemma 3.7.3. For y le nodes of size 4 the harge is one.
Proof.
(I) All lower tree edges are in ident to v1 .
1. The ba k ar a0 goes beyond the 4- y le to y2 6= y1. Note that v1 is not a ut
vertex so there is a ba k ar e0 from one of v2 ; v3 or v4 .
(a) The ba k edge e0 is in ident to y1. Add e0 and remove e. The ase where e0 is
in ident to v3 is shown in Figure 3-23.1)a). The harge to the 4- y le is just
one.
(b) The ba k edge e0 goes to y2 from either v2 or v4 . Suppose it is from v4 . Add e0
and remove (v1; v4 ). See Figure 3-23.1)b). The other ase is similar.
( ) The ba k edge e0 goes from v3 to y2. Sin e v2 and v4 are not Beta-verti es there
must be an edge (v2 ; v4). Add e0 and (v2 ; v4). Remove (v1 ; v4) and (v2; v3 ). See
Figure 3-23.1) ).
1)a) y2 1)b) y2 1)c) y2

y1 y1 y1
e e’ e
e
v4 e’ v4 v4 e’
v1 v3 v1 v3 v1 v3
a’ a’ a’
v2 v2 v2
a a a
z1 z2 z1 z2 z1 z2
2. The ba k ar a0 goes beyond the 4- y le to y1. This ase is dealt with in a similar
fashion to ase 1.
(II) There is a lower tree edge a in ident to v3, but none in ident to v2 or v4.
1. The ba k ar a0 goes beyond the 4- y le to y2 6= y1. Then we are already 2-vertex
onne ted beyond the 4- y le. See Figure 3-24.1).
2. The ba k ar a0 goes beyond the 4- y le to y1.
83
(a) There is a ba k edge e0 from v2 or v4 to y2. We show the latter ase in Figure
3-24.2)a). Add e0 and remove e and (v3 ; v4). This leads to zero net harge to
the 4- y le.
(b) There is a ba k edge e0 from v2 or v4 to y1. We show the latter ase in Figure
3-24.2)b). Add e0 and remove e and (v3; v4). We still have one edge to play
with. We use this edge to onne t us up to y2.
1) y2 2)a) 2)b)
y1 y1 y2 y1 y2
e e e’ e e’
v4 v1 v4 v1 v4
v1
a’ v2 a’ v2 v3 a’ v2 v3
v3
a a a
z1 z2 z1 z2 z1
z2
3)a) 3)b)
y1 y1
y2 y2
e e’ e e’
v1 v1
v4 v4
a’ v2 v3 a’ v2 v3
a a
z2 z1 z2 z1
3. The ba k edge a0 is in ident to v1 .

(a) There is a ba k edge e0 from v2 or v4 to y2. Both ases are similar. For the
latter ase add e0 and remove (v1; v4). See Figure 3-24.3)a).
(b) There is a ba k edge e0 from v2 or v4 to y1. For the latter ase add e0 , remove
e and (v1 ; v4 ). We then still have one edge to play with. We pi k any suitable
ba k edge that onne ts us up to y2. See Figure 3-24.3)b).
84
( ) Otherwise sin e v2 and v4 are not Beta-verti es there must be an edge (v2 ; v4).
This we deal with as before.
(III) There is a lower tree edge a in ident to v2 or v4. These ases are symmetri so
we will deal with the former ase.
1) y2 2) y1 y2 3) y2
y1 y1
e e e
v1 v4 v1 v4 v1 v4
a’
a’
v2 v3 v2 v3 v2
a’ v3
a a a
z2 z1 z2 z1 z1
z2
4)a)i) y2 4)a)ii) y2 4)a)iii) y1 y2

y1 y1
e e
e e’ v4 v4 e’
v4 v1
v1 v1
e’ v2 v3
v2 v3 v3 a’
a’ v2 a’
a a
a
z1 z2 z1 z2
z1 z2
4)a)iv) y1 y2 4)b) y1 y2 4)c) y1 y2
e e’ e e’ e
v4 v4 e’
v1 b v4
v1 v1
v2 b
v2 v3 v2 v3
v3 a’ a’ a’
a a a
z1 z2 z1 z2 z1 z2
1. The ba k ar a0 goes beyond the 4- y le to y2 6= y1. Then we are already 2-vertex

onne ted beyond the 4- y le. See Figure 3-25.1). In fa t we an also remove the
edge (v1 ; v2).
2. The ba k ar a0 goes beyond the 4- y le to y1. Remove (v1 ; v2). See Figure 3-25.2).
85
We may then add any appropriate ba k edge to onne t us up to y2.
3. The ba k ar a0 is in ident to v1 or v3. Here we may remove one edge from the
4- y le and then hoose any appropriate ba k edge to y2. The ase where a0 is in ident
to v1 is shown in Figure 3-25.3).
4. The ba k ar a0 is in ident to v4. Consider how we may rea h y2.
(a) Let e0 be a ba k edge into y2. The ases where e0 is from below the 4- y le
or where it originates from v2 ; v3 or v4 are easily dealt with. See Figures 3-
25.4)a)i)-iv).
(b) Then let e0 be in ident to v1. Suppose there is a ba k edge b from v2 to y1.
Then add b and e0 , remove e and (v1 ; v2). See Figure 3-25.4)b). The situation
in whi h there is a ba k edge from v4 to y1 is similar.
( ) If there is a ba k edge b from v3 to y1 then add b and e0 . Remove e, (v1 ; v4) and
(v2; v3 ). See Figure 3-25.4) ).
(d) Otherwise there must be a ba k edge b into y1 originating from below the 4- y le.
Here we add b and e0 and remove e and (v1; v2).
To deal with the ase of 5- y les we need the following lemma.

Lemma 3.7.4. Suppose a beta-free graph G ontains two verti es, v2 and v5 , of degree
2 in a 5- y le C = fv1 ; : : : ; v5 g. Then G ontains an optimal 2-vertex onne ted
subgraph that does not ontain the edge (v3 ; v4 ).
Proof. Let S be a minimum 2-vertex onne ted subgraph. Suppose that S ontains
(v3 ; v4). Now S must ontain two disjoint edges, e1 and e2, from C to G n C . If e1 and
e2 are in ident to v3 and v4 then S n (v3 ; v4 ) is also 2-vertex onne ted, a ontradi tion.
So, sin e v2 and v5 are of degree 2, without loss of generality e1 is in ident to v1 and
e2 is in ident to v3 . But G ontains no Beta verti es and so there must be an edge e
from v4 to G n C . Now S n (v3 ; v4) [ e is also 2-vertex onne ted and the result follows.
This is shown in Figure 3-26.
86
v5 v4
v1 C
e
v2 v3
e1 e2
G-C
Figure 3-26: Easy to deal with 5 y les.
Lemma 3.7.4 implies that if we nd a pentagon that has two su h degree two
verti es then we may re urse on the graph G n (v3; v4 ). In fa t we an generalise this
result.
Lemma 3.7.5. Let G be a beta-free graph that ontains a 5- y le C = fv1; : : : ; v5g.
Suppose v2 has degree 2 and that v5 has no neighbours outside C . Then G ontains
an optimal 2-vertex onne ted subgraph that does not ontain the edge (v3 ; v4 ).
Proof. Similar to the proof of Lemma 3.7.4.

Lemma 3.7.6. For y le nodes of size 5 the harge is one.
Proof.
(I) All lower tree edges are in ident to v1 . Note that v1 is not a ut vertex.
1. The ba k edge a0 is in ident to y2.
(a) There is a ba k edge e0 into y1 from v2 ; v3; v4 or v5 . We add e0 and drop e. An
example is shown in Figure 3-27.1)a).
(b) There is a ba k edge e0 into y2 from v2 or v5 . We illustrate the latter ase in
Figure 3-27.1)b). Add e0 and remove (v1; v5 ).
( ) There is a ba k edge e0 into y2 from v3 . The ase from v4 is symmetri . The
situation in whi h there is an edge (v2; v5) is shown in Figure 3-27.1) ). The
situation in whi h there is an edge (v2; v4 ) is similar. Otherwise v2 is of degree
87
2 and v5 has only neighbours within the 5- y le. Hen e we may apply Lemma
3.7.5 and onsider the graph G n (v3; v4 ).
2. The ba k edge a0 is in ident to y1. This ase is dealt with in a similar manner to
the previous ase.
1)a) y2 1)b) y2 1)c)

y1 y1 y2 y1
e’
e e’ e e
v5 v4 e’ v5 v4 v5 v4
v1 v1 v1
a’ a’ a’
v2 v3 v2 v3 v2 v3
a a a
z2 z1 z2 z1 z2 z1
(II) There is a lower tree edge a in ident to v3 or v4 , but none in ident to v2 or v4 .

These ases are symmetri so we will deal with the former ase.
1. The ba k edge a0 is in ident to y2. Then we are already done. See Figure 3-28a).
2. The ba k edge a0 is in ident to y1.
(a) There is a ba k edge e0 into y2 from any of the verti es in the 5- y le ex ept v3 .
Add e0 and remove e. An example is given in Figure 3-28.2)a).
(b) There is a ba k edge into y1 from either v2 or v4 . Both ases are similar. The
latter is shown in Figure 3-28.2)b). Add e0 and remove e and (v3; v4 ). We then
add an appropriate edge in ident to y2.
( ) The situations in whi h there are edges (v2 ; v5) or (v2 ; v4) are easy to deal with.
(d) So v2 has degree 2. It follows that there must be an edge e0 = (v5; y1), otherwise
we apply Lemma 3.7.5 and onsider Gn(v3; v4 ). In addition there must be a ba k
edge b into v4 , otherwise we may apply Lemma 3.7.5 and onsider G n (v1; v5). If
b is in ident to z1 then we may add b and e0 and remove a and (v4 ; v5 ) as shown
88
in Figure 3-28.2)d). Hen e we have one free edge left in whi h to onne t us up
to y2.
(e) Finally onsider the situation in whi h b is not in ident to z1 . The ba k edge
in ident to y2 must ome from v3 or below the 5- y le. Both ases are similar.
The former ase is shown in Figure 3-28.2)e). Add b and b0. Remove e and
(v3; v4 ).
1) y2 2)a) y1 y2 2)b) y1 y2
y1
e e e’
e v5 e’ v5
v5 v4 v4 v4
v1 v1 v1
v3 v2 v3 v2 v3
v2 a’ a’
a’ a a
a
z1 z2 z1 z2 z1
z2
2)d) y1 y2 2)e) y1 y2 4)a) y1 y2
e e b’ e
e’ e’ e’
v4 v4 v4
v1 v5 v1 v5 v1 v5
v2 v3 b v2 v3 b v2 v3
a’ a’ a’
a a a
z2 z1 z2 z1 z2 z1
y1 y2 y1 y2
4)b) y1 y2 4)d)i) 4)d)ii)
e e
e e’ e’ e’
v5 v4 v4
v4 b
v5 v1 v5
v1 v1
v3 v2 v3
v2 v3 v2
a’ a’ a
a’ a a
z1 z2 z1
z2 z1 z2
3. The ba k edge a0 is in ident to v4. This ase follows from the observation that we
89
may remove the edge (v3; v4 ).
(a) There is a ba k edge e0 to y2 from either v2 or v5. We show the later ase in
Figure 3-28.4)a).
(b) There is a ba k edge e0 to y1 from v2 . Add e0 , remove e and (v2; v3 ). Then add
an appropriate ba k edge to y2. See Figure 3-28.4)b).
( ) The situations in whi h there are edges (v2 ; v5) or (v2 ; v4) are easy to deal with.
(d) There is a ba k edge e0 to y1 from v5. If there is also a ba k edge b from v4 to
y2 then we pro eed as in Figure 3-28.4)d)i)). Suppose then that there is a ba k
edge b to y2 from v1 ; v2 or from below the 5- y le. These ases are similar. We
illustrate the last of these ases in Figure 3-28.4)d)ii).
(e) Suppose then the only ba k edge b into y2 omes from v3 . Here the situation
is more omplex. Noti e rst that removing v3 and y1 dis onne ts the graph
into two omponents C1 and C2. This is shown in Figure 3-29.4)e). Noti e that
this implies that in any optimal solution there is a path from y1 to v3 whose
internal verti es are all in C1 . In addition these internal verti es an only be
adja ent to other verti es on the path. Finally, note that, sin e v2 is of degree
2 any solution uses the edges (v1; v2) and (v2 ; v3). This for es some path of the
type shown in
Figure 3-29 to be in any optimal solution. Observe that this implies, by a
similar argument to that given in Lemma 3.7.4, that there is some optimal
solution whi h does not ontain the edge (v1; y1). Hen e we may onsider the
graph G n (v1 ; y1).
(f) Otherwise we apply Lemma 3.7.5 and onsider G n (v3; v4 ).
(III) There is a lower tree edge a in ident to v2 or v5. These ases are symmetri so
we will deal with the former ase.
90
4)e) C1
y1 y2
e
b
v1
v2 v3
C2 z2 z1
Figure 3-29: A tri ky 5 y le.
1. The ba k edge a0 is in ident to y2. Then we are already done. See Figure 3-30.1).
We may also remove the edge (v1 ; v2).
2. The ba k edge a0 is in ident to y1. Remove the edge (v1 ; v2). We have one spare
edge whi h we use to onne t up to y2. See Figure 3-30.2).
3. The ba k edge a0 is in ident to v1 or v3 . In either ase we may remove one edge
from the y le whi h gives us a spare edge to play with.
4. The ba k edge a0 is in ident to v5 . By onsideration of the ba k edges in ident to
y2 we get to remove either (v1 ; v2 ) or (v1 ; v5 ).
(a) The ases in whi h there is a ba k edge e0 to y2 from v2; v3 or v5 are all easy to
deal with.
(b) If there is a ba k edge e0 from below the 5- y le to y2 we then may add e0 and
remove (v1 ; v2).
( ) Suppose there is an edge e0 = (v1; y2) along with a ba k edge into y1 from any
of v2; v3 or v5 . These are dealt with in a similar manner to those in ase 5a).
Similarly if there is a ba k edge into y1 from below the 5- y le.
(d) Hen e the only vertex in ident to y2 is v4. Again the situations where there is
a ba k edge into y1 from any of v2 ; v3 or v5 are easy to deal with.
91
1) y2 2) y2
y1 y1
e e’ e e’
v5 v5
v1 v1
v4 v4
v2 v2
v3 v3
a’ a a’ a
z2 z1 z1
z2
5)f) y1 y2 5)g) y1 y2
e e
v5 e’ v5 e’
v1 v1
b v4 v4
v2
v3 a’
a b
z1
z1 z2 z2
(e) We use the usual methods to deal with the edges (v1; v3 ); (v3; v5) or (v2 ; v5), if
they are present.
(f) Suppose there is a ba k edge b into v5 . Then add b and e0, remove (v1; v2 ) and
(v4; v5 ). This is shown in Figure 3-30.5)f)).
(g) There must then be a ba k edge b into y1 sin e v5 is not a Beta-vertex. In
addition, note that v5 is of degree 2. Observe also that the removal of y1 and
v4 dis onne ts the graph. It follows, by a similar argument to the one given in
ase II)4)e), that we may onsider the graph G n (v1; y1). See 3-30.5)g)).
Theorem 3.7.7. The algorithm is a 34 -approximation algorithm for 2V C .
92
3.8 Strong Conne tivity
3.8.1 Tools and Operations
Let a be a dire ted y le onsisting of just two ar s, a = (u1; u2) and its
digon
omplement a = (u2; u1 ). We have the following lemma.
Lemma 3.8.1. Let G be a strongly onne ted graph ontaining at least three ver-
ti es. Then, if G ontains no ut verti es, there exists a minimum strongly onne ted
subgraph whi h does not ontain any digons.
Proof. Take a minimum strongly onne ted subgraph, S , ontaining the fewest pos-
sible number of digons. Suppose it ontains a digon a = (u1; u2) and a = (u2; u1).
Noti e there is no dire ted path from u1 to u2 in S n fag. Otherwise a is super uous
and S is not minimal. Similarly there is no dire ted path from u2 to u1 in S nfag. So,
onsider an out-tree, T1, rooted at u1 in S n fag. Clearly, u2 2= T1 . Similarly, u1 2= T2 ,
where T2 is an out-tree rooted at u2 in S n fag.
Now A = T1 \ T2 is empty. Otherwise A indu es a ut Æ(A) in whi h the only
ar s in S \ Æ(A) are all dire ted into A. Hen e S is not a strongly onne ted ar set.
Similarly V = T1 [ T2 . Suppose B = V n (T1 [ T2) is non-empty. Then S \ Æ(B )
onsists only of ar s dire ted out of B . Again, this ontradi ts the fa t that S is
strongly onne ted. So there are no ar s between T1 n fv1 g and T2 n fv2g. It follows
that both G n T1 and G n T2 are strongly onne ted when we onsider S restri ted to
G n T1 and S restri ted to G n T2 , respe tively. Now a and a are not the only ar s in
D onne ting one vertex in T1 and another in T2 . Otherwise u1 and/or u2 would be
ut verti es. So let a0 be an ar rossing from T1 to T2 . Without loss of generality we
may assume it is dire ted out of T1 and into T2 . In addition, it is in ident to exa tly
one of u1 or u2. Then S n fag [ fa0g is a minimum strongly onne ted subgraph with
one less digon than S , a ontradi tion.
Corollary 3.8.2. If an ar a is in an optimal solution to G then we may onsider
the graph G n fag instead.
93
Proof. If a is in an optimal solution then by Lemma 3.8.1 its omplement a (if it is
present in G) is not in some optimal solution. Hen e we may solve the problem on
the smaller graph G n fag.
We say that an ar a that is not present in some optimal solution is redundant.
If ever we nd a redundant ar then we may instead onsider the graph G n fag.
For example, if an ar is in a digon and we an show that its omplement is for ed
to be in every optimal solution then by Corollary 3.8.2 the ar is redundant. More
generally, redundant ar s may also o ur in the presen e of for ed paths. We say
that dire ted path, P , onsisting of the verti es v0; v1; : : : vr , is for ed if it is in some
optimal solution. In addition, suppose that we have a vertex u and ar s (u; v0) and
(u; vr ). Choosing the ar (u; v0) will always be preferable to (u; vr ). This follows as,
given that P is present, (u; v0) also reates a path from u to vk as well as from u
to v0. Hen e (u; vr ) is redundant and we may onsider the smaller graph G n (u; vr ).
Similarly, given su h a for ed path P and ar s (v0; u) and (vr ; u), we see that (v0 ; u)
is redundant.
Now onsider a y le in G onsisting of three ar s, (v0 ; v1); (v1; v2) and (v2 ; v0).
First, note that the nodes v0 ; v1 and v2 may indu e up to six ar s in G, the three
y le ar s and their omplements. Suppose there is an optimal minimum strongly
onne ted subgraph ontaining at least 2 of these six possible ar s. Then we may
hoose the three y le ar s, ontra t the y le, and nd a 32 -optimal subgraph on the
resultant graph. The following lemma shows that this approa h will be suÆ ient to
produ e a 32 -approximation fa tor.
Lemma 3.8.3. If three nodes indu e two ar s is some optimal solution then we may
hoose any 3- y le through the nodes and onsider the resultant sub-problem.
Proof. Let the optimal solution have opt ar s, two of whi h are indu ed by v0 ; v1 and
v2 . So k = opt 2 ar s are used on the ar set indu ed by G0 = G=fv0 ; v1 ; v2 g, the
graph formed by ontra ting the 3- y le. Let opt0 be the size of the optimal solution
to G0. Clearly opt0 k. So if we an nd 32 -optimal solution to G0 and then add the
three ar s in the y le we will have a solution to G. This solution will use at most
94
3 opt0 + 3 3 k + 3 = 3 (k + 2) = 3 opt ar s. Hen e we have a 23 -optimal solution to
2 2 2 2
G.
We will apply Lemma 3.8.3 in the following manner. If we ever nd a 3- y le that

indu es two ar s in some optimal solution then we will onsider the graph resultant
from ontra ting the y le.
There is one other operation that we will use in our algorithm, that of omponent
merging. Suppose we nd a y le C 0 in G that strongly onne ts together several
omponents (whi h do not in lude the root omponent). In su h ir umstan es we
may treat the group formed by su h additions as one mega- omponent. We may dene
a new omponent graph G 0 on the same original omponents, with the ex eption that
we will use the mega- omponent rather than its many base- omponents. This pro ess
of rening our omponent graph will be termed omponent merging. We will then
grow a new tree T 0 on G 0 and begin again. Here we must bear in mind that we must
also harge the ar set C 0 to the the mega- omponent, on top of any other ar s that
may subsequently be harged to it.
Note that ontra tions, deletions and omponent mergings an only be arried
out a polynomial number of times. As su h, repeatedly arrying out these operations
(and solving on the resultant graphs) will not ae t the polynomial nature of our
algorithm.
3.8.2 The Lo al Improvement Method
We are now ready to des ribe how to implement our version of the y le ontra tion
algorithm. Re all that we wish to eÆ iently ontra t a y le (or ir uit), in the
omponent graph, through the root and the target omponent, and repeat. Given
su h a y le, C in G , suppose that the ontra tion of C results in the ontra tion of a
set of verti es V 0 V . In order to show our algorithm has the desired performan e,
it is suÆ ient to prove the following theorem.
Theorem 3.8.4. Suppose that the algorithm alls for the ontra tion of a y le C .
Let this indu e the ontra tion of the vertex set V 0 . If V 0 indu es r ar s in our nal
95
solution then V 0 indu es at least 23 r ar s in the lower bound.
Proof. Our proof of Theorem 3.8.4 will be split into several parts. We onsider rst
the ase in where the y le, C in G , that we ontra t passes only through strongly
onne ted omponents. We then partition our proof depending upon whether the the
target omponent ontains at least 4 ar s or is a digon or 3- y le. Finally we examine
the situation in whi h the root omponent, target omponent or other omponent
ontained in C is a path.
So we wish to ontra t a y le through the root and the target omponent. In
addition, we desire the property that the verti es ontra ted in this pro ess indu e
enough ar s in the lower bound to over, with our 23 performan e guarantee, the num-
ber of ar s they indu e in the nal solution. We have seen that this is easy when
the target omponent is strongly onne ted and ontains at least four ar s, sin e this
gives us two extra ar s to play with. We now show that Theorem 3.8.4 holds for the
other types of target omponent.
Digons. Suppose the target omponent is a digon D. We have seen that it is

too ostly to harge 2 extra ar s to this omponent. Re all that we begin with the
assumption that all of the omponents whi h we en ounter in this se tion are strongly
onne ted. The situation in whi h some of the omponents may be paths will be dealt
with subsequently. We begin with some easy situations.
a) There is a ba k ar b emanating from the subtree T1 . Suppose b is in ident to v1 ,
see Figure 3-31a)i). In this ase we may add the ar b and remove the ar (v1 ; v0).
Noti e that the removal of the ar (v1 ; v0) implies that the net harge to the digon
is just one ar , giving our desired 23 fa tor. Suppose, instead, that b emanates from
a omponent C in T1 fv1g, where T1 is the subtree in T rooted at v1 . See Figure
3-31a)ii). In this ase we may add the ar b and remove the ar (v1 ; v0). Along with
the appropriate tree ar s this reates a y le C to ontra t. This y le also onne ts
into the root the all omponents between D and C in the subtree T1 . Note ea h su h
omponent re eives a net harge of one ar . Sin e these omponents ontain at least
two ar s, we have at worst a 32 blow-up fa tor.
96
a)i) ii)
r r
b b
t t
v1
D D
v0 v1 v0
b)i)
r ii) r
b t b t
v1
D D
v0 v1 v0
Figure 3-31: A target digon.
b) There is a ba k ar b emanating from v0. Now sin e v0 is not a ut vertex there

must be a path from r to v1 that does not pass through v0 . This path may onsist
of a single ar from r to v1, see Figure 3-31a)i). The general ase is shown in Figure
3-31a)ii). Combining this path with the ar b allows us to drop the ar (v0 ; v1) and
still reate a y le C . Again the net harge to the digon (and any other omponent
in the y le) is just one.
Observe that Figures 3-31a)ii) and b)ii) generalise Figures 3-31a)ii) and b)ii).
In parti ular, Figure 3-31a)ii) ontains a path from v1 to the root passing through
various omponents of G . In Figure 3-31a)i) this path is a trivial path onsisting of a
single ar from v1 to the root. The situation with regards to Figures 3-31b)i) and ii) is
similar, with the path, this time, being dire ted from the root to v1. Note that (sin e
we are assuming that ea h node in G is strongly onne ted and ontains at least two
ar s) if non-trivial paths are found then the harge asso iated with ea h additional
omponent is just one ar , giving a blow-up fa tor of at most 32 . Observe that the
general ase of non-trivial paths is just a ombination of the trivial path ase and a
path through these additional omponents. Hen e, it will be the ase that the trivial
path ase provides a worst ase analysis. If we an show that the blow-up asso iated
with the trivial path ase is at most 23 , then it will also follow for the general ase.
97
As a result, unless otherwise stated, we will make this worst ase assumption that
onne tions between spe i verti es will be via a single ar rather than via a path
in the omponent graph.
) The only possible situation remaining is the ase of a ba k ar b going from T0 nfv0 g
to the root omponent. Again, for a worse ase s enario we may assume it emanates
from C0 a hild omponent of v0 in T0 . In su h ir umstan es we will be for ed to
ontra t into the root several omponents together. We will nd a ir uit, rather than
a y le, to ontra t. In parti ular we will ontra t two y les (one of whi h does not
ontain the root omponent) and argue that the pro ess an be done eÆ iently.
Now suppose there is a ba k ar path from v1 to v0 , or vi e versa, via C1. In the
worst ase this path onsists of just two ar s. See Figure 3-33a). So by adding both
these path ar s whilst removing (v0 ; v1) we onne t up, into the root omponent, the
digon and C0 and C1. Thus if C0 and C1 both ontain at least three ar s then we use
12 ar s in total for the three omponents. Sin e these three omponents ontain 8
ar s this gives a 23 blow-up. Noti e that we do indeed ontra t a ir uit whi h onsists
of two y les that interse t at D.
What, though, if C0 or C1 are themselves digons? We will eliminate digons via
a re ursive pro ess. In order to show how this will be done we rst introdu e some
more notation. Given a digon Dj , label the verti es of the digon as follows. Let the
vja 2 Dj be in ident to ar s that urrently onne t TD to G n TD . Let the vjb be the
j j
other vertex in Dj . We say that a hild omponent C of Dj is a rung if there is a path

in the subtree TD from vja to vjb, or vi e versa, that passes through C . In the worst
j
ase this path will ontain just the single omponent C . In this ase we will label the
path ar s p1 and p2, see Figure 3-32a) and b). Note that the use of su h paths allows
the removal of either the ar (vja ; vjb) or the ar (vjb; vja). Hen e the path forms a
3- y le through the digon Dj , and its three ar s will be harged to the digon. This
observation will be used in a form of re ursion that enables us to \ignore" digons.
In addition, we say that a rung C is a base rung if either C ontains at least 3 ar s
or if C is a digon Dj+1 and the path ar s are in ident to dierent verti es in Dj+1.
Noti e that in the latter ase there is a 4- y le through the verti es of Dj and Dj+1.
98
a) b) t Dj c) t Dj
t Dj
vja vjb vja vjb
vja vjb
p1 p1 p2 p1
p2 p2
C C
D j+ 1
Figure 3-32: Rungs.
One su h possibility is shown in Figure 3-32 ). Should we nd a digon Dj+1 that

is a base rung and a hild of another digon Dj we will then perform our omponent
merging operation. We repla e these two digon omponents in the omponent graph
G by the 4- y le that passes through them.
a) b)
r r
b
t
D C0 = D0 D
C0 t’ v1 v1
v0 v0
C1 D1
D2
Figure 3-33: Growing a Pair of Digon Ladders
Thus in the previous ase, where we assumed that C0 and C1 both ontain at least
three ar s, the omponent C1 is a base rung hild of the digon D (and C0, whilst not
te hni ally a base rung, plays a similar role).
We now illustrate the basi re ursive argument that will be applied on a spe i
example.. Let C1 be a digon D1. Hen e C1 is a rung. Suppose also that it is not a
base rung. Suppose there are ar s p1 = (v1 ; v1a) and p2 = (v1a ; v0). The reader may
observe that were we to ontra t the 3- y le formed by p1 ; p2 and (v0; v1 ) then we
would essentially be in the same situation as in Figure 3-33a) but with the digon D
99
repla ed by the digon D1. Suppose then that D1 has a hild omponent C2 that is a
base rung. Now if C2 ontains at least three ar s then the blow-up fa tor of merging
the ir uit illustrated is again at most 23 . If C2 = D2 is a digon that is a base rung
then there is a 4- y le through D1 and D2 and we onsider a new omponent graph
G 0 with the 4- y le as a node. There is one nal ase to onsider. If the digon D2
has a ba k ar that goes beyond D1 and into D then it is easy to verify that su h a
ir umstan e an only be bene ial.
We now return to the general ase and formalise our arguments. We will su es-
sively eliminate digons, until we rea h a base rung. Assume that C0 is a digon and
onsider the pi ture shown in Figure 3-33b).
We take a hild omponent of D that is a rung (if there are many su h hild
omponents we may hoose the one with the highest priority in T ). If this omponent
is a base rung then we stop. Otherwise, if this rung is a digon D1 that is not a base
rung, we re urse on the hildren of D1. This pro ess forms a \ladder of digons",
D; D1 ; D2 ; : : : Dp. Similarly we grow another ladder of digons, by the same method,
beneath D0. Su h ladders are shown in Figure 3-33b). If both of these pro esses
terminate su essfully then we have a guarantee of the desired blow-up fa tor.
So suppose one of the ladders does not terminate with a base rung. Consider the
last digon Dp in the ladder. Then either Dp has no hildren or it has no hild
omponents that are rungs. Consider the former ase. Now a ba k ar b must
emanate from vpb otherwise the ar (vpb; vpa) is for ed and we may onsider the graph
G n (vpa ; vpb). In addition, the head of b must also be in ident to the tail of p1 otherwise
Dp is a base rung. Again its is only bene ial if b goes beyond Dp 1. Now there must
be a down ar d entering into vpb, otherwise the ar (vpa; vpb) is for ed and we may
onsider the graph G n (vpb; vpa). In the worst ase d emanates from the digon Dp 1.
Here we may onstru t a 4 y le through the verti es of Dp 1 and Dp. This is shown
in Figure 3-34a) and b). We then apply the merging omponents method, using the
4- y le instead of the two digons.
Similar arguments apply in the latter ase where Dp has hild omponents but
none are rungs. Hen e the ladder growing pro ess terminates and we are done.
100
a) Dp-1 b) Dp-1
p2 p2
p1 p1
d
b d b
vpb vpa vpb vpa
Dp Dp
Figure 3-34: Merging digons.
Finally, note that any omponent that is in ident to an ar a that emanates from
the root omponent may be hosen as a target node. This follows as we may for e a
to be hosen as the rst ar in T . In fa t, any omponent that is adja ent to the root
(via an in-ar or an out-ar ) may be hosen as the target node as we may as well have
hosen T to be an in-tree rooted at r. By symmetry, the arguments that follow with
respe t to an out-tree T would also apply to an in-tree. This observation allows us
to now assume that no omponents adja ent to the root omponent are digons (or,
in fa t, strongly onne ted with at least 4 ar s).
3-Cy les. Let our target omponent be a 3- y le. Again, we will assume that
none of the omponents whi h we en ounter in this se tion are paths. In addition,
should we en ounter a omponent that represents a digon then the \ladder of digons"
argument may be applied in the same manner as before. This observation allows us
to assume that all of the omponents whi h we en ounter are strongly onne ted and
ontain at least three ar s.
We may assume that all y les in omponent graph G that pass through the root
omponent and the target omponent are digons. Suppose otherwise and su h a y le
passes through three or more omponents. The two omponents adja ent to the root
omponent ontain at least three ar s ea h. The blow-up fa tor is then at most 23 .
This fa tor o urs for the ase in whi h there are three omponents (i.e. three ar s are
added) and the non-root omponents are both 3- y les. This is illustrated in Figure
3-35i).
Let the 3- y le onsist of the nodes v0; v1 and v2 with the tree ar , from the root
101
i) r ii)
r
v0 v0
b
v1 v2 b
v1 v2
u0
u1 u2
Figure 3-35: Some 3- y les.
to the y le, being in ident to v0 . By the arguments given above, there must be a
ba k ar b emanating from the 3- y le. If b is in ident to v2 then we an hoose b and
remove the ar (v2 ; v0). This situation is shown in Figure 3-35ii). Thus, we need only
4 ar s in total and they an all be a ounted for by the 3 ar s in the y le. Hen e,
sin e v0 is not a ut vertex there must be a ba k ar b emanating from v1. This
implies that there is no ar from the root omponent to v2, otherwise we may omit
the ar (v1; v2) in a similar manner to that of Figure 3-35ii).
There is another relatively straightforward situation. Suppose a tree ar t em-
anates from the node v2 and that it leads to a hild omponent C2 . If for su h a
subtree, T2, there emanates a ba k ar b0 into v0 then we an onne t up the 3- y le
and omponent C2 for a ost of 3 ar s. This follows as in hoosing d and b0 we may
remove ar (v2 ; v0). Sin e C2 ontains at least 3 ar s this gives the desired blow-up
fa tor. Note that if b0 emanates from a omponent lower than C2 in T2 then this only
helps us, as we onne t up any additional omponent for a harge of just one ar .
This situation is shown in Figure 3-36i). The orresponding ases for subtrees T1 and
T0 , with hild omponents C1 and C0 , are shown in Figures 3-36ii) and iii).
i) r ii) r iii) r
v0 v0 v0
b’ v1 v2 t
v1 v2 t v1 v2
C0
C2 t
b’ b’
C1
Figure 3-36: More 3- y les.
102
Now onsider the remaining possibilities. Re all there is a ba k ar b out of v1 . By
the ordering given by DFS method, there are no ar s from verti es in T0 to verti es
in either T1 or T2. There are also no ar s from verti es in T2 to verti es in T1 .
a) There is a ba k ar b0 oming into v2 from T0 . Choosing b0 allows us to drop the
ar (v1 ; v2). The net harge is 4 1 = 3 ar s whi h we harge to the y le and C0.
See Figure 3-37a). Sin e we an deal with the ase of ba k ar s entering into v1 , or
going beyond the 3- y le, we may now assume that for all the C0, every ba k ar goes
into v0.
b) There is a ba k ar b0 oming into v1 from T2. This situation is dealt with analo-
gously to the previous part, and is shown in Figure 3-37b).
a) r b)
r
v0
t v0
C0
v1 v2
v2 b’ v1
t
b’
C2
Figure 3-37: Yet more 3- y les.
) There is an ar a leaving T2 and entering T0 . Sin e the only ba k ar s leaving T0

enter into v0 itself, a reates a path between T2 and v0. Thus a a ts like an ar entering
v0 and an be dealt with in the same way (Figure 3-36i)), allowing the removal of
(v2 ; v0). Hen e we may assume that for all the C2, all the ba k ar s go into v2 .
d) There is an ar a leaving T1 and entering T2 . Now a a ts like an ar entering v2
and an be dealt with in the same way (Figure 3-36ii)), thus allowing the removal of
(v1 ; v2).
e) In what follows we are interested in the possible routes from v1 to v2 . Suppose
there is an ar a from v2 to a vertex u 2 T0 n fv0g. Now in any optimal solution there
must be a dire ted path P from u to v0 onsisting only of verti es in that subtree. As
su h, there must be an optimal solution that does not ontain the ar (v2 ; v0). This
follows, from a for ed path argument, as the ar a will also onne t v2 to v0 via P .
103
Using a then is at least as good as using (v2; v0) sin e it also provides a path from v2
to the verti es of P . Hen e we may remove the ar (v2; v0 ) and resolve the problem
on G n (v2; v0). By a similar argument there is no ar a from v1 to a vertex in T2 .
f) Otherwise the only way to rea h v2 from v1 is via the ar (v1 ; v2), or if it is present
(v0 ; v2). In addition, the only way up the tree from v2 is via the ar (v2 ; v0) or, if it
is present, (v2 ; v1). Hen e, at least two ar s indu ed by the verti es v0 ; v1 and v2 are
for ed to be in an optimal solution. So by Lemma 3.8.3 we may ontra t the 3- y le
and re urse on the smaller graph.
Paths. We now onsider the situation in whi h a omponent may represent a path P .
There are three ases to analyse. The ases where the path is the root omponent, the
target omponent or a omponent in a ontra ted y le C asso iated with a dierent
target omponent. Let P onsist of the verti es vk ; vk 1; : : : v0 with k ar s of the
form (vi+1; vi). We will draw the path so that the ar s are dire ted from right to
left. We will all an ar (vi; vj ); i r < j; a forward ar wrt vr . We all an ar
(vj ; vi); i r < j; a ba k ar wrt vr . We aim to onne t up the path with at most a
fa tor 23 blow-up. To do this we will start at the leftmost vertex of the path and add
ar s to strongly onne t it up to its immediate neighbours. We ontinue this pro ess
rightwards until all the verti es on the path are strongly onne ted. An example will
be of use here, see Figure 3-38a). Let v0 be the leftmost vertex. Suppose there is a
forward ar d going to v2 . Then adding d strongly onne ts v0 ; v1 and v2 . Now these
three verti es indu e two path ar s, (v2 ; v1) and (v1 ; v0). We may harge d to these
two ar s to give a blow-up fa tor 32 . We may then think of v0 ; v1 and v2 as being
ontra ted together to form a super-node on the path. Thus our general pi ture
remains the same as that of Figure 3-38a) after ea h step.
We will show that the harge asso iated, in strongly onne ting up the path, is
at most half of the length of the path. Consider the rst ase, where the omponent
representing the path is the root omponent. Hen e we only need to ensure that the
verti es on the path be ome strongly onne ted, and not worry about onne ting the
path upwards into the root omponent. We will also assume that for ea h Ti the
104
Ci will represent strongly onne ted omponents. Again, the ase where the Ci are
paths will be dealt with later.
During the following onne tion pro ess the vertex v0 may represent multiple
nodes but v1; v2 : : : vk an not. From now on, let d be the forward ar rea hing deepest
along the path from v0 . Sin e v0 may represent multiple nodes, there may be several
ar s rea hing the same node. Whether the hoi e between the ar s is important will
be made apparent. Note also that, by the DFS ordering, there are no ar s from a
vertex in Ti to a vertex in Tj , where i < j .
b) v1 v2 v3
d v0
a)
a
v0 v1 v2 v3 d
T0
c) d v1 v2 v3
v0
a
b
T1
Figure 3-38: Some paths.
a) Take v0 and onsider d. If d goes to v2 or beyond then we add d and harge it to

the (at least) two path ar s. See Figure 3-38a).
b) So suppose d only rea hes v1 . Now if d omes from T0 then we onsider the
omponent C0 . Suppose that C0 ontains at least three ar s. Now adding d strongly
onne ts together v0 , v1 and T0 . See Figure 3-38b). We need to harge the tree ar
a0 and d. We harge them to C0 and (v1 ; v0 ). Hen e we harge these two ar s to at
least four ar s ontained in the lower bound, a 32 fa tor. Suppose instead that C0
is a digon. Here we may apply the \ladder of digons" argument to obtain the same
fa tor.
) There is a ba k ar b from T1 into T0 . See Figure 3-38 ). We may add d, b and the
tree ar a1 whilst removing (v1; v0). Similar arguments to those of part b) now apply.
d) Otherwise unravel our pro ess one step. For larity we use negative indi es to label
the verti es of the path after unravelling the pro ess. If the previous deepest ar d0
105
overed at least 3 path ar s then we an harge d and d0 to the (at least) four path
ar s.
e) If the previous deepest ar d0 overed just one path ar then d is a for ed ar as it
is the only way in whi h to onstru t a dire ted path from v0 to v1. Hen e we may
onsider the graph G n (v1 ; v0).
f) So assume that d0 overed 2 path ar s. In addition, there is an ar a from v 1 to
v1 . Otherwise d is for ed and we again onsider G n (v1 ; v0 ). See Figure 3-39f). Note
also that sin e v1 is not a ut vertex, there must be a dire ted path P 0, that does not
ontain v1 , from a vertex vr ; r 2 to a vertex vs; s 0. First assume that P 0 onsists
of a single ar b between verti es in P .
f) d’ a d’ a
f)i)
v-2 v0 v1 v2 v-2 v0 v1 v2 vr
v-1 v-1
d d
d’ a f)iii) d’ a
f)ii)
v-2 v0 v1 v2 vr v-2 v0 v1 v2 vr
v-1 v-1 d
d
b b
Figure 3-39: Unravelling a path.
i) If there is an ar b into v 2 then use the ar s b; d0; (v0; v 1) and a. See Figure
3-39f)i). These four ar s are overed by the lower bound of the three path ar s
from v 2 to v1 . We don't yet have a path from v1 to v0 but later in onne ting
up the rest of the path we must nd a dire ted path from v1 to vr . This will
omplete a y le through v1 and v0. In fa t, the next stage in the pro ess will
then be similar to our basi ase ex ept that, in ee t, we get the edge b for free
(sin e it has already been harged to the ar s on the path between v 2 and v1 ).
This, of ourse, an only help us and our iterative pro ess ontinues as before.
ii) If there is an ar b into v 1 then we pro eed as in Figure 3-39f)ii) and again use
4 ar s asso iated with the three path ar s.
106
iii) Finally suppose that all the type b ar s are in ident to v0. See Figure 3-39f)iii).
Consider one su h ar (vr ; v0). Now the only way to rea h vr from a vertex
vs ; s 0 in the path P is via the vertex v1 . Hen e in any optimal solution
there is a dire ted path from v1 to vr that does not use any of the verti es
vs ; s 0. Hen e, by a for ed path argument, using the ar (vr ; v0 ) will always
be preferable to using the ar (v1 ; v0). Thus (v1 ; v0) is a redundant ar and we
may onsider the graph G n (v1; v0).
Now onsider the ase where P 0 does not onsist of a single ar but goes through
other omponents in the omponent graph G . The arguments above still apply with
the other omponents in ident to P 0 being harged just one ar .
It follows that in the spe i ase where the path is the root omponent we use
at most an additional b k2 ar s.
Now onsider the general ase where P is a non-root omponent i.e. it is the
target omponent or is a omponent in a y le C ontra ted for a dierent target
omponent. Let P be the root of a subtree T . For a worst ase analysis, we may
assume that there are no ba k ar s from a omponent in T n P to G n T . First remove
the DFS tree ar onne ting G n T and P . Add an ar a from P to G n T with the
property that a originates as far to the left on P as possible, say from vj . Figure
3-40i) shows the resulting situation. Here the other possible ar s onne ting P to
G n T are labelled b; b0 and d. This notation is used as the b ar s may be used as ba k
ar s and the d ar s may be used as forward ar s as before. The b0 ar s are slightly
dierent. Adding su h an ar does not reate a \ba k path" via G n T unless an ar
from G n T to P (i.e. a type b or type d ar ) has already been hosen.
This allows us to pro eed as before adding forward and ba k ar s to strongly
onne t up the path. We make one proviso. Should a b0 ar from P to G n T be
required as a ba k ar , and no ar from G n T to P has urrently been sele ted, then
we add an appropriate ar and remove a. An example where the addition of b0 leads
to the removal of a and addition of b is shown in Figure 3-40ii). Note that su h a
method still leads to a net harge of one when b0 is added.
The previous arguments, used for the spe i ase of the path being the root
107
i) G-T ii) G-T b’
b b’
d b a
a
v0 vj v0 v1 v2 v3
Figure 3-40: A non-root path.
omponent, may now be applied to this general ase. Hen e b k2 ar s are required to
internally strongly onne t the path. Now we have two possible situations after we
have internally onne ted up the path.
Type I) No ar s between P and G n T were hosen in the pro ess (ex ept for our
original hoi e of a). In this ase we need to add an extra ar a0 to onne t the path
upwards. So our total ost is at most b k2 +2 (in luding one ea h for a and a0 ). Noti e
that a and a0 play no part in internally onne ting up the path. This observation will
be of use later.
Type II) Some ar s, apart from a, between P and G n T were hosen in the pro ess.
These ar s are in luded in the b k2 harge to the path but they also su eed in strongly
onne ting the path upwards. Hen e the harge is b k2 + 1 (in luding one for a).
Re all that asso iated with ea h path were two \free" ar s and so we a tually have
k + 2 ar s to harge these additional ar s to. It follows that if the path is a target
omponent then the total number of ar s used is at most k + b k2 + 2 < 23 (k + 2), as
required.
So onsider the ase in whi h the path is a omponent in a y le C for whi h the
target omponent was a digon or 3- y le. Again asso iated with the target path are
two free ar s. Thus, we have at least three ar s (due to the 23 performan e guarantee
one extra ar is also allowed with the two free ar s) whi h we an use in strongly
onne ting up a path omponent. It now follows easily that it would only have been
bene ial to en ounter any paths in any of the ir uits we found in the previous
se tions, where the target omponents were digons and 3- y les.
Finally, onsider the ase in whi h the path is a omponent in a ir uit for whi h
the target omponent is another path. Re all that this o urs when we reallo ate path
108
ar s to a hild omponent Ci below. We postponed the dis ussion of the situation in
whi h these hild omponents were themselves paths. We deal with this topi here.
Call the upper path P 0 and the lower path P . In this situation we allo ate one ar
from P 0 to the P and for e two other ar s to be hosen that strongly onne t upwards
P . A total of three extra ar s. Now, if the lower path, P , is of type I) the ar s a and
a0 are super uous so the harge to P is b k2 + 3. Whereas, if P is of type II) then the
ar a was needed in onne ting up the path so with these extra three ar s we have a
harge of b k2 + 4. Given that we also reassigned one ar from P 0 to P we now have
k + 3 ar s asso iated with P . Now sin e P has a single parent su h extra ar s an be
reallo ated from P 0 to P at most on e. This ompletes the proof that the blow-up
fa tor for paths is at most 32 sin e k + b k2 + 4 < 23 (k + 3). The theorem follows.
Corollary 3.8.5. The algorithm is a 23 -approximation algorithm for the minimum
strongly onne ted subgraph problem.
109
110
Bibliography
[1℄ V. Auletta, Y. Dinitz, Z. Nutov and D. Parente, \A 2-approximation algorithm

for nding an optimum 3-vertex- onne ted spanning subgraph", Journal of Al-
gorithms, 32, pp21-30, 1999.
[2℄ J. Cheriyan, T. Jordan and Z. Nutov, \On rooted node- onne tivity problems",
Algorithmi a, 30, pp353-375, 2001.
[3℄ A. Czumaj and A. Lingas, \On approximability of the minimum- ost k- onne ted
spanning subgraph problem", Pro . 10th Annual SODA, pp281-90, 1999.
[4℄ R. Carr and R. Ravi, \A new bound for the 2-edge onne ted subgraph problem",
Integer programming and ombinatorial optimization, Springer-Verlag, Le ture
Notes in Computer S ien e, 1412, pp112-25, 1998.
[5℄ J. Cheriyan, A. Sebo and Z. Sigeti, \Improving on the 1.5-approximation of
a smallest 2-edge onne ted spanning subgraph", SIAM J. Dis rete Math, 14,
pp170-180, 2001.
[6℄ J. Cheriyan and S. Vempala, \Edge overings of setpairs and the iterative round-
ing method", Pro . 9th Annual International IPCO, pp30-44, 2001.
[7℄ J. Cheriyan, S. Vempala and A. Vetta, \Approximation algorithms for minimum-
ost k-vertex onne ted subgraphs", to appear Pro . 34th Annual STOC, 2002.
[8℄ J. Cheriyan, S. Vempala and A. Vetta, \Network design via iterative rounding
of setpair relaxations", submitted for publi ation, 2001.
111
[9℄ C. Fernandes, \A better approximation ratio for the minimum k-edge- onne ted
spanning subgraph problem", Pro . 8th Annual SODA, pp629-638, 1997.
[10℄ A. Frank, \Augmenting graphs to meet edge- onne tivity requirements", SIAM
J. Dis rete Math., 5, pp25-53, 1992.
[11℄ A. Frank, \Conne tivity augmentation problems in network design", in J. Birge

and K. Murty (eds.), Mathemati al Programming: State of the Art 1994, Uni-
versity of Mi higan, Ann Arbor, pp34-63, 1994.
[12℄ A. Frank and T. Jordan, \Minimal edge- overings of pairs of sets", Journal of
Combinatorial Theory, Series B, 65, pp73-110, 1995.
[13℄ A. Frank and E. Tardos, \An appli ation of submodular ows", Linear Algebra
and its Appli ations, 114/115, pp329-348, 1989.
[14℄ L. Fleis her, K. Jain, and D. Williamson, \An iterative rounding 2-approximation
algorithm for the element onne tivity problem", Pro . 42nd FOCS, pp339-47,
2001.
[15℄ H. Gabow, \A representation for rossing set families with appli ations to sub-
modular ow problems", Pro . 4th Annual SODA, pp202-211, 1993.
[16℄ H. Gabow, \Using expander graphs to nd vertex onne tivity", Pro . 41st
FOCS, pp410-20, 2000.
[17℄ M. Garey and D. Johnson, Computers and Intra tability : A Guide to the Theory
of NP- ompleteness, Freeman, San Fran is o, 1979.
[18℄ N. Garg, A. Singla and S. Vempala, \Improved approximations for bi onne ted
subgraphs via better lower bounding te hniques", Pro . 6th Annual SODA,
pp103-11, 1993.
[19℄ M. Goemans, \Worst- ase Comparison of Valid Inequalities for the TSP", Math-
emati al Programming, 69, pp335-49, 1995.
112
[20℄ M. Goemans, A. Goldberg, S. Plotkin, D. Shmoys, E. Tardos and D. Williamson,
\Improved approximation algorithms for network design problems", Pro . 5th
Annual SODA, pp223-232, 1994.
[21℄ M. Goemans and D. Williamson, \A general approximation te hnique for on-

strained forest problems", SIAM Journal on Computing, 24, pp296-317, 1995.
[22℄ M. Grots hel, L. Lovasz and A. S hrijver, Geometri algorithms and ombinato-
rial optimization, Springer-Verlag, New York, 1988.
[23℄ D. Hartvigsen, \Extensions of Mat hing Theory", Ph.D. thesis, GSIA, Carnegie
Mellon University, 1984.
[24℄ H. Hsu, \An Algorithm for Finding a Minimal Equivalent Graph of a Digraph",
Journal of the Asso iation for Computing Ma hinery, 22(1), pp11-16, 1975.
[25℄ K. Jain, \A fa tor 2 approximation algorithm for the generalized Steiner network
problem", Combinatori a, 21(1), pp39-60, 2001.
[26℄ K. Jain, I. Mandoiu, V. Vazirani and D. Williamson, \A primal-dual s hema
based approximation algorithm for the element onne tivity problem", Pro .
10th Annual SODA, pp484-489, 1999.
[27℄ S. Khuller and B. Raghava hari, \Improved approximation algorithms for uni-
form onne tivity problems", Journal of Algorithms, 21, pp434-450, 1996.
[28℄ S. Khuller, B. Raghava hari and N. Young, \Approximating the Minimum Equiv-
alent Digraph", Pro . 5th Annual SODA, 41(4), pp859-872, 1995.
[29℄ S. Khuller, B. Raghava hari and N. Young, \On Strongly Conne ted Digraphs
with Bounded Cy le Length", Dis rete Applied Mathemati s, 69(3), pp281-289,
1996.
[30℄ S. Khuller and U. Vishkin, \Bi onne tivity approximations and graph arvings",
Journal of the ACM, 41, pp214-35, 1994.
113
[31℄ G. Kortsarz and Z. Nutov, \Approximating node onne tivity problems via set
overs", in K. Jansen and S. Khuller (eds.), Approximation Algorithms for Com-
binatorial Optimization, Springer-Verlag, Le ture Notes in Computer S ien e,
1913, pp194-205, 2000.
[32℄ W. Mader, \Endli hkeitsatze fur k-kritis he Graphen", Math. Ann., 229, pp143-
153, 1977.
[33℄ W. Mader, \Conne tivity and edge- onne tivity in nite graphs", Surveys in
Combinatori s, London Mathemati al So iety Le ture Note Series, 38, Cam-
bridge University Press, pp66-95, 1979.
[34℄ W. Mader, \On k- riti ally n- onne ted graphs", in J. Bondy and K. Murty
(eds.), Progress in Graph Theory, A ademi Press, Toronto, pp389-398, 1984.
[35℄ C. Monma, B. Munsun and W. Pulleyblank, \Minimum-weight two- onne ted
spanning networks", Math. Programming, 46, No.2 (Series A), pp153-71, 1990.
[36℄ D. Moyles and G. Thompson, \An Algorithm for Finding a Minimum Equiva-
lent Graph of a Digraph", Journal of the Asso iation for Computing Ma hinery,
16(3), pp455-460, 1969.
[37℄ R. Ravi and D. Williamson, \An approximation algorithm for minimum- ost
vertex- onne tivity problems", Algorithmi a, 18, pp21-43, 1997.
Erratum, submitted to Algorithmi a, 2001.
[38℄ E. Tardos, \A strongly polynomial algorithm to solve ombinatorial linear pro-
grams", Operations Resear h, 34, pp250-256, 1986.
[39℄ S. Vempala and A. Vetta, \Fa tor 43 Approximations for Minimum 2-Conne ted
Subgraphs", in K. Jansen and S. Khuller (eds.), Approximation Algorithms for
Combinatorial Optimization, Springer-Verlag, Le ture Notes in Computer S i-
en e, 1913, pp262-73, 2000.
114
[40℄ A. Vetta, \Approximating the minimum strongly onne ted subgraph via a
mat hing lower bound", Pro . 12th Annual SODA, pp417-26, 2001.
115

Thesis

Uploaded by

Copyright:

Available Formats

Thesis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thesis

Uploaded by

Copyright:

Available Formats

Graph Conne

tivity: Relaxations and Algorithms

Submitted to the Department of Mathemati s

2 Vertex Conne tivity in Weighted Graphs 13

2.1 Setpair Relaxations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3.1 Two Simple Approximation Algorithms . . . . . . . . . . . . . . . . . 47

Vertex Conne tivity in Weighted

2.1 Setpair Relaxations

2.1.1 The Cut Covering Model

Now onsider the following fundamental theorem of Menger.

2.1.2 The Setpair Model

A more general integer programming formulation, developed by Frank and Jordan

2.2.1 The k-Out onne ted Subgraph Problem

Proof. In the dire ted version Gb of G, ea h edge e of G is repla ed by two oppositely

2.2.2 Augmentation Algorithm for k-Vertex Conne tivity

Figure 2-1: An Example

Figure 2-2: Phase I

Figure 2-3: Phase II

Suppose that Hi has p edges in Æ(W ). Then

Thus if p  ` + 1 jW 0j we are done. Moreover, sin e Hi is `-vertex onne ted,

Lower Bounds on the Integrality Gap

2.3.1 Setpairs and Skew Bisupermodularity

Two setpairs W and Y are said to be omparable if

Figure 2-4: Comparable setpairs.

Two setpairs W and Y are said to be non- rossing if they are

Figure 2-5: Crossing setpairs.

Bisubmodular Fun tions.

Given two setpairs W = (Wt ; Wh) and Y = (Yt; Yh), let W

A non-negative, integer-valued fun tion f on S is alled rossing bisupermodular if

Skew Bisupermodular Fun tions.

A non-negative, integer-valued fun tion f on S is alled skew bisupermodular if for

2.3.2 Chara terising a Basi Solution via a Non-Overlapping

Proof. Assume that f (W )+ f (Y )  f (W

Hen e, all the inequalities hold as equations, therefore, W

The main result of this se tion is as follows.

Hen e, the fun tion f is rossing bisupermodular (if either W

Fo us on a nonzero basi solution x. By Theorem 2.3.3, x orresponds to a

Proof of Theorem 2.3.5.

so   n ). We have jLj = m   , hen e, by by Lemma 2.3.8, L has at least m6

setpairs of one of the types I, II, or III.

x for whi h ea h edge e has xe 

2.3.4 Iterative Rounding and Element Conne tivity

Proof. We need to show that the fun tion f = g p is skew bisupermodular. We

Observe that ea h of the terms g(W ), g(Y ), g(W

Figure 2-6: An illustration for skew bisupermodularity.

the edge representing g(W

that (p; q) 2 Æ(W ).

(Æ(W ) O (Æ(Y 1) [ [ Æ(Y `))) n (Yt1; : : : ; Yt`)

either (p; q ) or (q; p) is assigned to W by Rule 1 or Rule 2.

Figure 2-7: Ar assignment rules.

ar a1 in Figure 2-7 is an example.℄

Proof. Let x 6= 0 be a basi solution su h that xe < 12 for ea h edge e. Then

Bi-Conne tivity in Unweighted

3.1.1 The DFS Tree Method

3.1.2 The Cy le Contra tion Method

The se ond algorithm is a \ y le- ontra tion" algorithm. Find a y le C in G that

3.2 The Lower Bound

3.2.1 The Lower Bound in Undire ted Graphs

3.2.2 The Lower Bound in Dire ted Graphs

Thus if p ` + 1 jW 0j we are done. Moreover, sin e Hi is `-vertex onne ted,

Proof. Assume that f (W )+ f (Y ) f (W

so n ). We have jLj = m , hen e, by by Lemma 2.3.8, L has at least m6

x for whi h ea h edge e has xe

the edge representing g(W

(a) k10 + k20 + 2 opt(G) k10 + k20 + 3

Theorem 3.3.5. Given a 2-edge onne ted graph, l(G) opt(G).

(ii) l(G1) l(G01) + 3: thus we have