kernelization and complexity results for connectivity ... · np-completeness of both problems if...

To appear in Networks, 2009

Kernelization and Complexity Results for Connectivity

Augmentation Problems∗

Jiong Guo†and Johannes Uhlmann‡

Abstract

Connectivity augmentation problems ask for adding a set of at most k

edges (called links) whose insertion makes a given graph satisfy a spec-ified connectivity property, such as bridge-connectivity or biconnectiv-ity. A bridge-connected (biconnected) graph is a connected graph thatdoes not possess an edge (a vertex) whose removal results in a discon-nected graph. We show that, for bridge-connectivity and biconnectivity,the respective connectivity augmentation problems admit problem kernelswith O(k2) vertices and links. Moreover, we study partial connectivityaugmentation problems, naturally generalizing connectivity augmentationproblems. Here, we do not require that, after adding the edges, the en-tire graph should satisfy the connectivity property, but a large subgraph.In this setting, three polynomial-time solvable connectivity augmentationproblems behave differently, namely, the partial bridge-connectivity aug-mentation problem and the partial biconnectivity augmentation problemremain polynomial-time solvable whereas the partial strong connectivityaugmentation problem becomes W[2]-hard with respect to k.

Keywords: Combinatorial problem, NP-complete problem, graph algo-rithm, fixed-parameter (in)tractability, data reduction.

1 Introduction

Connectivity augmentation problems on undirected and directed graphs haveas input a graph G = (V, E), a set E′ of edges, and a non-negative integer k,and ask for a set E′′ of at most k edges from E′ such that (V, E ∪ E′′) satisfies

∗A preliminary version appeared in Proc. of the 10th Workshop on Algorithms and DataStructures (WADS 07), volume 4619 in LNCS, pages 484-495, Springer 2007.

†Institut fur Informatik, Friedrich-Schiller-Universitat Jena, Ernst-Abbe-Platz 2, D-07743Jena, Germany. Email: [email protected]. Supported by the Deutsche Forschungsgemein-schaft (DFG), Emmy Noether research group PIAF (fixed-parameter algorithms), NI 369/4,and research project DARE (data reduction and problem kernels), GU 1023/1.

‡Institut fur Informatik, Friedrich-Schiller-Universitat Jena, Ernst-Abbe-Platz 2, D-07743 Jena, Germany. Email: [email protected]. Supported by the DeutscheForschungsgemeinschaft (DFG), research projects OPAL (optimal solutions for hard problemsin computational biology), NI 369/2, and PABI (parameterized algorithmics for bioinformat-ics), NI 369/7.

1


a specified connectivity property. The edges in E′ are called the links. Eswaranand Tarjan [5] introduced connectivity augmentation problems and describedtheir numerous applications.

We use G = (V, E) and D = (V, A) to denote undirected and directed graphs.A path from vertex u1 to vertex ul in G = (V, E) (or D = (V, A)) is a sequence ofedges {u1, u2}, {u2, u3}, . . ., {ul−1, ul} (or arcs (u1, u2), (u2, u3), . . ., (ul−1, ul)).A cycle is a path with u1 = ul.

A vertex u in an undirected graph is called a cut-vertex if there are twovertices v, w with v 6= u and w 6= u such that every path from v to w con-tains u. If an undirected graph G is connected, has at least three vertices, andhas no cut-vertex, then G is biconnected. A bridge in an undirected graph is anedge {u, v} that is the only path between u and v. If G is connected and hasno bridge, then G is bridge-connected. A directed graph D = (V, A) is stronglyconnected if, for all pairs of vertices u and v, there is a path from u to v. Theconnected (biconnected, bridge-connected, strongly connected) components of agraph are its maximal connected (biconnected, bridge-connected, strongly con-nected) subgraphs. We consider here two connectivity augmentation problems,namely, Bridge-Connectivity Augmentation (BCA) and Biconnectiv-

ity Augmentation (BIA), where we are asked to add at most k links to makethe given graph bridge-connected or biconnected.

There is a long history of research dealing with BCA and BIA starting in1976 with the work of Eswaran and Tarjan [5]. They showed that, in the casethat E′ is complete, that is, graph (V, E′) is a complete graph, both problemsare polynomial-time solvable. In 1981, Frederickson and JaJa [10] proved theNP-completeness of both problems if E′ is incomplete. Motivated by the NP-completeness, the polynomial-time approximability of the optimization versionsof these problems has been extensively studied in the literature. Frederick-son and JaJa [10] gave polynomial-time factor-2 approximation algorithms forboth problems. For BCA, Nagamochi [16] improved the approximation factorto 1.875. Later, Even et al. [7] presented a polynomial-time factor-1.8 approx-imation algorithm for BCA.1 In case of BIA, Khuller and Thurimella [13] im-proved the running time of the factor-2 approximation algorithm in [10]. Onthe negative site, Kortsarz et al. [14] proved that there exists an ǫ > 0 for whichit is NP-hard to approximate BCA and BIA within a factor of 1 + ǫ.

Concerning the parameterized complexity of these problems, we are onlyaware of one result due to Nagamochi [16]. Since the bridge-connected compo-nents of a graph form a tree, we may assume that the input graph of a BCA-instance is a tree by contracting these components. Nagamochi [16] showedthat BCA is fixed-parameter tractable with respect to the number of leaves ℓin the given tree. More precisely, there is an algorithm solving this problemin O(ℓℓ+1 log ℓ · (|V | + |E′|)) time. Since the number of leaves provides a lowerbound on the solution size k, BCA is fixed-parameter tractable with respectto k.

1In the conference version Even et al. [6] presented and sketched a proof of a polynomial-time factor-1.5 approximation algorithm for BCA.

2


Kernelization is a core concept of parameterized algorithmics [11, 12, 17].However, nothing has been known so far concerning the kernelization of BCAand BIA. Kernelization is considered as one of the theoretically and practi-cally most interesting techniques of parameterized algorithmics [8, 11, 12, 17].Informally speaking, kernelization means polynomial-time preprocessing (datareduction) with provable performance guarantee. More specifically, a kernel-ization is a polynomial-time algorithm that transforms a given instance I withparameter k of a problem P into a new instance I ′ with parameter k′ ≤ k of Psuch that the original instance I is a yes-instance with parameter k if and onlyif the new instance I ′ is a yes-instance with parameter k′ and |I ′| ≤ g(k) for afunction g. The instance I ′ is called the problem kernel. That is, the goal is toderive in polynomial-time an instance equivalent to the original one, but whosesize is bounded by a function only depending on the parameter k. Next, we pro-vide some arguments for the usefulness of kernelization. First of all, there mightbe input instances for which preprocessing yields already an optimal solution forthe problem or where preprocessing results in an instance that is small enoughsuch that the use of simple exhaustive search is affordable (in terms of runningtime). Thus, in general, there is no doubt that polynomial-time data reductionor preprocessing is useful when dealing with NP-hard problems. However, al-though preprocessing is ubiquitous, so far only the kernelization technique ofparameterized algorithmics provides a theoretical framework for the mathemat-ical analysis of provable performance guarantees of a proposed data reduction.Although kernelization has its origins in parameterized algorithmics, it is a farmore general concept not restricted to this area. Since kernelization results im-ply efficient preprocessing algorithms, they are useful for any problem solvingapproach, be it exact, approximative, or heuristic. For example, after data re-duction the instance size might be small enough to exploit a search-tree strategyor an Integer Linear Program to solve the problem exactly. Moreover, it is con-ceivable that for many instances better approximative solutions are found afterpreprocessing. The practical usefulness of kernelization has been proven by suc-cessful empirical studies. For instance, experimental work based on kernelizationhas been done for Dominating Set [1, 15] and Cluster Editing [2, 3].

Our contributions. We complement the result of Nagamochi with a problemkernel consisting of O(k2) vertices and links for BCA and BIA. To this end, weprovide several polynomial-time data reduction rules for BCA and BIA. Fur-thermore, we study partial connectivity augmentation problems, a natural gen-eralization of the connectivity augmentation problems, where we have as inputa graph G = (V, E), a set E′ of links, and two non-negative integers k, Φ and askfor a subset E′′ of E′ with |E′′| ≤ k such that graph (V, E∪E′′) has a subgraphthat contains at least Φ vertices and satisfies the given connectivity property.Clearly, if Φ = |V |, then we have the connectivity augmentation problems. Weconsider three partial connectivity augmentation problems, Partial Bridge-

Connectivity Augmentation, Partial Biconnectivity Augmentation,and Partial Strong Connectivity Augmentation. For all three con-

3


nectivity properties, their corresponding non-partial connectivity augmentationproblems are solvable in polynomial-time if E′ is complete [5]. In Sect. 5, weshow that Partial Bridge-Connectivity Augmentation and Partial Bi-

connectivity Augmentation with a complete link set remain polynomial-time solvable but Partial Strong Connectivity Augmentation with acomplete link set is W[2]-hard (and NP-hard), that is, it is very unlikely thatthis problem is fixed-parameter tractable with respect to the parameter k.2

Some technical details omitted from this work for the sake of readability can befound in the second author’s diploma thesis [20].

2 Preliminaries

Parameterized complexity is a two-dimensional framework for studying the com-putational complexity of problems [4, 9, 17]. One dimension is the input size n(as in classical complexity theory) and the other one the parameter k (usually apositive integer). A problem is called fixed-parameter tractable (fpt) if it can besolved in f(k)·nO(1) time, where f is a computable function only depending on k.A core tool in the development of fixed-parameter algorithms is polynomial-timepreprocessing by data reduction rules, often yielding a kernelization. Herein, thegoal is, given any problem instance I with parameter k, to transform it in poly-nomial time into a new instance I ′ with parameter k′ such that the size of I ′ isbounded from above by some function only depending on k, k′ ≤ k, and (I, k)is a yes-instance if and only if (I ′, k′) is a yes-instance. A data reduction ruleis correct if the new instance after an application of this rule is a yes-instanceif and only if the original instance is a yes-instance. Throughout this paper, wecall a problem instance reduced if the corresponding data reduction rules cannotbe applied anymore.

A formal framework to show fixed-parameter intractability was developedby Downey and Fellows [4] who introduced the concept of parameterized reduc-tions. A parameterized reduction from a parameterized problem L to anotherparameterized problem L′ is a function that, given an instance (I, k), computesin f(k) · nO(1) time an instance (I ′, k′) (with k′ only depending on k) such that(I, k) ∈ L ⇔ (I ′, k′) ∈ L′. The basic complexity class for fixed-parameter in-tractability is W[1] as there is good reason to believe that W[1]-hard problemsare not fixed-parameter tractable [4]. There is a whole hierarchy W[1] ⊆ W[2]⊆ W[3] ⊆ . . . W[P] of presumable fixed-parameter intractability, called the W-Hierarchy. A parameterized problem P is called W[t]-hard if all problems inW[t] can be reduced to P by means of a parameterized reduction. Note thatthe notion of W[t]-hardness is defined analogously to NP-hardness and that inparameterized complexity the classes of the W-Hierarchy play the role of theclass NP in classical complexity.

Throughout this paper, we set n := |V | for a given graph G = (V, E)and m := |E′| for a given link set E′. For a graph G, we also use V (G) and E(G)

2There exists a whole hierarchy of presumable fixed-parameter intractability and the firsttwo levels of this hierarchy are captured by the complexity classes W[1] and W[2] (see Sect. 2).

4


Figure 1: Example of contracting bridge-connected components of a given graph.The links are drawn as dashed lines.

to denote its vertex and edge set, respectively. The open neighborhood N(v) ofa vertex v ∈ V is the set of vertices that are adjacent to v. The degree of avertex v, denoted by deg(v), is the size of N(v).

3 The Bridge-Connectivity Augmentation Prob-

lem

The main result of this section is a data reduction for Bridge-Connectivity

Augmentation (BCA) that leads to a quadratic-size problem kernel. Givenan instance of BCA, we can assume that the input graph G is a tree [5, 10, 7]:Each bridge-connected component of G = (V, E) can be contracted into a singlevertex by contracting all edges in this component, resulting in a tree. The setof links has to be adapted accordingly. The contraction of the bridge-connectedcomponents can be done in O(|V | + |E|) time [19]. See Fig. 1 for an example.Hence, in the following, the input instance is always denoted by T .

In contrast to the tree edges, denoted by {u, v}, we denote links by (u, v).We use pu,v to denote the uniquely determined path between two vertices uand v in T . In the course of the data reduction process, if a link (u, v) ∈ E′

is added to a solution E′′, then the vertices from the path pu,v form a bridge-connected component and we contract all edges in this component, obtaininga tree again. We say a link (u, v) covers an edge e if e lies on pu,v. For anedge e ∈ E, we use l(e) to denote the set of links covering e. A link (u, v) ∈ E′

is called a shadow if there exists a link (x, y) ∈ E′ with V (pu,v) ( V (px,y), thatis, the path between u and v is entirely contained in the path between x and yand {u, v} 6= {x, y}. See Fig. 2 for an illustration of the shadow definition.Let NE′(u) := {v | (u, v) ∈ E′}. For a vertex v ∈ V and an edge e ∈ E, weuse Tv,e to denote the subtree of (V, E \ {e}) that contains v.

The following observation provides the starting point for our kernelization:

Lemma 3.1 ([5]). Let L(T ) be the set of leaves of the tree T of a BCA-instance.Every solution of this instance contains at least |L(T )|/2 many links, that is,k ≥ |L(T )|/2.

We can conclude that every yes-instance of BCA contains at most 2k leaves

5


uu v xx y v = y

Figure 2: Illustration of the shadow definition. Herein, edges are drawn assolid lines and links are drawn as dashed lines. In both examples, link (u, v)is a shadow since the path between u and v is a proper subpath of the pathbetween x and y.

and, also, at most 2k− 1 internal vertices with degree at least three. It remainsto upper-bound the number of internal vertices of degree two. If we can boundthe maximum length of the paths that consist solely of degree-2 vertices, thenwe can achieve an upper bound on the number of degree-2 vertices. To this end,we apply four data reduction rules. We begin with three data reduction rulesthat are also used in [7, 16] and whose correctness is easy to verify.

Shadow Deletion: Delete all shadows in E′.

Unit Link: If there is an edge e ∈ E with l(e) = {(u, v)}, then con-tract pu,v and decrease the parameter k by one.

Covered Edge: If l(e1) ⊆ l(e2) for two edges e1, e2 ∈ E, then con-tract e2.

The polynomial running time of these rules is not hard to see [20].

Lemma 3.2 ([20]). The above three rules can be executed in O(n ·m3 + n3 ·m)time.

Before we present the fourth data reduction rule, we show some structuralproperties of a BCA-instance that is reduced with respect to the above threerules. In particular, we show that, in a reduced instance, the links over apath consisting solely of degree-2 vertices have some “consecutiveness” property,which provides the basis for the fourth data reduction rule.

The first property concerns the links with a degree-2 vertex as one of theirendpoints.

Lemma 3.3. Let (T = (V, E), E′, k) be a reduced instance with respect to theabove three rules, let v ∈ V be a degree-2 vertex in T , and let e, e′ be the edgesincident to v. Then, there exists at least one link (v, x) in E′ with x ∈ V (Tv,e)and at least one link (v, y) in E′ with y ∈ V (Tv,e′).

Proof. This lemma follows directly from the fact that the instance is reducedwith respect to the Covered Edge rule: Suppose that there is no link between vand the vertices in Tv,e. Then, all links in l(e′) have to be between verticesin Tv,e and Tv,e′ and have to cover both e and e′. This means l(e′) ⊆ l(e) andthe Covered Edge rule would apply.

6


The next lemma shows that every link has to cover at least two edges.

Lemma 3.4. Let (T = (V, E), E′, k) be a reduced instance with respect to theabove three rules. Then, for every link (u, v) ∈ E′, it holds that |E(pu,v)| ≥ 2.

Proof. Suppose that the lemma is not true. Then, we have a link (u, v) ∈ E′

with {u, v} ∈ E. Then, either (u, v) is the only link covering {u, v} or it is ashadow of other links. In the former case, the Unit Link rule would apply; inthe latter case, the Shadow Deletion rule would apply.

The next lemma concerns the paths which consist exclusively of degree-2vertices. More specifically, we show that the links with both endpoints on sucha path P fulfill a certain “consecutiveness” property. That is, these links canbe ordered in such a way that in comparison to its direct predecessor a link isshifted by exactly one vertex, that is, compared with its direct preprocessor itstarts and ends at one vertex later on P .

Lemma 3.5. Let (T = (V, E), E′, k) be a reduced instance with respect tothe above three rules. Consider a path P = {u, v1}, {v1, v2}, . . . , {vl, w} in Twith deg(vi) = 2 for all 1 ≤ i ≤ l, deg(u) ≥ 3, and deg(w) ≥ 3. Let E′

v denotethe set of links with both endpoints in {v1, v2, . . . , vl}. If E′

v 6= ∅, then there ex-ists an integer N with 1 ≤ N ≤ l−1 such that E′

v = {(vi, vi+N ) | 1 ≤ i ≤ l−N}and there exists no link (x, y) with x ∈ V (Tu,{u,v1}) and y ∈ V (Tw,{vl,w}).

Proof. If E′v 6= ∅, then there is no link (x, y) with x ∈ V (Tu,{u,v1}) and y ∈

V (Tw,{vl,w}); otherwise, all links in E′v would be shadows of (x, y) and the

Shadow Deletion rule would apply.By Lemma 3.3, there exists at least one link (v1, x) with x ∈ V (Tv2,{v1,v2}).

Then, x has to be one of v2, . . . , vl; otherwise, all links in E′v would be shadows

of (v1, x). Furthermore, there can be only one vertex from v2, . . . , vl whichtogether with v1 forms a link; otherwise, we would have a shadow. Let x = vj

with 2 ≤ j ≤ l. If j = l, then we set N := l − 1. Obviously, since all links withboth endpoints in v1, . . . vl are shadows of (v1, vl), we have Ev = {(v1, vl)} andthe lemma follows.

In the case that j < l, we first show that (vi, vj) ∈ E′v implies (vi+1, vj+1) ∈

E′v, for 1 ≤ i < j < l. By Lemma 3.3, there exists a link (z, vj+1) ∈ E′

with z ∈ V (Tvj ,{vj ,vj+1}). Since all shadows are deleted by the Shadow Deletionrule and, by Lemma 3.4, every link covers at least two edges, there must exista vertex vr with i < r < j and z = vr. To show that (vi+1, vj+1) ∈ E′

v, itsuffices to show that r = i + 1. Suppose that r > i + 1. By Lemma 3.3, weknow that there is a link (vi+1, z

′) with z′ ∈ V (Tvi+2,{vi+1,vi+2}). Vertex z′ canonly be from {vi+2, . . . , vj}; otherwise, (z, vj+1) would be a shadow of (vi+1, z

′).Moreover, z′ cannot be a vertex with index smaller than j +1, since, otherwise,link (vi+1, z

′) would be a shadow of (vi, vj). Thus, r = i + 1 and (vi+1, vj+1) ∈E′

v.Then, by setting N := j − 1 and using the facts that (v1, vN+1) ∈ E′

v

and (vi, vj) ∈ E′v ⇒ (vi+1, vj+1) ∈ E′

v, we can show that {(vi, vi+N )|1 ≤ i ≤l − N} ⊆ E′

v. Moreover, it is easy to observe that including any other link

7


Figure 3: Example for the Degree-2-Path rule for N = 3, l = 11, c = 2, andd = 2.

between two vertices from {v1, . . . , vl} in E′v would contradict the fact that

the given instance is reduced with respect to the Shadow Deletion rule. Thiscompletes the proof.

The fourth data reduction rule restricts the length of paths that consist solelyof degree-2 vertices. By Lemma 3.5, the links with both endpoints from sucha path admit a consecutiveness property. By making use of this property, thenext data reduction rule replaces a long degree-2 path by a shorter “equivalent”degree-2 path.

Degree-2-Path: Let (T = (V, E), E′, k) be a reduced instance withrespect to the above three rules. Let P = {u, v1}, {v1, v2}, . . . , {vl, w} be apath in T such that deg(vi) = 2 for all 1 ≤ i ≤ l, deg(u) ≥ 3, and deg(w) ≥3 and let E′

v denote the set of links with both endpoints from v1, v2, . . . , vl.If there exists an integer N with l ≥ 2N such that E′

v = {(vi, vi+N ) | 1 ≤i ≤ l−N}, then proceed as follows: Let c := ⌊ l

N⌋−1 and d := (l mod N).

Replace P by a path P ′ = {u, x1}, {x1, x2}, . . . , {xN+d, w}. Remove alllinks in E′

v from E′ and add the links (xi, xi+N ) for 1 ≤ i ≤ d to E′.Replace every link (vi, y) with y ∈ V (Tu,{u,v1}) by the link (xi, y) andreplace every link (vi, y) with y ∈ V (Tw,{vl,w}) by the link (xN+d−(l−i), y).Finally, decrease parameter k by c.

See Fig. 3 for an example of the application of the Degree-2-Path rule. Next,we prove that this rule is correct.

Lemma 3.6. The Degree-2-Path rule is correct and can be executed in O(n2 +nm) time.

Proof. Let (T = (V, E), E′, k) be a BCA-instance reduced with respect to thefirst three rules and let (Ta, E′

a, ka) be the resulting instance after one applicationof the Degree-2-Path rule. Let P = {u, v1}, {v1, v2}, . . . , {vl, w} be the path forwhich the conditions of the Degree-2-Path rule are fulfilled and let E′

v be theset of links with both endpoints from v1, . . . , vl. By Lemma 3.5, either E′

v = ∅or there is an integer N with E′

v = {(vi, vN+i) | 1 ≤ i ≤ l − N} and there isno link between a vertex in V (Tu,{u,v1}) and a vertex in V (Tw,{vl,w}). Since theDegree-2-Path rule is applicable to P , we know E′

v 6= ∅ and l ≥ 2N . Then, wehave c = ⌊ l

N⌋− 1, d = (l mod N), and ka = k− c. We show that (T, E′, k) is a

yes-instance if and only if (Ta, E′a, ka) is a yes-instance. The correctness of the

Degree-2-Path rule then follows by induction on the number of applications ofthe rule.

8


“⇒”: Let E′′ ⊆ E′ with |E′′| ≤ k be a solution for the original instance. Wefirst show some properties of E′′ ∩E′

v and then construct a solution E′′a for the

new instance with |E′′a | ≤ ka from E′′.

We can assume that E′′∩E′v contains only pairwise “non-overlapping” links,

that is, there are no two links (vi, vi+N ), (vj , vj+N ) ∈ E′′∩E′v with i < j < i+N :

If there are two links (vi, vi+N ), (vj , vj+N ) ∈ E′′ ∩E′v with i < j < i + N , then

we construct another solution from E′′ by replacing (vj , vj+N ) by (vi+N , z),where z = vi+2N or z is a vertex from Tw,{vl,w}. Link (vi+N , z) exists due toLemma 3.3. Obviously, this yields a solution of the same size (or even of smallersize if (vi+N , z) is already part of the solution).

Next, we show that c ≤ |E′′ ∩ E′v| ≤ c + 1. On the one hand, since every

link in E′v covers exactly N edges and the links in E′′ ∩ E′

v are pairwise non-

overlapping, there can be at most ⌊ l−1N

⌋ = ⌊ (c+1)N+d−1N

⌋ = c + 1 pairwisenon-overlapping links in E′′∩E′

v. On the other hand, all edges of path P whichlie between vertex vN and vertex vl−N+1 have to be covered by links in E′

v. This

means that E′′ ∩ E′v has cardinality at least ⌈ l−2N+1

N⌉ = ⌈ (c+1)N+d−2N+1

N⌉ =

(c − 1) + ⌈d+1N

⌉ = c.In the following, let il := max{i | (y, vi) ∈ E′′ ∧ y ∈ V (Tu,{u,v1})} and ir :=

min{i | (vi, z) ∈ E′′ ∧ z ∈ V (Tw,{vl,w})}. We distinguish two cases, namely,|E′′ ∩E′

v| = c and |E′′ ∩E′v| = c + 1, and construct in both cases a solution for

the new instance.In the first case, we can assume that E′′∩E′

v = {(vil, vil+N ), (vil+N , vil+2N ),

. . ., (vil+(c−1)N , vil+cN )}. We construct the new solution E′′a from E′′ as fol-

lows: We replace every link (y, vi) ∈ E′′ by a link (y, xi). Note that, sincethe given instance is reduced with respect to the first three rules, the exis-tence of the link (y, vi) implies that i < N ≤ N + d and, thus, the link (y, xi)exists in E′

a. The links (vi, z) ∈ E′′ with z ∈ V (Tw,{vl,w}) are also replacedby links (xN+d−(l−i), z). With the same reason as above, links (xi, z) existin E′

a. Finally, we remove all links in E′′ ∩ E′v from E′′. The resulting set E′′

a

is then a solution for the new instance. Since |E′′ ∩ E′v| = c in this case,

we have |E′′a | ≤ ka. Obviously, all edges of Ta that are not between the new

vertices x1, . . . , xN+d are covered. To show that the edges between the newvertices are also covered, observe that the edges on the path between x1 and xil

and the edges on the path between xN+d−(l−ir) and xN+d are covered by thelinks (y, xil

) with y ∈ V (Tu,{u,v1}) and (xN+d−(l−ir), z) with z ∈ V (Tw,{vl,w})that replace the links (y, vil

) and (vir, z), respectively. Then, it suffices to show

that N + d − (l − ir) ≤ il. Since E′′ is a solution of the non-reduced instance,we get il + cN ≥ ir. This is equivalent to

il ≥ ir − cN = N + d − cN − N − d + ir = N + d − (l − ir).

In the second case, we assume that E′′ ∩E′v = {(vil

, vil+N), (vil+N , vil+2N ),. . ., (vil+cN , vil+(c+1)N)}. We construct the solution E′′

a for the new instancefrom E′′ as follows: We replace every link (y, vi) ∈ E′′ with y ∈ V (Tu,{u,v1}) bya link (y, xi). The links (vi, z) ∈ E′′ with z ∈ V (Tw,{vl,w}) are also replaced bylinks (xN+d−(l−i), z). We remove all links in E′′ ∩E′

v from E′′. Finally, we add

9


link (xil, xil+N ) to E′′. With a similar argument as in the first case, we can

show that the resulting set is a solution of the new instance.“⇐”: Let E′′

a with |E′′a | ≤ ka be a solution of the reduced instance. Let E′

x

denote the set of links from E′a that have both their endpoints from {x1, x2, . . . , xN+d}.

With a similar argument as in the proof of the reverse direction, we can assumethat all links in E′′

a ∩ E′x are pairwise non-overlapping. By the construction of

the links in E′a, we can conclude that 0 ≤ |E′′

a ∩ E′x| ≤ 1. Next, we distinguish

again two cases based on whether |E′′a ∩ E′

x| = 0 or |E′′a ∩ E′

x| = 1.In the first case, we let il := max{i | (y, xi) ∈ E′′

a ∧ y ∈ V (Tu,{u,x1})}.We replace the links (z, xi) with z /∈ {x1, x2, . . . , xN+d} in E′′

a by their cor-responding links in E′ and, then, add c links (vil

, vil+N ), (vil+N , vil+2N ), . . .,(vil+(c−1)N , vil+cN ) to E′′

a . In analog to the proof of the reverse direction, wecan show that the resulting set is a solution E′′ of the non-reduced instancewith |E′′| ≤ k.

In the second case, let E′′a ∩ E′

x = {(xi, xi+N )}. We replace the links (z, xi)with z /∈ {x1, x2, . . . , xN+d} in E′′

a by their corresponding links in E′. Then, wereplace (xi, xi+N ) by (vi, vi+N ) and add c links (vi+N , vi+2N ), . . ., (vi+(c−1)N , vi+cN ),(vi+cN , vi+(c+1)N ) to E′′

a . Analogously, we can show that the resulting set is asolution of the non-reduced instance.

Altogether, we have shown that the two instances are equivalent and, thus,the Degree-2-Path rule is correct. Next, we prove that this rule can be exe-cuted in O(n2 + n · m) time. Using a depth-first traversal of the input tree T ,we can determine in O(n) time all paths of the form P = {u, v1}, . . . , {vl, w}with deg(u) ≥ 3, deg(w) ≥ 3, and deg(vi) = 2 for all 1 ≤ i ≤ l. Next, we checkfor every such path whether the Degree-2-Path rule can be applied, that is, wehave to determine an integer N with l ≥ 2N such that it holds for the set E′

v oflinks with both endpoints from {v1, . . . , vl} that E′

v = {(vi, vi+N ) | 1 ≤ i ≤ l}.This is clearly doable in O(n+m) time. If this rule is applicable for a path, thenthe replacing operation described in this rule can be carried out in O(n+m) time:Replacing path P by the path P ′ = {u, x1}, . . . , {xN+d, w} can be executedin O(n) time. Next, to find the links with exactly one endpoints in {v1, . . . , vl}and to replace them with links with exactly one endpoints in {x1, . . . , xN+d}can be done in O(n + m) time. Therefore, the overall running time of theDegree-2-Path rule is O(n2 + n · m).

Next, we show that a BCA-instance reduced with respect to the four datareduction rules has O(k2) vertices and O(k2) links. As stated in the introductionof this section, we can easily bound the number of the leaves and the number ofthe internal vertices with degree at least three of a reduced instance by O(k).Thus, the key point in the following is to upper-bound the number of the internalvertices with degree two. Herein, we consider the paths formed by degree-twovertices. The next two lemmas are used to show the upper bounds on thenumber and the length of such paths. The first one is due to Even et al. [7]and shows that there is no such degree-two vertex path between a leaf and aninternal vertex of degree at least three.

10


Lemma 3.7 ([7]). Let (T, E′, k) be a BCA-instance to which the Shadow Dele-tion rule, the Unit Link rule, and the Covered Edge rule cannot be applied. Let vbe a leaf of T and u be the parent of v. Then, deg(u) ≥ 3.

In the next lemma, we upper-bound the length of a path in a reduced instancethat consists of degree-two vertices. Herein, we use L(T ) to denote the set ofleaves in tree T .

Lemma 3.8. Let (T, E′, k) be a reduced BCA-instance and let P = {u, v1},{v1, v2}, . . ., {vl, w} be a path in T with deg(u) ≥ 3, deg(w) ≥ 3, and deg(vi) = 2for all 1 ≤ i ≤ l. Let Lu := L(Tu,{u,v1}) and Lw := L(Tw,{vl,w}). Then,

(1) there are at most 2|Lu| + 2|Lw| links that have exactly one endpointin VP := {v1, v2, . . . , vl}.

(2) l ≤ 4 · min(|Lu|, |Lw|).

Proof. In the following, we use Tu and Tw to denote Tu,{u,v1} and Tw,{vl,w},respectively, and consider them as two rooted trees with roots u and w, respec-tively. For a vertex x in Tu or Tw, we use Tx to denote the subtree of Tu or Tw

rooted at x. Moreover, we use Au and Aw to denote the sets of internal verticesof degree at least 3 in Tu and Tw. Recall that VP := {v1, v2, . . . , vl}.

The key for proving the lemma is the following observation: There exist atmost |Au| + |Lu| links in E′ with one endpoint in Tu and one endpoint in VP

and there exist at most |Aw|+ |Lw| links in E′ with one endpoint in Tw and oneendpoint in VP . Here, we prove this observation for Tu. Consider a degree-2vertex x in Tu. By Lemma 3.3, there exists a link (x, y) ∈ E′ with y ∈ V (Tx).The existence of the link (x, y) then excludes any link (a, b) with a ∈ V (Ty)and b ∈ VP , since, otherwise, (x, y) would be a shadow and the Shadow Deletionrule would be applied. In this way, every degree-2 vertex x in Tu “blocks” atleast one vertex from Tx from building links with the vertices in VP . Thus, bya simple calculation, we arrive at the |Au|+ |Lu|-bound on the number of linksbetween the vertices in Tu and the vertices in VP .

From the above observation, we know that there are at most |Au| + |Lu| +|Aw| + |Lw| links with exactly one endpoint in VP . Since |Au| ≤ |Lu| − 1and |Aw| ≤ |Lw| − 1, the first part of the lemma follows.

To prove the second part of the lemma, we distinguish two cases. First,suppose that there exists at least one link (x, y) ∈ E′ with x ∈ V (Tu) and y ∈V (Tw). Then, since the instance is reduced with respect to the Shadow Deletionrule, there is no link in E′ between two vertices in VP . By Lemma 3.3, for everyvertex vi in VP , there are at least two links (vi, a), (vi, b) ∈ E′ with a ∈ V (Tu)and b ∈ V (Tw). According to the above observation, there are at most 2|Lu|−1(or 2|Lw| − 1) links between VP -vertices and the vertices in V (Tu) (or V (Tw)).Therefore, |VP | ≤ min(2|Lu| − 1, 2|Lw| − 1).

In the second case, there is no link (x, y) ∈ E′ with x ∈ V (Tu) and y ∈V (Tw). Let vil

be the vertex in VP such that there exist a link (vil, z) ∈ E′

with z ∈ V (Tu) and, for all il ≤ i ≤ l, there is no link (vi, z) ∈ E′ with z ∈V (Tu). From the above observation, we know il ≤ 2|Lu| − 1. By Lemma 3.3

11


and the fact that the instance is reduced with respect to the Degree-2-Path rule,for every vertex vi with il ≤ i ≤ l, there is a link (vi, vj) ∈ E′ with 1 ≤ j ≤ il.Since the instance is reduced with respect to the Shadow Deletion rule, therecannot be two vertices from {vil+1, . . . , vl} which form two links with one vertexfrom {v1, . . . , vil

}, we can conclude l ≤ 2il ≤ 4|Lu| − 2. Obviously, the sameargument can also be applied to obtain l ≤ 4|Lw| − 2. The second part of thelemma follows.

Now, we prove the size bound of the problem kernel for BCA.

Theorem 3.1. Bridge-Connectivity Augmentation admits a problem ker-nel with O(k2) vertices and O(k2) links.

Proof. Let (T, E′, k) be a reduced BCA-instance. Let L be the set of leavesof T , A be the set of internal vertices of degree at least three of T , and Bbe the set of internal vertices of degree two. Clearly, V = L ∪ A ∪ B. ByLemma 3.1, we know that |L| ≤ 2k and |A| ≤ 2k − 1. It remains to upper-bound |B|. By Lemma 3.7, we know that all degree-two vertices form pathsbetween vertices in A. Thus, there can be at most 2k−2 such paths. Moreover,according to the second part of Lemma 3.8, each of these paths contains at

most 4 · |L|2 = 2L ≤ 4k vertices. Thus, we have |B| ≤ (2k − 2) · 4k. Altogether,

|V | ≤ 2k + 2k − 1 + (2k − 2) · 4k = O(k2).In order to show |E′| = O(k2), observe that there can be at most O(k2) links

with both endpoints in L. Furthermore, for every vertex u ∈ A, there can be atmost |L| links with one endpoint being u: Root T at u. If there is a link (u, v) ∈E′, then there exists no link (u, w) with w being in the subtree of T rooted at v;otherwise, there would be a shadow. Then, there can be at most O(k2) links withat least one endpoint from A. Finally, consider a path P = {u, v1}, . . . , {vl, w}with deg(u) ≥ 3, deg(w) ≥ 3, and deg(vi) = 2 for all 1 ≤ i ≤ l. By the firstpart of Lemma 3.8, we know that there exists at most 2|L| links with exactlyone endpoint from {v1, . . . , vl}. Let Ev be the set of links with both endpointsfrom {v1, . . . , vl}. By Lemma 3.5, these links in Ev are of a “consecutive”form, that is, there exists an integer N such that Ev = {(vi, vi+N ) | 1 ≤ i ≤ l}.Moreover, from the second part of Lemma 3.8, we know that l ≤ 2|L|. Hence, wecan conclude |Ev| ≤ 2|L|. Since, due to Lemma 3.7, there are at most 2k−2 suchdegree-two-vertex paths, there can be at most (2k − 2) · (2|L| + 2|L|) = O(k2)links in E′ with at least one endpoint from B. Altogether, we arrive at theclaimed upper bound on the number of links in E′, namely, |E′| = O(k2).

4 The Biconnectivity Augmentation Problem

Recall the definition of the Biconnectivity Augmentation (BIA) problem:Given a graph G = (V, E), a set of links E′ on V , and a non-negative integer k,BIA asks for a set E′′ of at most k links from E′ whose insertion biconnects G. Inthis section, by studying Biconnectivity Augmentation (BIA), we deal witha more general problem setting than in the previous section. Hence, based on the

12


Figure 4: A graph together with its block tree. The cut-vertices are coloredgray and the block-vertices are drawn as rectangles.

previous section, we extend and refine our kernelization technique introducedthere. As shown by Frederickson and JaJa [10], we can assume that the inputgraph is a so-called block tree. A block tree T = (VT , ET ) is a tree over thevertex set VT := B ∪ C with B ∩ C = ∅ where the leaves of T form a subsetof B and the edges in ET have one endpoint from B and one endpoint from C.

A maximal connected induced subgraph that has no cut-vertex is calleda block.3 We can easily compute a block tree from a given undirected andconnected graph G = (V, E): Identify B as the set of blocks of G and C as theset of cut-vertices of G. Insert an edge between a block and a cut-vertex into ET

if the cut-vertex belongs to the block. In the following, the vertices in B arecalled block-vertices and the vertices in C are called cut-vertices. See Fig. 4 foran example of a graph and its block tree.

Bridge-Connectivity Augmentation (BCA) is a special case of BIA:For a given BCA-instance (T, E′, k), we construct a BIA-instance by replacingevery vertex in T by a block-vertex and every edge e in T by a degree-two cut-vertex that is adjacent to the endpoints of e. The link set and the parameterremain the same. In the following, we generalize the kernelization for BCA inSect. 3 to a kernelization for BIA which leads to a quadratic-size problem kernel.

Eswaran and Tarjan [5] gave a lower bound on the size of the solutions of aBIA-instance.

Lemma 4.1 ([5]). If (T, E′, k) is a yes-instance of BIA, then k ≥ ⌈|L|/2⌉where L is the set of leaves of T .

By Lemma 4.1, the number of leaves and the number of internal vertices ofdegree at least three of a given block tree can be easily bounded from aboveby 2k and 2k − 1. Again, we focus on the internal vertices of degree two of T .The decisive difference between a BIA-instance and a BCA-instance lies in thepartition of the tree vertices into two subsets, the block-vertices and the cut-vertices. A block-vertex can only have cut-vertices as neighbors and vice versa.In the following, we present first a preprocessing, which ensures that the linksin E′ are all between block-vertices.

3Note that a block of a connected graph is either a maximal biconnected subgraph or abridge.

13


T1T1 T2

T2

T3

T3 T4T4

uc1

c1

c2c2v

K

Figure 5: An illustration of the modification made after adding a link (u, v) tothe solution set.

Preprocessing: While there exists a link (u, v) ∈ E′ with u ∈ Cand v ∈ B ∪ C, replace (u, v) by the link (w, v) where w ∈ B is theneighbor of u that lies on the path between u and v in T . Finally,for all u ∈ B ∪ C, remove all links (u, u) from E′.

To see the correctness of the preprocessing, the following equivalent formu-lation of BIA is helpful: Given a block tree T = (B ∪ C, ET ), a set of links E′,and a non-negative integer k, find a subset E′′ of links with |E′′| ≤ k such that,for every c ∈ C, if c is removed from the graph (B ∪ C, ET ∪ E′′), then theresulting graph is still connected.

Lemma 4.2. The preprocessing is correct and can be executed in O(n ·m) time.

Proof. Based on the equivalent formulation of BIA, observe that, in order toensure that a cut-vertex c is not a cut-vertex anymore after adding the linksin E′′, we have to add at least one link (u, v) that “covers” c, that is, c lies onthe path between u and v in T , c 6= u, and c 6= v. Based on this observation,it is easy to see that a link (u, v) with one endpoint u being a cut-vertex doesnot really cover u. This link covers exactly the same set of cut-vertices as thelink (v, w) with w being the neighbor of u that lies on the path between u and v.Thus, the preprocessing is correct.

Concerning the running time, there are m links in E′. For each link, weexamine its endpoints and, in the case that at least one of its endpoints is acut-vertex, we compute the unique determined path between its endpoints in T .This is clearly doable in O(n) time.

Next, we present the data reduction rules for BIA which generalize the datareduction rules in Sect. 3. Herein, if we add a link (u, v) to the solution E′′,then, following Rosenthal and Goldner [18], we modify the instance as follows:Let P denote the path in T between u, v, let C be the set of cut-vertices on Pthat have degree at least three, and let N be the set of cut-vertices which areneighbors of the block-vertices in P and do not lie on P . Replace P by a singleblock-vertex K. Every link (u, v) with at least one endpoint being in P , say u,is replaced by link (K, v). For every vertex v ∈ N , add edge {K, v} and, forevery c ∈ C, add edge {K, c}. An illustration is given in Fig. 5.

The data reduction rules use the following terms and notations: Recall thata link (u, v) is called a shadow, if there exists a link (x, y) with {x, y} 6= {u, v}

14


such that, in T , the path between u and v is entirely contained in the pathbetween x and y. For a vertex u, we use E′

u to denote the links in E′ whichcover u or have u as one of its endpoints. We call a path between two cut-vertices a degree-2-cut-path if all vertices on this path are degree-two vertices.A degree-2-cut-path is maximal if it is not a proper subpath of another degree-2-cut-path.

Shadow Deletion: Delete all shadows.

Unit Link: If there exists a cut-vertex u with E′u = {(x, y)}, then add (x, y)

to E′′ and decrease the parameter k by one.

Covered Cut-Vertex: If there are two cut-vertices u and v with E′u ⊆ E′

v

and N(v) = {w1, w2}, then add a new block-vertex w and make it adjacentto the vertices in (N(w1) ∪ N(w2)) \ {v} and replace every link of theform (w1, x) or (w2, x) by a link (w, x). Finally, remove v, w1, w2 from T .

Degree-2-Cut-Path: Let (T, E′, k) be a BIA-instance to which the first threerules do not apply, let P = {c1, b1}, {b1, c2}, {c2, b2}, . . ., {cl, bl}, {bl, cl+1}be a maximal degree-2-cut-path in T with {b1, . . . , bl} ⊆ B and {c1, . . . , cl+1} ⊆C, and E′

b be the set of links with both endpoints from {b1, . . . , bl}. Ifthere exists an integer N with 2N ≤ l such that E′

b = {(bi, bi+N ) | 1 ≤ i ≤l − N}, then proceed as follows: Let c := ⌊ l

N⌋ − 1 and d := (l mod N).

Replace P by a path P ′ = {c′1, b′1}, . . . , {c

′N+d, b

′N+d}, {b

′N+d, c

′N+d+1}.

Remove all links in E′b from E′ and add links (b′i, b

′i+N ) for 1 ≤ i ≤ d

to E′. Replace every link (bi, y) with y ∈ V (Tc1,{c1,b1}) by link (b′i, y) andreplace every link (bi, y) with y ∈ V (Tcl+1,{bl,cl+1}) by link (b′N+d−(l−i), y).Finally, decrease the parameter k by c.

The correctness of the first three data reduction rules is easy to verify. Thecorrectness of the Degree-2-Cut-Path rule can be shown in a similar way as thethe Degree-2-Path rule in Sect. 3. Due to the similarity of the four rules to theones in Sect. 3, the running time follows from Lemmas 3.2 and 3.6. All detailsomitted here can be found in the corresponding diploma thesis [20].

Lemma 4.3. The four data reduction rules are correct and can be executedin O(n · m3 + n3 · m) time.

The proof of the next theorem works in analog to the proof of Theorem 3.1.

Theorem 4.1. Biconnectivity Augmentation admits a problem kernelwith O(k2) vertices and O(k2) links.

5 Partial Augmentation Problems

In this section, we study partial augmentation problems. Here, given a graph, aset of links, and two non-negative integers Φ, k, one asks for a set of at most klinks whose insertion in the graph results in a graph that has a subgraph with

15


at least Φ vertices that satisfies a given connectivity property. The three con-nectivity properties considered here are bridge-connectivity, biconnectivity, andstrong connectivity. We show that, in the case that the link set is complete, thatis, it contains all possible edges or arcs, Partial Bridge-Connectivity Aug-

mentation (PBCA) and Partial Biconnectivity Augmentation (PBIA)are polynomial-time solvable and Partial Strong Connectivity Augmen-

tation (PSCA) is W[2]-hard with respect to k. Note that the non-partial ver-sions of all three problems, where, after adding at most k links, the whole graphshould satisfy the given connectivity property, are polynomial-time solvable ifthe link set is complete.

5.1 Partial Bridge-Connectivity Augmentation

The Partial Bridge-Connectivity Augmentation problem (PBCA) is de-fined as follows: Given an undirected and connected graph G = (V, E), a setof links E′, and two non-negative integers Φ, k, find a set E′′ of at most klinks such that the graph (V, E ∪ E′′) contains a bridge-connected componentwith at least Φ vertices. Note that in the case that E′ is incomplete, Bridge-

Connectivity Augmentation is NP-complete, which implies that PBCA isalso NP-complete in this case. We show here that PBCA becomes polynomial-time solvable if E′ is complete. This extends a result by Eswaran and Tarjan [5]saying that BCA is polynomial-time solvable in the case of a complete linkset. Our solving strategy consists of two steps: The first step reduces the aug-mentation problem to a special subtree problem and the second step appliesa dynamic programming approach to solve the subtree problem. The specialsubtree problem, Maximum d-Leaves Subtree (MLST), is defined as follows:Given a tree T = (VT , ET ), two non-negative integers N, d, and a weight func-tion w : VT → N, find a subtree of T with at most d leaves such that the totalweight of the vertices in this subtree is at least N .

Lemma 5.1. Partial Bridge-Connectivity Augmentation with completelink sets can be reduced to Maximum d-Leaves Subtree in O(|V |+ |E|) time.

Proof. The reduction from the augmentation problem to MLST works as fol-lows: Given a PBCA-instance (G, E′, k, Φ), we contract the bridge-connectedcomponents of G into vertices, resulting in a tree T as described in Sect. 3.See Fig. 1 for an illustration. The weight w(v) of a tree vertex v is set equalto the number of vertices in the corresponding bridge-connected component.Finally, we set N := Φ and d := 2k. The resulting MLST-instance consistsof (T, N, d, w). Since the contraction of the bridge-connected components isdoable in O(|V |+ |E|) time [19], the same time bound holds for this reduction.

Suppose that we have a solution set E′′ of the PBCA-instance with |E′′| ≤ k.It is easy to observe that every link in E′′ is between two bridge-connectedcomponents of G which correspond to leaves of T . Let L′ denote the set ofthese leaves and let T ′ be the subtree of T with the leaf set L′. Then, theresulting bridge-connected component C of graph (V, E ∪ E′′) with |C| ≥ Φcontains the vertices which are contained in the bridge-connected components

16


of G which correspond to the vertices in T ′. Therefore, the total weight of thevertices of T ′ is at least Φ and since |E′′| ≤ k, we have |L′| ≤ 2k = d.

The other direction is easier. Given a subtree T ′ of T with at most d leavesand total vertex weight at least N , we can add d/2 links to make T ′ bridge-

connected: Eswaran and Tarjan [5] showed that at least ⌈ |L|2 ⌉ links are needed,

where L is the set of leaves of T ′. Moreover, they proved that this lower boundis achievable in linear time in the case of a complete link set. Therefore, wecan find a set of d/2 = k links to get a bridge-connected subgraph with N = Φvertices.

Next, we describe the dynamic programming approach solving MLST.

Theorem 5.1. Maximum d-Leaves Subtree can be solved in O(|VT | · d2)time.

Proof. We root the input tree T at an arbitrary vertex r and use Tv to denotethe subtree rooted at a vertex v. The bottom-up dynamic program maintains atable Av for each vertex v, which has d entries, Av(1), . . . , Av(d). Entry Av(i)stores the maximal total vertex weight achievable by a subtree of Tv that isrooted at v and has at most i leaves.The goal of the dynamic programmingprocedure is to compute Av(d) for all vertices v ∈ VT . Then, maxv∈VT

Av(d),by definition of Av(d), equals the total vertex weight achievable by a subtreeof T .

The initialization of the table Av for a leaf v is trivial, that is, we set Av(0) =0 and Av(i) = w(v) for 1 ≤ i ≤ d. At an internal vertex v, let c(v) = {u1, . . . , uj}denote the set of v’s children. We define table Bv, where for 1 ≤ i ≤ dand 0 ≤ l ≤ j table entry Bv(i, l) is defined to be the maximal total vertexweight achievable by a subtree of Tv that is rooted at v, has at most i leaves,and contains no vertex from {u1, . . . , ul}. Note that by this definition tableentry Bv(i, 0) equals Av(i). Indeed, Bv is just an “auxiliary” table needed forthe computation of Av. From l := j to 0, we compute Bv(i, l). For l = j, weset Bv(i, j) = w(v) for all i. This is correct, since the subtree cannot contain avertex from c(v) and, thus, no vertex below v. For l < j, we set

Bv(i, l) := max0≤i′≤i

(Bv((i − i′), l + 1) + Aul+1(i′)).

This is correct because every subtree rooted at v and containing no vertexfrom {u1, . . . , ul} can be decomposed into two trees, one is rooted at v andcontains vertices from subtrees Tul+2

, . . . , Tujand the other is rooted at ul+1

and is a subtree of Tul+1. Finally, we set Av(i) := Bv(i, 0) for all i.

At the root r, we know that the maximal total vertex weight achievable bya subtree of T with at most d leaves is then maxv∈VT

Av(d). Using a simpletraceback, this subtree can also be computed in the same time bound.

Concerning the running time, the dynamic program goes over all vertices. Ateach vertex v, it fills the table Av. To do this, it computes the functions Bv(i, l)with 1 ≤ i ≤ d and 0 ≤ l ≤ j where j denotes the number of v’s children. Foreach of these functions, it computes again at most d values (see the recursion

17


above). Altogether, it needs O(j · d2) time for a vertex having j children.Summing up over all vertices in T , we arrive at a total running time of O(|VT | ·d2).

Combining Lemma 5.1 and Theorem 5.1, we arrive at the main result of thissubsection.

Theorem 5.2. In the case of a complete link set, Partial Bridge-Connectivity

Augmentation is solvable in O(|V | · k2) time.

5.2 Partial Biconnectivity Augmentation

In this section, we introduce the Partial Biconnectivity Augmentation

problem (PBIA). Given an undirected and connected graph G = (V, E), a set oflinks E′, and two non-negative integers Φ, k, PBIA asks whether there exists aset E′′ ⊆ E′ of size at most k such that (V, E ∪ E′′) contains a biconnectedcomponent with at least Φ vertices. We show that, as Partial Bridge-

Connectivity Augmentation, this problem is polynomial-time solvable ifthe link set is complete.

The idea of the solving strategy is to find, by dynamic programming on theblock tree representation of G, a subgraph of G with at least Φ vertices thatcan be made biconnected by inserting at most k links.

Recall that the block tree TG = (VT , ET ) of an undirected and connectedgraph G is a tree over the vertex set VT = B ∪ C, where B are the blocksand C are the cut-vertices of G (see Sect. 4). Moreover, a cut-vertex c ∈ Cis adjacent to a block b ∈ B if and only if the cut-vertex c is contained in theblock b. Given a graph G and the block tree TG = (C ∪ B, ET ) of G, theweight of a subtree T ′ = (B′ ∪ C′, E′

T ) with C′ ⊆ C, B′ ⊆ B, and E′T ⊆ ET of

the block tree TG is defined as the size of the union of the blocks in B′, thatis, w(T ′) := |

⋃b∈B′ b|. See Fig. 6 for an example.

We show that PBIA reduces to a constrained subtree problem on the blocktree of the input graph.

Lemma 5.2. If E′ is complete, then (G, E′, k, Φ) is a yes-instance for Partial

Biconnectivity Augmentation if and only if the block tree TG = (C∪B, ET )of G contains a subtree T ′ = (C′ ∪ B′, E′

T ) such that

• the leaves of T ′ are contained in B,

• the number of leaves of T ′ is at most 2k,

• the maximum degree of a cut-vertex in T ′ is at most k + 1, and

• the weight w(T ′) = |⋃

b∈B′ b| of T ′ is at least Φ.

Proof. The proof is similar to the proof of Lemma 5.1. It is based on the resultby Eswaran and Tarjan [5] that for a connected graph one needs to add atleast max{⌈|LTG

|/2⌉, ∆c − 1} links to turn it biconnected, where LTGdenotes

the set of leaves of the block tree of G and ∆c denotes the maximum degree of

18


a) b) c)

Figure 6: An example for a graph a) its block tree b) and a subtree of the blocktree c). The vertices that are contained in the subtree, either as cut-vertices orcontained in a block, are colored gray in a). The weight of the subtree in c)is 14 in accordance with the number of gray vertices in the graph shown in a).

a cut-vertex in the block tree of G. Moreover, they showed that, in the case ofa complete link set, a set of max{⌈|LTG

|/2⌉, ∆c − 1} links whose addition turnsthe graph biconnected can be found in polynomial time. Next, we prove thecorrectness of the lemma.

“⇐”: Consider a subtree T ′ = (C′ ∪ B′, E′T ) fulfilling the conditions of the

lemma. Let V ′ :=⋃

b∈B′ b. Note that T ′ is the block tree of G[V ′] and, hence,by the above mentioned result it can be turned biconnected by the addition of aset E′′ of at most k links. Thus, V ′ forms a biconnected component in (V, E∪E′′)of size at least Φ.

“⇒”: Given a set E′′ of at most k link such that (V, E ∪ E′′) contains abiconnected component with at least Φ vertices. Moreover, let V ′ be the set ofvertices in such a biconnected component and let T ′ = (C′ ∪B′, E′

T ) denote theblock tree of G[V ′]. It is easy to verify that T ′ is a subtree of the block treeof G (indeed, C′ ⊆ C, B′ ⊆ B, and, hence, T ′ = TG[C′ ∪ B′]). It is well knownthat all leaves of a block tree are blocks. Hence, all leaves of T ′ are containedin B. Moreover, the weight |

⋃b∈B′ b| of T ′ is at least Φ since every vertex in V ′

occurs in at least one block in B′. Finally, note that T ′ has at most 2k leavesand a maximum cut-vertex degree of k +1. Otherwise, since the addition of E′′

turns G[V ′] biconnected, by the above mentioned result, E′′ would contain atleast k + 1 links.

The subtree problem introduced in Lemma 5.2 can be solved in O(k3 · (|C|+|B|)) time by a straightforward adaption of the dynamic programming for Max-

imum d-Leaves Subtree. The details can be found in the correspondingdiploma thesis [20]. Since the block tree of a graph can be computed in lineartime [19], we arrive at the following theorem.

Theorem 5.3. In the case of a complete link set, Partial Biconnectivity

Augmentation can be solved in O(k3 · |V | + |E|) time.

19


5.3 Partial Strong Connectivity Augmentation

Here, we show that Partial Strong Connectivity Augmentation (PSCA)is W[2]-hard, which is defined as follows: Given a directed graph D = (V, A), aset of links A′ ⊆ V ×V , and two non-negative integers Φ, k, find a subset A′′ of A′

with |A′′| ≤ k such that (V, A ∪ A′′) contains a strongly connected componentwith at least Φ vertices. We give a parameterized reduction from the W[2]-hard Set Cover problem [4], which is defined as follows: Given a set S ={s1, s2, . . . , sn}, a collection of subsets of S, that is, C := {S1, S2, . . . , Sm}with Si ⊆ S for all i, and a non-negative integer k, decide whether there is asubset C′ of C with |C′| ≤ k such that

⋃S′∈C′ S′ = S.

We show first that, in the case that the link set A′ is incomplete, that is, A′

does not contain all possible arcs, PSCA is W[2]-hard. Then, we extend theproof to the case that A′ is complete.

Theorem 5.4. Partial Strong Connectivity Augmentation in the caseof incomplete link sets is W[2]-hard with respect to k.

Proof. Given a Set Cover-instance (S, C, k), we construct a PSCA-instance(D = (V, A), A′, Φ, k′) as follows (see Fig. 7 for an example): We add foreach element s ∈ S an element-vertex vs and for each subset Si in C a subset-vertex vSi

to V . Moreover, a special vertex r is added to V . Then, we add narcs to A, which connect all element-vertices to r. In addition, an arc froma subset-vertex vSi

to an element-vertex vs is added to A if s ∈ Si. Then,we set A′ := {(r, vSi

) | Si ∈ C}, Φ := k + n + 1, and k′ := k. Next, weprove that each Set Cover-solution C′ := {Sj1 , Sj2 , . . . , Sjk

} with 1 ≤ jl ≤m for all 1 ≤ l ≤ k one-to-one corresponds to a PSCA-solution with A′′ :={(r, vSj1

), (r, vSj2), . . . , (r, vSjk

)}.Let C′ := {Sj1 , Sj2 , . . . , Sjk

} be a solution of the Set Cover-instance.Clearly, adding the arc (r, vSjl

) for 1 ≤ l ≤ k results in a strongly connectedcomponent consisting of r, vSjl

, and all element-vertices that correspond to theelements contained in Sjl

. Since⋃

S′∈C′ S′ = S, we get a strongly connectedcomponent with vertex r, subset-vertices vSj1

, vSj2, . . . , vSjk

, and all element-vertices. Altogether, this component has 1 + k + n vertices.

Consider a PSCA-solution A′′ := {(r, vSj1), (r, vSj2

), . . . , (r, vSjk)}. Vertex r

is in the resulting strongly connected component. Moreover, since we can with-out loss of generality assume that no subset in C is an empty set, the subset-vertices vSj1

, vSj2, . . . , vSjk

are in this component. Note that no other subset-vertices can be in this connected component. Therefore, the remaining n verticesof this component are all element-vertices. Since an element-vertex vs in D isonly reachable from the subset-vertices that correspond to the subsets contain-ing element s, the subcollection C := {Sj1 , Sj2 , . . . , Sjk

} covers all elementsfrom S.

Next, we show PSCA remains W[2]-hard even if A is complete.

Theorem 5.5. Partial Strong Connectivity Augmentation in the caseof complete link sets is W[2]-hard with respect to k.

20


va vb vc vd ve

vS1vS2

vS3vS4

r

Figure 7: Example of the reduction from Set Cover to PSCA. Herein, S1 ={a, b, c}, S2 = {a, c, d}, S3 = {b, e}, and S4 = {d, e}. The arcs are drawn assolid lines and the links as dashed lines. In the case of an incomplete link set(Theorem 5.4), only the shown links are allowed to be part of the solution. Inthe case of a complete links set (Theorem 5.5), all possible links could be partof the solution. However, as shown in the proof of Theorem 5.5, we may assumethat an optimal solution for the PSCA instance consists only of a subset of theshown links, too.

Proof. The construction of a PSCA-instance (D = (V, A), A′, Φ, k′) from a Set

Cover-instance (S, C, k) is almost the same as in the proof of Theorem 5.4 (seeFig. 7 for an example). The only difference is that here A′ = V × V . Next weshow the equivalence between the solutions. The direction that a Set Cover-solution corresponds to a PSCA-solution can be shown in the same way as inthe proof of Theorem 5.4.

Given a PSCA-solution A′′ with |A′′| ≤ k, we assume without loss of gen-erality that A′′ is minimal, that is, there is no proper subset of A′′ whose ad-dition to A results in a strongly connected component with at least Φ ver-tices. Let Ar be the set of links in A′ that go from r to subset-vertices. Weclaim that A′′ ⊆ Ar. To show this claim, we introduce the following notations.Let r(u) for a vertex u be the set of vertices reachable from u in D includ-ing u. We say a link (u, v) is a shadow if there exists another link (u′, v′) ∈ A′

with u′ ∈ r(u) and v ∈ r(v′). We use Ap to denote the set of links (u, v) in A′

with v ∈ r(u) and As to denote the set of shadows. We can use the followingargument to show that A′′ ∩ (Ap ∪ As) = ∅: Suppose that there is some linkin Ap∪As contained in A′′. Let V ′ denote the vertex set of the resulting stronglyconnected component after adding A′′ to A with |V ′| ≥ Φ. Since A′′ is minimal,there are two vertices u, v ∈ V ′ such that the two paths between them contain alink (u, v) from Ap ∪As. In the case that (u, v) ∈ Ap, we can replace (u, v) by apath from u to v in D and remove (u, v) from A′′. If (u, v) ∈ As, then we can finda link (u′, v′) ∈ A′ with u′ ∈ r(u) and v ∈ r(v′). Then, we can replace (u, v)by (u′, v′). We can still find a path from u to v by passing through (u′, v′).

21


Thus, A′′ ∩ (Ap ∪ As) = ∅.In order to show A′′ ⊆ Ar, it remains to show A′ \ (As ∪Ap) = Ar. Clearly,

the links (vs, r) and (vSi, r) for element-vertices vs and subset-vertices vSi

areall shadows. Clearly, a link (r, s) for an element-vertex vs is a shadow of alink (r, vSi

) for a subset-vertex vSiwith s ∈ Si. Moreover, for all 1 ≤ i < j ≤ m,

link (vSi, vSj

) is a shadow of link (r, vSj), since r ∈ r(vSi

). Analogously, forall 1 ≤ i ≤ n and 1 ≤ j ≤ m, links (vsi

, vSj) are shadows of links (r, vSj

). Itis also easy to see that links (vsi

, vsj) for all 1 ≤ i < j ≤ n are shadows of

links (r, vsj). In the case that 1 ≤ i ≤ n and 1 ≤ j ≤ m, the links (vSi

, vsj) are

shadows of links (r, vsj). These are all possible links of (V ×V )\Ar. Therefore,

A′′ ⊆ Ar.With A′′ ⊆ Ar, the proof of the corresponding Set Cover-solution follows

analogously to the proof of Theorem 5.4.

6 Open Problems

The most interesting open problem is to study the parameterized complexity ofthe Strong Connectivity Augmentation problem, where we are given a di-rected graph D = (V, A), a set of links A ( V ×V , and a non-negative integer kand ask for a subset A′′ of links such that graph (V, A∪A′′) is strongly connected.Improving the size bounds of the problem kernels for Bridge-Connectivity

Augmentation and Biconnectivity Augmentation to a linear function in kis another interesting open problem. Further opportunities for future work in-clude investigating the approximability and fixed-parameter tractability of thePartial Bridge-Connectivity Augmentation and Partial Biconnec-

tivity Augmentation problems for the case that the link set E′ is incomplete.

Acknowledgement: We thank Rolf Niedermeier for inspiring discussion andhelpful comments. We are grateful to an anonymous Networks referee for provid-ing constructive feedback that helped to significantly improve the presentationof this work.

References

[1] J. Alber, N. Betzler, and R. Niedermeier. Experiments on data reduc-tion for optimal domination in networks. Annals of Operations Research,146(1):105–117, 2006.

[2] S. Bocker, S. Briesemeister, and G. W. Klau. Exact algorithms for clusterediting: Evaluation and experiments. Algorithmica, 2009. Accepted forpublication.

[3] F. Dehne, M. A. Langston, X. Luo, S. Pitre, P. Shaw, and Y. Zhang. Thecluster editing problem: Implementations and experiments. In Proc. 2ndIWPEC, volume 4196 of LNCS, pages 13–24. Springer, 2006.

22


[4] R. G. Downey and M. R. Fellows. Parameterized Complexity. Springer,1999.

[5] K. P. Eswaran and R. E. Tarjan. Augmentation problems. SIAM Journalon Computing, 5(4):653–665, 1976.

[6] G. Even, J. Feldman, G. Kortsarz, and Z. Nutov. A 3/2-approximationalgorithm for augmenting the edge-connectivity of a graph from 1 to 2using a subset of a given edge set. In Proc. 4th APPROX, volume 2129 ofLNCS, pages 90–101. Springer, 2001.

[7] G. Even, J. Feldman, G. Kortsarz, and Z. Nutov. A 1.8 approximationalgorithm for augmenting edge-connectivity of a graph from 1 to 2. ACMTransactions on Algorithms, 5(2), 2009.

[8] M. R. Fellows. The lost continent of polynomial time: Preprocessing andkernelization. In Proc. 2nd IWPEC, volume 4169 of LNCS, pages 276–277.Springer, 2006.

[9] J. Flum and M. Grohe. Parameterized Complexity Theory. Springer, 2006.

[10] G. N. Frederickson and J. JaJa. Approximation algorithms for severalgraph augmentation problems. SIAM Journal on Computing, 10(2):270–283, 1981.

[11] J. Guo and R. Niedermeier. Invitation to data reduction and problemkernelization. ACM SIGACT News, 38(1):31–45, 2007.

[12] F. Huffner, R. Niedermeier, and S. Wernicke. Techniques for practicalfixed-parameter algorithms. The Computer Journal, 51(1):7–25, 2008.

[13] S. Khuller and R. Thurimella. Approximation algorithms for graph aug-mentation. Journal of Algorithms, 14(2):214–225, 1993.

[14] G. Kortsarz, R. Krauthgamer, and J. R. Lee. Hardness of approximationfor vertex-connectivity network design problems. SIAM Journal on Com-puting, 33(3):704–720, 2004.

[15] M. Marzban, Q.-P. Gu, and X. Jia. Computational study on dominatingset problem of planar graphs. In Proc. 2nd COCOA, volume 5165 of LNCS,pages 89–102. Springer, 2008.

[16] H. Nagamochi. An approximation for finding a smallest 2-edge-connectedsubgraph containing a specified spanning tree. Discrete Applied Mathemat-ics, 126:83–113, 2003.

[17] R. Niedermeier. Invitation to Fixed-Parameter Algorithms. Oxford Univer-sity Press, 2006.

[18] A. Rosenthal and A. Goldner. Smallest augmentations to biconnect a graph.SIAM Journal on Computing, 6(1):55–66, 1977.

23


[19] R. E. Tarjan. Depth-first search and linear graph algorithms. SIAM Journalon Computing, 1(2):146–160, 1972.

[20] J. Uhlmann. Parameterized Algorithmics for Connectivity Augmenta-tion Problems in Networks. Diploma thesis, WSI fur Informatik, Uni-versitat Tubingen, Germany, 2007. URL: http://theinf1.informatik.uni-jena.de/∼uhlmann/diploma-thesis.pdf.

24

kernelization and complexity results for connectivity ... · np-completeness of both problems if...

Documents