Approximation Algorithm and FPT Algorithm for Connected-k-Subgraph Cover on Minor-Free Graphs

Abstract

Given a graph G, the minimum Connected-k-Subgraph Cover problem (MinCkSC) is to find a minimum vertex subset C of G such that every connected subgraph of G on k vertices has at least one vertex in C. If furthermore the subgraph of G induced by C is connected, then the problem is denoted as MinCkSC$_{con}$. In this paper, we first present a PTAS for MinCkSC on an H-minor-free graph, where H is a graph with a constant number of vertices. Then, we design an $O((\omega+1)(2(k-1)(\omega+2))^{3\omega+3})|V|$-time FPT algorithm for MinCkSC$_{con}$ on a graph with treewidth $\omega$, based on which we further design an $O(2^{O(\sqrt{t}\log t)}|V|^{O(1)})$ time subexponential FPT algorithm for MinCkSC$_{con}$ on an H-minor-free graph, where t is an upper bound of solution size.

1. Introduction

This paper studies the minimum connected-k-subgraph cover problem (MinCkSC) and its connected version (MinCkSC $_{con}$ ).

The MinCkSC problem was proposed because of a security consideration in a wireless sensor network (WSN, see e.g. Zhang et al. (2014): if an attacker knows at least k related information fragments in a WSN, then he can decode the full information; to protect information, we have to select some protected vertices such that each connected group on k vertices has at least one protected vertex; since protected vertices cost more, it is desired that the number of protected vertices is as small as possible. In the language of graph theory, given a graph $G=(V,E)$ and an integer k, the goal of MinCkSC is to find a minimum vertex subset C such that each connected component of $G-C$ contains at most $k-1$ vertices. The MinCkSC problem, under the name of the minimum k-vertex separator problem (MinkVS), was also studied for the sake of parallel computing by Guruswami et al. (2017), Lee (2017).

If furthermore it is required that the connected-k-subgraph cover C induces a connected subgraph, then the problem is denoted as MinCkSC $_{con}$ . Considering connectivity in this problem is because of information sharing issues in a WSN, see e.g. Li et al. (2016).

In this paper, we study approximation algorithms and FPT algorithms for MinCkSC and MinCkSC $_{con}$ on H-minor-free graphs, where $|V(H)|$ is upper bounded by a constant.

1.1 Related works

Note that there is a problem which is closely related with MinCkSC: the minimum k-path vertex cover problem (MinVCP $_k$ ). A vertex set C is called a VCP $_k$ set if every path on k vertices has at least one vertex in C. The MinVCP $_k$ problem was first proposed by Novotný (2010) also because of a consideration on the security of WSNs.

For $k=2,3$ , MinCkSC and MinVCP $_k$ coincide. In particular, both MinC2SC and MinVCP $_2$ are the classic minimum vertex cover problem (MinVC) in Bar-Yehuda et al. (1981); Hochbaum (1982), which still receives a lot of attention in recent years in Bar-Yehuda et al. (2017); Cabello et al. (2015); Cai et al. (2019); Khot et al. (2018); Pourhassan et al. (2019). Since MinCkSC and MinVCP $_k$ and their connected variants are all NP-hard for $k\geq 2$ , most researches with theoretical guarantees focus on approximation algorithms and FPT algorithms.

In the following, we distinguish the weighted version and the cardinality version of these problems by whether using “W” in their abbreviations, such as MinWCkSC denotes the weighted version while MinCkSC denotes the cardinality version.

Note that both MinWCkSC and MinWVCP $_k$ are special cases of the minimum weight set cover problem (MinWSC) by Liu et al. (2020); Zhang et al. (2014), so they admit k-approximation. Improvement on the ratio needs deeper exploration of graph structures. Compared with the extensive studies on the MinVCP $_k$ problem, see e.g. Bresar et al. (2011); Cabello et al. (2015); Kardoš et al. (2011); Ran et al. (2019); Tu et al. (2011); Tu et al. (2011); Zhang et al. (2018); Zhang et al. (2017), studies on MinCkSC are relatively less. The terminology MinWCkSC was proposed in Zhang et al. (2014), and a $(k-1)$ -approximation algorithm was designed on graphs with girth at least k. Recently, Liu et al. (2020) relaxed the girth requirement from k to $2k/3$ . For MinCkSC (the cardinality version), Lee (2017) presented an $O(\log k)$ -approximation algorithm which runs in time $2^{O(k^3 \log k)}n^2 \log n + n^{O(1)}$ . For a more general class of problems called $\mathcal H$ -free node-deletion problem which includes both MinVCP $_k$ and MinCkSC, Li et al. (2018) present a PTAS on disk graphs. Gupta et al. (2019) presented a PTAS for MinCkSC on H-minor-free graphs, using a bidimensionality method. Li et al. (2016) were the first to study the connectivity version of the problem, and provided a k-approximation algorithm for MinCkSC $_{con}$ .

Parameterized algorithms for MinVCP $_k$ mainly focus on small k. There are a lot of works on MinVCP $_3$ , see e.g. Chang et al. (2016); Katrenič (2016); Tu (2015); Xiao et al. (2017) (recall that MinVCP $_3$ coincides with MinC3SC). Currently, the best known running time of FPT algorithms for MinVC and MinVCP $_3$ in terms of solution size t are $O^*(1.2738^t)$ by Chen et al. (2010) and $O^*(1.713^t)$ by Tsur (2019), respectively. For the connected version with $k=2$ , Cygan (2012) presented an FPT algorithm for MinCVC with running time $O(2^t n^{O(1)})$ . Bai et al. (2019) studied parameterized algorithm in terms of treewidth $\omega$ , presenting a $3^\omega n^{O(1)}$ -time algorithm for MinVCP $_3$ and a $4^\omega n^{O(1)}$ -time randomized algorithm for MinCVCP $_3$ . Parameterized algorithms for MinVCP $_k$ with $k=4,5,6,7$ are emerging recently in Červený et al. (2019); Tsur (2021, 2019); Tu et al. (2016), but works on general k are rare. Note that MinVCP $_k$ and MinCkSC can be considered as special cases of a so-called graph modification problem studied by Cai (1996), which has been shown to have an $O^*(k^t)$ time FPT algorithm. To improve the time complexity, one has to explore graph structures of these problems. As far as we know, there is no study on parameterized algorithm for MinCkSC $_{con}$ or MinCVCP $_k$ for general k.

1.2 Contribution

In this paper, we study MinCkSC and MinCkSC $_{con}$ for general constant k. The contributions are summarized as follows.

1. (1) Using a local search method, we present a simple but effective approximation algorithm for MinCkSC on H-minor-free graphs, where the number of vertices in H is upper bounded by a constant, achieving approximation ratio $(1+\varepsilon)$ in time $n^{O(1/\varepsilon^2)}$ , where $\varepsilon$ is an arbitrary constant in (0,1).

2. (2) We present an $O(\omega(2(k-1)\omega)^{3\omega})|V|$ -time algorithm for MinCkSC $_{con}$ , where $\omega$ is the treewidth of the graph. The algorithm is a dynamic programming, based on a nice tree-decomposition. Note that connectivity is a global feature. The challenge of designing the algorithm lies in how to extend partial solutions step by step, keeping enough information to guarantee global connectivity, while at the same time limiting the table size of the dynamic programming.

3. (3) Based on the above result, we prove that MinCkSC $_{con}$ has an $O\Big(\sqrt{t}^{O(\sqrt{t})}|V|\Big)$ -time algorithm on H-minor-free graphs, where t is an upper bound for the solution size.

The remaining parts of this paper are organized as follows. A PTAS for MinCkSC is presented in Section 2. Section 3 designs a dynamic programming for MinCkSC $_{con}$ whose running time depends on the treewidth. Section 4 obtains an FPT algorithm for MinCkSC $_{con}$ on H-minor-free graphs. And Section 5 concludes the paper with some discussion on further work.

2. Approximation Algorithm for MinC $\boldsymbol{k}$ SC

Before presenting the PTAS for MinCkSC, we introduce some terminologies and give some preliminary results.

Definition 1. If graph H is obtained from G via vertex deletions, edge deletions, and edge contractions, then H is a minor of G. If H is not a minor of G, then G is called an H-minor-free graph.

Definition 2. Let $\mathcal S=\{S_1,\ldots,S_q\}$ be a collection of subsets of V(G) such that $V(G)=S_1\cup\cdots\cup S_q$ . A vertex $v\in S_i$ is said to be a boundary vertex of $S_i$ if v also belongs to some $S_j$ with $j\neq i$ . Denote by $B(S_i)$ the set of all boundary vertices of $S_i$ , and $int(S_i)=S_i\setminus B(S_i)$ the set of interior vertices of $S_i$ . Note that $int(S_i)$ is exactly the set of vertices in $S_i$ whose neighbors are completely contained in $S_i$ .

Definition 3. For a graph $G=(V,E)$ and a number $r>0$ , a collection $\mathcal{S}=\{S_1,S_2,\ldots,S_q\}$ with $V(G)=S_1\cup\cdots\cup S_q$ is called an r-division of G if it satisfies the following properties:

1. (a) every edge of G is completely contained in some induced subgraph $G[S_i]$ ;

2. (b) $|S_i|\leq r$ for every $i=1,\ldots,q$ ;

3. (c) $q\leq c_{\mathcal{S}}\frac{|V(G)|}{r}$ and $\sum\limits_{i=1}^q|B(S_i)|\leq c_{\mathcal{S}}\frac{|V(G)|}{\sqrt{r}}$ , where $c_{\mathcal{S}}$ is a number related to the structure of the graph, but is independent of $|V(G)|$ and r.

According to Alon et al. (1990) and Le et al. (2020), the following result holds.

Lemma 4. Any H-minor-free graph G has an r-division $\mathcal S$ with $c_{\mathcal{S}}=$ poly $(|V(H)|)$ .

For an H-minor-free graph, which is a generalization of planar graph, if $|V(H)|$ has a constant upper bound, which is the case in this paper, then $c_{\mathcal S}$ is a constant. Note that $c_{\mathcal S}$ can be computed in polynomial time by Cabello et al. (2015).

Algorithm 1 Local Search Algorithm for MinCkSC (LS-MinCkSC).

Theorem. Algorithm 1 runs in time $n^{O(\frac{1}{\varepsilon^2})}$ for an H-minor-free graph, where $|V(H)|$ is upper bounded by a constant.

Proof. Since in each iteration, the size of current solution will be reduced by at least 1, the while-loop will be executed at most n rounds. By the assumption that $|V(H)|$ is upper bounded by a constant, the number $c_{\mathcal S}$ is a constant. So, each round of the while-loop takes $n^{O(r)}$ time to enumerate F’ and F” whose sizes are upper bounded by r, and the total time complexity is $n^{O(\frac{1}{\varepsilon^2})}$ . The lemma is proved.

The following theorem analyzes the performance ratio of our algorithm. Let C denote the output solution of Algorithm 1, $C^*$ denote an optimal solution. The idea of the analysis is to construct an auxiliary graph $G_{a}$ and use its division to construct a feasible solution C’ such that $|C|\leq|C'|\leq(1+\varepsilon)|C^*|$ .

Theorem For any $\varepsilon\in(0,1)$ , Algorithm 1 outputs a $(1+\varepsilon)$ -approximate solution for MinCkSC on H-minor-free graphs, where $|V(H)|$ is upper bounded by a constant.

Proof We construct an auxiliary graph $G_{a}$ in the following way: first, let $G_1$ be the graph obtained from $G-(C\cap C^*)$ by removing all those connected components of cardinality smaller than k; then let $G_a=G[V(G_1)\cup(C\cap C^*)]$ .

Claim 1 $|V(G_a)|\leq k|C^*|$ .

Let $C(G_1)$ be a connected component of $G_1$ . Then, $|V(C(G_1))|\geq k$ . Since every connected component of $G-C$ has cardinality at most $k-1$ , and because $V(G_1)\cap (C\cap C^*)=\emptyset$ , we have $V(C(G_1)\cap(C\setminus C^*)\neq\emptyset$ . Similarly, $V(C(G_1))\cap(C^*\setminus C)\neq\emptyset$ . Let $C'(G_1)=C(G_1)-(V(C(G_1))\cap(C^*\setminus C))$ , then $C'(G_1)\subseteq G-C^*$ , and thus, $ |V(C'(G_1))|\leq k-1$ . It follows that

\begin{align*}|V(G_1)|&=\sum\limits_{C(G_1):~\mbox{component of}~G_1}|V(C(G_1))|\\&\leq(k-1)\omega (G_1)+\sum\limits_{C(G_1):~\mbox{component of}~G_1}|V(C(G_1))\cap(C^*\setminus C)|\\&\leq(k-1)\omega(G_1)+|C^*\setminus C|,\end{align*}

where $\omega(G_1)$ is the number of connected components of $G_1$ . Since $V(C(G_1))\cap(C^*\setminus C)\neq\emptyset$ holds for any connected component $C(G_1)$ of $G_1$ , we have $\omega (G_1)\leq|C^*\setminus C|$ . Hence,

\[|V(G_1)|\leq(k-1)|C^*\setminus C|+|C^*\setminus C|=k|C^*\setminus C|.\]

It follows that

\[|V(G_a)|\leq|V(G_1)|+|C^*\cap C|\leq k|C^*|.\]

Claim 1 is proved.

Obviously, $G_{a}$ is also H-minor-free. Let $\{S_1,S_2,\ldots,S_k\}$ be an r-division of $G_{a}$ and let

$$C'_i=(C\setminus S_i)\cup(C^*\cap S_i)\cup B(S_i)\ \mbox{for}\ i=1,\ldots,q.$$

Claim 2: Every $C'_i$ is a CkSC of G.

We prove the claim by contradiction. Suppose that there exists a connected-k-vertex subgraph G’ of G with $V(G')\cap C'_i=\emptyset$ . Since both C and $C^*$ are CkSC of G, we have $V(G')\cap C\neq\emptyset$ and $V(G')\cap C^*\neq\emptyset$ . Suppose $u\in V(G')\cap C$ and $v\in V(G')\cap C^*$ , because $(C\cap C^*)\subseteq C'_i$ and $B(S_i)\subseteq C'_i$ , we must have $u\in \big(C\cap int(S_i)\big)\setminus C^*$ and $v\in C^*\setminus(C\cup S_i)$ . Since G’ is a connected subgraph of G, there is a path $P=(u,v_1,\ldots,v_t,v)$ in G’. Since $u\in int(S_i)$ and $v\in C^*\setminus(C\cup S_i)$ is not in $S_i$ , there exists some index $j\in\{1,\ldots,t\}$ with $v_j\in B(S_i)$ . But then $V(G')\cap C_i'\supseteq V(G')\cap B(S_i)\neq\emptyset$ , contradicting the assumption on G’. So $C'_i$ is a CkSC of G. Claim 2 is proved.

Note that $C_i'$ can be written as $C_i'=(C\setminus F_i')\cup F_i''$ , where $F_i'=C\cap int(S_i)$ and $F_i''=C^*\cap S_i$ . Since $F_i'\subseteq S_i$ , we have $|F_i'|\leq |S_i|\leq r$ . So, the reason why C cannot be improved to $C_i'$ by the algorithm is because of $|C_i'|\geq |C|$ . Combining this with

$$|C'_i|\leq|C|-|C\cap S_i|+|(C^*\cap S_i)\cup B(S_i)|,$$

we have

(1) \begin{equation}|C\cap S_i|\leq|C^*\cap S_i|+|B(S_i)|.\end{equation}

It follows that

\begin{align*}|C|& \leq \sum\limits_{i=1}^q|C\cap S_i|\\& \leq \sum\limits_{i=1}^q|C^*\cap S_i|+ \sum\limits_{i=1}^q|B(S_i)|\\&\leq|C^*|+2\sum\limits_{i=1}^q|B(S_i)|\\&\leq|C^*|+2c_{\mathcal{S}}\frac{k|C^*|}{\sqrt{r}}\\&=(1+2c_{\mathcal{S}}\frac{k}{\sqrt{r}})|C^*|\\&\leq(1+\varepsilon)|C^*|,\end{align*}

where the first inequality holds because $V(G)=S_1\cup \cdots \cup S_q$ ; the second inequality comes from (1); the third inequality holds because $|C^*\cap S_i|=|C^*\cap int(S_i)|+|C^*\cap B(S_i)|\leq |C^*\cap int(S_i)|+|B(S_i)|$ and the sets $\{int(S_i)\}_{i=1,\ldots,q}$ are mutually disjoint; the fourth inequality makes use of condition (c) in the definition of r-division and $|V(G_a)|\leq k|C^*|$ by Claim 1; and the last inequality comes from the choice of $r\geq \left(\frac{2c_{\mathcal{S}}k}{\varepsilon}\right)^2$ .

3. Algorithm for MinC $\boldsymbol{k}$ SC $_{\boldsymbol{con}}$ in Terms of Treewidth

In this section, we provide a dynamic programming for MinCkSC $_{con}$ whose running time depends on the treewidth $\omega$ of graph G.

3.1 Basic notations

Firstly, we introduce the definition of tree decomposition.

Definition 5. (Tree decomposition) For a given graph $G=(V,E)$ , a pair $(\{X_i:i\in I\},T=(I,F))$ is a tree decomposition, if the following conditions hold:

1. (i) $\bigcup_{i\in I}X_i=V$ ,

2. (ii) every edge $e\in E$ belongs to some $G[X_i]$ ,

3. (iii) T is a tree on vertex set I and edge set F, such that $X_i\cap X_k\subseteq X_j$ for any j lying on the path from i to k in T.

Each $X_i$ is called a bag. The treewidth of G is defined as

$$\omega(G)=\min\limits_{\scriptsize \mbox{all tree decomposition}\atop (\{X_i:i\in I\},T=(I,F))}\max\limits_{i\in I}\big(\{|X_i|\}-1\big).$$

Our algorithm needs nice tree decomposition defined as follows. For the sake of clarity, we use nodes to refer to vertices of T and use vertices to refer to vertices of G.

Definition 6. (Nice tree decomposition) For a graph $G=(V,E)$ , a tree decomposition is nice if nodes in I can be partitioned into four types as follows:

1. (i) leaf node: a node in T without children;

2. (ii) introduce node: a node i which has exactly one child j such that $X_j\subseteq X_i$ and $X_i\setminus X_j=\{v\}$ for some vertex $v\in V$ ;

3. (iii) forget node: a node i which has exactly one child j such that $X_j=X_i\cup\{v\}$ for some vertex $v\in V$ ;

4. (iv) join node: a node i which has exactly two children j,l such that $X_i=X_j=X_l$ .

A nice tree decomposition is illustrated by Figure 1. Bodlaender proved in Bodlaender (1996) that for any graph with treewidth $\omega$ , a tree decomposition can be computed in time $O(2^{\Theta(\omega^3)}|V|)$ , and given a tree decomposition, one can transform it into a nice tree decomposition in time $O(|V|+|E|)$ . Note that each non-leaf bag is a separator of graph G by Diestel (2005).

Figure 1. An illustration of nice tree decomposition. The left most node is the root, and we use symbols JN,IN,FN and LN to indicate join node, introduce node, forget node, and leaf node, respectively.

3.2 MinCkSC $_{con}$ on bounded treewidth graphs

In this section, we present a dynamic programming for the computation of MinCkSC $_{con}$ .

For a nice tree decomposition $(\{X_i:i\in I\},T)$ , a subtree of T rooted at i which contains all descendants of i is denoted as $T_i$ . The subgraph of G induced by all vertices of those bags corresponding to those nodes of $T_i$ is denoted as $G_i=G[\bigcup\limits_{j\in{V(T_i)}}X_j]$ .

Status variables: For each node $i\in I$ , the dynamic programming will record a set of optimal values with respect to all possible tokens of the form $(F,P_F,P_{X_i\setminus F})$ , where $F\subseteq X_i$ , $P_F$ is a partition of F, $P_{X_i\setminus F}$ consists of a partition of $X_i\setminus F$ , and each part of $P_{X_i\setminus F}$ is associated with an integer smaller than k. Define status variable $c(i,F,P_F,P_{X_i\setminus F})$ to be the cardinality of a minimum vertex set $C\subseteq V(G_i)$ which is compatible with token $(F,P_F,P_{X_i\setminus F})$ , where compatible means

1. (i) $C\cap X_i=F$ ;

2. (ii) $P_F$ is the partition of F corresponding to those connected components of $G_i[C]$ ;

3. (iii) $P_{X_i\setminus F}$ consists of the partition of ${X_i\setminus F}$ corresponding to those connected components of $G_i-C$ which has non-empty intersection with $X_i$ , and for each part of the partition, the integer associated with this part is the cardinality of the connected component of $G_i-C$ containing this part.

We shall use $P_F(v)$ (resp. $P_{X_i\setminus F}(v)$ ) to denote the part of $P_F$ (resp. $P_{X_i\setminus F}$ ) which contains vertex v, and use $k_D$ to denote the integer associated with a part D of the partition of $P_{X_i\setminus F}$ .

Initial condition: The dynamic programming starts computing from leaf nodes. For a leaf node $i\in I$ , if F is compatible with token set $(F,P_F,P_{X_i\setminus F})$ , then $c(i,F,P_F,P_{X_i\setminus F})=|F|$ , otherwise $c(i,F,P_F,P_{X_i\setminus F})=\infty$ .

Transition formula: We distinguish the transition formula into three cases depending on the type of the internal node.

Introduce node: Suppose i is an introduce node. Let j be the only child of i and denote by v the unique vertex in $X_i\setminus X_j$ . A triple $(F,P_F,P_{X_i\setminus F})$ is said to be a valid token for node i if

$\bf (a^{(in)})$ every part of $P_{X_i\setminus F}$ has cardinality at most $k-1$ ,

and exactly one of the following two conditions is satisfied:

$\bf (b_1^{(in)})$ if $v\in F$ , then $N_G(v)\cap F\subseteq P_F(v)$ ;

$\bf (b_2^{(in)})$ if $v\not\in F$ , then $N_G(v)\cap (X_i\setminus F)\subseteq P_{X_i\setminus F}(v)$ .

Superscript (in) indicates that the condition is for an introduce node. Condition $(a^{(in)})$ is necessary for C to be a feasible CkSC $_{con}$ . The reason for condition $(b_1^{(in)})$ is because if $v\in F$ , then any neighbor of v in F must belong to the connected component of $G_i[C]$ containing v. Similar reason for condition $(b_2^{(in)})$ .

Note that in the case of introduce node, $G_j=G_i-v$ due to condition (3) in the definition of tree decomposition (Definition 5). For a valid token $(F,P_F,P_{X_i\setminus F})$ , we say that token $(F',P_{F'},P_{X_j\setminus F'})$ for node j is compatible with $(F,P_F,P_{X_i\setminus F})$ if one of the following two conditions are met:

$\bf (c_1^{(in)})$ If $v\in F$ , then $F'=F\setminus\{v\}$ , $P_F=\big(P_{F'}\setminus \{P_{F'}(u)\colon u\in F'\cap N_G(v)\}\big)\cup \{D\}$ where $D=\left(\bigcup_{u\in F'\cap N_G(v)}P_{F'}(u)\right)\cup\{v\}$ , and $P_{X_j\setminus F'}=P_{X_i\setminus F}$ .

$\bf (c_2^{(in)})$ If $v\not\in F$ , then $F'=F$ , $P_F=P_{F'}$ , the partition of $P_{X_i\setminus F}$ is $\big(P_{X_j\setminus F'}\setminus\{P_{X_j\setminus F'}(u)\colon u\in (X_j\setminus F')\cap N_G(v)\}\big)\cup\{D'\}$ , where $D'=\left(\bigcup_{u\in (X_j\setminus F')\cap N_G(v)}P_{X_j\setminus F'}(u)\right)\cup\{v\}$ , and $k_{D'}=1+\sum_{D\in\{P_{X_j\setminus F'}(u)\colon u\in (X_j\setminus F')\cap N_G(v)\}}k_D$ must satisfy $k_{D'}\leq k-1$ .

The compatibility condition is illustrated by Fig. 2. In this figure, $P_{F'}=\{D_1,D_2,D_3\}$ (note that there are two patches in F which are both labeled $D_1$ since they belong to a same connected component of $G_j[C]$ ) and the partition of $P_{X_i\setminus F'}$ is $\{D_4,D_5,D_6\}$ . For the case when $v\in F$ , the inclusion of vertex v merges the two connected components of $G_j[C]$ containing $D_1$ and $D_2$ into one connected component of $G_i[C]$ , the new partition $P_F=\{D,D_3\}$ where $D=D_1\cup D_2\cup\{v\}$ . While $P_{X_i\setminus F}$ remains the same as $P_{X_j\setminus F'}$ . For the case when $v\not\in F$ , the inclusion of vertex v will not affect F and $P_F$ , but may merge $D_4$ and $D_5$ into a new part $D'=D_4\cup D_5\cup \{v\}$ . So the new partition of $P_{X_i\setminus F}$ should be $\{D',D_6\}$ , and the integer associated with D’ equals $k_{D_4}+k_{D_5}+1$ , which is the cardinality of the merged connected component of $G_i-C$ . It should be emphasized that if $k_{D'}\geq k$ , then $G_i-C$ contains a connected subgraph of cardinality at least k, and thus C cannot be a feasible solution to the CkSC $_{con}$ instance. This is why the compatibility condition in $\bf(c_2^{(in)})$ requires $k_{D'}\leq k-1$ .

Figure 2. An illustration of compatibility condition for introduce node. The whole figure shows $G_i$ . Solid closed curves are connected components of $G_i[C]$ , dashed closed curves are connected components of $G_i-C$ .

The transition formula for a valid token $(F,P_F,P_{X_i\setminus F})$ of an introduce node i is

where $1_{v\in F}=1$ if $v\in F$ and $1_{v\in F}=0$ if $v\not\in F$ .

Forget node: Suppose i is a forget node, j is the only child of i, and v is the only vertex in $X_j\setminus X_i$ . A triple $(F,P_F,P_{X_i\setminus F})$ is said to be a valid token for node i if

$\bf (a^{(for)})$ each part of $P_{X_i\setminus F}$ has at most $k-1$ vertices,

and one of the following two conditions is satisfied:

$\bf (b_1^{(for)})$ all vertices in $N_G(v)\cap F$ belong to a same part of $P_F$ , or

$\bf (b_2^{(for)})$ all vertices in $N_G(v)\cap (X_i\setminus F)$ belong to a same part of $P_{X_i\setminus F}$ .

Superscript (for) indicates that the condition is for a forget node. Note that different from the case for introduce node, in which exactly one condition of $(b_1^{(in)})$ and $(b_2^{(in)})$ can occur, for the case of forget node, conditions $(b_1^{(for)})$ and $(b_2^{(for)})$ can both occur.

For the case of forget node, $G_i=G_j$ . For a valid token $(F,P_F,P_{X_i\setminus F})$ , we say that a token $(F',P_{F'},P_{X_i\setminus F'})$ for node j is compatible with $(F,P_F,P_{X_i\setminus F})$ if one of the following two conditions are met:

$\bf (c_1^{(for)})$ All vertices of $N_G(v)\cap F$ belong to a same part of $P_F$ , $F'=F\cup\{v\}$ , $P_F=\big(P_{F'}\setminus \{P_{F'}(v)\}\big)\cup \{P_{F'}(v)\setminus \{v\}\}$ , and $P_{X_i\setminus F}=P_{X_j\setminus F'}$ .

$\bf (c_2^{(for)})$ All vertices of $N_G(v)\cap (X_i\setminus F)$ belong to a same part of $P_{X_i\setminus F}$ , $F'=F$ , $P_{F'}=P_F$ , $P_{X_i\setminus F}=\big(P_{X_j\setminus F'}\setminus\{P_{X_j\setminus F'}(v)\}\big)\cup \{P_{X_j\setminus F'}(v)\setminus\{v\}\}$ , and $k_{P_{X_j\setminus F'}(v)}\leq k-1$ .

Compatibility for a forget node is illustrated by Fig. 3. For a valid token $(F,P_F,P_{X_i\setminus F})$ with $F=D_1\cup D_2$ , $P_F=\{F\}$ and $P_{X_i\setminus F}=\{D_3,D_4\}$ , a compatible token for node j is $(F',P_{F'},P_{X_j\setminus F'})$ with $F'=F\cup \{v\}$ , $P_{F'}=\{F'\}$ and $P_{X_j\setminus F'}=\{D_3,D_4\}$ . For a valid token $(F,P_F,P_{X_i\setminus F})$ with $F=D_1\cup D_2$ , $P_F=\{D_1,D_2\}$ and $P_{X_i\setminus F}=\{D_3\cup D_4\}$ , if $k_{D_3}+k_{D_4}+1\leq k-1$ , then a compatible token for node j is $(F',P_{F'},P_{X_j\setminus F'})$ with $F'=F$ , $P_{F'}=P_F$ , the partition of $P_{X_j\setminus F}$ is $\{D_3\cup D_4\cup\{v\}\}$ , and the integer associated part $D=D_3\cup D_4\cup \{v\}$ is $k_D=k_{D_3}+k_{D_4}+1$ .

Figure 3. An illustration of compatibility condition for forget node.

The transition formula for a valid token $(F,P_F,P_{X_i\setminus F})$ of a forget node i is

Join node: Suppose $i\in I$ is a join node with two children j and l. A triple $(F,P_F,P_{X_i\setminus F})$ is a valid token if each part of $P_{X_i\setminus F}$ has at most $k-1$ vertices.

Figure 4. An illustration of compatibility condition for join node. Note that $G_j-X_i$ and $G_l-X_i$ are disjoint.

In this case, $X_i=X_j=X_l$ and $G_i=G_j\cup G_l$ . Furthermore, by condition $(iii)$ of tree decomposition, it can be observed that

(2) \begin{equation}\mbox{$V(G_j)\cap V(G_l)$ is exactly $X_i$.}\end{equation}

For a valid token $(F,P_F,P_{X_i\setminus F})$ of node i, we say that a token $(F',P_{F'},P_{X_i\setminus F'})$ of node j and a token $(F'',P_{F''},P_{X_i\setminus F''})$ of node l are compatible with $(F,P_F,P_{X_i\setminus F})$ if the following conditions are satisfied:

$\bf (c^{(join)})$ $F=F'=F''$ ;

$\bf (d^{(join)})$ Every part of $P_{F'}$ (as well as every part of $P_{F''}$ ) is completely contained in some part of $P_F$ . Furthermore, if there are a part D’ of $P_{F'}$ and a part D” of $P_{F''}$ which have non-empty intersection, then D’ and D” belong to a same part of $P_F$ ;

$\bf (e^{(join)})$ Every part of $P_{X_j\setminus F'}$ (as well as every part of $P_{X_l\setminus F''}$ ) is completely contained in some part of $P_{X_i\setminus F}$ . If there is a part D’ of $P_{X_j\setminus F'}$ and a part D” of $P_{X_l\setminus F''}$ with $D'\cap D''\neq\emptyset$ , then $D=D'\cup D''$ is a part of $P_{X_i\setminus F}$ , and furthermore, the integer associated with D satisfies $k_D=k_{D'}+k_{D''}-|D'\cap D''|\leq k-1$ .

The compatibility of join node is illustrated in Fig. 4. In this figure, partitions $P_F=\{D_1\cup D_2\cup D_3\}$ , $P_{F'}=\{D_1\cup D_2,D_3\}$ and $P_{F''}=\{D_1,D_2\cup D_3\}$ satisfy condition $\bf (d^{(join)})$ . Note that $D_1$ and $D_2$ belong to a same part of $P_{F'}$ , indicating that $D_1$ and $D_2$ are connected in $G_j$ . Since $G_j$ is a subgraph of $G_i$ , so $D_1$ and $D_2$ are also connected in $G_i$ , and thus $D_1\cup D_2$ should be completely contained in some part of $P_F$ . The same is true for $D_2\cup D_3$ by considering $G_l$ . Furthermore, the part $D_1\cup D_2\in P_{F'}$ and the part $D_2\cup D_3\in P_{F''}$ have non-empty intersection $D_2$ , so they must belong to a same part of $P_F$ which contains $D_2$ , indicating that they are connected through $D_2$ into a connected component of $G_i=G_j\cup G_l$ . Similarly, partition $P_{X_j\setminus F'}=\{D_4\}$ , $P_{X_l\setminus F''}=\{D_4\cup D_5\}$ , and $P_{X_i\setminus F}=\{D_4\cup D_5\}$ satisfy $\bf (e^{(join)})$ as long as $k_{D_4}^{(j)}+k_{D_4\cup D_5}^{(l)}-|D_4|\leq k-1$ , where $k_{D_4}^{(j)}$ is the integer associated with part $D_4$ in the partition of $P_{X_j\setminus F'}$ and $k_{D_4\cup D_5}^{(l)}$ is the integer associated with part $D_4\cup D_5$ in the partition of $P_{X_l\setminus F''}$ . It should be noted that if we denote by $H_j$ the connected component of $G_j$ containing part $D_4\cup D_6$ , denote by $H_l$ the connected component of $G_l$ containing part $D_4\cup D_5$ , then $H_j$ and $H_l$ are merged into one connected component of $G_i$ , denoted as $H_i$ . Because of observation (2), we have $V(H_j)\cap V(H_l)=D_4$ , and thus the size of $H_i$ is $|H_j|+|H_l|-|D_4|=k_{D_4}^{(j)}+k_{D_4\cup D_6}^{(l)}-|D_4|$ . For the feasibility of the solution, we must require this size $\leq k-1$ .

The transition formula for a valid token $(F,P_F,P_{X_i\setminus F})$ of a join node i is

In the above, we only give out transition formula for valid tokens. If $(F,P_F,P_{X_i\setminus F})$ is an invalid token for node i, or there are no tokens for his child (or children) which are compatible with $(F,P_F,P_{X_i\setminus F})$ , then set

$$c(i,F,P_F,P_{X_i\setminus F})=\infty.$$

Theorem Given a graph $G=(V,E)$ with treewidth $\omega$ , starting from leaf nodes of a nice tree decomposition of G, recursively use the above transition formulas in a bottom-up manner, when the dynamic programming arrives at the root node r, the minimum vertex set over all possible tokens $(F,P_F,P_{X_r\setminus F})$ for r is an optimal solution for the MinCkSC $_{con}$ instance. The time complexity is $O((2^{\Theta(\omega^3)}+(\omega+1)(2(k-1)(\omega+2))^{3\omega+3})|V|)$ .

Proof. The correctness of the dynamic programming is already shown when we discuss compatibility for the transition formulas.

To study the time complexity, regard each node i to maintain a table $\mbox{Tab}(i)$ , which is used to record values $c(i,F,P_F,P_{X_i\setminus F})$ for all tokens $(F,P_F,P_{X_i\setminus F})$ . We have to estimate

1. (i) the number of tokens for node i, and

2. (ii) for each token $(F,P_F,P_{X_i\setminus F})$ , the time to compute $c(i,F,P_F,P_{X_i\setminus F})$ .

For task (i), it can be seen that there are at most $(2(\omega+2))^{\omega+1}$ different partitions: for each vertex $v\in X_i$ , assign a label $(l_1(v),l_2(v))$ to v, where $l_1(v)$ indicates whether v belongs to F or $X_i\setminus F$ , and $l_2(v)$ indicates which part does v belongs to; there are at most $|X_i|$ parts for $P_F$ and at most $|X_i|$ parts for $P_{X_i\setminus F}$ ; then the number of partitions follows from the observation that there are two choices for $l_1(v)$ , at most $|X_i|+1$ choices for $l_2(v)$ (note that either F or $X_i\setminus F$ might be empty), and $|X_i|\leq \omega+1$ . The set F is simply the union of those parts of $P_F$ . For each part D of $P_{X_i\setminus F}$ , there are at most $k-1$ choices for $k_D$ . So, the number of different tokens for node i is at most $O((2(k-1)(\omega+2))^{\omega+1})$ .

For task (ii), the time to compute $c(i,F,P_F,P_{X_i\setminus F})$ depends on the number of tokens compatible with $(F,P_F,P_{X_i\setminus F})$ , which in turn is also upper bounded by the number for task (i) estimated in the above. As to checking compatibility given a token for node i’s child (or two tokens for its two children), it can be accomplished in linear time. So, the time to compute $c(i,F,P_F,P_{X_i\setminus F})$ is $O((\omega+1)(2(k-1)(\omega+2))^{2\omega+2})$ .

To sum up, for each node i in a nice tree decomposition of G, the time to compute all elements in table $\mbox{Tab}(i)$ is $O((\omega+1)(2(k-1)(\omega+2))^{3\omega+3})$ . Since there are $O(|V|)$ nodes in a nice tree decomposition of G by Kloks (1994), and the time to compute a nice tree decomposition is $O(2^{\Theta(\omega^3)}|V|)$ in Bodlaender (1996), the total time for the dynamic programming follows.

As a consequence, we have the following result.

Corollary 7. The MinCkSC $_{con}$ problem is polynomial time solvable on a graph whose treewidth is bounded above by a constant.

4. Sub-exponential FPT algorithm for MinC $\boldsymbol{k}$ SC $_{\boldsymbol{con}}$ on $\boldsymbol{H}$ -minor-free graphs

In this section, we present an FPT algorithm for the following parameterized version of the MinCkSC $_{con}$ problem on an H-free graph G: For a fixed graph H, given an H-free graph G and an integer t, is there a CkSC $_{con}$ in G with size at most t? If the answer is yes, then the algorithm is able to find out an optimal solution in time f(t)p(n), where p(n) is a polynomial on the input size n, and f(t) is a function depending on t and H (but is independent of n).

The algorithm employs the following result which was due to Demaine and Hajiaghayi.

Lemma 8. (Demaine et al. (2008)). Any H-minor-free graph with treewidth at least $\omega$ has an $c(H)\omega\times c(H)\omega$ grid as a minor, where c(H) is a parameter dependending on H.

We could prove the following upper bound for the size of a CkSC on a $g\times g$ grid.

Lemma 9. For a moderately large g (compared with k), any CkSC on a $g\times g$ grid has size larger than $g^2/(2k)$ .

Proof. Suppose C is a CkSC of a $g\times g$ grid. Note that each $\lceil \sqrt{k}\rceil\times\lceil \sqrt{k}\rceil$ grid must contain a vertex of C. A $g\times g$ grid has at least $\lfloor\frac{g}{\lceil\sqrt k\rceil}\rfloor\times \lfloor\frac{g}{\lceil\sqrt k\rceil}\rfloor$ disjoint $\lceil k\rceil\times\lceil k\rceil$ grids. When g is moderately large, $\lfloor\frac{g}{\lceil\sqrt k\rceil}\rfloor\times \lfloor\frac{g}{\lceil\sqrt k\rceil}\rfloor > \frac{g^2}{2k}$ . Then the lemma follows.

As a consequence of Lemma 9, if the minimum CkSC $_{con}$ of graph G has at most $g^2/2k$ vertices, then G cannot have $g\times g$ grid as a minor. So, for the parameterized version of the CkSC $_{con}$ problem with parameter t, if the answer is yes, then G is $(\sqrt{2kt}\times\sqrt{2kt})$ -grid minor-free. By Lemma 8, the treewidth of G is smaller than $\sqrt{2kt}/c(H)$ . Combining this with Theorem 3.2, we could find an optimal solution in time $O\left(\left(2^{\Theta((\sqrt{2kt}/{c(H)})^3)}+\frac{\sqrt{2kt}}{c(H)}\left(2(k-1)\frac{\sqrt{2kt}}{c(H)}\right)^{3\frac{\sqrt{2kt}}{c(H)}}\right)|V|\right)$ . On the other hand, if the computed treewidth is larger than $\sqrt{2kt}/c(H)$ , then the answer to the instance must be no.

In particular, if $|V(H)|$ has a constant upper bound, then c(H) is a constant. Combining this with our assumption that k is a constant, we have the following result.

Theorem Suppose H is a graph with a constant number of vertices and k is a constant integer. The parameterized version of the MinCkSC $_{con}$ problem on an H-minor-free graph can be solved in time $O\Big(\sqrt{t}^{O(\sqrt{t})}|V|\Big)$ .

5. Conclusion

In this paper, we designed a local search algorithm for MinCkSC on H-minor-free graph, which achieves $(1+\varepsilon)$ -approximation; presented a dynamic programming for MinCkSC $_{con}$ the running time of which depends on the treewidth; and obtained an FPT algorithm for MinCkSC $_{con}$ on H-minor-free graphs. Our results hold for general constant k.

Note that this paper only focuses on the cardinality version of MinCkSC and MinCkSC $_{con}$ . Studies on their weighted variants remain to be further explored.