This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

Kyoto University, Kyoto, [email protected] https://orcid.org/0000-0003-3244-6915 JSPS KAKENHI Grant Numbers JP20H05793, JP20K19742, JP21H03499 Kyoto University, Kyoto, [email protected] Nagoya University, Nagoya, Japan and https://www.math.mi.i.nagoya-u.ac.jp/~otachi/ [email protected] https://orcid.org/0000-0002-0087-853X JSPS KAKENHI Grant Numbers JP18H04091, JP18K11168, JP18K11169, JP20H05793, JP21K11752. \CopyrightKobayashi, Nagano, Otachi \ccsdesc[100]Mathematics of computing Discrete mathematics Combinatorics Combinatorial algorithms \relatedversion \EventEditorsJohn Q. Open and Joan R. Access \EventNoEds2 \EventLongTitle42nd Conference on Very Important Topics (CVIT 2016) \EventShortTitleCVIT 2016 \EventAcronymCVIT \EventYear2016 \EventDateDecember 24–27, 2016 \EventLocationLittle Whinging, United Kingdom \EventLogo \SeriesVolume42 \ArticleNo23

Finding shortest non-separating and non-disconnecting paths

Yasuaki Kobayashi    Shunsuke Nagano    Yota Otachi
Abstract

For a connected graph G=(V,E)G=(V,E) and s,tVs,t\in V, a non-separating ss-tt path is a path PP between ss and tt such that the set of vertices of PP does not separate GG, that is, GV(P)G-V(P) is connected. An ss-tt path is non-disconnecting if GE(P)G-E(P) is connected. The problems of finding shortest non-separating and non-disconnecting paths are both known to be NP-hard. In this paper, we consider the problems from the viewpoint of parameterized complexity. We show that the problem of finding a non-separating ss-tt path of length at most kk is W[1]-hard parameterized by kk, while the non-disconnecting counterpart is fixed-parameter tractable parameterized by kk. We also consider the shortest non-separating path problem on several classes of graphs and show that this problem is NP-hard even on bipartite graphs, split graphs, and planar graphs. As for positive results, the shortest non-separating path problem is fixed-parameter tractable parameterized by kk on planar graphs and polynomial-time solvable on chordal graphs if kk is the shortest path distance between ss and tt.

keywords:
Non-separating path, Parameterized complexity

1 Introduction

Lovász’s path removal conjecture states the following claim: There is a function f:f\colon\mathbb{N}\to\mathbb{N} such that for every f(k)f(k)-connected graph GG and every pair of vertices uu and vv, GG has a path PP between uu and vv such that GV(P)G-V(P) is kk-connected. This claim still remains open, while some spacial cases have been resolved [4, 15, 16, 22]. Tutte [22] proved that f(1)=3f(1)=3, that is, every triconnected graph satisfies that for every pair of vertices, there is a path between them whose removal results a connected graph. Kawarabayashi et al. [15] proved a weaker version of this conjecture: There is a function f:f\colon\mathbb{N}\to\mathbb{N} such that for every f(k)f(k)-connected graph GG and every pair of vertices uu and vv, GG has an induced path PP between uu and vv such that GE(P)G-E(P) is kk-connected.

As a practical application, let us consider a network represented by an undirected graph GG, and we would like to build a private channel between a specific pair of nodes ss and tt. For some security reasons, the path used in this channel should be dedicated to the pair ss and tt, and hence all other connections do not use all nodes and/or edges on this path while keeping their connections. In graph-theoretic terms, the vertices (resp. edges) of the path between ss and tt does not form a separator (resp. a cut) of GG. Tutte’s result [22] indicates that such a path always exists in triconnected graphs, but may not exist in biconnected graphs. In addition to this connectivity constraint, the path between ss and tt is preferred to be short due to the cost of building a private channel. Motivated by such a natural application, the following two problems are studied.

Definition 1.1.

Given a connected graph GG, s,tV(G)s,t\in V(G), and an integer kk, Shortest Non-Separating Path asks whether there is a path PP between ss and tt in GG such that the length of PP is at most kk and GV(P)G-V(P) is connected.

Definition 1.2.

Given a connected graph GG, s,tV(G)s,t\in V(G), and an integer kk, Shortest Non-Disconnecting Path asks whether there is a path PP between ss and tt in GG such that the length of PP is at most kk and GE(P)G-E(P) is connected.

We say that a path PP is non-separating (in GG) if GV(P)G-V(P) is connected and is non-disconnecting (in GG) if GE(P)G-E(P) is connected.


Related work. The shortest path problem in graphs is one of the most fundamental combinatorial optimization problems, which is studied under various settings. Indeed, our problems Shortest Non-Separating Path and Shortest Non-Disconnecting Path can be seen as variants of this problem. From the computational complexity viewpoint, Shortest Non-Separating Path is known to be NP-hard and its optimization version cannot be approximated with factor |V|1ε|V|^{1-\varepsilon} in polynomial time for ε>0\varepsilon>0 unless P == NP [23]. Shortest Non-Disconnecting Path is shown to be NP-hard on general graphs and polynomial-time solvable on chordal graphs [18].


Our results. We investigate the parameterized complexity of both problems. We show that Shortest Non-Separating Path is W[1]-hard and Shortest Non-Disconnecting Path is fixed-parameter tractable parameterized by kk. A crucial observation for the fixed-parameter tractability of Shortest Non-Disconnecting Path is that the set of edges in a non-disconnecting path can be seen as an independent set of a cographic matroid. By applying the representative family of matroids [11], we show that Shortest Non-Disconnecting Path can be solved in 2ωk|V|O(1)2^{\omega k}|V|^{O(1)} time, where ω\omega is the exponent of the matrix multiplication. We also show that Shortest Non-Disconnecting Path is OR-compositional, that is, there is no polynomial kernelization unless coNP \subseteq NP//poly. To cope with the intractability of Shortest Non-Separating Path, we consider the problem on planar graphs and show that it is fixed-parameter tractable parameterized by kk. This result can be generalized to wider classes of graphs which have the diameter-treewidth property [9]. We also consider Shortest Non-Separating Path on several classes of graphs. We can observe that the complexity of Shortest Non-Separating Path is closely related to that of Hamiltonian Cycle (or Hamiltonian Path with specified end vertices). This observation readily proves the NP-completeness of Shortest Non-Separating Path on bipartite graphs, split graphs, and planar graphs. For chordal graphs, we devise a polynomial-time algorithm for Shortest Non-Separating Path for the case where kk is the shortest path distance between ss and tt.

2 Preliminaries

We use standard terminologies and known results in matroid theory and parameterized complexity theory, which are briefly discussed in this section. See [6, 20] for details.


Graphs. Let GG be a graph. The vertex set and edge set of GG are denoted by V(G)V(G) and E(G)E(G), respectively. For vV(G)v\in V(G), the open neighborhood of vv in GG is denoted by NG(v)N_{G}(v) (i.e., NG(v)={uV(G):{u,v}E(G)}N_{G}(v)=\{u\in V(G):\{u,v\}\in E(G)\}) and the closed neighborhood of vv in GG is denoted by NG[v]N_{G}[v] (i.e., NG[v]=NG(v){v}N_{G}[v]=N_{G}(v)\cup\{v\}). We extend this notation to sets: NG(X)=vXNG(v)XN_{G}(X)=\bigcup_{v\in X}N_{G}(v)\setminus X and NG[X]=NG(X)XN_{G}[X]=N_{G}(X)\cup X for XV(G)X\subseteq V(G). For u,vV(G)u,v\in V(G), we denote by distG(u,v){\rm dist}_{G}(u,v) the length of a shortest path between uu and vv in GG, where the length of a path is defined to the number of edges in it. We may omit the subscript of GG from these notations when no confusion arises. For XV(G)X\subseteq V(G), we write G[X]G[X] to denote the subgraph of GG induced by XX. For notational convenience, we may use GXG-X instead of G[V(G)X]G[V(G)\setminus X]. For FEF\subseteq E, we also use GFG-F to represent the subgraph of GG consisting all vertices of GG and all edges in EFE\setminus F. For vertices uu and vv, a path between uu and vv is called a uu-vv path. A vertex is called a pendant if its degree is exactly 11.


Matroids and representative sets. Let EE be a finite set. If 2E\mathcal{I}\subseteq 2^{E} satisfies the following axioms, the pair =(E,)\mathcal{M}=(E,\mathcal{I}) is called a matroid: (1) \emptyset\in\mathcal{I}; (2) YY\in\mathcal{I} implies XX\in\mathcal{I} for XY2EX\subseteq Y\subseteq 2^{E}; and (3) for X,YX,Y\in\mathcal{I} with |X|<|Y||X|<|Y|, there is eYXe\in Y\setminus X such that X{e}X\cup\{e\}\in\mathcal{I}. Each set in \mathcal{I} is called an independent set of \mathcal{M}. From the third axiom of matroids, it is easy to observe that every (inclusion-wise) maximal independent set of \mathcal{M} have the same cardinality. The rank of \mathcal{M} is the maximum cardinality of an independent set of \mathcal{M}. A matroid =(E,)\mathcal{M}=(E,\mathcal{I}) of rank nn is linear (or representable) over a field 𝔽\mathbb{F} if there is a matrix M𝔽n×|E|M\in\mathbb{F}^{n\times|E|} whose columns are indexed by EE such that XX\in\mathcal{I} if and only if the set of columns indexed by XX is linearly independent in MM.

Let G=(V,E)G=(V,E) be a graph. A cographic matroid of GG is a matroid (G)=(E,)\mathcal{M}(G)=(E,\mathcal{I}) such that FEF\subseteq E is an independent set of (G)\mathcal{M}(G) if and only if GFG-F is connected. Every cographic matroid is linear and its representation can be computed in polynomial time [20].

Our algorithmic result for Shortest Non-Disconnected Path is based on representative families due to [11].

Definition 2.1.

Let =(E,)\mathcal{M}=(E,\mathcal{I}) be a matroid and let \mathcal{F}\subseteq\mathcal{I} be a family of independent sets of \mathcal{M}. For an integer q0q\geq 0, we say that ^\widehat{\mathcal{F}}\subseteq\mathcal{F} is qq-representative for \mathcal{F} if the following condition holds: For every YEY\subseteq E of size at most qq, if there is XX\in\mathcal{F} with XY=X\cap Y=\emptyset such that XYX\cup Y\in\mathcal{I}, then there is X^^\widehat{X}\in\widehat{\mathcal{F}} with X^Y=\widehat{X}\cap Y=\emptyset such that X^Y\widehat{X}\cup Y\in\mathcal{I}.

Theorem 2.2 ([11, 17]).

Given a linear matroid =(E,)\mathcal{M}=(E,\mathcal{I}) of rank nn that is represented as a matrix M𝔽n×|E|M\in\mathbb{F}^{n\times|E|} for some field 𝔽\mathbb{F}, a family \mathcal{F}\subseteq\mathcal{I} of independent sets of size pp, and an integer qq with p+qnp+q\leq n, there is a deterministic algorithm computing a qq-representative family ^\widehat{\mathcal{F}}\subseteq\mathcal{F} of size np(p+qp)np\binom{p+q}{p} with

O(||((p+qp)p3n2+(p+qq)ω1(pn)ω1))+(n+|E|)O(1)\displaystyle O\left(|\mathcal{F}|\cdot\left(\binom{p+q}{p}p^{3}n^{2}+\binom{p+q}{q}^{\omega-1}\cdot(pn)^{\omega-1}\right)\right)+(n+|E|)^{O(1)}

field operations, where ω<2.373\omega<2.373 is the exponent of the matrix multiplication.


Parameterized complexity. A parameterized problem is a language LΣ×L\subseteq\Sigma^{*}\times\mathbb{N}, where Σ\Sigma is a finite alphabet. We say that LL is fixed-parameter tractable (parameterized by kk) if there is an algorithm deciding if (I,k)L(I,k)\in L for given (I,k)Σ×(I,k)\in\Sigma^{*}\times\mathbb{N} in time f(k)|I|O(1)f(k)|I|^{O(1)}, where ff is a computable function. A kernelization for LL is a polynomial-time algorithm that given an instance (I,k)Σ×(I,k)\in\Sigma^{*}\times\mathbb{N}, computes an “equivalent” instance (I,k)Σ×(I^{\prime},k^{\prime})\in\Sigma^{*}\times\mathbb{N} such that (1) (I,k)L(I,k)\in L if and only if (I,k)L(I^{\prime},k^{\prime})\in L and (2) |I|+kg(k)|I^{\prime}|+k^{\prime}\leq g(k) for some computable function gg. We call (I,k)(I^{\prime},k^{\prime}) a kernel. If the function gg is a polynomial, then the kernelization algorithm is called a polynomial kernelization and its output (I,k)(I^{\prime},k^{\prime}) is called a polynomial kernel. An OR-composition is an algorithm that given pp instances (I1,k),(Ip,k)Σ×(I_{1},k),\ldots(I_{p},k)\in\Sigma^{*}\times\mathbb{N} of LL, computes an instance (I,k)Σ×(I^{\prime},k^{\prime})\in\Sigma^{*}\times\mathbb{N} in time (1ip|Ii|+k)O(1)(\sum_{1\leq i\leq p}|I_{i}|+k)^{O(1)} such that (1) (I,k)L(I^{\prime},k^{\prime})\in L if and only if (Ii,k)L(I_{i},k)\in L for some 1ip1\leq i\leq p and (2) k=kO(1)k^{\prime}=k^{O(1)}. For a parameterized problem LL, its unparameterized problem is a language L={x#1k:(x,k)L}L^{\prime}=\{x\#1^{k}:(x,k)\in L\}, where#Σ\#\notin\Sigma is a blank symbol and 1Σ1\in\Sigma is an arbitrary symbol.

Theorem 2.3 ([3]).

If a parameterized problem LL admits an OR-composition and its unparameterized version is NP-complete, then LL does not have a polynomial kernelization unless coNP \subseteq NP//poly.

3 Shortest Non-Separating Path

We discuss our complexity and algorithmic results for Shortest Non-Separating Path.

3.1 Hardness on graph classes

We obverse that, in most cases, Shortest Non-Separating Path is NP-hard on classes of graphs for which Hamiltonian Path (with distinguished end vertices) is NP-hard. Let G=(V,E)G=(V,E) be a graph and s,tVs,t\in V be distinct vertices of GG. We add a pendant vertex pp adjacent to ss and denote the resulting graph by GG^{\prime}. Then, we have the following observation.

{observation}

For every non-separating path PP between ss and tt in GG^{\prime}, V(G)V(P)={p}V(G)\setminus V(P)=\{p\}.

Suppose that for a class 𝒞\mathcal{C} of graphs,

  • the problem of deciding whether given graph G𝒞G\in\mathcal{C} has a Hamiltonian path between specified vertices ss and tt in GG is NP-hard and

  • G𝒞G\in\mathcal{C} implies G𝒞G^{\prime}\in\mathcal{C}.

By Section 3.1, GG^{\prime} has a non-separating ss-tt path if and only if GG has a Hamiltonian path between ss and tt. This implies that the problem of finding a non-separating path between specified vertices is NP-hard on class 𝒞\mathcal{C}.

Theorem 3.1.

The problem of deciding if an input graph has a non-separating ss-tt path is NP-complete even on planar graphs, bipartite graphs, and split graphs.

The classes of planar graphs and bipartite graphs are closed under the operation of adding a pendant. Recall that a graph GG is a split graph if the vertex set V(G)V(G) can be partitioned into a clique CC and an independent set II. Thus, for the class of split graphs, we need the assumption that the pendant added is adjacent to a vertex in CC.

As the problem trivially belongs to NP, it suffices to show that Hamiltonian Path (with distinguished end vertices) is NP-hard on these classes of graphs111These results for bipartite graphs and planar graphs seem to be folklore but we were not able to find particular references.. For split graphs, it is known that Hamiltonian Path is NP-hard even if the distinguished end vertices are contained in the clique CC [19]. Let GG be a graph and let vV(G)v\in V(G). We add a vertex vv^{\prime} that is adjacent to every vertex in NG(v)N_{G}(v), that is, vv and vv^{\prime} are twins. The resulting graph is denoted by GG^{\prime}. It is easy to verify that GG has a Hamiltonian cycle if and only if GG^{\prime} has a Hamiltonian path between vv and vv^{\prime}. The class of bipartite graphs is closed under this operation, that is, GG^{\prime} is bipartite. For planar graphs, GG^{\prime} may not be planar in general. However, Hamiltonian Cycle is NP-complete even if the input graph is planar and has a vertex of degree 22 [14]. We apply the above operation to this degree-22 vertex, and the resulting graph GG^{\prime} is still planar. As the problem of finding a Hamiltonian cycle is NP-hard even on bipartite graphs [19] and planar graphs [14], Theorem 3.1 follows.

3.2 W[1]-hardness

Next, we show that Shortest Non-Separating Path is W[1]-hard parameterized by kk. The proof is done by giving a reduction from Multicolored Clique, which is known to be W[1]-complete [10]. In Multicolored Clique, we are given a graph GG with a partition {V1,V2,,Vk}\{V_{1},V_{2},\ldots,V_{k}\} of V(G)V(G) and asked to determine whether GG has a clique CC such that |ViC|=1|V_{i}\cap C|=1 for each 1ik1\leq i\leq k.

From an instance (G,{V1,,Vk})(G,\{V_{1},\ldots,V_{k}\}) of Multicolored Clique, we construct an instance of Shortest Non-Separating Path as follows. Without loss of generality, we assume that GG contains more than kk vertices. We add two vertices ss and tt and edges between ss and all vV1v\in V_{1} and between tt and all vVkv\in V_{k}. For any pair of uViu\in V_{i} and vVjv\in V_{j} with |ij|2|i-j|\geq 2, we do the following. If {u,v}E\{u,v\}\in E, then we remove it. Otherwise, we add a path Pu,vP_{u,v} of length 22 and a pendant vertex that is adjacent to the internal vertex ww of Pu,vP_{u,v}. Finally, we add a vertex vv^{*}, add an edge between vv^{*} and each original vertex vV(G)v\in V(G), and add a pendant vertex pp adjacent to vv^{*}. The constructed graph is denoted by HH. See Figure 1 for an illustration of the graph HH.

Refer to caption
Figure 1: The left figure depicts an instance GG of Multicolored Clique and the right figure depicts the graph HH constructed from GG. Some vertices and edges in HH are not drawn in this figure for visibility. The edges of a clique CC and the corresponding non-separating ss-tt path PP are drawn as thick lines.
Lemma 3.2.

There is a clique CC in GG such that |CVi|=1|C\cap V_{i}|=1 for 1ik1\leq i\leq k if and only if there is a non-separating ss-tt path of length at most k+1k+1 in HH.

Proof 3.3.

Suppose first that GG has a clique CC with CVi={vi}C\cap V_{i}=\{v_{i}\} for 1ik1\leq i\leq k. Then, P=s,v1,v2,,vk,tP={\langle s,v_{1},v_{2},\ldots,v_{k},t\rangle} is an ss-tt path of length k+1k+1 in HH. To see the connectivity of HV(P)H-V(P), it suffices to show that every vertex is reachable to vv^{*} in HV(P)H-V(P). By the construction of HH, each vertex in V(G)V(P)V(G)\setminus V(P) is adjacent to vv^{*} in HV(P)H-V(P). Each vertex zz in V(H)(V(G){v,p})V(H)\setminus(V(G)\cup\{v^{*},p\}) is either the internal vertex ww of Pu,vP_{u,v} for some u,vV(G)u,v\in V(G) or the pendant vertex adjacent to ww. In both cases, at least one of uu and vv is not contained in PP as V(P){s,t}V(P)\setminus\{s,t\} is a clique in GG, implying that zz is reachable to vv^{*}.

Conversely, suppose that HH has a non-separating ss-tt path PP of length at most k+1k+1 in HH. By the assumption that GG has more than kk vertices, there is a vertex of GG that does not belong to PP. Observe that PP does not contain any internal vertex ww of some Pu,vP_{u,v} as otherwise the pendant vertex adjacent to ww becomes an isolated vertex by deleting V(P)V(P), which implies HV(P)H-V(P) has at least two connected components. Similarly, PP does not contain vv^{*}. These facts imply that the internal vertices of PP belong to V(G)V(G), and we have |V(P)Vi|=1|V(P)\cap V_{i}|=1 for 1ik1\leq i\leq k. Let C=V(P){s,t}C=V(P)\setminus\{s,t\}. We claim that CC is a clique in GG. Suppose otherwise. There is a pair of vertices u,vCu,v\in C that are not adjacent in GG. This implies that HH contains the path Pu,vP_{u,v}. However, as PP contains both uu and vv, the internal vertex of Pu,vP_{u,v} together with its pendant vertex forms a component in HV(P)H-V(P), yielding a contradiction that PP is a non-separating path in HH.

Thus, we have the following theorem.

Theorem 3.4.

Shortest Non-Separating Path is W[1]-hard parameterized by kk.

3.3 Graphs with the diameter-treewidth property

By Theorem 3.4, Shortest Non-Separating Path is unlikely to be fixed-parameter tractable on general graphs. To overcome this intractability, we focus on sparse graph classes. We first note that algorithmic meta-theorems for FO Model Checking [12, 13] does not seem to be applied to Shortest Non-Separating Path as we need to care about the connectivity of graphs, while it can be expressed by a formula in MSO logic, which is as follows. The property that vertex set XX forms a non-separating ss-tt path can be expressed as:

𝚌𝚘𝚗𝚗(VX)𝚑𝚊𝚖𝚙𝚊𝚝𝚑(X,s,t),\displaystyle{\tt conn}(V\setminus X)\land{\tt hampath}(X,s,t),

where 𝚌𝚘𝚗𝚗(Y){\tt conn}(Y) and 𝚑𝚊𝚖𝚙𝚊𝚝𝚑(Y,s,t){\tt hampath}(Y,s,t) are formulas in MSO2 that are true if and only if the subgraph induced by YY is connected and has a Hamiltonian path between ss and tt, respectively. We omit the details of these formulas, which can be found in [6] for example222In [6], they give an MSO2 sentence 𝚑𝚊𝚖𝚒𝚕𝚝𝚘𝚗𝚒𝚌𝚒𝚝𝚢{\tt hamiltonicity} expressing the property of having a Hamiltonian cycle, which can be easily transformed into a formula expressing 𝚑𝚊𝚖𝚙𝚊𝚝𝚑(X,s,t){\tt hampath}(X,s,t).. By Courcelle’s theorem [5] and its extension due to Arnborg et al. [1], we can compute a shortest non-separating ss-tt path in O(f(tw(G))n)O(f({\rm tw}(G))n) time, where nn is the number of vertices and tw(G){\rm tw}(G) is the treewidth333We do not give the definition of treewidth and (the optimization version of) Courcelle’s theorem. We refer to [6] for details. of GG. As there is an O(tw(G)tw(G)3n)O(\mathrm{tw}(G)^{\mathrm{tw}(G)^{3}}n)-time algorithm for computing the treewidth of an input graph GG [2], we have the following theorem.

Theorem 3.5.

Shortest Non-Separating Path is fixed-parameter tractable parameterized by the treewidth of input graphs.

A class 𝒞\mathcal{C} of graphs is minor-closed if every minor of a graph G𝒞G\in\mathcal{C} also belongs to 𝒞\mathcal{C}. We say that 𝒞\mathcal{C} has the diameter-treewidth property if there is a function f:f\colon\mathbb{N}\to\mathbb{N} such that for every G𝒞G\in\mathcal{C}, the treewidth of GG is upper bounded by f(diam(G))f({\rm diam}(G)), where diam(G){\rm diam}(G) is the diameter of GG. It is well known that every planar graph GG has treewidth at most 3diam(G)+13\cdot{\rm diam}(G)+1 [21]444More precisely, the treewidth of a planar graph is upper bounded by 3r+13r+1, where rr is the radius of the graph., which implies that the class of planar graphs has the diameter-treewidth property. This can be generalized to more wider classes of graphs. A graph is called an apex graph if it has a vertex such that removing it makes the graph planar.

Theorem 3.6 ([7, 9]).

Let 𝒞\mathcal{C} be a minor-closed class of graphs. Then, 𝒞\mathcal{C} has the diameter-treewidth property if and only if it excludes some apex graph.

For CV(G)C\subseteq V(G) that induces a connected subgraph G[C]G[C], we denote by GCG_{C} the graph obtained from GG by contracting all edges in G[C]G[C] and by vCv_{C} the vertex corresponding to CC in GCG_{C}. Since G[C]G[C] is connected, vertex vCv_{C} is well-defined.

Lemma 3.7.

Let CV(G)C\subseteq V(G) be a vertex subset such that G[C]G[C] is connected. Let PP be an ss-tt path in GG with V(P)C=V(P)\cap C=\emptyset. Then, PP is non-separating in GG if and only if it is non-separating in GCG_{C}.

Proof 3.8.

Suppose first that PP is non-separating in GG. Let u,vV(G)V(P)u,v\in V(G)\setminus V(P) be arbitrary. As PP is non-separating, there is a uu-vv path PP^{\prime} in GV(P)G-V(P). Let uu^{\prime} be the vertex of GCG_{C} such that u=uu^{\prime}=u if uCu\notin C and u=vCu^{\prime}=v_{C} if uCu\in C. Let vv^{\prime} be the vertex defined analogously. We show that there is a uu^{\prime}-vv^{\prime} path in GCV(P)G_{C}-V(P) as well. If PP^{\prime} does not contain any vertex in CC, it is also a uu^{\prime}-vv^{\prime} path in GCG_{C}, and hence we are done. Suppose otherwise. Let xx and yy be the vertices in V(P)CV(P^{\prime})\cap C that are closest to uu and vv, respectively. Note that xx and yy can be the end vertices of PP^{\prime}, that is, CC may contain uu and vv. Let Pu,xP_{u,x} and (resp. Py,vP_{y,v}) be the subpath of PP^{\prime} between uu and xx (resp. yy and vv). Then, the sequence of vertices obtained by concatenating Pu,xP_{u,x} after Py,v{y}P_{y,v}-\{y\} and replacing exactly one occurrence of a vertex in CC with vCv_{C} forms a path between uu^{\prime} and vv^{\prime}. Since we choose u,vu,v arbitrarily, there is a path between any pair of vertices in GCV(P)G_{C}-V(P) as well. Hence, PP is non-separating in GCG_{C}.

Conversely, suppose that PP is non-separating in GCG_{C}. For u,vV(GC)V(P)u,v\in V(G_{C})\setminus V(P), there is a path PP^{\prime} in GCV(P)G_{C}-V(P). Suppose that neither u=vCu=v_{C} nor v=vCv=v_{C}. Then, we can construct a uu-vv path in GV(P)G-V(P) as follows. If vCV(P)v_{C}\notin V(P^{\prime}), PP^{\prime} is also a path in GV(P)G-V(P) and hence we are done. Otherwise, vCV(P)v_{C}\in V(P^{\prime}). Let PuP_{u} and PvP_{v} be the subpaths in P{vC}P^{\prime}-\{v_{C}\} containing uu and vv, respectively. From PuP_{u} and PvP_{v}, we have a uu-vv path in GG by connecting them with an arbitrary path in G[C]G[C] between the end vertices other than uu and vv. Note that such a bridging path in G[C]G[C] always exists since G[C]G[C] is connected. Moreover, as V(P)C=V(P^{\prime})\cap C=\emptyset and V(P)C=V(P)\cap C=\emptyset, this is also a uu-vv path in GV(P)G-V(P). Suppose otherwise that either u=vCu=v_{C} or v=vCv=v_{C}, say u=vCu=v_{C}. In this case, we can construct a path between every vertex ww in CC and vv by concatenating PP^{\prime} and an arbitrary path in G[C]G[C] between ww and the end vertex of the subpath P{vC}P^{\prime}-\{v_{C}\} other than vv. Therefore, PP is non-separating in GG.

Now, we are ready to state the main result of this subsection.

Theorem 3.9.

Suppose that a minor-closed class 𝒞\mathcal{C} of graphs has the diameter-treewidth property. Then, Shortest Non-Separating Path on 𝒞\mathcal{C} is fixed-parameter tractable parameterized by kk.

Proof 3.10.

Let G𝒞G\in\mathcal{C}. We first compute B={vV(G):dist(s,v)k}B=\{v\in V(G):{\rm dist}(s,v)\leq k\}. This can be done in linear time. If tBt\notin B, then the instance (G,s,t,k)(G,s,t,k) is trivially infeasible. Suppose otherwise that tBt\in B. Let CC be a component in GBG-B. By definition, every non-separating ss-tt path PP of length at most kk does not contain any vertex of CC. Let GG^{\prime} be the graph obtained from GG by contracting all edges in E(GB)E(G-B). For each component CC in GBG-B, we denote by vCv_{C} the vertex of GG^{\prime} corresponding to CC (i.e., vCv_{C} is the vertex obtained by contracting all edges in CC). Then, we have diam(G)2k+2{\rm diam}(G^{\prime})\leq 2k+2 as diam(G[B])k{\rm diam}(G[B])\leq k and every vertex in V(G)BV(G^{\prime})\setminus B is adjacent to a vertex in BB. By Lemma 3.7, GG has a non-separating ss-tt path of length at most kk if and only if so does GG^{\prime}. Since 𝒞\mathcal{C} is minor-closed, we have G𝒞G^{\prime}\in\mathcal{C} and hence the treewidth of GG^{\prime} is upper bounded by f(2k+2)f(2k+2) for some function ff. By Theorem 3.5, we can check whether GG^{\prime} has a non-separating ss-tt path of length at most kk in O(g(k)|V(G)|)O(g(k)|V(G^{\prime})|) time for some function gg.

3.4 Chordal graphs with k=dist(s,t)k={\rm dist}(s,t)

In Section 3.1, we have seen that Shortest Non-Separating Path is NP-complete even on split graphs (and thus on more general chordal graphs as well). To overcome this intractability, we restrict ourselves to finding a non-separating ss-tt path of length dist(s,t){\rm dist}(s,t) on chordal graphs.

A graph GG is choral if it has no cycles of length at least 44 as an induced subgraph. In the following, fix a connected chordal graph GG.

Lemma 3.11.

Let SV(G)S\subseteq V(G) be a vertex set such that G[S]G[S] is connected. For u,vSu,v\in S, every induced uu-vv path PP in GG satisfies that V(P)N[S]V(P)\subseteq N[S].

Proof 3.12.

Suppose to the contrary that an induced uu-vv path PP contains a vertex xN[S]x\notin N[S]. Since PP starts and ends in SS, it contains a subpath Q=a,,x,,bQ={\langle a,\dots,x,\dots,b\rangle} such that a,bN(S)a,b\in N(S) and all other vertices in QQ belong to VN[S]V-N[S]. As a,bN(S)a,b\in N(S) and G[S]G[S] is connected, GG has an induced aa-bb path RR with all internal vertices belonging to SS. Since the internal vertices of QQ have no neighbors in SS, QRQ\cup R induces a cycle. Both aa-bb paths QQ and RR have length at least 22 as aa and bb are not adjacent, and thus the cycle G[QR]G[Q\cup R] has length at least 44. This contradicts that GG is chordal.

For u,vV(G)u,v\in V(G), a set of vertices SV(G){u,v}S\subseteq V(G)\setminus\{u,v\} is called a uu-vv separator of GG if there is no uu-vv path in GSG-S. An inclusion-wise minimal uu-vv separator of GG is called a minimal uu-vv separator. A minimal separator of GG is a minimal uu-vv separator for some u,vGu,v\in G. Dirac’s well-know characterization [8] of chordal graphs states that a graph is chordal if and only of every minimal separator induces a clique.

Lemma 3.13.

Let s,tV(G)s,t\in V(G) be such that {s,t}E(G)\{s,t\}\notin E(G). If vV(G){s,t}v\in V(G)\setminus\{s,t\} is an internal vertex of a shortest ss-tt path PP, then N[v]{s,t}N[v]\setminus\{s,t\} is an ss-tt separator of GG.

Proof 3.14.

Let d=dist(s,t)d={\rm dist}(s,t). For 0id0\leq i\leq d, let

Di={vV(G):dist(s,v)=idist(v,t)=di}\displaystyle D_{i}=\{v\in V(G):{\rm dist}(s,v)=i\land{\rm dist}(v,t)=d-i\}

and V(P)Di={ui}V(P)\cap D_{i}=\{u_{i}\}. Let jj (0<j<d0<j<d) be the index such that v=ujv=u_{j}.

Suppose to the contrary that there is an induced ss-tt path QQ such that V(Q)(N[uj]{s,t})=V(Q)\cap(N[u_{j}]\setminus\{s,t\})=\emptyset. By Lemma 3.11, V(Q)N[V(P)]=0idN[ui]V(Q)\subseteq N[V(P)]=\bigcup_{0\leq i\leq d}N[u_{i}] holds. Since QQ starts in N[u0]N[u_{0}] and ends in N[ud]N[u_{d}], there are indices ii and kk with 0i<j<kd0\leq i<j<k\leq d such that QQ consecutively visits a vertex viN[ui]v_{i}\in N[u_{i}] and then a vertex vkN[uk]v_{k}\in N[u_{k}] in this order. Since dist(ui,uk)=ki2{\rm dist}(u_{i},u_{k})=k-i\geq 2 and {vi,vk}E\{v_{i},v_{k}\}\in E, at least one of viuiv_{i}\neq u_{i} and vkukv_{k}\neq u_{k} holds. By symmetry, we assume that viuiv_{i}\neq u_{i}.

If vk=ukv_{k}=u_{k}, then viN(ui)N(uk)v_{i}\in N(u_{i})\cap N(u_{k}). In this case, we have i=j1i=j-1 and k=j+1k=j+1 since otherwise PP admits a shortcut using the subpath ui,vi,uk{\langle u_{i},v_{i},u_{k}\rangle}. This implies that dist(s,vi)dist(s,ui)+1=i+1=j{\rm dist}(s,v_{i})\leq{\rm dist}(s,u_{i})+1=i+1=j and dist(vi,t)1+dist(vk,t)=1+dist(uk,t)=1+(dk)=dj{\rm dist}(v_{i},t)\leq 1+{\rm dist}(v_{k},t)=1+{\rm dist}(u_{k},t)=1+(d-k)=d-j. Since dist(s,vi)+dist(vi,t)d{\rm dist}(s,v_{i})+{\rm dist}(v_{i},t)\geq d, we have dist(s,vi)=j{\rm dist}(s,v_{i})=j and dist(vi,t)=dj{\rm dist}(v_{i},t)=d-j. This implies that viDjN[uj]{s,t}v_{i}\in D_{j}\subseteq N[u_{j}]\setminus\{s,t\}, a contradiction

Next we consider the case vkukv_{k}\neq u_{k}. Recall that we also have viuiv_{i}\neq u_{i} as an assumption. In this case, we have ki3k-i\leq 3 as ui,vi,vk,uk{\langle u_{i},v_{i},v_{k},u_{k}\rangle} is not a shortcut for PP. Assume first that ki=3k-i=3. By symmetry, we may assume that i=j1i=j-1 and k=j+2k=j+2. Since dist(s,vi)dist(s,ui)+1=j{\rm dist}(s,v_{i})\leq{\rm dist}(s,u_{i})+1=j and dist(vi,t)2+dist(uk,t)2+(dk)=dj{\rm dist}(v_{i},t)\leq 2+{\rm dist}(u_{k},t)\leq 2+(d-k)=d-j, we have viDjN[uj]{s,t}v_{i}\in D_{j}\subseteq N[u_{j}]\setminus\{s,t\}, a contradiction. Next assume that ki=2k-i=2. That is, i=j1i=j-1 and k=i+1k=i+1. Since vi,vkN[uj]{s,t}v_{i},v_{k}\notin N[u_{j}]\setminus\{s,t\} and PP is shortest, the vertices viv_{i}, uiu_{i}, uju_{j}, uku_{k}, vkv_{k} are distinct and form a cycle of length 55. Observe that vi{s,t}v_{i}\notin\{s,t\} since otherwise vi=s,vk,uk{\langle v_{i}=s,v_{k},u_{k}\rangle} or ui,vi=t{\langle u_{i},v_{i}=t\rangle} is a shortcut. Similarly, vk{s,t}v_{k}\notin\{s,t\}. Hence, vi,vkN[uj]v_{i},v_{k}\notin N[u_{j}]. Therefore, the possible chords for the cycle vi,ui,uj,uk,vk{\langle v_{i},u_{i},u_{j},u_{k},v_{k}\rangle} are {ui,vk}\{u_{i},v_{k}\} and {uk,vi}\{u_{k},v_{i}\}. In any combination of them, the graph has an induced cycle of length at least 44.

Let dd and DiD_{i} be defined as in the proof of Lemma 3.13, and let D=0idDiD=\bigcup_{0\leq i\leq d}D_{i}. Note that each DiD_{i} is a clique: if i{0,d}i\in\{0,d\}, then it is a singleton; otherwise, it is a minimal ss-tt separator of the chordal graph G[D]G[D]. Observe that if |Di|=1|D_{i}|=1 for all 0id0\leq i\leq d, then GG contains a unique shortest ss-tt path, and thus the problem is trivial. Otherwise, we define \ell to be the minimum index such that |D|>1|D_{\ell}|>1 and rr to be the maximum index such that |Dr|>1|D_{r}|>1. Since |D0|=|Dd|=1|D_{0}|=|D_{d}|=1, we have 0<r<d0<\ell\leq r<d.

Our algorithm works as follows.

  1. 1.

    If GG contains a unique shortest ss-tt path PP, then test if PP is non-separating.

  2. 2.

    Otherwise, find a shortest ss-tt path PP satisfying the following conditions.

    1. (a)

      V(P)V(P) does not contain a minimal aa-bb separator for aDa\in D_{\ell} and bVDb\in V\setminus D.

    2. (b)

      V(P)V(P) does not contain a minimal aa-bb separator for aDa\in D_{\ell} and bDrb\in D_{r}.

Lemma 3.15.

The algorithm is correct.

Proof 3.16.

The first case is trivial. In the following, we prove the correctness of the second case.

First we show that the condition 2a is necessary. Let aDa\in D_{\ell} and bVDb\in V\setminus D. Since |D|>1|D_{\ell}|>1 and |V(P)D|=1|V(P)\cap D_{\ell}|=1, there is a vertex aDV(P)a^{\prime}\in D_{\ell}\setminus V(P), where aa^{\prime} may be aa itself. Since V(P)DV(P)\subseteq D, it holds that bV(P)b\notin V(P). Hence, aa^{\prime} and bb belong to the same connected component of GV(P)G-V(P). Since DD_{\ell} is a clique, aN[a]a\in N[a^{\prime}]. Thus, aa and bb belong to the same connected component of G(V(P){a,b})G-(V(P)\setminus\{a,b\}). Therefore, V(P)V(P) does not contain any aa-bb separator.

Next we show that the condition 2b is necessary. Let aDa\in D_{\ell} and bDrb\in D_{r}. As before, it suffices to show that aa and bb belong to the same connected component of G(V(P){a,b})G-(V(P)\setminus\{a,b\}). By the same reasoning in the previous case, there are vertices aDV(P)a^{\prime}\in D_{\ell}\setminus V(P) and bDrV(P)b^{\prime}\in D_{r}\setminus V(P) and they belong to the same connected component of GV(P)G-V(P). Now, since aN[a]a\in N[a^{\prime}] and bN[b]b\in N[b^{\prime}], aa and bb belong to the same connected component of G(V(P){a,b})G-(V(P)\setminus\{a,b\}).

Finally we show that the conditions 2a and 2b together form a sufficient condition for PP to be non-separating. Assume that a shortest ss-tt path PP satisfies the conditions 2a and 2b. Since |D|>1|D_{\ell}|>1 and |V(P)D|=1|V(P)\cap D_{\ell}|=1, there is a connected component CC of GV(P)G-V(P) that contains at least one vertex of DD_{\ell}. Now the condition 2a implies that VDV(C)V\setminus D\subseteq V(C) (recall that V(P)DV(P)\subseteq D), and the condition 2b implies that (DDr)V(P)V(C)(D_{\ell}\cup D_{r})\setminus V(P)\subseteq V(C) holds. To complete the proof, it suffices to show that DiV(P)V(C)D_{i}\setminus V(P)\subseteq V(C) for all ii. If i<i<\ell or i>ri>r, then DiV(P)=D_{i}\setminus V(P)=\emptyset. Let vDiV(P)v\in D_{i}\setminus V(P) for some ii with ir\ell\leq i\leq r. Observe that vv is an internal vertex of a shortest path from the unique vertex uD1u\in D_{\ell-1} to the unique vertex wDr+1w\in D_{r+1}. By Lemma 3.13, N[v]{u,w}N[v]\setminus\{u,w\} is a uu-ww separator. Since CC is connected and u,wu,w have neighbors in CC, G[V(C){u,w}]G[V(C)\cup\{u,w\}] contains a uu-ww path QQ. Since N[v]{u,w}N[v]\setminus\{u,w\} is a uu-ww separator, QQ contains a vertex qq such that

qV(Q)(N[v]{u,w})=(V(Q){u,w})N[v]V(C)N[v].q\in V(Q)\cap(N[v]\setminus\{u,w\})=(V(Q)\setminus\{u,w\})\cap N[v]\subseteq V(C)\cap N[v].

Therefore, vv has a neighbor (i.e., qq) in V(C)V(C), and thus vv itself belongs to CC.

Lemma 3.17.

The algorithm has a polynomial-time implementation.

Proof 3.18.

Since GG is chordal, each minimal separator of GG is a clique. Since PP is a shortest path, the size of a clique in G[V(P)]G[V(P)] is at most 22. Therefore, every minimal separator of GG contained in V(P)V(P) has size at most 22. Furthermore, every size-22 minimal separator {u,v}\{u,v\} is an edge of GG. This observation gives us the following implementation of the algorithm that clearly runs in polynomial time.

For i{1,2}i\in\{1,2\}, let i\mathcal{F}_{i} be the set of size-ii minimal aa-bb separators of GG such that aDa\in D_{\ell} and b(VD)Drb\in(V\setminus D)\cup D_{r}. It suffices to find a shortest ss-tt path PP such that no element of 12\mathcal{F}_{1}\cup\mathcal{F}_{2} is a subset of V(P)V(P). To forbid the elements of 1\mathcal{F}_{1}, we just remove the vertices that form the size-11 separators in 1\mathcal{F}_{1}. Similarly, to forbid the elements of 2\mathcal{F}_{2}, we remove the edges corresponding to the size-22 separators in 2\mathcal{F}_{2}. Now we find a shortest ss-tt path PP in the resultant graph. If PP has length d=distG(s,t)d={\rm dist}_{G}(s,t), then PP is a non-separating shortest ss-tt path in GG. Otherwise, GG does not have such a path.

Theorem 3.19.

There is a polynomial-time algorithm for Shortest Non-Separating Path on chordal graphs, provided that kk is equal to the shortest path distance between ss and tt.

4 Shortest Non-Disconnecting Path

The goal of this section is to establish the fixed-parameter tractability and a conditional lower bound on polynomial kernelizations for Shortest Non-Disconnecting Path.

4.1 Fixed-parameter tractability

Theorem 4.1.

Shortest Non-Disconnecting Path can be solved in time 2ωknO(1)2^{\omega k}n^{O(1)}, where ω\omega is the matrix multiplication exponent and nn is the number of vertices of the input graph GG.

To prove this theorem, we give a dynamic programming algorithm with the aid of representative families of cographic matroids. Let (G,s,t,k)(G,s,t,k) be an instance of Shortest Non-Disconnecting Path. For 0ik0\leq i\leq k and vV(G)v\in V(G), we define 𝖽𝗉(i,v){\sf dp}(i,v) as the family of all sets of edges FF satisfying the following two conditions: (1) FF is the set of edges of an ss-vv path of length ii and (2) GFG-F is connected. An edge set FF is legitimate if FF forms a path and GFG-F is connected. For a family of edge sets \mathcal{F} and an edge ee, we define e:={F{e}:F}\mathcal{F}\Join e:=\{F\cup\{e\}:F\in\mathcal{F}\} and 𝗅𝖾𝗀(){\sf leg}(\mathcal{F}) as the subfamily of \mathcal{F} consisting of all legitimate FF\in\mathcal{F}. The following simple recurrence correctly computes 𝖽𝗉(i,v){\sf dp}(i,v).

𝖽𝗉(i,v)=\displaystyle{\sf dp}(i,v)= {}\displaystyle\{\emptyset\} i=0i=0 and s=vs=v (1)
𝖽𝗉(i,v)=\displaystyle{\sf dp}(i,v)= \displaystyle\emptyset i=0i=0 and svs\neq v (2)
𝖽𝗉(i,v)=\displaystyle{\sf dp}(i,v)= 𝗅𝖾𝗀(uN(v)(𝖽𝗉(i1,u){u,v}))\displaystyle{\sf leg}\left(\displaystyle\bigcup_{u\in N(v)}({\sf dp}(i-1,u)\Join\{u,v\})\right) i>0i>0 . (3)

A straightforward induction proves that 𝖽𝗉(i,t){\sf dp}(i,t)\neq\emptyset if and only if GG has a non-disconnecting ss-tt path of length exactly ii and hence it suffices to check whether 𝖽𝗉(i,t){\sf dp}(i,t)\neq\emptyset for 0ik0\leq i\leq k. However, the running time to evaluate this recurrence is nO(k)n^{O(k)}. To reduce the running time of this algorithm, we apply Theorem 2.2 to each 𝖽𝗉(i,v){\sf dp}(i,v). Now, instead of (3), we define

𝖽𝗉(i,v)=𝗋𝖾𝗉ki(𝗅𝖾𝗀(uN(v)(𝖽𝗉(i1,u){u,v}))),\displaystyle{\sf dp}(i,v)={\sf rep}_{k-i}\left({\sf leg}\left(\displaystyle\bigcup_{u\in N(v)}({\sf dp}(i-1,u)\Join\{u,v\})\right)\right), (3’)

where 𝗋𝖾𝗉ki(){\sf rep}_{k-i}(\mathcal{F}) is a (ki)(k-i)-representative family of \mathcal{F} for the cographic matroid =(E(G),)\mathcal{M}=(E(G),\mathcal{I}) defined on GG. In the following, we abuse the notation of 𝖽𝗉{\sf dp} to denote the families of legitimate sets that are computed by the recurrence composed of (1), (2), and (3’).

Lemma 4.2.

The recurrence composed of (1), (2), and (3’) is correct, that is, GG has a non-disconnecting ss-tt path of length at most kk if and only if 0ik𝖽𝗉(i,t)\bigcup_{0\leq i\leq k}{\sf dp}(i,t)\neq\emptyset.

Proof 4.3.

It suffices to show that 𝖽𝗉(k,t){\sf dp}(k^{\prime},t)\neq\emptyset if GG has a non-disconnecting ss-tt path PP of length kkk^{\prime}\leq k. Let P=(v0=s,v1,,vk=t)P=(v_{0}=s,v_{1},\ldots,v_{k^{\prime}}=t) be a non-disconnecting path in GG. For 0ik0\leq i\leq k^{\prime}, we let Pi=(vi,vi+1,,vk)P_{i}=(v_{i},v_{i+1},\ldots,v_{k}). In the following, we prove, by induction on ii, a slightly stronger claim that there is a legitimate set F𝖽𝗉(i,vi)F\in{\sf dp}(i,v_{i}) such that FE(Pi)F\cup E(P_{i}) forms a non-disconnecting ss-tt path in GG for all 0ik0\leq i\leq k^{\prime}. As 𝖽𝗉(0,s)={}{\sf dp}(0,s)=\{\emptyset\} and P0=PP_{0}=P itself is a non-disconnecting path, we are done for i=0i=0. Suppose that i>0i>0. By the induction hypothesis, there is a legitimate F𝖽𝗉(i1,vi1)F\in{\sf dp}(i-1,v_{i-1}) such that FE(Pi1)F\cup E(P_{i-1}) forms a non-disconnecting ss-tt path in GG. Let =𝗅𝖾𝗀(uN(vi)(𝖽𝗉(i1){u,vi}))\mathcal{F}={\sf leg}(\bigcup_{u\in N(v_{i})}({\sf dp}(i-1)\Join\{u,v_{i}\})). Since FE(Pi1)F\cup E(P_{i-1}) is legitimate, F{{vi1,vi}}F\cup\{\{v_{i-1},v_{i}\}\} is also legitimate, implying that \mathcal{F} is nonempty. Let ^=𝗋𝖾𝗉ki()\widehat{\mathcal{F}}={\sf rep}_{k-i}(\mathcal{F}) be (ki)(k-i)-representative for \mathcal{F} and let Y={{vj,vj+1}:ij<k}Y=\{\{v_{j},v_{j+1}\}:i\leq j<k^{\prime}\}. As |Y|ki|Y|\leq k-i, XY=X\cap Y=\emptyset, and XYX\cup Y\in\mathcal{I}, ^\widehat{\mathcal{F}} contains an edge set X^\widehat{X} with X^Y\widehat{X}\cap Y and X^Y\widehat{X}\cup Y\in\mathcal{I}, implying that there is X^𝖽𝗉(i,vi)\widehat{X}\in{\sf dp}(i,v_{i}) such that X^E(Pi)\widehat{X}\cup E(P_{i}) forms a non-disconnecting ss-tt path in GG. Thus, the lemma follows.

Lemma 4.4.

The recurrence can be evaluated in time 2ωknO(1)5.18knO(1)2^{\omega k}n^{O(1)}\subset 5.18^{k}n^{O(1)}, where ω<2.373\omega<2.373 is the exponent of the matrix multiplication.

Proof 4.5.

By Theorem 2.2, 𝖽𝗉(i,v){\sf dp}(i,v) contains at most 2kkn2^{k}kn sets for 0ik0\leq i\leq k and vV(G)v\in V(G) and can be computed in time 2ωknO(1)2^{\omega k}n^{O(1)} by dynamic programming.

Thus, Theorem 4.1 follows.

4.2 Kernel lower bound

It is well known that a parameterized problem is fixed-parameter tractable if and only if it admits a kernelization (see [6], for example). By Theorem 4.1, Shortest Non-Disconnecting Path admits a kernelization. A natural step next to this is to explore the existence of polynomial kernelizations for Shortest Non-Disconnecting Path. However, the following theorem conditionally rules out the possibility of polynomial kernelization. To prove this, we first show the following lemma.

Lemma 4.6.

Let HH be a connected graph. Suppose that HH has a cut vertex vv. Let CC be a component in H{v}H-\{v\} and let FE(H[C{v}])F\subseteq E(H[C\cup\{v\}]). Then, HFH-F is connected if and only if H[C{v}]FH[C\cup\{v\}]-F is connected.

Proof 4.7.

If HFH-F is connected, then all the vertices in C{v}C\cup\{v\} are reachable from vv in HFH-F without passing through any vertex in V(H)({C}{v})V(H)\setminus(\{C\}\cup\{v\}). Thus, such vertices are reachable from vv in H[C{v}]FH[C\cup\{v\}]-F. Conversely, suppose H[C{v}]FH[C\cup\{v\}]-F is connected. Then, every vertex in CC is reachable from vv in HFH-F. Moreover, as FF does not contain any edge outside H[C{v}]H[C\cup\{v\}], every other vertex is reachable from vv in HFH-F as well.

Theorem 4.8.

Unless coNP \subseteq NP//poly, Shortest Non-Disconnecting Path does not admit a polynomial kernelization (with respect to parameter kk).

Proof 4.9.

We give an OR-composition for Shortest Non-Disconnecting Path. Let (G1,s1,t1,k),,(Gp,sp,tp,k)(G_{1},s_{1},t_{1},k),\ldots,(G_{p},s_{p},t_{p},k) be pp instances of Shortest Non-Disconnecting Path. We assume that for 1ip1\leq i\leq p, tit_{i} is not a cut vertex in GiG_{i}. To justify this assumption, suppose that tit_{i} is a cut vertex in GiG_{i}. Let CC be the component in Gi{ti}G_{i}-\{t_{i}\} that contains sis_{i}. By Lemma 4.6, for any sis_{i}-tit_{i} path, it is non-disconnecting in GiG_{i} if and only if so is in Gi[C{ti}]G_{i}[C\cup\{t_{i}\}]. Thus, by replacing GiG_{i} with Gi[C]G_{i}[C], we can assume that tit_{i} is not a cut vertex in GiG_{i}.

From the disjoint union of G1,,GpG_{1},\ldots,G_{p}, we construct a single instance (G,s,t,k)(G,s,t,k^{\prime}) as follows. We first add a vertex ss and an edge between ss and sis_{i} for each 1ip1\leq i\leq p. Then, we identify all tit_{i}’s into a single vertex tt. See Figure 2 for an illustration.

Refer to caption
Figure 2: An illustration of the graph GG obtained from q=4q=4 instances.

In the following, we may not distinguish tt from tit_{i}. Now, we claim that (G,s,t,k+1)(G,s,t,k+1) is a yes-instance if and only if (Gi,si,ti,k)(G_{i},s_{i},t_{i},k) is a yes-instance for some ii.

Consider an arbitrary ss-tt path in GG. Observe that all edges in the path except for that incident to ss are contained in a single subgraph GiG_{i} for some 1ip1\leq i\leq p as {s,t}\{s,t\} separates V(Gi){ti}V(G_{i})\setminus\{t_{i}\} from V(Gj){tj}V(G_{j})\setminus\{t_{j}\} for jij\neq i. Moreover, the path PP forms P=(s,si,v1,,vq,t)P=(s,s_{i},v_{1},\ldots,v_{q},t), meaning that the subpath P=(s1,v1,,vq,ti)P^{\prime}=(s_{1},v_{1},\ldots,v_{q},t_{i}) is an sis_{i}-tit_{i} path in GiG_{i}. This conversion is reversible: for any sis_{i}-tit_{i} path PP^{\prime} in GiG_{i}, the path obtained from PP^{\prime} by attaching ss adjacent to sis_{i} is an ss-tt path in GG. Thus, it suffices to show that for FE(Gi)F\subseteq E(G_{i}), F{{s,si}}F\cup\{\{s,s_{i}\}\} is a cut of GG if and only if FF is a cut of GiG_{i}. Since tt is a cut vertex in G{{s,si}}G-\{\{s,s_{i}\}\}, by Lemma 4.6, the claim holds.

References

  • [1] Stefan Arnborg, Jens Lagergren, and Detlef Seese. Easy problems for tree-decomposable graphs. J. Algorithms, 12(2):308–340, 1991. doi:10.1016/0196-6774(91)90006-K.
  • [2] Hans L. Bodlaender. A linear-time algorithm for finding tree-decompositions of small treewidth. SIAM J. Comput., 25(6):1305–1317, 1996. doi:10.1137/S0097539793251219.
  • [3] Hans L. Bodlaender, Rodney G. Downey, Michael R. Fellows, and Danny Hermelin. On problems without polynomial kernels. J. Comput. Syst. Sci., 75(8):423–434, 2009. doi:10.1016/j.jcss.2009.04.001.
  • [4] Guantao Chen, Ronald J. Gould, and Xingxing Yu. Graph connectivity after path removal. Comb., 23(2):185–203, 2003. doi:10.1007/s003-0018-z.
  • [5] Bruno Courcelle. The Monadic Second-Order Logic of Graphs. I. Recognizable Sets of Finite Graphs. Inf. Comput., 85(1):12–75, 1990.
  • [6] Marek Cygan, Fedor V. Fomin, Lukasz Kowalik, Daniel Lokshtanov, Daniel Marx, Marcin Pilipczuk, Michal Pilipczuk, and Saket Saurabh. Parameterized Algorithms. Springer Publishing Company, Incorporated, 1st edition, 2015.
  • [7] Erik D. Demaine and Mohammad Taghi Hajiaghayi. Diameter and treewidth in minor-closed graph families, revisited. Algorithmica, 40(3):211–215, 2004. doi:10.1007/s00453-004-1106-1.
  • [8] G. A. Dirac. On rigid circuit graphs. Abhandlungen aus dem Mathematischen Seminar der Universität Hamburg, 25:71–76, 1961.
  • [9] David Eppstein. Diameter and treewidth in minor-closed graph families. Algorithmica, 27(3):275–291, 2000. doi:10.1007/s004530010020.
  • [10] Michael R. Fellows, Danny Hermelin, Frances A. Rosamond, and Stéphane Vialette. On the parameterized complexity of multiple-interval graph problems. Theor. Comput. Sci., 410(1):53–61, 2009. doi:10.1016/j.tcs.2008.09.065.
  • [11] Fedor V. Fomin, Daniel Lokshtanov, Fahad Panolan, and Saket Saurabh. Efficient computation of representative families with applications in parameterized and exact algorithms. J. ACM, 63(4):29:1–29:60, 2016.
  • [12] Martin Grohe and Stephan Kreutzer. Methods for algorithmic meta theorems. In AMS-ASL Joint Special Session, volume 558 of Contemporary Mathematics, pages 181–206. American Mathematical Society, 2009.
  • [13] Martin Grohe, Stephan Kreutzer, and Sebastian Siebertz. Deciding first-order properties of nowhere dense graphs. J. ACM, 64(3):17:1–17:32, 2017. doi:10.1145/3051095.
  • [14] Alon Itai, Christos H. Papadimitriou, and Jayme Luiz Szwarcfiter. Hamilton paths in grid graphs. SIAM J. Comput., 11(4):676–686, 1982. doi:10.1137/0211056.
  • [15] Ken-ichi Kawarabayashi, Orlando Lee, Bruce A. Reed, and Paul Wollan. A weaker version of Lovász’ path removal conjecture. J. Comb. Theory, Ser. B, 98(5):972–979, 2008. doi:10.1016/j.jctb.2007.11.003.
  • [16] Matthias Kriesell. Induced paths in 5-connected graphs. J. Graph Theory, 36(1):52–58, 2001. doi:10.1002/1097-0118(200101)36:1<52::AID-JGT5>3.0.CO;2-N.
  • [17] Daniel Lokshtanov, Pranabendu Misra, Fahad Panolan, and Saket Saurabh. Deterministic truncation of linear matroids. ACM Trans. Algorithms, 14(2):14:1–14:20, 2018.
  • [18] Xiao Mao. Shortest non-separating st-path on chordal graphs. CoRR, abs/2101.03519, 2021. arXiv:2101.03519.
  • [19] Haiko Müller. Hamiltonian circuits in chordal bipartite graphs. Discret. Math., 156(1-3):291–298, 1996.
  • [20] James G. Oxley. Matroid Theory (Oxford Graduate Texts in Mathematics). Oxford University Press, Inc., USA, 2006.
  • [21] Neil Robertson and Paul D. Seymour. Graph minors. III. planar tree-width. J. Comb. Theory, Ser. B, 36(1):49–64, 1984. doi:10.1016/0095-8956(84)90013-3.
  • [22] William T. Tutte. How to draw a graph. Proceedings of the London Mathematical Society, s3-13(1):743–767, 1963. doi:https://doi.org/10.1112/plms/s3-13.1.743.
  • [23] Bang Ye Wu and Hung-Chou Chen. The approximability of the minimum border problem. In The 26th Workshop on Combinatorial Mathematics and Computation Theory, 2009.