Cycles
2
Akanksha Agrawal
3
Institute of Computer Science and Control, Hungarian Academy of Sciences (MTA SZTAKI),
4
Budapest, Hungary
5
akanksha@sztaki.mta.hu
6
Pallavi Jain
7
Institute of Mathematical Sciences, HBNI, Chennai, India
8
pallavij@imsc.res.in
9
Lawqueen Kanesh
10
Institute of Mathematical Sciences, HBNI, Chennai, India
11
lawqueen@imsc.res.in
12
Pranabendu Misra
13
University of Bergen, Bergen, Norway
14
Pranabendu.Misra@uib.no
15
Saket Saurabh
16
Institute of Mathematical Sciences, HBNI, Chennai, India
17
saket@imsc.res.in
18
Abstract
19
A generalization of classical cycle hitting problems, called conflict version of the problem, is
20
defined as follows. An input is undirected graphsGandHon the same vertex set, and a positive
21
integer k, and the objective is to decide whether there exists a vertex subsetX ⊆ V(G) such
22
that it intersects all desired “cycles” (all cycles or all odd cycles or all even cycles) andX is an
23
independent set inH. In this paper we study the conflict version of classicalFeedback Vertex
24
Set, andOdd Cycle Transversalproblems, from the view point of kernelization complexity.
25
In particular, we obtain the following results, when the conflict graphH belongs to the family
26
ofd-degenerate graphs.
27
1. CF-FVSadmits aO(kO(d)) kernel.
28
2. CF-OCTdoes not admit polynomial kernel (even whenHis 1-degenerate), unlessNP⊆ coNPpoly.
29
For our kernelization algorithm we exploit ideas developed for designing polynomial kernels for
30
the classicalFeedback Vertex Setproblem, as well as, devise new reduction rules that exploit
31
degeneracy crucially. Our main conceptual contribution here is the notion of “k-independence
32
preserver”. Informally, it is a set of “important” vertices for a given subset X ⊆ V(H), that
33
is enough to capture the independent set property in H. We show that ford-degenerate graph
34
independence preserver of sizekO(d) exists, and can be used in designing polynomial kernel.
35
2012 ACM Subject Classification Theory of computation→Design and analysis of algorithms
36
→Parameterized complexity and exact algorithms
37
Keywords and phrases Parameterized Complexity, Kernelization, Conflict-free problems, Feed-
38
back Vertex Set, Even Cycle Transversal, Odd Cycle Transversal
39
Digital Object Identifier 10.4230/LIPIcs.IPEC.2018.14
40
Funding This research has received funding from the European Research Council under ERC
41
grant no. 306992 PARAPPROX, ERC grant no. 715744 PaPaALG and ERC grant no. 725978
42
SYSTEMATIC-GRAPH, and DST, India for SERB-NPDF fellowship [PDF/2016/003508].
43
© A. Agrawal and P. Jain and L. Kanesh and P. Misra and S. Saurabh;
1 Introduction
44
Reducing the input data, in polynomial time, without altering the answer is one of the
45
popular ways in dealing with intractable problems in practice. While such polynomial
46
time heuristics can not solve NP-hard problems exactly, they work well on input instances
47
arising in real-life. It is a challenging task to assess the effectiveness of such heuristics
48
theoretically. Parameterized complexity, via kernelization, provides a natural way to quantify
49
the performance of such algorithms. In parameterized complexity each problem instance
50
comes with a parameterk and the parameterized problem is said to admit a polynomial
51
kernelif there is a polynomial time algorithm, called a kernelizationalgorithm, that reduces
52
the input instance down to an instance with size bounded by a polynomialp(k) ink, while
53
preserving the answer. The reduced instance is called ap(k) kernel for the problem.
54
The quest for designing polynomial kernels for “hitting cycles” in undirected graphs has
55
played significant role in advancing the field of polynomial time pre-processing – kernelization.
56
Hitting all cycles, odd cycles and even cycles correspond to well studied problems ofFeedback
57
Vertex Set(FVS),Odd Cycle Transversal(OCT) and Even Cycle Transversal
58
(ECT), respectively. Alternatively,FVS,OCTandECTcorrespond to deleting vertices such
59
that the resulting graph is a forest, a bipartite graph and an odd cactus graph, respectively.
60
All these problems,FVS,OCT, andECT, have been extensively studied in parameterized
61
algorithms and kernelization. The earliest knownFPTalgorithms forFVS go back to the
62
late 80’s and the early 90’s [4, 11] and used the seminal Graph Minor Theory of Robertson
63
and Seymour. On the other hand the parameterized complexity ofOCTwas open for long
64
time. Only, in 2003, Reed et al. [24] gave a 3knO(1) time algorithm forOCT. This is also
65
the paper which introduced themethod of iterative compressionto the field of parameterized
66
complexity. However, the existence of polynomial kernel, for FVSandOCT were open
67
questions for long time. ForFVS, Burrage et al. [7] resolved the question in the affirmative
68
by designing a kernel of sizeO(k11). Later, Bodlaender [5] reduced the kernel size toO(k3),
69
and finally Thomassé [25] designed a kernel of sizeO(k2). The kernel of Thomassé [25] is
70
best possible under a well known complexity theory hypothesis. It is important to emphasize
71
that [25] popularized the method ofexpansion lemma, one of the most prominent approach
72
in designing polynomial kernels. While, the kernelization complexity ofFVSwas settled
73
in 2006, it took another 6 years and a completely new methodology to design polynomial
74
kernel forOCT. Kratsch and Wahlström [16] resolved the question of existence of polynomial
75
kernel forOCTby designing a randomized kernel of sizeO(k4.5) using matroid theory.1 As
76
a counterpart toOCT, Misra et al. [20] studiedECTand designed anO(k3) kernel.
77
Fruitful and productive research on FVS andOCT have led to the study of several
78
variants and generalizations of FVS andOCT. Some of these admit polynomial kernels
79
and for some one can show that none can exist, unless some unlikely collapse happens in
80
complexity theory. In this paper we study the following generalization ofFVS, andOCT,
81
from the view-point of kernelization complexity.
82
Conflict Free Feedback Vertex Set(CF-FVS) Parameter: k Input:An undirected graphG, a conflict graphHon vertex setV(G) and a non-negative integerk.
Question: Does there existS⊆V(G), such that|S| ≤k,G−S is a forest andH[S] is edgeless?
83
1 This foundational paper has been awarded the Nerode Prize for 2018.
One can similarly define Conflict Free Odd Cycle Transversal(CF-OCT).
84
Motivation. On the outset, a natural thought is “why does one care” about such an
85
esoteric (or obscure) problem. We thoughtexactly the same in the beginning, till we realized
86
the modeling power the problem provides and the rich set of questions one can ask. In the
87
course of this paragraph we will try to explain this. First observe that, if one wants to model
88
“independent” version of these problems (where the solution is suppose to be an independent
89
set), then one takes conflict graph to be same as the input graph. An astute reader will figure
90
out that the problem as stated above is W[1]-hard – a simple reduction fromMulticolor
91
Independent Set with each color class being modeled as cycle and the conflict graph
92
being the input graph. Thus, a natural question is: when does the problem become FPT? To
93
state the question formally, letF andG be two families of graphs. Then, (G,F)-CF-FVSis
94
same problem asCF-FVS, but the input graphGand the conflict graphH are restricted
95
to belong toG and H, respectively. It immediately brings several questions: (a) for which
96
pairs of families the problem isFPT; (b) can we obtain some kind of dichotomy results; and
97
(c) what could we say about the kernelization complexity of the problem. We believe that
98
answering these questions for basic problems such as FVS, OCT, andDominating Set
99
will extend both the tractability as well as intractability tools in parameterized complexity
100
and led to some fruitful and rewarding research. It is worth to note that initially we were
101
inspired to define these problems by similar problems in computational geometry. See related
102
results for more on this.
103
Our Results and Methods. A graphGis calledd-degenerateif every subgraph of G
104
has a vertex of degree at most d. For a fixed positive integerd, letDd denote the set of
105
graphs ofdegeneracyat most d. In this paper we study the (?,Dd)-CF-FVS (Dd-CF-FVS)
106
problem. The symbol ? denotes that the input graph G is arbitrary. One can similarly
107
define Dd-CF-OCT. In fact, we study, CF-OCT for a very restricted family of conflict
108
graphs, a family of disjoint union of paths of length at most three and at most two star
109
graphs. We denote this family as P≤3?? and this variant of CF-OCT as P≤3??-CF-OCT.
110
Starting point of our research is the recent study of Jain et al. [14], who studied conflict-free
111
graph modification problems in the realm of parameterized complexity. As a part of their
112
study they gaveFPTalgorithms forDd-CF-FVS,Dd-CF-OCTandDd-CF-ECTusing the
113
independence covering families [17]. Their results also imply similarFPTalgorithm when the
114
conflict graph belongs to nowhere dense graphs. In this paper we focus on the kernelization
115
complexity ofDd-CF-FVS, andP≤3??-CF-OCTobtain the following results.
116
1. Dd-CF-FVSadmits aO(kO(d)) kernel.
117
2. P≤3??-CF-OCTdoes not admit polynomial kernel, unlessNP⊆ coNPpoly.
118
Note that D0 denotes edgeless graphs and henceD0-CF-FVS, andD0-CF-OCT are
119
essentially FVS, andOCT, respectively. Thus, any polynomial kernel forDd-CF-FVS, and
120
P≤3??-CF-OCT, must generalize the known kernels for these problems. We remark that the
121
above result imply thatCF-FVSadmits polynomial kernels, when the conflict graph belong
122
to several well studied graph families, such as planar graphs, graphs of bounded degree, graphs
123
of bounded treewidth, graphs excluding some fixed graph as a minor, a topological minor
124
and graphs of bounded expansion etc. (all these graphs classes have bounded degeneracy).
125
Strategy for CF-FVS.Our kernelization algorithm forCF-FVSconsists of the following
126
two steps. The first step of our kernelization algorithm is a structural decomposition of the
127
input graphG. This does not depend on the conflict graphH. In this phase of the algorithm,
128
given an instance (G, H, k) of CF-FVS we obtain an equivalent instance (G0, H0, k0) of
129
CF-FVSsuch that:
130
The minimum degree ofG0 is at least 2.
131
The number of vertices of degree at least 3 inG0 is upper bounded byO(k3). LetV≥3
132
denote the set of vertices of degree at least 3 inG0.
133
The number of maximal degree 2 paths in G0 is upper bounded by O(k3). That is,
134
G0−V≥3 consists ofO(k3) connected components where each component is a path.
135
We obtain this structural decomposition using reduction rules inspired by the quadratic
136
kernel forFVS[25]. As stated earlier, this step can be performed for any graphH. Thus the
137
problem reduces to designing reduction rules that bound the number of vertices of degree 2
138
in the reduced graph. Note that we can not do this for any arbitrary graphH as the problem
139
is W[1]-hard. Once the decomposition is obtained we can not use the knownreduction rules
140
forFVS. This is for a simple reason that inG0 the only vertices that are not bounded have
141
degree exactly 2 inG0. On the other hand forFVSwe can do simple “short-circuit” of degree
142
2 vertices (remove the vertex and add an edge between its two neighbors) and assume that
143
there is no vertices of degree two in the graph. So our actual contributions start here.
144
The second step of our kernelization algorithm bounds the degree two vertices in the
145
graph G0. Here we must use the properties of the graphH. We propose new reduction
146
rules for bounding degree two vertices, whenH belongs to the family ofd-degenerate graphs.
147
Towards this we use the notion ofd-degeneracy sequence, which is an ordering of the vertices
148
inH such that any vertex can have at mostdforward neighbors. This is used in designing a
149
marking scheme for the degree two vertices. Broadly speaking our marking scheme associates
150
a set with every vertexv. Here, set consists of “ paths and cycles ofG0 on which the forward
151
neighbors of v are”. Two vertices are called similar if their associated sets are same. We
152
show that if some vertex is not marked then we can safely contract this vertex to one of its
153
neighbors. We then upper bound the degree two vertices byO(kO(d)dO(d)), and thus obtain
154
a kernel of this size forDd-CF-FVS.
155
At the heart of our kernelization algorithm is a combinatorial tool of “k-independence
156
preserver”. Informally, it is a set of “important” vertices for a given subsetX ⊆V(H), that
157
is enough to capture the independent set property inH. We show that for d-degenerate
158
graph independence preserver of sizekO(d)exists, and can be used in designing polynomial
159
kernel. This is our main conceptual contribution.
160
Strategy for CF-OCT.The kernelization lower bound is obtained by the method of
161
cross-composition [6]. We first define a conflict version of thes-t-Cutproblem, whereH
162
belongs toP≤3??. Then, we show that the problem is NP-hard and cross composes to itself.
163
Finally, we give a parameter preserving reduction from the problem to P≤3??-CF-OCT, and
164
obtain the desired kernel lower bound.
165
Related Work. In the past, the conflict free versions of some classical problems have
166
been studied, e.g. forShortest Path[15], Maximum Flow[21, 22],Knapsack [23],Bin
167
Packing[12],Scheduling[13],Maximum MatchingandMinimum Weight Spanning
168
Tree[10, 9]. It is interesting to note that some of these problems areNP-hard even when
169
their non-conflicting version is polynomial time solvable. The study of conflict free problems
170
has also been recently initiated in computational geometry motivated by various applications
171
(see [1, 2, 3]).
172
2 Preliminaries
173
Throughout the paper, we follow the following notions. LetGbe a graph,V(G) andE(G)
174
denote the vertex set and the edge set of graphG, respectively. Letn andmdenote the
175
number of vertices and the number of edges of G, respectively. Let G be a graph and
176
X⊆V(G), thenG[X] is the graph induced onX andG−X is graphGinduced onV(G)\X.
177
Let ∆ denotes the maximum degree of graph G. We useNG(v) to denote the neighborhood
178
ofv inGandNG[v] to denoteNG(v)∪ {v}. LetE0 be subset of edges of graphG, byG[E0]
179
we mean the graph with the vertex setV(G) and the edge set E0. Let X ⊆E(G), then
180
G−X is a graph with the vertex setV(G) and the edge set E(G)\X. LetY be a set of
181
edges on vertex setV(G), thenG∪Y is graph with the vertex setV(G) and the edge set
182
E(G)∪Y. Degree of a vertexv in graphGis denoted by degG(v). For an integer `, we
183
denote the set{1,2, . . . , `}by [`]. ApathP ={v1, . . . , vn}is an ordered collection of vertices
184
such that there is an edge between every consecutive vertices inP andv1, vn areendpoints of
185
P. For a pathP byV(P) we denote set of vertices inP and byE(P) we denote set of edges
186
inP. AcycleC={v1, . . . , vn}is a path with an edgev1vn. We define amaximal degree two
187
induced path inGas an induced path of maximal length such that all vertices in path are of
188
degree exactly two inG. Anisolated cyclein graphGis defined as an induced cycle whose
189
all the vertices are of degree exactly two inG. LetG0 andGbe graphs,V(G0)⊆V(G) and
190
E(G0)⊆E(G), then we say thatG0 is asubgraph of G. The subscript in the notations will
191
be omitted if it is clear from the context.
192
A graphGhasdegeneracydif every subgraph ofGhas a vertex of degree at mostd. An
193
ordering of verticesσ:V(G)→ {1,· · ·, n}is is called ad-degeneracy sequenceof graphG, if
194
every vertexv has at mostdneighborsuwithσ(u)> σ(v). A graphGisd-degenerate if and
195
only if it has ad-degeneracy sequence. For a vertexvind-degenerate graphG, the neighbors
196
of v which comesafter (before)v ind-degeneracy sequence are calledforward (backward)
197
neighborsofvin the graphG. Given ad-degenerate graph, we can findd-degeneracy sequence
198
in linear time [18].
199
3 A Tool for Our Kernelization Algorithm
200
In this section, we give a tool, which we believe might be useful in obtaining kernelization
201
algorithm for “conflict free” versions of various parameterized problems (admitting kernels),
202
when the conflict graph belongs to the family ofd-degenerate graphs. We particularly use
203
this tool to obtain kernel for Dd-CF-FVS (Section 4). For a parameterized problem Π,
204
consider an instance (G, H, k) of its conflict free variant,Conflict FreeΠ. Then in the
205
kernelization step where we want to bound the number of vertices, it is seemingly useful to
206
be able to obtain a set of “important” vertices for a given subsetX ⊆V(H) that will be
207
enough to capture the independent set property inH. The above intuition becomes clear
208
when we describe the kernelization algorithm forDd-CF-FVS.
209
To formalize the notion of “important” set of vertices, we give the following definition.
210
IDefinition 1. For ad-degenerate graphH and a setX ⊆V(H), ak-independence preserver
211
for (H, X) is a setX0⊆X, such that for any independent setS inH of size at mostk, if
212
there isv∈(S∩X)\X0, then there isv0∈X0\S, such that (S\ {v})∪ {v0}is an independent
213
set inH.
214
Throughout this section, we work with a (fixed)d, which is the degeneracy of the input
215
graph. The goal of this section will be to obtain an algorithm for computing ak-independence
216
preserver for (H, X) of “small” size. To quantify the “small” size, we need the following
217
definition.
218
IDefinition 2. For eachq∈[d], we define an integernq as follows.
219
1. Ifq= 1, thennq =kd+k+ 1, and
220
2. nq =knq−1+kd+k+ 1, otherwise.
221
Next, we formally define the problem for which we want to design a polynomial time
222
algorithm. We call this problemd-Bounded Independence Preserver(d-BIP, for short).
223
d-Bounded Independence Preserver(d-BIP)
Input: Ad-degenerate graphH, a setX⊆V(H), and an integerk.
Output: A setX0⊆X of size at mostnd+1, such thatX0 is akindependence preserver for (H, X).
224
In the following, let (H, X, k) be an instance ofd-BIP. We work with a (fixed)d-degeneracy
225
sequence,σofH. We recall that such a sequence can be computed in polynomial time [18].
226
Forward and backward neighbors of a vertexv are also defined with respect to the ordering
227
σ. Ifσ(u)< σ(v), thenuis a backward neighbor ofv andv is a forward neighbor ofu. By
228
NHf(v) (NHb(v)) we denote the set of forward (backward) neighbors of the vertexvin H.
229
To design our polynomial time algorithm for d-BIP, we need the notion ofq-reducible
230
sets, which is formally defined below.
231
IDefinition 3. A setY ⊆V(H) is q-reducible, if for every setU ⊆Y, for which there is a
232
setZ⊆V(H), such that: (i)Z is of size exactlyd−q+ 1 and (ii) for eachu∈U, we have
233
Z⊆NHf(u), it holds that |U| ≤nq.
234
Now, we give our polynomial time algorithm ford-BIPin Algorithm 1.
235
Algorithm 1Algo1(H, X)
Require: d-degenerate graphH, X⊆V(H), and an integerk.
Ensure: X0 ⊆X of size at mostnd+1, which is a k-independence preserver of (H, X).
1: Forq∈[d], setnq =kd+ 1, whenq= 1, andnq =knq−1+kd+k+ 1, otherwise.
2: q= 1.
3: whileq≤ddo
4: while X is notq-reducibledo
5: FindU ⊆X of sizenq+ 1, for which there isZ ⊆V(H) of size exactlyd−q+ 1, such that for eachu∈U, we haveZ⊆NHf(u).
6: Letvbe an arbitrary vertex in U. 7: X =X\ {v}.
8: end while 9: q=q+ 1.
10: end while
11: while|X|> nd+1 do
12: Letv be an arbitrary vertex inX. 13: X =X\ {v}.
14: end while 15: SetX0 =X. 16: returnX0
To prove the correctness of our algorithm, we state an observation, the proof of which
236
follows from the fact that any vertex can have at mostdforward neighbors inH.
237
IObservation 1. LetH be ad-degenerate graph and S be an independent set ofH of size
238
at mostk. Then, for any setU ⊆V(H), such that for each vertexu∈U,NHb(u)∩S6=∅,
239
we have that|U| ≤kd.
240
Now we are ready to prove the correctness of our algorithm (Algorithm 1) ford-BIP.
241
ILemma 2. Algorithm 1 is correct.
242
Proof. Let (H, X, k) be an instance ofd-BIP, andX0 be the output returned by Algorithm 1
243
with it as the input. Clearly,X0⊆X as we do not add any new vertex to obtain the setX0,
244
and size of X0 is bounded bynd+1, since at Step 10-13 of the algorithm we reduce its size
245
to (at most)nd+1. Therefore, it remains to show thatX0 is ak-independence preserver of
246
(H, X). To this end, we consider the following cases.
247
Case 1: X isq-reducible, for eachq∈[d]. In this case, the algorithm arbitrarily deletes
248
vertices (if required) from X to obtain X0. If X = X0, then the claim trivially holds.
249
Therefore, we assume thatX0 is a strict subset ofX. To show thatX0 is a k-independence
250
preserver for (H, X), consider an independent setS inH of size at mostk. Furthermore,
251
consider a vertexv∈(S∩X)\X0 (again, if such a vertex does not exists, the claim follows).
252
To prove the desired result, we want to find a replacement vertex forvinX0 which can be
253
added toS (after removing v) to obtain an independent set inH. To this end, we mark
254
some vertices in X0. Firstly, mark all the forward neighbors of each s∈S in the setX0.
255
That is, we letXM0 to be the set (∪s∈SNHf(s))∩X0. Also, we add all vertices inS∩X0
256
to the set XM0 . By the property of d-degeneracy sequence, we have that |XM0 | ≤ kd+k
257
(see Observation 1). Next, we will mark some more vertices inXM0 with the hope to find
258
a replacement vertex forv inX0\XM0 to add toS. Recall that by our assumption X is
259
q-reducible, for each q∈[d], and in particular, it isd-reducible. Thus, for eachs∈S, the
260
setXs={x∈X |s∈NHf(x)} ⊆X has size at mostnd. Based on the above observation,
261
we describe our second level of marking of vertices inX0. For each s ∈ S, we add each
262
vertex inXstoXM0 . From the discussions above, we have that|XM0 | ≤kd+k+knd. Since
263
|X0|=nd+1, and by definition,nd+1=knd+kd+k+ 1, we haveX0\XM0 6=∅. Moreover,
264
no vertex in X0 has a neighbor in S\ {v}. Therefore, for v0 ∈ X0\XM0 , we have that
265
S0= (S\ {v})∪ {v0}is an independent set inH.
266
Case 2: X is notq-reducible, for someq∈[d]. Letq0 be the smallest integer for which
267
X is notq0-reducible. SinceX is notq0-reducible, there is a setU ⊆X of size at leastnq+ 1,
268
for which there is a setZ ⊆V(H) of size exactlyd−q+ 1, such that for each u∈U, we
269
haveZ⊆NHf(u). Consider (first) such pair of setsU, Z considered by the algorithm in Step
270
4. Furthermore, letv∈U be the vertex deleted by the algorithm in Step 6. Let ˆU =U\ {v}.
271
To prove the claim, it is enough to show that for an independent setS of size at most k
272
containing v inH, there isv0 ∈Uˆ such that (S\ {v})∪ {v0} is an independent set in H.
273
Here, we will use the fact that deleting a vertex from a set does not change a set from being
274
˜
q-reducible to a set which is not ˜q-reducible, where ˜q ∈[d]. In the following, consider an
275
independent set S of size at mostk containing v inH. We construct a marked set ˆUM,
276
of vertices in ˆU. Firstly, we add all the vertices in (∪s∈S\{v}NHf(s))∩Uˆ to ˆUM. Also, we
277
add all vertices inS∩Uˆ to ˆUM. Notice that at the end of above marking scheme, we have
278
|XˆM| ≤kd+k. We will mark some more vertices in ˆU. Before stating the second level of
279
marking, we remark thatS∩Z=∅. For eachs∈S\ {v}, letZs=Z∪ {s}. SinceS∩Z=∅,
280
we have that|Zs|=d−(q−1) + 1. Fors∈S\ {v}, let ˆUs={u∈Uˆ |Zs⊆NHf(u)}. SinceX
281
isq∗-reducible for eachq∗< q0, we have|Uˆs| ≤nq−1, for eachs∈S\ {v}. Now we are ready
282
to describe our second level of marking. For eachs∈S\ {v}, add all vertices inUsto the set
283
UˆM. Notice that|UˆM| ≤kd+k+knq−1. Moreover,|Uˆ| ≥nq andnq=knq−1+kd+k+ 1.
284
Thus, there is a vertexv0 ∈Uˆ\UˆM, such that (S\ {v})∪ {v0}is an independent set inH. J
285
ILemma 3. (?)2 Algorithm 1 runs in timenO(d).
286
Using Lemma 2 and Lemma 3 we obtain the following theorem.
287
2 The proofs of results marked with?will appear in the full version of the paper.
ITheorem 4. d-Bounded Independence Preserver admits an algorithm running in
288
timenO(d).
289
4 A Polynomial Kernel for D
d-CF-FVS
290
In this section, we design a kernelization algorithm forDd-CF-FVS.
291
To design a kernelization algorithm forDd-CF-FVS, we define another problem calledDd-
292
Disjoint-CF-FVS(Dd-DCF-FVS, for short). We first define the problemDd-DCF-FVS
293
formally, and then explain its uses in our kernelization algorithm.
294
Dd-Disjoint-CF-FVS(Dd-DCF-FVS) Parameter: k Input: An undirected graph G, a graph H ∈ Dd such that V(G) =V(H), a subset R⊆V(G), and a non-negative integerk.
Question: Is there a setS⊆V(G)\Rof size at mostk, such thatG−Sdoes not have any cycle andS is an independent set inH?
295
Notice thatDd-CF-FVS is a special case of Dd-DCF-FVS, whereR = ∅. Given an
296
instance ofDd-CF-FVS, the kernelization algorithm creates an instance ofDd-DCF-FVS
297
by settingR=∅. Then it applies a kernelization algorithm forDd-DCF-FVS. Finally, the
298
algorithm takes the instance returned by the kernelization algorithm forDd-DCF-FVSand
299
generates an instance ofDd-CF-FVS. Before moving forward, we note that the purpose
300
of having set R is to be able to prohibit certain vertices to belong to a solution. This is
301
particularly useful in maintaining the independent set property of the solution, when applying
302
reduction rules which remove vertices from the graph (with an intention of it being in a
303
solution).
304
We first focus on designing a kernelization algorithm forDd-DCF-FVS, and then give
305
a polynomial time linear parameter preserving reduction fromDd-DCF-FVStoDd-CF-
306
FVS. If the kernelization algorithm forDd-DCF-FVSreturns that (G, H, R, k) is aYES
307
(NO) instance ofDd-DCF-FVS, then conclude that (G, H, k) is aYES(NO) instance of
308
Dd-CF-FVS. In the following, we describe a kernelization algorithm forDd-DCF-FVS. Let
309
(G, H, R, k) be an instance ofDd-DCF-FVS. The algorithm starts by applying the following
310
simple reduction rules.
311
IReduction Rule 1.
312
(a) Ifk≥0 andGis acyclic, then return that (G, H, R, k) is aYESinstance ofDd-DCF-
313
FVS.
314
(b) Return that (G, H, R, k) is a NO instance of Dd-DCF-FVS, if one of the following
315
conditions is satisfied:
316
(i) k≤0 andGis not acyclic,
317
(ii) Gis not acyclic andV(G)⊆R, or
318
(iii) There are more than kisolated cycles inG.
319
IReduction Rule 2.
320
(a) Letv be a vertex of degree at most 1 inG. Then deletev from the graphsG, H and the
321
setR.
322
(b) If there is an edge inG(H) with multiplicity more than 2 (more than 1), then reduce
323
its multiplicity to 2 (1).
324
(c) If there is a vertex v with self loop in G. Ifv /∈R, delete v from the graphs Gand
325
H, and decrease k by one. Furthermore, add all the vertices inNH(v) to the set R,
326
otherwise return that (G, H, R, k) is aNOinstance ofDd-DCF-FVS.
327
(d) If there are parallel edges between (distinct) verticesu, v∈V(G) inG:
328
(i) Ifu, v∈R, then return that (G, H, R, k) is aNOinstance ofDd-DCF-FVS.
329
(ii) Ifu∈R (v∈R), delete v (u) from the graphsGandH, and decreasek by one.
330
Furthermore, add all the vertices inNH(v) (NH(u)) to the setR.
331
It is easy to see that the above reduction rules are correct, and can be applied in
332
polynomial time. In the following, we define some notion and state some known results,
333
which will be helpful in designing our next reduction rules.
334
IDefinition 4. For a graphG, a vertexv∈V(G), and an integert∈N, at-flower atv is a
335
set oft vertex disjoint cycles whose pairwise intersection is exactly{v}.
336
IProposition 1. [8, 19, 25] For a graphG, a vertexv∈V(G) without a self-loop inG, and
337
an integerk, the following conditions hold.
338
(i) There is a polynomial time algorithm, which either outputs a (k+ 1)-flower atv, or it
339
correctly concludes that no such (k+1)-flower exists. Moreover, if there is no (k+1)-flower
340
at v, it outputs a setXv⊆V(G)\ {v} of size at most 2k, such thatXv intersects every
341
cycle passing throughv in G.
342
(ii) If there is no (k+ 1)-flower at v in Gand the degree of v is at least 4k+ (k+ 2)2k.
343
Then using a polynomial time algorithm we can obtain a setXv ⊆V(G)\ {v} and a
344
set Cv of components ofG[V(G)\(Xv∪ {v})], such that each component inCv is a tree,
345
v has exactly one neighbor in C∈ Cv, and there exist at leastk+ 2 components inCv
346
corresponding to each vertex x∈Xv such that these components are pairwise disjoint
347
and vertices inXv have an edge to each of their associated components.
348
IReduction Rule 3. Consider v∈V(G), such that there is a (k+ 1)-flower atv inG. If
349
v∈R, then return that (G, H, R, k) is aNOinstance ofDd-DCF-FVS. Otherwise, deletev
350
fromG, H and decrease kby one. Furthermore, add all the vertices inNH(v) toR.
351
The correctness of the above reduction rule follows from the fact that such a vertex must
352
be part of every solution of size at mostk. Moreover, the applicability of it in polynomial
353
time follows from Proposition 1 (item (i)).
354
I Reduction Rule 4. Let v ∈ V(G), Xv ⊆V(G)\ {v}, and Cv be the set of components
355
which satisfy the conditions in Proposition 1(ii) (inG), then delete edges betweenv and the
356
components of the setCv, and add parallel edges betweenv and every vertexx∈Xv in G.
357
The polynomial time applicability of Reduction Rule 4 follows from Proposition 1. And,
358
in the following lemma, we prove the safeness of this reduction rule.
359
ILemma 5. (?)Reduction Rule 4 is safe.
360
In the following, we state an easy observation, which follows from non-applicability of
361
Reduction Rule 1 to 4.
362
IObservation 6. Let (G, H, R, k) be an instance ofDd-DCF-FVS, where none of Reduction
363
Rule 1 to 4 apply. Then the degree of each vertex inGis bounded byO(k2).
364
Proof. As Reduction Rule 3 is not applicable, then there is nok+ 1-flower inG. Now, if
365
there is v ∈ V(G) with degree at least 4k+ (k+ 2)2k, then Reduction Rule 4 would be
366
applicable. J
367
To design our next reduction rule, we construct an auxiliary graph G?. Intuitively
368
speaking, G? is obtained from Gby shortcutting all degree two vertices. That is, vertex
369
set ofG? comprises of all the vertices of degree at least three in 3. From now on, vertices
370
of degree at least 3 (inG) will be referred to as high degree vertices. For eachuv∈E(G),
371
whereu, v are high degree vertices, we add the edgeuvinG?. Furthermore, for an induced
372
maximal pathPuv, betweenuandv where all the internal vertices of Puv are degree two
373
vertices inG, we add the (multi) edgeuvtoE(G?). Next, we will use the following result to
374
bound the number of vertices and edges inG?.
375
IProposition 2. [8] A graphGwith minimum degree at least 3, maximum degree ∆, and a
376
feedback vertex set of size at mostk has at most (∆ + 1)kvertices and 2∆k edges.
377
The above result (together with the construction of G?) gives us the following (safe)
378
reduction rule.
379
IReduction Rule 5. If|V(G?)| ≥4k2+ 2k2(k+ 2) or|E(G?)| ≥8k2+ 4k2(k+ 2), then return
380
NO.
381
ILemma 7. Let(G, H, R, k)be an instance ofDd-DCF-FVS, where none of the Reduction
382
Rules 1 to 5 are applicable. Then we obtain the following bounds:
383
The number of vertices of degree at least3 in Gis bounded byO(k3).
384
The number of maximal degree two induced paths inGis bounded byO(k3).
385
Having shown the above bounds, it remains to bound the number of degree two vertices
386
inG. We start by applying the following simple reduction rule to eliminate vertices of degree
387
two inG, which are also inR.
388
IReduction Rule 6. Letv ∈R, dG(v) = 2, and x, y be the neighbors ofv in G. Delete v
389
from the graphsG, H and the setR. Furthermore, add the edgexyin G.
390
The correctness of this reduction rule follows from the fact that vertices inRcan not be part
391
of any solution and all the cycles passing throughvalso passes through its neighbors.
392
In the polynomial kernel for the Feedback Vertex Set problem (with no conflict
393
constraints), we can short-circuit degree two vertices. But in our case, we cannot perform
394
this operation, since we also need the solution to be an independent set in the conflict
395
graph. Thus to reduce the number of degree two vertices inG, we exploit the properties
396
of ad-degenerate graph. To this end, we use the tool that we developed in Section 3. This
397
immediately gives us the following reduction rule.
398
IReduction Rule 7. LetP be a maximal degree two induced path inG. If|V(P)| ≥nd+1+ 1,
399
apply Algorithm 1 with input (H, V(P)\R). LetVb(P) be the set returned by Algorithm 1.
400
Letv∈(V(P)\R)\Vb(P), andx, ybe the neighbors ofv inG. Deletev from the graphs
401
G, H. Furthermore, add edgexy inG.
402
ILemma 8. Reduction Rule 7 is safe.
403
Proof. Let (G, H, R, k) be an instance of Dd-DCF-FVSandv be a vertex in a maximal
404
degree two pathP with neighborsxandy, with respect to which Reduction Rule 8 is applied.
405
Furthermore, let (G0, H0, R, k) be the resulting instance after application of the reduction
406
rule. We will show that (G, H, R, k) is a YES instance of Dd-DCF-FVS if and only if
407
(G0, H0, R, k) is aYES instance ofDd-DCF-FVS.
408
In the forward direction, let (G, H, R, k) be aYESinstance ofDd-DCF-FVS andS be
409
one of its minimal solution. Consider the case whenv /∈S. In this case, we claim that S
410
is also a solution ofDd-DCF-FVSfor (G0, H0, R, k). Suppose not then eitherS is not an
411
independent set inH0 orG0−Scontains a cycle. Since,H0 is an induced subgraph ofH, we
412
have thatS0 is also an independent set inH0. So we assume thatG0−S has a cycle, sayC.
413
IfC does not contain the edgexy, thenC is also a cycle inG−S. Therefore, we assume
414
thatC contains the edgexy. But then (C\ {xy})∪ {xv, vy} is a cycle inG−S. Next, we
415
consider the case whenv ∈S. By Lemma 2 we have a vertex v0 ∈ V(P)\ {v} such that
416
(S\ {v})∪ {v0}is an independent set inH0. By using the fact that any cycle that passes
417
throughvalso contains all vertices in P (together with the discussions above) imply that
418
(S\ {v})∪ {v0}is a solution ofDd-DCF-FVSfor (G0, H0, R, k).
419
In the reverse direction, let (G0, H0, R, k) be aYESinstance ofDd-DCF-FVSandS0
420
be one of its minimal solution. We claim thatS0 is also a solution of Dd-DCF-FVSfor
421
(G, H, R, k). Suppose not, then eitherS is not an independent set inH orG−S contains a
422
cycle. Since,H0 is an induced subgraph ofH, we have thatS0is also an independent set inH.
423
Next, assume that there is a cycleC inG−S. The cycleCmust containv, otherwise,Cis
424
also a cycle inG0−S0. Sincev is a degree two vertex inG, therefore any cycle that contains
425
v, must also containxandy. As observed before, G− {xv, vy} is identical to G0− {xy}.
426
But then, (C\ {xv, vy})∪ {xy}is a cycle inG0−S0, a contradiction. This concludes thatS0
427
is a solution ofDd-DCF-FVSfor (G, H, R, k). J
428
I Lemma 9. (?) Let (G, H, R, k) be an instance of Dd-DCF-FVS, where none of the
429
Reduction Rules 1 to 7 are applicable. Then the number of vertices in a degree two induced
430
path in Gis bounded byO(kO(d)).
431
ITheorem 10. Dd-DCF-FVS admits a kernel with O(kO(d))vertices.
432
ILemma 11. (?)There is a polynomial time parameter preserving reduction fromDd-DCF-
433
FVStoDd-CF-FVS.
434
By Theorem 10 and Lemma 11, we obtain the following result.
435
ITheorem 12. Dd-CF-FVS admits a kernel withO(kO(d))vertices.
436
5 Kernelization Complexity of P
≤3??- CF-OCT
437
In this section, we show thatCF-OCTdoes not admit a polynomial kernel when the conflict
438
graph belongs to the familyP≤3??. LetP≤3 denotes the family of disjoint union of paths of
439
length at most three, andP≤3? denotes the family of disjoint union of paths of length at most
440
three and a star graph. We give parameter preserving reduction fromP≤3? -Conflict Free
441
s-t Cut(P≤3? -CF-s-t Cut) toP≤3??-CF-OCT.
442
We first prove that P≤3? -CF-s-t Cut is NP-hard. Then, we prove thatP≤3? -CF-s-t
443
Cut does not admit a polynomial compression, unless NP ⊆ coNPpoly using the method of
444
cross-composition.
445
I Theorem 13 (?). P≤3? -CF-s-t Cut does not admit a polynomial compression unless
446
NP⊆ coNPpoly.
447
Lower Bound for Kernel of P≤3??-CF-OCT.In this subsection, we prove the main
448
result of this section. We show that there does not exist a polynomial kernel of P≤3??-
449
CF-OCT. Towards this we give a parameter preserving reduction fromP≤3? -CF-s-t Cut
450
toP≤3??-CF-OCT. Given an instance (G, H, s, t, k) of P≤3? -CF-s-t Cut, we construct an
451
instance (G0, H0, k+ 1) of P≤3??-CF-OCT as follows. Initially, we haveV(G0) =V(H0) =
452
V(G)∪ {z, a, b}. Now, for each edgeei∈E(G), add a vertexwi toV(G0) andV(H0). Now,
453
we define the edge set ofG0. Letxi, yi be end points ofei∈E(G). For eachei∈E(G), add
454
edgesxiwi andyiwi toE(G0). Also, add a self loop onz inG0 and edges sa, abandbtto
455
E(G0). To construct the edge set ofH0, we setE(H0) =E(H− {s, t}). Additionally, we add
456
zs, zt, za, zt, andzwifor each wi ∈V(H0) toE(H0). Figure 1 describes the construction of
457
G0 andH0.
458
V(G)
GraphG0 GraphH0
H xi
yi xj yj
s
t
wi
wj z
a
b
... ... ......
z a b
w1 w2
...
w|E(G)|
Figure 1An illustration of construction of graphG0andH0 in reduction fromP≤3? -CF-s-tCut toP≤3??-CF-OCT.
Clearly,H0 belongs toP3?? and this construction can be carried out in the polynomial
459
time. Now, we prove the equivalence between the instances (G, H, s, t, k) ofP≤3? -CF-s-t
460
Cutand (G0, H0, k+ 1) ofP≤3??-CF-OCTin the following lemma.
461
ILemma 14. (G, H, s, t, k)is a yes-instance ofP≤3? -CF-s-tCutif and only if(G0, H0, k+1)
462
is a yes-instance ofP≤3??-CF-OCT.
463
Proof. In the forward direction, let (G, H, s, t, k) be a yes-instance ofP≤3? -CF-s-t Cut
464
andS be one of its solution. We claim that S∪ {z} is a solution toP≤3??-CF-OCT in
465
(G0, H0, k+ 1). In the graphG0, since we subdivide each edge, all the paths froms−t are of
466
even length. Since, we subdivide each edge ofG,G0− {a, b, z} is a bipartite graph. Hence,
467
an odd cycle inG0−zconsists of ans−tpath inG0− {a, b}and edgessa,abandbt. Clearly,
468
by the construction ofG0, (G0− {a, b})\S does not contain ans−tpath and henceG0−z
469
does not contain an odd cycle. Since,H[S] is edgeless,S∪ {z} is an independent set inH0.
470
This completes the proof in the forward direction.
471
In the reverse direction, letS be a solution toP≤3??-CF-OCTin (G0, H0, k+ 1). Since,
472
z ∈ S, therefore, s, t, a, b, wi ∈/ S for any wi ∈ V(H0). We claim that S0 = S \ {z}
473
is a solution to P≤3? -CF-s-t Cut in (G, H, s, t, k). Suppose not, then there exists a
474
s−t path (s, x1, x2,· · · , xl, t) in G\ S0. Correspondingly, there exists a s−t path
475
(s, w1, x1, w2, x2,· · · , xl, wl+1, t) in G0 of even length which results into an odd cycle
476
(s, w1, x1, w2, x2,· · · , xl, wl+1, t, b, a) inG0\S, a contradiction. This completes the proof.
477
J
478
Now, we present the main result of this section in the following theorem.
479
ITheorem 15. P≤3??-CF-OCTdoes not admit a polynomial kernel. unlessNP⊆ coNPpoly.
480
6 Conclusion
481
In this paper we studied kernelization complexity ofDd-CF-FVS andDd-CF-OCT. We
482
showed that the former admits a polynomial kernel of sizekO(d), whileDd-CF-OCTdoes not
483
admit any polynomial kernel unlessNP⊆ coNPpoly. In fact, the later does not admit polynomial
484
kernel even for much more specialized problem, namelyP≤3??-CF-OCT. Using much more
485
involved marking scheme we can show thatDd-CF-ECT admits polynomial kernel of size
486
kO(d). Similarly, we can extend the known polynomial kernel forOCTtoCF-OCTwhen
487
the conflict graphH has maximum degree at most one. Two most interesting questions that
488
still remain open form our work are following: (a) doesCF-FVSadmit uniform polynomial
489
kernel on graphs of bounded expansion; and (b) doesCF-OCTadmit a polynomial kernel
490
whenH is disjoint union of paths of length at most 2.
491