Kernelization algorithms for graph and other structure modiﬁcation problems

Kernelization algorithms for graph and other
structure modiﬁcation problems

Anthony P EREZ
´
Supervisors: Stephane B ESSY and Christophe PAUL
(AlGCo Team)

November 14

I NTRODUCTION

(Graph) Modification problems

Input: A graph (or another structure) and a (graph) property G.
Output: A minimum number of modification of the graph in order to
satisfy G.

modification: adding edges, deleting edges, deleting vertices, ...

2 / 42

I NTRODUCTION


C LUSTER E DITING
Input: A graph G = (V , E).
Output: A set F ⊆ (V × V ) of minimum size such that the graph
H = (V , E F ) is a disjoint union of cliques.

3 / 42

I NTRODUCTION


Cover a broad range of NP-Hard problems:
VERTEX COVER
FEEDBACK VERTEX SET
More general: F - MINOR DELETION
EDGE - MULTICUT

4 / 42

I NTRODUCTION


Find applications in various domains:
bioinformatics
machine learning
relational databases
image processing

4 / 42

I NTRODUCTION

Different approaches

Most modiﬁcation problems are NP-hard.
How to solve them efﬁciently?

Approximation algorithms
Exact exponential algorithms
Preprocessing steps (heuristics)

5 / 42

I NTRODUCTION




How to measure the efﬁciency of heuristics?

5 / 42

I NTRODUCTION




Exploit the fact that the number of modiﬁcations needed should be
small compared to the instance size n.

5 / 42

Outline of the talk

1 Parameterized complexity

Part I. Graph Modiﬁcation Problems

2 Branches and generic reduction rules

3 P ROPER I NTERVAL C OMPLETION

Part II. Different modiﬁcation problems

4 Considered problems

5 F EEDBACK A RC S ET IN TOURNAMENTS

PARAMETERIZED COMPLEXITY

Parameterized problem

G-M ODIFICATION
Input: A graph G = (V , E), k ∈ N.
Parameter: k .
Output: A set F ⊆ (V × V ) of size at most k such that the graph
H = (V , E F ) belongs to G.

Idea. Measure the complexity of a problem with respect to
some parameter k.

7 / 42


Parameterized problem

G-M ODIFICATION
Parameter: k .
Output: A set F ⊆ (V × V ) of size at most k such that the graph
H = (V , E F ) belongs to G.

Parameterized algorithm
A problem parameterized by some k ∈ N admits a parameterized
algorithm if it can be solved in time f (k ) · nO(1) .

7 / 42


Kernels

Given an instance (I, k ) of a parameterized problem L,
a kernelization algorithm:

runs in time Poly (|I| + k)
and outputs an instance (I , k ) such that:
(i) (I, k ) ∈ YES ⇔ (I , k ) ∈ YES
(ii) |I | h(k ) and k k

(I , k ) (I , k )
Poly (|I | + k )
|I | h(k )
k k
8 / 42


Kernels


(i) (I, k ) ∈ YES ⇔ (I , k ) ∈ YES

Theorem (Folklore)
Parameterized algorithm ⇔ Kernelization algorithm

8 / 42


Kernels


(i) (I, k ) ∈ YES ⇔ (I , k ) ∈ YES

Size: super-polynomial

8 / 42


Kernels


(i) (I, k ) ∈ YES ⇔ (I , k ) ∈ YES

Size: super-polynomial

Do all parameterized problems admit polynomial kernels?

8 / 42


Lower bounds for kernels

There exist some parameterized problems that do not admit polynomial
kernels. (under a complexity-theoretic assumption)
(i) Or-composition [Bodlaender et al., 2008 - Fortnow and Santhanam, 2008]
(ii) Polynomial time and parameter transformations
[Bodlaender et al., 2009]
(iii) Cross-composition [Bodlaender et al., 2011]

9 / 42

Graph modiﬁcation problems



G-M ODIFICATION
Parameter: k.
Output: A set F ⊆ (V × V ) of size at most k s.t. the graph H = (V , E F ) belongs to G.

B RANCHES AND GENERIC REDUCTION RULES P ROPER I NTERVAL C OMPLETION

Generic reduction rules

Connected component.

If G is hereditary and closed under disjoint union, remove any
connected component C that belongs to G.

11 / 42


Sunﬂower.
Consider a ﬁnite forbidden induced subgraph of G (obstruction).
For any pair e ⊆ (V × V ) that belongs to a set of k + 1 obstructions
pairwise intersecting exactly in e, transform G into (V , E {e}).

12 / 42


Critical clique.
Assume G is hereditary and closed under true twin addition.
For any critical clique T with |T | > k + 1, remove |T | − (k + 1)
arbitrary vertices from T .

u

v

13 / 42


Critical clique.
Assume G is hereditary and closed under true twin addition.
For any critical clique T with |T | > k + 1, remove |T | − (k + 1)
arbitrary vertices from T .

k =1

Lemma [Bessy, Paul and P., 2010]
There always exists an optimal edition
that preserves the critical cliques.

k =1

13 / 42


Branches: a natural idea

Reduce set of vertices that induce a graph belonging to G.
The Connected Component rule is a Branch reduction rule.

14 / 42




Context: can be used on problems where G admits a so-called
adjacency decomposition.

Branch: set of vertices B ⊆ V
such that:

(i) G[B] ∈ G and,
(ii) B is connected properly
to the rest of the graph.

14 / 42




Context: can be used on problems where G admits a so-called
adjacency decomposition.
G [B ] ∈ G

Branch: set of vertices B ⊆ V B
such that:

(i) G[B] ∈ G and,
(ii) B is connected properly
to the rest of the graph.
GB

14 / 42

Outline

Branches

Deﬁnition and known results
Branches
Reducing the branches


Deﬁnition
P ROPER I NTERVAL C OMPLETION
Parameter: k .
Output: A set F ⊆ (V × V ) E of size at most k such that
H = (V , E ∪ F ) is a proper interval graph.

16 / 42


Deﬁnition
Parameter: k .

NP-Complete [Golumbic et al., 1994]
FPT : O(24k m) (motivated by applications in genomic research)
[Kaplan, Shamir and Tarjan, 1994]
Polynomial kernel?

16 / 42


Deﬁnition
Parameter: k .

Theorem [Bessy and P., 2011]
The P ROPER I NTERVAL C OMPLETION problem admits a kernel with
O(k 4 ) vertices.

16 / 42


Some useful results

A graph is a proper interval graph if and only if:

it does not contain any of the following graphs as an induced
subgraph.

claw 3-sun net p-cycle (p ≥ 4)

[Wegner, 1967]

17 / 42


Some useful results

A graph is a proper interval graph if and only if:

its vertices admit an ordering v1 . . . vn such that:

vi vj ∈ E i < j ⇒ vi vl , vl vj ∈ E, i < l < j

[Looges and Olartu, 1993]

17 / 42



Remarks. Proper interval graphs are hereditary and:
(i) closed under disjoint union:
the Connected Component rule can be applied.
(ii) do not admit any claw or C4 as an induced subgraph:
the Sunﬂower rule can be applied.
(iii) closed under true twin addition:
the Critical Clique rule can be applied.

18 / 42



Remarks. Proper interval graphs are hereditary and:
(i) closed under disjoint union:
the Connected Component rule can be applied.
(ii) do not admit any claw or C4 as an induced subgraph:
the Sunﬂower rule can be applied.
(iii) closed under true twin addition:
the Critical Clique rule can be applied.

What about branches?

18 / 42


Adjacency decomposition

1 3

6 8
(a) 2

4 5 7 9

4
6 7
(b) 3
8
2 1 2 3 4 5 6 7 8 9
5 9
1

19 / 42


Adjacency decomposition

1 3

6 8
(a) 2

4 5 7 9

4
6 7
(b) 3
8
2 1 2 3 4 5 6 7 8 9
5 9
1

Branches can be used on P ROPER I NTERVAL C OMPLETION.

19 / 42


How to deﬁne a branch?

Consider the structure of a solution.
Look at unaffected vertices.

20 / 42


Branches
B
B1 BR B2
b1 bl
L R C

A subset B of V is a branch if:
(i) G[B] is a connected PIG with umbrella ordering σB = b1 , . . . , b|B| ,
(ii) The vertex set V B can be partitioned into sets L, R and C with:
no edges between B and C
every vertex in L (resp. R) has a neighbor in B
NB (L) ⊂ NB [b1 ] = {b1 , . . . , bl }
NB (R) ⊂ NB [b|B| ] = {bl , . . . , b|B| }
NL (bi+1 ) ⊆ NL (bi ) for every 1 ≤ i < l and NR (bi ) ⊆ NR (bi+1 ) for
every l ≤ i < |B|

21 / 42


Branches
B
B1 BR B2
b1 b|B|
L R C

NB (L) ⊂ NB [b1 ] = {b1 , . . . , bl }
NB (R) ⊂ NB [b|B| ] = {bl , . . . , b|B| }
every l ≤ i < |B|

21 / 42


Branches
B
B1 BR B2
b1 bl bl b|B|
L R C

NB (L) ⊂ NB [b1 ] = {b1 , . . . , bl }
NB (R) ⊂ NB [b|B| ] = {bl , . . . , b|B| }
every l ≤ i < |B|

21 / 42


Branches

B
B1 BR B2
b1 bl
L R C

If L = ∅ (or R = ∅), B is a 1-branch, otherwise B is a 2-branch
If B is a clique, we call B a K-join

22 / 42


Reducing the K -joins

Cannot be done directly.

x y z t

24 / 42




A clean K -join does not intersect any claw or C4 .

24 / 42




A clean K -join does not intersect any claw or C4 .

Assuming the graph is reduced by the generic rules, we can remove
O(k 3 ) vertices from any K -join to obtain a clean K -join.

24 / 42


Reducing the clean K -joins

Let B be a clean K -join of size at least 2k + 2. Let Bf be the k + 1 ﬁrst
vertices of B, Bl be its k + 1 last vertices and M = B (Bf ∪ Bl ).
Remove the set of vertices M from G.

Bf (k + 1 vertices) M Bl (k + 1 vertices)

25 / 42


Reducing the clean K -joins

Let B be a clean K -join of size at least 2k + 2. Let Bf be the k + 1 ﬁrst
vertices of B, Bl be its k + 1 last vertices and M = B (Bf ∪ Bl ).
Remove the set of vertices M from G.
Can be carried out in polynomial time!

25 / 42



In polynomial time, the 1- and 2-branches can be reduced to O(k 3 )
vertices.

Remove 2k + 1 vertices

BR B1
R G (B ∪ R )
B

26 / 42



In polynomial time, the 1- and 2-branches can be reduced to O(k 3 )
vertices.

Remove 2k + 1 vertices

BR B1
R G (B ∪ R )
B
2k + 1 vertices Remove 2k + 1 vertices

B1 B1 BR B2 B2
L R
B

26 / 42


Main result

O(k 4 ) vertices.

1-branch K -join K -join 2-branch K -join 1-branch

27 / 42


Main result

O(k 4 ) vertices.

1-branch K -join K -join 2-branch K -join 1-branch
O (k 3 ) O (k 3 ) O (k 3 ) O (k 3 ) O (k 3 ) O (k 3 )

27 / 42


Main result

O(k 4 ) vertices.

Related result [Bessy, Paul and P., 2010]
The C LOSEST 3-L EAF P OWER problem admits a kernel with O(k 3 )
vertices.

27 / 42

Different modification problems



Π-E DITION
Input: A dense set R of p-sized relations defined over an universe V , an integer k ∈ N.
Parameter: k.
Output: A set F ⊆ R of size at most k whose modification satisfies Π.

C ONSIDERED PROBLEMS F EEDBACK A RC S ET IN TOURNAMENTS

F EEDBACK A RC S ET IN TOURNAMENTS (FAST)

Input: A tournament T = (V , A) and an integer k ∈ N.
Parameter: k .
Output: A set at most k arcs whose reversal results in an acyclic
tournament.
1 4

3 1 4 2
2 3

29 / 42


F EEDBACK A RC S ET IN TOURNAMENTS (FAST)

Input: A tournament T = (V , A) and an integer k ∈ N.
Parameter: k .
Output: A set at most k arcs whose reversal results in an acyclic
tournament.

NP-Complete [Charbit et al., 2007]
Admits constant-factor approximation algorithms [Kenyon-Mathieu and
Schudy, 2007]

29 / 42


D ENSE R OOTED T RIPLET I NCONSISTENCY (RTI)
Input: A set of leaves V and a dense collection R of rooted binary
trees over three leaves of V .
Parameter: k .
Output: A set of at most k triplets whose modiﬁcation leads to a
collection admitting a consistent rooted binary tree deﬁned over V .
t1 t2 t3 t4

a b c c d b a b d a c d

R := {t1 , t2 , t3 , t4 }
R := {ab|c, cd|b, ab|d, ac|d}

30 / 42


Parameter: k .
t1 t2 t3 t4

a b c c d b a b d a c d

R := {t1 , t2 , t3 , t4 }
R := {ab|c, cd|b, ab|d, ac|d}

T is not consistent with R
a b c d
30 / 42


Parameter: k .
t1 t2 t3 t4

a b c c d b a b d c d a

R := {t1 , t2 , t3 , t4 }
R := {ab|c, cd|b, ab|d, cd|a}

T is consistent with R
a b c d
30 / 42


Parameter: k .

NP-Complete [Barky et al., 2010]
Does not admit a constant-factor approximation algorithm yet

30 / 42

Outline

F EEDBACK A RC S ET IN TOURNAMENTS
D ENSE R OOTED T RIPLET I NCONSISTENCY
Conﬂict Packing

Reduction rules
Conﬂict Packing


Consistency

FAST (folklore)
The following properties are equivalent:
(i) T is acyclic
(ii) T does not contain any directed triangle

32 / 42


Consistency

RTI [Guillemot and Mnich, 2010]
The following properties are equivalent:
(i) R is consistent
(ii) R does not contain any conﬂict on four leaves

Conﬂict. Set of vertices C ⊆ V that does not admit a consistent rooted binary tree.

32 / 42


Parameterized complexity

√
FAST RTI
1/3
FPT O ∗ (2 k log k ) a FPT O ∗ (2k log k ) b
Kernel with O(k 2 ) verticesa Kernel with O(k 2 ) verticesb
Linear vertex-kernel c No such result known before.
a
[Alon et al., 2009]
b
[Guillemot and Mnich, 2010]
c
[Bessy et al., 2009]

33 / 42


Parameterized complexity

√
FAST RTI
1/3
FPT O ∗ (2 k log k ) a FPT O ∗ (2k log k ) b
Kernel with O(k 2 ) verticesa Kernel with O(k 2 ) verticesb
Linear vertex-kernel c No such result known before.
a
[Alon et al., 2009]
b
[Guillemot and Mnich, 2010]
c

The linear vertex-kernel for FAST described by [Bessy et al., 2009]
uses a constant-factor approximation algorithm.
Their reduction rules can be adapted to RTI.
But no constant-factor approximation!

33 / 42


Conflict Packing

´
[Paul, P. and Thomasse, 2011]
works on problems characterized by some finite conflicts.
maximal collection of p-uplets disjoint conflits C.
provides a lower bound on the number of modification required.
implies that the instance induced by V V (C) is consistent.

34 / 42


Reduction rules

Remove any vertex that is not part of any directed triangle. a .
a
can be carried out in polynomial time.

35 / 42


Reduction rules

Safe partition
Assume V (T ) is ordered under some ordering σ, and let P be a
partition of σ into intervals.

V1 V2 Vl

AI := {uv ∈ A | ∃ i u , v ∈ Vi }
AO := A AI

B is the set of backward arcs of AO (arcs vi vj with i > j).

35 / 42


Reduction rules

Safe partition
P is safe if there exist |B| arc-disjoint conﬂicts using arcs of AO
only.

35 / 42


Safe Partition Reduction Rule

Let P be a safe partition of an ordered tournament T = (V , A, σ).
Reverse every arc of B and decrease k accordingly.

Use constant-factor approximation algorithm.
Use Conﬂict Packing.

36 / 42


Safe Partition Reduction Rule

Let P be a safe partition of an ordered tournament T = (V , A, σ).
Reverse every arc of B and decrease k accordingly.

Main question
How to compute a safe partition in polynomial time?
Use constant-factor approximation algorithm.
Use Conﬂict Packing.

36 / 42


Conﬂict Packing

A conﬂict packing of a tournament is a maximal collection of
arc-disjoint directed triangles.

37 / 42


Conﬂict Packing


Can be computed greedily (i.e. in polynomial time).
Let C be a conﬂict packing. If T = (V , A) is a positive instance then
|V (C)| 3k.

37 / 42


Conﬂict Packing


Conﬂict Packing Lemma [Paul, P. and Thomasse, 2011]
´
Let T = (V , A) be a tournament. There exists an ordering of T whose
backward arcs uv are such that u, v ∈ V (C).

37 / 42


Conﬂict Packing


Lemma [Paul, P. and Thomasse, 2011]
´
Let T = (V , A) be a tournament such that |V | > 4k. There exists a safe
partition that can be computed in polynomial time.

proof

37 / 42


Conﬂict Packing


Corollary [Paul, P. and Thomasse, 2011]
´
F EEDBACK A RC S ET IN TOURNAMENTS admits a kernel with at most 4k
vertices.

37 / 42


Application to the RTI problem

Remove vertices that do not belong to any conflict
Safe Partition reduction rule
Conflict Packing allows to find a Safe Partition

Theorem [Paul, P. and Thomasse, 2011]
´
D ENSE R OOTED T RIPLET I NCONSISTENCY admits a kernel with at most
5k vertices.

38 / 42

Conclusion

6 Our results

7 Open problems

O UR RESULTS O PEN PROBLEMS

Main results

Polynomial kernels
First polynomial kernels:
(i) C LOSEST 3-L EAF P OWER
(ii) P ROPER I NTERVAL C OMPLETION
(iii) C OGRAPH E DGE -E DITION
Improved polynomial kernels:
(i) F EEDBACK A RC S ET IN TOURNAMENTS
(ii) D ENSE R OOTED T RIPLET I NCONSISTENCY
(iii) D ENSE B ETWEENNESS and D ENSE C IRCULAR O RDERING

joint works with: S. Bessy, F. Fomin, S. Gaspers, S. Guillemot, F. Havet, C. Paul,
S. Saurabh and S. Thomasse. ´

40 / 42


Main results

Lower bounds on kernelization:
(i) For any l 7, the Pl -F REE E DGE -D ELETION problem
does not admit a polynomial kernel.
(ii) For any l 4, the Cl -F REE E DGE -D ELETION problem
does not admit a polynomial kernel.

joint work with: S. Guillemot, F. Havet and C. Paul.

40 / 42


Open problems

Do the F EEDBACK V ERTEX S ET IN TOURNAMENTS and C LUSTER
V ERTEX D ELETION problems admit linear vertex-kernels?
Characterize lower bounds for modification problems. details

Can we use branches on other problems?
(e.g. C HORDAL D ELETION)
Can we use Conflict Packing on other problems?
(e.g. (weakly)-fragile constraint modification problems)

41 / 42

Kernelization algorithms for graph and other structure modiﬁcation problems

More Related Content

What's hot (20)

Similar to Kernelization algorithms for graph and other structure modiﬁcation problems (20)

Recently uploaded (20)

Kernelization algorithms for graph and other structure modiﬁcation problems