A rounding algorithm for approximating minimum Manhattan networks
Victor Chepoi, Karim Nouioua, Yann Vaxès
LIF - Université de la Méditerranée Marseille - France
Rectilinear paths
A rectilinear (Manhattan) path between two points p,q of the plane is a path consisting of vertical and horizontal segments.
p
q
l
1- Paths
|py – qy|
p
q
An l1-path is a shortest rectilinear path in the l1-metric, i.e. a rectilinear path of length ║p-q║1= |px - qx|+|py - qy|
Manhattan networks
A Manhattan network on a set of n terminals T = { t1 , …, tn } is a network containing an l1-
path
between every pair of terminals of T, i.e. a 1-spanner for T in the l1-plane.
Minimum Manhattan networks
A minimum Manhattan network (MMN) on T is a Manhattan network of minimum total length.
The MMN problem is to find such a network (introduced by Gudmundsson, Levcopoulos and Narasimhan, APPROX 99).
It is not known whether it is in P or not.P
A MMN on 100
terminals
Applications of MMN
VLSI Circuit Design :
A MMN ensures optimal transmission delay at minimum cost.
Computational biology :
Finding efficient search space in Pair Hidden Markov Models alignment algorithm
[2003] F. Lam, M. Alexanderson and L. Pachter, J. Comput. Biology All areas in which geometric spanners occur (e.g., parallel
computations, distributed, communication networks, wireless networks, etc., etc.)
Previous results
[1999] J. Gudmundsson, C. Levcopoulos and G. Narasimhan Factor 8 approximation algorithm , O(n log n)
Factor 4 approximation algorithm, O(n3) [2002] R. Kato, K. Imai and T. Asano
Factor 2 approximation algorithm, O(n3) [2004] M. Benkert, T. Shirabe and A. Wolf
Factor 3 approximation algorithm, O(n log n) [2004] K. Nouioua
Factor 3 approximation algorithm, O(n log n)
Incomplete proof
• The complete grid on T contains at least one MMN on T.
• The part Г=(V,E) of complete grid inside the Pareto enveloppe contains at least one MMN on T.
• The edges on the boundary of belong to all MMN inside .
Properties of
MMN
Let le be the length of the edge e.
1. For each edge e in E, we introduce a 0-1 variable xe :
- xe = 1 if e belongs to the solution - xe = 0 otherwise
2. The objective is to minimize the total length of the network:
Integer Programming Formulation
m i n i m i z e
e ee E
l x
å
Ît
it
je
1
e c
x
Î
å ³
Integer Programming Formulation
3. Constraints : for every pair (ti,tj) of terminals and for every (ti,tj)-cut c of the oriented subgraph Г ij = Г ∩ Rij :
Integer programming formulation
minimize
subject to 1
{0,1},
e e e E
e ij F ij
e c e
l x
x c C
x e E
Î Î Î
³ Î
Î Î
å
å
Let Cij be the collection of all (ti,tj)-cuts in Гij.
Integer programming formulation
( ) ( )
( )
minimize
subject to , , { , }
1,
0 , ,
{0,1},
ij ij
ij i
e e e E
ij ij
e e ij i j
e v e v
ij e
e t
ij
e e ij
e
l x
f f ij v V t t
f ij
f x ij e E
x e E
Î
Î Î
Î
Î
Î
Î Î
å
å å å
+ -
-
Г Г
Г
{ 0 , 1 , }
x
eÎ e Î E
0 x
e 1 , e Î E
Linear Relaxation
0,7 0,7
0.3 0.
3 0.
3 0.
Integer programming
formulation
Fractional optimum = 27.5 Integer optimum = 28
Integer programming
formulation
A generating set is a subset F of pairs of terminals such
that a rectilinear network containing l1-paths for all pairs ij in F is a Manhattan network on T.
Generating sets
For example, the set of all pairs ij such that the rectangle Rij defined by ti and tj is empty is a generating set.
t
tj tk
Horizontal
strips
Vertical strips
Strips and
Staircases
t
it
i't
j't
jt
kStaircase
s
t
it
jStrips and
staircases
1. Solve the linear relaxation of the integer program
2. Execute the procedure Round_Strip to connect all pairs in strips.
3. Execute the procedure Round_Staircase to connect all pairs in staircases.
The algorithm
2. Procedure Round_Strip
x
e+ x
e’≥ 1
e e’
t
it
The algorithm
2. Procedure Round_Strip
0.3 0.4 0.7
0.6 0.6 0.6 0.4
0.4
t
it
jThe algorithm
p’ p
2. Procedure Round_Strip
t
it
j+0. 5 0.5
The algorithm
p’ p
3. Procedure Round_Staircase
t
kt
j’
t
it
jThe algorithm
3. Procedure Round_Staircase
0. 3
0. 1 0. 1
t
kt
j’
t
it
jThe algorithm
3. Procedure Round_Staircase
t
kt
j’
t
lt
it
jThe algorithm
Analys is
Lemma. The network N(T) returned by the algorithm is a
Manhattan network.
Proof idea.
• N(T) contains an l1-path between all pairs in strips and staircases.
• The set of all pairs in strips and staircases is a generating set.
Analys is
Lemma. The length of the network N(T) is at most twice the
cost of an optimal fractional solution.
Proof idea. To every rounded up edge e of N(T), we assign
a set Ee of edges parallel to e such that :
' '
(i) 1
e
2
e e E
x
Î
å ³
(ii) E
e E
f
for any two rounded up edges e, f of N(T),
Main result
Theorem.
Theorem. The rounding algorithm described in this talk achieves an approximation
guarantee of 2 for the Minimum Manhattan
Network problem.
Open
Questions
• Is the MMN problem NP-hard ?
• What is the worst integrality gap ?
• Does there exist an integrality gap in the case when the terminals define a
staircase ?
F-restricted MMN problem
Let F be a subset of pairs of terminals in T.
The F-restricted MMN problem is to find a rectilinear network of minimum length containing an l1-path between every pair of terminals in F.
If (T,F) is a complete graph, then we obtain the MMN problem.
The NP-hard Minimum Rectilinear Arborescence Problem
Integrality gap for F-restricted MMN
Fractional optimum = 2
t
1t
3t
2t
41/2 1/2
1/2
1/2
Integer optimum = 3
t
1t
3t
2t
41
1 1
A simple example shows that, in this case, the integrality
gap is at least 3/2 :