MIT OpenCourseWare

http://ocw.mit.edu

6.854J / 18.415J Advanced Algorithms

Fall 2008

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.
 18.415/6.854 Advanced Algorithms November 3, 2008

Lecture 16: Approximation Algorithms

Michel X. Goemans

1 Approximation Algorithms
Many optimizations problems arising in practice are N P hard. Under the widely accepted conjecture
that P =6 N P , we cannot compute eciently and exactly an optimal solution for all possible instances
of these problems. Several approaches have been used to deal with this intractability. On one
hand, dynamic programming, branch and bound, and implicit enumeration algorithms always nd
an optimal solution by navigating the space of feasible solutions in a more ecient way than an
exhaustive search, but their running time is not guaranteed to be polynomial in the inputs size. On
the other hand, heuristic algorithms provide a sub-optimal solution to the problem, but their running
time is polynomial in the size of the input problem. In this lecture we will focus on approximation
algorithms, which are heuristics that always nd a solution whose objective value is guaranteed to
be within a certain factor of the optimum solution.

Denition 1 (Approximation Algorithm) Let P be a minimization (resp. maximization) prob

lem with instances I I. An -approximation factor for 1 (resp. 1) algorithm for P is
an algorithm A whose running time is polynomial in the size of the given instance I, and outputs a
feasible solution of cost cA such that cA OP TI (resp. cA OP TI ), where OP TI is the cost
of the optimal solution for instance I.

In this lecture, we will discuss three general techniques of designing approximation algorithms
for NP-hard problems:

1. Using optimal value in the analysis without explicitly knowing it.

2. Linear programming relaxation and rounding.
3. Primal-dual technique.

2 A 3/2-Approximation Algorithm for the Metric TSP

The Traveling Salesman Problem is one of the most extensively studied problems in combinatorial
optimization. In the metric version of the problem, an instance is a complete undirected graph
G = (V, E) and c : E R+ , where c satises the metric property: c(u, v) = c(v, u) for all u, v V ,
and the triangle inequality, c(u, v) c(u, w) + c(w, v), for all u, v, w V . The objective is to nd
tour, that is a cycle visiting every vertex exactly once (also called a tour ) minimum cost .
A 32 approximation algorithm for this problem by Christodes [1] is as follows.

1. Find a minimum spanning tree T of G.

2. Compute a minimum cost perfect matching M on the set of odd-degree vertices Vodd T .
3. Find an Eulerian tour C (a cycle visiting all the edges exactly once) in M T .
4. Output the tour C that visits the vertices of G in the order of their rst appearance in the C .

Lec 16-1
 2 2 2

1 1 1
5 7 5 7 5 7
4 4 4

8 8 8
6 6 6
9 9 9
3 11 3 11 3 11
10 10 10

Figure 1: Execution of Christodes algorithm on an instance. The rst gure shows a minimum
cost spanning tree. The second gure shows the addition of a minimum cost matching on odd degree
vertices in the tree, and the third gure shows a cycle obtained after shortcutting an Eulerian
tour in the previous graph, starting from vertex 1.

Theorem 1 The above algorithm is a 3/2-approximation algorithm for the metric TSP.

Proof: It is clear that all steps in the algorithm can be implemented in polynomial time. The
minimum spanning tree can be found using a greedy algorithm, and the minimum cost matching
for Vodd can be found in polynomial time using the ellipsoid algorithm, as discussed in one of the
previous lectures (or by a purely combinatorial algorithm also based on the linear program we
discussed). Note that c(T ) OP T , because the optimal tour without an edge becomes a tree. Also,
c(M ) OP T /2. To see this, consider any optimal tour, and then short-cut it to get a cycle visiting
only vertices in Vodd with cost at most OP T . Since the cycle induces two matchings consisting of
alternating edges, at least one of them will have cost at most OP T /2. From this, the total cost of the
Eulerian cycle, an upper bound of the cost of the algorithm, is at most OP T + OP T /2 = 3/2 OP T .

Note that in the analysis of the algorithm, we used the value of OP T even without explicitly
computing it exactly, or getting a lower bound on it. Figure 1 shows an instance of the metric TSP,
and the execution of the algorithm on this instance.
A few remarks:
The above analysis for the algorithm is tight, i.e. > 0 there is an instance I such that the
algorithm returns a solution which is 3/2 times the optimal solution.
Currently, no algorithm with an approximation factor better than 3/2 is known for metric
TSP.
TSP is known to be MAX-SNP hard [5] even for the case when distances are either 1 or 2.
Also, Papadimitriou and Vempala [4] have proved that a 1.01 approximation algorithm for the
metric TSP will imply P = N P .

3 Designing Approximation Algorithms via Relaxations

One of the most important paradigms in the design of approximation algorithms are relaxations.
Consider the following (hard) minimization problem.

min f (x)
s.t. x S.

Lec 16-2
One approach to solve this problem is to extend S to a bigger set P S where the same problem
is easier to solve. Namely, we extend the function f to a function g : P R satisfying g(x) =
f (x), x S (or g(x) f (x)). If this condition holds, then

min f (x) min g(x),

xS xP

which gives a lower bound for the value of the optimal solution. Therefore, if an algorithm gives a
solution x S which satises f (x ) minxP g(x), then this is an - approximation algorithm
for the problem.
For example, many combinatorial optimization problems can be expressed as

min cT x
s.t. Ax = b,
x {0, 1}n.

A natural relaxation is to replace the integrality constraint xi {0, 1} by the linear constraint
0 xi 1, we obtain the LP relaxation of the integer program above.

min cT x
s.t. Ax = b,
0 xi 1, i = 1, . . . , n.

In some cases, the polytope corresponding to the LP relaxation has all integral extreme points.
In such cases, it is sucient to solve the LP relaxation to solve the original problem exactly. But
this is not true in general.

3.1 LP Relaxation for the Vertex Cover Problem

Given an undirected graph G = (V, E), a vertex cover of G is a collection of vertices C V such
that all edges e = (u, v) in E satisfy C {u, v} = 6 . The Vertex Coverproblem on an instance
G = (V, E), c : E R+ is to nd a cover C of G of minimum cost c(C) = vC c(v). This is known
to be an NP-hard problem.
A natural formulation using integer variables and linear constraints is the following. We dene a
variable xu {0, 1} which takes value 1 if it is in the vertex cover, 0 otherwise. Then the following
is an integer programming formulation for the vertex cover problem.

min cv xv (1a)
vV
s.t. xu + xv 1, (u, v) E, (1b)
xu {0, 1}, u V. (1c)

The LP relaxation for the vertex cover problem is

min cv xv (2a)
vV
s.t. xu + xv 1, (u, v) E, (2b)
xu 0, u V. (2c)

Note that we removed the xu 1 constraints, since if xu > 1 we can change it to xu = 1 without
increasing the cost, and still have a feasible solution.

Lec 16-3
 x1 = 1/2

x2 = 1/2 x3 = 1/2

Figure 2: An example where the LP relaxation for the Vertex Cover does not have an integer optimal
solution.

The LP relaxation does not necessarily have an optimal integral solution in general. For example,
consider the graph given in Figure 3.1 with all costs equal to 1. The optimal solution for this instance
has cost OP T = 2 , but the optimal solution for the LP relaxation has cost LP = 3/2, as shown in
the gure. What this example shows is not only that LP < OP T in general, but also an interesting
fact about the strength of this relaxation. Suppose that we are going to use LP as a lower bound
on OP T in order to prove an approximation guarantee. As we will see in the next subsection, we
will be able to nd a cover C with cost at most 2LP . Therefore, we can say

c(C) 2LP 2OP T

to prove an approximation guarantee of 2, However, the example proves that we will not be able to
decrease this factor beyond 4/3. This follows from the fact that

OP T c(C) LP OP T OP T /LP
then the best we can hope for is at most 4/3 by using this relaxation. This important property of
the bad examples is captured in the concept of integrality gap.

Denition 2 (Integrality gap) Given a relaxation LP () for an integer program IP () that

formulates a combinatorial (minimization) optimization problem on a collection of instances {},
the integrality gap of the linear program relaxation is the largest ratio between the optimal solution
of both formulations, namely:

IP ()
Integrality gap = sup
LP ()

For the Vertex Cover LP relaxation, the integrality gap is exactly 2. To see that it is at least
2, consider the complete graph G = Kn , with unitary costs. The minimum vertex cover has cost
n 1, while the linear program relaxation can assign 1/2 to all variables, which gives a total cost
of n/2. Therefore, the integrality gap is at least 2(nn1) 2. The upper bound follows from the
2-approximation algorithm we will see in the next subsection.

3.2 A 2-approximation Algorithm for Vertex Cover

A natural approach to get an integral solution from a fractional solution is to round the fractional
values. A simple rounding scheme for the vertex cover is as follows.

1. Solve the linear programming relaxation given by (2a)-(2c), to get the fractional solution x .

Lec 16-4
 2. Compute the vertex cover as C = {v V, xv 1/2} (i.e., round each fractional variable to the
nearest integer).

Theorem 2 The above rounding scheme is a 2-approximation algorithm for the Vertex Cover prob
lem.

Proof: First, we need to check that C is indeed a vertex cover. For each e = (u, v) E, xu +xv 1,
so at least one of xu , xv has value at least 1/2, and is in C. Next, the cost of this vertex cover
satises
c(C) = cv 2 cv xv = 2LP 2OP T,
v:x
v 1/2
vV

hence the LP rounding is a 2-approximation algorithm for the vertex cover problem.
This is a very basic (the simplest) example of rounding; more sophisticated rounding procedures
have been used to design approximation algorithms; well see some in coming lectures.

4 The Primal Dual Technique

Yet another way of designing approximation algorithms for intractable problems is the primal dual
method. The basic idea of the primal dual scheme is this: At every point of the algorithm, we
keep a feasible dual solution, and a corresponding infeasible integer primal solution. The dual
variables are then modied at every step and so is the infeasible primal solution, so as to achieve
primal feasibility. At this point, the dual gives a lower bound (for minimization problems) on the
optimal primal objective function value, which is used to derive the approximation factor for the
algorithm. The interesting thing about this technique is that we do not need to explicitly solve the
linear program (as is the case in rounding); the linear program is used only in the analysis of the
algorithm.
We illustrate this method for the vertex cover problem. The linear program for the vertex cover
problem is given by (2a)-(2c). The dual of this linear program is given by

max ye
eE

s.t. ye cv v V, (3)
e(v)
ye 0 e E.

The primal dual algorithm for the vertex cover problem is as follows. In the algorithm, C
corresponds to the set of vertices in the (supposed to be) vertex cover, and F is the set of edges in
the graph not yet covered by C.

1. y(v) 0 v V, C , F E.
6
2. While F =
3. Let e = (u, v) be any edge in F .
4. Increase ye until the constraint (3) becomes tight for u or v.
5. Add that corresponding vertex (say it is v) to C.
6. F F \ (v).

Theorem 3 The above algorithm achieves an approximation ratio of 2 for the vertex cover problem.

Lec 16-5
 3 4 3 4

3 2 3 2
3 y=3 4 3 y=3 4

y=1

3 2 3 2
3 y=3 4

y=1

3 y=1 2

Figure 3: Illustration of the primal-dual algorithm for the vertex cover problem. The cost of the
vertices are indicated next to each vertex. Dotted edge denotes the edge currently under considera
tion, thick edges denote those already covered by the current vertex cover. The vertices in the cover
are shown as solid circles.

Proof: First of all, it is clear that the set C returned by the algorithm is a vertex cover. Let
y be the dual solution returned. Observe that by construction, this solution is dual feasible (we
dual feasibility throughout the execution). Furthermore, for any v C, we have that
maintain
cv = e(v) ye . Let us now compute the cost of the vertex cover returned by the algorithm.

cv = ( ye ) = e ye 2 ye
vC vC e(v) eE eE

2LP (4a)
2OP T, (4b)
where e = 2, for edge e = (u, v) if both u, v C, 1 otherwise. The inequality (4a) follows from
weak duality, and inequality (4b) follows from the fact that the primal LP is a relaxation of the
vertex cover problem.
Figure 3 illustrates the execution of the primal-dual algorithm on a graph. For this instance,
the algorithm returns a vertex cover of cost 9, whereas the optimal solution in this instance has

Lec 16-6
cost 7 (corresponding to the two vertices on the diagonal edge). The lower bound given by the dual
solution has value 3 + 1 + 1 = 5.
A few nal remarks:

Dinur and Safra [2] have proved that it is NP-hard to approximate to the vertex cover with a
factor better than 1.36.
Currently, there is no algorithm for the vertex cover problem which achieves an approximation
ratio better than 2. So the two (simple!) algorithms presented here are, in fact, the present
best known approximation algorithms for this problem.

Khot and Regev [3] have proved that it is UGC-hard to approximate vertex cover within a
factor 2 , for any > 0.

References
[1] Christodes, N. (1976). Worst-case analysis of a new heuristic for the travelling salesman problem,
Report 388, Graduate School of Industrial Administration, CMU.

[2] Dinur, I. and S. Safra (2002). The importance of being biased. In Proceedings of the 34th ACM
Symposium on Theory of Computing, pp. 33-42.
[3] Khot, S. and O. Regev (2008). Vertex cover might be hard to approximate to within 2 .
Journal of Computer and System Sciences, 74:335-349.
[4] Papadimitriou, C.H. and S. Vempala (2000). On the approximability of the travelling salesman
problem. In Proceedings of the 32nd ACM Symposium on Theory of Computing, pp. 126-133.
[5] Papadimitriou, C.H. and M. Yannakakis (1993). The travelling salesman problem with distances
one and two. Mathematics of Operations Research, 18:1-11.

Lec 16-7