28.11.2014 Views

Lecture Notes Discrete Optimization - Applied Mathematics

Lecture Notes Discrete Optimization - Applied Mathematics

Lecture Notes Discrete Optimization - Applied Mathematics

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

mizes the ratio c(C)/|C|. A minimum mean cost cycle can be computed in time O(nm).<br />

Using this idea, one can show that the resulting algorithm has an overall running time of<br />

O(n 2 m 3 logn).<br />

6.4 Successive Shortest Path Algorithm<br />

We derive an alternative optimality condition. Suppose we associate a potential π(u) with<br />

every node u∈ V. Define the reduced cost c π (u,v) of an edge(u,v) as<br />

c π (u,v)=c(u,v)−π(u)+π(v).<br />

Note that this definition is applicable to both the original network and the residual graph.<br />

Theorem 6.4 (Reduced cost optimality conditions). A feasible flow f of G is a minimum<br />

cost flow if and only if there exist some node potentials π : V →R such that c π (u,v)≥0<br />

for every edge(u,v)∈E f of G f .<br />

Proof. Suppose that there exist node potentials such that c π (u,v) ≥ 0 for every edge<br />

(u,v) ∈ E f of the residual graph G f . Let C be an arbitrary simple directed cycle in G f .<br />

Then<br />

∑<br />

e∈C<br />

c(e)= ∑ c π (e)≥0.<br />

e∈C<br />

We conclude that G f does not contain any negative cost cycle. By Theorem 6.1, f is a<br />

minimum cost flow.<br />

Let f be a minimum cost flow. By Theorem 6.1, G f contains no negative cycle. Let<br />

δ : V → R be the shortest path distances from an arbitrarily chosen source node s ∈ V<br />

to every other node u ∈ V (with respect to c). Note that δ is well-defined because G f<br />

contains no negative cycles. The distance function δ must satisfy the triangle inequality<br />

(see Lemma 4.2), i.e., for every edge (u,v) ∈ E f , δ(v) ≤ δ(u)+c(u,v). Define π(u)=<br />

−δ(u) for every node u∈ V. With this definition, we have for every(u,v)∈E f :<br />

which concludes the proof.<br />

c π (u,v)=c(u,v)−π(u)+π(v)= c(u,v)+δ(u)−δ(v)≥ 0,<br />

We next introduce the notion of a pseudoflow. A pseudoflow x of G is a function x : E →<br />

R + that satisfies the nonnegativity and capacity constraints; it need not satisfy the flow<br />

balance constraints. Given a pseudoflow x, define the excess of a node u∈ V as<br />

exs(u)=b(u)+ ∑ x(v,u)− ∑ x(u,v).<br />

(v,u)∈E (u,v)∈E<br />

Intuitively, exs(u)>0means that node u has an excess of exs(u) units of flow; exs(u)

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!