04.02.2015 Views

Stochastic Programming - Index of

Stochastic Programming - Index of

Stochastic Programming - Index of

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

266 STOCHASTIC PROGRAMMING<br />

pos W<br />

pol pos W<br />

Figure 3 Finding the generators <strong>of</strong> pol pos W .<br />

There is another important aspect <strong>of</strong> the polar cone pos W ∗ that we have<br />

not yet discussed. It is indicated in Figure 3 by showing that the generators<br />

are pairwise normals. However, that is slightly misleading, so we have to turn<br />

to a three-dimensional figure to understand it better. We shall also need the<br />

term facet. Let a cone pos W have dimension k. Then every cone K positively<br />

spanned by k−1 generators from pos W , such that K belongs to the boundary<br />

<strong>of</strong> pos W , is called a facet. Consider Figure 4.<br />

What we note in Figure 4 is that the generators are not pairwise normals,<br />

but that the facets <strong>of</strong> one cone have generators <strong>of</strong> the other as normals. This<br />

goes in both directions. Therefore, when we state that h ∈ pos W if and only<br />

if h T y ≤ 0 for all generators <strong>of</strong> pol pos W , we are in fact saying that either h<br />

represents a feasible problem because it is a linear combination <strong>of</strong> columns in<br />

W or because it satisfies the inequality implied by the facets <strong>of</strong> pos W . In still<br />

other words, the point <strong>of</strong> finding W ∗ is not so much to describe a new cone,<br />

but to replace the description <strong>of</strong> pos W in terms <strong>of</strong> generators with another<br />

in terms <strong>of</strong> inequalities.<br />

This is useful if the number <strong>of</strong> facets is not too large. Generally speaking,<br />

performing an inner product <strong>of</strong> the form b T y is very cheap. In parallel<br />

processing, an inner product can be pipelined on a vector processor and the<br />

different inner products can be done in parallel. And, <strong>of</strong> course, as soon as we<br />

find one positive inner product, we can stop—the given recourse problem is<br />

infeasible.<br />

Readers familiar with extreme point enumeration will see that going from

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

Saved successfully!

Ooh no, something went wrong!