Topic 19 - Intro to Lagrangian Duality

Note this theorem is proved in the textbook it is a

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: etermine that no duality gap exists. Toward that end, we must first define the Lagrangian function as Saddle Point Optimality Theorem - SPOT (Theorem 6.2.5): If there exists an ̅ and ̅ such that ̅ ̅̅ ̅ for all and , then ̅ solves problem P and ̅ solves problem D with no duality gap; that is, ̅ ̅. Conceptually: ̅ ̅ is a high point with respect to all the u’s but a low point with respect to all x’s. Proof: We are given that ̅ ̅̅ ̅ ̅ ̅ For this to hold true for any ̅ ̅ ̅ ̅ , we must have 218 ̅ , so ̅ is feasible to P. Since we have implies that ̅ ̅ ̅ But we had ̅ ̅ ̅ , it must also be true for , which . and ̅ , so that must mean that we have ̅ ̅ . Returning to what we were given in the theorem, the second half of the inequality says that ̅̅ ̅ ̅ ̅ Since we just showed that ̅ ̅ ̅ ̅ , we have ̅ From weak duality, we had ̅ ̅ ̅ ̅ ̅ . Therefore, we must have ̅ ̅ ̅. Relationship between SPOT and KKT Conditions If ̅ is a KKT point and , If ̅ ̅ satisfy SPOT, ̅ are convex for ̅ . Then ̅ ̅ satisfy SPOT. , and , are differentiable, then ̅ is a KKT point. Conceptually: For convex problems, the Lagrange multipliers in the KKT system serve as the multipliers in SPOT. Summary The Lagrangian dual problem can be shown to have an optimal objective value that is equal to that of problem P under certain conditions. was shown to be a concave function, which implies that a local maximum is a global maximum. As a result, maximizing is “relatively” easy, so it could be more appealing than solving problem P. However, it is not always easy to optimize , since a closed form solution for it is frequently unavailable. Usually is obtained by evaluating at a point , which requires the minimization of a separate problem. Another issue is that is not always differentiable, so we need to use algorithms that will work on non-differentiable functions. We have now had a brief introduction to the important aspects of Lagrangian Duality. In the next topic, we will address some techniques for solving the Lagrangian Dual. 219...
View Full Document

This document was uploaded on 11/28/2013.

Ask a homework question - tutors are online