Brian's Digest: Convexity

SCI.OP-RESEARCH Digest Mon, 1 Jul 96 Volume 3: Issue 27

Date: Mon, 24 Jun 1996 16:52:01 -0400
From: Joao Luis Cardoso Soares <jls55@columbia.edu>
Subject: A convex analysis question
Hi

I am trying to figure out what what would be the convex hull of an infinite union of sets. According to most books in Convex Analysis it should be the set of all the _finite_ convex combinations with points in some of those sets.

I cannot understand why does it have to be finite.

Any help on making this clear is appreciated.

Thanks,

Joao

Date: 25 Jun 1996 10:29:35 -0500
From: hennebry@plains.nodak.edu (Michael J. Hennebry)
Subject: A convex analysis question
If one is attempting constrained optimization, then one usually needs closed sets. The union of a finite set of closed sets is closed, but the union of an infinite set of closed sets might not be. In fact, every open set, except the empty set, is the union of an infinite set of distinct non-intersecting closed sets.

Mike hennebry@plains.NoDak.edu Lennier does it with kindness. -- Garibaldi Date: 27 Jun 1996 22:03:28 GMT
From: rubin@msu.edu (Paul A. Rubin)
Subject: A convex analysis question
Probably to avoid dealing with infinite sums. If you define "infinite convex combination" to be an infinite series in which the weights are nonnegative and satisfy the property that the infinite series of weights converges to unity, then the question becomes whether the infinite combination of points converges - which isn't guaranteed, unless you have some additional property working in your favor (such as the sets in the union being uniformly bounded and the space being closed).

By the way, Rockafellar defines the convex hull of a set S to be the intersection of all convex sets containing S. He defines this for the case S subset of R^n, but the definition may well pertain more generally, since the property that the intersection of convex sets is convex seems to be universal.

-- Paul

Date: Thu, 27 Jun 1996 14:59:37
From: JMJ6@pge.com (Jonathan M. Jacobs)
Subject: A convex analysis question
The convex hull of X should be the minimal convex set containing X. Clearly it has to include finite convex combinations of points in X to be convex. That's enough to make it convex so by minimality you're done.

Jonathan Jacobs jmj6@pge.com Pacific Gas & Electric Co. (PG&E) *** Any views or opinions expressed herein are my own and do not *** *** necessarily reflect those of PG&E *** Date: 28 Jun 1996 10:07:49 -0500
From: hennebry@plains.nodak.edu (Michael J. Hennebry)
Subject: A convex analysis question
Not only might infinite convex combinations not converge, they might be hard to define. Consider the weights 1/(J+1) for J=1... . If the corresponding points are singletons J+1 for J=1... , then the series does not converge although the non-negative weights add to 1. Summing over a countably infinite set is bad enough, summing over an uncountably infinte set may be imposible to define, even if the set of points is bounded.

More importantly, infinite convex combinations are not necessary to enforce the convexity property: If P and Q are in a convex set and w in [0, 1], then wP + (1-w)Q is also in that set.

For the N dimensional set R^N, one need only consider convex combinations of up to N+1 points.

Even in Hilbert space, only finite convex combinations need be considered.

Please ignore my previous post. I'd left my head in my other hat.

Mike hennebry@plains.NoDak.edu Lennier does it with kindness. -- Garibaldi Date: 28 Jun 1996 15:53:08 -0500
From: hennebry@plains.nodak.edu (Michael J. Hennebry)
Subject: A convex analysis question
Not only might infinite convex combinations not converge, they might be hard to define. Consider the weights 2^-(J+1) for J=1... . If the corresponding points are the scalars 2^(J+1) for J=1... , then the series does not converge although the non-negative weights add to 1. Summing over a countably infinite set is bad enough, summing over an uncountably infinte set may be imposible to define, even if the set of points is bounded.

More importantly, infinite convex combinations are not necessary to enforce the convexity property: If P and Q are in a convex set and w in [0, 1], then wP + (1-w)Q is also in that set.

For the N dimensional set R^N, one need only consider convex combinations of up to N+1 points.

Even in Hilbert space, only finite convex combinations need be considered.

My previous post wasn't as bad as my first post on this subject, but is still worthy of ignore-ance. I'd given a series of weights that didn't sum to one.

Mike hennebry@plains.NoDak.edu Lennier does it with kindness. -- Garibaldi
SCI.OP-RESEARCH Digest Mon, 8 Jul 96 Volume 3: Issue 28

Date: 2 Jul 1996 11:07:01 -0500
From: rudi3964@utdallas.edu
Subject: A convex analysis question
Proof by induction:
In 1-space (a line), it's clear that the convex hull is simply the set of all convex combinations of the lowest and highest points. 2 is clearly finite.

Now consider point X in the convex hull of the infinite union. Bisect the n-space with an n-1 hyperplane. Consider A(n-1) to be the portion of the convex hull in the n-1 hyperplane. Each extreme point of A(n-1) is on a line between two extreme points in the original complex hull. So however many points are needed in n-1 space, only twice as many are needed in n-space.

By induction, then, any point in the convex hull in the plane can be expressed as a linear combination of four points in the infinite sets. (Actually, three is enough, but that proof doesn't lead to an easy induction). Similarly, with eight points you can find any point in the convex hull in 3-space, etc. In finite dimensions, then, any point in the convex hull can be expressed as a linear combination of a finite set of points from the original sets.

The point is not that infinite combinations aren't allowed, but that every point can be represented as a combination of a finite set of points.

I'm reasonably certain that it will never take more than n+1 points to specify any point in the convex hull, but I haven't constructed a proof.

Jay Rudin
University of Texas at Dallas

Date: 2 Jul 1996 21:40:40 GMT
From: rubin@msu.edu (Paul A. Rubin)
Subject: A convex analysis question
It's Caratheodory's Theorem. See, for instance, Rockafellar's _Convex Analysis_, Thm. 17.1, page 155.

SCI.OP-RESEARCH Digest Mon, 21 Oct 96 Volume 3: Issue 42

Date: 18 Oct 1996 17:33:48 GMT
From: spellucci@mathematik.th-darmstadt.de (Peter Spellucci)
Subject: convex image
hi , you all ,

here comes a nice problem:

obviously a linaer map map's convex sets to convex ones. What about sufficient conditions for a nonlinear map F going from R^n to R^n to map

(Clearly if F(x)=Ax+phi(x) and phi is diagonal and continuous, then a) holds but b) not) Any pointer to the literature will be highly appreciated.

Thanks in advance

peter

SCI.OP-RESEARCH Digest Mon, 28 Oct 96 Volume 3: Issue 43

Date: 21 Oct 96 18:44:45 GMT
From: chinneck@halley.sce.carleton.ca (John Chinneck)
Subject: convex? concave? both?
If you work with nonlinear functions for optimization or modelling, you might be interested in MProbe. MProbe is new software for empirically determining the "shape" of a nonlinear function of many variables in a specified region. By "shape" is meant whether the function is almost linear, convex, almost convex, concave, almost concave, or both convex and concave.

An empirical approach is used because a function whose analytic form appears horribly nonlinear may in fact turn out to be simpler (e.g. almost linear) in a limited region of interest.

MProbe also provides various analyses of function range, a multidimensional analog of "slope", simple model navigation, and the ability to plot a function value along a straight line between any two points in multidimensional space.

For further information on MProbe, or to download a student/demo version of MProbe, please visit the web page at:

http://www.sce.carleton.ca/faculty/chinneck/mprobe.html

If you do not have web access, you can obtain the student/demo version of MProbe via ftp at this address:

http://ftp.sce.carleton.ca

Login as "anonymous", then change to the directory pub/MProbe.

I look forward to your feedback on MProbe. If there is sufficient interest in the software, then I intend to continue adding functionality.

John W. Chinneck Systems and Computer Engineering tel: (613) 520-5733 Carleton University fax: (613) 520-5727 1125 Colonel By Drive email: chinneck@sce.carleton.ca Ottawa, Ontario K1S 5B6 CANADA