Monotone and maximal monotone affine subspaces
Introduction
Variational inequality problems and equilibrium problems occur in a large field of applications arising from various domains such as physics, mechanics, economics and operations research. They encompass optimization problems and related problems such as complementarity problems and saddle-point problems.
With a subset of we associate the multivalued maps and such that The set and the maps and are said to be monotone if
They are said to be maximal monotone if they are monotone and for all monotone subsets of one has .
(Maximal) Monotone maps play in variational inequality problems the role of (lower semicontinuous) convex functions in optimization problems. We say that a multivalued map is affine if its graph is an affine subspace, i.e., where and are two matrices and . Affine (monotone) maps play in variational inequality problems the role of quadratic (convex) functions in optimization problems. An important class of affine variational inequality problems is the class of linear complementarity problems. A slightly more general form of these problems is as follows: Linear (monotone) complementarity problems encompass quadratic (convex) optimization problems and Nash bimatrix equilibrium problems. They can be solved in a finite number of steps by an algorithm due to Lemke [6] which uses, as the simplex algorithm, pivotal rules. For linear complementarity problems see the text book by Cottle, Pang and Stone [3] and for quadratic programming see the seminal paper [5] of Frank and Wolfe.
The purpose of this paper is the analysis of monotone affine maps. In Section 2 we characterize monotone and maximal monotone affine subsets of . Section 3 analyses affine monotone subspaces under linear transformations. Such transformations occur along the completion of Lemke’s algorithm. In Section 4 we provide an algorithm that builds a maximal monotone affine extension of a monotone affine subspace.
Section snippets
Characterization of affine monotone subsets
Without loss of generality, we assume that the representation given in (2) is not redundant. In other terms, the matrix has rank . Also, we assume that is nonempty and not reduced to a singleton. Thus .
We start with a characterization of affine monotone (maps) sets in terms of matrices and .
Theorem 2.1 is monotone if and only if and the matrix has exactly positive eigenvalues. Proof Let us introduce the matrix and the matrix defined as follows
Monotonicity under linear transformations
We consider the case where is an affine subspace and is the image of under a linear transformation, i.e., there exists a matrix such that . Thus, where . If is monotone, is not necessarily monotone as seen below. Example 3.1 Let us consider the four matrices: Then From Theorem 2.2 one deduces that is maximal monotone. Set and
Affine maximal extension of a monotone affine subspace
A classical result (see for instance [7]) states that given any monotone but not maximal monotone set , there is a (not uniquely defined) maximal set such that . The classical proof uses Zorn’s lemma and thereby is not constructive. Recently, we have proposed in [4] a direct and constructive proof. The problem is that in the particular case where is an affine set, the maximal extension which is obtained is not necessarily affine.
Example 4.1 Let us consider , where and
References (7)
- et al.
Definiteness and semi-definiteness of quadratic forms revisited
Linear Algebra and its Applications
(1984) Manifestations of the Schur complement
Linear Algebra and its Applications
(1974)- et al.
The linear complementarity problem
(1992)