e-mail: finzel@mi.uni-erlangen.de

Martina Finzel-Hoffmann:

Discrete Piecewise Affine Approximation
[Diskrete stückweise affine Approximation]

June 1997

Abstract. The paper investigates classical problems in approximation, the metric projection P of the normed R^n onto a subset G of R^n, corresponding selections, and their construction when G is a linear subspace or a polyhedral subset of R^n. We consider as well l_p(n)-norms, 1 <= p <= oo, as well as polyhedral norms of R^n, and arrange more recent results about the in general set-valued mapping P. Moreover, the metric projection is piecewise polyhedral and globally Lipschitz continuous with respect to the Hausdorff metric.
In l_p(n), 1 < p < oo, the metric projection is piecewise affine and uniformly Lipschitz continuous with respect to p, i.e., for fixed G the Lipschitz constant does not depend on p.

If the projection is set-valued we ask for particular best approximants and continuous selections. The existence of continuous selections is obvious, however, their constructions need detailed investigations.
Here we study continuous extremal point selections which turn out to be piecewise affine and globally Lipschitz continuous in this polyhedral setting, and also classical selections. These are the strict best approximation which we consider in the more general situation of polyhedral subsets G of l_oo(n), and the least norm selection. They are all piecewise affine and globally Lipschitz continuous, and, in addition, if G is a linear subspace, the strict best approximation is quasi-linear with respect to G.
An elementary example in l_oo(3) demonstrates that in general a selection does not exist which unifies the three properties continuity, extremal point selection, and quasi linearity. Nevertheless, the l_1(n) is an exception where a Chebyshev perturbation always allows to construct a globally Lipschitz continuous and piecewise affine selection which also satisfies the three properties above.
The strict best approximation picks up this idea to construct continuous selections by imitating the behavior of the metric projection onto unicity subsets G. In l_oo(n) one cannot apply perturbations, but we can carry out finitely many iterations, each of them based on a description of the metric projection in terms of Plücker-Graßmann coordinates, a classification of these coordinates, and elementary vectors. Elementary vectors of a subspace are vectors therein which have minimal support. If G is a linear subspace the tuples of the classification are exactly the supports of the elementary vectors of the orthogonal complement of G.

Hence there is a tool explicitly available which makes an excellent match for the approximation problem in l_1(n) and l_oo(n), and which also provides numerical approaches to selections. Central are representations of the strict best approximation in l_oo(n) and of continuous extremal point selections in l_1(n) in terms of Plücker-Graßmann coordinates. While the first case is iteratively described, the description of the latter one is based on the interpolatory behavior of these selections. Corresponding algorithms are formulated.
There is an elegant method to determine the Plücker-Graßmann coordinates by a single application of the Gauß process followed by a recursive and parallel process. From that the elementary vectors are given by a generalized vector product which almost does not require any further calculation. If one is willing to determine the Plücker-Graßmann coordinates of the subspace under consideration then the algorithms of the corresponding selections take only few calculations what makes them effective under repeated projections onto the same subspace.

References

M. F.-H. 28. 8. 97