Saddle points and overdetermined complex equations (Q2266193)

It is known that the best uniform norm solution of overdetermined complex valued systems of equations satisfying the Haar condition for matrices is also a best weighted \(\ell_ p\) norm solution for each \(p\geq 1\), for some weight vector depending on p. This paper presents an alternative proof of this result which is valid for arbitrary matrices A. The proof relies on the fundamental theorem of game theory. It is shown that a saddle point \((z^*,\lambda^*)\) of a certain function gives a uniform norm solution, \(z^*\), of \(Az=b\) and a weight vector \(\lambda^*\) of the equivalent weighted \(\ell_ p\) norm problem. With appropriate qualifications concerning the weights, it follows that the worst (i.e., largest) possible weighted least \(\ell_ p\) norm error is also the best (i.e., least) possible Chebyshev error. For \(p=2\), it is shown that the weight vector \(\lambda^*\) solves a nonlinear optimization problem which can be posed without reference to solution vectors of \(Az=b\). In other words, the problem of finding the best uniform norm solution of \(Az=b\), when stated as a convex optimization problem, has a convex dual which for \(p=2\) can be posed independently of the primal variables z. The dual variables are the weights \(\lambda\).