Some improved bounds in sampling discretization of integral norms (Q6040816): Difference between revisions

In many areas of approximation theory and applications such as sampling, discretisation of integrals and especially of Euclidean and \(L^1\)-norms that are defined by integrals is of the essence. Sums instead of integrals are used for that. A key question is the distribution and \textit{the number of} the evaluation points that defined the finite sums which are employed to approximate the integral. In other words, one wants to know which distributions are optimal in some sense, minimising the number of points and maximising the accuracy by enlarging the lower bound and minimising the upper bound on the sum that discretises the integral. In this paper, two very interesting results are presented which give bounds on the accuracy of these discretisations. They demand that the subspaces on which the upper and lower bounds to the mentioned finite sums by the actual integrals (\(L^p\)-norms) are realised to satisfy a so-called Nikol'skii inequality on the supremums-norm of the elements (functions) by Euclidean norms. The results give an upper bound on the number of sampling points that are needed to provide multiplying constants to the said upper and lower bounds to be \(1\mp\epsilon\) (theoretically of course \(\epsilon=0\) would be good, but this is beyond reach). Calling the number of sampling points \(m\), the upper bound depends by a factor of \(n\log n\) on the dimensionof the subspaces and on \(1/\epsilon^2\) and on the constant \(K\) in the Nikol'skii inequality on the supremums-norm of the elements. A second theorem is offered, where the upper bound to \(m\) depends on \(p\) is formed, where not only the Euclidean integral norm is bounded above and below, but general \(p\)-norms with \(p\in(1,2)\). The upper bounds on \(m\) are slightly larger, depend on \(p\) of course and have terms of extra factors \((\log\log (Kn)-\log\epsilon)^2\log(1/\epsilon)\).