Approximation of generalized ridge functions in high dimensions (Q2315030)

Ridge functions are an important tool in the approximation of multivariable functions, closely related to the so-called radial basis functions. A ridge function \(f\) is defined on any (high-)dimensional real space and depends only on a univariate function composed with an inner product \(\langle\cdot,a\rangle\) with a vector \(a\) (same dimension, \(N\) say; compare with radial basis functions which are univariate functions composed with a [Euclidean] norm). In more generality, such as in this paper, the inner product can be replaced by the multiple of \(f\)'s argument \(x\) with a matrix \(A\) (again, same square dimension \(N\times N\)). In this set-up, the paper considers so-called sleeve functions that are defined on a tube (or ``sleeve'') about a manifold. Call this sleeve \(L\). The author presents methods to compute (reconstruct) this sleeve \(L\) by (i) adaptive algorithms ATPE, and by (ii) optimisation methods OGM. Theorems on error bounds for these computations are offered too.