Fundamentality of ridge functions (Q1318241)

For \(d,n\in \mathbb{N}\) denote by \(\Omega\) the set of all real \(d\times n\) matrices and let \(M(\Omega)= \text{span} (g(A.x)\): \(A\in\Omega\), \(g\in C(\mathbb{R}^ d, \mathbb{R})\}\), where \(C(\mathbb{R}^ d, \mathbb{R})\) denotes the space of all continuous functions \(g: \mathbb{R}^ d\to \mathbb{R}\) For \(d=1\) the functions in \(M(\Omega)\) are called ridge functions and they are relevant in problems of tomography, neural networks, approximation theory etc. The main question treated in this paper is the density of \(M(\Omega)\) in \(C(\mathbb{R}^ n, \mathbb{R})\), with respect to the topology of uniform convergence of compacta. For \(A\in\Omega\) let \(L(A)\) be the span of the \(d\) rows of \(A\) and let \(L(\Omega)= \bigcup\{ L(A)\): \(A\in\Omega\}\). Let \(H_ n^ k\) denote the set of homogeneous polynomials \(p\) of \(n\) variables \[ -p(s)= \sum_{| m|=k} c_ m s^ m, \qquad m=(m_ 1,\dots, m_ n)\in \mathbb{Z}^ n, \] \(| m|= m_ 1+\dots +m_ n\), \(s^ m= s_ 1^{m_ 1}\dots s_ n^{m_ n}\), and let \(H^ n= \bigcup_{n=0^ \infty} H_ k^ n\). Then \(M(\Omega)\) is dense in \(C(\mathbb{R}^ n, \mathbb{R})\) iff the only polynomial in \(H^ n\) vanishing identically on \(L(\Omega)\) is the zero polynomial (Theorem 2.1). This is equivalent to the assertion that \(M(\Omega)\) explicitly contains the polynomials. If \(k\in \mathbb{N}\) is fixed, then \[ M_ k(\Omega)= \Biggl\{ \sum_{i=1}^ k a_ i g(A_ i x):\;a_ i\in \mathbb{R},\;A_ i\in \Omega,\;i=1,\dots, k,\;g\in C(\mathbb{R}^ d, \mathbb{R}) \Biggr\} \] is not dense in \(C(\mathbb{R}^ d, \mathbb{R})\) (Theorem 5.1). Theorems 2.1 and 4.1 in this paper are obtained (in the case \(d=1\)) by \textit{B. A. Vostrecov} and \textit{M. A. Krejnes}, Dokl. Akad. Nauk SSSR 140, 1237-1240 (1961; Zbl 0106.273) and ibid. 144, 1212-1214 (1962; Zbl 0196.439).