Shape-preserving multivariate polynomial approximation in \(C[-1,1]^m\) (Q1774662)

Let \(f:[-1,1]^m\mapsto\mathbb R\), and a multi-index of integers \((s_1,\dots,s_m)\) where \(s_i\geq-1\), \(1\leq i\leq m\). For each \(1\leq i\leq m\), let \(\{x^{(i)}_j\}_{j=1}^{s_i+2}\) be distinct points in \([-1,1]\), and define the \(m\)-folded divided difference of \(f\) with respect to the above array of points as \[ [x^{(1)}_1,\dots,x^{(1)}_{s_1+2};[x^{(2)}_1,\dots,x^{(2)}_{s_2+2};\dots [x^{(m)}_1,\dots,x^{(m)}_{s_m+2};f]\dots]], \] where \([x^{(i)}_1,\dots,x^{(i)}_{s_i+2};\cdot]\) is the univariate divided difference with respect to the \(i\)th variable. The authors call \(f\) convex of order \((s_1,\dots,s_m)\), if the \(m\)-folded divided difference is nonnegative for all choices of arrays as above. Let \(\varphi(t):=\sqrt{1-t^2}\), and denote \( \Delta^2_{h_1\varphi(x_1),\dots,h_m\varphi(x_m)}f(x_1,\dots,x_m):= \sum_{k=0}^2(-1)^k\binom2kf(x_1+(1-k)h_1\varphi(x_1),\dots,x_m+(1-k)h_m \varphi(x_m))\), if \((x_1\pm h_1\varphi(x_1),\dots,x_m\pm h_m\varphi(x_m))\in[-1,1]^m\), and \(=0\) otherwise. Finally, set \[ \begin{multlined} \omega^\varphi_2(f;\delta_1,\dots,\delta_m):=\sup\{| \Delta^2_ {h_1\varphi(x_1),\dots,h_m\varphi(x_m)}f(x_1,\dots,x_m)| :\\ 0\leq h_i\leq\delta_i,1\leq i\leq m,\quad x_1,\dots,x_m\in[-1,1]\}. \end{multlined} \] It is proved that if \(f\in C([-1,1]^m)\), \(m\geq2\), then for any \(m\)-tuple \((n_1,\dots,n_m)\) of natural numbers, there exists a polynomial \(P_{n_1,\dots,n_m}(f)(x_1,\dots,x_m)\) of the degree \(n_i\) in the respective variable \(x_i\), such that \[ \| f-P_{n_1,\dots,n_m}(f)\| \leq C_m\omega^\varphi_2(f;1/n_1,\dots,1/n_m), \] where \(C_m\) is an absolute constant. Moreover if \(f\) is convex of order \((s_1,\dots,s_m)\) for various possibilities of \(m\) tuples \((s_1,\dots,s_m)\), then so are the polynomials. The result extends recent result of the second author for the case \(m=2\).