Linear and Nonlinear Methods of Relief Approximation

In this article we compare the effectiveness of free (nonlinear) relief approximation, equidistant relief approximation, and polynomial approximation $\mathscr R^{\mathrm{fr}}_N[f]$, $\mathscr R^{\mathrm{eq}}_N[f]$, $\mathscr E_N[f]$ of an individual function $f(\mathbf{x})$ in the metric $\mathscr L^2(\mathbb B^2)$, where $\mathbb B^2$ is the unit ball $|\mathbf{x}|\le1$ in the plane $\mathbb R^2$. The notation we use is the following
\begin{gather*} \mathscr R^{\mathrm{fr}}_N[f] :=\inf_{R\in\mathscr W^{\mathrm{fr}}_N}\|f-R\|, \quad \mathscr R^{\mathrm{eq}}_N[f]:=\min_{R\in\mathscr W^{\mathrm{eq}}_N}\|f-R\|, \\ \mathscr E_N[f]:=\min_{P\in\mathscr{P}^2_{N-1}}\|f-P\|. \end{gather*}
Here $\mathscr W^{\mathrm{fr}}_N$ is the set of all $N$-term linear combinations of functions of the plane wave type
$$ R(\mathbf{x})=\sum_1^N W_j(\mathbf{x}\cdot\boldsymbol\theta_j) $$
with arbitrary profiles $W_j(x)$, $x\in\mathbb R^1$ and transmission directions $\{\boldsymbol\theta_j\}_1^N$; $\mathscr W^{\mathrm{eq}}_N$ is the subset of $\mathscr W^{\mathrm{fr}}_N$ associated with $N$ equidistant directions;
$$ \mathscr{P}^2_{N-1}:=\operatorname{Span}\{x_1^kx_2^l\}_{k+l<N} $$
denotes the subspace of algebraic polynomials of degree less or equal to $N-1$ in two real variables. Obviously, inequalities $\mathscr R^{\mathrm{fr}}_N[f] \le\mathscr R^{\mathrm{eq}}_N[f]\le\mathscr E_N[f]$ hold.
We state the following model problem. What are the functions which satisfy the relation $\mathscr R^{\mathrm{fr}}_N[f]=o(\mathscr R^{\mathrm{eq}}_N[f])$, i.e., where nonlinear approximation $\mathscr R^{\mathrm{fr}}$ is more effective than linear one? This effect have been proved for harmonic functions, namely, for any $\varepsilon>0$ there exists $c_\varepsilon>0$ such that if $\Delta f(\mathbf{x})=0$, $|\mathbf{x}|<1$, $f\in\mathscr L^2(\mathbb B^2)$, then
$$ \mathscr R^{\mathrm{fr}}_N[f] \le c_\varepsilon\big(\mathscr R^{\mathrm{eq}}_N[f]\exp(-N^\varepsilon)+\mathscr R^{\mathrm{eq}}_{N^{2-3\varepsilon}}[f]\big). $$
On the other hand, $\mathscr R^{\mathrm{fr}}_N[f]\ge\frac1c\mathscr R^{\mathrm{eq}}_{N^2}[f]$. Thus $\mathscr R^{\mathrm{fr}}_N[f]$ has an “almost squared effectiveness” of $\mathscr R^{\mathrm{eq}}_N[f]$ for $f=f_{\mathrm{harm}}$. However, this ultra-high order of approximation is obtained via a collaps of wave vectors.
On the other hand, the nonlinearity of $\mathscr R^{\mathrm{fr}}$ which corresponds to the freedom of choice of wave vectors, does not much improve the order of approximation, for instance, for all the radial functions. If $f(\mathbf{x})=f(|\mathbf{x}|)$, then $\mathscr E_{2N}[f]\ge\mathscr R^{\mathrm{eq}}_N[f]\ge\sqrt{\dfrac23}\mathscr E_{2N}(f)$ and $\mathscr R^{\mathrm{fr}}_N[f]\ge\sup\limits_{\varepsilon>0}\sqrt{\dfrac\varepsilon{3(1+\varepsilon)}}\mathscr R^{\mathrm{eq}}_{(1+\varepsilon)N}[f]$.
The technique we use is the Fourier–Chebyshev analysis (which is related to the inverse Radon transform on $\mathbb B^2$) and a duality between the relief approximation problem and the optimization of quadrature formulas in the sense of Kolmogorov–Nikolskii [1] for trigonometric polynomials classes.