Generalised Chebyshev-like Forms in Approximation

• Generalised Chebyshev-like forms are analytic, algebraic, and optimization constructs that extend classical Chebyshev approximations to broader function classes via rational and non-polynomial bases.
• They utilize minimax formulation, quasi-convex and pseudo-convex properties, along with equioscillation theorems to ensure optimal uniform error control.
• Applications span numerical approximation, spectral analysis, signal processing, and PDE solutions, offering enhanced accuracy and computational efficiency over classical methods.

A generalised Chebyshev-like form refers to a class of analytic, algebraic, and optimization constructs that extend the classical Chebyshev polynomial and rational forms. These generalisations are characterized by their minimax, equioscillation, structural, and convexity properties, and appear in approximation theory, orthogonal polynomial systems, optimization, and spectral analysis. The unifying feature is the adaptation of Chebyshev-type “best” approximation (minimax, equioscillation, or extremal property) from polynomials to a broader class of functions, often involving ratios of linear forms over general basis sets, multivariate extensions, endpoint or measure modifications, or algebraic/combinatorial identities.

1. Analytic and Algebraic Definition of Generalised Chebyshev-like Forms

Let $\{\varphi_0,\ldots,\varphi_n\}$ and $\{\psi_0,\ldots,\psi_m\}$ be finite families of continuous, real-valued basis functions on a compact interval $I\subset\mathbb R$. A generalised Chebyshev-like form is any function

$$R(x) = \frac{\sum_{i=0}^n a_i\varphi_i(x)}{\sum_{j=0}^m b_j\psi_j(x)}$$

with coefficients $(a_0,\dots,a_n)\in\mathbb R^{n+1}$, $(b_0,\dots,b_m)\in\mathbb R^{m+1}$, and normalization constraint $\sum_j b_j\psi_j(x)\ge 1$ for all $x\in I$ (or any equivalent constraint to ensure positivity and avoid denominator sign changes). This generalizes classical rational (Chebyshev) forms, which use monomials for $\varphi_j$ and $\psi_k$, but admits any continuous basis, including non-polynomial functions such as $\exp$, $\sin$, or user-selected families (Millán et al., 2020).

For multivariate settings, the generalised Chebyshev-like form can be written

$$R(x) = \frac{\sum_{j=1}^m a_j\varphi_j(x)}{\sum_{k=0}^m b_k\psi_k(x)}$$

with $x\in\mathbb R^\ell$ and the coefficients indexed correspondingly (Millán et al., 2021).

Other generalisations include endpoint-mass perturbations (Koornwinder-type), two-parameter cosines (e.g., $(\beta,\gamma)$-Chebyshev functions), or composition over Julia sets (dynamical Chebyshev-like forms).

2. Minimax Formulation and Convexity Structure

Given a continuous target $f:I\to\mathbb R$, the best generalised rational approximation is the solution to the uniform error minimax problem: $$\Psi(a,b) := \sup_{x\in I} |f(x) - R(x)|, \qquad \min_{(a,b)\in C}\Psi(a,b)$$ where $$C=\{(a,b): \sum_j b_j\psi_j(x)\ge 1,\;\forall x\in I\}$$ is a closed convex set in $\mathbb R^{n+m+2}$ (Millán et al., 2020). The set of approximants includes the classical Chebyshev polynomial case ($b_0=1$, $m=0$) as a special instance.

For multivariate or grid-based approximation, one minimizes over $(a,b)$ to solve

$$E^* = \min_{a,b} \max_{i=1,\dots,N} \left| f(x_i) - R(x_i) \right|$$

where $x_i$ form a finite grid in $\mathbb R^\ell$, and the problem keeps the same quasi-convexity properties (Millán et al., 2021).

Convexity properties:

• For each fixed $x$, the map $$(a,b)\mapsto |f(x) - (a\cdot\varphi(x))/(b\cdot\psi(x))|$$ is quasi-convex. The supremum over $x$ of such quasi-convex functions is quasi-convex, so $\Psi(a,b)$ is quasi-convex on $C$. (Millán et al., 2020, Millán et al., 2021)
• $\Psi$ is not merely quasi-convex; it is pseudo-convex in the sense of Penot–Quang: any stationary point $(a^*,b^*)$ with $0\in\partial\Psi(a^*,b^*)$ is a global minimizer (Millán et al., 2020).

3. Equioscillation and Alternation Theorem

The alternation (equioscillation) property characterizes optimality of Chebyshev-type approximations in both classical and generalised settings. In the generalised rational context:

• A feasible $(a^*,b^*)\in C$ is optimal if and only if there exist at least $n+m+2$ points $x_0<\dots<x_{n+m+1}\in I$ such that

$$f(x_k) - R^*(x_k) = (-1)^kE^*,\qquad k=0,\dots,n+m+1$$

where $E^*=\Psi(a^*,b^*)$ (Millán et al., 2020, Millán et al., 2021). Equivalently, $0\in\partial\Psi(a^*,b^*)$ precisely when such an equioscillation set exists.

This generalises both the Chebyshev alternation theorem and the foundational result for rational uniform approximation.

4. Projection-Based Optimization Algorithms

The pseudo-convexity and strict structural properties allow efficient optimization algorithms:

• Algorithm F (Díaz Millán–Sukhorukova–Ugon) iterates projections against the Clarke subdifferential of the uniform error, using a feasible-direction linesearch (Line Search F) and projection onto the intersection of convex feasibility sets and supporting halfspaces.
• Each iteration involves subgradient evaluation at active points (where the approximation error attains the extremum) and projections onto $C$ and possibly halfspaces.
• Linesearch F always terminates unless already optimal; the overall sequence is Fejér monotone with respect to the solution set.
• Variant 2 (direct projection onto $C\cap H$) contracts fastest in practice and is numerically superior.
• Convergence is linear under non-degeneracy (Millán et al., 2020).

Algorithmic workflow (summarized):

Stage	Operation	Features
Subgradient eval	Compute $\partial\Psi(a,b)$ at active points	O((n+m)
Projection	Project onto $C$ and $C\cap H$	Often normalization + box
Linesearch	Quasi-Armijo; backtracking; ensures sufficient descent	Pseudo-monotone operator
Termination	Stop if stationary; otherwise project and iterate	Fejér monotonicity

For grid problems, bisection on $t$ with linear programming feasibility tests at each step is favored (Millán et al., 2021, Peiris et al., 2020).

5. Multivariate and Extended Forms

Generalised Chebyshev-like forms extend to higher dimensions and broader algebraic situations:

• Multivariate Generalised Forms: For $f:X\rightarrow\mathbb R$, $X\subset\mathbb R^\ell$ on a finite grid, one defines numerator and denominator as sums over any basis $\{\varphi_j\},\{\psi_k\}$ and imposes $Q(x_i)>0$ for all grid points. Quasiconvexity is retained, and the minimax error is solved by sequential LPs via bisection (Millán et al., 2021).
• Generalised Polynomial Systems: Endpoint-mass perturbations (Koornwinder/AlQudah), $(\beta,\gamma)$-Chebyshev families, non-autonomous Julia sets, and hypergeometric invariants provide further generalisations (AlQudah, 2015, AlQudah, 2015, Bedratyuk et al., 2019, Dressler et al., 2024).
• Diagonally-determined domains: For $\Omega\subset\mathbb R^d$ with a “diagonal” structure, the best multivariate Chebyshev polynomial coincides (up to affine rescaling) with the univariate Chebyshev solution on the corresponding diagonal, independent of ambient dimension (Dressler et al., 2024).
• Algebraic and Combinatorial Extensions: Generalised Chebyshev-like identities arise via derivations and kernel invariants (Bedratyuk et al., 2019), generating functions (Szabłowski, 2017, Reynolds et al., 2022), and in the context of modular forms and curve-counting (Andrews et al., 2010).

6. Structural and Functional Properties

These generalised forms share several analytic, algebraic, and structural hallmarks with classical Chebyshev theory:

• Recurrence relations and closed forms: Three-term recurrences, explicit binomial or hypergeometric expansions, and continued-fraction representations extend in the generalised classes (Bedratyuk et al., 2019, AlQudah, 2015, Marchi et al., 2023).
• Orthogonality and weight modifications: Orthogonality with respect to modified measures, including trimmed intervals and endpoint masses, remains a key feature (AlQudah, 2015, AlQudah, 2015, Marchi et al., 2023).
• Lebesgue constant and interpolation stability: Generalised nodes (e.g., $(\beta,\gamma)$-Chebyshev points) control growth of the Lebesgue constants, interpolation stability, and can be selected to interpolate between classical Chebyshev and Lobatto distributions (Marchi et al., 2021, Marchi et al., 2023).
• Generalised inequalities and functional analysis: Chebyshev-type inequalities extend to quasi-linear and non-additive integrals, encapsulating and generalising nonconstructive bounds for broad function systems (Boczek et al., 2019).

7. Applications and Computational Aspects

Generalised Chebyshev-like forms provide practical and conceptual advances in several domains:

• Numerical approximation: Achieve smaller minimax errors versus polynomial-only schemes for rapidly varying or nonsmooth functions, especially via rational or non-polynomial bases (Millán et al., 2020, Millán et al., 2021).
• Matrix and operator theory: Companion-matrix and generalized complex-unit constructions produce spectral and combinatorial results (Babusci et al., 2012).
• Signal processing and PDEs: Bivariate/double Chebyshev series, with gamma-function coefficients, serve for explicit integral transforms and PDE solutions (Reynolds et al., 2022, Szabłowski, 2017).
• Enumerative geometry and modular forms: Chebyshev-like expansions encode arithmetic and geometric generating functions, realized as quasi-modular forms (Andrews et al., 2010).
• Optimization: Unified convex programming frameworks cover both rational and more general quasilinear minimax approximation problems, enabling algorithmic deployment using LPs, SDP hierarchies, and splitting/projection methods (Peiris et al., 2020, Dressler et al., 2024).
• Symbolic identities and numerical stability: Multivariate and trigonometric Chebyshev-like identities support numerically robust algorithms in settings where cancellation-unstable binomial sums arise (e.g., atomic form factor evaluation) (Sen et al., 2020).

Generalised Chebyshev-like forms thus play a central role in the contemporary landscape of approximation theory and computational mathematics, merging deep algebraic structures with convex optimization methodologies and uniform error control in high-dimensional or structured domains.