Complex Remez Algorithm Overview

• The complex Remez algorithm is an iterative procedure computing the Chebyshev minimax polynomial for complex functions, extending the classical equioscillation and alternance principles to complex normed spaces.
• It updates extremal nodes, phases, and weights through solving linear systems and maximization steps, ensuring quadratic convergence under mild nondegeneracy conditions.
• The method supports approximations with general bases under linear constraints, with applications in spectral analysis, signal processing, and advanced engineering contexts.

The complex Remez algorithm, also referred to as Tang's generalization of the Remez algorithm, is a rigorous iterative procedure for computing the Chebyshev (minimax) polynomial approximation to a continuous complex-valued function on a compact set in the complex plane. Unlike its classical real-variable variant, the complex Remez algorithm extends the equioscillation principle and alternance theories to accommodate the substantial differences of best uniform approximation in complex normed spaces. This approach enables high-precision computation of Chebyshev polynomials and their uniform norms—crucial both for theoretical spectral analysis and practical engineering contexts—while also supporting uniform approximation by more general basis functions under arbitrary linear constraints (Rubin, 2024, Protasov et al., 2024).

1. Formulation of the Complex Minimax Approximation Problem

Let $K\subset\mathbb{C}$ be a compact set and $C(K)$ the Banach space of continuous complex-valued functions on $K$, equipped with the supremum norm

$\|f\|_K = \max_{z\in K} |f(z)|.$

For fixed $n\geq1$, denote by $\mathcal{P}_n$ the affine space of monic polynomials of degree $n$: $$\mathcal{P}_n = \{P(z) = z^n + a_{n-1}z^{n-1} + \cdots + a_0\}.$$ The Chebyshev (minimax) polynomial $T_n^K$ solves

$T_n^K = \arg\min_{P\in\mathcal{P}_n} \|P\|_K.$

For general approximation of a target $f\in C(K)$ by a real-linear subspace $V\subset C(K)$ of dimension $n$, the objective is

$$\varphi^* = \arg\min_{\varphi\in V} \|f - \varphi\|_K.$$

2. Duality and the Complex Equioscillation Principle

Standard Hahn-Banach duality yields a variational characterization: $$\min_{\varphi\in V}\|f-\varphi\|_K = \max_{L\in V^\perp,\,\|L\|\leq1} |L(f)|,$$ where $V^\perp$ is the annihilator of $V$ in $C(K)^*$. The Remez–Zuhovickii dual form restricts to functionals of the type

$$L(f) = \sum_{j=0}^n r_j\,\Re\{e^{-i\alpha_j} f(z_j)\}$$

subject to

$$z_j\in K,\,\, \alpha_j\in[0,2\pi),\,\, r_j\ge 0,\,\, \sum_{j=0}^n r_j = 1,$$

and

$$L(\varphi_k) = \sum_{j=0}^n r_j\,\Re\{e^{-i\alpha_j}\varphi_k(z_j)\} = 0, \quad k=0,\ldots,n-1.$$

This leads to the "phase alternation" condition, a generalization of alternance: there must exist $n+1$ nodes $(z_j,\alpha_j)$ so that

$$\Re\{e^{-i\alpha_j} E(z_j)\} = \|E\|_K, \quad E(z) = f(z)-\varphi(z).$$

In the complex setting, the strict sign alternance of the real case is replaced by alternation in phase, augmenting the classical alternation theorem to the unit circle in the complex plane (Rubin, 2024).

3. Algorithmic Description and Iterative Update

The complex Remez algorithm proceeds iteratively, maintaining a tuple of extremal points $t_j^{(\nu)}$, phases $\alpha_j^{(\nu)}$, and weights $r_j^{(\nu)}$ satisfying admissibility and annihilation constraints.

Core iteration (cf. (Rubin, 2024)):

• Linear system solution: Solve $A(t^{(\nu)}, \alpha^{(\nu)}) r^{(\nu)} = e_1$, where $A(t,\alpha)$ encodes the real parts of the basis at the current nodes with phase shifts. Adjust phases $\alpha_j \mapsto \alpha_j + \pi$ to ensure $r_j \ge 0$.
• Update of the approximant: Solve

$$A(t^{(\nu)},\alpha^{(\nu)})^T \begin{pmatrix} h^{(\nu)} \ \lambda^{(\nu)} \end{pmatrix} = c_f\bigl(t^{(\nu)},\alpha^{(\nu)}\bigr),$$

returning the trial approximant $\varphi^{(\nu)}(z)$.

• Supremum error calculation: Compute $\|f-\varphi^{(\nu)}\|_K$ versus $h^{(\nu)}$ and assess convergence via the relative error gap.
• Node and phase update: If not converged, find new extremal point $x\in K$ and associated phase $\vartheta$, then update one node-entry triple based on a direction computed by an additional linear solve.
• Iteration cost: Each step demands two solves of $(n+1)\times(n+1)$ real linear systems and one global maximization of the error (Rubin, 2024).

A broader generalization supports approximation by non-polynomial bases (e.g., complex exponentials) under arbitrary linear constraints, where up to $2n+1$ alternance points in $\mathbb{R}^{2n}$ are maintained, updating via convex geometry rules (Protasov et al., 2024).

4. Convergence Properties and Complexity

Under mild nondegeneracy conditions (eventually all $r_j^{(\nu)}>0$), Tang's convergence theorem shows that the lower bound functionals $h^{(\nu)}$ increase monotonically to the true minimax error, while the actual error $\|f-\varphi^{(\nu)}\|_K$ decreases. Locally, the gap decays quadratically: $$\|f-\varphi^{(\nu)}\|_K - h^{(\nu)} = \mathcal{O}((\|f-\varphi^{(\nu-1)}\|_K - h^{(\nu-1)})^2),$$ once the iterates are sufficiently close (Rubin, 2024). For generalized bases and constraints, the gap $B_k-b_k$ (see (Protasov et al., 2024)) may be shown to decay linearly under a uniform barycentric weight bound $\mu>0$.

Each iteration is computationally dominated by two linear solves and a global maximization of the error functional. The practical cost per iteration is $O(n^3 + n^2M)$ for a grid of size $M$, with low-rank matrix updates reducing the per-iteration solve cost to $O(n^2)$ when efficiently implemented. For non-polynomial bases and constraints, the dimension of the relevant real systems scales accordingly (Rubin, 2024, Protasov et al., 2024).

5. Numerical Phenomena, Experimental Results, and Applications

Recent studies employing the complex Remez algorithm have achieved computation of Chebyshev polynomials for degrees as high as $n\sim 100$ on intricate planar sets, marking a significant extension beyond the regimes accessible by other methods. Several notable numerical phenomena arise (Rubin, 2024):

• Widom factors: For a wide family of compact sets (regular polygons, hypocycloids, circular lunes, lemniscates), Widom factors

$W_n(K) = \frac{\|T_n^K\|_K}{\mathrm{Cap}(K)^n}$

tend algebraically toward 1, often at rate $O(n^{-1})$.

• Faber–Chebyshev convergence: On equipotential curves $K_r$ defined by a conformal map $\Phi$, the minimax polynomial approaches the Faber polynomial $F_n^K$, i.e.,

$$\|T_n^{K_r} - F_n^K\|_\infty \to 0 \quad \text{as} \; r \to \infty.$$

• Zero distributions: The zeros of $T_n^K$ cluster along analytic arcs and converge in the weak-star sense to equilibrium measures, in patterns mirroring those for Faber polynomials. Outward cusps, corners, or intersections act as attractors for zeros.

The flexibility of the complex Remez framework, accommodating arbitrary bases under linear constraints, extends its utility to signal processing, linear ordinary differential equations, and approximation by systems such as complex exponentials or Gaussian functions (Protasov et al., 2024).

6. Stability and Implementation Considerations

Ill-conditioning is a central challenge: the system matrices can become numerically unstable if extremal points coalesce or if the basis functions vary wildly in scale. Remedies include high-precision arithmetic, QR-factorization, scaling, or restriction of the domain for basis functions with rapidly decaying magnitudes (e.g., negative real exponential components). Exploiting set symmetries further reduces computational requirements (Rubin, 2024, Protasov et al., 2024).

For the more general algorithmic setting involving constraints, regularization of nearly degenerate convex hulls is advised: switching the update point to maximize minimal distances in the moment space preserves convergence stability. All linear solves are implemented via real arithmetic to sidestep issues with complex phases (Protasov et al., 2024).

7. Context and Theoretical Implications

The complex Remez algorithm provides a concrete computational tool unifying several classical threads—best uniform polynomial approximation, phase- alternation, convex geometry—within a single iterative framework applicable to complex analysis and approximation theory. The observed Faber–Chebyshev convergence and monotonicity of Widom factors provide experimental support for deeper structural relationships between Chebyshev and Faber polynomials of compact sets. The extension to exponentials and general constraints opens new avenues in constrained and nonstandard approximation, with applications across analysis, dynamical systems, and signal processing (Rubin, 2024, Protasov et al., 2024).