Chebyshev Polynomial Expansion

Updated 9 March 2026

Chebyshev polynomial expansion is a method that represents functions using a series of Chebyshev polynomials, capitalizing on orthogonality and near-minimax properties.
It leverages recurrence relations and FFT-based techniques for rapid, numerically stable coefficient computation and exponential convergence for analytic functions.
Applications include quantum simulation, signal processing, and numerical linear algebra, offering robust tools for approximation theory and operator analysis.

A Chebyshev polynomial expansion represents a function as a series in the basis of Chebyshev polynomials of the first kind, exploiting their orthogonality, recurrence, and near-minimax properties on compact intervals. This method is powerful for achieving efficient, accurate, and numerically stable approximations of both analytic and limited-regularity functions, and it underpins a wide array of numerical algorithms in approximation theory, computational physics, quantum simulation, signal processing, and machine learning.

1. Formal Definition and Core Properties

A Chebyshev polynomial of the first kind of degree $n$ is defined by $T_n(x) = \cos(n \arccos x)$ for $x\in[-1,1]$ . The Chebyshev expansion of a sufficiently smooth function $f$ on $[-1,1]$ takes the form

$f(x) = \frac{a_0}{2} + \sum_{n=1}^\infty a_n\,T_n(x),$

where the coefficients are

$a_n = \frac{2}{\pi}\int_{-1}^1 f(x) T_n(x) \frac{dx}{\sqrt{1-x^2}}, \quad n \geq 1,$

with $a_0$ defined similarly but with normalization $1/\pi$ for the constant term. The $\{T_n\}$ are orthogonal with respect to the weight $w(x) = (1-x^2)^{-1/2}$ :

$\int_{-1}^1 T_n(x) T_m(x) w(x) dx = \begin{cases} \pi, & n=m=0\ \pi/2, & n=m>0\ 0, & n\neq m \end{cases}$

This basis yields near-minimax polynomial approximation, providing close to minimal uniform error for a given degree $N$ (Akansha, 2024).

The three-term recurrence

$T_0(x) = 1, \quad T_1(x) = x, \quad T_{n+1}(x) = 2x T_n(x) - T_{n-1}(x)$

enables efficient evaluation and underpins stable numerical algorithms such as Clenshaw’s method.

2. Algorithmic Construction and Fast Computation

The Chebyshev expansion coefficients can be computed by numerical quadrature using Gauss–Chebyshev nodes $x_k = \cos\left[\frac{(2k-1)\pi}{2N}\right]$ , for $k=1,\dotsc,N$ , or via FFT-based techniques. For analytic $f$ , the expansion converges exponentially fast, with the error decaying as $O(\rho^{-N})$ for $f$ analytic in a Bernstein ellipse of parameter $\rho>1$ (Wang et al., 2014).

The discrete cosine transform (DCT) and FFT facilitate $O(N \log N)$ computation of coefficients from function samples at Chebyshev nodes, making the approach suited for high-degree expansions and large-scale settings (Wang et al., 2014, Akansha, 2024).

For accurate numerical evaluation, the Clenshaw algorithm leverages the three-term recurrence, computing $\sum_{n=0}^N c_n T_n(x)$ in $O(N)$ time with bounded error propagation. Conversion between Chebyshev and Laurent expansions further allows interval-arithmetic enclosures and avoids ill-conditioning near the endpoints (Aurentz et al., 2024).

3. Approximation Theory and Error Estimates

Chebyshev expansions provide the following approximation properties:

Analytic functions: Exponential convergence of truncation error, governed by the analyticity strip (Bernstein ellipse), i.e., $|f(x) - S_N[f](x)| \lesssim \rho^{-N}$ (Wang et al., 2014).
Functions of bounded variation: The best possible algebraic decay of Chebyshev coefficients for $f^{(k)}$ of bounded variation $V_k$ , given by

$|c_n| \leq \frac{2V_k}{\pi} \left( \frac{b-a}{2} \right)^{k+1} n^{-(k+1)},$

leading to $O(1/N)$ convergence for discontinuous derivatives (Akansha, 2024). The Xiang-type bounds further sharpen constants for limited regularity cases.

Minimax and probabilistic error: For monomials $x^n$ , the truncation error of Chebyshev expansion (degree $m<n$ ) admits both exact expressions and concentration-inequality bounds. Specifically, the supremum norm error relates to the binomial tail probability, with $E(n,m) = 2\,\Pr[X \geq (n+m)/2]$ for $X \sim \text{Binomial}(n,1/2)$ , yielding the explicit decay rate in degree (Saibaba, 2021).

These sharp error formulations inform adaptive choices of polynomial degree and provide guidance for non-smooth applications, particularly in signal processing and machine learning (Akansha, 2024).

4. Specialized Expansions and Recurrences

4.1 Chebyshev–Bessel Expansion

The exponential function admits an explicit Chebyshev expansion:

$e^t = e^d\,e^{\rho s} = e^d\left[I_0(\rho) + 2 \sum_{k=1}^\infty I_k(\rho)\,T_k(s)\right],$

with $s = (2t - a - b)/(b-a)$ , $d = (a + b)/2$ , $\rho = (b - a)/2$ , and $I_k$ the modified Bessel function of the first kind. Truncating at order $N$ with $N \approx \rho + \sqrt{2\rho \log(1/\epsilon)}$ controls the relative error (Hen, 28 Dec 2025).

This expansion is critical for exponential divided differences, underpinning stable $O(qN)$ algorithms for time-evolution kernels and efficient incremental updates when nodes are modified (Hen, 28 Dec 2025).

4.2 Recurrences for Differential Equations

For $f$ satisfying a linear ODE $L f = 0$ with $L$ a polynomial differential operator, the Chebyshev coefficients persistently satisfy a linear recurrence derived via a symbolic morphism and Ore algebra. Fast divide-and-conquer algorithms and FFT-based products enable subcubic computation in large order $k$ (0906.2888).

This symbolic-numeric framework unifies and generalizes classical recurrence-based spectral methods for differential equations, allowing direct manipulation of high-order expansions.

5. Applications in Computational and Physical Sciences

5.1 Quantum Many-Body and Spectral Calculations

Chebyshev expansions of Green’s functions, resolvents, and density of states replace direct diagonalization or inversion by stable sparse matrix-vector multiplications:

Localization properties: The inverse localization length $\kappa(E)$ , as well as the density of states in Hermitian and non-Hermitian random chains, are efficiently computed by Chebyshev recursions (Hatano et al., 2016).
Quantum transport: The Landauer conductance of large-scale systems is evaluated through Chebyshev expansions of tight-binding Green’s functions, using Jackson damping and complex absorbing potentials (CAPs) to manage spectral features and open boundary conditions, with favorable scaling to millions of sites (Castro et al., 2022).
Quantum kernel approximation: Flat Chebyshev approximations of the exponential function enable polynomial-time learning of quantum Hamiltonians from thermal states, with explicit bit complexity and spectral error control, crucial to quantum computational complexity (Wodecki et al., 2024).

5.2 Numerical Linear Algebra and Operator Functions

Chebyshev series underpin matrix-function approximations (e.g., $e^{-A}$ , $\log(A)$ ) in the context of applied numerical linear algebra, offering stability and near-minimax accuracy via truncated expansions.

5.3 Signal Processing and Graph Filtering

Chebyshev polynomial filters efficiently approximate frequency responses and spectral multipliers, with explicit error bounds for non-smooth and bounded-variation functions, crucial in digital signal processing and spectral graph neural networks (Akansha, 2024, Roy et al., 9 Apr 2025).

5.4 Machine Learning and Deep Learning Architectures

Convolutional neural networks are augmented with Chebyshev polynomial expansions to enhance their spectral representation capacity, enabling explicit control over filter regularity and efficient high-frequency feature extraction. Chebyshev-CNNs demonstrate marked improvements over conventional CNNs in medical imaging and can be generalized for graph-based models (Roy et al., 9 Apr 2025).

6. Advanced Topics and Extensions

6.1 Chebyshev Expansions for Orthogonal Polynomial Families

Chebyshev expansions generalize to other polynomial systems, including explicit connection coefficients (closed-form and holonomic recurrences) for Zernike polynomials, with computational methods scaling linearly in the degree for large expansions (Mathar, 23 Sep 2025).

6.2 Validated Numerics and Laurent-Horner Evaluation

Validated interval enclosures for Chebyshev expansions employ transformation to the Laurent basis via the Joukowski map, allowing interval Horner’s method to produce guaranteed enclosures in $O(N)$ time, with resilience to endpoint ill-conditioning—a significant advantage over eigenvalue-based interval evaluation (Aurentz et al., 2024).

6.3 Inverse Problems and Integral Transforms

Paired Chebyshev expansions are utilized to approach inverse problems, such as recovery from truncated Hilbert transforms (THT), leveraging SVD-like properties and iterative extrapolation schemes with robust error control and convergence, even under partial data and noise (You, 2020).

7. Limitations and Best-Practice Recommendations

For low-degree expansions ( $N\lesssim 50$ ), direct trigonometric or barycentric methods may be more efficient than recurrence-based or Laurent methods.
Eigenvalue-based interval evaluation methods suffer near the endpoints due to diverging condition numbers; Laurent-Horner evaluation remedies this (Aurentz et al., 2024).
For non-analytic targets, error decay is limited to $O(1/N)$ , but the optimal constants provided by Majidian- and Xiang-type bounds ensure that Chebyshev expansions remain near-optimal even for non-smooth cases (Akansha, 2024).
Algorithmic design often incorporates Jackson damping to suppress Gibbs oscillations arising from truncation.
In stochastic estimation of traces (e.g., large quantum Hamiltonians or spectral densities), error scales as $1/\sqrt{N R}$ , motivating careful balancing of expansion order and sampling (Castro et al., 2022).

Chebyshev polynomial expansions offer a theoretically grounded, computationally efficient, and broadly applicable toolset for high-fidelity function approximation and operator analysis, with deep integration across mathematics, physical sciences, and data-driven computational paradigms. The full scope and recent algorithmic advances—including incremental update schemes (Hen, 28 Dec 2025), bit-complexity-optimal flat approximations (Wodecki et al., 2024), and high-dimensional validated evaluation (Aurentz et al., 2024)—demonstrate the enduring and evolving impact of this foundational construct.