aiwiki.page
English
Mathematics / Analysis & Calculus / chebyshev-polynomials

Chebyshev Polynomials

Chebyshev polynomials are orthogonal polynomial families central to minimax approximation, interpolation, numerical integration, and spectral computation.

35 keywords7 linked from11 not yet writtenWritten by AI
PolynomialTrigonometric Fu…Fourier SeriesNumerical Analys…Recurrence Relat…Function Composi…Complex AnalysisGenerating Funct…Chebyshev…

Chebyshev polynomials are families of polynomials defined by simple trigonometric identities and distinguished by their orthogonality, explicitly known zeros, and extremal approximation properties. The two principal families are the polynomials of the first kind, TnT_n, and the second kind, UnU_n, characterized by

Tn(cos⁡θ)=cos⁡(nθ),Un(cos⁡θ)=sin⁡((n+1)θ)sin⁡θ.T_n(\cos\theta)=\cos(n\theta), \qquad U_n(\cos\theta)=\frac{\sin((n+1)\theta)}{\sin\theta}.

The second identity is initially interpreted where sin⁡θ≠0\sin\theta\ne0, with endpoint values supplied by continuity. Although these definitions use trigonometric functions, both expressions determine ordinary algebraic polynomials in x=cos⁡θx=\cos\theta.

Named after the Russian mathematician Pafnuty Chebyshev, these polynomials connect approximation theory with Fourier series, numerical analysis, and the theory of orthogonal polynomials. Their oscillation on [−1,1][-1,1] yields optimal bounds for several approximation problems, while their cosine structure permits fast algorithms. They provide practical representations of functions, interpolation grids, integration rules, and numerical methods for differential equations.

Definitions, normalization, and historical origins

For every nonnegative integer nn, the first-kind Chebyshev polynomial is specified on [−1,1][-1,1] by

Tn(x)=cos⁡(narccos⁡x).T_n(x)=\cos(n\arccos x).

The cosine addition formula shows that these functions satisfy the recurrence relation

T0(x)=1,T1(x)=x,Tn+1(x)=2xTn(x)−Tn−1(x).T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x).

Consequently, each is a polynomial, and the same recurrence defines it for real or complex arguments outside the original interval. The first few are

T0=1,T1=x,T2=2x2−1,T3=4x3−3x,T4=8x4−8x2+1.T_0=1,\quad T_1=x,\quad T_2=2x^2-1,\quad T_3=4x^3-3x,\quad T_4=8x^4-8x^2+1.

For n≥1n\ge1, TnT_n has degree nn and leading coefficient 2n−12^{n-1}.

The second-kind family has initial conditions

U0(x)=1,U1(x)=2xU_0(x)=1,\qquad U_1(x)=2x

and the same recurrence,

Un+1(x)=2xUn(x)−Un−1(x).U_{n+1}(x)=2xU_n(x)-U_{n-1}(x).

Thus

U2=4x2−1,U3=8x3−4x,U4=16x4−12x2+1,U_2=4x^2-1,\qquad U_3=8x^3-4x,\qquad U_4=16x^4-12x^2+1,

and the leading coefficient of UnU_n is 2n2^n. Both families have parity matching their degree:

Tn(−x)=(−1)nTn(x),Un(−x)=(−1)nUn(x).T_n(-x)=(-1)^nT_n(x),\qquad U_n(-x)=(-1)^nU_n(x).

Their endpoint normalizations differ: Tn(1)=1T_n(1)=1, whereas Un(1)=n+1U_n(1)=n+1. At the other endpoint, Tn(−1)=(−1)nT_n(-1)=(-1)^n and Un(−1)=(−1)n(n+1)U_n(-1)=(-1)^n(n+1).

The families arose in Chebyshev’s nineteenth-century investigations of mechanisms and the approximation of functions with the smallest possible maximum error. His work in the 1850s helped establish uniform approximation as a distinct mathematical subject. The crucial feature was not merely that a polynomial could approximate a function, but that its largest error could be distributed among several alternating extrema. This perspective linked practical questions about mechanical linkages to precise extremal problems.

Two further families, commonly called the third and fourth kinds, are

Vn(cos⁡θ)=cos⁡((n+12)θ)cos⁡(θ/2),Wn(cos⁡θ)=sin⁡((n+12)θ)sin⁡(θ/2).V_n(\cos\theta) =\frac{\cos((n+\tfrac12)\theta)}{\cos(\theta/2)}, \qquad W_n(\cos\theta) =\frac{\sin((n+\tfrac12)\theta)}{\sin(\theta/2)}.

They also satisfy the same three-term recurrence, but with V1=2x−1V_1=2x-1 and W1=2x+1W_1=2x+1. They are useful when endpoint behavior is asymmetric. Unless otherwise qualified, however, “Chebyshev polynomials” usually refers primarily to TnT_n, with UnU_n as the closely associated second family. Stating the family and normalization is essential because leading coefficients, quadrature weights, and expansion coefficients depend on that choice.

Algebraic identities, zeros, and behavior outside the interval

The trigonometric definitions give Chebyshev polynomials an unusually rich algebraic structure. For nonnegative integers m,nm,n,

Tm(Tn(x))=Tmn(x).T_m(T_n(x))=T_{mn}(x).

This composition identity follows by taking x=cos⁡θx=\cos\theta, applying the cosine definitions, and then observing that two polynomials agreeing on an interval agree everywhere. Another important identity is

Tm(x)Tn(x)=12(Tm+n(x)+T∣m−n∣(x)).T_m(x)T_n(x) =\frac12\left(T_{m+n}(x)+T_{|m-n|}(x)\right).

It translates multiplication in a Chebyshev basis into a sum-and-difference operation on indices, paralleling the multiplication of cosine waves. Such identities simplify nonlinear polynomial calculations without conversion to powers of xx.

For n≥1n\ge1, all zeros of TnT_n are simple and lie strictly inside (−1,1)(-1,1):

xk=cos⁡((2k−1)π2n),k=1,…,n.x_k=\cos\left(\frac{(2k-1)\pi}{2n}\right), \qquad k=1,\ldots,n.

Its extrema occur at

ξk=cos⁡(kπn),k=0,…,n,\xi_k=\cos\left(\frac{k\pi}{n}\right), \qquad k=0,\ldots,n,

where Tn(ξk)=(−1)kT_n(\xi_k)=(-1)^k. The zeros of UnU_n are

cos⁡(kπn+1),k=1,…,n.\cos\left(\frac{k\pi}{n+1}\right), \qquad k=1,\ldots,n.

Uniform spacing in the angle therefore becomes nonuniform spacing in xx, with increasingly dense points near the endpoints. This geometry is fundamental to Chebyshev interpolation.

Useful relations between the two families include

Tn′(x)=nUn−1(x),Tn(x)=Un(x)−xUn−1(x)T_n'(x)=nU_{n-1}(x), \qquad T_n(x)=U_n(x)-xU_{n-1}(x)

for n≥1n\ge1, and

Tn(x)2+(1−x2)Un−1(x)2=1.T_n(x)^2+(1-x^2)U_{n-1}(x)^2=1.

The last formula is a polynomial version of the identity cos⁡2(nθ)+sin⁡2(nθ)=1\cos^2(n\theta)+\sin^2(n\theta)=1. It explains simultaneously why ∣Tn(x)∣≤1|T_n(x)|\le1 on [−1,1][-1,1] and why the first-kind extrema are tied to zeros of a second-kind polynomial.

Outside the interval, oscillation changes into rapid growth. For x≥1x\ge1,

Tn(x)=cosh⁡(n arcosh⁡x).T_n(x)=\cosh(n\,\operatorname{arcosh}x).

More generally, if x=(z+z−1)/2x=(z+z^{-1})/2 with z≠0z\ne0, then

Tn(x)=zn+z−n2.T_n(x)=\frac{z^n+z^{-n}}2.

Equivalently,

Tn(x)=12[(x+x2−1)n+(x−x2−1)n].T_n(x)=\frac12\left[ (x+\sqrt{x^2-1})^n+(x-\sqrt{x^2-1})^n \right].

The expression is independent of the choice between the two square roots. These formulas connect the polynomials to complex analysis and explain both their strong approximation properties inside an interval and the possible instability of extrapolation beyond it.

Their generating functions are rational:

∑n=0∞Tn(x)tn=1−xt1−2xt+t2,∑n=0∞Un(x)tn=11−2xt+t2,\sum_{n=0}^{\infty}T_n(x)t^n =\frac{1-xt}{1-2xt+t^2}, \qquad \sum_{n=0}^{\infty}U_n(x)t^n =\frac1{1-2xt+t^2},

for tt sufficiently close to zero. Multiplying by the denominator and matching powers of tt recovers the recurrence relations directly.

Orthogonality and differential equations

Chebyshev polynomials are classical orthogonal polynomials, but their orthogonality requires a specified weight. For the first kind,

∫−11Tm(x)Tn(x)1−x2 dx={0,m≠n,π,m=n=0,π/2,m=n≥1.\int_{-1}^{1} \frac{T_m(x)T_n(x)}{\sqrt{1-x^2}}\,dx = \begin{cases} 0,&m\ne n,\\ \pi,&m=n=0,\\ \pi/2,&m=n\ge1. \end{cases}

The substitution x=cos⁡θx=\cos\theta transforms this integral into

∫0πcos⁡(mθ)cos⁡(nθ) dθ.\int_0^\pi \cos(m\theta)\cos(n\theta)\,d\theta.

Thus first-kind orthogonality is exactly cosine orthogonality expressed in a different coordinate. The endpoint singularities of the weight (1−x2)−1/2(1-x^2)^{-1/2} are integrable and form an essential part of the structure.

For the second kind,

∫−11Um(x)Un(x)1−x2 dx=π2δmn,\int_{-1}^{1} U_m(x)U_n(x)\sqrt{1-x^2}\,dx =\frac{\pi}{2}\delta_{mn},

where δmn\delta_{mn} equals one when m=nm=n and zero otherwise. Here the same substitution produces the ordinary sine orthogonality relation. The relevant weights therefore distinguish the two families even though their recurrence relations are identical. Neither family is generally orthogonal for the unweighted integral on [−1,1][-1,1]; that role belongs to a different classical family, the Legendre polynomials.

These integrals define weighted inner products. With the corresponding weighted square-integrable function spaces, the Chebyshev polynomials form complete orthogonal systems. Completeness means that finite linear combinations are dense in the relevant Hilbert space, so square-integrable functions admit expansions converging in the weighted mean-square norm. It does not imply convergence at every point or uniform convergence. In particular, a continuous function automatically belongs to the first-kind weighted space, but pointwise properties of its expansion require additional analysis.

The first-kind polynomial satisfies the differential equation

(1−x2)y′′−xy′+n2y=0,(1-x^2)y''-xy'+n^2y=0,

whereas the second-kind polynomial satisfies

(1−x2)y′′−3xy′+n(n+2)y=0.(1-x^2)y''-3xy'+n(n+2)y=0.

The first equation has the self-adjoint form

−ddx(1−x2 y′)=n21−x2 y.-\frac{d}{dx}\left(\sqrt{1-x^2}\,y'\right) =\frac{n^2}{\sqrt{1-x^2}}\,y.

Similarly, the second can be written

−ddx((1−x2)3/2y′)=n(n+2)1−x2 y.-\frac{d}{dx}\left((1-x^2)^{3/2}y'\right) =n(n+2)\sqrt{1-x^2}\,y.

These are examples of Sturm–Liouville problems, with singular endpoints.

For polynomial solutions, the boundary terms arising in integration by parts vanish. Subtracting the equations for two different degrees then proves orthogonality whenever their eigenvalues differ. This places the elementary cosine identities within a broader spectral framework: the polynomials are eigenfunctions of weighted differential operators, and the corresponding eigenvalues are n2n^2 or n(n+2)n(n+2). Their derivative identities and weighted orthogonality are therefore related consequences of one underlying differential structure.

Extremal properties and minimax approximation

The first-kind family solves a fundamental extremal problem. A monic polynomial has leading coefficient one. Among all real monic polynomials pp of degree n≥1n\ge1, the unique polynomial minimizing

∥p∥∞=max⁡−1≤x≤1∣p(x)∣\|p\|_\infty=\max_{-1\le x\le1}|p(x)|

is

p∗(x)=21−nTn(x),∥p∗∥∞=21−n.p_*(x)=2^{1-n}T_n(x), \qquad \|p_*\|_\infty=2^{1-n}.

This is not an assertion that TnT_n is universally the best polynomial approximation to a given function. Rather, it identifies the smallest possible maximum magnitude when the highest-degree coefficient is fixed.

The proof is based on alternating signs. Suppose a different monic polynomial qq had strictly smaller norm. At the n+1n+1 extrema of TnT_n, the difference q−p∗q-p_* would have alternating signs opposite to those of p∗p_*. It would therefore have a zero between each consecutive pair, giving at least nn distinct zeros. Yet the leading terms cancel, so q−p∗q-p_* has degree at most n−1n-1, which is impossible. The same alternation structure, with a refinement of the sign argument, establishes uniqueness.

The broader minimax approximation problem seeks a polynomial pp of degree at most nn minimizing ∥f−p∥∞\|f-p\|_\infty for a given continuous real function ff on a compact interval. The equioscillation theorem characterizes the unique optimum: there must be at least n+2n+2 ordered points at which the error attains its maximum absolute value with alternating signs. Chebyshev polynomials are the model example of this equal-ripple behavior. For instance, the best linear approximation to x2x^2 on [−1,1][-1,1] is the constant 1/21/2, because

x2−12=12T2(x)x^2-\frac12=\frac12T_2(x)

alternates between errors 1/2,−1/2,1/21/2,-1/2,1/2.

The same family is extremal in derivative bounds. Markov’s inequality for polynomials states that every real polynomial of degree at most nn satisfies

∥p′∥∞≤n2∥p∥∞\|p'\|_\infty\le n^2\|p\|_\infty

on [−1,1][-1,1]. Equality is attained by TnT_n, since Tn′(1)=n2T_n'(1)=n^2. The related Bernstein inequality is

∣p′(x)∣≤n1−x2∥p∥∞,−1<x<1.|p'(x)|\le \frac{n}{\sqrt{1-x^2}}\|p\|_\infty, \qquad -1<x<1.

Together these results quantify how rapidly a bounded polynomial can vary, including the stronger endpoint variation.

For general functions, minimax coefficients are usually not available in closed form. The Remez algorithm, introduced by Evgeny Remez in 1934, iteratively adjusts alternating extremal points to construct a minimax approximant. Truncated Chebyshev series often supply useful initial approximations, but truncation and true minimax approximation are distinct procedures: the latter explicitly optimizes the maximum error.

Chebyshev series and convergence rates

A first-kind Chebyshev expansion is conventionally written

f(x)∼a02+∑n=1∞anTn(x),f(x)\sim\frac{a_0}{2}+\sum_{n=1}^{\infty}a_nT_n(x),

where

an=2π∫−11f(x)Tn(x)1−x2 dx=2π∫0πf(cos⁡θ)cos⁡(nθ) dθ.a_n=\frac{2}{\pi}\int_{-1}^{1} \frac{f(x)T_n(x)}{\sqrt{1-x^2}}\,dx =\frac{2}{\pi}\int_0^\pi f(\cos\theta)\cos(n\theta)\,d\theta.

The factor 1/21/2 on the constant term compensates for the different norm of T0T_0. Other conventions absorb that factor into the constant coefficient; mixing conventions produces a common normalization error. The coefficients are those of the cosine series of g(θ)=f(cos⁡θ)g(\theta)=f(\cos\theta), extended as an even, 2π2\pi-periodic function.

For a function in the weighted square-integrable space, orthogonal projection onto the first N+1N+1 polynomials gives the best approximation in that weighted mean-square norm. The associated Parseval identity is

∫−11∣f(x)∣21−x2 dx=π∣a0∣24+π2∑n=1∞∣an∣2.\int_{-1}^{1}\frac{|f(x)|^2}{\sqrt{1-x^2}}\,dx =\frac{\pi|a_0|^2}{4} +\frac{\pi}{2}\sum_{n=1}^{\infty}|a_n|^2.

This least-squares optimality concerns an integrated error, not the maximum error. The distinction explains why a truncated Chebyshev series can be computationally excellent without being exactly minimax.

Convergence depends on regularity. Since ∣Tn(x)∣≤1|T_n(x)|\le1 on the interval, absolute summability of the coefficients implies uniform convergence. For the degree-NN truncation SNS_N,

∥f−SN∥∞≤∑n>N∣an∣\|f-S_N\|_\infty \le\sum_{n>N}|a_n|

whenever the series represents ff and this tail is summable. Smoothness of the periodic function f(cos⁡θ)f(\cos\theta) leads to algebraic coefficient decay through repeated integration by parts. Conversely, interior corners or jumps prevent geometric decay. A jump also produces the persistent local overshoot associated with the Gibbs phenomenon in Fourier analysis.

For an analytic function, convergence can be geometric. A Bernstein ellipse EρE_\rho, with parameter ρ>1\rho>1, is parametrized by

z=12(w+w−1),∣w∣=ρ.z=\frac12(w+w^{-1}),\qquad |w|=\rho.

Its foci are −1-1 and 11, and its semiaxes are (ρ+ρ−1)/2(\rho+\rho^{-1})/2 and (ρ−ρ−1)/2(\rho-\rho^{-1})/2. If ff is analytic on and inside this ellipse and bounded there by MM, then

∣an∣≤2Mρ−n,∥f−SN∥∞≤2Mρ−Nρ−1.|a_n|\le2M\rho^{-n}, \qquad \|f-S_N\|_\infty \le\frac{2M\rho^{-N}}{\rho-1}.

The nearest complex singularities restrict the admissible ellipse and hence determine the asymptotic geometric rate. For example, f(x)=1/(1+25x2)f(x)=1/(1+25x^2) has poles at ±i/5\pm i/5. The ellipse through those poles has parameter

ρ∗=15+1+125≈1.2198.\rho_*=\frac15+\sqrt{1+\frac1{25}}\approx1.2198.

Bounds can use any 1<ρ<ρ∗1<\rho<\rho_*, but not an ellipse containing a pole. This explains why apparently smooth functions on a real interval may nevertheless require many coefficients: convergence is governed partly by their behavior away from the real axis.

Interpolation at Chebyshev points

Chebyshev points are widely used for polynomial interpolation because they combine explicit formulas with controlled error amplification. Two grids are standard. The N+1N+1 roots of TN+1T_{N+1} are

xj=cos⁡((2j+1)π2N+2),j=0,…,N.x_j=\cos\left(\frac{(2j+1)\pi}{2N+2}\right), \qquad j=0,\ldots,N.

The N+1N+1 extrema of TNT_N, often called Chebyshev–Lobatto points, are

xj=cos⁡(jπN),j=0,…,N,x_j=\cos\left(\frac{j\pi}{N}\right), \qquad j=0,\ldots,N,

for N≥1N\ge1. The latter include both endpoints and are therefore convenient for boundary conditions.

Their endpoint clustering addresses a major weakness of equally spaced interpolation. The classical Runge example, f(x)=1/(1+25x2)f(x)=1/(1+25x^2), exhibits divergence near the endpoints for increasingly high-degree interpolation on equally spaced grids. Chebyshev grids do not make interpolation converge uniformly for every continuous function, but they avoid that example’s instability and provide strong convergence results for sufficiently smooth functions. Their interpolation Lebesgue constants, which measure worst-case amplification of sample errors in the uniform norm, satisfy

ΛN=2πlog⁡N+O(1).\Lambda_N=\frac{2}{\pi}\log N+O(1).

The corresponding constants for equally spaced points grow exponentially.

For the root grid, the extremal monic-polynomial property yields

∏j=0N(x−xj)=2−NTN+1(x).\prod_{j=0}^{N}(x-x_j)=2^{-N}T_{N+1}(x).

If ff has N+1N+1 continuous derivatives, the interpolation remainder gives

∥f−pN∥∞≤2−N(N+1)!∥f(N+1)∥∞.\|f-p_N\|_\infty \le \frac{2^{-N}}{(N+1)!} \|f^{(N+1)}\|_\infty.

The root grid minimizes the maximum magnitude of the nodal product among all choices of N+1N+1 real interpolation nodes in the interval. However, this derivative-based bound can be pessimistic or uninformative for some functions; complex-analytic estimates often describe actual convergence more effectively.

The interpolating polynomial is efficiently evaluated using the barycentric formula

pN(x)=∑j=0Nλjf(xj)x−xj∑j=0Nλjx−xj.p_N(x)= \frac{\displaystyle\sum_{j=0}^{N} \frac{\lambda_j f(x_j)}{x-x_j}} {\displaystyle\sum_{j=0}^{N} \frac{\lambda_j}{x-x_j}}.

At a node, the prescribed sample value is returned directly. For Chebyshev–Lobatto points, a common normalization is

λj=(−1)jδj,δ0=δN=12,δj=1(1≤j≤N−1).\lambda_j=(-1)^j\delta_j,\qquad \delta_0=\delta_N=\frac12,\qquad \delta_j=1\quad(1\le j\le N-1).

Only relative weights matter. For the root grid, weights are proportional to (−1)jsin⁡θj(-1)^j\sin\theta_j, where θj=(2j+1)π/(2N+2)\theta_j=(2j+1)\pi/(2N+2).

Discrete cosine transforms convert sampled values into coefficients of the interpolating polynomial. These coefficients are not generally identical to continuous-series coefficients: unresolved high-degree modes alias into lower modes on a finite grid. This distinction matters in adaptive approximation and nonlinear calculations, where insufficient sampling can make a computed tail appear smaller than the unresolved error actually is.

Gaussian and Clenshaw–Curtis quadrature

Chebyshev orthogonality leads directly to Gaussian quadrature. With nn roots of TnT_n, the first-kind Gauss–Chebyshev rule is

∫−11f(x)1−x2 dx≈πn∑k=1nf(cos⁡(2k−1)π2n).\int_{-1}^{1}\frac{f(x)}{\sqrt{1-x^2}}\,dx \approx \frac{\pi}{n} \sum_{k=1}^{n} f\left(\cos\frac{(2k-1)\pi}{2n}\right).

It is exact whenever ff is a polynomial of degree at most 2n−12n-1. All weights are equal, an unusually simple feature among Gaussian rules. Under x=cos⁡θx=\cos\theta, the formula becomes a midpoint rule for the integral of f(cos⁡θ)f(\cos\theta) on [0,π][0,\pi].

For the second kind,

∫−11f(x)1−x2 dx≈πn+1∑k=1nsin⁡2(kπn+1)f(cos⁡kπn+1).\int_{-1}^{1}f(x)\sqrt{1-x^2}\,dx \approx \frac{\pi}{n+1} \sum_{k=1}^{n} \sin^2\left(\frac{k\pi}{n+1}\right) f\left(\cos\frac{k\pi}{n+1}\right).

This rule also has polynomial exactness through degree 2n−12n-1. The differing weights reflect the different orthogonality measures. Applying the first rule directly to an unweighted integral would therefore solve the wrong integration problem. An unweighted integral can be rewritten to fit a weighted rule, but the transformed integrand may have less favorable endpoint regularity.

For ordinary integration,

∫−11f(x) dx,\int_{-1}^{1}f(x)\,dx,

a widely used alternative is Clenshaw–Curtis quadrature, developed by Charles Clenshaw and Alan Curtis in 1960. It samples at Chebyshev–Lobatto points, constructs the degree-NN interpolant, and integrates that polynomial exactly. The basic moments are

∫−11Tk(x) dx={2,k=0,0,k odd,21−k2,k≥2 even.\int_{-1}^{1}T_k(x)\,dx = \begin{cases} 2,&k=0,\\ 0,&k\text{ odd},\\ \dfrac{2}{1-k^2},&k\ge2\text{ even}. \end{cases}

These identities permit the integral to be assembled directly from the interpolant’s Chebyshev coefficients.

An (N+1)(N+1)-point Clenshaw–Curtis rule is automatically exact through degree NN, with an additional degree of exactness when symmetry makes the next degree odd. This is lower than the 2N+12N+1 exactness of (N+1)(N+1)-point Gauss–Legendre quadrature. Nevertheless, degree of polynomial exactness alone does not determine performance for general functions. For smooth or analytic integrands, Clenshaw–Curtis often converges at rates comparable in practical significance to Gaussian quadrature because high-order Chebyshev coefficients are already small.

A further advantage is nestedness. Replacing NN by 2N2N retains every previous Lobatto point, so refinement reuses existing function evaluations. Fast transforms also make weight or coefficient construction inexpensive. These properties are valuable when integrand evaluations are costly or when adaptive refinement is required. Both Gaussian and Clenshaw–Curtis methods remain sensitive to singularities, discontinuities, and extreme localization; their convergence follows the regularity of the integrand rather than a universal fixed rate.

Stable evaluation, transforms, and polynomial arithmetic

The three-term recurrence permits evaluation of all Tk(x)T_k(x) up to degree NN using O(N)O(N) arithmetic operations and a small amount of working memory. It also avoids explicit expansion into monomials, whose coefficients can become large even when the polynomial is bounded by one on [−1,1][-1,1]. The identity Tn(x)=cos⁡(narccos⁡x)T_n(x)=\cos(n\arccos x) is useful mathematically, but direct trigonometric evaluation is not always the best computational method, particularly near endpoints or for large degree.

For an entire expansion, the standard method is Clenshaw’s algorithm, published by Charles Clenshaw in 1955. Using the convention

p(x)=c0+∑k=1NckTk(x),p(x)=c_0+\sum_{k=1}^{N}c_kT_k(x),

set bN+1=bN+2=0b_{N+1}=b_{N+2}=0, and compute backward:

bk=ck+2xbk+1−bk+2,k=N,N−1,…,1.b_k=c_k+2xb_{k+1}-b_{k+2}, \qquad k=N,N-1,\ldots,1.

Then

p(x)=c0+xb1−b2.p(x)=c_0+xb_1-b_2.

This is the counterpart of Horner’s method for the Chebyshev basis. It evaluates the expansion in linear time without separately storing every basis polynomial.

The method generally has favorable numerical stability on the approximation interval, but stability must be distinguished from conditioning. If terms cancel so that p(x)p(x) is very small, relative error in the final value may still be large. Likewise, extrapolation outside [−1,1][-1,1] can magnify coefficient errors because the basis polynomials grow rapidly there. Near endpoints, specialized rearrangements can improve accuracy in cases where cancellation affects the standard final expression.

Coefficient transforms are particularly efficient. Samples on a Lobatto grid correspond to a type-I discrete cosine transform, while root grids lead to related type-II or type-III conventions depending on transform direction and normalization. These transforms can be implemented through the fast Fourier transform, reducing the conversion cost to O(Nlog⁡N)O(N\log N). Care with endpoint weights and the constant coefficient is essential: different software libraries often use different normalizations even when their sample grids agree.

Differentiation can also be performed directly in coefficient space. If

p(x)=∑k=0NckTk(x),p′(x)=∑k=0N−1dkTk(x),p(x)=\sum_{k=0}^{N}c_kT_k(x), \qquad p'(x)=\sum_{k=0}^{N-1}d_kT_k(x),

set dN=dN+1=0d_N=d_{N+1}=0, compute

dk−1=dk+1+2kck,k=N,N−1,…,1,d_{k-1}=d_{k+1}+2kc_k, \qquad k=N,N-1,\ldots,1,

and finally replace d0d_0 by d0/2d_0/2. Antiderivatives follow from

∫Tn(x) dx=Tn+1(x)2(n+1)−Tn−1(x)2(n−1),n≥2,\int T_n(x)\,dx =\frac{T_{n+1}(x)}{2(n+1)} -\frac{T_{n-1}(x)}{2(n-1)}, \qquad n\ge2,

with n=0,1n=0,1 handled separately. Multiplication uses the product identity or transform-based evaluation and resampling. In nonlinear computations, the sampling grid must be enlarged sufficiently to prevent higher-degree products from aliasing into retained coefficients.

Spectral methods, matrix algorithms, and filters

Chebyshev representations are central to spectral methods for differential equations on bounded, nonperiodic intervals. A solution is approximated by

uN(x)=∑k=0NckTk(x),u_N(x)=\sum_{k=0}^{N}c_kT_k(x),

and its coefficients are chosen so that a differential equation and its boundary conditions are satisfied in a prescribed discrete sense. In collocation methods, the equation is enforced at selected Chebyshev points. In Galerkin or tau formulations, residual conditions are imposed through weighted projections or coefficient constraints.

At a Lobatto grid, differentiation can be represented by a matrix. If λj\lambda_j are barycentric weights, its off-diagonal entries are

Dij=λjλi(xi−xj),i≠j,D_{ij}=\frac{\lambda_j}{\lambda_i(x_i-x_j)}, \qquad i\ne j,

and diagonal entries can be defined by

Dii=−∑j≠iDij.D_{ii}=-\sum_{j\ne i}D_{ij}.

The row-sum condition ensures that constants differentiate to zero. Boundary conditions replace or supplement selected equations. For smooth solutions, accuracy can increase much faster with degree than in a fixed-order local discretization; for analytic solutions, geometric convergence is possible.

Endpoint clustering is both useful and costly. The smallest grid spacing is of order N−2N^{-2}, which helps resolve rapid endpoint variation but can impose restrictive explicit time steps for time-dependent equations. Differentiation also amplifies high-degree errors, so accurately approximating a function does not automatically guarantee equally accurate high derivatives. Singular solutions, discontinuous coefficients, and complex geometries often require domain decomposition, mappings, or other adaptations rather than a single global expansion.

In numerical linear algebra, Chebyshev polynomials provide controlled polynomial filters. Suppose a symmetric positive-definite matrix AA has spectrum in [a,b][a,b], with 0<a<b0<a<b. Mapping that interval to [−1,1][-1,1] gives the normalized residual polynomial

qn(λ)=Tn ⁣(b+a−2λb−a)Tn ⁣(b+ab−a),qn(0)=1.q_n(\lambda)= \frac{T_n\!\left(\dfrac{b+a-2\lambda}{b-a}\right)} {T_n\!\left(\dfrac{b+a}{b-a}\right)}, \qquad q_n(0)=1.

It minimizes the maximum magnitude on [a,b][a,b] among degree-at-most-nn polynomials with this normalization. Writing κ=b/a\kappa=b/a,

max⁡λ∈[a,b]∣qn(λ)∣≤2(κ−1κ+1)n.\max_{\lambda\in[a,b]}|q_n(\lambda)| \le 2\left(\frac{\sqrt{\kappa}-1}{\sqrt{\kappa}+1}\right)^n.

This property underlies Chebyshev semi-iterative methods. Their recurrence coefficients can be prepared from spectral bounds, avoiding the inner products required by some other iterative methods, although inaccurate bounds can impair convergence.

Chebyshev polynomials also shape equal-ripple responses in signal processing. A type-I Chebyshev filter has normalized squared magnitude

∣H(iΩ)∣2=11+ε2Tn(Ω)2.|H(i\Omega)|^2= \frac{1}{1+\varepsilon^2T_n(\Omega)^2}.

For ∣Ω∣≤1|\Omega|\le1, bounded oscillation produces passband ripple controlled by ε\varepsilon; beyond that range, rapid polynomial growth produces attenuation. The same mathematics appears in Dolph–Chebyshev array weighting and windows, where equal-ripple sidelobes implement a minimax trade-off under specified constraints. These applications exploit the contrast between controlled oscillation within an interval and strong growth outside it.