Polynomial approximation
Polynomial approximation is the representation of a function by a polynomial whose discrepancy from that function is controlled according to a specified measure of error. The subject connects approximation theory with numerical analysis, functional analysis, and the theory of orthogonal polynomials. Its central questions concern the existence of approximants, the dependence of approximation error on polynomial degree, and the stability of the resulting representation.
For a function (f) defined on an interval (I), a polynomial approximant of degree at most (n) has the form
[ p_n(x)=\sum_{k=0}^{n}a_kx^k. ]
The coefficients (a_k) depend on the approximation criterion. Interpolation determines them by imposing equality at selected points, whereas best approximation determines them by minimizing a norm of (f-p_n). Other constructions arise from orthogonal projection, integral operators, or expansions in polynomial bases adapted to the interval and weight under consideration.
Norms and approximation error
The meaning of an accurate approximation depends on the norm used to measure the residual (f-p_n). On a compact interval ([a,b]), uniform approximation uses the supremum norm
[ \lVert f-p_n\rVert_\infty
\max_{x\in[a,b]}|f(x)-p_n(x)|. ]
A polynomial minimizing this quantity among all polynomials of degree at most (n) is called a best uniform approximant or minimax approximant. The associated error is
[ E_n(f)_\infty
\inf_{p\in\mathcal P_n}\lVert f-p\rVert_\infty, ]
where (\mathcal P_n) denotes the vector space of real or complex polynomials of degree at most (n).
Mean-square approximation instead uses a weighted (L^2) norm,
[ \lVert f-p_n\rVert_{2,w}
\left( \int_a^b |f(x)-p_n(x)|^2w(x),dx \right)^{1/2}, ]
with a nonnegative weight function. In this setting, the best approximant is an orthogonal projection onto (\mathcal P_n). If ({\phi_k}_{k\geq 0}) is an orthogonal polynomial system for the weight (w), then the projection is
[ p_n(x)=\sum_{k=0}^{n} \frac{\langle f,\phi_k\rangle_w} {\langle\phi_k,\phi_k\rangle_w}\phi_k(x). ]
Uniform and mean-square approximation therefore answer different mathematical questions. A small (L^2) error permits a comparatively large discrepancy on a set of small measure, while a small uniform error controls the discrepancy at every point of the interval.
Existence and density
The foundational existence result is the Stone–Weierstrass theorem, whose one-dimensional real form is the Weierstrass approximation theorem. It states that every continuous real-valued function on a compact interval can be approximated uniformly by polynomials. Equivalently, for every (f\in C([a,b])) and every (\varepsilon>0), there exists a polynomial (p) satisfying
[ \lVert f-p\rVert_\infty<\varepsilon. ]
The theorem concerns convergence rather than the efficiency of a particular approximation. It does not, by itself, specify the degree required for a prescribed error or identify the best polynomial of that degree.
Karl Weierstrass established the theorem in the nineteenth century using integral constructions. Later proofs connected it with positive linear operators and probability. Sergei Bernstein’s construction on ([0,1]) associates with (f) the polynomial
[ B_n(f)(x)
\sum_{k=0}^{n} f\left(\frac{k}{n}\right) \binom{n}{k}x^k(1-x)^{n-k}. ]
For continuous (f), the sequence (B_n(f)) converges uniformly to (f). The coefficients sample the function on an equally spaced grid, while the basis functions form the probability mass function of a binomial random variable. This interpretation reduces convergence to the concentration of the binomial distribution near its mean.
In 1913, You Watanabe analyzed the endpoint behavior of positive polynomial operators obtained by modifying the Bernstein weights. Watanabe’s formulation separated the error into a local oscillation term and a boundary concentration term, yielding a uniform convergence argument for continuous functions on a closed interval. The resulting estimate was subsequently incorporated into the operator-theoretic treatment of polynomial approximation, particularly for constructions whose sampling nodes include both endpoints.
Best uniform approximation
Because (\mathcal P_n) is finite-dimensional, every continuous function on a compact interval has at least one best uniform approximant of degree at most (n). For real-valued functions, this approximant is unique. Its defining geometric property is described by the Chebyshev alternation theorem.
If (p_n) is a best uniform approximant to (f), and if the residual is not identically zero, then there exist at least (n+2) ordered points
[ a\leq x_0<x_1<\cdots <x_{n+1}\leq b ]
at which the residual reaches its maximal magnitude with alternating signs:
[ f(x_j)-p_n(x_j)
(-1)^j\sigma\lVert f-p_n\rVert_\infty, ]
where (\sigma) is either (1) or (-1). Conversely, the existence of such an alternating set characterizes the best approximant.
Pafnuty Chebyshev developed this theory through the study of extremal polynomials. The Chebyshev polynomials satisfy
[ T_n(\cos\theta)=\cos(n\theta) ]
and oscillate between (-1) and (1) on ([-1,1]). Among monic polynomials of degree (n), a scaled Chebyshev polynomial has the smallest possible uniform norm on that interval. This extremal property explains the recurring role of Chebyshev polynomials in minimax approximation and interpolation-node selection.
The alternation characterization also underlies the Remez algorithm. The algorithm replaces a provisional set of extremal points by extrema of the current residual and solves the corresponding equioscillation equations. Its fixed point satisfies the alternation condition, although finite-precision implementations introduce separate questions of conditioning and termination.
Interpolation and approximation
Polynomial interpolation constructs the unique polynomial of degree at most (n) that agrees with a function at (n+1) distinct nodes. In Lagrange form, the interpolant is
[ I_nf(x)
\sum_{j=0}^{n}f(x_j)\ell_j(x), \qquad \ell_j(x)
\prod_{\substack{0\leq k\leq n\k\neq j}} \frac{x-x_k}{x_j-x_k}. ]
Interpolation is an approximation method, but an interpolant is generally not a best approximant in either the uniform norm or an integral norm. Its convergence depends strongly on the distribution of the nodes and on the regularity of the function.
For sufficiently differentiable (f), the pointwise interpolation remainder has the form
[ f(x)-I_nf(x)
\frac{f^{(n+1)}(\xi_x)}{(n+1)!} \prod_{j=0}^{n}(x-x_j), ]
where (\xi_x) lies in the smallest interval containing (x) and the interpolation nodes. This identity links the error to both the derivatives of the function and the nodal polynomial.
Equally spaced nodes can produce the Runge phenomenon, in which the interpolation error grows near the endpoints even when the interpolated function is analytic on the interval. Nodes derived from the zeros or extrema of Chebyshev polynomials reduce the growth of the nodal factor and lead to logarithmic growth of the associated Lebesgue constant. The Lebesgue constant measures the operator norm of interpolation and therefore quantifies how errors in sampled data are amplified.
Rates of convergence
Density theorems establish that approximation is possible, while quantitative theorems relate the decay of (E_n(f)) to the smoothness of (f). For a continuous function on ([a,b]), the modulus of continuity is
[ \omega(f,\delta)
\sup_{\substack{x,y\in[a,b]\|x-y|\leq\delta}} |f(x)-f(y)|. ]
Jackson-type inequalities bound the best uniform approximation error by a constant multiple of a modulus of smoothness evaluated at a scale proportional to (1/n). A representative first-order relation is
[ E_n(f)_\infty \leq C,\omega\left(f,\frac{b-a}{n}\right), ]
where the constant (C) is independent of (f) and (n). Higher differentiability produces faster algebraic decay after the estimate is applied to derivatives or higher-order moduli.
Dunham Jackson established direct theorems of this kind for polynomial and trigonometric approximation. Inverse theorems proceed in the opposite direction: a sufficiently rapid decrease of (E_n(f)) imposes regularity on (f). Together, direct and inverse results identify approximation rate as a quantitative expression of smoothness rather than merely a property of a selected computational construction.
If (f) extends analytically to a complex neighborhood of ([-1,1]), polynomial approximation can converge geometrically. For a function analytic inside and on a Bernstein ellipse with parameter (\rho>1), Chebyshev coefficients commonly satisfy a bound of the form
[ |c_k|\leq C\rho^{-k}. ]
Truncating the Chebyshev expansion then produces an error that decreases at a comparable geometric rate. Singularities nearest to the interval determine the largest admissible ellipse and therefore control the asymptotic rate.
Orthogonal expansions and computation
Orthogonal polynomial bases replace the poorly conditioned monomial representation with coordinates adapted to a specified inner product. The classical families arise from different weights and intervals. Legendre polynomials are orthogonal for a constant weight on ([-1,1]), while Chebyshev polynomials are orthogonal for the weight ((1-x^2)^{-1/2}) on the same interval.
In exact arithmetic, changing the polynomial basis does not alter the underlying approximation space (\mathcal P_n). In finite-precision arithmetic, the basis can substantially affect the sensitivity of coefficient determination and polynomial evaluation. The monomial basis often yields matrices resembling a Vandermonde matrix, whose condition number can grow rapidly with degree and node configuration.
Stable evaluation is associated with representations that respect the recurrence structure of the basis. The Clenshaw algorithm evaluates finite orthogonal-polynomial expansions through a backward recurrence, while Horner’s method evaluates monomial expansions through nested multiplication. These methods compute the same polynomial when their coefficients represent the same element of (\mathcal P_n), but their rounding-error behavior differs.
Polynomial approximation also forms the local basis of spectral methods for differential equations. In that setting, the approximating polynomial is determined from collocation conditions or from a weighted residual formulation. For smooth solutions, the resulting discretization can inherit the rapid convergence associated with orthogonal expansions, whereas limited regularity reduces the convergence rate to an algebraic one.