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

Orthogonal Polynomials

Orthogonal polynomials are polynomial sequences mutually orthogonal under an inner product, linking approximation, integration, spectral theory, and probability.

43 keywords16 linked from12 not yet writtenWritten by AI
PolynomialInner productMathematical Ana…ProbabilityMeasure TheoryGram–Schmidt Pro…Orthogonal proje…Gram matrixOrthogonal…

Orthogonal polynomials are sequences of polynomials in which distinct members have zero inner product with respect to a specified measure or weight. Typically, the sequence contains one polynomial of each nonnegative degree. Familiar examples include the Legendre, Chebyshev, Jacobi, Laguerre, and Hermite polynomials. Their orthogonality plays a role analogous to the perpendicularity of coordinate axes: it separates polynomial components that would otherwise interact when functions are expanded or approximated.

These systems connect algebraic structure with analysis. Orthogonality forces a short recurrence relation, constrains the location of zeros, and produces integration formulas of exceptionally high exactness. Particular families also satisfy differential or difference equations. Consequently, orthogonal polynomials are fundamental tools in approximation theory, numerical integration, probability, mathematical physics, and the spectral analysis of operators. Their theory extends beyond continuous weights on intervals to discrete measures, complex contours, several variables, and matrix-valued systems.

Orthogonality, measures, and normalization

The standard setting begins with a positive measure μ\mu on the real line, having infinite support and finite moments of every order:

mk=∫Rxk dμ(x),k=0,1,2,….m_k=\int_{\mathbb R}x^k\,d\mu(x),\qquad k=0,1,2,\ldots.

More precisely, absolute moments are assumed finite, so every polynomial product is integrable. The measure defines

⟨f,g⟩μ=∫Rf(x)g(x)‾ dμ(x).\langle f,g\rangle_\mu =\int_{\mathbb R}f(x)\overline{g(x)}\,d\mu(x).

The language of measure theory accommodates continuous distributions, discrete point masses, and mixtures within the same definition. When dμ(x)=w(x) dxd\mu(x)=w(x)\,dx on an interval II, the function ww is called the weight.

An orthogonal polynomial sequence {pn}n≥0\{p_n\}_{n\geq0} satisfies

deg⁡pn=n,⟨pn,pm⟩μ=0(n≠m).\deg p_n=n,\qquad \langle p_n,p_m\rangle_\mu=0\quad(n\ne m).

Infinite support ensures that every nonzero polynomial has positive squared norm: a polynomial has only finitely many zeros and therefore cannot vanish almost everywhere for such a measure. Applying the Gram–Schmidt process to 1,x,x2,…1,x,x^2,\ldots consequently constructs the sequence. If the support consists of exactly NN distinct points, the corresponding construction instead gives NN nonzero orthogonal polynomials, of degrees 00 through N−1N-1.

Orthogonality alone leaves an arbitrary nonzero multiplier at every degree. Two standard conventions remove this freedom. A monic polynomial has leading coefficient 11; an orthonormal polynomial has norm 11, conventionally with positive leading coefficient. Classical named families often use other normalizations, chosen for convenient endpoint values, generating functions, or differentiation formulas. Thus the statement that a family is orthogonal does not by itself specify its squared norms or recurrence coefficients.

If PnP_n is monic and hn=⟨Pn,Pn⟩μh_n=\langle P_n,P_n\rangle_\mu, orthogonality gives an important extremal characterization. Every other monic degree-nn polynomial has the form Pn+rP_n+r, with deg⁡r<n\deg r<n, and hence

∥Pn+r∥μ2=hn+∥r∥μ2.\|P_n+r\|_\mu^2=h_n+\|r\|_\mu^2.

Therefore PnP_n uniquely minimizes the weighted squared norm among monic polynomials of its degree. This is an instance of orthogonal projection, rather than a separate optimization phenomenon.

The moments also define a Gram matrix,

Gn=(mj+k)j,k=0n.G_n=(m_{j+k})_{j,k=0}^{n}.

Its positive definiteness is equivalent to positivity of the polynomial inner product through degree nn. Determinant formulas derived from these matrices express orthogonal polynomials in terms of moments. They establish an algebraic connection between the sequence and its measure, although they are often poor formulas for numerical computation because moment matrices can be severely ill-conditioned.

Three-term recurrence and spectral structure

Multiplication by the independent variable is the central operation behind the theory. Let pnp_n be orthonormal with positive leading coefficient, and set p−1=0p_{-1}=0. Then

xpn(x)=an+1pn+1(x)+bnpn(x)+anpn−1(x),xp_n(x) =a_{n+1}p_{n+1}(x)+b_np_n(x)+a_np_{n-1}(x),

where an>0a_n>0 for n≥1n\geq1, bnb_n is real, and the term involving p−1p_{-1} is omitted at n=0n=0. This three-term recurrence replaces a degree-nn polynomial calculation with a fixed number of operations per step.

The absence of lower-degree terms follows directly from orthogonality. If k≤n−2k\leq n-2, then

⟨xpn,pk⟩μ=⟨pn,xpk⟩μ=0,\langle xp_n,p_k\rangle_\mu =\langle p_n,xp_k\rangle_\mu=0,

because xpkxp_k has degree at most n−1n-1. Only degrees n+1,n,n−1n+1,n,n-1 can remain. The coefficients have the integral expressions

bn=∫xpn(x)2 dμ(x),an=∫xpn(x)pn−1(x) dμ(x).b_n=\int xp_n(x)^2\,d\mu(x),\qquad a_n=\int xp_n(x)p_{n-1}(x)\,d\mu(x).

For a measure symmetric about zero, the polynomials alternate between even and odd parity, and every bnb_n vanishes.

In monic normalization the recurrence becomes

Pn+1(x)=(x−bn)Pn(x)−βnPn−1(x),βn=hnhn−1>0.P_{n+1}(x)=(x-b_n)P_n(x)-\beta_nP_{n-1}(x), \qquad \beta_n=\frac{h_n}{h_{n-1}}>0.

Favard’s theorem gives a converse: a real polynomial sequence satisfying such a recurrence with positive βn\beta_n is orthogonal for some positive measure with infinite support and finite moments. The theorem guarantees existence, but not necessarily uniqueness, of that measure.

The recurrence can be represented by a symmetric tridiagonal Jacobi matrix:

J=(b0a10⋯a1b1a2⋯0a2b2⋱⋮⋮⋱⋱).J= \begin{pmatrix} b_0&a_1&0&\cdots\\ a_1&b_1&a_2&\cdots\\ 0&a_2&b_2&\ddots\\ \vdots&\vdots&\ddots&\ddots \end{pmatrix}.

In the polynomial basis this is the matrix of multiplication by xx. Its leading n×nn\times n block JnJ_n satisfies

det⁡(xI−Jn)=Pn(x).\det(xI-J_n)=P_n(x).

Thus polynomial zeros are eigenvalues of finite symmetric tridiagonal matrices.

For a compactly supported measure, multiplication by xx is a bounded self-adjoint operator on the relevant function space, and the spectral theorem provides a direct operator interpretation of the measure. On unbounded supports, domains and self-adjoint extensions require greater care. In particular, a formally symmetric infinite Jacobi matrix need not determine a unique self-adjoint operator. This distinction is closely related to whether a moment sequence determines a unique measure.

Classical continuous families

The classical continuous families are distinguished by a combination of orthogonality, second-order differential equations, and differentiation identities. Their weights satisfy a Pearson equation

(σw)′=τw,(\sigma w)'=\tau w,

where σ\sigma has degree at most two and τ\tau has degree at most one. With suitable endpoint conditions, the differential operator

Lf=1wddx(σwf′)=σf′′+τf′Lf=\frac{1}{w}\frac{d}{dx}\left(\sigma w f'\right) =\sigma f''+\tau f'

is symmetric on polynomials. Its polynomial eigenfunctions satisfy

σy′′+τy′+λny=0,λn=−nτ′−n(n−1)2σ′′.\sigma y''+\tau y'+\lambda_n y=0, \qquad \lambda_n=-n\tau'-\frac{n(n-1)}2\sigma''.

The endpoint conditions are essential: they remove boundary terms when integration by parts is used to prove orthogonality.

Jacobi polynomials Pn(α,β)P_n^{(\alpha,\beta)} are orthogonal on [−1,1][-1,1] with

w(x)=(1−x)α(1+x)β,α,β>−1.w(x)=(1-x)^\alpha(1+x)^\beta,\qquad \alpha,\beta>-1.

They satisfy

(1−x2)y′′+[β−α−(α+β+2)x]y′+n(n+α+β+1)y=0.(1-x^2)y'' +\bigl[\beta-\alpha-(\alpha+\beta+2)x\bigr]y' +n(n+\alpha+\beta+1)y=0.

The parameter restrictions make the weight integrable at both endpoints. Legendre polynomials are the case α=β=0\alpha=\beta=0, conventionally normalized by Pn(1)=1P_n(1)=1, with

∫−11Pn(x)Pm(x) dx=22n+1δnm.\int_{-1}^{1}P_n(x)P_m(x)\,dx =\frac{2}{2n+1}\delta_{nm}.

Their first terms are 1,x,(3x2−1)/21,x,(3x^2-1)/2, and (5x3−3x)/2(5x^3-3x)/2.

Chebyshev polynomials of the first kind are characterized by

Tn(cos⁡θ)=cos⁡(nθ).T_n(\cos\theta)=\cos(n\theta).

They are orthogonal for w(x)=(1−x2)−1/2w(x)=(1-x^2)^{-1/2}. Those of the second kind satisfy

Un(cos⁡θ)=sin⁡((n+1)θ)sin⁡θU_n(\cos\theta)=\frac{\sin((n+1)\theta)}{\sin\theta}

and use the weight (1−x2)1/2(1-x^2)^{1/2}. Both are proportional to special cases of Jacobi polynomials. Their trigonometric representations account for their explicit zeros and their close connection with cosine transforms.

Generalized Laguerre polynomials Ln(α)L_n^{(\alpha)} use the half-line weight

w(x)=xαe−x,x>0,α>−1,w(x)=x^\alpha e^{-x},\qquad x>0,\quad \alpha>-1,

and satisfy

xy′′+(α+1−x)y′+ny=0.xy''+(\alpha+1-x)y'+ny=0.

Their squared norms in the conventional normalization are

∫0∞xαe−xLn(α)(x)Lm(α)(x) dx=Γ(n+α+1)n!δnm,\int_0^\infty x^\alpha e^{-x} L_n^{(\alpha)}(x)L_m^{(\alpha)}(x)\,dx =\frac{\Gamma(n+\alpha+1)}{n!}\delta_{nm},

where Γ\Gamma is the gamma function.

Hermite polynomials occur in two common conventions. The physicists’ polynomials HnH_n are orthogonal for e−x2e^{-x^2} and satisfy

Hn′′−2xHn′+2nHn=0.H_n''-2xH_n'+2nH_n=0.

The probabilists’ polynomials He⁡n\operatorname{He}_n are orthogonal for e−x2/2e^{-x^2/2}, with

Hn(x)=2n/2He⁡n(2 x).H_n(x)=2^{n/2}\operatorname{He}_n(\sqrt2\,x).

For example, H2(x)=4x2−2H_2(x)=4x^2-2, whereas He⁡2(x)=x2−1\operatorname{He}_2(x)=x^2-1. Confusing these conventions changes both recurrence coefficients and quadrature scaling. Classical families also admit Rodrigues formulas; for instance,

Pn(x)=12nn!dndxn(x2−1)n.P_n(x)=\frac{1}{2^n n!} \frac{d^n}{dx^n}(x^2-1)^n.

Such formulas provide explicit constructions, but repeatedly differentiating them is rarely the best computational procedure.

Zeros, reproducing kernels, and asymptotic behavior

A degree-nn orthogonal polynomial for a positive real measure has nn distinct real zeros, all lying in the interior of the convex hull of the support. For a measure supported on an interval, every zero is therefore strictly between the extreme support points. If the support has gaps, some zeros may lie in those gaps; the theorem does not require every zero to belong to the support itself.

A sign-change argument explains the result. Form the product qq of the factors corresponding to sign changes of pnp_n inside the support’s convex hull. If there were fewer than nn such changes, then deg⁡q<n\deg q<n, while pnqp_nq would have one sign on the support and would not vanish almost everywhere. Its integral could not be zero, contradicting orthogonality. This excludes repeated zeros, nonreal zeros, and zeros outside the permitted interval.

Zeros of consecutive degrees strictly interlace. Writing the zeros of pnp_n as x1,n<⋯<xn,nx_{1,n}<\cdots<x_{n,n}, the pattern is

x1,n+1<x1,n<x2,n+1<⋯<xn,n<xn+1,n+1.x_{1,n+1}<x_{1,n}<x_{2,n+1}<\cdots <x_{n,n}<x_{n+1,n+1}.

This can be proved from the recurrence or from eigenvalue interlacing for the leading blocks of the Jacobi matrix. Interlacing supplies effective root brackets and helps explain the regular distribution of quadrature nodes.

For an orthonormal sequence, define the kernel

Kn(x,y)=∑k=0npk(x)pk(y).K_n(x,y)=\sum_{k=0}^{n}p_k(x)p_k(y).

It reproduces every polynomial qq of degree at most nn:

q(x)=∫Kn(x,y)q(y) dμ(y).q(x)=\int K_n(x,y)q(y)\,d\mu(y).

The Christoffel–Darboux formula compresses this sum to

Kn(x,y)=an+1pn+1(x)pn(y)−pn(x)pn+1(y)x−y.K_n(x,y)=a_{n+1} \frac{p_{n+1}(x)p_n(y)-p_n(x)p_{n+1}(y)}{x-y}.

At x=yx=y, the quotient is interpreted by taking a limit. The formula follows by telescoping the three-term recurrences, and it is central to estimates for expansions and zeros.

The reciprocal diagonal,

λn(x)=1Kn(x,x),\lambda_n(x)=\frac{1}{K_n(x,x)},

is the Christoffel function. It has a variational interpretation:

λn(x)=min⁡deg⁡q≤nq(x)=1∫∣q(t)∣2 dμ(t).\lambda_n(x) =\min_{\substack{\deg q\leq n\\q(x)=1}} \int |q(t)|^2\,d\mu(t).

It measures how cheaply a polynomial can attain unit value at a specified point while remaining small in weighted mean square.

Large-degree behavior depends on both the support and the weight. For Jacobi weights, the normalized zero-counting measures converge to the arcsine distribution,

dxπ1−x2,−1<x<1,\frac{dx}{\pi\sqrt{1-x^2}},\qquad -1<x<1,

independently of the fixed admissible parameters α,β\alpha,\beta. On unbounded intervals, zeros must instead be rescaled: Hermite zeros have scale n\sqrt n, while Laguerre zeros have scale nn. Precise asymptotic formulas distinguish interior oscillations, endpoint behavior, and regions outside the zeros.

Gaussian quadrature

Orthogonal polynomials determine Gaussian quadrature, which approximates a weighted integral using function values:

∫f(x) dμ(x)≈∑i=1nωif(xi).\int f(x)\,d\mu(x) \approx\sum_{i=1}^{n}\omega_i f(x_i).

The Gaussian nodes xix_i are the zeros of the degree-nn orthogonal polynomial. The weights ωi\omega_i are chosen so that the formula integrates every polynomial of degree at most n−1n-1 exactly. Orthogonality then raises the degree of exactness to 2n−12n-1.

To see why, divide any polynomial ff of degree at most 2n−12n-1 by PnP_n:

f=Pnq+r,deg⁡q,deg⁡r≤n−1.f=P_nq+r,\qquad \deg q,\deg r\leq n-1.

The integral of PnqP_nq is zero. At every Gaussian node, PnqP_nq also vanishes, so both the integral and the quadrature reduce to the remainder rr. Exactness for rr follows from ordinary polynomial interpolation. The extra accuracy is therefore a direct consequence of orthogonality rather than a fortuitous choice of nodes.

All Gaussian weights are positive. If ℓi\ell_i is the degree-(n−1)(n-1) cardinal polynomial equal to 11 at xix_i and 00 at the other nodes, exactness gives

ωi=∫ℓi(x)2 dμ(x)>0.\omega_i=\int \ell_i(x)^2\,d\mu(x)>0.

Equivalently, in orthonormal notation,

ωi=1Kn−1(xi,xi).\omega_i=\frac{1}{K_{n-1}(x_i,x_i)}.

Positivity prevents cancellation between weights and makes Gaussian rules especially useful for positive measures.

The degree 2n−12n-1 is optimal for an nn-node rule with freely chosen real nodes. If exactness extended through degree 2n2n, applying the rule to

r(x)=∏i=1n(x−xi)2r(x)=\prod_{i=1}^{n}(x-x_i)^2

would give zero, although its integral is strictly positive for an infinite-support positive measure. Under suitable smoothness assumptions on a compact interval, a sharper error statement is

∫f dμ−∑i=1nωif(xi)=f(2n)(ξ)(2n)!∫Pn(x)2 dμ(x)\int f\,d\mu-\sum_{i=1}^{n}\omega_i f(x_i) =\frac{f^{(2n)}(\xi)}{(2n)!}\int P_n(x)^2\,d\mu(x)

for some point ξ\xi in the interval.

The standard named rules follow their weights: Gauss–Legendre for unit weight on a finite interval, Gauss–Jacobi for algebraic endpoint factors, Gauss–Laguerre for exponentially decaying half-line integrals, and Gauss–Hermite for Gaussian weights. A change of variables adapts them to other scales and locations.

The Golub–Welsch algorithm, published in 1969, computes the nodes as eigenvalues of JnJ_n. If an eigenvector associated with xix_i has unit Euclidean norm and first component vi,1v_{i,1}, then

ωi=m0∣vi,1∣2.\omega_i=m_0|v_{i,1}|^2.

This transforms quadrature construction into symmetric tridiagonal eigenvalue computation. Related Gauss–Radau and Gauss–Lobatto formulas prescribe one or both interval endpoints, sacrificing some degree of exactness to accommodate those constraints.

Expansions, approximation, and numerical computation

An orthonormal system provides polynomial projections of functions. For ff in the space L2(μ)L^2(\mu), define

cn=∫f(x)pn(x) dμ(x),SNf=∑n=0Ncnpn.c_n=\int f(x)p_n(x)\,d\mu(x),\qquad S_Nf=\sum_{n=0}^{N}c_np_n.

The function SNfS_Nf uniquely minimizes ∥f−q∥L2(μ)\|f-q\|_{L^2(\mu)} among polynomials of degree at most NN. Orthogonality yields

∥f−SNf∥L2(μ)2=∥f∥L2(μ)2−∑n=0N∣cn∣2.\|f-S_Nf\|_{L^2(\mu)}^2 =\|f\|_{L^2(\mu)}^2-\sum_{n=0}^{N}|c_n|^2.

This identity concerns weighted mean-square approximation; it does not by itself imply small maximum error or pointwise convergence.

Whether the projections converge to every L2L^2 function depends on completeness. On a compact real support, continuous functions can be uniformly approximated by polynomials, and polynomials are dense in L2(μ)L^2(\mu). The orthonormal sequence therefore forms an orthonormal basis. For unbounded supports, completeness requires additional hypotheses and cannot be inferred merely from the existence of all moments. Thus the phrase “orthogonal polynomial expansion” must distinguish a formal series from an expansion known to converge to the original function.

Smoothness and complex analyticity govern convergence rates. For Chebyshev series, if ff is analytic inside a Bernstein ellipse with parameter ρ>1\rho>1 and bounded there by MM, its coefficients satisfy bounds proportional to Mρ−nM\rho^{-n}. Such geometric decay explains the efficiency of polynomial approximation for analytic functions. Limited differentiability typically gives algebraic rather than geometric decay. Discontinuities can produce persistent local oscillations analogous to those in Fourier series.

Chebyshev expansions are particularly convenient because

f(cos⁡θ)=∑n≥0cncos⁡(nθ)f(\cos\theta) =\sum_{n\geq0}c_n\cos(n\theta)

converts the polynomial expansion into a cosine expansion. Values sampled at suitable cosine-spaced nodes can be transformed efficiently using algorithms related to the fast Fourier transform. These nodes are much denser near the endpoints than uniformly spaced nodes, reducing the severe endpoint oscillations that can occur in high-degree interpolation at equally spaced points.

For direct evaluation, the three-term recurrence usually avoids the expense and cancellation associated with monomial coefficients. A finite orthogonal expansion can be evaluated by the Clenshaw algorithm, a backward recurrence generalizing Horner’s rule. Its effectiveness depends on the family, evaluation region, scaling, and arithmetic; recurrence formulas are not automatically stable in every setting.

Constructing recurrence coefficients from raw moments is a separate numerical problem. Although moment determinant formulas are exact symbolically, their condition numbers may be enormous. Floating-point arithmetic can then destroy the positive-definite structure on which orthogonality depends. More reliable approaches include carefully implemented Stieltjes procedures, modified moments relative to a known reference family, and Lanczos-type reductions of discretized multiplication operators. Numerical orthogonality must often be monitored rather than assumed.

Applications in probability, differential equations, and spectral algorithms

In probability, orthogonal polynomials organize functions of a random variable according to its distribution. If XX is standard normal, the probabilists’ Hermite polynomials satisfy

E[He⁡m(X)He⁡n(X)]=n! δmn.\mathbb E[\operatorname{He}_m(X)\operatorname{He}_n(X)] =n!\,\delta_{mn}.

Thus He⁡n/n!\operatorname{He}_n/\sqrt{n!} is orthonormal. Under the appropriate completeness conditions, a square-integrable function of XX can be expanded in this basis, and its variance becomes a sum of squared nonconstant coefficients.

This idea extends to polynomial chaos expansions for uncertainty quantification. Independent input variables with suitable marginal distributions lead to product bases: Hermite polynomials for Gaussian inputs, Jacobi polynomials for beta-type inputs, and Laguerre polynomials for gamma-type inputs. For an orthonormal expansion

Y=∑αcαΨα,Y=\sum_{\boldsymbol{\alpha}} c_{\boldsymbol{\alpha}}\Psi_{\boldsymbol{\alpha}},

the mean is c0c_{\boldsymbol0} and the variance is

Var⁡(Y)=∑α≠0∣cα∣2.\operatorname{Var}(Y) =\sum_{\boldsymbol{\alpha}\ne\boldsymbol0} |c_{\boldsymbol{\alpha}}|^2.

The usefulness of the decomposition depends on the input measure and on controlling the number of retained multivariate terms.

Classical polynomials also arise through separation of variables in differential equations. Legendre polynomials describe the axisymmetric angular dependence of solutions to Laplace’s equation in spherical coordinates. In quantum mechanics, the one-dimensional harmonic oscillator has wavefunctions proportional to

e−ξ2/2Hn(ξ),e^{-\xi^2/2}H_n(\xi),

where ξ\xi is a dimensionless position. Associated Laguerre polynomials occur in hydrogenic radial wavefunctions. In these examples, polynomial orthogonality is embedded in the orthogonality of physical eigenfunctions, with additional weight factors supplied by the coordinate system and exponential envelopes.

In spectral methods, a solution of a differential equation is approximated by a global polynomial expansion. Galerkin methods require the residual to be orthogonal to a chosen test space, while collocation methods enforce the equation at selected nodes. Legendre bases are natural for unweighted L2L^2 formulations, and Chebyshev bases are attractive for fast transforms and well-developed interpolation formulas. Boundary conditions may be imposed by modifying basis functions, replacing selected equations, or adding constrained terms.

Orthogonal polynomials also connect iterative numerical linear algebra with approximation on spectra. The Lanczos algorithm produces a tridiagonal matrix from a symmetric matrix and a starting vector. Its recurrence is an orthogonal polynomial recurrence for the discrete spectral measure determined by that vector. This connection explains why Gaussian quadrature can approximate quantities such as vTf(A)vv^\mathsf{T}f(A)v without diagonalizing the full matrix.

A further connection appears in random matrix theory. For unitary invariant ensembles with eigenvalue density proportional to

∏i<j(xi−xj)2∏iw(xi),\prod_{i<j}(x_i-x_j)^2\prod_i w(x_i),

orthogonal polynomials for the weight ww express eigenvalue correlation functions through a determinantal kernel. The Christoffel–Darboux identity then reduces many large-degree questions to asymptotic analysis of only two consecutive polynomials.

Discrete systems, moment problems, and broader generalizations

Orthogonality does not require integration against a continuous density. On a discrete support {xj}\{x_j\}, it takes the form

⟨f,g⟩=∑jwjf(xj)g(xj)‾,wj>0.\langle f,g\rangle =\sum_j w_jf(x_j)\overline{g(x_j)},\qquad w_j>0.

Important discrete families include Charlier polynomials for Poisson weights, Krawtchouk polynomials for binomial weights, and Meixner polynomials for negative-binomial weights. Hahn polynomials provide another major finite-support family. Their role parallels that of the classical continuous families, but differential equations are replaced by second-order difference equations.

Finite support changes the length of the system. For a binomial distribution supported on 0,1,…,N0,1,\ldots,N, there are N+1N+1 mutually orthogonal polynomial functions, which may be represented by degrees 00 through NN. Any higher-degree polynomial function on those nodes reduces to a lower-degree representative. This is a statement about the dimension of the function space, not a failure of the orthogonalization procedure.

A foundational issue is the moment problem: does a sequence {mk}\{m_k\} arise from a positive measure, and if so, is that measure unique? The Hamburger problem concerns measures on the whole real line, while the Stieltjes problem restricts them to the nonnegative half-line. Compactly supported positive measures are determined by their moments. On unbounded supports, different measures can have identical moments and hence identical monic orthogonal polynomials.

Carleman’s condition is a useful sufficient criterion for uniqueness in the Hamburger problem:

∑n=1∞m2n−1/(2n)=∞.\sum_{n=1}^{\infty}m_{2n}^{-1/(2n)}=\infty.

It is not necessary. The lognormal distribution illustrates why finite moments alone are insufficient. If X=eZX=e^Z, where ZZ is standard normal, then

E[Xksin⁡(2πlog⁡X)]=0\mathbb E[X^k\sin(2\pi\log X)]=0

for every nonnegative integer kk. Multiplying the lognormal density by

1+εsin⁡(2πlog⁡x),∣ε∣≤1,1+\varepsilon\sin(2\pi\log x),\qquad |\varepsilon|\leq1,

therefore produces distinct probability measures with the same moments. The same identity supplies a nonzero function orthogonal to every polynomial in the original L2L^2 space, showing directly that polynomial completeness can fail.

Orthogonal polynomials on the unit circle form another major theory. They use 1,z,z2,…1,z,z^2,\ldots with an inner product over ∣z∣=1|z|=1. Because multiplication by zz is unitary rather than self-adjoint, their basic recurrence has a different structure and involves reversed polynomials. Its parameters are Verblunsky coefficients, lying in the open unit disk for a nontrivial infinite system. Real-line three-term recurrences should not be transferred to this setting unchanged.

In several real variables, each total degree contains multiple independent monomials, so there is no single canonical polynomial at every degree. Product measures give simple tensor-product constructions; more general measures lead to vector or block recurrences. Matrix-valued weights similarly produce matrix orthogonal polynomials. Other extensions include multiple orthogonality with respect to several measures, Sobolev inner products containing derivatives, and qq-orthogonal systems governed by multiplicative rather than additive shifts.

Historical development and modern research

Orthogonal polynomial theory developed from several practical and analytical problems rather than from a single definition. Eighteenth-century work on gravitational potentials and spherical geometry led to Legendre polynomials. Their generating function,

(1−2xt+t2)−1/2=∑n=0∞Pn(x)tn,(1-2xt+t^2)^{-1/2} =\sum_{n=0}^{\infty}P_n(x)t^n,

arises naturally when expanding inverse-distance potentials. This origin explains the family’s enduring association with celestial mechanics and potential theory.

In 1814, Carl Friedrich Gauss introduced the principle of choosing quadrature nodes to achieve the highest possible polynomial exactness. The nodes were linked to what are now recognized as orthogonal polynomials. This established one of the subject’s most important connections: integral approximation could be improved by allowing the evaluation points themselves to vary, rather than fixing equally spaced nodes.

During the nineteenth century, Jacobi, Chebyshev, Hermite, and Laguerre developed families associated with special functions, approximation, and differential equations. Chebyshev’s work in the 1850s connected polynomial approximation with practical mechanical problems and with broader questions of optimality. Thomas Stieltjes subsequently developed deep connections among orthogonal polynomials, continued fractions, and moments. Continued fractions for transforms such as

F(z)=∫dμ(x)z−xF(z)=\int\frac{d\mu(x)}{z-x}

encode the same coefficients that appear in the three-term recurrence.

The abstract theory consolidated in the early twentieth century. Salomon Bochner’s 1929 classification identified the classical families under the standard second-order differential-operator assumptions and positive real orthogonality conditions. Favard’s 1935 theorem established the recurrence-based existence principle. Gábor Szegő’s Orthogonal Polynomials, first published in 1939, organized much of the theory and became a central reference for zeros, expansions, and asymptotics. These developments shifted attention from separate named sequences toward structural properties shared by whole classes of measures.

Electronic computation created a further phase. The 1969 Golub–Welsch formulation made Gaussian quadrature a practical application of symmetric eigenvalue algorithms. Later work emphasized stable recurrence-coefficient construction, fast polynomial transforms, high-order differential-equation solvers, and the interaction between orthogonality and finite-precision arithmetic. The relevant distinction is often between a formula that is mathematically exact and an algorithm that preserves its structure numerically.

Modern asymptotic research uses complex-variable methods, logarithmic potential theory, and matrix Riemann–Hilbert problems. A Riemann–Hilbert formulation developed in the early 1990s, together with nonlinear steepest-descent methods, made it possible to derive detailed large-degree formulas for broad classes of weights. Applications include eigenvalue statistics, varying exponential weights, and transitions near singular endpoints or merging intervals. Active problems also concern nonclassical measures, multiple and matrix orthogonality, and the behavior of recurrence coefficients when the support or weight develops increasingly complicated structure.