Exposition

Research

Contact

  • Andrius Kulikauskas
  • ms@ms.lt
  • +370 607 27 665
  • Eičiūnų km, Alytaus raj, Lithuania

Support me!

Thank you!

edit SideBar

I am learning, contemplating and writing up a derivation of the classification of Sheffer polynomials.

Here are steps in the derivation.

Preliminaries

Understand what are orthogonal polynomial sets.

The coefficients of the polynomials are presumably rational numbers.

A polynomial set is a sequence of polynomials {$\{P_n(x):n=0,1,2...\}$} in which the degree of {$P_n(x)=n$} for all {$n$}. Polynomial sets are considered equivalent if they are obtainable from each other by a linear change of the independent variable {$x$} and/or a rescaling of the dependent variable (What is meant by dependent variable?)

A polynomial set is orthogonal with respect to a postive measure {$\textrm{d}\alpha$} if {$\int_{-\infty}^{\infty}P_n(x)P_m(x)\textrm{d}\alpha(x)=h_n\delta_{nm}$} where the moments {$\mu_n$} of {$\textrm{d}\alpha$}, {$\mu_n=\int_{-\infty}^{\infty}x^n\textrm{d}\alpha$}, exist for all {$n$}.

Understand the significance of orthogonal polynomials

How do orthogonal polynomials express the spectral theorem? If A is Hermitian, there exists an orthonormal basis of V consisting of eigenvectors of A. Each eigenvalue is real.

Recurrence relation

This recurrence relation for orthogonal polynomials is derived by expressing {$xP_n(x)$} in terms of the orthogonal polynomials as they form a basis for all polynomials. We observe that {$xP_n(x)$} has degree {$n+1$} and thus the expression cannot have any terms of higher degree. Nor can it have any terms of degree {$k<n-1$} as can be seen by linearizing {$xP_n(x)P_k(x)$}, which must yield {$0$}.

It is meaningful to focus on {$xP_n(x)$} and thus contemplate the following expression:

{$$xP_n(x) = P_{n+1}(x) + A_n P_n(x) + B_n P_{n-1}(x)$$}

Quadratic fetters for exponentials

{$P_n{x}$} does not depend on terms lower than {$n-2$}. This apparently imposes a quadratic growth on what otherwise could be growing exponentially in size and complexity.

We get thus a chain where each polynomial is related to the polynomials directly above and below. This gives a way of thinking about a Dynkin diagram (especially their chains) in terms of raising operators {$x$} and lowering operators {$\frac{\textrm{d}}{\textrm{dx}}$}. A raising operator adds a new cell. A lowering operator removes any one of the existing {$n$} cells, thus marking it.

Weights and moments

Given a weight, it is possible to calculate its moments. Then the orthogonal polynomials associated to the weight can be calculated by a determinant of the moments.

This suggests why the combinatorics of orthogonal polynomials are based on permutations and their statistics. I should investigate that.

It also shows that the problem of classifying orthogonal polynomials is really the matter of classifying the weights by which the originate. The weight is the relevant space.

I wonder what is the combinatorial interpretation of the moments?

Sheffer Polynomials

I want to classify orthogonal polynomial sets {$\{P_n(x)\}$} of the form

{$$\sum_{n=0}^{\infty}P_n(x)t^n=A(t)e^{xu(t)}$$}

where {$A(t)=\sum_{n=0}^{\infty}a_nt^n$}, {$a_0=1$}, and {$u(t)=\sum_{n=0}^{\infty}u_nt^n$}, {$u_0=0$}, {$u_1=1$}.

The generating function {$e^{xu(t)}$} is precisely that for polynomials of binomial type. The coefficients are precisely those given by Bell polynomials on scalars {$a_1, a_2, \cdots$} so that {$p_n(x)=\sum_{k=1}^{n=\infty}B_{n,k}(a_1,\cdots,a_{n-k+1})x^k$}. Furthermore, if we set {$P(t)=\sum_{n=1}^{\infty}\frac{a_n}{n!}t^n$}, then {$P^{-1}(\frac{\textrm{d}}{\textrm{dx}})$} is the delta operator of this sequence.

The form {$A(t)e^{xu(t)}$} is consequential in that it establishes an additional constraint above and beyond orthogonality which, working together, the two constraints yield very narrow possibilities. This form is remarkably constricting in how it links together the various powers of {$x$} and {$t$} when combined with orthogonality. Perhaps this is because orthogonality is a quadratic relationship, and here it is being combined with an exponential relationship. The weights are exponential, they have an analytic symmetry, they are reproduced by differentiation, along with an additional factor, and they overwhelm any polynomial factors as we go out to infinity. The quadratic constraint is a general theme, as with the tridiagonality of the Jacobi operator. I suspect the same theme is behind the restrictiveness of the Dynkin diagrams, which could be understood as chains of dimensions on which raising and lowering operators act. A might exponential horse is fettered by quadratic orthogonal shackles which keep one leg from extending too far from another.

Elements of Lie groups and Lie algebras

How does {$A(t)e^{xu(t)}$} relate to Lie groups and Lie algebras?

This suggests that {$xu(t)$} is an element of a Lie algebra. The Lie group is compact if {$A(t)=1$} and not compact otherwise. Thus {$A(t)$} is the factor that expresses noncompactness.

An aside: Classifying the classical orthogonal polynommials

It is interesting to consider the difference between the Sheffer polynomials and the classical orthogonal polynomials.

  • The Hermite polynomials and Laguerre polynomials ar both Sheffer and classical.
  • The Jacobi polynomials are classical but not Sheffer.
  • The Meixner polynomials and Meixner-Pollacek polynomials are Sheffer but not classical.

The classical orthogonal polynomials are A helpful paper is available: Kil H. Kwon, Lance L. Littlejohn. Classification of Classical Orthogonal Polynomials. 1997. Journal of Korean Math Society.

The classical orthogonal polynomials are solutions to the differential equation:

{$$L[y](x) = l_2(x)y''(x) + l_1(x)y' (x) = \lambda_n y(x)$$}

This is part of Sturm-Liouville theory, which deals with equations of the form:

{$$\frac{\textrm{d}}{\textrm{dx}}[p(x)\frac{\textrm{dy}}{\textrm{dx}}] + q(x)y = - \lambda w(x)y$$}

The main result concerns the situation where {$p(x)$} and {$w(x)$} are continuous functions over the finite interval {$[a,b]$} with some boundary requirements. It states that the eigenvalues {$\lambda_n$} are real and distinct, that the eigenfunction {$y_n(x)$} for {$\lambda_n$} has exactly {$n-1$} zeroes in {$(a,b)$}, and the normalized eigenfunctions form an orthonormal basis of the Hilbert space {$L^2([a,b],w(x)dx)$} under the w-weighted inner product.

The time-independent Schroedinger equation is handled by this theory.

The Fivefold Classification

Recurrence relation for Sheffer polynomials

{$\{P_n(x)\}$} is a Sheffer orthogonal polynomial set iff it satisfies the recurrence relation:

{$$P_{n+1}(x)=[x-(ln+f)]P_n(x)-n(kn+c)P_{n-1}(x)$$}

This recurrence relation is determined by {$a_1$}, {$a_2$}, {$h_2$}, {$h_3$}. Namely, {$f=-a_1$}, {$l=-2h_2$}, {$c=a_1^2-2a_2+2a_1h_2$}, {$k=4h_2^2-3h_3$}.

It is meaningful to focus on {$xP_n(x)$} and thus contemplate the following expression:

{$$xP_n(x) = P_{n+1}(x) + (nl + f) P_n(x) + n(nk + c) P_{n-1}(x)$$}

Differential equations

{$\{P_n(x)\}$} is an orthogonal polynomial set iff {$u(t)$} and {$A(t)$} satisfy the differential equations where {$(1-\alpha t)(1-\beta t)=1-m_1t+m_2t^2$} is a polynomial with real coefficients but possibly complex roots:

{$$u'(t)=\frac{1}{(1-\alpha t)(1-\beta t)}=\frac{A'(t)}{\gamma t A(t)}$$}

I suppose {$\gamma$} is real.

I am solving for {$A(t)$} by getting the recursion relation yielded by the Taylor expansion and studying that.

I will ask at Math Stack Exchange how to solve this Pearson differential equation for {$A(t)$} in the usual way.

I should read Differential equation and learn more about them. I can also study Bachioua Lahcene. On Pearson families of distributions and its applications.

Here the classification is a straightforward analysis of the possibilities for {$u'(t)$}:

polynomials{$\alpha$}, {$\beta$}{$u'(t)$}{$u(t)$}
Hermite{$\alpha=\beta$}, {$\beta=0$}{$1$}{$t$}
Charlier{$\alpha\neq\beta$}, {$\beta=0$}{$\frac{1}{1-\alpha t}$}{$-\frac{1}{\alpha}\textrm{ln}(1-\alpha t)$}
Laguerre{$\alpha=\beta$}, {$\beta\neq 0$}{$\frac{1}{(1-\alpha t)^2}$}{$\frac{t}{(1 - \alpha t)}$}
Meixner{$\alpha \neq \beta$}, both real, nonzero{$\frac{1}{(1-\alpha t)(1-\beta t)}$}{$\frac{1}{\beta - \alpha}\textrm{ln}(1-\alpha t)+\frac{1}{\alpha - \beta}\textrm{ln}(1-\beta t)$}
Meixner-Pollaczek{$\alpha=\overline{\beta}$}, complex conjugates, nonzero{$\frac{1}{1 + |\alpha|^2t^2}$}{$\frac{1}{|\alpha|}\textrm{tan}^{-1}|\alpha|t$}

Note that {$u(0)=0$} determines the constant of integration. For the Laguerre polynomials, it is {$-\frac{1}{\alpha}$}, and in every other case it is {$0$}.

polynomials{$\textrm{ln}A(t)$}A(t){$e^{xu(t)}$}
Hermite{$\frac{1}{2}\frac{\gamma}{\alpha ^2}(\alpha t)^2$}{$(e^{\frac{1}{2}(\alpha t)^2})^{\frac{\gamma}{\alpha^2}}$}{$(e^{\alpha t})^{\frac{x}{\alpha}}$}
Charlier{$\frac{\gamma}{\alpha^2}[\alpha t - \textrm{ln}(1-\alpha t)]$}{$(\frac{e^{\alpha t}}{(1-\alpha t)})^{\frac{\gamma}{\alpha^2}}$}{$(\frac{1}{1-\alpha t})^{\frac{x}{\alpha}}$}
Laguerre{$\frac{\gamma}{\alpha^2}[\textrm{ln}(1-\alpha t)+\frac{\alpha t}{1-\alpha t}]$}{$((1-\alpha t)e^{(\frac{\alpha t}{1 - \alpha t})})^{\frac{\gamma}{\alpha^2}}$}{$(e^{\frac{\alpha t}{(1 - \alpha t)}})^{\frac{x}{\alpha}}$}
Meixner{$\frac{\gamma}{\alpha \beta}[\frac{\alpha}{\alpha - \beta} \textrm{ln}(1-\beta t) - \frac{\beta}{\alpha - \beta} \textrm{ln}(1 - \alpha t)]$}{$(\frac{(1-\beta t)^\alpha}{(1-\alpha t)^\beta})^{\frac{\gamma}{\alpha\beta(\alpha - \beta)}}$}{$(\frac{1-\beta t}{1-\alpha t})^{\frac{x}{\alpha-\beta}}$}
Meixner-Pollaczek{$\frac{\gamma}{2|\alpha|^2}\textrm{ln}(1+|\alpha|^2t^2)$}{$(\sqrt{1+|\alpha|^2t^2})^{\frac{\gamma}{|\alpha|^2}}$}{$(e^{\textrm{tan}^{-1}|\alpha|t})^{\frac{x}{|\alpha|}}$}

Compare with Galiffa & Riston

polynomials{$h_2$}, {$h_3$}kinks {$k$}, links {$l$}
Hermite{$h_2^2=h_3=0$}{$k=0$}, {$l=0$}
Charlier{$h_2^2=\frac{3}{4}h_3$}{$k=0$}, {$l\neq 0$}
Laguerre{$h_2^2=h_3$}{$l^2-4k=0$}
Meixner{$h_2^2>h_3$}{$l^2-4k>0$}
Meixner-Pollaczek{$h_2^2<h_3$}{$l^2-4k<0$}

Note that in the Meixner-Pollaczek case we have {$\alpha=a+bi$} and {$|\alpha|=a^2+b^2$} and we are dealing with a circle, as with {$\textrm{tan} \theta$}. What is the meaning of {$\textrm{tan}^{-1}$}?

Compare with the Pearson system for classifying distributions as solutions to the differential equation

{$$\frac{\textrm{dy}}{\textrm{dx}} = \frac{y(m-x)}{a+bx+cx^2}$$}


Viennot's Combinatorial Interpretation


Relating the five cases

Understand how the various cases relate to each other, for example, in taking limits, or in perturbing degeneracies in ways so that they are nondegenerate.

Further steps

Further steps I can take...

  • I will write to India to ask how to contact Xavier Viennot and who is working on related combinatorics.
  • How can I relate Zeng's interpretation of {$P_n(x)$} with pairs of combinatorial interpretations for {$A(t)$} and {$e^{xu(t)}$}?
  • Learn about the Pearson system of distributions.
  • Explore what happens when we successively integrate {$\textrm{ln} x$}.

Sources








Xavier Viennot overviews the combinatorics of the moments of orthogonal polynomials: Hermite (involutions), Laguerre (permutations), Charlier (partitions), Meixner (ordered partitions), Meixner-Pollaczek (secant = even alternating permutations), Chebyshev I (ways of choosing half), Chebyshev II (Catalan).

Keisti - Įkelti - Istorija - Spausdinti - Naujausi keitimai -
Search:
Šis puslapis paskutinį kartą keistas May 05, 2021, at 10:30 AM