Small eigenvalues of large Hankel matrices:The indeterminate case

Christian Berg, Yang Chen, Mourad E. H. Ismail

TL;DR

This paper provides a definitive link between the "determinacy" of a moment problem and the smallest eigenvalue of its associated Hankel matrix. The authors prove that a moment problem is indeterminate (meaning multiple measures share the same moments) if and only if the smallest eigenvalue of the large Hankel matrix remains bounded above zero, providing an explicit lower bound for these eigenvalues.

What Problem Does This Paper Address?

In classical analysis, the moment problem asks: given a sequence of numbers $s_n$, can we find a positive measure $\alpha$ such that $s_n = \int x^n d\alpha(x)$?

There are two flavors of this problem:

  1. Determinate: Only one unique measure $\alpha$ fits the moments.
  2. Indeterminate: Infinitely many different measures share the exact same moments.

Mathematically, these moments are often organized into a Hankel matrix $H_N$, where the entry at row $j$ and column $k$ is $s_{j+k}$. Previous research by giants like Szegö and Widom showed that for determinate problems, the smallest eigenvalue ($\lambda_N$) of this matrix shrinks to zero as the matrix size $N$ grows. The "gap" addressed here is the indeterminate case: what happens to the eigenvalues when the measure isn't unique? The authors set out to prove that in these cases, the eigenvalues don't just vanish—they hit a "floor" and stay there.

Methods

The authors employ a mix of classical matrix analysis and the theory of orthogonal polynomials.

  • Rayleigh Quotients: They use the Rayleigh quotient to characterize the smallest eigenvalue $\lambda_N$ as a minimization problem.
  • Orthonormal Polynomials: By expanding polynomials in terms of an orthonormal system $\{p_k\}$ associated with the measure $\alpha$, they relate the reciprocal of the eigenvalue ($1/\lambda_N$) to the sum of the squares of these polynomials.
  • Nevanlinna Parametrization: For indeterminate problems, the solutions can be described using a $2 \times 2$ matrix of entire functions (the Nevanlinna matrix). The authors use these functions ($B$ and $D$) to calculate the "radius of the Weyl circle," which provides the geometric intuition for the eigenvalue's lower bound.
  • Numerical Extrapolation: To validate their theoretical lower bounds, they perform numerical computations on specific polynomial weights (like the Stieltjes-Wigert weight) to see how closely their predicted "floor" matches actual matrix behavior.

Datasets

As a theoretical mathematics paper, this work does not use external empirical datasets. Instead, it evaluates its theorems against several well-known "test cases" of indeterminate moment problems:

  • Stieltjes-Wigert Polynomials: Related to log-normal distributions.
  • Al-Salam-Carlitz Polynomials: A staple in $q$-series and basic hypergeometric functions.
  • Freud-like weights: Measures involving exponential decays like $\exp(-|x|^{1/2})$.
  • $q^{-1}$-Hermite Polynomials: A $q$-analog of the classic Hermite polynomials.

Key Insights and Ideas

  • The Determinacy Litmus Test: The most significant "aha moment" is Theorem 1.1: A moment problem is determinate if and only if $\lim_{N \to \infty} \lambda_N = 0$. This turns an abstract property of measures into a concrete spectral property of matrices.
  • The Spectral Floor: In the indeterminate case, the eigenvalues are bounded below by the harmonic mean of a function $\rho$ along the unit circle. This $\rho$ is essentially the reciprocal of the sum of squares of all orthonormal polynomials at a given point.
  • Independence from Shifts: Unlike previous criteria (like Hamburger’s theorem), this new characterization only requires looking at the original moment sequence, not shifted versions (like $s_{n+2}$), making it more elegant and direct.
  • Trace Class Connection: The authors reformulate the problem in terms of operator theory, noting that the infinite Hankel operator is "trace class" if and only if a specific sum of orthonormal polynomial coefficients converges.

Key Results

  • Theorem 1.1: Established the definitive link: $\lambda_N \to 0$ is the signature of a unique (determinate) measure.
  • Theorem 1.2: Provided an explicit lower bound for $\lambda_N$ in indeterminate cases.
  • Numerical Accuracy: For the Stieltjes-Wigert case with $q=1/2$, the authors found the actual smallest eigenvalue $s \approx 0.3605$, while their theoretical lower bound $l \approx 0.3435$. The error was small (under 5% for most values of $q$), proving that their theoretical "floor" is a very tight estimate of the actual spectral behavior.
  • Closed-form Expressions: The paper successfully derived complex series expressions for the eigenvalue bounds for the Al-Salam-Carlitz and $q^{-1}$-Hermite cases, which involve theta functions and basic hypergeometric series.

Limitations and Future Work

  • Tightness of Bounds: While the lower bound is explicit, it is not always the exact limit. The authors note a small percentage error between their bound and the numerical limit, suggesting room for an even more refined asymptotic theory.
  • Computational Complexity: For some cases (like Example 4), the final expression for the bound is a "5-fold sum." While mathematically rigorous, these are computationally heavy to evaluate and lack the simple elegance of the Stieltjes-Wigert result.
  • Generalization: The paper focuses on measures on the real line ($\mathbb{R}$). Extending these spectral characterizations to multidimensional moment problems remains an open challenge.

Summary updated

Abstract

In this paper we characterise the indeterminate case by the eigenvalues of the Hankel matrices being bounded below by a strictly positive constant. An explicit lower bound is given in terms of the orthonormal polynomials and we find expresions for this lower bound in a number of indeterminate moment problems.
BibTeX
@inproceedings{berg2002small,
  author = {Christian Berg and Yang Chen and Mourad E. H. Ismail},
  title = {{Small eigenvalues of large Hankel matrices:The indeterminate case}},
  booktitle = {MATHEMATICA},
  year = {2002},
  doi = {10.7146/math.scand.a-14379},
}

Loading related content...

Figure