Asymptotic Expansions of Gaussian and Laguerre Ensembles at the Soft Edge III: Generating Functions

This paper concludes a series on asymptotic expansions at the soft edge for Gaussian and Laguerre ensembles by deriving gap-probability generating functions, demonstrating that correction terms follow a universal multilinear structure with rational coefficients independent of the generating variable, and validating these results for orthogonal ensembles through extensive numerical simulations.

The Gist

The Big Picture: Zooming in on the Edge of Chaos

Imagine you have a giant crowd of people (representing the "levels" or eigenvalues in a random matrix). In mathematics, we often study how these crowds behave when they get very large.

Most of the time, we look at the middle of the crowd, where things are predictable and calm. But this paper focuses on the edge of the crowd—specifically, the very last person standing at the "soft edge." This is the person with the highest value. In the world of random matrices, this edge is where things get wild, unpredictable, and mathematically fascinating.

The author, Folkmar Bornemann, is the third in a series of papers trying to understand exactly how this edge behaves as the crowd size ($n$) grows toward infinity.

The Main Tool: The "Magic Remote Control"

To understand the crowd, the paper uses a special mathematical tool called a Generating Function. Think of this as a Magic Remote Control for the crowd.

• The Button ($\xi$): The remote has a dial or button labeled $\xi$ (xi).
• The Effect: When you turn this dial, it doesn't just count the people; it changes the rules of the game.
  • If you set it to 0, it tells you the average number of people at the edge.
  • If you set it to 1, it tells you the probability that the edge is empty (a "gap").
  • If you set it to other numbers, it tells you the probability of having exactly 1, 2, or 3 people at the edge.

The paper's goal is to figure out the exact formula for this remote control as the crowd gets infinitely large.

The Discovery: A Universal Recipe

The paper's main discovery is that this "Magic Remote Control" follows a very specific, neat pattern as the crowd grows.

Imagine you are baking a cake (the main result).

1. The Base Cake: There is a perfect, standard cake that represents the main behavior. In math terms, this is the "leading-order term."
2. The Frosting and Sprinkles: As the crowd gets bigger, the cake isn't quite perfect yet. You need to add corrections (frosting, sprinkles) to make it accurate.

The paper proves that for the Unitary Ensembles (a specific type of random matrix, like a perfectly balanced deck of cards), these corrections follow a strict recipe:

• The corrections are not random. They are built by taking the Base Cake and applying a specific set of multipliers to its "flavors" (mathematical derivatives).
• These multipliers are like pre-made spice mixes. They are fixed recipes (polynomials) that depend only on the size of the crowd and the type of matrix, not on which button ($\xi$) you pressed on the remote.

The Analogy:
Think of the "Base Cake" as a song. The "corrections" are like adding harmonies. The paper shows that no matter what song you start with, the harmonies are always added using the same set of musical rules (the polynomial coefficients). You don't need to invent new rules for every new song; you just apply the same rulebook.

The "Linearly Induced" Family

The paper points out that this recipe is so powerful that it applies to any question you can ask about the crowd, as long as you ask it in a "linear" way.

• Question A: "What is the chance the highest level is below $X$?"
• Question B: "What is the chance the second highest level is below $X$?"
• Question C: "What is the chance the tenth highest level is below $X$?"

Because the "Magic Remote Control" contains all the answers, and because the corrections follow that strict recipe, all these different questions get the same type of correction. If you know how to fix the answer for the highest level, you automatically know how to fix the answer for the 10th highest level. You just use the same spice mix on a different part of the cake.

The Mystery of the Other Crowds (Orthogonal and Symplectic)

The paper handles three types of crowds:

1. Unitary ($\beta=2$): The "perfect" crowd. The author proves the recipe works 100% here.
2. Orthogonal ($\beta=1$) and Symplectic ($\beta=4$): These are slightly "messier" crowds (like crowds with different social rules).

For these two messier crowds, the author hypothesizes (guesses with strong reasoning) that the exact same recipe applies.

• The Guess: The corrections for these crowds use the same spice mixes (polynomials) as the perfect crowd, just with a slight twist in how they are applied.
• The Evidence: The author didn't prove it with a rigid mathematical chain yet, but they checked it against computer simulations. They simulated crowds of size 10 and 100, calculated the "10th highest level," and compared it to the recipe. The recipe matched the simulation data perfectly, even when they had to add four layers of "frosting" (correction terms) to get it right.

The "Duality" Surprise

One of the coolest findings is a "mirror effect" between the Orthogonal and Symplectic crowds.

• The paper finds that the "spice mixes" (polynomial coefficients) for the Orthogonal crowd are identical to those for the Symplectic crowd.
• It's as if two different types of crowds, which seem totally different on the surface, are actually wearing the exact same hidden uniform underneath.

Summary

In short, this paper says:

1. We have a "Magic Remote" that controls the statistics of the edge of random crowds.
2. For the most standard crowd, we have a proven formula showing that all corrections are built from the main result using a fixed set of rules.
3. For the other two types of crowds, we strongly suspect the same rules apply.
4. We have tested this suspicion with computers, and it works perfectly, even for very specific, hard-to-predict scenarios.

The paper essentially provides a universal instruction manual for calculating how these random crowds behave at their edges, turning a chaotic problem into a predictable, step-by-step recipe.

Technical Summary

Technical Summary: Asymptotic Expansions of Gaussian and Laguerre Ensembles at the Soft Edge III: Generating Functions

Problem Statement
This paper concludes a series of works by the author on asymptotic expansions at the soft edge for classical $n$-dimensional Gaussian and Laguerre random matrix ensembles (orthogonal $\beta=1$, unitary $\beta=2$, and symplectic $\beta=4$). While previous studies focused on the distribution of the largest level and the level density, this work extends the analysis to the gap-probability generating functions, defined as:
$$E(x; \xi) := \mathbb{E}\left[(1 - \xi)^{N(x, \infty)}\right],$$
where $N(x, \infty)$ is the number of eigenvalues exceeding $x$. The generating function encodes various statistics, including gap probabilities, the distribution of the $k$-th largest level, and the expected number of levels above a threshold. The objective is to derive asymptotic expansions for these generating functions as $n \to \infty$ under soft-edge scaling, specifically identifying the structure of the correction terms beyond the leading-order limit law.

Methodology
The paper employs a combination of rigorous operator-theoretic analysis for the unitary case and structural hypotheses supported by numerical verification for the orthogonal and symplectic cases.

1. Unitary Ensembles ($\beta=2$):

   • The generating function is expressed as a Fredholm determinant involving the correlation kernel $K_n$: $E_{2,n}(x; \xi) = \det(I - \xi K_n)|_{L^2(x, \infty)}$.
   • The proof follows the methodology established in prior works [7, 8] for the largest level distribution. It involves expanding the scaled kernel $K_n$ in powers of a parameter $h_n \asymp n^{-2/3}$.
   • The expansion of the kernel is lifted to the Fredholm determinant using perturbation theory.
   • Crucially, the author utilizes Tracy–Widom theory and the properties of the Painlevé II transcendent $q(s; \xi)$. By representing the logarithmic derivatives of the generating function in terms of $q$ and its derivatives, the author demonstrates that the nonlinear operator-theoretic expressions can be converted into a multilinear form involving derivatives of the leading term. This relies on the algebraic independence of $q$ and $q'$ over the field of rational functions.

2. Orthogonal and Symplectic Ensembles ($\beta=1, 4$):

   • Since a direct proof analogous to the unitary case is not provided, the author relies on structural hypotheses regarding the algebraic relations between the three symmetry classes.
   • The approach utilizes the Forrester–Rains interrelations, which express unitary ensembles as even decimations of superpositions of orthogonal ensembles, and symplectic ensembles as even decimations of orthogonal ensembles.
   • Hypothesis I: Posits that the generating functions for the "even" and "odd" components of the orthogonal ensembles admit asymptotic expansions similar to the unitary case.
   • Hypothesis II: Posits that the correction terms for these components can be represented as multilinear forms of the derivatives of the leading term, with polynomial coefficients independent of the dummy variable $\xi$.
   • By solving the linear systems arising from these interrelations, the author derives the expansion coefficients for $\beta=1$ and $\beta=4$.

Key Contributions and Results

1. Multilinear Structure of Correction Terms:
   The primary result is the demonstration that the asymptotic expansion of the generating function $E_{\beta,n}(x; \xi)$ at the soft edge takes the form:
   $$E_{\beta,n}(x; \xi)|_{x=\mu_{n'}+\sigma_{n'}s} = F_\beta(s; \xi) + \sum_{j=1}^m G_{\beta,j}(s, \tau_{n'}; \xi) h_{n'}^j + O(h_{n'}^{m+1}),$$
   where the correction terms $G_{\beta,j}$ are multilinear forms of the higher-order derivatives of the leading term $F_\beta(s; \xi)$:
   $$G_{\beta,j}(s, \tau; \xi) = \sum_{k=1}^{2j} P_{\beta,j,k}(s, \tau) \cdot \partial_s^k F_\beta(s; \xi).$$
   Here, $P_{\beta,j,k}(s, \tau)$ are rational polynomials in the scaling variable $s$ and the Laguerre parameter $\tau$, which are independent of the dummy variable $\xi$.

2. Inheritance by Linearly Induced Quantities:
   Because the correction terms are multilinear with $\xi$-independent coefficients, any quantity obtained by applying a constant-coefficient linear differential operator to the generating function (e.g., the distribution of the $k$-th largest level, gap probabilities) inherits the exact same expansion structure. The same polynomials $P_{\beta,j,k}$ apply to these derived quantities.

3. Explicit Polynomial Coefficients:
   The paper provides explicit formulas for the polynomials $P_{\beta,j,k}$ up to order $j=4$.

   • For $\beta=2$ (Unitary), the coefficients are derived with proof (Table 1).
   • For $\beta=1$ (Orthogonal) and $\beta=4$ (Symplectic), the coefficients are derived under the stated hypotheses (Table 2).
   • A notable duality is observed: $P_{1,j,k} = P_{4,j,k}$.

4. Validation of Hypotheses:
   For the orthogonal and symplectic cases, the paper does not claim a rigorous proof of the hypotheses but offers substantial evidence:

   • Consistency: The derived expansions for level densities (obtained by differentiating the generating function) match previously proven expansions for $\beta=1, 4$ up to order $m=10$.
   • Algebraic Reconstruction: Using algebraic independence results for the Airy function, the author reconstructs the functional form of the polynomials for $j=1, 2$ from the known density expansions.
   • Numerical Verification: The expansions are tested against simulation data for the $k$-th largest level in orthogonal ensembles. The results show that including up to four correction terms ($m=4$) yields approximations within plotting accuracy for dimensions as small as $n=10$ and $n=100$, even for large $k$ where the leading-order limit law is inaccurate.

Significance and Claims
The paper claims to establish a unified framework for finite-size corrections in random matrix theory at the soft edge. Its significance lies in:

• Generalization: Extending the asymptotic expansion theory from specific statistics (like the largest eigenvalue) to the entire generating function, thereby covering all linearly induced statistics simultaneously.
• Structural Insight: Revealing that the complex correction terms are not arbitrary but possess a rigid multilinear structure governed by a small set of universal polynomial coefficients.
• Practical Utility: Providing explicit coefficients that allow for high-accuracy approximations of finite-$n$ distributions, which is crucial for applications requiring precision beyond the leading-order Tracy–Widom limit.

The author explicitly notes that while the results for $\beta=2$ are proven, the results for $\beta=1$ and $\beta=4$ are conditional on the structural hypotheses, though these are strongly supported by consistency checks and numerical experiments. The paper does not propose new applications or future experimental directions but focuses on the theoretical completion of the asymptotic expansion program for these ensembles.