🔢 mathematics

On Trace Averages and Sharp Asymptotics for Powers of Haar-Unitary Matrices

This paper establishes exact decompositions and sharp asymptotic limit laws for the trace moments and eigenvalue statistics of powers of Haar-distributed unitary matrices, utilizing tools such as Rains's power-map theorem, Charlier-Ewens convolutions, and Fredholm factorizations to characterize their behavior across various regimes of matrix dimension and power.

Original authors: Ayesha Irfan

Published 2026-10-26

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ayesha Irfan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the hidden architecture of complex systems, from the energy levels of heavy atoms to the zeros of prime-number functions, there exists a surprising order. Scientists study these systems by looking at a special kind of mathematical object called a unitary matrix, which can be thought of as a vast, rotating wheel with many spokes. When this wheel is spun randomly according to the strict rules of symmetry, the positions of its spokes follow a pattern known as the Circular Unitary Ensemble. In this pattern, the spokes repel one another, refusing to cluster together, creating a spacing that is far more even than if they were placed by pure chance. This repulsion is a hallmark of quantum chaos and random matrix theory, a field that seeks to find universal laws in the behavior of complex, unpredictable systems.

A natural question arises: what happens if we take this spinning wheel and raise it to a power, effectively spinning it faster or folding its circle over itself multiple times? Intuitively, one might expect this action to scramble the delicate repulsion between the spokes into a chaotic mess. However, a deep mathematical insight revealed that the structure does not collapse; instead, it splits cleanly. The spectrum of the powered matrix breaks apart into several smaller, independent groups of spokes. Within each group, the original repulsion remains, but the groups themselves do not interact. This discovery sets the stage for a new investigation into how these groups behave as the system grows larger and the power increases.

A recent study by Ayesha Irfan takes this splitting phenomenon and measures exactly how close the resulting system comes to behaving like a simple, random collection of points. The research focuses on two specific ways of observing the system. The first is a global view, looking at the sum of all the spokes, which acts as a single number representing the entire matrix. The second is a local view, zooming in on a tiny segment of the circle to count how many spokes land there. The goal is to determine the precise distance between the actual behavior of these powered matrices and two idealized models: a standard bell-curve distribution for the global sum, and a Poisson process, which describes events happening completely independently at random, for the local counts.

The study finds that the answer depends entirely on the relationship between the size of the matrix and the power to which it is raised. When the power is small compared to the size of the matrix, the system behaves in a way that is very close to the bell curve, but with a specific, measurable deviation. The researchers calculated the exact size of this deviation, finding that it is not just a vague approximation but a precise quantity that can be written down. They discovered that this deviation is shaped by a coherent sum of many small corrections, which combine to produce a specific factor involving the number e. This result is significant because previous methods could only estimate the size of the error, whereas this work pinpoints the exact leading term, revealing that the first few moments of the distribution are not enough to determine the full picture; the entire band of corrections must be summed together to get the right answer.

As the power increases to match the size of the matrix, the behavior shifts. The system transitions from having large, repulsive groups to having many tiny, independent groups. In this regime, the global sum still approaches the bell curve, but the rate at which it gets there changes. The study provides a sharp formula for this transition, showing that the error decays at a specific rate determined by the number of these tiny groups. When the power becomes larger than the matrix size, the system reaches a state of complete independence. Here, the spokes are no longer repelling each other at all; they are distributed exactly like independent random points. The study confirms that in this final stage, the distance to the ideal random distribution is extremely small, shrinking at a rate proportional to the inverse of the matrix size, and provides the exact constant that governs this rapid convergence.

On the local scale, where researchers count the number of spokes in a fixed interval, the story is equally precise. As the matrix grows and the power increases, the repulsive groups become so small and numerous that they effectively act as rare, independent sources of points. The study proves that the distribution of these local counts converges to a Poisson process, which is the mathematical model for completely random, independent events. The researchers calculated the exact leading error term for this convergence. They found that the error is determined by the sum of the squares of the sizes of the independent groups. This means that the more uneven the sizes of these groups are, the larger the deviation from perfect randomness. The study also quantifies a subtle "under-dispersion," showing that the local counts are slightly less variable than a purely random process would predict, a residual effect of the original repulsion that has not yet fully vanished.

The paper achieves this clarity by developing new mathematical tools that separate the stable, predictable part of the system from the finite-size corrections. It uses a technique that can be likened to peeling an onion, where each layer reveals a specific correction band. By summing these layers coherently, the author avoids the inaccuracies that come from looking at them in isolation. The results are not simulations or rough estimates; they are rigorous proofs that hold true for any size of the matrix and any power, provided both are large enough. The work bridges the gap between the rigid, repulsive world of quantum mechanics and the fluid, independent world of pure randomness, showing exactly how and when one transforms into the other.

In the end, the research provides a complete map of the landscape. It tells us that whether we are looking at a small power or a massive one, the path to randomness is not a sudden jump but a smooth, calculable journey. The study rules out the idea that the transition is vague or that the error terms are too complex to pin down. Instead, it demonstrates that the deviations from randomness follow strict, predictable laws that can be written down with exact constants. This level of precision allows scientists to understand not just that a system becomes random, but exactly how much it differs from randomness at any given stage of its evolution, offering a deeper understanding of the fundamental mechanisms that govern complex, high-dimensional systems.

Technical Summary: On Trace Averages and Sharp Asymptotics for Powers of Haar-Unitary Matrices

Problem Statement
The paper investigates the statistical behavior of powers of Haar-distributed unitary matrices, $U \in U(N)$ . Specifically, it addresses two distinct regimes of convergence as $N \to \infty$ :

Global Statistics: The distribution of the normalized trace $X_{N,m} = \text{Tr}(U^m)/\sqrt{m \wedge N}$ and its distance to the standard complex Gaussian distribution in total variation ( $d_{TV}$ ).
Local Statistics: The distribution of the unfolded eigenvalue spectrum of $U^m$ on a fixed interval, specifically its distance to a unit-rate Poisson process.

A central motivation is Rains's power-map theorem, which states that the spectrum of $U^m$ decomposes into a superposition of independent, lower-dimensional Circular Unitary Ensemble (CUE) spectra. The paper seeks to quantify how this spectral decomposition translates into precise asymptotic laws for trace moments and point processes, particularly in regimes where the power $m$ scales with $N$ (fixed, growing, or exceeding $N$ ).

Methodology
The author employs a synthesis of representation theory, analytic combinatorics, and random matrix theory techniques:

Exact Trace Identities: The paper utilizes Frobenius's character formula and Schur orthogonality to express mixed trace moments as finite character sums. To handle the finite- $N$ corrections beyond the "stable range" (where moments match Gaussian moments), the author introduces a Charlier–Ewens convolution. This separates the stable Wick term from a finite correction series indexed by "correction bands." Additionally, a displacement determinant formula is derived, expressing mixed moments as a finite sum of determinants of permutation matrices, bypassing the complexity of character sums.
Analytic Tools:
- Bessel and Fredholm Representations: The generating functions for trace moments are identified with Bessel determinants. These are further factorized using the Borodin–Okounkov–Geronimo–Case identity into Fredholm determinants involving Hankel operators.
- Fourier–Laguerre Transfer: To move from characteristic functions to probability densities, the paper uses a transfer mechanism where Laguerre polynomials serve as the basis for expanding the density correction. This allows for the precise calculation of $L^1$ norms required for total variation bounds.
- Poisson–Charlier Expansion: For local statistics, the superposition of independent CUE blocks is approximated by a sum of sparse Bernoulli processes. The error is controlled using Poisson–Charlier expansions and bounds on the total variation distance between point processes.

Key Contributions and Results

1. Exact Mixed Trace Moments and Finite- $N$ Corrections
The paper provides an exact evaluation of mixed trace averages $C_N(\mathbf{p}, \mathbf{q})$ for arbitrary exponents.

Charlier–Ewens Correction: Theorem 6 establishes that the deviation from the Gaussian Wick term is a finite sum of "correction bands." The number of bands is determined by the geometry of the partition of the total degree $D$ relative to $N$ .
Displacement Determinant: Theorem 7 offers a computationally efficient formula for these moments as a sum of determinants, valid for any $N$ .

2. Sharp Global Gaussian Asymptotics
The paper determines the sharp leading-order asymptotics for the total variation distance $d_{TV}(X_{N,m}, Z)$ , where $Z$ is a standard complex Gaussian. The results are uniform across four regimes defined by the relationship between $m$ and $N$ :

Fixed Powers ( $m=p$ ): Theorem 1 provides the exact leading term, which includes a coherent summation factor of $e^{2/p}$ . This factor arises from the coherent summation of Laguerre modes in the first correction band, a detail missed by analyzing only the first non-Gaussian moment.
Joint Growth ( $m \to \infty, N/m \to \infty$ ): Theorem 3 shows that the leading term has the same form as the fixed-power case, with the $e^{2/p}$ factor converging to 1 as $p \to \infty$ .
Fixed Block Dimension ( $N/m \to q$ ): Theorem 4 analyzes the regime where the block dimensions remain bounded. The leading correction involves a linear combination of two adjacent Laguerre modes ( $L_{q+1}$ and $L_{q+2}$ ).
Independent-Phase Regime ( $m \ge N$ ): Theorem 2 proves that when $m \ge N$ , the eigenvalues become independent uniform points. The paper derives an explicit constant for the $O(N^{-1})$ convergence rate, sharpening previous bounds.

3. Sharp Local Poisson Approximation
Theorem 5 addresses the local spectral statistics of $U^m$ on a fixed interval $B$ of length $L$ .

Poisson Limit: As $N, m \to \infty$ , the unfolded spectrum converges to a Poisson process.
Leading Error Term: The paper identifies the exact leading term for the total variation distance between the local eigenvalue count and a Poisson variable. The error is proportional to $\beta_{N,m}$ , the sum of squared block intensities.
Under-dispersion: The first-order correction, derived via the second Charlier polynomial, quantifies the "residual under-dispersion" of local eigenvalue counts, showing that the variance is strictly less than the Poisson variance for large $N, m$ .

Significance and Claims
The paper claims to provide the first sharp asymptotics for these problems, moving beyond order-of-magnitude bounds to explicit constants and precise functional forms.

Precision: Unlike previous works that established super-exponential convergence or $O(N^{-1})$ rates with unspecified constants, this paper derives the exact leading constants (e.g., involving $e^{2/p}$ and specific integrals of Laguerre polynomials).
Unification: The results unify the behavior of trace statistics across all scaling regimes of $m$ and $N$ , demonstrating how the spectral decomposition into independent blocks governs both global Gaussian fluctuations and local Poisson limits.
Mechanism of Correction: A key insight is that the leading total variation constant is not determined solely by the first non-Gaussian moment but by the coherent summation of an entire band of Laguerre corrections. This explains why simple moment-matching arguments fail to capture the precise convergence rate.

The work relies on rigorous analytic estimates (Bessel/Fredholm) and combinatorial identities (Charlier/Ewens) to bridge the gap between exact finite- $N$ identities and sharp infinite- $N$ limits.

Technical Summary: On Trace Averages and Sharp Asymptotics for Powers of Haar-Unitary Matrices

More like this