🔢 mathematics

On non-monotonicity of logarithmic energy for random matrices

This paper refutes the conjectured dimensional monotonicity of the quadratically penalized logarithmic energy for mean empirical spectral distributions by constructing finite-energy Wigner and Gaussian-regularized Bernoulli counterexamples.

Original authors: Theodoros Assiotis

Published 2026-07-28

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Theodoros Assiotis

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

Imagine you are a detective trying to solve a mystery about how chaos settles down into order. In the world of mathematics and physics, there is a special kind of "messiness" called entropy. Think of it like a pile of laundry: if you throw clothes on the floor, they are messy; if you fold them neatly, they are organized. Scientists have long believed that as you add more and more items to a system (like adding more socks to the pile), the system naturally becomes more "organized" in a very specific, predictable way. This idea is called monotonicity. It suggests that as you increase the size of a system, a certain measure of its "energy" or "disorder" should always go down, step by step, until it hits a perfect, smooth bottom.

To understand this, imagine a crowd of people at a concert. If you look at just two people, they might be standing in a weird, jagged pattern. If you look at a hundred people, they might start to look like a smooth wave. The "logarithmic energy" is a fancy math tool used to measure how spread out or clumped together these people are. For decades, researchers thought that as you added more people to the crowd (increasing the dimension), this energy would always decrease smoothly, like a ball rolling down a perfectly smooth hill. This seemed like a universal rule for how random things behave, whether they were spinning coins, vibrating atoms, or numbers in a giant spreadsheet.

But what if the hill isn't smooth? What if, just for a moment, the ball has to roll up a tiny bump before it can go down again? That is exactly what this paper discovers. The authors, led by Theodoros Assiotis, decided to test this "smooth hill" theory using giant grids of random numbers, known as random matrices. They didn't just guess; they built specific, concrete examples to see if the rule held true. And they found a surprise: the hill has a bump. The energy doesn't always go down as the system gets bigger. In fact, for certain types of random matrices, the energy actually increases when you go from a small size to a slightly larger one, breaking the long-held belief that the process is always monotonic.

The Story of the Bumpy Hill

The paper is a mathematical detective story that sets out to prove a suspected rule wrong. The rule in question was proposed by other scientists (Chafaï, Dadoun, and Youssef) and suggested that for a wide variety of random systems, a specific "penalized energy" score would always get lower as the system grew larger. Think of this score as a "chaos meter." The hypothesis was that if you have a small, chaotic system and you make it bigger, the chaos meter should always drop, indicating the system is becoming more orderly and stable.

The authors of this paper, however, suspected that this smooth decline wasn't the whole story. To test this, they constructed two very specific "counterexamples"—mathematical scenarios designed to break the rule.

The First Counterexample: The Wigner Matrix
First, they looked at a type of matrix called a "Wigner matrix," which is like a giant grid where the numbers are chosen randomly but with a specific symmetry (the top-right matches the bottom-left). They imagined a scenario where the numbers in the grid are drawn from a specific distribution (a "semicircle" shape).
They calculated the energy for a tiny $1 \times 1$ grid and then for a slightly larger $2 \times 2$ grid.

The Result: The energy for the $1 \times 1$ grid was at the expected "low" point. But when they calculated it for the $2 \times 2$ grid, the energy went up.
The Proof: They didn't just guess this; they did the math and showed that the increase was strictly positive. They proved that the "chaos meter" jumped higher, meaning the system became less orderly when it got bigger, at least for this specific step. This shattered the idea that the energy always decreases monotonically for these types of matrices.

The Second Counterexample: The Bernoulli Matrix
Next, they tackled an even more complex case: matrices where every single number is chosen independently from a mix of two types of values (like a coin flip between a very small number and a very large number), but then "smoothed out" with a little bit of Gaussian noise (like adding a tiny bit of static to a radio signal).
They compared the energy of a $2 \times 2$ grid against a $3 \times 3$ grid.

The Result: Again, the energy went up. The $3 \times 3$ grid had a higher energy score than the $2 \times 2$ grid.
The Mechanism: The authors explained this using the idea of "collisions." In the smaller grid, the random numbers tended to spread out nicely. But in the $3 \times 3$ grid, the specific way the numbers interacted caused them to "bunch up" or "collide" in a way that created a temporary spike in energy. They calculated the exact probability of these bunching events and showed that the math guaranteed the energy would rise for small values of the smoothing parameter.

Why This Matters

The paper doesn't just say "the rule is broken"; it provides the exact mathematical proof that the rule is broken for these specific cases. The authors are very careful to say that this doesn't mean the rule is broken everywhere. It might still be true for very large systems or under different conditions. But it proves that the "smooth hill" theory is not a universal law for all random matrices.

The key takeaway is that nature (or at least, the math of random numbers) is more unpredictable than we thought. Sometimes, when you add more pieces to the puzzle, the picture doesn't get clearer immediately; it might get a little messier first. The authors have shown that the path to order isn't always a straight, downward slope. It can have bumps, and sometimes, you have to climb a little bit before you can slide down again.

In the end, the paper is a triumph of precision. The authors didn't rely on computer simulations that might have errors; they built rigorous mathematical arguments to show that for these specific setups, the energy must increase. They have successfully constructed a "finite-energy Wigner counterexample" and a "Gaussian-regularised Bernoulli family" that serve as undeniable proof that the conjecture of dimensional monotonicity is false in its general form. The smooth hill has a bump, and the ball has to roll up it.

Technical Summary: Non-Monotonicity of Logarithmic Energy for Random Matrices

Problem Statement
This paper addresses a conjecture regarding the dimensional monotonicity of quadratically penalised logarithmic energy for mean empirical spectral distributions. Specifically, it investigates whether the penalised energy functionals $F_R$ (for Wigner matrices) and $F_C$ (for matrices with independent and identically distributed entries) decrease monotonically as the matrix dimension $n$ increases, converging to the energy of their respective equilibrium measures (the semicircle law $\sigma$ and the circular law $\omega$ ).

The conjecture, formulated by Chafaï, Dadoun, and Youssef [5], posits that for a sequence of $n \times n$ matrices $M_n$ , the quantity $F(E[\mu_{M_n/\sqrt{n}}])$ should be non-increasing in $n$ and converge to the equilibrium value. While this monotonicity was proven for specific ensembles (Gaussian Unitary, Complex Ginibre, and Square Laguerre Unitary) via explicit computation, the general validity for arbitrary Wigner matrices or general i.i.d. entry matrices remained an open question.

Methodology
The author constructs explicit counterexamples to disprove the general monotonicity conjecture. The methodology relies on two distinct constructions:

Wigner Counterexample: A real symmetric Wigner sequence is constructed where the entry distribution is the standard semicircle law $\sigma$ itself. The paper analyzes the mean spectral distribution $\rho_n = E[\mu_{M_n/\sqrt{n}}]$ for small dimensions ( $n=1$ and $n=2$ ).
- The analysis utilizes a decomposition of the penalised energy $F_R(\mu) - F_R(\sigma)$ derived in Proposition 2.2. This decomposition expresses the energy difference as the sum of a "semicircle gap" term involving the potential $v_{sc}$ outside the support $[-2, 2]$ and a non-negative Fourier term.
- By showing that $\rho_2$ places positive mass outside the support of the semicircle law (specifically in $(2, \infty)$ ), the author demonstrates that the energy gap term is strictly positive, leading to $F_R(\rho_2) > F_R(\rho_1)$ .
i.i.d. Counterexample (Gaussian-Regularised Bernoulli): A one-parameter family of matrices with i.i.d. entries is constructed. The entries are defined as $Y_\varepsilon = (X + \varepsilon G)/\sqrt{1+\varepsilon^2}$ , where $X$ is a discrete Bernoulli variable taking values $-1/\sqrt{31}$ and $\sqrt{31}$ with specific probabilities, and $G$ is a standard complex Gaussian.
- The parameter $\varepsilon$ acts as a regularisation to ensure the existence of a bounded density and finite logarithmic energy.
- The analysis focuses on the asymptotic behavior of the penalised energy $F_C(\nu_{n,\varepsilon})$ as $\varepsilon \to 0$ .
- The author employs Lemma 2.3 and Lemma 2.4 to bound the logarithmic energy of the regularised measures from above and below using the "collision mass" (the sum of squared point masses) of the limiting unregularised atomic measures $\nu^0_n$ .
- Specifically, the paper computes the exact collision mass $C_2$ for the dimension $n=2$ case and establishes a lower bound $L_3$ for the dimension $n=3$ case based on the spectral clustering of zero eigenvalues in the unregularised Bernoulli matrices.

Key Results
The paper establishes two main theorems:

Theorem 1.1 (Wigner Case): For the Wigner sequence with semicircle-distributed entries, the mean spectral distribution $\rho_1$ is the semicircle law $\sigma$ , while $\rho_2$ has a strictly larger penalised energy. Specifically, $F_R(\rho_2) - F_R(\rho_1) \geq \frac{p_0^3}{2}(\frac{3}{4} - \log 2) > 0$ , where $p_0 = \sigma((3/2, 2))$ . This proves that monotonicity fails even for Wigner matrices with compactly supported, bounded densities.
Theorem 1.2 (i.i.d. Case): For the Gaussian-regularised Bernoulli family with specific parameters ( $p=1/32, q=31/32$ ), there exists a threshold $\varepsilon_0 > 0$ such that for all $0 < \varepsilon < \varepsilon_0$ , the penalised energy in dimension 3 is strictly greater than in dimension 2 ( $F_C(\nu_{3,\varepsilon}) > F_C(\nu_{2,\varepsilon})$ ).
- This result is derived by showing that the asymptotic coefficient of the energy growth as $\varepsilon \to 0$ for $n=3$ (denoted $L_3$ ) is strictly larger than the corresponding coefficient for $n=2$ (denoted $C_2$ ).
- The calculation yields $L_3 - C_2 > 0$ , confirming the non-monotonicity.

Significance and Claims
The paper claims to definitively show that the proposed dimensional monotonicity of quadratically penalised logarithmic energy does not hold for general Wigner matrices or general matrices with i.i.d. entries. The author notes that while the monotonicity holds for specific Gaussian-based ensembles, it is not a universal property of random matrix limits.

The work suggests that the question of whether such monotonicity holds under more restrictive assumptions or only after a sufficiently large dimension threshold remains an interesting open problem. The paper does not propose new applications or future experimental directions but focuses strictly on the theoretical disproof of the conjecture for the general classes of matrices defined.

Technical Tools
The proofs rely on:

Variational characterisations of logarithmic energy and large deviation rate functionals.
Explicit computation of Stieltjes transforms and potential theory for the semicircle law.
Analysis of collision masses for atomic measures and their Gaussian convolutions.
Perturbation theory (Rouché's theorem and Riesz projections) to track eigenvalue branches in the regularised Bernoulli matrices as the noise parameter $\varepsilon$ vanishes.
Precise combinatorial enumeration of eigenvalue configurations for small matrix dimensions ( $n=2, 3$ ).

The Story of the Bumpy Hill

Why This Matters

More like this