High-dimensional central limit theorems for eigenvalue distributions of generalized Wishart processes

Jianfeng Yao; Jian Song; Wangjun Yuan

REVIEW 3 major objections 4 minor 17 references

High-dimensional central limit theorems for eigenvalue distributions of generalized Wishart processes

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read For a broad class of matrix diffusions, the eigenvalue histogram's N-scaled centered fluctuations converge to a Gaussian process, with covariance fixed by the limit measure and the local diffusion coefficient G(x,x).

desk verdict A sound but conditional CLT for generalized Wishart eigenvalue fluctuations, with a genuinely new OU application; the main proof holds up, but the paper leans on an unproved convergence hypothesis and omits proofs for the particle-system half. read the letter →

arxiv 1908.03304 v1 pith:3YTV64HJ submitted 2019-08-09 math.PR

Jian Song , Jianfeng Yao , Wangjun Yuan This is my paper · ORCID

classification math.PR MSC 60H1560F0560B20

keywords Dyson'sBrownianmotionWishartprocessgeneralizedsquaredBesselparticlesystemcentrallimittheoremempiricalmeasurefluctuationsOrnstein-Uhlenbeckmatrixrandom

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper establishes central limit theorems for eigenvalue fluctuations of generalized Wishart processes and a wider class of particle systems. It shows that, after subtracting an initial value and a deterministic drift, the N-scaled centered fluctuation process $Q_t^N(f)=N\langle f,L^N(t)-\mu_t\rangle$ converges to a Gaussian process. The limiting covariance is $2\int_0^{t\wedge s}\langle f_i'(x)f_j'(x)G(x,x),\mu_u\rangle\,du$, so the random spread of eigenvalues around their high-dimensional limit has size $1/\sqrt{N}$ and is described entirely by the limit measure $\mu_t$ and the coefficient limit $G(x,x)$. The same argument recovers earlier fluctuation results for Dyson's Brownian motion and the Wishart process under more general initial conditions, and produces a new fluctuation theorem for the eigenvalues of a symmetric Ornstein-Uhlenbeck matrix process.

What carries the argument

The mechanism is Itô's formula applied to linear statistics of the eigenvalues, which splits $\langle f,L^N(t)\rangle$ into an initial value, a martingale $M_f^N(t)$ with quadratic variation $(4/N)\int_0^t\langle (f'g_Nh_N)^2,L^N(s)\rangle\,ds$, and drift terms. The proof shows $Q_t^N(f)-N M_f^N(t)$ tends to zero uniformly in $t$, then reads the Gaussian limit off the limit of the quadratic variation via a martingale central limit theorem (Lemma 4.1). A secondary mechanism is the comparison principle of Section 3.1, which sandwiches a target process between two comparison processes with known moment bounds and thereby extends the CLT to polynomial test functions and to drifts that merely converge to a constant.

What would settle it

Simulate Dyson's Brownian motion (3.25) for large $N$, take $f(x)=x$, and estimate the variance of the centered fluctuation $L_t(x)-L_0(x)$ at $t=1$; the theorem predicts this variance is exactly $2$, so a statistically significant deviation from $2$ would refute the covariance formula. Alternatively, find a coefficient pair satisfying (2.2) for which the limit equation (1.5) has two distinct solutions with the same initial measure; subsequences selecting different limits would then show no single Gaussian limit describing the full sequence.

Watch

Extended reading notes

Core claim

The central discovery is structural: for eigenvalues of the matrix stochastic differential equation (1.1), the fluctuation field $Q_t^N(f)$ has a Gaussian limit whose covariance depends only on the product $f_i'(x)f_j'(x)$, the coefficient limit $G(x,x)$, and the limit measure $\mu_t$, not on the finer dynamics. The proof identifies $Q_t^N(f)$ with a martingale term $N M_f^N(t)$ up to a remainder that vanishes in $L^p$ uniformly in time; the quadratic variation of that martingale converges to $2\int_0^t\langle (f')^2G(x,x),\mu_s\rangle\,ds$, so a martingale central limit theorem forces the Gaussian limit. In the three applications, restricting to polynomial test functions gives more: the limiting fluctuation processes $L_t(x^n)$ satisfy explicit linear recursions driven by Gaussian processes, yielding distributional descriptions such as $L_t(x)=L_0(x)+G_t(x)$ for the Wishart and Dyson cases.

Load-bearing premise

The load-bearing premise is that the eigenvalue histogram converges, as the dimension grows, to one deterministic limit curve $\mu_t$; the paper assumes this convergence rather than proving it, and for general coefficients uniqueness of such a curve is left open, so if convergence fails or only occurs along a subsequence the stated Gaussian limit is not asserted.

Editorial extensions

If this is right

For the Wishart, Dyson, and symmetric Ornstein-Uhlenbeck flows, the limiting fluctuation process is Gaussian on the polynomial basis and satisfies explicit recursions; for example $L_t(1)=0$ and $L_t(x)$ is the initial fluctuation plus a Gaussian martingale.
The covariance formula gives a direct computation rule for asymptotic covariances of linear spectral statistics: integrate $f_i'f_j'G(x,x)$ against the limit measure over the time overlap.
The comparison principle extends the CLT to generalized particle systems without a matrix representation or explicit stationary density, as long as the drift converges uniformly at rate $o(1/N)$.
Initial data need only be convergent and polynomially moment-bounded; the null-initial-condition restriction of earlier CLTs is removed.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the empirical measures converge only along a subsequence, the same proof gives a subsequential Gaussian CLT; uniqueness of solutions to the limit equation (1.5) would automatically upgrade this to a full statement, so resolving that uniqueness is the natural next step.
The covariance formula suggests a broader universality principle: for interacting particle systems with a deterministic mean-field limit, centered fluctuations of smooth statistics should be Gaussian with covariance driven by the local diffusion coefficient and the limit measure, independent of the detailed interaction kernel.
The recursive moment equations yield an algorithm to simulate the limiting fluctuation processes for Wishart, Dyson, and OU ensembles without simulating $N$ particles; finite-$N$ Monte Carlo checks of variance predictions such as $2t$ for Dyson's Brownian motion would test the theory directly.
For the Ornstein-Uhlenbeck matrix process, the explicit time-change relation with Brownian motion identifies its eigenvalue fluctuations as a deterministic time change of Dyson fluctuations, which predicts the covariance (3.41) and may extend to other matrix flows with explicit solutions.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The useful part of this paper is the general CLT for eigenvalue fluctuations of generalized Wishart processes, and especially the application to the symmetric OU matrix process, which I don't think was in the literature. The Dyson and Wishart applications mostly recover known results with slightly more general initial conditions, which is honest and useful. The proof of Theorem 2.1 is the real meat: Itô's formula, a martingale decomposition, and Rebolledo's martingale CLT, with the covariance falling out as 2∫⟨f'g'G, μ⟩. That argument is coherent and, as far as I can tell, correct.

The comparison principle in Section 3.1 is a nice addition. It gives moment bounds that let the authors push the test functions from the bounded class F to polynomials. That is what makes the recursive formulas in Theorems 3.2–3.4 possible. Credit where due: the paper is also upfront that the central theorem is conditional on the empirical measures converging to a deterministic limit μ, and that uniqueness of such limits is open. That is not circular; the covariance is computed from the assumed limit, not fitted.

The soft spots are real but not fatal. First, the central theorem is genuinely conditional, and the phrase "without loss of generality" before assuming full-sequence convergence is a bit loose: passing to a subsequence could change the covariance. The applications get around this via the companion paper Song–Yao–Yuan, but the general statement is weaker than it looks. Second, the entire particle-system section, Theorem 2.2 and its corollaries, is stated without proof. The paper says the proofs are similar, and they probably are, but for a standalone contribution that is a notable omission. Third, the formula for Lt(x) in the OU theorem looks off to me — there seems to be a sign error in the deterministic integral term. I would not bet heavily on that without checking carefully, but it deserves a look. There are also minor wording issues in Lemma 3.2 and Corollary 3.3, but they do not affect the main line.

Who is this for? People working on random matrix diffusions and eigenvalue fluctuations. The paper is worth a serious referee: the central theorem is plausible and mostly proved, the OU application is new, and the comparison principle is independently useful. I would send it to review, with the expectation that the authors fill in the Section 2.2 proofs and fix the OU formula before publication.

Referee Report

3 major / 4 minor

Summary. The paper studies fluctuations of the empirical eigenvalue measure of generalized Wishart processes and of related interacting particle systems. The main theoretical result, Theorem 2.1, states that under rate conditions on the coefficients and under the assumption that the full sequence of empirical measures converges weakly to a deterministic limit μ_t, the N-scaled centered fluctuation process Q_t^N(f) converges in distribution to a centered Gaussian process whose covariance is 2∫_0^{t∧s}⟨f_i'(x)f_j'(x)G(x,x), μ_u⟩du. The proof uses Itô's formula, a martingale decomposition, and Rebolledo's martingale CLT. Section 2.2 states an analogous CLT for a more general particle system, with proofs omitted. The applications in Section 3 treat the Wishart process, Dyson's Brownian motion, and a symmetric Ornstein-Uhlenbeck matrix process; for these cases the authors prove moment bounds via comparison principles and obtain recursive equations for fluctuation functionals of polynomial test functions.

Significance. If the results are valid, the paper gives a useful and fairly general CLT for eigenvalue fluctuations of matrix diffusions, extending earlier work of Cabanal-Duvillard and Anderson-Guionnet-Zeitouni to more general initial conditions and to a new Ornstein-Uhlenbeck example. The stochastic-calculus proof of Theorem 2.1 is clean and gives an explicit covariance formula in terms of the limit measure and the coefficient limit G. The paper also provides a comparison principle and moment bounds that are of independent interest for non-colliding particle systems. The main qualification is that the general theorem is conditional on full-sequence convergence of the empirical measures, and the uniqueness of the limit equation is left open; in addition, the particle-system CLT of Section 2.2 is stated without proof.

major comments (3)

[Section 1, Eq. (1.5); Theorem 2.1] The manuscript says that 'up to considering a subsequence, the theory is here developed, without loss of generality, by assuming the convergence of the whole sequence'. This reduction is not a genuine WLOG because uniqueness of solutions to the limit equation (1.5) is explicitly left open. A subsequential limit μ may depend on the subsequence, and since the covariance (2.3) is a functional of μ, different subsequences could lead to different Gaussian limits. Theorem 2.1 should therefore be formulated either as a conditional CLT under full-sequence convergence or as a subsequential CLT, with the dependence of the covariance on the chosen limit made explicit.
[Section 2.2, Theorem 2.2 and Corollaries 2.3-2.4, Proposition 2.2] Theorem 2.2 is a central advertised result, but it is stated with the sentence that its proof and the proofs of the associated corollaries are 'similar to those of Theorem 2.1 ... and thus omitted'. The particle system (2.21) has a different drift normalization, a different noise coefficient, and a different test-function class than the eigenvalue system (1.2), so Theorem 2.2 is not a formal corollary of Theorem 2.1. A rigorous treatment requires either a full proof or a detailed indication of how the covariance ∫⟨f_i'f_j'σ̃^2, μ_u⟩du emerges from the quadratic variation of the relevant martingale.
[Section 3.2, proof of Theorem 3.2, around Eq. (3.18)] The induction argument for convergence of (L_t^N(x^k),...,L_t^N(x)) has a joint-convergence gap. The proof shows separately that (̃Q_t^N(x^k),...,̃Q_t^N(x)) converges to a Gaussian vector and, by induction, that (L_t^N(x^{k-1}),...,L_t^N(x)) converges. But separate convergence does not imply joint convergence of ̃Q_t^N(x^k) with the lower-order L^N vector. Since L_t^N(x^k) is defined as ̃Q_t^N(x^k) plus a function of lower-order L^N terms, the argument needs an explicit joint-tightness or Cramér-Wold step. The same issue appears in the inductive parts of Theorems 3.3 and 3.4.

minor comments (4)

[Throughout] There are numerous typographical errors, including 'repsectively', 'ﬂunctuations', 'Scetion 2.2', and 'eignenvalues'; these should be corrected in revision.
[Throughout] The author name 'Ma/suppress lecki' appears to be a LaTeX corruption of 'Maślęcki' or 'Malecki' and should be typeset consistently and correctly in the text and references.
[Corollary 2.2, proof] The phrase 'Choosing p = ln2 N' is ambiguous: it should be written as (ln N)^2 or as a differently named exponent variable, especially because p is also used as the moment order in condition (2.15).
[Theorem 2.1, condition (2.2)] The condition lim N‖N G_N(x,y)-G(x,y)‖_{L∞(R²)}=0 should be stated with care when G is unbounded, as in the Wishart example where G(x,y)=x+y; the norm is understood in the extended sense, and the text would benefit from a remark clarifying this.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Theorem 2.1 is an explicit conditional CLT whose covariance is computed from the assumed limit measure, with self-citations providing independent support rather than substituting for the proof.

full rationale

The central claim is not circular. Theorem 2.1 explicitly takes as hypotheses the coefficient limits (2.2) and the weak convergence of the empirical measures {L^N(t)} to a deterministic limit {μ_t}; it then derives, via Itô's formula and Rebolledo's martingale CLT, that the centered fluctuation process Q^N converges to a Gaussian process with covariance 2∫⟨f_i'f_j'G, μ_u⟩du. That covariance is not fitted or assumed separately: it is computed from the quadratic variation (2.6) of the N-scaled martingale N M_f^N, with the weak convergence hypothesis used only to replace L^N by μ in the final integral. No parameter is fitted to fluctuation data and then renamed a prediction. The main self-citation, to Song et al. (2019), is used to justify the limiting measure equation (2.7) and, in the applications, to verify tightness/convergence of L^N; this supports an explicitly stated hypothesis rather than the Gaussian conclusion itself, and the paper openly records after (1.5) that uniqueness of the limit μ is unknown in general. The 'without loss of generality' subsequence remark is a genuine scope limitation — the theorem is subsequential unless uniqueness is established — but it is an honest stated hypothesis, not a hidden use of the target result. Accordingly, there is no circular step.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

No fitted parameters appear in the central theorem; the constants c, C(T), and α are limits or technical bounds inherited from the model assumptions. The main axioms are existence of the particle system, convergence of the empirical measures, N-rate coefficient convergence, and standard martingale or comparison results. The paper does not introduce new entities such as particles, forces, or conserved quantities beyond the generalized Wishart and particle-system models already in the literature.

assumptions (5)

domain assumption The eigenvalue SDE (1.2) has a unique non-exploding, non-colliding strong solution with infinite collision time.
Assumed in Theorem 2.1 and taken from Graczyk and Malecki (2013, 2014); not reproved in this paper.
domain assumption The empirical measures L^N(t) converge weakly to a deterministic limit {μ_t} as N goes to infinity.
This is the convergence input to Theorem 2.1; the paper notes uniqueness of μ_t is open in general and applications use the companion paper Song et al. (2019).
domain assumption The coefficients converge at N-rate: N||b_N(x)-b(x)||∞→0 and N||N G_N(x,y)-G(x,y)||∞→0.
Equation (2.2); the proof uses these rates to kill the drift and interaction error terms.
standard math Rebolledo's martingale CLT and the Geiss-Manthey comparison principle are valid.
Stated as Lemma 4.1 and Lemma 4.2 and cited to the literature; used as black boxes in the main proofs.
domain assumption The stationary densities P_N and Q_N for Wishart and Dyson eigenvalue processes exist and give exponential tail bounds for the largest eigenvalue.
Used in Lemma 3.1, Lemma 3.2, and Theorem 3.3 to certify moment bounds; these facts are taken from earlier Wishart and Dyson process literature.

how reviews work

0 comments

Cite this review

Pith. "Pith review of High-dimensional central limit theorems for eigenvalue distributions of generalized Wishart processes." pith.science (2026). https://pith.science/paper/3YTV64HJ

@misc{pith2026190803304,
  author       = {Pith},
  title        = {Pith review of: High-dimensional central limit theorems for eigenvalue distributions of generalized Wishart processes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3YTV64HJ}},
  note         = {Machine review of arXiv:1908.03304}
}

read the original abstract

We consider eigenvalues of generalized Wishart processes as well as particle systems, of which the empirical measures converge to deterministic measures as the dimension goes to infinity. In this paper, we obtain central limit theorems to characterize the fluctuations of the empirical measures around the limit measures by using stochastic calculus. As applications, central limit theorems for the Dyson's Brownian motion and the eigenvalues of the Wishart process are recovered under slightly more general initial conditions, and a central limit theorem for the eigenvalues of a symmetric Ornstein-Uhlenbeck matrix process is obtained.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

[1]

W., Guionnet, A., and Zeitouni, O

Anderson, G. W., Guionnet, A., and Zeitouni, O. (2010). An I ntroduction to R andom M atrices . Cambridge University Press, Cambridge New York

work page 2010
[2]

Bru, M.-F. (1991). W ishart processes. Journal of Theoretical Probability , 4(4):725--751

work page 1991
[3]

Cabanal-Duvillard, T. (2001). Fluctuations de la loi empirique de grandes matrices al\' e atoires. Ann. Inst. H. Poincar\' e Probab. Statist. , 37(3):373--402

work page 2001
[4]

and Guionnet, A

Cabanal-Duvillard, T. and Guionnet, A. (2001). Large deviations upper bounds for the laws of matrix-valued processes and non-communicative entropies. Ann. Probab. , 29(3):1205--1261

work page 2001
[5]

and L\' e pingle, D

C\' e pa, E. and L\' e pingle, D. (1997). Diffusing particles with electrostatic repulsion. Probab. Theory Related Fields , 107(4):429--449

work page 1997
[6]

Chan, T. (1992). The W igner semi-circle law and eigenvalues of matrix-valued diffusions. Probab. Theory Related Fields , 93(2):249--272

work page 1992
[7]

Dyson, F. J. (1962). A B rownian-motion model for the eigenvalues of a random matrix. J. Mathematical Phys. , 3:1191--1198

work page 1962
[8]

and Manthey, R

Gei , C. and Manthey, R. (1994). Comparison theorems for stochastic differential equations in finite and infinite dimensions. Stochastic Process. Appl. , 53(1):23--35

work page 1994

Show all 17 references

[9]

and Ma ecki, J

Graczyk, P. and Ma ecki, J. (2013). Multidimensional Y amada- W atanabe theorem and its applications to particle systems. J. Math. Phys. , 54(2):021503, 15

2013
[10]

and Ma ecki, J

Graczyk, P. and Ma ecki, J. (2014). Strong solutions of non-colliding particle systems. Electron. J. Probab. , 19:no. 119, 21

2014
[11]

and Ma ecki, J

Graczyk, P. and Ma ecki, J. (2019). On squared B essel particle systems. Bernoulli , 25(2):828--847

2019
[12]

and Shreve, S

Karatzas, I. and Shreve, S. E. (1991). B rownian M otion and S tochastic C alculus , volume 113 of Graduate Texts in Mathematics . Springer-Verlag, New York, second edition

1991
[13]

and P \'e rez , J

Ma ecki , J. and P \'e rez , J. L. (2019). Universality classes for general random matrix flows . arXiv e-prints , page arXiv:1901.02841

2019 arXiv
[14]

Rogers, L. C. G. and Shi, Z. (1993). Interacting B rownian particles and the W igner law. Probab. Theory Related Fields , 95(4):555--570

1993
[15]

Rudin, W. (1991). F unctional A nalysis . International Series in Pure and Applied Mathematics. McGraw-Hill, Inc., New York, second edition

1991
[16]

\' S niady, P. (2002). Random regularization of B rown spectral measure. J. Funct. Anal. , 193(2):291--313

2002
[17]

Song, J., Yao, J., and Yuan, W. (2019). High-dimensional limits of eigenvalue distributions for general W ishart process. arXiv preprint arXiv:1901.02190

2019 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

W., Guionnet, A., and Zeitouni, O

Anderson, G. W., Guionnet, A., and Zeitouni, O. (2010). An I ntroduction to R andom M atrices . Cambridge University Press, Cambridge New York

work page 2010

[2] [2]

Bru, M.-F. (1991). W ishart processes. Journal of Theoretical Probability , 4(4):725--751

work page 1991

[3] [3]

Cabanal-Duvillard, T. (2001). Fluctuations de la loi empirique de grandes matrices al\' e atoires. Ann. Inst. H. Poincar\' e Probab. Statist. , 37(3):373--402

work page 2001

[4] [4]

and Guionnet, A

Cabanal-Duvillard, T. and Guionnet, A. (2001). Large deviations upper bounds for the laws of matrix-valued processes and non-communicative entropies. Ann. Probab. , 29(3):1205--1261

work page 2001

[5] [5]

and L\' e pingle, D

C\' e pa, E. and L\' e pingle, D. (1997). Diffusing particles with electrostatic repulsion. Probab. Theory Related Fields , 107(4):429--449

work page 1997

[6] [6]

Chan, T. (1992). The W igner semi-circle law and eigenvalues of matrix-valued diffusions. Probab. Theory Related Fields , 93(2):249--272

work page 1992

[7] [7]

Dyson, F. J. (1962). A B rownian-motion model for the eigenvalues of a random matrix. J. Mathematical Phys. , 3:1191--1198

work page 1962

[8] [8]

and Manthey, R

Gei , C. and Manthey, R. (1994). Comparison theorems for stochastic differential equations in finite and infinite dimensions. Stochastic Process. Appl. , 53(1):23--35

work page 1994