Pith. sign in

REVIEW 2 major objections 4 minor 37 references

Randomly coupled differential equations with elliptic correlations

T0 review · 2 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 large random network with elliptic correlations between reciprocal connections, the averaged squared solution at critical coupling decays as $t^{-1/2}$ with an explicit, corrected prefactor.

desk verdict Substantial paper with a real factor-N blunder in the main kernel formula; the proof actually proves the corrected version. read the letter →

arxiv 1908.05178 v4 pith:BYW3YRZC submitted 2019-08-14 math.PR

László Erdős , Torben Krüger , David Renfrew This is my paper · ORCID
classification math.PR MSC 60B2015B52
keywords non-Hermitianrandommatricesellipticensembleelliptic-typeextraspectralDysonequationlocallawneuralnetworkspowerdecayresolventproducts
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

This paper proves a precise asymptotic description of the long-time decay for a large linear system $\partial_t u=-u+gXu$, where the random matrix $X$ has elliptic correlations: reciprocal pairs $(x_{ij},x_{ji})$ may be correlated with a complex parameter $\varrho$, with $|\varrho|<1$, and different pairs are independent. The central result is that at the critical coupling $g=(1+|\varrho|^2+2\Re\varrho)^{-1/2}$, the expectation of the squared norm of the solution over random initial data decays like $t^{-1/2}$ with an explicit prefactor that depends only on $\varrho$, for every elliptic ensemble with $0<|\varrho|<1$. This corrects an earlier non-rigorous formula for the same quantity. For the more general elliptic-type ensemble, with entrywise non-negative reciprocal-correlation matrix $T$, the paper proves the same $t^{-1/2}$ law with a coefficient $A(S,T)$ built from the variance profile $S$ and $T$. The proof is powered by a deterministic approximation of $\operatorname{Tr}_N f(X)g(X^*)$ by a double contour integral whose kernel is expressed through the solution of the extraspectral Dyson equation.

What carries the argument

The load-bearing object is the self-consistent pseudo-resolvent $b(\zeta)\in\mathbb{C}^N$, defined as the maximal holomorphic solution of the extraspectral Dyson equation $$1+(\zeta+Tb(\zeta))b(\zeta)=0,\qquad b(\zeta)\sim -1/\zeta\ \text{as }\zeta\to\infty,$$ subject to the spectral-radius condition $r(D_{|b(\zeta)|^2}S)<1$. It plays the role of the diagonal of the resolvent of $X$ outside the pseudospectrum. The proof machinery is a Hermitization: the product of resolvents $(X-\zeta_1)^{-1}(X^*-\zeta_2)^{-1}$ is recovered as an $\alpha$-derivative of the resolvent of a $4N\times 4N$ Hermitian block matrix $H^Z_\alpha$ at $\alpha=0$. The resolvent of $H^Z_\alpha$ is compared, with an optimal local law, to the solution of the corresponding matrix Dyson equation; the stability operator of that equation is shown invertible with control $\Delta_Z^{-1}$, and analytic perturbation theory extracts the singularity of $K$ at $\zeta_1=\zeta_2=\zeta_*$ that produces the power-law decay.

What would settle it

Simulate the elliptic ensemble at critical coupling with a fixed $0<|\varrho|<1$ and large $N$, averaging $\|u_t\|^2$ over initial conditions and comparing the coefficient of $t^{-1/2}$ with $[1+|\varrho|^2-2\Re((\varrho+|\varrho|^2)/(\varrho+1))]/(2\sqrt{\pi})$. A systematic mismatch in the prefactor or a different decay exponent as $t$ grows within the admissible window $t\le N^c$ would disprove the central asymptotic formula.

Watch

Extended reading notes

Core claim

The paper's main theorem is that for elliptic-type random matrices satisfying assumptions (A), (B), and (2.C-F), the normalized trace of $f(X)g(X^*)$ converges with overwhelming probability to $$\frac{1}{(2\pi i)^2}\oint_\gamma\oint_\gamma f(\zeta_1)g(\zeta_2)K(\zeta_1,\zeta_2)\,d\zeta_1\,d\zeta_2,$$ where $K(\zeta_1,\zeta_2)=N^{-2}\sum_{i,j}[(D^{-1}_{b(\zeta_1)b(\zeta_2)}-S)^{-1}]_{ij}$ and $b(\zeta)$ is the unique holomorphic solution of the extraspectral Dyson equation $1+(\zeta+Tb(\zeta))b(\zeta)=0$ with $b(\zeta)\sim -1/\zeta$ at infinity. The singularities of this kernel near the rightmost point of the deterministic pseudospectrum control the long-time behavior: at critical coupling the double integral is dominated by a residue that produces the $t^{-1/2}$ decay and the explicit prefactor in the elliptic case. In the elliptic case, the kernel simplifies to $K(\zeta_1,\zeta_2)=b(\zeta_1)b(\zeta_2)/(1-b(\zeta_1)b(\zeta_2))$, which reduces the problem to a Bessel-function series and yields formulas (2.3) and (2.5). The paper also states the elliptic-type variant with coefficient $A(S,T)$ given explicitly in (7.16).

Load-bearing premise

The load-bearing premise is that the known Wigner-type approximation of a resolvent by a deterministic matrix still works for the $4N\times 4N$ block matrix built from $X$ and $X^*$, even though the strongly correlated pairs $x_{ij}$ and $x_{ji}$ lie in widely separated positions within that block matrix; the proof relies on the general correlation conditions from [15] for this transfer.

Editorial extensions

If this is right

  • For every $0<|\varrho|<1$, the averaged squared norm at critical coupling decays as $t^{-1/2}$, the same exponent as the fully asymmetric case, while the fully symmetric case decays as $t^{-3/2}$.
  • The explicit formula (2.3) replaces the earlier non-rigorous expression and can be used as a benchmark for finite-$N$ numerical studies of random linear networks.
  • For elliptic-type matrices with non-negative $T$, the decay exponent $t^{-1/2}$ is universal and the coefficient $A(S,T)$ is computable from $S$, $T$, and the Perron eigenvectors of $D_{|b(\zeta_*)|^2}S$.
  • The trace formula for $f(X)g(X^*)$ provides a deterministic equivalent for correlated non-Hermitian matrices that holds with high probability, not just in expectation.

Reading between the lines

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

  • If the same kernel controls eigenvector overlaps, the method may yield rigorous overlap statistics for elliptic-type matrices, where only Gaussian results are known.
  • The critical-decay exponent being independent of $\varrho$ for all $|\varrho|<1$ suggests a wider universality class for partially symmetric random connectivity; one testable extension is to sparse or heavy-tailed connectivity matrices satisfying the same moment and correlation assumptions.
  • The formulas (2.3) and (2.5) could be compared directly to simulations of nonlinear neural rate models near instability, since the linearized dynamics at critical coupling is the regime the paper analyzes.
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.

Referee Report

2 major / 4 minor

Summary. The paper studies the large-N linear system ∂_t u_t = -u_t + gX u_t with a random connectivity matrix X whose pairs (x_ij, x_ji) are correlated. It introduces an elliptic-type ensemble with general variance profile S and correlation matrix T, defines the self-consistent pseudospectrum and the extraspectral Dyson equation, and proves a high-probability limit for the normalized trace Tr_N f(X)g(X*) in Theorem 2.7. This trace asymptotics is applied to the elliptic ensemble to obtain an explicit Bessel-function expansion and the t^{-1/2} decay at critical coupling, correcting the formula of [28], and to general elliptic-type matrices to obtain the same power-law decay with a model-dependent coefficient A(S,T). The proof is based on Hermitization, the matrix Dyson equation, and the optimal local law for correlated Hermitian matrices from [15].

Significance. If the stated results are correct, they substantially generalize the earlier paper [14] from independent entries to general elliptic correlations, give the first rigorous treatment of the partially symmetric neural-network model, and pinpoint a concrete error in the final formula of [28]. Strengths of the manuscript include the high-probability (rather than expectation-only) statements, the treatment of non-identical variance profiles, and the explicit and internally consistent Bessel-function reduction in Section 7.1. The authors also honestly flag the delicate point that the 4N×4N Hermitized matrix H^Z_α satisfies only the weaker assumptions (C),(D) of [15] rather than (CD). However, the central kernel formula in Theorem 2.7 is stated with an incorrect normalization, and this must be corrected before the results as printed are valid.

major comments (2)
  1. [Theorem 2.7, Eq. (2.15); Remark 2.8; proof of (6.8)] The normalization of the kernel K is off by a factor N. For the elliptic ensemble, S = (1/N)11^T and b is a scalar. Writing c = (b1 b2)^{-1}, one has sum_{ij} [(c I - (1/N)11^T)^{-1}]_{ij} = N/(c - 1), so the right-hand side of (2.15) equals N^{-1} b1 b2/(1 - b1 b2), whereas (2.17) and the proof of Theorem 2.1 require K = b1 b2/(1 - b1 b2). The same mismatch occurs in the proof of (6.8): the (3,1) block of ∂_α M^Z_α|_{\alpha=0} is D_r with r = (D^{-1}_{b1 b2} - S)^{-1} 1, and Tr_N of that block equals N^{-1} sum_i r_i = N^{-1} sum_{ij} [(D^{-1}_{b1 b2} - S)^{-1}]_{ij}, which is N times the quantity appearing in (2.15). Thus Theorem 2.7 as stated would give a vanishing limit for the normalized trace in the elliptic and Ginibre cases, contradicting the standard limit 1/(ζ1 ζ2 - 1) as well as the paper's own derivation. The intended definition is evidently K = N^{-1} sum_{ij} [(D^{-1}_{b1 b2} - S)^{-1}]_{ij}; this correction should be made in Theorem 2.7, in Remark 2.8, and in the independent-entry formula (2.16).
  2. [Sections 5.1 and 6, application of [15] to H^Z_α] The proof of Theorem 6.1 depends on applying [15, Theorem 2.1 and Corollary 2.3] to the 4N×4N Hermitian matrix H^Z_α, in which the strongly correlated entries x_ij and x_ji appear in widely separated blocks. The authors state that H^Z_α satisfies assumptions (C) and (D) of [15] via Example 2.11, but the verification is not carried out in the text. Since this is the load-bearing transfer of the local law to the block linearization, please add an explicit check of the correlation-decay and block-structure conditions, or quote precisely the part of [15, Example 2.11] that covers this case.
minor comments (4)
  1. [Theorem 2.1, Eq. (2.2)] The displayed probability statement has the inequality "(\ge 1 - C_{\epsilon,\nu}/N^\nu)" placed after the closing parenthesis in a nonstandard way; it should read P( ... ) \ge 1 - C_{\epsilon,\nu}/N^\nu.
  2. [Proof of (6.8)] In the sentence defining K in the proof of (6.8), the reference "(2.16)" should be "(2.15)" once the normalization correction is made.
  3. [Section 7.1, after Eq. (7.6)] The phrase "using the relationship I_j(z) = I_j(z)" appears to be a typo; the intended statement is presumably the conjugation identity for Bessel functions of complex argument.
  4. [Corollary 7.6] The second sentence says "This bound follows from Δ_ζ ≥ Δ_{|ζ|} by (7.4)"; the cross-reference should be to Lemma 7.4 rather than Eq. (7.4).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the kernel and decay prefactors are derived from the independent Dyson-equation solution, not fitted to the target trace limit; prior same-author results are used as independent theorems. A separate factor-N kernel inconsistency is a correctness issue, not circularity.

full rationale

The derivation is self-contained in the relevant sense: the limiting kernel K(ζ1,ζ2) is built from b(ζ), the solution of the extraspectral Dyson equation (2.9), which is defined independently of the trace-limit formula (2.14); no parameter is fitted to the target observable, and the elliptic ODE decay is obtained by contour integration after Theorem 2.7 rather than by assuming the decay. The prior works [14] and [15] are used as black-box theorems (local law, spectral concentration, stability of the matrix Dyson equation), not as assertions of the product-resolvent limit being proven; [15] in particular addresses Hermitian matrices with slow correlation decay and does not assume the target trace formula. The authors explicitly flag the delicate transfer of [15] to the 4N×4N hermitizations H^Z_α, which violate condition (CD) but fall under the more general conditions (C)–(D); this is an applicability and limitation point, not a circular reduction. Separately, there appears to be an internal numerical factor-N mismatch between (2.15) and the elliptic specialization (2.17) and the (3,1)-block computation in (6.8); this is a correctness concern, not a circularity. No step reduces by construction to its inputs.

Assumptions & free parameters 0 free parameters · 8 assumptions · 0 invented entities

The central claims rest on the stated random-matrix hypotheses and on prior local-law theorems; no free parameters are fitted to make the derivation work. The self-consistent pseudo-resolvent b and kernel K are defined from the data, not assumed.

assumptions (8)
  • domain assumption Assumption (A): E[x_ij]=0 for all entries.
    Stated in Section 2.1; used in all main theorems.
  • domain assumption Assumption (B): finite moments with E|x_ij|^p ≤ φ_p N^{-p/2}.
    Moment bound needed for the local law and concentration estimates.
  • domain assumption Assumption (2.C): independence across pairs (x_ij,x_ji).
    Defines the elliptic-type structure; used for the MDE self-energy.
  • domain assumption Assumption (2.D): |t_ij| ≤ |ϱ| sqrt(s_ij s_ji) with |ϱ|<1.
    Genuinely non-Hermitian condition ensuring r(D_{|b|^2} T)<|ϱ|<1 (Section 4).
  • domain assumption Assumption (2.E): uniform primitivity of S and S*S.
    Provides the effective spectral gap and Perron-Frobenius estimates in Lemma 4.9.
  • domain assumption Assumption (2.F): piecewise Hölder continuity of T.
    Needed for uniform entrywise bounds on b via the quadratic vector equation theory of [3].
  • domain assumption Assumption t_ij ≥ 0 for Theorem 2.6.
    Technical non-negativity to use the QVE framework and characterize E; flagged by the authors in Section 7.2.
  • standard math Local law of [15, Theorem 2.1 and Corollary 2.3] applies to H^Z_α under conditions (C),(D).
    The proof in Section 6 replaces random resolvents by MDE solutions using this theorem; the authors cite [15, Example 2.11] for the block-matrix setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Randomly coupled differential equations with elliptic correlations." pith.science (2026). https://pith.science/paper/BYW3YRZC

@misc{pith2026190805178,
  author       = {Pith},
  title        = {Pith review of: Randomly coupled differential equations with elliptic correlations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BYW3YRZC}},
  note         = {Machine review of arXiv:1908.05178}
}
abstract

We consider the long time asymptotic behavior of a large system of $N$ linear differential equations with random coefficients. We allow for general elliptic correlation structures among the coefficients, thus we substantially generalize our previous work [14] that was restricted to the independent case. In particular, we analyze a recent model in the theory of neural networks [27] that specifically focused on the effect of the distributional asymmetry in the random connectivity matrix $X$. We rigorously prove and slightly correct the explicit formula from [28] on the time decay as a function of the asymmetry parameter. Our main tool is an asymptotically precise formula for the normalized trace of $f(X) g(X^*)$, in the large $N$ limit, where $f$ and $g$ are analytic functions.

Figures

Figures reproduced from arXiv: 1908.05178 by the authors.

Figure 7.1
Figure 7.1. Contours of integration We set δ1 = 3√  0 and will close the ζ2-contour to pick up a residue. From Lemma 7.7, ∂z2(τ ) = −1. In particular, z2(ζ ∗ + ∆) = ζ ∗ − ∆ +O(|∆| 2 ) and Re(z2(ζ1)−ζ ∗ ) = O( 0 ) for any ζ1 ∈ Γ ,3 √  0. Thus we choose δ2 ≥ C √  0 for sufficiently large C > 0 and can close the ζ2-contour by adding the line τ [eiδ2 , e −iδ2 ], once again adding an exponentially small in t error term. Then t… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 35 canonical work pages

  1. [15]

    Erdős, T

    L. Erdős, T. Krüger, and D. Schröder. Random matrices with slow correlation decay. Forum Math. Sigma, 7:e8, 89, 2019

    work page 2019

  2. [28]

    Mehlig and J

    B. Mehlig and J. T. Chalker. Statistical properties of eigenvectors in non-Hermitian Gaus- sian random matrix ensembles.J. Math. Phys., 41(5):3233–3256, 2000

    work page 2000

  3. [14]

    Erdős, T

    L. Erdős, T. Krüger, and D. Renfrew. Power law decay for systems of randomly coupled differential equations.SIAM J. Math. Anal., 50(3):3271–3290, 2018

    work page 2018

  4. [1]

    Abramowitz.Handbook of Mathematical Functions, With Formulas, Graphs, and Math- ematical Tables,

    M. Abramowitz.Handbook of Mathematical Functions, With Formulas, Graphs, and Math- ematical Tables,. Dover Publications, Inc., New York, NY, USA, 1974

    work page 1974

  5. [2]

    O. H. Ajanki, L. Erdős, and T. Krüger. Universality for general Wigner-type matrices. Probab. Theory Related Fields, 169(3-4):667–727, 2017

    work page 2017

  6. [3]

    O. H. Ajanki, L. Erdős, and T. Krüger. Quadratic vector equations on complex upper half-plane. Mem. Amer. Math. Soc., 261(1261):v+133, 2019

    work page 2019

  7. [4]

    O. H. Ajanki, L. Erdős, and T. Krüger. Stability of the matrix Dyson equation and random matrices with correlations.Probab. Theory Related Fields, 173(1-2):293–373, 2019

    2019

  8. [5]

    J. Alt, L. Erdős, and T. Krüger. The Dyson equation with linear self-energy: spectral bands, edges and cusps. arXiv:1804.07752, Apr. 2018

    arXiv 2018

Show all 37 references
  1. [6]

    J. Alt, L. Erdős, and T. Krüger. Local inhomogeneous circular law.Ann. Appl. Probab., 28(1):148–203, 2018

    2018

  2. [7]

    J. Alt, L. Erdős, T. Krüger, and Y. Nemish. Location of the spectrum of Kronecker random matrices. Ann. Inst. Henri Poincaré Probab. Stat., 55(2):661–696, 2019

    2019

  3. [8]

    Altenberg

    L. Altenberg. A sharpened condition for strict log-convexity of the spectral radius via the bipartite graph. Linear Algebra and Its Applications, 438(9):3702–3718, 2013

    2013

  4. [9]

    S. T. Belinschi, P. Śniady, and R. Speicher. Eigenvalues of non-Hermitian random matri- ces and Brown measure of non-normal operators: Hermitian reduction and linearization method. Linear Algebra Appl., 537:48–83, 2018

    2018

  5. [10]

    Bordenave and D

    C. Bordenave and D. Chafaï. Around the circular law.Probab. Surveys, 9:1–89, 2012

    2012

  6. [11]

    Bourgade and G

    P. Bourgade and G. Dubach. The distribution of overlaps between eigenvectors of Ginibre matrices. Probab. Theory Related Fields, 177(1-2):397–464, 2020. 35

    2020

  7. [12]

    J. T. Chalker and B. Mehlig. Eigenvector statistics in non-Hermitian random matrix ensembles. Phys. Rev. Lett., 81:3367–3370, Oct 1998

    1998

  8. [13]

    N. Cook, W. Hachem, J. Najim, and D. Renfrew. Non-Hermitian random matrices with a variance profile (I): deterministic equivalents and limiting ESDs.Electron. J. Probab., 23:Paper No. 110, 61, 2018

    2018

  9. [16]

    Y. V. Fyodorov. On statistics of bi-orthogonal eigenvectors in real and complex Ginibre ensembles: Combining partial Schur decomposition with supersymmetry.Comm. Math. Phys., 363(2):579–603, Oct 2018

    2018

  10. [17]

    V. L. Girko. Circular law.Teor. Veroyatnost. i Primenen., 29:669–679, 1984

    1984

  11. [18]

    V. L. Girko. Elliptic law.Theory Probab. Appl., 30:677–690, 1986

    1986

  12. [19]

    V. L. Girko.Theory of Stochastic Canonical Equations: Volumes I and II. Mathematics and Its Applications. Springer Netherlands, 2012

    2012

  13. [20]

    Grilli, T

    J. Grilli, T. Rogers, and S. Allesina. Modularity and stability in ecological communities. Nature communications, 7:12031, June 2016

    2016

  14. [21]

    J. W. Helton, R. Rashidi Far, and R. Speicher. Operator-valued semicircular elements: solving a quadratic matrix equation with positivity constraints. Int. Math. Res. Not. IMRN, (22):Art. ID rnm086, 15, 2007

    2007

  15. [22]

    Hennequin, T

    G. Hennequin, T. Vogels, and W. Gerstner. Optimal control of transient dynamics in balanced networks supports generation of complex movements.Neuron, 82(6):1394 – 1406, 2014

    2014

  16. [23]

    Hennequin, T

    G. Hennequin, T. P. Vogels, and W. Gerstner. Non-normal amplification in random bal- anced neuronal networks.Phys. Rev. E, 86:011909, Jul 2012

    2012

  17. [24]

    Kuczala and T

    A. Kuczala and T. O. Sharpee. Eigenvalue spectra of large correlated random matrices. Phys. Rev. E, 94:050101, Nov 2016

    2016

  18. [25]

    Lim and M

    S. Lim and M. S. Goldman. Balanced cortical microcircuitry for maintaining information in working memory.Nature Neuroscience, 16(9):1306–1314, 2013

    2013

  19. [26]

    MacNeil and C

    D. MacNeil and C. Eliasmith. Fine-tuning and the stability of recurrent neural networks. PLOS ONE, 6(9):1–16, 09 2011

    2011

  20. [27]

    Martí, N

    D. Martí, N. Brunel, and S. Ostojic. Correlations between synapses in pairs of neurons slow down dynamics in randomly connected neural networks.Phys. Rev. E, 97:062314, Jun 2018

    2018

  21. [29]

    H. H. Nguyen and S. O’Rourke. The elliptic law.Int. Math. Res. Not. IMRN, (17):7620– 7689, 2015. 36

    2015

  22. [30]

    Rajan and L

    K. Rajan and L. F. Abbott. Eigenvalue spectra of random matrices for neural networks. Phys. Rev. Lett., 97:188104, Nov 2006

    2006

  23. [31]

    E. Seneta. Non-negative matrices and Markov chains. Springer Series in Statistics. Springer, New York, 2006. Revised reprint of the second (1981) edition [Springer-Verlag, New York; MR0719544]

    1981

  24. [32]

    Sompolinsky, A

    H. Sompolinsky, A. Crisanti, and H.-J. Sommers. Chaos in random neural networks.Phys. Rev. Lett., 61(3):259–262, 1988

    1988

  25. [33]

    S. Song, P. J. Sjöström, M. Reigl, S. Nelson, and D. B. Chklovskii. Highly nonrandom features of synaptic connectivity in local cortical circuits.PLOS Biology, 3(3), 03 2005

    2005

  26. [34]

    T. Tao, V. Vu, and M. Krishnapur. Random matrices: Universality of ESDs and the circular law.Ann. Probab., 38(5):2023–2065, 09 2010

    2023

  27. [35]

    van Vreeswijk and H

    C. van Vreeswijk and H. Sompolinsky. Chaotic balanced state in a model of cortical circuits. Neural Comput., 10:1321–1371, 1998

    1998

  28. [36]

    Walters and S

    M. Walters and S. Starr. A note on mixed matrix moments for the complex Ginibre ensemble. J. Math. Phys., 56(1):013301, 20, 2015

    2015

  29. [37]

    Y. Wang, H. Markram, P. H. Goodman, T. K. Berger, J. Ma, and P. S. Goldman-Rakic. Heterogeneity in the pyramidal network of the medial prefrontal cortex.Nature Neuro- science, 9(4):534, 2006. 37

    2006

Pith tools

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