Pith. sign in

REVIEW 2 major objections 5 minor 21 references

Edge statistics for singular values of products of rectangular complex Ginibre matrices

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that the depth-to-width ratio $\Delta_{M,N}=\sum_{j=0}^M 1/(N+v_j)$ is a sharp threshold: if it tends to infinity, edge singular values of products of rectangular Ginibre matrices show Gaussian fluctuations; if it tends…

desk verdict Plausible sharp-threshold edge classification for rectangular product Ginibre, but the proof of the Airy-side theorem is incomplete because Lemma 2.2's key derivative is wrong as printed. read the letter →

arxiv 2507.07431 v1 pith:ZU5ES5GX submitted 2025-07-10 math.PR

Yandong Gu This is my paper · ORCID
classification math.PR MSC 60B2015B52
keywords productsofrandommatricessingularvaluesedgestatisticsdepth-to-widthratioGinibreAirykernelTracy-Widomdistributiondeterminantalpointprocess
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 studies the edge statistics of the logarithmic singular values of a product $Y_M=X_M\cdots X_1$ of independent rectangular complex Ginibre matrices, where $X_j$ has size $(N+v_j)\times(N+v_{j-1})$. Its central claim is that the depth-to-width ratio $\Delta_{M,N}=\sum_{j=0}^M 1/(N+v_j)$ is a sharp threshold: when $\Delta_{M,N}\to\infty$, the rescaled correlation functions converge to a Gaussian one-point density and all higher correlation functions vanish; when $\Delta_{M,N}\to 0$, the same correlation functions converge to the Airy kernel determinant, giving Tracy-Widom GUE edge statistics. Combined with the moderate regime where a critical kernel appears, this gives a complete three-way classification of the soft edge by a single scalar parameter. The interest is that the individual rectangularities of the chain do not matter separately; only their aggregate $\Delta_{M,N}$ decides the universality class.

What carries the argument

The load-bearing object is the double integral representation (1.2) of the correlation kernel, obtained from the exact joint densities of the singular values: an integral over a Hankel contour $C$ and a closed contour $\Sigma$, with ratios of gamma functions encoding all $M+1$ matrix dimensions. The asymptotic analysis is a steepest-descent argument: after a change of variables and Stirling expansion, the kernel is written through a phase $f_{M,N}$ whose critical point is $q_0=z_0\rho_{M,N}^{-3/2}$, with $z_0$ the root of (1.10). The contours are split into local pieces near $q_0$, where a cubic Taylor expansion produces the Airy kernel, and global pieces whose decay is controlled by the monotonicity of $\Re f_{M,N}$ along the deformed contours. The depth-to-width ratio $\Delta_{M,N}$, the sum of the reciprocal widths, is the parameter that selects the regime: it controls the variance scale $\rho_{M,N}(k)$ in the Gaussian theorem and the small parameter in the Airy theorem.

What would settle it

Evaluate the coefficient of the linear term in $y$ of $d\,\Re f_{M,N}(x_0+iy)/dy$ at $y=0$, namely $\sum_{j=0}^M \rho_{M,N}^3/(N+v_j+\rho_{M,N}^{3/2}x_0)^2 - 1/x_0^2$; Lemma 2.2 requires this coefficient to be negative for $x_0>q_0$. A concrete check is $M=1$, $N=100$, $v_1=0$ with $x_0$ just above $q_0$. If the sign is wrong for any profile, the global estimate behind Theorem 1.2 fails, even if the local Airy computation survives.

Watch

Extended reading notes

Core claim

On its own terms, the paper proves two limit theorems for the determinantal point process of $\log(Y_M^*Y_M)$. Theorem 1.1 states that if $\Delta_{M,N}\to\infty$, then for fixed $k\in\mathbb{N}$ the $n$-point correlation functions, centred at $\sum_{j=0}^M\psi(N+v_j+1-k)$ and scaled by $\rho_{M,N}(k)=\left(\sum_{j=0}^M\psi'(N+v_j+1-k)\right)^{1/2}$, converge uniformly on compacts to a single Gaussian density for $n=1$ and to zero for $n>1$. Theorem 1.2 states that if $\Delta_{M,N}\to 0$, then with scale $\rho_{M,N}$ and centre $\log\lambda_M$ determined by the unique positive solution $z_0$ of $\sum_{j=0}^M 1/(N+v_j+z)=1/z$, the $n$-point correlation functions converge uniformly on compacts to the determinant $\det(K_{\mathrm{Ai}}(\xi_i,\xi_j))_{i,j=1}^n$. The author presents this as completing the phase diagram: the critical kernel occupies the moderate regime, and $\Delta_{M,N}$ interpolates between Gaussian and Airy statistics at the two ends.

Load-bearing premise

The load-bearing premise is that the global pieces of the deformed integration contours decay at the stated exponential rates; in particular, the sign of the decay rate of the phase near the critical point must be correct for every rectangularity profile.

Editorial extensions

If this is right

  • If Theorem 1.1 holds, the edge statistics in the high-DWR regime collapse to a single Gaussian density: the $n$-point correlation functions vanish for $n>1$.
  • If Theorem 1.2 holds, the edge statistics in the low-DWR regime match the universal soft edge of the GUE, encoded by the Airy kernel.
  • Together with the moderate-DWR critical kernel, the two theorems give a complete classification: Gaussian, critical, and Airy statistics correspond to $\Delta_{M,N}\to\infty$, $\Delta_{M,N}\to\gamma\in(0,\infty)$, and $\Delta_{M,N}\to 0$.
  • The low-DWR theorem handles chains whose matrices have different sizes, with the centre $\log\lambda_M$ and scale $\rho_{M,N}$ determined by all the $v_j$'s through the critical equation.
  • The paper's stated motivation is that these edge laws matter for stability analysis in deep neural networks and communication systems, where products of rectangular random matrices arise.

Reading between the lines

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

  • A question the paper leaves open is the speed of the transition between the Airy and Gaussian regimes; the uniform formulation suggests one could test whether $\Delta_{M,N}$ decaying like $1/\log N$ still gives Airy statistics and where the crossover begins.
  • A practical reading is that for finite $N$ the single number $\Delta_{M,N}$, not the individual widths $v_j$, tells a user whether edge fluctuations are Gaussian or Airy-like.
  • The mechanism points to the same DWR trichotomy for neighbouring product ensembles with similar kernel structures, such as products of truncated unitary matrices, where a modified depth-to-width ratio already appears in related work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper studies the edge statistics of singular values of products of independent rectangular complex Ginibre matrices with dimension sequence (N+v_j) x (N+v_{j-1}) and v_0=0. Building on the exact correlation kernel (1.2) and on the depth-to-width ratio (DWR) Δ_{M,N}=Σ_{j=0}^M 1/(N+v_j) introduced in [15], it proves two theorems. Theorem 1.1 states that when Δ_{M,N}→∞, after centering by Σ_j ψ(N+v_j+1-k) and scaling by ρ_{M,N}(k), the correlation functions converge to a Gaussian density for n=1 and vanish for n>1. Theorem 1.2 states that when Δ_{M,N}→0, with scaling ρ_{M,N} defined by (1.8) and centering log λ_{M,N} defined by (1.9)-(1.10), the correlation functions converge uniformly on compacts to the Airy kernel determinant. The proofs use steepest-descent analysis on the kernel (1.2), with local Taylor expansions and global decay estimates. Together with the moderate-DWR critical kernel from [15], the parameter Δ_{M,N} is claimed to be a sharp threshold between Gaussian and Airy edge statistics.

Significance. If the proof gaps are closed, the paper would establish a sharp, parameter-free phase transition for edge statistics of rectangular Ginibre products, extending the square-case results of [15] to non-square multiplicative chains. The claims are anchored by several external consistency checks: the M=2 square edge parameter λ_M reduces to the known Fuss-Catalan edge, the one-rectangular-factor limit matches the Wishart edge (√(N+v_1)+√N)^2, and the Gaussian variance in Theorem 1.1 matches the known Lyapunov variance. The derivation starts from the exact kernel and contains no fitting parameters or target-dependent normalization, which is a genuine strength. However, the global steepest-descent estimates that carry the main theorems are currently under-verified, so the paper is not yet at the standard of a journal publication.

major comments (2)
  1. [§2.2, Lemma 2.2, Eq. (2.58)] The derivative expansion (2.58) drops the term coming from Im log z in d/dy Re f_{M,N}(x0+iy). The correct first-order coefficient is 1/x0 - ρ^{3/2} Σ_j 1/(N+v_j+ρ^{3/2}x0), not -ρ^{3/2} Σ_j ... . At x0=q0, equation (1.10) makes this coefficient vanish identically, so the printed formula does not even give the correct sign at the critical point; for x0>q0 the sign depends on a nontrivial comparison between 1/x0 and ρ^{3/2}Σ_j(N+v_j+ρ^{3/2}x0)^{-1} that is not supplied. Since Lemma 2.2 is the only justification for the decay of Re f along C1_global and hence for the global bound (2.56)-(2.57), the uniform Airy limit in Theorem 1.2 is not rigorously established as written.
  2. [§2.1, Step 3, Eqs. (2.22)-(2.25)] The Gaussian theorem rests on uniform linear-decay estimates that are asserted rather than proved. Inequality (2.22) postulates an ε>0 with no argument controlling Im ψ uniformly on the vertical contour; (2.23) states an endpoint value (1/2)√N ρ^{1/2}(k)(1+o(1)) without derivation; and Lemma 2.1's inequality (2.25) is justified by a sketch on finitely many contour segments, with the uniformity of the error in (2.29) not quantified. These estimates are plausible and likely fixable, but they are load-bearing for the conclusion (2.27) that I2 and I_1^{global} are negligible, so Theorem 1.1 also needs a more complete proof.
minor comments (5)
  1. [Abstract] The abstract's first paragraph reverses the nomenclature used in Section 1.1 and in the theorems: Gaussian fluctuations occur for Δ→∞ (high DWR) and the Airy kernel for Δ→0 (low DWR), while the abstract states the opposite.
  2. [§2.1, Eq. (2.6)] The subscript in ρ_{M,n}(k) should be ρ_{M,N}(k); the same inconsistency appears in (2.11).
  3. [Throughout] There are several typos: 'digama' (page 3), 'makig' (before (2.50)), 'maximun' and 'leftend' (page 9).
  4. [§2.2, Eq. (2.52)] The notation C_{ρ^{1/20}_{M,N}} and Σ_{ρ^{1/20}_{M,N}} for the truncated contours is hard to parse; it would help to define these explicitly as scaled versions of the contours in (2.53)-(2.54).
  5. [References] Reference [4] contains a stray page number '255202 (2014)' after the EPL volume and page range.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the correlation kernel derivation is self-contained steepest descent, and the only self-citation is contextual rather than load-bearing.

full rationale

The paper's central claims are derived from the exact double-integral kernel (1.2), not from a fitted model or from a target-dependent normalization. In Theorem 1.2, the scaling parameters rho_{M,N} and lambda_M are determined by the stationary-point equation (1.10) and the cubic Taylor expansion f'''_{M,N}(q0)=2+o(1); the Airy determinant then emerges from the resulting cubic phase, matching the standard Airy kernel against external benchmarks. Theorem 1.1 similarly produces a Gaussian limit from the quadratic term of F(t;k) with rho_{M,N}(k) defined by F''(1-k;k), so no fitted quantity is later renamed as a prediction. The claimed sharp threshold combines the present extreme-DWR results with the moderate-DWR critical kernel from [15], which is an external published result with a proof, and the DWR itself was introduced there; this is legitimate reliance on prior work, not circular importation. The author's own [12] is cited only in the introduction for a related phase transition in truncated unitary products and plays no role in the proofs. The only substantive concern is a proof gap in Lemma 2.2: the expansion (2.58) drops the -1/x_0 term in the derivative of Re f_{M,N}, so the sign and the global decay used in (2.56)-(2.57) are not rigorously established. That is a correctness/completeness issue, not a circularity, because the intended Airy conclusion is not assumed anywhere in the derivation. There is no self-definitional reduction, no fitted input called a prediction, and no load-bearing self-citation chain.

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

There are no fitted constants: rho_{M,N}(k), rho_{M,N}, lambda_M and z_0 are determined by the model through saddle point conditions (1.5) and (1.8)-(1.10), not tuned to the target Airy or Gaussian kernels. The nontrivial inputs are external: the exact kernel (1.2), the moderate-regime result [15], and standard asymptotics for the digamma and Gamma functions. The proof sketches presuppose the unshown global decay estimates described in the weakest_assumption field.

assumptions (5)
  • domain assumption The exact correlation kernel (1.2) from Kuijlaars-Zhang [14], built on the joint density of [5], describes the log-eigenvalues of Y^*Y as a determinantal point process for all M, N, v_j.
    Invoked at the start of Section 2; the paper does not rederive the kernel.
  • standard math Existence and uniqueness of the positive solution z_0 of equation (1.10) for M >= 1.
    The paper states 'the unique positive solution' before (1.8) without proof; monotonicity of f(z) = sum 1/(N+v_j+z) - 1/z from -infinity to 0+ gives it, but the argument is omitted in the text.
  • standard math Stirling and digamma expansions (2.2)-(2.3) hold uniformly in the sectors used, including when v_j grow with N.
    Used throughout Section 2 for dominant terms and error bounds.
  • domain assumption The moderate-DWR critical kernel result of [15], with partially overlapping authorship via D.-Z. Liu, is taken as established and used as the middle phase of the phase diagram.
    Section 1.1, fourth paragraph; the paper's sharp-threshold framing leans on [15] without reproving it.
  • domain assumption Model conventions: v_j >= 0, v_0 = 0, M >= 1, and complex Ginibre entries with the standard scaling.
    Model definition (1.1) and Section 1.1; the normalization convention for Ginibre entries is standard but not stated explicitly.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Edge statistics for singular values of products of rectangular complex Ginibre matrices." pith.science (2026). https://pith.science/paper/ZU5ES5GX

@misc{pith2026250707431,
  author       = {Pith},
  title        = {Pith review of: Edge statistics for singular values of products of rectangular complex Ginibre matrices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZU5ES5GX}},
  note         = {Machine review of arXiv:2507.07431}
}
read the original abstract

We investigate the edge statistics of singular values for products of independent rectangular complex Ginibre matrices. Building on work \cite{LWW23}, which introduced the depth-to-width ratio (DWR) and established the critical kernel, we show that the DWR acts as a sharp threshold: local statistics exhibit Gaussian fluctuations in the low-DWR regime, while the Airy kernel emerges at the soft edge in the high-DWR regime.

Figures

Figures reproduced from arXiv: 2507.07431 by the authors.

Figure 1
Figure 1. Schematic contours in proof of Theorem 1.1 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Schematic contours in proof of Theorem 1.2 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 20 canonical work pages

  1. [15]

    Liu, D.-Z., Wang, D., Wang, Y.: Lyapunov exponent, universality and phase transition for products of random matrices. Comm. Math. Phys. 399(3), 1811–1855 (2023)

    work page 2023

  2. [1]

    Ahn, A.: Fluctuations of β-Jacobi product processes. Probab. Theory Relat. Fields 183, 57–123 (2022)

    work page 2022

  3. [2]

    Ahn, A., Peski, V.R.: Lyapunov exponents for truncated unitary and Ginibre matrices. Ann. Inst. Henri Poincar´ e Probab. Stat. 59(2), 1029–1039 (2023). 10

    work page 2023

  4. [3]

    Journal of Physics A 47(39), 395202, 35 (2014)

    Akemann, G., Burda, Z., Kieburg, M.: Universal distribution of Lyapunov exponents for products of Ginibre matrices. Journal of Physics A 47(39), 395202, 35 (2014)

    work page 2014

  5. [4]

    EPL (Europhys

    Akemann, G., Burda, Z., Kieburg, M.: From integrable to chaotic systems: Universal local statistics of Lyapunov exponents. EPL (Europhys. Lett.) 126(4), 40001 (2019). 255202 (2014)

    work page 2019

  6. [5]

    Akemann, G., Kieburg, M., Wei, L.: Singular value correlation functions for products of Wishart random matrices. J. Phys. A 46(27), 275205 (2013)

    work page 2013

  7. [6]

    Cam- bridge Studies in Advanced Mathematics 118

    Anderson, G.W., Guionnet, A., Zeitouni, O.: An Introduction to Random Matrices. Cam- bridge Studies in Advanced Mathematics 118. Cambridge University Press, Cambridge, 2010

    work page 2010

  8. [7]

    Bellman, R.: Limit theorems for non-commutative operations. I. Duke Math. J. 21, 491–500 (1954)

    work page 1954

Show all 21 references
  1. [8]

    arXiv preprint arXiv:2503.22801 (2025)

    Berezin, S., Strahov, E.: Last-passage percolation and product-matrix ensembles. arXiv preprint arXiv:2503.22801 (2025)

    2025 arXiv

  2. [9]

    Springer Series in Solid-State Sciences 104

    Crisanti, A., Paladin, G., Vulpiani, A.: Products of Random Matrices in Statistical Physics. Springer Series in Solid-State Sciences 104. Springer, Berlin, 1993

    1993

  3. [10]

    Furstenberg, H., Kesten, H.: Products of random matrices. Ann. Math. Stat. 31, 457–469 (1960)

    1960

  4. [11]

    Gorin, V., Sun, Y.: Gaussian fluctuations for products of random matrices. Amer. J. Math. 144 (2), 287–393 (2022)

    2022

  5. [12]

    arXiv preprint

    Gu, Y., Liu, D.-Z.: Local Statistics of Singular Values for Products of Truncated Unitary Matrices. arXiv preprint. arXiv:2506.14193(2025)

    2025 arXiv

  6. [13]

    Jiang, T., Qi, Y.: Spectral radii of large non-Hermitian random matrices. J. Theor. Probab. 30(1), 326–364 (2017)

    2017

  7. [14]

    Kuijlaars, A.B.J., Zhang, L.: Singular values of products of Ginibre random matrices, multiple orthogonal polynomials and hard edge scaling limits. Commun. Math. Phys. 332(2), 759–781 (2014)

    2014

  8. [16]

    Liu, D.-Z., Wang, D., Zhang, L.: Bulk and soft-edge universality for singular values of products of Ginibre random matrices. Ann. Inst. Henri Poincar´ e Probab. Stat. 52(4), 1734– 1762 (2016)

    2016

  9. [17]

    Liu, D.-Z., Wang, Y.: Phase transitions for infinite products of large non-Hermitian random matrices. Ann. Inst. Henri Poincar´ e Probab. Stat. 60(4), 2813–2848 (2024)

    2024

  10. [18]

    Fields Institute Mono- graphs 35

    Mingo, J.A., Speicher, R.: Free Probability and Random Matrices. Fields Institute Mono- graphs 35. Springer, New York, 2017

    2017

  11. [19]

    (eds.): NIST handbook of math- ematical functions

    Olver, F.W.J., Lozier, D.W., Boisvert, R.F., Clark, C.W. (eds.): NIST handbook of math- ematical functions. U.S. Department of Commerce, National Institute of Standards and Technology, Washington, DC; Cambridge University Press, Cambridge 2010. With 1 CD- ROM (Windows, Macinto...

    2010

  12. [20]

    Tulino, A.M., Verd´ u, S.: Random matrix theory and wireless communications. Found. Trends Commun. Inf. Theory. 1(1), 1-182 (2004)

    2004

  13. [21]

    Tracy, C.A., Widom, H.: Level-spacing distributions and the Airy kernel. Commun. Math. Phys. 159(1), 151–174 (1994). 11

    1994

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.