Pith. sign in

REVIEW 4 minor 32 references

Top Singular Value in Sum-Products of Random Matrices

T0 review · 0 major / 4 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read The top singular value of a sum of random matrix products equals a random-energy-model free energy at temperature set by N, n and m.

desk verdict Clean non-asymptotic REM reduction for the top singular value of sum-products in the open triple-scaling regime; the only soft spot is a flagged technical range restriction that does not break the theorems. read the letter →

arxiv 2607.04047 v1 pith:KHA2PVQK submitted 2026-07-04 math.PR math-phmath.MP

Kevin Han Huang , Boris Hanin This is my paper · ORCID
classification math.PRmath-phmath.MP MSC 60B2015B5282B44
keywords randommatrixproductstopsingularvalueenergymodelLyapunovexponenttriplescalingmoderatedeviationsphasetransition
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 largest singular value of a sum of m independent products of N Gaussian matrices of size n. When all three dimensions grow, that singular value is shown to coincide with the free energy of a classical random-energy model whose inverse temperature is fixed by the combination sqrt of 2(N-1)/(n log m). The energies themselves are non-Gaussian but become Gaussian in the scaling limit, so the well-known REM phase transition at critical temperature sqrt(2) appears as a sharp change in the singular-value asymptotics. High-temperature regimes are dominated by many typical matrix products; low-temperature regimes are dominated by a few extreme products. Explicit non-asymptotic error bounds make the correspondence quantitative under mild growth conditions, giving a single phase diagram that interpolates between free-probability, ergodic-theory and moderate-deviation regimes previously treated separately.

What carries the argument

Reduction of the top Lyapunov exponent to the log-partition function of a non-Gaussian REM, followed by a Cramer moderate-deviation approximation of the energies by Gaussians and Laplace-method concentration of that partition function.

What would settle it

Compute the top singular value of the sum-product matrix for a sequence of triples (m,n,N) that violate log m = o(N^{1/3}) and check whether the observed value still tracks the REM free-energy formula within the claimed error; a systematic deviation would falsify the approximation.

Watch

Extended reading notes

Core claim

When m, n and N tend to infinity simultaneously, the top singular value of the sum-product matrix X is asymptotically identical to the log-partition function of a random energy model whose inverse temperature is beta = sqrt(2(N-1)/(n log m)) and whose energies depend on the ratio N/n. The identification is made precise by two non-asymptotic theorems that control the approximation error both above and below the critical temperature sqrt(2).

Load-bearing premise

The Gaussian approximation of the individual energies is valid only when log m grows slower than N to the power one-third; if that growth condition fails the error bounds no longer hold.

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

0 major / 4 minor

Summary. The paper studies the top singular value s1(X) (equivalently the top Lyapunov exponent) of the sum-product matrix X = m^{-1/2} sum_{i=1}^m X_i, where each X_i is a product of N i.i.d. n imes n Gaussian matrices with N(0,n^{-1}) entries. In the triple-scaling regime m,n,N o∞ the authors show that log s1(X) and log∥Xθ∥ (for fixed unit θ) are approximated by the log-partition function of a random energy model (REM) whose inverse temperature is β=√[2(N-1)/(n log m)] and whose energies are non-Gaussian but approximately Gaussian with law depending on N/n. Theorems 1 and 2 give explicit non-asymptotic error bounds that exhibit a high-temperature/low-temperature phase transition at β=√2, recovering the classical Gaussian REM free energy after recentering. The argument proceeds by reducing the matrix problem to a non-Gaussian REM (Lemmas 4–5, Corollary 6), approximating the energies via a Cramér moderate-deviation theorem (Proposition 10), and controlling the REM partition function by Markov and Laplace-method arguments (Propositions 7–8).

Significance. The work supplies the first non-asymptotic description of the top singular value of a sum of many random matrix products in the simultaneous limit m,n,N o∞. The identification with a REM free energy at a temperature that depends on all three parameters is conceptually clean and yields a transparent phase diagram (Figures 2–3) that interpolates between free-probability, ergodic-theory and double-scaling regimes previously studied only for m=1. The proofs are fully quantitative, track error terms explicitly, and rest only on classical concentration and moderate-deviation tools; the authors themselves flag the technical restrictions log m=o(N^{1/3}) and N=o(n^3) as improvable. The results therefore constitute a solid, self-contained contribution to non-asymptotic random matrix theory and to the interface with disordered systems.

minor comments (4)
  1. [Remark 1] Remark 1 correctly notes that the hypothesis log m=o(N^{1/3}) originates from the classical Cramér range x=o(N^{1/6}). A short additional sentence indicating whether sharper moderate-deviation results (e.g., under sub-exponential tails of the log-chi-squared variables) could remove the restriction would help the reader assess optimality.
  2. [Theorem 2] In the statement of Theorem 2 the additive error (1+2α)log n/(2β^{2} log m) appears only after the sphere-supremum removal (Lemma 4). It would be clearer to display this term already in the high-temperature regime of Theorem 1, or to note explicitly that it is an artifact of the net argument rather than of the REM approximation.
  3. [Figures 1–3] Figures 1–3 are informative but the captions are dense. A one-line summary of the phase transition (β=√2) placed above each figure would improve readability.
  4. [Section 3.1] The comparison with the classical REM literature (Section 3.1) is accurate, yet a brief remark on the difference between almost-sure convergence along the subsequence m=2^M and the in-probability statements of Propositions 7–8 would prevent possible confusion for readers coming from statistical physics.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the REM approximation for the sum-product top singular value is derived from concentration, Cramér moderate deviations and Laplace’s method; self-citations of prior single-product results are technical scaffolding, not load-bearing for the central claim.

full rationale

The derivation chain is self-contained. Lemmas 4–5 and Corollary 6 reduce logs1(X) and log∥Xθ∥ to a non-Gaussian REM log-partition function by Gaussian concentration (Hanson–Wright) and distributional invariance under Haar orthogonal multiplications; the single-product identities used (Lemma 3, the argument of Lemma 4) are special cases or verbatim extensions of Hanin–Paouris 2021 that hold for the same reason for the sum, and are re-proved or re-checked in place. Proposition 10 supplies a location-dependent Gaussian approximation of the energies via the classical Cramér moderate-deviation theorem (external, 1938). Propositions 7–8 then obtain quantitative concentration of the REM free energy by Markov’s inequality (ultra-high temperature) and an explicit Laplace-method covering argument (moderate-to-low temperature), with all error terms written out. The limiting expression Z is compared to the classical Gaussian REM free energy of Derrida/Bovier et al., which is used only for interpretation, not as an input that forces the answer. There is no parameter fitting, no uniqueness theorem imported from the authors, no ansatz smuggled via citation, and no self-definitional loop. The sole self-citations (Hanin–Paouris, Hanin–Nica, Hanin–Jiang) supply background techniques or comparison regimes; the triple-scaling REM connection and the non-asymptotic bounds of Theorems 1–2 are new and independently derived. Score 1 reflects only the presence of non-load-bearing self-citations of prior technique.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The argument rests only on standard probabilistic tools (Hanson–Wright, Cramér moderate deviations, digamma asymptotics of log-χ^{2}) and the classical Gaussian REM free-energy formula used for comparison. No free parameters are fitted; the inverse temperature eta is defined from the model parameters. No new physical entities are postulated.

assumptions (4)
  • domain assumption Entries of each factor matrix are i.i.d. N(0,1/n)
    Defines the probability space; used from equation (1) onward.
  • standard math Cramér-type moderate deviation theorem for i.i.d. sums with exponential moments (Cramér 1938)
    Invoked in Proposition 10 to approximate non-Gaussian energies by Gaussians at locations ~√log m.
  • standard math Hanson–Wright concentration for quadratic forms of Gaussians
    Used in Lemma 5 to replace the last matrix layer by its conditional expectation.
  • standard math Asymptotic expansion of digamma and polygamma functions for log-χ^{2} moments
    Lemma 9; supplies mean and variance of the elementary energies Yij.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Top Singular Value in Sum-Products of Random Matrices." pith.science (2026). https://pith.science/paper/KHA2PVQK

@misc{pith2026260704047,
  author       = {Pith},
  title        = {Pith review of: Top Singular Value in Sum-Products of Random Matrices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KHA2PVQK}},
  note         = {Machine review of arXiv:2607.04047}
}
abstract

We study the top singular value for a sum of $m$ independent $n \times n$ random matrices, each of which is a product of $N$ i.i.d. $n\times n$ Gaussian matrices. Our main conceptual observation is that when $m,n,N\rightarrow \infty$, the top singular value coincides with the partition function in a random energy model at the inverse temperature $\beta=\sqrt{2(N-1)/(n\log m)}$, with energies depending on the ratio $N/n$. We provide several non-asymptotic results making this approximation precise.

Figures

Figures reproduced from arXiv: 2607.04047 by the authors.

Figure 1
Figure 1. Known results on the value of log s1(X) when m = 1 under the single scaling regime, where only one of n and N grows with the other fixed. ψ denotes the digamma function. For the double-scaling regime, where N/n → γ ∈ (0,∞), results are only known for (i) the complex Gaussian case (Theorem 1.2 of Liu et al. [2023]): log s1(X) ≈ log n 2 + F(γ) + oP(1), where F(γ) is a generic quantity that depends only on γ; (ii) for … view at source ↗
Figure 2
Figure 2. Our results about the value of log ∥Xθ∥ under the triple scaling where m grows with n and N, which is determined not by lim N n = γ but by the inverse temperature β = √ 2(N−1) √ n log m . We now interpret the results in both the high and low temperature regimes: High temperature regime with β ≤ √ 2. In this case, the limiting approximation is Z = 0, and Theo￾rems 1 and 2 imply the following about log ∥Xθ∥ and log s1… view at source ↗
Figure 3
Figure 3. Our results about the value of log s1(X) under the triple scaling where m grows with n and N, which is determined not by lim N n = γ but by the inverse temperature β = √ 2(N−1) √ n log m . in which case s1 (X) ∈ [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 1 linked inside Pith

  1. [1]

    Non-asymptotic results for singular values of

    Hanin, Boris and Paouris, Grigoris , journal=. Non-asymptotic results for singular values of. 2021 , publisher=

    2021

  2. [2]

    International Journal of Modern Physics B , volume=

    Large deviations and the random energy model , author=. International Journal of Modern Physics B , volume=. 2001 , publisher=

    2001

  3. [3]

    Fluctuations of the free energy in the

    Bovier, Anton and Kurkova, Irina and L. Fluctuations of the free energy in the. The Annals of Probability , volume=. 2002 , publisher=

    2002

  4. [4]

    Physical Review B , volume=

    Random-energy model: An exactly solvable model of disordered systems , author=. Physical Review B , volume=. 1981 , publisher=

    1981

  5. [5]

    arXiv preprint arXiv:2503.07872 , year=

    Global Universality of Singular Values in Products of Many Large Random Matrices , author=. arXiv preprint arXiv:2503.07872 , year=

    arXiv

  6. [6]

    Communications in Mathematical Physics , volume=

    Products of many large random matrices and gradients in deep neural networks , author=. Communications in Mathematical Physics , volume=. 2020 , publisher=

    2020

  7. [7]

    Sur un nouveau th

    Cram. Sur un nouveau th. Actualités Scientifiques et Industrielles , pages=

  8. [8]

    1975 , publisher=

    Sums of independent random variables , author=. 1975 , publisher=

    1975

Show all 32 references
  1. [9]

    2023 , publisher=

    Liu, Song-Hao and Zhang, Zhuo-Song , journal=. 2023 , publisher=

    2023

  2. [10]

    2018 , publisher=

    High-dimensional probability: An introduction with applications in data science , author=. 2018 , publisher=

    2018

  3. [11]

    2023 , publisher=

    Liu, Dang-Zheng and Wang, Dong and Wang, Yanhui , journal=. 2023 , publisher=

    2023

  4. [12]

    Probability Theory and Related Fields , volume =

    Ahn, Andrew , title =. Probability Theory and Related Fields , volume =

  5. [13]

    From integrable to chaotic systems: Universal local statistics of

    Akemann, Gernot and Burda, Zdzislaw and Kieburg, Mario , journal=. From integrable to chaotic systems: Universal local statistics of. 2019 , publisher=

    2019

  6. [14]

    The Annals of Probability , pages=

    A limit theorem for the norm of random matrices , author=. The Annals of Probability , pages=. 1980 , publisher=

    1980

  7. [15]

    Probability theory and related fields , volume=

    On the limit of the largest eigenvalue of the large dimensional sample covariance matrix , author=. Probability theory and related fields , volume=. 1988 , publisher=

    1988

  8. [16]

    Physical Review E—Statistical, Nonlinear, and Soft Matter Physics , volume=

    Products of rectangular random matrices: singular values and progressive scattering , author=. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics , volume=. 2013 , publisher=

    2013

  9. [17]

    A simple proof of almost sure convergence for the largest singular value of a product of

    Saada, Thiziri Nait and Naderi, Alireza , journal=. A simple proof of almost sure convergence for the largest singular value of a product of

  10. [18]

    The Annals of Mathematical Statistics , volume=

    Products of random matrices , author=. The Annals of Mathematical Statistics , volume=. 1960 , publisher=

    1960

  11. [19]

    A multiplicative ergodic theorem

    Oseledets, Valery Iustinovich , journal=. A multiplicative ergodic theorem. 1968 , publisher=

    1968

  12. [20]

    The distribution of

    Newman, Charles M , journal=. The distribution of. 1986 , publisher=

    1986

  13. [21]

    The Annals of Probability , pages=

    The stability of large random matrices and their products , author=. The Annals of Probability , pages=. 1984 , publisher=

    1984

  14. [22]

    The triangle law for

    Isopi, Marco and Newman, Charles M , journal=. The triangle law for. 1992 , publisher=

    1992

  15. [23]

    On the largest

    Kargin, Vladislav , journal=. On the largest. 2014 , publisher=

    2014

  16. [24]

    Universal distribution of

    Akemann, Gernot and Burda, Zdzislaw and Kieburg, Mario , journal=. Universal distribution of. 2014 , publisher=

    2014

  17. [25]

    Physical Review E , volume=

    Universality of local spectral statistics of products of random matrices , author=. Physical Review E , volume=. 2020 , publisher=

    2020

  18. [26]

    Extremal singular values of random matrix products and

    Ahn, Andrew , journal =. Extremal singular values of random matrix products and

  19. [27]

    A new application of random matrices:

    Haagerup, Uffe and Thorbj. A new application of random matrices:. The Annals of Mathematics , pages=. 2005 , publisher=

    2005

  20. [28]

    Non-commutative polynomials of independent

    Schultz, Hanne , journal=. Non-commutative polynomials of independent. 2005 , publisher=

    2005

  21. [29]

    Current Developments in Mathematics, 2025 , pages =

    van Handel, Ramon , title =. Current Developments in Mathematics, 2025 , pages =

    2025

  22. [30]

    Random Matrices: Theory and Applications , volume=

    Asymptotic spectra of matrix-valued functions of independent random matrices and free probability , author=. Random Matrices: Theory and Applications , volume=. 2015 , publisher=

    2015

  23. [31]

    Limiting Spectral Distributions of Sums of Products of Non-Hermitian Random Matrices , journal =

    K. Limiting Spectral Distributions of Sums of Products of Non-Hermitian Random Matrices , journal =. 2018 , doi =

    2018

  24. [32]

    Electronic Communications in Probability , volume =

    Bordenave, Charles , title =. Electronic Communications in Probability , volume =

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.