Pith. sign in

REVIEW 5 major objections 7 minor 61 references

An RDT based confirmation of Lehner's formula for Kronecker-Gaussian matrices

T0 review · 5 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The asymptotic largest eigenvalue of a Kronecker–Gaussian matrix is exactly a deterministic max-over-positive-semidefinite value, and a comparison-based random-duality argument proves it without spectral methods.

desk verdict A careful RDT reproof of a known spectral-edge formula; the upper bound is solid, but the lower bound relies on an unproved matrix-valued replica tightness principle, so the main equality is not established. read the letter →

arxiv 2607.26551 v1 pith:E5OTJVU7 submitted 2026-07-29 math.PR cs.ITmath.ITmath.STstat.TH

Mihailo Stojnic This is my paper · ORCID
classification math.PRcs.ITmath.ITmath.STstat.TH MSC 60B2046L5415B5260G15
keywords Kronecker–GaussianmatricesstrongasymptoticfreenessrandomdualitytheoryspectraledgeGaussiancomparisonprinciplereplicaoverlaptightnesssemicircularfreeoperatorsmatrixconcentration
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 takes on a known deterministic formula: for a Kronecker–Gaussian matrix $H=A_0\otimes I+\frac{1}{\sqrt{n}}\sum_{i=1}^l A_i\otimes \bar G_i$ with symmetric positive-semidefinite $A_i$ and symmetric Gaussian $\bar G_i$, the limiting largest eigenvalue should be $\rho_n=\max_{S\succeq 0,\ \mathrm{tr}(S)=1}\left(\mathrm{tr}(A_0S)+2\,\mathrm{tr}\sqrt{\sum_{i=1}^l (A_iS)^2}\right)$. The paper's objective is to prove this equality without free probability or spectral methods, using only comparison principles from Random Duality Theory. A sympathetic reading is that RDT's random-dual upper bound and its two-replica-cannot-double lower-bound principle are enough to pin the edge exactly, thereby rederiving the strong asymptotic freeness edge results of [25,43] as a byproduct.

What carries the argument

The machinery is the four-step RDT protocol adapted to the matrix-valued setting. First, $\lambda_n(H)$ is rewritten as $\max_{\mathrm{tr}(R^TR)=1}(\mathrm{tr}(F_0)+\frac{\sqrt{2}}{\sqrt{n}}\xi)$ with $F_i=R A_i R^T$ and $\xi=\max_{X^TX=I}\sum_i \mathrm{tr}(X^TG_iXF_i)$. Second, a Gaussian-process comparison inequality bounds $\xi$ by the random dual $L=\max_{X^TX=I}\sum_i\sqrt{2}\,\mathrm{tr}(G_i^{(1)}XF_i)$, giving the upper bound. Third, the dual is handled by the Lagrangian identity $L=\sqrt{2}\min_{\Gamma=\Gamma^T}\left(\frac14\mathrm{tr}(\sum_i F_iF_i^T\Gamma^{-1})+\mathrm{tr}(\Gamma)\right)$, whose large-$n$ limit is $\min_{\Gamma}(\mathrm{tr}(K\Gamma^{-1})+\mathrm{tr}(\Gamma))$ with $K=\sum_i F_iF_i^T$ and optimizer $\tilde\Gamma=\sqrt{K}$. Fourth, the missing lower bound is reduced to condition (92): $\min_{\Gamma,\Lambda}\bar L^{(2)}<2\min_\Gamma\bar L$ over admissible overlaps $Q$, which the paper proves by contradiction using the $\Lambda$-derivative at $(\tilde\Gamma/2,0)$. The load-bearing object is the two-replica overlap $Q$ and the strict inequality that certifies the two-replica-cannot-double principle.

What would settle it

Evaluate condition (92) for a concrete instance: fix $k,l$, positive-semidefinite $A_i$, an admissible overlap $Q\in\mathcal{Q}$, and $t\in(0,1)$, and compute $\min_{\Gamma=\Gamma^T,\Lambda=\Lambda^T}\bar L^{(2)}$ and $2\min_{\Gamma=\Gamma^T}\bar L$ from (90) and (35). If the former is not strictly smaller than the latter, the proof's tightness condition fails. A sharper check: find any admissible $Q$ and $t$ for which the $\Lambda$-derivative in (106) vanishes at $(\tilde\Gamma/2,0)$, since Theorem 5's contradiction argument requires that derivative to be nonzero.

Watch

Extended reading notes

Core claim

The central claim is that $\lim_{n\to\infty}\mathbb{E}\lambda_n(H)=\rho_n$ (and consequently $\lim_{n\to\infty}\mathbb{E}\lambda_1(H)=\rho_1$ by Gaussian sign symmetry), where $\rho_n$ is the deterministic maximum over $S\succeq 0$, $\mathrm{tr}(S)=1$, of $\mathrm{tr}(A_0S)+2\,\mathrm{tr}\sqrt{\sum_i(A_iS)^2}$, matching the formula of [31] for the semicircular free counterpart. The proof is a sandwich: the upper bound comes from comparing the Gaussian process defining $\lambda_n(H)$ with a simpler random-dual process and minimizing a Lagrangian over a dual variable $\Gamma$; the matching lower bound comes from the RDT tightness principle that a two-copy system with any nontrivial overlap $Q$ has strictly less than twice the single-copy free energy, which forces the interpolated upper and lower limits in (41) to coincide. The paper therefore claims to reconfirm the spectral-edge formula of [31] and to reprove the strong-asymptotic-freeness edge statements of [25,43] without spectral theory.

Load-bearing premise

The argument assumes that for these matrix-valued Gaussian processes, a two-copy system with any non-trivial overlap cannot achieve twice the optimal value of a single copy, and that this failure forces the two comparison limits to coincide—a principle the paper takes from earlier work and extends to the $k$-fold matrix setting rather than deriving from a more basic argument.

Editorial extensions

If this is right

  • The exact spectral edge of Kronecker–Gaussian matrices is established without free probability, Stieltjes transforms, or matrix Dyson equations; the comparison argument alone carries the proof.
  • The strong asymptotic freeness edge statements of [25,43] follow as corollaries of the RDT comparison route, not as inputs.
  • By Gaussian symmetry the same result covers the smallest eigenvalue $\lambda_1(H)$, so both edges of the spectrum are pinned deterministically.
  • The deterministic edge $\rho_n$ is a convex SDP, so the theorem turns an asymptotic spectral statement into a finite-dimensional optimization that can be computed in principle.
  • The proof extends the replica-overlap tightness principle to the $k$-fold matrix-valued setting, which the paper presents as a generic mechanism for other Gaussian-process comparisons.

Reading between the lines

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

  • If the overlap-tightness principle tolerates noncommuting $F_i$ and indefinite $A_i$, the comparison route could prove spectral-edge formulas for ensembles outside the reach of current free-probability tools, such as rectangular or weakly dependent Gaussian models; that extension is not in the paper.
  • Condition (92) is a finite-dimensional inequality for fixed $k,l$ and concrete $A_i$, so it could be certified by numerical optimization for specific instances, giving a practical check of the lower bound even where the analytic contradiction argument is hard to run.
  • The proof implicitly restricts the dual variable $\Gamma$ to the positive-definite cone in (32)–(37); a fully explicit statement of that restriction and a justification of the square-root optimizer $\tilde\Gamma=\sqrt{K}$ inside the cone would close a small rigor gap the paper leaves open.
  • The author flags the indefinite-$A_i$ case as future work; if RDT handles it, the same two-sided comparison would reprove broader polynomial-norm convergence results that currently rely on free probability.
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

5 major / 7 minor

Summary. The paper proposes a proof, based on Random Duality Theory (RDT), of Lehner's deterministic spectral-edge formula for Kronecker–Gaussian matrices of the form H = A0⊗I + n^{-1/2} Σ Ai⊗G_i. The main claim is that lim_{n→∞} Eλ_n(H) = ρ_n, where ρ_n is the deterministic SDP value appearing in (5)–(6) and (38). The upper bound lim Eλ_n(H) ≤ ρ_n is obtained through a Slepian/Gordon comparison of Gaussian processes followed by a Lagrangian dual calculation. The lower bound is attempted through a replicated-system argument, Section 3.5, whose key step is an RDT 'tightness' equivalence stated in (48)–(49). The paper concludes that this also reproves the strong asymptotic freeness edge results of [25,43].

Significance. If the proof were complete, the paper would provide a genuinely different route to a known and important result: Lehner's formula for the spectral edges of Kronecker–Gaussian matrices would follow from RDT rather than from free probability or spectral methods. The upper-bound half is essentially rigorous and the algebra in Theorem 5 is internally coherent. The paper contains no fitted parameters and the final formula is explicit and falsifiable. However, the lower-bound half, which is the load-bearing part, depends on a strong RDT tightness principle that is imported from self-cited preprints and asserted to extend to matrix-valued overlaps without proof. As written, the central equality is therefore not established.

major comments (5)
  1. [§3.5, Eqs. (48)–(49)] The lower-bound argument rests on the equivalence min_{Q∈Q} (lim 2·n^{-1/2} E D(0) − lim n^{-1/2} E D^(2)(t)) > 0 ⇔ lim n^{-1/2} E D(1) = lim n^{-1/2} E D(0). This is asserted, not proved. The text says that 'the remaining parts of the [53,54] methodologies automatically extend', but for k>1 the overlap Q = (X^(1))^T X^(2) is a k×k contraction rather than a scalar, replica symmetry breaking could occur at matrix level, and the second-moment control in (22) does not by itself rule out overlap dependence. Since (93)–(94), and hence the final equality lim Eλ_n(H) = ρ_n, follow only through this equivalence, the central claim is unsupported as written. A self-contained proof of (49) in this matrix-valued Stiefel setting, or a precise reduction to the scalar case, is required.
  2. [§3.3, Eqs. (32)–(37)] The dual minimization is over Γ = Γ^T with no positive-definiteness restriction. For the scalar analogue, min_{γ≠0} (a/γ + γ) is unbounded below, and for matrices the same phenomenon occurs when Γ is allowed to be indefinite. The optimality condition (37) and the substitution Γ → √n Γ in Theorem 4's proof presuppose Γ ≻ 0 and invertible. The domain must be explicitly restricted to Γ ≻ 0 (or Γ ⪰ 0 with a limiting argument), and the existence of the minimizer should be stated.
  3. [§2.1, Eqs. (4)–(7) and (6)] The main theorem as stated is not dimensionally consistent. H ∈ R^{nk×nk} has nk eigenvalues, but (4) lists λ_1 ≤ ... ≤ λ_n, and (6) writes lim_{n→∞} λ_n(H) = ρ_n. With λ_i defined as the i-th smallest eigenvalue, λ_n(H) is the n-th smallest among nk eigenvalues, not the maximum. The object in (7) is max over S^{nk}, i.e., the largest eigenvalue λ_{nk}(H). The statements in (6) and throughout should use the correct eigenvalue index, or λ_n(H) should be explicitly redefined as the maximal eigenvalue.
  4. [§2.1, Eq. (5)] The displayed definition of ρ_n and ρ_1 is not a well-defined matrix expression as written: A0⊗I is nk×nk, while Z and A_i Z^{-1} A_i (with A_i being k×k) cannot be added to it. The later SDP formula in (38), taken from [18], is clear, but (5) needs to be corrected to that representation or to the intended Schur-complement form so that the central object is defined precisely.
  5. [§3.3 and §3.5.2, Eqs. (33)–(34), (95)–(96)] The limit n→∞ is interchanged with the minimization over Γ (and Λ) without justification. The law of large numbers gives convergence of (1/n)Σ (F_i G_i)(F_i G_i)^T to K, but the minimizer depends on n, and the parameter domain is unbounded. Before writing the displayed limits, one needs uniform concentration over Γ (and Λ), or an epsilon-net argument. Without this, the upper bound is not fully rigorous.
minor comments (7)
  1. [§3.2, Eq. (23)] In the sentence following (23), '(X^(1))^T X^(1) = 0' should be '= I'.
  2. [§3.5, Eq. (42)] The set Q_F is garbled: the condition '>0∈R^{n×k}' is not meaningful, and the set should be defined as a subset of R^{k×k} (or with a precise trace inequality).
  3. [§3.5.1, Eq. (61)] The second labeled line for θ̄_i,2 should be θ̄_i,4; the current label duplicates the second term.
  4. [§3.5.3, Eq. (107)] The notation ∥(I−Q)F_i∥²_F is not literally the same as the preceding trace tr((I−Q^T)F_i(I−Q)F_i); the latter equals ∥F_i^{1/2}(I−Q)F_i^{1/2}∥²_F. Please clarify the notation.
  5. [§2.1, after Eq. (5)] The sentence 'we assume that A_i's are such that inf can be replaced by max' is vague; specify the nondegeneracy condition that guarantees attainment.
  6. [Throughout] There are several typographical issues, e.g., 'eignevalue' after (4), 'pne obtains' in §3.5.1, and the limiting notation in (41) and (65) without explicit existence statements.
  7. [Abstract and §4] The claim of 'effectively reproving the key asymptotic freeness results obtained via spectral methods in [25,43]' is stronger than what is shown: the paper treats the specific Kronecker–Gaussian sum (2) and its spectral edge, not general polynomials in independent Gaussian matrices.

Circularity Check

1 steps flagged · score 4.0 of 10

Lower-bound equality hinges on the RDT tightness principle imported from the author's own prior preprints; the upper bound is independently derived.

  1. self citation load bearing [Section 3.5, around equations (48)-(49) and the paragraph after (49)]
    "Similarly to [51], showing (48) establishes complete k-fold matrix analogues to Theorem 2.4 in [54] and Theorem 5.2 in [53]. The only additional thing that one has to ensure is that Q-overlap is adequately chosen. ... Since k and l are fixed they don’t impact concentrations and the remaining parts of the [53,54] methodologies automatically extend and ensure lim_{n→∞} 1/√n ED(1) = lim_{n→∞} 1/√n ED(0)."

    The lower-bound half of the main theorem, lim Eλ_n(H) = ρ_n, requires the interpolation limits to coincide: equations (93)-(94) use exactly lim (1/√n) ED(1) = lim (1/√n) ED(0). The only route supplied to this equality is the imported RDT tightness equivalence (48)-(49), which is not proved in this paper but taken from the author's own earlier preprints [47,48,51] and from Talagrand [53,54]. The extension to the present matrix-valued Stiefel setting, where the overlap Q is a k×k contraction and the F_i need not commute, is asserted in one sentence ('automatically extend') rather than derived.

full rationale

The paper does not fit parameters, does not rename a known result, and does not assume Lehner's formula in deriving the upper bound. The upper bound is obtained by a Slepian/Gordon comparison and a deterministic algebra identity cited from [18], which is independent external support. The main circularity concern is the lower bound: the equality of interpolation limits, which is necessary to conclude lim Eλ_n(H) = ρ_n, is imported from the author's own RDT preprints and from Talagrand, with the matrix-valued extension asserted rather than proved. This is a load-bearing self-citation, but the central claim still has substantial independent content (the upper-bound proof and the algebraic verification of condition (92) in Theorem 5). The score therefore reflects partial, not total, circularity.

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

No free parameters and no invented entities are introduced. The proof relies instead on two ad hoc technical assumptions: the imported RDT replicated-system tightness principle and an unstated positive-definiteness constraint in the dual minimization, plus standard comparison and concentration assumptions.

assumptions (4)
  • standard math Slepian-Gordon comparison theorem applies to centered Gaussian processes indexed by the Stiefel manifold X^T X = I.
    Invoked in Theorem 1 and Theorem 3 to compare the primal process with its dual, via equations (18)-(24) and (56)-(64); the index set is nonconvex but the theorem is used as the engine.
  • ad hoc to paper Strong duality holds for the Lagrangian (27)-(32) for the Stiefel-constrained linear optimization, with Gamma implicitly positive definite.
    Equation (28) says 'strong duality gives' without proof. The unrestricted Gamma = Gamma^T version is actually unbounded, so an unstated Gamma positive definite restriction is doing real work in the derivation.
  • ad hoc to paper RDT tightness: for Q in the overlap set, 2-Q-rep < 2 times (1-rep) is equivalent to equality of the interpolated limits in equations (48)-(49).
    This is the load-bearing lower-bound step. The paper says 'machineries of [47,48,51,53] give' and 'practically means' rather than proving the equivalence for the current matrix-valued process.
  • domain assumption Concentration of xi, L, D(t), and D^(2)(t) around their expectations as n goes to infinity for fixed k and l.
    Remark 1 and several inline comments assert concentration, but no quantitative concentration bounds are supplied.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An RDT based confirmation of Lehner's formula for Kronecker-Gaussian matrices." pith.science (2026). https://pith.science/paper/E5OTJVU7

@misc{pith2026260726551,
  author       = {Pith},
  title        = {Pith review of: An RDT based confirmation of Lehner's formula for Kronecker-Gaussian matrices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5OTJVU7}},
  note         = {Machine review of arXiv:2607.26551}
}
read the original abstract

Remarkable breakthroughs [25,43] established the so-called strong asymptotic freeness between classical Gaussian random ensembles and their semicircular free counterparts. Along the same lines, the Lehner formula [31], associated with the free counterpart, precisely determines the spectral edges of Kronecker-Gaussian matrices. We here revisit and study this formula without the utilization of random matrix theory and spectral methods. In particular, relying on concepts utilized within \emph{Random Duality Theory} (RDT) [45,46,51], we reconfirm Lehner's formula and effectively reprove the key asymptotic freeness results obtained via spectral methods in [25,43].

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 51 canonical work pages

  1. [31]

    F. Lehner. Computing norms of free operators with matrix coefficients.American Journal of Mathe- matics, 121(3):453–486, 1999

    work page 1999

  2. [18]

    Collins, A

    B. Collins, A. Guionnet, and F. Parraud. On the operator norm of non-commutative polynomials in deterministic matrices and iid GUE matrices.Cambridge Journal of Mathematics, 10(1):195–260, 2022

    work page 2022

  3. [1]

    Adamczak, A

    R. Adamczak, A. E. Litvak, A. Pajor, and N. Tomczak-Jaegermann. Quantitative estimates of the convergence of the empirical covariance matrix in log-concave ensembles.Journal of the American Mathematical Society, 23(2):535–561, 2010

    work page 2010

  4. [2]

    Adamczak, A

    R. Adamczak, A. E. Litvak, A. Pajor, and N. Tomczak-Jaegermann. Sharp bounds on the rate of convergence of the empirical covariance matrix.Comptes Rendus Mathématique, 349(3–4):195–200, 2011

    work page 2011

  5. [3]

    G. W. Anderson. Convergence of the largest singular value of a polynomial in independent Wigner matrices.Annals of Probability, 41(3B):2103–2181, 2013

    work page 2013

  6. [4]

    J. Baik, G. Ben Arous, and S. Peche. Phase transition of the largest eigenvalue for non-null complex sample covariance matrices.The Annals of Probability, 33(5):1643–1697, 2005

    work page 2005

  7. [5]

    A. S. Bandeira, G. Cipolloni, D. Schroder, and R. van Handel. Matrix concentration inequalities and free probability ii. Two-sided bounds and applications. 2024. available online athttp://arxiv.org/ abs/2406.11453

    arXiv 2024

  8. [6]

    Bandeira, M.T

    A.S. Bandeira, M.T. Boedihardjo, and R. van Handel. Matrix concentration inequalities and free probability.Inventiones Mathematicae, 234:419–487, 2023

    work page 2023

Show all 61 references
  1. [7]

    Barbier, N

    J. Barbier, N. Macris, and L. Miolane. The layered structure of tensor estimation and its mutual information. In2017 55th Annual Allerton Conference on Communication, Control, and Computing (Allerton), pages 1056–1063. IEEE, 2017. 20

    2017

  2. [8]

    J. K. Behne and G. Reeves. Fundamental limits for rank-one matrix estimation with groupwise het- eroskedasticity. InProceedings of The 25th International Conference on Artificial Intelligence and Statis- tics, volume 151 ofProceedings of Machine Learning Research, pages 8650–8...

    2022

  3. [9]

    S. T. Belinschi and M. Capitaine. Spectral properties of polynomials in independent wigner and deter- ministic matrices.Journal of Functional Analysis, 273(12):3901–3963, December 2017

    2017

  4. [10]

    Bordenave and B

    C. Bordenave and B. Collins. Eigenvalues of random lifts and polynomials of random permutation matrices.Ann. of Math. (2), 190(3):811–875, 2019

    2019

  5. [11]

    Bordenave and B

    C. Bordenave and B. Collins. Norm of matrix-valued polynomials in random unitaries and permutations

  6. [12]

    Bourgain

    J. Bourgain. Random points in isotropic convex sets. InConvex Geometric Analysis (Berkeley, CA, 1996), volume 34 ofMathematical Sciences Research Institute Publications, pages 53–58. Cambridge University Press, 1999

    1996

  7. [13]

    Brailovskaya and R

    T. Brailovskaya and R. van Handel. Universality and sharp matrix concentration inequalities.Geometric and Functional Analysis, 34:1003–1045, 2024

    2024

  8. [14]

    E. Cassidy. Random permutations acting onk-tuples have near-optimal spectral gap fork= poly(n)

  9. [15]

    C.-F. Chen, J. Garza-Vargas, J. A. Tropp, and R. van Handel. A new approach to strong convergence. Annals of Mathematics, 203(2):555–602, 2026

    2026

  10. [16]

    available online athttp://arxiv.org/abs/2412.13941

  11. [17]

    Collins and C

    B. Collins and C. Male. The strong asymptotic freeness of Haar and deterministic matrices.Annales scientifiques de l’Ecole Normale Superieure, 47(1):147–163, 2014

    2014

  12. [19]

    D. L. Donoho, M. Gavish, and I. M. Johnstone. Optimal shrinkage of eigenvalues in the spiked covariance model.The Annals of Statistics, 46(4):1742–1778, 2018

    2018

  13. [20]

    Collins and Y

    B. Collins and Y. Yamagishi. A Sudakov-Fernique proof of Lehner-type edge bounds for matrix-valued GUE sums. 2026. available online athttp://arxiv.org/abs/2606.21137

    2026 arXiv

  14. [21]

    Fernique

    X. Fernique. Regularite des trajectoires des fonctions aleatoires Gaussiens.Springer Lecture notes, 480:1–96, 1975

    1975

  15. [22]

    Fernique

    X. Fernique. Des resultats nouveaux sur les processus Gaussiens.C.R. Acad. Sci. Paris Ser A-B, 278:A363–A365, 1974

    1974

  16. [23]

    Y. Gordon. Some inequalities for Gaussian processes and applications.Israel Journal of Mathematics, 50(4):265–289, 1985

    1985

  17. [24]

    Giannopoulos, M

    A. Giannopoulos, M. Hartzoulaki, and A. Tsolomitis. Random points in isotropic unconditional convex bodies.Journal of the London Mathematical Society, 72(3):779–798, 2005

    2005

  18. [25]

    Haagerup and S

    U. Haagerup and S. Thorbjornsen. A new application of random matrices: Ext(c∗ red(f2))is not a group. Annals of Mathematics, 162(2):711–775, 2005

    2005

  19. [26]

    Guionnet and D

    A. Guionnet and D. Shlyakhtenko. Free diffusions and matrix models with strictly convex interaction. Geometric and Functional Analysis, 18(6):1875–1916, 2009

    1916

  20. [27]

    Hide and M

    W. Hide and M. Magee. Near optimal spectral gaps for hyperbolic surfaces.Annals of Mathematics, 198(2):791–824, 2023. 21

    2023

  21. [28]

    B. Hayes. A random matrix approach to the Peterson–Thom conjecture.Indiana Univ. Math. J., 71(3):1243–1297, 2022

    2022

  22. [29]

    Koltchinskii and K

    V. Koltchinskii and K. Lounici. Concentration inequalities and moment bounds for sample covariance operators.Bernoulli, 23(1):110–133, 2017

    2017

  23. [30]

    Kannan, L

    R. Kannan, L. Lovasz, and M. Simonovits. Random walks and ano∗(n5)volume algorithm for convex bodies.Random Structures & Algorithms, 11(1):1–50, 1997

    1997

  24. [32]

    D. Kunisky. Lehner’s operator norm formulas, semidefinite programming, and spiked matrix models

  25. [33]

    Lubetzky and Y

    E. Lubetzky and Y. Peres. Cutoff on all Ramanujan graphs.Geometric and Functional Analysis, 26(4):1148–1177, 2016

    2016

  26. [34]

    M. Magee. Strong convergence of unitary and permutation representations of discrete groups. 2025. available online athttp://arxiv.org/abs/2503.21619

    2025 arXiv

  27. [35]

    Lesieur, L

    T. Lesieur, L. Miolane, M. Lelarge, F. Krzakala, and L. Zdeborová. Statistical and computational phase transitions in spiked tensor estimation. In2017 IEEE International Symposium on Information Theory (ISIT), pages 511–515, 2017

    2017

  28. [36]

    Montanari and E

    A. Montanari and E. Richard. A statistical model for tensor PCA.Advances in Neural Information Processing Systems, 27, 2014

    2014

  29. [37]

    A. Pak, J. Ko, and F. Krzakala. Optimal algorithms for the inhomogeneous spiked Wigner model. In Advances in Neural Information Processing Systems, volume 36, pages 5557–5586, 2023

    2023

  30. [38]

    C. Male. The norm of polynomials in large random and deterministic matrices.Probability Theory and Related Fields, 154(3–4):659–750, 2012

    2012

  31. [39]

    F. Parraud. On the operator norm of non-commutative polynomials in deterministic matrices and iid Haar unitary matrices.Probability Theory and Related Fields, 182(3-4):1019–1073, 2022

    2022

  32. [40]

    Perry, A

    A. Perry, A. S. Wein, and A. S. Bandeira. Statistical limits of spiked tensor models.Annales de l’Institut Henri Poincare, Probabilites et Statistiques, 56(1):238–295, 2020

    2020

  33. [41]

    G. Paouris. Concentration of mass on convex bodies.Geometric and Functional Analysis, 16(5):1021– 1049, 2006

    2006

  34. [42]

    Rudelson

    M. Rudelson. Random vectors in the isotropic position.Journal of Functional Analysis, 164(1):60–72, 1999

    1999

  35. [43]

    H. Schultz. Non-commutative polynomials of independent Gaussian random matrices: The real and symplectic cases.Probability Theory and Related Fields, 131(2):261–309, 2005

    2005

  36. [44]

    Pourkamali, J

    F. Pourkamali, J. Barbier, and N. Macris. Matrix inference in growing rank regimes.IEEE Trans. Inf. Theory, 70(11):8133–8163, 2024

    2024

  37. [45]

    M. Stojnic. Various thresholds forℓ 1-optimization in compressed sensing. 2009. available online at http://arxiv.org/abs/0907.3666

    2009 arXiv

  38. [46]

    M. Stojnic. Regularly random duality. 2013. available online athttp://arxiv.org/abs/1303.7295

    2013 arXiv

  39. [47]

    D. Slepian. The one sided barier problem for Gaussian noise.Bell System Tech. Journal, 41:463–501, 1962

    1962

  40. [48]

    M. Stojnic. Generic and lifted probabilistic comparisons – max replaces minmax. 2016. available online athttp://arxiv.org/abs/1612.08506. 22

    2016 arXiv

  41. [49]

    M. Stojnic. A CLuP algorithm to practically achieve∼0.76SK–model ground state free energy.Journal of Statistical Mechanics: Theory and Experiment, (11):123302, 2025

    2025

  42. [50]

    M. Stojnic. Fully bilinear generic and lifted random processes comparisons. 2016. available online at http://arxiv.org/abs/1612.08516

    2016 arXiv

  43. [51]

    M. Stojnic. Precise sample covariance spectral norm error – an RDT view. 2026. available online at http://arxiv.org/abs/2607.14460

    2026 arXiv

  44. [52]

    V. N. Sudakov. Gaussian random processes and measures of solid angles in Hilbert space.Soviet Math. Dokl., 12(1):412–415, 1971

    1971

  45. [53]

    M. Stojnic. Binary perceptron computational gap – a parametric fl-RDT view.Journal of Statistical Mechanics: Theory and Experiment, (4):043301, 2026

    2026

  46. [54]

    Talagrand

    M. Talagrand. The Parisi formula.Annals of mathematics, 163:221–263, 01 2006

    2006

  47. [55]

    J. A. Tropp. User-friendly tail bounds for sums of random matrices.Foundations of Computational Mathematics, 11(4):373–434, 2011

    2011

  48. [56]

    Talagrand

    M. Talagrand. Free energy of the spherical mean field model.Probability Theory and Related Fields, 134:339–382, 3 2006

    2006

  49. [57]

    Vershynin

    R. Vershynin. Introduction to the non-asymptotic analysis of random matrices. InCompressed Sensing, pages 210–268. Cambridge University Press, 2012

    2012

  50. [58]

    R. A. Vitale. Some comparisons for Gaussian processes.Proceedings of the American Mathematical Society, 128(10):3043–3046, 2000. 23

    2000

  51. [59]

    van Handel

    R. van Handel. Strong convergence: A short survey. InProceedings of the International Congress of Mathematicians, pages 145–165. SIAM, 2026

    2026

  52. [2024]

    available online athttp://arxiv.org/abs/2304.05714

    arXiv

  53. [2026]

    available online athttp://arxiv.org/abs/2606.14687

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.