arxiv: 2604.11922 · v1 · submitted 2026-04-13 · 🧮 math.PR · cs.LG· math.CO

Recognition: unknown

FlowBoost Reveals Phase Transitions and Spectral Structure in Finite Free Information Inequalities

Baran Hashemi

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:37 UTC · model grok-4.3

classification 🧮 math.PR cs.LGmath.CO

keywords finite free convolutionStam inequalityphase transitionsHermite polynomialsspectral structurecentral limit theoremdeep generative optimizationextremal problems

0 comments

The pith

FlowBoost optimization shows p=2 is the sharp critical exponent for finite free Stam inequalities, with the Hermite pair as unique equality case and a conjectured n-independent spectrum for the coupling matrix.

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

The paper deploys FlowBoost, a closed-loop deep generative optimization framework, to locate extremal real-rooted polynomials in ell p generalizations of the finite free Stam inequality under finite free additive convolution. At p=2 the optimizer recovers the Hermite pair as the sole equality case and extracts the spectral structure of the linearized convolution map at that point. This produces the conjecture that the singular values of the doubly stochastic coupling matrix E_n on the mean-zero subspace equal 2 to the power of minus k over two for k from 1 to n minus 1, independent of degree n. Accepting the conjecture immediately yields a sharp local stability constant and a finite free central limit theorem convergence rate that are both uniform in n. For p greater than 2 the Hermite pair itself violates the inequality with deficit controlled by the ell p contraction ratio of E_n, while for p less than 2 the extremal configurations bifurcate into non-matching pairs with bimodal root structure that recover the Hermite diagonal only in the limit as p approaches 2 from below.

Core claim

Using FlowBoost we find that the Hermite pair is the unique equality case of the p-Stam inequality at p=2 and that the singular values of the doubly stochastic coupling matrix E_n on the mean-zero subspace are conjectured to be 2^{-k/2} for k=1 to n-1 independently of n; conditional on this the local stability is sharp and the CLT converges uniformly; the inequality is violated by the Hermite pair for all p>2 with deficit controlled by the p-contraction ratio of E_n; for p<2 the extremal configurations bifurcate to non-matching bimodal pairs that converge to the Hermite diagonal only as p approaches 2 from below. Systematic FlowBoost computations support p^*=2 as the sharp critical exponent.

What carries the argument

The FlowBoost closed-loop deep generative optimization applied to extremal problems under finite free additive convolution, together with the doubly stochastic coupling matrix E_n whose singular values on the mean-zero subspace are conjectured to be 2^{-k/2}.

If this is right

The finite free central limit theorem converges at a rate that is uniform in the degree n.
A sharp local stability constant holds for the inequality at the critical value p=2.
The Hermite pair produces a positive deficit for every p greater than 2, governed by the ell p contraction ratio of E_n.
For p less than 2 the equality cases are non-matching pairs with bimodal root distributions rather than Hermite pairs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Verification of the singular value conjecture would supply exact constants for stability questions in other finite free inequalities.
The phase transition at p=2 indicates that the ell two norm is distinguished for stability under finite free convolution.
Generative optimization methods could be applied to discover equality cases and critical exponents in other extremal problems in analysis and probability.

Load-bearing premise

The FlowBoost closed-loop optimizer is assumed to locate the true global extremal structures and equality cases, including uniqueness at p=2 and the bifurcation for p less than 2, without missing other configurations.

What would settle it

A direct numerical computation of the singular values of E_n for moderate n that finds any deviation from the sequence 2^{-k/2} would disprove the spectral conjecture and the derived uniform stability and CLT rates.

Figures

Figures reproduced from arXiv: 2604.11922 by Baran Hashemi.

**Figure 1.** Figure 1: E-value-based AHT as used in this study. FlowBoost first produces elite extremal samples En,p. These are compressed into summary statistics T(En,p), compared against a library of low-complexity candidate families H, and screened against the Hermite baseline through the conditional e-values of (13). A candidate family is promoted only when its population-level fit remains stable across the tested grid and … view at source ↗

**Figure 2.** Figure 2: High-precision numerical evidence for the conjectured dyadic spectrum of the Hermite coupling operator En|W . Panel (a) plots log2 σk for the first ten singular values at the representative degrees n = 20, 40, 60, 90, together with the reference line log2 σk = −k/2. Panel (b) shows the maximum relative error among the first ten singular values across the high-precision restricted audit n ∈ {10, 20, 30, 40… view at source ↗

**Figure 3.** Figure 3: Representative convergence trajectories for the variance-normalized finite free central limit iteration at n = 20. Each colored curve tracks the perturbation norm dH(fk) for a different Hermite initialization, while the dashed line gives the reference decay (1/ √ 2)k predicted by the conjectural second singular value. After a short transient, the observed slopes are consistent with geometric contraction a… view at source ↗

**Figure 4.** Figure 4: Global computational phase diagram for the p-Stam problem on the tested (n, p) grid. The color in each cell is sgn(g min p ) log10(1 + |g min p |), so the sign records whether the best observed deficit is positive or negative and the magnitude records its logarithmic size. The sign transition occurs at p = 2 on the tested grid, while the post-processed numerical reporting in the text and tables is given in… view at source ↗

**Figure 5.** Figure 5: Root-configuration atlas for the best observed configurations at two representative degrees, n = 12 and n = 20, across the transition in p. In each panel, the blue and red rows show the jointly normalized root locations of α and β, respectively, while the annotated value gives the best observed additive deficit gp at that parameter pair. Read from left to right, the atlas exhibits a clear geometric progres… view at source ↗

**Figure 6.** Figure 6: Population-level order-statistic bands and gap analysis for the top100 elite samples at each displayed (n, p) pair. In the first three columns, each polynomial is normalized independently to mean 0 and unit variance, and then the median together with the 25–75% and 10–90% bands of the i-th sorted root are plotted separately for α and β. A two-block population appears here as a stable jump between root ind… view at source ↗

**Figure 7.** Figure 7: Compact summary of the subcritical bifurcation. Panel (a) contrasts a representative low-p root distribution with the quadratic Hermite geometry; Panel (b) tracks the best pair-mismatch statistic dP Q across p; and Panel (c) tracks the best Hermite distance D. Together the panels summarize the two main structural signatures of the transition: mismatch and non-Hermite geometry both persist well below p = 2 … view at source ↗

**Figure 8.** Figure 8: Stage-wise results for the p = 2 FlowBoost pipeline at representative degrees n = 6, 8, 10, 12, 20. SRP first identifies a high-quality feasible region, supervised CFM learns a proposal distribution from that region, and reward-guided fine-tuning sharpens the proposal distribution further so that the subsequent final push attains the best values. References [1] Mohammed Abouzaid, Andrew J. Blumberg, Marti… view at source ↗

read the original abstract

Using FlowBoost, a closed-loop deep generative optimization framework for extremal structure discovery, we investigate $\ell^p$-generalizations of the finite free Stam inequality for real-rooted polynomials under finite free additive convolution $\boxplus_n$. At $p=2$, FlowBoost finds the Hermite pair as the unique equality case and reveals the spectral structure of the linearized convolution map at this extremal point. As a result, we conjecture that the singular values of the doubly stochastic coupling matrix $E_n$ on the mean-zero subspace are ${2^{-k/2}:k=1,\ldots,n-1}$, independent of $n$. Conditional on this conjecture, we obtain a sharp local stability constant and the finite free CLT convergence rate, both uniform in $n$. We introduce a one-parameter family of $p$-Stam inequalities using $\ell^p$-Fisher information and prove that the Hermite pair itself violates the inequality for every $p>2$, with the sign of the deficit governed by the $\ell^p$-contraction ratio of $E_n$. Systematic computation via FlowBoost supports the conjecture that $p^*\!=2$ is the sharp critical exponent. For $p<2$, the extremal configurations undergo a bifurcation, meaning that they become non-matching pairs with bimodal root structure, converging back to the Hermite diagonal only as $p\to 2^-$. Our findings demonstrate that FlowBoost, can be an effective tool of mathematical discovery in infinite-dimensional extremal problems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper proves the Hermite pair violates the p-Stam inequality for p>2 and ties the deficit to the contraction ratio of E_n, but the spectral conjecture and p*=2 sharpness claim rest on trusting FlowBoost to find global extremals.

read the letter

The paper's clearest contribution is the proof that the Hermite pair violates the proposed p-Stam inequality for every p greater than 2. They link the sign of the deficit directly to the ell^p contraction ratio of the coupling matrix E_n. That part stands without needing the computations. They also introduce a one-parameter family of these inequalities and run FlowBoost to explore the extremal root structures. At p=2 it recovers the Hermite pair as unique equality case, and for p less than 2 it finds a bifurcation to non-matching bimodal pairs. From the p=2 case they conjecture that the singular values of E_n on the mean-zero subspace are exactly 2 to the power of minus k over 2, for k from 1 to n-1, and that this gives uniform stability and CLT rates. The soft spot is exactly where the stress-test note points: the spectral conjecture and the claim that p star equals 2 is sharp both rest on FlowBoost locating the global extremals reliably. The abstract gives no convergence diagnostics, error bars, or independent verification for small n. If the optimizer is missing other configurations, the conditional sharp constants fall apart. The proved violation is fine, but the bigger claims are more tentative. This work is for researchers in free probability who want to see how computational search can generate conjectures about finite free convolutions and information inequalities. A reader already thinking about stability in free CLTs might pick up the p-family idea or the bifurcation observation. I would send it to peer review. The violation proof is new and self-contained, so referees can evaluate that directly and comment on whether the numerics support the conjectures or need more rigor. It is worth the time even if revisions are needed on the computational side.

Referee Report

3 major / 3 minor

Summary. The manuscript introduces FlowBoost, a closed-loop deep generative optimization framework, to investigate ℓ^p-generalizations of the finite free Stam inequality for real-rooted polynomials under finite free additive convolution ⊞_n. It rigorously proves that the Hermite pair violates the inequality for every p>2, with the sign of the deficit determined by the ℓ^p-contraction ratio of the coupling matrix E_n. At p=2, FlowBoost computations identify the Hermite pair as the unique equality case and reveal spectral structure, leading to the conjecture that the singular values of the doubly stochastic matrix E_n on the mean-zero subspace are 2^{-k/2} for k=1 to n-1, independent of n. Conditional on this conjecture, sharp local stability constants and uniform finite-free CLT convergence rates are derived. Further computations indicate a bifurcation to non-matching bimodal pairs for p<2, supporting the claim that p^*=2 is the sharp critical exponent.

Significance. If the spectral conjecture holds, the work would deliver sharp, n-uniform constants for local stability and finite-free CLT rates, advancing finite free probability and information inequalities. The explicit proof of the p>2 violation (independent of numerics) is a solid contribution. The computational discovery of phase transitions and equality cases illustrates FlowBoost's potential as a discovery tool in infinite-dimensional extremal problems, though its global optimality requires additional validation.

major comments (3)

[Abstract and spectral conjecture section] Abstract and the section presenting the spectral conjecture: the claim that singular values of E_n are exactly 2^{-k/2} (k=1,...,n-1) independent of n, and the resulting sharp local stability constant and finite-free CLT rate, rest solely on FlowBoost outputs without reported error bars, convergence diagnostics, or independent verification (e.g., exhaustive enumeration for small n or analytic bounds).
[p<2 bifurcation and sharpness section] The section on p<2 extremal configurations: the bifurcation to non-matching bimodal pairs and the assertion that p^*=2 is sharp (supported by uniqueness of the Hermite pair at p=2) depend on FlowBoost reliably locating global extremals; no certificate rules out missed configurations or local minima in the closed-loop optimizer.
[Computational experiments] The computational experiments supporting the equality cases: while the p>2 violation proof is load-bearing and independent, the overall central claims on sharpness and spectral structure are conditional on unverified global optimality of FlowBoost, which is an invented entity without external reproducibility guarantees.

minor comments (3)

[Introduction] The introduction could include a brief reminder of the definition of finite free additive convolution ⊞_n and the standard Stam inequality to improve accessibility.
[Notation and preliminaries] Notation for the coupling matrix E_n and the mean-zero subspace should be defined explicitly at first use, with a reference to prior finite free probability literature.
[Abstract and conclusion] The abstract and conclusion could more clearly separate the rigorously proved violation for p>2 from the conjectural elements.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the careful reading and constructive comments, which highlight the numerical foundations of several claims. We address each major point below, agreeing to add diagnostics, reproducibility measures, and explicit caveats on the conditional nature of the results. The rigorous p>2 violation proof remains independent of numerics.

read point-by-point responses

Referee: [Abstract and spectral conjecture section] Abstract and the section presenting the spectral conjecture: the claim that singular values of E_n are exactly 2^{-k/2} (k=1,...,n-1) independent of n, and the resulting sharp local stability constant and finite-free CLT rate, rest solely on FlowBoost outputs without reported error bars, convergence diagnostics, or independent verification (e.g., exhaustive enumeration for small n or analytic bounds).

Authors: We acknowledge that the spectral conjecture and derived constants rest on numerical evidence. In revision we will add error bars from repeated runs with varied seeds, optimizer convergence diagnostics, and independent verification for small n (n≤5) via direct computation of E_n and exhaustive search over low-degree polynomials. The text will explicitly state that the conjecture and sharp constants are conditional on the numerical observations, pending analytic confirmation. revision: partial
Referee: [p<2 bifurcation and sharpness section] The section on p<2 extremal configurations: the bifurcation to non-matching bimodal pairs and the assertion that p^*=2 is sharp (supported by uniqueness of the Hermite pair at p=2) depend on FlowBoost reliably locating global extremals; no certificate rules out missed configurations or local minima in the closed-loop optimizer.

Authors: We agree that the observed bifurcation and sharpness of p^*=2 lack a global optimality certificate. We will revise the section to report results from multiple independent runs with randomized initializations, discuss the risk of local minima, and qualify the sharpness claim as numerically supported by the combination of the proven p>2 violation, the p=2 equality case, and the consistent bifurcation pattern, rather than asserted as proven. revision: partial
Referee: [Computational experiments] The computational experiments supporting the equality cases: while the p>2 violation proof is load-bearing and independent, the overall central claims on sharpness and spectral structure are conditional on unverified global optimality of FlowBoost, which is an invented entity without external reproducibility guarantees.

Authors: We note the distinction drawn for the rigorous p>2 proof. In revision we will release the full FlowBoost implementation, hyperparameters, and experimental scripts for reproducibility. A new subsection will detail the methodology and validation procedures. The abstract and main text will be updated to state clearly that claims on spectral structure, uniqueness at p=2, and the phase transition are conditional on the numerical findings. revision: yes

standing simulated objections not resolved

A rigorous analytic proof or certificate establishing global optimality of FlowBoost or exactness of the spectral conjecture on the singular values of E_n.

Circularity Check

0 steps flagged

No significant circularity; conjectures are computational discoveries, not reductions by construction

full rationale

The paper introduces FlowBoost as an external optimization tool for discovering extremal structures and equality cases in the finite free Stam inequalities. It then explicitly conjectures the singular values of E_n from those numerical outputs and derives conditional stability constants and convergence rates. The violation of the p-Stam inequality for p>2 is proved directly for the Hermite pair without relying on the conjecture. No step equates a claimed prediction or first-principles result to its own inputs by definition, fitted parameters, or self-referential closure. The computational support for p^*=2 is presented as empirical evidence for a conjecture, not as an algebraic identity or load-bearing self-citation. The derivation chain therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The paper rests on standard background from finite free probability (real-rooted polynomials and finite free convolution) and introduces a new computational framework whose reliability is not independently verified in the abstract.

axioms (2)

domain assumption Real-rooted polynomials are closed under finite free additive convolution and admit well-defined Fisher information functionals
Invoked throughout the study of ℓ^p-Stam inequalities and equality cases.
domain assumption The linearized convolution map at the Hermite pair admits a spectral decomposition whose singular values can be extracted numerically
Used to motivate the conjecture on the values 2^{-k/2}.

invented entities (1)

FlowBoost no independent evidence
purpose: Closed-loop deep generative optimization framework for discovering extremal structures in infinite-dimensional problems
New method introduced to locate equality cases and phase transitions that are otherwise intractable.

pith-pipeline@v0.9.0 · 5568 in / 1679 out tokens · 94823 ms · 2026-05-10T15:37:35.684959+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 4 canonical work pages · 1 internal anchor

[1]

Blumberg, Martin Hairer, Joe Kileel, Tamara G

Mohammed Abouzaid, Andrew J. Blumberg, Martin Hairer, Joe Kileel, Tamara G. Kolda, Paul D. Nelson, Daniel Spielman, Nikhil Srivastava, Rachel Ward, Shmuel Weinberger, and Lauren Williams. First Proof,
[2]

arXiv:2602.05192.https://arxiv.org/abs/2602.05192

work page arXiv
[3]

Flow-based extremal mathematical structure discovery,

Gergely Bérczi, Baran Hashemi, and Jonas Klüver. Flow-based extremal mathematical structure discovery,
[4]

Can transformers do enumerative geome- try? InThe Thirteenth International Conference on Learning Representations, 2025.https://openreview

Baran Hashemi, Roger Guigo Corominas, and Andrea Giacchetto. Can transformers do enumerative geome- try? InThe Thirteenth International Conference on Learning Representations, 2025.https://openreview. net/forum?id=4X9RpKH4Ls

2025
[5]

Finite free information inequalities, 2026

Jorge Garza-Vargas, Nikhil Srivastava, and Zoe Stier. Finite free information inequalities, 2026. arXiv:2602.15822

work page arXiv 2026
[6]

Cumulants for finite free convolution.Journal of Combinatorial The- ory, Series A, 155:244–266, 2018

Octavio Arizmendi and Daniel Perales. Cumulants for finite free convolution.Journal of Combinatorial The- ory, Series A, 155:244–266, 2018

2018
[7]

Marcus, Daniel A

Adam W. Marcus, Daniel A. Spielman, and Nikhil Srivastava. Finite free convolutions of polynomials.Prob- ability Theory and Related Fields, 182:807–848, 2022

2022
[8]

J. L. Walsh. On the location of the roots of certain types of polynomials.Transactions of the American Mathematical Society, 24:163–180, 1922

1922
[9]

Bauschke, Osman Güler, Adrian S

Heinz H. Bauschke, Osman Güler, Adrian S. Lewis, and Hristo S. Sendov. Hyperbolic polynomials and convex analysis.Canadian Journal of Mathematics, 53:470–488, 2001

2001
[10]

N. M. Blachman. The convolution inequality for entropy powers.IEEE Transactions on Information Theory, 11:267–271, 1965

1965
[11]

Bobkov and Gennadii P

Sergey G. Bobkov and Gennadii P. Chistyakov. Entropy power inequality for the Rényi entropy.IEEE Transactions on Information Theory, 61:708–714, 2015

2015
[12]

Bobkov and Augusto Marsiglietti

Sergey G. Bobkov and Augusto Marsiglietti. Variants of the entropy power inequality.IEEE Transactions on Information Theory, 63:7747–7752, 2017

2017
[13]

The analogues of entropy and of Fisher’s information measure in free probability theory, I

Dan Voiculescu. The analogues of entropy and of Fisher’s information measure in free probability theory, I. Communications in Mathematical Physics, 155:71–92, 1993

1993
[14]

Szarek and Dan Voiculescu

Stanisław J. Szarek and Dan Voiculescu. Volumes of restricted Minkowski sums and the free analogue of the entropy power inequality.Communications in Mathematical Physics, 178:563–570, 1996

1996
[15]

A free analogue of Shannon’s problem on monotonicity of entropy.Advances in Math- ematics, 208:824–833, 2007

Dimitri Shlyakhtenko. A free analogue of Shannon’s problem on monotonicity of entropy.Advances in Math- ematics, 208:824–833, 2007

2007
[16]

Gomez, Łukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems 30, pages 5998–6008, 2017

2017
[17]

Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling, 2022. arXiv:2210.02747. FLOWBOOST AND PHASE TRANSITIONS IN FREE INFORMATION INEQUALITIES 21

work page internal anchor Pith review Pith/arXiv arXiv 2022
[18]

Testing by betting: a strategy for statistical and scientific communication.Journal of the Royal Statistical Society: Series A, 184:407–431, 2021

Glenn Shafer. Testing by betting: a strategy for statistical and scientific communication.Journal of the Royal Statistical Society: Series A, 184:407–431, 2021

2021
[19]

E-values: calibration, combination, and applications.The Annals of Statistics, 49:1736–1754, 2021

Alexander Vovk and Ruodu Wang. E-values: calibration, combination, and applications.The Annals of Statistics, 49:1736–1754, 2021

2021
[20]

Howard, Aaditya Ramdas, Jon McAuliffe, and Jasjeet Sekhon

Steven R. Howard, Aaditya Ramdas, Jon McAuliffe, and Jasjeet Sekhon. Time-uniform, nonparametric, nonasymptotic confidence sequences.The Annals of Statistics, 49:1055–1080, 2021

2021
[21]

mpmath: a Python library for arbitrary-precision floating-point arithmetic (version 1.4.0), 2026.http://mpmath.org/

The mpmath development team. mpmath: a Python library for arbitrary-precision floating-point arithmetic (version 1.4.0), 2026.http://mpmath.org/

2026
[22]

Adam W. Marcus. Polynomial convolutions and (finite) free probability, 2021. arXiv:2108.07054

work page arXiv 2021
[23]

Ball, Franck Barthe, and Assaf Naor

Shiri Artstein, Keith M. Ball, Franck Barthe, and Assaf Naor. Solution of Shannon’s problem on the mono- tonicity of entropy.Journal of the American Mathematical Society, 17:975–982, 2004

2004
[24]

Andrew R. Barron. Entropy and the central limit theorem.The Annals of Probability, 14:336–342, 1986. Max Planck Institute for Mathematics in the Sciences, Leipzig, Germany Email address:baran.hashemi@mis.mpg.de

1986