Hierarchical Maximum Entropy via the Renormalization Group

Amir R. Asadi

REVIEW 4 major objections 3 minor 34 references

Hierarchical Maximum Entropy via the Renormalization Group

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Hierarchical maximum entropy is solved by iterating coarse-graining and escort renormalization, with the optimum equal to a product of partition functions.

desk verdict The central variational identity in Theorem 7 is false as stated (a sigma_1 factor is dropped), but the parameter-flow examples in Section 5 are correct and the hierarchical maximum entropy idea is genuinely novel. read the letter →

arxiv 2509.01424 v1 pith:HRPW25A5 submitted 2025-09-01 cs.IT cs.LGmath.IT

Amir R. Asadi This is my paper · ORCID

classification cs.ITcs.LGmath.IT MSC 94A17

keywords hierarchicalmaximumentropyrenormalizationgroupGibbsvariationalprincipleDonsker–VaradhanrepresentationParetooptimalityparameterflowescortdistributionminimumrelative

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 introduces the "hierarchical maximum entropy" problem: given a sequence of coarse-graining maps, find a distribution that maximizes entropy at every level simultaneously, subject to a constraint on the expected loss. Its central claim is that the Pareto-optimal solutions are produced by a renormalization-group-like iteration: start from a Gibbs distribution, coarse-grain, renormalize with an escort transformation, and repeat; then disintegrate the resulting level distributions into one joint law. The paper proves a multilevel Gibbs-variational identity, $H(\sigma,T)(X) - \lambda E[L(X)] = \log Z(\lambda) - D(\sigma,T)(P_X \parallel \tilde{P}[\lambda]_X)$, which identifies the constructed law as the unique optimizer and the product of partition functions as the maximum value. This matters because it extends the classical Gibbs-Boltzmann characterization of maximum entropy to multiscale systems and reduces a multi-objective problem to a one-parameter root-finding condition. The paper also exhibits three model families—modular Gaussians, Dirichlet log-loss, and the nearest-neighbor Ising model—where the iteration collapses to a closed-form parameter flow, making the computation efficient.

What carries the argument

The engine is the renormalization operator $R(P; \theta)$, which replaces a density $p$ by $p^\theta / Z$—the escort transform—and Algorithm 1, which interleaves this with push-forward coarse-graining $T_i^\#$. The accumulated coefficients $\bar{\sigma}_i$ turn the hierarchical entropy into a weighted sum of conditional entropies, so that each induction step applies Lemma 4, a one-level variational identity for escort renormalization, and Lemma 5, a chain rule for relative entropy along a deterministic map. In the three invariant examples, the whole iteration reduces to low-dimensional recurrences: the $(A_i, B_i)$ sub-blocks of a modular precision matrix, the Dirichlet parameter $\alpha^{(i)}$, and the Ising coupling $\theta_i$. These recu

What would settle it

Take $d=2$ on a finite alphabet, choose any coarse-graining $T$ and coefficients $\sigma$ with $\bar{\sigma}_1 / \bar{\sigma}_2 \neq 1$, and run Algorithm 1. Construct $\tilde{P}[\lambda]$ by eq. (4.1); then compute the left and right sides of Theorem 7 directly. A mismatch—which occurs whenever the level-2 marginal of $\tilde{P}[\lambda]$ equals the pre-renormalization pushforward $U^{(1)}$ instead of $P^{(2)}$—settles whether the identity holds generally.

Watch

Extended reading notes

Core claim

The core discovery is Theorem 7, a hierarchical version of the Gibbs variational principle and the Donsker–Varadhan representation. For any weights $\sigma$ and trial distribution $P_X$, hierarchical entropy minus $\lambda$ times expected loss equals $\log Z(\lambda)$ minus a hierarchical KL divergence to a constructed distribution $\tilde{P}[\lambda]_X$. Algorithm 1 builds it: start from the Gibbs density at temperature $\lambda/\sigma_1$, then at each level push forward under the coarse-graining map and renormalize by raising the density to the power $\bar{\sigma}_i/\bar{\sigma}_{i+1}$. Disintegrating the level densities gives $\tilde{P}[\lambda]_X$, and $Z(\lambda)$ is the product of all normalization constants. Since hierarchical KL divergence is nonnegative, $\tilde{P}[\lambda]_X$

Load-bearing premise

The central identity assumes that the distributions produced by the iterative renormalization can be disintegrated into one joint law whose coarse-grained marginals coincide with those same renormalized distributions; if that consistency fails, the optimizer and the KL term in the variational identity are not matched.

Editorial extensions

If this is right

The hierarchical max-entropy problem becomes a constructive recipe: run Algorithm 1, disintegrate, and the result is the unique Pareto-optimal distribution; no separate multi-objective solver is needed.
The mean constraint is enforced by a one-dimensional condition, d log Z/dλ = −μ, so finding λ is root-finding rather than a full variational optimization.
For modular Gaussian, Dirichlet log-loss, and nearest-neighbor Ising models, computing the optimal distribution costs O(d) parameter-flow steps instead of operations on the full product space.
The same iteration solves hierarchical minimum relative entropy, with the generalized escort and the condition d log Z/dλ = μ.
Since the classical d=1 case is recovered, the framework is a strict generalization of the Gibbs-Boltzmann maximum entropy principle.

Reading between the lines

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

One consequence the author leaves implicit: any coarse-graining scheme with a known renormalization flow can be turned into a variational principle, so the framework suggests a recipe for scale-adapted regularizers in multiscale or deep models.
The parameter-flow recurrences resemble fixed-point equations; one could test whether hierarchical max-entropy distributions exhibit critical scaling, for example whether the Ising coupling θ_i flows to the same fixed point as the temperature is varied.
A finite-sample extension would replace entropies by plug-in estimates and use the parameter flows to define a hierarchical maximum-entropy prior, giving a sampler that avoids burn-in because the conditional layers factorize.
A reader extending the framework should first verify the consistency condition behind equation (4.1)—that the disintegrated joint's own level-i pushforwards match the renormalized P^{(i)}—since the theorem's KL term is defined against those objects.

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.

Desk Editor's Note

read the letter

Bottom line: the central variational identity, Theorem 7, is false as stated. The d=1 base case drops the sigma_1 factor on log Z_1, and the same factor error propagates through the induction, so Corollary 8 and Theorem 9 inherit it. That said, the paper is not empty: the hierarchical maximum entropy formulation is genuinely new, and the three parameter-flow examples in Section 5 are real, correctly computed RG calculations.

What's new: the Pareto formulation of multi-level entropy maximization, the idea of constructing optimizers via iterative renormalization plus disintegration, and the observation that under hierarchical invariance the RG iterations reduce to low-dimensional parameter flows. The three examples—modular Gaussians, Dirichlet aggregation, and Ising decimation—are presented cleanly and the recursions (5.1), (5.4), and Theorem 13's alpha-rule are correct. These are standard enough as RG facts, but having them laid out in one framework is useful, and the link to ResNet Hessian structure is a nice pointer.

The soft spots are structural. First, applying Lemma 2 to f = L/sigma_1 yields sigma_1 H - lambda E[L] = sigma_1 log Z_1 - sigma_1 D, not log Z_1 - sigma_1 D as in the proof. Because Z_1 is defined with exponent -lambda L/sigma_1, the factor cannot be absorbed into a normalization convention. The induction repeats the mistake via Lemma 4, so the claimed maximum log Z(lambda) and the lambda* condition are off by the accumulated sigma factors. This is not a typo.

Second, the construction of tilde P in (4.1) does not produce a joint whose level-i marginals are the P^(i)'s from Algorithm 1. Each renormalization step replaces the pushforward U^(i) with a different P^(i+1), so the reverse conditionals of the P^(i)'s do not assemble into a single joint consistent with those marginals. The hierarchical KL in Theorem 7 is then not a divergence against the claimed optimizer; the proof's final identification with Corollary 6 is unjustified.

Third, Theorem 3's "if and only if" overclaims: linear scalarization gives the Pareto front only under properness assumptions; the blanket statement is too strong.

The paper is a promising draft, not a finished paper. The parameter-flow section is a contribution worth keeping, and the hierarchical ME idea deserves pursuit. My recommendation: send it to a capable referee—the correct examples and the clear formal setup outweigh the false theorem for the purpose of getting a useful report. A referee should tell the author to fix the factor error, give a consistent disintegration-based construction, and resubmit.

Referee Report

4 major / 3 minor

Summary. The paper introduces "hierarchical maximum entropy" as a weighted sum of entropies of successive coarse-grained representations, and proposes to solve the scalarized problem (3.4) by a renormalization-group procedure (Algorithm 1) combined with a disintegration/reconstruction step (4.1). Its central result, Theorem 7, claims a hierarchical Gibbs variational identity; Corollary 8 then identifies the maximizer and maximum value log Z(λ), and Theorem 9 gives a Lagrange-multiplier condition d/dλ log Z(λ) = -μ. Sections 5 and 6 present parameter-flow recursions for quadratic, logarithmic, and nearest-neighbor loss functions, and a relative-entropy analogue.

Significance. If Theorem 7 were correct, the paper would establish a novel variational connection between hierarchical maximum entropy and renormalization-group theory, with explicit parameter flows in three nontrivial model classes. The Section 5 invariance arguments for modular Gaussians, Dirichlet distributions, and the one-dimensional Ising model are interesting and may be of independent value. However, the central identity is false as stated, and the reconstruction (4.1) does not produce the joint distribution whose marginals are claimed. The advertised variational principle and its corollaries are therefore not established, so the paper cannot be recommended in its current form.

major comments (4)

[§4, Theorem 7 and Eq. (4.5)] The base case of the induction is algebraically wrong. Applying Lemma 2 to f = L/σ1 gives H(X) − (λ/σ1)E[L] = log Z1 − D(P ∥ P^(1)); multiplying by σ1 yields σ1H − λE[L] = σ1 log Z1 − σ1 D(P ∥ P^(1)). The proof instead writes log Z1 − σ1D(P ∥ P^(1)). The same factor is lost inductively: Lemma 4 contributes σ_{k+1} log Z_{k+1}, not log Z_{k+1}. Thus Theorem 7, Corollary 8, and Theorem 9 are false for generic positive σ; even the form of Z(λ) would need revision.
[§4, Lemma 4] Lemma 4's statement is inconsistent with its own proof. Since q̃ = q^{θ/(1+θ)}/Z̃, D(P ∥ Q̃) = ∫ p log(p/q^{θ/(1+θ)}) + log Z̃. Substitution gives −(1+θ)∫ p log(p/q^{θ/(1+θ)}) = (1+θ)log Z̃ − (1+θ)D(P ∥ Q̃), not log Z̃ − (1+θ)D(P ∥ Q̃). The stated identity is therefore false, and because this lemma supplies the induction step of Theorem 7, the error is load-bearing.
[§4, Eq. (4.1) and Algorithm 1] The joint distribution P̃[λ]_X is constructed by chaining disintegrations of the iterates P^(i), but these P^(i) are not marginals of any single joint distribution. After coarse-graining, U^(i) = T_i#P^(i) is renormalized to P^(i+1) ≠ U^(i); the reverse kernel of P^(i) under T_i pushes P^(i) to U^(i), not to P^(i+1). Hence the level-(i+1) marginal of P̃ is not P^(i+1), and the final identification with D(σ,T)(P ∥ P̃) in Eq. (4.8) is unjustified. The same reconstruction defect propagates to Section 6, Theorem 17.
[§3, Theorem 3] The 'if and only if' claim with strictly positive weights is false in general for Pareto optimality of concave multiobjective problems. For d=2 with f1(x)=x and f2(x)=−x on [0,1], x=0 is Pareto-optimal, but every supporting weight vector has a zero component. The theorem should be restricted to properly Pareto-optimal points in the sense of Geoffrion, or to the Pareto points generated by positive scalarization, and the hierarchical max-entropy claims adjusted accordingly.

minor comments (3)

[§4, Eq. (4.5)] The index in the conditional-KL sum is garbled: for i=1 it refers to X^(0), which is undefined. The later line (4.8) uses D(P_{X^{(i+1)}|X^{(i)}} ∥ P^{(i)}_{X^{(i+1)}|X^{(i)}} | P_{X^{(i)}}); Eq. (4.5) should match that.
[§5.3] The sentence 'where σ̄_d = ··· = σ̄_1 > 0' is impossible because the accumulated coefficients are strictly increasing when all σ_i > 0. Presumably σ_1 = ··· = σ_d was intended.
[§2, Lemma 2] Regularity conditions on f and λ are not stated uniformly; the discrete and continuous cases are treated separately but the common hypothesis '∫ exp(−λf) dx < ∞' should also specify the range of λ and the reference measure.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the main hierarchical variational identity is a mathematical claim derived from stated lemmas, not a fitted or self-referential prediction. The only self-citation, to [2], is contextual and non-load-bearing.

full rationale

The paper's central derivation chain is not circular. Theorem 7 is an asserted identity of the form H(σ,T)(X) − λE[L(X)] = log Z(λ) − D(σ,T)(PX ∥ P̃[λ]_X), and it is derived from Lemma 2 (Gibbs variational principle) and Lemma 4 (a renormalization identity), both of which are stated and proved in the paper (Lemma 2 as a known external result, Lemma 4 with a direct proof). The distribution P̃[λ]_X is explicitly constructed via Algorithm 1 and disintegration (eq. 4.1), and the theorem claims this candidate is the maximizer (Corollary 8); constructing a candidate and proving it optimizes the objective is not circular. No parameter is fitted to data and then renamed as a prediction; no empirical quantity is predicted from its own fitting subset. The parameter-flow results in Section 5 are derived from standard closure properties of Gaussians, Dirichlet distributions, and the one-dimensional Ising model, with external citations for the classical identities used. The only self-citation is in Section 6: “In [2], it was shown that when T is a sequence of decimation transformations, the optimization problem (6.2) has a unique minimizer, which can be efficiently computed using the Marginalize-Tilt (MT) algorithm.” This is a prior result by the author (with Abbe), but it is used for context and motivation; the extension to general hierarchical transformations is proved in the present paper using Lemmas 15 and 16, which are cited to the external source [31] (Van Erven and Harremos). Thus the citation is not load-bearing. A separate reviewer concern, namely that the k=1 base case of the proof in Theorem 7 appears algebraically inconsistent with Lemma 2 (the σ1 factor is dropped from log Z1), would be a mathematical correctness issue, not a circularity issue: the proof does not assume its conclusion, and the alleged error is an internal mismatch, not an equivalence between inputs and outputs by construction. Therefore the circularity score is low.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. The free parameters are the user-chosen hierarchical weights σ_i. The main unstated assumptions are the correctness of the Pareto scalarization, the consistency of the disintegration construction, and the proper application of the classical variational principle.

free parameters (1)

hierarchical coefficients σ_i
Chosen by hand; they weight the entropy at each level and parameterize the claimed Pareto front. The λ* formulas depend explicitly on these weights.

assumptions (4)

domain assumption Standard Borel spaces with regular conditional probabilities and reverse random transformations exist
Invoked before eq. (4.1) to define disintegrations and \tilde P.
domain assumption Linear scalarization with positive weights constructs the entire Pareto front for concave objectives
Used in Theorem 3; generally requires proper Pareto optimality (Geoffrion efficiency), which is not stated or proven.
ad hoc to paper The iterated renormalized marginals P(i) can be reassembled into a single distribution on the original space
This is assumed by the definition of \tilde P in eq. (4.1) and is not true in general because renormalization changes the pushforward.
standard math Gibbs variational principle / Donsker-Varadhan representation (Lemma 2)
Standard background result used as the induction base; the paper misapplies it by dropping the weight factor.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hierarchical Maximum Entropy via the Renormalization Group." pith.science (2026). https://pith.science/paper/HRPW25A5

@misc{pith2026250901424,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Maximum Entropy via the Renormalization Group},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HRPW25A5}},
  note         = {Machine review of arXiv:2509.01424}
}

read the original abstract

Hierarchical structures, which include multiple levels, are prevalent in statistical and machine-learning models as well as physical systems. Extending the foundational result that the maximum entropy distribution under mean constraints is given by the exponential Gibbs-Boltzmann form, we introduce the framework of "hierarchical maximum entropy" to address these multilevel models. We demonstrate that Pareto optimal distributions, which maximize entropies across all levels of hierarchical transformations, can be obtained via renormalization-group procedures from theoretical physics. This is achieved by formulating multilevel extensions of the Gibbs variational principle and the Donsker-Varadhan variational representation of entropy. Moreover, we explore settings with hierarchical invariances that significantly simplify the renormalization-group procedures, enhancing computational efficiency: quadratic modular loss functions, logarithmic loss functions, and nearest-neighbor loss functions. This is accomplished through the introduction of the concept of parameter flows, which serves as an analog to renormalization flows in renormalization group theory. This work connects ideas from probability theory, information theory, and statistical mechanics.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 33 canonical work pages

[1]

User-friendly introduction to PAC-Bayes bounds

Pierre Alquier. User-friendly introduction to PAC-Bayes bounds. Foundations and Trends® in Machine Learning , 17(2):174–303, 2024. 18

work page 2024
[2]

Asadi and Emmanuel Abbe

Amir R. Asadi and Emmanuel Abbe. Chaining meets chain rule: Multilevel entropic regularization and training of neural networks. Journal of Machine Learning Research , 21(139):1–32, 2020

work page 2020
[3]

A simple probabilistic construction yielding generalized entropies and divergences, escort distributions and q-gaussians

Jean-Francois Bercher. A simple probabilistic construction yielding generalized entropies and divergences, escort distributions and q-gaussians. Physica A: Statistical Mechanics and its Applications, 391(19):4460–4469, 2012

work page 2012
[4]

A maximum entropy approach to natural language processing

Adam L Berger, Vincent J Della Pietra, and Stephen A Della Pietra. A maximum entropy approach to natural language processing. Computational Linguistics , 22(1):39–71, 1996

work page 1996
[5]

PAC-Bayesian supervised classification: The thermodynamics of statistical learning

Olivier Catoni. PAC-Bayesian supervised classification: The thermodynamics of statistical learning. Lecture Notes-Monograph Series, 56:i–163, 2007

work page 2007
[6]

I-divergence geometry of probability distributions and minimization prob- lems

Imre Csisz´ ar. I-divergence geometry of probability distributions and minimization prob- lems. The Annals of Probability , pages 146–158, 1975

work page 1975
[7]

Diffusion Schr¨ odinger bridge with applications to score-based generative modeling

Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion Schr¨ odinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems , 34:17695–17709, 2021

work page 2021
[8]

W. E. Principles of Multiscale Modeling . Cambridge University Press, 2011

work page 2011

Show all 34 references

[9]

Arnold M. Faden. The existence of regular conditional probabilities: necessary and sufficient conditions. The Annals of Probability , pages 288–298, 1985

1985
[10]

On the phase space approach to complexity.Journal of Statistical Physics, 69(1-2):411–425, 1992

Hans C Fogedby. On the phase space approach to complexity.Journal of Statistical Physics, 69(1-2):411–425, 1992

1992
[11]

Proper efficiency and the theory of vector maximization

Arthur M Geoffrion. Proper efficiency and the theory of vector maximization. Journal of Mathematical Analysis and Applications , 22(3):618–630, 1968

1968
[12]

Gr¨ unwald and A

Peter D. Gr¨ unwald and A. Philip Dawid. Game theory, maximum entropy, minimum discrepancy and robust bayesian decision theory. Annals of Statistics , 32(4):1367–1433, 2004

2004
[13]

Multiple objective decision making – methods and applications: a state-of-the-art survey , volume 164

C-L Hwang and Abu Syed Md Masud. Multiple objective decision making – methods and applications: a state-of-the-art survey , volume 164. Springer Science & Business Media, 2012

2012
[14]

Information theory and statistical mechanics

Edwin T Jaynes. Information theory and statistical mechanics. Physical Review, 106(4):620, 1957

1957
[15]

Information theory and statistical mechanics

Edwin T Jaynes. Information theory and statistical mechanics. ii. Physical Review , 108(2):171, 1957

1957
[16]

Renormalization group and probability theory

Giovanni Jona-Lasinio. Renormalization group and probability theory. Physics Reports, 352(4-6):439–458, 2001

2001
[17]

Numerical evaluations of the critical properties of the two-dimensional ising model

Leo P Kadanoff and Anthony Houghton. Numerical evaluations of the critical properties of the two-dimensional ising model. Physical Review B , 11(1):377, 1975

1975
[18]

Theory of Probability and Random Processes

Leonid Koralov and Yakov G Sinai. Theory of Probability and Random Processes. Springer Science & Business Media, 2007

2007
[19]

Demystifying resnet

Sihan Li, Jiantao Jiao, Yanjun Han, and Tsachy Weissman. Demystifying resnet. arXiv preprint arXiv:1611.01186, 2016. 19

2016 arXiv
[20]

Teaching the renormalization group

Humphrey J Maris and Leo P Kadanoff. Teaching the renormalization group. American Journal of Physics , 46(6):652–657, 1978

1978
[21]

Multiscale complexity of correlated gaussians

Richard Metzler and Yaneer Bar-Yam. Multiscale complexity of correlated gaussians. Phys- ical Review E , 71(4):046114, 2005

2005
[22]

Nonlinear multiobjective optimization , volume 12

Kaisa Miettinen. Nonlinear multiobjective optimization , volume 12. Springer Science & Business Media, 1999

1999
[23]

Recursion equations in gauge field theories

Alexander A Migdal. Recursion equations in gauge field theories. Sov. Phys. JETP , 42(3):413–418, 1975

1975
[24]

Envelope theorems for arbitrary choice sets

Paul Milgrom and Ilya Segal. Envelope theorems for arbitrary choice sets. Econometrica, 70(2):583–601, 2002

2002
[25]

Levinson and fast choleski algorithms for toeplitz and almost toeplitz matrices

Bruce Ronald Musicus. Levinson and fast choleski algorithms for toeplitz and almost toeplitz matrices. Research Laboratory of Electronics Technical Report, Massachusetts In- stitute of Technology https: // dspace. mit. edu/ handle/ 1721. 1/ 4954, 1988

1988
[26]

Using maximum entropy for text classification

Kamal Nigam, John Lafferty, and Andrew McCallum. Using maximum entropy for text classification. In IJCAI-99 workshop on machine learning for information filtering , vol- ume 1, pages 61–67. Stockholom, Sweden, 1999

1999
[27]

Entropic optimal transport: Convergence of potentials

Marcel Nutz and Johannes Wiesel. Entropic optimal transport: Convergence of potentials. Probability Theory and Related Fields , 184(1):401–424, 2022

2022
[28]

A maximum entropy approach to species distribution modeling

Steven J Phillips, Miroslav Dud ´ ık, and Robert E Schapire. A maximum entropy approach to species distribution modeling. In Proceedings of the twenty-first International Conference on Machine Learning , page 83, 2004

2004
[29]

A maximum entropy model for part-of-speech tagging

Adwait Ratnaparkhi. A maximum entropy model for part-of-speech tagging. In Conference on Empirical Methods in Natural Language Processing , 1996

1996
[30]

Principal information theoretic approaches

Ehsan S Soofi. Principal information theoretic approaches. Journal of the American Sta- tistical Association, 95(452):1349–1353, 2000

2000
[31]

R´ enyi divergence and Kullback-Leibler divergence

Tim Van Erven and Peter Harremos. R´ enyi divergence and Kullback-Leibler divergence. IEEE Transactions on Information Theory , 60(7):3797–3820, 2014

2014
[32]

The renormalization group and critical phenomena

Kenneth G Wilson. The renormalization group and critical phenomena. Reviews of Modern Physics, 55(3):583, 1983

1983
[33]

The renormalization group and the ϵ expansion

Kenneth G Wilson and John Kogut. The renormalization group and the ϵ expansion. Physics reports, 12(2):75–199, 1974

1974
[34]

Complexity and 1 /f noise: A phase space approach

Yi-Cheng Zhang. Complexity and 1 /f noise: A phase space approach. Journal de Physique I, 1(7):971–977, 1991. 20

1991

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

User-friendly introduction to PAC-Bayes bounds

Pierre Alquier. User-friendly introduction to PAC-Bayes bounds. Foundations and Trends® in Machine Learning , 17(2):174–303, 2024. 18

work page 2024

[2] [2]

Asadi and Emmanuel Abbe

Amir R. Asadi and Emmanuel Abbe. Chaining meets chain rule: Multilevel entropic regularization and training of neural networks. Journal of Machine Learning Research , 21(139):1–32, 2020

work page 2020

[3] [3]

A simple probabilistic construction yielding generalized entropies and divergences, escort distributions and q-gaussians

Jean-Francois Bercher. A simple probabilistic construction yielding generalized entropies and divergences, escort distributions and q-gaussians. Physica A: Statistical Mechanics and its Applications, 391(19):4460–4469, 2012

work page 2012

[4] [4]

A maximum entropy approach to natural language processing

Adam L Berger, Vincent J Della Pietra, and Stephen A Della Pietra. A maximum entropy approach to natural language processing. Computational Linguistics , 22(1):39–71, 1996

work page 1996

[5] [5]

PAC-Bayesian supervised classification: The thermodynamics of statistical learning

Olivier Catoni. PAC-Bayesian supervised classification: The thermodynamics of statistical learning. Lecture Notes-Monograph Series, 56:i–163, 2007

work page 2007

[6] [6]

I-divergence geometry of probability distributions and minimization prob- lems

Imre Csisz´ ar. I-divergence geometry of probability distributions and minimization prob- lems. The Annals of Probability , pages 146–158, 1975

work page 1975

[7] [7]

Diffusion Schr¨ odinger bridge with applications to score-based generative modeling

Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion Schr¨ odinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems , 34:17695–17709, 2021

work page 2021

[8] [8]

W. E. Principles of Multiscale Modeling . Cambridge University Press, 2011

work page 2011