Some observations on the ambivalent role of symmetries in Bayesian inference problems

Guilhem Semerjian

REVIEW 1 major objections 3 minor 80 references

Some observations on the ambivalent role of symmetries in Bayesian inference problems

T0 review · 1 major / 3 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read In Bayesian inference, unobservable signal symmetries force a quotiented error metric; in extensive-rank matrix factorization, Bayes-optimal estimation becomes a three-level problem with a linear-assignment inner step.

desk verdict A clear, honest note that correctly identifies the quotient-distance issue in extensive-rank matrix factorization, but its strong claim that local statistical mechanics tools fail rests on an unproven no-spontaneous-breaking assumption. read the letter →

arxiv 2501.07975 v1 pith:NZ4UMFPY submitted 2025-01-14 cond-mat.dis-nn cond-mat.stat-mechcs.ITmath.ITmath.PRmath.STstat.TH

Guilhem Semerjian This is my paper · ORCID

classification cond-mat.dis-nncond-mat.stat-mechcs.ITmath.ITmath.PRmath.STstat.TH

keywords Bayesianinferencesymmetriesinproblemsquotienteddistanceextensive-rankmatrixfactorizationstatisticalmechanicsofdisorderedsystemslinearassignmentproblemequivariantestimatorslow-degreepolynomialmethod

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 argues that symmetries in Bayesian inference cut both ways. A symmetry acting strongly on the observations can be exploited: restricting estimators to equivariant functions reduces the computational dimension without hurting accuracy. A symmetry acting only on the signal is detrimental: all points on the same orbit are indistinguishable, so the error measure must be quotiented by the group. The paper shows that for extensive-rank matrix factorization this quotiented error turns Bayes-optimal estimation into a three-level problem that is not expressible through local posterior marginals. The upshot is that the standard replica, cavity, and I-MMSE machinery, which is built on local marginal probabilities and the mutual information, is not the right tool for this regime.

What carries the argument

The load-bearing object is the quotiented distance $d_G(S,\hat S)=\inf_{g\in G} d(g\cdot S,\hat S)$, which identifies all signals that differ by an unobservable symmetry. For extensive-rank matrix factorization with $G=\mathrm{S}_r$ acting on the columns of $X$, this distance becomes the squared Frobenius norms minus twice the maximum over column permutations of the trace of a matching matrix; that inner maximum is the bipartite matching (linear assignment) problem. It is this inner optimization that turns estimation into a 'three-level problem': minimize over the estimator, average over the posterior, and maximize over the group element. The paper also uses the rearrangement inequality in the $\mathrm{S}_N$ example to show that for some groups the quotiented distance collapses to an ordinary distance on ordered coordinates, which is why that case remains tractable.

What would settle it

Take $Y=XX^T+Z$ with $r=\alpha n$ and i.i.d. prior entries, and compare, at growing $n$, the risk of the estimator from Eq. (26) (computed by solving the inner assignment problem numerically) with the risk predicted by the symmetric replica, cavity, or approximate-message-passing solution. If the local-marginal solution already achieves the same quotiented risk, the paper's claim that local posterior marginals are inadequate is refuted; alternatively, show that the posterior overlap distribution conditioned on one observation is a single peak rather than a mixture over column permutations, and the assumed absence of spontaneous symmetry breaking would be false.

Watch

Extended reading notes

Core claim

The paper's central claim is that unobservable invariances must be gauged away before defining the distance between signal and estimator, and that doing so changes the structure of the Bayes-optimal estimator. For a signal $S$ with symmetry group $G$ leaving observations invariant, the correct distance is $d_G(S,\hat S)=\inf_{g\in G} d(g\cdot S,\hat S)$, and the optimal estimator minimizes $\mathbb{E}[d_G(S,x)\mid Y]$. In the extensive-rank matrix factorization problem $Y=XX^T+Z$ with column-i.i.d. prior, this gives the three-level formula of Eq. (26): minimize over $x$ the quantity $\|x\|^2-2\,\mathbb{E}[\max_{\pi\in\mathrm{S}_r}\sum_{i,\mu} X_{i,\mu}x_{i,\pi(\mu)}\mid Y]$. The inner maximization over permutations is a bipartite matching problem, so the estimator cannot be reduced to posterior expectations of simple functions of $X$. The paper concludes that the mutual information $I(S;Y)$ no longer controls the average risk through an I-MMSE-type relation, and that the graphical-model view of the posterior is an inadequate starting point in the extensive-rank regime.

Load-bearing premise

The conclusion that local statistical-mechanics tools fail for extensive rank rests on the assumption that the posterior distribution does not spontaneously split into components related by column permutations in the thermodynamic limit; if it did, symmetry-broken replica or cavity solutions could still compute the quotiented-distance estimator.

Editorial extensions

If this is right

For any inference model with an unobservable symmetry, using the naive square-error or Hamming distance makes the Bayes-optimal estimator trivial (for example, the posterior mean vanishes), so the quotiented distance is mandatory to extract any useful estimate.
In the symmetric stochastic block model, finite symmetry groups allow the computation through symmetry-broken pure states; the paper's framework explains why this route is available there but doubtful when the symmetry group grows with system size.
The mutual information $I(S;Y)$ and the minimum mean-square error are no longer linked by the I-MMSE relation once the distance is quotiented, so information-theoretic thresholds do not directly give the quotiented-distance risk.
The Bayes-optimal estimator for extensive-rank matrix factorization must solve a bipartite matching problem inside a posterior average; the paper suggests a message-passing solution of the assignment problem as the inner step of a multi-level statistical-mechanics treatment.
Symmetries acting on the observations remain an asset: restricting to equivariant low-degree polynomials reduces the dimension of the variational space without increasing the risk, as in matrix denoising.

Reading between the lines

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

This analysis implies that current approximate-message-passing state evolutions for extensive-rank matrix factorization, which track only low-dimensional order parameters, may be missing a genuinely matching-like order parameter; a testable consequence is that algorithms initialized with a permutation-aligned guess will outperform symmetric AMP at the same noise level.
The rearrangement-inequality example suggests a general design rule: when the symmetry group is a sorting group with totally ordered orbits, the quotiented distance becomes explicit and the optimal estimator is a posterior mean of an orbit representative; for groups like $\mathrm{S}_r$ acting on the columns of a matrix with $n>1$, no such ordering exists, which is a structural reason for the diffi
The same quotiented-distance prescription should apply to supervised learning with label-invariant symmetries: if a loss function compares predictions with unquotiented distances, the symmetry of the data-generation process makes the learning target inconsistent; quotienting the loss by the group is a direct, testable modification of equivariant-network training.
The three-level structure is likely to appear also in dictionary learning and blind source separation, where column permutation and sign ambiguities are unobservable; this suggests the difficulty is not specific to matrix factorization but generic to extensive-rank inverse problems.

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.

Desk Editor's Note

read the letter

Dear colleague,

Quick take: this is a clear, honest note, and the main observation is right. When a symmetry acts trivially on the observations, the error metric must be quotiented by that symmetry, and for extensive-rank matrix factorization that yields a three-level optimization with an assignment-problem inner max, not a local posterior-marginal computation. The paper's stronger methodological conclusion—that the replica/cavity/AMP machinery is inadequate—is not proven. It rests on the assumption that the extensive symmetry group Sr does not spontaneously break into pure states in the thermodynamic limit. The paper itself flags this as open in Sec. V, so the reader's conditional verdict is fair.

What is genuinely useful: the explicit formulation of Eq. (26) and its connection to the linear assignment problem is a good reformulation, and the order-statistics/permutation example is a neat illustration of a quotiented distance that still admits a closed-form estimator. The beneficial-symmetry discussion follows Hunt-Stein/equivariance arguments and is correctly credited. The paper does not oversell: Eq. (24) is attributed to [32], and the examples are mostly pedagogical. No parameters are fitted and no circularity burdens the argument.

Soft spots, in proportion: the statement that the optimal estimator is "not expressible in terms of local marginal laws" is asserted, not demonstrated. As the paper itself notes, in the finite-symmetry SBM case the same difficulty is resolved by spontaneous breaking: local marginals in one pure state do give the quotiented-distance estimator. Whether that can happen when r is proportional to n is genuinely open. No free-energy barrier or cluster-size bound is given. So the central claim is conditional on a negative answer to that question. That does not make the paper wrong; it makes it a programmatic note rather than a theorem.

Who should read it: anyone working on extensive-rank matrix factorization or on statistical mechanics of inference with unobservable symmetries. It deserves a serious referee, because the reformulation is useful and the open problem is well-posed. I would recommend acceptance as a note after the author either explicitly frames the local-marginal claim as conditional or adds a discussion of what a spontaneous-breaking scenario would imply.

Best, [Your name]

Referee Report

1 major / 3 minor

Summary. This note examines the role of symmetries in Bayesian inference, distinguishing beneficial symmetries that reduce the complexity of low-degree polynomial estimators from detrimental unobservable symmetries that require quotiented distances. The paper derives the Bayes-optimal estimator for the quotiented distance dG in a generic setting and specializes to the extensive-rank matrix factorization problem, where the column-permutation symmetry Sr leads to a three-level optimization problem (Eq. (26)) involving a bipartite matching inner problem. It argues that this estimator is not expressible in terms of local posterior marginals and therefore challenges the standard replica/cavity approach, while leaving the possibility of spontaneous symmetry breaking as an open question.

Significance. If the conclusions hold, the paper clarifies why the extensive-rank matrix factorization problem has resisted standard statistical mechanics treatments and provides a concrete starting point for future work. The derivations in Sec. IV are clean and correct, the examples are instructive, and the paper is honest about the open issues. The main value is conceptual; the paper does not prove the impossibility of local-marginal formulations, but it formulates the correct objective and highlights the role of the growing symmetry group.

major comments (1)

[Sec. IV (Eq. (26)) and Sec. V] The central methodological claim that the Bayes-optimal estimator for the quotiented distance is 'not expressible in terms of local marginal laws of the posterior probability PX|Y' is stated categorically in Sec. IV but depends on an unproven premise: the absence of spontaneous breaking of the extensive symmetry group Sr. The paper itself acknowledges in Sec. V that 'It is much less clear that such a phenomenon can occur when the symmetry group grows.' In the SBM case, spontaneous breaking of the finite symmetry group is precisely what allows the quotiented-distance estimator to be recovered from local marginals of a pure state. If an analogous phenomenon occurs for Sr with r proportional to n, the statement in Sec. IV would be false. The author should either provide an argument (e.g., a free-energy barrier estimate or a cluster-size bound) that spontaneous breaking cannot occur, or explicitly reformulate the claim as conditional on the symmetric posterior and discuss how symmetry-broken solutions of the replica/cavity equations would alter the conclusion.

minor comments (3)

[Sec. IV, last paragraph] The phrase 'local marginal laws of the posterior probability PX|Y' is ambiguous: it could mean the marginals of the full (symmetric) posterior or the marginals of a pure state. Clarifying this distinction would prevent a misreading, since the SBM example shows that marginals of a pure state can suffice.
[Abstract] The phrase 'the consequences this implies for the statistical mechanics treatment' is stronger than the exploratory tone of the note supports; consider softening to 'the potential consequences' or 'the challenges this poses.'
[Sec. IV, example with power sums] The observation model in Eq. (21), Y_p = \sum_i (S_i)^p, is artificial; a brief remark that it is chosen for illustrative purposes would help the reader calibrate the scope of the example.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper's central quotiented-distance estimator is derived from its own definitions, and its self-citation is not load-bearing.

full rationale

The paper makes no fitted predictions and reports no numerical benchmarks; its main formulas are obtained by direct substitution into definitional equations. In particular, Eq. (24) is the special case of the quotiented distance dG defined in Eq. (10) for the group Sr acting by column permutations, and Eq. (26) is obtained by inserting that distance into the general Bayes-optimal estimator formula (3); the paper explicitly says 'the Bayes-optimal estimator for the quotiented distance takes its generic form (12), which reads here (26)'. This is a derivation from prior definitions, not a prediction that reduces to an input. The attribution of Eq. (24) to [32] is a pointer to prior discussion, not independent support for the paper's own claim, and the formula is rederived in the text from Eq. (10). The self-citation of [31] occurs in Section III, where the Hunt-Stein reduction of low-degree polynomial estimators is cited alongside [42]; that discussion concerns the beneficial-symmetry side and is not load-bearing for the central claim about extensive-rank matrix factorization. The paper also explicitly flags its main open premise in Section V: 'It is much less clear that such a phenomenon can occur when the symmetry group grows in the thermodynamic limit'. That is an acknowledged limitation or correctness gap, but it is not a circular reduction of the paper's result to its assumptions. No step in the claimed derivation chain is equivalent to its inputs by construction, and no load-bearing conclusion is supported only by a self-citation chain.

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

No free parameters are fitted. The axioms listed are the background assumptions the paper invokes; the most fragile is the unproven possibility of spontaneous breaking of an extensive symmetry group, which the author acknowledges as open.

assumptions (5)

domain assumption The observer knows the prior PS and the channel PY|S.
This is the Bayesian inference framing established in Sec. II A; all optimal estimator computations assume these laws are known.
domain assumption The error metric and group action satisfy d(g*S, Shat) = d(S, g^{-1}*Shat) so that Eq. (10) is symmetric.
Stated in Sec. IV before Eq. (10); SE and Hamming distances with isometric actions satisfy it.
standard math Rearrangement inequality determines the permutation maximizing the scalar product in Eq. (19).
Used to derive Eq. (20); cited to theorem 368 in [67].
domain assumption For finite symmetry groups, the posterior measure decomposes into pure states related by the group, allowing the quotiented distance to be computed from symmetry-breaking replica/cavity solutions.
Used for SBM in Sec. IV; the extension to growing groups is left open in Sec. V.
standard math I-MMSE relation connects mutual information to SE risk for Gaussian channels.
Invoked in Sec. V to explain why the usual statistical mechanics toolbox computes SE risk; the paper argues this link fails for quotiented distances.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Some observations on the ambivalent role of symmetries in Bayesian inference problems." pith.science (2026). https://pith.science/paper/NZ4UMFPY

@misc{pith2026250107975,
  author       = {Pith},
  title        = {Pith review of: Some observations on the ambivalent role of symmetries in Bayesian inference problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NZ4UMFPY}},
  note         = {Machine review of arXiv:2501.07975}
}

read the original abstract

We collect in this note some observations on the role of symmetries in Bayesian inference problems, that can be useful or detrimental depending on the way they act on the signal and on the observations. We emphasize in particular the need to gauge away unobservable invariances in the definition of a distance between a signal and its estimator, and the consequences this implies for the statistical mechanics treatment of such models, taking as a motivating example the extensive rank matrix factorization problem.

Figures

Figures reproduced from arXiv: 2501.07975 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

80 extracted references · 74 canonical work pages

[32]

Matrix Inference in Growing Rank Regimes

F. Pourkamali, J. Barbier, and N. Macris. Matrix Infere nce in Growing Rank Regimes. arXiv:2306.01412, (2023)

work page Pith review arXiv 2023
[1]

Fulton and J

W. Fulton and J. Harris. Representation theory, a ﬁrst course . Springer, 1991

work page 1991
[2]

internal

= ||S||2 2 + || ˆS||2 2 − 2|⟨S, ˆS⟩| . (16) At variance with the N = 1 case the absolute value of the scalar product S1 ˆS1 + · · ·+ SN ˆSN cannot be factored as the product of a function of S and one of ˆS. As a consequence in the expression of the optimal estimator ˆSBO(Y ) ∈ arginf x∈ RN ( ||x||2 2 − 2 E [ |⟨S, x ⟩| ⏐ ⏐ ⏐ Y ]) , (17) the minimization o...

work page
[3]

inner problem

The solution of this “inner problem” would then provide a function o f ( S, x ), which should be averaged over S according to the posterior law S|Y , and ﬁnally minimized over x. Multi-level problems of a similar kind were studied with statistical mechanics methods in [70, 71]: the inner problem is so lved by a message passing algorithm (for the assignmen...

work page
[4]

Goodman and N

R. Goodman and N. R. Wallach. Symmetry, Representations, and Invariants . Springer, 2009

work page 2009
[5]

A. Zee. Group theory in a nutshell for physicists . Princeton University Press, 2016

work page 2016
[6]

Toulouse

G. Toulouse. Theory of Frustration Eﬀect in Spin-Glasse s: I. Comm. Phys. , 2, 115–119 (1977)

work page 1977
[7]

Toulouse

G. Toulouse. Symmetry and topology concepts for spin gla sses and other glasses. Phys. Rep. , 49, 267 (1979)

work page 1979

Show all 80 references

[8]

M. L. Eaton. Group Invariance Applications in Statistic s. Regional Conference Series in Probability and Statistics , 1, i–133 (1989)

1989
[9]

R. A. Wijsman. Invariant Measures on Groups and Their Use in Statistics. Lecture Notes-Monograph Series , 14, i–218 (1990)

1990
[10]

E. L. Lehmann and G. Casella. Theory of Point Estimation, 2d edition . Springer, 1998

1998
[11]

Shawe-Taylor

J. Shawe-Taylor. Building symmetries into feedforward networks. In 1989 First IEE International Conference on Artiﬁcial Neural Networks, (Conf. Publ. No. 313) , pages 158–162, 1989

1989
[12]

M. M. Bronstein, J. Bruna, T. Cohen, and P. Velickovic. G eometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges. arXiv:2104.13478, (2021)

2021 arXiv
[13]

Villar, D

S. Villar, D. W. Hogg, W. Yao, G. A. Kevrekidis, and B. Sch ¨ olkopf. Towards fully covariant machine learning.Transactions on Machine Learning Research , (2024)

2024
[14]

Cohen and M

T. Cohen and M. Welling. Group Equivariant Convolution al Networks. In M. F. Balcan and K. Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machin e Learning , volume 48 of Proceedings of Machine Learning Research, pages 2990–2999, New York, New York...

2016
[15]

S. Chen, E. Dobriban, and J. H. Lee. A Group-Theoretic Fr amework for Data Augmentation. Journal of Machine Learning Research, 21(245), 1–71 (2020)

2020
[16]

Simsek, F

B. Simsek, F. Ged, A. Jacot, F. Spadaro, C. Hongler, W. Ge rstner, and J. Brea. Geometry of the Loss Landscape in Overparameterized Neural Networks: Symmetries and Invari ances. In M. Meila and T. Zhang, editors, Proceedings of the 38th International Conference on Machine Lear...

2021
[17]

Nishimori

H. Nishimori. Statistical Physics of Spin Glasses and Information Proces sing: An Introduction . Oxford University Press, Oxford, UK, 2001

2001
[18]

Engel and C

A. Engel and C. Van den Broeck. Statistical Mechanics of Learning . Cambridge University Press, 2001

2001
[19]

M´ ezard and A

M. M´ ezard and A. Montanari. Physics, Information, Computation . Oxford University Press, Oxford, 2009

2009
[20]

Zdeborov´ a and F

L. Zdeborov´ a and F. Krzakala. Statistical physics of i nference: thresholds and algorithms. Advances in Physics , 65(5), 453–552 (2016)

2016
[21]

Charbonneau, E

P. Charbonneau, E. Marinari, M. M´ ezard, G. Parisi, F. R icci-Tersenghi, G. Sicuro, and F. Zamponi, editors. Spin Glass Theory and Far Beyond . World Scientiﬁc, 2023

2023
[22]

S. E. Fienberg and S. S. Wasserman. Categorical data ana lysis of single sociometric relations. Sociological methodology, 12, 156–192 (1981)

1981
[23]

P. W. Holland, K. B. Laskey, and S. Leinhardt. Stochasti c blockmodels: First steps. Social networks, 5(2), 109–137 (1983)

1983
[24]

Bollob´ as, S

B. Bollob´ as, S. Janson, and O. Riordan. The phase trans ition in inhomogeneous random graphs. Random Structures & Algorithms, 31(1), 3–122 (2007)

2007
[25]

Decelle, F

A. Decelle, F. Krzakala, C. Moore, and L. Zdeborov´ a. As ymptotic analysis of the stochastic block model for modular networks and its algorithmic applications. Phys. Rev. E , 84, 066106 (2011)

2011
[26]

C. Moore. The Computer Science and Physics of Community Detection: Landscapes, Phase Transitions, and Hardness. Bulletin of EATCS , 1(121) (2017)

2017
[27]

E. Abbe. Community Detection and Stochastic Block Mode ls: Recent Developments. Journal of Machine Learning Research, 18(177), 1–86 (2018)

2018
[28]

J. Bun, R. Allez, J.-P. Bouchaud, and M. Potters. Rotati onal Invariant Estimator for General Noisy Matrices. IEEE Transactions on Information Theory , 62(12), 7475–7490 (2016)

2016
[29]

Maillard, F

A. Maillard, F. Krzakala, M. M´ ezard, and L. Zdeborov´ a . Perturbative construction of mean-ﬁeld equations in exte nsive- rank matrix factorization and denoising. Journal of Statistical Mechanics: Theory and Experiment , 2022(8), 083301 (2022)

2022
[30]

Barbier and N

J. Barbier and N. Macris. Statistical limits of diction ary learning: Random matrix theory and the spectral replica method. Phys. Rev. E , 106, 024136 (2022)

2022
[31]

Troiani, V

E. Troiani, V. Erba, F. Krzakala, A. Maillard, and L. Zde borov´ a. Optimal denoising of rotationally invariant rect angular matrices. Proceedings of Mathematical and Scientiﬁc Machine Learnin g (MSML), PMLR , 190, 97–112 (2022)

2022
[33]

Semerjian

G. Semerjian. Matrix denoising: Bayes-optimal estima tors via low-degree polynomials. J. Stat. Phys. , 191, 139 (2024)

2024
[34]

Barbier, F

J. Barbier, F. Camilli, J. Ko, and K. Okajima. On the phas e diagram of extensive-rank symmetric matrix denoising bey ond 13 rotational invariance. arXiv:2411.01974, (2024)

2024 arXiv
[35]

Olshausen and D

B. Olshausen and D. Field. Emergence of simple-cell rec eptive ﬁeld properties by learning a sparse code for natural images. Nature, 381, 607–609 (1996)

1996
[36]

Zibulevsky and B

M. Zibulevsky and B. A. Pearlmutter. Blind source separ ation by sparse decomposition in a signal dictionary. Neural computation, 13(4), 863–882 (2001)

2001
[37]

Kreutz-Delgado, J

K. Kreutz-Delgado, J. F. Murray, B. D. Rao, K. Engan, T.- W. Lee, and T. J. Sejnowski. Dictionary learning algorithms for sparse representation. Neural computation, 15(2), 349–396 (2003)

2003
[38]

D. A. Spielman, H. Wang, and J. Wright. Exact Recovery of Sparsely-Used Dictionaries. In Proceedings of the 25th Annual Conference on Learning Theory , volume 23, pages 37.1–37.18, 2012

2012
[39]

Rangan and A

S. Rangan and A. K. Fletcher. Iterative estimation of co nstrained rank-one matrices in noise. In 2012 IEEE International Symposium on Information Theory Proceedings , pages 1246–1250, 2012

2012
[40]

Barbier, M

J. Barbier, M. Dia, N. Macris, F. Krzakala, T. Lesieur, a nd L. Zdeborov´ a. Mutual information for symmetric rank-on e matrix estimation: A proof of the replica formula. Advances in Neural Information Processing Systems 29 (NIPS 2016), pages 424–432 (2016)

2016
[41]

Lesieur, F

T. Lesieur, F. Krzakala, and L. Zdeborov´ a. Constraine d low-rank matrix estimation: phase transitions, approxim ate message passing and applications. Journal of Statistical Mechanics: Theory and Experiment , 2017(7), 073403 (2017)

2017
[42]

Lelarge and L

M. Lelarge and L. Miolane. Fundamental limits of symmet ric low-rank matrix estimation. Probab. Theory Relat. Fields , 173, 859–929 (2019)

2019
[43]

Barbier and N

J. Barbier and N. Macris. The adaptive interpolation me thod: a simple scheme to prove replica formulas in Bayesian inference. Probab. Theory Relat. Fields , 174, 1133–1185 (2019)

2019
[44]

Montanari and A

A. Montanari and A. S. Wein. Equivalence of Approximate Message Passing and Low-Degree Polynomials in Rank-One Matrix Estimation. arXiv:2212.06996, (2022)

2022 arXiv
[45]

Sakata and Y

A. Sakata and Y. Kabashima. Statistical mechanics of di ctionary learning. Europhysics Letters, 103(2), 28008 (2013)

2013
[46]

Kabashima, F

Y. Kabashima, F. Krzakala, M. M´ ezard, A. Sakata, and L. Zdeborov´ a. Phase Transitions and Sample Complexity in Bayes-Optimal Matrix Factorization. IEEE Transactions on Information Theory , 62(7), 4228–4265 (2016)

2016
[47]

H. C. Schmidt. Statistical Physics of Sparse and Dense M odels in Optimization and Inference. In PhD thesis , 2018

2018
[48]

Barbier, J

J. Barbier, J. Ko, and A. A. Rahman. A multiscale cavity m ethod for sublinear-rank symmetric matrix factorization. arXiv:2403.07189, (2024)

2024
[49]

Pourkamali and N

F. Pourkamali and N. Macris. Bayesian extensive-rank m atrix factorization with rotational invariant priors. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, pages 24025–24073, 2024

2024
[50]

Pourkamali and N

F. Pourkamali and N. Macris. Rectangular Rotational In variant Estimator for General Additive Noise Matrices. arXiv:2304.12264, (2023)

2023 arXiv
[51]

I. D. Landau, G. C. Mel, and S. Ganguli. Singular vectors of sums of rectangular random matrices and optimal estimati on of high-rank signals: The extensive spike model. Phys. Rev. E , 108, 054129 (2023)

2023
[52]

Camilli and M

F. Camilli and M. M´ ezard. Matrix factorization with ne ural networks. Phys. Rev. E , 107, 064308 (2023)

2023
[53]

Camilli and M

F. Camilli and M. M´ ezard. The decimation scheme for sym metric matrix factorization. Journal of Physics A: Mathematical and Theoretical, 57(8), 085002 (2024)

2024
[54]

Hastie, R

T. Hastie, R. Tibshirani, and J. Friedman. The elements of statistical learning, 2d edition . Springer, 2009

2009
[55]

L. P. Barnes, A. Dytso, J. Liu, and H. V. Poor. Multivaria te Priors and the Linearity of Optimal Bayesian Estimators under Gaussian Noise. arXiv:2401.16701, (2024)

2024 arXiv
[56]

I. G. Macdonald. Symmetric functions and Hall polynomials . Oxford University Press, 1998

1998
[57]

H. Weyl. The Classical Groups: Their Invariants and Representation s. Princeton University Press, 1966

1966
[58]

Schramm and A

T. Schramm and A. S. Wein. Computational barriers to est imation from low-degree polynomials. The Annals of Statistics , 50(3), 1833 – 1858 (2022)

2022
[59]

S. B. Hopkins and D. Steurer. Eﬃcient Bayesian Estimati on from Few Samples: Community Detection and Related Problems. In 2017 IEEE 58th Annual Symposium on Foundations of Computer S cience (FOCS), pages 379–390, 2017

2017
[60]

Kunisky, A

D. Kunisky, A. S. Wein, and A. S. Bandeira. Notes on Compu tational Hardness of Hypothesis Testing: Predictions Usin g the Low-Degree Likelihood Ratio. In Mathematical Analysis, its Applications and Computation , pages 1–50. Springer International Publishing, 2022

2022
[61]

Kunisky, C

D. Kunisky, C. Moore, and A. S. Wein. Tensor cumulants fo r statistical inference on invariant distributions. arXiv:2404.18735, (2024)

2024 arXiv
[62]

S. Mei, T. Misiakiewicz, and A. Montanari. Learning wit h invariances in random features and kernel models. In M. Bel kin and S. Kpotufe, editors, Proceedings of Thirty Fourth Conference on Learning Theory , volume 134 of Proceedings of Machine Learning Research, pages 3351–34...

2021
[63]

E. Abbe, J. M. Pereira, and A. Singer. Estimation in the G roup Action Channel. In 2018 IEEE International Symposium on Information Theory (ISIT) , pages 561–565, 2018

2018
[64]

A. S. Bandeira, B. Blum-Smith, J. Kileel, J. Niles-Weed , A. Perry, and A. S. Wein. Estimation under group actions: Recovering orbits from invariants. Applied and Computational Harmonic Analysis , 66, 236–319 (2023)

2023
[65]

B¨ urgisser, M

P. B¨ urgisser, M. L. Do˘ gan, V. Makam, M. Walter, and A. W igderson. Complexity of Robust Orbit Problems for Torus Actions and the abc-Conjecture. In R. Santhanam, editor, 39th Computational Complexity Conference (CCC 2024) , volume 300 of Leibniz International Proceedings i...

2024
[66]

Adrian, J

M. Adrian, J. Dubochet, J. Lepault, and A. W. McDowall. C ryo-electron microscopy of viruses. Nature, 308, 32 (1984)

1984
[67]

Singer and Y

A. Singer and Y. Shkolnisky. Three-Dimensional Struct ure Determination from Common Lines in Cryo-EM by Eigenvect ors 14 and Semideﬁnite Programming. SIAM Journal on Imaging Sciences , 4(2), 543–572 (2011)

2011
[68]

Sigworth

F. Sigworth. A Maximum-Likelihood Approach to Single- Particle Image Reﬁnement. Journal of Structural Biology , 122(3), 328–339 (1998)

1998
[69]

Hardy, J

G. Hardy, J. Littlewood, and G. P´ olya. Inequalities. Cambridge Mathematical Library. Cambridge University Pr ess, 1934

1934
[70]

Edmonds and R

J. Edmonds and R. M. Karp. Theoretical Improvements in A lgorithmic Eﬃciency for Network Flow Problems. J. ACM , 19(2), 248—-264 (1972)

1972
[71]

D. Guo, S. Shamai, and S. Verdu. Mutual information and m inimum mean-square error in Gaussian channels. IEEE Transactions on Information Theory , 51(4), 1261–1282 (2005)

2005
[72]

Altarelli, A

F. Altarelli, A. Braunstein, A. Ramezanpour, and R. Zec china. Stochastic optimization by message passing. Journal of Statistical Mechanics: Theory and Experiment , 2011(11), P11009 (2011)

2011
[73]

Castellana and L

M. Castellana and L. Zdeborov´ a. Adversarial satisﬁab ility problem. Journal of Statistical Mechanics: Theory and Experi- ment, 2011(03), P03023 (2011)

2011
[74]

Bayati, D

M. Bayati, D. Shah, and M. Sharma. Max-Product for Maxim um Weight Matching: Convergence, Correctness, and LP Duality. IEEE Trans. Inform. Theory , 54(3), 1241–1251 (2008)

2008
[75]

S. Mei, A. Montanari, and P.-M. Nguyen. A mean ﬁeld view o f the landscape of two-layer neural networks. Proceedings of the National Academy of Sciences , 115(33), E7665–E7671 (2018)

2018
[76]

Sarao Mannelli, G

S. Sarao Mannelli, G. Biroli, C. Cammarota, F. Krzakala , P. Urbani, and L. Zdeborov´ a. Marvels and Pitfalls of the Langevin Algorithm in Noisy High-Dimensional Inference. Phys. Rev. X , 10, 011057 (2020)

2020
[77]

Ben Arous, R

G. Ben Arous, R. Gheissari, and A. Jagannath. Online sto chastic gradient descent on non-convex losses from high- dimensional inference. Journal of Machine Learning Research , 22(106), 1–51 (2021)

2021
[78]

Gluch and R

G. Gluch and R. Urbanke. Noether: The More Things Change , the More Stay the Same. arXiv:2104.05508, (2021)

2021 arXiv
[79]

Hajjar and L

K. Hajjar and L. Chizat. On the symmetries in the dynamic s of wide two-layer neural networks. Electronic Research Archive, 31(4), 2175–2212 (2023)

2023
[80]

Dandi, E

Y. Dandi, E. Troiani, L. Arnaboldi, L. Pesce, L. Zdeboro v´ a, and F. Krzakala. The beneﬁts of reusing batches for grad ient descent in two-layer networks: breaking the curse of inform ation and leap exponents. In Proceedings of the 41st International Conference on Machine Lea...

2025

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [32]

Matrix Inference in Growing Rank Regimes

F. Pourkamali, J. Barbier, and N. Macris. Matrix Infere nce in Growing Rank Regimes. arXiv:2306.01412, (2023)

work page Pith review arXiv 2023

[2] [1]

Fulton and J

W. Fulton and J. Harris. Representation theory, a ﬁrst course . Springer, 1991

work page 1991

[3] [2]

internal

= ||S||2 2 + || ˆS||2 2 − 2|⟨S, ˆS⟩| . (16) At variance with the N = 1 case the absolute value of the scalar product S1 ˆS1 + · · ·+ SN ˆSN cannot be factored as the product of a function of S and one of ˆS. As a consequence in the expression of the optimal estimator ˆSBO(Y ) ∈ arginf x∈ RN ( ||x||2 2 − 2 E [ |⟨S, x ⟩| ⏐ ⏐ ⏐ Y ]) , (17) the minimization o...

work page

[4] [3]

inner problem

The solution of this “inner problem” would then provide a function o f ( S, x ), which should be averaged over S according to the posterior law S|Y , and ﬁnally minimized over x. Multi-level problems of a similar kind were studied with statistical mechanics methods in [70, 71]: the inner problem is so lved by a message passing algorithm (for the assignmen...

work page

[5] [4]

Goodman and N

R. Goodman and N. R. Wallach. Symmetry, Representations, and Invariants . Springer, 2009

work page 2009

[6] [5]

A. Zee. Group theory in a nutshell for physicists . Princeton University Press, 2016

work page 2016

[7] [6]

Toulouse

G. Toulouse. Theory of Frustration Eﬀect in Spin-Glasse s: I. Comm. Phys. , 2, 115–119 (1977)

work page 1977

[8] [7]

Toulouse

G. Toulouse. Symmetry and topology concepts for spin gla sses and other glasses. Phys. Rep. , 49, 267 (1979)

work page 1979