Stochastic Localization with Non-Gaussian Tilts and Applications to Tensor Ising Models

Arianna Piana; Dan Mikulincer

REVIEW 3 major objections 4 minor 38 references

Stochastic Localization with Non-Gaussian Tilts and Applications to Tensor Ising Models

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper extends stochastic localization to non-Gaussian tilts and proves that tensor Ising measures satisfy the explicit bound $C_P(\mu) \le 1/(1-56n\|T\|_{\mathrm{inj}})$, giving rapid mixing for Glauber dynamics whenever…

desk verdict A genuinely new localization tool with explicit tensor Ising bounds; the central proof survives the stress-test, though the statement needs cleanup. read the letter →

arxiv 2412.12720 v2 pith:LF4LSY7O submitted 2024-12-17 math.PR math-phmath.FAmath.MP

Dan Mikulincer , Arianna Piana This is my paper · ORCID

classification math.PRmath-phmath.FAmath.MP MSC 60K3560J2782B2060H10

keywords stochasticlocalizationnon-GaussiantiltstensorIsingmodelspectralgapGlauberdynamicsPoincaréinequalitymixingtimeBooleanhypercube

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 asks whether stochastic localization, a technique that simplifies a complicated probability measure by tilting it with Gaussians, can work when the tilt is not Gaussian, and whether that helps for genuinely non-quadratic models. It answers yes for quartic potentials on the Boolean hypercube: a new tensorized localization process decomposes any tensor Ising measure into a mixture of simple low-rank pieces, up to an arbitrarily small error. From that decomposition it proves an explicit variance bound, $C_P(\mu) \le 1/(1 - 56 n \|T\|_{\mathrm{inj}})$ whenever $\|T\|_{\mathrm{inj}} \le 1/(56n)$. Since the Poincaré constant is the reciprocal of the spectral gap of Glauber dynamics, this yields rapid mixing throughout a high-temperature regime with explicit constants. The interest is that explicit, checkable spectral-gap bounds for non-quadratic spin models were previously available only with implicit constants or in special cases.

What carries the argument

The load-bearing object is the tensorized stochastic localization process: a continuous-time evolution of relative densities $dF_t(x) = \langle x^{\otimes 2} - v_t, C_t\, dW_t\rangle F_t(x)$, where $W_t$ is a Brownian motion in the space of symmetric matrices, $C_t$ is a smoothed projection onto the image of the remaining tensor potential $T_t = T - \tfrac12\int_0^t C_s^2\, ds$, and $v_t$ is an adapted drift chosen so that the process $X_t$ solving $dX_t = C_t\, dW_t - C_t^2 v_t\, dt$ stays inside a ball of radius $\delta$. This drift cancels the unwanted quadratic tilt that would otherwise destroy the spectral gap, and the projection keeps the normalized variance nearly constant along the flow. Stopping when the remaining tensor has rank one converts the measure into a mixture of low-rank quartic measures, and the error $\delta$ is then absorbed into the final bound.

What would settle it

Take the rank-one tensor $T(x) = \beta(\sum_i x_i)^4/n^3$ on $\{\pm1\}^n$ with $\beta = 0.016$, so that $\|T\|_{\mathrm{inj}} = 0.016/n$ is below $1/(56n)$, and compute the true Poincaré constant of Glauber dynamics by exact diagonalization for $n$ up to around 20. If the constant exceeds $1/(1 - 56n\|T\|_{\mathrm{inj}}) \approx 9.6$ at any such $n$, Theorem 1.2 is false; if it stays below, the bound survives this direct test.

Watch

Extended reading notes

Core claim

The central claim, stated as Theorem 1.2, is that for a measure $\mu(x) \propto \exp(T(x))$ on $\{\pm 1\}^n$ with $T$ a positive definite symmetric fourth-order tensor, the Poincaré constant obeys $C_P(\mu) \le 1/(1 - 56 n \|T\|_{\mathrm{inj}})$ provided $\|T\|_{\mathrm{inj}} \le 1/(56 n)$. The argument runs through a decomposition theorem: the measure is split, up to an arbitrarily small error, into a mixture of non-negative measures of the form $\exp(\langle u,x\rangle^2 \langle v,x\rangle^2 + \langle w,x\rangle^2 + \langle \ell,x\rangle + \psi(x))$, with the vectors bounded by $2\sqrt{\|T\|_{\mathrm{inj}}}$ or $4n\|T\|_{\mathrm{inj}}$ and $|\psi| \le \delta$ almost surely. Each low-rank piece is shown to satisfy a variance bound by a derivative-matrix version of the influence-matrix criterion, and concavity of the Glauber Dirichlet form reassembles the pieces. The same machinery yields an explicit high-temperature bound for Gaussian degree-4 spin glasses and a transition statement for the rank-one tensor Curie-Weiss model, where mixing becomes exponentially slow above $\beta \approx 0.504$.

Load-bearing premise

The proof rests on the existence of an adapted drift $v_t$ that keeps the auxiliary process $X_t$ inside a ball of radius $\delta$ forever; if that drift construction fails, the leftover term in the decomposition is no longer small and the spectral gap bound collapses.

Editorial extensions

If this is right

For every positive definite degree-4 tensor with $\|T\|_{\mathrm{inj}} \le 1/(56n)$, Glauber dynamics on $\{\pm1\}^n$ has spectral gap at least $1 - 56n\|T\|_{\mathrm{inj}}$ and therefore polynomial mixing time.
For Gaussian degree-4 spin glasses with $N(0,1/n^3)$ entries, the same route gives $C_P(\mu) \le 1/(1 - 200.928\,\beta)$ at inverse temperatures up to about $1/200.928$.
The rank-one tensor Curie-Weiss model is covered on both sides: rapidly mixing for $\beta \le 1/56$ and exponentially slow above $\beta \approx 0.504$, so the explicit constant is within roughly a factor of 28 of the true threshold.
The decomposition transfers to the sphere and to higher-degree tensors, with the spectral-gap guarantee degrading rapidly in the degree.
Because the decomposition reduces arbitrary tensors to low-rank pieces, any future modified log-Sobolev or mixing result proved for the low-rank pieces would automatically extend to the full tensor Ising model.

Reading between the lines

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

The explicit constant $1/56$ is likely improvable: the rank-one tensor Curie-Weiss example suggests the true threshold is about 28 times larger, so a sharper influence-matrix bound for quartic forms would immediately improve Theorem 1.2.
The bounded-drift construction is the fragile step; if a more robust or simpler way to keep the localization path inside a small ball were found, the method would probably extend to other non-quadratic potentials beyond quartic tensors.
The decomposition is not tied to spectral gaps: it supplies a generic transfer principle, so inequalities of any kind proved on the low-rank pieces lift to the original high-temperature measure.
For higher degrees, the rank of the decomposition grows so fast that the method's guarantees deteriorate; a different stopping rule or a coarser decomposition would be needed to keep spectral-gap bounds competitive in that regime.

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

The paper extends stochastic localization to non-Gaussian tilts in a way that yields the first explicit spectral-gap constant for tensor Ising models. The central argument holds up, and the novel machinery is real. The tensorized SL process, the removal of the barycenter, and the drift that suppresses the dangerous linear term are all genuinely new constructions, and the decomposition of quartic potentials into low-rank pieces is clever. The explicit 1/56 constant is a concrete improvement over the implicit bounds from [3].

It does what it claims. The iterative decomposition with normalized variance for unnormalized measures is carefully handled, and the Dobrushin-based spectral gap argument is clean. The tensor Curie-Weiss lower bound and the spin-glass corollary are reasonable applications, and the comparison with [3] is honest.

The soft spots are real but mostly cosmetic. Theorem 1.1 assumes T is positive definite with zero diagonal entries, which is impossible for a positive definite matrix; this needs a fix, likely replacing it with a shifted or semidefinite version. The application of Theorem 1.5 to a negative definite quadratic form in the proof of Theorem 1.1 is not justified in the text; a standard hypercube shift probably repairs it but should be stated. Proposition 3.3 and Lemma A.1 use opposite sign conventions for v_t; they are consistent under a sign flip, but as written it is confusing and will trip readers. There is also a factor-of-2 typo in the Itô calculation in Lemma A.1: the drift coefficient should be -8n, not -4n. The sign remains confining, so the Bessel-type argument still goes through, but the constants need correction.

On the stress-test claim: it does not hold up on reading. Substituting the printed v_t into the printed SDE in Lemma A.1 (which has a plus sign) gives a confining drift toward the origin, not a repulsive one. The bounded-drift lemma is therefore plausible, and the main decomposition theorem is not undermined.

This paper is for researchers in stochastic localization, Markov chain mixing, and spin glasses. It deserves a serious referee. I would send it out, asking for fixes to Theorem 1.1's statement, the negative-definite application, and the sign/constant conventions in the appendix. None of these look load-bearing, but they must be cleaned up before the paper is citable as a reference.

Referee Report

3 major / 4 minor

Summary. The paper extends stochastic localization to non-Gaussian tilts by introducing a tensorized process driven by Dyson Brownian motion, with the tilt applied to x^\otimes 2 rather than to x. The main structural result is Theorem 1.1, which decomposes a quartic potential on the Boolean hypercube into a mixture of low-rank measures of the form exp(\langle u,x\rangle^2\langle v,x\rangle^2 + \langle x,w\rangle^2 + \langle \ell,x\rangle + \psi(x)) plus a small uniformly bounded perturbation \psi. This is then used to prove Theorem 1.2, an explicit Poincar\'e-constant bound for tensor Ising models with \|T\|_{inj} \le 1/(56n), yielding rapid mixing for Glauber dynamics. The paper also gives a slow-mixing lower bound for the tensor Curie-Weiss model (Proposition 5.1, Lemma 1.3) and sketches extensions to tensors of higher degree.

Significance. If correct, the result provides the first fully explicit spectral-gap bound for quartic spin systems under Glauber dynamics, replacing the implicit constant obtained from spectral independence with the explicit constant 1/56. The proof is constructive and first-principles: the constants are not fitted, and the decomposition theorem is a genuine extension of the rank-one decomposition of [22] to quartic potentials. Corollary 1.4 gives a quantitative high-temperature bound for degree-4 spin glasses. The main technical novelty is the bounded-drift construction that controls the non-Gaussian remainder term, and this is also the most fragile part of the paper. The lower-bound example in Section 5 is a useful sanity check showing that the high-temperature regime cannot be substantially enlarged without additional ideas.

major comments (3)

[Appendix A.2, Lemma A.1 and Eq. (A.2)] The bounded-process construction used in Proposition 3.3 is internally inconsistent. Lemma A.1 states dX_t = C_t dB_t + C_t^2 v_t dt and defines v_t := -4n X_t/(\delta - \|X_t\|^2); substituting gives dX_t = C_t dB_t - 4n C_t^2 X_t/(\delta - \|X_t\|^2) dt. Proposition 3.3 instead defines X_t by dX_t = C_t dW_t - C_t^2 v_t dt, so with the printed v_t the drift would be +4n C_t^2 X_t/(\delta - \|X_t\|^2) dt, which is repulsive rather than confining. The displayed It\^o computation for d\|X_t\|^2 also has a factor-of-two mismatch: the coefficient should be 8n, not 4n, and the later expression for df(Y_t) does not follow from the preceding formula. Since X_t is exactly the remainder R in the proof of Theorem 3.1 and controls the smallness of \psi(x) in Theorem 1.1 and hence the 56n\|T\|_{inj} coefficient in Theorem 1.2, this is a load-bearing gap in the proof as written. The argument appears repairable by choosing v_t with the opposite sign and redoing the Bessel-type estimate, but the corrected derivation is not present. In addition, Proposition 3.3 concerns matrix-valued X_t, whereas Lemma A.1 is stated for vectors in R^n; the state-space dimension should be n^2 in the matrix case.
[Section 3.3, proof of Theorem 1.1 before Eq. (3.18)] Theorem 1.5 is invoked for the measure \bar{\mu}_M(x) \propto \exp(-2n\sqrt{\|T\|_{inj}}\langle x,\tilde{M}x\rangle), i.e. for a quadratic form that is negative semidefinite, whereas Theorem 1.5 requires the quadratic form in the exponent to be positive definite. The application can be repaired by using the fact that \|x\|_2^2 = n is constant on C_n and shifting \tilde{M} by a suitable multiple of the identity, but this step is not stated. As written, the derivation of the decomposition (3.18) is not justified.
[Theorem 1.1 and Section 2.1] The standing assumption of Theorem 1.1 is inconsistent. The theorem requires T to be a positive definite symmetric fourth-order tensor with zero diagonal entries. Positive definiteness is defined in Section 2.2 in the n^2 \times n^2 matrix sense, while Section 2.1 defines a diagonal entry as a multi-index with any repetition, so T_{ijij} = 0 for all i,j. Under this definition all diagonal entries of the n^2 \times n^2 matrix vanish, which is impossible for a positive definite matrix. Thus the theorem as stated is vacuous. The statement should either redefine "zero diagonal" (for example, only T_{iiii}=0) or be formulated for positive semidefinite T with an explicit recentering; the proofs of Theorem 1.2 and Corollary 1.4 already rely on such recentering steps.

minor comments (4)

[Theorem 2.3] The displayed convergence "P(|n^{p/2-1}\|T\|_{inj} - E_0(p)| \le \varepsilon) \to 0" should be convergence to 1; as printed it contradicts the concentration statement and its use in Corollary 1.4.
[Proof of Corollary 1.4] The line "\mu(x) \propto (\beta \tilde{T}(x))" is missing the exponential; it should be \mu(x) \propto \exp(\beta \tilde{T}(x)).
[Proof of Theorem 1.2, Section 4.2] The bound \|D_{\psi}\|_{op} \le n\delta is off by a constant factor: since each entry of D_{\psi} is at most 4\delta, a more natural bound is 4n\delta. Because \delta is arbitrary, this does not change the final conclusion.
[Proof of Proposition 3.3] The iteration over the decreasing sequence of image subspaces is only sketched in the final paragraph of Appendix A.2; if the base Bessel-type estimate is corrected, the stopping-time induction and the behavior of the process at degeneracy times should be spelled out in detail.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the spectral-gap theorem is derived from a self-contained localization decomposition, with all constants obtained by explicit estimates; the flagged appendix drift issue is a correctness risk rather than circularity.

full rationale

The derivation chain is not circular. Theorem 1.2 is obtained from Theorem 1.1's decomposition together with the derivative-matrix bounds (Lemmas 4.1-4.3) and Dobrushin's condition; none of these inputs contains the Poincare constant being proved. Theorem 1.1 is built from Theorem 3.1 and Theorem 3.7, whose proofs use Ito calculus on the tensorized localization process (Lemmas 3.4-3.6) and the external rank-one decomposition theorem [22, Thm 12]. The bounded-drift input (Proposition 3.3, Lemma A.1) is a separate stochastic-control construction: the drift v_t is chosen as a feedback function of X_t so that X_t stays in a ball, and the bound sup_t ||X_t|| <= delta is proved by a Bessel-type argument, not assumed. The paper cites [22] and [5,11] for the smoothed projection and injective-norm concentration, but these are independent external results; no load-bearing premise is justified only by a self-citation, and the authors' own prior works [13] and [23] appear only in related-work lists. The appendix contains a possible sign inconsistency between Proposition 3.3's SDE and Lemma A.1's SDE, and the constants in the Ito expansion are not fully clean; these are correctness risks that would affect the validity of Proposition 3.3, not cases where a target quantity is defined or fitted in terms of itself. Since no equation is equivalent to its input by construction and no fitted parameter is renamed as a prediction, the circularity score is 0.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard stochastic calculus, the Dobrushin condition, and cited concentration results. The only ad hoc assumption is the positive definiteness (or shift thereof) of the tensor, and the implicit diagonal shift when applying Theorem 1.5 to indefinite quadratic forms. No parameters are fitted to data.

assumptions (6)

standard math Stochastic calculus (Ito's formula, martingale properties) is valid.
Used throughout Section 3, e.g., Lemmas 3.4 and 3.5.
domain assumption Dobrushin condition implies CP(µ) ≤ 1/(1-‖Aµ‖_op).
Used in Section 4.1 to bound spectral gaps of decomposed measures.
domain assumption Concentration of the injective norm of Gaussian tensors (Theorem 2.3), cited from [5,11].
Used in Corollary 1.4 to estimate ‖T‖_inj for Gaussian spin glass tensors.
standard math The decomposition theorem of [22, Theorem 12] (Theorem 1.5 here) is valid; the paper extends it without reproving it.
Used in the final step of Theorem 1.1 to decompose the remaining quadratic part.
domain assumption On the Boolean hypercube, x_i^2=1 so diagonal shifts of quadratic forms only change the normalization, allowing indefinite quadratic forms to be shifted to positive definite ones.
Implicit in the application of Theorem 1.5 to the negative definite quadratic form in the proof of Theorem 1.1; not explicitly stated.
ad hoc to paper Positive definiteness of T as an n^2×n^2 matrix, or a shift making it so, is assumed for the decomposition.
Theorem 1.2 assumes T positive definite; the spin glass tensor is shifted by ‖T‖_inj I to ensure this.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Stochastic Localization with Non-Gaussian Tilts and Applications to Tensor Ising Models." pith.science (2026). https://pith.science/paper/LF4LSY7O

@misc{pith2026241212720,
  author       = {Pith},
  title        = {Pith review of: Stochastic Localization with Non-Gaussian Tilts and Applications to Tensor Ising Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LF4LSY7O}},
  note         = {Machine review of arXiv:2412.12720}
}

read the original abstract

We present generalizations and modifications of Eldan's Stochastic Localization process, extending it to incorporate non-Gaussian tilts, making it useful for a broader class of measures. As an application, we introduce new processes that enable the decomposition and analysis of non-quadratic potentials on the Boolean hypercube, with a specific focus on quartic polynomials. Using this framework, we derive new spectral gap estimates for tensor Ising models under Glauber dynamics, resulting in rapid mixing.

Figures

Figures reproduced from arXiv: 2412.12720 by the authors.

**Figure 1.** Decomposition of a fourth-order tensor. where T is a positive definite symmetric fourth-order tensor with zero diagonal entries. Then, for every δ > 0, there exists a decomposition of µ as µ = Z µu,v,w,ℓ,ψη(du, dv, dw, dℓ, dψ), where with u, v, w, ℓ ∈ R n and ψ : Cn → R, and µu,v,w,ℓ,ψ is a non-negative measure of the form µu,v,w,ℓ,ψ(x) ∝ exp hu, xi 2 hv, xi 2 + hx, wi 2 + hℓ, xi + ψ(x) . Moreover, we have the fol… view at source ↗

**Figure 2.** illustrates the procedure we just described. T8 T (1) 4 ⊗ T (2) 4 + T (3) 4 [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗

**Figure 3.** Decomposition of a tensor of degree 16, T16. First, it is decomposed into T (1) 8 ⊗ T (2) 8 + T (3) 8 . Each T (i) 8 is then further decomposed as described in the previous section. The final decomposition consists of rank-1 tensors T (i,j,k) 1 with i, j, k = 1, 2, 3. recursion n2 p−1 ≤ 2 √ n2 p , and we need to understand the value n1 which satisfies ku (j)k 2 ≤ n1. Since n2 p := kT2 p kinj, a calculation produces … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 31 canonical work pages

[3]

Universality of Spectral Independence with Applications to Fast Mixing in Spin Glasses

Nima Anari, Vishesh Jain, Frederic Koehler, Huy Tuan Pha m, and Thuy-Duong Vuong. “Universality of Spectral Independence with Applications to Fast Mixing in Spin Glasses”. In: Proceedings of the 2024 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA) (2024), pp. 5029–5056

work page 2024
[22]

A sp ectral condition for spectral gap: fast mixing in high-temperature Ising models

Ronen Eldan, Frederic Koehler, and Ofer Zeitouni. “A sp ectral condition for spectral gap: fast mixing in high-temperature Ising models”. In: Probability theory and related ﬁelds 182.3-4 (2022), pp. 1035–1051

work page 2022
[1]

Spectral gap estimates for mixed p-spin models at high temperature

Arka Adhikari, Christian Brennecke, Changji Xu, and Hor ng-Tzer Yau. “Spectral gap estimates for mixed p-spin models at high temperature”. In: Probab. Theory Related Fields 189.3-4 (2024), pp. 879–907

work page 2024
[2]

Entropic independence: optimal mixing of down-up random w alks

Nima Anari, Vishesh Jain, Frederic Koehler, Huy Tuan Pha m, and Thuy-Duong Vuong. “Entropic independence: optimal mixing of down-up random w alks”. In: STOC ’22— Proceedings of the 54th Annual ACM SIGACT Symposium on Theory of Co mputing. ACM, New York, 2022, pp. 1418–1430. isbn: 978-1-4503-9264-8

work page 2022
[4]

Tr ickle-Down in Localization Schemes and Applications

Nima Anari, Frederic Koehler, and Thuy-Duong Vuong. “Tr ickle-Down in Localization Schemes and Applications”. In: Proceedings of the 56th Annual ACM Symposium on The- ory of Computing . New York, NY, USA: ACM, 2024, pp. 1094–1105

work page 2024
[5]

Random Matrices and Complex- ity of Spin Glasses

Antonio Auﬃnger, G´ erard Ben Arous, and Jiˇ r ´ ıˇCern´ y. “Random Matrices and Complex- ity of Spin Glasses”. In: Communications on pure and applied mathematics 66.2 (2013), pp. 165–201. 30

work page 2013
[6]

Polynomials and multi linear mappings in topological vector-spaces

Jacek Bochnak and J´ ozef Siciak. “Polynomials and multi linear mappings in topological vector-spaces”. eng. In: Studia Mathematica 39.1 (1971), pp. 59–76

work page 1971
[7]

Sudakov–Fernique post-AMP, and a n ew proof of the local convexity of the TAP free energy

Michael Celentano. “Sudakov–Fernique post-AMP, and a n ew proof of the local convexity of the TAP free energy”. In: The Annals of probability 52.3 (2024)

work page 2024

Show all 38 references

[8]

An Almost Constant Lower Bound of the Isope rimetric Coeﬃcient in the KLS Conjecture

Yuansi Chen. “An Almost Constant Lower Bound of the Isope rimetric Coeﬃcient in the KLS Conjecture”. In: Geometric and functional analysis 31.1 (2021), pp. 34–61

2021
[9]

Hit-and-run mixing via localization schemes

Yuansi Chen and Ronen Eldan. Hit-and-run mixing via localization schemes . 2022. arXiv: 2212.00297

2022 arXiv
[10]

Localization Schemes: A F ramework for Proving Mixing Bounds for Markov Chains

Yuansi Chen and Ronen Eldan. “Localization Schemes: A F ramework for Proving Mixing Bounds for Markov Chains”. In: 2022 IEEE 63rd Annual Symposium on Foundations of Computer Science (FOCS) . IEEE, 2022, pp. 110–122

2022
[11]

Injective norm of real and complex random tensors I: From spin glasses to geometric entanglement

Stephane Dartois and Benjamin McKenna. Injective norm of real and complex random tensors I: From spin glasses to geometric entanglement . 2024. arXiv: 2404.03627

2024 arXiv
[12]

The Description of a Random Field by M eans of Conditional Probabil- ities and Conditions of Its Regularity

P. L. Dobruschin. “The Description of a Random Field by M eans of Conditional Probabil- ities and Conditions of Its Regularity”. In: Theory of probability and its applications 13.2 (1968), pp. 197–224

1968
[13]

Fast relaxation of the random ﬁeld Ising dynamics

Ahmed El Alaoui, Ronen Eldan, Reza Gheissari, and Arian na Piana. “Fast relaxation of the random ﬁeld Ising dynamics”. In: Annals of Probability (2024), to appear

2024
[14]

An Information- Theoretic View of Stochastic Localization

Ahmed El Alaoui and Andrea Montanari. “An Information- Theoretic View of Stochastic Localization”. In: IEEE Transactions on Information Theory 68.11 (2022), pp. 7423–7426

2022
[15]

Sampling from Mean-Field Gibbs Measures via Diﬀusion Processes

Ahmed El Alaoui, Andrea Montanari, and Mark Sellke. Sampling from Mean-Field Gibbs Measures via Diﬀusion Processes . 2023. arXiv: 2310.08912

2023
[16]

Sa mpling from the Sherrington- Kirkpatrick Gibbs measure via algorithmic stochastic loca lization

Ahmed El Alaoui, Andrea Montanari, and Mark Sellke. “Sa mpling from the Sherrington- Kirkpatrick Gibbs measure via algorithmic stochastic loca lization”. In: 2022 IEEE 63rd Annual Symposium on Foundations of Computer Science—FOCS 202 2. IEEE Computer Soc., Los Alamitos, CA, 20...

2022
[17]

Analysis of high-dimensional distributions using pathwis e methods

Ronen Eldan. Analysis of high-dimensional distributions using pathwis e methods . Ed. by Dmitry Beliaev and Stanislav Smirnov. Vol. 6. EMS Press, 202 2, pp. 4246–4271
[18]

Gaussian-width gradient complexity, re verse log-Sobolev inequalities and nonlinear large deviations

Ronen Eldan. “Gaussian-width gradient complexity, re verse log-Sobolev inequalities and nonlinear large deviations”. In: Geometric and functional analysis 28.6 (2018), pp. 1548– 1596

2018
[19]

Taming correlations through entropy-eﬃ cient measure decompositions with applications to mean-ﬁeld approximation

Ronen Eldan. “Taming correlations through entropy-eﬃ cient measure decompositions with applications to mean-ﬁeld approximation”. In: Probability theory and related ﬁelds 176.3-4 (2020), pp. 737–755

2020
[20]

Thin Shell implies spectral gap up to Poly log via a Stochastic Localization scheme

Ronen Eldan. “Thin Shell implies spectral gap up to Poly log via a Stochastic Localization scheme”. In: Geometric and functional analysis 23.2 (2013), pp. 532–569

2013
[21]

Concentration on the Bool ean hypercube via pathwise stochastic analysis

Ronen Eldan and Renan Gross. “Concentration on the Bool ean hypercube via pathwise stochastic analysis”. In: Inventiones mathematicae 230.3 (2022), pp. 935–994

2022
[23]

Stability of the Shann on-Stam inequality via the F¨ ollmer process

Ronen Eldan and Dan Mikulincer. “Stability of the Shann on-Stam inequality via the F¨ ollmer process”. In:Probab. Theory Related Fields 177.3-4 (2020), pp. 891–922

2020
[24]

Log concavity and concent ration of Lipschitz functions on the Boolean hypercube

Ronen Eldan and Omer Shamir. “Log concavity and concent ration of Lipschitz functions on the Boolean hypercube”. In: Journal of functional analysis 282.8 (2022), pp. 109392–. 31

2022
[25]

An introduction to probability theory and its applications

William Feller. An introduction to probability theory and its applications . Vol. I . Third. John Wiley & Sons, Inc., New York-London-Sydney, 1968

1968
[26]

Sampling with ﬂows, diﬀusion, and autoregressive neural networks from a spin-gl ass perspective

Davide Ghio, Yatin Dandi, Florent Krzakala, and Lenka Z deborov´ a. “Sampling with ﬂows, diﬀusion, and autoregressive neural networks from a spin-gl ass perspective”. In: Proceed- ings of the National Academy of Sciences - PNAS 121.27 (2024), e2311810121–

2024
[27]

Weak Poincar´ e Inequalities, Simulated Annealing, and Sampling from Sphe rical Spin Glasses

Brice Huang, Sidhanth Mohanty, Amit Rajaraman, and Dav id X Wu. “Weak Poincar´ e Inequalities, Simulated Annealing, and Sampling from Sphe rical Spin Glasses”. In: (2024). arXiv: 2411.09075

2024 arXiv
[28]

Sampling from Spherical Spin Glasses in Total Variation via Algorithmic Stochastic Localization

Brice Huang, Andrea Montanari, and Huy Tuan Pham. Sampling from Spherical Spin Glasses in Total Variation via Algorithmic Stochastic Localization. 2024. arXiv: 2404.15651

2024 arXiv
[29]

Isoperimetr ic problems for convex bodies and a localization lemma

R. Kannan, L. Lov´ asz, and M. Simonovits. “Isoperimetr ic problems for convex bodies and a localization lemma”. In: Discrete & computational geometry 13.3-4 (1995), pp. 541–559

1995
[30]

Bourgain’s slicing pr oblem and KLS isoperimetry up to polylog

Bo’az Klartag and Joseph Lehec. “Bourgain’s slicing pr oblem and KLS isoperimetry up to polylog”. In: Geometric and functional analysis 32.5 (2022), pp. 1134–1159

2022
[31]

Spectral monotonici ty under Gaussian convolution

Bo’az Klartag and Eli Putterman. “Spectral monotonici ty under Gaussian convolution”. In: Annales de la Facult´ e des sciences de Toulouse: Math´ ematiques 32.5 (2024), pp. 939– 967

2024
[32]

Eldan’s stochastic localization and the KLS con- jecture: Isoperimetry, concentration and mixing

Yin Tat Lee and Santosh S. Vempala. “Eldan’s stochastic localization and the KLS con- jecture: Isoperimetry, concentration and mixing”. In: Annals of mathematics 199.3 (2024)

2024
[33]

Markov chains and mixing times

David Levin, Yuval Peres, and Elizabeth L Wilmer. Markov chains and mixing times . 1st ed. Vol. 58. American Mathematical Society, 2009

2009
[34]

Kuikui Liu, Sidhanth Mohanty, Amit Rajaraman, and Davi d X. Wu. Fast Mixing in Sparse Random Ising Models . 2024. arXiv: 2405.06616

2024 arXiv
[35]

Sampling, Diﬀusions, and Stochastic Localization

Andrea Montanari. Sampling, Diﬀusions, and Stochastic Localization . 2023. arXiv: 2305.10690

2023 arXiv
[36]

Posterior Sampling in High Dimension via Diﬀusion Processes

Andrea Montanari and Yuchen Wu. Posterior Sampling in High Dimension via Diﬀusion Processes. 2023. arXiv: 2304.11449

2023 arXiv
[37]

Mathematical Aspects of Mixing Times in Markov Chains

Ravi Montenegro and Prasad Tetali. Mathematical Aspects of Mixing Times in Markov Chains. 1st ed. Vol. 1. Foundations and Trends in Theoretical Compu ter Science 3. 2006, pp. 237–354

2006
[38]

Mixing phases of the Glauber dynamics for the p-spin Curie-Weiss model

Ramkrishna Jyoti Samanta, Somabha Mukherjee, and Jian g Zhang. Mixing phases of the Glauber dynamics for the p-spin Curie-Weiss model . 2024. arXiv: 2412.16952. A Components of TSL A.1 Smoothed projections the construction of the smoothed projection from Lemma 3.2 is based on ...

2024 arXiv

Pith tools

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

[1] [3]

Universality of Spectral Independence with Applications to Fast Mixing in Spin Glasses

Nima Anari, Vishesh Jain, Frederic Koehler, Huy Tuan Pha m, and Thuy-Duong Vuong. “Universality of Spectral Independence with Applications to Fast Mixing in Spin Glasses”. In: Proceedings of the 2024 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA) (2024), pp. 5029–5056

work page 2024

[2] [22]

A sp ectral condition for spectral gap: fast mixing in high-temperature Ising models

Ronen Eldan, Frederic Koehler, and Ofer Zeitouni. “A sp ectral condition for spectral gap: fast mixing in high-temperature Ising models”. In: Probability theory and related ﬁelds 182.3-4 (2022), pp. 1035–1051

work page 2022

[3] [1]

Spectral gap estimates for mixed p-spin models at high temperature

Arka Adhikari, Christian Brennecke, Changji Xu, and Hor ng-Tzer Yau. “Spectral gap estimates for mixed p-spin models at high temperature”. In: Probab. Theory Related Fields 189.3-4 (2024), pp. 879–907

work page 2024

[4] [2]

Entropic independence: optimal mixing of down-up random w alks

Nima Anari, Vishesh Jain, Frederic Koehler, Huy Tuan Pha m, and Thuy-Duong Vuong. “Entropic independence: optimal mixing of down-up random w alks”. In: STOC ’22— Proceedings of the 54th Annual ACM SIGACT Symposium on Theory of Co mputing. ACM, New York, 2022, pp. 1418–1430. isbn: 978-1-4503-9264-8

work page 2022

[5] [4]

Tr ickle-Down in Localization Schemes and Applications

Nima Anari, Frederic Koehler, and Thuy-Duong Vuong. “Tr ickle-Down in Localization Schemes and Applications”. In: Proceedings of the 56th Annual ACM Symposium on The- ory of Computing . New York, NY, USA: ACM, 2024, pp. 1094–1105

work page 2024

[6] [5]

Random Matrices and Complex- ity of Spin Glasses

Antonio Auﬃnger, G´ erard Ben Arous, and Jiˇ r ´ ıˇCern´ y. “Random Matrices and Complex- ity of Spin Glasses”. In: Communications on pure and applied mathematics 66.2 (2013), pp. 165–201. 30

work page 2013

[7] [6]

Polynomials and multi linear mappings in topological vector-spaces

Jacek Bochnak and J´ ozef Siciak. “Polynomials and multi linear mappings in topological vector-spaces”. eng. In: Studia Mathematica 39.1 (1971), pp. 59–76

work page 1971

[8] [7]

Sudakov–Fernique post-AMP, and a n ew proof of the local convexity of the TAP free energy

Michael Celentano. “Sudakov–Fernique post-AMP, and a n ew proof of the local convexity of the TAP free energy”. In: The Annals of probability 52.3 (2024)

work page 2024