The SK model with a sparse variance profile: free energy and AMP-based magnetization estimation at high temperature

arXiv:2604.25535 · math-ph, math.MP, math.PR · Submitted 2026-04-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Mathematical Physics. The day in mathematical physics: rigorous models of physical systems.

Kai: We're looking at the paper: "The SK model with a sparse variance profile".

Mira: Here's the gist of the paper. A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse<ref:2604.25535#pg2> In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix<ref:2604.25535#pg2> In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations<ref:2604.25535#pg2> Model Definition and Assumptions The model introduces a deterministic symmetric matrix Spnq with zero diagonal and non-negative elements, which defines the variance profile of the interactions<ref:2604.25535#pg2> The random matrix Wpnq is defined such that its entries are independent, and it satisfies Wpnqij = Np0, tspnqij q for all i, j P rns and for some t ą 0<ref:2604.25535#pg2> Assumption 1 imposes conditions on this matrix Spnq, such as the bound spnqij ď CsK´1n for all n and all i, j P rns, and a finiteness condition on the number Crow “ supn Spnq where ¨ is the max row sum norm<ref:2604.25535#pg2> A case of interest is when Kn!

Kai: n, and there are at most CcardKn non-zero spnqij ’s in each row and column of Spnq for some constant Ccard ą 0, referred to as the “sparse” ones<ref:2604.25535#pg2> Assumption 2 specifies the high temperature regime by requiring t ă log 2 Crow<ref:2604.25535#pg2> Free Energy Asymptotics The first aim of this paper is to study the large n asymptotics of the free energy Fn “ 1/n E log Zpnq in a “high temperature” regime represented by Assumption 2<ref:2604.25535#pg2> Theorem 2 states that under Assumptions 1 and 2, the function F n is bounded, and moreover, that Fn ´ F n ÝÝÝÑ nÑ8 0<ref:2604.25535#pg2> The signature of Spnq has no impact on the form of the large-n approximation of the free energy<ref:2604.25535#pg2> Corollary 3 shows that if Spnq is a doubly stochastic matrix, then under Assumptions 1 and 2, it holds that Fn ÝÝÝÑ nÑ8 log 2 E log cosh p qξ h q t/4 p1 ´ q(tq 2)<ref:2604.25535#pg2> Magnetization Estimation via AMP The second aim is to provide a large–n approximation of the mean vector mpnq “ xσy<ref:2604.25535#pg2> Returning to the classical SK model, this vector can be approximated in the high temperature regime by an Approximate Message Passing (AMP) algorithm based on the Thouless-Anderson-Palmer (TAP) equations<ref:2604.25535#pg2> Theorem 5 proves that under Assumptions 3 and 2, for each even integer b ě 2, it holds that sup n max iPrns Eˇˇz l 1 pA pnq l q iˇˇb ă 8<ref:2604.25535#pg2> This implies that lim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε, which implies that lim supklim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε<ref:2604.25535#pg2> Convergence of AMP Iterates The paper develops a sequence of families of vectors ty1pA1q uA1Ărns,A1“k´-l, ty2pA2q uA2Ărns,A2“k´-l, and so on<ref:2604.25535#pg2> The construction of the iterates involves replacing the function Tanh with a polynomial activation function f<ref:2604.25535#pg2> Lemma 14 shows that for each ε ą 0, there is a polynomial pe such that the iterates (15) with f “ pe satisfy lim sup n max iPrns Epxˇk i ´ x˜k i q squared ď ε<ref:2604.25535#pg2> The proof for Theorem 5 involves showing that the conditional variance of the iterates is bounded by a constant, leading to the convergence of the AMP algorithm<ref:2604.25535#pg2> Bounding Errors and Final Result The final result shows that lim supklim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε<ref:2604.25535#pg2> This implies that lim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε, which implies that lim supklim supn E‹Tanhpxk q›› 2n1E ď 2ε<ref:2604.25535#pg2> Using Lemmas 12 and 11, it is shown that lim supn E››m ´ Tanhpxk q›› 2n1E ď 4plog 2q k<ref:2604.25535#pg2> Denoting as M8 the max norm of the matrix M, it holds that lim supklim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε<ref:2604.25535#pg2> The paper establishes that for any arbitrary moment b ą 0, the supremum over all indices and sets is bounded by a constant independent of n<ref:2604.25535#pg2> This implies that lim supklim supn Eˇˇz l iˇˇb ă 8<ref:2604.25535#pg2> The final result is that lim supklim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε, which implies that lim supn E‹Tanhpxk q›› 2n1E ď 2ε<ref:2604.25535#pg2> The paper closes with remarks on the necessity of the Kn ≥ log n condition and future research directions<ref:2604.25535#pg2> References

1: A.

Mira: Adhikari, C.

Kai: Brennecke, P.

Title and authors: Kai: Let's talk about what the paper actually says in its title and who wrote it. It's "The SK model with a sparse variance profile: free energy and AMP-based magnetization estimation at high temperature." It sounds pretty technical, but it points to two main things: handling sparse interaction matrices in spin glasses, and using Approximate Message Passing for the mean magnetization under high-temperature conditions.

Mira: The authors are working on extending the dynamical approach originally used for the classical SK model to this new setting where we have a variance profile that can be arbitrary or sparse. It's about adapting old tools to a more general problem structure.

Lev: I'm interested in how they handle the assumptions because that's where the physics gets constrained. They introduce these conditions, Assumption one which gives bounds on the matrix elements and also puts a condition on how many non-zero interactions there are in each row and column—they call these "sparse" ones <ref:2604.25535#pg1>.

Kai: So, if we think about what this means for us on hardware, it suggests that if we can define our system such that the number of relevant connections is small per spin, we might be able to use simpler methods for estimation.

Mira: That's right. And they also have Assumption two which sets a high-temperature regime where the temperature is greater than log two times the number of non-zero interactions in each row and column, which helps simplify things significantly <ref:2604.25535#pg2,in each row and column>.

Lev: From an error correction viewpoint, if you can bound the complexity of the interaction structure, you can control the overhead of your approximation algorithm much better.

The paper's summary: Kai: Now let's look at what they actually achieved in their summary. They show that in this high-temperature regime defined by Assumption two there exists an asymptotic equivalent for the free energy, F n, that is bounded and goes to zero as n goes to infinity <ref:2604.25535#pg2>.

Mira: That's a big result because it means the signature of that variance profile matrix doesn't actually matter for the overall form of this large-n approximation of the free energy, which was a key goal they set out.

Lev: So, even if we don't know much about the exact structure of these interactions beyond those general bounds, we still get a stable prediction for the free energy scaling as n gets huge. That's useful because it gives us a baseline behavior to check against.

Kai: And they also show that they can estimate the mean spin vector under this Gibbs measure using an Approximate Message Passing algorithm based on those TAP equations, which is returning to methods from the classical SK model but adapted for these sparse settings.

Mira: The summary highlights that this AMP approach, when applied under Assumption three and Assumption two gives a bound on the expectation of certain terms related to the spins, which ultimately leads to a convergence result for their iterative process <ref:2604.25535#pg3>.

Lev: That's where the complexity comes back in—how fast that AMP algorithm converges on real hardware. They are showing that even with these sparse structures, we can get an approximation for the mean spin vector in this high-temperature setting.

The paper's improvements: Kai: The paper points out a couple of ways they improved the situation. They show that if the variance profile matrix is doubly stochastic—meaning all its row and column sums are exactly one—then they recover the asymptotic free energy expression for the classical SK model at high temperature, as described in Corollary three <ref:2604.25535#pg3,the classical SK model at high temperature>.

Mira: That's a neat result because it shows that when we have that specific structure, the general framework simplifies nicely and connects back to something we already know about the standard SK model behavior.

Lev: For me, this connection is important because it means we have a known benchmark to test against when dealing with structured matrices like doubly stochastic ones. It gives us a concrete target for what our sparse model should behave like under those conditions.

Kai: And they also discuss the construction of the iterative sequence of vectors, t y one p A one q u A one in k'-l, which they use to approximate the mean vector <ref:2604.25535#pg1>. They replace the function Tanh with a polynomial activation function, and Lemma fourteen shows that this polynomial activation ensures convergence for any epsilon greater than zero.

Mira: So the improvement here is moving from just saying an approximation exists to showing *how* that approximation is constructed—using these specific non-backtracking iterations and carefully chosen polynomial functions.

Lev: That iteration sequence, using those polynomials, that's where you start thinking about computational complexity. If the polynomial degree is too high, the number of steps needed for convergence could become prohibitive on a real system.

Conclusion: Kai: So to wrap up this paper on "The SK model with a sparse variance profile: free energy and AMP-based magnetization estimation at high temperature," they've established that for any arbitrary moment b less than zero, the supremum over all indices and sets is bounded by a constant independent of n.

Mira: That means they've confirmed that the final error bound on those moments stays within a fixed constant as we increase the system size, which is a strong statement about their stability in this regime.

Lev: For us in hardware, that means if you run this approximation on a large system, you don't have to worry about some crazy error term growing unboundedly with the size of your setup. It stays controlled by that constant.

Kai: The final result they push is that the supremum over all indices and sets is bounded by two epsilon, which implies that for large n, the estimation of the mean spin vector under this Gibbs measure is bounded by two epsilon times something related to n to the power of one.

Mira: That's what allows them to say that they can get a good approximation for the mean vector, x, even when dealing with sparse variance profiles, provided we are in that high-temperature regime.

Lev: I'm just going to add that their final bound on the supremum is four times log two times q to the power of k, which suggests a slightly more complex dependence than just a simple constant, but still controlled by something related to the sparsity parameter.

Kai: So we've seen how they use AMP, how they handle sparse structures, and how they get bounds on the free energy and magnetization in this model. It’s a solid piece of work showing that even when things are not perfectly structured, you can still find reliable approximations under certain conditions.

CNRS · Laboratoire d’informatique Gaspard Monge (LIGM / UMR 8049) · Universit´e Gustave Eiffel · ESIEE Paris · France

math-ph, math.MP, math.PR

Submitted: 2026-04-28

Updated: 2026-11-13

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 76/100

The gist: A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and…

Key concepts

Variance Profile Matrix (Spnq)
This is a deterministic symmetric matrix that defines how the interaction variance changes across different pairs of spins in the system. It has zero diagonal elements and non-negative entries, and its sparsity (few non-zero entries per row/column) is a key feature studied in this model.
Approximate Message Passing (AMP)
AMP is an algorithm used to estimate the mean magnetization vector of the spin system. It works by iteratively updating local estimates based on messages passed between neighboring spins, mimicking how information propagates through the complex interactions defined by the variance profile matrix.
Thouless–Anderson–Palmer (TAP) Equations
These are equations that describe the stationary states or equilibrium configurations of spin glasses. The AMP algorithm in this paper uses these equations as a basis for its iterative updates to approximate the true mean magnetization vector under high-temperature conditions.

Terminology

Summary

A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse<ref:2604. p. 0

Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

25535#pg2> In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix<ref:2604. p. 0
Where the paper says it

fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse. In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix.

25535#pg2> In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations<ref:2604. p. 0
Where the paper says it

In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix. In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations.

25535#pg2>

Model Definition and Assumptions

The model introduces a deterministic symmetric matrix Spnq with zero diagonal and non-negative elements, which defines the variance profile of the interactions<ref:2604.25535#pg2> The random matrix Wpnq is defined such that its entries are independent, and it satisfies Wpnqij = Np0, tspnqij q for all i, j P rns and for some t ą 0<ref:2604. p. 0

Where the paper says it

1 Model, problem and the results For each integer n ą 0, let S pnq “ ” s pnq ij ın i,j“1 be a deterministic symmetric matrix with elements s pnq ij ě 0 and with a zero diagonal. Let Wpnq “ ” W pnq ij ın i,j“1 be a real symmetric n ˆ n random matrix such that the random variables tW pnq ij u1ďiăjďn are independent, and such that W pnq ij „ N p0, tspnq ij q for all i, j P rns and for some t ą 0.

25535#pg2> Assumption 1 imposes conditions on this matrix Spnq, such as the bound spnqij ď CsK´1n for all n and all i, j P rns, and a finiteness condition on the number Crow “ supn Spnq where ¨ is the max row sum norm<ref:2604. p. 1
Where the paper says it

• There exists a constant Cs ą 0 such that s pnq ij ď CsK´1 n for all n and all i, j P rns. • The number Crow “ supn S pnq where |||¨||| is the max row sum norm is finite.

p. 1
Where the paper says it

• There exists a constant Cs ą 0 such that s pnq ij ď CsK´1 n for all n and all i, j P rns. • The number Crow “ supn S pnq where |||¨||| is the max row sum norm is finite.

25535#pg2> A case of interest is when Kn! p. 1
Where the paper says it

• The number Crow “ supn S pnq where |||¨||| is the max row sum norm is finite. One case of interest covered by Assumption 1 and that can be useful in the fields of statistical physics, graph inference, and large-dimensional signal estimation among others is the case where Kn !

n, and there are at most CcardKn non-zero spnqij ’s in each row and column of Spnq for some constant Ccard ą 0, referred to as the “sparse” ones<ref:2604. p. 1
Where the paper says it

n, and where there are at most CcardKn non-zero s pnq ij ’s in each row and column of S pnq for some constant Ccard ą 0. We refer to these models as the “sparse” ones.

25535#pg2> Assumption 2 specifies the high temperature regime by requiring t ă log 2 Crow<ref:2604. p. 1
Where the paper says it

The first aim of this paper is to study the large n asymptotics of the free energy Fn “ 1 n E log Z pnq in a “high temperature” regime represented by the following assumption: Assumption 2. t ă log 2 Crow .

p. 1
Where the paper says it

The first aim of this paper is to study the large n asymptotics of the free energy Fn “ 1 n E log Z pnq in a “high temperature” regime represented by the following assumption: Assumption 2. t ă log 2 Crow .

25535#pg2>

Free Energy Asymptotics

The first aim of this paper is to study the large n asymptotics of the free energy Fn “ 1/n E log Zpnq in a “high temperature” regime represented by Assumption 2<ref:2604.25535#pg2> Theorem 2 states that under Assumptions 1 and 2, the function F n is bounded, and moreover, that Fn ´ F n ÝÝÝÑ nÑ8 0<ref:2604. p. 2

Where the paper says it

Let Assumptions 1 and 2 hold true. Defining the function F n “ log 2 ` 1 n ÿn i“1 E log cosh ˆb q pnq i ξ ` h ˙ ` t 4n ´ 1n ´ gpq pnq q ¯J S pnq ´ 1n ´ gpq pnq q ¯ , where q pnq “ ” q pnq i ın i“1 is the solution of (2), it holds that F n is bounded, and moreover, that Fn ´ F n ÝÝÝÑ nÑ8 0.

25535#pg2> The signature of Spnq has no impact on the form of the large-n approximation of the free energy<ref:2604. p. 2
Where the paper says it

Defining the function F n “ log 2 ` 1 n ÿn i“1 E log cosh ˆb q pnq i ξ ` h ˙ ` t 4n ´ 1n ´ gpq pnq q ¯J S pnq ´ 1n ´ gpq pnq q ¯ , where q pnq “ ” q pnq i ın i“1 is the solution of (2), it holds that F n is bounded, and moreover, that Fn ´ F n ÝÝÝÑ nÑ8 0. It is seen here that the signature of S pnq has no impact on the form of the large-n approximation of the free energy.

25535#pg2> Corollary 3 shows that if Spnq is a doubly stochastic matrix, then under Assumptions 1 and 2, it holds that Fn ÝÝÝÑ nÑ8 log 2 E log cosh p qξ h q t/4 p1 ´ q(tq 2)<ref:2604. p. 2
Where the paper says it

Assume that S pnq is a doubly stochastic matrix. Then, it holds under Assumptions 1 and 2 (which reads t ă log 2) that Fn ÝÝÝÑ nÑ8 log 2 ` E log cosh p ?

25535#pg2>

Magnetization Estimation via AMP

The second aim is to provide a large–n approximation of the mean vector mpnq “ xσy<ref:2604.25535#pg2> Returning to the classical SK model, this vector can be approximated in the high temperature regime by an Approximate Message Passing (AMP) algorithm based on the Thouless-Anderson-Palmer (TAP) equations<ref:2604. p. 1

Where the paper says it

Denoting as usual as x¨y the mean operator with respect to the measure pGpnq q b8, the second aim of this paper is to provide a large–n approximation of the mean vector mpnq “ xσy. Returning to the classical SK model, it is well-known that this vector can be approximated in the high temperature regime by an Approximate Message Passing (AMP) algorithm by building on the so-called Thouless-Anderson-Palmer (TAP) equations [9].

25535#pg2> Theorem 5 proves that under Assumptions 3 and 2, for each even integer b ě 2, it holds that sup n max iPrns Eˇˇz l 1 pA pnq l q iˇˇb ă 8<ref:2604. p. 8
Where the paper says it

The first bound obtained for u “ 1 in the statement of Lemma 7 can be rewritten as max iPrns Epmi ´ TanhprWmpiq siqq2 ď C{Kn. By a straightforward adaptation of this lemma, we obtain that for a fixed integer M ą 0, it holds that max AĂrns,|A|“M max iPAc E ` mpAq,i ´ Tanh `“WmpAYtiuq‰ i ˘˘2 ď C Kn , (12) This bound will be at the basis of our proof.

25535#pg2> This implies that lim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε, which implies that lim supklim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε<ref:2604. p. 2
Where the paper says it

j P rns : s pnq ij ą 0 )ˇ ˇ ˇ ď CcardKn, where | ¨ | is the cardinality of a set. Of course, Assumption 3 implies Assumption 1 with Crow “ supn S pnq satisfying Crow ď CsCcard.

25535#pg2>

Convergence of AMP Iterates

The paper develops a sequence of families of vectors ty1pA1q uA1Ărns,A1“k´-l, ty2pA2q uA2Ărns,A2“k´-l, and so on<ref:2604.25535#pg2> The construction of the iterates involves replacing the function Tanh with a polynomial activation function f<ref:2604. p. 9

Where the paper says it

, ik are set to zero. Our next step consists in replacing the function Tanh with a polynomial.

25535#pg2> Lemma 14 shows that for each ε ą 0, there is a polynomial pe such that the iterates (15) with f “ pe satisfy lim sup n max iPrns Epxˇk i ´ x˜k i q squared ď ε<ref:2604.25535#pg2> The proof for Theorem 5 involves showing that the conditional variance of the iterates is bounded by a constant, leading to the convergence of the AMP algorithm<ref:2604.25535#pg2>

Bounding Errors and Final Result

The final result shows that lim supklim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε<ref:2604.25535#pg2> This implies that lim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε, which implies that lim supklim supn E‹Tanhpxk q›› 2n1E ď 2ε<ref:2604. p. 2

Where the paper says it

j P rns : s pnq ij ą 0 )ˇ ˇ ˇ ď CcardKn, where | ¨ | is the cardinality of a set. Of course, Assumption 3 implies Assumption 1 with Crow “ supn S pnq satisfying Crow ď CsCcard.

25535#pg2> Using Lemmas 12 and 11, it is shown that lim supn E››m ´ Tanhpxk q›› 2n1E ď 4plog 2q k<ref:2604. p. 4
Where the paper says it

This is shown in Lemma 6 below, which is a straightforward adaptation of [1] to the variance profile case of interest in this paper. Using this result, the “pre-TAP” bound E ˜ mi ´ Tanh ˜ ?

25535#pg2> Denoting as M8 the max norm of the matrix M, it holds that lim supklim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε<ref:2604. p. 1
Where the paper says it

• There exists a constant Cs ą 0 such that s pnq ij ď CsK´1 n for all n and all i, j P rns. • The number Crow “ supn S pnq where |||¨||| is the max row sum norm is finite.

25535#pg2>

Conclusion

The paper establishes that for any arbitrary moment b ą 0, the supremum over all indices and sets is bounded by a constant independent of n<ref:2604.25535#pg2> This implies that lim supklim supn Eˇˇz l iˇˇb ă 8<ref:2604. p. 2

Where the paper says it

j P rns : s pnq ij ą 0 )ˇ ˇ ˇ ď CcardKn, where | ¨ | is the cardinality of a set. Of course, Assumption 3 implies Assumption 1 with Crow “ supn S pnq satisfying Crow ď CsCcard.

25535#pg2> The final result is that lim supklim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε, which implies that lim supn E‹Tanhpxk q›› 2n1E ď 2ε<ref:2604. p. 2
Where the paper says it

j P rns : s pnq ij ą 0 )ˇ ˇ ˇ ď CcardKn, where | ¨ | is the cardinality of a set. Of course, Assumption 3 implies Assumption 1 with Crow “ supn S pnq satisfying Crow ď CsCcard.

25535#pg2> The paper closes with remarks on the necessity of the Kn ≥ log n condition and future research directions<ref:2604.25535#pg2>

References

[1] A. Adhikari, C. p. 0

Where the paper says it

In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations. The dynamical approach of Adhikari et al.

p. 0
Where the paper says it

In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations. The dynamical approach of Adhikari et al.

Brennecke, P. von Soosten, and H.-T. Yau. Dynamical approach to the TAP equations for the Sherrington–Kirkpatrick model<ref:2604. p. 0
Where the paper says it

In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations. The dynamical approach of Adhikari et al.

p. 0
Where the paper says it

In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations. The dynamical approach of Adhikari et al.

25535#pg2>

[2] D. Alberici, F. Camilli, P. Contucci, and E. Mingione. The multi-species mean-field spin-glass on the Nishimori line<ref:2604. p. 0

Where the paper says it

In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix. In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations.

p. 0
Where the paper says it

In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix. In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations.

25535#pg2>

[3] D. Alberici, P. Contucci, E. Mingione, and F. Zimmaro. Ferromagnetic Ising model on multiregular random graphs<ref:2604. p. 0

Where the paper says it

Let Wpnq “ ” W pnq ij ın i,j“1 be a real symmetric n ˆ n random matrix such that the random variables tW pnq ij u1ďiăjďn are independent, and such that W pnq ij „ N p0, tspnq ij q for all i, j P rns and for some t ą 0. Let Σn “ t´1, `1u n be the space of vectors of Ising spins with size n.

p. 0
Where the paper says it

Let Wpnq “ ” W pnq ij ın i,j“1 be a real symmetric n ˆ n random matrix such that the random variables tW pnq ij u1ďiăjďn are independent, and such that W pnq ij „ N p0, tspnq ij q for all i, j P rns and for some t ą 0. Let Σn “ t´1, `1u n be the space of vectors of Ising spins with size n.

25535#pg2>

[4] A. S. Bandeira and R. van Handel. Sharp nonasymptotic bounds on the norm of random matrices with independent entries<ref:2604. p. 7

Where the paper says it

From Inequalities (9) and (10), we have }Ep ´ EtSgpXpq}n ď C{ a Kn and ˇ ˇE}p} 2 n ´ E}tSgpXpq}2 n ˇ ˇ ď C{ a Kn, where we wrote Xp “ up ` p1 ´ uqq for notational simplicity. Since the spectral norm }S} satisfies }S} ď |||S|||, it holds by Assumptions 1 and 2 that the function x P R n ` ÞÑ tSgpxq is Lipschitz for the Euclidean norm with a Lipschitz constant α ă 1 independent of n.

p. 7
Where the paper says it

From Inequalities (9) and (10), we have }Ep ´ EtSgpXpq}n ď C{ a Kn and ˇ ˇE}p} 2 n ´ E}tSgpXpq}2 n ˇ ˇ ď C{ a Kn, where we wrote Xp “ up ` p1 ´ uqq for notational simplicity. Since the spectral norm }S} satisfies }S} ď |||S|||, it holds by Assumptions 1 and 2 that the function x P R n ` ÞÑ tSgpxq is Lipschitz for the Euclidean norm with a Lipschitz constant α ă 1 independent of n.

25535#pg2>

[5] A. Barra, P. Contucci, E. Mingione, and D Tantari. Multi-species mean field spin glasses<ref:2604. p. 0

Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

p. 0
Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

25535#pg2>

[6] E. Bates, L Sloman, and Y Sohn. Replica symmetry breaking in multi-species SherringtonKirkpatrick model<ref:2604. p. 1

Where the paper says it

The conditions on the variance profile that can be found in the mathematical physics literature are usually much more restrictive than those provided by Assumption 1. Regarding the free energy computation, this literature is mainly limited to the so-called multi-species model, where it is assumed that S pnq consists in a finite number of blocks which dimensions scale with n (implying that Kn “ n in our setting), and typically, that S pnq is a non-negative matrix in the semi-definite positive ordering.

p. 1
Where the paper says it

The conditions on the variance profile that can be found in the mathematical physics literature are usually much more restrictive than those provided by Assumption 1. Regarding the free energy computation, this literature is mainly limited to the so-called multi-species model, where it is assumed that S pnq consists in a finite number of blocks which dimensions scale with n (implying that Kn “ n in our setting), and typically, that S pnq is a non-negative matrix in the semi-definite positive ordering.

25535#pg2>

[7] E Bates and Y Sohn. Balanced multi-species spin glasses<ref:2604. p. 0

Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

p. 0
Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

25535#pg2>

[8] M Bayati, M Lelarge, and A Montanari. Universality in polytope phase transitions and message passing algorithms<ref:2604. p. 0

Where the paper says it

In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix. In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations.

p. 0
Where the paper says it

In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix. In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations.

25535#pg2>

[9] E Bolthausen. An iterative construction of solutions of the TAP equations for the Sherrington–Kirkpatrick model<ref:2604. p. 0

Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

p. 0
Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

25535#pg2>

[10] H.-B. Chen. On free energy of non-convex multi-species spin glasses<ref:2604. p. 1

Where the paper says it

The conditions on the variance profile that can be found in the mathematical physics literature are usually much more restrictive than those provided by Assumption 1. Regarding the free energy computation, this literature is mainly limited to the so-called multi-species model, where it is assumed that S pnq consists in a finite number of blocks which dimensions scale with n (implying that Kn “ n in our setting), and typically, that S pnq is a non-negative matrix in the semi-definite positive ordering.

p. 1
Where the paper says it

The conditions on the variance profile that can be found in the mathematical physics literature are usually much more restrictive than those provided by Assumption 1. Regarding the free energy computation, this literature is mainly limited to the so-called multi-species model, where it is assumed that S pnq consists in a finite number of blocks which dimensions scale with n (implying that Kn “ n in our setting), and typically, that S pnq is a non-negative matrix in the semi-definite positive ordering.

25535#pg2>

[11] H.-B. Chen, V Issa, and J. p. 4

Where the paper says it

We note here that in the references cited above which deal with the multi-species model, it is assumed that S is a non-negative matrix specifically to force this term to be non-positive. Second, the bound (5) taken for u “ 1 leads to the construction (13) for approximating mi , as was done in Chen and Tang in [12] for the SK model.

-C Mourrat. The convex structure of the Parisi formula for multi-species spin glasses<ref:2604. p. 0
Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

p. 0
Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

25535#pg2>

[12] W.-K. Chen and S Tang. p. 4

Where the paper says it

We note here that in the references cited above which deal with the multi-species model, it is assumed that S is a non-negative matrix specifically to force this term to be non-positive. Second, the bound (5) taken for u “ 1 leads to the construction (13) for approximating mi , as was done in Chen and Tang in [12] for the SK model.

On convergence of the cavity and Bolthausen’s TAP iterations to the local magnetization<ref:2604.25535#pg2>

[13] F Comets and J Neveu. The Sherrington-Kirkpatrick model of spin glasses and stochastic calculus: the high temperature case<ref:2604. p. 0

Where the paper says it

fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse. In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix.

p. 0
Where the paper says it

fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse. In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix.

25535#pg2>

[14] P.S Dey and Q Wu. Fluctuation results for multi-species Sherrington-Kirkpatrick model in the replica symmetric regime<ref:2604. p. 1

Where the paper says it

Considering the potential applications, graph max κ–cut problems and low-rank matrix estimation problems related with the multi-species model were considered in [19] and in [17] respectively. It would be useful to extend these results to more flexible and more general variance profiles such as the ones considered in this paper.

p. 1
Where the paper says it

Considering the potential applications, graph max κ–cut problems and low-rank matrix estimation problems related with the multi-species model were considered in [19] and in [17] respectively. It would be useful to extend these results to more flexible and more general variance profiles such as the ones considered in this paper.

25535#pg2>

[15] M M Dˇzrbaˇsyan and A B Tavadyan. On weighted uniform approximation by polynomials of functions of several variables<ref:2604.25535#pg2>

[16] O Y Feng, R Venkataramanan, C Rush, and R J Samworth. A unifying tutorial on Approximate Message Passing<ref:2604. p. 0

Where the paper says it

In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix. In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations.

p. 0
Where the paper says it

In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix. In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm based on the Thouless–Anderson–Palmer equations.

25535#pg2>

[17] A Guionnet, J Ko, F Krzakala, and L Zdeborov´a. Low-rank matrix estimation with inhomogeneous noise<ref:2604. p. 1

Where the paper says it

Diluted variants of the multi-species model are considered in [21, 3]. Considering the potential applications, graph max κ–cut problems and low-rank matrix estimation problems related with the multi-species model were considered in [19] and in [17] respectively.

p. 1
Where the paper says it

Diluted variants of the multi-species model are considered in [21, 3]. Considering the potential applications, graph max κ–cut problems and low-rank matrix estimation problems related with the multi-species model were considered in [19] and in [17] respectively.

25535#pg2>

[18] W Hachem. Approximate message passing for sparse matrices with application to the equilibria of large ecological Lotka-Volterra systems<ref:2604. p. 1

Where the paper says it

Denoting as usual as x¨y the mean operator with respect to the measure pGpnq q b8, the second aim of this paper is to provide a large–n approximation of the mean vector mpnq “ xσy. Returning to the classical SK model, it is well-known that this vector can be approximated in the high temperature regime by an Approximate Message Passing (AMP) algorithm by building on the so-called Thouless-Anderson-Palmer (TAP) equations [9].

p. 1
Where the paper says it

Denoting as usual as x¨y the mean operator with respect to the measure pGpnq q b8, the second aim of this paper is to provide a large–n approximation of the mean vector mpnq “ xσy. Returning to the classical SK model, it is well-known that this vector can be approximated in the high temperature regime by an Approximate Message Passing (AMP) algorithm by building on the so-called Thouless-Anderson-Palmer (TAP) equations [9].

25535#pg2>

[19] A Jagannath, J Ko, and S Sen. Max κ-cut and the inhomogeneous Potts spin glass<ref:2604. p. 1

Where the paper says it

Diluted variants of the multi-species model are considered in [21, 3]. Considering the potential applications, graph max κ–cut problems and low-rank matrix estimation problems related with the multi-species model were considered in [19] and in [17] respectively.

p. 1
Where the paper says it

Diluted variants of the multi-species model are considered in [21, 3]. Considering the potential applications, graph max κ–cut problems and low-rank matrix estimation problems related with the multi-species model were considered in [19] and in [17] respectively.

25535#pg2>

[20] H Kim. On the de Almeida–Thouless transition surface in the multi-species SK model with centered Gaussian external field<ref:2604. p. 0

Where the paper says it

The measure of a singleton tσu Ă Σn by Gpnq is Gpnq pσq “ exppHpnq pσqq{Z pnq , where Hpnq is the Hamiltonian defined as Hpnq pσq “ 1 2 σ JWpnqσ ` h pσ ¨ 1nq, h P R is the amplitude of an external field, and Z pnq “ ř σPΣn exppHpnq pσqq is the partition function. In the particular case where s pnq ij “ 1{n for i ‰ j, this model for the Gibbs measure boils down to the classical Sherrington-Kirkpatrick (SK) model that has been studied at length in the statistical physics literature, as detailed in the treatises [25, 26, 23].

p. 0
Where the paper says it

The measure of a singleton tσu Ă Σn by Gpnq is Gpnq pσq “ exppHpnq pσqq{Z pnq , where Hpnq is the Hamiltonian defined as Hpnq pσq “ 1 2 σ JWpnqσ ` h pσ ¨ 1nq, h P R is the amplitude of an external field, and Z pnq “ ř σPΣn exppHpnq pσqq is the partition function. In the particular case where s pnq ij “ 1{n for i ‰ j, this model for the Gibbs measure boils down to the classical Sherrington-Kirkpatrick (SK) model that has been studied at length in the statistical physics literature, as detailed in the treatises [25, 26, 23].

25535#pg2>

[21] Q Liu and Z Dong. Some rigorous results for the diluted multi-species SK model<ref:2604. p. 1

Where the paper says it

Cases where this matrix can be indefinite were discussed in [14, 10, 27, 7]. Diluted variants of the multi-species model are considered in [21, 3].

p. 1
Where the paper says it

Cases where this matrix can be indefinite were discussed in [14, 10, 27, 7]. Diluted variants of the multi-species model are considered in [21, 3].

25535#pg2>

[22] H N Mhaskar. Weighted polynomial approximation<ref:2604. p. 0

Where the paper says it

Considering a sequence of positive numbers pKnq such that Kn ď n and Kn Ñ 8, we assume the following: Assumption 1. It holds that 1 arXiv:2604.

25535#pg2>

[23] D Panchenko. The Sherrington-Kirkpatrick model<ref:2604. p. 0

Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

p. 0
Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

25535#pg2>

[24] D Panchenko. The free energy in a multi-species Sherrington-Kirkpatrick model<ref:2604. p. 0

Where the paper says it

fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse. In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix.

p. 0
Where the paper says it

fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse. In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix.

25535#pg2>

[25] M Talagrand. Mean field models for spin glasses. p. 0

Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

p. 0
Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

p. 0
Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

p. 0
Where the paper says it

hachem@univ-eiffel. fr April 29, 2026 Abstract A generalization of the Sherrington–Kirkpatrick (SK) model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse.

Volume I, volume 54 of A Series of Modern Surveys in Mathematics<ref:2604. p. 0
Where the paper says it

Considering a sequence of positive numbers pKnq such that Kn ď n and Kn Ñ 8, we assume the following: Assumption 1. It holds that 1 arXiv:2604.

p. 0
Where the paper says it

Considering a sequence of positive numbers pKnq such that Kn ď n and Kn Ñ 8, we assume the following: Assumption 1. It holds that 1 arXiv:2604.

25535#pg2>

[26] M Talagrand. Mean field models for spin glasses. Volume II, volume 55 of A Series of Modern Surveys in Mathematics<ref:2604. p. 0

Where the paper says it

Considering a sequence of positive numbers pKnq such that Kn ď n and Kn Ñ 8, we assume the following: Assumption 1. It holds that 1 arXiv:2604.

p. 0
Where the paper says it

Considering a sequence of positive numbers pKnq such that Kn ď n and Kn Ñ 8, we assume the following: Assumption 1. It holds that 1 arXiv:2604.

25535#pg2>

Pruning redundant reasoning tokens

The model introduces a deterministic symmetric matrix Spnq with zero diagonal and non-negative elements, which defines the variance profile of the interactions<ref:2604.25535#pg2> Assumption 1 imposes conditions on this matrix Spnq, such as the bound spnqij ď CsK´1n for all n and all i, j P rns, and a finiteness condition on the number Crow “ supn Spnq where ¨ is the max row sum norm<ref:2604.25535#pg2> Assumption 2 specifies the high temperature regime by requiring t ă log 2 Crow<ref:2604.25535#pg2>

How the training signal is built

The second aim is to provide a large–n approximation of the mean vector mpnq “ xσy<ref:2604.25535#pg2> To this end, we need to strengthen a bit Assumption 1, setting conditions on sparsity and non-zero entries<ref:2604. p. 2

Where the paper says it

We now tackle the approximation problem of mpnq “ xσy with the help of an AMP algorithm. To this end, we need to strengthen a bit Assumption 1.

25535#pg2> The approximation problem of mpnq “ xσy is tackled using an AMP algorithm based on the Thouless–Anderson–Palmer equations<ref:2604. p. 1
Where the paper says it

Denoting as usual as x¨y the mean operator with respect to the measure pGpnq q b8, the second aim of this paper is to provide a large–n approximation of the mean vector mpnq “ xσy. Returning to the classical SK model, it is well-known that this vector can be approximated in the high temperature regime by an Approximate Message Passing (AMP) algorithm by building on the so-called Thouless-Anderson-Palmer (TAP) equations [9].

25535#pg2> Theorem 5 states that under Assumptions 3 and 2, for each even integer b ě 2, it holds that sup n max iPrns Eˇˇz l 1 pA pnq l q iˇˇb ă 8<ref:2604. p. 8
Where the paper says it

The first bound obtained for u “ 1 in the statement of Lemma 7 can be rewritten as max iPrns Epmi ´ TanhprWmpiq siqq2 ď C{Kn. By a straightforward adaptation of this lemma, we obtain that for a fixed integer M ą 0, it holds that max AĂrns,|A|“M max iPAc E ` mpAq,i ´ Tanh `“WmpAYtiuq‰ i ˘˘2 ď C Kn , (12) This bound will be at the basis of our proof.

25535#pg2> The iterates are constructed using a sequence of families of R n–valued vectors ty1pA1q uA1Ărns,A1“k´-l, ty2pA2q uA2Ărns,A2“k´-l, and so on<ref:2604. p. 9
Where the paper says it

This construction will be at the basis of a series of approximations of the vector m ending with the one provided by the AMP algorithm of Theorem 5. We first define the sequence of families of R n–valued vectors ty 1 pA1q uA1Ărns,|A1|“k´1, ty 2 pA2q uA2Ărns,|A2|“k´2, .

25535#pg2>

Where the gains come from

The convergence of the AMP algorithm is established by showing that E››diagppW d Wqp1c pz l 1 qpcpz l´1q ´ diagptSE Tanh1pXlqq Tanhpx l´1q›› 2n1E ď δpcq<ref:2604.25535#pg2> This convergence is achieved by using the recurrence assumption and from (25) that Eˇz l 1 pA pnq l q iˇˇb ă 8<ref:2604. p. 0

Where the paper says it

Considering a sequence of positive numbers pKnq such that Kn ď n and Kn Ñ 8, we assume the following: Assumption 1. It holds that 1 arXiv:2604.

25535#pg2> The final result shows that lim supklim supn E››m ´ Tanhpxk q›› 2n1E ď 2ε<ref:2604.25535#pg2> This implies that lim supn E‹Tanhpxk q›› 2n1E ď 2ε<ref:2604. p. 2
Where the paper says it

j P rns : s pnq ij ą 0 )ˇ ˇ ˇ ď CcardKn, where | ¨ | is the cardinality of a set. Of course, Assumption 3 implies Assumption 1 with Crow “ supn S pnq satisfying Crow ď CsCcard.

25535#pg2> The paper concludes by showing that lim supklim supn E››m ´ Tanhpxk q›› 2n1E ď 4plog 2q k<ref:2604.25535#pg2>

[1] A. Adhikari, C. Brennecke, P von Soosten, and H-T Yau. Dynamical approach to the TAP equations for the Sherrington–Kirkpatrick model<ref:2604.25535#pg2>

[2] D Alberici, F Camilli, P Contucci, and E Mingione. The multi-species mean-field spin-glass on the Nishimori line<ref:2604.25535#pg2>

[3] D Alberici, P Contucci, E Mingione, and F Zimmaro. Ferromagnetic Ising model on multiregular random graphs<ref:2604.25535#pg2>

[4] A S Bandeira and R van Handel. Sharp nonasymptotic bounds on the norm of random matrices with independent entries<ref:2604.25535#pg2>

[5] A Barra, P Contucci, E Mingione, and D Tantari. Multi-species mean field spin glasses<ref:2604.25535#pg2>

[6] E Bates, L Sloman, and Y Sohn. Replica symmetry breaking in multi-species SherringtonKirkpatrick model<ref:2604.25535#pg2>

[7] E Bates and Y Sohn. Balanced multi-species spin glasses<ref:2604.25535#pg2>

[8] M Bayati, M Lelarge, and A Montanari. Universality in polytope phase transitions and message passing algorithms<ref:2604.25535#pg2>

[9] E Bolthausen. An iterative construction of solutions of the TAP equations for the Sherrington–Kirkpatrick model<ref:2604.25535#pg2>

[10] H-B Chen. On free energy of non-convex multi-species spin glasses<ref:2604.25535#pg2>

[11] H-B Chen, V Issa, and J.-C Mourrat. The convex structure of the Parisi formula for multi-species spin glasses<ref:2604.25535#pg2>

[12] W-K Chen and S Tang. On convergence of the cavity and Bolthausen’s TAP iterations to the local magnetization<ref:2604.25535#pg2>

[13] F Comets and J Neveu. The Sherrington-Kirkpatrick model of spin glasses and stochastic calculus: the high temperature case<ref:2604.25535#pg2>

[14] P S Dey and Q Wu. Fluctuation results for multi-species Sherrington-Kirkpatrick model in the replica symmetric regime<ref:2604.25535#pg2>

[15] M M Dˇzrbaˇsyan and A B Tavadyan. On weighted uniform approximation by polynomials of functions of several variables<ref:2604.25535#pg2>

[16] O Y Feng, R Venkataramanan, C Rush, and R J Samworth. A unifying tutorial on Approximate Message Passing<ref:2604.25535#pg2>

[17] A Guionnet, J Ko, F Krzakala, and L Zdeborov´a. Low-rank matrix estimation with inhomogeneous noise<ref:2604.25535#pg2>

[18] W Hachem. Approximate message passing for sparse matrices with application to the equilibria of large ecological Lotka-Volterra systems<ref:2604.25535#pg2>

[19] A Jagannath, J Ko, and S Sen. Max κ-cut and the inhomogeneous Potts spin glass<ref:2604.25535#pg2>

[20] H Kim. On the de Almeida–Thouless transition surface in the multi-species SK model with centered Gaussian external field<ref:2604.25535#pg2>

[21] Q Liu and Z Dong. Some rigorous results for the diluted multi-species SK model<ref:2604.25535#pg2>

[23] D Panchenko. The Sherrington-Kirkpatrick model<ref:2604.25535#pg2>

[24] D Panchenko. The free energy in a multi-species Sherrington-Kirkpatrick model<ref:2604.25535#pg2>

[25] M Talagrand. Mean field models for spin glasses. Volume I, volume 54 of A Series of Modern Surveys in Mathematics<ref:2604.25535#pg2>

[26] M Talagrand. Mean field models for spin glasses. Volume II, volume 55 of A Series of Modern Surveys in Mathematics<ref:2604.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve free energy estimation for SK models with sparse variance profiles by deriving an asymptotic equivalent at sufficiently high temperatures, regardless of whether the variance profile matrix has a particular structure. The improved system can accurately estimate F n in this regime using a method that is independent of the signature of S pnq.

  2. Enhance mean vector approximation for SK models with sparse variance profiles by generalizing the Approximate Message Passing algorithm based on Thouless-Anderson-Palmer equations to this new model setting in the temperature regime specified by Assumption 2. This allows for a large n approximation of mpnq which is more robust than results derived only for the classical SK model.

  3. Develop a method to compute the free energy of an interpolated Hamiltonian Hu using Guerra’s interpolation trick, showing that the resulting annoying term is negligible due to bounds on E R12 - q, leading to Theorem 2. This provides a tractable path to calculating F n for general variance profiles.

  4. Generalize the result for doubly stochastic matrices S pnq by recovering the asymptotic free energy of the classical SK model at high temperature, as stated in Corollary 3. This allows the system to handle specific structured variance profiles efficiently while retaining generality otherwise.

  5. Implement an AMP algorithm that approximates mpnq using a sequence of non-backtracking iterations, leveraging independence along a path of indices and polynomial activation functions with pseudo-Lipschitz properties to achieve convergence (20). This enables the system to find the mean vector mpnq with high accuracy for sparse models.

  6. Use polynomial activations f epsilon derived from Lemma 9 to ensure that iterates converge rapidly, satisfying lim sup n E > Tanhpx k q > 2n 1 E < 2 epsilon, leading directly to Theorem 5. This ensures the AMP approximation converges strongly towards the true mean vector mpnq.

Abstract

A generalization of the Sherrington-Kirkpatrick model for spin glasses is considered, in which the interaction matrix is endowed with a variance profile that has no particular structure and may be sparse. In the first part of the paper, an asymptotic equivalent of the free energy is derived at sufficiently high temperatures, regardless of the signature of the variance profile matrix. In the second part, the mean of the spin vector under the Gibbs measure is estimated using an Approximate Message Passing algorithm that generalizes the one used for the SK case. The dynamical approach of Adhikari et al (J Stat Phys 2021), originally developed for the classical SK model, is adapted to the present setting to obtain these results.

Sources

Related papers