Synchronous and asynchronous Monte Carlo dynamics of mean-field spin glasses with rotationally invariant couplings: what they share and where they differ

arXiv:2610.24942 · cond-mat.dis-nn · Submitted 2026-10-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: We're looking at the paper: "Synchronous and asynchronous Monte Carlo dynamics of mean-field spin glasses with rotationally invariant couplings".

Mira: Here's the gist of the paper. Synchronous and asynchronous Monte Carlo dynamics of mean-field spin glasses with rotationally invariant couplings: what they share and where they differ compares two Markov chain Monte Carlo methods for Ising spin glasses with rotationally invariant couplings, revealing that while their slow-time-scale equations coincide, their specific update rules lead to differences in relaxation times and wall-clock efficiency <ref:2610.24942#pg2> Common Effective Dynamics The effective single-spin processes of the two dynamics share a retarded self-interaction R(G)−κ1 and a Gaussian noise with covariance Σϕ

Eqs. [Kai: (15) and (20)] <ref:2610.24942#pg4>.

Mira: These shared terms represent the common structure of the dynamics for both methods <ref:2610.24942#pg4>.

Kai: Furthermore, the fluctuation–dissipation theorem (FDT) at inverse temperature βe states that G(τ) = −βe C′(τ) Θ(τ), which holds for any spectrum <ref:2610.24942#pg9>.

Paper summary: Kai: So, we're looking at this paper, "Synchronous and asynchronous Monte Carlo dynamics of mean-field spin glasses with rotationally invariant couplings: what they share and where they differ". Mira It’s comparing two ways to simulate Ising spin glasses: the synchronous Monte Carlo method, SyncMC, which updates everything at once with some auxiliary fields, against the standard asynchronous Metropolis–Hastings dynamics. Kai The main point here is figuring out what's common between these two methods and what makes them different in terms of how they relax. Mira They claim that while the slow-time-scale equations for both systems end up being the same, SyncMC has specific extra terms that only affect the fastest time scale. Lev That's interesting because on real hardware, we always worry about those fast dynamics; if those extra terms don't show up in our measurements, they might be negligible.

Kai: Right, so what does this mean for us? The paper says the effective single-spin processes share a retarded self-interaction and a Gaussian noise. Mira That shared structure is key because it's governed by the fluctuation–dissipation theorem at inverse temperature beta e, which holds for any spectrum they tested <ref:2610.24942#pg9>. Lev We need that FDT to hold if we want to trust the slow dynamics description.

Kai: But SyncMC has these specific terms: an instantaneous self-coupling of strength kappa one plus c, and a Gaussian noise whose covariance depends only on the response function. Mira These terms are what make the two methods behave differently in practice, particularly in how fast they relax. Lev If we're thinking about running this on real hardware with finite size N, those specific SyncMC terms might introduce different relaxation times compared to standard asynchronous steps.

Kai: The paper shows that because of time-scale separation, these syncmc-specific terms don't change the slow equations when you look at long time separations. Mira That means we can use the same slow description for both dynamics, even though their specific update rules are different in the short term. Lev So, from a simulation perspective, we can use either method to study aging or slow relaxation without worrying about which one is fundamentally better for the long run of the physics.

Kai: The results they show are quite concrete when comparing them. Mira In the paramagnetic phase, SyncMC relaxes much more slowly than asynchronous MH by about one order of magnitude. Lev That's a significant slowdown, maybe ten times slower in that regime for this specific model setup <ref:2610.24942#pg3>.

Kai: And when we look at the aging regime of the random-over-mean ROM, one Monte Carlo step in asynchronous MH corresponds to about three hundred forty to four hundred seventy SyncMC steps. Mira That number is quite large compared to what we might expect for simpler models. Lev If we were trying to build an error correction scheme, that overhead of hundreds of steps per simulation would be a major concern for wall-clock efficiency <ref:2610.24942#pg3>.

Kai: They also give a comparison to the Sherrington–Kirkpatrick model. Mira For that SK model in the aging regime, one asynchronous Monte Carlo step corresponds to about fifteen to eighteen SyncMC steps <ref:2610.24942#pg3>.

Lev: That difference in scaling between the ROM and the SK model is what really matters for hardware deployment. Kai It suggests that a purely synchronous approach, if it can be implemented efficiently on a GPU, might still have some wall-clock overhead compared to sequential asynchronous updates. Mira The practical consequence they point out is that for the ROM in its aging regime, the advantage of SyncMC in terms of raw speed shrinks to about a factor of ten <ref:2610.24942#pg3>.

Kai: So what does this mean for someone just listening? It means that if you're trying to understand how these spin glasses behave over long times, you can use either simulation method because the physics is governed by the same slow equations <ref:2610.24942#pg3>. Mira The paper shows that even though the details of their update rules differ, they converge on the same fluctuation–dissipation relations at inverse temperature beta e.

Lev: From an engineering standpoint, it's reassuring to know that we have a clear picture of where the overhead lands—it's not a massive slowdown for all regimes, just specific ones <ref:2610.24942#pg3>.

Kai: The authors stress that these results hold for any rotationally invariant ensemble. Mira That’s broad, which is good because it means this isn't just tied to one specific physical system but applies generally to many spin glass problems <ref:2610.24942#pg1>.

Lev: If we look at the practical implications for building these things, the paper notes they found no finite-size effect for N greater than four thousand ninety-six and only a weak indication at N equals two thousand forty-eight for SyncMC <ref:2610.24942#pg3>. Kai That suggests that we can get pretty good results on larger systems if we use the SyncMC approach, even though it's slower in some cases.

Mira: Overall, the paper shows a clear division: the physics is described by a common slow description, but the way you actually simulate it—synchronously or asynchronously—has different costs in terms of time and efficiency <ref:2610.24942#pg1>.

Lev: It seems like the main contribution is clarifying that distinction, which helps us decide how much overhead to tolerate when we move from theory to actual error correction protocols <ref:2610.24942#pg3>.

Conclusion: Kai: So we've been looking at how these two Monte Carlo methods, synchronous versus asynchronous dynamics, behave when simulating mean-field spin glasses with rotationally invariant couplings.

Mira: The main thing the authors are showing us is that even though they use different update rules for their simulations, the slow time-scale physics—the long-term behavior of the system—actually ends up being governed by the same set of equations.

Lev: That's what we expected because they both share those fundamental terms like a retarded self-interaction and some kind of Gaussian noise that follows the fluctuation–dissipation theorem at inverse temperature beta e.

Kai: Exactly, but then they point out that SyncMC has these extra terms specific to its update rule, like an instantaneous self-coupling of strength kappa one plus c.

Mira: And those specific terms only show up on the fastest time scale—the very short bursts of action in the simulation.

Lev: So when we talk about long aging or slow relaxation, those extra SyncMC terms don't change the outcome because they vanish at slow time separations and leave the local equilibrium unchanged.

Kai: That means we can use either method to study the physics of these spin glasses over a long time, whether it's for theoretical insight or for checking what we build on hardware.

Mira: The paper really boils down to this: you don't need to pick one simulation technique if you only care about the final state of the system in terms of its slow evolution.

Lev: The practical part is that the authors show specific overheads, like how many steps it takes for one Monte Carlo step in asynchronous versus SyncMC in certain regimes.

Kai: They give us conversion factors, like one asynchronous step equaling hundreds of SyncMC steps for some models, which tells you where the computational cost actually lands.

Mira: It sets up a clear picture of the trade-off between simulation speed and capturing the correct physical dynamics.

Lev: So what this means for our work is that we need to be careful about which method we use depending on whether we're looking at fast relaxation or long aging.

Kai: It’s a good reminder that even in complex systems, the underlying physics often has a common language, but the practical execution still matters a lot.

Institute for Physics of Intelligence, the University of Tokyo · Trans-scale Quantum Science Institute, The University of Tokyo

cond-mat.dis-nn

Submitted: 2026-10-18

Updated: 2026-11-03

Comments: 46 pages, 5 figures

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

Importance score: 96/100

The gist: Synchronous and asynchronous Monte Carlo dynamics of mean-field spin glasses with rotationally invariant couplings…

Key concepts

Slow Time Scale Equations
These are simplified equations describing how the system evolves over very long time scales, ignoring fast, rapid fluctuations. Both synchronous and asynchronous methods yield the same slow-time evolution because they both adhere to local equilibrium conditions and the fluctuation–dissipation theorem on fast time scales.
Fluctuation–Dissipation Theorem (FDT)
The FDT relates the response of a system to external changes (like a spin flip) with its spontaneous fluctuations. In this context, it states that the correlation function G(τ) is directly related to the noise C'(τ) via an inverse temperature betae, ensuring consistency across both dynamics.
SyncMC Specific Terms
Synchronous Monte Carlo introduces unique terms into its update rule, such as an instantaneous self-coupling and a specific type of Gaussian noise. These terms only affect the fastest time scale but do not change the slow-time equations because they vanish at slow separations.

Terminology

Summary

Synchronous and asynchronous Monte Carlo dynamics of mean-field spin glasses with rotationally invariant couplings: what they share and where they differ compares two Markov chain Monte Carlo methods for Ising spin glasses with rotationally invariant couplings, revealing that while their slow-time-scale equations coincide, their specific update rules lead to differences in relaxation times and wall-clock efficiency <ref:2610. p. 0

Where the paper says it

Synchronous and asynchronous Monte Carlo dynamics of mean-field spin glasses with rotationally invariant couplings: what they share and where they differ Yoshiyuki Kabashima1, 2 1 Institute for Physics of Intelligence, The University of Tokyo, Tokyo 113-0033, Japan 2Trans-Scale Quantum Science Institute, The University of Tokyo, Tokyo 113-0033, Japan (Dated: October 20, 2026) We compare two Markov chain Monte Carlo dynamics for Ising spin glasses with rotationally invariant couplings: the synchronous Monte Carlo method (SyncMC), which updates all spins in parallel with the help of Gaussian auxiliary fields, and asynchronous Metropolis–Hastings dynamics, which updates one spin at a time. Using the Guionnet–Ma¨ıda formula, we derive the dynamical mean-field theory of both dynamics in terms of the R-transform of the coupling spectrum.

24942#pg2>

Common Effective Dynamics

The effective single-spin processes of the two dynamics share a retarded self-interaction R(G)−κ1 and a Gaussian noise with covariance Σϕ[G, C] [Eqs. (15) and (20)] <ref:2610. p. 0

Where the paper says it

Such dynamics satisfy detailed balance with respect to the Boltzmann distribution. On modern parallel hardware, however, updating one arXiv:2610.

24942#pg4>. These shared terms represent the common structure of the dynamics for both methods <ref:2610. p. 9
Where the paper says it

The dynamics of the spin itself is different, however. Sompolinsky and Zippelius treated soft spins obeying a Langevin equation, whereas here an Ising spin flips with the MH rate; the resulting C and G are therefore different functions, and only the structure of the effective field is shared.

24942#pg4>. Furthermore, the fluctuation–dissipation theorem (FDT) at inverse temperature βe states that G(τ) = −βe C′(τ) Θ(τ), which holds for any spectrum <ref:2610. p. 3
Where the paper says it

The inverse temperature is β. For any constant c, the spectrum of J + cI has R-transform R(z) + c; that is, only κ1 is shifted by c.

24942#pg9>. This property is a key ingredient of the slow-time-scale analysis <ref:2610. p. 0
Where the paper says it

The two effective single-spin processes share a retarded self-interaction and a Gaussian noise; SyncMC has, in addition, an instantaneous self-coupling and a noise that depends only on the response function. These additional terms act only on the fastest time scale, so that the equations for slow time scales, including the fluctuation–dissipation ratio and, for the Sherrington–Kirkpatrick model, the Parisi equation, are the same for the two dynamics.

24942#pg9>.

SyncMC Specific Terms

Synchronous Monte Carlo (SyncMC) has additional terms that are specific to its update rule, namely an instantaneous self-coupling of strength κ1 + c, where c is the constant added to the diagonal of the coupling matrix in SyncMC <ref:2610.24942#pg4>. It also introduces a Gaussian noise whose covariance β−1D[G] depends only on the response function <ref:2610. p. 2

Where the paper says it

SyncMC has, in addition, an instantaneous self-coupling of strength κ1 + c, where c is the constant added to the diagonal of the coupling matrix in SyncMC, and a Gaussian noise whose covariance β −1D[G] depends only on the response function [Eq. These SyncMC-specific terms affect only the fastest time scale.

24942#pg4>. These SyncMC-specific terms act only on the fastest time scale, so that the equations for slow time scales coincide with asynchronous MH <ref:2610. p. 0
Where the paper says it

The two effective single-spin processes share a retarded self-interaction and a Gaussian noise; SyncMC has, in addition, an instantaneous self-coupling and a noise that depends only on the response function. These additional terms act only on the fastest time scale, so that the equations for slow time scales, including the fluctuation–dissipation ratio and, for the Sherrington–Kirkpatrick model, the Parisi equation, are the same for the two dynamics.

24942#pg4>.

Slow Time Scale Invariance

The slow-time-scale equations of the two dynamics coincide because they are determined by local equilibrium of a spin in its field and the fluctuation–dissipation theorem on fast time scales, both following detailed balance <ref:2610.24942#pg9>. The SyncMC-specific terms do not alter these two quantities because the instantaneous self-coupling leaves the local equilibrium unchanged, and the covariance of noise beyond its white part vanishes at slow time separations. p. 2

Where the paper says it

SyncMC has, in addition, an instantaneous self-coupling of strength κ1 + c, where c is the constant added to the diagonal of the coupling matrix in SyncMC, and a Gaussian noise whose covariance β −1D[G] depends only on the response function [Eq. These SyncMC-specific terms affect only the fastest time scale.

Differences in Relaxation and Efficiency

The two dynamics differ in the shape of the fast relaxation CF, GF, which depends on c for SyncMC. In the paramagnetic phase, SyncMC relaxes much more slowly than asynchronous MH by about one order of magnitude. p. 2

Where the paper says it

In the paramagnetic phase the effective processes reproduce the finite-N data of both dynamics. Below the transition, the fluctuation– dissipation relations of both dynamics approach the same predictions: the one-stepaging form with q = 0.

In the aging regime of the ROM, one MCS corresponds to 340–470 SyncMC steps, whereas for SK model it is 15–18 steps. p. 2
Where the paper says it

1, and the Parisi solution for the SK model at T = 0. For the same coupling matrices, the two dynamics give the same integrated response at fixed correlation when their waiting times are related by the time needed to reach the same energy; one MCS of asynchronous MH then corresponds to about 102–103 SyncMC steps for the ROM at T = 0.

Numerical Implementation and Conclusions

The simulations support this picture, showing that the two dynamics approach the same fluctuation–dissipation relations. The practical consequence is that slow relaxation and aging of these models can be studied with either dynamics. p. 2

Where the paper says it

The update rule thus sets how fast a system moves through its configuration space, but not which states it visits on slow time scales. In practice, slow relaxation and aging of these models can be studied with either dynamics.

The conclusions rest on several assumptions, including the separation of time scales, which are consistent with the data observed. p. 0
Where the paper says it

The two effective single-spin processes share a retarded self-interaction and a Gaussian noise; SyncMC has, in addition, an instantaneous self-coupling and a noise that depends only on the response function. These additional terms act only on the fastest time scale, so that the equations for slow time scales, including the fluctuation–dissipation ratio and, for the Sherrington–Kirkpatrick model, the Parisi equation, are the same for the two dynamics.

The analysis applies to any rotationally invariant ensemble. p. 1
Where the paper says it

For SyncMC, the DMFT was derived for the SK model [2] and for the random orthogonal model (ROM) [1, 9]. In this paper we derive the DMFT of SyncMC and of asynchronous MH in a common form, for Ising spin glasses whose coupling matrix is drawn from a rotationally invariant ensemble.

p. 1
Where the paper says it

For SyncMC, the DMFT was derived for the SK model [2] and for the random orthogonal model (ROM) [1, 9]. In this paper we derive the DMFT of SyncMC and of asynchronous MH in a common form, for Ising spin glasses whose coupling matrix is drawn from a rotationally invariant ensemble.

p. 1
Where the paper says it

For SyncMC, the DMFT was derived for the SK model [2] and for the random orthogonal model (ROM) [1, 9]. In this paper we derive the DMFT of SyncMC and of asynchronous MH in a common form, for Ising spin glasses whose coupling matrix is drawn from a rotationally invariant ensemble.

p. 1
Where the paper says it

For SyncMC, the DMFT was derived for the SK model [2] and for the random orthogonal model (ROM) [1, 9]. In this paper we derive the DMFT of SyncMC and of asynchronous MH in a common form, for Ising spin glasses whose coupling matrix is drawn from a rotationally invariant ensemble.

p. 1
Where the paper says it

For SyncMC, the DMFT was derived for the SK model [2] and for the random orthogonal model (ROM) [1, 9]. In this paper we derive the DMFT of SyncMC and of asynchronous MH in a common form, for Ising spin glasses whose coupling matrix is drawn from a rotationally invariant ensemble.

Practical Consequences

The conversion factor between SyncMC steps and MCS depends on the model and regime, being about 10 in the paramagnetic phase of the ROM and 15–18 in the aging regime of the SK model. The advantage of SyncMC in wall-clock time shrinks to a factor of about 10 for the ROM in its aging regime, suggesting that an asynchronous MH with sequential updates fused into a single GPU kernel could reduce or remove this advantage. p. 2

Where the paper says it

In practice, slow relaxation and aging of these models can be studied with either dynamics. SyncMC needs more updates to reach a given state but performs each update for all spins in parallel; on a GPU this gives a shorter wallclock time in our implementation, by a factor that depends on the model and the regime (Sec.

The paper finds no finite-size effect for N ≥ 4096 and only a weak indication at N = 2048 for SyncMC. p. 1
Where the paper says it

III B and Appendix A of the present paper show how random sequential trials become a continuous-time single-spin process as N → ∞. For SyncMC, the DMFT was derived for the SK model [2] and for the random orthogonal model (ROM) [1, 9].

The analysis applies to any rotationally invariant ensemble. The data and code are available at Zenodo [15]. p. 2
Where the paper says it

Simulations of both dynamics, together with numerical solutions of the effective processes, support this picture (Sec. In the paramagnetic phase the effective processes reproduce the finite-N data of both dynamics.

p. 2
Where the paper says it

Simulations of both dynamics, together with numerical solutions of the effective processes, support this picture (Sec. In the paramagnetic phase the effective processes reproduce the finite-N data of both dynamics.

p. 2
Where the paper says it

Simulations of both dynamics, together with numerical solutions of the effective processes, support this picture (Sec. In the paramagnetic phase the effective processes reproduce the finite-N data of both dynamics.

Appendix Details

The discrete-time FDT of SyncMC is given by Eq. p. 2

Where the paper says it

In practice, slow relaxation and aging of these models can be studied with either dynamics. SyncMC needs more updates to reach a given state but performs each update for all spins in parallel; on a GPU this gives a shorter wallclock time in our implementation, by a factor that depends on the model and the regime (Sec.

p. 2
Where the paper says it

In practice, slow relaxation and aging of these models can be studied with either dynamics. SyncMC needs more updates to reach a given state but performs each update for all spins in parallel; on a GPU this gives a shorter wallclock time in our implementation, by a factor that depends on the model and the regime (Sec.

(25), which reduces to the continuous-time form for ω≪ 1. p. 1
Where the paper says it

Answering this question requires a description in which both dynamics can be compared term by term. For mean-field models, dynamical mean-field theory (DMFT) provides such a description: in the limit of a large number of spins N, it reduces the dynamics of the whole system to that of a single spin in an effective field.

The noise covariance Σϕ is given by Eq. p. 9
Where the paper says it

(15) and (16) give ΓMH = G and Σϕ = C. The effective field then has the same structure as in the effective process of Sompolinsky and Zippelius [3]: a retarded self-interaction given by the response function and a Gaussian noise whose covariance is the correlation function.

(21). The derivation of Eq. (26) shows that the common noise satisfies the FDT at the same inverse temperature βe. p. 3
Where the paper says it

We write R(x) = R x 0 R(v) dv. The inverse temperature is β.

The Langevin equation for the slow components of the local field is given by Eq. p. 9
Where the paper says it

The dynamics of the spin itself is different, however. Sompolinsky and Zippelius treated soft spins obeying a Langevin equation, whereas here an Ising spin flips with the MH rate; the resulting C and G are therefore different functions, and only the structure of the effective field is shared.

(F8). The stationary distribution satisfies detailed balance and has the temperature vk/wk = T/xk. p. 0
Where the paper says it

INTRODUCTION Markov chain Monte Carlo (MCMC) sampling of spin systems is usually carried out with asynchronous updates: one spin is chosen at a time and updated according to the Metropolis–Hastings (MH) or heat-bath rule. Such dynamics satisfy detailed balance with respect to the Boltzmann distribution.

p. 0
Where the paper says it

INTRODUCTION Markov chain Monte Carlo (MCMC) sampling of spin systems is usually carried out with asynchronous updates: one spin is chosen at a time and updated according to the Metropolis–Hastings (MH) or heat-bath rule. Such dynamics satisfy detailed balance with respect to the Boltzmann distribution.

The plateau values follow from qk = E(∂yΦk(yk))2. p. 7
Where the paper says it

The subscript k labels the T time points of Sec. III A, and σk, uk, ˆuk, θk are the values at these points.

The analysis applies to any rotationally invariant ensemble. The data and code are available at Zenodo [15].

Appendix G

This appendix describes a way to solve the effective process of asynchronous MH that avoids discretization of flip times in the Bernoulli scheme. It involves generating candidate times from a Poisson process with rate one and sampling the spin at these times. p. 3

Where the paper says it

In one MCS each spin is chosen once on average, but the number of times a given spin is chosen fluctuates: it is Poisson distributed with mean one as N → ∞. In this limit, t becomes a continuous variable and each spin becomes a continuous-time jump process with flip rate W(σ, h) = min(1, e−2βσh).

The cost per evaluation of the retarded term does not depend on any time step. p. 4
Where the paper says it

In both cases one unit of t contains N single-spin update opportunities, so we compare the two dynamics with one MCS of asynchronous MH corresponding to one SyncMC step. The difference lies in how these opportunities are distributed: in SyncMC every spin is updated exactly once per unit time, all at the same instant, whereas in asynchronous MH each spin is chosen at random instants whose number per unit time is Poisson distributed with mean one.

p. 4
Where the paper says it

In both cases one unit of t contains N single-spin update opportunities, so we compare the two dynamics with one MCS of asynchronous MH corresponding to one SyncMC step. The difference lies in how these opportunities are distributed: in SyncMC every spin is updated exactly once per unit time, all at the same instant, whereas in asynchronous MH each spin is chosen at random instants whose number per unit time is Poisson distributed with mean one.

The correlation function on the grid is computed exactly from the lists of flip times. p. 8
Where the paper says it

9 where τ − j is the instant just before the j-th flip. This is a density with respect to dτ1 · · · dτn; correspondingly, the sum over trajectories on the grid becomes X σ(0) X n≥0 Z 0<τ1<···<τn<L dτ1 · · · dτn.

The noise at the candidate times is obtained by conditional sampling from its values on this grid, and this behaves locally like an Ornstein–Uhlenbeck process. The representation by flip times is analogous to the kink (segment) representation of continuous-time quantum Monte Carlo methods. p. 0
Where the paper says it

Synchronous and asynchronous Monte Carlo dynamics of mean-field spin glasses with rotationally invariant couplings: what they share and where they differ Yoshiyuki Kabashima1, 2 1 Institute for Physics of Intelligence, The University of Tokyo, Tokyo 113-0033, Japan 2Trans-Scale Quantum Science Institute, The University of Tokyo, Tokyo 113-0033, Japan (Dated: October 20, 2026) We compare two Markov chain Monte Carlo dynamics for Ising spin glasses with rotationally invariant couplings: the synchronous Monte Carlo method (SyncMC), which updates all spins in parallel with the help of Gaussian auxiliary fields, and asynchronous Metropolis–Hastings dynamics, which updates one spin at a time. Using the Guionnet–Ma¨ıda formula, we derive the dynamical mean-field theory of both dynamics in terms of the R-transform of the coupling spectrum.

The cost per evaluation of the retarded term does not depend on any time step.

Appendix B

The saddle point for asynchronous MH yields the coupling −ihˆ⊤R(G)σ and a quadratic term −1/2 hˆ⊤Σϕhˆ in the single-site weight. The matrix function RA(KQ) is given by Eq. p. 9

Where the paper says it

(15) and (16) give ΓMH = G and Σϕ = C. The effective field then has the same structure as in the effective process of Sompolinsky and Zippelius [3]: a retarded self-interaction given by the response function and a Gaussian noise whose covariance is the correlation function.

(C4). The noise covariance Σϕ in Eq. p. 1
Where the paper says it

The main results are as follows. The effective processes of the two dynamics share a retarded self-interaction R(G)−κ1 and a Gaussian noise with covariance Σϕ[G, C] [Eqs.

(16) is given by the upper-right block of R(M) = Pnκn+1Mn. p. 9
Where the paper says it

III A with v = σ, vˆ = xˆ and K =   0 −iIT −iIT −β −1 IT   . (18) The new feature compared with MH is the lower-right block −β −1 IT , which comes from the noise term in Eq.

The effective process is given by Eq. p. 9
Where the paper says it

(15) and (16) give ΓMH = G and Σϕ = C. The effective field then has the same structure as in the effective process of Sompolinsky and Zippelius [3]: a retarded self-interaction given by the response function and a Gaussian noise whose covariance is the correlation function.

p. 9
Where the paper says it

(15) and (16) give ΓMH = G and Σϕ = C. The effective field then has the same structure as in the effective process of Sompolinsky and Zippelius [3]: a retarded self-interaction given by the response function and a Gaussian noise whose covariance is the correlation function.

(20). The noise covariance D is given by Eq. p. 9
Where the paper says it

(15) and (16) give ΓMH = G and Σϕ = C. The effective field then has the same structure as in the effective process of Sompolinsky and Zippelius [3]: a retarded self-interaction given by the response function and a Gaussian noise whose covariance is the correlation function.

p. 9
Where the paper says it

(15) and (16) give ΓMH = G and Σϕ = C. The effective field then has the same structure as in the effective process of Sompolinsky and Zippelius [3]: a retarded self-interaction given by the response function and a Gaussian noise whose covariance is the correlation function.

(21).

Appendix C

The saddle point for SyncMC yields the conjugate order parameters Qˆ = 0 −iRA(G⊤) −iRA(G) −iL⊤ − β−1RA(G⊤). The noise covariance N is given by Eq. p. 9

Where the paper says it

(15) and (16) give ΓMH = G and Σϕ = C. The effective field then has the same structure as in the effective process of Sompolinsky and Zippelius [3]: a retarded self-interaction given by the response function and a Gaussian noise whose covariance is the correlation function.

(21). The effective process is given by Eq. (20). The noise covariance D is given by Eq. (C9).

Appendix D

The Fourier representation of the common noise is derived in Eq. (D3). The discrete-time FDT of SyncMC is given by Eq. (25). The continuous-time form for the noise covariance is given by Eq. p. 9

Where the paper says it

(15) and (16) give ΓMH = G and Σϕ = C. The effective field then has the same structure as in the effective process of Sompolinsky and Zippelius [3]: a retarded self-interaction given by the response function and a Gaussian noise whose covariance is the correlation function.

p. 9
Where the paper says it

(15) and (16) give ΓMH = G and Σϕ = C. The effective field then has the same structure as in the effective process of Sompolinsky and Zippelius [3]: a retarded self-interaction given by the response function and a Gaussian noise whose covariance is the correlation function.

(D5). The derivation of Eq. (26) is shown in Eq. (D7). The discrete-time FDT of SyncMC is given by Eq. (25). The continuous-time form for the noise covariance is given by Eq. (D5).

Appendix E

The breakdown of the fluctuation–dissipation theorem is derived from the effective process of asynchronous MH, Eqs. (15) and (16). The equilibrium solution is given by Eq. p. 2

Where the paper says it

679 for the ROM at T = 0. 1, and the Parisi solution for the SK model at T = 0.

(E3). The stability condition is λ > 1 stated in Sec. p. 1
Where the paper says it

For Ising spins, the DMFT of single-spin-flip dynamics of the Sherrington–Kirkpatrick (SK) model [6] was formulated by Sommers [7], and Eissfeller and Opper [8] solved the effective single-spin process of the deterministic parallel dynamics Si(t+ 1) = sgn hi(t) (zero temperature, no Boltzmann sampling) by Monte Carlo sampling of single-spin trajectories. They noted that it was not clear whether a similar single-spin equation can be derived for random sequential updates; for the Metropolis–Hastings rule, Sec.

VII.

Appendix F

The Langevin equations for the slow components of the local field are described by Eq. (F8). The stationary distribution satisfies detailed balance and has the temperature vk/wk = T/xk. The plateau values follow from qk = E(∂yΦk(yk))2, where the average is over yk distributed according to the static field z−1. p. 7

Where the paper says it

III A, and σk, uk, ˆuk, θk are the values at these points. The single-site path weight w[σ|u], the probability of the spin trajectory given the history of the local field, is where the two dynamics differ.

The analysis applies to any rotationally invariant ensemble.

The noise and retarded self-interaction of band k have the same time dependence and satisfy the FDT at the effective temperature T/xk. The overall sizes follow from wk ≡ Z ∞ 0 Γk(τ) dτ = R(χk) − R(χk+1). The plateau values follow from qk = E(∂yΦk(yk))2, where the average is over yk distributed according to the static field z−1 and the reweighted distributions Pj of the slower bands j < k. p. 7

Where the paper says it

III A, and σk, uk, ˆuk, θk are the values at these points. The single-site path weight w[σ|u], the probability of the spin trajectory given the history of the local field, is where the two dynamics differ.

The analysis applies to any rotationally invariant ensemble. The data and code are available at Zenodo [15].

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. Improved AI can perform accelerated spin glass simulations by using synchronous Monte Carlo dynamics (SyncMC) for fast wall-clock time calculations, achieving a factor of three to four orders of magnitude speedup over asynchronous Metropolis-Hastings dynamics in the paramagnetic phase.

  2. Improved AI can accurately predict slow relaxation and aging behavior in spin glasses like the Sherrington-Kirkpatrick (SK) model by using either SyncMC or asynchronous MH dynamics, as both approach the same slow time scale equations, allowing for study of slow relaxation and aging of these models.

  3. Improved AI can determine the fluctuation dissipation ratio and the Parisi solution for mean-field spin glasses in the aging regime by utilizing either dynamics, as the slow-time-scale equations coincide despite differences in update rules.

  4. Improved AI can develop more efficient numerical solvers for effective single-spin processes by employing matrix function iterations, such as Newton's method described in Appendix C, to compute terms like R(G) up to high orders needed for accurate results.

  5. Improved AI can estimate response functions G and correlation functions C from finite-N simulations using score function estimators, which are unbiased and require no matrix inversion for the SyncMC case.

Abstract

We compare two Markov chain Monte Carlo dynamics for Ising spin glasses with rotationally invariant couplings: the synchronous Monte Carlo method (SyncMC), which updates all spins in parallel with the help of Gaussian auxiliary fields, and asynchronous Metropolis-Hastings dynamics, which updates one spin at a time. Using the Guionnet-Maïda formula, we derive the dynamical mean-field theory of both dynamics in terms of the R-transform of the coupling spectrum. The two effective single-spin processes share a retarded self-interaction and a Gaussian noise; SyncMC has, in addition, an instantaneous self-coupling and a noise that depends only on the response function. These additional terms act only on the fastest time scale, so that the equations for slow time scales, including the fluctuation-dissipation ratio and, for the Sherrington-Kirkpatrick model, the Parisi equation, are the same for the two dynamics. Simulations of the random orthogonal and Sherrington-Kirkpatrick models confirm that both dynamics approach the same fluctuation-dissipation relations, differing only in how many updates they need to do so.

Sources

Related papers