acceptodds
Under review as a conference paper at ICLR 2027

Retrieval Metastability and Centroid Hopping in Continuous Log-Sum-ReLU Dense Associative Memory

Abstract

We study the retrieval state of Dense Associative Memory with the log-sum-ReLU (LSR) energy on the -sphere as a function of the load and temperature . The compact support of the LSR kernel partitions the retrieval region into three regimes. At low , even the nearest spurious pattern is separated from the target by a free-energy barrier above , and retrieval is stable on any accessible timescale. Above a threshold , the state thermally wandering near the retrieval minimum already fills a spurious pattern basin; the two basins merge into a centroid and retrieval is lost. Between these limits lies the metastable regime: spurious pattern basins exist nearby but are separated from the retrieval basin by an entropic barrier. To escape, the state must drift beyond its typical thermal range and reach a spurious basin, a rare event whose rate we identify as a multi-channel Kramers process. After escape, the system hops between two-pattern centroids. Monte Carlo simulations validate this mechanism. The disorder-averaged escape rate, with a single calibrated prefactor, reproduces the measured escape probabilities across the simulated range of load and temperature. At large the exponential growth of the number of escape channels is exactly compensated by their entropic barriers, so the escape rate falls as at every temperature, and the metastable regime extends to . Simulations at fixed load and increasing system size show this suppression. The apparent phase boundary observed in finite- simulations lies inside the metastable regime: it is a kinetic crossover, reached where the escape time equals the simulation time, not a thermodynamic transition. Since the escape time grows exponentially with , the metastable state becomes indistinguishable from thermodynamic stability at system sizes of practical interest. The kinetic boundary is relevant only for the moderate accessible to Monte Carlo simulations.

Primary area:
learning theory
Keywords:
dense associative memory, modern Hopfield networks, storage capacity, metastability, Kramers escape, entropic barrier, energy landscape, Monte Carlo simulation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_eX2KOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies retrieval metastability in continuous log-sum-ReLU dense associative memory with random spherical patterns. The authors propose that retrieval loss occurs through an entropic barrier into a two-pattern centroid state, followed by further centroid hopping. Using the exact spherical overlap distribution, they construct a multi-channel Kramers model. Within this model, the disorder-averaged escape rate is proportional to exp⁡[−N(αth−α)]\exp[-N(\alpha_{\mathrm{th}}-\alpha)], up to subexponential factors, where αth\alpha_{\mathrm{th}} is the large-system threshold for another pattern to lie within the target pattern’s kernel support. Monte Carlo simulations examine escape probabilities and finite-size effects. The authors interpret the apparent retrieval boundary as a finite-time kinetic crossover rather than a thermodynamic transition.

Critical strengths

S1. A useful distinction between retrieval stability and finite-time retention.The paper addresses a discrepancy between a theoretical retrieval threshold and finite-size simulations. The interpretation based on the lifetime of a metastable retrieval state is physically meaningful. It clarifies what a finite observation window measures. This insight remains useful even if the kinetic model is approximate.

S2. A geometric analysis that produces specific predictions. The barrier calculation and the two-pattern equilibrium analysis connect geometry with observable alignments. The saddle-point analysis balances the number of escape channels against their barriers. This gives a scaling prediction for the disorder-averaged escape rate within the proposed model. It goes beyond a qualitative explanation of metastability.

S3. Targeted numerical investigations. The authors compare all-pattern and nearest-pattern escape detection. They also test the sensitivity of the escape criterion and examine finite-size effects. A single calibrated prefactor is used across different conditions. Although the agreement is limited, these experiments provide useful support for parts of the proposed mechanism.

Critical weaknesses

Overall, I find the physical interpretation interesting and the geometric analysis useful. However, the current evidence does not sufficiently support the central dynamical claims.

W1. The dynamical exponent needs further justification. The free-energy calculation in Eq. (9) does not establish an exact one-dimensional reduction of the Metropolis first-passage problem. The escape-rate model in Eq. (18) requires additional dynamical assumptions. The union bound supports escape suppression if the per-channel estimates are valid. It does not establish the predicted exponent for the disorder-averaged escape rate of the full dynamics.

W2. The predominance of two-pattern states needs more direct evidence. Equation (10) describes equilibrium with only two active patterns. Agreement with the observed alignments does not establish that such states dominate the dynamics. Direct statistics of active-pattern counts and residence times would strengthen the centroid-hopping explanation.

W3. Quantitative agreement is limited. The model uses measured diffusion coefficients and one calibrated prefactor. Agreement is mainly obtained at low load, while escape is overpredicted at higher load. At α=0.20\alpha=0.20 and T=0.40T=0.40, the escaped fraction decreases approximately as e−0.07Ne^{-0.07N}, versus e−0.12Ne^{-0.12N} in the model over the tested sizes. The test at α=0.28\alpha=0.28 covers only three sizes and mainly supports slower escape as NN increases. These results do not rule out the asymptotic prediction, but they do not yet validate its exponent.

Questions for authors

Q1. What assumptions connect the support-entry barrier to the rate of confirmed escape? Please distinguish results established for the Metropolis dynamics from predictions of the rate model. What analytical or numerical evidence supports treating returns to retrieval and inter-channel correlations as prefactor corrections?

Q2. What numerical evidence supports the predominance of two-pattern states after escape? Active-pattern counts at a few representative settings would help assess how often states with three or more active patterns occur.

Q3. How well does the predicted survival curve match the existing escape-time data? A comparison at a few representative settings, with uncertainty estimates, would be useful.

3 comments
AuthorsOpenReview ↗

Response to Reviewer eX2K

We thank the reviewer for a careful and technically precise review. We agree with W1 to W3, and the revision addresses each of them.

W1. We agree that Eq. (9) does not reduce the Metropolis first-passage problem exactly to one dimension, and that the union bound with the Kramers rate of each channel, Eq. (18), rests on this reduction. The revision replaces the Kramers rate by the probability that one Metropolis step, taken in the Boltzmann equilibrium of the retrieval basin, brings a given spurious pattern into support. For a state started in that equilibrium, the escape probability within tt steps is at most tt times the sum of these probabilities (Levin and Peres 2017, proof of Thm. 7.4), whatever the correlations between channels. Each probability follows from the equilibrium distribution and the Metropolis proposal, with no fitted constant, and equals e−ΔF/Te^{-\Delta F/T} of Eq. (9) up to a factor polynomial in NN. The disorder average of the sum is exact by linearity and has the exponent of Eq. (14). This exponent is therefore an upper bound for the Metropolis dynamics itself (App. B). The rate of Eq. (13) adds that escape reaches this exponent, with the prefactor AA calibrated on the simulations.

W2 and Q2. App. C now gives these statistics for 1000 trials at each of (α,T)=(0.20,0.40)(\alpha, T) = (0.20, 0.40), (0.22,0.325)(0.22, 0.325), (0.22,0.40)(0.22, 0.40) and (0.24,0.30)(0.24, 0.30). After the first escape, three patterns are in support during 0.2 to 4.1% of the time and four or more during at most 0.03%. Two patterns are in support during 33 to 75% of the time, and one pattern, the target or a spurious pattern, during the rest. In 99.4 to 100% of the hops the new two-pattern state shares one pattern with the previous one, so the transitions are pair-to-pair. The median residence in a two-pattern state is 2750 to 18400 steps, and the alignment there matches Eq. (10) with median deviation 0.002.

W3. The over-prediction at high load came mostly from the error in the confirmation test described in the general comment. The revision counts escape at the barrier crossing at all points of Fig. 3, with new runs, which is the event that the rate of Eq. (13) describes (see Q1). The measured diffusion coefficient sets only the time unit of the rate, and AA is calibrated at one point and then kept fixed for Fig. 3, Table 1 and App. C. At α≥0.26\alpha \geq 0.26 measured and predicted probabilities now differ by at most 0.03. Over the 25 points of Fig. 3 with predicted probability between 0.01 and 0.99, the median deviation is 0.016 and the largest 0.087. Table 1 also comes from new runs with the same count, read every 64 steps. The submitted table counted a trial as escaped when its target alignment at the end of the run was below 0.86, which also counts thermal dips below this level, and these dips flattened the measured decay at large NN. Now the measured escaped fraction is 0.9 to 1.1 times the predicted one at all twelve entries, and 0.94 to 1.26 times with AA calibrated at the same 64-step interval. Its decay with NN is e−0.107Ne^{-0.107N} against e−0.111Ne^{-0.111N} predicted at T=0.40T = 0.40, and e−0.114Ne^{-0.114N} against e−0.119Ne^{-0.119N} at T=0.25T = 0.25. We agree that α=0.28\alpha = 0.28 shows the direction of the NN dependence rather than the exponent. Escape there takes at most one relaxation time, so it is not a rare event. The runs now cover six sizes, N=35N = 35 to 56, and the median time to the barrier crossing grows from 560 to 3420 steps (the submitted times were the first drop of the target alignment below 0.5). The decay with NN predicted at these sizes is tested at α=0.20\alpha = 0.20.

Q1. Escape is now counted at the barrier crossing ϕμ≥φc\phi_\mu \geq \varphi_c itself, so the count needs no assumption that connects the crossing to a confirmed escape. At α=0.20\alpha = 0.20, counting escape instead when the spurious alignment first reaches the target alignment lowers the escaped fraction by at most 0.031 at every entry of Table 1 and changes its slope in NN by less than 0.002. Returns therefore change the prefactor and not the decay with NN. Correlations between channels do not enter the bound of App. B, and the recounted data agree with the model without a correction for them. Independent of the rate of Eq. (13) are the threshold αth\alpha_{\rm th} (exact density), the barrier of Eq. (9) (Boltzmann equilibrium) and the upper bound of App. B (Metropolis dynamics). The rate itself, its prefactor AA and the kinetic boundary come from the rate of Eq. (13) and are tested in Fig. 3, Table 1 and App. C.

Q3. App. C compares predicted and measured survival curves at eight settings, with bootstrap bands for the data and the calibration band of AA for the model. Over the whole window the two curves differ by at most 0.05 at seven settings and by at most 0.10 at the eighth, (α,T)=(0.26,0.25)(\alpha, T) = (0.26, 0.25).

Thank you for the substantial revisions and for correcting the simulation issues. The added active-pattern statistics support the centroid-hopping mechanism. The revised finite-size tests and survival curves also substantially address my concerns about the quantitative evidence.

The new escape-probability bound strengthens the theoretical contribution, although it does not establish that the actual dynamics attains the predicted exponent. Given the revised numerical evidence, I consider it reasonable to assume that the remaining dynamical corrections are subexponential in NN. In this context, I do not regard the absence of a rigorous proof of this assumption or an analytical determination of AA as a major concern. Please state this assumption explicitly when applying Eq. (14) to the Metropolis dynamics at fixed load and temperature, while retaining the distinction between the bound and the rate-model prediction.

Overall, the geometric analysis and improved numerical validation provide a useful account of retrieval metastability. My main concerns have been sufficiently addressed, and I am raising my rating from 2 to 3.

Thank you for the reassessment and for raising the rating.

We will state this assumption in Section 3, after the bound: "Applying (14) to the Metropolis dynamics at fixed α\alpha and TT assumes, beyond this bound, that the escape rate of the dynamics differs from the rate (13) by a factor subexponential in NN. The simulations support this assumption."

Reviewer_UGeGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper is primarily about a theoretical investigation into why simulations of log-sum-ReLU (LSR) dense associative memory lose retrieval below the threshold predicted using the exact distribution of random spherical patterns. The authors propose that these failures reflect finite-time escape from a metastable retrieval basin into two-pattern centroid states, rather than the disappearance of retrieval itself. Their geometric analysis and Monte Carlo experiments relate escape to storage load, temperature, and dimension, predicting exponentially increasing retrieval lifetimes with dimension. The main contribution is an explanation of retrieval dynamics within LSR, rather than a new retrieval method or a demonstrated improvement in practical storage capacity.

Critical strengths

Dense associative memories while interesting in theory have been hard to put into practice for storing large scale data. The main problem seems to be metastability which is usually explained as inaccurate recovery due to the mixture of two patterns. This happens particularly, when the patterns are not as separable, particularly in modern Hopfield networks. This paper sees metastability from the time standpoint, i.e. for how long does the retrieval recall the same pattern before flipping over of migrating to another basin. This viewpoint could be useful for people in the machine learning community who are trying to retrieve with these memories.

The paper addresses a concrete discrepancy between theory and simulation and offers a physically motivated mechanism supported by targeted experiments. Its distinction between asymptotic capacity and finite-time retrieval reliability is relevant to evaluating associative memories: successful storage should depend on how long retrieval must remain reliable under noise. The geometric analysis, escape diagnostics, and finite-size experiments provide complementary evidence for this perspective.

Critical weaknesses

My main concerns is the significance of this result from a practical consideration. The broader significance for machine learning remains questionable, particularly in the light of other retrieval framework such as Modern Hopfield Networks. The study focuses on one energy function, random spherical patterns, and thermal dynamics, without establishing whether the proposed mechanism predicts failures for realistic image representations or commonly used deterministic retrieval procedures. Matched comparisons with log-sum-exp memories and controlled changes in pattern separability would clarify which findings are specific to LSR and which generalize across associative memories. The finite-size results support increasing retrieval lifetime more clearly than the precise claimed asymptotic exponent. While the focused analysis is potentially valuable, a stronger connection to practical storage decisions and evaluation would make its contribution to an ICLR audience more compelling.

Questions for authors

Prior work, including Hopfield Encoding Networks presented at UniReps 2024, links mixed retrieval states to poor pattern separability and shows that better-separated representations can reduce their occurrence. How does this perspective relate to the finite-temperature metastability studied here? In particular, how would changing pattern separability at fixed load, dimension, and temperature affect the escape barriers and retrieval lifetimes? A controlled comparison would help clarify whether the proposed mechanism explains an additional source of instability beyond pattern crowding, or quantifies its dynamical consequences.

1 comment
AuthorsOpenReview ↗

Response to Reviewer UGeG

We thank the reviewer for the positive assessment of the mechanism and for the questions about its scope.

We see the contribution as the reviewer summarizes it, an explanation of retrieval dynamics rather than a new retrieval method. We do not fully share the concern about its significance. The LSR memory studied here is itself a Modern Hopfield Network, the name now used for dense associative memories with the LSE, LSR and other kernels. As the reviewer notes, the distinction between asymptotic capacity and finite-time retrieval reliability matters when associative memories are evaluated. Here it shows that the boundary near 0.25 seen in finite-NN simulations is a kinetic crossover, which moves toward 0.347 as NN grows. For a given NN, noise level and required retention time, the escape rate of Eq. (13) gives the largest load at which retrieval survives, and Fig. 3 (right) shows this boundary for the simulation time.

Deterministic retrieval is the limit T→0T \to 0, where no escape occurs below the threshold of Eq. (3). In the escape rate, the stored patterns enter only through the distribution of their alignments with the target (App. C). The runs with controlled separability described below change this distribution, and the model follows them with no new parameter.

We have carried out the same analysis for the LSE energy on the NN-sphere, with the exact alignment density and Monte Carlo up to large NN. That study is currently under review elsewhere, so we summarize its results here without a reference. With the exact density, the retrieval boundary at β=1\beta = 1 and T=0T = 0 is α=0.623\alpha = 0.623, the threshold of Lucibello and Mézard (2024), and extends to T>0T > 0. The boundary seen in finite-NN Monte Carlo is kinetic, as in LSR. It lies below the N→∞N \to \infty boundary and approaches it as NN grows. In LSE every pattern contributes to the energy. The contribution of the spurious patterns, as a distribution over their alignment with the target, has its maximum either at the edge, on the best-aligned patterns, or at an interior saddle alignment, on exponentially many patterns. The switch is at a load α∗\alpha^{\ast} that depends on β\beta, α∗=0.241\alpha^{\ast} = 0.241 at β=1\beta = 1. Retrieval in LSE is therefore lost to one of the best-aligned patterns below α∗\alpha^{\ast} and to a collective state of many patterns above it. In LSR it is lost to a two-pattern centroid, which does not exist in LSE. The compact support of the LSR kernel keeps the energy constant along the escape path up to the support of the spurious pattern and lowers it beyond, toward the centroid. The barrier is purely entropic, and the large-NN exponent of the escape rate is the same at every temperature. In LSE every escape route passes an energy barrier, so escape slows exponentially as TT decreases and stops at T=0T = 0.

The revision gives the exponent as an upper bound for the Metropolis dynamics itself (response to eX2K, W1), and the decay with NN at α=0.20\alpha = 0.20 now follows the predicted one, e−0.107Ne^{-0.107N} against e−0.111Ne^{-0.111N} at T=0.40T = 0.40.

We ran the controlled comparison at fixed (α,N,T)=(0.24,41,0.40)(\alpha, N, T) = (0.24, 41, 0.40). The spurious patterns were redrawn so that their alignment with the target does not exceed 0.56, 0.52, 0.48 or 0.45, and compared with unrestricted patterns. The measured escape probability is 0.933 (unrestricted), 0.812, 0.410, 0.060 and 0.012, against 0.950, 0.847, 0.440, 0.087 and 0.016 predicted with the same AA. Better separation raises the barrier ΔF(φ1μ)\Delta F(\varphi_{1\mu}) of the nearest patterns and lengthens the retrieval lifetime. The mechanism therefore quantifies the dynamical consequences of pattern crowding, the effect discussed for Hopfield Encoding Networks (Kashyap et al. 2024), rather than adding a separate source of instability (App. C).

We also revised the presentation to make the paper easier to follow. The introduction now states the theory and its results before the description of the simulations. Sections 2 and 3 now have headings that name the steps of the argument. Section 4 states which results hold for the Metropolis dynamics and which come from the rate of Eq. (13). Repeated statements were removed.

Reviewer_h2cmOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

Considering dense associative memories, this paper builds on the work of Petrova (2026) and considers a thermodynamic view of retrieval dynamics. The core idea is that when there is too much noise, retrieval basins merge into a centroid and retrieval is lost. But there is also a metastable regime where spurious pattern basins exist near a retrieval basin but are separated by an entropic barrier. For escape, a multi-channel Kramers process is identified, after which the system hops between two-pattern centroids. This is developed theoretically as well as through simulations.

Critical strengths

Studying the retrieval dynamics of dense associative memories is an important subject in its own right, going back to Hopfield and many others, but also for its potential insight into Transformer and diffusion models that have strong mathematical relations to associative memories.

Taking a thermodynamics approach is compelling, and the exponential escape rate result is clean and compelling, as the phase diagram.

The joint mathematical and simulation development appears to be correct.

Critical weaknesses

The authors initially discuss associative memories with exponential capacity. They may be interested in earlier work in this direction too that drew on information-theoretic/coding-theoretic ideas, e.g. [A. Karbasi, A. H. Salavati, A. Shokrollahi, and L. R. Varshney, “Noise-Enhanced Associative Memories,” NeurIPS 2013.], to get to exponential capacity. Moreover, there are several other related works in the literature of associative memories that draw on related physics methods, which would be helpful to discuss.

The rhetorical structure is hard to follow, especially in the intro: going from background work into Monte Carlo simulations in a lot of detail and then only very briefly saying that a theory is developed. The theory, as the main contribution, should be centered in the introduction.

More generally, the argumentation structure is hard to follow, and the relationship between the mathematical and the simulation (and how they relate to one another) would be helpful to have better explication of.

Questions for authors

Please do clarify the specific relationships between the simulations and the mathematical development.

2 comments
AuthorsOpenReview ↗

Response to Reviewer h2cm

We thank the reviewer for the positive assessment and for the comments on the presentation.

We added Karbasi et al. (2013) to the introduction as a separate route to exponential capacity that uses no energy function, together with the LSE capacity obtained by Lucibello and Mézard (2024) with the methods of statistical physics applied to pattern disorder.

We understand that describing simulations in this detail before the theory is unusual. We kept this order because the simulations show the problem directly. They place the retrieval boundary near α=0.25\alpha = 0.25 rather than at the exact threshold 0.3470.347, and as the load increases, more trials settle at target alignments of about 0.83 and 0.48 (Fig. 1). The theory explains both. Right after the discrepancy, before the description of the simulations, a new paragraph now states the theory and its results and points to where each is tested. Sections 2 and 3 now have headings that name the steps of the argument.

The theory and the simulations relate as follows. The threshold αth\alpha_{\rm th}, Eq. (3), is exact for N→∞N \to \infty and follows from the exact density of the alignments. The entropic barrier, Eq. (9), and the centroid alignment, Eq. (10), are derived from the Boltzmann equilibrium (Section 2, App. A). For the Metropolis dynamics started in the equilibrium of the retrieval basin, the exponent of the escape rate is an upper bound on the disorder-averaged escape probability (App. B). The rate of Eq. (13) adds that escape reaches this exponent, with one calibrated prefactor. The simulations test it through the escape probabilities (Fig. 3, left), their dependence on NN (Table 1) and the survival curves (App. C). They also test the centroid alignment (Fig. 1) and the pair-to-pair hops (App. C). The green contour of Fig. 3 (right), drawn over the alignment measured in the simulations, marks where the disorder-averaged escape time equals the simulation time. Typical samples escape more slowly than this average, so most trials leave retrieval only near 0.25. Section 4 states which results hold for the Metropolis dynamics and which come from the rate of Eq. (13).

Reviewer_h2cmOpenReview ↗

thanks

Thank you to the authors for the further explanation on how the simulations and theory are related. The broader embedding in the exponential-capacity associative memory literature will also be helpful. I also appreciate the responses to the other reviewers.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision

We thank the reviewers for their careful reading and their constructive comments.

The paper makes two contributions. First, it identifies how retrieval is lost at finite temperature in the LSR memory, a Modern Hopfield Network with the log-sum-ReLU energy. The state escapes over an entropic barrier into the support of a spurious pattern, settles into a two-pattern centroid, and then hops between centroids. Second, it corrects the retrieval threshold of Petrova et al. (2026) from α=0.25\alpha = 0.25, obtained with the Gaussian approximation of the inter-pattern alignments, to 0.3470.347, obtained with their exact density. It shows that the boundary near 0.250.25 seen in earlier Monte Carlo simulations is a kinetic crossover, which moves toward 0.3470.347 as NN grows. The questions of the reviews concern the evidence for the escape rate and its large-NN exponent (eX2K, UGeG), the scope of the results (UGeG), and the presentation (UGeG, h2cm).

The revision adds this evidence. New text is in bold, except for editorial changes. The large-NN exponent is now an upper bound on the disorder-averaged escape probability of the Metropolis dynamics itself. It follows from the Boltzmann equilibrium of the retrieval basin and the Metropolis rule, without the Kramers rate (App. B).

While preparing the revision we found two errors in the escape detectors of the submitted Fig. 3. The detector used for the points with α≥0.24\alpha \geq 0.24 confirmed each barrier crossing by comparing the diffusion toward the spurious pattern after the crossing with the diffusion of the target alignment before it. Because one variable stored both coordinates, the test included a false jump from one to the other at the crossing and rejected most escapes that occurred early in a run. At high load most escapes are early, which caused the over-prediction at α≥0.26\alpha \geq 0.26 in the submitted figure. In addition, both detectors read the crossing from the proposed Monte Carlo move, accepted or not, which raised the escaped fraction at T≥0.6T \geq 0.6 by up to 0.12. We now count escape at the barrier crossing ϕμ≥φc\phi_\mu \geq \varphi_c of Section 2 itself, on the accepted state, the event that the rate of Eq. (13) describes, with no confirmation step, and we reran all points of Fig. 3 and Table 1 with this count. At T≥0.6T \geq 0.6 the new count reads the alignments every 16 steps, against every 4 steps in the submitted runs, which lowers the escaped fraction there by up to 0.06. The submitted Table 1 counted a trial as escaped when its target alignment at the end of the run was below 0.86, and the submitted escape times at α=0.28\alpha = 0.28 were the first drop of the target alignment below 0.5. Both are now counted at the barrier crossing. The model is unchanged, and its single constant is calibrated at the same point as before. The rerun of that point gives A=40A = 40, against 33 in the submission. All points of Fig. 3 now lie within 0.09 of the model, and the decay with NN in Table 1 follows the predicted one. The code and data of the new runs are in rev1/ of the supplementary material.

We removed the submitted comparison of escape times with all-pattern and nearest-pattern detection, because its all-pattern times came from a drop of the target alignment, the criterion that Section 3 calls biased. In the new runs both counts come from the same trajectories, and the nearest pattern makes 31 to 96% of the first escapes, as the model predicts (Section 3, App. C). The test of the confirmation threshold was removed with the confirmation step.

A new appendix gives the number of patterns in support after escape, survival curves with uncertainties, and runs with controlled pattern separability (App. C).

To make the paper easier to follow, the introduction now states the theory and its results before the description of the simulations. Sections 2 and 3 now have headings that name the steps of the argument. Section 4 states which results hold for the Metropolis dynamics and which come from the rate of Eq. (13). Repeated statements were removed.

The main conclusions of the paper are unchanged.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.