Searcharxiv⌕ Search

arXiv · 2611.14727

Susceptibility Bottlenecks in Equilibrium Propagation Using Oscillator Ising Machines

Abstract

Equilibrium propagation (EP) enables learning in physical systems by exploiting the displacement between free and weakly perturbed equilibrium states. Here, we investigate how the underlying oscillator dynamics govern the transmission of learning signals in oscillator Ising machines (OIMs). We show that the hidden-layer response can be decomposed into two distinct factors: the nudging force applied at the output and the susceptibility governing its transmission to the hidden layer. Using a bipartite OIM with interactions interpolating between phase-difference and phase-sum coupling, we investigate these factors while ensuring that all coupling forms realize the same Ising Hamiltonian at binary phase-locked states. We demonstrate that second-harmonic injection primarily influences the nudging force through phase binarization, whereas the coupling form controls the cross-layer susceptibility through the mixed curvature of the energy landscape. At equal mixing, the mixed curvature vanishes on the binary manifold, strongly suppressing hidden-layer learning signals. Numerical simulations validate these mechanisms and reveal their consequences for classification performance. Our findings establish that realizing a desired energy landscape alone is insufficient to ensure effective physical learning; the system must also support the transmission of perturbations required for learning.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Malihe Farasat, Oshada Rathnayake, Nikhil Shukla. 2026-11-10. Susceptibility Bottlenecks in Equilibrium Propagation Using Oscillator Ising Machines. https://searcharxiv.com/abs/2611.14727

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Graph neural networks and the energetic cavity method for combinatorial optimization

We study the use of graph neural networks (GNNs) for finding approximate ground states of Ising models. Efficiently finding these ground states is of broad significance because many combinatorial optimization problems can be formulated as an Ising model with the appropriate choice of couplings and fields. Exactly solving these problems is hard but there are many good heuristic methods. A lineage of these heuristics build from mean-field approximations: one approach uses the leading eigenvector of an appropriately defined matrix, another is the min-sum algorithm, also known as the energetic cavity method. Without modification, GNNs perform worse than both of these methods. We consider small modifications to the GNN to incorporate these heuristics and find that this considerably improves performance. While the modified approach is much stronger than other deep learning approaches, we still find that simulated annealing is generally as good for the same computational cost.

cond-mat.dis-nn↗

Contact-number hyperuniformity at glass close packing in infinite dimensions

We investigate contact-number fluctuations of harmonic spheres near glass close packing (GCP), the amorphous jamming endpoint reached by compressing an ideal glass. Using replica theory, we show that the Gardner-unstable region shrinks toward jamming as the spatial dimension increases, allowing GCP to be approached along a stable one-step replica symmetry breaking branch in the high-dimensional limit. By coupling an external field to the local contact number, we derive the variance of the total contact number and the wave-number-dependent structure factor directly from the microscopic interactions. At GCP, the structure factor vanishes quadratically at small scaled wave numbers, demonstrating contact-number hyperuniformity. Upon weak compression above GCP, the structure factor develops a low-wave-number plateau whose height is proportional to the excess coordination. These results provide an analytical description of contact-number hyperuniformity and its breakdown upon compression.

cond-mat.dis-nn↗

Stochastic Dynamics of Domain Walls on a Racetrack: Impact of Inhomogeneity of Magnetic Anisotropy

Racetrack memory is a spintronic technology that stores information in current-driven magnetic domain walls (DW), offering a promising route toward non-volatile, high-density, and high-speed memory devices. In ultrathin magnetic films, atomic-scale interface disorder can produce spatially inhomogeneous magnetic anisotropy, creating a complex spatially varying energy landscape that strongly influences DW motion. In this work, we model anisotropy inhomogeneities using a two-dimensional fractal Weierstrass--Mandelbrot function, which provides independent control over the disorder amplitude and characteristic spatial scale. Micromagnetic simulations reveal pronounced stochasticity of current-driven domain-wall transport even in the absence of thermal fluctuations. Transport is quantified by the transmission fraction, defined as the fraction of DWs that successfully reach a detector after propagating through a disordered racetrack. We investigate its dependence on current density, disorder amplitude, and the spatial scale of the anisotropy inhomogeneity. The resulting transmission statistics are described by Gaussian error-function profiles and exhibit a universal scaling behavior upon normalization by the corresponding median values. Statistical analysis of DW trajectories further characterizes the disorder-induced fluctuations of the propagation process. The variance of the random DW positions shows a pronounced dependence on disorder amplitude and is well described by the proposed phenomenological model. These results provide a quantitative framework for describing stochastic DW transport in structurally disordered racetracks and establish disorder-induced reliability limits in racetrack-memory devices.

cond-mat.dis-nn↗