Computation beyond energy descent: a theory for asymmetric, oscillatory, and chaotic recurrent neural dynamics, and principled escape from poor attractors in analog optimization networks

The Hopfield-Tank framework (Science, 1986) established that graded-response recurrent neural circuits with effectively symmetric synapses descend a Lyapunov energy function and settle into attractors, turning network design into energy engineering (demonstrated by an analog-to-digital converter and a traveling-salesman network) and mapping directly onto analog electronic hardware. Its boundary is the descent property itself: strong synaptic asymmetry yields oscillations and chaos with no computational-level theory; solution quality stops at local minima with no principled escape mechanism or performance guarantee; rhythmic and sequenced computation is excluded. What survives as a precise, coupled open core is: (1) a theory, with design power comparable to the energy function, of when and how asymmetric, oscillatory, or chaotic recurrent dynamics perform reliable computation (including sequenced and rhythmic computation), and (2) principled escape from poor local minima for analog energy-descent networks on hard combinatorial problems, with theoretical performance guarantees compatible with physical hardware. Modern work (transient-chaos hardness of analog constraint satisfaction, trap-free asymmetric networks, nonequilibrium landscape theory, chaotic sampling dynamics, coherent Ising machines) attacks fragments of this core without closing it.

Background

In 1986 Hopfield and Tank proposed a framework for computation in model neural circuits (Hopfield and Tank 1986). Each neuron is a graded-response element: an input node with capacitance $C_i$ and resistance $R_i$ whose potential $u_i$ evolves continuously and produces a monotonic, bounded output $V_i=f_i(u_i)$, a sigmoid standing for a firing rate or an amplifier's transfer curve. The network dynamics are

$$C_i\frac{du_i}{dt}=\sum_{j=1}^{N} T_{ij}f_j(u_j)-\frac{u_i}{R_i}+I_i,$$

where $T_{ij}$ is the synaptic strength from neuron $j$ to neuron $i$ and $I_i$ is an external input. The central result is a Lyapunov (energy) function - a scalar quantity that decreases monotonically along trajectories - which exists when the synapses are effectively symmetric ($T_{ij}=T_{ji}$), the transfer functions are monotonic and bounded, and the inputs vary slowly. The state then flows downhill in the energy $E$ and converges to a local minimum (a nearby low-energy configuration), which precludes sustained oscillations and chaos. Computation thereby becomes energy engineering: choose a state-space representation of candidate solutions and construct $E$ so that valid configurations have energy equal to the cost to minimize. Hopfield and Tank demonstrated this with a 7-neuron circuit for four-bit analog-to-digital conversion and an $n^2$-neuron network for the traveling salesman problem (TSP - the NP-hard problem, for which no polynomial-time algorithm is known, of finding the shortest tour through $n$ cities), and they gave an explicit mapping of the equations onto electronic hardware (amplifiers and resistive synapses). The paper also delimited the framework. The descent property fails for strongly asymmetric synapses - which most biological circuits have - and networks with substantial disorganized asymmetry can oscillate or behave chaotically (a limit cycle is an isolated periodic trajectory; chaos is bounded aperiodic dynamics with sensitive dependence on initial conditions), so circuits that compute with rhythms, such as central pattern generators (the oscillatory neural circuits that produce rhythmic motor behaviors like locomotion) and oscillatory olfactory processing, fell outside the theory. Descent ends in a local rather than global minimum, so solution quality depends on the energy landscape and the starting state, and the paper contained no principled escape mechanism or performance guarantee.

Four decades of later work built on these pieces. Escape by randomness is classically understood in idealized settings: simulated annealing (Kirkpatrick, Gelatt and Vecchi 1983) explores a discrete search space with temperature-controlled random acceptance of uphill moves; the Boltzmann machine (Ackley, Hinton and Sejnowski 1985) is the stochastic analogue of a Hopfield network whose equilibrium distribution concentrates on low-energy states; Langevin dynamics - deterministic descent plus white noise - with logarithmically slow cooling provably converges in probability to global optima (Geman and Hwang 1986; Hajek 1988). These asymptotic guarantees require impractically slow schedules and give no guarantees for the deterministic analog circuit. The TSP mapping itself proved fragile under parameter changes (Wilson and Pawley 1988; see also the 1987 Science exchange involving Kohonen and Oja, Palm, and the response of Hopfield and Tank). Asymmetric dynamics were characterized statistically by mean-field theory, which replaces a large random network by self-consistent equations for a few averaged quantities and located the onset of chaos in random asymmetric networks (Sompolinsky, Crisanti and Sommers 1988) and in balanced excitatory-inhibitory networks (van Vreeswijk and Sompolinsky 1996); this describes typical dynamics but is not a design tool. Sequence storage exploits asymmetry (Sompolinsky and Kanter 1986); timing-based computation was developed by Tank and Hopfield (1987) and Hopfield and Herz (1995); the oscillatory olfactory bulb was modeled by Li and Hopfield (1989). Reservoir computing (Maass, Natschläger and Markram 2002; Bertschinger and Natschläger 2004) computes by feeding inputs into an untrained random - hence asymmetric - dynamical reservoir and training only a static readout, a paradigm that uses rather than designs or explains the internal dynamics. Hardware matured: the memristor (Strukov, Snider, Stewart and Williams 2008) is a compact two-terminal electrically programmable resistance, that is, a writable analog synapse; in-memory computing architectures (Sebastian, Le Gallo, Khaddam-Aljameh and Eleftheriou 2020) integrate dense crossbar arrays of such devices; and Hopfield optimization networks with manufacturable programmable synapses have been demonstrated (Yi, Kumar and Williams 2021, 2022). At large scale, analog optimization is now pursued as Ising machines - networks of coupled oscillators or spins that settle toward low-energy states of an Ising (pairwise binary-variable) energy - realized optically and electronically (Inagaki et al. 2016; reviewed by Mohseni, McMahon and Byrnes 2022). Reverse-engineering real neural populations into energy functions became the inverse Ising problem of fitting pairwise maximum-entropy models to recorded activity (Schneidman, Berry, Segev and Bialek 2006; Nguyen, Zecchina and Berg 2017). Two threads of the original limitation list remain open and have partially merged: Ercsey-Ravasz and Toroczkai (2011) showed that trapping of analog constraint-satisfaction networks corresponds to transient chaos - chaotic behavior that eventually terminates - whose duration grows rapidly with problem hardness; trap-free asymmetric network variants exist (Molnár and Ercsey-Ravasz 2013); chaotic simulated annealing via transiently chaotic networks (Chen and Aihara 1995) empirically escapes poor minima but without general guarantees. Coherent Ising machines are still trapped by local minima in a way that worsens with problem size (Kako, Leleu, Inui, Khoyratee, Reifenstein and Yamamoto 2020), local-minima effects and delayed instabilities limit their performance (Sato 2022), and only recently has a first convergence proof appeared, under restrictive assumptions, for one oscillator-based design (Pramanik, Chatterjee and Oza 2023). Nonequilibrium landscape theory (Yan, Zhao, Hu, Wang, Wang and Wang 2013) replaces the single energy by a landscape-flux decomposition for asymmetric networks, a candidate descriptive formalism rather than a design-level computational theory. Recent work states the residual gap in the original 1986 terms: computation in asymmetric networks beyond gradient-like flows still lacks a general framework (Rutishauser, Slotine and Douglas 2015; Terada and Toyoizumi 2024; Nartallo-Kaluarachchi, Lambiotte and Goriely 2026).

Problem Statement

Extend computation in graded-response recurrent neural circuits beyond symmetric energy descent, with rigor and design power comparable to the Hopfield-Tank energy-function framework. This is the surviving open core of the limitation list attached to Hopfield and Tank (1986), and it has two linked facets sharing this single objective:

(1) Dynamics facet. Develop a mathematical theory of computation in recurrent networks whose synaptic organization is asymmetric, oscillatory, or chaotic: conditions under which such dynamics perform well-defined, reliable computations - including sequenced and rhythmic computation - and design principles by which circuits can be engineered to compute through limit cycles, transient chaos, or attractor-to-attractor transitions, the regime of central pattern generators and oscillatory olfactory circuits, rather than only analyzed statistically. In the source's own words, nonsymmetric circuits can exhibit oscillatory trajectories that "are undoubtedly important in neural computations, but as yet we lack the mathematical tools to manipulate and understand them at a computational level" (Hopfield and Tank 1986).

(2) Solution-quality facet. Provide a principled account of solution quality for analog energy-descent computation, whose structural limitation in the source is that trajectories "reach a state in which $E$ is a local minimum, and stop changing with time", with no escape mechanism or performance guarantee present in the framework: mechanisms for escaping poor local minima that are compatible with the graded-response dynamics and with physical hardware, together with theoretical performance guarantees for such mechanisms on classes of hard combinatorial optimization problems.

The two facets are related: the mechanisms that escape poor attractors (noise, chaos, asymmetry) are precisely dynamics outside gradient descent in $E$, and the hardness of analog optimization has been identified with transient chaos (Ercsey-Ravasz and Toroczkai 2011). Each facet mirrors the energy framework by pairing an analysis requirement (conditions on reliable computation; mechanisms for escaping poor minima) with a matching design requirement (design principles for engineering circuits in that regime; theoretical performance guarantees for the mechanisms).

References

1. Hopfield J.J., Tank D.W. (1986). Computing with Neural Circuits: A Model. Science 233(4764):625-633. DOI 10.1126/science.3755256. https://doi.org/10.1126/science.3755256
2. Kirkpatrick S., Gelatt C.D., Vecchi M.P. (1983). Optimization by Simulated Annealing. Science. DOI 10.1126/science.220.4598.671. https://doi.org/10.1126/science.220.4598.671
3. Ackley D.H., Hinton G.E., Sejnowski T.J. (1985). A Learning Algorithm for Boltzmann Machines. Cognitive Science. DOI 10.1207/s15516709cog0901_7. https://doi.org/10.1207/s15516709cog0901_7
4. Geman S., Hwang C. (1986). Diffusions for Global Optimization. SIAM Journal on Control and Optimization. DOI 10.1137/0324060. https://doi.org/10.1137/0324060
5. Sompolinsky H., Kanter I. (1986). Temporal Association in Asymmetric Neural Networks. Physical Review Letters. DOI 10.1103/physrevlett.57.2861. https://doi.org/10.1103/physrevlett.57.2861
6. Tank D.W., Hopfield J.J. (1987). Neural computation by concentrating information in time. Proceedings of the National Academy of Sciences. DOI 10.1073/pnas.84.7.1896. https://doi.org/10.1073/pnas.84.7.1896
7. Kohonen T., Oja E. (1987). Computing with Neural Networks (Technical Comment). Science. DOI 10.1126/science.235.4793.1227-a. https://doi.org/10.1126/science.235.4793.1227-a
8. Palm G. (1987). Computing with Neural Networks (Technical Comment). Science. DOI 10.1126/science.235.4793.1227-b. https://doi.org/10.1126/science.235.4793.1227-b
9. Hopfield J.J., Tank D.W. (1987). Response: Computing with Neural Networks. Science. DOI 10.1126/science.235.4793.1228. https://doi.org/10.1126/science.235.4793.1228
10. Hajek B. (1988). Cooling Schedules for Optimal Annealing. Mathematics of Operations Research 13(2). DOI 10.1287/moor.13.2.311. https://doi.org/10.1287/moor.13.2.311
11. Sompolinsky H., Crisanti A., Sommers H.J. (1988). Chaos in Random Neural Networks. Physical Review Letters 61. DOI 10.1103/physrevlett.61.259. https://doi.org/10.1103/physrevlett.61.259
12. Wilson G.V., Pawley G.S. (1988). On the stability of the Travelling Salesman Problem algorithm of Hopfield and Tank. Biological Cybernetics. DOI 10.1007/bf00363956. https://doi.org/10.1007/bf00363956
13. Li Z., Hopfield J.J. (1989). Modeling the olfactory bulb and its neural oscillatory processings. Biological Cybernetics. DOI 10.1007/bf00200803. https://doi.org/10.1007/bf00200803
14. Chen L., Aihara K. (1995). Chaotic simulated annealing by a neural network model with transient chaos. Neural Networks. DOI 10.1016/0893-6080(95)00033-v. https://doi.org/10.1016/0893-6080(95)00033-v
15. Hopfield J.J., Herz A.V. (1995). Rapid local synchronization of action potentials: toward computation with coupled integrate-and-fire neurons. Proceedings of the National Academy of Sciences 92(15). DOI 10.1073/pnas.92.15.6655. https://doi.org/10.1073/pnas.92.15.6655
16. van Vreeswijk C., Sompolinsky H. (1996). Chaos in Neuronal Networks with Balanced Excitatory and Inhibitory Activity. Science 274. DOI 10.1126/science.274.5293.1724. https://doi.org/10.1126/science.274.5293.1724
17. Maass W., Natschläger T., Markram H. (2002). Real-Time Computing Without Stable States: A New Framework for Neural Computation Based on Perturbations. Neural Computation. DOI 10.1162/089976602760407955. https://doi.org/10.1162/089976602760407955
18. Bertschinger N., Natschläger T. (2004). Real-Time Computation at the Edge of Chaos in Recurrent Neural Networks. Neural Computation. DOI 10.1162/089976604323057443. https://doi.org/10.1162/089976604323057443
19. Schneidman E., Berry M.J., Segev R., Bialek W. (2006). Weak pairwise correlations imply strongly correlated network states in a neural population. Nature. DOI 10.1038/nature04701. https://doi.org/10.1038/nature04701
20. Strukov D.B., Snider G.S., Stewart D.R., Williams R.S. (2008). The missing memristor found. Nature. DOI 10.1038/nature06932. https://doi.org/10.1038/nature06932
21. Ercsey-Ravasz M., Toroczkai Z. (2011). Optimization hardness as transient chaos in an analog approach to constraint satisfaction. Nature Physics. DOI 10.1038/nphys2105. https://doi.org/10.1038/nphys2105
22. Yan H., Zhao L., Hu L., Wang X., Wang E., Wang J. (2013). Nonequilibrium landscape theory of neural networks. Proceedings of the National Academy of Sciences. DOI 10.1073/pnas.1310692110. https://doi.org/10.1073/pnas.1310692110
23. Molnár B., Ercsey-Ravasz M. (2013). Asymmetric Continuous-Time Neural Networks without Local Traps for Solving Constraint Satisfaction Problems. PLoS ONE. DOI 10.1371/journal.pone.0073400. https://doi.org/10.1371/journal.pone.0073400
24. Rutishauser U., Slotine J.-J., Douglas R. (2015). Computation in Dynamically Bounded Asymmetric Systems. PLOS Computational Biology. DOI 10.1371/journal.pcbi.1004039. https://doi.org/10.1371/journal.pcbi.1004039
25. Inagaki T., Haribara Y., Igarashi K., Sonobe T., Tamate S., Honjo T., Marandi A., McMahon P.L., Umeki T., Enbutsu K., Tadanaga O., Takenouchi H., Aihara K., Kawarabayashi K.-i., Inoue K., Utsunomiya S., Takesue H. (2016). A coherent Ising machine for 2000-node optimization problems. Science. DOI 10.1126/science.aah4243. https://doi.org/10.1126/science.aah4243
26. Nguyen H.C., Zecchina R., Berg J. (2017). Inverse statistical problems: from the inverse Ising problem to data science. Advances in Physics. DOI 10.1080/00018732.2017.1341604. https://doi.org/10.1080/00018732.2017.1341604
27. Kako S., Leleu T., Inui Y., Khoyratee F., Reifenstein S., Yamamoto Y. (2020). Coherent Ising Machines with Error Correction Feedback. Advanced Quantum Technologies. DOI 10.1002/qute.202000045. https://doi.org/10.1002/qute.202000045
28. Sebastian A., Le Gallo M., Khaddam-Aljameh R., Eleftheriou E. (2020). Memory devices and applications for in-memory computing. Nature Nanotechnology. DOI 10.1038/s41565-020-0655-z. https://doi.org/10.1038/s41565-020-0655-z
29. Yi S.-I., Kumar S., Williams R.S. (2021). Improved Hopfield Network Optimization Using Manufacturable Three-Terminal Electronic Synapses. IEEE Transactions on Circuits and Systems I. DOI 10.1109/tcsi.2021.3119648. https://doi.org/10.1109/tcsi.2021.3119648
30. Yi S.-I., Kumar S., Williams R.S. (2022). Combinatorial Optimization in Hopfield Networks with Noise and Diagonal Perturbations. IEEE International Symposium on Circuits and Systems. DOI 10.1109/iscas48785.2022.9937983. https://doi.org/10.1109/iscas48785.2022.9937983
31. Mohseni N., McMahon P.L., Byrnes T. (2022). Ising machines as hardware solvers of combinatorial optimization problems. Nature Reviews Physics. DOI 10.1038/s42254-022-00440-8. https://doi.org/10.1038/s42254-022-00440-8
32. Sato S. (2022). Effects of local minima and bifurcation delay on combinatorial optimization with continuous variables. arXiv:2211.00822. https://arxiv.org/abs/2211.00822
33. Pramanik S., Chatterjee S., Oza H. (2023). Convergence Analysis of Opto-Electronic Oscillator based Coherent Ising Machines. arXiv:2312.04290. https://arxiv.org/abs/2312.04290
34. Terada Y., Toyoizumi T. (2024). Chaotic neural dynamics facilitate probabilistic computations through sampling. Proceedings of the National Academy of Sciences. DOI 10.1073/pnas.2312992121. https://doi.org/10.1073/pnas.2312992121
35. Nartallo-Kaluarachchi R., Lambiotte R., Goriely A. (2026). Drift-Diffusion Matching: Embedding dynamics in latent manifolds of asymmetric neural networks. arXiv:2602.14885. https://arxiv.org/abs/2602.14885