Hidden Symmetry Found in the Brain’s Messy Head-Direction Circuits

For decades, theoretical neuroscientists have been enchanted by an idea of almost mathematical purity: that the brain represents continuous quantities—head direction, spatial position, angular velocity—using so-called continuous attractor networks. In the classical picture, neurons are wired together with exquisite symmetry, and this symmetry produces a manifold of stable activity states, any one of which can represent a different value of the variable. A bump of neural activity slides smoothly around an abstract ring, tracking the angle of an animal’s head as it turns. The catch has always been that real neural circuits refuse to look like the theory. Recordings from rodent head-direction cells reveal a riot of heterogeneity: tuning curves of wildly different widths, depths and shapes, nothing like the identical, merely shifted profiles that symmetric ring models demand. A new study published in Nature Neuroscience by David G. Clark, L. F. Abbott and Haim Sompolinsky now shows that this apparent contradiction may be an illusion, and that the messiness of mammalian circuits can coexist with—perhaps even conceal—the elegant machinery of continuous attractors.

The team’s strategy was unusual and, in many ways, audacious. Rather than building an idealized model and hoping the brain matches it, they started from data itself. Working with large-scale electrophysiological recordings of 1,533 head-direction cells in the mouse postsubiculum, collected across 31 mice, they constructed recurrent network models directly from the measured tuning curves. The resulting networks, containing as many units as recorded neurons, behaved as what the authors call quasicontinuous attractors: systems that retain the essential dynamics of a ring attractor despite the pronounced disorder visible in their connectivity. This data-driven route sidesteps a long-standing gap between idealized theory and experimental reality, and it demonstrates that the heterogeneity long seen as evidence against attractor models is, in fact, fully compatible with them.

To understand how disorder and order can coexist, the researchers turned to a generative approach. They built a statistical process capable of reproducing the heterogeneous tuning of head-direction cells—capturing the distributions of tuning widths, peak rates and correlations—while keeping the resulting connectivity mathematically tractable even at large network sizes. The key insight is that a weight matrix can look disordered, with connections scattered apparently at random, and yet carry a hidden circular symmetry. When the authors analyzed the eigenvalues of their optimized weight matrices, they found degeneracies—repeated eigenvalues—that are the fingerprint of rotational symmetry. In a purely random network, such degeneracies would be vanishingly unlikely. Their presence in a seemingly disordered connectome could therefore serve as a diagnostic signature, a way of detecting continuous-attractor organization hidden inside complex biological wiring diagrams.

The mathematical machinery the authors deployed is dynamical mean-field theory, a set of techniques borrowed from statistical physics that describes the collective behavior of large recurrent networks. Applied to their data-derived circuits, the theory recovered precisely the ingredients of the classical ring attractor. The effective interactions took the form of a Mexican hat profile: local excitation surrounded by broader inhibition, the canonical circuit motif that stabilizes a localized bump of activity. Even more strikingly, the analysis revealed a spontaneously broken symmetry. Just as a magnet cooled below its critical temperature picks a direction and thereby breaks rotational symmetry, the network’s activity settles into a single bump at a particular angle, selecting one state out of the continuous family that the underlying symmetry makes available. The classical picture, it turns out, is not wrong—it is simply hidden one level below the surface disorder.

This result carries real weight for a debate that has animated computational neuroscience for years. Continuous attractor models have enjoyed spectacular support in the insect brain. In the fruit fly’s central complex, the ellipsoid body circuit implements a ring-like architecture with remarkable regularity, and connectome studies have confirmed the predicted wiring motifs. Zebrafish heading circuits show similar organization. But the mammalian brain, with its cortex, its vastly larger neuron counts and its apparently haphazard connectivity, has resisted such clean interpretations. Head-direction cells were discovered in the rat postsubiculum in the 1980s, and subsequent work in mice has documented their heterogeneity in detail. Critics could reasonably ask whether the tidy ring model was a theorist’s fantasy. The new study answers that a ring attractor need not look tidy. Symmetry can be present in the statistics of the connectivity rather than in its explicit, neuron-by-neuron arrangement.

The implications extend beyond head direction. Grid cells in the entorhinal cortex, which fire at the vertices of hexagonal lattices as an animal explores its environment, have been shown to exhibit low-dimensional continuous attractor dynamics, with population activity confined to a toroidal manifold. Yet grid cell tuning is also notoriously heterogeneous, and the new framework offers a principled way to reconcile that diversity with attractor organization. Indeed, the authors extended their approach to build a grid-cell model derived from heterogeneous tuning curves, suggesting that the same data-to-network methodology could illuminate other cognitive circuits where idealized models and messy recordings seem to disagree. The lesson is general: representational heterogeneity and dynamical order are not mutually exclusive, and the search for structure should target the statistics of connectivity, not its superficial appearance.

There is also a practical payoff for the emerging era of connectomics. Projects mapping the complete wiring of brain regions are producing weight matrices of staggering size, and a central challenge is to read function off structure. If continuous-attractor organization announces itself through eigenvalue degeneracies—outliers in the spectrum of the connectivity matrix that cannot be explained by randomness—then theorists have handed experimentalists a concrete, computable test. One could scan a connectome for these spectral signatures and identify, purely from anatomy, which circuits are likely to implement continuous variables. Given that the study’s methods are fully reproducible—the authors released an installable Python package, eleven Jupyter notebooks generating every figure, and cached intermediate results archived on Zenodo—the community can immediately build on the approach.

The work also resonates with a broader shift in how neuroscientists think about trained and biological networks. Recent theory has emphasized that recurrent networks performing a computation often combine a structured, low-rank component with a high-dimensional disordered background, and that the structure lives in a few dominant modes of the connectivity. The new study pushes this idea further by showing that the structured component need not be a simple rank-one perturbation; it can be a full symmetry group, with the degeneracies that rotational invariance implies, embedded in an otherwise disordered matrix. In this view, the classical ring attractor and the disordered biological network are two descriptions of the same object, one coarse-grained and idealized, the other detailed and realistic, connected by the tools of mean-field theory.

Of course, important questions remain. The study builds networks from tuning curves, which constrain connectivity indirectly; direct experimental verification of the predicted weight statistics will require connectome data from the postsubiculum or related circuits, which do not yet exist at sufficient scale in mammals. Whether the eigenvalue degeneracies survive biological noise, cell-type mixing and the realities of cortical microcircuitry is an open empirical question. And while the framework explains how heterogeneity and attractor dynamics coexist, it does not yet explain how such circuits develop or learn during an animal’s lifetime. Still, the conceptual advance is hard to overstate: a core theoretical construct of neuroscience, long dismissed as too clean for real brains, has been shown to survive contact with real data in disguise. The brain, it seems, does not abandon its symmetries—it buries them where only careful mathematics can find them.

Subject of Research: Continuous attractor dynamics and hidden symmetry in disordered mammalian head-direction neural circuits

Article Title: Symmetries and continuous attractors in disordered neural circuits

Article References: Symmetries and continuous attractors in disordered neural circuits. (n.d.). https://doi.org/10.1038/s41593-026-02452-4

Image Credits: AI Generated

DOI: 10.1038/s41593-026-02452-4

Keywords: neuroscience, continuous attractors, head-direction cells, ring attractor, neural circuits, computational neuroscience, dynamical mean-field theory, connectomics, postsubiculum, neural tuning heterogeneity, spontaneous symmetry breaking, recurrent networks