Ring Attractors: Circular Dynamics

• Ring attractors are dynamical systems with a circular topology that support a continuous activity bump for encoding angular, phase, or directional variables.
• They use distance-dependent connectivity with localized excitation and global inhibition to stabilize the activity bump and allow smooth transitions.
• These models are applied in neuroscience for head-direction circuits and in deep reinforcement learning to enhance spatial decision-making.

A ring attractor is a dynamical system whose attractor manifold has the topology of a circle, supporting a continuous, neutrally stable activity “bump” that can persist or move along the ring. This architecture enables the persistent representation of angular, phase, or directional variables in biological and artificial systems. Ring attractors arise in fields ranging from theoretical neuroscience—where they model head-direction or orientation selectivity circuits—to general-relativistic dynamics in the guise of equatorial rings in Poynting–Robertson effect flows, and have recently been adapted as architectural modules in deep reinforcement learning (RL) for spatially structured decision-making (Falco et al., 2019, Saura et al., 2024).

1. Mathematical Foundations in Dynamical Systems and Neural Fields

A ring attractor’s defining feature is its continuous symmetry under rotation, with a “bump” of activity stabilized by the connectivity and nonlinearity of the underlying system. In neural-field models, the ring topology is realized by parameterizing the network coordinate as an angular variable $\theta\in[0,2\pi)$ and employing distance-dependent recurrent connectivity that ensures local excitation and broad inhibition (typically a "Mexican-hat" profile):

$$\tau\,\frac{\partial v(\theta,t)}{\partial t} = -v(\theta,t) + \int_{0}^{2\pi} w(\theta-\theta')\,f(v(\theta',t))\,d\theta' + I(\theta)$$

Here, $w(\phi)$ is usually of the form $$J_e\,e^{-\phi^2/(2\sigma_e^2)}-J_i\,e^{-\phi^2/(2\sigma_i^2)}$$ with $J_e,J_i>0$ and $\sigma_e<\sigma_i$, inducing localized self-excitation and global inhibition. For a sufficiently strong central peak and adequately suppressed flanks, this network supports stable bump solutions that can drift along the ring in response to perturbations or inputs (Saura et al., 2024).

In discrete network implementations, the ring is composed of $N$ excitatory units $\{v_n\}$ (each representing a preferred angle $\alpha_n=2\pi(n-1)/N$) and one global inhibitory unit $u$. The lateral connectivity $w^{E\to E}_{mn} = \exp(-d_{mn}^2/\lambda^2)$ (where $d_{mn}=\min\{|m-n|,N-|m-n|\}$) implements the periodic topology, while the balance against global inhibition ensures single-bump stability.

2. Ring Attractors in General-Relativistic Poynting–Robertson Effect

In the context of the three-dimensional general-relativistic Poynting–Robertson (PR) effect, ring attractors emerge as geometric configurations—specifically, the equatorial ring and critical hypersurface—defining the long-term fate of infalling test particles subject to gravitational and radiation forces in a Kerr geometry (Falco et al., 2019). The equations of motion govern a six-dimensional phase space for such particles, encompassing velocity components and spatial coordinates:

$$\frac{d\nu}{d\tau},\,\,\frac{d\psi}{d\tau},\,\,\frac{d\alpha}{d\tau},\,\,\frac{dr}{d\tau},\,\,\frac{d\theta}{d\tau},\,\,\frac{d\varphi}{d\tau}$$

The critical hypersurface $\mathcal{H}$, defined by the solution to balance conditions for velocity and location (Eqs. I–II in (Falco et al., 2019)), forms a compact 3D manifold in phase space, diffeomorphic to an axisymmetric ellipsoid. Within $\mathcal{H}$, all non-equatorial orbits are unstable to latitudinal drift, with the unique equatorial ring $\mathcal{E}$ serving as the ultimate attractor. This situation rigorously demonstrates a ring-type manifold as the global basin of attraction for almost all bounded solutions, with asymptotic stability proved via Lyapunov functionals (see Section 4 below).

3. Ring Attractors as Architectures in Deep Reinforcement Learning

Recent work has imported ring attractor models into deep RL architectures, exploiting their ability to capture spatial structure and action continuity. In such systems, discrete actions $a\in\{1,\dots,A\}$ are mapped onto ring positions $\theta_a=2\pi(a-1)/A$, while the ring network receives state-dependent value signals $Q(s,a)$ projected as Gaussian input profiles centered at each $\theta_a$:

$$x_n = \sum_{a=1}^A \frac{Q(s,a)}{\sqrt{2\pi\sigma_a^2}} \exp\left(-\frac{1}{2}\left(\frac{\alpha_n-\theta_a}{\sigma_a}\right)^2\right)$$

The network dynamics then evolve as:

$$\tau\,\frac{dv_n}{dt} = f\left(\sum_{m=1}^N w^{E\to E}_{mn} v_m + x_n + w^{I\to E}_n u\right) - v_n$$

$$\tau\,\frac{du}{dt} = f\left(w^{I\to I} u + \sum_{m=1}^N w^{E\to I}_m v_m\right) - u$$

Action selection is decoded by identifying the neuron with maximal activation ($n^* = \operatorname{argmax}_n v_n$) and mapping back to its corresponding action. This mechanism enforces smooth transitions between adjacent actions and preserves spatial or rotational symmetries required by the task environment (Saura et al., 2024).

When implemented as a recurrent layer within a DL agent, the ring topology is maintained by distance-dependent input and recurrent weights, yielding a ring-RNN update:

$h_t = \tanh(V(s_t) + U(h_{t-1}))$

with $V_{m,n}$ and $U_{m,n}$ parameterized by decaying exponentials of ring distance.

4. Stability Analysis and Lyapunov Functions

The stability of ring attractor dynamics can be established by exhibiting suitable Lyapunov functions. In the PR system, three Lyapunov functionals are constructed:

• Kinetic + potential-energy difference:

$$\mathbb{K}(\boldsymbol x)=\frac{m}{2}[\nu-\nu_\mathrm{crit}(\theta)]^2 + (A-M)\left(\frac{1}{r}-\frac{1}{r_\mathrm{crit}(\theta)}\right)$$

• Angular-momentum difference:

$$\mathbb{L}(\boldsymbol x)=m\left(r\nu\sin\psi\cos\alpha - r_\mathrm{crit}(\theta)\nu_\mathrm{crit}(\theta)\right)$$

• Rayleigh dissipative potential difference:

$$\mathbb{F}(\boldsymbol x)=\tilde{\sigma}\,\mathcal{I}^2\left[\ln\left(\frac{\mathbb{E}_\mathrm{crit}}{E_p}\right) - \ln\left(\frac{\mathbb{E}(U)}{E_p}\right)\right]$$

Each is strictly decreasing along trajectories outside the critical hypersurface $\mathcal{H}$ and vanishes only on $\mathcal{H}$, establishing $\mathcal{H}$ as an asymptotically stable attractor. Within $\mathcal{H}$, the dynamics drive all orbits to the equatorial ring $\mathcal{E}$ (Falco et al., 2019). The Lyapunov approach carries directly to neural ring attractors, where bump stability and resistance to perturbations are analyzed by linearization and spectral properties, with lateral inhibition ensuring all modes except translation decay.

5. Practical Applications and Empirical Performance

In RL, ring attractor modules enhance the expressivity and performance of agents operating in spatially structured or rotationally symmetric action spaces. Empirical evaluations on Atari-100k, Super Mario Bros, and Highway-env show substantial improvements:

Model	Benchmark	Mean Human-Normalized Score	Performance Gain
EfficientZero	Atari-100k	0.959	—
EffZeroRA	Atari-100k	1.454	+53%
DDQN	Mario/Highway	—	Baseline
DDQNRA	Mario/Highway	—	+15–20%

(Table adapted from (Saura et al., 2024), Table 1)

Ablation studies show that removing the spatial mapping or distance-dependent weights negates the performance gains, underscoring the specific contribution of the ring topology. Qualitative analyses confirm that ring attractors facilitate smooth action interpolation, robust representation of head direction, and generalized spatial reasoning. The architecture is biologically plausible, paralleling ring-attractor circuits observed in Drosophila and mammalian cortex.

6. Basin of Attraction, Topology, and Phase-Space Geometry

The basin of attraction for ring attractors in both physical and neural systems is characterized by the property that nearly all trajectories (except those escaping to spatial infinity) converge to the attractor manifold. In the PR context, the critical hypersurface $\mathcal{H}$ is topologically a compact, axisymmetric ellipsoid; its internal equatorial ring $\mathcal{E}$ is pointwise attracting, and the boundary between capture and escape can exhibit sensitive dependence on initial data—manifesting as fractal-like basin boundaries at high luminosity (Falco et al., 2019). In neural or DL implementations, the attractor basin corresponds to the network’s ability to robustly recover and maintain the bump in the face of input noise or perturbations, so long as global inhibition and excitation are appropriately balanced.

Verbal descriptions (from (Falco et al., 2019)):

• The 3D critical hypersurface appears as a flattened egg-shaped surface of revolution.
• Cross-sectional slices reveal manifest symmetry about the equatorial plane.
• Trajectories starting near, but not on, $\mathcal{H}$ spiral in as Lyapunov functionals decrease monotonically to zero at capture, with trajectories eventually drifting toward the equatorial ring.

7. Extensions and Broader Implications

Ring attractor models have been extended to multi-agent systems via coupled rings, continuous-action variants employing continuum fields, and uncertainty-aware exploration mechanisms in RL via variable-width Gaussian input profiles. Biological observations support the plausibility of ring-based architectures for encoding head direction and orientation variables in both invertebrate and vertebrate species. In engineering, the explicit spatial embedding and smooth continuity afforded by ring attractors enable more robust decision-making in robotic and navigation contexts, particularly where actions or states are naturally circular or periodic.

A plausible implication is that the ring attractor motif provides a canonical solution to tasks requiring stable, continuous representation and manipulation of angular variables or phase information across diverse domains.