Definition

A recurrent network with update rule $\mathbf{h}_{t+1}=F(\mathbf{h}_t,\mathbf{x}_t)$ exhibits a limit cycle when its trajectory converges, from a range of initial conditions and after perturbation, to the same closed (periodic) orbit in state space — distinct from a fixed point (a single state) and from a line attractor (Line Attractor, a 1-manifold of marginally-stable fixed points along which the state integrates evidence, not a periodic orbit it returns to). Establishing a genuine limit cycle, rather than merely noting cyclical-looking PCA trajectories, requires: (1) evidence the orbit is contractive — nearby trajectories converge toward it, typically via a finite-time Lyapunov-index (FTLI) analysis showing $\lambda_{FTLI}<0$ for the trained dynamics (contrasted against chaotic, $\lambda_{FTLI}\ge0$, untrained/random-weight controls); and (2) evidence the orbit is recovered after perturbation rather than merely traced once.

Relative to line-attractor and manifolds-circle

Line Attractor is a different dynamical object with the same general role (a property of the network’s own update dynamics, not a static appearance-only shape): a 1D manifold of marginally-stable fixed points along which state moves to accumulate scalar evidence, with $\lambda\approx 0$ locally rather than genuine contraction toward a periodic orbit. Circle is a different kind of claim entirely — a static, appearance-only topological/geometric fact about how activations are arranged across many different inputs at one fixed point in time (e.g. a genuinely periodic input variable laid out on a ring), established by parametrization/PCA-loop inspection rather than by demonstrating the network’s own dynamics contract toward and recover that loop. A closed-orbit shape can in principle appear in either role; this node is reserved specifically for the dynamical-object case, where the network’s own trained update rule is shown to settle into and recover the cycle.

Key evidence

Li & Zhan (2026) train recurrent policies (RNN/GRU/Mamba, via evolution strategies or PPO) on partially-observable navigation and Procgen tasks (Jumper, Bigfish, Bossfight). Hidden states converge, in PCA projection, to a stable low-dimensional closed orbit that persists across episodes and recovers after perturbation; finite-time Lyapunov-index analysis confirms optimized policies are contractive ($\lambda_{FTLI}<0$) versus chaotic random-weight controls. Canonical Correlation Analysis (CCA) between the recurrent hidden state and a “Behavioral Potential Field” embedding of physical trajectories shows sustained high canonical correlation (>0.7 across 10+ modes for trained agents, versus rapid collapse after 3 dimensions for randomized controls). Causally, selectively randomizing only the top CCA-aligned dimensions of an optimal hidden state collapses navigation behavior back to baseline (“amnesia,” peak convergence time reverting from ~50 to ~250 steps), while randomizing every other dimension preserves fast convergence — proving the discovered CCA subspace, and by extension the limit-cycle geometry it is aligned to, is both necessary and sufficient for the learned navigational behavior. See trained-recurrent-policies-hidden-states-converge-to-a-stable-low-dimensional-limit-cycle-and-selectively-randomizing-the-top-cca-aligned-dimensions-causally-collapses-navigation-behavior-to-baseline.

Key papers

• Li, J. & Zhan, X. (2026). Unraveling the Hidden Dynamical Structure in Recurrent Neural Policies. arXiv:2602.01196 — origin of this map’s limit-cycle characterization, with a causal CCA-subspace ablation tying the discovered geometry to behavior.