Lienard Multi-Cycle Recurrent Cell
Implementation & benchmark of arXiv:2608.19542 — The number of limit cycles of piecewise linear Liénard systems
Source paper: The number of limit cycles of piecewise linear Liénard systems arXiv:2608.19542 ⓘ · analyzed Sep 1, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Replace the generic nonlinear drift in a two-dimensional continuous-time recurrent cell by a learnable piecewise-linear Lienard restoring force. Fold breakpoints and jump breakpoints become explicit architectural controls for creating multiple oscillatory attractors, allowing hidden states to encode phase, mode, or periodic memory. Weak input coupling can select or perturb attractors while preserving the autonomous cycle structure.
Formulas
Mathematical statement
Use the latent dynamics dx/dt = F(x) - y + u_x(t) and dy/dt = x + u_y(t), where x and y are hidden states and u_x,u_y are input drives. Define F(x) = a + b x + sum_i a_i abs(x-x_i) + sum_i b_i sign(x-x_i), where x_i are breakpoints, a_i control slope changes or folds, and b_i control jumps. The paper gives constructions with n-m folds and m jumps that produce n+m limit cycles. The outermost slopes a_l and a_r determine a compactified phase-portrait transition: each side has zero equilibria at infinity when the absolute slope is below 2, one at equality, and two above 2. In a neural implementation, this supplies a falsifiable attractor-count prediction and an asymptotic transition threshold.
Implementation notes
Integrate the cell as a small continuous-time recurrent module inside a Neural ODE or as an explicit recurrent update. For each channel maintain x and y, compute F(x) from learnable breakpoint locations x_i and coefficients a, b, a_i, b_i, and apply an RK4 step or small-step Euler update: x_new = x + dt times (F(x)-y+W_x h_in), y_new = y + dt times (x+W_y h_in). Use the exact sign function in the forward pass with a straight-through gradient estimator, or warm-start with tanh((x-x_i)/epsilon) and anneal epsilon. Initialize breakpoints from quantiles of expected latent activations. The first cheap test is an autonomous experiment with n=3 using three folds, two folds plus one jump, and three jumps, followed by a weakly driven sequence task with three periodic modes. Estimate attracting cycles from many random initial states using Poincare-section crossings, period clustering, amplitude clustering, and basin frequencies. The paper predicts that suitable parameters can realize 3, 4, and 6 cycles for these three cases. Separately compute the slopes of the two outermost linear pieces and sweep each through absolute value 2. The quantitative prediction is a transition in the number of large-radius equilibria or escaping trajectories near absolute slope 2, with the measured threshold within 20 percent after discretization effects. Repeat with half the integration step; genuine cycles should retain their count and period, whereas numerical Euler instability should show step-size-dependent norm growth.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.