Wilson--Cowan and the Stuart--Landau Limit: A Bogdanov--Takens View

The Wilson-Cowan model contains the Stuart-Landau oscillator as a special case — and a single codimension-2 point in parameter space marks exactly where it stops being one.

The Wilson-Cowan (WILCO) model describes a cortical circuit as two interacting populations: excitatory neurons exciting each other and driving inhibitory neurons, which push back. It's been called a "biologically grounded" version of the Stuart-Landau (SL) oscillator — the universal mathematical description of any system near a smooth oscillation onset (a Hopf bifurcation). This paper makes that claim precise, and shows when it breaks down.

The first result is a clean negative: if you remove recurrent self-excitation (set the excitatory-to-excitatory weight κ_xx = 0), WILCO cannot oscillate at all. The math is direct — the trace of the Jacobian, which must pass through zero for oscillation to appear, is pinned strictly negative by the leak terms regardless of any other parameter. Driving the system harder, tuning the sigmoid, adjusting the inhibitory bias — none of it helps. You need κ_xx > 0, and you need it above a threshold. This corrects a common intuition that "positioning the operating point on the steep part of the sigmoid" is sufficient; it's necessary but not sufficient.

Once κ_xx clears that threshold, what happens next depends on a second quantity: the cross-coupling product κ_xy · κ_yx (how strongly excitation drives inhibition and vice versa). When this product is small, WILCO produces a clean "Hopf bubble" — the system oscillates between two Hopf bifurcations, the cycle is nearly sinusoidal, and the local math at either end is exactly the Stuart-Landau normal form. This is WILCO behaving as SL. When the cross-coupling product is large, a special codimension-2 point called a Bogdanov-Takens (BT) point enters the picture. Three bifurcation curves — a saddle-node fold, a SNIC (saddle-node on an invariant circle), and a Hopf — all emanate from this single organizing point. The fold brings bistability; the SNIC brings Class I excitability, where oscillation frequency goes to zero at onset and the waveform looks like neural spikes rather than sine waves. This is WILCO being genuinely richer than SL.

The paper also draws a useful comparison to second-order neural mass models like Jansen-Rit and PING. Those models achieve a clean Hopf bubble — structurally identical to weak-coupling WILCO — without needing recurrent self-excitation, because their second-order synaptic kernel supplies the destabilizing phase shift internally. Same bifurcation skeleton, different mechanism. Both count as "Stuart-Landau realizations" in a model taxonomy, but the biology underneath is distinct.

The practical upshot: whether a WILCO unit behaves like a smooth resonator (Class II, SL-like) or a pulse-generating integrator (Class I, near SNIC) is not a structural feature of the model — it's a parameter regime. The hinge between them is a single locatable point in the (κ_xx, P_x) plane. That's a concrete, testable handle for anyone building or fitting neural mass models.