What Is an Oscillation, Really? — Rosetta Stone of Neural Mass Models

1Two different meanings, one word

The paper is careful about this (§1.1, concept (2)): "oscillator" and "oscillation" aren't quite the same claim. An oscillator is a system built to sustain rhythmic motion, usually through an attracting limit cycle, or, in the idealized undamped case, a neutrally stable center. An oscillation is a property of data — a signal with a dominant timescale, typically visible as a narrow peak in a power spectrum. Usually these line up. But not always: a damped system driven by noise can throw off "noise-sustained quasi-cycles" — spectral peaks with no limit cycle underneath them at all — while a chaotic system can be aperiodic and still show a prominent spectral bump. Every model on this site is an oscillator in the first sense. Whether its output counts as an oscillation in the second sense is, in the end, a question you ask of the data, not the equations.

2The compression test

"In algorithmic information theory terms, we would say that a signal is an oscillation if it can be efficiently compressed by exploiting its approximate periodicity." — Castaldo et al., §1.1, concept (2)

If a signal roughly repeats, you don't need to store every value — one cycle plus a short note on how each repeat differs from the last is enough to reconstruct the rest. The paper's Appendix J formalizes exactly this: a dataset counts as an oscillation when a description built from a "periodic template" plus a small correction is dramatically shorter than any description that doesn't assume periodicity. No reference model required. The test is just whether assuming a repeating pattern makes the data cheaper to describe than not assuming it.

Left: stored value-by-value, a signal has no shortcuts. Right: if it's roughly periodic, one clean cycle plus a small correction reconstructs the whole thing — same signal, far less to write down. The paper's compression test (Appendix J) is just this idea made precise.

3The circle underneath everything

Push the idea one step further (Appendix J) and every oscillator on this ladder — however many state variables, however much biological detail — reduces to the same topological object: a phase circling a loop.

Start with the barest possible oscillator: a phase clock, θ̇ = ω, tracing a circle with no amplitude at all — that's the Kuramoto rung on this site. Add amplitude and you get the undamped harmonic oscillator, z = z₀e^iωt: still a perfect circle, now with a radius that never changes. Neither of these has any mechanism to self-correct — nudge them and they stay nudged.

Real oscillators settle back to a fixed rhythm after a disturbance. That takes a nonlinearity. When the paper derives Stuart–Landau's cubic term from first principles (normal-form reduction near a Hopf bifurcation — the point where a pair of complex-conjugate eigenvalues crosses the imaginary axis, flipping a fixed point's stability), it doesn't start by assuming the equation should respect circular symmetry. It starts from a generic nonlinear correction to the harmonic oscillator and asks which terms survive the reduction. Almost everything cancels. The one term that can't be removed is −(γ+iβ)|z|²z — and that term happens to be exactly the one that respects the same rotational symmetry the bare circle already had.

Three rungs, one shape. Phase-only traces a bare circle; the harmonic oscillator adds a radius that never changes; Stuart–Landau adds the one nonlinear term that survives normal-form reduction — the term that pulls a perturbed trajectory (dashed) back onto the ring.

The circle is structural — it's the one piece of geometry that persists across every model in the ladder, which is why it shows up again, in some more detailed costume, at every single rung above it.

See the model ladder → Back to the intro →