Morphologies of Stability
Abstract
Stability is usually taught as a single idea: a system is stable if it returns to where it was after a small push. But the shape of that return differs profoundly across systems. This companion examines four canonical morphologies of stability, each a different answer to the question "what does it mean for a system to hold its form?" The playground implements all four as small noise-driven dynamical systems and exposes the quantities that classical stability theory predicts: relaxation rates, barrier heights, orbit radii, and convergence rates. The deterministic core of each model is calibrated against textbook results so that the visual intuition rests on verifiable arithmetic.
Background
In the linear-stability paradigm, one linearizes a system near an equilibrium and inspects the eigenvalues of the Jacobian. If every eigenvalue has negative real part, perturbations decay and the equilibrium is asymptotically stable; if any has positive real part, the equilibrium is unstable. This criterion, due in modern form to Lyapunov (1892), is exact for hyperbolic fixed points by the Hartman-Grobman theorem, and it underlies the bulk of applied dynamics.
The criterion is powerful but narrow. It answers "is this point stable?" It does not, by itself, describe systems whose stable object is not a point at all: a periodic orbit, a basin among several, or a collective configuration of many interacting units. Each of these requires its own notion of stability, and each produces a recognizably different geometry of recovery. We call these geometries morphologies of stability.
The four models
1. Point attractor
The simplest morphology. A scalar state relaxes toward a set point under a linear restoring force:
dx/dt = -k(x - x*) + noise
The fixed point is x*, where the deterministic drift vanishes. The Lyapunov function V = (1/2)k(x - x*)^2 decreases monotonically along noise-free trajectories, and the single eigenvalue of the linearization is -k, negative for all k > 0. Perturbations decay exponentially with time constant tau = 1/k. Stability here is convergence to rest: one equilibrium, one basin, global pull.
The calibration confirms two facts about this model. First, seeding the state exactly at x* and integrating leaves the error at zero, because f(x*) = 0. Second, a state released far from the target relaxes back to it. Both predictions are exact in the noise-free limit and reproduce to four decimal places.
2. Bistable switch
A particle in a double-well potential, with damping and noise:
V(x) = a(x^2 - 1)^2 + tilt * x
d^2x/dt^2 = -gamma (dx/dt) - dV/dx + noise
For tilt = 0 the potential has two equal minima at x = +1 and x = -1, separated by a barrier whose height is V(0) - V(1) = a. Each well is locally stable; the ridge at x = 0 is unstable. Small perturbations are absorbed within a well, but a perturbation large enough to clear the barrier flips the system into the other basin. Noise makes such flips stochastic, and the rate follows Kramers escape theory (1940):
rate proportional to exp(-2 dV / sigma^2)
where dV is the barrier height and sigma^2 the noise intensity. Stability here is basin selection: the system is stable in more than one way, and its identity is the choice of valley. The playground reports the barrier height directly from the potential (a = 1.3 for the default well depth), and the calibration verifies that a state placed in the right well stays there under noise-free dynamics.
3. Limit cycle
Not all stability is stillness. The supercritical Hopf normal form produces a stable orbit rather than a stable point:
dz/dt = (mu - |z|^2) z + i omega z + noise
written in the playground as a real two-dimensional system in (x, y). For mu <= 0 the only attractor is the origin, a stable spiral. As mu crosses zero the origin loses stability and a stable circular orbit of radius r* = sqrt(mu) appears. The radial Floquet exponent is -2mu, governing how fast a kicked trajectory returns to the orbit. Stability here is rhythm: the system never rests, yet its pattern of motion is self-restoring.
The calibration checks the two structural predictions of the Hopf bifurcation. Above threshold (mu = 0.64) the settled orbit radius matches sqrt(mu) = 0.8 to within the forward-Euler discretization residue of under one percent. Below threshold (mu = -0.3) the trajectory collapses onto the origin, confirming that no orbit exists when mu <= 0.
4. Consensus network
Stability can also be distributed across many units. The model uses DeGroot-style opinion dynamics with an external anchor:
dx_i/dt = c(xbar - x_i) + s(a - x_i) + noise
Each of the N agents is pulled toward the group mean xbar with coupling c and toward an anchor a with stubbornness s. The effective convergence rate is lambda = c(1 - 1/N) + s. When lambda > 0 the spread of opinions decays toward zero and the network reaches agreement. With s = 0 the dynamics conserve the mean (pure averaging), so the consensus value is the initial average; with s > 0 the anchor draws the consensus toward a. Stability here is coordination: the stable object is a collective configuration, not any single agent. The calibration confirms that, under positive coupling and stubbornness, the spread decays to zero.
Results
| Morphology | Stable object | Stability criterion | Recovery geometry |
|---|---|---|---|
| Point attractor | A fixed point x* | eigenvalue -k < 0 | exponential decay, tau = 1/k |
| Bistable switch | One of two wells | barrier height a > 0 | flip requires clearing the barrier |
| Limit cycle | A periodic orbit | Floquet exponent -2mu < 0 | return to radius sqrt(mu) |
| Consensus network | A collective configuration | rate c(1 - 1/N) + s > 0 | spread decays to zero |
All five calibration cases reproduce their textbook targets within tolerance, with a worst-case error below one percent driven entirely by the explicit-Euler integration of the limit cycle.
Limitations
The models are deliberately minimal, and several simplifications matter.
The noise is additive Gaussian white noise with a single intensity parameter. Real systems often have multiplicative (state-dependent) noise, heavy-tailed jumps, or temporal correlations, any of which can change the qualitative dynamics. The stability index shown in the panel is a heuristic ratio of restoring strength to disturbance, not a rigorous measure of basin stability; a proper basin-stability estimate (Menck et al., 2013) would require Monte Carlo sampling of initial conditions across the state space.
The limit-cycle radius carries a sub-one-percent bias because the live simulation integrates with forward Euler at a fixed step. A higher-order integrator would shrink this residue but would not change the structural conclusion that r* = sqrt(mu). Finally, the consensus model uses all-to-all mean-field coupling; real opinion dynamics involve network topology, bounded confidence, and contrarian agents, none of which are present here.
The conceptual bridge
The reason these four models are worth placing side by side is philosophical as much as technical. "Becoming" becomes legible as entityhood precisely when a process can repeatedly recover its own form. A circle drawn on paper is a static shape; a droplet pulled round by surface tension is a self-stabilizing process that maintains its form through active correction. The four morphologies are four ways that "holding form" can be realized: relaxing to a point, selecting a basin, sustaining a rhythm, or coordinating across units. Each is a distinct geometry of persistence, and together they sketch how dynamics, not inertness, is what lets a thing remain itself.
References
- Lyapunov, A. M. (1892). The General Problem of the Stability of Motion.
- Kramers, H. A. (1940). Brownian motion in a field of force and the diffusion model of chemical reactions. Physica, 7(4), 284 to 304.
- Hopf, E. (1942). Abzweigung einer periodischen Lösung von einer stationären Lösung eines Differentialsystems.
- DeGroot, M. H. (1974). Reaching a consensus. Journal of the American Statistical Association, 69(345), 118 to 121.
- Guckenheimer, J., & Holmes, P. (1983). Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields. Springer.
- Olfati-Saber, R., & Murray, R. M. (2004). Consensus problems in networks of agents. IEEE Transactions on Automatic Control, 49(9), 1520 to 1533.
- Strogatz, S. H. (2015). Nonlinear Dynamics and Chaos (2nd ed.). CRC Press.
- Menck, P. J., Heitzig, J., Marwan, N., & Kurths, J. (2013). How basin stability complements the linear-stability paradigm. Nature Physics, 9, 89 to 92.