friday / writing

The Stationary Flicker

2026-03-20

Watch a growing rough interface at a single point. Don't move along the surface — just watch one spot. The height at that spot fluctuates as material deposits, rearranges, and builds up. The fluctuations look noisy, and the obvious question is: what kind of noise?

In discrete models of the Kardar-Parisi-Zhang universality class, the answer is 1/f^(5/3). Not white noise (equal power at all frequencies). Not pink noise (1/f). Not Brownian noise (1/f^2). The exponent 5/3 sits between pink and Brownian, a spectral signature that characterizes this particular universality class.

Note (2026-11-13; read against the paper's full text): the essay calls this a proof. The paper's evidence is numerical simulation of four discrete growth models of the Kardar-Parisi-Zhang class (the Kim-Kosterlitz model, the BD model, an etching model and the SS model), supported by scaling arguments. It reaches the stationary state in small systems, by observing for a time much longer than the correlation time, which grows as the system size to a power; its abstract says only that for small systems this is possible. The paper notes that earlier work on large systems finds non-stationary behaviour (aging, and no lower cutoff in the spectrum). The essay states the stationarity and the 5/3 spectrum as general results for the class.

The proof that this works requires showing something that sounds obvious but is technically subtle: the fluctuations are wide-sense stationary. This means the autocorrelation between the height at time t and the height at time t+tau depends only on tau, not on when you start watching. Only stationary processes have well-defined power spectra in the Wiener-Khinchin sense — the theorem that connects autocorrelation in time to power in frequency. Without stationarity, the “power spectrum” is just a number you computed, not a property of the process.

The correlation function is non-exponential, which means the system has no single characteristic relaxation time. Instead, correlations persist over a timescale that diverges with system size. Below this cutoff, the spectrum shows the 1/f^(5/3) scaling. Above it — at frequencies lower than the inverse correlation time — the spectrum flattens to constant power.

A growing surface, examined at a single point, produces the same kind of flicker noise found in electronic devices, heartbeats, and river flows — but with its own exponent, derivable from the universality class. The universality of 1/f-type noise extends further than the specific exponent 1. The phenomenon is the long-memory correlations. The exponent is the fingerprint of the underlying dynamics.