Nonstationarity

heteroskedasticity, regime switches, geometric non-stationarity
dynamicaldim trajectory4 metrics

What It Measures

How the signal's local geometric character changes over time.

Computes a 5D descriptor (entropy, lag-1 autocorrelation, variance, kurtosis, permutation entropy) on sliding windows and tracks its trajectory through descriptor space. The other geometries compute static, full-sequence summaries. This one measures the derivative: how fast the local character is changing, how bursty that change is, and how much of the descriptor space the trajectory explores.

Metrics

metric_volatility

Mean speed of movement through descriptor space (z-scored). Triangle Wave (3.31) and Clipped Sine (3.27) score highest because their periodic structure creates rapid, repeated transitions between distinct geometric regimes. Logistic Period-5 (2.95) is similar. Devil's Staircase scores 0.0: its local geometry is constant within each plateau, and the jumps between plateaus are too rare to raise the mean speed.

regime_persistence

How long do geometric regimes last? Measured by the autocorrelation decay time of the descriptor trajectory. Rossler Hyperchaos, Quantum Walk, and Lotka-Volterra all score 1.0 (maximum persistence — once they enter a geometric regime, they stay). Rossler Attractor scores 0.033 (regimes change rapidly as the trajectory spirals between lobes). High persistence signals piecewise-stationary dynamics.

trajectory_dim

PCA participation ratio of the descriptor cloud, normalized by 5. How many independent descriptor axes does the trajectory use? Zipf Distribution (0.904) and Noisy Sine (0.896) explore nearly the full 5D space. Devil's Staircase scores 0.0 (the trajectory is confined to a single point in descriptor space). High trajectory_dim means the signal's local geometry changes in multiple independent ways simultaneously.

vol_of_vol

Coefficient of variation of the descriptor speed. Is the rate of geometric change itself stable or bursty? Gaussian Collatz (2.08) and Thue-Morse (2.01) score highest: their geometric changes come in bursts separated by calmer intervals. This is the actual "volatility of volatility" — a second-order nonstationarity measure. Van der Pol (1.74) scores high because its relaxation oscillations create alternating fast and slow geometric evolution.

Atlas Rankings

metric_volatility
SourceDomainValue
Triangle Wavewaveform3.3135
Clipped Sinewaveform3.2654
Logistic r=3.74 (Period-5 Window)chaos2.9489
···
Constant 0xFFnoise0.0000
Constant 0x00noise0.0000
Devil's Staircaseexotic0.0000
regime_persistence
SourceDomainValue
Rössler Hyperchaoschaos1.0000
Quantum Walkquantum1.0000
Lotka-Volterrabio1.0000
···
Constant 0xFFnoise0.0000
Constant 0x00noise0.0000
Rossler Attractorchaos0.0333
trajectory_dim
SourceDomainValue
Zipf Distributionexotic0.9039
Noisy Sine (SNR 3 dB)waveform0.8957
Gaussian Noisenoise0.8907
···
Constant 0xFFnoise0.0000
Constant 0x00noise0.0000
Devil's Staircaseexotic0.0000
vol_of_vol
SourceDomainValue
Gaussian Collatz Orbitnumber_theory2.0839
Thue-Morseexotic2.0135
Van der Pol Oscillatorexotic1.7395
···
Constant 0xFFnoise0.0000
Constant 0x00noise0.0000
Devil's Staircaseexotic0.0000

When It Lights Up

Nonstationarity detects regime switching and concatenation that static metrics miss entirely. A signal made by splicing together two different sources will score high on vol_of_vol (bursty regime changes) and trajectory_dim (multiple descriptors change) while possibly looking unremarkable to any single static geometry. In the atlas, regime_persistence separates the dynamical view's "coherent chaos" cluster (Rossler Hyperchaos, Lotka-Volterra: chaotic but geometrically stable) from "incoherent chaos" (Rossler Attractor: chaotic and geometrically unstable).

# Scale Geometries

Open in Atlas
← Gottwald-MelbourneLogarithmic Spiral →