p-Variation

Path roughness, variation index, regularity
scaledim path space4 metrics

What It Measures

How rough is the signal's path, as characterized by the critical exponent where cumulative variation transitions from finite to infinite.

Computes the sum of |increments|^p for multiple values of p. For smooth signals, all p-variations are finite. For Brownian motion, the 2-variation is finite but the 1-variation diverges. For rougher paths (Levy flights), even the 2-variation diverges. The variation index — the critical p at the transition — is a fundamental path invariant that characterizes roughness independently of amplitude.

Metrics

var_p2

The 2-variation (sum of squared increments, normalized). Fibonacci Word (0.764) scores highest: its binary substitution sequence produces increments of size 1 at almost every step, maximizing the normalized quadratic variation. Thue-Morse (0.667) and Collatz Parity (0.663) are similar — all are binary-valued with frequent transitions. Persistent fBm scores 0.000005 (nearly zero — its smooth, correlated increments have tiny squared variation). This metric separates "jumpy" binary signals from "smooth" continuous ones.

variation_index

The estimated critical p. Accelerometer Jog, Accelerometer Walk, and Wind Speed all hit 4.0 (maximum — their paths are extremely rough, with variation diverging even at high p). Mu-law Sine scores 1.0 (smooth path with finite 1-variation — the slope transition happens at p=1). The variation index is invariant under affine transformations (shift and scale) of the signal but not under general monotone transforms.

increment_persistence

Lag-1 autocorrelation of increments. Clipped Sine (0.98) and Mackey-Glass (0.97) have strongly persistent increments (smooth, trending dynamics). Logistic period-2 (-1.0) has perfectly anti-persistent increments (each step reverses the previous one). This is the path-roughness analog of the Hurst exponent: positive = smooth trending, negative = rapidly alternating.

volatility_clustering

Autocorrelation of absolute increments (|Δx_t|). Lotka-Volterra (0.95) and Clipped Sine (0.94) score highest — their increment magnitudes are strongly correlated over time (large changes cluster together). Phyllotaxis (-0.62) and Circle Map QP (-0.62) score most negative — their increment sizes anti-alternate. This is the classic GARCH-like volatility clustering signature, measured directly from the path's p-variation structure.

Atlas Rankings

increment_persistence
SourceDomainValue
Clipped Sinewaveform0.9718
Mackey-Glassmedical0.9689
Ocean Swellgeophysics0.9603
···
Logistic r=3.2 (Period-2)chaos-1.0000
Logistic r=3.5 (Period-4)chaos-0.9761
Noisy Period-2chaos-0.9670
var_p2
SourceDomainValue
Fibonacci Wordexotic0.7639
Thue-Morseexotic0.6666
Collatz Paritynumber_theory0.6627
···
Constant 0xFFnoise0.0000
fBm (Persistent)noise0.0000
PID Controllerexotic0.0000
variation_index
SourceDomainValue
LIGO Livingstonastro4.0000
Wind Speedclimate4.0000
Nikkei Returnsfinancial4.0000
···
μ-law Sinewaveform1.0001
Van der Pol Oscillatorexotic1.0003
Sine Wavewaveform1.0003
volatility_clustering
SourceDomainValue
Lotka-Volterrabio0.9495
Van der Pol Oscillatorexotic0.9318
Clipped Sinewaveform0.9310
···
Phyllotaxisbio-0.6178
Circle Map Quasiperiodicchaos-0.6178
Hilbert Walkexotic-0.5335

When It Lights Up

p-Variation captures path roughness in a way that's complementary to Holder Regularity. Holder measures local smoothness point by point; p-Variation measures global path roughness as a summability property. In the atlas, var_p2 strongly separates binary/symbolic signals (Fibonacci, Thue-Morse, Collatz Parity: high var_p2) from continuous oscillations (fBm, sine waves: near-zero var_p2). The variation_index provides the roughness invariant that connects the framework to stochastic analysis: Brownian motion has index 2, and deviations from 2 indicate either smoother-than-Brownian (index < 2) or rougher-than-Brownian (index > 2) dynamics.

Open in Atlas
← Multi-Scale WassersteinLogarithmic Spiral →