Surrogate Time Series Testing

Description

Compare a time-series metric on real data with its distribution across surrogate series that keep chosen properties but satisfy a null hypothesis, showing whether it reflects real structure.

Algorithms for dimension, entropy, Lyapunov exponents, synchrony and similar metrics return a number even for random or linear data, so a value alone is no evidence of structure. Surrogate testing generates many series from the measured data that preserve some properties but are consistent with a null hypothesis, computes the metric on the real series and on every surrogate, and rejects the null if the real value falls outside the surrogate distribution. For a purely random null a random shuffle of the values suffices, which keeps the distribution but destroys all temporal order. For a stationary linear stochastic null, a Fourier transform followed by phase randomisation and inverse transform keeps the power spectrum and autocorrelation but destroys higher-order timing. Rejecting linearity makes nonlinearity plausible but does not demonstrate chaos.

Attributes

Loss type
Unmigrated
Effort
Days

Steps

  1. State the null hypothesis: a random process, or a stationary linear stochastic process.
  2. Generate many surrogate series from the measured data (for example 99 or more): random shuffles for the first null, Fourier transform with randomised phases and inverse transform for the second.
  3. Compute the metric of interest on the measured series and on every surrogate.
  4. Locate the measured value in the surrogate distribution and reject the null if it lies outside (for example beyond the 2.5th or 97.5th percentile).
  5. Report the null used, so readers know which structure the result rules out.

Sources

  • Signal Processing for Neuroscientists (van Drongelen, 2018)

Also known as

  • Surrogate data test
  • Phase-randomised surrogates
  • Shuffle surrogates
  • Test for nonlinearity

Categories

  • Information Theory, Physics & Mathematics