Time Rescaling for Non-Stationary Process Analysis

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 217 papers23 conceptsfrozen 2026-10-04

Most statistical models assume a steady world. Whether it is a Gaussian Process expecting a consistent length scale or a Poisson process expecting a constant rate, these tools struggle when the underlying "rhythm" of the data changes. In reality, neurons fire in bursts, patients are tested more frequently as they get sicker, and consumer shopping habits fluctuate with seasonal sales.

The core idea shared across these papers is temporal rescaling (or time-warping): a mathematical transformation that stretches or compresses the time axis to "flatten" a non-stationary process into a stationary one. By mapping irregular real-world time into a uniform "operational time," researchers can apply standard statistical tools to complex, time-varying data.

The Intuition: The Rubber Band Metaphor

Imagine a sequence of events—like heartbeats—recorded on a rubber band. When the heart beats faster, the events are crowded together; when it slows down, they spread out. This inconsistency makes it hard to predict the next beat or detect an anomaly.

Temporal rescaling treats the time axis like that rubber band. If we know the "intensity" or "volatility" of the process, we can stretch the crowded sections and compress the sparse ones until the heartbeats appear to occur at a perfectly steady, constant rate. Once the data is "homogenized," we can use simple models that assume a constant rate. To get a prediction in the real world, we simply "un-warp" the result back to the original time axis.

Normalizing Event Sequences

In the study of point processes (sequences of discrete events), the Time-Rescaling Theorem is the foundational tool. It states that any point process can be transformed into a simple, unit-rate Poisson process if we rescale time according to its cumulative intensity function.

Jonathan W. Pillow demonstrated in Time-rescaling methods for the estimation and assessment of non-Poisson neural encoding models how this transformation allows for rigorous goodness-of-fit tests. If a model accurately captures a neuron's behavior, the rescaled spikes should look like a uniform Poisson process; any deviation from this uniformity signals a model failure. Similarly, Shuai Zhang and colleagues used this theorem in Conformal Anomaly Detection in Event Sequences to detect system failures by identifying when event "rhythms" deviate from their rescaled normal state.

This approach is not limited to simple rates. Cina Aghamohammadi and Tatiana A. Engel utilized time rescaling in A doubly stochastic renewal framework for partitioning spiking variability to separate the inherent "noise" of a neuron from the fluctuations driven by external stimuli, revealing stable signatures of neural activity across the brain.

Warping Continuous Signals and Kernels

The same logic applies to continuous signals, such as physiological readings or mechanical oscillations. Standard Gaussian Processes (GPs) use "kernels" that assume a fixed length scale—essentially, they expect the signal to change at a consistent speed.

When a signal's frequency varies, a standard GP will either overfit the slow parts or blur the fast parts. Mu Niu and Dirk Husmeier addressed this in Statistical inference in mechanistic models: time warping for improved gradient matching by introducing a time-warping layer. This layer homogenizes the signal's rhythm before smoothing, allowing a single-length-scale kernel to accurately capture complex dynamics in non-linear differential equations.

For massive datasets, Jan Graßhoff developed warpSKI in Scalable Gaussian Process Separation for Kernels with a Non-Stationary Phase. By placing "inducing points" in a warped temporal space, the method can handle non-stationary signals like heartbeats with over 100,000 observations, a feat previously impossible for standard GP regression.

Applications Across Domains

The power of time rescaling lies in its versatility across disparate fields:

Why It Works

Time rescaling works because it separates what is happening from when it is happening. By defining a bijective mapping between "real time" and "operational time," researchers can isolate the structural properties of a process (like the shape of an inter-spike interval or the form of a periodic function) from the external factors that speed it up or slow it down. As Shinsuke Koyama noted in Coding efficiency and detectability of rate fluctuations with non-Poisson neuronal firing, this separation is not just a mathematical convenience; it often reflects the underlying efficiency of how biological and physical systems encode information.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 13 papers
  2. Version 1 · 2026-10-04 · 17 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_temporal_rescaling_transformation_2c11bcea,
  title = {temporal rescaling transformation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-rescaling-transformation/rdr_2c11bceaf316440ba6a9d8c69e40f83e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_time_rescaling_for_non_stationary_proces_2c11bcea_v1,
  title = {Time Rescaling for Non-Stationary Process Analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/time-rescaling-for-non-stationary-process-analysis/rdr_2c11bceaf316440ba6a9d8c69e40f83e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}