The Irregular Pulse of Time: How Uneven Measurement Rhythms Skew Conclusions and What to Do About It
Context
Analyzing irregular time series remains a challenge for the sciences of complex systems. Related papers in the archive follow quantum algorithms and numerical methods for equations, where noise and instability are also tackled, but with different tools. In geophysics and paleoclimatology, uneven measurement spacing is not just a nuisance: it carries information about processes, such as stalagmite growth rate. The edit-distance method, borrowed from text processing, allows comparing segments of different lengths, but until now it was unclear how variable measurement density affects it.
Methods
The foundation is (m)Edit-distance: a similarity measure between two segments of a series, computed as the minimum cost of transforming one into the other via shift, deletion, and insertion operations. The authors systematically studied how this cost depends on the number of points in the segments, using synthetic series (white noise, AR(1), sinusoid) with gamma-distributed intervals. They then proposed a correction scheme: an ensemble of surrogate series is generated that preserves the real segment sizes and interval distributions but destroys the dynamics. Comparing the determinism measure (DET) of the real series with the distribution of DET from surrogates allows subtracting the contribution of sampling frequency.
Results
They found that the transformation cost increases with segment size: longer segments appear more dissimilar. This effect varies across dynamics: for AR(1) it is weaker than for noise or sinusoid. When sampling frequency changed abruptly in a synthetic series, DET showed a false transition to higher predictability. After correction with SRC surrogates, the false transition disappeared, and the linear increase in predictability built into the model was restored. On the real stalagmite record from Niue (6.4–5.4 thousand years ago), the correction revealed two periods of reduced seasonal predictability, coinciding with intensified El Niño and tropical cyclones. Without correction, these periods were masked by the artifact of high stalagmite growth rate.
Implications
This work fills an important gap: it shows that even 'honest' metrics for irregular series are not free from bias when sampling frequency changes. The proposed correction method is universal: it does not depend on the specific data type and can be applied to any measure based on edit distance. This is a step forward compared to related works, where irregularity is often bypassed by interpolation or ignored. Researchers now have a tool for reliably comparing episodes with different measurement densities, which is especially important for paleoclimate, where the time step is set by natural processes.
Future development
The SRC-surrogate method can be extended to other similarity measures and to tasks such as clustering, classification, or network construction. The authors plan to study in more detail how the dependence on segment size manifests for different signal types. A promising application is to adaptive windows, where segment length is chosen not by time but by a physical parameter, such as the cardiac cycle. This would allow analyzing electrocardiograms or seismic records without prior interpolation.
Impact
The method is directly applicable in paleoclimatology, geophysics, astronomy, and finance, where series are often irregular. It is also useful for analyzing data with gaps in sociology and medicine.
Next steps
Test the method on other paleoclimate archives, such as ice cores or lake sediments. Investigate the sensitivity of the correction to the choice of surrogate generation parameters.
Key open problems
In time series analysis, the question of how to separate internal nonstationarity of the system from observational artifacts remains open. This work contributes to solving this problem for irregular series, but does not exhaust it: for example, cases where the process governing sampling frequency is itself nonstationary are not considered.
The stalagmite from Niue grew faster during wet periods, so its layers contained more measurement points per year, and without correction this looked like an increase in climate predictability, when in fact it was just an artifact.
Key numbers
- paleoclimate record period: 1000 years (6.4–5.4 thousand years ago)
- number of surrogates in the ensemble: 200
- recurrence fraction in the analysis: 15%
- sliding window size on synthetic data: 200 intervals
- sliding window size on paleo data: 200 years