Imagine an interpreter who gets distracted by every rustle in the hall. The more rustles, the more errors in the translation, even though the meaning of the speech hasn't changed. The same happens with data: if measurements are unevenly spaced, standard analytical tools start "hearing" changes that never occurred in the underlying process. Irregular time series are everywhere, from stock tick data to layers of stalagmite in a cave. Direct application of familiar methods is impossible, and attempts to "even out" the series by interpolation add their own distortions.

One popular way to compare segments of such a series is a similarity measure based on edit-distance (the cost of transforming one segment into another through shifts, deletions, and insertions of points). It comes from text processing and is convenient because it does not require equal lengths. But the authors of the study discovered that the transformation cost systematically increases with segment size. A longer segment appears more dissimilar to others simply because it contains more points. It's as if the interpreter confused words more often the longer the speech went on, even if the speaker was equally clear throughout.

To separate real changes from artifacts of measurement frequency, the researchers proposed a method of surrogate data with constraints on sampling frequency (SRC-surrogates). The idea is simple: generate many artificial copies of the series that preserve segment sizes and the distribution of intervals but destroy the dynamics. Then the predictability measure (DET) of the real series is compared with the distribution of DET from surrogates. If the real series falls outside this distribution, its behavior cannot be explained by uneven sampling alone.

On a synthetic series with an abrupt change in sampling frequency, the method worked: the spurious shift to higher predictability disappeared, and the linear trend built into the model was recovered. The method was then applied to real data: a stalagmite from Niue Island that grew for 1000 years (6.4–5.4 thousand years ago). Without correction, it seemed that the predictability of seasonal rainfall was increasing, but after subtracting the artifact, two periods of reduced predictability emerged. These coincide with intensified El Niño and tropical cyclones.

Now researchers have a tool for honest comparison of episodes with different measurement densities. This is especially important for paleoclimate, where the time step is set by natural processes, not by the experimenter's choice. The method is universal and can be applied to any edit-distance-based measure, and in the future to clustering, classification, and networks. The interpreter is no longer thrown off by rustles: they have learned to distinguish them from the speaker's voice.