This work sits at the intersection of nonlinear dynamics and climatology, alongside studies of quantum measurements and extracting laws from noise. Methods for distinguishing chaos from stochasticity in short records have been developed, but for El Niño the question remained open. Some models describe ENSO as low-dimensional chaos driven by the seasonal cycle; others treat it as a random process. The authors test this directly on observational data rather than relying solely on models.
They applied the Kaplan–Glass test to the monthly Niño 3 index (average sea surface temperature in the region 5°S–5°N, 150°W–90°W) from 1871 to 2008. The test is based on phase space reconstruction using Takens' time-delay embedding: from the scalar time series they construct m-dimensional vectors with delay τ, partition the space into boxes, and for each box compute the average length of the displacement vectors of the trajectory. For a deterministic system, the tangent directions within a box are nearly parallel, so this length is close to 1; for a random walk it decays as n^(-1/2). They compared the original series with surrogates having random Fourier phases, which preserve the spectrum and autocorrelation but destroy nonlinear correlations. Additionally, they simulated a delayed equation for thermocline depth with and without white noise.
For the original Niño 3 series, the quantity Λ(τ) was noticeably higher than for phase-randomized surrogates, indicating a low-dimensional deterministic component. However, after subtracting the climatology (the mean annual cycle), Λ(τ) dropped sharply and the difference from surrogates vanished. This means the apparent determinism was an artifact of the seasonal cycle, and the residual signal is stochastic. A delayed model with seasonal forcing but no noise showed high determinism even after removing seasonality, i.e., it exhibited low-dimensional chaos. Adding Gaussian white noise with amplitude σ=0.7 made the model series after removing seasonality indistinguishable from a random process, matching the data. Histograms and superposed epoch analysis showed good agreement between the noisy model and observations: the positive tail of the distribution is heavier due to stronger El Niño events, and the average episode shape is reproduced up to 15 months after onset.
This work differs from neighboring studies by applying a direct test for determinism to the instrumental ENSO record rather than limiting itself to model experiments. It is the first test of the low-dimensional chaos hypothesis for Niño 3 using the Kaplan–Glass method. The result suggests that stochastic forcing of delayed equatorial wave dynamics may be a more appropriate description than pure chaos. For climate modeling, this implies that probabilistic ENSO forecasts should account for the random nature of atmosphere–ocean interaction.
The same test could be applied to other ENSO indices, such as the Southern Oscillation, or to longer paleoreconstructions. It would be interesting to see how stochasticity changes under different regimes (for example, during strong El Niño events). Bayesian methods could also be used to estimate parameters of stochastic delayed models and compare their predictive skill with chaotic models.
The results will influence the development of statistical and hybrid ENSO forecasting models, as well as the interpretation of paleoclimate data, where the seasonal cycle is often removed.
Test the method on other ENSO proxies and on high-resolution model data to see whether stochasticity persists under different conditions.
This connects to unresolved problems in climatology: there is still no consensus on whether ENSO is a self-sustaining oscillator or is forced by noise. This work supports the latter view, but does not rule out a weak deterministic component that the short record cannot reveal.
The Kaplan–Glass test was developed in the early 1990s for analyzing chaotic signals in medicine and physics, but it has also found applications in geophysics, for example in studying magnetospheric substorms.