Establishing causal relationships from observational time series is a central challenge in many sciences, from climatology to neuroscience. Granger causality (GC) is the most common tool, but it assumes linearity, which rarely holds in real systems. Kernel methods have allowed for nonlinearity, yet existing approaches do not account for explicit cross-relationships between variables in feature space. This work fills that gap, building on reproducing kernel Hilbert space (RKHS) theory and statistical guarantees based on Rademacher complexity. Neighboring papers on arXiv explore quantum gravity, gene networks, and black holes, but this article stands apart by focusing on causal analysis of time series in Earth sciences.
The authors construct three variants of kernel Granger causality: the standard stacked (concatenating variables into one vector), the summed (separate kernels for each variable), and the new cross-kernel (XKGC), where the joint mapping includes cross-terms between time series. For the cross-kernel, they introduce three linear transformations in RKHS, yielding a kernel with four terms: two autocorrelation and two cross-correlation terms. Parameter estimation is done via kernel ridge regression (KRR) with an RBF kernel, and hyperparameters are tuned by cross-validation. Statistical significance is determined using surrogate time series (100 permutations) and a threshold equal to the maximum causal strength on surrogates. For soil moisture analysis, a nonlinear complex PCA (ROCK-PCA) is used to extract the interannual component.
On a nonlinear coupled system with strong autoregression, XKGC showed higher detection power and fewer false positives compared to linear GC and standard KGC. In a system of coupled logistic maps, XKGC demonstrated improved power across the entire range of coupling parameter, including the synchronization region. In the task of determining the arrow of time on a coupled Rössler system, only XKGC correctly recovered the causal direction when time delay was varied. Theoretically, the Rademacher complexity bound for the cross-kernel is shown to be always no worse than for stacked and summed kernels, and this is confirmed empirically. For El Niño and soil moisture data, XKGC revealed that about 50% of interannual soil moisture variability is driven by El Niño, with the method distinguishing three phases of the 2015–2016 event and detecting influence in previously unreported regions, such as Northwestern Europe.
This work differs from neighboring arXiv studies by shifting the focus from physical objects (black holes, quantum gravity) to methodology for analyzing observational data in Earth sciences. Unlike previous kernel approaches that ignored cross-relationships, here they are introduced explicitly, yielding tighter generalization bounds and increased sensitivity. This is the first test of a cross-kernel on real climate data, demonstrating its practical value for detecting teleconnections.
The method can be extended to multivariate time series, complex-valued signals, and spatiotemporal structures. It may also incorporate prior information about noise and develop adaptive kernels tailored to specific physical processes. Promising applications include other climate indices and systems with time delays.
The method will impact climatology, neuroscience, econometrics, and any field requiring causal inference from nonlinear time series. It may also improve understanding of teleconnection mechanisms in the Earth system.
Test the method on other climate indices (e.g., the Madden–Julian Oscillation) and on data with higher temporal resolution. Investigate the influence of kernel choice and hyperparameters on the robustness of conclusions.
In Earth sciences, the question of separating causality from correlation in strongly coupled climate variables remains open. This work directly addresses the problem of nonlinear causality but does not account for hidden common factors and nonstationarity, which remain challenges for future research.
Using data from the SMOS satellite (measuring soil moisture in the top 5 cm), the XKGC method found that El Niño explains half of the interannual variability in soil moisture on a global scale, with the influence visible even in Northwestern Europe, where it had not been detected before.
The idea of detecting causal links in time series echoes Isaac Asimov's Foundation series, where psychohistory predicts the future based on analysis of mass data, although there causality is statistical rather than individual.