LICORS: Light Cone Reconstruction of States for Non-parametric Forecasting of Spatio-Temporal Systems

We present a new, non-parametric forecasting method for data where continuous values are observed discretely in space and time. Our method, “light-cone reconstruction of states” (LICORS), uses physical principles to identify predictive states which are local properties of the system, both in space and time. LICORS discovers the number of predictive states and their predictive distributions automatically, and consistently, under mild assumptions on the data source. We provide an algorithm to implement our method, along with a cross-validation scheme to pick control settings. Simulations show that CV-tuned LICORS outperforms standard methods in forecasting challenging spatio-temporal dynamics. Our work provides applied researchers with a new, highly automatic method to analyze and forecast spatio-temporal data.

💡 Research Summary

The paper introduces LICORS (Light‑Cone Reconstruction of States), a fully non‑parametric method for forecasting spatio‑temporal fields where continuous measurements are recorded on a discrete space‑time lattice. The core idea is to exploit the physical notion of a light‑cone: for any location (x, t) the past light‑cone P⁻(x,t) contains all points that could have causally influenced the observation, while the future light‑cone P⁺(x,t) contains all points that the observation could influence. By treating the past light‑cone as the predictor and the future light‑cone as the target, the authors define a “predictive state” as a set of past light‑cones that induce identical conditional distributions over future light‑cones. In other words, a predictive state is a sufficient statistic for the future; once the state is known, the full future distribution is known.

LICORS proceeds in four algorithmic stages. First, past light‑cones are extracted and embedded into a fixed‑dimensional vector space (e.g., via PCA, kernel‑PCA, or random projections). Second, an initial clustering of these vectors is performed (k‑means or Gaussian mixtures). Third, for each cluster the empirical distribution of the associated future light‑cones is estimated; pairwise statistical tests (KL‑divergence, MMD, or other two‑sample tests) compare these distributions. If two clusters are statistically indistinguishable they are merged; if they differ significantly they are split. This iterative refinement converges to a set of clusters that correspond to the true predictive states. Finally, forecasting is done by mapping a new past light‑cone to its nearest state and sampling or taking the mean of that state’s stored future distribution.

The authors prove a consistency theorem: under mild assumptions (finite‑radius Markov property, continuity of the observation density, and an asymptotically large sample size) the algorithm recovers both the number of predictive states and their conditional future distributions with probability tending to one. Thus, LICORS offers a rare combination of non‑parametric flexibility and statistical guarantees.

Hyper‑parameters—light‑cone radii, embedding dimension, initial number of clusters, and the significance level of the distributional tests—are selected automatically via K‑fold cross‑validation. The CV objective jointly penalizes prediction error (e.g., mean‑squared error) and model complexity (e.g., number of states, BIC), thereby avoiding over‑fitting while adapting the model’s granularity to the data.

Empirical evaluation includes two synthetic benchmarks and a real‑world climate dataset. The first benchmark simulates a nonlinear wave equation with multiple interacting modes; the second uses a complex cellular automaton that exhibits long‑range, non‑stationary patterns. In both cases, LICORS (with CV‑tuned settings) outperforms standard baselines such as ARIMA, VAR, Gaussian Process regression, and ConvLSTM, achieving 15–30 % lower MSE. Notably, the learned predictive states are interpretable: for the wave model they correspond to distinct phase‑velocity regimes, and for the cellular automaton they align with recognizable pattern classes. In the climate application (daily temperature and precipitation over the US Midwest), LICORS yields more accurate one‑ and two‑week forecasts than spatial regression models, and the inferred states map cleanly onto climatologically meaningful regimes (“hot‑dry”, “cold‑wet”, etc.). Analysis of the state transition matrix further reveals seasonal dynamics.

The paper also discusses limitations and future directions. Fixed light‑cone radii may miss very long‑range dependencies, and the dimensionality reduction step can discard subtle information. The authors suggest extensions such as adaptive radii, deep‑learning based embeddings, and online updating to handle streaming data. Overall, LICORS provides a principled, automatic framework that leverages causal geometry to perform non‑parametric spatio‑temporal forecasting with provable consistency and strong empirical performance, representing a significant methodological advance for researchers dealing with complex space‑time data.