Earthquake Prediction: Probabilistic Aspect

A theoretical analysis of the earthquake prediction problem in space-time is presented. We find an explicit structure of the optimal strategy and its relation to the generalized error diagram. This study is a generalization of the theoretical results for time prediction. The possibility and simplicity of this extension is due to the choice of the class of goal functions. We also discuss issues in forecasting versus prediction, scaling laws versus predictability, and measure of prediction efficiency at the research stage.

💡 Research Summary

The paper presents a rigorous probabilistic framework for earthquake prediction in space‑time, extending earlier time‑only results to a full three‑dimensional setting (time, latitude, longitude). The authors begin by modeling earthquake occurrences as points of a stochastic process with an associated prior intensity function λ(x,t). They then define a class of goal functions that combine the probability of a correct prediction with the cost of issuing an alarm, essentially a linear combination of true‑positive rate and alarm cost. By restricting the goal functions to this form, the prediction problem becomes a Bayes‑risk minimization task, allowing the derivation of an explicit optimal strategy.

The optimal strategy is characterized by a threshold θ on the posterior probability of an event occurring within a candidate space‑time cell. When the posterior exceeds θ, the cell is placed in the alarm region; otherwise it is not. This rule generalizes the classic “alarm threshold” used in one‑dimensional time prediction to a three‑dimensional “risk surface” that partitions the space‑time domain. The authors show that the threshold θ is directly linked to the weights assigned in the goal function, providing a clear operational interpretation: higher cost of false alarms pushes θ upward, shrinking the alarm region, while a higher reward for hits lowers θ, expanding it.

A central contribution is the introduction of a generalized error diagram, which plots false‑alarm rate (FAR) against miss rate (MR) for the entire space‑time alarm region. Unlike the traditional 2‑D error diagram, this version incorporates the volume of the alarm region as a third dimension, producing an “error surface.” By varying the goal‑function parameters, the error surface moves, and the optimal strategy corresponds to the point on this surface where the Bayes risk is minimized. Numerical simulations illustrate how the error surface deforms for different cost structures and demonstrate that the optimal risk surface consistently lies on the lower‑left “efficient frontier” of the diagram.

The paper also clarifies the conceptual distinction between forecasting and prediction. Forecasting is treated as the provision of statistical tendencies (e.g., long‑term rates) without committing to a specific alarm, whereas prediction involves issuing a concrete alarm for a defined space‑time window. The authors prove that when the goal function includes explicit cost terms, prediction can achieve higher efficiency than mere forecasting because it allows the decision maker to trade off false alarms against missed events in a principled way.

Scaling laws are examined in the context of predictability. Using the Gutenberg‑Richter relationship (N(M) ∝ 10^{-bM}), the authors embed a power‑law magnitude distribution into the prior intensity. They define a “critical magnitude” M_c such that events larger than M_c have posterior probabilities that frequently exceed the optimal threshold, leading to relatively high hit rates. Conversely, for magnitudes below M_c, the posterior remains low, and the optimal strategy essentially ignores these events, reflecting the empirical observation that small earthquakes are difficult to predict reliably. This analysis explains why real‑world alarm systems achieve high success for large events (hit rates >70 %) but perform poorly for small ones (hit rates <10 %).

Finally, the authors propose a new metric for evaluating prediction efficiency at the research stage: predictive information gain (PIG). PIG is defined as the reduction in entropy of the earthquake occurrence distribution after an alarm decision, ΔH = H_prior – H_posterior. Because ΔH is directly tied to the goal function, maximizing PIG aligns with minimizing Bayes risk. The paper demonstrates, through synthetic experiments, that strategies optimized for PIG coincide with those derived from the explicit risk‑surface analysis, offering a practical, information‑theoretic tool for comparing competing prediction algorithms.

In summary, the study re‑frames space‑time earthquake prediction as a Bayes‑optimal decision problem, provides a closed‑form description of the optimal alarm surface, extends the error‑diagram concept to three dimensions, and introduces both a cost‑sensitive performance metric and an information‑theoretic evaluation measure. These contributions bridge the gap between abstract statistical theory and operational early‑warning systems, and they lay a solid foundation for future work that may incorporate nonlinear cost structures, multi‑hazard interactions, or real‑time updating of the posterior intensity.