Abstract
This paper introduces a spatiotemporal resonator model and an inference method for detection and estimation of nearly periodic temporal phenomena in spatiotemporal data. The model is derived as a spatial extension of a stochastic harmonic resonator model, which can be formulated in terms of a stochastic differential equation. The spatial structure is included by introducing linear operators, which affect both the oscillations and damping, and by choosing the appropriate spatial covariance structure of the driving time-white noise process. With the choice of the linear operators as partial differential operators, the resonator model becomes a stochastic partial differential equation, which is compatible with infinite-dimensional Kalman filtering. The resulting infinite-dimensional Kalman filtering problem allows for a computationally efficient solution as the computational cost scales linearly with measurements in the temporal dimension. This framework is applied to weather prediction and to physiological noise elimination in functional magnetic resonance imaging brain data.
Introduction
Many natural phenomena exhibit "quasi-periodic" behavior—patterns that repeat with slight variations over time. Examples range from the diurnal cycles of global temperatures to the rhythmic physiological signals of the human heart and lungs. Modeling these signals across both space and time (spatiotemporal data) presents a significant computational challenge known as the "curse of dimensionality."
Standard statistical tools like Gaussian Processes (GPs) are highly effective at capturing these structures but suffer from cubic scaling ($O(T^3)$) relative to the number of time steps. This makes them impractical for large-scale datasets. Conversely, discrete models like Vector Autoregressive Moving Average (VARMA) models often fail to generalize to new spatial locations. This research proposes a Spatio-Temporal Resonator Model that treats space and time as continuous while maintaining the computational efficiency of state-space models.
Methodology
The Spatio-Temporal Resonator Model
The core innovation lies in framing the resonator as a Stochastic Partial Differential Equation (SPDE). The authors begin with a stochastic harmonic resonator—a standard differential equation for a noisy pendulum—and extend it into the spatial domain by replacing scalar coefficients with linear operators, such as the Laplacian $\nabla^2$.
Infinite-Dimensional Bayesian Filtering
By formulating the model as an SPDE, the researchers can apply an infinite-dimensional version of the Kalman Filter and Rauch–Tung–Striebel (RTS) smoother. To make this computationally tractable, the infinite-dimensional problem is projected onto a finite set of basis functions (eigenfunctions of the Laplacian) using a Hilbert space approximation. This transformation results in a finite-dimensional state-space model that scales linearly ($O(T)$) with time, allowing for the processing of millions of data points.
Applications and Results
Global Weather Modeling
The framework was tested on the NESDIS dataset, comprising over 5.6 million hourly temperature readings from 11,344 stations globally for July 2011. The model successfully decomposed the data into a "slow bias" representing long-term trends and an "oscillatory component" representing the daily cycle.
As shown in the figure above, the model accurately captured the summer heat in the Northern Hemisphere and the day-night variations. Notably, the model provides uncertainty estimates (hatched zones) in regions with sparse data, demonstrating its robustness as a spatial interpolator.
Physiological Noise in fMRI
The method was also applied to a 30-second fast-sampled fMRI scan to detect and isolate physiological noise. The filter successfully separated cardiac and respiratory signals from the underlying neural data.
The resulting "cardiac map" identified strong oscillations near major cerebral arteries, while the "respiratory map" highlighted artifacts near the eyes caused by breathing-induced movement. This separation is critical for improving the signal-to-noise ratio in neuroimaging studies.
Discussion and Future Work
The primary advantage of this framework is its efficiency; it enables the application of Gaussian Process-like modeling to datasets that were previously too large to handle. By "baking" physical properties like damping and wave propagation into the linear operators, the model naturally aligns with the underlying physics of the phenomena.
However, limitations remain. The truncated basis expansion used for computational speed can struggle with very fine-grained spatial variations. Furthermore, the current model assumes spatial stationarity—that the rules of oscillation do not change across the domain. Future research may explore non-linear oscillators (e.g., the Van der Pol oscillator) to model more complex biological rhythms and non-stationary spatial kernels to better reflect real-world heterogeneity.
Summary updated