Nonparametric regression of spatio-temporal data using infinite-dimensional covariates

1. Problem Statement

The paper addresses a critical gap in spatio-temporal analysis where data is collected at irregular time intervals and varying spatial locations (e.g., sensor failures, dynamic deployment), while the explanatory variables (covariates) are infinite-dimensional or functional in nature.

• The Challenge: Existing methods typically fail in this setting because:
  • Standard functional data analysis assumes independent and identically distributed (i.i.d.) data, ignoring temporal dependence.
  • Standard functional time series methods assume regular time grids and fixed spatial designs.
  • Nonparametric spatio-temporal regression usually relies on finite-dimensional covariates and regular grids.
• The Specific Context: The authors aim to model a response $Y_t(s)$ (e.g., air pollution levels) at time $t$ and location $s$, driven by a covariate process $X_t$ that is infinite-dimensional (e.g., historical time-series of multiple pollutants across all locations) and exhibits complex temporal dependence.
• The Model:
  $$Y_t(s) = \mu(X_t, s) + \sigma(X_t, s)\epsilon_t(s)$$
  Where $\mu$ and $\sigma$ are unknown smooth functions, and $\epsilon_t(s)$ represents spatially correlated but temporally independent errors.

2. Methodology

A. Mathematical Framework & Assumptions

• Infinite-Dimensional Covariates: The covariate $X_t$ takes values in an abstract space $\chi \subseteq \mathbb{R}^\infty$. The authors assume $\chi$ is a subspace of a weighted $L^2$ Banach space to define scaled norms.
• Dependence Structure (The Key Innovation): Instead of relying on standard mixing conditions (like $\alpha$-mixing or $\beta$-mixing), which often fail for infinite-dimensional processes with infinite-order lags, the authors introduce a Polynomially Decaying Moment Contraction (PMC) condition.
  • $X_t$ is modeled as a functional of past i.i.d. innovations ($X_t = G_t(\dots, Z_{t-1}, Z_t)$).
  • The dependence decays polynomially: $\theta^2(m) = \mathcal{O}(m^{-\tau})$ for $\tau > 2$. This is a weaker condition than Geometric Moment Contraction (GMC) and allows for processes that are not mixing.
• Spatial Correlation: The error term $\epsilon_t(s)$ has a spatial covariance structure $\rho(s_1, s_2)$ that decays polynomially with distance.
• Series Expansion: The mean and variance functions are decomposed using a countable orthonormal basis $\{b_k(s)\}$ (e.g., splines, wavelets, or polynomials):
  $$\mu(x, s) = \sum_{k=1}^\infty \mu_k(x)b_k(s), \quad \sigma(x, s) = \sum_{k=1}^\infty \sigma_k(x)b_k(s)$$

B. Estimation Procedure

The estimation is a two-stage process:

1. Spatial Aggregation (Modified Monte-Carlo):
   Since $Y_t(s)$ is observed at irregular locations, the continuous integral $\int Y_t(s)b_k(s)ds$ cannot be computed directly. The authors propose a Modified Monte-Carlo procedure:

   • The spatial domain is divided into a grid of hypercubes based on the "effective spatial resolution" (density of observed points).
   • A representative point is chosen for each grid cell.
   • The integral is approximated by a weighted sum over these representative points, ensuring uniform coverage and controlling covariance bias.

2. Kernel Regression:
   Once the spatially aggregated response $Y^*_{tk}$ is estimated, the problem reduces to estimating the component functions $\mu_k(x)$ in a nonlinear regression model with infinite-dimensional covariates.

   • A Nadaraya-Watson type kernel estimator is used, adapted for the infinite-dimensional space using a scaled distance metric and a bandwidth matrix.
   • The estimator for the $k$-th component is:
     $$\hat{\mu}_k(x) = \frac{\sum_{i=1}^n \mathcal{K}(H_n^{-1}(x - X_{t_i})) \hat{Y}^*_{t_i k}}{\sum_{i=1}^n \mathcal{K}(H_n^{-1}(x - X_{t_i}))}$$

C. Inference

• Asymptotic Normality: The authors derive the asymptotic distribution of the estimator, proving it converges to a normal distribution after appropriate normalization.
• Simultaneous Confidence Bands: Using extreme value theory (Gumbel distribution), they construct simultaneous confidence intervals for the entire mean function $\mu(x, s)$ over a dense subset of the covariate space and spatial domain.

3. Key Contributions

1. Novel Model Framework: Proposes the first nonparametric regression model for spatio-temporal data that simultaneously handles irregular sampling (in time and space) and infinite-dimensional covariates.
2. Weaker Dependence Assumptions: Replaces the standard, often uncheckable, mixing conditions with the PMC condition. This allows the theory to cover processes with infinite-order lags and polynomial decay of dependence, which are common in functional time series but excluded by mixing assumptions.
3. Reduced Moment Requirements: The methodology requires only second-order stationarity for the covariates and does not demand higher-order moments (e.g., 4th or $(2+\delta)$-th moments) that previous infinite-dimensional regression studies required.
4. Simultaneous Inference: Provides a rigorous framework for constructing simultaneous confidence bands, enabling hypothesis testing for the shape of the mean surface, not just point-wise inference.

4. Results

Theoretical Results

• Consistency: The estimators for the mean components $\hat{\mu}_k(x)$ and the full surface $\hat{\mu}(x, s)$ are proven to be consistent ($o_P(1)$) as the sample size $n \to \infty$.
• Asymptotic Normality: The normalized estimator converges to a standard normal distribution.
• Variance Estimation: A consistent estimator for the covariance operator is derived, enabling the construction of confidence bands.

Empirical Applications

1. Air Pollution in Delhi:

   • Data: PM2.5 measurements from 38 stations in Delhi (2015–2020) with irregular missingness.
   • Covariates: Infinite-dimensional history of multiple pollutants across all stations.
   • Findings: The model successfully captured spatial heterogeneity (higher pollution in central Delhi vs. suburbs) and temporal patterns (seasonal variations). It demonstrated that central stations require fewer spatial covariates for accurate prediction compared to peripheral stations, likely due to land-use homogeneity.
   • Validation: Sensitivity analysis showed the model performs equally well for in-training and out-of-training locations (Diebold-Mariano test).

2. Soccer Shot Analysis (Supplementary):

   • Data: Expected Goals (xG) for Arsenal, Liverpool, and Manchester City.
   • Covariates: Team performance metrics over the season.
   • Findings: The model identified distinct spatial attacking profiles. For instance, Arsenal's 2024/25 season showed a broader region of moderate xG outside the penalty box (long-range shooting), correlating with their improved league standing, while Liverpool showed a concentrated high-xG pattern inside the box.

3. Simulation Studies:

   • Simulations confirmed that the Mean Absolute Percentage Error (MAPE) decreases as the number of spatial grid points increases, validating the consistency of the spatial aggregation method.
   • The estimator remained robust even when covariates were lagged or partially observed.

5. Significance

This paper bridges the gap between functional data analysis and spatio-temporal statistics. By relaxing the dependence assumptions to PMC, it opens the door for analyzing complex, real-world datasets (like environmental monitoring or high-frequency financial data) where:

• Sensors fail or move (irregular spatial sampling).
• Data arrives at irregular times.
• Predictors are entire curves or high-dimensional histories.

The ability to provide simultaneous confidence bands without strong mixing assumptions makes this a powerful tool for rigorous statistical inference in modern, complex data environments.