Effective Sample Size for Functional Spatial Data

Technical Summary: Effective Sample Size for Functional Spatial Data

Problem Statement
The concept of Effective Sample Size (ESS) is a well-established metric in geostatistics for scalar random fields, quantifying the amount of independent information in a dataset by accounting for redundancy caused by spatial correlation. However, the extension of ESS to functional spatial data—where observations are functions indexed by spatial location—remains largely unexplored. Existing methodologies for scalar data cannot directly address the dual complexity of functional spatial datasets: the intrinsic dependence structure within the functional curves (e.g., smoothness over time or depth) and the spatial dependence between observations at different locations. Without a formal definition, researchers lack a tool to quantify redundancy or determine the effective number of independent function-valued observations in such datasets.

Methodology
The authors propose a novel definition of functional ESS within a Hilbert space framework, specifically $L^2([0, 1])$. The methodology proceeds through the following steps:

1. Theoretical Framework: The authors define a functional random field $\{\chi_s : s \in \mathbb{R}^d\}$ where each observation is a square-integrable function. They assume weak stationarity and isotropy, describing spatial dependence via a covariance operator.
2. Trace-Covariogram: To reduce the infinite-dimensional covariance operator to a tractable scalar summary, the authors utilize the trace-covariogram, defined as $\sigma_{tr}(h) = \int_0^1 \sigma_h(t, t) dt$. This scalar measure captures the aggregate covariance between functional observations separated by distance $h$.
3. Definition of Functional ESS: Adapting Griffith's (2005) scalar definition, the functional ESS ($ESS_F$) is defined as:
   $$ESS_F = \frac{n^2 \sigma_{tr}(0)}{\sum_{i=1}^n \sum_{j=1}^n \sigma_{tr}(\|s_i - s_j\|)}$$
   where $n$ is the nominal sample size and $\sigma_{tr}(0)$ is the trace-variance.
4. Estimation: In practice, the trace-covariogram is unknown. The authors propose estimating it via an empirical trace-variogram (using the relationship $\gamma_{tr}(h) = \sigma_{tr}(0) - \sigma_{tr}(h)$) and fitting a parametric model (e.g., Exponential, Spherical, Gaussian) using a least-squares approach. A plug-in estimator is then computed using the fitted parameters.
5. Theoretical Properties: The paper establishes that under non-negative trace-covariograms, $1 \leq ESS_F \leq n$. It further demonstrates that $ESS_F$ behaves intuitively: it equals $n$ under independence and $1$ under perfect positive correlation. Additionally, the authors show that adding an independent observation strictly increases the functional ESS.

Key Contributions and Illustrative Results

• Theoretical Extension: The primary contribution is the formal definition of ESS for functional spatial data, bridging the gap between functional data analysis (FDA) and geostatistics.
• FAR(1) Model Analysis: To illustrate the behavior of the proposed measure, the authors analyze a Functional Autoregressive process of order 1 (FAR(1)). They derive a closed-form expression showing that the functional ESS is the weighted harmonic mean of the marginal ESSs of the scalar processes projected onto the eigenfunctions of the covariance operator. Numerical studies demonstrate that:
  • Stronger temporal dependence (higher autoregressive coefficients) reduces the functional ESS.
  • A slower decay in noise variance across eigen-directions (increasing the number of active components) increases the functional ESS.
• Real Data Application: The methodology is applied to a dataset of geometric vertical velocities in the Pacific Ocean (NCEP GODAS data), consisting of 600 spatial locations with velocity profiles measured at 22 depth levels.
  • Three parametric models (Exponential, Spherical, Gaussian) were fitted to the empirical trace-variogram.
  • The estimated $ESS_F$ values ranged from approximately 42 to 105, suggesting that only 17% to 25% of the original sample is required to capture the essential statistical information.
  • Validation: The authors validated these findings by constructing functional boxplots and comparing the full dataset against random subsamples of size 106 (the maximum estimated ESS). Numerical metrics (median discrepancies, central region width differences, and inclusion proportions) confirmed that the subsamples closely resembled the full dataset, visually and statistically supporting the redundancy quantification.

Significance and Claims
The paper claims that the proposed functional ESS provides an interpretable and meaningful measure for quantifying redundancy in functional spatial datasets. Its significance lies in its ability to:

1. Guide Subsampling: It offers a principled method for selecting a reduced set of observations, which is particularly valuable in practical settings where data collection is costly or sparse (e.g., oceanographic monitoring).
2. Preserve Information: The authors demonstrate that selecting a fraction of the data (e.g., 17% in the ocean velocity case) can retain the key statistical characteristics of the complete dataset, including functional location and dispersion.
3. Handle Functional Structure: The approach naturally accommodates non-stationarities along the functional domain (e.g., varying velocity fluctuations with ocean depth) by representing profiles as functions rather than discrete vectors.

The authors conclude by noting that while the current work addresses isotropic univariate functional fields, future research directions include extending the framework to multivariate spatial functional fields (accounting for cross-correlations), anisotropic fields, and developing uncertainty quantification (asymptotic properties) for the estimator.