A Kernel-Based Nonparametric Test for Conditional Independence of Functional Data

1. Problem Statement

The paper addresses the challenge of testing conditional independence (CI) in the context of functional data.

• Context: In modern applications (e.g., wearable devices, macroeconomic time series), data often arrive as random functions (curves) rather than finite-dimensional vectors.
• The Gap: While nonparametric CI tests exist for multivariate data (e.g., Zhang et al., 2011), there were no rigorous extensions to the functional setting. Existing methods could not handle the infinite-dimensional nature of random functions $X(t), Y(t), Z(t)$.
• Objective: To develop a kernel-based test to determine if two random functions $X$ and $Y$ are conditionally independent given a third random function $Z$ (i.e., $X \perp\!\!\!\perp Y \mid Z$).

2. Methodology

2.1 Theoretical Framework: Reproducing Kernel Hilbert Spaces (RKHS)

The authors utilize the RKHS framework to characterize conditional independence without parametric assumptions.

• Conjoined Conditional Covariance Operator (CCCO): The core of the method is the construction of a modified operator, denoted as $\Sigma_{\ddot{X}\ddot{Y}|Z}$.
  • Standard Conditional Covariance Operators (CCO) are insufficient for testing CI because they only capture linear conditional dependencies.
  • The authors define $\ddot{X} = (X, Z)$ and $\ddot{Y} = (Y, Z)$ and construct the CCCO as:
    $$\Sigma_{\ddot{X}\ddot{Y}|Z} = \Sigma_{\ddot{X}\ddot{Y}} - \Sigma_{\ddot{X}Z}\Sigma_{ZZ}^{\dagger}\Sigma_{Z\ddot{Y}}$$
    where $\Sigma^{\dagger}$ denotes the Moore-Penrose inverse.
• Theoretical Equivalence: Theorem 1 establishes that under specific assumptions (characteristic kernels and smoothness conditions), $\Sigma_{\ddot{X}\ddot{Y}|Z} = 0$ if and only if $X \perp\!\!\!\perp Y \mid Z$.

2.2 Test Statistic Construction

• Estimation: The population operator is estimated using sample data. Due to the singularity of the empirical covariance operator $\hat{\Sigma}_{ZZ}$, Tikhonov regularization is applied: $(\hat{\Sigma}_{ZZ} + \epsilon_n I)^{-1}$.
• Statistic: The test statistic is based on the squared Hilbert-Schmidt norm of the estimated CCCO:
  $$T_n = n \|\hat{\Sigma}_{\ddot{X}\ddot{Y}|Z}\|_{HS}^2$$
  Under the null hypothesis ($H_0: X \perp\!\!\!\perp Y \mid Z$), $T_n$ should be small; under the alternative, it should be large.

2.3 Asymptotic Theory (Key Theoretical Contribution)

A major hurdle in previous work (Zhang et al., 2011) was the lack of rigorous proof for the asymptotic distribution of the test statistic, primarily due to the slow convergence rate of the inverse covariance operator in functional settings.

• Convergence Rate Improvement: The authors leverage a recently established sharpened convergence rate for regression operators (Choi et al., 2026).
• Conditions: They derive specific conditions on the smoothness parameter $\beta$ and the eigenvalue decay rate $\alpha$ of the covariance operator. Specifically, if $\beta > \frac{\alpha-1}{2\alpha}$ and the regularization parameter $\epsilon_n$ is tuned optimally, the convergence rate of the estimator exceeds $n^{-1/4}$.
• Result: This faster rate ensures that the error terms vanish sufficiently fast, allowing the derivation of the asymptotic distribution:
  $$n \|\hat{\Sigma}_{\ddot{X}\ddot{Y}|Z}\|_{HS}^2 \xrightarrow{d} \sum_{k=1}^{\infty} \lambda_k Z_k^2$$
  where $\lambda_k$ are eigenvalues of the limiting operator and $Z_k$ are i.i.d. standard normals. This provides the first rigorous justification of the asymptotic null distribution for kernel-based CI tests, applicable to both functional and multivariate data.

2.4 Practical Implementation

• Curve Reconstruction: Since functional data are observed at discrete points, the authors use RKHS smoothers (with Tikhonov regularization) to reconstruct continuous trajectories.
• Inner Product Approximation: Numerical integration (Simpson's rule) is used to approximate inner products and distances required for the second-layer kernels.
• Tuning Parameters: Bandwidths for Gaussian kernels and regularization parameters ($\delta_n$ for curve smoothing, $\epsilon_n$ for CCCO estimation) are selected via Generalized Cross-Validation (GCV).
• Computational Acceleration: To handle the high computational cost of eigen-decomposition for $n^2 \times n^2$ matrices, the authors employ a rank-reduction algorithm (based on Tang and Li, 2024) that reduces the problem to finding eigenvalues of an $n \times n$ matrix.

3. Key Contributions

1. Extension to Functional Data: The first nonparametric kernel-based test for conditional independence specifically designed for random functions.
2. Rigorous Asymptotic Justification: The paper resolves a long-standing theoretical gap by providing a complete proof of the asymptotic null distribution, correcting the incomplete proof in Zhang et al. (2011). This relies on new convergence rates for regression operators in RKHS.
3. CCCO Characterization: The formal introduction and proof of the Conjoined Conditional Covariance Operator as a valid measure for functional conditional independence.
4. Algorithmic Efficiency: Development of a computationally feasible algorithm for large-scale functional data, including curve reconstruction and accelerated eigenvalue computation.

4. Results

4.1 Simulation Studies

• Setup: The authors simulated five models with varying dependence structures (additive, multiplicative, and non-linear transformations) and sample sizes ($n=50$ to $500$). Both balanced and unbalanced observation schemes were tested.
• Findings:
  • Type I Error: Under the null hypothesis (Model 1), p-values were uniformly distributed, with most exceeding 0.05, indicating correct size control.
  • Power: Under alternative hypotheses (Models 2–5), the test demonstrated high power, with p-values dropping significantly below 0.05 as sample size increased.
  • Robustness: The method performed consistently well under both balanced and unbalanced observation schedules.

4.2 Real Data Applications

• WISDM Dataset (Activity Recognition):
  • Task: Test if x- and y-axis accelerations are conditionally independent given the z-axis acceleration for various activities.
  • Result: For dynamic activities (walking, jogging, stairs), the null hypothesis was not rejected (p > 0.05), implying the z-axis captures the relationship between x and y. For sitting, the null was rejected (p < 0.05), indicating a residual dependency between x and y not explained by z, consistent with physical observations of body sway.
• WDI Dataset (Macroeconomics):
  • Task: Test if pairs of socioeconomic indicators (e.g., life expectancy, inflation, unemployment) are conditionally independent given GDP.
  • Result: A conditional dependency graph was constructed. Notably, Life Expectancy was found to be conditionally dependent on most other factors even after controlling for GDP, whereas the Real Interest Rate appeared conditionally independent of others given GDP.

5. Significance

This paper significantly advances the field of functional data analysis and causal inference.

• Theoretical Impact: By rigorously establishing the asymptotic distribution, it validates the use of kernel-based CI tests in high-dimensional and infinite-dimensional settings, correcting previous theoretical oversights.
• Practical Utility: It provides a powerful, nonparametric tool for researchers dealing with complex functional data (e.g., medical signals, climate data, economic time series) to uncover hidden conditional dependencies that linear or parametric methods might miss.
• Broad Applicability: The theoretical results regarding convergence rates and asymptotic distributions are not limited to functional data but also strengthen the theoretical foundation of existing multivariate CI tests.