A Kernel-Based Nonparametric Test for Conditional Independence of Functional Data
Yin Tang, Bing Li · 2026
Read A Kernel-Based Nonparametric Test for Conditional Independence of Functional Data by Yin Tang, Bing Li (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-13704.txt.
Introduction
Yin Tang and Bing Li build the first test for conditional independence of random functions. The idea is core to dimension reduction, causal inference, and graphical models. All prior tests work only for multivariate data, not for data in the form of functions. Their test rests on the conjoined conditional covariance operator, and they derive its limit law. A statistic comes from the eigen split of that operator. They show the method on activity, biometrics, and macro-level data.
Catalog record
| CREATOR | Yin Tang, Bing Li |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 105496 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.13704
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Kernel-Based Nonparametric Test for Conditional Independence of Functional Data Yin Tang, Bing Li Abstract Conditional independence is a fundamental concept in many areas of statistical research, including, for example, sufficient dimension reduction, causal inference, and statistical graphical models. In many modern applications, data arise in the form of random functions, making it important to determine whether two random functions are conditionally independent given a third. However, to the best of our knowledge, existing conditional independence tests in the literature apply only to multivariate data, and extensions to the functional setting are not available. To fill this gap, we develop a kernel-based test for conditional independence of random functions based on the conjoined conditional covariance operator (CCCO). We rigorously derive the asymptotic distribution of the CCCO estimator using a recently established sharpened convergence rate for the regression operator ( Choi et al. 2026 ) . Based on this result, we construct a test statistic using the spectral decomposition of the operator appearing in the asymptotic distribution. The proposed method is illustrated through applications to an activity and biometrics dataset and a macroeconomic dataset. Keywords : conditional independence, functional data, conjoined conditional covariance operator, reproducing kernel Hilbert space 1 Introduction Conditional independence is a fundamental concept underlying a wide range of multivariate data analysis methods. For instance, in Sufficient Dimension Reduction, we need to find \beta such that Y\perp\!\!\!\perp X|\beta^{\mathsf{T}}X ( Li 2018b ) . In graphical models, nodes i and j are disconnected if and only if \displaystyle X_{i}\perp\!\!\!\perp X_{j}|\{X_{1},\ldots,X_{p}\}\setminus\{X_{i},X_{j}\}, where X_{1},\ldots,X_{p} represent the random variables on nodes 1,\ldots,p . See Lauritzen 1996 and Koller and Friedman 2009 . In causal inference, the common cause principle indicates that two dependent random variables X and Y may have a common cause Z such that X\perp\!\!\!\perp Y|Z ( Peters et al. 2017 ) . Conditional independence plays a central role in all three applications. It is thus of great importance to test whether two random variables are conditionally independent given a third. Li and Fan 2020 provides a comprehensive review of nonparametric conditional independence (CI) tests for continuous variables. For example, under the Gaussian assumption, partial correlation and conditional correlation can be used to characterize the conditional dependency between two random variables. In fact, the Gaussian assumption can be relaxed as indicated in Baba et al. 2004 . Representative works in this area include the discretization-based test ( Huang 2010 ) , metric-based tests ( Su and White 2008 ; Huang et al. 2016 ) , the permutation-based two-sample test ( Doran et al. 2014 ) , and the mutual independence test under transformations ( Cai et al. 2022 ) . Also, some regression-based methods can be employed for this purpose under some additivity or linearity assumptions (see, for example, Zhang et al. 2017 ). To further relax modeling assumptions, kernel-based methods have recently gained prominence by embedding probability distributions into feature spaces induced by kernels. Within the reproducing-kernel Hilbert space (RKHS) framework, conditional independence can be characterized via the Conjoined Conditional Covariance Operator (CCCO); see, for example, Fukumizu et al. 2004 ; Fukumizu et al. 2007b ; Lee et al. 2016 . Building on this characterization, Zhang et al. 2011 proposed a kernel-based conditional independence test whose statistic is closely related to the Hilbert–Schmidt norm of the CCCO introduced in Fukumizu et al. 2007b . Related developments include Huang et al. 2022 , which introduces a kernel partial correlation coefficient as a measure of conditional independence, and Sheng and Sriperumbudur 2023 , which investigates the connection between distance-based and kernel-based measures of conditional independence. In this paper, we extend kernel-based conditional independence tests to the functional data setting, where the three random elements involved in the test are random functions or vectors of random functions. Such functional observations are increasingly prevalent in modern data analysis, particularly with the rapid growth of applications involving smart wearable devices. See, for example, Ramsay and Silverman 2007 and Acar‐Denizli and Delicado 2025 . A flexible conditional independence testing procedure would constitute a valuable addition to the toolbox of functional data analysis, enabling the study of complex interconnections among functional variables. Such interconnections arise in a wide range of problems, including function-on-function regression ( Sun et al. 2018 ; Sang and Li 2026 ) , functional graphical models ( Qiao et al. 2019 ; Li and Solea 2018 ) , and functional sufficient dimension reduction ( Ferré and Yao 2003 ; Li and Song 2017 ; Li and Song 2022 ) . In addition to the importance of conditional independence testing for functional data, we would like to highlight a particularly important theoretical contribution of our extension. As the pioneering work on kernel-based conditional independence testing, Zhang et al. 2011 provided only a sketch of the proof for the asymptotic null distribution, without a detailed list of conditions under which the asymptotic distribution holds. A key difficulty arises from the fact that the test statistic involves the inverse of a covariance operator. In practice, the inverse must be estimated using Tikhonov regularization, which slows down the convergence rate of a related regression operator. Until recently, the best known convergence rate for this operator was \displaystyle\epsilon_{n}^{\beta\wedge 1}+n^{-1/2}\epsilon_{n}^{-1}, (1) where n is the sample size, \epsilon_{n} is the Tikhonov regularization parameter, and \beta>0 characterizes the smoothness of the regression relationship between the two random variables involved. See, for example, Li and Song 2017 . The optimal rate implied by ( 1 ) is n^{-1/4} , which can be achieved when \beta\geq 1 with the optimal tuning \epsilon_{n}\asymp n^{-1/4} . However, our calculations (see Section 5 ) indicate that this rate is not sufficiently fast to guarantee the asymptotic distribution claimed in Zhang et al. 2011 . Only recently has the convergence rate of the regression operator been sharpened to \displaystyle n^{-1/2}\epsilon_{n}^{(\beta\wedge 1)-1}+\epsilon_{n}^{\beta\wedge 1}+n^{-1}\epsilon_{n}^{-(3\alpha+1)/(2\alpha)}+n^{-1/2}\epsilon_{n}^{-(\alpha+1)/(2\alpha)}, (2) where \alpha>1 characterizes the decay rate of the eigenvalues of the covariance operator; see Choi et al. 2026 . As shown in Theorem 2 , if \displaystyle\beta>\frac{\alpha-1}{2\alpha},\qquad\frac{\alpha(\beta\wedge 1)}{2\alpha(\beta\wedge 1)+\alpha+1}>\frac{1}{4},\qquad\epsilon_{n}\asymp n^{-\frac{\alpha(\beta\wedge 1)}{2\alpha(\beta\wedge 1)+\alpha+1}}, then the convergence rate in ( 2 ) is faster than n^{-1/4} , which is sufficient to establish the asymptotic distribution of the test statistic in Zhang et al. 2011 . Although our analysis is carried out in the functional setting, the above result also applies to the multivariate setting. Consequently, Theorem 2 provides the first rigorous justification of the asymptotic null distribution for the kernel-based conditional independence test. Given the widespread use of this test, establishing its asymptotic validity is both important and significant. The rest of the paper is organized as follows. In Section 2 , we review the background on RKHS and several linear operators needed for constructing the test statistic at the population level. In Section 3 , we introduce the test statistic for conditional independence. In Section 4 , we rigorously develop the asymptotic distribution of the proposed kernel-based test statistic for conditional independence for functional data. We also carefully discuss the key conditions needed in the asymptotic development. In Section 5 , we introduce practical methods for approximating random functions involved in constructing the test statistic. Section 6 presents the sample-level implementation, including estimation of the eigenvalues appearing in the asymptotic distribution using an acceleration method,