Abstract
In this paper, we prove that functional sliced inverse regression (FSIR) achieves the optimal (minimax) rate for estimating the central space in functional sufficient dimension reduction problems. First, we provide a concentration inequality for the FSIR estimator of the covariance of the conditional mean. Based on this inequality, we establish the root-$n$ consistency of the FSIR estimator of the image of covariance of the conditional mean. Second, we apply the most widely used truncated scheme to estimate the inverse of the covariance operator and identify the truncation parameter that ensures that FSIR can achieve the optimal minimax convergence rate for estimating the central space. Finally, we conduct simulations to demonstrate the optimal choice of truncation parameter and the estimation efficiency of FSIR. To the best of our knowledge, this is the first paper to rigorously prove the minimax optimality of FSIR in estimating the central space for multiple-index models and general $Y$ (not necessarily discrete).
Introduction
In the era of high-dimensional data, Functional Data Analysis (FDA) has emerged as a critical framework for analyzing data points represented as continuous curves or functions—such as daily temperature fluctuations or high-frequency financial indices. To extract meaningful patterns from these infinite-dimensional objects, researchers employ Sufficient Dimension Reduction (SDR). The goal of SDR is to identify a low-dimensional "central space" that captures all the information in a predictor function $X$ relevant to a response variable $Y$.
Sliced Inverse Regression (SIR) has long been a staple for dimension reduction in multivariate settings. While its functional counterpart, FSIR, was introduced in 2003, its theoretical limits remained poorly understood for decades. Specifically, it was unknown whether FSIR could achieve the minimax convergence rate—the theoretical "speed limit" for any estimation algorithm. This paper bridges that gap, providing the first rigorous proof of FSIR's minimax optimality for multiple-index models and general response variables.
The Challenge of Functional Inversion
The primary difficulty in functional SDR lies in the nature of functional covariance operators. Unlike finite-dimensional matrices, these operators are "compact," making their inversion an ill-posed problem. Standard estimation techniques often fail because small perturbations in the data can lead to massive errors in the estimated inverse. To manage this, researchers typically use regularization or truncation methods, but identifying the mathematically optimal level of truncation has remained an open question.
Methodology: FSIR-OT and the WSSC
The authors introduce an improved framework termed Optimal Truncated FSIR (FSIR-OT). This approach relies on two primary innovations:
- Weak Sliced Stable Condition (WSSC): The authors propose this new, milder stability condition for functional data. By ensuring that the average variance within "slices" of the data remains controlled, the WSSC allows the algorithm to handle continuous response variables $Y$. This is a significant advancement over previous proofs that were restricted to discrete or categorical responses.
- Optimal Truncation Strategy: To stabilize the inversion of the covariance operator, the authors project the data onto the top $m$ eigenfunctions, a process similar to Principal Component Analysis (PCA). The critical contribution here is the identification of the optimal truncation parameter $m$. The authors prove that $m$ should scale with the sample size $n$ according to the relationship $m \propto n^{1/(\alpha+2\beta)}$, where $\alpha$ represents the decay rate of the eigenvalues and $\beta$ represents the smoothness of the underlying link function.
Key Insights: Symmetry and Supervision
A striking discovery in this research is the "Convergence Link." The authors found that the optimal convergence rate for FSIR is $n^{-(2\beta-1)/(\alpha+2\beta)}$. This is identical to the minimax rate found in functional linear regression, suggesting a deep theoretical symmetry between these two classes of functional problems.
Furthermore, the paper highlights a crucial distinction between supervised and unsupervised truncation. While standard PCA-based methods choose $m$ based solely on the variance of the predictors (unsupervised), FSIR-OT is supervised—it considers the relationship with the response $Y$. The authors demonstrate that this supervised approach allows for a much more efficient, lower-dimensional representation, significantly reducing prediction errors compared to traditional methods.
Empirical Validation
The theoretical findings were validated through both synthetic simulations and real-world application:
- Synthetic Models: Testing across linear, polynomial, and exponential models confirmed the theoretical scaling laws. In log-log plots of the optimal $m$ versus sample size $n$, the empirical slope (approximately 0.2) closely matched the theoretical prediction (2/11 for the parameters used).
- Bike Sharing Dataset: Using data from the UCI Machine Learning Repository, the authors modeled daily bike rental counts based on hourly temperature functions. FSIR-OT outperformed existing methods such as Functional Cumulative Slicing Estimation (FCSE) and Regularized FSIR (RFSIR), providing a more parsimonious and accurate model.
Conclusion
By establishing the root-$n$ consistency of the conditional mean covariance estimator and identifying the optimal truncation parameter, this paper provides the missing mathematical foundation for FSIR. It proves that FSIR is not just a useful heuristic, but a fundamentally optimal estimator for functional dimension reduction. Future work may extend these results to scenarios where the structural dimension of the central space is unknown or where the signal-to-noise ratio varies significantly.
Summary updated