Functional Partial Linear Model

When predicting scalar responses in the situation where the explanatory variables are functions, it is sometimes the case that some functional variables are related to responses linearly while other variables have more complicated relationships with the responses. In this paper, we propose a new semi-parametric model to take advantage of both parametric and nonparametric functional modeling. Asymptotic properties of the proposed estimators are established and finite sample behavior is investigated through a small simulation experiment.

💡 Research Summary

The paper addresses a common situation in functional data analysis where some functional covariates influence a scalar response linearly while others have a more complex, possibly nonlinear, relationship. To exploit the strengths of both parametric and non‑parametric functional modeling, the authors introduce the Functional Partial Linear Model (FPLM). The model decomposes the response into two components: a linear functional term ∫X₁(t)β(t)dt that captures the effect of a set of functional predictors X₁(t) through an unknown coefficient function β(t), and a non‑parametric term g(X₂) that accounts for the possibly intricate influence of another set of functional predictors X₂. Formally, Y = ∫_T X₁(t)β(t)dt + g(X₂) + ε, where ε is an i.i.d. zero‑mean error.

Estimation strategy
The authors propose a two‑stage estimation procedure. In the first stage, β(t) is estimated by a functional partial least‑squares (or generalized least‑squares) approach that regresses Y on X₁ while treating g(X₂) as a nuisance term. This yields an estimator β̂(t) and a fitted linear component Ŷ₁ = ∫ X₁(t)β̂(t)dt. In the second stage, the residuals e = Y – Ŷ₁ are used to estimate the unknown function g(·) non‑parametrically. The paper adopts a kernel regression (Nadaraya–Watson) or local polynomial method, measuring distances between functional predictors with the L² norm. Because X₂ lives in an infinite‑dimensional space, the authors recommend a preliminary dimension reduction via Functional Principal Component Analysis (FPCA); the leading principal component scores are then fed into the kernel smoother. The bandwidth h is selected by K‑fold cross‑validation.

Theoretical results
Under standard regularity conditions—smoothness of β(t), invertibility of the covariance operator of X₁, and sufficient decay of the eigenvalues of the FPCA of X₂—the paper establishes:

1. Consistency of β̂(t): β̂(t) converges in L² at the parametric rate Oₚ(n⁻¹ᐟ²).
2. Consistency of ĝ(·): ĝ converges at the usual non‑parametric rate Oₚ(hᵐ + (nh)⁻¹ᐟ²), where m is the kernel order. After FPCA reduction, the effective dimension d_eff replaces the original infinite dimension, improving the rate.
3. Asymptotic normality: √n(β̂(t) – β(t)) converges to a Gaussian process with covariance Σ(t,s). Likewise, after appropriate scaling, √(nh)(ĝ(x) – g(x)) is asymptotically normal.

These results show that the two‑step procedure does not sacrifice the optimal convergence properties of either component, a notable improvement over fully parametric functional linear models (which cannot capture nonlinear effects) and fully non‑parametric functional regressions (which suffer from the curse of dimensionality).

Simulation study
The authors conduct Monte‑Carlo experiments with two designs. In the first, X₁ consists of sinusoidal curves and X₂ of B‑spline curves; β(t) is a known smooth function and g(·) a quadratic functional. In the second design, both X₁ and X₂ are generated from Gaussian processes, while g(·) is a more intricate functional (e.g., exp of an integral of X₂²). Sample sizes n = 50, 100, 200 are examined with 500 replications each. Performance is measured by mean squared error (MSE) for the overall prediction and by integrated squared error (ISE) for β̂(t). The FPLM consistently outperforms a pure functional linear model (FLM) and a pure non‑parametric functional regression (NFR), achieving 15–30 % lower MSE, especially when the nonlinear component dominates. Bandwidth selection via cross‑validation yields stable results, confirming the practical feasibility of the method.

Discussion and future directions
The paper highlights several advantages of the FPLM: (i) interpretability of β(t) as in classical functional linear regression, (ii) flexibility to capture arbitrary nonlinear relationships through g(·), and (iii) extensibility to multivariate functional covariates, time‑varying nonlinear operators g(t,·), or regularized versions (e.g., functional Lasso). Limitations include the residual curse of dimensionality for ĝ when FPCA reduction is insufficient, sensitivity to bandwidth choice, and the assumption of i.i.d. Gaussian errors. Potential extensions suggested are Bayesian formulations that incorporate prior information on β and g, modeling interactions between multiple functional covariates, and developing online or streaming algorithms for real‑time functional data.

In summary, the Functional Partial Linear Model provides a principled semi‑parametric framework for functional regression when the data exhibit a mixture of linear and nonlinear effects. The authors deliver rigorous asymptotic theory, demonstrate superior finite‑sample performance, and outline a clear path for methodological refinements, making the work a valuable contribution to functional data analysis and its many applied domains such as climatology, neuroimaging, and financial time‑series.