[Paper Review] Local Quadratic Estimation of the Curvature in a Functional Single Index Model

Zi Ye, Giles Hooker|arXiv (Cornell University)|Mar 25, 2018

Ecology and Vegetation Dynamics StudiesEnvironmental Science15 references4 citations

TL;DR

This paper proposes a local quadratic estimation method for the curvature (second derivative) of the link function in a functional single index model, where both the coefficient function and link function are unknown. The method uses nested optimization with bandwidth selection, achieving a convergence rate of $ O(h_n^4 + \frac{1}{n h_n^4}) $, and shows that the argument of the link function can be estimated root-$ n $ consistently, though performance is sensitive to initial values and bandwidth choice.

ABSTRACT

The nonlinear effects of environmental variability on species abundance plays an important role in the maintenance of ecological diversity. Nonetheless, many common models use parametric nonlinear terms pre-determining ecological conclusions. Motivated by this concern, we study the estimate of the second derivative (curvature) of the link function g in a functional single index model. Since the coefficient function and the link function are both unknown, the estimate is expressed as a nested optimization. For a fixed and unknown coefficient function, the link function and its second derivative are estimated by local quadratic approximation, then the coefficient function is estimated by minimizing the MSE of the model. In this paper, we derive the rate of convergence of the estimation. In addition, we prove that the argument of g, can be estimated root-n consistently. However, practical implementation of the method requires solving a nonlinear optimization problem, and our results show that the estimates of the link function and the coefficient function are quite sensitive to the choices of starting values.

Motivation & Objective

• To estimate the second derivative (curvature) of the link function in a functional single index model where both the coefficient function and link function are unknown.
• To develop a nested optimization procedure that jointly estimates the coefficient function and the curvature of the link function.
• To derive theoretical convergence rates for the curvature estimator under regularity conditions.
• To investigate the sensitivity of the curvature estimation to initial values and bandwidth selection in practical implementation.
• To provide a framework for assessing ecological responses to environmental variability through nonparametric curvature estimation.

Proposed method

• Uses local quadratic approximation to estimate the second derivative of the link function $ g $ at each design point.
• For a fixed coefficient function $ \beta $, estimates $ g $ and $ g'' $ via local polynomial regression with bandwidth $ h_n $.
• Performs nested optimization: first estimates $ g'' $ for a given $ \beta $, then minimizes the mean squared error (MSE) over $ \beta $ to estimate $ \beta^0 $.
• Employs a bandwidth $ h_n $ that decreases with sample size $ n $, ensuring theoretical convergence.
• Applies cross-validation (10-fold and GCV) to select optimal bandwidths, with a heuristic post-cross-validation adjustment.
• Uses Lipschitz continuity of $ g'' $ and root-$ n $ consistency of $ \int X_i \hat{\beta} $ to bound estimation error.

Experimental results

Research questions

• RQ1 What is the theoretical convergence rate of the local quadratic estimator for the second derivative of the link function in a functional single index model?
• RQ2 How does the estimation of the curvature $ g'' $ depend on the choice of initial values in the nonlinear optimization procedure?
• RQ3 Can the argument $ \int X_i \beta^0 $ be estimated consistently at the root-$ n $ rate despite uncertainty in $ \beta^0 $?
• RQ4 How do different bandwidth selection strategies affect the accuracy of curvature estimation?
• RQ5 What is the impact of rescaling bandwidths and using different starting values on the performance of curvature estimation?

Key findings

• The estimation error for $ g'' $ satisfies $ \frac{1}{n} \sum_{i=1}^n \mathbb{E} \left[ \hat{g}''\left( \int X_i \hat{\beta} \right) - g''\left( \int X_i \beta^0 \right) \right]^2 = O\left( h_n^4 + \frac{1}{n h_n^4} \right) $, establishing the convergence rate.
• The argument $ \int X_i \hat{\beta} $ is estimated at the root-$ n $ rate, enabling consistent curvature estimation under regularity.
• The curvature estimator is highly sensitive to the choice of initial values in the nonlinear optimization, with poor starting values leading to suboptimal solutions.
• Rescaling the bandwidth significantly improves estimation accuracy, especially for $ g'' $, as shown in simulation results with RASE2 values reduced by up to 90%.
• Cross-validation (10-fold and GCV) provides reliable bandwidth selection, but post-cross-validation tuning is necessary for optimal curvature estimation.
• Simulation results show that even with optimal bandwidths, curvature estimation remains challenging without good initial values, as evidenced by high RASE2 values when starting from random or suboptimal points.