Abstract
We study a scalar-on-function historical linear regression model which assumes that the functional predictor does not influence the response when the time passes a certain cutoff point. We approach this problem from the perspective of locally sparse modeling, where a function is locally sparse if it is zero on a substantial portion of its defining domain. In the historical linear model, the slope function is exactly a locally sparse function that is zero beyond the cutoff time. We propose a nested group bridge penalty that is able to specifically shrink the tail of a function. Combined with the B-spline basis expansion and penalized least squares, the nested group bridge approach can identify the cutoff time and produce a smooth estimate of the slope function simultaneously. The proposed method is demonstrated with an application of determining the effect of the past engine acceleration on the current particulate matter emission.
Introduction: The "Memory" of Functional Data
In Functional Data Analysis (FDA), researchers often treat data points as continuous curves rather than discrete numbers. A common task is "scalar-on-function" regression: predicting a single response based on an entire predictor curve. In many real-world scenarios, the "history" of the curve doesn't matter indefinitely. This leads to the historical functional linear model, where the predictor's influence ceases after a certain cutoff point in time.
The challenge is twofold: identifying exactly when the past stops mattering (the cutoff time $\delta$) and maintaining a smooth, physically plausible estimate of the effect over time. Previous methods often resulted in "jumpy" or discontinuous estimates at the cutoff point, which does not align with physical realities where effects usually fade out gradually.
Methodology: B-Splines and Nested Group Bridge
The proposed method combines B-spline basis expansion with a novel nested group bridge penalty.
B-Spline Representation
The unknown slope function $\beta(t)$, which represents the "effect" of the past, is expanded using B-splines. B-splines are piecewise polynomials known for their smoothness and compact support, meaning each basis function only affects a small part of the curve.
The Nested Group Bridge Penalty
The core innovation is the nested group bridge penalty. By organizing the B-spline coefficients into nested groups ($A_1 \supset A_2 \supset \dots \supset A_M$), the penalty hits the "tail" of the function (the most distant past) harder than the "head" (the recent past). This forces the function to become exactly zero beyond a certain point, effectively "shrinking" the tail until only the relevant historical window remains. This approach treats the historical cutoff problem as one of "local sparsity."
Optimization
Since the resulting objective function is non-convex, the authors developed an iterative algorithm that re-frames the problem as a series of Lasso problems, which are computationally efficient to solve.
Real-World Application: Truck Emissions
The model was evaluated using a real-world dataset from the Coordinating Research Councils E55/E59 project, tracking the engine acceleration of diesel trucks and their resulting particulate matter emissions.
As shown in the results above, the model identified a cutoff of approximately 20 seconds. This suggests that acceleration older than 20 seconds has no significant impact on current particulate emissions. The resulting estimate is smoother and more interpretable than those produced by previous aggressive truncation methods.
Key Results and Insights
- Superior Accuracy: In simulations, the Nested Group Bridge (NGR) consistently achieved lower Mean Integrated Squared Error (MISE) compared to standard smoothing splines and existing truncation methods.
- Consistency: The authors mathematically proved that the estimator is consistent, meaning the estimate is guaranteed to converge to the true historical effect and cutoff time as more data is collected.
- Smoothness Matters: Unlike previous methods that simply cut the function off, this approach allows the effect to taper off to zero smoothly, which is more statistically stable and physically plausible.
Conclusion
The nested group bridge approach provides a robust framework for estimating the "memory" of historical functional data. By combining local sparsity with smoothness constraints, it offers a significant improvement over traditional truncation techniques. Future work may extend this approach to function-on-function models where the response is also a curve over time.
Summary updated