Abstract
The problem of curve registration appears in many different areas of applications ranging from neuroscience to road traffic modeling. In the present work, we propose a nonparametric testing framework to determine if two noisy signals are identical up to a simple translation. We derive the asymptotic distribution of a penalized likelihood-ratio test statistic, demonstrating a nonparametric version of Wilks’ phenomenon.
Introduction
In fields such as neuroscience and road traffic modeling, researchers frequently analyze functional data—observations represented as continuous signals or curves. A prevalent assumption is the Deformed Mean Template (DMT) model, which posits that observed signals are noisy, shifted versions of a single underlying template. While significant effort has been devoted to estimating these shifts (registration), the fundamental question of goodness-of-fit often remains unaddressed: is the DMT model actually appropriate for the data? If two curves are not related by a simple translation, registration becomes mathematically ill-posed. This paper introduces a formal hypothesis test to determine if two curves coincide up to a translation when the template is unknown.
Methodology
The proposed framework utilizes a Penalized Likelihood-Ratio (PLR) approach. To handle the infinite-dimensional nature of nonparametric curve registration, the authors transform the functional data into the Fourier domain, using Fourier coefficients as sufficient statistics.
To prevent overfitting to high-frequency noise, an $l_2$ penalization term is introduced, corresponding to Sobolev regularity. The choice of shrinkage weights—how higher frequencies are dampened—is critical. The authors evaluate three types:
- Projection weights: A "hard" cutoff of high frequencies.
- Tikhonov weights: A smooth dampening based on frequency.
- Pinsker weights: An optimal weighting scheme for specific regularity classes.
The core theoretical contribution is proving Wilks’ phenomenon: as the noise level $\sigma$ approaches zero, the test statistic $\Delta_\sigma$ converges to a standard Gaussian distribution. This allows for the calculation of $p$-values and the establishment of a natural threshold for hypothesis testing without knowing the true underlying template.
Results and Numerical Validation
The test was evaluated using synthetic datasets, specifically the HeaviSine function—a benchmark in signal processing with sharp jumps and oscillations—subjected to independent Brownian motion noise.
Numerical simulations confirmed the theoretical results. As shown in Figure 1, the proportion of true negatives converges to the nominal 95% level as the noise level decreases (represented by increasing $n = -2 \log_2(\sigma)$). Notably, Pinsker and Tikhonov weights exhibit faster convergence than simple projection weights.
To assess the power of the test, the authors introduced perturbations $\gamma \varphi(t)$ to the template, where $\varphi(t)$ represented cosine waves or Cauchy-like functions.
As illustrated in Figure 2, the test is consistent: the probability of correctly rejecting the null hypothesis (true positive rate) approaches 1 as the perturbation magnitude $\gamma$ increases. Even small deviations from a pure translation are detectable when noise levels are sufficiently low ($\sigma \approx 0.005$).
Discussion and Future Work
This work establishes a rigorous foundation for testing the validity of the DMT model. By marginalizing the unknown template as a nuisance parameter, the test remains robust even when the "true" signal is unknown.
However, several avenues for future research remain. The current framework is restricted to translations; extending the mathematical proofs to cover more complex deformations like scaling, rotations, or non-rigid warping is a logical next step. Additionally, while the noise variance $\sigma$ is assumed known in the proofs, practical applications would benefit from a version that estimates $\sigma$ directly from the data. Finally, the authors note that the "rate of separation" (the smallest detectable difference) might be slightly suboptimal due to the composite nature of the null hypothesis, suggesting further refinement of the test's sensitivity.
Summary updated