A New Class of Asymptotically Distribution-Free Smooth Tests

Technical Summary: A New Class of Asymptotically Distribution-Free Smooth Tests

Problem Statement
The paper addresses the challenge of constructing smooth goodness-of-fit tests that remain asymptotically distribution-free, even when model parameters are estimated and sample sizes are only moderately large. Traditional smooth tests, such as the score test (Neyman, 1937) and order selection tests (Ledwina, 1994), rely on the asymptotic distribution-freeness of specific test statistics. However, this property often requires very large sample sizes to be realized in practice, particularly for complex distributions or when computationally intensive parameter estimation is involved. Furthermore, the choice of basis functions in these classical tests is often arbitrary, and their limiting distributions may vary depending on the specific parametric model being tested, necessitating model-specific simulations (e.g., parametric bootstrapping) to determine critical values.

Methodology
The authors propose a framework that shifts the perspective from ensuring specific statistics are distribution-free to constructing the underlying empirical process itself to be asymptotically distribution-free by design. This is achieved through two primary methodological components:

1. The Khmaladze-2 (K2) Transform:
   The core innovation is the application of the K2 transform (Khmaladze, 2016) to the construction of the orthonormal basis functions used in smooth tests.

   • Mechanism: The transform utilizes unitary operators to map model-dependent projections of the empirical process into a "standard projection" defined by a user-chosen reference distribution, $F_\gamma$.
   • Construction: The authors define a K2 orthonormal basis $\{h_{j\beta}\}$ for the space $L^2(G_\beta)$ by mapping a basis from the reference space $L^2(F_\gamma)$ via a sequence of operators: a multiplication operator $l_{\gamma,\beta}$, a reflection operator $K$, and a unitary operator $U_p$.
   • Result: This construction ensures that the residuals of the basis functions, $\{\tilde{h}_{j\beta}\}$, possess the same mean and covariance structure under the null hypothesis $G_\beta$ as the residuals of the basis functions under the reference distribution $F_\gamma$. Consequently, the empirical process indexed by these residuals, and any functional of it (including order and subset selection statistics), shares the same limiting distribution under $G_\beta$ as it does under $F_\gamma$. Since $F_\gamma$ can be chosen to be simple (e.g., a standard normal or uniform), the limiting distribution becomes distribution-free.

2. Projected Bootstrap:
   To address the computational burden of simulating null distributions, particularly when parameters are estimated, the authors utilize the "projected bootstrap" (Algeri, 2022).

   • Mechanism: Instead of re-estimating the unknown parameter $\beta$ on every bootstrap replicate (as in the classical parametric bootstrap), this method exploits the projection structure induced by parameter estimation.
   • Efficiency: Based on Proposition 1 (Khmaladze, 1980), the empirical process indexed by estimated residuals is asymptotically equivalent to the process indexed by the true residuals. This allows the user to estimate the parameter once on the observed data and then simulate the null distribution without re-estimating $\beta$ for each bootstrap sample, significantly reducing computational time.

Key Results
The paper validates the proposed methodology through theoretical proofs and simulation studies:

• Theoretical Validity: Propositions 2 through 5 formally demonstrate that the K2-transformed basis functions yield empirical processes that are asymptotically distribution-free, even when parameters are estimated via Z-estimators (including maximum likelihood).
• Simulation Studies (Section 4.2):
  • The authors simulated data from a complex mixture distribution and tested it against three different parametric null models ($G_{\beta,1}, G_{\beta,2}, G_{\beta,3}$).
  • Distribution-Freeness: When using standard basis functions (compositions of Legendre polynomials and the null CDF), the null distributions of the test statistics varied significantly across the different models. In contrast, when using the K2-transformed basis functions, the simulated null distributions under all three models were indistinguishable from the distribution under the reference model $F_\gamma$, even with a sample size of $n=100$.
  • Power: The K2-based tests demonstrated competitive or superior power compared to classical smooth tests in specific scenarios (e.g., detecting deviations in $G_{\beta,2}$ and $G_{\beta,3}$), though the authors note that power improvements are not guaranteed for all alternatives.
• Computational Efficiency (Section 3.1): In a simulation involving an asymmetric Laplace distribution, the projected bootstrap reduced CPU time by approximately 75% (9 minutes vs. 37 minutes for 100,000 replicates) compared to the classical parametric bootstrap, while accurately recovering the true null distribution.
• Case Study (Section 5): The methodology was applied to an X-ray spectrum from the star RT Cru. The K2-transformed order and subset selection tests confirmed the validity of a parametric model describing background and three spectral lines (Fe XXV, Fe XXVI, and Fe K$\alpha$), with p-values of 0.463 and 0.454, respectively. The QQ-plots confirmed the distribution-free property in this real-world application.

Significance and Claims
The paper claims to provide a new family of smooth tests that are asymptotically distribution-free by design, rather than by the specific choice of a test statistic.

• Robustness: The distribution-free property is preserved even when parameters are estimated and sample sizes are moderately large, addressing a known limitation of classical smooth tests.
• Generality: The approach allows applied scientists to use a single, standard limiting distribution (derived from a simple reference model) for testing a wide variety of complex parametric models, eliminating the need for model-specific asymptotic derivations or computationally expensive resampling for every new model.
• Computational Advantage: The integration of the projected bootstrap offers a practical, efficient alternative to the classical parametric bootstrap, making the procedure feasible for complex models where parameter estimation is time-consuming.

The authors conclude that this framework extends the utility of the K2 transform beyond goodness-of-fit testing into the realm of smooth tests, offering a middle ground between general goodness-of-fit tests and directional Neyman-Pearson tests. They acknowledge that while the current work focuses on univariate settings, the framework is adaptable to multivariate models and regression contexts.