High-Dimensional Asymptotics & Feature Testing

• High-dimensional asymptotics is defined by regimes where data dimension p is comparable to or exceeds sample size n, necessitating new testing methodologies.
• The topic addresses constructing norm-based test statistics, such as quadratic U-statistics and L2-norm tests, with calibrated bootstrap and Studentized techniques for error control.
• Feature testing methods adapt to both dense and sparse signals, using adaptive, dimension-agnostic procedures that ensure valid inference in genomics, neuroimaging, and machine learning.

High-dimensional asymptotics and feature testing comprise the theoretical and methodological core of modern statistical inference when the dimensionality of data, $p$, is comparable to, or much larger than, the sample size $n$. These regimes arise routinely in genomics, neuroimaging, finance, and machine learning, demanding new approaches for both global testing and local (feature-wise) inference that depart from classical fixed-$p$ or $p/n \to 0$ theory. The following exposition synthesizes key developments spanning analytic techniques, test construction, adaptivity, calibration, and the trade-offs peculiar to high-dimensional inference.

1. Defining the High-dimensional Regime and Main Testing Problem Classes

High-dimensional asymptotics govern models in which $p \gg n$, $p/n \to \gamma \in (0,\infty]$, or other non-classical growth rates, necessitating techniques robust across a spectrum of $p$ to $n$ ratios. Core inferential tasks include:

• Global mean vector testing: $H_0: \mu=0$ vs $H_1: \mu\neq0$ for $X_i \sim N_p(\mu, \Sigma)$.
• Two-sample location tests: $H_0: \mu_1=\mu_2$ for high-dimensional populations.
• Feature-wise (marginal) testing: Individually assess $H_{0j}: \mu_j=0$ for $j=1,\ldots,p$.
• Regression/testability in high-dimensional models: Assess parameters (e.g., a single $\beta_j$) in regression when the overall parameter vector is non-sparse.

In all cases, classic procedures (Hotelling's $T^2$, likelihood ratio tests) fail to control size or power appropriately due to singularity or instability of the sample covariance, spurious accumulation of noise, and the breakdown of Gaussian approximations without alternative regularity conditions or calibration.

2. Limit Theorems, U-statistics, and L2/Norm-based Test Statistics

A fundamental approach is to construct test statistics based on U-statistics or norm-based summaries which possess analyzable high-dimensional limiting distributions. Notable forms and results include:

• Quadratic/inner-product U-statistics (Chen–Qin style): For one-sample means,

$U_n = \frac{2}{n(n-1)} \sum_{i<j} X_i' X_j,$

and similarly adapted for two-sample contrasts (Li, 2022).

• $L^2$-norm of the sample mean:

$$R_n = \frac{n \|\bar X_n\|_2^2 - \operatorname{tr}(\Sigma)}{\|\Sigma\|_F},$$

with limiting distribution a weighted sum of chi-squared variables or, under a weak spike condition, $N(0,2)$ (Xu et al., 2014).

• General U-statistics for $\ell_p$ norms: Summing feature-wise symmetric kernel estimators $K_\ell$ over all features to estimate $\|\theta\|_a^a = \sum_\ell \theta_\ell^a$ (He et al., 2018, Zhang et al., 2023).
• Asymptotic limiting distributions: Under moment and covariance regularity (e.g., $$\operatorname{tr}(\Sigma^4) = o(\operatorname{tr}(\Sigma^2)^2)$$), these statistics exhibit Gaussian or $t$-type convergence as $p \to \infty$, with possible adjustments for finite $n$ (Li, 2022, Xu et al., 2014). For more complex objects like maxima, limiting extreme-value distributions (Gumbel) apply.
• Uniform-over-dimension central limit theorems: Limit results that hold simultaneously and uniformly in $p$ (no constraint on how $p$ and $n$ relate), enabling procedures valid for arbitrary relative growth (Chowdhury et al., 2024, Karmakar et al., 10 Dec 2025).

3. Calibration, Adaptivity, and Error Control in High-dimensional Feature Testing

Direct application of classical asymptotic critical values is often invalid at high dimensions. Recent frameworks emphasize:

• Studentization by sample variance: Closed-form estimators for the variance term in quadratic forms enable a t-distribution (with dimension-free degrees of freedom) to arise even for fixed small $n$ but diverging $p$ (Li, 2022).
• Multiplier/bootstrap calibration: Tests based on general normed U-statistics employ multiplier bootstrap schemes (with resampling or simulation based on estimated covariances) to approximate null laws, robust to unknown or heavy-tailed distributions (Zhou et al., 2018, Karmakar et al., 10 Dec 2025).
• Dimension-agnostic procedures: Uniform-over-dimension theory provides calibration valid for low, moderate, or ultra-high $p$, bridging the gap between traditional and high-dimensional regimes (Chowdhury et al., 2024, Karmakar et al., 10 Dec 2025).
• Adaptive testing for arbitrary sparsity: Modern methods combine a family of $L_p$-norm-based tests (for $p=2$, max-type, etc.), and use adaptive (min-p) strategies to attain optimal power rate-adaptively across dense and sparse alternative regimes (He et al., 2018, Zhang et al., 2023, Zhou et al., 2018). Independence among studentized $U(a)$ for distinct $a$ and independence with maximum-type statistics are rigorously established.

Table: Overview of Adaptive High-dimensional Testing Frameworks

Statistic/Approach	Calibration	Optimal regime
$L^2$-norm (quadratic)	Plug-in/bias-corrected/Student t	Dense signals, strong mean
Max-type (max-$|z_j|$)	Extreme value (Gumbel), bootstr.	Sparse, strong per-feature
Adaptive min-$p$	Combination, use independence	Both dense and sparse

4. Feature Selection, Multiple Testing, and Control of Error Rates

Feature-wise inference, that is, the simultaneous testing and identification of impactful features among thousands, poses unique challenges:

• Familywise and FDR control: For massive multiple testing, per-feature tests can be adjusted using Bonferroni, Holm, or FDR methodology, using dimension-adapted critical values justified by high-dimensional or uniform-over-dimension theory (Li, 2022, Romanes et al., 2018).
• High-dimensional discriminant analysis and penalized selection: EBIC-penalized likelihood-ratio frameworks integrate feature testing and selection, achieving Chernoff consistency (type I and II errors vanish as $n$ increases for $p,n\to\infty$ with $\log p/n \to 0$), and dominate fixed-threshold/FDR approaches in asymptotic error (Romanes et al., 2018).
• U-statistics for multiple test statistics: Frameworks allow for simultaneous computation and joint limiting law for a wide class of feature-oriented statistics (means, variances, correlations), and for testing a global null across all features (He et al., 2018, Zhang et al., 2023).

5. Power, Detectability, and Structural Constraints

Power analysis in high dimensions is governed by the signal-to-noise ratio, the sparsity level of the alternative, and the covariance structure:

• Detection boundaries: The minimum detectable signal scales as $\sqrt{\log p / n}$ in sparse regimes and $1/\sqrt{n}$ for dense alternatives. Adaptive and scan-based tests (over $L_p$ norms) achieve minimax rates across these (Zhou et al., 2018).
• Dense vs sparse trade-off: $L^2$-type statistics are most powerful for alternatives spread across many coordinates (dense signals), whereas max-type or higher-order $\ell_p$-norm statistics target alternatives where only a few features deviate (sparse signals) (He et al., 2018, Zhang et al., 2023).
• Critical structural determinants in regression: In high-dimensional regression, the detectability of a coefficient depends on the sparsity of the corresponding column (or row) of the precision matrix (decorrelating vector), rather than the coefficient vector's own sparsity. Minimax lower bounds of order $n^{-1/2} + s n^{-1}\log p$ are sharp; weak correlation (small $s$) allows $\sqrt{n}$-rate testability even for dense $\beta$ (Bradic et al., 2018).

6. Applications, Empirical Validation, and Practical Guidance

Empirical evaluations and applications to real datasets confirm the theoretical insights:

• Simulation studies: New t-statistics and uniformly valid U-statistics maintain size accuracy and competitive power across a broad range of $p,n$ and under non-Gaussian or heavy-tailed settings. Classical tests fail in small $n$, large $p$ setups, inflating type I error or losing power (Li, 2022, Karmakar et al., 10 Dec 2025).
• Neuroimaging and genomics: Case studies (e.g., fMRI ROI analysis with few subjects but many voxels) demonstrate the validity of finite-sample t-distributions for region-level tests, with multiple comparison control resulting in interpretable discoveries (Li, 2022).
• Implementation: Modern testing frameworks scale linearly with $p$ (test statistic computation, bootstrap), or are combinatorially optimized (dynamic programming for monotone-index U-statistics), making large-scale application feasible (Zhang et al., 2023).
• Practical recommendations: For $n \geq 3$ and large $p$, finite-sample t-tests or U-statistics with appropriate bootstrapping are preferred. When exact covariance is unknown or heavy-tailed data suspected, use spatial sign/kernels. For feature selection, EBIC-penalized multiDA offers asymptotically vanishing type I/II rates, provided $\log p/n \to 0$ (Romanes et al., 2018).

7. Methodological Extensions and Open Directions

Current research continues to expand the boundaries of high-dimensional asymptotics and feature testing:

• Uniform-over-dimension theory provides a theoretical basis for inferential procedures that are valid for all $p$, unifying previously distinct low- and high-dimensional regimes (Chowdhury et al., 2024, Karmakar et al., 10 Dec 2025).
• Structured covariance and spiked PCA: Extensions of principal subspace and subsphericity tests accommodate extreme $p/n$ ratios, under mild spectral growth assumptions (Virta, 2021).
• Adaptive and computationally efficient combinations: Low-cost bootstrap and dynamic programming render fully adaptive test classes practical and scalable (Zhou et al., 2018, Zhang et al., 2023).
• Feature learning and neural representations: High-dimensional asymptotic analysis of early training dynamics and random features in neural networks reveals phase transitions in testability and feature detectability, calibrating when learning enters a regime distinguishable from random weights (Hu et al., 2024, Ba et al., 2022).

The theoretical landscape underscores that the interplay between dimensionality, sample size, and signal structure is paramount in governing both the statistical validity and efficiency of feature testing in modern data analysis.