Abstract
The family of rank estimators, including Han's maximum rank correlation (Han, 1987) as a notable example, has been widely exploited in studying regression problems. For these estimators, although the linear index is introduced for alleviating the impact of dimensionality, the effect of large dimension on inference is rarely studied. This paper fills this gap via studying the statistical properties of a larger family of M-estimators, whose objective functions are formulated as U-processes and may be discontinuous in increasing dimension set-up where the number of parameters, $p_{n}$, in the model is allowed to increase with the sample size, $n$. First, we find that often in estimation, as $p_{n}/n\rightarrow 0$, $(p_{n}/n)^{1/2}$ rate of convergence is obtainable. Second, we establish Bahadur-type bounds and study the validity of normal approximation, which we find often requires a much stronger scaling requirement than $p_{n}^{2}/n\rightarrow 0.$ Third, we state conditions under which the numerical derivative estimator of asymptotic covariance matrix is consistent, and show that the step size in implementing the covariance estimator has to be adjusted with respect to $p_{n}$. All theoretical results are further backed up by simulation studies.
Introduction
Rank-based estimators are a cornerstone of robust statistics and econometrics, prized for their "distribution-free" properties. By relying on the relative ranks of observations rather than their exact values, these estimators—such as Han’s Maximum Rank Correlation (MRC)—avoid the need for stringent distributional assumptions regarding error terms. Traditionally, the theoretical validity of these estimators has been established under the assumption of a fixed dimension ($p$) as the sample size ($n$) approaches infinity.
However, modern data analysis frequently involves "increasing dimensions," where the number of parameters $p_n$ grows alongside $n$. While high-dimensional theory is well-developed for smooth, differentiable objective functions, rank estimators present a unique challenge: their objective functions are often formulated as discontinuous U-processes (e.g., involving indicator functions like $1\{Y_i > Y_j\}$). This paper provides a rigorous theoretical framework to bridge this gap, establishing the conditions under which rank estimators remain reliable in high-dimensional settings.
The Challenge of Discontinuous U-Processes
Standard M-estimation theory typically relies on Taylor expansions, which require the objective function to be smooth. Rank estimators, however, often utilize step functions that are non-differentiable. To analyze these, the authors treat the estimators as a class of M-estimators where the objective functions are U-processes—statistics based on pairs or tuples of data points.
To overcome the lack of smoothness, the authors utilize the Hoeffding Decomposition, which allows a complex U-process to be broken down into a manageable linear term and a "degenerate" remainder term. A central technical contribution of this work is a new Maximal Inequality (Theorem 2.1) that provides the exact order of magnitude for this remainder term as the dimension increases. This inequality is crucial for proving that the stochastic "noise" inherent in the U-process does not overwhelm the statistical "signal" as $p_n$ grows.
Key Theoretical Findings
1. Convergence Rates
The research establishes that as long as the ratio of parameters to observations vanishes ($p_n/n \to 0$), the estimators achieve a convergence rate of $O_P(\sqrt{p_n/n})$. This is significant because it matches the minimax optimal rate for this class of problems, confirming that rank estimators are efficient even as the model complexity increases.
2. The Inference Gap and Bahadur-Type Bounds
A critical insight of this paper is the distinction between estimation and inference. While $p_n/n \to 0$ is sufficient for the estimator to converge to the true parameter, valid statistical inference (such as constructing confidence intervals) requires much stricter conditions. The authors derive Bahadur-type representations, which show how closely the estimator can be approximated by a sum of independent variables—a prerequisite for Asymptotic Normality.
They find that for rank estimators, the normal approximation often requires a scaling requirement significantly stronger than the common $p_n^2/n \to 0$ rule-of-thumb. If these stricter conditions are not met, the standard normal approximation fails, leading to unreliable confidence intervals.
3. Covariance Matrix Estimation
Because rank objective functions are not differentiable, researchers typically use numerical derivatives to estimate the asymptotic covariance matrix. The authors prove that the "step size" ($\epsilon_n$) used in these numerical calculations must be specifically tuned to the dimension $p_n$. They provide a framework for adjusting this step size to ensure the consistency of the covariance estimator in high dimensions.
Simulation Evidence
The authors validated their theoretical results using Monte Carlo simulations based on a binary choice model ($Y = 1$ if $X^\top \beta + \epsilon \geq 0$). Even in relatively low-to-moderate dimensions (varying $p$ from 1 to 4 with $n=400$), the simulations demonstrated a clear "inference breakdown." As $p$ increased, the coverage probability of 95% confidence intervals dropped significantly, confirming that the requirements for valid inference are far more sensitive to dimensionality than the requirements for point estimation.
Conclusion and Practical Implications
This paper provides the first comprehensive treatment of rank estimators in increasing dimensions. For practitioners, the message is twofold:
- Robustness with a Cost: Rank estimators remain powerful, distribution-free tools in high dimensions, but their inferential validity is fragile.
- Tuning is Essential: When performing inference in high-dimensional settings, one cannot use standard software defaults for numerical derivatives; the step size must be adjusted to account for $p_n$.
Future work remains to be done in the realm of "ultra-high" dimensions ($p \gg n$), where sparsity-inducing penalties like the Lasso would be required to manage the dimensionality. Additionally, the $O(n^2)$ computational complexity of rank estimators remains a hurdle for massive datasets, suggesting a need for more computationally efficient approximations that preserve these theoretical guarantees.
Summary updated