Quantile Regression Function Analysis

• Quantile regression function is a parameter that defines the conditional quantiles of an outcome, capturing the full distribution beyond the mean.
• It uses series approximations to model complex, high-dimensional covariate effects, ensuring uniform estimation across quantile indices.
• Uniform inference methods—including pivotal, bootstrap, and Gaussian approaches—enable robust confidence bands and hypothesis tests for function-valued parameters.

A quantile regression function is a function-valued parameter that describes the conditional quantiles of a response variable given covariates and serves as a principal object for analyzing the impact of covariates across the entire conditional distribution, not solely its mean. In contrast to classical regression methods, which typically target the conditional mean, quantile regression systematically studies the effect of covariates on the conditional quantile process for all quantile indices. This approach enables a detailed analysis of distributional heterogeneity, tail behavior, and nonlinear covariate effects, especially relevant in settings with high-dimensional or nonparametric covariate representations, as well as in inference for linear functionals and derivatives of the quantile function.

1. Definition and Role of the Conditional Quantile Regression Function

The conditional quantile regression function $Q_{Y|X}(u \mid x)$ is defined, for a response variable $Y$, covariate vector $x \in \mathbb{R}^p$, and quantile index $u \in (0,1)$, as the functional inverse of the conditional cumulative distribution function: $$Q_{Y|X}(u \mid x) = \inf \{y: F_{Y|X}(y \mid x) \geq u\}$$ This function characterizes, for fixed $x$, the $u$-th quantile of $Y$ conditional on $X=x$, providing a nonparametric description of the entire conditional distribution. The quantile regression function is paramount for function-valued inference tasks, such as estimation of conditional quantile curves, partial derivatives, average partial derivatives, and other linear functionals in both low- and high-dimensional or infinite-dimensional (series) settings (Belloni et al., 2011).

2. Series Approximation and High-Dimensional Quantile Regression

When modelling the entire conditional quantile function in settings with many regressors or with $X$ in a high/infinite-dimensional space, the function is commonly approximated via a linear combination of basis (“series”) expansions: $$Q_{Y|X}(u \mid x) \approx \sum_{j=1}^m b_j(u) Z_j(x)$$ Here, $Z_j(x)$ are known basis functions (e.g., polynomials, splines, or other suitable transformations), $b_j(u)$ are unknown coefficient functions to be estimated, and $m$ is the dimension of the approximation, which may increase with sample size $n$. This framework, known as the QR-series approach, enables a nonparametric yet efficiently estimable representation of the conditional quantile function (Belloni et al., 2011).

A central challenge in this regime is that as $m$ increases (potentially with $n$), uniform empirical control of the Gram matrix (the matrix of empirical second moments of the $Z_j$) is required to guarantee the uniform consistency and efficiency of the estimates. The following result, adapted from Guédon and Rudelson, expresses this key empirical process control:

Let $(Z_1, ..., Z_n)$ be $n$ i.i.d. vectors in $\mathbb{R}^m$, and define

$$\delta^2 := \frac{\log n}{n} \cdot \frac{\mathbb{E} \max_{1 \le i \leq n} \|Z_i\|^2}{\max_{\alpha \in \mathcal{S}^{m-1}} \mathbb{E}[(\alpha^\top Z_1)^2]}$$

If $\delta^2 < 1$, then

$$\mathbb{E} \left\{ \max_{\alpha \in \mathcal{S}^{m-1}} \left| \frac{1}{n} \sum_{i=1}^n (\alpha^\top Z_i)^2 - \mathbb{E}\left[(\alpha^\top Z_1)^2\right] \right| \right\} \leq 2\delta \max_{\alpha \in \mathcal{S}^{m-1}} \mathbb{E}\left[(\alpha^\top Z_1)^2\right]$$

This ensures that, provided $m$ and the magnitude of $Z_i$ do not grow too quickly relative to $n$, the worst-case discrepancy (over all unit directions $\alpha$) between the empirical variance and the population variance is controlled by a small multiple of the maximal variance, a result that is crucial for the validity of strong approximation theory for coefficient estimation in high-dimensional QR-series models (Belloni et al., 2011).

3. Uniform Inference and Resampling Methods

Estimation of the quantile regression function in the QR-series framework enables uniform strong approximation of the full vector-valued coefficient process (i.e., as a stochastic process in $u$ and $x$) by pivotal and Gaussian processes. On this basis, a suite of four resampling methods has been established for performing inference uniformly in $u$ and $x$:

• Pivotal method: Uses estimated pivotal processes as an approximation to the distribution of the estimator.
• Gradient bootstrap: Employs a model-based (gradient) bootstrap for approximating the sampling distribution of the estimator.
• Gaussian method: Approximates the coefficient process by a properly constructed Gaussian process.
• Weighted bootstrap: Utilizes resampling weights to construct the empirical distribution of the coefficient function.

All these methods are designed to provide uniform inference for function-valued (or curve) parameters—such as the entire conditional quantile function, its derivatives, and linear functionals—across both quantile index $u$ and regressor $x$. Pointwise inference is obtained as a special case (Belloni et al., 2011).

4. Inference for Functionals of the Quantile Process

The QR-series approach and its coupling (strong approximation) theory enable inference for a large class of linear functionals of the quantile function:

• The conditional quantile curve $Q_{Y|X}(u \mid x)$ itself.
• Partial derivatives with respect to regressors, corresponding to quantile-specific “marginal effects”.
• Average partial derivatives (integrated over $x$).
• Conditional average partial derivatives, i.e., effects averaged over $x$ at a given $u$.

Uniform rates of convergence are derived for these functionals (in the sup-norm), as well as techniques for constructing confidence bands and hypothesis tests for the entire function-valued object (Belloni et al., 2011).

5. Uniform Asymptotics and Rates

Under suitable growth conditions for $m$ and regularity assumptions for the Gram matrix, the QR-series estimators achieve uniform rates of convergence over both the quantile index $u$ and the regressor $x$. The deviation of the empirical process (and hence the estimator) from its population counterpart can be controlled uniformly, facilitating statistical inference on the entire conditional quantile function (not just pointwise). This uniformity is essential in high-dimensional settings for providing valid coverage for bands and confidence sets for function-valued parameters (Belloni et al., 2011).

6. Practical Application: Empirical Example

The methodological developments are illustrated in an empirical application that estimates the price elasticity function and tests the Slutsky condition for the individual demand for gasoline, indexed by the individual's unobserved propensity for gasoline consumption. This use case exemplifies the power of quantile regression functions for:

• Constructing uniform confidence and inference bands for entire functional relationships.
• Testing shape and economic regularity conditions (e.g., monotonicity, Slutsky negativity) that are inherently function-valued. Here, the function-valued inference is key, as the object of interest is a distributional feature (e.g., elasticity curve) varying across quantiles rather than a single scalar summary (Belloni et al., 2011).

7. Empirical Process Control in High Dimensions

The accurate estimation and inference for quantile regression functions in nonparametric and high-dimensional series settings fundamentally depend on empirical process control over the Gram matrix. The lemma provided gives explicit, non-asymptotic bounds for the discrepancy between empirical and population second moments, crucial for ensuring the validity of uniform limit theorems and for justifying the coupling (strong approximation) used in the resampling methods. The bounds require that the model complexity (dimension $m$) and the norms of the basis functions are regulated so that $\delta^2$ remains below 1, ensuring uniform convergence properties and the reliability of inferential procedures (Belloni et al., 2011).

---

The quantile regression function thus serves as a cornerstone in modern regression analysis for modeling and inference on conditional quantile processes, particularly in high-dimensional, nonparametric, and functional data settings. Its amenability to uniform inference, strong approximation, and flexible modeling using series expansions enables statistically rigorous analysis of covariate effects on a wide array of function-valued characteristics of the outcome distribution.