Efficient QRGMM: Quantile Regression for Generative Modeling

• The paper introduces E-QRGMM, achieving efficient conditional generative modeling via quantile regression and cubic Hermite interpolation to accelerate sample generation.
• It optimally reduces the quantile-regression grid size from O(n^(1/2)) to O(n^(1/5)) while maintaining a sup-norm convergence rate of O_p(n^(-1/2)) for accurate distributional fit.
• E-QRGMM extends to nonlinear and deep architectures, offering practical solutions for high-dimensional risk estimation in finance, e-commerce, and simulation-based inference.

Efficient Quantile-Regression-Based Generative Metamodeling (E-QRGMM) is a framework for constructing conditional generative models of simulator outputs or real-world stochastic systems via quantile regression. E-QRGMM achieves tractable, distributionally faithful, and computationally efficient generation of samples conditioned on covariates, enabling downstream uncertainty quantification, risk estimation, and decision analysis in high-dimensional or high-stakes domains, including supply chain finance, simulation-based inference, and e-commerce credit risk management (Liang et al., 27 Jan 2026, Zhang et al., 18 Jun 2025, Hong et al., 2023).

1. Foundation: Quantile-Regression-Based Generative Metamodeling (QRGMM)

At its core, E-QRGMM builds upon Quantile-Regression-Based Generative Metamodeling (QRGMM), which models the conditional distribution of a real-valued output $Y$ given covariates $x$ via the quantile function $Q(\tau|x) = \inf\{ y : F_Y(y|x) \ge \tau \}$ for $\tau \in (0,1)$. The central principle is that the entire conditional distribution can be generated via the inverse-transform principle: for $U \sim \mathrm{Unif}(0,1)$, $Y \approx \hat Q(U|x)$, where $\hat Q$ is an estimator of $Q$ (Hong et al., 2023).

QRGMM operationalizes this via the following scheme:

• Offline-stage: Fit quantile regressions at $m$ grid points $\tau_j = j/m$, $j=1,\dots,m-1$ to obtain $\hat Q(\tau_j|x)$, typically using linear or nonlinear regression models and the pinball loss.
• Online-stage: For $K$ i.i.d. samples, draw $U_k \sim \mathrm{Unif}(0,1)$, and set
  • $\hat Y_k(x) = \hat Q(\tau_1|x)$ if $U_k < \tau_1$
  • $\hat Y_k(x) =$ linearly (or more generally) interpolated value between adjacent $\hat Q(\tau_j|x)$ for $\tau_j \leq U_k < \tau_{j+1}$
  • $\hat Y_k(x) = \hat Q(\tau_{m-1}|x)$ if $U_k \ge \tau_{m-1}$

This yields a feed-forward generator where sampling reduces to uniform noise generation and fast function evaluation, resulting in sample complexity and speed orders of magnitude superior to typical adversarial or diffusion-based models (Zhang et al., 18 Jun 2025, Hong et al., 2023).

2. E-QRGMM: Algorithmic Enhancements and Hermite Interpolation

E-QRGMM introduces algorithmic innovations that substantially improve the efficiency of QRGMM while preserving its fidelity:

• Cubic Hermite Interpolation: In the "central" region of the quantile grid, E-QRGMM employs Hermite cubic interpolation rather than linear, utilizing both quantile values and gradients with respect to $\tau$. For $\tau_j \leq \tau < \tau_{j+1}$,

$$\hat Q_C(\tau|x) = h_{00}(\xi) Q_j + h_{10}(\xi) (\tau_{j+1}-\tau_j) D_j + h_{01}(\xi) Q_{j+1} + h_{11}(\xi) (\tau_{j+1}-\tau_j) D_{j+1}$$

where $Q_j=\hat Q(\tau_j|x)$, $D_j=\hat D(\tau_j|x)$ is an estimator of $\partial_\tau Q(\tau|x)$, and $h_{00}, h_{01}, h_{10}, h_{11}$ are the Hermite basis functions (Liang et al., 27 Jan 2026).

• Pathwise Sensitivity Gradient Estimation: The derivative $D(\tau|x) = \partial_\tau Q(\tau|x)$ is estimated via

$$\hat D(\tau|x^*) = (x^*)^\top \hat \Lambda(\tau)^{-1} \bar{x}$$

with

$$\hat\Lambda(\tau) = \frac{1}{2 \delta_n} \frac{1}{n} \sum_{i=1}^n x_i x_i^\top \mathbf{1}\{ |y_i - x_i^\top \hat \beta(\tau)| < \delta_n \}$$

and $\bar{x} = \frac{1}{n} \sum_i x_i$, where $\delta_n = O(n^{-1/5})$ (Liang et al., 27 Jan 2026).

• Adaptive Grid Design: E-QRGMM utilizes a hybrid interpolation scheme: fine uniform grid and linear interpolation for tails (where gradient estimation is numerically unstable due to sparse data), and a very lean, coarse grid in the central region where cubic Hermite interpolation is applied. The overall quantile-regression grid size is thus reduced from $O(n^{1/2})$ to $O(n^{1/5})$ while still achieving the optimal statistical convergence rate (Liang et al., 27 Jan 2026).

3. Theoretical Guarantees and Optimality

E-QRGMM's accuracy is characterized by the sup-norm approximation error between the estimated and true conditional quantile functions over $\tau$:

$$\sup_{\tau} |\hat Q(\tau|x) - Q(\tau|x)| = O_p(n^{-1/2})$$

when the grid size $m = O(n^{1/5})$. Hermite interpolation in the central region introduces an interpolation error $O(m^{-4})$ under fourth-order differentiability, and gradient estimation error $O_p(n^{-3/10})/m$, both negligible for chosen $m$ (Liang et al., 27 Jan 2026). The convergence rates are derived under standard regularity: linear quantile regression is a correct model, the conditional density is bounded, and design matrices are nondegenerate (Hong et al., 2023).

Fundamental lower bounds for the conditional quantile estimation problem, established using regression reduction and Assouad's lemma, show that minimax rates $n^{-\beta/(2\beta+d)}$ (with smoothness $\beta$, covariate dimension $d$) are achieved by E-QRGMM under Hölder regularity classes (Schmidt-Hieber et al., 2024). This suggests statistical efficiency is essentially optimal under nonparametric smoothness assumptions.

4. Generative Architecture: Nonlinear and Deep Extensions

While the canonical E-QRGMM relies on linear or basis-augmented quantile regression, the method extends naturally to nonlinear and deep parametrizations. For high-dimensional, combinatorially rich input spaces (e.g., e-commerce data), Deep Factorization Machines (DeepFM) are leveraged as the backbone for quantile regression neural networks:

$$\hat Q_{\tau_j}(x) = \mathrm{DeepFM}(x;\, \theta(\tau_j))$$

DeepFM combines:

• Embedding layers for categorical covariates
• Second-order pairwise factorization machines ($\sum_{i<j} \langle v_i, v_j \rangle x_i x_j$)
• Deep multilayer perceptrons (MLP) with ReLU activations

All quantile levels are trained jointly using the pinball loss with stochastic optimization. Online sampling then follows the inverse-transform procedure as in the linear case (Zhang et al., 18 Jun 2025).

Alternative Bayesian and nonparametric quantile-regression models, such as IQ-BART (Implicit Quantile Bayesian Additive Regression Trees), encode the quantile function via a sum-of-trees prior, sampling the conditional quantile surface $(X,\tau) \mapsto Q(\tau|X)$ and supporting nonparametric inference with minimax posterior concentration rates (O'Hagan et al., 5 Jul 2025). Local-polynomial estimators are also developed as efficient E-QRGMM constructions in the nonparametric regime (Schmidt-Hieber et al., 2024).

5. Functional Risk Estimation and Uncertainty Quantification

A prominent application of E-QRGMM is in functional risk quantification where outputs $Y$ are mapped to risk measures as a function of scenario–e.g., loan amount $l$ in supply chain finance:

• Probability of Default: $r_1(l) = \Pr(Y < l/r)$
• Expected Loss: $r_2(l) = \mathbb{E}[(l - rY)^+]$
• Generalized Loss: $r_3(l) = \mathbb{E}[g_l(Y) \mathbf{1}[Y < a(l)]]$

Given the generative model, Monte Carlo estimates of all such functionals over a continuum of loan levels can be evaluated with a single forward pass and large-scale sampling, exploiting the efficiency of E-QRGMM (Zhang et al., 18 Jun 2025). This enables flexible, covariate-dependent estimation of risk under a unified theoretical framework, as opposed to single-point estimators.

A key virtue of E-QRGMM is that efficient retraining facilitates bootstrap-based construction of covariate-conditional confidence intervals for arbitrary estimands $\psi(Y(x))$ (mean, quantile, tail probability, etc.), overcoming the limitations of conformal prediction and naive bootstrap in conditional settings (Liang et al., 27 Jan 2026).

6. Computational Complexity and Empirical Performance

E-QRGMM is engineered for both scalability and accuracy:

• Offline Cost: Fitting $O(n^{1/5})$ quantile-regression problems of size $n$ (per grid point) in moderate dimension, possibly parallelized or distributed.
• Gradient estimation at quantile grid knots in the central region is computationally negligible relative to quantile regression itself.
• Online Sampling: Evaluating $m$ quantiles per input $x$ for $O(\log m)$ bin location and $O(1)$ interpolation or cubic computation; $K$ samples in $O(K + m \cdot \text{eval cost})$ per $x$. With DeepFM, feed-forward neural evaluation is the dominant per-query cost, but scale advantages persist (Zhang et al., 18 Jun 2025, Liang et al., 27 Jan 2026).

Empirically, E-QRGMM achieves substantially improved tradeoffs between grid size, runtime, and statistical accuracy versus both classical QRGMM and state-of-the-art deep generative baselines (e.g., GANs, DDIM, RectFlow). On benchmark tasks (synthetic and real data; including inventory simulators, e-commerce sales, and supply chain finance), E-QRGMM attains superior distributional fit (KS $\approx 0.01$, WD $\approx 0.03$ in $0.1$s) and confidence interval coverage, while allowing rapid, repeated bootstrapping (Liang et al., 27 Jan 2026, Zhang et al., 18 Jun 2025). Performance remains robust when risk measures or functionals of the output distribution are the estimation target.

Table: Grid Complexity vs. Convergence Rates in E-QRGMM

Method	Grid Size $m$	Convergence Rate
QRGMM	$O(n^{1/2})$	$O_p(n^{-1/2})$
E-QRGMM	$O(n^{1/5})$	$O_p(n^{-1/2})$

Hermite interpolation and gradient estimation enable this order-of-magnitude reduction in quantile-regression grid size without degrading the rate.

7. Extensions, Limitations, and Future Directions

E-QRGMM is compatible with a range of quantile-regression backbones:

• Nonlinear function approximation: Neural nets, random forests, basis expansions
• Local-polynomial estimators: Achieve minimax optimal rates under nonparametric smoothness, with tractable $L_1$ loss for conditional generation (Schmidt-Hieber et al., 2024).
• Multivariate outputs: Extension is possible via transport maps (e.g., Knothe–Rosenblatt rearrangement) and sequential quantile modeling, but faces challenges in high-dimensional output spaces due to complexity and monotonicity constraints.

Practical limitations arise due to instability of derivative estimates in tail regions (remedied by hybrid interpolation), curse of dimensionality for very high-dimensional covariates or outcomes, and the need for careful grid/truncation choices. The method is less appropriate when quantile regression is badly misspecified or data are extremely sparse in certain regions.

A plausible implication is that further advances may incorporate adaptive knot placement, nonparametric or Bayesian quantile processes (e.g., IQ-BART), or multivariate/structured quantile models to mitigate these issues. The framework’s amenability to large-scale parallelization and efficient bootstrap unlocks practical covariate-dependent uncertainty quantification for complex simulators and decision environments (Liang et al., 27 Jan 2026, Schmidt-Hieber et al., 2024, O'Hagan et al., 5 Jul 2025).

---

References:

• (Liang et al., 27 Jan 2026) E-QRGMM: Efficient Generative Metamodeling for Covariate-Dependent Uncertainty Quantification
• (Zhang et al., 18 Jun 2025) Conditional Generative Modeling for Enhanced Credit Risk Management in Supply Chain Finance
• (Hong et al., 2023) Learning to Simulate: Generative Metamodeling via Quantile Regression
• (O'Hagan et al., 5 Jul 2025) Generative Regression with IQ-BART
• (Schmidt-Hieber et al., 2024) Generative Modelling via Quantile Regression