Quantile Huber Loss: A Robust, Smooth Loss Function

• Quantile Huber loss is a hybrid loss function that unifies asymmetric quantile and symmetric Huber objectives through a tunable threshold parameter.
• It continuously interpolates between L₁ (quantile) and L₂ (expectile) losses, providing smooth gradients and strong convexity for efficient optimization.
• Its applications span robust regression, high-dimensional inference, and distributional reinforcement learning, enabling precise uncertainty quantification and parameter tuning.

The quantile Huber loss, also termed the asymmetric Huber, Huberized quantile, or hybrid L₁–L₂ quantile loss, is a parametric loss function that unifies the robustifying qualities of the Huber loss with the distributional sensitivity and asymmetry of the pinball (quantile) loss. Predominantly used in robust regression, distributional reinforcement learning, probabilistic forecasting, and high-dimensional inference, the quantile Huber loss interpolates continuously between the classical quantile (L₁-type) and expectile (asymmetric L₂-type) objectives via a threshold parameter. Its widespread adoption is driven by advantageous smoothness for optimization, convexity properties, and algorithmic flexibility in modern statistical and deep learning frameworks.

1. Mathematical Definition and Limiting Regimes

The core formulation of the quantile Huber loss for scalar residual $u = y - \hat{y}$ and quantile level $\tau \in (0,1)$ with Huber threshold $\delta>0$ is: $$L_{\tau,\delta}(u) = \begin{cases} \tau |u| - \frac{\delta \tau^2}{2}, & u < -\tau \delta\ \frac{1}{2\delta}u^2, & -\tau\delta \leq u \leq (1-\tau)\delta\ (1-\tau)|u| - \frac{\delta (1-\tau)^2}{2}, & u > (1-\tau)\delta \end{cases}$$ This asymmetric loss generalizes both the symmetric Huber loss (recovered at $\tau=0.5$) and the classical quantile ("check-function" or pinball) loss ($\delta \to 0$). As $\delta \to 0^+$, the quadratic region shrinks, yielding: $L_{\tau,0}(u) = | \tau - 1_{u<0} | |u|$ As $\delta \to \infty$, the loss transitions to a rescaled asymmetric squared-error (expectile) loss: $$L_{\tau,\infty}(u) = | \tau - 1_{u<0} | \frac{1}{2}u^2$$ The tuning parameter $\delta$ determines the width of the central quadratic regime. This trait enables practitioners to balance robustness to outliers (linear tails) and differentiability (smooth fit near the prediction).

2. Analytical Properties

The quantile Huber loss is continuous and everywhere differentiable in $u$: $$\frac{d}{du} L_{\tau,\delta}(u) = \begin{cases} -\tau, & u < -\tau\delta \ u/\delta, & -\tau\delta \leq u \leq (1-\tau)\delta\ 1-\tau, & u > (1-\tau)\delta \end{cases}$$ This piecewise-linear-quadratic (PLQ) structure underlies key algorithmic benefits:

• The loss is convex in $u$ for all $\tau$ and $\delta>0$ (Ramamurthy et al., 2015, Tyralis et al., 2023, Taggart, 2021).
• The gradient is Lipschitz with constant $1/\delta$ in the quadratic regime and globally bounded by $\max(\tau,1-\tau)$.
• Strong convexity in the joint (model, quantile parameter) space is obtained for the normalized loss $$\bar{\rho}_{\tau,\delta}(r) = \rho_{\tau,\delta}(r) + \log c(\tau)$$ where $c(\tau)$ is the log-partition function. This admits biconvex alternating-minimization schemes (Ramamurthy et al., 2015).

In the context of distributional RL, the asymmetry $| \tau - 1_{u<0} |$ ensures that the loss targets conditional quantile functionals, while the quadratic region provides improved gradient flow and variance reduction in optimization (Jullien et al., 2023, Malekzadeh et al., 2024).

3. Statistical Role and Elicitability

For a random variable $Y$ with cumulative distribution $F$, the minimizer of the expected quantile Huber loss defines the so-called "Huber functional": $$H_{\delta}^\tau(F) = \arg\min_x \mathbb{E}[L_{\tau,\delta}(Y-x)]$$ This functional always exists and is elicitable, yielding a non-empty closed interval that interpolates between the $\tau$-quantile set (as $\delta \to 0$) and the unique $\tau$-expectile (as $\delta \to \infty$) (Taggart, 2021, Tyralis et al., 2023). The scoring function is strictly consistent for this functional. This property persists in both the classical (frequentist) and Bayesian settings (Soomro et al., 2023).

The mixture decomposition of scoring functions consistent for $H_{\delta}^\tau$ admits a representation as non-negative mixtures over capped pinball-type elementary losses, which grounds interpretability and supports forecast comparison in practical applications (Taggart, 2021).

4. Practical Use in Machine Learning and High-Dimensional Regression

The quantile Huber loss is employed in regression, sparse estimation, gradient-boosted machines, and deep neural networks. In sparse regression,

$$\hat{x} = \arg\min_{x \in \mathbb{R}^p} \sum_{i=1}^n L_{\tau, \delta}(b_{i} - A_i^T x) + \lambda \|x\|_1$$

delivers robustness to outliers, convexity for efficient optimization, and scalability via specialized interior-point methods (Aravkin et al., 2014). In high-dimensional settings, generalized Orthogonal Matching Pursuit (OMP) leverages the differentiability of the loss for variable selection, with finite-sample convergence and statistical consistency guarantees.

In gradient boosting, quantile Huber loss enables approximate second-order optimization and stable tree splits, overcoming the zero-Hessian issue of the non-differentiable pinball loss (Yin et al., 2023). Empirical evaluation on summary metrics—coverage, prediction interval width, and coverage-width criterion—demonstrates that quantile Huber loss yields more efficient uncertainty estimates than non-smoothed quantile boosting.

In deep learning, architectures with the quantile Huber objective (e.g., Deep Huber Quantile Regression Networks) interpolate smoothly between quantile regression and expectile regression networks, affording strict consistency for Huber quantile functionals and conferring well-calibrated interval predictions (Tyralis et al., 2023).

5. Role in Distributional Reinforcement Learning

Quantile Huber loss has become foundational in distributional RL algorithms such as QR-DQN, IQN, FQF, and D4PG-QR, where the return distribution is approximated via quantiles. Loss computation over pairs of predicted and target quantiles proceeds as: $$L_{\text{QR}}(\psi) = \frac{1}{N^2} \sum_{i=1}^N \sum_{j=1}^N |\hat{\tau}^{(i)} - 1_{u^{(i,j)}<0}| \, \frac{L_H^k(u^{(i,j)})}{k}$$ with $u^{(i,j)} = y^{(j)} - \theta^{(i)}$ and $L_H^k(u)$ as the (symmetric) Huber loss with threshold $k$.

Empirical results demonstrate that using the Huberized quantile loss accelerates training and improves score convergence; however, theoretical contraction in Wasserstein distance is lost relative to pure pinball loss, and pathologies such as distributional collapse (return distribution spiking to the mean) can occur (Jullien et al., 2023). The dual expectile–quantile approach circumvents this by jointly learning expectile and quantile representations, thus restoring contraction guarantees as the number of fractions diverges.

Further, generalizations of the quantile Huber loss—e.g., losses directly derived from the 1-Wasserstein distance between Gaussians, with adaptive threshold parameters reflecting prediction noise—improve robustness, convergence stability, and allow interpretable, data-driven parameter tuning without exhaustive search (Malekzadeh et al., 2024).

6. Bayesian Formulations and Extensions

Recent work has derived Bayesian regularized quantile Huber regression models, representing the likelihood as a scale-mixture of normals with adaptive robustness (shape) and scale parameters: $$L_{\tau, \eta, \rho}(u) = \sqrt{ \eta \left[ \eta + (u/\rho^2) (\tau - 1_{u<0}) \right] } - \eta$$ Posterior inference via Gibbs sampling exploits the hierarchy of conjugate priors on regression coefficients and hyperparameters. Data-driven adaptation of the robustness parameter $\eta$ enables full probabilistic quantification of uncertainty and outlier control. Empirical results show improved parameter recovery and interval calibration, especially in outlier-heavy or heavy-tailed regimes (Soomro et al., 2023).

7. Implementation Strategies and Empirical Recommendations

Best practices emerging from the literature for applying the quantile Huber loss include:

• Setting the threshold parameter $\delta$ or its generalizations to the standard deviation of inlier noise or adaptively estimating it from data (e.g., $b = |\sigma_1 - \sigma_2|$ as in distributional RL) (Aravkin et al., 2014, Malekzadeh et al., 2024).
• Cross-validating $\delta$ or analogous parameters alongside regularization hyperparameters for optimal prediction-uncertainty tradeoff (Yin et al., 2023).
• Exploiting the PLQ structure to leverage efficient gradient-based or interior-point optimization, including in high-dimensional and distributed computing contexts (Aravkin et al., 2014).
• Implementing the loss and its derivatives for custom objectives in tree-based ensemble methods and deep neural architectures is direct, with pseudocode and PyTorch-like templates available in the distributional RL literature (Malekzadeh et al., 2024).

Empirical validation spans synthetic, semi-synthetic, and real-world datasets—encompassing regression, time-series forecasting, high-dimensional genomics, and RL domains—consistently confirming the theoretical prediction that quantile Huber losses enhance estimator robustness, promote stable and rapid convergence, enable principled uncertainty quantification, and, where required, interpolate seamlessly between classical quantile and expectile regimes (Ramamurthy et al., 2015, Aravkin et al., 2014, Tyralis et al., 2023, Soomro et al., 2023, Yin et al., 2023, Jullien et al., 2023, Malekzadeh et al., 2024).