Rescaled Huberized Pinball Loss

• RHPL is a smooth, non-convex, and asymmetric loss function that generalizes classical pinball and quantile Huber losses for robust prediction.
• It mitigates noise and outlier effects through exponential tail clipping and adaptive scaling, ensuring bounded influence in training.
• RHPL shows superior performance in both support vector machines and distributional reinforcement learning by combining theoretical guarantees with practical adaptivity.

The Rescaled Huberized Pinball Loss (RHPL) is a smooth, non-convex, and asymmetric loss function that generalizes the classical pinball (quantile) loss and the quantile Huber loss. Originally developed to address robustness and stability issues in learning under noise and outlier contamination, RHPL provides bounded influence, strong theoretical guarantees, and practical adaptivity to noise. It has been successfully embedded in both support vector machines for classification (RHPSVM) and in distributional reinforcement learning as a quantile regression loss, demonstrating empirical and theoretical superiority over standard alternatives (Diao, 27 Nov 2025, Malekzadeh et al., 2024).

1. Formal Definition and Functional Formulation

The RHPL modifies standard quantile and Huber losses by incorporating exponential (“correntropy”) tail clipping and adaptive scaling. In its classification setting for a sample $(x, y)$, with output $f(x)$ and $u = y \cdot f(x) - 1$, the RHPL is parameterized by the quantile (pinball) parameter $\tau \in (0,1)$, Huber smoothing width $h > 0$, and rescaling $\eta > 0$ (commonly chosen as $\tau$ or 1):

$$L_{\tau,h}(u) = \begin{cases} \eta[1 - \exp(-(u-h/2)^2/2h^2)] & \text{if } u > h \ \eta[1 - \exp(-u^2/2h^2)] & \text{if } 0 \leq u \leq h \ \eta[1 - \exp(-\tau u^2/2h^2)] & \text{if } -h < u < 0 \ \eta[1 - \exp{(-(-u-h/2)^2/2h^2)}] & \text{if } u \leq -h \end{cases}$$

In distributional reinforcement learning, RHPL is derived from the 1-Wasserstein distance between Gaussians, with adaptive threshold $b = |\sigma_p - \sigma_t|$ determined online from the predicted and target quantile noise scales:

$$C_{GL}^b(u) = |u|[1-2\Phi(-|u|/b)] + b\sqrt{2/\pi} \exp(-u^2/2b^2) - b\sqrt{2/\pi}$$

The full RHPL for quantiles $f(x)$0 and $f(x)$1 is then:

$f(x)$2

where $f(x)$3 and $f(x)$4 are midpoint quantile weights (Malekzadeh et al., 2024).

2. Mathematical Properties and Theoretical Guarantees

The RHPL possesses several properties central to robust machine learning:

• Asymmetry: $f(x)$5 unless $f(x)$6, enabling differential penalization of over- and underestimations.
• Smoothness: Constructed from exponentials and quadratics, RHPL is $f(x)$7 and infinitely differentiable, with no non-differentiable corners.
• Non-convexity with Local Convexity: The loss is globally non-convex due to saturation in the tails ($f(x)$8), but convex within the central Huber region ($f(x)$9).
• Bounded Influence: The gradient of $u = y \cdot f(x) - 1$0 is bounded by $u = y \cdot f(x) - 1$1, giving bounded sensitivity to individual outliers.
• Fisher Consistency: The minimizer of the expected RHPL risk for any $u = y \cdot f(x) - 1$2 and $u = y \cdot f(x) - 1$3 recovers the correct Bayes rule, i.e., $u = y \cdot f(x) - 1$4 (Diao, 27 Nov 2025).
• Generalization Bound: Under $u = y \cdot f(x) - 1$5-Lipschitzness and an RKHS kernel bounded by $u = y \cdot f(x) - 1$6, the generalization error is controlled by an explicit bound involving empirical loss and terms scaling as $u = y \cdot f(x) - 1$7.

3. Related Losses and Limit Regimes

RHPL generalizes several loss families:

Classical Loss	Limit of RHPL	Asymmetry	Outlier Behavior
Pinball (Quantile)	$u = y \cdot f(x) - 1$8, $u = y \cdot f(x) - 1$9	Yes	Linear, unbounded
Absolute/Huber Loss	$\tau \in (0,1)$0, $\tau \in (0,1)$1	No	Capped by $\tau \in (0,1)$2
Quantile-Huber	$\tau \in (0,1)$3, using $\tau \in (0,1)$4 in RL setting	Yes	Bounded by $\tau \in (0,1)$5

In the case $\tau \in (0,1)$6 and $\tau \in (0,1)$7 in RL, RHPL smoothly degenerates to the pure quantile loss $\tau \in (0,1)$8 (Malekzadeh et al., 2024).

4. Algorithmic Embedding and Optimization

Classification (RHPSVM Model)

In support vector classification, the RHPSVM minimizes a regularized empirical risk:

$\tau \in (0,1)$9

Slack variables and dualization yield a quadratic program with coordinate-wise variable box constraints reflecting the Huber region status of each sample. Optimization leverages the concave-convex procedure (CCCP) to decompose non-convexity, solving at each iteration a convex quadratic subproblem using the ClipDCD coordinate-descent algorithm. Convergence is guaranteed by monotonic CCCP progress and DCD convergence properties (Diao, 27 Nov 2025).

Distributional Reinforcement Learning

In QR-DQN, IQN, or FQF, RHPL replaces the standard quantile-Huber loss. For each gradient step, residuals $h > 0$0 are computed, per-sample standard deviations $h > 0$1 and $h > 0$2 estimated, and the adaptive threshold $h > 0$3 selected. The exact loss or its piecewise quadratic/linear approximation is used, and training proceeds analogously to classical quantile regression (Malekzadeh et al., 2024).

5. Empirical Performance and Hyperparameter Roles

Extensive experiments confirm RHPL’s advantages:

• Classification Under Noise: On synthetic and UCI datasets with label flips or outliers, RHPSVM outperforms hinge-SVM, pinball-SVM, and other robust SVM variants by 5–10% in classification accuracy. It remains competitive or superior in clean data scenarios.
• High-Dimensional Small-Sample Regimes: On tasks such as crop-leaf image classification with $h > 0$4, RHPSVM achieves 3–5% higher test accuracy, attributable to its tail-bounded outlier robustness and stability in support-vector selection (Diao, 27 Nov 2025).
• Distributional RL: On Atari benchmarks, substituting quantile-Huber with RHPL in QR-DQN or FQF yields higher mean human-normalized scores (934% vs 902%) and faster convergence. In option hedging, D4PG-QR with RHPL auto-selects optimal $h > 0$5 and matches or exceeds hand-tuned alternatives (Malekzadeh et al., 2024).

Parameter roles are sharply delineated:

• $h > 0$6 governs asymmetry: $h > 0$7 for symmetric noise; $h > 0$8 for positive-label noise robustness; $h > 0$9 for recall on minority classes.
• $\eta > 0$0 (or $\eta > 0$1) controls the extent of smoothing and tail saturation: smaller values sharpen the quadratic region, reducing outlier resistance, while larger values enforce stronger capping at the cost of optimization speed or capacity. RHPL is stable for $\eta > 0$2 and $\eta > 0$3 across datasets.

6. Interpretability, Adaptivity, and Extensions

The rescaling (editor’s term: dynamic thresholding) via $\eta > 0$4 or $\eta > 0$5 gives RHPL the ability to adapt to noise encountered in practice, eliminating reliance on manual hyperparameter search. In RL, $\eta > 0$6 can be efficiently estimated online, directly linking the quadratic region width to the distribution shift between predicted and target quantiles.

RHPL forms a universal smooth loss architecture that maintains recovery of the Bayes rule, retains strong regularization and generalization guarantees, and exhibits noise/self-calibrated adaptivity. Its formula subsumes traditional losses as limit cases, supporting extensibility to various advanced SVM variants and robust regression paradigms (Diao, 27 Nov 2025, Malekzadeh et al., 2024).

7. Significance and Contemporary Usage

RHPL and its instantiations (RHPSVM in classification, RHPL in distributional RL) represent an overview of robust statistics (influence functions, Huberization), asymmetric cost design (quantile losses), and modern kernel and deep learning optimization. By combining smoothness, asymmetry, and tail capping, RHPL achieves superior resistance to outliers, improved empirical stability, and faster convergence in both classic and contemporary machine learning pipelines.

The loss structure’s explicit link between adaptivity (via noise scale estimation), theoretical regularity (with generalization and stability guarantees), and empirical robustness marks RHPL and its descendants as central components in modern robust learning research (Diao, 27 Nov 2025, Malekzadeh et al., 2024).