Heavy-Tailed Linear Bandits: Huber Regression with One-Pass Update

We study the stochastic linear bandits with heavy-tailed noise. Two principled strategies for handling heavy-tailed noise, truncation and median-of-means, have been introduced to heavy-tailed bandits. Nonetheless, these methods rely on specific noise assumptions or bandit structures, limiting their applicability to general settings. The recent work [Huang et al.2024] develop a soft truncation method via the adaptive Huber regression to address these limitations. However, their method suffers undesired computational cost: it requires storing all historical data and performing a full pass over these data at each round. In this paper, we propose a \emph{one-pass} algorithm based on the online mirror descent framework. Our method updates using only current data at each round, reducing the per-round computational cost from $\mathcal{O}(t \log T)$ to $\mathcal{O}(1)$ with respect to current round $t$ and the time horizon $T$, and achieves a near-optimal and variance-aware regret of order $\widetilde{\mathcal{O}}\big(d T^{\frac{1-\varepsilon}{2(1+\varepsilon)}} \sqrt{\sum_{t=1}^T \nu_t^2} + d T^{\frac{1-\varepsilon}{2(1+\varepsilon)}}\big)$ where $d$ is the dimension and $\nu_t^{1+\varepsilon}$ is the $(1+\varepsilon)$-th central moment of reward at round $t$.

In online decision-making with streaming data, such as in financial markets and online advertising, observations are often affected by heavy-tailed noise, making learning and decision-making particularly challenging. Existing methods rely on storing and processing all past data at each step to achieve strong performance, which is time-consuming and inefficient. This paper introduces a new algorithm that makes decisions using only the current observation, drastically reducing computational cost while maintaining high performance—both theoretically proven and empirically validated.