1: \begin{abstract}
2: Algorithms for bandit convex optimization and online learning often
3: rely on constructing noisy gradient estimates, which are then used
4: in appropriately adjusted first-order algorithms, replacing
5: actual gradients.
6: Depending on the properties of the function to be optimized and the nature of ``noise'' in the bandit feedback,
7: the bias and variance of gradient estimates exhibit various tradeoffs.
8: % and, correspondingly,
9: %the algorithms were proven to achieve various convergence rates.
10: In this paper we propose a novel framework that replaces the specific gradient estimation
11: methods with an abstract oracle.
12: % capturing the controlleable bias-variance tradeoff
13: %of existing gradient estimators.
14: With the help of the new framework we unify previous works,
15: reproducing their results in a clean and concise fashion,
16: while, perhaps more importantly, the framework also allows us to formally show that to achieve the optimal
17: root-$n$ rate %in online/stochastic bandit convex optimization
18: either the algorithms that use existing gradient estimators,
19: or the proof techniques used to analyze them
20: have to go beyond what exists today.
21: \if0
22: We present a novel noisy gradient oracle model for convex optimization. The model allows to explicitly address the bias-variance tradeoff of gradient estimation methods typical in the literature, as well as to prove upper and lower bounds for the minimax error. As such the oracle model allows a clean way of addressing the limits of achievable performance
23: when biased, noisy gradient estimators with controllable bias-variance tradeoffs are employed.
24: When considering the model for online/stochastic convex optimization,
25: one consequence of our results is that the currently used gradient estimates and the proof techniques used to analyze them cannot
26: achieve the optimal square-root regret rate for online/stochastic bandit convex optimization.
27: \fi
28: % when the function cannot be queried outside the domain.
29: \end{abstract}
30: