1: \begin{abstract}

2: This work studies minimization problems with zero-order noisy oracle information under the assumption that the objective function is highly smooth and possibly satisfies additional properties. We consider two kinds of zero-order projected gradient descent algorithms, which differ in the form of the gradient estimator. The first algorithm algorithm uses a gradient estimator based on randomization on the $\ell_2$ sphere. The precise form that we consider is due to \cite{BP2016} and it has been used for zero-order optimization of strongly convex functions. We present an improved analysis of this algorithm for the same class of functions and we derive rates of convergence for more general function classes. In particular, we consider functions which satisfies the Polyak-Łojasiewicz condition instead of strong convexity, and the larger class of highly smooth non-convex functions. The second algorithm is based on $\ell_1$-type randomization. We show that this novel algorithm enjoys similar theoretical guarantees than the first one and, in the case of noiseless oracle, it enjoys better bounds. % on both its bias and variance.

3: The improvements are achieved by new bounds on bias and variance for both algorithms, which are obtained via Poincaré type inequalities for uniform distributions on $\ell_1$ or $\ell_2$ spheres. The optimality of the upper bounds is discussed and a slightly more general lower bound than the state-of-the art bound in \cite{akhavan2020} is presented. These results imply that the proposed algorithms are nearly minimax optimal.\evg{for which class?}

4: 

5: 

6: \end{abstract}

7: