1: \begin{abstract}

2: The problem of finding an optimum using noisy evaluations of a smooth

3: cost function arises in many contexts, including economics, business,

4: medicine, experiment design, and foraging theory.  We derive an

5: asymptotic bound

6: \begin{math}

7:  E[ (x_t-x^*)^2 ] \geq O(t^{-1/2})

8: \end{math}

9: on the rate of convergence of a sequence $(x_0, x_1, \ldots)$

10: generated by an unbiased feedback process observing noisy evaluations

11: of an unknown quadratic function maximised at $x^*$.  The bound is

12: tight, as the proof leads to a simple algorithm which meets it.  We

13: further establish a bound on the total regret,

14: \begin{math}

15:  E\bigl[ \sum_{\tau=1}^{t} (x_{\tau} - x^*)^2 \bigr] \geq O(t^{1/2}) .

16: \end{math}

17: These bounds may impose practical limitations on an agent's

18: performance, as $O(\epsilon^{-4})$ queries are made before the queries

19: converge to $x^*$ with $\epsilon$ accuracy.

20: \end{abstract}

21: