Title: Learning low-degree functions from a logarithmic number of random queries

Abstract: We prove that for any integer $n\in\mathbb{N}$, $d\in\{1,\ldots,n\}$ and any $\varepsilon,\delta\in(0,1)$, a bounded function $f:\{-1,1\}^n\to[-1,1]$ of degree at most $d$ can be learned with probability at least $1-\delta$ and $L_2$-error $\varepsilon$ using $\log(\tfrac{n}{\delta})\,\varepsilon^{-d-1} C^{d^{3/2}\sqrt{\log d}}$ random queries for a universal finite constant $C>1$.