Lifting randomized query complexity to randomized communication complexity (1703.07521v5)

Abstract: We show that for a relation $f\subseteq {0,1}^n\times \mathcal{O}$ and a function $g:{0,1}^{m}\times {0,1}^{m} \rightarrow {0,1}$ (with $m= O(\log n)$), $$\mathrm{R}{1/3}(f\circ g^n) = \Omega\left(\mathrm{R}{1/3}(f) \cdot \left(\log\frac{1}{\mathrm{disc}(M_g)} - O(\log n)\right)\right),$$ where $f\circ g^n$ represents the composition of $f$ and $g^n$, $M_g$ is the sign matrix for $g$, $\mathrm{disc}(M_g)$ is the discrepancy of $M_g$ under the uniform distribution and $\mathrm{R}{1/3}(f)$ ($\mathrm{R}{1/3}(f\circ g^n)$) denotes the randomized query complexity of $f$ (randomized communication complexity of $f\circ g^n$) with worst case error $\frac{1}{3}$. In particular, this implies that for a relation $f\subseteq {0,1}^n\times \mathcal{O}$, $$\mathrm{R}{1/3}(f\circ \mathrm{IP}_m^n) = \Omega\left(\mathrm{R}{1/3}(f) \cdot m\right),$$ where $\mathrm{IP}_m:{0,1}^m\times {0,1}^m\rightarrow {0,1}$ is the Inner Product (modulo $2$) function and $m= O(\log(n))$.