Brassard et al. [Phys. Rev. Lett. 96, 250401 (2006)] showed that shared nonlocal boxes with the CHSH probability greater than $\frac{3+\sqrt{6}}6$ yields trivial communication complexity. There still exists the gap with the maximum CHSH probability $\frac{2+\sqrt{2}}4$ achievable by quantum mechanics. It is an interesting open question to determine the exact threshold for the trivial communication complexity. Brassard et al.'s idea is based on the recursive bias amplification by the 3-input majority function. It was not obvious if other choice of function exhibits stronger bias amplification. We show that the 3-input majority function is the unique optimal, so that one cannot improve the threshold $\frac{3+\sqrt{6}}6$ by Brassard et al.'s bias amplification. In this work, protocols for computing the function used for the bias amplification are restricted to be non-adaptive protocols or particular adaptive protocol inspired by Pawłowski et al.'s protocol for information causality [Nature 461, 1101 (2009)]. We first show a new adaptive protocol inspired by Pawłowski et al.'s protocol, and then show that the new adaptive protocol is better than any non-adaptive protocol. Finally, we show that the 3-input majority function is the unique optimal for the bias amplification if we apply the new adaptive protocol to each step of the bias amplification.