TR19-169 Authors: Lijie Chen, Ron Rothblum, Roei Tell, Eylon Yogev

Publication: 24th November 2019 13:50

Downloads: 315

Keywords:

The Exponential-Time Hypothesis ($ETH$) is a strengthening of the $\mathcal{P} \neq \mathcal{NP}$ conjecture, stating that $3\text{-}SAT$ on $n$ variables cannot be solved in time $2^{\epsilon\cdot n}$, for some $\epsilon>0$. In recent years, analogous hypotheses that are ``exponentially-strong'' forms of other classical complexity conjectures (such as $\mathcal{NP}\not\subseteq\mathcal{BPP}$ or $co\text{-}\mathcal{NP}\not\subseteq \mathcal{NP}$) have also been considered. These Exponential-Time Hypotheses have been widely influential across different areas of complexity theory. However, their connections to *derandomization and circuit lower bounds* have yet to be systematically studied. Such study is indeed the focus of the current work, and we prove a sequence of results demonstrating that *the connections between exponential-time hypotheses, derandomization, and circuit lower bounds are remarkably strong*.

First, we show that if $3\text{-}SAT$ (or even $TQBF$) cannot be solved by probabilistic algorithms that run in time $2^{n/\mathrm{polylog}(n)}$, then $\mathcal{BPP}$ can be deterministically simulated ``on average case'' in (nearly-)polynomial-time (i.e., in time $n^{\mathrm{polyloglog}(n)}$). This result addresses a long-standing lacuna in uniform ``hardness-to-randomness'' results, which did not previously extend to such parameter settings. Moreover, we extend this result to support an ``almost-always'' derandomization conclusion from an ``almost-always'' lower bound hypothesis.

Secondly, we show that *disproving* certain exponential-time hypotheses requires proving breakthrough circuit lower bounds. In particular, if $CircuitSAT$ for circuits over $n$ bits of size $\mathrm{poly}(n)$ can be solved by *probabilistic algorithms* in time $2^{n/\mathrm{polylog}(n)}$, then $\mathcal{BPE}$ does not have circuits of quasilinear size. The main novel feature of this result is that we only assume the existence of a *randomized* circuit-analysis algorithm, whereas previous similar results crucially relied on the hypothesis that the circuit-analysis algorithm does not use randomness.

Thirdly, we show that a very weak exponential-time hypothesis is closely-related to the classical question of whether derandomization and circuit lower bounds are *equivalent*. Specifically, we show two-way implications between the hypothesis that the foregoing equivalence holds and the hypothesis that $\mathcal{E}$ cannot be decided by ``small'' circuits that are *uniformly generated* by relatively-efficient non-deterministic machines. This highlights a sufficient-and-necessary path for progress towards proving that derandomization and circuit lower bounds are indeed equivalent.