eng
Schloss Dagstuhl – Leibniz-Zentrum für Informatik
Leibniz International Proceedings in Informatics
1868-8969
2019-12-04
22:1
22:14
10.4230/LIPIcs.FSTTCS.2019.22
article
More on AC^0[oplus] and Variants of the Majority Function
Limaye, Nutan
1
Srinivasan, Srikanth
2
Tripathi, Utkarsh
2
Department of Computer Science and Engineering, IIT Bombay, India
Department of Mathematics, IIT Bombay, India
In this paper we prove two results about AC^0[oplus] circuits.
(1) We show that for d(N) = o(sqrt(log N/log log N)) and N <= s(N) <= 2^(dN^(1/4d^2)) there is an explicit family of functions {f_N:{0,1}^N - > {0,1}} such that
- f_N has uniform AC^0 formulas of depth d and size at most s;
- f_N does not have AC^0[oplus] formulas of depth d and size s^epsilon, where epsilon is a fixed absolute constant.
This gives a quantitative improvement on the recent result of Limaye, Srinivasan, Sreenivasaiah, Tripathi, and Venkitesh, (STOC, 2019), which proved a similar Fixed-Depth Size-Hierarchy theorem but for d << log log N and s << exp(N^(1/2^Omega(d))).
As in the previous result, we use the Coin Problem to prove our hierarchy theorem. Our main technical result is the construction of uniform size-optimal formulas for solving the coin problem with improved sample complexity (1/delta)^O(d) (down from (1/delta)^(2^O(d)) in the previous result).
(2) In our second result, we show that randomness buys depth in the AC^0[oplus] setting. Formally, we show that for any fixed constant d >= 2, there is a family of Boolean functions that has polynomial-sized randomized uniform AC^0 circuits of depth d but no polynomial-sized (deterministic) AC^0[oplus] circuits of depth d.
Previously Viola (Computational Complexity, 2014) showed that an increase in depth (by at least 2) is essential to avoid superpolynomial blow-up while derandomizing randomized AC^0 circuits. We show that an increase in depth (by at least 1) is essential even for AC^0[oplus].
As in Viola’s result, the separating examples are promise variants of the Majority function on N inputs that accept inputs of weight at least N/2 + N/(log N)^(d-1) and reject inputs of weight at most N/2 - N/(log N)^(d-1).
https://drops.dagstuhl.de/storage/00lipics/lipics-vol150-fsttcs2019/LIPIcs.FSTTCS.2019.22/LIPIcs.FSTTCS.2019.22.pdf
AC^0[oplus]
Coin Problem
Promise Majority