.png){kind=icon}
.png){kind=icon}
Return
Bootstrapping variables in algebraic circuits
.png){kind=icon}
OA
AI
DOI：10.1073/pnas.1901272116.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
We show that for the blackbox polynomial identity testing (PIT) problem it suffices to study circuits that depend only on the first extremely few variables. One needs only to consider size-s degree-s circuits that depend on the first log(oc) s variables (where c is a constant and composes a logarithm with itself c times). Thus, the hitting-set generator (hsg) manifests a bootstrapping behavior-a partial hsg against very few variables can be efficiently grown to a complete hsg. A Boolean analog, or a pseudorandom generator property of this type, is unheard of. Our idea is to use the partial hsg and its annihilator polynomial to efficiently bootstrap the hsg exponentially w.r.t. variables. This is repeated c times in an efficient way. Pushing the envelope further we show that (i) a quadratic-time blackbox PIT for 6,913-variate degree-s size-s polynomials will lead to a near-complete derandomization of PIT and (ii) a blackbox PIT for n-variate degree-s size-s circuits in s(n delta) time, for delta < 1/2, will lead to a near-complete derandomization of PIT (in contrast, s(n) time is trivial). Our second idea is to study depth-4 circuits that depend on constantly many variables. We show that a polynomial-time computable, O(s(1.49))-degree hsg for trivariate depth-4 circuits bootstraps to a quasipolynomial time hsg for general polydegree circuits and implies a lower bound that is a bit stronger than that of Kabanets and Impagliazzo [Kabanets V, Impagliazzo R (2003) Proceedings of the Thirty-Fifth Annual ACM Symposium on Theory of Computing STOC '03].
Keywords:
hitting-set
depth-4
derandomization
identity testing
lower bound
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
P
IF:
9.1
Papers:
10.8W
Citations:
73.5W