When are large codes possible for AVCs?

We study a general Omniscient Arbitrarily Varying Channel (AVC) problem where Alice wishes to communicate a message to receiver Bob by inputting a length-n vector x to a channel. Jammer James observes x, and as a function of x chooses a state sequence s. Bob observes y (such that channel inputs and outputs are related component-wise as y
_i
= w(x
_i
,s
_i
) for some deterministic function w(.,.)) from which he must estimate m with no error. Input and state constraints determine feasible inputs x and s for Alice and James respectively. In this work we characterize when a positive communication rate is possible.We first show that the capacity of any such AVC completely depends upon the relationship between a confusability set, and the set of completely-positive-self-couplings (both are convex sets of certain single-letter probability distributions). Our main result provides essentially matching necessary and sufficient conditions for capacity positivity; we show that the zero-error capacity of an AVC is positive if there are completely-positive-self-couplings outside the confusability set of the given AVC; and that the AVC capacity is zero if all completely-positive-self couplings are in the interior of this confusability set. Our achievability uses a novel code construction based on completely-positive-self-couplings called cloud codes which are strict generalizations of all known Gilbert-Varshamov (GV) type codes. Our converse is based upon Ramsey-theoretic ideas, a generalization of the Plotkin bound leveraging a known result on the duality of completely positive matrices and copositive matrices, and a Fourier-analytic proof of the non-existence of certain sequences of random variables.