Interaction is Necessary for Distributed Learning with Privacy or Communication Constraints

Local differential privacy (LDP) is a model where users send privatized data to an untrusted central server whose goal it to solve some data analysis task. In the non-interactive version of this model the protocol consists of a single round in which a server sends requests to all users then receives their responses. This version is deployed in industry due to its practical advantages and has attracted significant research interest. Our main result is an exponential lower bound on the number of samples necessary to solve the standard task of learning a large-margin linear separator in the non-interactive LDP model. Via a standard reduction this lower bound implies an exponential lower bound for stochastic convex optimization and specifically, for learning linear models with a convex, Lipschitz and smooth loss. These results answer the questions posed by Smith, Thakurta, and Upadhyay (IEEE Symposium on Security and Privacy 2017) and Daniely and Feldman (NeurIPS 2019). Our lower bound relies on a new technique for constructing pairs of distributions with nearly matching moments but whose supports can be nearly separated by a large margin hyperplane. These lower bounds also hold in the model where communication from each user is limited and follow from a lower bound on learning using non-adaptive statistical queries.

Exponential Lower Bound for Berge-Ramsey Problems

We give an exponential lower bound for the smallest $$N$$ N such that no matter how we c-color the edges of a complete $$r$$ r -uniform hypergraph on $$N$$ N vertices, we can always find a monochromatic Berge-$$K_n$$ K n .

Sizes of spaces of triangulations of 4-manifold and balanced presentations of the trivial group

Let [Formula: see text] be any compact four-dimensional PL-manifold with or without boundary (e.g. the four-dimensional sphere or ball). Consider the space [Formula: see text] of all simplicial isomorphism classes of triangulations of [Formula: see text] endowed with the metric defined as follows: the distance between a pair of triangulations is the minimal number of bistellar transformations required to transform one of the triangulations into the other. Our main result is the existence of an absolute constant [Formula: see text] such that for every [Formula: see text] and all sufficiently large [Formula: see text] there exist more than [Formula: see text] triangulations of [Formula: see text] with at most [Formula: see text] simplices such that pairwise distances between them are greater than [Formula: see text] ([Formula: see text] times). This result follows from a similar result for the space of all balanced presentations of the trivial group. (“Balanced” means that the number of generators equals to the number of relations). This space is endowed with the metric defined as the minimal number of Tietze transformations between finite presentations. We prove a similar exponential lower bound for the number of balanced presentations of length [Formula: see text] with four generators that are pairwise [Formula: see text]-far from each other. If one does not fix the number of generators, then we establish a super-exponential lower bound [Formula: see text] for the number of balanced presentations of length [Formula: see text] that are [Formula: see text]-far from each other.