Document Type

Technical Report

Publication Date


Technical Report Number



The notion of distributed functional monitoring was recently introduced by Cormode, Muthukrishnan and Yi to initiate a formal study of the communication cost of certain fundamental problems arising in distributed systems, especially sensor networks. In this model, each of k sites reads a stream of tokens and is in communication with a central coordinator, who wishes to continuously monitor some function f of \sigma, the union of the k streams. The goal is to minimize the number of bits communicated by a protocol that correctly monitors f(\sigma), to within some small error. As in previous work, we focus on a threshold version of the problem, where the coordinator's task is simply to maintain a single output bit, which is 0 whenever f(\sigma) \leq \tau(1 - \epsilon) and 1 whenever f(\sigma) \geq \tau. Following Cormode et al., we term this the (k, f, \tau, \epsilon) functional monitoring problem. In previous work, some upper and lower bounds were obtained for this problem, with f being a frequency moment function, e.g., F_0, F_1, F_2. Importantly, these functions are monotone. Here, we further advance the study of such problems, proving three new classes of results. First, we prove new lower bounds on this problem when f = F_p, for several values of p. Second, we study the effect of non-monotonicity of f on our ability to give nontrivial monitoring protocols, by considering f = F_p with deletions allowed, as well as f = H, the empirical Shannon entropy of a stream. Third, we provide nontrivial monitoring protocols when f is either H, or any of a related class of entropy functions (Tsallis entropies). These are the first nontrivial algorithms for distributed monitoring of non-monotone functions.