In a recent work (Ghazi et al., SODA 2016), the authors with Komargodski and Kothari initiated the study of communication with contextual uncertainty, a setup aiming to understand how efficient communication is possible when the communicating parties imperfectly share a huge context. In this setting, Alice is given a function f and an input string x , and Bob is given a function g and an input string y . The pair ( x y ) comes from a known distribution and f and g are guaranteed to be close under this distribution. Alice and Bob wish to compute g ( x y ) with high probability. The previous work showed that any problem with one-way communication complexity k in the standard model has public-coin communication O ( k (1 + I )) bits in the uncertain case, where I is the mutual information between x and y . A lower bound of ( I ) bits on the public-coin uncertain communication was also shown.

An important question that was left open is the power that public randomness brings to uncertain communication. Can Alice and Bob achieve efficient communication amid uncertainty without using public randomness? How powerful are public-coin protocols in overcoming uncertainty? Motivated by these two questions:

- We prove the first separation between private-coin uncertain communication and public-coin uncertain communication. We exhibit a function class for which the communication in the standard model and the public-coin uncertain communication are O (1) while the private-coin uncertain communication is a growing function of the length n of the inputs.