Limits on Gradient Compression for Stochastic Optimization

We consider stochastic optimization overl p spaces using access to a first-order oracle. We ask: What is the minimum precision required for oracle outputs to retain the unrestricted convergence ratesƒ We characterize this precision for every p≥ 1 by deriving information theoretic lower bounds and by providing quantizers that (almost) achieve these lower bounds. Our quantizers are new and easy to implement. In particular, our results are exact for p = 2 and p = ∞, showing the minimum precision needed in these settings are Θ(d) and Θ(log d), respectively. The latter result is surprising since recovering the gradient vector will require Ω(d) bits.