The communication complexity of distributed estimation
By: Parikshit Gopalan , Raghu Meka , Prasad Raghavendra and more
Potential Business Impact:
Makes computers guess better with less information.
We study an extension of the standard two-party communication model in which Alice and Bob hold probability distributions $p$ and $q$ over domains $X$ and $Y$, respectively. Their goal is to estimate \[ \mathbb{E}_{x \sim p,\, y \sim q}[f(x, y)] \] to within additive error $\varepsilon$ for a bounded function $f$, known to both parties. We refer to this as the distributed estimation problem. Special cases of this problem arise in a variety of areas including sketching, databases and learning. Our goal is to understand how the required communication scales with the communication complexity of $f$ and the error parameter $\varepsilon$. The random sampling approach -- estimating the mean by averaging $f$ over $O(1/\varepsilon^2)$ random samples -- requires $O(R(f)/\varepsilon^2)$ total communication, where $R(f)$ is the randomized communication complexity of $f$. We design a new debiasing protocol which improves the dependence on $1/\varepsilon$ to be linear instead of quadratic. Additionally we show better upper bounds for several special classes of functions, including the Equality and Greater-than functions. We introduce lower bound techniques based on spectral methods and discrepancy, and show the optimality of many of our protocols: the debiasing protocol is tight for general functions, and that our protocols for the equality and greater-than functions are also optimal. Furthermore, we show that among full-rank Boolean functions, Equality is essentially the easiest.
Similar Papers
The communication complexity of distributed estimation
Computational Complexity
Lets computers estimate data with less talking.
Sequential 1-bit Mean Estimation with Near-Optimal Sample Complexity
Machine Learning (Stat)
Helps computers guess averages with less data.
Lower Bounds on Relative Error Quantum Compression and Classical Shadows
Quantum Physics
Connects two computer problems to save information.