Gaussian distributions maximize information cost for correlation targets
A Self-Coupling Gaussian Extremal Inequality for Information-Constrained Optimal Transport with Applications to Rate-Distortion-Perception
Summary
The paper investigates how much information is needed to create a certain level of similarity or correlation between two identical sets of data from a known distribution. It finds that when considering distributions with fixed averages and spreads, the normal (Gaussian) distribution requires the most information to achieve these similarity goals. The authors prove this using a novel mathematical construction involving stochastic processes called Ornstein–Uhlenbeck bridges. They also apply this result to problems in data compression, balancing how much data is lost versus how closely the output resembles the original.
What this means in practice
- •For data compression engineers: Improve rate-distortion-perception trade-offs when compressing data under Gaussian assumptions using this new extremal inequality.
- •For information theory practitioners: Design encoding schemes with known mutual information bounds for correlated sources involving limited or unlimited shared randomness.
A theory result. No direct application yet.