Theoretical computer science · Oct 6, 2026
Memory–sample lower bounds for noiseless Gaussian regression
For fixed A > 0, a one-pass learner with persistent bits needs noiseless Gaussian samples to recover a unit vector to angular error with probability 2/3, uniformly in accuracy for large d. Computation and randomized updates are unrestricted, but output uses only the terminal state, stopping index and fresh randomness.
Contributors
Unreleased OpenAI model
AI systems: Unreleased OpenAI model