Gap-free Differentially Private PCA for Gaussian Data
It tackles differentially private PCA on Gaussian data without relying on an eigengap assumption.
The authors focus on PCA where the input vectors are drawn from a Gaussian distribution and the output must satisfy differential privacy. Traditional private PCA analyses require a non‑trivial eigenvalue gap to bound error, which limits applicability when the spectrum is flat or unknown.
They construct a private mechanism that removes the eigengap dependence entirely. The design leverages properties of Gaussian data to calibrate noise and uses a refined sensitivity analysis, yielding a procedure that remains differentially private while avoiding the usual spectral‑gap term in the error bound.
The resulting algorithm provably achieves privacy with error guarantees that do not contain any gap‑related factor. Consequently, the method matches the best known non‑private rates up to the privacy cost, demonstrating that a gap‑free guarantee is attainable for Gaussian PCA.
TakeawayA differentially private PCA algorithm can achieve gap‑free error bounds for Gaussian inputs.