Unlearning on Smooth Strongly Convex Stochastic Optimization (SCO)
1Excess Risk RateDP baseline
Evaluation Results
| Method | Links | |
|---|---|---|
| DP baselineReference / comments=Group DP and [ALD21]2026.06 | 1 | |
| Retrain from scratchReference / comments=Exact 0-unlearning; Theorem 52026.06 | 1 | |
| ERM / warm-start unlearningReference / comments=[AKGK25]; corrected in Theorem 72026.06 | 1 | |
| Newton-step unlearningReference / comments=[SAKS21]; sharpened in Theorem 8; requires Lipschitz Hessian2026.06 | 1 | |
| Our upper boundReference / comments=Optimal up to condition number; tight for mean estimation2026.06 | 1 | |
| Our lower boundReference / comments=Optimal up to condition number; tight for mean estimation2026.06 | 1 |