Root-Finding Bilevel Optimization (RF-BO) Theoretical Analysis
-3Sample ComplexityImplicit Gradient Methods
Evaluation Results
| Method | Links | |
|---|---|---|
| Implicit Gradient Methods(UL) h / Objective=Minimization of ∥h∥2, (LL) R / g=SC or PL, Noise Assumption=Bounded variance, Jacobian needed?=Yes, Single-Loop=Partial2026.06 | -3 | |
| Penalty-based Methods(UL) h / Objective=Minimization + penalty, (LL) R / g=SC, Noise Assumption=Bounded / small variance, Jacobian needed?=Yes, Single-Loop=Yes2026.06 | -3 | |
| Contextual Methods(UL) h / Objective=Minimization (contextual), (LL) R / g=SC, Noise Assumption=Small variance O(ϵ), Jacobian needed?=Yes, Single-Loop=Yes2026.06 | -3 | |
| RF-TTSA(UL) h / Objective=Direct root-finding h = 0, (LL) R / g=SC or PL, Noise Assumption=Bounded variance / Markovian, Jacobian needed?=No, Single-Loop=Yes2026.06 | -2 |