(A.1), p. 21 — (1/2L)‖∇f(x) − ∇f(x⋆)‖² ≤ f(x) − f(x⋆) at a minimizer x⋆ of an L-smooth f
ProvedSGDKaczmarz.SGD.eq_A_1gradient-lipschitzp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1smoothness
Let be a real Hilbert space and a differentiable function whose gradient is Lipschitz with constant :
If is a global minimizer of , then for every
This is inequality (A.1) of the paper, which the proof of the co-coercivity Lemma A.1 applies to two auxiliary convex functions.
Formalization Note The page writes (A.1) as a chain whose middle term adds , which is at a minimizer; only the outer inequality is stated. The hypothesis is needed because the page divides by . Convexity is not assumed: the inequality holds at any global minimizer.
Preamble
import Mathlib open scoped NNReal InnerProductSpace
Formal statement
namespace SGDKaczmarz.SGD
/-- (A.1), proof of Lemma A.1, p. 21 (outer inequality): if `∇f` is `L`-Lipschitz and `x⋆`
minimizes `f`, then `(1/2L)‖∇f(x) − ∇f(x⋆)‖² ≤ f(x) − f(x⋆)`. -/
theorem eq_A_1 {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
(f : H → ℝ) (hf : Differentiable ℝ f) (L : ℝ≥0) (hL0 : 0 < L)
(hL : LipschitzWith L (gradient f)) (xstar : H) (hmin : ∀ y, f xstar ≤ f y) (x : H) :
1 / (2 * (L : ℝ)) * ‖gradient f x - gradient f xstar‖ ^ 2 ≤ f x - f xstar := by sorry
end SGDKaczmarz.SGD
Source
Needell, Srebro & Ward, Stochastic gradient descent, weighted sampling, and the randomized Kaczmarz algorithm, arXiv:1310.5715v5, proof of Lemma A.1 (Appendix A.1), (A.1), p. 21
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.