Curvature Absorption Helper Lemmas #
Helper lemmas for the curvature absorption proof (h_curv_absorb_hyp in Basic.lean). These establish the key perturbation bounds:
- Step bound: ‖hn‖ ≤ η·‖gn‖ + √η·|ρ|·(‖vn‖ + √η·‖gn‖)
- MVT bound: ‖P(y-x) - (π y - π x)‖ ≤ ε₁·‖y-x‖
- Kills-normal bound: ‖P en‖ ≤ ε₁·‖en‖
theorem
PLAcceleratedNesterovLean.nesterov_step_bound
{d : ℕ}
(f : E d → ℝ)
(η ρ : ℝ)
(s : NesterovState d)
(hη : 0 ≤ η)
:
The Nesterov step satisfies hn = -η·gn + √η·ρ·(vn - √η·gn), so ‖hn‖ ≤ η·‖gn‖ + √η·|ρ|·(‖vn‖ + √η·‖gn‖).
theorem
PLAcceleratedNesterovLean.xi_bound_mvt
{d : ℕ}
(P : E d →L[ℝ] E d)
(π : E d → E d)
(x y c : E d)
(r : ℝ)
(hxs : x ∈ Metric.ball c r)
(hys : y ∈ Metric.ball c r)
(hπ_diff : ∀ z ∈ Metric.ball c r, DifferentiableAt ℝ π z)
(ε₁ : ℝ)
(hbound : ∀ z ∈ Metric.ball c r, ‖fderiv ℝ π z - P‖ ≤ ε₁)
:
MVT bound: if ‖fderiv π z - P‖ ≤ ε₁ for all z in a ball, then ‖P(y-x) - (π y - π x)‖ ≤ ε₁·‖y-x‖ for x, y in the ball.
theorem
PLAcceleratedNesterovLean.curv_absorption_algebraic
{E' : Type u_1}
[NormedAddCommGroup E']
[InnerProductSpace ℝ E']
(wn ξn gn Pen hn en : E')
(sm sa Ln : ℝ)
(hsm_nn : 0 ≤ sm)
(hsa_pos : 0 < sa)
(hLn_nn : 0 ≤ Ln)
(ε₁ : ℝ)
(hε₁_pos : 0 < ε₁)
(hε₁_le1 : ε₁ ≤ 1)
(hξ : ‖ξn‖ ≤ ε₁ * ‖hn‖)
(hPe : ‖Pen‖ ≤ ε₁ * ‖en‖)
(C_wh : ℝ)
(_hC_wh_nn : 0 ≤ C_wh)
(hwh : sm ^ 2 * ‖wn‖ ^ 2 + ‖hn‖ ^ 2 ≤ C_wh * Ln)
(C_h : ℝ)
(_hC_h_nn : 0 ≤ C_h)
(hh_sq : ‖hn‖ ^ 2 ≤ C_h * Ln)
(C_ge : ℝ)
(_hC_ge_nn : 0 ≤ C_ge)
(hge : ‖gn‖ * ‖en‖ ≤ C_ge * Ln)
(θ : ℝ)
(_hθ_pos : 0 < θ)
(habs : ε₁ * (C_wh / 2 + sm ^ 2 / 2 * C_h + sa * C_ge) ≤ θ * sa)
:
Core algebraic assembly for curvature absorption. Given that ‖ξ‖ and ‖Pe‖ each have an ε₁ factor from Dπ continuity, and all squared norms are bounded by constants times Ln, conclude δ_curv + proj_err ≤ a/2 * Ln.
The key algebraic steps are:
- Young's inequality: sm·‖w‖·‖h‖ ≤ (sm²·‖w‖² + ‖h‖²)/2 (absorbs √μ')
- ε₁² ≤ ε₁ (since ε₁ ≤ 1), so ‖ξ‖² ≤ ε₁·‖h‖²
- Everything factors as ε₁ · K · Ln, and ε₁ · K ≤ a/2 by hypothesis