Copyright (c) 2025. All rights reserved. Released under Apache 2.0 license as described in the file LICENSE.
PLMBBridge — Adapter/Bridge Lemmas for PLAcceleratedNesterovLeans Integration #
Overview #
This file provides adapter/bridge lemmas connecting PLMB's generic
Morse–Bott and tubular-neighborhood theory (over an arbitrary
finite-dimensional real inner-product space E) to the concrete
setting used by PLAcceleratedNesterovLeans (where E d := EuclideanSpace ℝ (Fin d)).
Bridge classification table #
| Bridge | Difficulty | Mechanism |
|---|---|---|
IsTubNeighSub.toIsTubularNeighborhood | Trivial | Field-for-field |
IsTubNeighSub.shrink | Short proof | Restriction to open subset |
IsTubNeighSub.radius | Short proof | Per-point ball from openness |
tubularProj @ Ed d | Trivial | Type specialization |
tubular_neighborhood_projection @ Ed d | Trivial | Type specialization |
‖Ext.gradient f x‖ = ‖fderiv ℝ f x‖ | One-liner | LinearIsometryEquiv.norm_map |
Ext.PL → MuPL | Already in Thm3 | ExternalThm3.pl_to_muPL |
hessian f x ξ ξ = Ext.hessianQuadForm | Already in Thm3 | hessianQuadForm_eq_hessian |
Ext.PL + argmin → IsMuMB | Composition | Pipeline of above bridges |
muPL_implies_muMB @ Ed d | Trivial | Type specialization |
References #
- Rebjock & Boumal, "Fast convergence to non-isolated minima: four equivalent conditions for C² functions"
- PLAcceleratedNesterovLeans:
PLAcceleratedNesterovLean/Core/Defs.lean,PLAcceleratedNesterovLean/Convergence/LocalGeometry/Main.lean
Nearest-point projection specialized to the Euclidean ambient space Ed d.
Equations
Instances For
The norm of the gradient equals the operator norm of the Fréchet derivative.
This follows from toDual ℝ E being a linear isometry.
Variant with Ext.gradient (BridgeDefs.lean).
Squared norms version.
Connection to Mathlib's gradient (from Analysis.Calculus.Gradient.Basic).
Mathlib defines gradient 𝕜 f x := (toDual 𝕜 E).symm (fderiv 𝕜 f x).
Our ExternalThm3.gradient has the same definition (specialized to ℝ).
They are definitionally equal.
Norm bridge for Mathlib's gradient.
Corollary 2.17 at Ed d: μ-PŁ ⟹ μ-Morse–Bott.
Theorem 2.16 at Ed d: μ-PŁ ⟹ local-min set is a C¹ submanifold.
Full pipeline at Ed d: global PL + global minimizer ⟹ μ-MB.
Given:
f : Ed d → ℝis C²fsatisfies the Polyak–Łojasiewicz condition onUwith constantμx₀is a global minimizer (i.e.x₀ ∈ ExternalThm3.argminSet f)U ∈ 𝓝 x₀(x₀ is in the interior of U)
Conclude: f satisfies μ-Morse–Bott at x₀.
This is the bridge lemma that PLAcceleratedNesterovLeans' local_fiberwise_geometry can
use to invoke PLMB's structural results.
Submanifold conclusion via global PL + tubular neighborhood.
Under the same hypotheses, localMinSet f x₀ is a C¹ submanifold
with tangent space hessianKer f x₀.