Approximate contacts, score gaps and parameter enclosures
DefinitionPGDRevision20261113_Contactsbregman-geometryconvex-optimizationproximal-gradient
This module specifies a positive curvature family on a compact parameter interval. The score is the actual parameter derivative of the logarithmic kernel, factors through an ordered state feature, and is jointly continuous. It defines approximate signed contacts, certified full-domain norms, logarithmic error bands, minimum score gaps, both orientation-dependent geometry radii and the scale enclosure. Its final target combines the general localization bounds with uniqueness in the admissible geometry specialization.
Definition code
import Definitions.Def_PGDRevision20261113_Geometry
set_option autoImplicit false
noncomputable section
open scoped ENNReal
namespace PGDRevision20261113
structure ContactFamily where
domain : Set ℝ
A : ℝ → ℝ
B : ℝ → ℝ
kernel : ℝ → ℝ → ℝ
feature : ℝ → ℝ
def score (D : ContactFamily) (c t : ℝ) : ℝ :=
deriv (fun s => Real.log (D.kernel s t)) c
def AdmissibleContactFamily (D : ContactFamily) (C : Set ℝ) : Prop :=
C.Nonempty ∧ IsCompact C ∧ Set.OrdConnected C ∧
(∀ t ∈ D.domain, 0 ≤ D.A t ∧ 0 < D.B t) ∧
(∀ c ∈ C, ∀ t ∈ D.domain,
0 < D.kernel c t ∧ DifferentiableAt ℝ (fun s => Real.log (D.kernel s t)) c) ∧
ContinuousOn (fun z : ℝ × ℝ => score D z.1 z.2) (C ×ˢ D.domain) ∧
(∀ c ∈ C, ∀ s ∈ D.domain, ∀ t ∈ D.domain,
D.feature s = D.feature t → score D c s = score D c t) ∧
∀ c ∈ C, ∀ s ∈ D.domain, ∀ t ∈ D.domain,
D.feature s < D.feature t → score D c s < score D c t
def contactError (D : ContactFamily) (c p t : ℝ) : ℝ :=
(p * D.kernel c t - D.A t) / (p * D.kernel c t + D.B t)
def contactRho (D : ContactFamily) (c p : ℝ) : ℝ≥0∞ :=
⨆ t ∈ D.domain, ENNReal.ofReal |contactError D c p t|
def ApproximateContacts (D : ContactFamily) (C : Set ℝ) (candidate : ℝ × ℝ)
(L U sigma : ℝ) (t : Fin 3 → ℝ) : Prop :=
candidate.1 ∈ C ∧ 0 < candidate.2 ∧ 0 < L ∧ L ≤ U ∧ U < 1 ∧
(sigma = 1 ∨ sigma = -1) ∧ (∀ i, t i ∈ D.domain) ∧
D.feature (t 0) < D.feature (t 1) ∧ D.feature (t 1) < D.feature (t 2) ∧
(∀ i, L ≤ alternatingSign sigma i * contactError D candidate.1 candidate.2 (t i)) ∧
contactRho D candidate.1 candidate.2 ≤ ENNReal.ofReal U ∧
∀ i, alternatingSign sigma i = -1 → U * D.B (t i) < D.A (t i)
def contactBeta (D : ContactFamily) (L U sigma : ℝ) (t : Fin 3 → ℝ) (i : Fin 3) : ℝ :=
if alternatingSign sigma i = 1 then
Real.log (upperBand (D.A (t i)) (D.B (t i)) U) -
Real.log (upperBand (D.A (t i)) (D.B (t i)) L)
else
Real.log (lowerBand (D.A (t i)) (D.B (t i)) L) -
Real.log (lowerBand (D.A (t i)) (D.B (t i)) U)
def scoreGap (D : ContactFamily) (C : Set ℝ) (s t : ℝ) : ℝ :=
sInf ((fun c => score D c t - score D c s) '' C)
def contactScoreBound (D : ContactFamily) (C : Set ℝ) (t : Fin 3 → ℝ) : ℝ :=
sSup {v | ∃ c ∈ C, ∃ i : Fin 3, v = |score D c (t i)|}
def lowerRadius (D : ContactFamily) (C : Set ℝ) (L U sigma : ℝ) (t : Fin 3 → ℝ) : ℝ :=
if sigma = 1 then
(contactBeta D L U sigma t 0 + contactBeta D L U sigma t 1) /
scoreGap D C (t 0) (t 1)
else
(contactBeta D L U sigma t 1 + contactBeta D L U sigma t 2) /
scoreGap D C (t 1) (t 2)
def upperRadius (D : ContactFamily) (C : Set ℝ) (L U sigma : ℝ) (t : Fin 3 → ℝ) : ℝ :=
if sigma = 1 then
(contactBeta D L U sigma t 1 + contactBeta D L U sigma t 2) /
scoreGap D C (t 1) (t 2)
else
(contactBeta D L U sigma t 0 + contactBeta D L U sigma t 1) /
scoreGap D C (t 0) (t 1)
def logScaleRadius (D : ContactFamily) (C : Set ℝ)
(L U sigma : ℝ) (t : Fin 3 → ℝ) : ℝ :=
sSup (Set.range (contactBeta D L U sigma t)) +
contactScoreBound D C t *
max (lowerRadius D C L U sigma t) (upperRadius D C L U sigma t)
def ContactOptimum (D : ContactFamily) (C : Set ℝ) (z : ℝ × ℝ) : Prop :=
z.1 ∈ C ∧ 0 < z.2 ∧ ∀ w : ℝ × ℝ, w.1 ∈ C → 0 < w.2 →
contactRho D z.1 z.2 ≤ contactRho D w.1 w.2
def ApproximateLocalizationBoundsClaim : Prop :=
∀ (D : ContactFamily) (C : Set ℝ), AdmissibleContactFamily D C →
∀ (candidate : ℝ × ℝ) (L U sigma : ℝ) (t : Fin 3 → ℝ),
ApproximateContacts D C candidate L U sigma t →
(0 < scoreGap D C (t 0) (t 1) ∧ 0 < scoreGap D C (t 1) (t 2)) ∧
(∀ z : ℝ × ℝ, z.1 ∈ C → 0 < z.2 → contactRho D z.1 z.2 ≤ ENNReal.ofReal U →
candidate.1 - lowerRadius D C L U sigma t ≤ z.1 ∧
z.1 ≤ candidate.1 + upperRadius D C L U sigma t ∧
|Real.log (z.2 / candidate.2)| ≤ logScaleRadius D C L U sigma t) ∧
∃ z : ℝ × ℝ, ContactOptimum D C z ∧
ENNReal.ofReal L ≤ contactRho D z.1 z.2 ∧
contactRho D z.1 z.2 ≤ ENNReal.ofReal U
def LocalizationGeometryUniquenessClaim : Prop :=
∀ (G : GeometryData) (C : Set ℝ), AdmissibleGeometry G →
IsCompact C → Set.OrdConnected C → C ⊆ Set.Ici (0 : ℝ) →
(∃ z : ℝ × ℝ, GeometryOptimum G z ∧ z.1 ∈ C) →
∃! z : ℝ × ℝ, RestrictedGeometryOptimum G C z
def ApproximateContactLocalizationClaim : Prop :=
ApproximateLocalizationBoundsClaim ∧ LocalizationGeometryUniquenessClaim
end PGDRevision20261113
Source
Author manuscript: Exact Residual Bounds and Optimal Geometry Selection for Proximal Gradient Methods, revised reading edition, 13 November 2026, Section 7, Theorem 7.1, page 18.