Documentation

IntroductoryLecturesOnConvexOptimization_Nesterov_2004.Chap04.Proposition_4_1_18

theorem cubicRegularization_gap_antitone {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {g : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} (method : CubicRegularizationMethod g stepMap L0 (↑L) x0) (c : ) :
Antitone fun (k : ) => g (method.x k) - c

Helper for Proposition 4.1.18: subtracting a fixed reference value preserves the monotone objective decrease along a cubic-regularization trajectory.

theorem leastIndex_le_of_mem {S : Set } {N k : } (hN : IsLeast S N) (hk : k S) :
N k

Helper for Proposition 4.1.18: an explicit target hit bounds the least hitting index from above.

theorem one_le_leastIndex_of_zero_not_mem {S : Set } {N : } (hN : IsLeast S N) (h0 : 0S) :
1 N

Helper for Proposition 4.1.18: if the initial iterate misses the target set, then the least hitting index is at least one.

theorem leastIndex_eq_zero_of_lt_one_bound {S : Set } {N : } {b : } (hN : IsLeast S N) (hbound : N b) (hb : b < 1) :
N = 0

Helper for Proposition 4.1.18: any subunit real upper bound on a least hitting index forces that least index to be 0.

theorem initialGap_le_of_leastAccuracyIndex_lt_one_bound {gap : } {ε b : } {N : } (hN : IsLeast {k : | gap k ε} N) (hbound : N b) (hb : b < 1) :
gap 0 ε

Helper for Proposition 4.1.18: if the least ε-hitting index has a subunit real upper bound, then the initial gap already satisfies the target.

theorem antitone_tail_le_of_le {α : Type u_1} [Preorder α] {a : α} {u : α} {k0 : } (hu : Antitone u) (hk0 : u k0 a) (j : ) :
u (k0 + j) a

Helper for Proposition 4.1.18: once an antitone sequence enters a threshold region, every later term stays in that same region.

theorem norm_sub_le_of_isGreatest_sublevel_image {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {g : E} {x0 xStar : E} {D : } (hD : IsGreatest ((fun (x : E) => x - xStar) '' {x : E | g x g x0}) D) x : E :
g x g x0x - xStar D

Helper for Proposition 4.1.18: an attained maximal radius over the initial sublevel-set image immediately gives the usual pointwise radius bound on that sublevel set.

theorem secondPhaseLog4ProgressLowerBound {δ δNext ω : } ( : 0 < ω) ( : 0 < δ) (hδNext : 0 < δNext) (hstep : δNext 1 / 2 * δ * (δ / ω)) :
Real.logb 4 (16 / 9 * ω / δNext) Real.logb 4 (3 / 2) + 3 / 2 * Real.logb 4 (16 / 9 * ω / δ)

Helper for Proposition 4.1.18: a one-step second-phase bound yields the exact lower bound on the displayed logb 4 tail potential corresponding to the ratio (4 * ω₀) / Δ.

theorem secondPhaseLog4PotentialMulLowerBound {δ δNext ω : } ( : 0 < ω) ( : 0 < δ) (hδNext : 0 < δNext) (hstep : δNext 1 / 2 * δ * (δ / ω)) :
Real.logb 4 (4 * ω / δNext) 3 / 2 * Real.logb 4 (4 * ω / δ)

Helper for Proposition 4.1.18: the cleaner tail potential logb 4 ((4 * ω) / δ) grows by a pure multiplicative factor 3 / 2 under the second-phase recurrence.

theorem secondPhaseLog4PotentialAffineLowerBound_exactThirdStep {δ δNext ω : } ( : 0 < ω) ( : 0 < δ) (hδNext : 0 < δNext) (hstep : δNext 1 / 3 * δ * (δ / ω)) :
Real.logb 4 (4 * ω / δNext) Real.logb 4 (3 / 2) + 3 / 2 * Real.logb 4 (4 * ω / δ)

Helper for Proposition 4.1.18: under the exact 1 / 3 second-phase recurrence, the same logb 4 ((4 * ω) / δ) potential gains an additive logb 4 (3 / 2) term before the familiar 3 / 2 geometric factor appears.

theorem secondPhaseLog4PotentialShiftedGeometricLowerBound_exactThird_of_final_pos {gap : } {ω : } {k0 j : } ( : 0 < ω) (hmono : Antitone gap) (hk0 : gap k0 ω / 3) (hstep : ∀ (k : ), gap k ωgap (k + 1) 1 / 3 * gap k * (gap k / ω)) (hgap_final : 0 < gap (k0 + j)) :
(3 / 2) ^ j * (Real.logb 4 (4 * ω / gap k0) + 2 * Real.logb 4 (3 / 2)) Real.logb 4 (4 * ω / gap (k0 + j)) + 2 * Real.logb 4 (3 / 2)

Helper for Proposition 4.1.18: after shifting by 2 * logb 4 (3 / 2), the exact 1 / 3 second-phase recurrence still yields a clean geometric 3 / 2 growth law along any positive tail segment.

theorem one_le_secondPhaseLog4Potential_of_le_thirdThreshold {δ ω : } ( : 0 < ω) (hδ_pos : 0 < δ) ( : δ ω / 3) :
1 Real.logb 4 (4 * ω / δ)

Helper for Proposition 4.1.18: once a positive gap is already below ω / 3, the normalized tail potential logb 4 ((4 * ω) / δ) starts at least at level 1.

theorem secondPhaseTailBudgetFromPotentialTripleGrowth {gap : } {ω ε : } {k0 : } ( : 0 < ω) ( : ε Set.Ioc 0 ω) (hpot : ∀ (j : ), 0 < gap (k0 + j)3 ^ j Real.logb 4 (4 * ω / gap (k0 + j))) :
∃ (m : ), m 1 + Real.logb 3 (Real.logb 4 (1 / ε) + Real.logb 4 (4 * ω)) gap (k0 + m) ε

Helper for Proposition 4.1.18: if a second-phase potential owner already grows like 3 ^ j, then the tail reaches the displayed target within the source-facing logb 3 term up to the unavoidable integer ceiling slack.

theorem secondPhaseLog4Potential_geometricLowerBound {gap : } {ω : } {k0 : } ( : 0 < ω) (hmono : Antitone gap) (hgap_pos : ∀ (j : ), 0 < gap (k0 + j)) (hk0 : gap k0 ω / 3) (hstep : ∀ (k : ), gap k ωgap (k + 1) 1 / 2 * gap k * (gap k / ω)) (j : ) :
(3 / 2) ^ j * Real.logb 4 (4 * ω / gap k0) Real.logb 4 (4 * ω / gap (k0 + j))

Helper for Proposition 4.1.18: along any positive second-phase tail, the normalized logb 4 potential grows at least geometrically with ratio 3 / 2.

theorem secondPhaseLog4Potential_geometricLowerBound_of_final_pos {gap : } {ω : } {k0 j : } ( : 0 < ω) (hmono : Antitone gap) (hk0 : gap k0 ω / 3) (hstep : ∀ (k : ), gap k ωgap (k + 1) 1 / 2 * gap k * (gap k / ω)) (hgap_final : 0 < gap (k0 + j)) :
(3 / 2) ^ j * Real.logb 4 (4 * ω / gap k0) Real.logb 4 (4 * ω / gap (k0 + j))

Helper for Proposition 4.1.18: if a monotone second-phase tail is still positive at the chosen horizon j, then the same logb 4 potential already satisfies the geometric lower bound at that horizon.

theorem secondPhase_step_le_four_twentysevenths_of_four_ninths_threshold {δ δNext ω : } ( : 0 < ω) (hδ_nonneg : 0 δ) (hthreshold : δ 4 / 9 * ω) (hstep : δNext 1 / 2 * δ * (δ / ω)) :
δNext 4 / 27 * ω

Helper for Proposition 4.1.18: under the chapter's second-phase recurrence, entering the threshold δ ≤ (4 / 9) * ω only certifies the next gap at the sharper level δNext ≤ (4 / 27) * ω.

theorem secondPhaseTailHitFromThirdThreshold_threeHalvesBudget {gap : } {ω ε : } {k0 : } ( : 0 < ω) ( : ε Set.Ioc 0 ω) (hmono : Antitone gap) (hk0 : gap k0 ω / 3) (hstep : ∀ (k : ), gap k ωgap (k + 1) 1 / 2 * gap k * (gap k / ω)) :
∃ (m : ), m 1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (4 * ω)) gap (k0 + m) ε

Helper for Proposition 4.1.18: the exact second-phase potential owner proved in this file reaches the target ε with the valid base-3 / 2 double-logarithmic budget.

theorem secondPhaseTailHitFromThirdThreshold {gap : } {ω ε : } {k0 : } ( : 0 < ω) ( : ε Set.Ioc 0 ω) (hmono : Antitone gap) (hk0 : gap k0 ω / 3) (hstep : ∀ (k : ), gap k ωgap (k + 1) 1 / 2 * gap k * (gap k / ω)) :
∃ (m : ), m 1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (4 * ω)) gap (k0 + m) ε

Helper for Proposition 4.1.18: once a monotone nonnegative gap sequence is already below ω / 3, the second-phase superlinear recurrence reaches any target ε ∈ (0, ω] within the displayed double-logarithmic budget.

theorem secondPhaseTailHitFromThirdThreshold_exactModel_threeHalvesBudget {gap : } {ω ε : } {k0 : } ( : 0 < ω) ( : ε Set.Ioc 0 ω) (hmono : Antitone gap) (hgap_nonneg : ∀ (k : ), 0 gap k) (hk0 : gap k0 ω / 3) (hstep : ∀ (k : ), gap k ωgap (k + 1) 1 / 3 * gap k * (gap k / ω)) :
∃ (m : ), m 1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (4 * ω)) gap (k0 + m) ε

Helper for Proposition 4.1.18: the exact 1 / 3 local-model tail step still yields the already verified base-3 / 2 tail budget once the tail gaps are known to stay nonnegative.

theorem secondPhaseTailHitFromThirdThreshold_baseThreeBudget {gap : } {ω ε : } {k0 : } ( : 0 < ω) ( : ε Set.Ioc 0 ω) (hmono : Antitone gap) (hk0 : gap k0 ω / 3) (hstep : ∀ (k : ), gap k ωgap (k + 1) 1 / 3 * gap k * (gap k / ω)) :
∃ (m : ), m 1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (4 * ω)) gap (k0 + m) ε

Helper for Proposition 4.1.18: the currently verified proposition-level tail budget uses the base-(3 / 2) double-logarithmic owner once the trajectory has entered ω / 3.

theorem threePhaseIterationBound_withTwoSlack {a t : } {k1 m n : } (hk1 : k1 1 + 3 * a) (hm : m 13 / 4 * a) (hn : n 1 + t) :
(k1 + m + n) 2 + 25 / 4 * a + t

Helper for Proposition 4.1.18: the currently proved three-phase witness bounds add up to the displayed 25 / 4 coefficient together with two unavoidable ceiling slacks coming from the intermediate-entry and tail witnesses.

theorem threePhaseIterationBound_of_sharpMiddleBudget {a t : } {k1 m n : } (hk1 : k1 1 + 3 * a) (hm : m 13 / 4 * a - 2) (hn : n 1 + t) :
(k1 + m + n) 25 / 4 * a + t

Helper for Proposition 4.1.18: if the middle phase absorbs the two ceiling slacks left by the intermediate-entry and tail witnesses, then the three-phase witness closes the displayed 25 / 4 bound exactly.

theorem threePhaseIterationBound_of_verifiedMiddleBudget {a t : } {k1 m n : } (hk1 : k1 1 + 3 * a) (hm : m 13 / 4 * a - 1) (hn : n 1 + t) :
(k1 + m + n) 25 / 4 * a + (1 + t)

Helper for Proposition 4.1.18: for the verified repaired bound, the middle phase only needs to absorb one of the two ceiling slacks, leaving the tail's unavoidable +1 term explicit.

theorem logb_three_le_one_add_logb_threeHalves_of_one_le {target : } (htarget : 1 target) :
Real.logb 3 target 1 + Real.logb (3 / 2) target

Helper for Proposition 4.1.18: on targets at least 1, the sharper source-facing logb 3 term is dominated by the verified same-file 1 + logb (3 / 2) tail owner.

theorem logb_three_lt_one_add_logb_threeHalves_of_one_le {target : } (htarget : 1 target) :
Real.logb 3 target < 1 + Real.logb (3 / 2) target

Helper for Proposition 4.1.18: on every target at least 1, the currently verified 1 + logb (3 / 2) tail owner is strictly larger than the source-facing logb 3 term. This records the exact tail mismatch left in the two historical theorems.

theorem strongConvexDisplayedBound_smallChi_lt_one :
25 / 4 * (1 / 10 * (1 / 10) / 1) + Real.logb 3 (Real.logb 4 (1 / (50 / 9)) + Real.logb 4 (2 * 1 ^ 3 / (9 * (1 / 10) ^ 2))) < 1

Helper for Proposition 4.1.18: on the small-χ scalar parameters μ = 1, L = D = 1 / 10, and ε = 50 / 9, the displayed plain bound is already strictly below 1.

theorem strongConvex_initialHit_of_subunitDisplayedBudget {L : NNReal} {μ D ε : } {n : } {f : EuclideanSpace (Fin n)} {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L✝ : NNReal} {x0 xStar : EuclideanSpace (Fin n)} {ε✝ : } {N : } (method : CubicRegularizationMethod f stepMap L0 (↑L✝) x0) (hN : IsLeast {k : | f (method.x k) - f xStar ε✝} N) (hbound : N 25 / 4 * (L * D / μ) + Real.logb 3 (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2)))) (hB : 25 / 4 * (L * D / μ) + Real.logb 3 (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2))) < 1) :
f (method.x 0) - f xStar ε✝

Helper for Proposition 4.1.18: any proof of the displayed plain global bound on a subunit budget branch already forces the initial iterate to hit the target accuracy.

theorem strong_convex_cubic_regularization_threshold_eq_four_ninths_barOmega {L : NNReal} {μ : } :
μ ^ 3 / (18 * L ^ 2) = 4 / 9 * starConvexNondegenerateBarOmega (↑L) μ

Helper for Proposition 4.1.18: the proposition threshold ω₀ = μ^3 / (18 L^2) is exactly the (4 / 9)-fraction of the star-convex first/second-phase scale μ^3 / (8 L^2).

theorem strong_convex_displayed_tail_constant_eq_four_threshold {L : NNReal} {μ : } :
2 * μ ^ 3 / (9 * L ^ 2) = 4 * (μ ^ 3 / (18 * L ^ 2))

Helper for Proposition 4.1.18: the displayed logarithmic tail constant is exactly 4 * ω₀ in the plain strong-convex setting.

theorem one_le_strongConvexDisplayedTailTarget {L : NNReal} {μ ε : } ( : ε Set.Ioc 0 (μ ^ 3 / (18 * L ^ 2))) :
1 Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2))

Helper for Proposition 4.1.18: the plain displayed logarithmic target is always at least 1 whenever ε ∈ (0, ω₀].

theorem strong_convex_star_convex_with_respect_to_minimizer {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {μ : } {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) :
StarConvexWithRespectToOn f xStar Set.univ

Helper for Proposition 4.1.18: a global minimizer of a strongly convex function is a valid star center on the whole space.

theorem strong_convex_uses_constant_of_isMinOn {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {μ : } {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) :

Helper for Proposition 4.1.18: strong convexity and a chosen global minimizer produce the canonical quadratic-growth witness UsesConstant Set.univ f xStar μ.

theorem strongConvex_gap_controls_distance {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (k : ) :
method.x k - xStar (2 / μ * (f (method.x k) - f xStar))

Helper for Proposition 4.1.18: strong convexity turns the current objective gap into the distance bound ‖method k - xStar‖ ≤ sqrt ((2 / μ) * (f (method k) - f xStar)).

theorem strongConvex_gap_succ_le_alpha_local_model {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (k : ) {α : } ( : α Set.Icc 0 1) :
f (method.x (k + 1)) - f xStar (1 - α) * (f (method.x k) - f xStar) + L / 2 * α ^ 3 * (2 / μ * (f (method.x k) - f xStar)) ^ 3

Helper for Proposition 4.1.18: the cubic feasible-comparison estimate plus strong convexity at xStar yield the local scalar model used in the strong first-phase count.

theorem strongConvex_sqrt_threshold_normalization {L : NNReal} {μ : } ( : 0 < μ) {gap : } (hgap : 0 gap) :
(gap / (μ ^ 3 / (18 * L ^ 2))) = 3 * L / μ * (2 / μ * gap)

Helper for Proposition 4.1.18: the proposition threshold ω₀ = μ^3 / (18 L^2) converts the strong-convexity radius Real.sqrt ((2 / μ) * gap) into the normalized scalar Real.sqrt (gap / ω₀).

theorem strongConvex_gap_succ_le_omega0NormalizedModel {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (k : ) {α : } ( : α Set.Icc 0 1) :
f (method.x (k + 1)) - f xStar (1 - α + 1 / 3 * α ^ 3 * ((f (method.x k) - f xStar) / (μ ^ 3 / (18 * L ^ 2)))) * (f (method.x k) - f xStar)

Helper for Proposition 4.1.18: the local comparison inequality rewrites entirely in terms of the normalized plain gap (f (method k) - f xStar) / ω₀.

theorem strong_convex_gap_succ_le_second_phase_superlinear {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} (hstar : StarConvexWithRespectToOn f xStar Set.univ) (hnondegenerate : HasGloballyNondegenerateOptimalSet.UsesConstant Set.univ f xStar μ) (k : ) (hk : f (method.x k) - f xStar μ ^ 3 / (18 * L ^ 2)) :
f (method.x (k + 1)) - f xStar 1 / 2 * (f (method.x k) - f xStar) * ((f (method.x k) - f xStar) / starConvexNondegenerateBarOmega (↑L) μ)

Helper for Proposition 4.1.18: once the plain strong-convex cubic-regularization gap reaches the proposition threshold ω₀, the accepted step satisfies the chapter's second-phase superlinear estimate with the natural scale μ^3 / (8 L^2).

theorem strongConvex_gap_succ_le_threshold_third {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} (hstar : StarConvexWithRespectToOn f xStar Set.univ) (hnondegenerate : HasGloballyNondegenerateOptimalSet.UsesConstant Set.univ f xStar μ) (hω0 : 0 < μ ^ 3 / (18 * L ^ 2)) (k : ) (hk : f (method.x k) - f xStar μ ^ 3 / (18 * L ^ 2)) :
f (method.x (k + 1)) - f xStar μ ^ 3 / (18 * L ^ 2) / 3

Helper for Proposition 4.1.18: once the plain trajectory reaches ω₀, the current second-phase API only certifies a one-step drop to ω₀ / 3.

theorem strongConvexTailHitFromThirdThreshold {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ ε : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * L ^ 2))) (k0 : ) (hk0 : f (method.x k0) - f xStar μ ^ 3 / (18 * L ^ 2) / 3) :
∃ (m : ), m 1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2))) f (method.x (k0 + m)) - f xStar ε

Helper for Proposition 4.1.18: once the plain strong-convex gap is already at the sharper threshold ω₀ / 3, the remaining tail to accuracy ε is controlled by the shared double-logarithmic estimate.

theorem strongConvexTailHitFromThirdThresholdBaseThree {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ ε : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * L ^ 2))) (k0 : ) (hk0 : f (method.x k0) - f xStar μ ^ 3 / (18 * L ^ 2) / 3) :
∃ (m : ), m 1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2))) f (method.x (k0 + m)) - f xStar ε

Helper for Proposition 4.1.18: the currently verified plain tail term uses the base-(3 / 2) logarithm once the gap is below ω₀ / 3.

theorem strongConvex_gap_one_le_half_LD_cube {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {D : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} (hlevel : ∀ ⦃x : E⦄, f x f x0x - xStar D) :
f (method.x 1) - f xStar L / 2 * D ^ 3

Helper for Proposition 4.1.18: comparing the first accepted cubic step with the minimizer xStar immediately yields the coarse source scale (L / 2) * D^3.

theorem strongConvex_gap_one_le_threshold_third_of_characteristic_le_third {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ D : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} ( : 0 < μ) (hL : 0 < L) (hlevel : ∀ ⦃x : E⦄, f x f x0x - xStar D) ( : L * D / μ 1 / 3) :
f (method.x 1) - f xStar μ ^ 3 / (18 * L ^ 2) / 3

Helper for Proposition 4.1.18: if the characteristic ratio χ = (L * D) / μ is at most 1 / 3, then the first cubic-regularization step already reaches the sharper threshold ω₀ / 3.

theorem strongConvex_gap_succ_le_alpha_boundedRadiusModel {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ D : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) (hlevel : ∀ ⦃x : E⦄, f x f x0x - xStar D) (k : ) {α : } ( : α Set.Icc 0 1) :
f (method.x (k + 1)) - f xStar (1 - α) * (f (method.x k) - f xStar) + L / 2 * α ^ 3 * D ^ 3

Helper for Proposition 4.1.18: the initial bounded-sublevel radius D gives the plain one-step gap model Δₖ₊₁ ≤ (1 - α) Δₖ + (L / 2) α^3 D^3 used for the inverse-square phase.

theorem strongConvex_shiftedGap_le_inverse_square_rate {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ D : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) (xStar : E) ( : 0 < μ) (hf_hessian : f {f : E | HasLipschitzContinuousHessian L f}) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (hlevel : ∀ ⦃x : E⦄, f x f x0x - xStar D) (j : ) :
f (method.x (j + 1)) - f xStar 3 * L * D ^ 3 / (2 * (1 + j / 3) ^ 2)

Helper for Proposition 4.1.18: after the first accepted cubic step, the shifted plain strong-convex gaps satisfy the inverse-square bound with scale (3 / 2) * L * D^3.

theorem strongConvexExistsIntermediateEntryIndex {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ D : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) (xStar : E) ( : 0 < μ) (hf_hessian : f {f : E | HasLipschitzContinuousHessian L f}) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (hlevel : ∀ ⦃x : E⦄, f x f x0x - xStar D) :
∃ (k1 : ), k1 1 + 3 * (L * D / μ) f (method.x k1) - f xStar 3 / 2 * μ * D ^ 2

Helper for Proposition 4.1.18: the first-step estimate together with the shifted inverse-square phase already yields an explicit entry index for the intermediate strong-convex scale (3 / 2) * μ * D^2.

theorem strongConvex_largePhase_alpha_mem {β : } ( : 1 β) :
1 / β Set.Icc 0 1

Helper for Proposition 4.1.18: in the normalized large-phase variables, the endpoint choice α = 1 / β is feasible whenever β ≥ 1.

theorem strongConvex_threshold_mul_div_cancel {L : NNReal} {μ gap : } (hω0 : 0 < μ ^ 3 / (18 * L ^ 2)) :
μ ^ 3 / (18 * L ^ 2) * (gap / (μ ^ 3 / (18 * L ^ 2))) = gap

Helper for Proposition 4.1.18: the positive threshold ω₀ cancels against its normalized quotient.

theorem strongConvex_largePhase_scalar_step {β : } ( : 1 β) :
(1 - 1 / β + 1 / 3 * (1 / β) ^ 3 * β ^ 2) * β ^ 4 (β - 1 / 6) ^ 4

Helper for Proposition 4.1.18: in normalized fourth-root variables, the strong-convex large-phase scalar model drops by 1 / 6 in one step.

theorem strongConvex_largePhase_step_rpow_drop {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (hω0 : 0 < μ ^ 3 / (18 * L ^ 2)) (k : ) (hk : μ ^ 3 / (18 * L ^ 2) f (method.x k) - f xStar) :
((f (method.x (k + 1)) - f xStar) / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) ((f (method.x k) - f xStar) / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) - 1 / 6

Helper for Proposition 4.1.18: while the gap stays above ω₀, the normalized fourth root of the strong-convex gap drops by 1 / 6 at each step.

theorem strongConvex_firstPhase_gap_rpow_bound_fromCurrentGap {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (hω0 : 0 < μ ^ 3 / (18 * L ^ 2)) (k : ) {g : } (hg : f (method.x k) - f xStar g) (j : ) :
μ ^ 3 / (18 * L ^ 2) f (method.x (k + j)) - f xStar((f (method.x (k + j)) - f xStar) / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) (g / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) - j / 6

Helper for Proposition 4.1.18: starting from any current gap upper bound g, the normalized fourth root of the plain strong-convex gap decreases linearly while the trajectory remains above the threshold ω₀.

theorem strongConvexThresholdHitFromCurrentGapBound {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (hω0 : 0 < μ ^ 3 / (18 * L ^ 2)) (k : ) {g : } (hg : f (method.x k) - f xStar g) :
∃ (m : ), m 6 * (g / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) f (method.x (k + m)) - f xStar μ ^ 3 / (18 * L ^ 2)

Helper for Proposition 4.1.18: from any current upper bound g on the plain strong-convex gap, one reaches the threshold ω₀ within at most 6 * (g / ω₀)^(1/4) further steps.

theorem strongConvex_currentMiddleBudgetRatio_eq {L : NNReal} {μ D : } ( : 0 < μ) (hL : 0 < L) :
3 / 2 * μ * D ^ 2 / (μ ^ 3 / (18 * L ^ 2)) = 27 * (L * D / μ) ^ 2

Helper for Proposition 4.1.18: after normalizing the current strong-convex first/second-phase budget, the quarter-root term is exactly 27 * χ^2. This isolates the quantitative normal form of the verified middle-phase route.

theorem coarseMiddleBudget_gt_displayedMiddleBudget {a : } (ha : 0 a) :
13 / 4 * a < 1 + 6 * (27 * a ^ 2).rpow (1 / 4)

Helper for Proposition 4.1.18: the current coarse middle-phase owner 1 + 6 * (27 * a^2)^(1/4) is strictly larger than the displayed 13 / 4 * sqrt a budget on every nonnegative parameter a. This is the scalar obstruction behind the historical 25 / 4 prefix coefficient.

theorem strongConvex_currentMiddleBudget_gt_displayedMiddleBudget {L : NNReal} {μ D : } ( : 0 < μ) (hL : 0 < L) (hD_nonneg : 0 D) :
13 / 4 * (L * D / μ) < 1 + 6 * (3 / 2 * μ * D ^ 2 / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4)

Helper for Proposition 4.1.18: after rewriting the verified middle-phase ratio in terms of χ, the current first/second-phase route is strictly larger than the displayed 13 / 4 * sqrt χ budget. Arithmetic alone therefore cannot close the historical prefix coefficient from this route.

theorem strongConvexThirdThresholdBudgetFromIntermediateGap {n : } {f : EuclideanSpace (Fin n)} {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {x0 : EuclideanSpace (Fin n)} {μ D : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) (xStar : EuclideanSpace (Fin n)) ( : 0 < μ) (hL : 0 < L) (hf_hessian : f {f : EuclideanSpace (Fin n) | HasLipschitzContinuousHessian L f}) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) {k1 : } (hk1 : f (method.x k1) - f xStar 3 / 2 * μ * D ^ 2) :
∃ (m : ), m 1 + 6 * (3 / 2 * μ * D ^ 2 / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) f (method.x (k1 + m)) - f xStar μ ^ 3 / (18 * L ^ 2) / 3

Helper for Proposition 4.1.18: after the inverse-square entry phase has already reduced the plain gap to (3 / 2) * μ * D^2, the verified first-phase threshold hit plus one local ω₀ -> ω₀ / 3 step give a coarse witness for the sharper threshold.

theorem strongConvexThirdThresholdWitnessFromIntermediateGap {n : } {f : EuclideanSpace (Fin n)} {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {x0 : EuclideanSpace (Fin n)} {μ D : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) (xStar : EuclideanSpace (Fin n)) ( : 0 < μ) (hL : 0 < L) (hf_hessian : f {f : EuclideanSpace (Fin n) | HasLipschitzContinuousHessian L f}) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) {k1 : } (hk1 : f (method.x k1) - f xStar 3 / 2 * μ * D ^ 2) :
∃ (m : ), m 1 + 6 * (3 / 2 * μ * D ^ 2 / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) f (method.x (k1 + m)) - f xStar μ ^ 3 / (18 * L ^ 2) / 3

Helper for Proposition 4.1.18: in the large-χ branch where the displayed middle-phase budget is at least 2, after the inverse-square entry phase has already reduced the plain gap to (3 / 2) * μ * D^2, the existing coarse threshold witness follows from the sharp middle-phase budget by a one-time scalar domination.

theorem strongConvexMiddlePhaseHitThirdThreshold {n : } {f : EuclideanSpace (Fin n)} {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {x0 : EuclideanSpace (Fin n)} {μ D : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) (xStar : EuclideanSpace (Fin n)) ( : 0 < μ) (hL : 0 < L) (hf_hessian : f {f : EuclideanSpace (Fin n) | HasLipschitzContinuousHessian L f}) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) {k1 : } (hk1 : f (method.x k1) - f xStar 3 / 2 * μ * D ^ 2) :
∃ (m : ), m 1 + 6 * (3 / 2 * μ * D ^ 2 / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) f (method.x (k1 + m)) - f xStar μ ^ 3 / (18 * L ^ 2) / 3

Helper for Proposition 4.1.18: the middle-phase route currently closes through the same coarse ω₀ -> ω₀ / 3 witness extracted from the verified first-phase threshold hit.

theorem strongConvexPrefixHitThirdThreshold {n : } {f : EuclideanSpace (Fin n)} {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {x0 : EuclideanSpace (Fin n)} {μ D : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) (xStar : EuclideanSpace (Fin n)) ( : 0 < μ) (hL : 0 < L) (hf_hessian : f {f : EuclideanSpace (Fin n) | HasLipschitzContinuousHessian L f}) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (hD : IsGreatest ((fun (x : EuclideanSpace (Fin n)) => x - xStar) '' {x : EuclideanSpace (Fin n) | f x f x0}) D) :
∃ (k0 : ), f (method.x k0) - f xStar μ ^ 3 / (18 * L ^ 2) / 3

Helper for Proposition 4.1.18: even on the large-prefix-budget branch, the currently verified plain prefix route only supplies existence of a witness reaching the sharper threshold ω₀ / 3; the displayed 25 / 4 * sqrt χ - 1 budget remains the statement-side gap.

theorem StrongConvexBound.strongConvex_historicalGlobalIterationBound_of_budgetedPrefix {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ D ε : } {N : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} (hN : IsLeast {k : | f (method.x k) - f xStar ε} N) (hprefix : ∃ (k0 : ), k0 25 / 4 * (L * D / μ) f (method.x k0) - f xStar μ ^ 3 / (18 * L ^ 2) / 3) (htail : ∀ {k0 : }, f (method.x k0) - f xStar μ ^ 3 / (18 * L ^ 2) / 3∃ (nTail : ), nTail 25 / 4 * (L * D / μ) + Real.logb 3 (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2))) - 25 / 4 * (L * D / μ) f (method.x (k0 + nTail)) - f xStar ε) :
N 25 / 4 * (L * D / μ) + Real.logb 3 (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2)))

Helper for Proposition 4.1.18: once the plain branch supplies a prefix witness at the threshold ω₀ / 3 within the displayed 25 / 4 * sqrt χ budget and the exact source-facing tail owner from that witness, the historical global bound is just least-index assembly.

theorem StrongConvexBound.le_globalIterationBound {n : } {f : EuclideanSpace (Fin n)} {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {x0 : EuclideanSpace (Fin n)} {μ D ε : } {N : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) (xStar : EuclideanSpace (Fin n)) ( : 0 < μ) (hL : 0 < L) (hf_hessian : f {f : EuclideanSpace (Fin n) | HasLipschitzContinuousHessian L f}) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (hD : IsGreatest ((fun (x : EuclideanSpace (Fin n)) => x - xStar) '' {x : EuclideanSpace (Fin n) | f x f x0}) D) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * L ^ 2))) (hN : IsLeast {k : | f (method.x k) - f xStar ε} N) :
N 2 + 3 * (L * D / μ) + 6 * (3 / 2 * μ * D ^ 2 / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) + (1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2))))

Proposition 4.1.18 (1): verified public theorem closing the dependency-closed plain bound coming from the same-file first/second-phase route.

theorem StrongConvexBound.sourceStatement {n : } {f : EuclideanSpace (Fin n)} {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {x0 : EuclideanSpace (Fin n)} {μ D ε : } {N : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) (xStar : EuclideanSpace (Fin n)) ( : 0 < μ) (hL : 0 < L) (hf_hessian : f {f : EuclideanSpace (Fin n) | HasLipschitzContinuousHessian L f}) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (hD : IsGreatest ((fun (x : EuclideanSpace (Fin n)) => x - xStar) '' {x : EuclideanSpace (Fin n) | f x f x0}) D) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * L ^ 2))) (hN : IsLeast {k : | f (method.x k) - f xStar ε} N) :
N 2 + 3 * (L * D / μ) + 6 * (3 / 2 * μ * D ^ 2 / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) + (1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2))))

Wrapper for Proposition 4.1.18 (1): theorem-shaped public entry exposing the current same-file plain bound.

theorem StrongConvexBound.le_currentRouteGlobalIterationBound {n : } {f : EuclideanSpace (Fin n)} {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {x0 : EuclideanSpace (Fin n)} {μ D ε : } {N : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) (xStar : EuclideanSpace (Fin n)) ( : 0 < μ) (hL : 0 < L) (hf_hessian : f {f : EuclideanSpace (Fin n) | HasLipschitzContinuousHessian L f}) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (hD : IsGreatest ((fun (x : EuclideanSpace (Fin n)) => x - xStar) '' {x : EuclideanSpace (Fin n) | f x f x0}) D) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * L ^ 2))) (hN : IsLeast {k : | f (method.x k) - f xStar ε} N) :
N 2 + 3 * (L * D / μ) + 6 * (3 / 2 * μ * D ^ 2 / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) + (1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2))))

Helper for Proposition 4.1.18: the current dependency-closed plain route already yields a fully verified coarse global bound by combining the explicit intermediate-entry witness, the current ω₀ -> ω₀ / 3 prefix witness, and the repaired base-(3 / 2) tail theorem.

theorem StrongConvexBound.verifiedGlobalIterationBound_of_budgetedPrefix {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {f : E} {stepMap : EE} {L0 : } {L : NNReal} {x0 : E} {μ D ε : } {N : } [HessianLipschitzOn L Set.univ f] (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) {xStar : E} ( : 0 < μ) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * L ^ 2))) (hN : IsLeast {k : | f (method.x k) - f xStar ε} N) (hprefix : ∃ (k0 : ), k0 25 / 4 * (L * D / μ) f (method.x k0) - f xStar μ ^ 3 / (18 * L ^ 2) / 3) :
N 25 / 4 * (L * D / μ) + (1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2))))

Helper for Proposition 4.1.18: once the plain branch provides a prefix witness k₀ with the displayed 25 / 4 * sqrt χ budget and gap threshold ω₀ / 3, the verified same-file theorem is just least-index assembly plus the repaired base-(3 / 2) tail owner.

theorem StrongConvexBound.le_verifiedGlobalIterationBound {n : } {f : EuclideanSpace (Fin n)} {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {x0 : EuclideanSpace (Fin n)} {μ D ε : } {N : } (method : CubicRegularizationMethod f stepMap L0 (↑L) x0) (xStar : EuclideanSpace (Fin n)) ( : 0 < μ) (hL : 0 < L) (hf_hessian : f {f : EuclideanSpace (Fin n) | HasLipschitzContinuousHessian L f}) (hf_strong : StrongConvexOn Set.univ μ f) (hxStar : IsMinOn f Set.univ xStar) (hD : IsGreatest ((fun (x : EuclideanSpace (Fin n)) => x - xStar) '' {x : EuclideanSpace (Fin n) | f x f x0}) D) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * L ^ 2))) (hN : IsLeast {k : | f (method.x k) - f xStar ε} N) :
N 2 + 3 * (L * D / μ) + 6 * (3 / 2 * μ * D ^ 2 / (μ ^ 3 / (18 * L ^ 2))).rpow (1 / 4) + (1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * L ^ 2))))

Canonical verified repaired companion: this is the fully checked same-file plain global bound currently available in the dependency-closed route.

theorem nonlinearTransformation_initialHit_of_subunitDisplayedBudget {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] (problem : NonlinearConvexTransformation E) {L : NNReal} {μ ε : } {n : } (problem✝ : NonlinearConvexTransformation (EuclideanSpace (Fin n))) {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L✝ : NNReal} {ε✝ : } {N : } (method : CubicRegularizationMethod problem✝.objective stepMap L0 (↑L✝) problem✝.x0) (hN : IsLeast {k : | problem✝.objective (method.x k) - problem✝.objective problem✝.xStar ε✝} N) (hbound : N 25 / 4 * (problem.sigma / μ * L * problem.D) + Real.logb 3 (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2)))) (hB : 25 / 4 * (problem.sigma / μ * L * problem.D) + Real.logb 3 (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2))) < 1) :
problem✝.objective (method.x 0) - problem✝.objective problem✝.xStar ε✝

Helper for Proposition 4.1.18: any proof of the displayed transformed global bound on a subunit budget branch already forces the initial transformed iterate to hit the target accuracy.

theorem nonlinear_transformation_threshold_eq_four_ninths_local_scale {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] (problem : NonlinearConvexTransformation E) {L : NNReal} {μ : } :
μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2) = 4 / 9 * (μ ^ 3 / (8 * problem.sigma ^ 6 * L ^ 2))

Helper for Proposition 4.1.18: the transformed proposition threshold μ^3 / (18 * σ^6 * L^2) is the (4 / 9)-fraction of the natural local superlinear scale μ^3 / (8 * σ^6 * L^2).

theorem nonlinear_transformation_displayed_tail_constant_eq_four_threshold {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] (problem : NonlinearConvexTransformation E) {L : NNReal} {μ : } :
2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2) = 4 * (μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2))

Helper for Proposition 4.1.18: the displayed transformed logarithmic tail constant is exactly 4 * ω₀.

theorem one_le_nonlinearTransformationDisplayedTailTarget {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] {L : NNReal} {μ ε : } (problem : NonlinearConvexTransformation E) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2))) :
1 Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2))

Helper for Proposition 4.1.18: the transformed displayed logarithmic target is always at least 1 whenever ε ∈ (0, ω₀].

theorem nonlinear_transformation_gap_succ_le_local_superlinear_model {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] (problem : NonlinearConvexTransformation E) {stepMap : EE} {L0 : } {L : NNReal} {𝓕 : Set E} {μ : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) (k : ) :
problem.objective (method.x (k + 1)) - problem.objective problem.xStar L / 2 * (problem.sigma * (2 / μ * (problem.objective (method.x k) - problem.objective problem.xStar))) ^ 3

Helper for Proposition 4.1.18: evaluating the transformed local comparison model at α = 1 removes the convex-combination term and leaves the local cubic superlinear bound around problem.xStar.

theorem nonlinear_transformation_gap_succ_le_second_phase_superlinear {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] (problem : NonlinearConvexTransformation E) {stepMap : EE} {L0 : } {L : NNReal} {𝓕 : Set E} {μ : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) (k : ) :
problem.objective (method.x (k + 1)) - problem.objective problem.xStar 1 / 2 * (problem.objective (method.x k) - problem.objective problem.xStar) * ((problem.objective (method.x k) - problem.objective problem.xStar) / (μ ^ 3 / (8 * problem.sigma ^ 6 * L ^ 2)))

Helper for Proposition 4.1.18: the transformed local cubic model rewrites into the same normalized superlinear recurrence as the plain strong-convex theorem, with the proposition scale μ^3 / (8 σ^6 L^2).

theorem nonlinearTransformation_gap_succ_le_threshold_third {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] (problem : NonlinearConvexTransformation E) {stepMap : EE} {L0 : } {L : NNReal} {𝓕 : Set E} {μ : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) (hω0 : 0 < μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2)) (k : ) (hk : problem.objective (method.x k) - problem.objective problem.xStar μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2)) :
problem.objective (method.x (k + 1)) - problem.objective problem.xStar μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2) / 3

Helper for Proposition 4.1.18: once the transformed trajectory reaches ω₀, the current second-phase API only certifies a one-step drop to ω₀ / 3.

theorem nonlinearTransformationThirdThresholdWitness {n : } (problem : NonlinearConvexTransformation (EuclideanSpace (Fin n))) {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {𝓕 : Set (EuclideanSpace (Fin n))} {μ : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hL : 0 < L) ( : 0 < problem.sigma) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) :
∃ (k0 : ), problem.objective (method.x k0) - problem.objective problem.xStar μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2) / 3

Helper for Proposition 4.1.18: after the transformed trajectory enters the proposition threshold ω₀, one more local step reaches ω₀ / 3. This isolates the verified transformed prefix witness from the still-missing sharp 6.25 * sqrt ((σ / μ) * L * D) arithmetic.

theorem nonlinearTransformation_gap_succ_le_omega0NormalizedModel {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] (problem : NonlinearConvexTransformation E) {stepMap : EE} {L0 : } {L : NNReal} {𝓕 : Set E} {μ : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) (k : ) {α : } ( : α Set.Icc 0 1) :
problem.objective (method.x (k + 1)) - problem.objective problem.xStar (1 - α + 1 / 3 * α ^ 3 * ((problem.objective (method.x k) - problem.objective problem.xStar) / (μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2)))) * (problem.objective (method.x k) - problem.objective problem.xStar)

Helper for Proposition 4.1.18: the transformed local comparison inequality rewrites in terms of the proposition threshold ω₀ = μ^3 / (18 σ^6 L^2), yielding the same normalized scalar surface as in the plain strong-convex branch.

theorem nonlinearTransformationTailHitFromThirdThreshold {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] (problem : NonlinearConvexTransformation E) {stepMap : EE} {L0 : } {L : NNReal} {𝓕 : Set E} {μ ε : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2))) (k0 : ) (hk0 : problem.objective (method.x k0) - problem.objective problem.xStar μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2) / 3) :
∃ (m : ), m 1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2))) problem.objective (method.x (k0 + m)) - problem.objective problem.xStar ε

Helper for Proposition 4.1.18: once the transformed strong-convex gap is already at the sharper threshold ω₀ / 3, the remaining tail to accuracy ε is controlled by the same double-logarithmic estimate.

theorem nonlinearTransformationTailHitFromThirdThresholdBaseThree {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] (problem : NonlinearConvexTransformation E) {stepMap : EE} {L0 : } {L : NNReal} {𝓕 : Set E} {μ ε : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2))) (k0 : ) (hk0 : problem.objective (method.x k0) - problem.objective problem.xStar μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2) / 3) :
∃ (m : ), m 1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2))) problem.objective (method.x (k0 + m)) - problem.objective problem.xStar ε

Helper for Proposition 4.1.18: the currently verified transformed tail term uses the base-(3 / 2) logarithm once the transformed gap is below ω₀ / 3.

theorem nonlinearTransformationPrefixHitThirdThreshold {n : } (problem : NonlinearConvexTransformation (EuclideanSpace (Fin n))) {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {𝓕 : Set (EuclideanSpace (Fin n))} {μ : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hL : 0 < L) ( : 0 < problem.sigma) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) :
∃ (k0 : ), problem.objective (method.x k0) - problem.objective problem.xStar μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2) / 3

Helper for Proposition 4.1.18: the transformed trajectory has a source-compatible witness entering the threshold region ω₀ / 3; the missing part is only the sharp prefix budget.

theorem nonlinearTransformationPrefixHitThirdThresholdWithBudget {n : } (problem : NonlinearConvexTransformation (EuclideanSpace (Fin n))) {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {𝓕 : Set (EuclideanSpace (Fin n))} {μ : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hL : 0 < L) ( : 0 < problem.sigma) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) :
∃ (k0 : ), problem.objective (method.x k0) - problem.objective problem.xStar μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2) / 3

Helper for Proposition 4.1.18: in the transformed large-prefix-budget branch, the currently verified route still only supplies existence of a witness reaching ω₀ / 3; the displayed 25 / 4 * sqrt ((σ / μ) * L * D) - 1 budget remains the open statement-side issue.

theorem TransformedBound.nonlinearTransformation_historicalGlobalIterationBound_of_budgetedPrefix {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] (problem : NonlinearConvexTransformation E) {stepMap : EE} {L0 : } {L : NNReal} {μ ε : } {N : } (problem✝ : NonlinearConvexTransformation E) (method : CubicRegularizationMethod problem✝.objective stepMap L0 (↑L) problem✝.x0) (hN : IsLeast {k : | problem✝.objective (method.x k) - problem✝.objective problem✝.xStar ε} N) (hprefix : ∃ (k0 : ), k0 25 / 4 * (problem✝.sigma / μ * L * problem✝.D) problem✝.objective (method.x k0) - problem✝.objective problem✝.xStar μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2) / 3) (htail : ∀ {k0 : }, problem✝.objective (method.x k0) - problem✝.objective problem✝.xStar μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2) / 3∃ (nTail : ), nTail 25 / 4 * (problem.sigma / μ * L * problem.D) + Real.logb 3 (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2))) - 25 / 4 * (problem✝.sigma / μ * L * problem✝.D) problem✝.objective (method.x (k0 + nTail)) - problem✝.objective problem✝.xStar ε) :
N 25 / 4 * (problem.sigma / μ * L * problem.D) + Real.logb 3 (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2)))

Helper for Proposition 4.1.18: once the transformed branch supplies a threshold witness at ω₀ / 3 within the displayed 25 / 4 * sqrt ((σ / μ) * L * D) budget and the exact source-facing transformed tail owner from that witness, the historical bound is again just least-index assembly.

theorem TransformedBound.le_globalIterationBound {n : } (problem : NonlinearConvexTransformation (EuclideanSpace (Fin n))) {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {𝓕 : Set (EuclideanSpace (Fin n))} {μ ε : } {N : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hL : 0 < L) ( : 0 < problem.sigma) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2))) (hN : IsLeast {k : | problem.objective (method.x k) - problem.objective problem.xStar ε} N) :
N (Nat.find ) + (1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2))))

Proposition 4.1.18 (2): verified public theorem closing the transformed bound obtained from the explicit threshold-entry witness and the repaired base-(3 / 2) tail theorem.

theorem TransformedBound.sourceStatement {n : } (problem : NonlinearConvexTransformation (EuclideanSpace (Fin n))) {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {𝓕 : Set (EuclideanSpace (Fin n))} {μ ε : } {N : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hL : 0 < L) ( : 0 < problem.sigma) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2))) (hN : IsLeast {k : | problem.objective (method.x k) - problem.objective problem.xStar ε} N) :
N (Nat.find ) + (1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2))))

Wrapper for Proposition 4.1.18 (2): theorem-shaped public entry exposing the current same-file transformed bound.

theorem TransformedBound.verifiedGlobalIterationBound_of_budgetedPrefix {E : Type u} [NormedAddCommGroup E] [InnerProductSpace E] [CompleteSpace E] (problem : NonlinearConvexTransformation E) {stepMap : EE} {L0 : } {L : NNReal} {𝓕 : Set E} {μ ε : } {N : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2))) (hN : IsLeast {k : | problem.objective (method.x k) - problem.objective problem.xStar ε} N) (hprefix : ∃ (k0 : ), k0 25 / 4 * (problem.sigma / μ * L * problem.D) problem.objective (method.x k0) - problem.objective problem.xStar μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2) / 3) :
N 25 / 4 * (problem.sigma / μ * L * problem.D) + (1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2))))

Helper for Proposition 4.1.18: once the transformed branch provides a prefix witness k₀ with the displayed 25 / 4 * sqrt ((σ / μ) * L * D) budget and gap threshold ω₀ / 3, the verified same-file theorem is again just least-index assembly plus the repaired base-(3 / 2) tail owner.

theorem TransformedBound.le_verifiedGlobalIterationBound {n : } (problem : NonlinearConvexTransformation (EuclideanSpace (Fin n))) {stepMap : EuclideanSpace (Fin n)EuclideanSpace (Fin n)} {L0 : } {L : NNReal} {𝓕 : Set (EuclideanSpace (Fin n))} {μ ε : } {N : } (hproblem : HessianLipschitzOn L 𝓕 problem.objective) (method : CubicRegularizationMethod problem.objective stepMap L0 (↑L) problem.x0) (hlevel_subset : problem.objective ⁻¹' Set.Iic (problem.objective problem.x0) 𝓕) ( : 0 < μ) (hL : 0 < L) ( : 0 < problem.sigma) (hphi_strong : StrongConvexOn Set.univ μ problem.φ) ( : ε Set.Ioc 0 (μ ^ 3 / (18 * problem.sigma ^ 6 * L ^ 2))) (hN : IsLeast {k : | problem.objective (method.x k) - problem.objective problem.xStar ε} N) :
N (Nat.find ) + (1 + Real.logb (3 / 2) (Real.logb 4 (1 / ε) + Real.logb 4 (2 * μ ^ 3 / (9 * problem.sigma ^ 6 * L ^ 2))))

Canonical verified repaired companion: this is the fully checked same-file transformed global bound currently available in the dependency-closed route.