Proof of Theorem 2.6, p. 19 — implies
ProvedFastRatesSVM.GeomNoise.tau_estimatelearning-theoryp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1real-analysis
Let , and , and put . For consider the equation in
There is , depending only on , and , such that for every the equation has a solution , and every solution satisfies
This is the choice of the splitting level in display (33): it balances the two terms there, and the resulting bound decays like , which yields every geometric noise exponent .
Formalization Note The page writes "Let us define by … For and small this definition implies …". "Small " is the threshold , chosen before and . The existence of a solution, presupposed by "define by", is stated as the first conjunct; the inequality is stated for every positive solution. For one has , so the logarithm is positive.
Preamble
import Mathlib
Formal statement
namespace FastRatesSVM.GeomNoise
theorem tau_estimate (γ q cγ : ℝ) (hγ : 0 < γ) (hq : 0 ≤ q) (hcγ : 0 < cγ) :
∃ t₀ : ℝ, 0 < t₀ ∧ ∀ t : ℝ, 0 < t → t < t₀ →
(∃ τ : ℝ, 0 < τ ∧ τ ^ (q + 1) = Real.exp (-(τ / cγ) ^ (2 / γ) * t⁻¹)) ∧
∀ τ : ℝ, 0 < τ → τ ^ (q + 1) = Real.exp (-(τ / cγ) ^ (2 / γ) * t⁻¹) →
τ ≤ (cγ ^ (2 / γ) * (q + 1) * γ / 2) ^ (γ / 2) *
(t * Real.log (1 / (cγ ^ (2 / γ) * (q + 1) * t))) ^ (γ / 2) := by sorry
end FastRatesSVM.GeomNoise
Source
Steinwart, Scovel, Fast Rates for Support Vector Machines Using Gaussian Kernels, arXiv:0708.1838v1, p. 19, §4.2, proof of Theorem 2.6, display after (33)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.