Convex gradient descent has a normalized final-value bound
AutoSamplingTheory.TechnicalLemmas.Analysis.GradientDescentRates.convex_value_le · theorem · Teaching coverage
Statement
E is a complete real inner-product space and f:E→R is C1. The gradient is the genuine Riesz gradient of f. For all x,y, f(y)<=f(x)+<gradient f(x),y-x>+(beta/2)||y-x||². The step satisfies h>0 and beta*h<=1; N is a positive natural number. x0 and z are arbitrary points; X_N is exactly N iterations of x ↦ x-h gradient f(x) starting at x0. No minimizer is required for the comparator bound. f is convex on the whole space. No positivity assumption on beta is needed beyond the upper model and admissible step. The following bound holds for the actual final iterate.
All objects and hypotheses
- E is a complete real inner-product space and f:E→R is C1. The gradient is the genuine Riesz gradient of f.
- For all x,y, f(y)<=f(x)+<gradient f(x),y-x>+(beta/2)||y-x||². The step satisfies h>0 and beta*h<=1; N is a positive natural number.
- x0 and z are arbitrary points; X_N is exactly N iterations of x ↦ x-h gradient f(x) starting at x0. No minimizer is required for the comparator bound.
- f is convex on the whole space. No positivity assumption on beta is needed beyond the upper model and admissible step.
Mathematical proof
1. Reuse the actual-iterate energy estimate
Write q=1-alpha h, R=||x0-z|| and e_N=f(X_N)-f(z). The existing weighted theorem already derives objective descent and the forced recurrence for the actual gradient iterates. Invoke it directly; its conclusion also allows negative comparator gaps.
Corresponding Lean step
GradientDescentValue.gradient_descent_weighted_value_bound.
2. Evaluate the zero-curvature sum
For a convex objective set alpha=0, hence q=1. Every summand equals one and there are exactly N summands. Thus the weighted estimate becomes 2hN e_N<=R².
Corresponding Lean step
strongConvexOn_zero; Finset.sum_const, card_range, one_pow and nsmul_eq_mul.
3. Divide on the positive domain
Since h>0 and N>0, the coefficient 2hN is positive. Dividing preserves the inequality. With beta>0, substituting h=1/beta gives exactly beta R²/(2N), the source prescribed-step convex rate. This substitution is exercised by a compiled consumer test.
Corresponding Lean step
le_div_iff₀; Tests.Shared.GradientDescentRates verifies the reciprocal-step specialization.
Lean statement · convex_value_le
Convex gradient descent has a normalized final-value bound.
Braces mark parameters Lean can infer; square brackets request structures such as a measurable space or probability measure. Named hypotheses are mathematical premises, not facts established by this declaration. Section parameters are described in the mathematical hypotheses above; the module link retains their exact source context.
theorem convex_value_le {f : E → ℝ} {β h : ℝ} (hf : ContDiff ℝ 1 f)
(hc : ConvexOn ℝ univ f) (hh : 0 < h) (hstep : β * h ≤ 1)
(hu : ∀ x y, f y ≤ f x + inner ℝ (gradient f x) (y - x) + β / 2 * ‖y - x‖ ^ 2)
(x₀ z : E) {N : ℕ} (hN : 0 < N) :
f ((fun x => x - h • gradient f x)^[N] x₀) - f z ≤ ‖x₀ - z‖ ^ 2 / (2 * h * N)Lean proof · convex_value_le
Write q=1-alpha h, R=||x0-z|| and e_N=f(X_N)-f(z). The existing weighted theorem already derives objective descent and the forced recurrence for the actual gradient iterates. Invoke it directly; its conclusion also allows negative comparator gaps. For a convex objective set alpha=0, hence q=1. Every summand equals one and there are exactly N summands. Thus the weighted estimate becomes 2hN e_N<=R². Since h>0 and N>0, the coefficient 2hN is positive. Dividing preserves the inequality. With beta>0, substituting h=1/beta gives exactly beta R²/(2N), the source prescribed-step convex rate. This substitution is exercised by a compiled consumer test.
Braces mark parameters Lean can infer; square brackets request structures such as a measurable space or probability measure. Named hypotheses are mathematical premises, not facts established by this declaration. Section parameters are described in the mathematical hypotheses above; the module link retains their exact source context.
theorem convex_value_le {f : E → ℝ} {β h : ℝ} (hf : ContDiff ℝ 1 f)
(hc : ConvexOn ℝ univ f) (hh : 0 < h) (hstep : β * h ≤ 1)
(hu : ∀ x y, f y ≤ f x + inner ℝ (gradient f x) (y - x) + β / 2 * ‖y - x‖ ^ 2)
(x₀ z : E) {N : ℕ} (hN : 0 < N) :
f ((fun x => x - h • gradient f x)^[N] x₀) - f z ≤ ‖x₀ - z‖ ^ 2 / (2 * h * N) := by
have he := gradient_descent_weighted_value_bound hf
(strongConvexOn_zero.mpr hc) hh.le hstep (by norm_num : (0 : ℝ) * h ≤ 1) hu x₀ z N
simp only [zero_mul, sub_zero, one_pow, sum_const, card_range, nsmul_eq_mul,
mul_one, one_mul] at he
apply (le_div_iff₀ (by positivity : 0 < 2 * h * (N : ℝ))).mpr
nlinarith [he]
/-- Strongly convex rational rate, and exact inverse-power form on its positive-base domain. -/Scope and omitted-condition boundaries
- These are public integrations of the existing weighted theorem and previously compiled test normalizations. The new representation edge is the exact inverse-power equality; no new descent or recurrence proof is claimed.
- C1 on a complete real Hilbert space and arbitrary comparator generalize the C2 Euclidean minimizer-based source rate. Source smoothness is represented by its global quadratic upper model. The condition beta*h<=1 agrees with h<=1/beta when beta>0.
- Finite normalized bounds require h>0 and N>0. No real-valued normalized h=0 or N=0 assertion, analytic alpha-to-zero limit, iteration-complexity theorem, minimum existence, Riemannian or stochastic result is claimed.
Source and reuse
ASTIS parents called
Mathlib API called (external library)
- strongConvexOn_zero
- Finset.sum_const
- Finset.card_range
- le_div_iff₀
Mathematical sources
- Chewi Theorem3.4 — Normalized source rate and weighted proof; endpoint interpretation explicitly separated.
- Previously compiled normalization proofs — The convex and strong rational algebra existed as test consumers before this public interface.
- Public interface consumer tests — Actual unit quadratic, arbitrary comparators, positive and zero geometric base, reciprocal step and singular inverse-power counterexample.
ASTIS prose is not a quotation or a source-equivalence certificate. Definitions and aliases are explained as constructions, not counted as new mathematical proofs.