High-contrast homogenization: Provider.Quenched.SmallContrastDilation #
Imported from the Apache-2.0 HighContrastHomogenization development at commit
7a13dbcd8d6609264a713373f5c69ceeac870472.
The dilation device: the burn-in clock #
Change variables so that the burn-in is one step, then apply the recursion inequality.
With G m := F (s * m) at s the recursion's start, the three groups transform:
- the drop history — the one place a factor could hide, and it does not:
grouping
k ∈ ((j-1)s, js]telescopes toF (s(j-1)) - F (s j) = G (j-1) - G j, ands*m - k ≥ s*(m-j)givesr ^ (s*m-k) ≤ (r ^ s) ^ (m-j), so the dilated sum dominates the original, which is the drop-sum bound below; - the quadratic index —
l' m := 3 * m / 4works, becauses * (3 * m / 4) ≤ 3 * (s * m) / 4andFis antitone andl'satisfies the iteration decay's free index hypothesism ≤ 2 * l' m + 1; - the source — exactly
delta * (r ^ s) ^ m, no bookkeeping at all.
So the general iteration decay applies to G unchanged, at the dilated rate
r ^ s, with delayFactor = 1 — and hsmall in the dilated variable is
law-free.
The cost the sketch does not name, and where it goes #
The lemma requires its rate at most 1/2, so the dilated rate must be taken as
min (alpha * s) (1/2). When alpha * s > 1/2 — which is the case, since
s = Θ(N₁) — the dilated conclusion is at rate 1/(640 A) per dilated step,
i.e. 1/(640 A s) per original generation: law-dependent. The binder fixes
alpha before the law, so the decay rate cannot absorb it, and a single dilated
pass does not close the endpoint.
It closes in two stages, and this is what "the burn-in rides in the starting
threshold" means. Stage 1 (dilated) is used only to reach smallness: it gives
F (s * m) ≤ K · delta · 3 ^ (-m/(640 A)), so at
m₁ := ⌈640 A · log₃ (1/eps)⌉ the excess is below any prescribed eps, at the
scale s * m₁ — law-dependent, but a starting scale, which is exactly the
Cdelay slot the binder sanctions and bounds. Stage 2 restarts the
undilated account at base N₀' := s * m₁, where hinit is supplied by stage
1 rather than by the bootstrap floor — so the recursion tolerance may be taken
as small as
the law requires, hsmall holds, and the conclusion is at the law-free rate
alpha / (320 A).
That is the resolution: the recursion tolerance's lower bound was the bootstrap
floor 5 * cStar, pinned before the law; stage 1 replaces that floor by a
quantity chosen after the law, and the pinning disappears.