A gapless (flat) head is never context stable: no fixed key budget serves all
ProvedAttentionBudget.not_ctxStable_uniformaether-catalogshared
A gapless (flat) head is never context stable: no fixed key budget serves all context lengths.
theorem AttentionBudget.not_ctxStable_uniform{τ : ℝ} (hτ0 : 0 < τ) (hτ : τ ≤ 1) :
¬ CtxStable (fun _ => (1 : ℝ)) τ := by sorry
Formalization Note Transplanted verbatim from the Aether Catalog source Shared/AttentionBudgetScaling.lean; the statement is byte-identical to the source declaration, elaborated with autoImplicit disabled in the platform environment.
Preamble
-- Thm stub generated from Shared/AttentionBudgetScaling.lean
import Mathlib
import Definitions.Def_Shared_AttentionBudgetKnee
import Definitions.Def_Shared_AttentionBudgetScaling
/-!
# Scaling laws for the attention budget: closed form, head merging, and the limits of
"flatness"
This is the second research cycle built on `Shared.AttentionBudgetKnee`. Three
questions left open there are settled here.
1. **Closed form for the universal budget.** `kstar_le_geometricBudget` replaces the
existence statement of cycle 1 by the explicit value
`K(r, τ) = max ⌈log((1-τ)(1-r)) / log r⌉ 1`, valid at *every* context length.
Solving the formula for `r` turns a measured knee into a bound on the decay ratio of
the attention profile, a model-internal quantity: the knee is a *spectrometer*.
2. **Is the flat chain literally flat?** No. `exact_flatness_refuted` computes the
knee of the ideal geometric profile `(1/2)^i` at two context lengths and finds
`k*(1) = 1 ≠ 2 = k*(2)`. The normaliser grows with the context, so the retained
fraction is *antitone* in `n` (`retained_antitone_context`). Hence an observed
`{16, 16}` cannot be upgraded to an equality law; the correct invariant is uniform
boundedness, characterised by `ctxStable_iff_uniform_pass`.
3. **What happens when heads are merged?** Context stability is closed under mixing
(`ctxStable_add`): the knee of a two-head mixture is sandwiched between the two
per-head knees (`min_le_kstar_add`, `kstar_add_le_max`), because retained mass is a
*mediant* of the per-head retained masses (`retained_add_ge_min`,
`retained_add_le_max`). A context-stable model therefore cannot be destabilised by
adding further context-stable heads — while by `kstar_ge_of_bounded_ratio` a single
gapless head already destroys stability (`not_ctxStable_uniform`). Stability is a
property of the *worst* head: a max law, not a sum law.
-- !-- Lab Notes -- !--
Hypothesizer (cycle 2):
(H6) The budget admits a closed form in `(r, τ)` alone — no dependence on context
length or on the head count. [BOLD]
(H7) Exact flatness `k*(2n) = k*(n)` is false even for the ideal geometric
profile; only boundedness survives.
(H8) Context stability is closed under head mixing and is decided by the worst
head (a max law, not a sum law). [BOLD]
Experimenter: H6 = `kstar_le_geometricBudget`; H7 = `exact_flatness_refuted`
(explicit computation: retained mass `2/3 < 3/4` at `k = 1`, `n = 2`, while the same
gate is cleared at `n = 1` by `k = 1`); H8 = `ctxStable_add` via the mediant
inequality. All proved, zero sorries.
Analyst: the mediant inequality is the structural reason a *max* law appears rather
than a sum law: `(A₁+A₂)/(B₁+B₂)` is squeezed between `A₁/B₁` and `A₂/B₂`, so mixing is
never worse than the worst head. This predicts that per-head knees measured separately
should bracket the model-level knee — a directly testable consequence.
Critic: `exact_flatness_refuted` is a genuine falsification, not a corner case: the
profile is the canonical geometric one and the gate `3/4` is interior. Its moral is
that any claimed "flat chain" must be reported as a bracket, exactly as `knee_bracket`
demands.
-/
open AttentionBudget
open Finset
/-! ## Context stability -/
/-! ## A closed form for the universal budget -/
/-! ## Retained mass decays with context length -/
/-! ## H7: exact flatness is false -/
/-! ## H8: merging heads — the mediant law -/
variable {w₁ w₂ : ℕ → ℝ} {τ : ℝ} {n : ℕ}Formal statement
theorem AttentionBudget.not_ctxStable_uniform{τ : ℝ} (hτ0 : 0 < τ) (hτ : τ ≤ 1) :
¬ CtxStable (fun _ => (1 : ℝ)) τ := by sorrySource