Provedrandomized-conversionconversion-theoremattention-compute-lawmedian-of-means

Randomized Conversion Theorem (assembly): E[χ²] ≤ (1+2ε)^L + δ (L+2)^L for the proposal synthesised from a randomized oracle

For all $\htmlData{mx=1}{\Omega \in \htmlData{mx=0}{\operatorname{Type}}}$ , $\htmlData{mx=3}{w, \mathit{chi} : \htmlData{mx=0}{\Omega} \mathrel{\htmlData{mx=2}{\to}} \htmlData{mx=1}{\mathbb{R}}}$ , $\htmlData{mx=1}{L \in \htmlData{mx=0}{\mathbb{N}}}$ , $\htmlData{mx=1}{\varepsilon, \delta \in \htmlData{mx=0}{\mathbb{R}}}$ and $\htmlData{mx=3}{\mathit{good} : \htmlData{mx=0}{\Omega} \mathrel{\htmlData{mx=2}{\to}} \htmlData{mx=1}{\operatorname{Prop}}}$ , if $0 \mathrel{\htmlData{mx=1}{\leq}} \htmlData{mx=0}{\varepsilon}$ , $\htmlData{mx=4}{\forall \omega,\ 0 \mathrel{\htmlData{mx=3}{\leq}} \htmlData{mx=2}{\htmlData{mx=0}{w}(\htmlData{mx=1}{\omega})}}$ , $\htmlData{mx=3}{\sum_{\omega} \htmlData{mx=2}{\htmlData{mx=0}{w}(\htmlData{mx=1}{\omega})}} \mathrel{\htmlData{mx=4}{=}} 1$ , $\htmlData{mx=13}{\forall \omega,\ \htmlData{mx=2}{\htmlData{mx=0}{\mathit{good}}(\htmlData{mx=1}{\omega})} \mathrel{\htmlData{mx=12}{\implies}} \htmlData{mx=5}{\htmlData{mx=3}{\mathit{chi}}(\htmlData{mx=4}{\omega})} \mathrel{\htmlData{mx=11}{\leq}} \htmlData{mx=10}{(\htmlData{mx=8}{1 + \htmlData{mx=7}{2\htmlData{mx=6}{\varepsilon}}})^{\htmlData{mx=9}{L}}}}$ , $\htmlData{mx=9}{\forall \omega,\ \htmlData{mx=2}{\htmlData{mx=0}{\mathit{chi}}(\htmlData{mx=1}{\omega})} \mathrel{\htmlData{mx=8}{\leq}} \htmlData{mx=7}{(\htmlData{mx=5}{\htmlData{mx=4}{\htmlData{mx=3}{L}} + 2})^{\htmlData{mx=6}{L}}}}$ and $\htmlData{mx=10}{\sum_{\omega \in \htmlData{mx=6}{\htmlData{mx=0}{\operatorname{filter}}(\htmlData{mx=5}{\omega \mapsto \htmlData{mx=4}{\lnot \htmlData{mx=3}{\htmlData{mx=1}{\mathit{good}}(\htmlData{mx=2}{\omega})}}})}} \htmlData{mx=9}{\htmlData{mx=7}{w}(\htmlData{mx=8}{\omega})}} \mathrel{\htmlData{mx=12}{\leq}} \htmlData{mx=11}{\delta}$ (given $\htmlData{mx=2}{\htmlData{mx=0}{\operatorname{Fintype}}\,\htmlData{mx=1}{\Omega}}$ and $\htmlData{mx=2}{\htmlData{mx=0}{\operatorname{DecidablePred}}\,\htmlData{mx=1}{\mathit{good}}}$ ), then

\htmlData{mx=7}{\sum_{\omega} \htmlData{mx=6}{\htmlData{mx=2}{\htmlData{mx=0}{w}(\htmlData{mx=1}{\omega})} \cdot \htmlData{mx=5}{\htmlData{mx=3}{\mathit{chi}}(\htmlData{mx=4}{\omega})}}} \mathrel{\htmlData{mx=21}{\leq}} \htmlData{mx=20}{\htmlData{mx=12}{(\htmlData{mx=10}{1 + \htmlData{mx=9}{2\htmlData{mx=8}{\varepsilon}}})^{\htmlData{mx=11}{L}}} + \htmlData{mx=19}{\htmlData{mx=13}{\delta} \cdot \htmlData{mx=18}{(\htmlData{mx=16}{\htmlData{mx=15}{\htmlData{mx=14}{L}} + 2})^{\htmlData{mx=17}{L}}}}}

What the notation means · or point at any symbol

Type: Type: a definition from the library, not special notation. Its name is its meaning; open it in Lean to see how it is defined. (Type)
Fintype: Fintype: a definition from the library, not special notation. Its name is its meaning; open it in Lean to see how it is defined. (Fintype)
f(x)Function application: Fintype applied to Ω. Lean writes f x; papers write f(x). (Fintype Ω)
⟹Implies: If Ω, then ℝ. (Between types, A → B is the type of functions from A to B.) (Ω → ℝ)
x : TVariables: w, chi range over Ω → ℝ: the statement must hold for every choice. ((w chi : Ω → ℝ))
NNatural numbers: The natural numbers 0, 1, 2, …. In Lean they start at 0. (ℕ)
x ∈ TVariables: ε, δ range over real numbers: the statement must hold for every choice. ((ε δ : ℝ))
DecidablePred: DecidablePred: a definition from the library, not special notation. Its name is its meaning; open it in Lean to see how it is defined. (DecidablePred)
≤At most: 0 is less than or equal to ε. (0 ≤ ε)
∀For all: The statement after the comma holds for every ω. (∀ ω, 0 ≤ w ω)
∑Summation: Add up w(ω) for every ω. (∑ ω, w ω)
=Equality: ∑ ω, w(ω) and 1 are the same. (∑ ω, w ω = 1)
RReal numbers: The real numbers: every point on the number line. (ℝ)
Prop: Prop: a definition from the library, not special notation. Its name is its meaning; open it in Lean to see how it is defined. (Prop)
filter: Finset.univ.filter: a definition from the library, not special notation. Its name is its meaning; open it in Lean to see how it is defined. (Finset.univ.filter)
(x : T)As real numbers: L read as an element of real numbers. Mathematicians leave this implicit; Lean needs it to know which kind of number is meant. ((L : ℝ))
aᵇPower: 1 + 2 · ε raised to the power L. ((1 + 2 * ε) ^ L)
+Addition: 1 plus 2 · ε. (1 + 2 * ε)
x ↦ f(x)Function (λ-abstraction): The function sending ω to ¬good(ω). Lean writes fun x => …; papers write x ↦ …. (fun ω => ¬ good ω)
·Multiplication: 2 times ε. (2 * ε)
¬Not: It is not the case that good(ω). (¬ good ω)

theorem rc-randomized-conversion-theorem {Ω : Type} [Fintype Ω] (w chi : Ω → ℝ) (L : ℕ) (ε δ : ℝ) (good : Ω → Prop) [DecidablePred good] (hε : 0 ≤ ε) (hw : ∀ ω, 0 ≤ w ω) (hsum : ∑ ω, w ω = 1) (hgood : ∀ ω, good ω → chi ω ≤ (1 + 2 * ε) ^ L) (hbad : ∀ ω, chi ω ≤ ((L : ℝ) + 2) ^ L) (hδ : ∑ ω ∈ Finset.univ.filter (fun ω => ¬ good ω), w ω ≤ δ) :
    ∑ ω, w ω * chi ω ≤ (1 + 2 * ε) ^ L + δ * ((L : ℝ) + 2) ^ L

On the good event (all boosted node estimates within factor 1+ε) the deterministic Conversion Theorem gives χ² ≤ (1+2ε)^L; off it the smoothing floor gives χ² ≤ (L+2)^L; the good event has mass ≥ 1−δ. Hence E[χ²] ≤ (1+2ε)^L + δ (L+2)^L. Part of the Randomized Conversion Theorem: the Conversion Theorem (counting ⇒ sampling descent) extended to every randomized bounded-variance attention algorithm via median-of-means boosting, a union bound over the 2n descent nodes, and ε-smoothing of each node; closes the residual scalability crux of the Attention Compute Law.

How it was proved

Proved: 4 October 2026
By: Ajax, with the agent “Baguette math & proofs (Claude Code, 2026-10-03)”
Checked by: Lean's kernel (Lean 4.33.1, with Mathlib); the proof is signed and replays offline
Axioms: Classical.choice, Quot.sound, propext
Proof: 35 lines of Lean

Randomized Conversion Theorem (assembly): E[χ²] ≤ (1+2ε)^L + δ (L+2)^L for the proposal synthesised from a randomized oracle

How it was proved

More proved statements