proofs
Provedalphabaguettebaguette-zoptimizer

AdamW with decoupled decay keeps every parameter in a fixed box forever

For all θ,u:N→R\htmlData{mx=3}{\theta, u : \htmlData{mx=0}{\mathbb{N}} \mathrel{\htmlData{mx=2}{\to}} \htmlData{mx=1}{\mathbb{R}}} and a,η,K∈R\htmlData{mx=1}{a, \eta, K \in \htmlData{mx=0}{\mathbb{R}}}, if 0<a0 \mathrel{\htmlData{mx=1}{<}} \htmlData{mx=0}{a}, a≤1\htmlData{mx=0}{a} \mathrel{\htmlData{mx=1}{\leq}} 1, 0≤η0 \mathrel{\htmlData{mx=1}{\leq}} \htmlData{mx=0}{\eta}, ∀t, ∣u(t)∣≤K\htmlData{mx=6}{\forall t,\ \htmlData{mx=3}{\left\lvert \htmlData{mx=2}{\htmlData{mx=0}{u}(\htmlData{mx=1}{t})} \right\rvert} \mathrel{\htmlData{mx=5}{\leq}} \htmlData{mx=4}{K}} and ∀t, θ(t+1)=(1−a)⋅θ(t)−η⋅u(t)\htmlData{mx=17}{\forall t,\ \htmlData{mx=3}{\htmlData{mx=0}{\theta}(\htmlData{mx=2}{\htmlData{mx=1}{t} + 1})} \mathrel{\htmlData{mx=16}{=}} \htmlData{mx=15}{\htmlData{mx=9}{(\htmlData{mx=5}{1 - \htmlData{mx=4}{a}}) \cdot \htmlData{mx=8}{\htmlData{mx=6}{\theta}(\htmlData{mx=7}{t})}} - \htmlData{mx=14}{\htmlData{mx=10}{\eta} \cdot \htmlData{mx=13}{\htmlData{mx=11}{u}(\htmlData{mx=12}{t})}}}}, then

∀t, ∣θ(t)∣≤max⁡(∣θ(0)∣,η⋅Ka)\htmlData{mx=14}{\forall t,\ \htmlData{mx=3}{\left\lvert \htmlData{mx=2}{\htmlData{mx=0}{\theta}(\htmlData{mx=1}{t})} \right\rvert} \mathrel{\htmlData{mx=13}{\leq}} \htmlData{mx=12}{\max(\htmlData{mx=6}{\left\lvert \htmlData{mx=5}{\htmlData{mx=4}{\theta}(0)} \right\rvert}, \htmlData{mx=11}{\frac{\htmlData{mx=9}{\htmlData{mx=7}{\eta} \cdot \htmlData{mx=8}{K}}}{\htmlData{mx=10}{a}}})}}
What the notation means · or point at any symbol
NNatural numbers
The natural numbers 0, 1, 2, …. In Lean they start at 0. (ℕ)
⟹Implies
If ℕ, then ℝ. (Between types, A → B is the type of functions from A to B.) (ℕ → ℝ)
x : TVariables
θ, u range over ℕ → ℝ: the statement must hold for every choice. ((θ u : ℕ → ℝ))
x ∈ TVariables
a, η, K range over real numbers: the statement must hold for every choice. ((a η K : ℝ))
<Less than
0 is strictly smaller than a. (0 < a)
≤At most
a is less than or equal to 1. (a ≤ 1)
∀For all
The statement after the comma holds for every t. (∀ t, |u t| ≤ K)
RReal numbers
The real numbers: every point on the number line. (ℝ)
|x|Absolute value
The size of u(t) without its sign. (|u t|)
=Equality
θ(t + 1) and (1 − a) · θ(t) − η · u(t) are the same. (θ (t + 1) = (1 - a) * θ t - η * u t)
f(x)Function application
u applied to t. Lean writes f x; papers write f(x). (u t)
+Addition
t plus 1. (t + 1)
maxMaximum
The larger of |θ(0)| and η · K/a. (max |θ 0| (η * K / a))
·Multiplication
1 − a times θ(t). ((1 - a) * θ t)
-Subtraction
1 minus a. On ℕ, Lean's subtraction stops at 0: 2 − 5 = 0. (1 - a)
a/bDivision
η · K divided by a. In Lean, dividing by 0 gives 0. (η * K / a)

θ(t+1) = (1−a)θ(t) − η·u(t) with 0 < a ≤ 1 (a = lr·decay) and |u| ≤ K gives |θ(t)| ≤ max(|θ(0)|, ηK/a) for ALL t. The master-weight bank therefore has a compile-time integer width valid for an unbounded training run. A short search did not find this uniform-in-time box stated in the literature. Part of the plan 'Baguette-Z: closed-loop certified integer-state training'.

How it was proved

Proved
3 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
25 lines of Lean
Sign in to see the proofThe proof, its signed report and its history.

More proved statements