OctoInformBeta

OpenAI preprint, September 2026 · §4. Matrix variation along the filter · §4.1

§4.1 (displayed): the scale energy (scale-energy): For a positive semidefinite start AA and total gain Γ\Gamma, any finite dyadic energy sum is at most 28(tr⁡A+2Γ)28(\operatorname{tr}A+2\Gamma)

A trace bound for dyadic energy sums of admissible runs.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Along the filter that chooses the signs in the Euclidean Steinitz-Bergström theorem, a matrix records the variation accumulated so far, and each step of the filter changes it at an energy cost. The same sequence of steps, a run, can be observed at every dyadic scale 2s2^s, with ss an integer, its energy weighted by 2s2^s. The question is whether these weighted energies, summed over many scales, stay below a fixed multiple of the traces of the starting matrix and of what the steps add, however many scales are taken.

Moves and Their Energy

Fix a finite set ι\iota; a state is a real ι×ι\iota \times \iota matrix. A move pp on ι\iota is a pair (C,G)(C, G) of real ι×ι\iota \times \iota matrices, its contraction CC and its gain GG, acting on a state MM by p(M)=CMCT+G.p(M) = C M C^{T} + G.(1)✦ A move (C,G)(C, G) is admissible if CC is a contraction, CCT≼IC C^{T} \preceq I, and its gain GG is positive semidefinite.✦

A run 𝒑=(p0,…,pk−1)\mathbf p = (p_0, \dots, p_{k-1}) is a finite sequence of moves; from a state DD it passes through the states D0=DD_0 = D and Di+1=pi(Di)D_{i+1} = p_i(D_i). The energy ℰ(D;𝒑)\mathcal E(D; \mathbf p) of a run 𝒑\mathbf p from a state DD is the sum of the costs of its moves at the states they act on, ℰ(D;𝒑)=∑i<kσpi(Di),\mathcal E(D; \mathbf p) = \sum_{i < k} \sigma_{p_i}(D_i),(2)✦ where σp(M)\sigma_p(M) is the cost of the move pp at the state MM.

The total gain Γ(𝒑)\Gamma(\mathbf p) of a run 𝒑\mathbf p is the sum of the traces of its gains, Γ(𝒑)=∑i<ktr⁡Gi\Gamma(\mathbf p) = \sum_{i<k} \operatorname{tr} G_i, where GiG_i is the gain of pip_i.✦ For a real number cc, the rescaling of a move p=(C,G)p = (C, G) by cc is the move ρc(p)=(C,cG)\rho_c(p) = (C, cG), and ρc(𝒑)\rho_c(\mathbf p) rescales every move of a run 𝒑\mathbf p by cc.✦

The Claim

Theorem 1. Let AA be a positive semidefinite real ι×ι\iota \times \iota matrix, let 𝒑\mathbf p be a run of admissible moves, and let SS be a finite set of integers. Then ∑s∈S2sℰ(2−sA;ρ2−s(𝒑))≤28(tr⁡A+2Γ(𝒑)).\sum_{s \in S} 2^{s}\, \mathcal E\bigl(2^{-s} A;\ \rho_{2^{-s}}(\mathbf p)\bigr) \le 28 \bigl(\operatorname{tr} A + 2\, \Gamma(\mathbf p)\bigr).(3)✦

In the paper, the result reads ∑θθEθ(0,n)≤Ktr⁡Q\sum_\theta \theta E_\theta(0,n) \le K \operatorname{tr} Q for a constant KK, the sum running over dyadic scales θ\theta, which stand where the page has 2s2^s, of the scale-weighted energies θEθ(0,n)\theta E_\theta(0,n).

What Is Checked

The claim (3) is stated over real square matrices indexed by a finite set ι\iota: a positive semidefinite start AA, a run of moves (1) that are admissible, the energy (2) of a run from a state, the rescaling ρc\rho_c of gains and the total gain Γ\Gamma, with the scales 2s2^s indexed by a finite set SS of integers.