OctoInformBeta

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

Lemma 4.3 [Endpoint comparison]: The weighted square of the change of the cutoff matrix over a filter stretch, measured against 230400230400 times differences of tr⁡p\operatorname{tr} p

Bounding the change of the cutoff by potential differences.

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

A filter moves a d×md\times m matrix RR in discrete time: the step at time tt sends it to CtR+btejtTC_tR + b_t\,e_{j_t}^{T}, multiplying it by a d×dd\times d matrix CtC_t and adding a vector btb_t to its column jtj_t. Over a stretch of steps, a positive definite matrix attached to the start is carried forward by the product of the step matrices and then gains a new part, which gives the matrix attached to the end.

Each end has a cutoff matrix: a fixed cutoff function applied to the spectrum of the matrix at that end, lifted to a second set of coordinates. The question is how far apart the two cutoff matrices are in a weighted Hilbert-Schmidt norm, compared with the changes of a spectral potential along the stretch.

The library states the comparison for real matrices indexed by finite sets, with the weight on the second set of coordinates split into an old part and a new part.

Spectral Functions and Lifts

The cutoff f0f_0 is the continuous piecewise linear function on ℝ\mathbb{R} given by f0(x)={0,x≤1/2,2x−1,1/2<x<1,1,1≤x≤2,2−x/2,2<x<4,0,x≥4.f_0(x) = \begin{cases} 0, & x \le 1/2,\\ 2x - 1, & 1/2 < x < 1,\\ 1, & 1 \le x \le 2,\\ 2 - x/2, & 2 < x < 4,\\ 0, & x \ge 4. \end{cases}(1)✦

The weight ww with exponent a=3/2a = 3/2 is the real function given by w(x)=min⁡{xa,x−1},w(x) = \min\{x^{a}, x^{-1}\},(2)✦

The potential pp is the real function given by p(x)=∫0xw(u)udu.p(x) = \int_0^x \frac{w(u)}{u}\,du.(3)✦ On positive definite matrices, f0f_0 and pp act by the functional calculus.

The lift of a real square matrix AA indexed by a finite set ι\iota along a real ι×κ\iota\times\kappa matrix RR, for a second finite set κ\kappa, is the square matrix RTARR^{T}AR indexed by κ\kappa.✦ For finite sets ι\iota and κ\kappa, a coisometry RR is a real ι×κ\iota\times\kappa matrix with RRT=IRR^{T} = I.

The weighted square of a real matrix DD with columns indexed by κ\kappa against a real square matrix WW indexed by κ\kappa is the number tr⁡(DWDT),\operatorname{tr}(DWD^{T}),(4)✦ which for positive semidefinite WW equals ∥DW1/2∥HS2\|DW^{1/2}\|_{\mathrm{HS}}^{2}.

The Claim

Theorem 1 (Endpoint comparison). Let ι\iota and κ\kappa be finite sets. Let XX, YY and ZZ be positive definite real matrices indexed by ι\iota, and let 𝒞\mathcal{C} be an invertible real matrix indexed by ι\iota with 𝒞𝒞T≼I\mathcal{C}\mathcal{C}^{T} \preceq I and Z=𝒞Y𝒞TZ = \mathcal{C}Y\mathcal{C}^{T}. Let TT and SS be ι×κ\iota\times\kappa coisometries, and let FX=TTf0(X)TF_X = T^{T}f_0(X)T and FY=STf0(Y)SF_Y = S^{T}f_0(Y)S be the cutoff matrices, the lifts of f0(X)f_0(X) along TT and of f0(Y)f_0(Y) along SS. Let QoldQ_{\mathrm{old}} and QnewQ_{\mathrm{new}} be positive semidefinite real matrices indexed by κ\kappa with FYQnew=0F_YQ_{\mathrm{new}} = 0 and SQoldST=Y,TQoldTT=Z,TQnewTT=X−Z,TST=𝒞,TQoldST=𝒞Y.\begin{gathered} SQ_{\mathrm{old}}S^{T} = Y, \qquad TQ_{\mathrm{old}}T^{T} = Z, \qquad TQ_{\mathrm{new}}T^{T} = X - Z,\\ TS^{T} = \mathcal{C}, \qquad TQ_{\mathrm{old}}S^{T} = \mathcal{C}Y. \end{gathered}(5) Then the weighted square of FX−FYF_X - F_Y against W=I+Qold+QnewW = I + Q_{\mathrm{old}} + Q_{\mathrm{new}} satisfies tr⁡((FX−FY)W(FX−FY)T)≤230400(tr⁡p(Y)−tr⁡p(Z)+tr⁡p(X)−tr⁡p(Z)).\begin{aligned} \operatorname{tr}\bigl((F_X - F_Y)W(F_X - F_Y)^{T}\bigr) &\le 230400\,\bigl(\operatorname{tr} p(Y) - \operatorname{tr} p(Z)\\ &\qquad + \operatorname{tr} p(X) - \operatorname{tr} p(Z)\bigr). \end{aligned}(6)✦

In the paper, the result reads: with Y=XsY = X_s and X=XtX = X_t the matrices at the start ss and the end tt of the stretch, 𝒞=Ct⋯Cs+1\mathcal{C} = C_t\cdots C_{s+1} and Z=𝒞Y𝒞TZ = \mathcal{C}Y\mathcal{C}^{T}, one has ∥(Ft−Fs)(Im+Q/θ)1/2∥HS2≤C{tr⁡(p(Y)−p(Z))+tr⁡(p(X)−p(Z))}\|(F_t - F_s)(I_m + Q/\theta)^{1/2}\|_{\mathrm{HS}}^{2} \le C\{\operatorname{tr}(p(Y) - p(Z)) + \operatorname{tr}(p(X) - p(Z))\} for a constant CC, where its cutoff matrices FtF_t and FsF_s at the two ends and its scaled weight Q/θQ/\theta, the weight QQ divided by the scale θ\theta, are FXF_X, FYF_Y and Qold+QnewQ_{\mathrm{old}} + Q_{\mathrm{new}} here.

What Is Checked

The claim (6) is stated over the cutoff (1), the potential (3) built from the weight (2) with exponent a=3/2a = 3/2, the lift RTARR^{T}AR and the weighted square (4), for positive definite XX, YY and Z=𝒞Y𝒞TZ = \mathcal{C}Y\mathcal{C}^{T} with 𝒞\mathcal{C} invertible and 𝒞𝒞T≼I\mathcal{C}\mathcal{C}^{T} \preceq I, coisometries TT and SS, and positive semidefinite QoldQ_{\mathrm{old}} and QnewQ_{\mathrm{new}} satisfying (5) and FYQnew=0F_YQ_{\mathrm{new}} = 0.