OctoInformBeta

OpenAI preprint, September 2026 · §5. Sampling and orthogonal packing · §5

Lemma 5.1 [Sampling a matrix block]: For pxppxp near a scalar block, τ(yx)\tau(yx) and ∥[x,y]∥2,τ2\|[x,y]\|_{2,\tau}^2 are cc times those at pp, with errors in η\eta and δτ\delta_\tau

Sampling a matrix block approximately scales traces and commutator norms.

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

Let DD be a C*^*-algebra and qq a projection of DD. A row over qq is a kk-tuple U=(u1,…,uk)U=(u_1,\dots,u_k) of elements of DD such that ui*ui=q,ui*ul=0(i≠l),u_i^*u_i=q,\qquad u_i^*u_l=0\quad(i\neq l),(1)✦ so that U∈M1,k(D)U\in M_{1,k}(D) satisfies U*U=1k⊗qU^*U=1_k\otimes q. The range of a row UU over qq is the projection p=UU*p=UU^*, and the row writes the corner pDppDp as k×kk\times k matrices over qDqqDq.

Traces and Matrix Blocks

A tracial state τ\tau of DD is a positive linear functional τ:D→ℂ\tau\colon D\to\mathbb{C} that is normalized, τ(1)=1\tau(1)=1, and tracial, τ(ab)=τ(ba)\tau(ab)=\tau(ba) for all a,b∈Da,b\in D; T(D)T(D) is the set of tracial states of DD.✦ The trace 22-norm of an element aa of DD at τ∈T(D)\tau\in T(D) is the real number ∥a∥2,τ=τ(a*a)1/2\|a\|_{2,\tau}=\tau(a^*a)^{1/2}.✦

Let UU be a row over qq and aa an element of DD. The (i,l)(i,l) matrix coefficient of aa along UU is uiaul*u_iau_l^*, for 1≤i,l≤k1\le i,l\le k.✦ For a scalar matrix H=(hil)∈Mk(ℂ)H=(h_{il})\in M_k(\mathbb{C}), the matrix block of HH over aa along UU is the element U(H⊗a)U*=∑i,l=1khiluiaul*U(H\otimes a)U^*=\sum_{i,l=1}^{k} h_{il}\,u_iau_l^*(2)✦ of DD, and the diagonal sum of aa along UU is U(1k⊗a)U*=∑i=1kuiaui*U(1_k\otimes a)U^*=\sum_{i=1}^{k}u_iau_i^*✦. The range pp of UU is the diagonal sum of qq.

The Claim

Theorem 1 (Sampling a matrix block). Let DD be a C*^*-algebra, τ∈T(D)\tau\in T(D) and xx a unitary of DD, and let η<1\eta<1 and c≥0c\ge 0 be real numbers.

  1. Let UU be a row over a projection qq of DD with range pp, and suppose that ∥pxp−U(H⊗q)U*∥<η\|pxp-U(H\otimes q)U^*\|<\eta for some scalar matrix H∈Mk(ℂ)H\in M_k(\mathbb{C}). Let e≤qe\le q be a projection, y=U(1k⊗e)U*y=U(1_k\otimes e)U^* and δτ=|τ(y)−cτ(p)|\delta_\tau=|\tau(y)-c\,\tau(p)|. Then |τ(yx)−cτ(px)|≤η(τ(y)+cτ(p))+2δτ,|∥[x,y]∥2,τ2−c∥[x,p]∥2,τ2|≤6η(τ(y)+cτ(p))+10δτ.\begin{aligned} |\tau(yx)-c\,\tau(px)| &\le \eta\,\bigl(\tau(y)+c\,\tau(p)\bigr)+2\delta_\tau,\\ \bigl|\|[x,y]\|_{2,\tau}^2-c\,\|[x,p]\|_{2,\tau}^2\bigr| &\le 6\eta\,\bigl(\tau(y)+c\,\tau(p)\bigr)+10\delta_\tau. \end{aligned}(3)✦

  2. Let y≤py\le p be projections of DD and zz an element of DD that commutes with pp and with yy, such that ∥pxp−z∥<η\|pxp-z\|<\eta and τ(zy)τ(p)=τ(zp)τ(y),τ(z*zy)τ(p)=τ(z*zp)τ(y).\tau(zy)\,\tau(p)=\tau(zp)\,\tau(y),\qquad \tau(z^*zy)\,\tau(p)=\tau(z^*zp)\,\tau(y).(4)✦ Then yy and pp satisfy the two inequalities (3), with δτ=|τ(y)−cτ(p)|\delta_\tau=|\tau(y)-c\,\tau(p)|.✦

  3. For every projection ff of DD, ∥[x,f]∥2,τ2=2τ(f)−2∥fxf∥2,τ2.\|[x,f]\|_{2,\tau}^2=2\tau(f)-2\|fxf\|_{2,\tau}^2.(5)✦

In the paper, the result reads: for a nonzero projection q∈Dq\in D, U∈M1,k(D)U\in M_{1,k}(D) with U*U=1k⊗qU^*U=1_k\otimes q and p=UU*p=UU^*, a unitary xx of DD and 0<η<10<\eta<1 with dist⁡(U*xU,Mk(ℂ)⊗q)<η\operatorname{dist}(U^*xU,M_k(\mathbb{C})\otimes q)<\eta, a projection e≤qe\le q and y=U(1k⊗e)U*y=U(1_k\otimes e)U^*, one has, for 0≤c≤10\le c\le 1, τ∈T(D)\tau\in T(D) and δτ=|τ(y)−cτ(p)|\delta_\tau=|\tau(y)-c\,\tau(p)|, the inequalities |τ(yx)−cτ(px)|≤η(τ(y)+cτ(p))+2δτ|\tau(yx)-c\,\tau(px)|\le\eta(\tau(y)+c\,\tau(p))+2\delta_\tau and |∥[x,y]∥2,τ2−c∥[x,p]∥2,τ2|≤6η(τ(y)+cτ(p))+10δτ\bigl|\|[x,y]\|_{2,\tau}^2-c\,\|[x,p]\|_{2,\tau}^2\bigr|\le 6\eta(\tau(y)+c\,\tau(p))+10\delta_\tau.

What Is Checked

The three statements are stated over the tracial states of a C*^*-algebra DD with its C*^*-order and their trace 22-norms, the unitaries and projections of DD, and rows over a projection with their matrix coefficients, matrix blocks (2) and diagonal sums.