OctoInformBeta

OpenAI preprint, September 2026 · Section 3. A common purification of effects

Lemma 3.1: A finite-dimensional target preserves cross products and satisfies an operator Jensen bound

Effects admit a common resolvent purification.

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 H=ℂnH=\mathbb C^n, where nn is a finite index set. The question is whether the resolvent kernels of different effects can share a finite-dimensional purification space while retaining their mutual Gram operators.

An effect TT on HH is a positive semidefinite operator satisfying T≤IT\leq I. Here II is the identity on HH, X*X^* denotes the Hermitian adjoint of XX, and X≤YX\leq Y is the Löwner order, meaning that Y−XY-X is positive semidefinite.

Resolvent Kernels and Their Common Target

For a complex matrix TT on HH and a real parameter ss, the resolvent RT(s)R_T(s) and purification kernel KT(s)K_T(s) are the operators in (1). RT(s)=(T+sI)−1,KT(s)=TRT(s).R_T(s)=(T+sI)^{-1}, \qquad K_T(s)=T R_T(s).(1) For a complex matrix TT on HH, the effect entropy f(T)f(T) is the operator obtained by continuous functional calculus from the scalar function in (2), with the convention 0log⁡0=00\log 0=0. f(T)=(t↦tlog⁡t)(T).f(T)=(t\mapsto t\log t)(T).(2)✦ All logarithms in (2) are natural.

The joint purification kernel 𝒦A(s):ℋJ→H\mathcal K_A(s):\mathcal H_J\to H of a finite family A=(Aj)j∈JA=(A_j)_{j\in J} of complex matrices on HH, at a real parameter ss, is the block row operator defined in (3), where ℋJ=⨁j∈JH\mathcal H_J=\bigoplus_{j\in J}H. 𝒦A(s)(uj)j∈J=∑j∈JKAj(s)uj.\mathcal K_A(s)(u_j)_{j\in J} =\sum_{j\in J}K_{A_j}(s)u_j.(3)✦ For a finite family A=(Aj)j∈JA=(A_j)_{j\in J} of complex matrices on HH, the common Gram operator GAG_A on ℋJ\mathcal H_J is the integral in (4). GA=∫0∞𝒦A(s)*𝒦A(s)ds.G_A=\int_0^\infty \mathcal K_A(s)^*\mathcal K_A(s)\,ds.(4)✦ Let ιj:H→ℋJ\iota_j:H\to\mathcal H_J be the inclusion into the jjth summand. The finite common purification FAj:H→ℋJF_{A_j}:H\to\mathcal H_J associated with a finite family A=(Aj)j∈JA=(A_j)_{j\in J} of complex matrices on HH and an index j∈Jj\in J is the column block of the positive square root of GAG_A in (5). FAj=GA1/2ιj.F_{A_j}=G_A^{1/2}\iota_j.(5)✦

The Claim

Theorem 1. Every positive semidefinite operator TT on HH satisfies the normalization identity (6): ∫0∞KT(s)*KT(s)ds=T.\int_0^\infty K_T(s)^*K_T(s)\,ds=T.(6)✦

Let (Aj)j∈J(A_j)_{j\in J} be a finite family of effects on HH, and let qj≥0q_j\geq0 satisfy ∑j∈Jqj=1\sum_{j\in J}q_j=1. For an effect B=∑j∈JqjAjB=\sum_{j\in J}q_jA_j, the kernel distance bound is (7): ∑j∈Jqj∫0∞(KAj(s)−KB(s))*(KAj(s)−KB(s))ds≤∑j∈Jqjf(Aj)−f(B).\sum_{j\in J}q_j\int_0^\infty \bigl(K_{A_j}(s)-K_B(s)\bigr)^* \bigl(K_{A_j}(s)-K_B(s)\bigr)\,ds \leq \sum_{j\in J}q_j f(A_j)-f(B).(7)✦

For every finite family A=(Aj)j∈JA=(A_j)_{j\in J} of positive semidefinite operators on HH, there are maps F̃Aj:H→ℋJ\widetilde F_{A_j}:H\to\mathcal H_J satisfying (8) for all j,ℓ∈Jj,\ell\in J: F̃Aj*F̃Aj=Aj,F̃Aj*F̃Aℓ=∫0∞KAj(s)*KAℓ(s)ds.\widetilde F_{A_j}^*\widetilde F_{A_j}=A_j, \qquad \widetilde F_{A_j}^*\widetilde F_{A_\ell} =\int_0^\infty K_{A_j}(s)^*K_{A_\ell}(s)\,ds.(8)✦

Suppose that every AjA_j is an effect. Let RR be a finite set, e:R→Je:R\to J, and j0∈Jj_0\in J. If qr≥0q_r\geq0, ∑r∈Rqr=1\sum_{r\in R}q_r=1, and ∑r∈RqrAe(r)=Aj0\sum_{r\in R}q_r A_{e(r)}=A_{j_0}, the finite common purifications defined in (5) satisfy (9): ∑r∈Rqr(FAe(r)−FAj0)*(FAe(r)−FAj0)≤∑r∈Rqrf(Ae(r))−f(Aj0).\sum_{r\in R}q_r \bigl(F_{A_{e(r)}}-F_{A_{j_0}}\bigr)^* \bigl(F_{A_{e(r)}}-F_{A_{j_0}}\bigr) \leq\sum_{r\in R}q_r f(A_{e(r)})-f(A_{j_0}).(9)✦

In the paper, writing FTF_T for the resolvent purification of an effect TT, the result reads: for every effect TT, FT*FT=TF_T^*F_T=T; for effects A1,…,ANA_1,\ldots,A_N and positive weights qjq_j summing to one, with B=∑jqjAjB=\sum_jq_jA_j, one has ∑jqj(FAj−FB)*(FAj−FB)≤∑jqjf(Aj)−f(B)\sum_jq_j(F_{A_j}-F_B)^*(F_{A_j}-F_B)\leq\sum_jq_jf(A_j)-f(B); and every finite collection of effects admits a common finite-dimensional target for its resolvent maps preserving every product FS*FTF_S^*F_T for effects SS and TT in that collection.

What Is Checked

The claims are stated over finite-dimensional complex matrices and their integrals for positive real ss; the Jensen bounds use finite convex combinations of effects whose weights may vanish.