OctoInformBeta

OpenAI preprint, September 2026 · 3. A thermal block entropy inequality · Section 3

Entropy concavity and the diagonal upper bound: The ordinary diagonal gives an entropy upper bound at arbitrary trace

Entropy is concave on the positive cone.

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

Fix a finite-dimensional complex Hilbert space with an orthonormal basis indexed by II. The ordinary diagonal d(P)d(P) of a positive semidefinite matrix PP is the nonnegative vector d(P)=(Re⁡Pii)i∈Id(P)=(\operatorname{Re}P_{ii})_{i\in I} of populations in this basis.

The von Neumann entropy S(P)S(P) of a complex matrix PP indexed by II is the real part of the trace of the spectral functional calculus of PP with the function η(x)=−xln⁡x\eta(x)=-x\ln x, with 0ln⁡0=00\ln0=0.✦ S(P)=Re⁡Tr⁡η(P)=−Tr⁡(Pln⁡P).S(P)=\operatorname{Re}\operatorname{Tr}\eta(P) =-\operatorname{Tr}(P\ln P).(1)✦ For a nonnegative vector x=(xi)i∈Ix=(x_i)_{i\in I}, its entropy H(x)H(x) is the scalar functional H(x)=−∑i∈Ixiln⁡xi.H(x)=-\sum_{i\in I}x_i\ln x_i.(2) The entropies in (1) and (2) are measured in nats.

The Claim

Theorem 1. Let PP and QQ be positive semidefinite complex matrices indexed by the finite set II, and let t∈[0,1]t\in[0,1]. Then S(P)≤H(d(P)).S(P)\le H(d(P)).(3)✦ Moreover, S(tP+(1−t)Q)≥tS(P)+(1−t)S(Q).S\bigl(tP+(1-t)Q\bigr)\ge tS(P)+(1-t)S(Q).(4)✦

In the paper, the result reads: for a state PP, S(P)≤H(d(P))S(P)\le H(d(P)), and for any finite probability mixture of states PiP_i with weights wiw_i, S(∑iwiPi)≥∑iwiS(Pi)S(\sum_i w_iP_i)\ge\sum_i w_iS(P_i).

What Is Checked

The inequalities (3) and (4) are stated over the positive semidefinite cone of finite square complex matrices, with no trace-normalization condition.