OctoInformBeta

OpenAI preprint, September 2026 · 2. An entropy inequality with one zero block · Section 2

Theorem 2.1: Normalized block weights and a homogeneous deficit bound for rectangular matrices

Entropy bounds for one zero block.

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 rectangular matrix with one zero block gives an unnormalized quantum state through its Gram matrix. The entropy question compares that state with the three occupied blocks, retaining the squared Hilbert-Schmidt mass of each block.

Let A∈ℂr×dA\in\mathbb C^{r\times d}, B∈ℂr×eB\in\mathbb C^{r\times e}, and C∈ℂs×dC\in\mathbb C^{s\times d}, where all dimensions are finite. The matrix with one zero block is M(A,B,C)M(A,B,C) in (1): M(A,B,C)=(ABC0).M(A,B,C)=\begin{pmatrix}A&B\\ C&0\end{pmatrix}.(1)✦ For a complex rectangular matrix TT, its Gram matrix is TT*TT^*, where T*T^* is the conjugate transpose✦, and its mass m(T)m(T) is the squared Hilbert-Schmidt mass defined in (2): m(T)=Re⁡Tr⁡(TT*),x=m(A),y=m(B),z=m(C).m(T)=\operatorname{Re}\operatorname{Tr}(TT^*), \qquad x=m(A),\quad y=m(B),\quad z=m(C).(2)✦

For a finite complex square matrix PP, its von Neumann entropy S(P)S(P) is the real part of the trace of the spectral function associated with t↦−tln⁡tt\mapsto-t\ln t, with 0ln⁡0=00\ln0=0. For positive semidefinite PP, it is given by (3), without a unit trace requirement. S(P)=−Tr⁡(Pln⁡P).S(P)=-\operatorname{Tr}(P\ln P).(3) The binary entropy h(q)h(q) for q∈[0,1]q\in[0,1] and the qubit entropy g(u)g(u) at determinant u∈[0,1/4]u\in[0,1/4] are the functions in (4); all entropies here are in nats. h(q)=−qln⁡q−(1−q)ln⁡(1−q),g(u)=h(1+1−4u2).h(q)=-q\ln q-(1-q)\ln(1-q), \qquad g(u)=h\!\left(\frac{1+\sqrt{1-4u}}{2}\right).(4)✦

For nonnegative masses x,y,zx,y,z, the larger and smaller block eigenvalues λ+(x,y,z)\lambda_+(x,y,z) and λ−(x,y,z)\lambda_-(x,y,z) are the scalars in (5): λ±(x,y,z)=x+y+z±(x+y+z)2−4yz2.\lambda_\pm(x,y,z) =\frac{x+y+z\ \pm\sqrt{(x+y+z)^2-4yz}}{2}.(5) For real x,y,zx,y,z, the scalar entropy deficit Δ(x,y,z)\Delta(x,y,z) is the real function defined in (6), with λ±=λ±(x,y,z)\lambda_\pm=\lambda_\pm(x,y,z) and 0ln⁡0=00\ln0=0: Δ(x,y,z)=−xln⁡x−yln⁡y−zln⁡z+λ+ln⁡λ++λ−ln⁡λ−.\Delta(x,y,z) =-x\ln x-y\ln y-z\ln z +\lambda_+\ln\lambda_++\lambda_-\ln\lambda_-.(6)

The Claim

Theorem 1. Let X,Y,ZX,Y,Z be complex matrices of the respective block dimensions in (1), each of mass one. For nonnegative real weights a,b,ca,b,c with a+b+c=1a+b+c=1, the normalized entropy bound is (7): aS(XX*)+bS(YY*)+cS(ZZ*)+g(bc)≤S(M(aX,bY,cZ)M(aX,bY,cZ)*).\begin{split} &aS(XX^*)+bS(YY^*)+cS(ZZ^*)+g(bc)\\ &\qquad\le S\!\left( M(\sqrt a\,X,\sqrt b\,Y,\sqrt c\,Z) M(\sqrt a\,X,\sqrt b\,Y,\sqrt c\,Z)^* \right). \end{split}(7)✦ For arbitrary complex blocks A,B,CA,B,C of those dimensions, with masses x,y,zx,y,z as in (2), the entropy deficit satisfies (8): S(AA*)+S(BB*)+S(CC*)−S(M(A,B,C)M(A,B,C)*)≤Δ(x,y,z).S(AA^*)+S(BB^*)+S(CC^*) -S\!\left(M(A,B,C)M(A,B,C)^*\right) \le\Delta(x,y,z).(8)✦

In the paper, the result reads: for complex d×ed\times e matrices A,B,CA,B,C with masses x,y,zx,y,z satisfying x+y+z=1x+y+z=1, the bound (9) holds, with every zero-weight term omitted and M=M(A,B,C)M=M(A,B,C): S(MM*)≥xS(AA*x)+yS(BB*y)+zS(CC*z)+g(yz).S(MM^*)\ge xS\!\left(\frac{AA^*}{x}\right) +yS\!\left(\frac{BB^*}{y}\right) +zS\!\left(\frac{CC^*}{z}\right)+g(yz).(9)

What Is Checked

Both claims are stated over finite complex rectangular matrices with the compatible block dimensions in (1); (7) additionally assumes unit block masses and nonnegative weights summing to one.