OctoInformBeta

OpenAI preprint, September 2026 · 2. An entropy inequality with one zero block · Section 2.4

Schur complement, sign averaging, and rectangular embedding: Diagonal blocks and entropy under rectangular isometric embeddings

Entropy bounds for a 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 matrix with one zero block couples three rectangular amplitudes. Its Gram entropy can be compared with the sum of the three separate Gram entropies, with the comparison expressed in terms of their squared Hilbert-Schmidt masses.

Let I,K,M,NI,K,M,N be finite index sets. For complex matrices A∈ℂI×KA\in\mathbb C^{I\times K}, B∈ℂI×MB\in\mathbb C^{I\times M} and C∈ℂN×KC\in\mathbb C^{N\times K}, the matrix X(A,B,C)X(A,B,C) with one zero block is (1). X(A,B,C)=(ABC0).X(A,B,C)=\begin{pmatrix}A&B\\ C&0\end{pmatrix}.(1)✦ The Gram matrix of a rectangular complex matrix AA is AA*AA^*.✦ The squared Hilbert-Schmidt mass of a rectangular complex matrix AA is Tr⁡(AA*)\operatorname{Tr}(AA^*), with the library taking the real part of this trace.✦

The entropy S(P)S(P) of a square complex matrix PP is the real trace of the spectral function η(P)\eta(P) in (2).✦ On positive semidefinite matrices it is the von Neumann entropy in nats. η(t)=−tln⁡t,0ln⁡0=0,S(P)=Re⁡Tr⁡η(P),S(P)=−Tr⁡(Pln⁡P)(P≥0).\begin{gathered} \eta(t)=-t\ln t,\qquad 0\ln0=0,\qquad S(P)=\operatorname{Re}\operatorname{Tr}\eta(P),\\ S(P)=-\operatorname{Tr}(P\ln P)\quad(P\geq0). \end{gathered}(2)✦

For real x,y,zx,y,z, the smaller and larger block eigenvalues are the functions λ−\lambda_- and λ+\lambda_+ in (3). 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}.(3) The scalar entropy deficit δ\delta is the real function of real variables x,y,zx,y,z given by (4). δ(x,y,z)=η(x)+η(y)+η(z)−η(λ+(x,y,z))−η(λ−(x,y,z)).\delta(x,y,z)=\eta(x)+\eta(y)+\eta(z) -\eta\bigl(\lambda_+(x,y,z)\bigr) -\eta\bigl(\lambda_-(x,y,z)\bigr).(4)✦

For a rectangular complex matrix ZZ and square complex matrices D,ED,E on its row and column spaces, the block log determinant increment L(Z;D,E)L(Z;D,E) is the real function in (5). L(Z;D,E)=ln⁡|det(D+ZE−1Z*)|−ln⁡|det⁡D|.L(Z;D,E)=\ln\left|\det(D+ZE^{-1}Z^*)\right|-\ln|\det D|.(5)✦ The coordinate pinching 𝒟(D)\mathcal D(D) of a square matrix DD on ℂI\mathbb C^I is the diagonal matrix diag⁡((Dii)i∈I)\operatorname{diag}((D_{ii})_{i\in I}).

The Claim

Theorem 1. For compatible finite complex matrices A,B,CA,B,C, the zero block determinant identity is (6). det⁡(𝟏+X(A,B,C)X(A,B,C)*)=det⁡(𝟏+CC*)det⁡(𝟏+BB*+A(𝟏+C*C)−1A*).\det\bigl(\mathbf 1+X(A,B,C)X(A,B,C)^*\bigr) =\det(\mathbf 1+CC^*)\, \det\bigl(\mathbf 1+BB^*+A(\mathbf 1+C^*C)^{-1}A^*\bigr).(6)✦

Let D,ED,E be positive definite matrices on ℂI\mathbb C^I and let z∈ℂIz\in\mathbb C^I. The diagonal log determinant bound reads (7). L(diag⁡(z);𝒟(D),𝒟(E))≤L(diag⁡(z);D,E).L\bigl(\operatorname{diag}(z);\mathcal D(D),\mathcal D(E)\bigr) \leq L\bigl(\operatorname{diag}(z);D,E\bigr).(7)✦

For real numbers sis_i indexed by II, the diagonal matrix AsA_s on ℂI\mathbb C^I is defined by As=diag⁡((si)i∈I)A_s=\operatorname{diag}((s_i)_{i\in I}). With B∈ℂI×MB\in\mathbb C^{I\times M} and C∈ℂN×IC\in\mathbb C^{N\times I}, the diagonal block entropy bound is (8). S(AsAs*)+S(BB*)+S(CC*)−S(X(As,B,C)X(As,B,C)*)≤δ(Tr⁡(AsAs*),Tr⁡(BB*),Tr⁡(CC*)).\begin{split} &S(A_sA_s^*)+S(BB^*)+S(CC^*) -S\bigl(X(A_s,B,C)X(A_s,B,C)^*\bigr)\\ &\hspace{2em}\leq \delta\bigl(\operatorname{Tr}(A_sA_s^*), \operatorname{Tr}(BB^*),\operatorname{Tr}(CC^*)\bigr). \end{split}(8)✦

The square block entropy bound (9) holds for every square complex matrix AA on ℂI\mathbb C^I and compatible B,CB,C. S(AA*)+S(BB*)+S(CC*)−S(X(A,B,C)X(A,B,C)*)≤δ(Tr⁡(AA*),Tr⁡(BB*),Tr⁡(CC*)).\begin{split} &S(AA^*)+S(BB^*)+S(CC^*)-S\bigl(X(A,B,C)X(A,B,C)^*\bigr)\\ &\hspace{2em}\leq \delta\bigl(\operatorname{Tr}(AA^*), \operatorname{Tr}(BB^*),\operatorname{Tr}(CC^*)\bigr). \end{split}(9)✦

Every square complex matrix AA admits a singular value decomposition (10), where U,VU,V are unitary and the numbers sis_i are real and nonnegative. A=Udiag⁡((si)i∈I)V*.A=U\operatorname{diag}((s_i)_{i\in I})V^*.(10)✦

Let R,A,TR,A,T be compatible rectangular complex matrices with R*R=𝟏R^*R=\mathbf 1 and TT*=𝟏TT^*=\mathbf 1. The isometric invariance of Gram entropy is (11). S((RAT)(RAT)*)=S(AA*).S\bigl((RAT)(RAT)^*\bigr)=S(AA^*).(11)✦

What Is Checked

The statements concern finite complex matrices, with positive definite auxiliary matrices in (7) and a square central block in the entropy bounds (8) and (9).