OctoInformBeta

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

Lemma 2.3: Positive definite Hermitian blocks with a fixed rectangular coupling ZZ

Joint convexity of a log determinant increment.

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

For positive definite Hermitian matrices DD and EE and a fixed rectangular matrix ZZ of matching size with conjugate transpose Z*Z^*, the log determinant increment measures the change produced by adding ZE−1Z*ZE^{-1}Z^* to DD.

Let II and JJ be finite index sets, and fix Z∈ℂI×JZ\in\mathbb C^{I\times J}. For square complex matrices D∈ℂI×ID\in\mathbb C^{I\times I} and E∈ℂJ×JE\in\mathbb C^{J\times J}, the block log determinant increment GZ(D,E)G_Z(D,E) is the real number defined by (1). GZ(D,E)=ln⁡|det⁡(D+ZE−1Z*)|−ln⁡|det⁡D|.G_Z(D,E) = \ln\bigl|\det(D+ZE^{-1}Z^*)\bigr| -\ln\bigl|\det D\bigr|.(1)✦

The Claim

Theorem 1. Let D0,D1∈ℂI×ID_0,D_1\in\mathbb C^{I\times I} and E0,E1∈ℂJ×JE_0,E_1\in\mathbb C^{J\times J} be positive definite Hermitian matrices, and let t∈[0,1]t\in[0,1]. The increment satisfies the joint convexity inequality (2): GZ((1−t)D0+tD1,(1−t)E0+tE1)≤(1−t)GZ(D0,E0)+tGZ(D1,E1).G_Z\bigl((1-t)D_0+tD_1,\,(1-t)E_0+tE_1\bigr) \leq (1-t)G_Z(D_0,E_0)+tG_Z(D_1,E_1).(2)✦

In the paper, the result reads: for each matrix size NN and each fixed complex N×NN\times N matrix ZZ, GZ(D,E)=ln⁡det⁡(D+ZE−1Z*)−ln⁡det⁡DG_Z(D,E)=\ln\det(D+ZE^{-1}Z^*)-\ln\det D is jointly convex on pairs of positive definite Hermitian matrices D,ED,E.

What Is Checked

The claim is stated over the real-valued increment (1), with a fixed complex rectangular matrix ZZ and positive definite Hermitian matrices D,ED,E of the matching sizes.