OctoInformBeta

OpenAI preprint, September 2026 · §3. A regularized matrix field · §3

Lemma 3.1: Bounds in squared Hilbert-Schmidt and matrix norms under eigenvalue and spectral Lipschitz conditions, with trace and unitary invariance

Lipschitz bounds for functional calculus on Hermitian matrices.

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

Take a finite index set SS, two complex Hermitian matrices AA and BB indexed by S×SS \times S, and a real function f:ℝ→ℝf : \mathbb{R} \to \mathbb{R}. When ff satisfies a Lipschitz bound with constant L≥0L \ge 0 on the eigenvalues or spectra of AA and BB, the question is how the difference f(A)−f(B)f(A) - f(B) under the functional calculus is bounded in terms of A−BA - B in the Hilbert-Schmidt norm and in the norm on the matrix space.

Norms and Matrix Flattening

Let MS(ℂ)M_S(\mathbb{C}) denote the space of complex matrices A=(Aij)i,j∈SA = (A_{ij})_{i,j \in S} equipped with the norm ∥⋅∥\|\cdot\|, and write A*∈MS(ℂ)A^* \in M_S(\mathbb{C}) for the conjugate transpose of AA, given by (A*)ij=Aji¯(A^*)_{ij} = \overline{A_{ji}}. A Hermitian matrix A∈MS(ℂ)A \in M_S(\mathbb{C}) is a matrix in MS(ℂ)M_S(\mathbb{C}) that satisfies A*=AA^* = A.

The squared Hilbert-Schmidt norm ∥A∥HS2\|A\|_{\mathrm{HS}}^2 of a matrix A∈MS(ℂ)A \in M_S(\mathbb{C}) is the real number given by (1): ∥A∥HS2=∑i∈S∑j∈S|Aij|2,\|A\|_{\mathrm{HS}}^2 = \sum_{i \in S} \sum_{j \in S} |A_{ij}|^2,(1)✦ The Hilbert-Schmidt norm ∥A∥HS\|A\|_{\mathrm{HS}} of a matrix A∈MS(ℂ)A \in M_S(\mathbb{C}) is the nonnegative square root of ∥A∥HS2\|A\|_{\mathrm{HS}}^2.

For a Hermitian matrix A∈MS(ℂ)A \in M_S(\mathbb{C}), write (λi(A))i∈S(\lambda_i(A))_{i \in S} for its SS-indexed family of real eigenvalues and σ(A)⊂ℝ\sigma(A) \subset \mathbb{R} for its real spectrum. For a Hermitian matrix A∈MS(ℂ)A \in M_S(\mathbb{C}) and a real function f:ℝ→ℝf : \mathbb{R} \to \mathbb{R}, denote by f(A)∈MS(ℂ)f(A) \in M_S(\mathbb{C}) the functional calculus of ff at AA.

Let ℓ2(S×S;ℂ)\ell^2(S \times S; \mathbb{C}) be the complex Euclidean space of arrays v=(v(i,j))(i,j)∈S×Sv = (v_{(i,j)})_{(i,j) \in S \times S} equipped with the ℓ2\ell^2 norm. The matrix flattening map Φ:MS(ℂ)→ℓ2(S×S;ℂ)\Phi : M_S(\mathbb{C}) \to \ell^2(S \times S; \mathbb{C}) is the continuous ℝ\mathbb{R}-linear map assigning to each A∈MS(ℂ)A \in M_S(\mathbb{C}) the array (Aij)(i,j)∈S×S(A_{ij})_{(i,j) \in S \times S}.✦ The matrix unflattening map Ψ:ℓ2(S×S;ℂ)→MS(ℂ)\Psi : \ell^2(S \times S; \mathbb{C}) \to M_S(\mathbb{C}) is the continuous ℝ\mathbb{R}-linear map assigning to each v∈ℓ2(S×S;ℂ)v \in \ell^2(S \times S; \mathbb{C}) the matrix (v(i,j))i,j∈S(v_{(i,j)})_{i,j \in S}.✦ Write ∥Φ∥\|\Phi\| and ∥Ψ∥\|\Psi\| for the operator norms of Φ\Phi and Ψ\Psi.

The Claim

Theorem 1. Let A,B∈MS(ℂ)A, B \in M_S(\mathbb{C}) be Hermitian matrices, let f:ℝ→ℝf : \mathbb{R} \to \mathbb{R} be a function, and let L≥0L \ge 0 be a real number. If |f(λi(A))−f(λj(B))|≤L|λi(A)−λj(B)||f(\lambda_i(A)) - f(\lambda_j(B))| \le L |\lambda_i(A) - \lambda_j(B)| for all i,j∈Si, j \in S, then the squared Hilbert-Schmidt norm satisfies (2): ∥f(A)−f(B)∥HS2≤L2∥A−B∥HS2.\|f(A) - f(B)\|_{\mathrm{HS}}^2 \le L^2 \|A - B\|_{\mathrm{HS}}^2.(2)✦ If |f(x)−f(y)|≤L|x−y||f(x) - f(y)| \le L |x - y| for all x∈σ(A)x \in \sigma(A) and y∈σ(B)y \in \sigma(B), then the norm on MS(ℂ)M_S(\mathbb{C}) satisfies (3): ∥f(A)−f(B)∥≤∥Ψ∥∥Φ∥L∥A−B∥.\|f(A) - f(B)\| \le \|\Psi\| \, \|\Phi\| \, L \, \|A - B\|.(3)✦

In the paper, the result reads: if a real function ff (written hh in the paper) is LL-Lipschitz on an interval containing the spectra of two Hermitian matrices AA and BB, then ∥f(A)−f(B)∥HS≤L∥A−B∥HS\|f(A) - f(B)\|_{\mathrm{HS}} \le L \|A - B\|_{\mathrm{HS}}.

What Is Checked

The theorem is stated over Hermitian matrices A,B∈MS(ℂ)A, B \in M_S(\mathbb{C}) indexed by a finite set SS, a real function f:ℝ→ℝf : \mathbb{R} \to \mathbb{R} and a constant L≥0L \ge 0 satisfying the pointwise Lipschitz condition on pairs of eigenvalues or points of the real spectra, the functional calculus on MS(ℂ)M_S(\mathbb{C}), the squared Hilbert-Schmidt norm, and the norm on MS(ℂ)M_S(\mathbb{C}) together with the continuous ℝ\mathbb{R}-linear flattening and unflattening maps Φ\Phi and Ψ\Psi.

Writing Tr\operatorname{Tr} for the trace on MS(ℂ)M_S(\mathbb{C}), the library also states the trace identity (4) for every A∈MS(ℂ)A \in M_S(\mathbb{C}): ∥A∥HS2=Re⁡Tr⁡(A*A).\|A\|_{\mathrm{HS}}^2 = \operatorname{Re}\operatorname{Tr}(A^* A).(4)✦ For every A∈MS(ℂ)A \in M_S(\mathbb{C}) and all unitary matrices U,V∈MS(ℂ)U, V \in M_S(\mathbb{C}), the squared Hilbert-Schmidt norm satisfies unitary invariance (5): ∥U*AV∥HS2=∥A∥HS2.\|U^* A V\|_{\mathrm{HS}}^2 = \|A\|_{\mathrm{HS}}^2.(5)✦