OctoInformBeta

OpenAI preprint, September 2026 · §2. A matrix energy and its Hessian

Lemma 2.1: Smoothness of the square root and first-order and second-order energy derivatives alongside spectral Hessian formulas and additive pairing bounds

Derivatives and Hessian bounds for the matrix energy.

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 index set ι\iota, and let Mι(ℂ)M_\iota(\mathbb{C}) be the space of complex square matrices indexed by ι\iota. On Mι(ℂ)M_\iota(\mathbb{C}), one considers the matrix energy and its gradient. At positive definite matrices, one seeks Fréchet and one-parameter derivatives of both the matrix square root and the energy on the positive semidefinite cone, while on the boundary one seeks one-parameter derivatives along positive semidefinite paths annihilating a fixed projection. One also seeks a spectral formula and unitary invariance for the Hessian quadratic cost, together with symmetry and pairing bounds for the associated bilinear forms.

Definitions

The cone 𝒞+\mathcal{C}_+ of positive semidefinite matrices in Mι(ℂ)M_\iota(\mathbb{C}) is the set of matrices Q∈Mι(ℂ)Q \in M_\iota(\mathbb{C}) satisfying Q≥0Q \ge 0. A matrix P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) is written P>0P > 0 when it is positive definite. For a map FF on Mι(ℂ)M_\iota(\mathbb{C}) and a matrix P∈𝒞+P \in \mathcal{C}_+, the Fréchet derivative D𝒞+F(P)D_{\mathcal{C}_+} F(P) of FF at PP within 𝒞+\mathcal{C}_+ is the Fréchet derivative over ℝ\mathbb{R} of FF at PP within 𝒞+\mathcal{C}_+. The real trace tr⁡ℝ(A)\operatorname{tr}_{\mathbb{R}}(A) of a matrix A∈Mι(ℂ)A \in M_\iota(\mathbb{C}) is the real number Re⁡tr⁡(A)\operatorname{Re}\operatorname{tr}(A).✦ The real Hilbert-Schmidt inner product ⟨X,Y⟩HS\langle X, Y \rangle_{\mathrm{HS}} of matrices X,Y∈Mι(ℂ)X, Y \in M_\iota(\mathbb{C}) is the real number tr⁡ℝ(X*Y)\operatorname{tr}_{\mathbb{R}}(X^* Y), where X*X^* denotes the conjugate transpose of XX.

The matrix energy Φ(P)\Phi(P) of a matrix P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) is the real number given by (1), Φ(P)=tr⁡ℝ(PP),\Phi(P) = \operatorname{tr}_{\mathbb{R}}(P \sqrt{P}),(1)✦ where P∈Mι(ℂ)\sqrt{P} \in M_\iota(\mathbb{C}) is the matrix square root. The energy gradient ∇ΦP\nabla\Phi_P of a matrix P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) is the functional on Mι(ℂ)M_\iota(\mathbb{C}) defined on K∈Mι(ℂ)K \in M_\iota(\mathbb{C}) by (2): ∇ΦP(K)=32tr⁡ℝ(PK).\nabla\Phi_P(K) = \frac{3}{2} \operatorname{tr}_{\mathbb{R}}(\sqrt{P} K).(2)

The Sylvester operator SA:Mι(ℂ)→Mι(ℂ)S_A : M_\iota(\mathbb{C}) \to M_\iota(\mathbb{C}) associated with A∈Mι(ℂ)A \in M_\iota(\mathbb{C}) is the map sending X∈Mι(ℂ)X \in M_\iota(\mathbb{C}) to (3): SA(X)=AX+XA.S_A(X) = A X + X A.(3)✦ When A>0A > 0, SAS_A is invertible with inverse SA−1S_A^{-1} on Mι(ℂ)M_\iota(\mathbb{C}). The inverse root Sylvester operator SP−1S_{\sqrt{P}}^{-1} of a positive definite matrix P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) is the inverse of SPS_{\sqrt{P}} on Mι(ℂ)M_\iota(\mathbb{C}).

The positive definite Hessian pairing ℬ̃P(K,L)\widetilde{\mathcal{B}}_P(K, L) of a positive definite matrix P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) and directions K,L∈Mι(ℂ)K, L \in M_\iota(\mathbb{C}) is the real number given by (4): ℬ̃P(K,L)=32⟨K,SP−1(L)⟩HS.\widetilde{\mathcal{B}}_P(K, L) = \frac{3}{2} \langle K, S_{\sqrt{P}}^{-1}(L) \rangle_{\mathrm{HS}}.(4) The Sylvester penalty pA(X)p_A(X) of matrices A,X∈Mι(ℂ)A, X \in M_\iota(\mathbb{C}) is the real number ⟨X,SA(X)⟩HS\langle X, S_A(X) \rangle_{\mathrm{HS}}, and the Sylvester variational functional VA,K(X)V_{A, K}(X) of matrices A,K,X∈Mι(ℂ)A, K, X \in M_\iota(\mathbb{C}) is the real number 2⟨K,X⟩HS−pA(X)2 \langle K, X \rangle_{\mathrm{HS}} - p_A(X). The root variational cost rA(K)r_A(K) of A,K∈Mι(ℂ)A, K \in M_\iota(\mathbb{C}) is the real number given by (5), rA(K)=32supX∈Mι(ℂ)VA,K(X),r_A(K) = \frac{3}{2} \sup_{X \in M_\iota(\mathbb{C})} V_{A, K}(X),(5)✦ The Hessian quadratic cost bP(K)b_P(K) of matrices P,K∈Mι(ℂ)P, K \in M_\iota(\mathbb{C}) is the real number given by (6), and the polarized Hessian bilinear form ℬP(K,L)\mathcal{B}_P(K, L) of matrices P,K,L∈Mι(ℂ)P, K, L \in M_\iota(\mathbb{C}) is the real number given by (7): bP(K)=rP(K),b_P(K) = r_{\sqrt{P}}(K),(6)✦ ℬP(K,L)=14(bP(K+L)−bP(K−L)).\mathcal{B}_P(K, L) = \frac{1}{4}\bigl(b_P(K + L) - b_P(K - L)\bigr).(7)✦

The spectral weight wD,m(i,j)∈ℝw_{D, m}(i, j) \in \mathbb{R} of D∈Mι(ℂ)D \in M_\iota(\mathbb{C}) and m:ι→ℝm : \iota \to \mathbb{R} at indices i,j∈ιi, j \in \iota is the real quantity (8) (with 0/0=00/0 = 0): wD,m(i,j)=|mi|2|mj|2|mi|+|mj||Dij|2.w_{D, m}(i, j) = \frac{|m_i|^2 |m_j|^2}{|m_i| + |m_j|} |D_{ij}|^2.(8)✦

The Claim

Theorem 1. Suppose P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) is positive definite. Then the matrix square root X↦XX \mapsto \sqrt{X} is C∞C^\infty over ℝ\mathbb{R} at PP within 𝒞+\mathcal{C}_+ with Fréchet derivative (9), D𝒞+(⋅)(P)=SP−1,D_{\mathcal{C}_+}(\sqrt{\cdot})(P) = S_{\sqrt{P}}^{-1},(9)✦ and the matrix energy Φ\Phi is Fréchet differentiable over ℝ\mathbb{R} at PP within 𝒞+\mathcal{C}_+ with derivative (10): D𝒞+Φ(P)=∇ΦP.D_{\mathcal{C}_+}\Phi(P) = \nabla\Phi_P.(10)✦

Let P:ℝ→Mι(ℂ)P : \mathbb{R} \to M_\iota(\mathbb{C}) be differentiable at x∈ℝx \in \mathbb{R} with derivative UU. If P(x)>0P(x) > 0 and P(t)∈𝒞+P(t) \in \mathcal{C}_+ for all tt in a neighborhood of xx, then t↦Φ(P(t))t \mapsto \Phi(P(t)) is differentiable at xx with (11): ddtΦ(P(t))|t=x=∇ΦP(x)(U).\left.\frac{d}{dt}\Phi(P(t))\right|_{t=x} = \nabla\Phi_{P(x)}(U).(11)✦ If K:ℝ→Mι(ℂ)K : \mathbb{R} \to M_\iota(\mathbb{C}) is also differentiable at xx with derivative VV and K(x)K(x) is Hermitian, then t↦∇ΦP(t)(K(t))t \mapsto \nabla\Phi_{P(t)}(K(t)) is differentiable at xx with (12): ddt∇ΦP(t)(K(t))|t=x=ℬ̃P(x)(K(x),U)+∇ΦP(x)(V).\left.\frac{d}{dt}\nabla\Phi_{P(t)}(K(t))\right|_{t=x} = \widetilde{\mathcal{B}}_{P(x)}(K(x), U) + \nabla\Phi_{P(x)}(V).(12)✦

Consider Q∈𝒞+Q \in \mathcal{C}_+ with Q2=QQ^2 = Q and a path P:ℝ→Mι(ℂ)P : \mathbb{R} \to M_\iota(\mathbb{C}) differentiable at x∈ℝx \in \mathbb{R} with derivative UU. Suppose P(x)+Q>0P(x) + Q > 0 while P(t)∈𝒞+P(t) \in \mathcal{C}_+ and P(t)Q=0P(t) Q = 0 in a neighborhood of xx. Then t↦Φ(P(t))t \mapsto \Phi(P(t)) is differentiable at xx with (13): ddtΦ(P(t))|t=x=∇ΦP(x)(U).\left.\frac{d}{dt}\Phi(P(t))\right|_{t=x} = \nabla\Phi_{P(x)}(U).(13)✦ If K:ℝ→Mι(ℂ)K : \mathbb{R} \to M_\iota(\mathbb{C}) is also differentiable at xx with derivative VV, while K(t)K(t) is Hermitian and K(t)Q=0K(t) Q = 0 in a neighborhood of xx, then t↦∇ΦP(t)(K(t))t \mapsto \nabla\Phi_{P(t)}(K(t)) is differentiable at xx with (14): ddt∇ΦP(t)(K(t))|t=x=ℬP(x)(K(x),U)+∇ΦP(x)(V).\left.\frac{d}{dt}\nabla\Phi_{P(t)}(K(t))\right|_{t=x} = \mathcal{B}_{P(x)}(K(x), U) + \nabla\Phi_{P(x)}(V).(14)✦

Given q:ι→ℝq : \iota \to \mathbb{R} with qi≥0q_i \ge 0 for all i∈ιi \in \iota and K∈Mι(ℂ)K \in M_\iota(\mathbb{C}) such that Kij=0K_{ij} = 0 whenever qi+qj=0q_i + q_j = 0, the diagonal matrix diag⁡(q)\operatorname{diag}(q) with diagonal entries qiq_i satisfies (15) (with 0/0=00/0 = 0): rdiag⁡(q)(K)=32∑i,j∈ι|Kij|2qi+qj.r_{\operatorname{diag}(q)}(K) = \frac{3}{2} \sum_{i, j \in \iota} \frac{|K_{ij}|^2}{q_i + q_j}.(15)✦ For any m:ι→ℝm : \iota \to \mathbb{R} and D∈Mι(ℂ)D \in M_\iota(\mathbb{C}), writing M=diag⁡(m)M = \operatorname{diag}(m) gives (16): bM2(MDM)=32∑i,j∈ιwD,m(i,j).b_{M^2}(M D M) = \frac{3}{2} \sum_{i, j \in \iota} w_{D, m}(i, j).(16)✦ Whenever U∈Mι(ℂ)U \in M_\iota(\mathbb{C}) is unitary, P∈𝒞+P \in \mathcal{C}_+, and K∈Mι(ℂ)K \in M_\iota(\mathbb{C}), the Hessian cost satisfies (17): bUPU*(UKU*)=bP(K).b_{U P U^*}(U K U^*) = b_P(K).(17)✦

For any positive definite matrix P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) and all K,L∈Mι(ℂ)K, L \in M_\iota(\mathbb{C}), the identities (18) and (19) and the bounds (20) and (21) hold: ℬ̃P(K,L)=ℬ̃P(L,K),\widetilde{\mathcal{B}}_P(K, L) = \widetilde{\mathcal{B}}_P(L, K),(18)✦ ℬ̃P(K,K)=bP(K),\widetilde{\mathcal{B}}_P(K, K) = b_P(K),(19)✦ 2ℬ̃P(K,L)≤bP(K)+bP(L),2 \widetilde{\mathcal{B}}_P(K, L) \le b_P(K) + b_P(L),(20)✦ 2|ℬ̃P(K,L)|≤bP(K)+bP(L).2 \lvert \widetilde{\mathcal{B}}_P(K, L) \rvert \le b_P(K) + b_P(L).(21)✦ Take P,Q∈𝒞+P, Q \in \mathcal{C}_+ satisfying Q2=QQ^2 = Q, P+Q>0P + Q > 0, and PQ=QP=0P Q = Q P = 0. If K∈Mι(ℂ)K \in M_\iota(\mathbb{C}) satisfies KQ=QK=0K Q = Q K = 0, then (22) holds, ℬP(K,K)=bP(K),\mathcal{B}_P(K, K) = b_P(K),(22)✦ and if L∈Mι(ℂ)L \in M_\iota(\mathbb{C}) also satisfies LQ=QL=0L Q = Q L = 0, then (23) holds: 2|ℬP(K,L)|≤bP(K)+bP(L).2 \lvert \mathcal{B}_P(K, L) \rvert \le b_P(K) + b_P(L).(23)✦

In the paper, the result reads: on the cone of matrices positive definite on a fixed support, Φ\Phi is smooth, with dΦP(K)=32tr⁡(PK)d\Phi_P(K) = \frac{3}{2}\operatorname{tr}(\sqrt{P}K) and d2ΦP(K,L)=ℬP(K,L)d^2\Phi_P(K, L) = \mathcal{B}_P(K, L) for Hermitian directions K,LK, L on that support; in an orthonormal eigenbasis with positive eigenvalues pip_i on the support, bP(K)=32∑i,j|Kij|2/(pi+pj)b_P(K) = \frac{3}{2}\sum_{i, j} |K_{ij}|^2 / (\sqrt{p_i} + \sqrt{p_j}); and in particular, ℬP\mathcal{B}_P is a positive definite symmetric bilinear form on the real space of supported Hermitian directions with |ℬP(K,L)|≤bP(K)1/2bP(L)1/2|\mathcal{B}_P(K, L)| \le b_P(K)^{1/2} b_P(L)^{1/2}. The theorem above represents a fixed support by a positive semidefinite projection QQ with PQ=0P Q = 0 and P+Q>0P + Q > 0 using Fréchet and one-parameter derivatives of Φ\Phi and ∇Φ\nabla\Phi, while the spectral formula splits into unitary invariance with diagonal evaluation and the pairing bounds take additive form.

What Is Checked

The claim is stated over the matrix energy Φ\Phi, its gradient ∇ΦP\nabla\Phi_P, the Sylvester operator SAS_A and inverse SP−1S_{\sqrt{P}}^{-1}, the root and Hessian costs rAr_A and bPb_P, the pairings ℬP\mathcal{B}_P and ℬ̃P\widetilde{\mathcal{B}}_P, and the spectral weight wD,mw_{D, m} on Mι(ℂ)M_\iota(\mathbb{C}), for positive definite and positive semidefinite matrices, unitary and diagonal matrices, and differentiable paths with a projection Q∈𝒞+Q \in \mathcal{C}_+. For these constructions, the real trace tr⁡ℝ\operatorname{tr}_{\mathbb{R}} and the energy gradient ∇ΦP\nabla\Phi_P are continuous real-linear functionals on Mι(ℂ)M_\iota(\mathbb{C}), every Sylvester operator SAS_A is real-linear, and both SAS_A (for A>0A > 0) and SP−1S_{\sqrt{P}}^{-1} (for P>0P > 0) are continuous automorphisms of Mι(ℂ)M_\iota(\mathbb{C}).