OctoInformBeta

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

§2 (setting): matrix energy, support, and Hessian quadratic form: Gradient and variational Hessian forms via the Sylvester operator, agreeing with the inverse Sylvester operator on positive definite matrices and invariant under kernel projections

Matrix energy and its variational Hessian on complex 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

Let ι\iota be a finite index set, and let Mι(ℂ)M_\iota(\mathbb{C}) be the real vector space of ι×ι\iota \times \iota complex matrices. Write A≥0A \ge 0 when A∈Mι(ℂ)A \in M_\iota(\mathbb{C}) is positive semidefinite and A>0A > 0 when AA is positive definite, and let X*X^* denote the conjugate transpose of X∈Mι(ℂ)X \in M_\iota(\mathbb{C}) and P\sqrt{P} denote the functional-calculus square root of P∈Mι(ℂ)P \in M_\iota(\mathbb{C}).

On Mι(ℂ)M_\iota(\mathbb{C}), the problem is to equip square complex matrices with a matrix energy and gradient functional together with a variational Hessian form applicable both on the positive definite cone and on positive semidefinite matrices annihilating an orthogonal projection.

Real Trace, Inner Product, and the Sylvester Operator

The real trace tr⁡ℝ(A)∈ℝ\operatorname{tr}_{\mathbb{R}}(A) \in \mathbb{R} of a matrix A∈Mι(ℂ)A \in M_\iota(\mathbb{C}) is the real-valued functional defined by tr⁡ℝ(A)=Re⁡tr⁡(A).\operatorname{tr}_{\mathbb{R}}(A) = \operatorname{Re}\operatorname{tr}(A).(1)✦ The real Hilbert-Schmidt inner product hsReal⁡(X,Y)∈ℝ\operatorname{hsReal}(X, Y) \in \mathbb{R} of two matrices X,Y∈Mι(ℂ)X, Y \in M_\iota(\mathbb{C}) is the real pairing given by hsReal⁡(X,Y)=tr⁡ℝ(X*Y).\operatorname{hsReal}(X, Y) = \operatorname{tr}_{\mathbb{R}}(X^* Y).(2)✦

The Sylvester operator SA:Mι(ℂ)→Mι(ℂ)S_A : M_\iota(\mathbb{C}) \to M_\iota(\mathbb{C}) of a matrix A∈Mι(ℂ)A \in M_\iota(\mathbb{C}) is the matrix map acting on X∈Mι(ℂ)X \in M_\iota(\mathbb{C}) by SA(X)=AX+XA.S_A(X) = A X + X A.(3)✦ When A∈Mι(ℂ)A \in M_\iota(\mathbb{C}) is positive definite, SAS_A is a continuous real-linear automorphism of Mι(ℂ)M_\iota(\mathbb{C}).✦ The inverse root Sylvester operator SP−1:Mι(ℂ)→Mι(ℂ)S_{\sqrt{P}}^{-1} : M_\iota(\mathbb{C}) \to M_\iota(\mathbb{C}) of a positive definite matrix P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) is the continuous real-linear inverse of SPS_{\sqrt{P}}, satisfying for every L∈Mι(ℂ)L \in M_\iota(\mathbb{C}) SP(SP−1(L))=L=SP−1(SP(L)).S_{\sqrt{P}}\bigl(S_{\sqrt{P}}^{-1}(L)\bigr) = L = S_{\sqrt{P}}^{-1}\bigl(S_{\sqrt{P}}(L)\bigr).(4)✦

The Variational Functional and Root Cost

The Sylvester penalty 𝒫A(X)∈ℝ\mathcal{P}_A(X) \in \mathbb{R} of a pair A,X∈Mι(ℂ)A, X \in M_\iota(\mathbb{C}) is the real quantity defined by 𝒫A(X)=hsReal⁡(X,SA(X)).\mathcal{P}_A(X) = \operatorname{hsReal}\bigl(X, S_A(X)\bigr).(5)✦ The variational functional VA(K,X)∈ℝV_A(K, X) \in \mathbb{R} of matrices A,K,X∈Mι(ℂ)A, K, X \in M_\iota(\mathbb{C}) is the real-valued function given by VA(K,X)=2hsReal⁡(K,X)−𝒫A(X).V_A(K, X) = 2\operatorname{hsReal}(K, X) - \mathcal{P}_A(X).(6)✦ The root variational cost cA(K)∈ℝc_A(K) \in \mathbb{R} attached to matrices A,K∈Mι(ℂ)A, K \in M_\iota(\mathbb{C}) is the real value defined by cA(K)=32supX∈Mι(ℂ)VA(K,X).c_A(K) = \frac{3}{2} \sup_{X \in M_\iota(\mathbb{C})} V_A(K, X).(7)✦

The Claim

Definition 1. The matrix energy Φ(P)∈ℝ\Phi(P) \in \mathbb{R} of a matrix P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) is the real value defined by Φ(P)=tr⁡ℝ(PP).\Phi(P) = \operatorname{tr}_{\mathbb{R}}(P\sqrt{P}).(8)✦ The energy gradient ∇Φ(P):Mι(ℂ)→ℝ\nabla\Phi(P) : M_\iota(\mathbb{C}) \to \mathbb{R} at P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) is the continuous real-linear functional acting on X∈Mι(ℂ)X \in M_\iota(\mathbb{C}) by ∇Φ(P)(X)=32tr⁡ℝ(PX).\nabla\Phi(P)(X) = \frac{3}{2}\operatorname{tr}_{\mathbb{R}}(\sqrt{P}\,X).(9)✦ The Hessian quadratic form bP(K)∈ℝb_P(K) \in \mathbb{R} of matrices P,K∈Mι(ℂ)P, K \in M_\iota(\mathbb{C}) is the real quantity given by bP(K)=cP(K).b_P(K) = c_{\sqrt{P}}(K).(10)✦ The Hessian bilinear form ℬP(K,L)∈ℝ\mathcal{B}_P(K, L) \in \mathbb{R} for P,K,L∈Mι(ℂ)P, K, L \in M_\iota(\mathbb{C}) is the real pairing defined by ℬP(K,L)=bP(K+L)−bP(K−L)4.\mathcal{B}_P(K, L) = \frac{b_P(K + L) - b_P(K - L)}{4}.(11)✦ The positive definite Hessian ℬPPD(K,L)∈ℝ\mathcal{B}_P^{\mathrm{PD}}(K, L) \in \mathbb{R} of a positive definite matrix P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) and matrices K,L∈Mι(ℂ)K, L \in M_\iota(\mathbb{C}) is the real pairing ℬPPD(K,L)=32hsReal⁡(K,SP−1(L)).\mathcal{B}_P^{\mathrm{PD}}(K, L) = \frac{3}{2}\operatorname{hsReal}\bigl(K, S_{\sqrt{P}}^{-1}(L)\bigr).(12)✦ If A∈Mι(ℂ)A \in M_\iota(\mathbb{C}) is positive semidefinite and K,Y∈Mι(ℂ)K, Y \in M_\iota(\mathbb{C}) satisfy SA(Y)=KS_A(Y) = K, then cA(K)=32hsReal⁡(K,Y).c_A(K) = \frac{3}{2}\operatorname{hsReal}(K, Y).(13)✦ Whenever P∈Mι(ℂ)P \in M_\iota(\mathbb{C}) is positive definite and K,L∈Mι(ℂ)K, L \in M_\iota(\mathbb{C}), one has ℬP(K,L)=ℬPPD(K,L).\mathcal{B}_P(K, L) = \mathcal{B}_P^{\mathrm{PD}}(K, L).(14)✦ If P,Q∈Mι(ℂ)P, Q \in M_\iota(\mathbb{C}) are positive semidefinite with Q2=QQ^2 = Q, PQ=QP=0PQ = QP = 0, and P+Q>0P + Q > 0, and K,L∈Mι(ℂ)K, L \in M_\iota(\mathbb{C}) satisfy KQ=QK=0KQ = QK = 0 and LQ=QL=0LQ = QL = 0, then ℬP+Q(K,L)=ℬP(K,L).\mathcal{B}_{P + Q}(K, L) = \mathcal{B}_P(K, L).(15)✦

In the paper, the result reads: for positive semidefinite PP, Φ(P)=tr⁡(P3/2)\Phi(P) = \operatorname{tr}(P^{3/2}), supp⁡P=(ker⁡P)⟂\operatorname{supp} P = (\ker P)^\perp, ⟨K,L⟩HS=tr⁡(K*L)\langle K, L \rangle_{\mathrm{HS}} = \operatorname{tr}(K^* L), LA(X)=AXL_A(X) = AX, and RA(X)=XAR_A(X) = XA, and for Hermitian KK and LL supported on PP, ℬP(K,L)=32Re⁡⟨K,(LP+RP)−1L⟩HS\mathcal{B}_P(K, L) = \frac{3}{2}\operatorname{Re}\langle K, (L_{\sqrt{P}} + R_{\sqrt{P}})^{-1} L \rangle_{\mathrm{HS}} and bP(K)=ℬP(K,K)b_P(K) = \mathcal{B}_P(K, K) (00 if P=0P = 0), with hsReal⁡(K,L)\operatorname{hsReal}(K, L) standing for Re⁡⟨K,L⟩HS\operatorname{Re}\langle K, L \rangle_{\mathrm{HS}} and SAS_A for LA+RAL_A + R_A.

What Is Checked

The matrix energy and Hessian constructions, along with their underlying operators and identities, are stated over a finite index set ι\iota, with arbitrary matrices in Mι(ℂ)M_\iota(\mathbb{C}) for the real trace, real Hilbert-Schmidt inner product, matrix energy, energy gradient, Sylvester operator, and variational constructions; with A>0A > 0 for the Sylvester automorphism and P>0P > 0 for the inverse root Sylvester operator, positive definite Hessian, and coincidence identity; with A≥0A \ge 0 and SA(Y)=KS_A(Y) = K for a Sylvester solution; and with P,Q≥0P, Q \ge 0 satisfying Q2=QQ^2 = Q, P+Q>0P + Q > 0, and PQ=QP=KQ=QK=LQ=QL=0PQ = QP = KQ = QK = LQ = QL = 0 for kernel-projection invariance. For these constructions, the library proves that tr⁡ℝ\operatorname{tr}_{\mathbb{R}} is additive and real-homogeneous and defines a continuous real-linear functional on Mι(ℂ)M_\iota(\mathbb{C}), and that for every A∈Mι(ℂ)A \in M_\iota(\mathbb{C}) the operator SAS_A is additive and real-homogeneous as a real-linear endomorphism of Mι(ℂ)M_\iota(\mathbb{C}).