OctoInformBeta

OpenAI preprint, September 2026 · §5. Rank-one limits and a norm identity

Proposition 5.1: A positive definite norm equality and inverse square root limits for positive semidefinite matrices

Normalized pencils satisfy the Choi-Lam norm identity.

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 two positive unital linear maps sending 4×44 \times 4 real symmetric matrices to real symmetric matrices of positive orders, together with a linear map from ℝ3\mathbb{R}^3 to rectangular matrices linked by a Schur-complement equivalence. A unit vector in ℝ4\mathbb{R}^4 acts through the difference between the identity matrix and its outer product to cut out orthogonal kernel projections in the two target spaces. For two unit vectors in ℝ4\mathbb{R}^4 and a parameter in ℝ3\mathbb{R}^3, the problem relates a biquadratic form to the Euclidean operator norm of the projected rectangular matrix and determines when that projected block vanishes. The Euclidean operator norm of a positive semidefinite matrix equals its maximum eigenvalue, and inverse square roots along affine rays of positive semidefinite matrices converge to the orthogonal projection onto the kernel.

Definitions

For a nonnegative integer nn, real nn-dimensional Euclidean space ℝn\mathbb{R}^n is the space of real nn-tuples with the standard Euclidean inner product✦, and the space 𝒮n\mathcal{S}^n of real symmetric matrices of order nn is the subspace of real n×nn \times n square matrices equal to their transpose.✦ Write A≽0A \succeq 0 when a real n×nn \times n matrix AA is positive semidefinite and A≻0A \succ 0 when AA is positive definite, and let InI_n denote the n×nn \times n identity matrix. Over a commutative ring RR, write RnR^n for the nn-tuples with entries in RR and Rm×nR^{m \times n} for the m×nm \times n matrices over RR (with Rn×nR^{n \times n} the n×nn \times n square matrices).

Let RR be a commutative ring, with indices in {1,2,3}\{1, 2, 3\} taken cyclically. For y∈R3y \in R^3, the quadratic matrix Q(y)∈R3×3Q(y) \in R^{3 \times 3} is the 3×33 \times 3 matrix over RR with (i,j)(i, j) entry Q(y)ij={yi2+yi+12if i=j,−yiyjif i≠j.Q(y)_{ij} = \begin{cases} y_i^2 + y_{i+1}^2 & \text{if } i = j, \\ -y_i y_j & \text{if } i \neq j. \end{cases}(1)✦ For z,y∈R3z, y \in R^3, the Choi-Lam biquadratic form b(z,y)∈Rb(z, y) \in R is the polynomial defined in terms of (1) by b(z,y)=∑i=13zi2yi2−2(z1y1z2y2+z1y1z3y3+z2y2z3y3)+∑i=13zi2yi+12=z⊤Q(y)z.b(z, y) = \sum_{i=1}^3 z_i^2 y_i^2 - 2(z_1 y_1 z_2 y_2 + z_1 y_1 z_3 y_3 + z_2 y_2 z_3 y_3) + \sum_{i=1}^3 z_i^2 y_{i+1}^2 = z^\top Q(y) z.(2)✦

Given y∈R3y \in R^3 over a commutative ring RR, the matrix map Φy:R4×4→R4×4\Phi_y : R^{4 \times 4} \to R^{4 \times 4} acts on a 4×44 \times 4 matrix over RR written in (1+3)×(1+3)(1 + 3) \times (1 + 3) block form (ar⊤sT)\begin{pmatrix} a & r^\top \\ s & T \end{pmatrix} with a∈Ra \in R, r,s∈R3r, s \in R^3, and T∈R3×3T \in R^{3 \times 3} by Φy(ar⊤sT)=(tr⁡(Q(y)T)−r⊤Q(y)−Q(y)saQ(y)).\Phi_y \begin{pmatrix} a & r^\top \\ s & T \end{pmatrix} = \begin{pmatrix} \operatorname{tr}(Q(y) T) & -r^\top Q(y) \\ -Q(y) s & a Q(y) \end{pmatrix}.(3)✦

A normalized pencil (D,E,B)(D, E, B) with integer dimensions a,c≥1a, c \ge 1 consists of linear maps D:𝒮4→𝒮aD : \mathcal{S}^4 \to \mathcal{S}^a, E:𝒮4→𝒮cE : \mathcal{S}^4 \to \mathcal{S}^c, and B:ℝ3→ℝa×cB : \mathbb{R}^3 \to \mathbb{R}^{a \times c} such that DD and EE are positive (D(X)≽0D(X) \succeq 0 and E(X)≽0E(X) \succeq 0 for every X∈𝒮4X \in \mathcal{S}^4 with X≽0X \succeq 0) and unital (D(I4)=IaD(I_4) = I_a and E(I4)=IcE(I_4) = I_c), and the Schur-complement equivalence holds: for all X,Z∈𝒮4X, Z \in \mathcal{S}^4 with X≻0X \succ 0 and all y∈ℝ3y \in \mathbb{R}^3, with Φy\Phi_y given by (3), Z−Φy(X−1)≽0⇔E(Z)−B(y)⊤D(X)−1B(y)≽0.Z - \Phi_y(X^{-1}) \succeq 0 \iff E(Z) - B(y)^\top D(X)^{-1} B(y) \succeq 0.(4)✦

Given a nonnegative integer nn and a vector v∈ℝnv \in \mathbb{R}^n, the outer product vv⊤∈ℝn×nv v^\top \in \mathbb{R}^{n \times n} is the symmetric n×nn \times n matrix with (i,j)(i, j) entry vivjv_i v_j.✦

For any real square matrix A∈ℝn×nA \in \mathbb{R}^{n \times n} (with n≥0n \ge 0), its kernel projection proj⁡ker⁡A∈ℝn×n\operatorname{proj}_{\ker A} \in \mathbb{R}^{n \times n} is the matrix representing the orthogonal projection of ℝn\mathbb{R}^n onto ker⁡A\ker A.✦

For a normalized pencil (D,E,B)(D, E, B) with dimensions a,c≥1a, c \ge 1 and vectors v,u∈ℝ4v, u \in \mathbb{R}^4, the left kernel projection P(v)∈ℝa×aP(v) \in \mathbb{R}^{a \times a} and the right kernel projection R(u)∈ℝc×cR(u) \in \mathbb{R}^{c \times c} are P(v)=proj⁡ker⁡D(I4−vv⊤),R(u)=proj⁡ker⁡E(I4−uu⊤).P(v) = \operatorname{proj}_{\ker D(I_4 - v v^\top)}, \qquad R(u) = \operatorname{proj}_{\ker E(I_4 - u u^\top)}.(5)

The inverse square root A−1/2∈ℝn×nA^{-1/2} \in \mathbb{R}^{n \times n} of a symmetric matrix A∈𝒮nA \in \mathcal{S}^n (for a nonnegative integer nn), given an orthogonal matrix U∈ℝn×nU \in \mathbb{R}^{n \times n} of eigenvectors of AA with eigenvalues λ1,…,λn\lambda_1, \dots, \lambda_n, is the matrix A−1/2=UΔU⊤,A^{-1/2} = U \Delta U^\top,(6)✦ where Δ∈ℝn×n\Delta \in \mathbb{R}^{n \times n} is the diagonal matrix with ii-th diagonal entry (λi)−1(\sqrt{\lambda_i})^{-1}.

The Euclidean operator norm ∥A∥op∈ℝ\|A\|_{\mathrm{op}} \in \mathbb{R} of a matrix A∈ℝm×nA \in \mathbb{R}^{m \times n} (for nonnegative integers mm and nn) is the norm of the linear map ℝn→ℝm\mathbb{R}^n \to \mathbb{R}^m defined by AA.✦

For vectors u=(u0,u′)u = (u_0, u') and v=(v0,v′)v = (v_0, v') in ℝ4\mathbb{R}^4 with u0,v0∈ℝu_0, v_0 \in \mathbb{R} and u′,v′∈ℝ3u', v' \in \mathbb{R}^3, the wedge coordinates of vv and uu form the vector v0u′−u0v′∈ℝ3v_0 u' - u_0 v' \in \mathbb{R}^3.✦

The Claim

Given an integer n≥1n \ge 1 and a positive semidefinite matrix A∈ℝn×nA \in \mathbb{R}^{n \times n}, write σℝ(A)={λ∈ℝ:A−λIn is not invertible}\sigma_{\mathbb{R}}(A) = \{\lambda \in \mathbb{R} : A - \lambda I_n \text{ is not invertible}\} for its real spectrum; then ∥A∥op=max⁡σℝ(A).\|A\|_{\mathrm{op}} = \max \sigma_{\mathbb{R}}(A).(7)✦

Take a nonnegative integer nn and a positive semidefinite matrix H∈𝒮nH \in \mathcal{S}^n. As t→∞t \to \infty in ℝ\mathbb{R}, the family (In+(t−1)H)−1/2(I_n + (t - 1) H)^{-1/2} defined by (6) satisfies limt→∞(In+(t−1)H)−1/2=proj⁡ker⁡H\lim_{t \to \infty} (I_n + (t - 1) H)^{-1/2} = \operatorname{proj}_{\ker H}(8)✦ in ℝn×n\mathbb{R}^{n \times n}, and limt→∞∥(In+(t−1)H)−1/2−proj⁡ker⁡H∥op=0.\lim_{t \to \infty} \bigl\|(I_n + (t - 1) H)^{-1/2} - \operatorname{proj}_{\ker H}\bigr\|_{\mathrm{op}} = 0.(9)✦

Suppose (D,E,B)(D, E, B) is a normalized pencil with dimensions a,c≥1a, c \ge 1, matrices X,Z∈𝒮4X, Z \in \mathcal{S}^4 satisfy X≻0X \succ 0 and Z≻0Z \succ 0 with D(X)≻0D(X) \succ 0 and E(Z)≻0E(Z) \succ 0, and y∈ℝ3y \in \mathbb{R}^3; then ∥Z−1/2Φy(X−1)Z−1/2∥op=∥D(X)−1/2B(y)E(Z)−1/2∥op2.\bigl\|Z^{-1/2} \Phi_y(X^{-1}) Z^{-1/2}\bigr\|_{\mathrm{op}} = \bigl\|D(X)^{-1/2} B(y) E(Z)^{-1/2}\bigr\|_{\mathrm{op}}^2.(10)✦

Theorem 1. Let (D,E,B)(D, E, B) be a normalized pencil with dimensions a,c≥1a, c \ge 1. For every pair of unit vectors u=(u0,u′)u = (u_0, u') and v=(v0,v′)v = (v_0, v') in ℝ4\mathbb{R}^4 and every y∈ℝ3y \in \mathbb{R}^3, b(v0u′−u0v′,y)=∥P(v)B(y)R(u)∥op2.b(v_0 u' - u_0 v', y) = \|P(v) B(y) R(u)\|_{\mathrm{op}}^2.(11)✦ For every unit vector v∈ℝ4v \in \mathbb{R}^4 and every y∈ℝ3y \in \mathbb{R}^3, P(v)B(y)R(v)=0.P(v) B(y) R(v) = 0.(12)✦

In the paper, the result reads: if a,c≥1a, c \ge 1, D:𝒮4→𝒮aD : \mathcal{S}^4 \to \mathcal{S}^a and E:𝒮4→𝒮cE : \mathcal{S}^4 \to \mathcal{S}^c are positive unital linear maps, and B:ℝ3→ℝa×cB : \mathbb{R}^3 \to \mathbb{R}^{a \times c} is linear such that Z−Φy(X−1)≽0⇔E(Z)−B(y)⊤D(X)−1B(y)≽0Z - \Phi_y(X^{-1}) \succeq 0 \iff E(Z) - B(y)^\top D(X)^{-1} B(y) \succeq 0 for every X≻0X \succ 0, Z∈𝒮4Z \in \mathcal{S}^4, and y∈ℝ3y \in \mathbb{R}^3, where Φy\Phi_y and bb stand for the paper’s block map and Choi-Lam form, and P(v)P(v) and R(u)R(u) stand for the paper’s kernel projections proj⁡ker⁡D(I4−vv⊤)\operatorname{proj}_{\ker D(I_4 - v v^\top)} and proj⁡ker⁡E(I4−uu⊤)\operatorname{proj}_{\ker E(I_4 - u u^\top)}, then for every pair of unit vectors u=(u0,u′)u = (u_0, u') and v=(v0,v′)v = (v_0, v') in ℝ4\mathbb{R}^4 and every y∈ℝ3y \in \mathbb{R}^3, b(v0u′−u0v′,y)=∥P(v)B(y)R(u)∥op2b(v_0 u' - u_0 v', y) = \|P(v) B(y) R(u)\|_{\mathrm{op}}^2.

What Is Checked

The library states (7) over integers n≥1n \ge 1 and positive semidefinite matrices A∈ℝn×nA \in \mathbb{R}^{n \times n}; (8) and (9) over nonnegative integers nn and positive semidefinite matrices H∈𝒮nH \in \mathcal{S}^n; (10) over a normalized pencil (D,E,B)(D, E, B), matrices X,Z∈𝒮4X, Z \in \mathcal{S}^4 with X≻0X \succ 0, Z≻0Z \succ 0, D(X)≻0D(X) \succ 0, and E(Z)≻0E(Z) \succ 0, and y∈ℝ3y \in \mathbb{R}^3; (11) over a normalized pencil (D,E,B)(D, E, B), unit vectors u,v∈ℝ4u, v \in \mathbb{R}^4, and y∈ℝ3y \in \mathbb{R}^3; and (12) over a normalized pencil (D,E,B)(D, E, B), a unit vector v∈ℝ4v \in \mathbb{R}^4, and y∈ℝ3y \in \mathbb{R}^3.