OctoInformBeta

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

§5: Kernel projections P(v) and R(u): Orthogonal projections onto kernels from outer products, together with the Euclidean operator norm and inverse square root

Left and right kernel projections of a normalized pencil.

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 normalized pencil of positive unital linear maps from 4×44 \times 4 real symmetric matrices to real symmetric matrices of positive orders, coupled by a Schur-complement condition. Rank-one limits and a norm identity on such pencils require the orthogonal projections onto the kernels obtained by evaluating the pencil maps at the identity minus a vector outer product, together with the Euclidean operator norm and the inverse square root of a real symmetric matrix. These notes go through those definitions and the matrix and pencil constructions they are built on.

Euclidean Space and Matrix Preliminaries

For a natural number nn and a set RR, the n×nn \times n square matrices over RR are the matrices with entries in RR whose rows and columns are indexed by {1,…,n}\{1, \dots, n\}.✦ For natural numbers mm and nn, write ℝm×n\mathbb{R}^{m \times n} for the space of real m×nm \times n matrices.

Euclidean space ℝn\mathbb{R}^n for a natural number nn is the real vector space of nn-tuples ℝn={(v1,…,vn):vi∈ℝ for 1≤i≤n},\mathbb{R}^n = \{(v_1, \dots, v_n) : v_i \in \mathbb{R} \text{ for } 1 \le i \le n\},(1)✦ equipped for u,v∈ℝnu, v \in \mathbb{R}^n with the Euclidean inner product ⟨u,v⟩=∑i=1nuivi\langle u, v \rangle = \sum_{i=1}^n u_i v_i and Euclidean norm ∥v∥2=⟨v,v⟩1/2\|v\|_2 = \langle v, v \rangle^{1/2}.

The space 𝒮n\mathcal{S}^n of real symmetric matrices is the subspace of self-adjoint matrices in ℝn×n\mathbb{R}^{n \times n}.✦ Write In∈𝒮nI_n \in \mathcal{S}^n for the n×nn \times n identity matrix; for A∈𝒮nA \in \mathcal{S}^n, write A≽0A \succeq 0 when AA is positive semidefinite and A≻0A \succ 0 when AA is positive definite.

For a natural number nn and a vector v∈ℝnv \in \mathbb{R}^n, the outer product vv⊤∈ℝn×nv v^\top \in \mathbb{R}^{n \times n} of vv is the real square matrix with entries (2). (vv⊤)ij=vivj(1≤i,j≤n).(v v^\top)_{ij} = v_i v_j \qquad (1 \le i, j \le n).(2)✦

The symmetric outer product vv⊤∈𝒮nv v^\top \in \mathcal{S}^n of a vector v∈ℝnv \in \mathbb{R}^n is the self-adjoint matrix having entries (2): vv⊤∈𝒮n,(vv⊤)⊤=vv⊤.v v^\top \in \mathcal{S}^n, \qquad (v v^\top)^\top = v v^\top.(3)✦

For a real square matrix A∈ℝn×nA \in \mathbb{R}^{n \times n} with kernel ker⁡A={x∈ℝn:Ax=0}\ker A = \{x \in \mathbb{R}^n : A x = 0\}, the kernel projection proj⁡ker⁡A∈ℝn×n\operatorname{proj}_{\ker A} \in \mathbb{R}^{n \times n} of AA is the matrix of the orthogonal projection of ℝn\mathbb{R}^n onto ker⁡A\ker A, satisfying (4). (proj⁡ker⁡A)x∈ker⁡Aand⟨x−(proj⁡ker⁡A)x,z⟩=0(x∈ℝn,z∈ker⁡A).(\operatorname{proj}_{\ker A}) x \in \ker A \quad\text{and}\quad \langle x - (\operatorname{proj}_{\ker A}) x, z \rangle = 0 \qquad (x \in \mathbb{R}^n,\; z \in \ker A).(4)✦

Normalized Pencils

For y=(y1,y2,y3)∈ℝ3y = (y_1, y_2, y_3) \in \mathbb{R}^3, the quadratic matrix Q(y)∈ℝ3×3Q(y) \in \mathbb{R}^{3 \times 3} of yy is the matrix with entries (5) for cyclic indices i,j∈{1,2,3}i, j \in \{1, 2, 3\}. 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 \ne j. \end{cases}(5)✦

For y∈ℝ3y \in \mathbb{R}^3, the linear map Φy:𝒮4→𝒮4\Phi_y : \mathcal{S}^4 \to \mathcal{S}^4 of yy is the linear map acting in (1+3)×(1+3)(1 + 3) \times (1 + 3) block form on any matrix in 𝒮4\mathcal{S}^4 with top-left entry α∈ℝ\alpha \in \mathbb{R}, off-diagonal column r∈ℝ3r \in \mathbb{R}^3, and lower-right block T∈𝒮3T \in \mathcal{S}^3 by (6). Φy(αr⊤rT)=(tr⁡(Q(y)T)−r⊤Q(y)−Q(y)rαQ(y)).\Phi_y\!\begin{pmatrix} \alpha & r^\top \\ r & T \end{pmatrix} = \begin{pmatrix} \operatorname{tr}(Q(y) T) & -r^\top Q(y) \\ -Q(y) r & \alpha Q(y) \end{pmatrix}.(6)✦

A normalized pencil (D,E,B)(D, E, B) consists of positive integers aa and cc together with 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}. For a normalized pencil (D,E,B)(D, E, B), the maps DD and EE are positive (D(X)≽0D(X) \succeq 0 and E(X)≽0E(X) \succeq 0 whenever X∈𝒮4X \in \mathcal{S}^4 satisfies X≽0X \succeq 0) and unital (D(I4)=IaD(I_4) = I_a and E(I4)=IcE(I_4) = I_c). Together with (6), these maps satisfy the Schur-complement equivalence: for all X,Z∈𝒮4X, Z \in \mathcal{S}^4 with X≻0X \succ 0 and all y∈ℝ3y \in \mathbb{R}^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.(7)

The Definition

Definition 1. Let mm and nn be natural numbers and let A∈ℝm×nA \in \mathbb{R}^{m \times n}. The Euclidean operator norm ∥A∥op∈ℝ\|A\|_{\mathrm{op}} \in \mathbb{R} of AA is the operator norm of the linear map x↦Axx \mapsto A x from ℝn\mathbb{R}^n to ℝm\mathbb{R}^m, ∥A∥op=supx∈ℝn∥x∥2≤1∥Ax∥2.\|A\|_{\mathrm{op}} = \sup_{\substack{x \in \mathbb{R}^n \\ \|x\|_2 \le 1}} \|A x\|_2.(8)✦

For a symmetric matrix A∈𝒮nA \in \mathcal{S}^n with spectral decomposition A=Udiag⁡(λ1,…,λn)U⊤A = U \operatorname{diag}(\lambda_1, \dots, \lambda_n) U^\top, where U∈ℝn×nU \in \mathbb{R}^{n \times n} is orthogonal and λ1,…,λn∈ℝ\lambda_1, \dots, \lambda_n \in \mathbb{R} are the eigenvalues of AA, the inverse square root A−1/2∈ℝn×nA^{-1/2} \in \mathbb{R}^{n \times n} of AA is the matrix A−1/2=Udiag⁡(d1,…,dn)U⊤,di={λi−1/2if λi>0,0if λi≤0.A^{-1/2} = U \operatorname{diag}(d_1, \dots, d_n) U^\top, \qquad d_i = \begin{cases} \lambda_i^{-1/2} & \text{if } \lambda_i > 0, \\ 0 & \text{if } \lambda_i \le 0. \end{cases}(9)✦

Given a normalized pencil (D,E,B)(D, E, B) with dimensions a,c≥1a, c \ge 1 and v∈ℝ4v \in \mathbb{R}^4, the left kernel projection P(v)∈ℝa×aP(v) \in \mathbb{R}^{a \times a} of vv is the matrix defined via (3) and (4) by P(v)=proj⁡ker⁡D(I4−vv⊤).P(v) = \operatorname{proj}_{\ker D(I_4 - v v^\top)}.(10)✦ The right kernel projection R(u)∈ℝc×cR(u) \in \mathbb{R}^{c \times c} associated with a normalized pencil (D,E,B)(D, E, B) of dimensions a,c≥1a, c \ge 1 and a vector u∈ℝ4u \in \mathbb{R}^4 is the matrix R(u)=proj⁡ker⁡E(I4−uu⊤).R(u) = \operatorname{proj}_{\ker E(I_4 - u u^\top)}.(11)✦

In the paper, the result reads: for unit vectors u,v∈ℝ4u, v \in \mathbb{R}^4, P(v)=proj⁡ker⁡D(I4−vv⊤)P(v) = \operatorname{proj}_{\ker D(I_4 - v v^\top)} acting on ℝa\mathbb{R}^a and R(u)=proj⁡ker⁡E(I4−uu⊤)R(u) = \operatorname{proj}_{\ker E(I_4 - u u^\top)} acting on ℝc\mathbb{R}^c (5.1). Here (10) and (11) are defined for arbitrary u,v∈ℝ4u, v \in \mathbb{R}^4 without requiring unit length.

What Is Checked

The definitions are stated over a set RR for square matrices, arbitrary natural numbers mm and nn, vectors in ℝn\mathbb{R}^n, real matrices in ℝm×n\mathbb{R}^{m \times n} and ℝn×n\mathbb{R}^{n \times n}, symmetric matrices in 𝒮n\mathcal{S}^n, arbitrary vectors u,v∈ℝ4u, v \in \mathbb{R}^4, and every normalized pencil (D,E,B)(D, E, B) built from Q(y)Q(y) and Φy\Phi_y for y∈ℝ3y \in \mathbb{R}^3 under the Schur-complement equivalence (7).