OctoInformBeta

OpenAI preprint, September 2026 · §6. The tangent obstruction (proof of Theorem 1.1)

§6: Local constant-rank smooth extension of R(u): The Gram projection, smooth local extensions for self-adjoint families and E(I−uu⊤)E(I - u u^\top), and derivatives of R(u)R(u) along spherical tangent curves at s=0s = 0

Kernel projections extend smoothly with constant rank near some point.

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

Consider a linear map EE from 4×44 \times 4 to c×cc \times c real symmetric matrices (c≥1c \ge 1) and the symmetric matrices E(I−uu⊤)E(I - u u^\top) formed from the 4×44 \times 4 identity matrix II and the outer product uu⊤u u^\top of a unit vector u=(u0,u′)∈ℝ4u = (u_0, u') \in \mathbb{R}^4 with u0∈ℝu_0 \in \mathbb{R} nonzero and u′∈ℝ3u' \in \mathbb{R}^3. Because orthogonal projections onto kernels can jump discontinuously wherever the rank drops, the question is when the rank stabilizes locally so that the orthogonal projection onto the kernel of E(I−uu⊤)E(I - u u^\top) extends smoothly to an open set in ℝ4\mathbb{R}^4 and differentiates along directions tangent to the sphere.

More generally, for any smooth family of self-adjoint operators on a finite-dimensional real inner product space parametrized by a subset of a real normed space, one asks for local rank constancy and a smooth local extension of the orthogonal projection onto the kernel.

Definitions

For a nonnegative integer nn and a set RR, the space Rn×nR^{n \times n} of n×nn \times n square matrices over RR is the set of matrices with entries in RR indexed by nn rows and nn columns.✦ For a nonnegative integer nn, Euclidean space ℝn\mathbb{R}^n is the nn-dimensional real Euclidean vector space with standard inner product ⟨⋅,⋅⟩\langle \cdot, \cdot \rangle and norm ∥⋅∥\|\cdot\|✦, and the space SnS^n of real symmetric matrices is the subspace of self-adjoint matrices in ℝn×n\mathbb{R}^{n \times n}.✦ For real normed spaces E1E_1 and E2E_2, the space ℒ(E1,E2)\mathcal{L}(E_1, E_2) is the real normed space of bounded linear operators from E1E_1 to E2E_2, and for a real normed space FF, the space ℒ(F)\mathcal{L}(F) of bounded linear operators is ℒ(F,F)\mathcal{L}(F, F). For a finite-dimensional real normed space FF and an operator L∈ℒ(F)L \in \mathcal{L}(F), the rank rank⁡L\operatorname{rank} L of LL is the dimension of the image of LL. For a nonnegative integer cc and a symmetric matrix M∈ScM \in S^c, the rank rank⁡M\operatorname{rank} M of MM is the rank of the bounded linear operator on ℝc\mathbb{R}^c induced by MM.

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 square matrix with (i,j)(i, j) entry vivjv_i v_j✦, and the symmetric outer product vv⊤∈Snv v^\top \in S^n is the symmetric matrix with (i,j)(i, j) entry vivjv_i v_j.✦ For a finite-dimensional real inner product space FF and an operator L∈ℒ(F)L \in \mathcal{L}(F), the orthogonal projection proj⁡ker⁡L∈ℒ(F)\operatorname{proj}_{\ker L} \in \mathcal{L}(F) onto the kernel of LL is the orthogonal projection of FF onto ker⁡L\ker L. For a nonnegative integer nn and a matrix M∈ℝn×nM \in \mathbb{R}^{n \times n}, the kernel projection proj⁡ker⁡M∈ℝn×n\operatorname{proj}_{\ker M} \in \mathbb{R}^{n \times n} is the square matrix representing the orthogonal projection of ℝn\mathbb{R}^n onto the kernel of the Euclidean linear map defined by MM.✦

Decompose vectors u∈ℝ4u \in \mathbb{R}^4 as u=(u0,u′)u = (u_0, u') with u0∈ℝu_0 \in \mathbb{R} and u′∈ℝ3u' \in \mathbb{R}^3. The unit sphere S3⊂ℝ4S^3 \subset \mathbb{R}^4 is the set {u∈ℝ4:∥u∥=1}\{u \in \mathbb{R}^4 : \|u\| = 1\}, and the domain Ω⊂S3\Omega \subset S^3 is the subset {u∈S3:u0≠0}\{u \in S^3 : u_0 \neq 0\}. For a vector v∈ℝ4v \in \mathbb{R}^4, the orthogonal complement v⟂⊂ℝ4v^\perp \subset \mathbb{R}^4 of vv is the subspace {h∈ℝ4:⟨v,h⟩=0}\{h \in \mathbb{R}^4 : \langle v, h \rangle = 0\}. For a nonnegative integer cc, a linear map E:S4→ScE : S^4 \to S^c, the 4×44 \times 4 identity matrix II, and a vector u∈ℝ4u \in \mathbb{R}^4, the kernel projection R(u)∈ℝc×cR(u) \in \mathbb{R}^{c \times c} is proj⁡ker⁡E(I−uu⊤)\operatorname{proj}_{\ker E(I - u u^\top)}. For vectors v,h∈ℝ4v, h \in \mathbb{R}^4 and a scalar s∈ℝs \in \mathbb{R}, the spherical tangent curve γv,h:ℝ→ℝ4\gamma_{v,h} : \mathbb{R} \to \mathbb{R}^4 is given by (1): γv,h(s)=(1+s2∥h∥2)−1/2(v+sh).\gamma_{v,h}(s) = (1 + s^2 \|h\|^2)^{-1/2} (v + s h).(1)✦

The Gram projection gram⁡(B)∈ℒ(H2)\operatorname{gram}(B) \in \mathcal{L}(H_2) of a bounded linear operator B∈ℒ(H1,H2)B \in \mathcal{L}(H_1, H_2) between real Hilbert spaces H1H_1 and H2H_2, with adjoint B*∈ℒ(H2,H1)B^* \in \mathcal{L}(H_2, H_1), is the bounded linear operator defined by (2): gram⁡(B)=B(B*B)invB*,\operatorname{gram}(B) = B (B^* B)^{\mathrm{inv}} B^*,(2)✦ where (B*B)inv∈ℒ(H1)(B^* B)^{\mathrm{inv}} \in \mathcal{L}(H_1) is the inverse of B*BB^* B when B*BB^* B is invertible and 00 otherwise.

The Claim

Theorem 1. Let VV be a real normed space, let FF be a finite-dimensional real inner product space, let S⊆VS \subseteq V be a nonempty subset, and let A:V→ℒ(F)A : V \to \mathcal{L}(F) be a smooth (C∞C^\infty) map on VV such that A(u)A(u) is self-adjoint for every u∈Su \in S. Then there exist v∈Sv \in S and an open neighborhood U⊆VU \subseteq V of vv satisfying (3): rank⁡A(u)≤rank⁡A(v)(u∈S),rank⁡A(u)=rank⁡A(v)(u∈U∩S).\operatorname{rank} A(u) \le \operatorname{rank} A(v) \quad (u \in S), \qquad \operatorname{rank} A(u) = \operatorname{rank} A(v) \quad (u \in U \cap S).(3)✦ There also exist v∈Sv \in S, an open neighborhood U⊆VU \subseteq V of vv, and a map P:V→ℒ(F)P : V \to \mathcal{L}(F) that is C∞C^\infty on UU such that (4) holds: P(u)=proj⁡ker⁡A(u)(u∈U∩S).P(u) = \operatorname{proj}_{\ker A(u)} \quad (u \in U \cap S).(4)✦ Simultaneously, there exist v∈Sv \in S, an open neighborhood U⊆VU \subseteq V of vv, and a map P:V→ℒ(F)P : V \to \mathcal{L}(F) that is C∞C^\infty on UU satisfying (5): rank⁡A(u)≤rank⁡A(v)(u∈S),rank⁡A(u)=rank⁡A(v)andP(u)=proj⁡ker⁡A(u)(u∈U∩S).\begin{aligned} \operatorname{rank} A(u) &\le \operatorname{rank} A(v) \quad (u \in S), \\ \operatorname{rank} A(u) &= \operatorname{rank} A(v) \quad \text{and} \quad P(u) = \operatorname{proj}_{\ker A(u)} \quad (u \in U \cap S). \end{aligned}(5)✦

For any nonnegative integer cc and linear map E:S4→ScE : S^4 \to S^c, there exist v∈Ωv \in \Omega and an open neighborhood U⊆ℝ4U \subseteq \mathbb{R}^4 of vv satisfying (6): rank⁡E(I−uu⊤)≤rank⁡E(I−vv⊤)(u∈Ω),rank⁡E(I−uu⊤)=rank⁡E(I−vv⊤)(u∈U∩Ω).\begin{aligned} \operatorname{rank} E(I - u u^\top) &\le \operatorname{rank} E(I - v v^\top) \quad (u \in \Omega), \\ \operatorname{rank} E(I - u u^\top) &= \operatorname{rank} E(I - v v^\top) \quad (u \in U \cap \Omega). \end{aligned}(6)✦ For an integer c≥1c \ge 1 and a linear map E:S4→ScE : S^4 \to S^c, there exist v∈Ωv \in \Omega, an open neighborhood U⊆ℝ4U \subseteq \mathbb{R}^4 of vv, and a map G:ℝ4→ℝc×cG : \mathbb{R}^4 \to \mathbb{R}^{c \times c} that is C∞C^\infty on UU such that (7) holds: G(u)=R(u)(u∈U∩S3).G(u) = R(u) \quad (u \in U \cap S^3).(7)✦ For an integer c≥1c \ge 1 and a linear map E:S4→ScE : S^4 \to S^c, there exist v∈Ωv \in \Omega and a linear map T:ℝ4→ℝc×cT : \mathbb{R}^4 \to \mathbb{R}^{c \times c} satisfying (8) for every h∈v⟂h \in v^\perp: dds|s=0R(γv,h(s))=T(h).\left.\frac{d}{ds}\right|_{s=0} R(\gamma_{v,h}(s)) = T(h).(8)✦

In the paper, the result reads: on Ω={u∈S3:u0≠0}\Omega = \{u \in S^3 : u_0 \neq 0\}, the matrix family E(I−uu⊤)E(I - u u^\top) (the paper’s A(u)A(u)) attains its maximum rank ρ\rho at some v∈Ωv \in \Omega, and on a neighborhood of vv in Ω\Omega it has constant rank ρ\rho and R(u)=I−W(u)(W(u)⊤W(u))−1W(u)⊤R(u) = I - W(u)(W(u)^\top W(u))^{-1} W(u)^\top (or R(u)=IR(u) = I if ρ=0\rho = 0) is smooth.

What Is Checked

The claims are stated over a smooth map A:V→ℒ(F)A : V \to \mathcal{L}(F) from a real normed space VV to bounded linear operators on a finite-dimensional real inner product space FF that is self-adjoint on a nonempty subset S⊆VS \subseteq V, a bounded linear operator B∈ℒ(H1,H2)B \in \mathcal{L}(H_1, H_2) between real Hilbert spaces, and a linear map E:S4→ScE : S^4 \to S^c for a nonnegative integer cc (with c≥1c \ge 1 in (7) and (8)) applied to I−uu⊤I - u u^\top on Ω\Omega and S3S^3 together with the kernel projections R(u)R(u) and the spherical tangent curves γv,h\gamma_{v,h} along v⟂v^\perp.