OctoInformBeta

OpenAI preprint, September 2026 · §5. A matrix path with prescribed endpoints · §5

Proposition 5.1 [Connecting path]: Existence of bounded paths and uniform factors on [l,r][l, r] for 0<σ<10 < \sigma < 1, positive kk, continuous orthonormal uu, and ε∈(0,1]\varepsilon \in (0, 1]

Connecting opposite diagonals along a regularized matrix flow.

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

For a natural number NN and a bounded interval [l,r][l, r], take an orthonormal family of NN real functions on [l,r][l, r] and NN positive weights, as arise in the spectral theory of one-dimensional Schrödinger operators. Balancing a regularized matrix field on N×NN \times N real symmetric matrices against a rank-one outer product scaled by a lower triangular matrix yields a singularly perturbed matrix flow with parameter ε∈(0,1]\varepsilon \in (0, 1]. The question is how to join opposite diagonal matrices at ll and rr by a symmetric matrix path along this flow while keeping the path and the lower triangular factor controlled uniformly as ε\varepsilon becomes small.

Matrix Spaces and the Regularized Field

Given a natural number NN, the symmetric matrix space Sym⁡(N)\operatorname{Sym}(N) is the space of N×NN \times N real symmetric matrices.✦ For a natural number NN, a lower triangular matrix CC is an N×NN \times N real matrix such that Cij=0C_{ij} = 0 whenever i<ji < j.✦ For a natural number NN, the lower triangular matrix space ℒN\mathcal{L}_N is the space of all lower triangular N×NN \times N real matrices.✦

To each vector k∈ℝNk \in \mathbb{R}^N (for a natural number NN) corresponds the diagonal symmetric matrix diag⁡(k)∈Sym⁡(N)\operatorname{diag}(k) \in \operatorname{Sym}(N), the N×NN \times N diagonal matrix with ii-th diagonal entry kik_i.✦ Any vector a∈ℝNa \in \mathbb{R}^N (for a natural number NN) determines the symmetric outer product aa⊤∈Sym⁡(N)a a^\top \in \operatorname{Sym}(N), the N×NN \times N symmetric matrix with (i,j)(i, j) entry aiaja_i a_j.✦

For a natural number NN, let INI_N be the N×NN \times N identity matrix, for a vector k∈ℝNk \in \mathbb{R}^N let DkD_k be the N×NN \times N diagonal matrix with diagonal entries ki2k_i^2, and for an N×NN \times N real matrix BB and s∈ℝs \in \mathbb{R}, let the matrix pencil Xk,B(s)X_{k,B}(s) be the N×NN \times N real matrix Xk,B(s)=Dk−2sB+s2IN.X_{k,B}(s) = D_k - 2s B + s^2 I_N.(1)✦ For real numbers σ\sigma, δ\delta, and yy, the regularized scalar primitive ϕσ,δ(y)∈ℝ\phi_{\sigma,\delta}(y) \in \mathbb{R}, with t+=max⁡(t,0)t_+ = \max(t, 0), is the real number defined by ϕσ,δ(y)=∫0y(t++δ)σ−3/2dt.\phi_{\sigma,\delta}(y) = \int_0^y (t_+ + \delta)^{\sigma - 3/2}\,dt.(2)✦ For a real parameter σ\sigma, the field normalization constant cσ∈ℝc_\sigma \in \mathbb{R} is the real number cσ=σ∫ℝ(1+s2)σ−3/2ds.c_\sigma = \frac{\sigma}{\int_{\mathbb{R}} (1 + s^2)^{\sigma - 3/2}\,ds}.(3)✦

For a natural number NN, parameters σ,δ,Λ∈ℝ\sigma, \delta, \Lambda \in \mathbb{R}, a vector k∈ℝNk \in \mathbb{R}^N, and an N×NN \times N real matrix BB, with ϕσ,δ(Xk,B(s))\phi_{\sigma,\delta}(X_{k,B}(s)) evaluated by the continuous functional calculus using (1), (2), and (3), the truncated matrix field Mσ,δ,k(Λ)(B)M_{\sigma,\delta,k}^{(\Lambda)}(B) is the N×NN \times N real matrix Mσ,δ,k(Λ)(B)=cσ∫−ΛΛ(ϕσ,δ(Xk,B(s))−ϕσ,δ(s2)IN)ds.M_{\sigma,\delta,k}^{(\Lambda)}(B) = c_\sigma \int_{-\Lambda}^{\Lambda} \bigl(\phi_{\sigma,\delta}(X_{k,B}(s)) - \phi_{\sigma,\delta}(s^2) I_N\bigr)\,ds.(4)✦ In terms of the truncated matrix field (4), the matrix field Mσ,δ,k(B)M_{\sigma,\delta,k}(B) for σ,δ∈ℝ\sigma, \delta \in \mathbb{R}, a natural number NN, k∈ℝNk \in \mathbb{R}^N, and an N×NN \times N real matrix BB is the N×NN \times N real matrix Mσ,δ,k(B)=limΛ→+∞Mσ,δ,k(Λ)(B).M_{\sigma,\delta,k}(B) = \lim_{\Lambda \to +\infty} M_{\sigma,\delta,k}^{(\Lambda)}(B).(5)✦ For a natural number NN, parameters σ<1\sigma < 1 and δ>0\delta > 0, and a vector k∈ℝNk \in \mathbb{R}^N, the symmetric matrix field ℳσ,δ,k\mathcal{M}_{\sigma,\delta,k} is the map from Sym⁡(N)\operatorname{Sym}(N) to Sym⁡(N)\operatorname{Sym}(N) sending each B∈Sym⁡(N)B \in \operatorname{Sym}(N) to the matrix Mσ,δ,k(B)M_{\sigma,\delta,k}(B) from (5).✦

For any natural number NN, a vector c∈ℝNc \in \mathbb{R}^N, and an N×NN \times N real matrix ZZ, the linearized endpoint map inverse Tc(Z)T_c(Z) is the N×NN \times N real matrix (Tc(Z))ij={0if i<j,Zij/(2ci)if i=j,Zij/cjif i>j.\bigl(T_c(Z)\bigr)_{ij} = \begin{cases} 0 & \text{if } i < j, \\ Z_{ij}/(2c_i) & \text{if } i = j, \\ Z_{ij}/c_j & \text{if } i > j. \end{cases}(6)✦

The Claim

Theorem 1. Let NN be a natural number, let σ,δ∈ℝ\sigma, \delta \in \mathbb{R} satisfy 0<σ<10 < \sigma < 1 and δ>0\delta > 0, and let l,r∈ℝl, r \in \mathbb{R} satisfy l≤rl \le r. Let k∈ℝNk \in \mathbb{R}^N satisfy ki>0k_i > 0 for every ii, and let u:ℝ→ℝNu : \mathbb{R} \to \mathbb{R}^N be continuous on [l,r][l, r] such that, for all ii and jj, ∫lrui(t)uj(t)dt\int_l^r u_i(t) u_j(t)\,dt equals 11 when i=ji = j and 00 when i≠ji \neq j. Then for each ε∈(0,1]\varepsilon \in (0, 1], there exist a lower triangular matrix C∈ℒNC \in \mathcal{L}_N and a path B:ℝ→Sym⁡(N)B : \mathbb{R} \to \operatorname{Sym}(N), differentiable within [l,r][l, r] at every t∈[l,r]t \in [l, r] with derivative B′(t)B'(t), satisfying (7), where CC is bounded entrywise by a constant R∈ℝR \in \mathbb{R} independent of ε\varepsilon. |Cij|≤R(for all i,j),B(l)=diag⁡(k),B(r)=−diag⁡(k),∥B(t)∥≤∥k∥(t∈[l,r]),B′(t)=ε−1(ℳσ,δ,k(B(t))−(Cu(t))(Cu(t))⊤)(t∈[l,r]).\begin{aligned} |C_{ij}| &\le R \quad (\text{for all } i, j), & B(l) &= \operatorname{diag}(k), & B(r) &= -\operatorname{diag}(k), \\ \|B(t)\| &\le \|k\| \quad (t \in [l, r]), & B'(t) &= \varepsilon^{-1}\bigl(\mathcal{M}_{\sigma,\delta,k}(B(t)) - (Cu(t))(Cu(t))^\top\bigr) \quad (t \in [l, r]). \end{aligned}(7)✦

In the paper, the result reads: for a bounded interval I=(x−,x+)I = (x_-, x_+), a function u∈H01(I;ℝN)u \in H^1_0(I;\mathbb{R}^N) satisfying ∫Iu(x)u(x)⊤dx=IN\int_I u(x) u(x)^\top\,dx = I_N, and K=diag⁡(k1,…,kN)K = \operatorname{diag}(k_1, \dots, k_N) with every ki>0k_i > 0, assuming MM and μ\mu have the regularity and properties (5.1) through (5.3), for every ε>0\varepsilon > 0 there exist a real lower triangular matrix CC and a path B∈C1(I‾;Sym⁡(N))B \in C^1(\bar{I};\operatorname{Sym}(N)) such that, with a=Cua = Cu, εB′=M(B)−aa⊤\varepsilon B' = M(B) - a a^\top, B(x−)=KB(x_-) = K, B(x+)=−KB(x_+) = -K, and ∥B(x)∥op≤∥K∥op\|B(x)\|_{\mathrm{op}} \le \|K\|_{\mathrm{op}} for x∈Ix \in I, where for fixed MM, KK, and II the matrices CC can be chosen bounded uniformly for 0<ε≤10 < \varepsilon \le 1. The theorem above answers that statement for the interval [l,r][l, r] with l≤rl \le r, K=diag⁡(k)K = \operatorname{diag}(k), and a=Cua = Cu, replacing H01(I;ℝN)H^1_0(I;\mathbb{R}^N) by continuity of uu on [l,r][l, r] and the abstract pair M,μM, \mu by ℳσ,δ,k\mathcal{M}_{\sigma,\delta,k}, while restricting ε\varepsilon to (0,1](0, 1] and stating (7) via differentiability within [l,r][l, r], an entrywise bound on CC, and the matrix and vector norms in place of the operator norm.

What Is Checked

The connecting path claim (7) is stated over a natural number NN, real parameters 0<σ<10 < \sigma < 1, δ>0\delta > 0, and l≤rl \le r, a positive vector k∈ℝNk \in \mathbb{R}^N, and a continuous orthonormal function uu on [l,r][l, r], using the lower triangular matrix space ℒN\mathcal{L}_N, the symmetric matrix space Sym⁡(N)\operatorname{Sym}(N), the diagonal matrix diag⁡(k)\operatorname{diag}(k), the symmetric outer product aa⊤a a^\top, and the symmetric matrix field ℳσ,δ,k\mathcal{M}_{\sigma,\delta,k} built from (1), (2), (3), (4), and (5), alongside the linearized endpoint map inverse (6) defined on vectors c∈ℝNc \in \mathbb{R}^N and N×NN \times N real matrices ZZ.

Suppose l,r,κ∈ℝl, r, \kappa \in \mathbb{R} and f,f′:ℝ→ℝf, f' : \mathbb{R} \to \mathbb{R} are such that ff is continuous on [l,r][l, r], differentiable at every x∈[l,r]x \in [l, r] with derivative f′(x)f'(x), satisfies f(l)≤κf(l) \le \kappa, and obeys f′(x)≤0f'(x) \le 0 whenever x∈[l,r]x \in [l, r] and f(x)>κf(x) > \kappa; then the upper barrier bound (8) holds: f(x)≤κfor all x∈[l,r].f(x) \le \kappa \qquad \text{for all } x \in [l, r].(8)✦

If NN is a natural number, CC is an N×NN \times N lower triangular real matrix, and d∈ℝNd \in \mathbb{R}^N satisfies di>0d_i > 0 for every ii and CC⊤=diag⁡(d)C C^\top = \operatorname{diag}(d), then the Gram diagonalization (9) gives Cij=0(i≠j)andCii2=difor all i.C_{ij} = 0 \quad (i \neq j) \qquad \text{and} \qquad C_{ii}^2 = d_i \quad \text{for all } i.(9)✦