OctoInformBeta

OpenAI preprint, September 2026 · §6. From the matrix path to the action · §6

§6 (proof of Theorem 2.1): penalized action identity and compactness limit: Differential identities, compact penalty bounds for small ε>0\varepsilon > 0, and the right-hand limit δ↓0\delta \downarrow 0 for 0<σ<10 < \sigma < 1

Regularized matrix paths give penalized action and endpoint bounds.

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, a spectral exponent γ∈(1/2,3/2)\gamma \in (1/2, 3/2) with σ=γ−1/2∈(0,1)\sigma = \gamma - 1/2 \in (0, 1), and a bounded interval I=(l,r)I = (l, r), consider an orthonormal NN-tuple of functions (u1,…,uN)(u_1, \dots, u_N) in H01(I;ℝ)H_0^1(I; \mathbb{R}) and positive threshold parameters k1,…,kN>0k_1, \dots, k_N > 0. Testing the finite-interval action on H01(I;ℝN)H_0^1(I; \mathbb{R}^N) along lower triangular linear combinations of (u1,…,uN)(u_1, \dots, u_N) asks whether an upper bound on this action implies the same upper bound on the spectral sum ∑i=1N∫−kiki(ki2−s2)σds\sum_{i=1}^N \int_{-k_i}^{k_i} (k_i^2 - s^2)^\sigma\,ds.

The passage from the finite-interval action to the spectral sum runs through paths of N×NN \times N real symmetric matrices connecting the diagonal matrix with entries k1,…,kNk_1, \dots, k_N to its negative under a regularized matrix field with regularization parameter δ>0\delta > 0 and penalty parameter ε>0\varepsilon > 0. One tracks the field primitive along a penalized flow of symmetric matrices for small ε>0\varepsilon > 0 and then passes to the regularization limit δ↓0\delta \downarrow 0.

Matrix Field and Action

For a natural number NN, the symmetric matrix space Sym⁡(N)\operatorname{Sym}(N) is the space of N×NN \times N real symmetric matrices✦, and the lower triangular matrix space 𝒯N\mathcal{T}_N is the space of N×NN \times N real matrices CC with Cij=0C_{ij} = 0 whenever i<ji < j.✦

The diagonal symmetric matrix diag⁡(k)∈Sym⁡(N)\operatorname{diag}(k) \in \operatorname{Sym}(N) of a vector k∈ℝNk \in \mathbb{R}^N (N∈ℕN \in \mathbb{N}) is the N×NN \times N diagonal matrix with ii-th diagonal entry kik_i.✦ For N∈ℕN \in \mathbb{N} and a∈ℝNa \in \mathbb{R}^N, the symmetric outer product aa⊤∈Sym⁡(N)a a^\top \in \operatorname{Sym}(N) is the N×NN \times N real symmetric matrix with (i,j)(i, j)-entry aiaja_i a_j.✦

Given a real parameter σ\sigma, the field normalization constant cσ∈ℝc_\sigma \in \mathbb{R} is the real number defined by (1): cσ=σ∫−∞∞(1+s2)σ−3/2ds.c_\sigma = \frac{\sigma}{\int_{-\infty}^{\infty} (1 + s^2)^{\sigma - 3/2}\,ds}.(1)✦

With real parameters σ\sigma, δ\delta, and yy, and writing t+=max⁡(t,0)t_+ = \max(t, 0), the regularized scalar primitive fσ,δ(y)∈ℝf_{\sigma,\delta}(y) \in \mathbb{R} is the real number defined by (2): fσ,δ(y)=∫0y(t++δ)σ−3/2dt.f_{\sigma,\delta}(y) = \int_0^y (t_+ + \delta)^{\sigma - 3/2}\,dt.(2)✦

For N∈ℕN \in \mathbb{N}, k∈ℝNk \in \mathbb{R}^N, an N×NN \times N real matrix BB, and s∈ℝs \in \mathbb{R}, the matrix pencil X(s)∈ℝN×NX(s) \in \mathbb{R}^{N \times N} is the N×NN \times N real matrix defined by (3), where INI_N is the N×NN \times N identity matrix: X(s)=diag⁡(ki2)−2sB+s2IN.X(s) = \operatorname{diag}(k_i^2) - 2sB + s^2 I_N.(3)✦

At cutoff R∈ℝR \in \mathbb{R}, for N∈ℕN \in \mathbb{N}, σ,δ∈ℝ\sigma, \delta \in \mathbb{R}, k∈ℝNk \in \mathbb{R}^N, and an N×NN \times N real matrix BB, the truncated matrix field MR(B)∈ℝN×NM_R(B) \in \mathbb{R}^{N \times N} (suppressing σ\sigma, δ\delta, and kk) is the N×NN \times N real matrix given by (4), with fσ,δ(X(s))f_{\sigma,\delta}(X(s)) evaluated via the continuous functional calculus: MR(B)=cσ∫−RR(fσ,δ(X(s))−fσ,δ(s2)IN)ds.M_R(B) = c_\sigma \int_{-R}^R \bigl(f_{\sigma,\delta}(X(s)) - f_{\sigma,\delta}(s^2) I_N\bigr)\,ds.(4)✦

The regularized matrix field M(B)∈ℝN×NM(B) \in \mathbb{R}^{N \times N} of an N×NN \times N real matrix BB, for N∈ℕN \in \mathbb{N}, σ,δ∈ℝ\sigma, \delta \in \mathbb{R}, and k∈ℝNk \in \mathbb{R}^N, is the N×NN \times N real matrix given by the cutoff limit (5): M(B)=limR→+∞MR(B).M(B) = \lim_{R \to +\infty} M_R(B).(5) When σ<1\sigma < 1 and δ>0\delta > 0, the symmetric matrix field M:Sym⁡(N)→Sym⁡(N)M : \operatorname{Sym}(N) \to \operatorname{Sym}(N) (for N∈ℕN \in \mathbb{N} and k∈ℝNk \in \mathbb{R}^N) is the map sending each B∈Sym⁡(N)B \in \operatorname{Sym}(N) to M(B)∈Sym⁡(N)M(B) \in \operatorname{Sym}(N).✦

The matrix field primitive Jσ,δ,k(B)∈ℝJ_{\sigma,\delta,k}(B) \in \mathbb{R} (written J(B)J(B) when σ\sigma, δ\delta, and kk are fixed) of an N×NN \times N real matrix BB, for N∈ℕN \in \mathbb{N}, σ,δ∈ℝ\sigma, \delta \in \mathbb{R}, and k∈ℝNk \in \mathbb{R}^N, is the real number defined by (6): Jσ,δ,k(B)=−∫01tr⁡(BM(tB))dt.J_{\sigma,\delta,k}(B) = -\int_0^1 \operatorname{tr}\bigl(B M(tB)\bigr)\,dt.(6)✦

For vectors k,a,v∈ℝNk, a, v \in \mathbb{R}^N (N∈ℕN \in \mathbb{N}) and a real parameter σ\sigma, the action density ℒ(a,v)∈ℝ\mathcal{L}(a, v) \in \mathbb{R} (suppressing σ\sigma and kk) is the real number defined by (7): ℒ(a,v)=∑i=1Nvi2+∑i=1Nki2ai2−(∑i=1Nai2)1+1/σ.\mathcal{L}(a, v) = \sum_{i=1}^N v_i^2 + \sum_{i=1}^N k_i^2 a_i^2 - \left(\sum_{i=1}^N a_i^2\right)^{1 + 1/\sigma}.(7)✦ For a bounded interval I=(l,r)I = (l, r) and a positive diagonal matrix K=diag⁡(k)K = \operatorname{diag}(k), the finite-interval action ℰK(a)\mathcal{E}_K(a) of a path a∈H01(I;ℝN)a \in H_0^1(I; \mathbb{R}^N) is the real number ∫lrℒ(a(t),a′(t))dt\int_l^r \mathcal{L}(a(t), a'(t))\,dt.

On a pair (B,a)∈Sym⁡(N)×ℝN(B, a) \in \operatorname{Sym}(N) \times \mathbb{R}^N with N∈ℕN \in \mathbb{N}, σ∈ℝ\sigma \in \mathbb{R}, and k∈ℝNk \in \mathbb{R}^N, the action defect Δact(B,a)∈ℝ\Delta_{\mathrm{act}}(B, a) \in \mathbb{R} is the real number given by (8): Δact(B,a)=(∑i=1Nai2)1+1/σ−∑i=1Nai((diag⁡(ki2)−B2)a)i.\Delta_{\mathrm{act}}(B, a) = \left(\sum_{i=1}^N a_i^2\right)^{1 + 1/\sigma} - \sum_{i=1}^N a_i \bigl((\operatorname{diag}(k_i^2) - B^2) a\bigr)_i.(8)✦

The constraint defect Δcon(B,a)∈ℝ\Delta_{\mathrm{con}}(B, a) \in \mathbb{R} of a pair (B,a)∈Sym⁡(N)×ℝN(B, a) \in \operatorname{Sym}(N) \times \mathbb{R}^N, for N∈ℕN \in \mathbb{N}, σ,δ∈ℝ\sigma, \delta \in \mathbb{R}, and k∈ℝNk \in \mathbb{R}^N, is the real number defined by the squared Hilbert-Schmidt norm (9): Δcon(B,a)=∥M(B)−aa⊤∥HS2=∑i=1N∑j=1N(M(B)ij−aiaj)2.\Delta_{\mathrm{con}}(B, a) = \|M(B) - a a^\top\|_{\mathrm{HS}}^2 = \sum_{i=1}^N \sum_{j=1}^N \bigl(M(B)_{ij} - a_i a_j\bigr)^2.(9)✦

The Claim

Theorem 1. (1) For N∈ℕN \in \mathbb{N}, real numbers σ<1\sigma < 1, δ>0\delta > 0, ε>0\varepsilon > 0, and x∈ℝx \in \mathbb{R}, a vector k∈ℝNk \in \mathbb{R}^N, paths a,a′:ℝ→ℝNa, a' : \mathbb{R} \to \mathbb{R}^N and B:ℝ→Sym⁡(N)B : \mathbb{R} \to \operatorname{Sym}(N), and B′∈Sym⁡(N)B' \in \operatorname{Sym}(N) such that each component aia_i has derivative ai′(x)a'_i(x) at xx, BB has derivative B′B' at xx in Sym⁡(N)\operatorname{Sym}(N), and εB′=M(B(x))−a(x)a(x)⊤\varepsilon B' = M(B(x)) - a(x) a(x)^\top, the differential identity (10) holds at xx: ddt|t=x(J(B(t))+a(t)⊤B(t)a(t))=2a′(x)⊤B(x)a(x)−Δcon(B(x),a(x))ε.\left.\frac{d}{dt}\right|_{t=x} \bigl(J(B(t)) + a(t)^\top B(t) a(t)\bigr) = 2\,a'(x)^\top B(x) a(x) - \frac{\Delta_{\mathrm{con}}(B(x), a(x))}{\varepsilon}.(10)✦

(2) For N∈ℕN \in \mathbb{N}, real parameters σ,δ,ε∈ℝ\sigma, \delta, \varepsilon \in \mathbb{R}, vectors k,a,v∈ℝNk, a, v \in \mathbb{R}^N, and B∈Sym⁡(N)B \in \operatorname{Sym}(N), the pointwise bound (11) holds: 2∑i=1Nvi(Ba)i−Δcon(B,a)ε≤∑i=1Nvi2+∑i=1Nki2ai2−(∑i=1Nai2)1+1/σ+(Δact(B,a)−Δcon(B,a)ε).2 \sum_{i=1}^N v_i (B a)_i - \frac{\Delta_{\mathrm{con}}(B, a)}{\varepsilon} \le \sum_{i=1}^N v_i^2 + \sum_{i=1}^N k_i^2 a_i^2 - \left(\sum_{i=1}^N a_i^2\right)^{1 + 1/\sigma} + \left(\Delta_{\mathrm{act}}(B, a) - \frac{\Delta_{\mathrm{con}}(B, a)}{\varepsilon}\right).(11)✦

(3) If EE is a topological space, 𝒦⊆E\mathcal{K} \subseteq E is compact, f,g:E→ℝf, g : E \to \mathbb{R} are continuous functions with g(z)≥0g(z) \ge 0 for all z∈𝒦z \in \mathcal{K} and f(z)≤0f(z) \le 0 whenever z∈𝒦z \in \mathcal{K} satisfies g(z)=0g(z) = 0, and η>0\eta > 0, then there exists ε0>0\varepsilon_0 > 0 such that every ε∈(0,ε0]\varepsilon \in (0, \varepsilon_0] and z∈𝒦z \in \mathcal{K} satisfy (12): f(z)−g(z)ε≤η.f(z) - \frac{g(z)}{\varepsilon} \le \eta.(12)✦

(4) Given N∈ℕN \in \mathbb{N}, 0<σ<10 < \sigma < 1, δ>0\delta > 0, k∈ℝNk \in \mathbb{R}^N, a compact set 𝒦⊆Sym⁡(N)×ℝN\mathcal{K} \subseteq \operatorname{Sym}(N) \times \mathbb{R}^N, and η>0\eta > 0, there exists ε0>0\varepsilon_0 > 0 such that every ε∈(0,ε0]\varepsilon \in (0, \varepsilon_0] and (B,a)∈𝒦(B, a) \in \mathcal{K} satisfy (13): Δact(B,a)−Δcon(B,a)ε≤η.\Delta_{\mathrm{act}}(B, a) - \frac{\Delta_{\mathrm{con}}(B, a)}{\varepsilon} \le \eta.(13)✦

(5) Suppose N∈ℕN \in \mathbb{N}, 0<σ<10 < \sigma < 1, δ>0\delta > 0, ε>0\varepsilon > 0, l≤rl \le r, η∈ℝ\eta \in \mathbb{R}, and k∈ℝNk \in \mathbb{R}^N. Let a,v:ℝ→ℝNa, v : \mathbb{R} \to \mathbb{R}^N be continuous on [l,r][l, r] with a(l)=a(r)=0a(l) = a(r) = 0 such that each component aia_i has derivative vi(t)v_i(t) at every t∈(l,r)t \in (l, r), and let B:ℝ→Sym⁡(N)B : \mathbb{R} \to \operatorname{Sym}(N) satisfy B(l)=diag⁡(k)B(l) = \operatorname{diag}(k) and B(r)=−diag⁡(k)B(r) = -\operatorname{diag}(k), with derivative within [l,r][l, r] in Sym⁡(N)\operatorname{Sym}(N) equal to ε−1(M(B(t))−a(t)a(t)⊤)\varepsilon^{-1}\bigl(M(B(t)) - a(t) a(t)^\top\bigr) at every t∈[l,r]t \in [l, r]. If Δact(B(t),a(t))−ε−1Δcon(B(t),a(t))≤η\Delta_{\mathrm{act}}(B(t), a(t)) - \varepsilon^{-1}\Delta_{\mathrm{con}}(B(t), a(t)) \le \eta for all t∈[l,r]t \in [l, r], then (14) holds: J(−diag⁡(k))−J(diag⁡(k))≤∫lrℒ(a(t),v(t))dt+(r−l)η.J(-\operatorname{diag}(k)) - J(\operatorname{diag}(k)) \le \int_l^r \mathcal{L}(a(t), v(t))\,dt + (r - l)\eta.(14)✦

(6) Given N∈ℕN \in \mathbb{N}, 0<σ<10 < \sigma < 1, δ>0\delta > 0, l≤rl \le r, S∈ℝS \in \mathbb{R}, and k∈ℝNk \in \mathbb{R}^N with ki>0k_i > 0 for all ii, suppose u,v:ℝ→ℝNu, v : \mathbb{R} \to \mathbb{R}^N are continuous on [l,r][l, r] with u(l)=u(r)=0u(l) = u(r) = 0, each component uiu_i has derivative vi(t)v_i(t) at every t∈(l,r)t \in (l, r), and ∫lrui(t)uj(t)dt\int_l^r u_i(t) u_j(t)\,dt equals 11 when i=ji = j and 00 otherwise. If ∫lrℒ(Cu(t),Cv(t))dt≤S\int_l^r \mathcal{L}(C u(t), C v(t))\,dt \le S for every C∈𝒯NC \in \mathcal{T}_N, then (15) holds: J(−diag⁡(k))−J(diag⁡(k))≤S.J(-\operatorname{diag}(k)) - J(\operatorname{diag}(k)) \le S.(15)✦

(7) Given N∈ℕN \in \mathbb{N}, 0<σ<10 < \sigma < 1, l≤rl \le r, S∈ℝS \in \mathbb{R}, and k∈ℝNk \in \mathbb{R}^N with ki>0k_i > 0 for all ii, suppose u,v:ℝ→ℝNu, v : \mathbb{R} \to \mathbb{R}^N are continuous on [l,r][l, r] with u(l)=u(r)=0u(l) = u(r) = 0, each component uiu_i has derivative vi(t)v_i(t) at every t∈(l,r)t \in (l, r), and ∫lrui(t)uj(t)dt\int_l^r u_i(t) u_j(t)\,dt equals 11 when i=ji = j and 00 otherwise. If ∫lrℒ(Cu(t),Cv(t))dt≤S\int_l^r \mathcal{L}(C u(t), C v(t))\,dt \le S for every C∈𝒯NC \in \mathcal{T}_N, then (16) holds: ∑i=1N∫−kiki(ki2−s2)σds≤S.\sum_{i=1}^N \int_{-k_i}^{k_i} (k_i^2 - s^2)^\sigma\,ds \le S.(16)✦

(8) For N∈ℕN \in \mathbb{N}, 0<σ<10 < \sigma < 1, and k∈ℝNk \in \mathbb{R}^N with ki≥0k_i \ge 0 for all ii, the endpoint jump Jσ,δ,k(−diag⁡(k))−Jσ,δ,k(diag⁡(k))J_{\sigma,\delta,k}(-\operatorname{diag}(k)) - J_{\sigma,\delta,k}(\operatorname{diag}(k)) satisfies the right-hand limit (17): limδ↓0(Jσ,δ,k(−diag⁡(k))−Jσ,δ,k(diag⁡(k)))=∑i=1N∫−kiki(ki2−s2)σds.\lim_{\delta \downarrow 0} \bigl(J_{\sigma,\delta,k}(-\operatorname{diag}(k)) - J_{\sigma,\delta,k}(\operatorname{diag}(k))\bigr) = \sum_{i=1}^N \int_{-k_i}^{k_i} (k_i^2 - s^2)^\sigma\,ds.(17)✦

In the paper, the result reads: along the path BB, (J(B)+a⊤Ba)′=2(a′)⊤Ba−ε−1∥M(B)−aa⊤∥HS2(J(B) + a^\top B a)' = 2(a')^\top B a - \varepsilon^{-1}\|M(B) - a a^\top\|_{\mathrm{HS}}^2, yielding J(−K)−J(K)≤ℰK(a)+∫IFε(B,a)dxJ(-K) - J(K) \le \mathcal{E}_K(a) + \int_I F_\varepsilon(B, a)\,dx with K=diag⁡(k)K = \operatorname{diag}(k), I=(l,r)I = (l, r), and Fε(B,a)=Δact(B,a)−ε−1Δcon(B,a)F_\varepsilon(B, a) = \Delta_{\mathrm{act}}(B, a) - \varepsilon^{-1}\Delta_{\mathrm{con}}(B, a), while compactness of the trajectory set 𝒦\mathcal{K} ensures max⁡(B,a)∈𝒦max⁡{Fε(B,a),0}→0\max_{(B, a) \in \mathcal{K}} \max\{F_\varepsilon(B, a), 0\} \to 0 as ε→0\varepsilon \to 0 and dominated convergence as δ→0\delta \to 0 yields the action inequality.

What Is Checked

The claims are stated over natural numbers NN, real parameters σ\sigma, δ\delta, ε\varepsilon, η\eta, l≤rl \le r, SS, and k∈ℝNk \in \mathbb{R}^N, the spaces Sym⁡(N)\operatorname{Sym}(N) and 𝒯N\mathcal{T}_N, the constructions diag⁡(k)\operatorname{diag}(k) and aa⊤a a^\top, the matrix fields MM and MRM_R built from cσc_\sigma, fσ,δf_{\sigma,\delta}, and the pencil X(s)X(s), the primitive Jσ,δ,kJ_{\sigma,\delta,k}, the action density ℒ\mathcal{L} together with Δact\Delta_{\mathrm{act}} and Δcon\Delta_{\mathrm{con}}, paths BB, aa, uu, and vv on [l,r][l, r] satisfying boundary and differential hypotheses (and orthonormality for uu), and continuous real functions ff and gg on a compact set 𝒦\mathcal{K} in a topological space EE or in Sym⁡(N)×ℝN\operatorname{Sym}(N) \times \mathbb{R}^N.