betaFeedback

WorksMathematics

Cubic local observability of Gaussian-blind sources in thermal ϕ24ϕ^4_2 fields

Abstract

We study external sources that are invisible to a spatially local observation in a massive Gaussian thermal field on a circle. A quadratic-reference calibration removes the first-order thermal mass-shift signal. For the resulting quartic model we prove a cutoff-independent, source-uniform C1C^1 expansion of local Wick cubic means. Its leading operator has the time-integrated cube of the thermal covariance as its kernel. A positive bilateral Laplace representation proves injectivity on the entire external source space, while analyticity gives exponential singular-value upper bounds. The expansion yields local Araki relative-entropy lower bounds on every fixed finite-dimensional source space, and a uniform separation theorem at every prescribed Sobolev resolution. Under an additional interval-length condition, a reflection-positivity identity yields an explicit, nonoptimal double-logarithmic conditional stability bound for this specific thermal kernel. Continuum probability estimates and the integrability assumptions of the Feynman–Kac–Nelson construction are verified in Section~3.

Introduction

A static source changes both a thermal state and the information available to a local observer. In a free massive field the source Bogoliubov–Kubo–Mori (BKM) metric is the inverse massive Laplacian. This global identity does not imply local identifiability: an exterior source can translate the field only outside the observation region. We study the local information carried by a calibrated quartic response and establish separate linearized and finite-coupling stability statements.

Fix a spatial circle TL\mathbb{T}_L, an inverse temperature β>0\beta> 0, and a mass m>0m > 0. Let II and OO be nonempty open intervals with disjoint closures. After choosing an origin and a lift to R\mathbb{R}, write

I=(0,ℓ),O‾⊂(ℓ−L,0),0<ℓ<L.(1)I = (0,\ell), \qquad\overline{O} \subset(\ell-L,0), \qquad0 < \ell< L. \tag*{(1)}

For A=−∂x2+m2A = -\partial_x^2 + m^2, every v∈C∞(I)v \in C^\infty(I) generates a Gaussian-blind source AvAv: the corresponding free mean is vv, and its restriction to OO vanishes. In fact, the entire equal-time local Gaussian state is unchanged, not only its mean.

A necessary qualification concerns thermal mass calibration. Vacuum and thermal Wick ordering differ by a quadratic term. The resulting first-order mean signal can already be produced by a shifted Gaussian mass. We therefore retain an explicit quadratic coefficient b≥0b \ge0 and use the calibrated source

Jλ,v=(A+2bλ)v.(2)J_{\lambda,v} = (A + 2b\lambda)v. \tag*{(2)}

The Gaussian comparison model with quadratic operator A+2bλA + 2b\lambda is then blind to Jλ,vJ_{\lambda,v} for every λ≥0\lambda\ge0. Thermal Wick ordering corresponds to b=0b = 0; the original vacuum-ordered quartic interaction corresponds, at this fixed temperature and up to a scalar, to b=6dβb = 6d_\beta, where dβd_\beta is defined in (2.6). This is an exact calibration of the specified quadratic reference, not an identification of the full interacting pole mass.

Let μ\mu be the centered Gaussian measure on Tβ×TL\mathbb{T}_{\beta} \times\mathbb{T}_{L} with covariance C=(−∂τ2+A)−1C = (-\partial_{\tau}^{2} + A)^{-1}. Thermal Wick powers are denoted by colons. The local observable

Wh=∫Oh(y):ϕ(0,y)3: dy,h∈L2(O),W_h = \int_{O} h(y):\phi(0,y)^{3}: \,dy,\qquad h \in L^{2}(O),

is a spatially smeared Wick cubic field, not the cube of a smeared linear field. Its mean in the calibrated source state defines a map Mλ:H01(I)→L2(O)\mathcal{M}_{\lambda}: H_{0}^{1}(I) \to L^{2}(O) by duality. All quantum-state assertions below are made for smooth sources; the Euclidean observation map extends to all of H01(I)H_{0}^{1}(I).

Principal conclusions

The main statements separate the continuum cubic response, the properties of its linearized kernel, and the resulting local information bounds.

Theorem A (Uniform continuum cubic response; see Theorem 4.1). Let H=H01(I)H = H_{0}^{1}(I) with ∥v∥H2=∫I(∣v′∣2+m2v2) dx\lVert v\rVert_{H}^{2} = \int_{I} (|v'|^{2} + m^{2}v^{2})\,dx and Y=L2(O)Y = L^{2}(O). For each R<∞R < \infty and fixed λ∗<∞\lambda_{*} < \infty, the calibrated observation map satisfies

Mλ(v)=−24λTv+Rλ(v),sup⁡∥v∥H≤R∥DRλ(v)∥H→Y≤CRλ2,0≤λ≤λ∗,(3)\mathcal{M}_{\lambda}(v) = -24\lambda\mathcal{T}v + R_{\lambda}(v),\qquad\sup_{\lVert v\rVert_{H}\leq R}\lVert DR_{\lambda}(v)\rVert_{H\to Y} \leq C_{R}\lambda^{2},\qquad0\leq\lambda\leq\lambda_{*}, \tag*{(3)}

where

(Tv)(y)=∫IKβ(x−y)v(x) dx,Kβ(z)=∫0βC(τ,z)3 dτ.(4)(\mathcal{T}v)(y) = \int_{I} K_{\beta}(x-y)v(x)\,dx,\qquad K_{\beta}(z) = \int_{0}^{\beta} C(\tau,z)^{3}\,d\tau. \tag*{(4)}

The coefficient 24=4⋅3!24 = 4\cdot3! is independent of the quadratic calibration parameter bb.

Theorem B (Injectivity and analytic smoothing; see Theorem 5.2 and Proposition 5.4). The operator T:L2(I)→L2(O)\mathcal{T}:L^{2}(I)\to L^{2}(O) is injective, compact, and of infinite rank. Its singular values satisfy sn(T)≤Ce−cns_{n}(\mathcal{T})\leq Ce^{-cn} for some C,c>0C,c>0. The injectivity proof uses positive Laplace atoms and a half-plane zero-set argument, while analyticity yields the singular-value bound.

Theorem C (Local information separation; see Theorems 6.2 and 6.3). For every fixed finite-dimensional smooth source space E⊂HE\subset H and fixed source radius RR, sufficiently small positive coupling gives

SO(ωλ,v∥ωλ,w)≥cE,Rλ2∥v−w∥H2(v,w∈E, ∥v∥H,∥w∥H≤R),S_{O}(\omega_{\lambda,v}\Vert\omega_{\lambda,w}) \geq c_{E,R}\lambda^{2}\lVert v-w\rVert_{H}^{2}\qquad(v,w\in E,\ \lVert v\rVert_{H},\lVert w\rVert_{H}\leq R),

where cE,R>0c_{E,R}>0 depends on EE, RR, and the fixed model parameters. For a compact infinite-dimensional Dirichlet spectral prior and each prescribed resolution r>0r>0, one sufficiently small positive coupling separates all source pairs at HH-distance at least rr. The admissible coupling λr\lambda_{r} depends on rr; for smooth pairs the separation also holds for local Araki relative entropy.

Theorem D (Conditional double-logarithmic stability; see Theorem 7.3). If ℓ<L/2\ell<L/2, then for s>0s>0 and u∈Xsu\in X_{s} with ∥u∥Xs≤M\lVert u\rVert_{X_{s}}\leq M,

∥u∥H≤CsML(M/∥Tu∥Y)s/(1+s),L(x)=1+log⁡(1+log⁡(1+x)).(5)\lVert u\rVert_{H}\leq\frac{C_{s}M}{\mathcal{L}(M/\lVert\mathcal{T}u\rVert_{Y})^{s/(1+s)}},\qquad\mathcal{L}(x)=1+\log(1+\log(1+x)). \tag*{(5)}

The rate follows from a reflected positive energy and a quantitative bound for the associated noninteger moment system. The space XsX_{s} is the Dirichlet spectral prior defined in § 6.

Relation to previous work

The BKM metric and canonical correlations are classical [20, 27]. Finite-temperature free-energy convexity also enters Hamiltonian learning [1]. Our setting concerns the information available in a local algebra from an exterior continuum source. The constructive P(ϕ)2P(\phi)_{2} framework and the relation between periodic Euclidean processes and KMS states are developed in [18, 30, 17, 19, 15, 16]. We use the positive-temperature perturbation theorem [15], Theorem 7.11 and Propositions 7.8–7.10 after verifying its sharp-time hypotheses.

Local relative entropy for free-field coherent excitations was studied in [23, 7, 5]. For interacting theories, Drago, Faldino and Pinamonti [9] analyze perturbative equilibrium-state relative entropy, while Fröb, Much and Papadopoulos [12] compute a first-order quartic-coupling contribution for vacuum and coherent states in a Rindler wedge. In contrast, we obtain lower bounds for constructive finite-temperature states from a measured local Wick cubic observable and data processing [3], without computing an interacting modular Hamiltonian.

Nonlinear inverse-source methods have precedents in hyperbolic and elliptic problems [21, 11, 22]. The inverse scattering analysis of Sasaki and Suzuki [29] concerns a linearly sourced Klein–Gordon field. The model-specific results here are injectivity of the calibrated thermal cubic kernel and local information separation from its continuum response.

Propagation of smallness for analytic functions and heat observability are studied in [32, 2]; we use the spectral heat kernel on the circle [8]. Analytic-kernel approximation and classical Müntz theory supply ingredients for the singular-value and moment estimates [31, 4]. Related quantitative moment and inverse-Laplace methods appear in [28, 13, 14, 6], and inverse-problem instability is illustrated in [25, 26]. § 7 identifies the kernel-specific reflected energy and the source of the double-logarithmic loss. The final regularization step uses standard Hilbert-space arguments [10].

Gaussian blindness and the calibrated interacting model

The two-moment obstruction

Proposition 2.1 (Exact Gaussian blindness). On the smooth source space Cc∞(I)C^\infty_c(I), the restriction of the free thermal state to the common equal-time local algebra on OO depends exactly on the two moments in (2.2). The kernel of this restriction, relative to the zero source, is

{Av:v∈Cc∞(I)}.\{Av : v \in C^\infty_c(I)\}.

In particular this kernel has codimension two in Cc∞(I)C^\infty_c(I).

Proof. Completing the square gives a Weyl displacement of the centered free thermal state by the mean A−1JA^{-1}J. The field covariance is source independent and the momentum mean is zero. Hence local Weyl expectations depend only on the restriction of A−1JA^{-1}J to OO. The two exponentials in (2.2) are linearly independent on every nonempty interval, and the two source moment functionals are independent.

If both moments vanish, the mean is zero on OO. It satisfies Au=0Au=0 on the component of TL∖supp⁡J\mathbb{T}_L\setminus\operatorname{supp}J containing TL∖I‾\mathbb{T}_L\setminus\overline{I}. ODE uniqueness propagates the zero solution through this component and into neighborhoods of both endpoints of II. Elliptic regularity gives u∈Cc∞(I)u\in C^\infty_c(I) and J=AuJ=Au. The converse follows either from (2.2) by integration by parts or from A−1Av=vA^{-1}Av=v. A displacement supported in II commutes with every field and momentum Weyl operator supported in OO, proving equality on the full common time-zero local algebra. □

The proposition holds for any positive mass and the connected separated interval geometry in (1).

For completeness, in physical source coordinates the free global identities are

g0(f,h)=β⟨f,A−1h⟩,D(ρ0,J∥ρ0,K)=β2∥J−K∥Hm−12.(6)g_0(f,h)=\beta\langle f,A^{-1}h\rangle,\qquad D(\rho_{0,J}\|\rho_{0,K})=\frac{\beta}{2}\|J-K\|_{H_m^{-1}}^2. \tag*{(6)}

They follow from log⁡Z0(J)−log⁡Z0(0)=β⟨J,A−1J⟩/2\log Z_0(J)-\log Z_0(0)=\beta\langle J,A^{-1}J\rangle/2.

Thermal covariance and Wick conventions

Put kn=2πn/Lk_n=2\pi n/L, ωn=(kn2+m2)1/2\omega_n=(k_n^2+m^2)^{1/2} and νj=2πj/β\nu_j=2\pi j/\beta. The covariance is

C(τ,x)=1βL∑j,n∈Zeiνjτ+iknxνj2+ωn2.(7)C(\tau,x)=\frac{1}{\beta L}\sum_{j,n\in\mathbb{Z}}\frac{e^{i\nu_j\tau+ik_nx}}{\nu_j^2+\omega_n^2}. \tag*{(7)}

Its time-zero kernel is

Cβ(x)=1L∑n∈Z12ωncoth⁡βωn2eiknx.(8)C_\beta(x)=\frac{1}{L}\sum_{n\in\mathbb{Z}}\frac{1}{2\omega_n}\coth\frac{\beta\omega_n}{2}e^{\mathrm{i}k_nx}. \tag*{(8)}

Although Cβ(0)C_\beta(0) diverges, Cβ∈Lp(TL)C_\beta\in L^p(\mathbb{T}_L) for every finite pp. Let PNP_N truncate only spatial Fourier modes, ϕN=PNϕ\phi_N=P_N\phi, and cN=EϕN(0,x)2c_N=\mathbb{E}\phi_N(0,x)^2. Thermal Wick powers use cNc_N at this cutoff and are passed to the limit as distributions. If c0,N=L−1∑∣n∣≤N(2ωn)−1c_{0,N}=L^{-1}\sum_{|n|\leq N}(2\omega_n)^{-1}, then

dβ=lim⁡N→∞(cN−c0,N)=1L∑n∈Z1ωn(eβωn−1)>0,(9)d_\beta=\lim_{N\to\infty}(c_N-c_{0,N})=\frac{1}{L}\sum_{n\in\mathbb{Z}}\frac{1}{\omega_n(e^{\beta\omega_n}-1)}>0, \tag*{(9)}

and

:ϕ4:0=:ϕ4:β+6dβ:ϕ2:β+3dβ2.(10):\phi^4:_0=:\phi^4:_\beta+6d_\beta:\phi^2:_\beta+3d_\beta^2. \tag*{(10)}

Thermal normal ordering defines the model at the fixed inverse temperature β\beta.

Source calibration and the exact shifted measure

Write J=Tβ×TL\mathcal{J}=\mathbb{T}_\beta\times\mathbb{T}_L and set

V4=∫J:ϕ4:β,V2=∫J:ϕ2:β,YJ=∫0βϕ(τ,J) dτ.V_4=\int_{\mathcal{J}}:\phi^4:_\beta,\qquad V_2=\int_{\mathcal{J}}:\phi^2:_\beta,\qquad Y_J=\int_0^\beta\phi(\tau,J)\,d\tau.

The physical Euclidean source measure is proportional to

exp⁡{−λ(V4+bV2)+YJλ,v} dμ,b≥0.(11)\exp\{-\lambda(V_4+bV_2)+Y_{J_{\lambda,v}}\}\,d\mu,\qquad b\geq0. \tag*{(11)}

The associated formal Hamiltonian is dΓ(A)+λ∫(:Φ4:β+b:Φ2:β)−Φ(Jλ,v)d\Gamma(\sqrt{A})+\lambda\int(:\Phi^4:_\beta+b:\Phi^2:_\beta)-\Phi(J_{\lambda,v}); its rigorous KMS interpretation is given in Section 3.

Extend v∈H=H01(I)v\in H=H^1_0(I) by zero to the circle. The static Cameron–Martin shift by vv has squared norm β∥v∥H2\beta\|v\|_H^2. For every local time-zero functional FF, F(ϕ+v)=F(ϕ)F(\phi+v)=F(\phi). Completing the quadratic tilt in (2.8) therefore gives the exact identity

Eλ,vF=E[Fe−λU(v)]E[e−λU(v)],U(v)=∫J:(ϕ+v)4:β+bV2.(12)\mathbb{E}_{\lambda,v}F=\frac{\mathbb{E}[Fe^{-\lambda U(v)}]}{\mathbb{E}[e^{-\lambda U(v)}]},\qquad U(v)=\int_{\mathcal{J}}:(\phi+v)^4:_\beta+bV_2. \tag*{(12)}

The terms 2bλ∫vϕ2b\lambda\int v\phi cancel after shifting. The remaining scalar factor is

exp⁡{β2∥v∥H2+bλβ∥v∥22},\exp\left\{\frac{\beta}{2}\|v\|_H^2+b\lambda\beta\|v\|_2^2\right\},

and also cancels in the quotient. We use ν^bλ,v\widehat{\nu}_{b\lambda,v} for the probability measure on the right of (2.9) and Pλ,v0P^0_{\lambda,v} for its local time-zero marginal.

The Gaussian model obtained by deleting λV4\lambda V_4 has quadratic operator A+2bλA+2b\lambda, mean vv, and source-independent covariance. Its local state is thus exactly independent of vv. The calibration removes the following possible false positive: with the uncalibrated source AvAv, the local mean at infinitesimal source amplitude has the term

−2bλ(A−1v)∣O+O(λ2).-2b\lambda(A^{-1}v)|_O+O(\lambda^2).

For vacuum ordering b=6dβb=6d_\beta, this is precisely the thermal mass-shift term. This motivates the calibrated source used throughout the interacting model.

The physical source distance and the displacement distance are uniformly comparable:

∥v−w∥H≤∥Jλ,v−Jλ,w∥H−1≤(1+2bλm2)∥v−w∥H.(13)\|v-w\|_H\leq\|J_{\lambda,v}-J_{\lambda,w}\|_{H^{-1}}\leq\left(1+\frac{2b\lambda}{m^2}\right)\|v-w\|_H. \tag*{(13)}

This follows by applying the Fourier multiplier I+2bλA−1I+2b\lambda A^{-1} to A1/2(v−w)A^{1/2}(v-w).

Continuum estimates and the thermal-state interface

We prove the estimates needed below instead of assuming source-uniform differentiability. The only reconstruction input is the standard positive-temperature FKN theorem, stated at the point where it is used.

Wick convergence and Nelson bounds

Define

UN(v)=∫G(: ⁣(ϕN+PNv)4 ⁣: ⁣cN+b: ⁣ϕN2 ⁣: ⁣cN).U_N(v)=\int_{\mathcal{G}}\left(:\!(\phi_N+P_Nv)^4\!:\!_{c_N}+b:\!\phi_N^2\!:\!_{c_N}\right).

The full Gaussian path space is retained. Only the interaction is truncated; the limiting local algebra is never identified with a spatial Fourier tensor factor.

Lemma 3.1 (Uniform polynomial and exponential bounds). For every R<∞R<\infty, 1≤p<∞1\le p<\infty and a∗<∞a_\ast<\infty,

sup⁡N, ∥v∥H≤R∥DjUN(v)[u1,…,uj]∥Lp(μ)≤Cp,R,j∏i=1j∥ui∥H,0≤j≤4,(14)\sup_{\substack{N,\ \lVert v\rVert_H\le R}}\left\lVert D^jU_N(v)[u_1,\ldots,u_j]\right\rVert_{L^p(\mu)} \le C_{p,R,j}\prod_{i=1}^{j}\lVert u_i\rVert_H,\qquad0\le j\le4, \tag*{(14)}
sup⁡N, ∥v∥H≤R, 0≤a≤a∗Ee−aUN(v)<∞.(15)\sup_{\substack{N,\ \lVert v\rVert_H\le R,\ 0\le a\le a_\ast}}\mathbb{E}e^{-aU_N(v)}<\infty. \tag*{(15)}

The polynomial quantities in (3.1) converge in every finite LpL^p, uniformly on the indicated source and direction balls. Their limits define U(v)U(v) and its derivatives. The normalized densities

r^λ,v=e−λU(v)Zλ(v),Zλ(v)=Ee−λU(v),(16)\hat r_{\lambda,v}=\frac{e^{-\lambda U(v)}}{Z_\lambda(v)},\qquad Z_\lambda(v)=\mathbb{E}e^{-\lambda U(v)}, \tag*{(16)}

satisfy sup⁡0≤λ≤λ∗, ∥v∥H≤R∥r^λ,v∥2=KR<∞\sup_{0\le\lambda\le\lambda_\ast,\ \lVert v\rVert_H\le R}\lVert\hat r_{\lambda,v}\rVert_2=K_R<\infty for every fixed λ∗<∞\lambda_\ast<\infty.

Proof. We first establish (3.2) at v=0v=0. Let CN=PNCC_N=P_NC. The Fourier coefficients in (2.4) give

∥CM−CN∥L2(G)≤CN−1,sup⁡N∥CN∥L6(G)<∞(M>N).\lVert C_M-C_N\rVert_{L^2(\mathcal G)}\le CN^{-1},\qquad\sup_N\lVert C_N\rVert_{L^6(\mathcal G)}<\infty\quad(M>N).

For the first estimate, sum (νj2+ωn2)−2(\nu_j^2+\omega_n^2)^{-2} over jj and ∣n∣>N|n|>N. The second follows from the two-dimensional Sobolev embedding, since C∈HaC\in H^a for every a<1a<1. Wick’s rule and nested projections imply

∥∫G(: ⁣ϕM4 ⁣:−: ⁣ϕN4 ⁣:)∥22=4! βL∫G(CM4−CN4)≤C∥CM−CN∥2sup⁡K∥CK∥63≤CN−1.\left\lVert\int_{\mathcal G}\left(:\!\phi_M^4\!:-:\!\phi_N^4\!:\right)\right\rVert_2^2 =4!\,\beta L\int_{\mathcal G}(C_M^4-C_N^4) \le C\lVert C_M-C_N\rVert_2\sup_K\lVert C_K\rVert_6^3\le CN^{-1}.

The same argument for the quadratic term is simpler. Gaussian hypercontractivity on each chaos consequently yields

∥UM(0)−UN(0)∥p≤Cb(p−1)2N−1/2,p≥2.(17)\lVert U_M(0)-U_N(0)\rVert_p\le C_b(p-1)^2N^{-1/2},\qquad p\ge2. \tag*{(17)}

Since cN≤Clog⁡(eN)c_N\le C\log(eN), the polynomial

z4+(b−6cN)z2+3cN2−bcNz^4+(b-6c_N)z^2+3c_N^2-bc_N

has a deterministic lower bound −Cb(1+cN)2-C_b(1+c_N)^2. Hence UN(0)≥−Cblog⁡2(eN)U_N(0)\ge-C_b\log^2(eN). For a sufficiently large tt, choose n(t)=⌊exp⁡(ϵt)⌋n(t)=\lfloor\exp(\epsilon\sqrt{t})\rfloor with ϵ>0\epsilon>0 small enough that Un(0)≥−t/2U_n(0)\ge-t/2 for n≤n(t)n\le n(t). If N>n(t)N>n(t), Markov’s inequality and (3.4) give

Pr⁡(UN(0)<−t)≤(2Cb(p−1)2n(t)−1/2t)p.\Pr(U_N(0)<-t)\le\left(\frac{2C_b(p-1)^2n(t)^{-1/2}}{t}\right)^p.

Taking an integer pp proportional to t n(t)1/4\sqrt{t}\,n(t)^{1/4}, with a sufficiently small proportionality constant, proves

sup⁡NPr⁡(UN(0)<−t)≤exp⁡{−ct ec′t}.\sup_N\Pr(U_N(0)<-t)\le\exp\{-c\sqrt{t}\,e^{c'\sqrt{t}}\}.

This proves every negative exponential moment uniformly in NN. Fatou’s lemma includes the limit, and

∣e−ax−e−ay∣≤a∣x−y∣(e−ax+e−ay)\left|e^{-ax}-e^{-ay}\right|\le a|x-y|\left(e^{-ax}+e^{-ay}\right)

with Hölder’s inequality proves convergence of weights in every finite LpL^p, uniformly for bounded a≥0a\ge0. This is the classical Nelson argument, included to specify the cutoff and constants’ dependencies; see [30, 17] for its general constructive context.

For source dependence, H1(TL)↪L∞H^{1}(T_{L}) \hookrightarrow L^{\infty} and ∥PNv∥∞≤C∥v∥H\lVert P_{N}v\rVert_{\infty} \leq C\lVert v\rVert_{H}. Expansion of :(ϕN+PNv)4::(\phi_{N}+P_{N}v)^{4}: gives finitely many Wick polynomials with bounded deterministic coefficients. If j≤4j \leq4 and gg is bounded, Wick’s rule gives

∥∫Tg:ϕNj:∥22≤j!∥g∥∞2βL∥CNj∥L1.\left\lVert\int_{\mathcal{T}} g :\phi_{N}^{j}: \right\rVert_{2}^{2} \leq j!\lVert g\rVert_{\infty}^{2}\beta L\lVert C_{N}^{j}\rVert_{L^{1}}.

Hypercontractivity proves (3.1). For convergence use CNj→CjC_{N}^{j}\to C^{j} in L1L^{1}, together with ∥(I−PN)v∥∞≤CN−1/2∥v∥H\lVert(I-P_{N})v\rVert_{\infty}\leq CN^{-1/2}\lVert v\rVert_{H}.

To obtain the shifted exponential bound, apply the static Cameron–Martin identity with the complete shift vv. It gives

Ee−aUN(v)=e−β∥v∥H2/2−abβ∥PNv∥22Ee−aUN(0)+YAv+2abPNv.(18)\mathbb{E}e^{-aU_{N}(v)}=e^{-\beta\lVert v\rVert_{H}^{2}/2-ab\beta\lVert P_{N}v\rVert_{2}^{2}}\mathbb{E}e^{-aU_{N}(0)+Y_{A}v+2abP_{N}v}. \tag*{(18)}

The prefactor is at most one. Cauchy–Schwarz bounds the last expectation by

(Ee−2aUN(0))1/2exp⁡{β(1+2abm2)2R2}.\left(\mathbb{E}e^{-2aU_{N}(0)}\right)^{1/2}\exp\left\{\beta\left(1+\frac{2ab}{m^{2}}\right)^{2}R^{2}\right\}.

This proves (3.2). Finally EUN(v)=β∫(PNv)4≤CR4\mathbb{E}U_{N}(v)=\beta\int(P_{N}v)^{4}\leq CR^{4}, so Jensen’s inequality gives ZN,λ(v)≥e−CλR4Z_{N,\lambda}(v)\geq e^{-C\lambda R^{4}}. Dividing the uniform exponential bounds by this lower bound proves the density estimate and its convergence. □

Local cubic fields and the quantum realization.

Lemma 3.2 (Sharp-time observables). For h∈L2(O)h\in L^{2}(O) the thermal Wick cubic WhW_{h} exists in every finite Lp(μ)L^{p}(\mu) and is measurable with respect to FO=σ{ϕ(0,h):h∈Cc∞(O)}\mathcal{F}_{O}=\sigma\{\phi(0,h):h\in C_{c}^{\infty}(O)\}. Moreover

EWhWk=6∫O×Oh(y)k(z)Cβ(y−z)3 dy dz,(19)\mathbb{E}W_{h}W_{k}=6\int_{O\times O}h(y)k(z)C_{\beta}(y-z)^{3}\,dy\,dz, \tag*{(19)}
∥Wh∥p≤cp∥h∥2,sup⁡λ≤λ∗,∥v∥H≤REν^λ,vWh2≤CW∥h∥22.(20)\lVert W_{h}\rVert_{p}\leq c_{p}\lVert h\rVert_{2},\qquad \sup_{\lambda\leq\lambda_{*},\lVert v\rVert_{H}\leq R}\mathbb{E}_{\widehat{\nu}_{\lambda,v}}W_{h}^{2}\leq C_{W}\lVert h\rVert_{2}^{2}. \tag*{(20)}

One admissible choice is CW=162KR∥Cβ3∥L1(TL)C_{W}=162K_{R}\lVert C_{\beta}^{3}\rVert_{L^{1}(T_{L})}.

Proof. The Fourier coefficients of CβC_{\beta} have size O((1+∣η∣)−1)O((1+|\eta|)^{-1}), hence Cβ∈LqC_{\beta}\in L^{q} for all finite qq, and the projected covariances converge in each such space. The covariance of the difference of two nested Wick cubics is 6(Cβ,M3−Cβ,N3)6(C_{\beta,M}^{3}-C_{\beta,N}^{3}), which tends to zero in L1L^{1}. Young’s inequality gives convergence as maps h↦Whh\mapsto W_{h} from L2L^{2} to Gaussian L2L^{2}. Hypercontractivity upgrades this to every finite LpL^{p} and proves (3.6).

For locality first take h∈Cc∞(O)h\in C_{c}^{\infty}(O) and use local mollifiers with support smaller than the distance of supp⁡h\operatorname{supp}h to ∂O\partial O. The same covariance argument shows that their Wick cubics have the same limit as the Fourier construction. They are FO\mathcal{F}_{O}-measurable. Density then treats all h∈L2(O)h\in L^{2}(O).

Since WhW_{h} lies in third Gaussian chaos, ∥Wh∥4≤33/2∥Wh∥2\lVert W_{h}\rVert_{4}\leq3^{3/2}\lVert W_{h}\rVert_{2}. Consequently

Eν^λ,vWh2≤KR∥Wh∥42≤27KREWh2≤162KR∥Cβ3∥1∥h∥22.\mathbb{E}_{\widehat{\nu}_{\lambda,v}}W_{h}^{2}\leq K_{R}\lVert W_{h}\rVert_{4}^{2}\leq27K_{R}\mathbb{E}W_{h}^{2}\leq162K_{R}\lVert C_{\beta}^{3}\rVert_{1}\lVert h\rVert_{2}^{2}.

Proposition 3.3 (Thermal-state interface). For v∈Cc∞(I)v\in C_{c}^{\infty}(I), the finite-volume free thermal system admits the FKN perturbation by the sharp-time potential

Vλ,v=λ∫TL(:Φ4:β+b:Φ2:β) dx−Φ(Jλ,v).\mathcal{V}_{\lambda,v}=\lambda\int_{T_{L}}\left(:\Phi^{4}:_{\beta}+b:\Phi^{2}:_{\beta}\right)\,dx-\Phi(J_{\lambda,v}).

The resulting states ωλ,v\omega_{\lambda,v} are normal states on a common observable algebra. Their restrictions to its time-zero field subalgebra have the periodic Euclidean law (2.8). In particular their equal-time restrictions to OO have law Pλ,vOP_{\lambda,v}^{O} from (2.9).

Proof. We first identify precisely the perturbation theorem. Theorem 7.11 of Gérard–Jäkel [15] applies to a stochastically positive β\beta-KMS system (F,U,τ,ω)(\mathcal{F},\mathcal{U},\tau,\omega) and to a real, selfadjoint time-zero potential VV affiliated with its abelian algebra U\mathcal{U}, provided that

V,e−βV∈L1(Q,Σ0,μ),V∈L2+ε(Q,Σ0,μ)for some ε>0.(21)V,e^{-\beta V}\in L^{1}(Q,\Sigma_{0},\mu),\qquad V\in L^{2+\varepsilon}(Q,\Sigma_{0},\mu)\quad\text{for some }\varepsilon>0. \tag*{(21)}

We apply the L2+εL^{2+\varepsilon} integrability criterion. Theorem 7.6 establishes the periodic FKN path measure, while Propositions 7.8–7.10 identify the time-zero abelian algebra, the observable algebra, and the perturbed dynamics under the canonical unitary identification.

For clarity, we verify the free compact-space input. With kn=2πn/Lk_n=2\pi n/L and ωn=(kn2+m2)1/2\omega_n=(k_n^2+m^2)^{1/2}, its time-periodic covariance is, for ∣t∣≤β|t|\le\beta,

C(t,x)=1L∑n∈Zeiknxe−ωn∣t∣+e−ωn(β−∣t∣)2ωn(1−e−βωn).(22)C(t,x)=\frac{1}{L}\sum_{n\in\mathbb{Z}}e^{ik_nx}\frac{e^{-\omega_n|t|}+e^{-\omega_n(\beta-|t|)}}{2\omega_n(1-e^{-\beta\omega_n})}. \tag*{(22)}

Each spatial Fourier mode is the covariance of a thermal harmonic oscillator. In particular, for 0≤s,t≤β/20\le s,t\le\beta/2, the time-reflected kernel of one mode is

e−ωn(s+t)+e−βωneωn(s+t)2ωn(1−e−βωn),\frac{e^{-\omega_n(s+t)}+e^{-\beta\omega_n}e^{\omega_n(s+t)}}{2\omega_n(1-e^{-\beta\omega_n})},

a sum of two positive rank-one kernels. Summing the modes gives reflection positivity for the spatially smeared fields. The Gaussian bridge conditional-independence property for each oscillator, followed by the finite-mode limit, gives the periodic Markov property and the standard stochastically positive free KMS path space. Equivalently, the circle heat semigroup is the explicitly defined spectral semigroup

(et∂x2f)(x)=1L∑n∈Ze−tkn2f^(n)eiknx,t>0,(e^{t\partial_x^2}f)(x)=\frac{1}{L}\sum_{n\in\mathbb{Z}}e^{-tk_n^2}\widehat{f}(n)e^{ik_nx},\qquad t>0,

which requires no noncompact-space estimate; see [8, 15, 19]. The general results of Section 7 of [15] are formulated for the abstract KMS path space and do not require the spatial coordinate to be R\mathbb{R} once these free-system hypotheses are checked.

Apply the theorem with

V=Vλ,v=λ∫TL(: ⁣Φ4 ⁣: ⁣β+b: ⁣Φ2 ⁣: ⁣β) dx−Φ(Jλ,v),v∈Cc∞(I).V=\mathcal{V}_{\lambda,v}=\lambda\int_{\mathbb{T}_L}\bigl(:\!\Phi^4\!:\!_{\beta}+b:\!\Phi^2\!:\!_{\beta}\bigr)\,dx-\Phi(J_{\lambda,v}),\qquad v\in C^\infty_c(I).

At sharp time, ∥Cβ,M−Cβ,N∥2≤CN−1/2\|C_{\beta,M}-C_{\beta,N}\|_2\le CN^{-1/2} and sup⁡N∥Cβ,N∥6<∞\sup_N\|C_{\beta,N}\|_6<\infty. Repeating the Wick-contraction and hypercontractivity argument in Lemma 3.1, with a single spatial slice instead of the full time–space cylinder, gives LpL^p convergence of the integrated Wick quartic for all finite pp, with error CpN−1/4C_pN^{-1/4}. The deterministic cutoff lower bound is −Clog⁡2(eN)-C\log^2(eN), and the same low/high-frequency Nelson argument gives

sup⁡NEexp⁡[−qβλ∫TL(: ⁣ΦN4 ⁣: ⁣β+b: ⁣ΦN2 ⁣: ⁣β) dx]<∞,0≤q<∞,λ≥0.(23)\sup_N\mathbb{E}\exp\left[-q\beta\lambda\int_{\mathbb{T}_L}\bigl(:\!\Phi_N^4\!:\!_{\beta}+b:\!\Phi_N^2\!:\!_{\beta}\bigr)\,dx\right]<\infty,\qquad0\le q<\infty,\quad\lambda\ge0. \tag*{(23)}

The smooth source Φ(Jλ,v)\Phi(J_{\lambda,v}) is a centered Gaussian linear functional at sharp time, hence EeqΦ(Jλ,v)<∞\mathbb{E}e^{q\Phi(J_{\lambda,v})}<\infty for every real qq. Hölder’s inequality, the LpL^p convergence of the Wick integral, and (3.10) imply that V∈LpV\in L^p for every finite pp and e−βV∈L1e^{-\beta V}\in L^1. Both the Wick integral and the smooth field functional are measurable with respect to the sharp-time sigma algebra; their limits therefore define a real affiliated multiplication operator. This verifies each hypothesis in (3.8) for the circle model.

The normalized Euclidean path measure supplied by the periodic FKN theorem has density (2.8). Its sharp-time law, translated by the static Cameron–Martin shift, is exactly (2.9). The common normal-state realization and the stated equal-time local restrictions follow from Theorem 7.11 and Propositions 7.8–7.10 of [15].

We compare all states on a fixed time-zero local algebra A0(O)\mathcal{A}_0(O) containing the Weyl field and momentum operators supported in OO. Write

SO(ωλ,v∥ωλ,ω)=S(ωλ,v∣A0(O)∥ωλ,ω∣A0(O)).S_O(\omega_{\lambda,v}\mathbin{\|}\omega_{\lambda,\omega})=S(\omega_{\lambda,v}|_{\mathcal{A}_0(O)}\mathbin{\|}\omega_{\lambda,\omega}|_{\mathcal{A}_0(O)}).

Every WhW_h is affiliated to its time-zero field subalgebra. Monotonicity of Araki relative entropy [3] gives

SO(ωλ,v∥ωλ,w)≥KL⁡(Pλ,vO∥Pλ,wO).(24)S_{\mathcal{O}}(\omega_{\lambda,v}\Vert\omega_{\lambda,w}) \ge\operatorname{KL}(P_{\lambda,v}^{\mathcal{O}}\Vert P_{\lambda,w}^{\mathcal{O}}). \tag*{(24)}

For nonsmooth v∈Hv \in H, the probability estimates refer to (12); the quantum interpretation applies to smooth sources as in Proposition 3.3.

A uniform cubic-response expansion

Theorem 4.1 (Source-uniform weak-coupling response). Let b≥0b \ge0 and λ∗<∞\lambda_\ast< \infty be fixed. For every R<∞R < \infty the functionals

h⟼Eνλ,vWhh \longmapsto\mathbb{E}_{\nu_{\lambda,v}}W_h

define Mλ(v)∈Y=L2(O)\mathcal{M}_{\lambda}(v) \in Y = L^2(\mathcal{O}) for v∈Hv \in H, ∥v∥H≤R\lVert v\rVert_H \le R, and 0≤λ≤λ∗0 \le\lambda\le\lambda_\ast. There is a constant CRC_R such that

Mλ(v)=−24λTv+Rλ(v),Rλ(0)=0,sup⁡∥v∥H≤R∥DvRλ(v)∥H→Y≤CRλ2.(25)\mathcal{M}_{\lambda}(v)=-24\lambda\mathcal{T}v+\mathcal{R}_{\lambda}(v),\qquad\mathcal{R}_{\lambda}(0)=0,\qquad\sup_{\lVert v\rVert_H\le R}\lVert D_v\mathcal{R}_{\lambda}(v)\rVert_{H\to Y}\le C_R\lambda^2. \tag*{(25)}

All statements are about the continuum model. At finite spatial cutoff the analogous estimate holds with its own leading kernel TN\mathcal{T}_N, with constants independent of the cutoff. Passing to the limit gives (25) with the full kernel T\mathcal{T}.

Proof. Put Fh(λ,v)=E[Whe−λU(v)]F_h(\lambda,v)=\mathbb{E}[W_h e^{-\lambda U(v)}] and Z(λ,v)=Ee−λU(v)Z(\lambda,v)=\mathbb{E}e^{-\lambda U(v)}. Riesz representation and Lemmas 3.1–3.2 define Mλ\mathcal{M}_{\lambda} because

∣Fh/Z∣≤KR∥Wh∥2≤CR∥h∥2.\lvert F_h/Z\rvert\le K_R\lVert W_h\rVert_2 \le C_R\lVert h\rVert_2.

We justify, in particular, the mixed derivative required for the remainder, Let Qv[u]=DvU(v)[u]Q_v[u]=D_vU(v)[u]. For j≤2j\le2,

∂λjFh=E[Wh(−U(v))je−λU(v)],DvFh[u]=−λE[WhQv[u]e−λU(v)].\partial_{\lambda}^{j}F_h=\mathbb{E}[W_h(-U(v))^j e^{-\lambda U(v)}],\qquad D_vF_h[u]=-\lambda\mathbb{E}[W_hQ_v[u]e^{-\lambda U(v)}].

In particular,

∂λ2DvFh[u]=E[WhQv[u](2U(v)−λU(v)2)e−λU(v)].(26)\partial_{\lambda}^{2}D_vF_h[u]=\mathbb{E}[W_hQ_v[u](2U(v)-\lambda U(v)^2)e^{-\lambda U(v)}]. \tag*{(26)}

The same formulas with WhW_h omitted hold for ZZ. For every fixed product in these expressions, Hölder’s inequality, (14), (15), and (20) give a uniform bound CR∥h∥2∥u∥HC_R\lVert h\rVert_2\lVert u\rVert_H. Taylor’s formula for the polynomial U(v+u)U(v+u) proves Fréchet differentiability. The same estimates dominate the difference quotients. Since Z≥e−Cλ∗R4Z\ge e^{-C\lambda_\ast R^4}, the quotient rule gives

sup⁡λ∈[0,λ∗], ∥v∥H≤R∣∂λ2Dv(Fh/Z)[u]∣≤CR∥h∥2∥u∥H.(27)\sup_{\lambda\in[0,\lambda_\ast],\,\lVert v\rVert_H\le R}\left\lvert\partial_{\lambda}^{2}D_v(F_h/Z)[u]\right\rvert\le C_R\lVert h\rVert_2\lVert u\rVert_H. \tag*{(27)}

The cutoff convergence in Lemma 3.1, again with Hölder, passes all these derivatives to the continuum uniformly on the balls. The derivatives in λ\lambda at zero are right derivatives; no convergent perturbation series is required.

At λ=0\lambda=0, M0(v)=0\mathcal{M}_0(v)=0. Expanding the shifted Wick polynomial gives

U(v)=V4+bV2+4∫Iv:ϕ3:+6∫Iv2:ϕ2:+4∫Iv3ϕ+β∫Iv4.(28)U(v)=V_4+bV_2+4\int_{\mathcal{I}}v:\phi^3:+6\int_{\mathcal{I}}v^2:\phi^2:+4\int_{\mathcal{I}}v^3\phi+\beta\int_{\mathcal{I}}v^4. \tag*{(28)}

Only the third-chaos term pairs with WhW_h. Wick’s rule yields

∂λ⟨h,Mλ(v)⟩∣0+=−E[WhU(v)]=−24∫O×Ih(y)v(x)(∫0βC(τ,x−y)3 dτ)dx dy=−24⟨h,Tv⟩.\begin{aligned} \left.\partial_{\lambda}\langle h,\mathcal{M}_{\lambda}(v)\rangle\right|_{0+}=-\mathbb{E}[W_hU(v)] \\ &=-24\int_{\mathcal{O}\times\mathcal{I}}h(y)v(x)\left(\int_0^\beta C(\tau,x-y)^3\,\mathrm{d}\tau\right)\mathrm{d}x\,\mathrm{d}y=-24\langle h,\mathcal{T}v\rangle. \end{aligned}

Taylor’s integral formula applied to DvMλD_v\mathcal{M}_{\lambda} and (27) proves (25). The measure at v=0v=0 is invariant under ϕ↦−ϕ\phi\mapsto-\phi, so its cubic mean is exactly zero and Rλ(0)=0\mathcal{R}_{\lambda}(0)=0. □\square

Remark 4.2 (What the cubic observable removes). The first coefficient in (25) is unchanged by bV2bV_2 or by an additional first-order quadratic counterterm accompanied by the corresponding quadratic-reference source calibration. Such terms have Gaussian chaos degree at most two and are orthogonal to WhW_h. Thus the coefficient is not a thermal mass-shift artifact. The finite-coupling expectation EWh\mathbb{E}W_h is a Wick cubic moment, not in general the exact connected third cumulant. Their first-order non-Gaussian response agrees here because the local free mean is zero.

For comparison, for a local linear test hh the same calculation gives, with v=suv=su,

Eλ,suΦ(h)=−4λsu3∫Iu(x)3(A−1h)(x) dx+OC1([−S,S])(λ2).(29)\mathbb{E}_{\lambda,su}\Phi(h)=-4\lambda s_u^3\int_I u(x)^3(A^{-1}h)(x)\,\mathrm{d}x+O_{C^1([-S,S])}(\lambda^2). \tag*{(29)}

The term linear in ss has disappeared after the calibration. Its derivative at s=0s=0 therefore cannot furnish the uniform first-order lower gain that an uncalibrated thermal mass term would supply. The cubic channel (25), by contrast, is linear in the entire displacement vv at its first nonzero order.

The thermal cubic kernel: injectivity, rank, and ill-posedness

A positive bilateral Laplace representation

Set κn=(m2+(2πn/β)2)1/2\kappa_n=(m^2+(2\pi n/\beta)^2)^{1/2}. On the cut circle 0<z<L0<z<L, the Matsubara expansion and its time integral give

C(τ,z)=1β∑n∈Ze2πinτ/βGκn(z).(30)C(\tau,z)=\frac{1}{\beta}\sum_{n\in\mathbb{Z}}e^{2\pi i n\tau/\beta}G_{\kappa_n}(z). \tag*{(30)}
Kβ(z)=1β2∑n,j∈ZGκn(z)Gκj(z)Gκn+j(z).(31)K_\beta(z)=\frac{1}{\beta^2}\sum_{n,j\in\mathbb{Z}}G_{\kappa_n}(z)G_{\kappa_j}(z)G_{\kappa_{n+j}}(z). \tag*{(31)}

For every closed substrip ϵ≤ℜz≤L−ϵ\epsilon\leq\Re z\leq L-\epsilon, the double sum and all its fixed-order derivatives converge absolutely and uniformly. Indeed each factor and its derivatives are bounded by a polynomial in κ\kappa times e−ϵκ/κe^{-\epsilon\kappa}/\kappa.

Lemma 5.1 (Positive atoms). There exists a positive, locally finite atomic measure σβ\sigma_\beta on R\mathbb{R} such that

Kβ(z)=∫Re−az dσβ(a),0<ℜz<L.(32)K_\beta(z)=\int_{\mathbb{R}}e^{-az}\,\mathrm{d}\sigma_\beta(a),\qquad0<\Re z<L. \tag*{(32)}

It has an atom of mass at least

bn=18β2mκn2(1−e−mL)(1−e−κnL)2(33)b_n=\frac{1}{8\beta^2m\kappa_n^2(1-e^{-mL})(1-e^{-\kappa_nL})^2} \tag*{(33)}

at an=2κn+ma_n=2\kappa_n+m, for every n≥1n\geq1. In particular bn≥c(1+n)−2b_n\geq c(1+n)^{-2}, an≍na_n\asymp n, and inf⁡n(an+1−an)>0\inf_n(a_{n+1}-a_n)>0.

Proof. Expand each Green function in (31) into its two positive exponential terms. This defines a positive atomic measure, allowing repeated exponents to add their masses. Finiteness of (31) at any real z∈(0,L)z\in(0,L) proves local finiteness and the bilateral exponential integrability required in (32). The term (n,j)=(n,0)(n,j)=(n,0), with the decaying exponential chosen in all three factors, gives (33). The remaining assertions follow directly from the formula for κn\kappa_n. □

Theorem 5.2 (Infinite-dimensional injectivity). For the separated geometry (1), T:L2(I)→L2(O)\mathcal{T}:L^2(I)\to L^2(O) is injective.

Proof. Suppose Tu=0\mathcal{T}u=0 on OO. For ℓ−L<ℜy<0\ell-L<\Re y<0 define

F(y)=∫IKβ(x−y)u(x) dx,U(a)=∫Ie−axu(x) dx.F(y)=\int_I K_\beta(x-y)u(x)\,\mathrm{d}x,\qquad U(a)=\int_I e^{-ax}u(x)\,\mathrm{d}x.

The kernel estimates above show that FF is holomorphic in this vertical strip. Its continuous restriction vanishes on the real interval OO, so FF vanishes on the entire strip. Fix a real y0y_0 in the strip. The signed measure

dζ(a)=eay0U(a) dσβ(a)\mathrm{d}\zeta(a)=e^{ay_0}U(a)\,\mathrm{d}\sigma_\beta(a)

has finite total variation: its integral is bounded by ∫R∣u(x)∣Kβ(x−y0) dx<∞\int_{\mathbb{R}} |u(x)|\mathcal{K}_{\beta}(x-y_0)\,dx<\infty. For real tt,

F(y0+it)=∫Reiat dζ(a)=0.F(y_0+it)=\int_{\mathbb{R}}e^{iat}\,d\zeta(a)=0.

Fourier uniqueness for finite measures gives ζ=0\zeta=0. Positivity of the atom at each ana_n implies U(an)=0U(a_n)=0.

The function U(z)=∫0ℓe−zxu(x) dxU(z)=\int_0^\ell e^{-zx}u(x)\,dx is bounded and holomorphic on the right half-plane. If it is nonzero, its positive real zeros obey the Blaschke condition ∑an/(1+an2)<∞\sum a_n/(1+a_n^2)<\infty. To recall the elementary reason, map the half-plane to the disk by ζ=(z−1)/(z+1)\zeta=(z-1)/(z+1) and apply Jensen’s formula to the resulting bounded holomorphic function (after removing a possible finite-order zero at the center). Its zeros must have ∑(1−∣ζn∣)<∞\sum(1-|\zeta_n|)<\infty, whereas for an≍na_n\asymp n this sum diverges. Therefore UU is identically zero. Fourier uniqueness for e−cxu(x)e^{-cx}u(x) with any c>0c>0 gives u=0u=0. Equivalently, the last step is the Müntz completeness mechanism [4]; the half-plane proof specifies the zero-set condition needed here. □

Finite observable designs and an analytic rank distinction

Proposition 5.3 (Strict sign regularity). If y1<⋯<yr<x1<⋯<xry_1<\cdots<y_r<x_1<\cdots<x_r and xr−y1<Lx_r-y_1<L, then

(−1)r(r−1)/2det⁡[Kβ(xj−yi)]i,j=1r>0.(34)(-1)^{r(r-1)/2}\det[\mathcal{K}_{\beta}(x_j-y_i)]_{i,j=1}^{r}>0. \tag*{(34)}

Consequently nonzero nonnegative smooth probes and source displacement profiles supported on respectively ordered, disjoint subintervals give a full-rank cubic response matrix of every fixed finite size.

Proof. The continuous determinant identity applied to (32) gives

det⁡[Kβ(xj−yi)]=1r!∫Rrdet⁡[eakyi]i,kdet⁡[e−akxj]k,j∏k=1rdσβ(ak).\det[\mathcal{K}_{\beta}(x_j-y_i)]=\frac{1}{r!}\int_{\mathbb{R}^r}\det[e^{a_k y_i}]_{i,k}\det[e^{-a_k x_j}]_{k,j}\prod_{k=1}^{r}d\sigma_{\beta}(a_k).

Absolute integrability follows by expanding the determinants and using the finite positive kernel products. For ordered distinct aka_k, the first determinant is positive and the second has sign (−1)r(r−1)/2(-1)^{r(r-1)/2}. This sign property of exponentials follows from their nonzero Wronskian and Rolle’s theorem; its sign is fixed by coalescing the ordered arguments. The measure has infinitely many distinct atoms, so a set with rr distinct atoms has positive measure. This proves strictness. Integrating (5.5) against the ordered nonnegative profiles proves the last assertion by multilinearity. □

For uncalibrated linear field means, the first-order mass-shift response on one exterior interval lies in the span of emye^{my} and e−mye^{-my}, so it has rank at most two. If only the zero Matsubara mode is kept in (31), the cubic kernel is β−2Gm(z)3\beta^{-2}G_m(z)^3 and its rank is at most four, from the exponents ±m,±3m\pm m,\pm3m. The full thermal kernel has infinite rank by Theorem 5.2. Effective rank additionally depends on the observation noise, spatial smoothing and sample size.

Proposition 5.4 (Analytic smoothing). The operators T:L2(I)→L2(O)\mathcal{T}:L^2(I)\to L^2(O) and T:H→L2(O)\mathcal{T}:H\to L^2(O) are compact. There are C,c>0C,c>0 such that their singular values satisfy sn(T)≤Ce−cns_n(\mathcal{T})\le Ce^{-cn}. No bound ∥u∥H≤C∥Tu∥Y\lVert u\rVert_H\le C\lVert\mathcal{T}u\rVert_Y holds on HH. On any nontrivial spectral prior ball ∥u∥Xs≤M\lVert u\rVert_{X_s}\le M, no uniform Hölder inverse estimate ∥u∥H≤C∥Tu∥Yα\lVert u\rVert_H\le C\lVert\mathcal{T}u\rVert_Y^\alpha with α>0\alpha>0 holds.

Proof. Because I‾\overline{I} and O‾\overline{O} are separated, Kβ(x−y)\mathcal{K}_{\beta}(x-y) is holomorphic and bounded in a fixed complex neighborhood of y∈O‾y\in\overline{O}, uniformly in x∈I‾x\in\overline{I}. Rescale the observation interval to [−1,1][-1,1] and choose a Bernstein ellipse of parameter ρ>1\rho>1 lying in that common neighborhood. The Chebyshev coefficients cj(x)c_j(x) of y↦Kβ(x−y)y\mapsto\mathcal{K}_{\beta}(x-y) satisfy

sup⁡x∈I∣cj(x)∣≤Cρ−j.\sup_{x\in I}|c_j(x)|\le C\rho^{-j}.

This follows by writing the coefficient as a contour integral on the ellipse (the classical analytic approximation estimate [31 Chap. 8]). Thus the truncated kernel ∑j=0n−1cj(x)Tj(y)\sum_{j=0}^{n-1}c_j(x)T_j(y) is a sum of nn separated functions of xx and yy, hence gives an operator of rank at most nn. The uniformly

convergent geometric tail has supremum norm at most Cρ−nC\rho^{-n} on I×OI \times O and therefore Hilbert–Schmidt and operator norm at most C′ρ−nC'\rho^{-n}. The approximation-number characterization of Hilbert-space singular values gives sn+1(T)≤C′ρ−ns_{n+1}(T) \le C'\rho^{-n}. Finally, composition with the bounded embedding H↪L2(I)H \hookrightarrow L^2(I) gives the same form for the map from HH to YY.

For the last assertion, let EnE_n be the first nn Dirichlet modes, normalized in HH. A rank-(n−1)(n-1) approximant has a unit vector un∈Enu_n \in E_n in its kernel, whence ∥Tun∥≤Ce−cn\|Tu_n\| \le Ce^{-cn} and ∥un∥Xs≤Cns\|u_n\|_{X_s} \le Cn^s. Scaling to the prior ball produces ∥vn∥H≥cMn−s\|v_n\|_H \ge cMn^{-s} and ∥Tvn∥≤CMn−se−cn\|Tv_n\| \le CMn^{-s}e^{-cn}. These contradict any such Hölder estimate. Compactness and infinite-dimensional injectivity also rule out a bounded inverse on the whole space.

A reproducible finite-grid singular-value illustration

Take β=4\beta= 4, m=1m = 1, L=2πL = 2\pi, I=(0,1)I = (0,1) and O=(−1.8,−1.0)O = (-1.8,-1.0). These intervals satisfy (1.1) and the short-source condition (7.1). Using the Matsubara expansion (5.1), retain ∣n∣≤35|n| \le35 and approximate the periodic τ\tau-integral in Kβ(z)=∫0βC(τ,z)3 dτK_\beta(z) = \int_0^\beta C(\tau,z)^3\,\mathrm{d}\tau by a 192-point midpoint trapezoidal rule. On uniform midpoint grids xj∈Ix_j \in I and yi∈Oy_i \in O of size NN, form the weighted matrix

Aij(N)=Δx Δy Kβ(xj−yi).(35)A_{ij}^{(N)} = \sqrt{\Delta x\,\Delta y}\,K_\beta(x_j-y_i). \tag*{(35)}

Figure 1 displays the normalized singular values for N=96N = 96. Their values at indices 2, 4, 6, 8 are approximately 2.22⋅10−22.22 \cdot10^{-2}, 4.79⋅10−54.79 \cdot10^{-5}, 3.69⋅10−83.69 \cdot10^{-8}, and 4.35⋅10−124.35 \cdot10^{-12}. An independent N=48N = 48 calculation differs by at most about 2.2% for the first six normalized singular values. These finite-matrix singular values illustrate Proposition 5.4; they are not certified bounds for the continuum operator and the smallest values remain sensitive to floating-point precision.

Figure 1. Normalized singular values of the 96 × 96 weighted finite-grid matrix (5.6). The thermal Matsubara sum is truncated at ∣n∣≤35|n| \le35; the independent 48 × 48 comparison is reported in the text. Reproduction code and numerical data are included in the accompanying ancillary files.

Local entropy bounds and Sobolev resolution

We use the convention H2(P,Q)=∫(dP−dQ)2H^2(P,Q) = \int(\sqrt{\mathrm{d}P} - \sqrt{\mathrm{d}Q})^2. It satisfies KL⁡(P∥Q)≥H2(P,Q)\operatorname{KL}(P\mathbin{\|}Q) \ge H^2(P,Q).

Lemma 6.1 (A second-moment information bound). For the local laws defined above and ∥v∥H,∥w∥H≤R\|v\|_H,\|w\|_H \le R,

KL⁡(Pλ,vO∥Pλ,wO)≥∥Mλ(v)−Mλ(w)∥Y24CW.(36)\operatorname{KL}(P_{\lambda,v}^{\mathcal{O}}\mathbin{\|}P_{\lambda,w}^{\mathcal{O}}) \ge\frac{\|\mathcal{M}_\lambda(v)-\mathcal{M}_\lambda(w)\|_Y^2}{4C_W}. \tag*{(36)}

For smooth sources the same right side is a lower bound for SO(ωλ,v∥ωλ,w)S_O(\omega_{\lambda,v}\mathbin{\|}\omega_{\lambda,w}).

Proof. For any hh, factor the difference of densities as (p−q)(p+q)(\sqrt{p}-\sqrt{q})(\sqrt{p}+\sqrt{q}). Cauchy–Schwarz and (3.7) yield

∣EPWh−EQWh∣2≤2(EPWh2+EQWh2)H2(P,Q)≤4CW∥h∥22H2(P,Q).|\mathbb{E}_PWh-\mathbb{E}_QWh|^2 \le2(\mathbb{E}_PW_h^2+\mathbb{E}_QW_h^2)H^2(P,Q) \le4C_W\|h\|_2^2H^2(P,Q).

Taking the supremum over unit hh proves (36); smooth compactly supported tests are dense. The quantum statement is (24).

Theorem 6.2 (Fixed finite-dimensional source spaces). Let E⊂HE \subset H be any fixed finite-dimensional subspace and set

γE=inf⁡u∈E,∥u∥H=1∥Tu∥Y>0.\gamma_E=\inf_{u\in E,\lVert u\rVert_H=1}\lVert\mathcal{T}u\rVert_Y>0.

For v,w∈Ev,w\in E of HH-norm at most RR and 0<λ≤min⁡{λ∗,12γE/CR}0<\lambda\leq\min\{\lambda_\ast,12\gamma_E/C_R\} (omit the second restriction if CR=0C_R=0),

KL⁡(Pλ,vO∥Pλ,wO)≥36λ2γE2CW∥v−w∥H2.(37)\operatorname{KL}(P_{\lambda,v}^{O}\Vert P_{\lambda,w}^{O})\geq\frac{36\lambda^2\gamma_E^2}{C_W}\lVert v-w\rVert_H^2. \tag*{(37)}

If E⊂Cc∞(I)E\subset C^\infty_c(I), this also holds with SOS_O on the left. By (13) the right side is at least

36λ2γE2CW(1+2bλ/m2)∥Jλ,v−Jλ,w∥Hm−12.\frac{36\lambda^2\gamma_E^2}{C_W(1+2b\lambda/m^2)}\lVert J_{\lambda,v}-J_{\lambda,w}\rVert_{H_m^{-1}}^2.

Proof. Injectivity and compactness of the unit sphere of EE give γE>0\gamma_E>0. The C1C^1 estimate in Theorem 4.1, integrated on the source segment, yields

∥Mλ(v)−Mλ(w)∥≥λ(24γE−CRλ)∥v−w∥H≥12λγE∥v−w∥H.\lVert\mathcal{M}_\lambda(v)-\mathcal{M}_\lambda(w)\rVert\geq\lambda(24\gamma_E-C_R\lambda)\lVert v-w\rVert_H\geq12\lambda\gamma_E\lVert v-w\rVert_H.

Apply Lemma 6.1.

An infinite-dimensional prior at prescribed resolution

Let μk=(πk/ℓ)2+m2\mu_k=(\pi k/\ell)^2+m^2 and

ek(x)=(2/ℓ)1/2μk−1/2sin⁡(πkx/ℓ),Dek=μkek.e_k(x)=(2/\ell)^{1/2}\mu_k^{-1/2}\sin(\pi kx/\ell),\qquad D e_k=\mu_k e_k.

These form an orthonormal basis of HH. For s>0s>0 define

Xs=D(Ds/2),∥v∥Xs=∥Ds/2v∥H,BM,s={v∈Xs:∥v∥Xs≤M}.(38)X_s=\mathcal{D}(D^{s/2}),\qquad\lVert v\rVert_{X_s}=\lVert D^{s/2}v\rVert_H,\qquad B_{M,s}=\{v\in X_s:\lVert v\rVert_{X_s}\leq M\}. \tag*{(38)}

For larger ss, the Dirichlet spectral definition fixes the boundary compatibility conditions. The ball is compact in HH, and its HH-radius is at most R=μ1−s/2MR=\mu_1^{-s/2}M.

Theorem 6.3 (Uniform separation at every prescribed resolution). Fix M,s>0M,s>0. For r>0r>0 such that the set below is nonempty, put

αM,s(r)=min⁡v,w∈BM,s∥v−w∥H≥r∥T(v−w)∥Y.(39)\alpha_{M,s}(r)=\min_{\substack{v,w\in B_{M,s}\\\lVert v-w\rVert_H\geq r}}\lVert\mathcal{T}(v-w)\rVert_Y. \tag*{(39)}

Then αM,s(r)>0\alpha_{M,s}(r)>0. For every

0<λ≤λr:=min⁡{λ∗,6αM,s(r)CRR}(40)0<\lambda\leq\lambda_r:=\min\left\{\lambda_\ast,\frac{6\alpha_{M,s}(r)}{C_RR}\right\} \tag*{(40)}

(with the second restriction omitted if CR=0C_R=0), all pairs in this set satisfy

KL⁡(Pλ,vO∥Pλ,wO)≥36λ2αM,s(r)2CW>0.(41)\operatorname{KL}(P_{\lambda,v}^{O}\Vert P_{\lambda,w}^{O})\geq\frac{36\lambda^2\alpha_{M,s}(r)^2}{C_W}>0. \tag*{(41)}

For its smooth pairs the same bound holds for the local Araki relative entropy.

Proof. The minimizing set in (39) is compact in H×HH\times H. A zero minimum would give v=wv=w by Theorem 5.2, contradicting the resolution constraint. Since ∥v−w∥H≤2R\lVert v-w\rVert_H\leq2R, Theorem 4.1 gives

∥Mλ(v)−Mλ(w)∥≥24λαM,s(r)−2CRRλ2≥12λαM,s(r).\lVert\mathcal{M}_\lambda(v)-\mathcal{M}_\lambda(w)\rVert\geq24\lambda\alpha_{M,s}(r)-2C_RR\lambda^2\geq12\lambda\alpha_{M,s}(r).

Lemma 6.1 proves (41).

Here λr\lambda_r depends on the prescribed resolution rr; fixed-coupling separation at all resolutions is Problem 8.2.

The weak-coupling information coefficient

Proposition 6.4 (Non-Gaussian local information at order λ2\lambda^{2}). For fixed v,w∈Hv,w\in H, set

Bv,w=−E[U(v)−U(w)∣FO]+E[U(v)−U(w)].B_{v,w}=-\mathbb{E}[U(v)-U(w)\mid\mathcal{F}_{O}]+\mathbb{E}[U(v)-U(w)].

Then

lim⁡λ↓0λ−2H2(Pλ,vO,Pλ,wO)=14∥Bv,w∥22≥144sup⁡h≠0∣⟨h,T(v−w)⟩∣2EWh2.(42)\lim_{\lambda\downarrow0}\lambda^{-2}H^{2}(P^{O}_{\lambda,v},P^{O}_{\lambda,w})=\frac{1}{4}\lVert B_{v,w}\rVert_{2}^{2}\geq144\sup_{h\neq0}\frac{\lvert\langle h,\mathcal{T}(v-w)\rangle\rvert^{2}}{\mathbb{E}W_{h}^{2}}. \tag*{(42)}

The supremum ranges over nonzero h∈C0∞(O)h\in C_{0}^{\infty}(O) with nonzero denominator. For v≠wv\neq w it is positive. For smooth sources the right side is also a lower bound for lim inf⁡λ↓0SO/λ2\liminf_{\lambda\downarrow0}S_{O}/\lambda^{2}.

Proof. Relative to the free local law, the density is pλ,v=E[e−λU(v)∣FO]/Zλ(v)p_{\lambda,v}=\mathbb{E}[e^{-\lambda U(v)}\mid\mathcal{F}_{O}]/Z_{\lambda}(v). Lemma 3.1 gives its expansion in L4L^{4}:

pλ,v=1−λ{E[U(v)∣FO]−EU(v)}+OL4(λ2).p_{\lambda,v}=1-\lambda\{\mathbb{E}[U(v)\mid\mathcal{F}_{O}]-\mathbb{E}U(v)\}+O_{L^{4}}(\lambda^{2}).

For x≥−1x\geq-1, the error in 1+x=1+x/2+O(x2)\sqrt{1+x}=1+x/2+O(x^{2}) has a uniform absolute bound x2/2x^{2}/2. Thus the square-root densities have the corresponding L2L^{2} expansion, proving the equality in (42). (28) and Wick orthogonality give E[WhBv,w]=−24⟨h,T(v−w)⟩\mathbb{E}[W_{h}B_{v,w}]=-24\langle h,\mathcal{T}(v-w)\rangle. Cauchy–Schwarz proves the lower bound; injectivity proves its strict positivity for a suitable test. Finally use (3.11) and KL≥H2\mathrm{KL}\geq H^{2}.

Equation (42) gives the measured Hellinger coefficient and a lower coefficient for local Araki entropy; the corresponding classical Fisher information and quantum BKM metric describe different quantities.

An explicit conditional stability estimate

Theorem 6.3 needs only compactness and injectivity. We now obtain an explicit, nonoptimal separation function for the actual kernel under the additional geometric condition

0<ℓ<L/2.(43)0<\ell<L/2. \tag*{(43)}

Reflection energy and analytic propagation

Choose d>0d>0 with 2ℓ+d<L2\ell+d<L and set O♯=(−d−ℓ,−d)O^{\sharp}=(-d-\ell,-d). This interval is an analytic-continuation target, not an additional measurement region. If F=TuF=\mathcal{T}u, define

Ed(u)=∫Iu(x′)F(−d−x′) dx′=∫Re−ad∣∫Ie−axu(x) dx∣2dσβ(a)≥0.(44)E_{d}(u)=\int_{I}u(x')F(-d-x')\,\mathrm{d}x'=\int_{\mathbb{R}}e^{-ad}\left|\int_{I}e^{-ax}u(x)\,\mathrm{d}x\right|^{2}\mathrm{d}\sigma_{\beta}(a)\geq0. \tag*{(44)}

The identity follows from Fubini and (5.3), since 0<d+x+x′<L0<d+x+x'<L. Absolute integrability holds even for sign-changing uu, by the boundedness of the kernel on this compact separation range. The positivity of the reflected quadratic energy permits restriction to any subset of positive atoms for sign-changing uu.

Lemma 7.1 (Propagation to the reflected interval). There are C>0C>0 and α∈(0,1)\alpha\in(0,1), depending only on the fixed geometry and thermal parameters, such that, for ∥u∥H≤B\lVert u\rVert_{H}\leq B and δ=∥Tu∥L2(O)≤B\delta=\lVert\mathcal{T}u\rVert_{L^{2}(O)}\leq B,

sup⁡y∈O♯∣F(y)∣≤CB1−αδα.(45)\sup_{y\in O^{\sharp}}|F(y)|\leq CB^{1-\alpha}\delta^{\alpha}. \tag*{(45)}

Consequently, with Un=∫Ie−anxu(x) dxU_{n}=\int_{I}e^{-a_{n}x}u(x)\,\mathrm{d}x,

∑n=1∞bne−dan∣Un∣2≤CB2−αδα.(46)\sum_{n=1}^{\infty}b_{n}e^{-da_{n}}|U_{n}|^{2}\leq CB^{2-\alpha}\delta^{\alpha}. \tag*{(46)}

Proof. The analyticity statements in Section 5 imply that FF is holomorphic throughout the strip S={z∈C:ℓ−L<ℜz<0}S=\{z\in\mathbb{C}:\ell-L<\Re z<0\}. Since the closures of OO and O♯O^{\sharp} lie strictly inside the real part of SS, choose a bounded open rectangle D0⊆SD_{0}\subseteq S whose interior contains both, and a slightly larger rectangle D1⋐SD_1 \Subset S with D0‾⊂D1\overline{D_0} \subset D_1. Uniform kernel bounds on D1‾×I‾\overline{D_1} \times\overline{I} and Cauchy–Schwarz give

sup⁡z∈D1‾∣F′(z)∣+sup⁡z∈D0‾∣F′′(z)∣≤C0B.(47)\sup_{z\in\overline{D_1}} |F'(z)|+\sup_{z\in\overline{D_0}} |F''(z)|\le C_0B. \tag*{(47)}

Fix closed real intervals J0⊂⊂int⁡J1⊂⊂OJ_0\subset\subset\operatorname{int}J_1\subset\subset O with positive lengths and put A0=max⁡J0∣F∣A_0=\max_{J_0}|F|. Assume first B>0B>0 and δ>0\delta>0. At a maximizer x0∈J0x_0\in J_0, ∣F(x)∣≥A0/2|F(x)|\ge A_0/2 whenever ∣x−x0∣≤A0/(2C0B)|x-x_0|\le A_0/(2C_0B), by (47). For a small enough numerical constant c>0c>0, J1J_1 contains a one-sided or two-sided interval of length at least cmin⁡{1,A0/B}c\min\{1,A_0/B\} on which this lower bound holds. Because A0≤C0BA_0\le C_0B, integration over O⊃J1O\supset J_1 yields

δ2≥cA02min⁡{1,A0/B}≥c′A03B.\delta^2\ge cA_0^2\min\{1,A_0/B\}\ge c'\frac{A_0^3}{B}.

Consequently

max⁡J0∣F∣≤C1B1/3δ2/3.(48)\max_{J_0}|F|\le C_1B^{1/3}\delta^{2/3}. \tag*{(48)}

To propagate the smallness from J0J_0 to the reflected interval, consider the slit domain D=D0∖J0D=D_0\setminus J_0. The bounded subharmonic function log⁡∣F∣\log|F| (with value −∞-\infty at zeros) obeys the usual two-constants inequality; equivalently one may apply the maximum principle to log⁡(∣F∣+ε)\log(|F|+\varepsilon) and let ε↓0\varepsilon\downarrow0 [32]. Let ω(z)\omega(z) denote the harmonic measure of the two sides of the slit J0J_0 in DD. On the outer rectangle boundary, ∣F∣≤C0B|F|\le C_0B; on either side of the slit, ∣F∣≤A0|F|\le A_0. Thus, for z∈Dz\in D,

∣F(z)∣≤(C0B)1−ω(z)A0ω(z).(49)|F(z)|\le(C_0B)^{1-\omega(z)}A_0^{\omega(z)}. \tag*{(49)}

The harmonic measure ω\omega is strictly positive at every interior point of the connected bounded domain DD, because the slit has positive logarithmic capacity. By continuity and compactness, its minimum on the part of O♯‾\overline{O^\sharp} outside the interior of J0J_0 and inside DD is some ω∗>0\omega_*>0; on J0J_0 the stronger bound (48) already holds. Shrinking ω∗\omega_* if necessary and combining (48) with (49) proves

sup⁡O♯‾∣F∣≤CB1−αδα,α=23ω∗∈(0,1).\sup_{\overline{O^\sharp}}|F|\le CB^{1-\alpha}\delta^\alpha,\qquad\alpha=\frac{2}{3}\omega_*\in(0,1).

If δ=0\delta=0, the identity theorem gives F≡0F\equiv0 throughout SS; if B=0B=0, then u=0u=0. The remaining cases, including δ\delta comparable to BB, follow by enlarging CC in (47). Finally (44) and ∥u∥L1(I)≤CB\|u\|_{L^1(I)}\le CB give 0≤Ed(u)≤CB2−αδα0\le\mathcal{E}_d(u)\le CB^{2-\alpha}\delta^\alpha. In the positive atomic representation of (44), the term at each ana_n contributes bne−dan∣Un∣2≥0b_ne^{-da_n}|U_n|^2\ge0. Keeping only these terms yields (46). □

Quantitative recovery of the noninteger moments

Put a=4π/βa=4\pi/\beta, r=e−axr=e^{-ax}, r0=e−aℓr_0=e^{-a\ell}, and

q(r)={u(−log⁡r/a)/(ar),r0<r<1,0,0<r≤r0.q(r)= \begin{cases} u(-\log r/a)/(ar), & r_0<r<1,\\ 0, & 0<r\le r_0. \end{cases}

The zero trace of uu at both ends gives q∈H1(0,1)q\in H^1(0,1) and

c∥u∥L2(I)≤∥q∥2≤C∥u∥L2(I),∥q∥H1(0,1)≤C∥u∥H.(50)c\|u\|_{L^2(I)}\le\|q\|_2\le C\|u\|_{L^2(I)},\qquad\|q\|_{H^1(0,1)}\le C\|u\|_H. \tag*{(50)}

Writing ζn=an/a\zeta_n=a_n/a, the moments are

Un=∫01q(r)rζn dr,ζn=n2+(mβ2π)2+mβ4π.U_n=\int_0^1q(r)r^{\zeta_n}\,dr,\qquad\zeta_n=\sqrt{n^2+\left(\frac{m\beta}{2\pi}\right)^2}+\frac{m\beta}{4\pi}.

In particular n+c−≤ζn≤n+c+n+c_-\le\zeta_n\le n+c_+ and ζn+1−ζn≥c0>0\zeta_{n+1}-\zeta_n\ge c_0>0.

Lemma 7.2 (A Müntz approximation bound). Let EN=span⁡{rζ1,…,rζN}E_N=\operatorname{span}\{r^{\zeta_1},\ldots,r^{\zeta_N}\} and let ΠN\Pi_N be its orthogonal projection in L2(0,1)L^2(0,1). Then

∥q−ΠNq∥2≤C∥q∥H1(0,1)log⁡(eN).(51)\|q-\Pi_Nq\|_2\le\frac{C\|q\|_{H^1(0,1)}}{\log(eN)}. \tag*{(51)}

The Gram matrix GN=(1/(ζi+ζj+1))i,j=1NG_N = (1/(\zeta_i+\zeta_j+1))_{i,j=1}^N satisfies

∥GN−1∥op⁡≤eCN2log⁡(eN),∥ΠNq∥22=U∗GN−1U.(52)\lVert G_N^{-1}\rVert_{\operatorname{op}} \le e^{CN^2\log(eN)}, \qquad\lVert\Pi_N q\rVert_2^2 = U^*G_N^{-1}U. \tag*{(52)}

Proof. The Cauchy determinant formula, applied to the Gram matrix after adjoining rkr^k, gives the exact distance identity

dist⁡(rk,EN)=12k+1∏j=1N∣k−ζj∣k+ζj+1,k=0,1,…(53)\operatorname{dist}(r^k,E_N)=\frac{1}{\sqrt{2k+1}}\prod_{j=1}^N\frac{\lvert k-\zeta_j\rvert}{k+\zeta_j+1}, \qquad k=0,1,\ldots \tag*{(53)}

A coincident exponent gives zero and causes no difficulty. For 0≤k≤K0\le k\le K, retain factors with j≥JKj\ge J_K, where JK≤C(K+1)J_K\le C(K+1) is large enough that ζj>k\zeta_j>k. The remaining factors obey

ζj−kζj+k+1≤exp⁡(−2k+1j+k+c++1).\frac{\zeta_j-k}{\zeta_j+k+1}\le\exp\left(-\frac{2k+1}{j+k+c_++1}\right).

Comparison of the harmonic sum with its integral proves, when N≥C(K+1)N\ge C(K+1),

dist⁡(rk,EN)≤(C(K+1)N+1)2k+1≤C(K+1)N+1.\operatorname{dist}(r^k,E_N)\le\left(\frac{C(K+1)}{N+1}\right)^{2k+1}\le\frac{C(K+1)}{N+1}.

Let pKp_K be the degree-KK shifted Legendre projection of qq. The Legendre eigenvalue identity and integration by parts give

∥q−pK∥2≤C∥q′∥2K+1.\lVert q-p_K\rVert_2\le\frac{C\lVert q'\rVert_2}{K+1}.

Indeed the orthonormal coefficients cjc_j satisfy ∑j(j+1)∣cj∣2≤∫01r(1−r)∣q′(r)∣2 dr\sum j(j+1)\lvert c_j\rvert^2\le\int_0^1r(1-r)\lvert q'(r)\rvert^2\,dr. The monomial coefficients of pK=∑k=0KAkrkp_K=\sum_{k=0}^K A_kr^k satisfy ∑∣Ak∣≤C1K+1∥q∥2\sum\lvert A_k\rvert\le C_1^{K+1}\lVert q\rVert_2 for a fixed C1>1C_1>1, by the binomial formula for shifted Legendre polynomials. Hence

dist⁡(q,EN)≤C∥q∥H1K+1+C(K+1)C1K+1N+1∥q∥2.\operatorname{dist}(q,E_N)\le\frac{C\lVert q\rVert_{H^1}}{K+1}+\frac{C(K+1)C_1^{K+1}}{N+1}\lVert q\rVert_2.

Choose K=⌊ϵlog⁡(N+1)⌋K=\lfloor\epsilon\log(N+1)\rfloor with ϵ\epsilon small enough that C1K+1≤CNC_1^{K+1}\le C\sqrt{N}. This proves (7.9). The distance identity (7.11) and the preceding polynomial projection estimates give (7.9).

Finally

det⁡GN=∏i<j(ζj−ζi)2∏i,j(ζi+ζj+1).\det G_N=\frac{\prod_{i<j}(\zeta_j-\zeta_i)^2}{\prod_{i,j}(\zeta_i+\zeta_j+1)}.

Uniform separation of the ζj\zeta_j and their linear growth give det⁡GN≥e−CN2log⁡(eN)\det G_N\ge e^{-CN^2\log(eN)}. Using λmax⁡(GN)≤Tr⁡GN≤CN\lambda_{\max}(G_N)\le\operatorname{Tr}G_N\le CN bounds its smallest eigenvalue from below by the same form after changing CC. The projection identity in (7.10) follows by solving its normal equations. □

Theorem 7.3 (Double-logarithmic Sobolev stability). Assume (1.1) and (7.1). For each s>0s>0 there is CsC_s such that, for all u∈Xsu\in X_s and M>0M>0 with ∥u∥Xs≤M\lVert u\rVert_{X_s}\le M,

∥u∥H≤CsML(M/∥Tu∥Y)θs,θs=s1+s,L(x)=1+log⁡(1+log⁡(1+x)).(54)\lVert u\rVert_H\le\frac{C_sM}{\mathcal{L}(M/\lVert Tu\rVert_Y)^{\theta_s}},\qquad\theta_s=\frac{s}{1+s},\qquad\mathcal{L}(x)=1+\log(1+\log(1+x)). \tag*{(54)}

At zero data the right side is interpreted as zero. Constants depend on the fixed intervals, mm, β\beta, LL, ss, not on the source dimension or sign.

Proof. Initially assume only ∥u∥H≤B\lVert u\rVert_H\le B. Equations (5.4) and (7.4) imply

∑n≤N∣Un∣2≤C(1+N)2eCNB2−αδα.\sum_{n\le N}\lvert U_n\rvert^2\le C(1+N)^2e^{CN}B^{2-\alpha}\delta^\alpha.

Combining Lemma 7.2 and (7.8) gives

∥u∥2≤CBlog⁡(eN)+exp⁡{C2N2log⁡(eN)}B1−α/2δα/2.(55)\lVert u\rVert_2\le\frac{CB}{\log(eN)}+\exp\{C_2N^2\log(eN)\}B^{1-\alpha/2}\delta^{\alpha/2}. \tag*{(55)}

For sufficiently small ϵ=δ/B\epsilon=\delta/B, take

N=c1⌈log⁡(1/ϵ)log⁡log⁡(1/ϵ)⌉N=c_1\left\lceil\sqrt{\frac{\log(1/\epsilon)}{\log\log(1/\epsilon)}}\right\rceil

with c>0c>0 small. The second term is at most CBϵα/4CB\epsilon^{\alpha/4} and log⁡(eN)≥c′log⁡log⁡(1/ϵ)\log(eN)\ge c'\log\log(1/\epsilon). Adjusting constants outside this range yields ∥u∥2≤CB/L(B/δ)\|u\|_2\le CB/\mathcal{L}(B/\delta).

In the Dirichlet basis, Hölder’s inequality gives the exact spectral interpolation

∥u∥H≤∥u∥2s/(1+s)∥u∥Xs1/(1+s).\|u\|_H\le\|u\|_2^{s/(1+s)}\|u\|_{X_s}^{1/(1+s)}.

Use B=μ1−s/2MB=\mu_1^{-s/2}M and absorb fixed rescalings of the argument of L\mathcal{L}. If δ=0\delta=0, send N→∞N\to\infty in (55) instead.

Theorem 7.3 establishes the conditional bound under (43).

Remark 7.4 (Origin of the double logarithm and possible improvement). The propagation step (45) converts the original data size δ\delta into a power B1−αδαB^{1-\alpha}\delta^\alpha on an auxiliary interval. This step does not, by itself, force a double logarithm. The principal quantitative loss occurs in the subsequent reconstruction of the noninteger moments: our approximation estimate (51) has only an inverse-log NN rate, while (52) allows an amplification as large as exp⁡(CN2log⁡(eN))\exp(CN^2\log(eN)). Balancing these two contributions in (55) selects NN of order log⁡(B/δ)/log⁡log⁡(B/δ)\sqrt{\log(B/\delta)}/\log\log(B/\delta) and yields the double logarithm. Sharper Müntz–Jackson approximation or structured moment inversion might improve the rate, as suggested by the related mechanisms in [13, 14, 6]; however, their estimates do not automatically apply to the present frequencies ζn\zeta_n together with the reflected thermal weights and Dirichlet-spectral prior.

Corollary 7.5 (An explicit resolution function). Under the same geometry, the separation function in (39) has the lower bound

αM,s(r)≥2Mexp⁡(exp⁡(qr−1)−1)−1,qr=max⁡{2,(2CsMr)1/θs}.(56)\alpha_{M,s}(r)\ge\frac{2M}{\exp(\exp(q_r-1)-1)-1},\qquad q_r=\max\left\{2,\left(\frac{2C_sM}{r}\right)^{1/\theta_s}\right\}. \tag*{(56)}

Inserting this number in (40) and (41) gives explicit sufficient coupling and information thresholds for every prescribed positive resolution.

Proof. Apply (54) to u=v−wu=v-w, whose XsX_s norm is at most 2M2M, and solve L(2M/∥Tu∥)≤(2CsM/r)1/θs\mathcal{L}(2M/\|Tu\|)\le(2C_sM/r)^{1/\theta_s}.

The doubly exponential denominator in (56) leads to small sufficient thresholds; the constants follow from the stated analytic bounds.

Continuous inversion and a spectral consequence

For t>0t>0 define

γs(t)=inf⁡u∈Xs∥u∥H=1(∥Tu∥Y2+t∥u∥Xs2).\gamma_s(t)=\inf_{\substack{u\in X_s\\\|u\|_H=1}}\left(\|Tu\|_Y^2+t\|u\|_{X_s}^2\right).

Theorem 7.3 implies, for small tt,

γs(t)≥cstL(1/t)2θs.(57)\gamma_s(t)\ge c_s t\mathcal{L}(1/t)^{2\theta_s}. \tag*{(57)}

To verify this explicitly, a failure would give tj↓0t_j\downarrow0, ϵj↓0\epsilon_j\downarrow0 and unit uju_j with δj2+tjqj2≤ϵjtjL(1/tj)2θs\delta_j^2+t_jq_j^2\le\epsilon_jt_j\mathcal{L}(1/t_j)^{2\theta_s}, where qj=∥uj∥Xs≥μ1s/2q_j=\|u_j\|_{X_s}\ge\mu_1^{s/2}. Then qj≤ϵjL(1/tj)θsq_j\le\sqrt{\epsilon_j}\mathcal{L}(1/t_j)^{\theta_s} and L(qj/δj)≥cL(1/tj)\mathcal{L}(q_j/\delta_j)\ge c\mathcal{L}(1/t_j). Applying (54) with prior size qjq_j contradicts ∥uj∥H=1\|u_j\|_H=1.

If ∥zδ−Tv∗∥Y≤δ\|z^\delta-Tv_*\|_Y\le\delta and ∥v∗∥Xs≤M\|v_*\|_{X_s}\le M, the continuous Tikhonov minimizer

vtδ=arg⁡min⁡v∈Xs{12∥Tv−zδ∥Y2+t2∥v∥Xs2}v_t^\delta=\mathop{\arg\min}_{v\in X_s}\left\{\frac{1}{2}\|Tv-z^\delta\|_Y^2+\frac{t}{2}\|v\|_{X_s}^2\right\}

exists uniquely. Its normal equation and Cauchy–Schwarz in the product energy norm give

∥vtδ−v∗∥H2≤δ2+tM2γs(t).\|v_t^\delta-v_*\|_H^2\le\frac{\delta^2+tM^2}{\gamma_s(t)}.

With t=(δ/M)2t=(\delta/M)^2, this has the rate in (54). The regularization argument is standard [10]; the kernel-specific input is (57).

Fixed-coupling identifiability and its limitations

The exact residual problem

For fixed λ>0\lambda>0 let ν^λ,0∝e−λ(V4+bV2)μ\hat{\nu}_{\lambda,0}\propto e^{-\lambda(V_4+bV_2)}\mu, and put

ΔVv=∫G(: ⁣(ϕ+v)4 ⁣:β−: ⁣ϕ4 ⁣:β).\Delta V_v=\int_{\mathcal{G}}\left(:\!(\phi+v)^4\!:_\beta-:\!\phi^4\!:_\beta\right).

The calibrated shifted representation gives the exact local density

dPλ,vOdPλ,0O=Eν^λ,0[e−λΔVv∣FO]Eν^λ,0e−λΔVv.(58)\frac{\mathrm{d}P^O_{\lambda,v}}{\mathrm{d}P^O_{\lambda,0}}=\frac{\mathbb{E}_{\hat{\nu}_{\lambda,0}}\left[e^{-\lambda\Delta V_v}\mid\mathcal{F}_O\right]}{\mathbb{E}_{\hat{\nu}_{\lambda,0}}e^{-\lambda\Delta V_v}}. \tag*{(58)}

This is a density on the equal-time commutative field algebra, not a density matrix for the whole local quantum algebra. Equality of the two normalized conditional expectations is equivalent to equality of the corresponding measured laws. Showing that this implies v=wv=w at one fixed coupling would be sufficient for local quantum-state injectivity, but is not established here.

For smooth directions uu, differentiating (58) gives the measured Fisher information

Iλ,vO(u,u)=λ2Var⁡ν^λ,v(Eν^λ,v[Qv[u]∣FO]),Qv[u]=4∫Gu: ⁣(ϕ+v)3 ⁣:β.(59)\mathcal{I}^O_{\lambda,v}(u,u)=\lambda^2\operatorname{Var}_{\hat{\nu}_{\lambda,v}}\left(\mathbb{E}_{\hat{\nu}_{\lambda,v}}[Q_v[u]\mid\mathcal{F}_O]\right),\qquad Q_v[u]=4\int_{\mathcal{G}}u:\!(\phi+v)^3\!:_\beta. \tag*{(59)}

The conditional-score identity is standard [24]. Its integrability here follows from Lemma 3.1. Positivity of the conditional variance in every nonzero source direction is an open problem. The term λ2Var⁡(Qv[u])\lambda^2\operatorname{Var}(Q_v[u]) in the parameter-dependent shifted full model generally differs from the physical BKM metric; the local observation law is invariant under that shift.

A limitation of perturbative inference

Proposition 8.1 (An analytic remainder can destroy injectivity). There are bounded compact operators T,RT,R on ℓ2\ell^2, with TT injective, such that Mλ=−24λT+24λ2RM_\lambda=-24\lambda T+24\lambda^2R has a nonzero kernel at arbitrarily small positive couplings, despite a uniform O(λ2)O(\lambda^2) derivative remainder bound.

Proof. Take Ten=e−nenTe_n=e^{-n}e_n and Ren=ne−nenRe_n=ne^{-n}e_n. Both are bounded and compact, and M1/nen=0M_{1/n}e_n=0. □\square

The example isolates the additional structure required to deduce fixed-coupling infinite-dimensional uniqueness from Theorem 4.1 and injectivity of TT. Similarly, the diagonal analytic kernels with singular values e−ne^{-n} and e−n2e^{-n^2} both have exponential upper bounds but different logarithmic inverse scales. An upper bound does not select a matching stability rate.

A sufficient structural condition

For a fixed coupling and reference source v0v_0, let

η(v)=dPλ,vOdPλ,v0O\eta(v)=\sqrt{\frac{\mathrm{d}P^O_{\lambda,v}}{\mathrm{d}P^O_{\lambda,v_0}}}

and A0=Dη(v0)A_0=D\eta(v_0), when these derivatives exist. If on a convex source set

⟨Dη(v)[u],A0u⟩≥c∥A0u∥22(c>0),(60)\langle D\eta(v)[u],A_0u\rangle\ge c\lVert A_0u\rVert_2^2\qquad(c>0), \tag*{(60)}

then line integration gives

H2(Pλ,vO,Pλ,wO)≥c2∥A0(v−w)∥22=c24Iλ,v0O(v−w,v−w).H^2(P^O_{\lambda,v},P^O_{\lambda,w})\ge c^2\lVert A_0(v-w)\rVert_2^2=\frac{c^2}{4}\mathcal{I}^O_{\lambda,v_0}(v-w,v-w).

Indeed pair η(v)−η(w)\eta(v)-\eta(w) with A0(v−w)A_0(v-w) and apply Cauchy–Schwarz. Whether (60) holds for the interacting field remains open.

Problem 8.2. For a fixed positive coupling, is the conditional law in (58) injective on an exterior Sobolev prior? If so, what uniform conditional stability modulus can be proved? Is there a useful field-specific estimate such as (60), or a different completeness mechanism for the conditional expectation?

Conclusion

Quadratic-reference calibration isolates a cubic local response from the thermal mass shift. Its continuum weak-coupling expansion, kernel injectivity and analytic smoothing yield local relative-entropy separation for finite-dimensional source spaces and for prescribed Sobolev resolutions. A reflected positive energy gives conditional stability under the additional interval-length assumption. Fixed-coupling infinite-dimensional identifiability is formulated in Problem 8.2.

Acknowledgments and funding

Rubin Zhang acknowledges support from the China Scholarship Council (CSC). Ning Zhang acknowledges support from the National Natural Science Foundation of China (Grant No. 62535007) and Huazhong University of Science and Technology (Grant No. 5003011074).

Declaration of AI assistance

During manuscript preparation, ChatGPT (OpenAI) assisted with literature review, drafting and editing, organization and checking of mathematical derivations, and software and figure generation. The authors take responsibility for the content and accuracy of the manuscript and accompanying materials.

Data and code availability

No external datasets were used. The Python script reproduce_figure.py, the CSV values figure1_data.csv, and reproduction instructions are supplied as supplementary material (and as ancillary files in the anc/ directory of the arXiv source). The vector Figure 1 is supplied as bkm_singular_values.pdf.

References

School of Mathematics and Statistics, Huazhong University of Science and Technology, Wuhan, Hubei 430074, China

Current address: Department of Mathematical and Statistical Sciences, University of Alberta, Edmonton, Alberta T6G 2G1, Canada

Email address: [email protected]

School of Mathematics and Statistics, Huazhong University of Science and Technology, Wuhan, Hubei 430074, China

Email address: [email protected]

References

  1. [1]A. Anshu, S. Arunachalam, T. Kuwahara and M. Soleimanifar, Sample-efficient learning of interacting quantum systems, Nature Phys. 17 (2021), 931–935. doi:10.1038/s41567-021-01232-0.DOI
  2. [2]J. Apraiz, L. Escauriaza, G. Wang and C. Zhang, Observability inequalities and measurable sets, J. Eur. Math. Soc. 16 (2014), 2433–2475. doi:10.4171/JEMS/490.DOI
  3. [3]H. Araki, Relative entropy of states of von Neumann algebras, Publ. Res. Inst. Math. Sci. 11 (1976), 809–833. doi:10.2977/prims/1195191148.DOI
  4. [4]P. Borwein and T. Erdélyi, The full Müntz theorem in C[0,1] and L₁[0,1], J. London Math. Soc. (2) 54 (1996), 102–110. doi:10.1112/jlms/54.1.102.DOI
  5. [5]H. Casini, S. Grillo and D. Pontello, Relative entropy for coherent states from Araki formula, Phys. Rev. D 99 (2019), 125020. doi:10.1103/PhysRevD.99.125020.DOI
  6. [6]M. Choulli, Various stability estimates for the problem of determining an initial heat distribution from a single measurement, Riv. Mat. Univ. Parma (N.S.) 7 (2016), no. 2, 279–307. arXiv:1512.07421.
  7. [7]F. Ciolli, R. Longo and G. Ruzzi, The information in a wave, Commun. Math. Phys. 379 (2020), 979–1000. doi:10.1007/s00220-019-03593-3.DOI
  8. [8]E. B. Davies, Heat Kernels and Spectral Theory, Cambridge Tracts in Mathematics 92, Cambridge University Press, Cambridge, 1989. doi:10.1017/CBO9780511566158.DOI
  9. [9]N. Drago, F. Faldino and N. Pinamonti, Relative entropy and entropy production for equilibrium states in pAQFT, Ann. Henri Poincaré 19 (2018), 3289–3319. doi:10.1007/s00023-018-0730-2.
  10. [10]H. W. Engl, M. Hanke and A. Neubauer, Regularization of Inverse Problems, Mathematics and Its Applications 375, Kluwer Academic Publishers, Dordrecht, 1996.
  11. [11]A. Feizmohammadi and L. Oksanen, An inverse problem for a semi-linear elliptic equation in Riemannian geometries, J. Differential Equations 269 (2020), 4683–4719. doi:10.1016/j.jde.2020.03.037.DOI
  12. [12]M. B. Fröb, A. Much and K. Papadopoulos, Relative entropy for λφ⁴ in the Rindler wedge, Phys. Rev. D 114 (2026), 085001. doi:10.1103/z7tjp-p1xx.DOI
  13. [13]M. A. García-Ferrero and A. Rüland, On two methods for quantitative unique continuation results for some nonlinear operators, Comm. Partial Differential Equations 45 (2020), 1512–1560. doi:10.1080/03605302.2020.1776323.
  14. [14]G. Gendron, Stability estimates for an inverse Steklov problem in a class of hollow spheres, Asymptot. Anal. 126 (2022), 323–377. doi:10.3233/ASY-211684.
  15. [15]C. Gérard and C. D. Jäkel, Thermal quantum fields with spatially cutoff interactions in 1 + 1 space–time dimensions, J. Funct. Anal. 220 (2005), 157–213. doi:10.1016/j.jfa.2004.08.003.DOI
  16. [16]C. Gérard and C. D. Jäkel, Thermal quantum fields without cut-offs in 1 + 1 space-time dimensions, Rev. Math. Phys. 17 (2005), 113–174. doi:10.1142/S0129055X05002003.DOI
  17. [17]J. Glimm and A. Jaffe, Quantum Physics: A Functional Integral Point of View, 2nd ed., Springer, New York, 1987. doi:10.1007/978-1-4612-4728-9.DOI
  18. [18]R. Høegh-Krohn, Relativistic quantum statistical mechanics in two-dimensional space-time, Comm. Math. Phys. 38 (1974), 195–224. doi:10.1007/BF01651542.
  19. [19]A. Klein and L. J. Landau, Stochastic processes associated with KMS states, J. Funct. Anal. 42 (1981), 368–428. doi:10.1016/0022-1236(81)90096-3.DOI
  20. [20]R. Kubo, Statistical-mechanical theory of irreversible processes. I, J. Phys. Soc. Japan 12 (1957), 570–586. doi:10.1143/JPSJ.12.570.
  21. [21]Y. Kurylev, M. Lassas and G. Uhlmann, Inverse problems for Lorentzian manifolds and non-linear hyperbolic equations, Invent. Math. 212 (2018), 781–857. doi:10.1007/s00222-017-0780-y.DOI
  22. [22]T. Liimatainen and Y.-H. Lin, Uniqueness results for inverse source problems for semilinear elliptic equations, Inverse Problems 40 (2024), 045030. doi:10.1088/1361-6420/ad3088.DOI
  23. [23]R. Longo, Entropy of coherent excitations, Lett. Math. Phys. 109 (2019), 2587–2600. doi:10.1007/s11005-019-01196-6.DOI
  24. [24]T. A. Louis, Finding the observed information matrix when using the EM algorithm, J. Roy. Statist. Soc. Ser. B 44 (1982), 226–233. doi:10.1111/j.2517-6161.1982.tb01203.x.DOI
  25. [25]N. Mandache, Exponential instability in an inverse problem for the Schrödinger equation, Inverse Problems 17 (2001), 1435–1444. doi:10.1088/0266-5611/17/5/313.DOI
  26. [26]F. Natterer, The Mathematics of Computerized Tomography, Classics in Applied Mathematics 32, SIAM, Philadelphia, 2001. doi:10.1137/1.9780898719284.DOI
  27. [27]D. Petz, Geometry of canonical correlation on the state space of a quantum system, J. Math. Phys. 35 (1994), 780–795. doi:10.1063/1.530611.DOI
  28. [28]S. Saitoh, V. K. Tuan and M. Yamamoto, Conditional stability of a real inverse formula for the Laplace transform, Z. Anal. Anwend. 20 (2001), 193–202. doi:10.4171/ZAA/1010.
  29. [29]H. Sasaki and A. Suzuki, An inverse scattering problem for the Klein–Gordon equation with a classical source in quantum field theory, Hokkaido Math. J. 40 (2011), 149–186. doi:10.14492/hokmj/1310042826.DOI
  30. [30]B. Simon, The P(φ)₂ Euclidean (Quantum) Field Theory, Princeton University Press, Princeton, 1974.
  31. [31]L. N. Trefethen, Approximation Theory and Approximation Practice, extended edition, SIAM, Philadelphia, 2019. doi:10.1137/1.9781611975949.DOI
  32. [32]S. Vessella, A continuous dependence result in the analytic continuation problem, Forum Math. 11 (1999), 695–703. doi:10.1515/form.1999.020.DOI

Paper details

vlnc:W_jeby3z74
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

ChatGPT (OpenAI)

View statement
Looking for related papers…
Contents