Introduction and main result
We consider the relaxation to equilibrium of the massive sine-Gordon model in two dimensions as the lattice spacing tends to zero. We prove a logarithmic Sobolev inequality which remains uniform in this limit for 0 < β < 8 π 0 < \beta< 8\pi 0 < β < 8 π . Bauerschmidt and Bodineau [3 Theorem 1.6] established the inequality for β < 6 π \beta< 6\pi β < 6 π using the Polchinski flow. To reach the range proposed in [4 Section 6.1, p. 262] , we need a sharper estimate of the curvature produced by the nonlinear part of this flow. We average the fluctuations up to a fixed physical length, compare the remaining Hamiltonian with a convex one, and transfer its inequality to the original measure.
Gaussian averaging reduces the original cosine but also produces interactions which depend on field differences. These are unchanged by a common field shift, so the same decay cannot hold for the whole potential. We estimate their curvature by the gradient energy and keep the Gaussian decay for the remaining oscillation. The oscillating part can generate new neutral terms at each integration, so both estimates must be preserved.
Let m , L > 0 m,L>0 m , L > 0 , L / ε ∈ N L/\varepsilon\in\mathbb{N} L / ε ∈ N , and Λ = ( Z / ( L / ε ) Z ) 2 \Lambda= (\mathbb{Z}/(L/\varepsilon)\mathbb{Z})^2 Λ = ( Z / ( L / ε ) Z ) 2 . A site x x x has physical position ε x \varepsilon x ε x . Write ( f , g ) = ∑ x f x g x (f,g) = \sum_x f_xg_x ( f , g ) = ∑ x f x g x , let e 1 , e 2 e_1,e_2 e 1 , e 2 be the coordinate unit vectors, and use the nearest-neighbour Laplacian Δ Λ φ x = ∑ i = 1 2 ( φ x + e i + φ x − e i − 2 φ x ) \Delta_\Lambda\varphi_x = \sum_{i=1}^{2}(\varphi_{x+e_i}+\varphi_{x-e_i}-2\varphi_x) Δ Λ φ x = ∑ i = 1 2 ( φ x + e i + φ x − e i − 2 φ x ) . For z ∈ R z \in\mathbb{R} z ∈ R and 0 < β < 8 π 0 < \beta< 8\pi 0 < β < 8 π , put
A = − Δ Λ + ε 2 m 2 i d , α = 2 − β 4 π , V 0 ( φ ) = 2 z ε α ∑ x cos ( β φ x ) . (1) A=-\Delta_\Lambda+\varepsilon^2m^2\mathrm{id}, \qquad\alpha=2-\frac{\beta}{4\pi}, \qquad V_0(\varphi)=2z\varepsilon^\alpha\sum_x\cos(\sqrt{\beta}\varphi_x). \tag*{(1)} A = − Δ Λ + ε 2 m 2 id , α = 2 − 4 π β , V 0 ( φ ) = 2 z ε α x ∑ cos ( β φ x ) . (1)
The sine-Gordon measure on R Λ \mathbb{R}^{\Lambda} R Λ is
d ν ε , L ( φ ) = Z ε , L − 1 e − 1 2 ( φ , A φ ) − V 0 ( φ ) ∏ x d φ x . (2) d\nu_{\varepsilon,L}(\varphi)=Z_{\varepsilon,L}^{-1}e^{-\frac{1}{2}(\varphi,A\varphi)-V_{0}(\varphi)}\prod_{x}d\varphi_x. \tag*{(2)} d ν ε , L ( φ ) = Z ε , L − 1 e − 2 1 ( φ , A φ ) − V 0 ( φ ) x ∏ d φ x . (2)
In V 0 V_0 V 0 , the factor ε α = ε 2 ε − β / ( 4 π ) \varepsilon^\alpha=\varepsilon^2\varepsilon^{-\beta/(4\pi)} ε α = ε 2 ε − β / ( 4 π ) combines the area element with the compensation for logarithmic Gaussian variance, as in [3 (1.15)–(1.16)] . Throughout, ∂ x \partial_x ∂ x , ∇ \nabla ∇ and Hess differentiate field values. We use the Euclidean norm ∥ v ∥ 2 = ( ∑ x ∣ v x ∣ 2 ) 1 / 2 \lVert v\rVert_2=(\sum_x|v_x|^2)^{1/2} ∥ v ∥ 2 = ( ∑ x ∣ v x ∣ 2 ) 1/2 for vectors and its induced matrix norm ∥ ⋅ ∥ o p \lVert\cdot\rVert_{\mathrm{op}} ∥ ⋅ ∥ op . For a field-dependent vector or matrix, ∥ ⋅ ∥ ∞ , 2 \lVert\cdot\rVert_{\infty,2} ∥ ⋅ ∥ ∞ , 2 or ∥ ⋅ ∥ ∞ , o p \lVert\cdot\rVert_{\infty,\mathrm{op}} ∥ ⋅ ∥ ∞ , op also takes the supremum over fields. In particular,
∥ Hess F ∥ ∞ , o p = sup φ ∈ R Λ sup ∥ v ∥ 2 = 1 ∥ Hess F ( φ ) v ∥ 2 , \lVert\operatorname{Hess}F\rVert_{\infty,\mathrm{op}}=\sup_{\varphi\in\mathbb{R}^{\Lambda}}\sup_{\lVert v\rVert_2=1}\lVert\operatorname{Hess}F(\varphi)v\rVert_2, ∥ Hess F ∥ ∞ , op = φ ∈ R Λ sup ∥ v ∥ 2 = 1 sup ∥ Hess F ( φ ) v ∥ 2 ,
where the first supremum is over all field configurations and the second over all Euclidean unit vectors in R Λ \mathbb{R}^{\Lambda} R Λ . For scalar functions, ∥ f ∥ ∞ = sup φ ∣ f ( φ ) ∣ \lVert f\rVert_\infty=\sup_\varphi|f(\varphi)| ∥ f ∥ ∞ = sup φ ∣ f ( φ ) ∣ . The letters C C C , C β C_\beta C β , C M C_M C M denote scalar bounds; a generic Gaussian covariance will be denoted by Σ \Sigma Σ .
For nonnegative F F F , define Ent ν F = E ν ( F log F ) − E ν F log E ν F \operatorname{Ent}_\nu F=\mathbb{E}_\nu(F\log F)-\mathbb{E}_\nu F\log\mathbb{E}_\nu F Ent ν F = E ν ( F log F ) − E ν F log E ν F and the physically normalized Dirichlet form
D ε , L ( F ) = ε − 2 ∑ x E ν ε , L [ ( ∂ x F ) 2 ] . (3) \mathcal{D}_{\varepsilon,L}(F)=\varepsilon^{-2}\sum_x\mathbb{E}_{\nu_{\varepsilon,L}}\left[(\partial_xF)^2\right]. \tag*{(3)} D ε , L ( F ) = ε − 2 x ∑ E ν ε , L [ ( ∂ x F ) 2 ] . (3)
Let γ ε , L \gamma_{\varepsilon,L} γ ε , L be the largest constant for which
Ent ν ε , L F ≤ 2 γ ε , L D ε , L ( F ) (4) \operatorname{Ent}_{\nu_{\varepsilon,L}}F\leq\frac{2}{\gamma_{\varepsilon,L}}\mathcal{D}_{\varepsilon,L}(\sqrt{F}) \tag*{(4)} Ent ν ε , L F ≤ γ ε , L 2 D ε , L ( F ) (4)
holds for every smooth positive F F F . The factor ε − 2 \varepsilon^{-2} ε − 2 gives γ ε , L = m 2 \gamma_{\varepsilon,L}=m^2 γ ε , L = m 2 in the Gaussian case z = 0 z=0 z = 0 . By the approximation used in [3 Section 2.2, p. 2075] , it suffices to prove (4) for smooth F F F bounded above and away from zero.
Theorem 1.1. Let 0 < β < 8 π 0<\beta<8\pi 0 < β < 8 π . For every fixed m m m , L > 0 L>0 L > 0 and z ∈ R z\in\mathbb{R} z ∈ R ,
inf ε > 0 L / ε ∈ N γ ε , L > 0. (5) \inf_{\substack{\varepsilon>0\\ L/\varepsilon\in\mathbb{N}}}\gamma_{\varepsilon,L}>0. \tag*{(5)} ε > 0 L / ε ∈ N inf γ ε , L > 0. (5) There are c β , C β > 0 c_\beta,C_\beta>0 c β , C β > 0 such that, if ∣ z ∣ m − α ≤ c β |z|m^{-\alpha}\leq c_\beta ∣ z ∣ m − α ≤ c β , then
γ ε , L ≥ m 2 e − C β ∣ z ∣ m − α ≥ m 2 − C β ∣ z ∣ m β / ( 4 π ) . (6) \gamma_{\varepsilon,L}\geq m^2e^{-C_\beta|z|m^{-\alpha}}\geq m^2-C_\beta|z|m^{\beta/(4\pi)}. \tag*{(6)} γ ε , L ≥ m 2 e − C β ∣ z ∣ m − α ≥ m 2 − C β ∣ z ∣ m β / ( 4 π ) . (6) The constants in this assertion are independent of L L L , m m m , z z z , ε \varepsilon ε .
The two assertions use different stopping lengths. Since α > 0 \alpha>0 α > 0 , the dimensionless coupling ∣ z ∣ r α |z|r^\alpha ∣ z ∣ r α is small at a sufficiently short physical length for any fixed z z z , independently of the cutoff. This gives (5) . Under the small-coupling assumption, we can stop at a fixed fraction of the mass length, or of the side length in a smaller torus, giving (6) .
For the relaxation interpretation, let f s f_s f s be the density relative to ν ε , L \nu_{\varepsilon,L} ν ε , L at physical time s s s for the reversible Langevin dynamics with Dirichlet form D ε , L \mathcal{D}_{\varepsilon,L} D ε , L . The entropy dissipation identity [4 Proposition 2.2] gives
d d s Ent ν ε , L f s = − 4 D ε , L ( f s ) ≤ − 2 γ ε , L Ent ν ε , L f s . \frac{d}{ds}\operatorname{Ent}_{\nu_{\varepsilon,L}}f_s=-4\mathcal{D}_{\varepsilon,L}(\sqrt{f_s})\leq-2\gamma_{\varepsilon,L}\operatorname{Ent}_{\nu_{\varepsilon,L}}f_s. d s d Ent ν ε , L f s = − 4 D ε , L ( f s ) ≤ − 2 γ ε , L Ent ν ε , L f s .
The entropy therefore decays exponentially at a rate independent of the lattice resolution. The same constant controls fluctuations. By the spectral-gap consequence [4 Proposition 2.5] , for a smooth test function f f f on the physical torus,
Var ν ε , L ( ε 2 ∑ x f ( ε x ) φ x ) ≤ γ ε , L − 1 ε 2 ∑ x f ( ε x ) 2 . \operatorname{Var}_{\nu_{\varepsilon,L}}\left(\varepsilon^2 \sum_x f(\varepsilon x)\varphi_x\right) \le\gamma_{\varepsilon,L}^{-1}\varepsilon^2 \sum_x f(\varepsilon x)^2. Var ν ε , L ( ε 2 x ∑ f ( ε x ) φ x ) ≤ γ ε , L − 1 ε 2 x ∑ f ( ε x ) 2 .
The Riemann sum on the right remains bounded as ε → 0 \varepsilon\to0 ε → 0 . The logarithmic Sobolev inequality also gives Gaussian concentration uniformly in the cutoff [2 Proposition 5.4.1 and (5.4.2)] .
The estimates needed along the flow
In physical units, the microscopic curvature can be as low as m 2 − 2 ∣ z ∣ ε − β / ( 4 π ) m^2-2\lvert z\rvert\varepsilon^{-\beta/(4\pi)} m 2 − 2 ∣ z ∣ ε − β / ( 4 π ) , and sup V 0 − inf V 0 ≤ 4 ∣ z ∣ L − β / ( 4 π ) \sup V_0-\inf V_0\leq4\lvert z\rvert L^{-\beta/(4\pi)} sup V 0 − inf V 0 ≤ 4 ∣ z ∣ L − β / ( 4 π ) . To obtain cutoff-uniform bounds from the convexity and bounded-perturbation criteria [1 , 16 ] , we apply them after smoothing. We must therefore control the renormalized potential at every background field, including fields far from equilibrium.
Already the product of two cosines contains cos ( β ( φ x − φ y ) ) \cos(\sqrt{\beta}(\varphi_x-\varphi_y)) cos ( β ( φ x − φ y )) , which is unchanged when the same constant is added to every field value. We call interactions with this invariance neutral. Differentiating twice in a field direction f f f shows the gain that this invariance can provide,
d 2 d s 2 cos ( β [ φ x + s f x − φ y − s f y ] ) ∣ s = 0 = − β cos ( β ( φ x − φ y ) ) ( f x − f y ) 2 . (7) \left.\frac{d^2}{ds^2}\cos\left(\sqrt{\beta}[\varphi_x+sf_x-\varphi_y-sf_y]\right)\right|_{s=0}=-\beta\cos\left(\sqrt{\beta}(\varphi_x-\varphi_y)\right)(f_x-f_y)^2. \tag*{(7)} d s 2 d 2 cos ( β [ φ x + s f x − φ y − s f y ] ) s = 0 = − β cos ( β ( φ x − φ y ) ) ( f x − f y ) 2 . (7)
For the local power counting, take f f f to be a heat-smoothed direction and measure the separation r r r in units of the smoothing length. Each difference then supplies one power of r r r . Together with the pair coefficient r − β / ( 2 π ) r^{-\beta/(2\pi)} r − β / ( 2 π ) and the radial area element, this gives the integral
∫ 0 1 r 3 − β / ( 2 π ) d r = 1 4 − β / ( 2 π ) < ∞ ⟺ β < 8 π . (8) \int_0^1 r^{3-\beta/(2\pi)}\,dr=\frac{1}{4-\beta/(2\pi)}<\infty\quad\Longleftrightarrow\quad\beta<8\pi. \tag*{(8)} ∫ 0 1 r 3 − β / ( 2 π ) d r = 4 − β / ( 2 π ) 1 < ∞ ⟺ β < 8 π . (8)
Bauerschmidt and Bodineau already prove this pair bound in [3 Proposition 3.6] . Their higher-order estimate uses a first spatial moment, which restricts it to β < 6 π \beta<6\pi β < 6 π [3 Lemmas 3.11 and A.4, Section 3.7] . The pair calculation therefore identifies the expected range. To obtain a logarithmic Sobolev inequality in this range, we must control the curvature of the exact interaction after all orders have been combined.
We keep a spatially weighted bound on the full Hessian. This bound ensures convexity of the conditional Hamiltonians used in the next integration and also controls the neutral part. Indeed, the Hessian of a neutral interaction has zero row and column sums, so its quadratic form can be written in differences of the test vector, as in (7) . Along the flow we apply this identity to heat kernels. At the stopping scale, spatial translation invariance and a shortest-path estimate bound it by the gradient energy, as shown in (59) . This contribution changes the coefficient of the gradient energy but leaves the Gaussian mass unchanged.
For the oscillating part, we compare Gaussian averaging with an average in which the original cosine phase fluctuates independently of the other variables. The latter has the exact cosine decay. Their difference involves derivatives with this phase held fixed. These derivatives vanish for the original cosine, and we control them by a quadratic bound in the coupling. Their sum over sites is zero. Subtracting the central heat-kernel value then produces a second spatial difference, by evenness of the kernel. The resulting time integral, computed in (43) , is finite throughout β < 8 π \beta<8\pi β < 8 π .
The main difficulty is to preserve this first-derivative bound. Applying the same comparison to a derivative would require second derivatives of the local term. We make the comparison at the midpoint, where the error decays as a power of the scale ratio. Differentiating the remaining Gaussian density gives a derivative bound with this decay. The incoming first derivatives give a second bound with stronger spatial decay. Interpolating the two derivative estimates retains both forms of decay. When the length is multiplied by M M M , the coefficient of the incoming quadratic bound grows at most as M 3 α / 2 M^{3\alpha/2} M 3 α /2 , whereas the allowed quadratic bound grows as M 2 α M^{2\alpha} M 2 α . This strict gain, made explicit in Lemma 3.6 , closes the induction for every α > 0 \alpha> 0 α > 0 .
Renormalization-group analyses of the sine-Gordon measure include [6 , 19 , 10 , 11 ] . Further constructions using flows below 6 π 6\pi 6 π appear in [5 , 15 , 23 ] . Pelaic [20 Proposition 1.1 and p. 2, footnote 2] constructs the measure throughout the subcritical range for arbitrary real coupling at finite volume. Meyer [18 Theorems 1.1–1.2 and Proposition 1.3] obtains order-by-order approximate potentials throughout the subcritical range, with polynomial field-gradient dependence, and a construction by forward–backward stochastic differential equations below ( 6 / 7 ) 8 π (6/7)8\pi ( 6/7 ) 8 π . The parameter denoted β 2 \beta^2 β 2 there is our β \beta β . These field-dependent estimates serve the construction of the measure. Our curvature criterion requires a Hessian bound uniform over all background fields. To apply it through an approximate potential, one would also have to control the Hessian of the remainder with the same uniformity.
For the dynamics, the cutoff-uniform relaxation estimate complements the well-posedness theory obtained by stochastic quantization. Local well-posedness in the full subcritical range is proved in [9 Theorem 1.1] , and global well-posedness below 6 π 6\pi 6 π in [7 Theorem 1.1] . Finite-volume a priori bounds throughout the subcritical range are obtained in [13 Theorem 1.1 and Remark 1.2] .
Section 2 deduces the theorem from estimates on the renormalized potential. Section 3 establishes them by iteration.
Reduction to estimates on the renormalized potential
We integrate Gaussian fluctuations in order of their diffusion length,
Q t = e − t A / 2 , C t = ∫ 0 t e − s A d s , e − V t ( ϕ ) = E C t e − V 0 ( ϕ + ζ ) . (9) Q_t = e^{-tA/2}, \qquad C_t = \int_0^t e^{-sA}\,\mathrm{d}s, \qquad e^{-V_t(\phi)} = \mathbb{E}_{C_t}e^{-V_0(\phi+\zeta)}. \tag*{(9)} Q t = e − t A /2 , C t = ∫ 0 t e − s A d s , e − V t ( ϕ ) = E C t e − V 0 ( ϕ + ζ ) . (9)
Here E Σ \mathbb{E}_{\Sigma} E Σ averages over a centred Gaussian vector ζ \zeta ζ with covariance Σ \Sigma Σ , while ϕ \phi ϕ is fixed. The heat kernel spreads over a distance of order 1 + t \sqrt{1+t} 1 + t in lattice units, which will be our length scale. At fixed cutoff the potentials are smooth with bounded derivatives. The convolution solves the Polchinski equation [21 (18)] , in the normalization of [4 Proposition 3.4] ,
∂ t V t = 1 2 Tr ( C ˙ t Hess V t ) − 1 2 ( ∇ V t , C ˙ t ∇ V t ) . (10) \partial_t V_t = \frac{1}{2}\operatorname{Tr}(\dot{C}_t \operatorname{Hess} V_t) - \frac{1}{2}(\nabla V_t,\dot{C}_t\nabla V_t). \tag*{(10)} ∂ t V t = 2 1 Tr ( C ˙ t Hess V t ) − 2 1 ( ∇ V t , C ˙ t ∇ V t ) . (10)
We iterate the Gaussian formula as the finite-time evolution of this equation. The Hessian which enters the logarithmic Sobolev criterion is weighted by C ˙ t \dot{C}_t C ˙ t . Since ( u , v ) C ˙ t = ( u , C ˙ t v ) = ( Q t u , Q t v ) (u,v)_{\dot{C}_t}=(u,\dot{C}_t v)=(Q_tu,Q_tv) ( u , v ) C ˙ t = ( u , C ˙ t v ) = ( Q t u , Q t v ) , as in [3 (1.2), (1.11)] , we can express it in the ordinary Euclidean coordinates as Hess [ V t ( Q t ϕ ) ] = Q t Hess V t ( Q t ϕ ) Q t \operatorname{Hess}[V_t(Q_t\phi)]=Q_t\operatorname{Hess}V_t(Q_t\phi)Q_t Hess [ V t ( Q t ϕ )] = Q t Hess V t ( Q t ϕ ) Q t . The same choice lets us repeat Gaussian integration without changing its form. Indeed,
Q t C r Q t = ∫ 0 r e − ( t + s ) A d s = C t + r − C t , Q t Q r = Q t + r . Q_t C_r Q_t = \int_0^r e^{-(t+s)A}\,\mathrm{d}s = C_{t+r}-C_t, \qquad Q_t Q_r = Q_{t+r}. Q t C r Q t = ∫ 0 r e − ( t + s ) A d s = C t + r − C t , Q t Q r = Q t + r .
Adding these independent Gaussian fluctuations therefore gives
e − V t + r ( Q t ϕ ) = E C r e − V t ( Q t ( Q r ϕ + ζ ) ) . (11) e^{-V_{t+r}(Q_t\phi)} = \mathbb{E}_{C_r}e^{-V_t(Q_t(Q_r\phi+\zeta))}. \tag*{(11)} e − V t + r ( Q t ϕ ) = E C r e − V t ( Q t ( Q r ϕ + ζ )) . (11)
We will need the following estimates up to a fixed physical length. Write ( D w ) x , k = w x + e k − w x (Dw)_{x,k}=w_{x+e_k}-w_x ( D w ) x , k = w x + e k − w x , so that D ∗ D = − Δ Λ D^*D=-\Delta_\Lambda D ∗ D = − Δ Λ , and osc f = sup f − inf f \operatorname{osc} f=\sup f-\inf f osc f = sup f − inf f .
Proposition 2.1. For each 0 < β < 8 π 0<\beta<8\pi 0 < β < 8 π there are c β , C β > 0 c_\beta,C_\beta>0 c β , C β > 0 with the following property. Let ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 , t = ℓ 2 − 1 t=\ell^2-1 t = ℓ 2 − 1 , and suppose
ε ℓ ≤ c β min { m − 1 , L } , ∣ z ∣ ( ε ℓ ) α ≤ c β . \varepsilon\ell\le c_\beta\min\{m^{-1},L\},\qquad|z|(\varepsilon\ell)^\alpha\le c_\beta. ε ℓ ≤ c β min { m − 1 , L } , ∣ z ∣ ( ε ℓ ) α ≤ c β . Then
∥ Hess [ V t ( Q t ⋅ ) ] ∥ ∞ , op ≤ C β ∣ z ∣ ε 2 ( ε ℓ ) − β / ( 4 π ) . (12) \lVert\operatorname{Hess}[V_t(Q_t\cdot)]\rVert_{\infty,\operatorname{op}}\le C_\beta|z|\varepsilon^2(\varepsilon\ell)^{-\beta/(4\pi)}. \tag*{(12)} ∥ Hess [ V t ( Q t ⋅ )] ∥ ∞ , op ≤ C β ∣ z ∣ ε 2 ( ε ℓ ) − β / ( 4 π ) . (12) Moreover, V t ( Q t φ ) = N ( φ ) + R ( φ ) V_t(Q_t\varphi)=N(\varphi)+R(\varphi) V t ( Q t φ ) = N ( φ ) + R ( φ ) , where N N N is invariant under constant shifts and, for all φ , w \varphi,w φ , w ,
∣ ( w , Hess N ( φ ) w ) ∣ ≤ C β ∣ z ∣ ( ε ℓ ) α ∥ D w ∥ 2 2 , (13) \left|(w,\operatorname{Hess}N(\varphi)w)\right|\le C_\beta|z|(\varepsilon\ell)^\alpha\lVert Dw\rVert_2^2, \tag*{(13)} ∣ ( w , Hess N ( φ ) w ) ∣ ≤ C β ∣ z ∣ ( ε ℓ ) α ∥ D w ∥ 2 2 , (13) osc R ≤ C β ∣ z ∣ L 2 ( ε ℓ ) − β / ( 4 π ) . (14) \operatorname{osc}R\le C_\beta|z|L^2(\varepsilon\ell)^{-\beta/(4\pi)}. \tag*{(14)} osc R ≤ C β ∣ z ∣ L 2 ( ε ℓ ) − β / ( 4 π ) . (14) The constants are independent of ℓ , ε , L , m , z \ell,\varepsilon,L,m,z ℓ , ε , L , m , z .
For small ∣ z ∣ ( ε ℓ ) α |z|(\varepsilon\ell)^\alpha ∣ z ∣ ( ε ℓ ) α , the bound (13) lets us absorb the neutral curvature into the gradient energy. The bounded part is then treated by (14) . In a large torus its oscillation grows with the volume. Under the small-coupling assumption we use (12) instead, to keep the whole Hamiltonian convex.
Proof of Theorem 1.1 assuming Proposition 2.1. For z = 0 z=0 z = 0 , the assertion is the Gaussian inequality [14 ] . For z ≠ 0 z\ne0 z = 0 , we use the finite-time transport criterion of [4 Theorem 5.4 and (5.26)] , following [22 ] . If μ t \mu_t μ t is the measure with density proportional to e − 1 2 ( φ , A φ ) − V t ( Q t φ ) e^{-\frac{1}{2}(\varphi,A\varphi)-V_t(Q_t\varphi)} e − 2 1 ( φ , A φ ) − V t ( Q t φ ) and γ ( μ t ) \gamma(\mu_t) γ ( μ t ) is its Euclidean logarithmic Sobolev constant, the criterion gives
γ ε , L ≥ ε − 2 γ ( μ t ) exp [ − ∫ 0 t ∥ Hess [ V s ( Q s ⋅ ) ] ∥ ∞ , op d s ] . (15) \gamma_{\varepsilon,L}\ge\varepsilon^{-2}\gamma(\mu_t)\exp\left[-\int_0^t\lVert\operatorname{Hess}[V_s(Q_s\cdot)]\rVert_{\infty,\operatorname{op}}\,ds\right]. \tag*{(15)} γ ε , L ≥ ε − 2 γ ( μ t ) exp [ − ∫ 0 t ∥ Hess [ V s ( Q s ⋅ )] ∥ ∞ , op d s ] . (15)
These are the rescaled measures in [4 (5.28)–(5.31)] . The factor ε − 2 \varepsilon^{-2} ε − 2 converts the Euclidean inequality to (3) .
Choose a physical length r ∗ > 0 r_*>0 r ∗ > 0 such that
r ∗ ≤ c β min { m − 1 , L } , ∣ z ∣ r ∗ α ≤ c β , r_*\le c_\beta\min\{m^{-1},L\},\qquad|z|r_*^\alpha\le c_\beta, r ∗ ≤ c β min { m − 1 , L } , ∣ z ∣ r ∗ α ≤ c β ,
reducing c β c_\beta c β if necessary. This choice is independent of the cutoff. Set
t ∗ = max { ( r ∗ / ε ) 2 − 1 , 0 } . t_*=\max\{(r_*/\varepsilon)^2-1,0\}. t ∗ = max {( r ∗ / ε ) 2 − 1 , 0 } .
For ε ≤ r ∗ \varepsilon\le r_* ε ≤ r ∗ , the assumptions of Proposition 2.1 hold up to t ∗ t_* t ∗ , since both the physical length and its coupling increase with time. Use its decomposition V t ∗ ( Q t ∗ ⋅ ) = N + R V_{t_*}(Q_{t_*}\cdot)=N+R V t ∗ ( Q t ∗ ⋅ ) = N + R . The loss in the transfer is
∫ 0 t ∗ ∥ Hess [ V s ( Q s ⋅ ) ] ∥ ∞ , op d s ≤ C β ∣ z ∣ ε ∫ 0 t ∗ ( 1 + s ) − β / ( 8 π ) d s = 2 C β α ∣ z ∣ ε α ( ( 1 + t ∗ ) α / 2 − 1 ) ≤ C β ∣ z ∣ r ∗ α . (16) \begin{aligned}
\int_0^{t_*}\lVert\operatorname{Hess}[V_s(Q_s\cdot)]\rVert_{\infty,\operatorname{op}}\,ds\le C_\beta|z|\varepsilon\int_0^{t_*}(1+s)^{-\beta/(8\pi)}\,ds \\
&=\frac{2C_\beta}{\alpha}|z|\varepsilon^\alpha\left((1+t_*)^{\alpha/2}-1\right)\le C_\beta|z|r_*^\alpha. \tag*{(16)}
\end{aligned} ∫ 0 t ∗ ∥ Hess [ V s ( Q s ⋅ )] ∥ ∞ , op d s ≤ C β ∣ z ∣ ε ∫ 0 t ∗ ( 1 + s ) − β / ( 8 π ) d s = α 2 C β ∣ z ∣ ε α ( ( 1 + t ∗ ) α /2 − 1 ) ≤ C β ∣ z ∣ r ∗ α . (16)
If ε > r ∗ \varepsilon>r_* ε > r ∗ , the integral is zero. In this case put N = 0 N=0 N = 0 and R = V 0 R=V_0 R = V 0 . The bare estimates give
osc V 0 ≤ 4 ∣ z ∣ L 2 ε − β / ( 4 π ) ≤ 4 ∣ z ∣ L 2 r ∗ − β / ( 4 π ) , \operatorname{osc}V_0\le4|z|L^2\varepsilon^{-\beta/(4\pi)}\le4|z|L^2r_*^{-\beta/(4\pi)}, osc V 0 ≤ 4∣ z ∣ L 2 ε − β / ( 4 π ) ≤ 4∣ z ∣ L 2 r ∗ − β / ( 4 π ) ,
∥ Hess V 0 ∥ ∞ , op ≤ 2 β ∣ z ∣ ε 2 r ∗ − β / ( 4 π ) . \lVert\operatorname{Hess}V_0\rVert_{\infty,\operatorname{op}}\le2\beta|z|\varepsilon^2r_*^{-\beta/(4\pi)}. ∥ Hess V 0 ∥ ∞ , op ≤ 2 β ∣ z ∣ ε 2 r ∗ − β / ( 4 π ) .
Thus, at every cutoff, V t ∗ ( Q t ∗ ⋅ ) = N + R V_{t_*}(Q_{t_*}\cdot)=N+R V t ∗ ( Q t ∗ ⋅ ) = N + R with the same stopping estimates. For the part containing N N N , the neutral form bound gives
A + Hess N ≥ ( 1 − C β ∣ z ∣ r ∗ α ) D ∗ D + ε 2 m 2 id ≥ ε 2 m 2 id , A+\operatorname{Hess}N\geq(1-C_{\beta}|z|^{r_*^{\alpha}})D^*D+\varepsilon^2m^2\operatorname{id}\geq\varepsilon^2m^2\operatorname{id}, A + Hess N ≥ ( 1 − C β ∣ z ∣ r ∗ α ) D ∗ D + ε 2 m 2 id ≥ ε 2 m 2 id ,
and the oscillation and full Hessian bounds give
osc R ≤ C β ∣ z ∣ L 2 r ∗ − β / ( 4 π ) , ∥ Hess [ V t ∗ ( Q t ∗ ⋅ ) ] ∥ ∞ , op ≤ C β ∣ z ∣ ε 2 r ∗ − β / ( 4 π ) . (17) \operatorname{osc}R\leq C_{\beta}|z|L^2r_*^{-\beta/(4\pi)},
\qquad
\left\|\operatorname{Hess}[V_{t_*}(Q_{t_*}\cdot)]\right\|_{\infty,\operatorname{op}}
\leq C_{\beta}|z|\varepsilon^2r_*^{-\beta/(4\pi)}. \tag*{(17)} osc R ≤ C β ∣ z ∣ L 2 r ∗ − β / ( 4 π ) , ∥ Hess [ V t ∗ ( Q t ∗ ⋅ )] ∥ ∞ , op ≤ C β ∣ z ∣ ε 2 r ∗ − β / ( 4 π ) . (17)
Bakry–Émery [4 Theorem 2.8] and the Holley–Stroock perturbation bound [2 Proposition 5.1.6] consequently give γ ( μ t ∗ ) ≥ ε 2 m 2 e − osc R \gamma(\mu_{t_*})\geq\varepsilon^2m^2e^{-\operatorname{osc}R} γ ( μ t ∗ ) ≥ ε 2 m 2 e − osc R . Together with (15) , this yields
γ ε , L ≥ m 2 exp [ − C β ∣ z ∣ ( r ∗ α + L 2 r ∗ − β / ( 4 π ) ) ] > 0. \gamma_{\varepsilon,L}\geq m^2\exp\left[-C_{\beta}|z|\left(r_*^{\alpha}+L^2r_*^{-\beta/(4\pi)}\right)\right]>0. γ ε , L ≥ m 2 exp [ − C β ∣ z ∣ ( r ∗ α + L 2 r ∗ − β / ( 4 π ) ) ] > 0.
For the volume-uniform assertion, first suppose L m ≥ 1 Lm\geq1 L m ≥ 1 and take r ∗ = c 1 / m r_*=c_1/m r ∗ = c 1 / m , with c 1 > 0 c_1>0 c 1 > 0 fixed and small. The assumption ∣ z ∣ m − α ≤ c β |z|m^{-\alpha}\leq c_{\beta} ∣ z ∣ m − α ≤ c β makes this choice admissible. The full Hessian bound in (17) now gives
A + Hess [ V t ∗ ( Q t ∗ ⋅ ) ] ≥ ε 2 m 2 ( 1 − C β ∣ z ∣ m − α ) id . A+\operatorname{Hess}[V_{t_*}(Q_{t_*}\cdot)]\geq\varepsilon^2m^2(1-C_{\beta}|z|m^{-\alpha})\operatorname{id}. A + Hess [ V t ∗ ( Q t ∗ ⋅ )] ≥ ε 2 m 2 ( 1 − C β ∣ z ∣ m − α ) id .
Bakry–Émery applies to the whole measure μ t ∗ \mu_{t_*} μ t ∗ , and (15) gives
γ ε , L ≥ m 2 ( 1 − C β ∣ z ∣ m − α ) e − C β ∣ z ∣ m − α ≥ m 2 e − C β ∣ z ∣ m − α , \gamma_{\varepsilon,L}\geq m^2(1-C_{\beta}|z|m^{-\alpha})e^{-C_{\beta}|z|m^{-\alpha}}\geq m^2e^{-C_{\beta}|z|m^{-\alpha}}, γ ε , L ≥ m 2 ( 1 − C β ∣ z ∣ m − α ) e − C β ∣ z ∣ m − α ≥ m 2 e − C β ∣ z ∣ m − α ,
after changing c β , C β c_{\beta},C_{\beta} c β , C β . If L m < 1 Lm<1 L m < 1 , take r ∗ = c 1 L r_*=c_1L r ∗ = c 1 L . The fixed-volume exponent now satisfies
r ∗ α + L 2 r ∗ − β / ( 4 π ) = ( c 1 α + c 1 − β / ( 4 π ) ) L α ≤ C β m − α , r_*^{\alpha}+L^2r_*^{-\beta/(4\pi)}
=(c_1^{\alpha}+c_1^{-\beta/(4\pi)})L^{\alpha}\leq C_{\beta}m^{-\alpha}, r ∗ α + L 2 r ∗ − β / ( 4 π ) = ( c 1 α + c 1 − β / ( 4 π ) ) L α ≤ C β m − α ,
which gives the same bound. Finally e − u ≥ 1 − u e^{-u}\geq1-u e − u ≥ 1 − u yields (6) .
Estimates for the renormalized potential
We now prove Proposition 2.1 by repeating the Gaussian integration (11) . The full Hessian bound will ensure that each conditional Hamiltonian remains uniformly convex.
The decomposition of the potential
Fix two length scales ℓ ≥ 1 \ell\geq1 ℓ ≥ 1 and ℓ ′ = M ℓ \ell'=M\ell ℓ ′ = M ℓ , with M ≥ 2 M\geq2 M ≥ 2 , and put
t = ℓ 2 − 1 , T = ℓ ′ 2 − ℓ 2 , Q t C T Q t = C t + T − C t , Q t Q T = Q t + T . (18) t=\ell^2-1,\qquad T=\ell'^2-\ell^2,\qquad Q_tC_TQ_t=C_{t+T}-C_t,\qquad Q_tQ_T=Q_{t+T}. \tag*{(18)} t = ℓ 2 − 1 , T = ℓ ′2 − ℓ 2 , Q t C T Q t = C t + T − C t , Q t Q T = Q t + T . (18)
Assume ε ℓ ′ ≤ c 0 min { m − 1 , L } \varepsilon\ell'\leq c_0\min\{m^{-1},L\} ε ℓ ′ ≤ c 0 min { m − 1 , L } , the heat regime of Lemma A.1, and write d ( i , j ) d(i,j) d ( i , j ) for periodic nearest-neighbour distance. Fix μ > 0 \mu>0 μ > 0 , depending only on β \beta β , so that max { 32 , 24 / α } μ ≤ ρ 0 \max\{32,24/\alpha\}\mu\leq\rho_0 max { 32 , 24/ α } μ ≤ ρ 0 .
At length ℓ \ell ℓ , write
V t ( Q t φ ) = N ( φ ) + R ( φ ) , R = ℓ − 2 ∑ x g x . (19) V_t(Q_t\varphi)=N(\varphi)+R(\varphi),\qquad R=\ell^{-2}\sum_xg_x. \tag*{(19)} V t ( Q t φ ) = N ( φ ) + R ( φ ) , R = ℓ − 2 x ∑ g x . (19)
Initially N = 0 N=0 N = 0 and g x = 2 z ε α cos ( β φ x ) g_x=2z\varepsilon^\alpha\cos(\sqrt{\beta}\varphi_x) g x = 2 z ε α cos ( β φ x ) . The function g x g_x g x will change under integration, while the index x x x continues to refer to this original cosine. Its smoothed phase is β ( Q t φ ) x \sqrt{\beta}(Q_t\varphi)_x β ( Q t φ ) x , which changes under a common field shift at rate
c t = β ∑ i Q t ( i , x ) = β e − ε 2 m 2 t / 2 , period = 2 π / c t . c_t=\sqrt{\beta}\sum_iQ_t(i,x)=\sqrt{\beta}e^{-\varepsilon^2m^2t/2},
\qquad\text{period}=2\pi/c_t. c t = β i ∑ Q t ( i , x ) = β e − ε 2 m 2 t /2 , period = 2 π / c t .
We require N N N to be neutral and each real g x g_x g x to have this period and zero period average. The logarithm couples the local terms. To retain the sum over x x x , we differentiate the logarithm with respect to a parameter multiplying R R R . For 0 ≤ λ ≤ 1 0 \leq\lambda\leq1 0 ≤ λ ≤ 1 , weight the Gaussian law of Q T φ ′ + ζ Q_T\varphi' + \zeta Q T φ ′ + ζ , with covariance C T C_T C T , by e − ( N + λ R ) e^{-(N+\lambda R)} e − ( N + λ R ) and normalize it. Denote this expectation by E λ \mathbf{E}_\lambda E λ and put
G x ( φ ′ ) = M 2 ∫ 0 1 E λ g x d λ . G_x(\varphi') = M^2 \int_0^1 \mathbf{E}_\lambda g_x\,d\lambda. G x ( φ ′ ) = M 2 ∫ 0 1 E λ g x d λ .
Here M 2 = ℓ ′ 2 / ℓ 2 M^2 = \ell'^2/\ell^2 M 2 = ℓ ′2 / ℓ 2 changes the area normalization, and integration in λ \lambda λ gives the exact decomposition,
− log E C T e − ( N + R ) ( Q T φ ′ + ζ ) + log E C T e − N ( Q T φ ′ + ζ ) = ∫ 0 1 E λ R d λ = ℓ − 2 ∑ x ∫ 0 1 E λ g x d λ = ℓ ′ − 2 ∑ x G x . (20) -\log\mathbf{E}_{C_T}e^{-(N+R)(Q_T\varphi'+\zeta)}+\log\mathbf{E}_{C_T}e^{-N(Q_T\varphi'+\zeta)}
= \int_0^1 \mathbf{E}_\lambda R\,d\lambda= \ell^{-2}\sum_x\int_0^1 \mathbf{E}_\lambda g_x\,d\lambda= \ell'^{-2}\sum_x G_x. \tag*{(20)} − log E C T e − ( N + R ) ( Q T φ ′ + ζ ) + log E C T e − N ( Q T φ ′ + ζ ) = ∫ 0 1 E λ R d λ = ℓ − 2 x ∑ ∫ 0 1 E λ g x d λ = ℓ ′ − 2 x ∑ G x . (20)
The periods match because
Q T ( φ ′ + 2 π c t + T − 1 1 ) = Q T φ ′ + 2 π c t − 1 1. Q_T(\varphi' + 2\pi c_{t+T}^{-1}\mathbf{1}) = Q_T\varphi' + 2\pi c_t^{-1}\mathbf{1}. Q T ( φ ′ + 2 π c t + T − 1 1 ) = Q T φ ′ + 2 π c t − 1 1 .
Thus G x G_x G x has period 2 π / c t + T 2\pi/c_{t+T} 2 π / c t + T . Its period average, which is invariant under every constant shift, is
Π f ( φ ′ ) = 1 2 π ∫ 0 2 π f ( φ ′ + θ c t + T − 1 1 ) d θ . \Pi f(\varphi') = \frac{1}{2\pi}\int_0^{2\pi} f(\varphi' + \theta c_{t+T}^{-1}\mathbf{1})\,d\theta. Π f ( φ ′ ) = 2 π 1 ∫ 0 2 π f ( φ ′ + θ c t + T − 1 1 ) d θ .
We put this average into the neutral interaction and retain the zero-average part in the local terms,
N ′ ( φ ′ ) = − log E C T e − N ( Q T φ ′ + ζ ) + ℓ ′ − 2 ∑ x Π G x ( φ ′ ) , (21) N'(\varphi') = -\log\mathbf{E}_{C_T}e^{-N(Q_T\varphi'+\zeta)}+\ell'^{-2}\sum_x\Pi G_x(\varphi'), \tag*{(21)} N ′ ( φ ′ ) = − log E C T e − N ( Q T φ ′ + ζ ) + ℓ ′ − 2 x ∑ Π G x ( φ ′ ) , (21)
g x ′ ( φ ′ ) = G x ( φ ′ ) − Π G x ( φ ′ ) . g'_x(\varphi') = G_x(\varphi')-\Pi G_x(\varphi'). g x ′ ( φ ′ ) = G x ( φ ′ ) − Π G x ( φ ′ ) .
Then (11) and (20) give V t + T ( Q t + T φ ′ ) = N ′ ( φ ′ ) + ℓ ′ − 2 ∑ x g x ′ ( φ ′ ) V_{t+T}(Q_{t+T}\varphi')=N'(\varphi')+\ell'^{-2}\sum_x g'_x(\varphi') V t + T ( Q t + T φ ′ ) = N ′ ( φ ′ ) + ℓ ′ − 2 ∑ x g x ′ ( φ ′ ) . The first term in N ′ N' N ′ is neutral, and Π g x ′ = 0 \Pi g'_x=0 Π g x ′ = 0 . Hence N ′ = Π [ V t + T ( Q t + T ⋅ ) ] N'=\Pi[V_{t+T}(Q_{t+T}\cdot)] N ′ = Π [ V t + T ( Q t + T ⋅ )] , an identity we shall use to bound its derivatives by those of the full potential.
The Gaussian average of the original cosine determines the leading local coefficient. We keep this coefficient exact and use the dimensionless coupling to bound the nonlinear terms,
z ℓ = z ε α e − β C t ( 0 , 0 ) / 2 , ζ ℓ = ∣ z ∣ ε α , ℓ 2 ∣ z ℓ ∣ ≤ C β ζ ℓ . (22) z_\ell= z\varepsilon^\alpha e^{-\beta C_t(0,0)/2},\qquad\zeta_\ell= |z|\varepsilon^\alpha,\qquad\ell^2|z_\ell|\leq C_\beta\zeta_\ell. \tag*{(22)} z ℓ = z ε α e − β C t ( 0 , 0 ) /2 , ζ ℓ = ∣ z ∣ ε α , ℓ 2 ∣ z ℓ ∣ ≤ C β ζ ℓ . (22)
The inequality follows from (A.5), and the exact Gaussian average is
E C t [ 2 z ε α cos ( β ( Q t φ + ζ ) x ) ] = 2 z ε α e − β C t ( 0 , 0 ) / 2 cos ( β ( Q t φ ) x ) . \mathbf{E}_{C_t}\left[2z\varepsilon^\alpha\cos\left(\sqrt{\beta}(Q_t\varphi+\zeta)_x\right)\right]
=2z\varepsilon^\alpha e^{-\beta C_t(0,0)/2}\cos\left(\sqrt{\beta}(Q_t\varphi)_x\right). E C t [ 2 z ε α cos ( β ( Q t φ + ζ ) x ) ] = 2 z ε α e − β C t ( 0 , 0 ) /2 cos ( β ( Q t φ ) x ) .
The factor ℓ − 2 \ell^{-2} ℓ − 2 in (19) makes the leading local coefficient 2 ℓ 2 z ℓ 2\ell^2 z_\ell 2 ℓ 2 z ℓ , of size ζ ℓ \zeta_\ell ζ ℓ . To measure the dependence of g x g_x g x on variables other than its original phase, subtract from a coordinate derivative the common shift producing the same phase change,
Θ t = c t − 1 ∑ j ∂ j , ∂ i ( x ) = ∂ i − β Q t ( i , x ) Θ t . (23) \Theta_t=c_t^{-1}\sum_j\partial_j,\qquad\partial_i^{(x)}=\partial_i-\sqrt{\beta}Q_t(i,x)\Theta_t. \tag*{(23)} Θ t = c t − 1 j ∑ ∂ j , ∂ i ( x ) = ∂ i − β Q t ( i , x ) Θ t . (23)
Indeed, Θ t cos ( β ( Q t φ ) x ) = − sin ( β ( Q t φ ) x ) \Theta_t\cos(\sqrt{\beta}(Q_t\varphi)_x)=-\sin(\sqrt{\beta}(Q_t\varphi)_x) Θ t cos ( β ( Q t φ ) x ) = − sin ( β ( Q t φ ) x ) , whereas ∂ i ( x ) cos ( β ( Q t φ ) x ) = 0 \partial_i^{(x)}\cos(\sqrt{\beta}(Q_t\varphi)_x)=0 ∂ i ( x ) cos ( β ( Q t φ ) x ) = 0 . Thus Θ t \Theta_t Θ t differentiates the angle, whereas ∂ i ( x ) \partial_i^{(x)} ∂ i ( x ) holds the phase fixed and detects dependence on field differences.
For g g g with period 2 π / c t 2\pi/c_t 2 π / c t under common shifts, θ ↦ g ( φ + θ c t − 1 1 ) \theta\mapsto g(\varphi+\theta c_t^{-1}\mathbf{1}) θ ↦ g ( φ + θ c t − 1 1 ) is 2 π 2\pi 2 π -periodic at each fixed φ \varphi φ . Thus e a Θ t 2 e^{a\Theta_t^2} e a Θ t 2 is the heat semigroup on the circle in this angle, with auxiliary time a a a . On functions of zero shift average, the supremum-norm estimate is ∥ e a Θ t 2 g ∥ ∞ ≤ C e − a ∥ g ∥ ∞ \lVert e^{a\Theta_t^2}g\rVert_\infty\le Ce^{-a}\lVert g\rVert_\infty ∥ e a Θ t 2 g ∥ ∞ ≤ C e − a ∥ g ∥ ∞ . For the induction we need decay with coefficient one, to retain the exact Gaussian cosine coefficient while allowing only quadratic errors. We therefore use the standard equivalent-norm construction for bounded semigroups [12 Lemma II.3.10] , applied to e a Θ t 2 e^{a\Theta_t^2} e a Θ t 2 , and set
∥ g ∥ ∗ = sup a ≥ 0 e a ∥ e a Θ t 2 g ∥ ∞ , e a Θ t 2 g ( φ ) = 1 4 π a ∫ R e − θ 2 / ( 4 a ) g ( φ + θ c t − 1 1 ) d θ , a > 0. (24) \lVert g\rVert_*=\sup_{a\ge0}e^a\lVert e^{a\Theta_t^2}g\rVert_\infty,
\qquad
e^{a\Theta_t^2}g(\varphi)=\frac{1}{\sqrt{4\pi a}}\int_{\mathbb{R}}e^{-\theta^2/(4a)}g(\varphi+\theta c_t^{-1}\mathbf{1})\,d\theta,\quad a>0. \tag*{(24)} ∥ g ∥ ∗ = a ≥ 0 sup e a ∥ e a Θ t 2 g ∥ ∞ , e a Θ t 2 g ( φ ) = 4 π a 1 ∫ R e − θ 2 / ( 4 a ) g ( φ + θ c t − 1 1 ) d θ , a > 0. (24)
For a ≤ 1 a\le1 a ≤ 1 use supremum-norm contraction, and for a ≥ 1 a\ge1 a ≥ 1 use ∑ n ≠ 0 e − n 2 a ≤ C e − a \sum_{n\ne0}e^{-n^2a}\le Ce^{-a} ∑ n = 0 e − n 2 a ≤ C e − a . These bounds give norm equivalence, while the semigroup property gives the required decay,
∥ g ∥ ∞ ≤ ∥ g ∥ ∗ ≤ C ∥ g ∥ ∞ , ∥ e b Θ t 2 g ∥ ∗ = e − b sup a ≥ b e a ∥ e a Θ t 2 g ∥ ∞ ≤ e − b ∥ g ∥ ∗ . \lVert g\rVert_\infty\le\lVert g\rVert_*\le C\lVert g\rVert_\infty,\qquad
\lVert e^{b\Theta_t^2}g\rVert_* = e^{-b}\sup_{a\ge b}e^a\lVert e^{a\Theta_t^2}g\rVert_\infty\le e^{-b}\lVert g\rVert_*. ∥ g ∥ ∞ ≤ ∥ g ∥ ∗ ≤ C ∥ g ∥ ∞ , ∥ e b Θ t 2 g ∥ ∗ = e − b a ≥ b sup e a ∥ e a Θ t 2 g ∥ ∞ ≤ e − b ∥ g ∥ ∗ .
In particular, ∥ cos ( β ( Q t φ ) x ) ∥ ∗ = 1 \lVert\cos(\sqrt{\beta}(Q_t\varphi)_x)\rVert_*=1 ∥ cos ( β ( Q t φ ) x ) ∥ ∗ = 1 . Constant-coefficient field derivatives commute with this semigroup, so the same norm applies to the derivatives of g x g_x g x . For functions of φ ′ \varphi' φ ′ , both derivatives and the norm use t + T t+T t + T in place of t t t . Inside E λ \mathbf{E}_\lambda E λ they use t t t .
We ask for linear bounds on the full Hessian and the derivative at x x x . The local norm may exceed the Gaussian cosine coefficient by a quadratic term, and the phase-fixed derivative is itself quadratic,
sup i ∑ j e μ d ( i , j ) / ( 2 ℓ ) sup φ ∣ ∂ i ∂ j ( N + R ) ∣ ≤ H ζ ℓ − 2 . (25) \sup_i\sum_j e^{\mu d(i,j)/(2\ell)}\sup_\varphi\left|\partial_i\partial_j(N+R)\right|\le H\zeta\ell^{-2}. \tag*{(25)} i sup j ∑ e μ d ( i , j ) / ( 2 ℓ ) φ sup ∣ ∂ i ∂ j ( N + R ) ∣ ≤ H ζ ℓ − 2 . (25)
and, at every field,
∥ g x ∥ ∗ ≤ 2 ℓ 2 ∣ z ℓ ∣ + B ζ 2 ℓ , ∥ ∂ i ( x ) g x ∥ ∗ ≤ B ζ 2 ℓ − 2 e − 6 μ d ( i , x ) / ℓ , ∣ ∂ x g x ∣ ≤ K ζ ℓ − 2 . (26) \begin{aligned}
\lVert g_x\rVert_*&\le2\ell^2|z_\ell|+B\zeta^2\ell,\\
\lVert\partial_i^{(x)}g_x\rVert_*&\le B\zeta^2\ell^{-2}e^{-6\mu d(i,x)/\ell},\\
|\partial_xg_x|&\le K\zeta\ell^{-2}. \tag*{(26)}
\end{aligned} ∥ g x ∥ ∗ ∥ ∂ i ( x ) g x ∥ ∗ ∣ ∂ x g x ∣ ≤ 2 ℓ 2 ∣ z ℓ ∣ + B ζ 2 ℓ , ≤ B ζ 2 ℓ − 2 e − 6 μ d ( i , x ) / ℓ , ≤ K ζ ℓ − 2 . (26)
We shall fix K , H ≥ 1 K,H\ge1 K , H ≥ 1 before B B B . Constants denoted by C C C depend only on β , μ \beta,\mu β , μ . Constants C M C_M C M may also depend on M M M and the fixed K , H K,H K , H , but not on B B B . We write C M , K C_{M,K} C M , K when independence of H H H is needed. All are independent of cutoff, volume and field.
Proposition 3.1. For each 0 < β < 8 π 0<\beta<8\pi 0 < β < 8 π , one can choose M ≥ 2 M\ge2 M ≥ 2 , K , H , B ≥ 1 K,H,B\ge1 K , H , B ≥ 1 and ζ ∗ > 0 \zeta_*>0 ζ ∗ > 0 , depending only on β \beta β , so that Gaussian integration (21) preserves the full Hessian bound (25) and the local value and first-derivative bounds (26) , with the same constants H , K , B H,K,B H , K , B , from length ℓ \ell ℓ to ℓ ′ = M ℓ \ell'=M\ell ℓ ′ = M ℓ . This holds for real C 2 C^2 C 2 functions N , g x N,g_x N , g x , with N N N invariant under constant shifts and each g x g_x g x periodic under common shifts with period 2 π / c t 2\pi/c_t 2 π / c t and zero period average, provided
ε ℓ ′ ≤ c 0 min { m − 1 , L } , ζ ℓ ′ ≤ ζ ∗ . \varepsilon\ell'\le c_0\min\{m^{-1},L\},\qquad\zeta\ell'\le\zeta_*. ε ℓ ′ ≤ c 0 min { m − 1 , L } , ζ ℓ ′ ≤ ζ ∗ . At the new scale, the derivatives and norm use t + T = ℓ ′ 2 − 1 t+T=\ell'^2-1 t + T = ℓ ′2 − 1 .
The proof is completed in Subsection 3.7, where we choose the constants. We first note that the derivative at x x x controls the angular derivative, since Q t ( x , x ) ≥ c β ℓ − 2 Q_t(x,x)\ge c_\beta\ell^{-2} Q t ( x , x ) ≥ c β ℓ − 2 . Indeed, by (23) ,
Θ t g x = ∂ x g x − ∂ x ( x ) g x β Q t ( x , x ) , ∥ Θ t g x ∥ ∞ ≤ C β ( K + 1 ) ζ ℓ . \Theta_tg_x=\frac{\partial_xg_x-\partial_x^{(x)}g_x}{\sqrt{\beta Q_t(x,x)}},\qquad\lVert\Theta_tg_x\rVert_\infty\le C_\beta(K+1)\zeta\ell. Θ t g x = β Q t ( x , x ) ∂ x g x − ∂ x ( x ) g x , ∥ Θ t g x ∥ ∞ ≤ C β ( K + 1 ) ζ ℓ .
under B ζ ℓ ≤ 1 B\zeta_\ell\le1 B ζ ℓ ≤ 1 . Since ∂ i g x = ∂ i ( x ) g x + β Q t ( i , x ) Θ t g x \partial_i g_x = \partial_i^{(x)} g_x + \sqrt{\beta}Q_t(i,x)\Theta_t g_x ∂ i g x = ∂ i ( x ) g x + β Q t ( i , x ) Θ t g x , the pointwise heat-kernel bound gives
∥ ∂ i g x ∥ ∗ ≤ C ( ∥ ∂ i ( x ) g x ∥ ∞ + β Q t ( i , x ) ∥ Θ t g x ∥ ∞ ) ≤ C β ( K + 1 ) ζ ℓ ℓ − 2 e − 6 μ d ( i , x ) / ℓ . (27) \|\partial_i g_x\|_* \le C\left(\|\partial_i^{(x)}g_x\|_\infty+\sqrt{\beta}Q_t(i,x)\|\Theta_tg_x\|_\infty\right)
\le C_\beta(K+1)\zeta_\ell\ell^{-2}e^{-6\mu d(i,x)/\ell}. \tag*{(27)} ∥ ∂ i g x ∥ ∗ ≤ C ( ∥ ∂ i ( x ) g x ∥ ∞ + β Q t ( i , x ) ∥ Θ t g x ∥ ∞ ) ≤ C β ( K + 1 ) ζ ℓ ℓ − 2 e − 6 μ d ( i , x ) / ℓ . (27)
We used equivalence of the circle and supremum norms only on these derivatives, all of which have zero shift average.
The same Hessian bound applies throughout the interpolation N + λ R N+\lambda R N + λ R . To see this, first use the period average to write
∂ i ∂ j N ( φ ) = 1 2 π ∫ 0 2 π ∂ i ∂ j ( N + R ) ( φ + θ c t − 1 1 ) d θ . \partial_i\partial_jN(\varphi)=\frac{1}{2\pi}\int_0^{2\pi}\partial_i\partial_j(N+R)(\varphi+\theta c_t^{-1}\mathbf{1})\,d\theta. ∂ i ∂ j N ( φ ) = 2 π 1 ∫ 0 2 π ∂ i ∂ j ( N + R ) ( φ + θ c t − 1 1 ) d θ .
Taking the field supremum inside the integral gives the same weighted row bound for N N N . The triangle inequality then gives it for N + λ R = ( 1 − λ ) N + λ ( N + R ) N+\lambda R=(1-\lambda)N+\lambda(N+R) N + λ R = ( 1 − λ ) N + λ ( N + R ) , 0 ≤ λ ≤ 1 0\le\lambda\le1 0 ≤ λ ≤ 1 . In particular, if ρ / ℓ ′ ≤ μ / ( 8 ℓ ) \rho/\ell'\le\mu/(8\ell) ρ / ℓ ′ ≤ μ / ( 8 ℓ ) ,
sup i ∑ j e ρ d ( i , j ) / ℓ ′ sup φ ∣ ∂ i ∂ j ( N + λ R ) ∣ ≤ H ζ ℓ ℓ − 2 . (28) \sup_i\sum_j e^{\rho d(i,j)/\ell'}\sup_\varphi|\partial_i\partial_j(N+\lambda R)|\le H\zeta_\ell\ell^{-2}. \tag*{(28)} i sup j ∑ e ρ d ( i , j ) / ℓ ′ φ sup ∣ ∂ i ∂ j ( N + λ R ) ∣ ≤ H ζ ℓ ℓ − 2 . (28)
Below we impose B ζ ℓ ≤ 1 B\zeta_\ell\le1 B ζ ℓ ≤ 1 and C M H ζ ℓ ≤ 1 / 8 C_MH\zeta_\ell\le1/8 C M H ζ ℓ ≤ 1/8 , fixing C M C_M C M after M M M , K K K , H H H and before B B B to cover the conditional estimates. The final choice of ζ ∗ \zeta_* ζ ∗ will ensure both conditions.
Conditional integration
We shall estimate the covariance terms produced by differentiating a Gaussian convolution. For a positive definite covariance matrix Σ \Sigma Σ , define
T F , Σ f ( φ ) = E Σ [ e − F ( φ + ζ ) f ( φ + ζ ) ] E Σ e − F ( φ + ζ ) . (29) T_{F,\Sigma}f(\varphi)=\frac{\mathbb{E}_{\Sigma}[e^{-F(\varphi+\zeta)}f(\varphi+\zeta)]}{\mathbb{E}_{\Sigma}e^{-F(\varphi+\zeta)}}. \tag*{(29)} T F , Σ f ( φ ) = E Σ e − F ( φ + ζ ) E Σ [ e − F ( φ + ζ ) f ( φ + ζ )] . (29)
For each fixed background φ \varphi φ , this is expectation over ψ = φ + ζ \psi=\varphi+\zeta ψ = φ + ζ with density proportional to e − F ( ψ ) − 1 2 ( ψ − φ , Σ − 1 ( ψ − φ ) ) e^{-F(\psi)-\frac{1}{2}(\psi-\varphi,\Sigma^{-1}(\psi-\varphi))} e − F ( ψ ) − 2 1 ( ψ − φ , Σ − 1 ( ψ − φ )) . We call this the conditional law at φ \varphi φ . Write Cov F , Σ ( f , g ) = T F , Σ ( f g ) − ( T F , Σ f ) ( T F , Σ g ) \operatorname{Cov}_{F,\Sigma}(f,g)=T_{F,\Sigma}(fg)-(T_{F,\Sigma}f)(T_{F,\Sigma}g) Cov F , Σ ( f , g ) = T F , Σ ( f g ) − ( T F , Σ f ) ( T F , Σ g ) . These are the Polchinski operators of [4 Definition 3.2 and Proposition 3.5] . For our scale step, E λ h = T N + λ R , C T h ( Q T φ ) \mathbb{E}_\lambda h=T_{N+\lambda R,C_T}h(Q_T\varphi) E λ h = T N + λ R , C T h ( Q T φ ) . If F r ( φ ) = − log E r Σ e − F ( φ + ζ ) F_r(\varphi)=-\log\mathbb{E}_{r\Sigma}e^{-F(\varphi+\zeta)} F r ( φ ) = − log E r Σ e − F ( φ + ζ ) , Gaussian covariance addition gives T F , Σ = T F r , ( 1 − r ) Σ T F , r Σ T_{F,\Sigma}=T_{F_r,(1-r)\Sigma}T_{F,r\Sigma} T F , Σ = T F r , ( 1 − r ) Σ T F , r Σ .
The background φ \varphi φ occurs only in the Gaussian density when we integrate in ψ \psi ψ . We can therefore differentiate a conditional average without differentiating its observable. For bounded h h h , differentiating the numerator and denominator gives
∂ w T F , Σ h = T F , Σ [ h ( Σ − 1 w , ψ − φ ) ] − ( T F , Σ h ) T F , Σ ( Σ − 1 w , ψ − φ ) = Cov F , Σ ( h , ( Σ − 1 w , ψ ) ) . \partial_wT_{F,\Sigma}h=T_{F,\Sigma}[h(\Sigma^{-1}w,\psi-\varphi)]-(T_{F,\Sigma}h)T_{F,\Sigma}(\Sigma^{-1}w,\psi-\varphi)
=\operatorname{Cov}_{F,\Sigma}(h,(\Sigma^{-1}w,\psi)). ∂ w T F , Σ h = T F , Σ [ h ( Σ − 1 w , ψ − φ )] − ( T F , Σ h ) T F , Σ ( Σ − 1 w , ψ − φ ) = Cov F , Σ ( h , ( Σ − 1 w , ψ )) .
If Hess F ≥ − Σ − 1 / 2 \operatorname{Hess}F\ge-\Sigma^{-1}/2 Hess F ≥ − Σ − 1 /2 , the Hessian of the conditional Hamiltonian is at least Σ − 1 / 2 \Sigma^{-1}/2 Σ − 1 /2 . Brascamp–Lieb [8 ] then gives
Var F , Σ ( ( Σ − 1 w , ψ ) ) ≤ T F , Σ ( Σ − 1 w , [ Σ − 1 + Hess F ] − 1 Σ − 1 w ) ≤ 2 ( w , Σ − 1 w ) . \operatorname{Var}_{F,\Sigma}((\Sigma^{-1}w,\psi))\le T_{F,\Sigma}(\Sigma^{-1}w,[\Sigma^{-1}+\operatorname{Hess}F]^{-1}\Sigma^{-1}w)\le2(w,\Sigma^{-1}w). Var F , Σ (( Σ − 1 w , ψ )) ≤ T F , Σ ( Σ − 1 w , [ Σ − 1 + Hess F ] − 1 Σ − 1 w ) ≤ 2 ( w , Σ − 1 w ) .
Using Var F , Σ ( h ) ≤ ∥ h ∥ ∞ 2 \operatorname{Var}_{F,\Sigma}(h)\le\|h\|_\infty^2 Var F , Σ ( h ) ≤ ∥ h ∥ ∞ 2 in Cauchy–Schwarz proves
∣ ∂ w T F , Σ h ∣ ≤ 2 ∥ h ∥ ∞ ( w , Σ − 1 w ) . (30) |\partial_wT_{F,\Sigma}h|\le\sqrt{2}\|h\|_\infty\sqrt{(w,\Sigma^{-1}w)}. \tag*{(30)} ∣ ∂ w T F , Σ h ∣ ≤ 2 ∥ h ∥ ∞ ( w , Σ − 1 w ) . (30)
The next lemma lets us differentiate a covariance using only first derivatives of its observables. This is what we need when the observables are a local term and a derivative of the interaction. The proof splits the Gaussian integration so that the background derivative falls on the outer density. The convolved-Hessian estimate keeps the corresponding conditional Hamiltonian convex. Related weighted covariance estimates appear in [17 Theorem 3.1 and Remark 3.2] .
Lemma 3.2. Let F ∈ C 2 ( R n ) F \in C^2(\mathbb{R}^n) F ∈ C 2 ( R n ) , Σ > 0 \Sigma> 0 Σ > 0 , and let W 1 , W 2 W_1, W_2 W 1 , W 2 be positive diagonal matrices such that
∥ W i Σ W i − 1 ∥ o p ∥ W i Hess F W i − 1 ∥ ∞ , o p ≤ 1 8 , i = 1 , 2. \lVert W_i\Sigma W_i^{-1}\rVert_{\mathrm{op}}\lVert W_i\operatorname{Hess}F W_i^{-1}\rVert_{\infty,\mathrm{op}} \le\frac{1}{8}, \qquad i=1,2. ∥ W i Σ W i − 1 ∥ op ∥ W i Hess F W i − 1 ∥ ∞ , op ≤ 8 1 , i = 1 , 2. For C 1 C^1 C 1 functions f , g f,g f , g with the indicated bounded gradients, every direction w w w , 0 ≤ r ≤ 1 0 \le r \le1 0 ≤ r ≤ 1 and i = 1 , 2 i=1,2 i = 1 , 2 , uniformly in the background,
∣ Cov F , Σ ( f , g ) ∣ ≤ 2 ∥ W 1 − 1 Σ W 2 − 1 ∥ o p ∥ W 1 ∇ f ∥ ∞ , 2 ∥ W 2 ∇ g ∥ ∞ , 2 , (31) \lvert\operatorname{Cov}_{F,\Sigma}(f,g)\rvert\le2\lVert W_1^{-1}\Sigma W_2^{-1}\rVert_{\mathrm{op}}\lVert W_1\nabla f\rVert_{\infty,2}\lVert W_2\nabla g\rVert_{\infty,2}, \tag*{(31)} ∣ Cov F , Σ ( f , g )∣ ≤ 2 ∥ W 1 − 1 Σ W 2 − 1 ∥ op ∥ W 1 ∇ f ∥ ∞ , 2 ∥ W 2 ∇ g ∥ ∞ , 2 , (31) ∥ W i Hess F r W i − 1 ∥ ∞ , o p ≤ 2 ∥ W i Hess F W i − 1 ∥ ∞ , o p , (32) \lVert W_i\operatorname{Hess}F_rW_i^{-1}\rVert_{\infty,\mathrm{op}} \le2\lVert W_i\operatorname{Hess}F W_i^{-1}\rVert_{\infty,\mathrm{op}}, \tag*{(32)} ∥ W i Hess F r W i − 1 ∥ ∞ , op ≤ 2 ∥ W i Hess F W i − 1 ∥ ∞ , op , (32) ∣ ∂ w Cov F , Σ ( f , g ) ∣ ≤ 8 2 ∥ W 1 − 1 Σ W 2 − 1 ∥ o p ( w , Σ − 1 w ) ∥ W 1 ∇ f ∥ ∞ , 2 ∥ W 2 ∇ g ∥ ∞ , 2 . (33) \lvert\partial_w\operatorname{Cov}_{F,\Sigma}(f,g)\rvert\le8\sqrt{2}\lVert W_1^{-1}\Sigma W_2^{-1}\rVert_{\mathrm{op}}\sqrt{(w,\Sigma^{-1}w)}\lVert W_1\nabla f\rVert_{\infty,2}\lVert W_2\nabla g\rVert_{\infty,2}. \tag*{(33)} ∣ ∂ w Cov F , Σ ( f , g )∣ ≤ 8 2 ∥ W 1 − 1 Σ W 2 − 1 ∥ op ( w , Σ − 1 w ) ∥ W 1 ∇ f ∥ ∞ , 2 ∥ W 2 ∇ g ∥ ∞ , 2 . (33)
Proof. Fix the background and write E \mathbf{E} E for conditional expectation. We first check convexity of the conditional Hamiltonian. For the symmetric matrices Σ \Sigma Σ and Hess F ( ψ ) \operatorname{Hess}F(\psi) Hess F ( ψ ) , the operator norm is the largest modulus of an eigenvalue. Conjugation by W i W_i W i preserves these eigenvalues, so the conjugated norms bound the original ones. The hypothesis therefore gives
∥ Σ ∥ o p ∥ Hess F ∥ ∞ , o p ≤ ∥ W i Σ W i − 1 ∥ o p ∥ W i Hess F W i − 1 ∥ ∞ , o p ≤ 1 8 . \lVert\Sigma\rVert_{\mathrm{op}}\lVert\operatorname{Hess}F\rVert_{\infty,\mathrm{op}} \le\lVert W_i\Sigma W_i^{-1}\rVert_{\mathrm{op}}\lVert W_i\operatorname{Hess}F W_i^{-1}\rVert_{\infty,\mathrm{op}} \le\frac{1}{8}. ∥ Σ ∥ op ∥ Hess F ∥ ∞ , op ≤ ∥ W i Σ W i − 1 ∥ op ∥ W i Hess F W i − 1 ∥ ∞ , op ≤ 8 1 .
In particular, id ≤ ∥ Σ ∥ o p Σ − 1 \operatorname{id} \le\lVert\Sigma\rVert_{\mathrm{op}}\Sigma^{-1} id ≤ ∥ Σ ∥ op Σ − 1 and
Σ − 1 + Hess F ≥ Σ − 1 − ∥ Hess F ∥ ∞ , o p id ≥ ( 1 − ∥ Σ ∥ o p ∥ Hess F ∥ ∞ , o p ) Σ − 1 ≥ 7 8 Σ − 1 . \begin{aligned}
\Sigma^{-1}+\operatorname{Hess}F \ge\Sigma^{-1}-\lVert\operatorname{Hess}F\rVert_{\infty,\mathrm{op}}\operatorname{id} \\
&\ge(1-\lVert\Sigma\rVert_{\mathrm{op}}\lVert\operatorname{Hess}F\rVert_{\infty,\mathrm{op}})\Sigma^{-1} \ge\frac{7}{8}\Sigma^{-1}.
\end{aligned} Σ − 1 + Hess F ≥ Σ − 1 − ∥ Hess F ∥ ∞ , op id ≥ ( 1 − ∥ Σ ∥ op ∥ Hess F ∥ ∞ , op ) Σ − 1 ≥ 8 7 Σ − 1 .
Consider the operator
L = Tr ( Σ ∇ 2 ) − ( ψ − φ + Σ ∇ F , ∇ ) . \mathcal{L}=\operatorname{Tr}(\Sigma\nabla^2)-(\psi-\varphi+\Sigma\nabla F,\nabla). L = Tr ( Σ ∇ 2 ) − ( ψ − φ + Σ∇ F , ∇ ) .
Integration by parts against the conditional density gives, for smooth h , k h,k h , k ,
E [ ( L h ) k ] = ∑ i , j Σ i j E [ k ∂ i ∂ j h − k ( ∂ i F + [ Σ − 1 ( ψ − φ ) ] i ) ∂ j h ] = − ∑ i , j Σ i j E [ ( ∂ i k ) ( ∂ j h ) ] = − E ( ∇ h , Σ ∇ k ) . \begin{aligned}
\mathbf{E}[(\mathcal{L}h)k]=\sum_{i,j}\Sigma_{ij}\mathbf{E}\left[k\partial_i\partial_jh-k\left(\partial_iF+[\Sigma^{-1}(\psi-\varphi)]_i\right)\partial_jh\right] \\
&=-\sum_{i,j}\Sigma_{ij}\mathbf{E}[(\partial_i k)(\partial_j h)]=-\mathbf{E}(\nabla h,\Sigma\nabla k).
\end{aligned} E [( L h ) k ] = i , j ∑ Σ ij E [ k ∂ i ∂ j h − k ( ∂ i F + [ Σ − 1 ( ψ − φ ) ] i ) ∂ j h ] = − i , j ∑ Σ ij E [( ∂ i k ) ( ∂ j h )] = − E ( ∇ h , Σ∇ k ) .
The curvature bound gives a Poincaré inequality by Brascamp–Lieb. Hence the mean-zero weak solution u u u of − L u = f − E f -\mathcal{L}u=f-\mathbf{E}f − L u = f − E f exists, and the preceding identity gives the covariance representation [17 Section 2, (2.5)–(2.9)] ,
Cov F , Σ ( f , g ) = E [ ( − L u ) g ] = E ( ∇ u , Σ ∇ g ) . \operatorname{Cov}_{F,\Sigma}(f,g)=\mathbf{E}[(-\mathcal{L}u)g]=\mathbf{E}(\nabla u,\Sigma\nabla g). Cov F , Σ ( f , g ) = E [( − L u ) g ] = E ( ∇ u , Σ∇ g ) .
The covariance is therefore controlled by a weighted L 2 L^2 L 2 bound on ∇ u \nabla u ∇ u . Differentiating the Poisson equation once gives
( id − L + Hess F Σ ) ∇ u = ∇ f . (\operatorname{id}-\mathcal{L}+\operatorname{Hess}F\Sigma)\nabla u=\nabla f. ( id − L + Hess F Σ ) ∇ u = ∇ f .
For either weight W = W i W=W_i W = W i , the hypothesis implies
∥ W Hess F Σ W − 1 ∥ ∞ , o p ≤ ∥ W Hess F W − 1 ∥ ∞ , o p ∥ W Σ W − 1 ∥ o p ≤ 1 8 . \lVert W\operatorname{Hess}F\Sigma W^{-1}\rVert_{\infty,\mathrm{op}} \le\lVert W\operatorname{Hess}F W^{-1}\rVert_{\infty,\mathrm{op}}\lVert W\Sigma W^{-1}\rVert_{\mathrm{op}} \le\frac{1}{8}. ∥ W Hess F Σ W − 1 ∥ ∞ , op ≤ ∥ W Hess F W − 1 ∥ ∞ , op ∥ W Σ W − 1 ∥ op ≤ 8 1 .
Test the gradient equation against W 2 ∇ u W^2\nabla u W 2 ∇ u and integrate. The diffusion contributes a nonnegative term, while the displayed bound absorbs the Hessian term,
E ( W ∇ u , W ∇ f ) = E ∥ W ∇ u ∥ 2 2 + ∑ k E ( ∇ ( W ∇ u ) k , Σ ∇ ( W ∇ u ) k ) + E ( W ∇ u , W Hess F Σ W − 1 W ∇ u ) ≥ 7 8 E ∥ W ∇ u ∥ 2 2 . \begin{aligned}
\mathbf{E}(W\nabla u,W\nabla f)=\mathbf{E}\lVert W\nabla u\rVert_2^2+\sum_k\mathbf{E}(\nabla(W\nabla u)_k,\Sigma\nabla(W\nabla u)_k) \\
&+\mathbf{E}(W\nabla u,W\operatorname{Hess}F\Sigma W^{-1}W\nabla u)
\ge\frac{7}{8}\mathbf{E}\lVert W\nabla u\rVert_2^2.
\end{aligned} E ( W ∇ u , W ∇ f ) = E ∥ W ∇ u ∥ 2 2 + k ∑ E ( ∇ ( W ∇ u ) k , Σ∇ ( W ∇ u ) k ) + E ( W ∇ u , W Hess F Σ W − 1 W ∇ u ) ≥ 8 7 E ∥ W ∇ u ∥ 2 2 .
For the weak solution this test is justified by local elliptic regularity and a smooth cutoff in field space. The cutoff term tends to zero since E ∥ ∇ u ∥ 2 2 < ∞ \mathbb{E}\|\nabla u\|_2^2 < \infty E ∥∇ u ∥ 2 2 < ∞ . Cauchy–Schwarz now gives ( E ∥ W ∇ u ∥ 2 2 ) 1 / 2 ≤ 2 ∥ W ∇ f ∥ ∞ , 2 (\mathbb{E}\|W\nabla u\|_2^2)^{1/2} \le2\|W\nabla f\|_{\infty,2} ( E ∥ W ∇ u ∥ 2 2 ) 1/2 ≤ 2∥ W ∇ f ∥ ∞ , 2 . Putting this in the covariance representation proves
∣ Cov F , Σ ( f , g ) ∣ = ∣ E ( W 1 ∇ u , W 1 − 1 Σ W 2 − 1 W 2 ∇ g ) ∣ ≤ ∥ W 1 − 1 Σ W 2 − 1 ∥ op ( E ∥ W 1 ∇ u ∥ 2 2 ) 1 / 2 ∥ W 2 ∇ g ∥ ∞ , 2 \begin{aligned}
|\operatorname{Cov}_{F,\Sigma}(f,g)| = |\mathbb{E}(W_1\nabla u,W_1^{-1}\Sigma W_2^{-1}W_2\nabla g)| \\
&\le\|W_1^{-1}\Sigma W_2^{-1}\|_{\operatorname{op}}(\mathbb{E}\|W_1\nabla u\|_2^2)^{1/2}\|W_2\nabla g\|_{\infty,2}
\end{aligned} ∣ Cov F , Σ ( f , g ) ∣ = ∣ E ( W 1 ∇ u , W 1 − 1 Σ W 2 − 1 W 2 ∇ g ) ∣ ≤ ∥ W 1 − 1 Σ W 2 − 1 ∥ op ( E ∥ W 1 ∇ u ∥ 2 2 ) 1/2 ∥ W 2 ∇ g ∥ ∞ , 2
≤ 2 ∥ W 1 − 1 Σ W 2 − 1 ∥ op ∥ W 1 ∇ f ∥ ∞ , 2 ∥ W 2 ∇ g ∥ ∞ , 2 . \le2\|W_1^{-1}\Sigma W_2^{-1}\|_{\operatorname{op}}\|W_1\nabla f\|_{\infty,2}\|W_2\nabla g\|_{\infty,2}. ≤ 2∥ W 1 − 1 Σ W 2 − 1 ∥ op ∥ W 1 ∇ f ∥ ∞ , 2 ∥ W 2 ∇ g ∥ ∞ , 2 .
To bound the convolved Hessian, differentiate with respect to the background while keeping ζ \zeta ζ as the integration variable. The quotient rule gives
∇ T F , r Σ f = T F , r Σ ∇ f − Cov F , r Σ ( f , ∇ F ) . \nabla T_{F,r\Sigma}f=T_{F,r\Sigma}\nabla f-\operatorname{Cov}_{F,r\Sigma}(f,\nabla F). ∇ T F , r Σ f = T F , r Σ ∇ f − Cov F , r Σ ( f , ∇ F ) .
We apply the covariance bound with reciprocal weights. If W W W is one of the weights in the hypothesis, symmetry and transposition give
∥ W − 1 Σ W ∥ op = ∥ W Σ W − 1 ∥ op , ∥ W − 1 Hess F W ∥ ∞ , op = ∥ W Hess F W − 1 ∥ ∞ , op . \|W^{-1}\Sigma W\|_{\operatorname{op}}=\|W\Sigma W^{-1}\|_{\operatorname{op}},\qquad\|W^{-1}\operatorname{Hess}FW\|_{\infty,\operatorname{op}}=\|W\operatorname{Hess}FW^{-1}\|_{\infty,\operatorname{op}}. ∥ W − 1 Σ W ∥ op = ∥ W Σ W − 1 ∥ op , ∥ W − 1 Hess F W ∥ ∞ , op = ∥ W Hess F W − 1 ∥ ∞ , op .
Thus W − 1 W^{-1} W − 1 satisfies the same hypothesis. For ∥ v ∥ 2 = 1 \|v\|_2=1 ∥ v ∥ 2 = 1 , use the weights W , W − 1 W,W^{-1} W , W − 1 and ∇ ( W v , ∇ F ) = Hess F W v \nabla(Wv,\nabla F)=\operatorname{Hess}FWv ∇ ( W v , ∇ F ) = Hess F W v to obtain
∣ ( v , W Cov F , r Σ ( f , ∇ F ) ) ∣ = ∣ Cov F , r Σ ( f , ( W v , ∇ F ) ) ∣ |(v,W\operatorname{Cov}_{F,r\Sigma}(f,\nabla F))|=|\operatorname{Cov}_{F,r\Sigma}(f,(Wv,\nabla F))| ∣ ( v , W Cov F , r Σ ( f , ∇ F )) ∣ = ∣ Cov F , r Σ ( f , ( W v , ∇ F )) ∣
≤ 2 r ∥ W − 1 Σ W ∥ op ∥ W ∇ f ∥ ∞ , 2 ∥ W − 1 Hess F W v ∥ ∞ , 2 \le2r\|W^{-1}\Sigma W\|_{\operatorname{op}}\|W\nabla f\|_{\infty,2}\|W^{-1}\operatorname{Hess}FWv\|_{\infty,2} ≤ 2 r ∥ W − 1 Σ W ∥ op ∥ W ∇ f ∥ ∞ , 2 ∥ W − 1 Hess F W v ∥ ∞ , 2
≤ 2 r ∥ W Σ W − 1 ∥ op ∥ W Hess F W − 1 ∥ ∞ , op ∥ W ∇ f ∥ ∞ , 2 . \le2r\|W\Sigma W^{-1}\|_{\operatorname{op}}\|W\operatorname{Hess}FW^{-1}\|_{\infty,\operatorname{op}}\|W\nabla f\|_{\infty,2}. ≤ 2 r ∥ W Σ W − 1 ∥ op ∥ W Hess F W − 1 ∥ ∞ , op ∥ W ∇ f ∥ ∞ , 2 .
Taking the supremum over v v v and the background proves
∥ W ∇ T F , r Σ f ∥ ∞ , 2 ≤ ( 1 + 2 r ∥ W Σ W − 1 ∥ op ∥ W Hess F W − 1 ∥ ∞ , op ) ∥ W ∇ f ∥ ∞ , 2 \|W\nabla T_{F,r\Sigma}f\|_{\infty,2}\le(1+2r\|W\Sigma W^{-1}\|_{\operatorname{op}}\|W\operatorname{Hess}FW^{-1}\|_{\infty,\operatorname{op}})\|W\nabla f\|_{\infty,2} ∥ W ∇ T F , r Σ f ∥ ∞ , 2 ≤ ( 1 + 2 r ∥ W Σ W − 1 ∥ op ∥ W Hess F W − 1 ∥ ∞ , op ) ∥ W ∇ f ∥ ∞ , 2
≤ 2 ∥ W ∇ f ∥ ∞ , 2 . \le2\|W\nabla f\|_{\infty,2}. ≤ 2∥ W ∇ f ∥ ∞ , 2 .
Since ∇ F r = T F , r Σ ∇ F \nabla F_r=T_{F,r\Sigma}\nabla F ∇ F r = T F , r Σ ∇ F , apply this estimate to ( W − 1 v , ∇ F ) (W^{-1}v,\nabla F) ( W − 1 v , ∇ F ) . For every ∥ v ∥ 2 = 1 \|v\|_2=1 ∥ v ∥ 2 = 1 this gives
∥ W Hess F r W − 1 v ∥ ∞ , 2 = ∥ W ∇ T F , r Σ ( W − 1 v , ∇ F ) ∥ ∞ , 2 \|W\operatorname{Hess}F_rW^{-1}v\|_{\infty,2}=\|W\nabla T_{F,r\Sigma}(W^{-1}v,\nabla F)\|_{\infty,2} ∥ W Hess F r W − 1 v ∥ ∞ , 2 = ∥ W ∇ T F , r Σ ( W − 1 v , ∇ F ) ∥ ∞ , 2
≤ 2 ∥ W Hess F W − 1 v ∥ ∞ , 2 ≤ 2 ∥ W Hess F W − 1 ∥ ∞ , op , \le2\|W\operatorname{Hess}FW^{-1}v\|_{\infty,2}\le2\|W\operatorname{Hess}FW^{-1}\|_{\infty,\operatorname{op}}, ≤ 2∥ W Hess F W − 1 v ∥ ∞ , 2 ≤ 2∥ W Hess F W − 1 ∥ ∞ , op ,
which proves (32) .
It remains to differentiate the covariance. We split the Gaussian integration, keeping the gradients of f , g f,g f , g in the inner average and differentiating the outer density with respect to the background. Put f r = T F , r Σ f f_r=T_{F,r\Sigma}f f r = T F , r Σ f and g r = T F , r Σ g g_r=T_{F,r\Sigma}g g r = T F , r Σ g . The Gaussian-convolution proof of the evolution identities in [4 (3.25)–(3.26)] applies here because the conditional Hamiltonians have the quadratic lower bound proved above. It gives
∂ r f r = 1 2 Tr ( Σ ∇ 2 f r ) − ( ∇ F r , Σ ∇ f r ) , \partial_r f_r=\frac{1}{2}\operatorname{Tr}(\Sigma\nabla^2 f_r)-(\nabla F_r,\Sigma\nabla f_r), ∂ r f r = 2 1 Tr ( Σ ∇ 2 f r ) − ( ∇ F r , Σ∇ f r ) ,
and the same equation for g r g_r g r . When differentiating T F r , ( 1 − r ) Σ ( f r g r ) T_{F_r,(1-r)\Sigma}(f_rg_r) T F r , ( 1 − r ) Σ ( f r g r ) , we also differentiate F r F_r F r and the remaining covariance ( 1 − r ) Σ (1-r)\Sigma ( 1 − r ) Σ . Their contribution cancels the separate evolution terms of f r f_r f r and g r g_r g r . The product rule leaves the cross derivative,
d d r T F r , ( 1 − r ) Σ ( f r g r ) = T F r , ( 1 − r ) Σ [ ∂ r ( f r g r ) − 1 2 Tr ( Σ ∇ 2 ( f r g r ) ) + ( ∇ F r , Σ ∇ ( f r g r ) ) ] = 1 2 T F r , ( 1 − r ) Σ [ g r Tr ( Σ ∇ 2 f r ) + f r Tr ( Σ ∇ 2 g r ) − Tr ( Σ ∇ 2 ( f r g r ) ) ] = − T F r , ( 1 − r ) Σ ( ∇ f r , Σ ∇ g r ) . \begin{aligned}
\frac{d}{dr}T_{F_r,(1-r)\Sigma}(f_rg_r)=T_{F_r,(1-r)\Sigma}\left[\partial_r(f_rg_r)-\frac{1}{2}\operatorname{Tr}(\Sigma\nabla^2(f_rg_r))+(\nabla F_r,\Sigma\nabla(f_rg_r))\right] \\
&=\frac{1}{2}T_{F_r,(1-r)\Sigma}\left[g_r\operatorname{Tr}(\Sigma\nabla^2f_r)+f_r\operatorname{Tr}(\Sigma\nabla^2g_r)-\operatorname{Tr}(\Sigma\nabla^2(f_rg_r))\right] \\
&=-T_{F_r,(1-r)\Sigma}(\nabla f_r,\Sigma\nabla g_r).
\end{aligned} d r d T F r , ( 1 − r ) Σ ( f r g r ) = T F r , ( 1 − r ) Σ [ ∂ r ( f r g r ) − 2 1 Tr ( Σ ∇ 2 ( f r g r )) + ( ∇ F r , Σ∇ ( f r g r )) ] = 2 1 T F r , ( 1 − r ) Σ [ g r Tr ( Σ ∇ 2 f r ) + f r Tr ( Σ ∇ 2 g r ) − Tr ( Σ ∇ 2 ( f r g r )) ] = − T F r , ( 1 − r ) Σ ( ∇ f r , Σ∇ g r ) .
At r = 0 r = 0 r = 0 the average is T F , Σ ( f g ) T_{F,\Sigma}(fg) T F , Σ ( f g ) ; at r = 1 r = 1 r = 1 it is ( T F , Σ f ) ( T F , Σ g ) (T_{F,\Sigma}f)(T_{F,\Sigma}g) ( T F , Σ f ) ( T F , Σ g ) . Integration therefore gives
Cov F , Σ ( f , g ) = ∫ 0 1 T F r , ( 1 − r ) Σ ( ∇ f r , Σ ∇ g r ) d r . (34) \operatorname{Cov}_{F,\Sigma}(f,g)=\int_{0}^{1}T_{F r,(1-r)\Sigma}(\nabla f_r,\Sigma\nabla g_r)\,dr. \tag*{(34)} Cov F , Σ ( f , g ) = ∫ 0 1 T F r , ( 1 − r ) Σ ( ∇ f r , Σ∇ g r ) d r . (34)
The conditional-gradient estimate bounds the observable by
∥ ( ∇ f r , Σ ∇ g r ) ∥ ∞ ≤ ∥ W 1 − 1 Σ W 2 − 1 ∥ o p ∥ W 1 ∇ f r ∥ ∞ , 2 ∥ W 2 ∇ g r ∥ ∞ , 2 ≤ 4 ∥ W 1 − 1 Σ W 2 − 1 ∥ o p ∥ W 1 ∇ f ∥ ∞ , 2 ∥ W 2 ∇ g ∥ ∞ , 2 . \lVert(\nabla f_r,\Sigma\nabla g_r)\rVert_{\infty}\leq\lVert W_1^{-1}\Sigma W_2^{-1}\rVert_{\mathrm{op}}\lVert W_1\nabla f_r\rVert_{\infty,2}\lVert W_2\nabla g_r\rVert_{\infty,2}
\leq4\lVert W_1^{-1}\Sigma W_2^{-1}\rVert_{\mathrm{op}}\lVert W_1\nabla f\rVert_{\infty,2}\lVert W_2\nabla g\rVert_{\infty,2}. ∥( ∇ f r , Σ∇ g r ) ∥ ∞ ≤ ∥ W 1 − 1 Σ W 2 − 1 ∥ op ∥ W 1 ∇ f r ∥ ∞ , 2 ∥ W 2 ∇ g r ∥ ∞ , 2 ≤ 4 ∥ W 1 − 1 Σ W 2 − 1 ∥ op ∥ W 1 ∇ f ∥ ∞ , 2 ∥ W 2 ∇ g ∥ ∞ , 2 .
For the outer conditional law, (32) with unit weights gives
Hess F r ≥ − 2 ∥ Hess F ∥ ∞ , o p id ≥ − 1 4 Σ − 1 , [ ( 1 − r ) Σ ] − 1 + Hess F r ≥ 1 2 [ ( 1 − r ) Σ ] − 1 . \operatorname{Hess}F_r\geq-2\lVert\operatorname{Hess}F\rVert_{\infty,\mathrm{op}} \operatorname{id}\geq-\frac{1}{4}\Sigma^{-1},\qquad[(1-r)\Sigma]^{-1}+\operatorname{Hess}F_r\geq\frac{1}{2}[(1-r)\Sigma]^{-1}. Hess F r ≥ − 2 ∥ Hess F ∥ ∞ , op id ≥ − 4 1 Σ − 1 , [( 1 − r ) Σ ] − 1 + Hess F r ≥ 2 1 [( 1 − r ) Σ ] − 1 .
For each r r r , the function ( ∇ f r , Σ ∇ g r ) (\nabla f_r,\Sigma\nabla g_r) ( ∇ f r , Σ∇ g r ) depends on the integration variable, not on the outer background. We can therefore apply (30) to the outer density without taking another derivative of f r f_r f r or g r g_r g r , obtaining
∣ ∂ w Cov F , Σ ( f , g ) ∣ ≤ 2 ( w , Σ − 1 w ) ∫ 0 1 ∥ ( ∇ f r , Σ ∇ g r ) ∥ ∞ 1 − r d r ≤ 8 2 ∥ W 1 − 1 Σ W 2 − 1 ∥ o p ( w , Σ − 1 w ) ∥ W 1 ∇ f ∥ ∞ , 2 ∥ W 2 ∇ g ∥ ∞ , 2 . \left|\partial_w\operatorname{Cov}_{F,\Sigma}(f,g)\right|\leq\sqrt{2}\sqrt{(w,\Sigma^{-1}w)}\int_{0}^{1}\frac{\lVert(\nabla f_r,\Sigma\nabla g_r)\rVert_{\infty}}{\sqrt{1-r}}\,dr
\leq8\sqrt{2}\lVert W_1^{-1}\Sigma W_2^{-1}\rVert_{\mathrm{op}}\sqrt{(w,\Sigma^{-1}w)}\lVert W_1\nabla f\rVert_{\infty,2}\lVert W_2\nabla g\rVert_{\infty,2}. ∣ ∂ w Cov F , Σ ( f , g ) ∣ ≤ 2 ( w , Σ − 1 w ) ∫ 0 1 1 − r ∥( ∇ f r , Σ∇ g r ) ∥ ∞ d r ≤ 8 2 ∥ W 1 − 1 Σ W 2 − 1 ∥ op ( w , Σ − 1 w ) ∥ W 1 ∇ f ∥ ∞ , 2 ∥ W 2 ∇ g ∥ ∞ , 2 .
The integrable factor ( 1 − r ) − 1 / 2 (1-r)^{-1/2} ( 1 − r ) − 1/2 justifies differentiation under the integral. The endpoint formulas for the undifferentiated interpolation follow by continuity. Mollification in field space preserves the weighted Hessian and gradient bounds. For sufficiently small mollification radii, the conditional Hamiltonians have a common quadratic lower bound, so dominated convergence passes the estimates to the stated C 2 C^2 C 2 , C 1 C^1 C 1 inputs. The constants in the estimates are unchanged. □
For the scale step, take F = N + λ R F=N+\lambda R F = N + λ R , 0 ≤ λ ≤ 1 0\leq\lambda\leq1 0 ≤ λ ≤ 1 , and Σ = C T \Sigma=C_T Σ = C T . Let e i e_i e i now denote the unit vector in the field coordinate i i i , so ( Q T e i ) a = Q T ( a , i ) (Q_Te_i)_a=Q_T(a,i) ( Q T e i ) a = Q T ( a , i ) , and let W x W_x W x be diagonal with ( W x ) a a = e ρ d ( a , x ) / ℓ ′ (W_x)_{aa}=e^{\rho d(a,x)/\ell'} ( W x ) aa = e ρ d ( a , x ) / ℓ ′ . Assume 0 < ρ ≤ ρ 0 0<\rho\leq\rho_0 0 < ρ ≤ ρ 0 and ρ / ℓ ′ ≤ μ / ( 8 ℓ ) \rho/\ell'\leq\mu/(8\ell) ρ / ℓ ′ ≤ μ / ( 8 ℓ ) . The triangle inequality gives
∣ ( W x Hess F W x − 1 ) a b ∣ ≤ e ρ d ( a , b ) / ℓ ′ sup φ ∣ ∂ a ∂ b F ∣ . \left|(W_x\operatorname{Hess}F W_x^{-1})_{ab}\right|\leq e^{\rho d(a,b)/\ell'}\sup_{\varphi}|\partial_a\partial_bF|. ( W x Hess F W x − 1 ) ab ≤ e ρ d ( a , b ) / ℓ ′ φ sup ∣ ∂ a ∂ b F ∣.
The expression on the right is symmetric in a , b a,b a , b . Its row sums, and hence its column sums, are at most H ζ ℓ ℓ − 2 H\zeta\ell\ell^{-2} H ζ ℓ ℓ − 2 by (28) . Schur’s inequality therefore gives ∥ W x Hess F W x − 1 ∥ ∞ , o p ≤ H ζ ℓ ℓ − 2 \lVert W_x\operatorname{Hess}F W_x^{-1}\rVert_{\infty,\mathrm{op}}\leq H\zeta\ell\ell^{-2} ∥ W x Hess F W x − 1 ∥ ∞ , op ≤ H ζ ℓ ℓ − 2 . The same argument for C T C_T C T , using (A.4), verifies the hypothesis of Lemma 3.2,
∥ W x C T W x − 1 ∥ o p ∥ W x Hess F W x − 1 ∥ ∞ , o p ≤ C ℓ ′ 2 H ζ ℓ ℓ − 2 ≤ 1 8 . \lVert W_xC_TW_x^{-1}\rVert_{\mathrm{op}}\lVert W_x\operatorname{Hess}F W_x^{-1}\rVert_{\infty,\mathrm{op}}\leq C\ell'^2H\zeta\ell\ell^{-2}\leq\frac{1}{8}. ∥ W x C T W x − 1 ∥ op ∥ W x Hess F W x − 1 ∥ ∞ , op ≤ C ℓ ′2 H ζ ℓ ℓ − 2 ≤ 8 1 .
For observables localized near x , y x,y x , y , the covariance bound retains their separation,
∥ W x − 1 C T W y − 1 ∥ o p ≤ ∥ W x − 1 C T W x ∥ o p sup a e − ρ [ d ( a , x ) + d ( a , y ) ] / ℓ ′ ≤ C ρ ℓ 2 e − ρ d ( x , y ) / ℓ ′ . \lVert W_x^{-1}C_TW_y^{-1}\rVert_{\mathrm{op}}\leq\lVert W_x^{-1}C_TW_x\rVert_{\mathrm{op}}\sup_a e^{-\rho[d(a,x)+d(a,y)]/\ell'}
\leq C_\rho\ell^{2}e^{-\rho d(x,y)/\ell'}. ∥ W x − 1 C T W y − 1 ∥ op ≤ ∥ W x − 1 C T W x ∥ op a sup e − ρ [ d ( a , x ) + d ( a , y )] / ℓ ′ ≤ C ρ ℓ 2 e − ρ d ( x , y ) / ℓ ′ .
The background derivative in (33) introduces an inverse covariance. Its square root contributes ℓ ′ − 2 \ell'^{-2} ℓ ′ − 2 even if only the last half of the Gaussian integration remains. Indeed, for 0 ≤ S ≤ T / 2 0\leq S\leq T/2 0 ≤ S ≤ T /2 ,
( Q T e j , ( C T − C S ) − 1 Q T e j ) = [ A e ( T − S ) A − 1 ] j j ≤ C T ( e − T A / 4 ) j j ≤ C ℓ ′ − 4 . (35) (Q_Te_j,(C_T-C_S)^{-1}Q_Te_j)=\left[\frac{A}{e^{(T-S)A}-1}\right]_{jj}
\leq\frac{C}{T}(e^{-TA/4})_{jj}\leq C\ell'^{-4}. \tag*{(35)} ( Q T e j , ( C T − C S ) − 1 Q T e j ) = [ e ( T − S ) A − 1 A ] j j ≤ T C ( e − T A /4 ) j j ≤ C ℓ ′ − 4 . (35)
The identity follows from C T − C S = A − 1 e − S A ( id − e − ( T − S ) A ) C_T-C_S=A^{-1}e^{-SA}(\operatorname{id}-e^{-(T-S)A}) C T − C S = A − 1 e − S A ( id − e − ( T − S ) A ) and Q T 2 = e − T A Q_T^2=e^{-TA} Q T 2 = e − T A . For the inequality use a / ( e ( T − S ) a − 1 ) ≤ C T − 1 e − T a / 4 a/(e^{(T-S)a}-1)\le C T^{-1}e^{-Ta/4} a / ( e ( T − S ) a − 1 ) ≤ C T − 1 e − T a /4 for a > 0 a>0 a > 0 , then the diagonal heat bound. At S = 0 S=0 S = 0 , the square root in (33) cancels the factor ℓ ′ 2 \ell^{\prime2} ℓ ′2 from the covariance. Once the smallness condition holds, these constants are independent of M M M , cutoff and volume.
We shall also need the gradient of ( Q T e i , ∇ F ) (Q_Te_i,\nabla F) ( Q T e i , ∇ F ) . The heat-kernel bound and its spatial sum give
∥ W i Q T e i ∥ 2 2 = ∑ a e 2 ρ d ( a , i ) / ℓ ′ Q T ( a , i ) 2 ≤ ( sup a e ρ d ( a , i ) / ℓ ′ Q T ( a , i ) ) ∑ a e ρ d ( a , i ) / ℓ ′ Q T ( a , i ) ≤ C ρ ℓ ′ − 2 . \lVert W_iQ_Te_i\rVert_2^2=\sum_a e^{2\rho d(a,i)/\ell'}Q_T(a,i)^2
\le\left(\sup_a e^{\rho d(a,i)/\ell'}Q_T(a,i)\right)\sum_a e^{\rho d(a,i)/\ell'}Q_T(a,i)\le C_\rho\ell'^{-2}. ∥ W i Q T e i ∥ 2 2 = a ∑ e 2 ρ d ( a , i ) / ℓ ′ Q T ( a , i ) 2 ≤ ( a sup e ρ d ( a , i ) / ℓ ′ Q T ( a , i ) ) a ∑ e ρ d ( a , i ) / ℓ ′ Q T ( a , i ) ≤ C ρ ℓ ′ − 2 .
Combining this with the weighted Hessian bound yields
∥ W i Hess F Q T e i ∥ ∞ , 2 ≤ ∥ W i Hess F W i − 1 ∥ ∞ , op ∥ W i Q T e i ∥ 2 ≤ C ρ H ζ ℓ ℓ − 2 ℓ ′ − 1 . (36) \lVert W_i\operatorname{Hess}FQ_Te_i\rVert_{\infty,2}\le\lVert W_i\operatorname{Hess}FW_i^{-1}\rVert_{\infty,\operatorname{op}}\lVert W_iQ_Te_i\rVert_2\le C_\rho H\zeta_\ell\ell^{-2}\ell'^{-1}. \tag*{(36)} ∥ W i Hess F Q T e i ∥ ∞ , 2 ≤ ∥ W i Hess F W i − 1 ∥ ∞ , op ∥ W i Q T e i ∥ 2 ≤ C ρ H ζ ℓ ℓ − 2 ℓ ′ − 1 . (36)
Similarly, for ρ ≤ 6 μ \rho\le6\mu ρ ≤ 6 μ , squaring (27) gives
∥ W x ∇ g x ∥ ∞ , 2 2 ≤ C ( K + 1 ) 2 ζ ℓ 2 ℓ − 4 ∑ a e − 6 μ d ( a , x ) / ℓ ≤ C ( K + 1 ) 2 ζ ℓ 2 ℓ − 2 . \lVert W_x\nabla g_x\rVert_{\infty,2}^2\le C(K+1)^2\zeta_\ell^2\ell^{-4}\sum_a e^{-6\mu d(a,x)/\ell}\le C(K+1)^2\zeta_\ell^2\ell^{-2}. ∥ W x ∇ g x ∥ ∞ , 2 2 ≤ C ( K + 1 ) 2 ζ ℓ 2 ℓ − 4 a ∑ e − 6 μ d ( a , x ) / ℓ ≤ C ( K + 1 ) 2 ζ ℓ 2 ℓ − 2 .
These sums involve the diffusion area, rather than the number of sites. The first derivative of G x G_x G x contains a covariance with [ Q T ∇ F ] i [Q_T\nabla F]_i [ Q T ∇ F ] i . The Hessian of G x G_x G x contains the background derivative of this covariance. Write Cov λ \operatorname{Cov}_\lambda Cov λ for covariance under E λ \mathbb{E}_\lambda E λ , and take ρ = 6 μ \rho=6\mu ρ = 6 μ , which is admissible for M ≥ 48 M\ge48 M ≥ 48 . By (31) and the preceding gradient bounds,
∣ Cov λ ( g x , [ Q T ∇ F ] i ) ∣ ≤ C ℓ ′ 2 e − 6 μ d ( i , x ) / ℓ ′ [ ( K + 1 ) ζ ℓ ℓ − 1 ] [ H ζ ℓ ℓ − 2 ℓ ′ − 1 ] ≤ C M , K H ζ ℓ 2 ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ . (37) \left|\operatorname{Cov}_\lambda(g_x,[Q_T\nabla F]_i)\right|\le C\ell'^2e^{-6\mu d(i,x)/\ell'}[(K+1)\zeta_\ell\ell^{-1}][H\zeta_\ell\ell^{-2}\ell'^{-1}]
\le C_{M,K}H\zeta_\ell^2\ell'^{-2}e^{-6\mu d(i,x)/\ell'}. \tag*{(37)} ∣ Cov λ ( g x , [ Q T ∇ F ] i ) ∣ ≤ C ℓ ′2 e − 6 μ d ( i , x ) / ℓ ′ [( K + 1 ) ζ ℓ ℓ − 1 ] [ H ζ ℓ ℓ − 2 ℓ ′ − 1 ] ≤ C M , K H ζ ℓ 2 ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ . (37)
A derivative in the outgoing background φ j ′ \varphi'_j φ j ′ is a derivative of the conditional law in direction Q T e j Q_Te_j Q T e j . Thus (33) has the same product of weighted gradients, with the additional factor bounded by C ℓ ′ − 2 C\ell'^{-2} C ℓ ′ − 2 in (35) . We obtain
∣ ∂ j Cov λ ( g x , [ Q T ∇ F ] i ) ∣ ≤ C M , K H ζ ℓ 2 ℓ ′ − 4 e − 6 μ d ( i , x ) / ℓ ′ . (38) \left|\partial_j\operatorname{Cov}_\lambda(g_x,[Q_T\nabla F]_i)\right|\le C_{M,K}H\zeta_\ell^2\ell'^{-4}e^{-6\mu d(i,x)/\ell'}. \tag*{(38)} ∣ ∂ j Cov λ ( g x , [ Q T ∇ F ] i ) ∣ ≤ C M , K H ζ ℓ 2 ℓ ′ − 4 e − 6 μ d ( i , x ) / ℓ ′ . (38)
Here F = N + λ R F=N+\lambda R F = N + λ R , and both bounds hold uniformly in λ ∈ [ 0 , 1 ] \lambda\in[0,1] λ ∈ [ 0 , 1 ] and the background, with C M , K C_{M,K} C M , K independent of H H H , B B B , cutoff and volume.
The neutral convolution
The Hessian of the neutral convolution has an averaged Hessian term and a covariance term. The latter is quadratic in H ζ ℓ H\zeta_\ell H ζ ℓ . For the former, zero row sums supply two spatial differences, giving a linear bound with coefficient independent of M M M .
Proposition 3.3. Suppose N N N is invariant under constant shifts and satisfies the row bound (25) with N N N in place of N + R N+R N + R . For M ≥ 16 M\ge16 M ≥ 16 , in the heat regime and under the standing smallness conditions, its Gaussian convolution is neutral and
sup i ∑ j e μ d ( i , j ) / ( 2 ℓ ′ ) sup φ ′ ∣ ∂ i ∂ j [ − log E C T e − N ( Q T φ ′ + ζ ) ] ∣ ≤ [ C μ H + C β , μ , M H 2 ζ ℓ ] ζ ℓ ℓ ′ − 2 . (39) \sup_i\sum_j e^{\mu d(i,j)/(2\ell')}\sup_{\varphi'}\left|\partial_i\partial_j\left[-\log\mathbb{E}C_Te^{-N(Q_T\varphi'+\zeta)}\right]\right|
\le[C_\mu H+C_{\beta,\mu,M}H^2\zeta_\ell]\zeta_\ell\ell'^{-2}. \tag*{(39)} i sup j ∑ e μ d ( i , j ) / ( 2 ℓ ′ ) φ ′ sup ∂ i ∂ j [ − log E C T e − N ( Q T φ ′ + ζ ) ] ≤ [ C μ H + C β , μ , M H 2 ζ ℓ ] ζ ℓ ℓ ′ − 2 . (39) The constant C μ C_\mu C μ is independent of M M M , K K K , H H H , B B B .
Proof. Differentiate N ( φ + s 1 ) = N ( φ ) N(\varphi+s1)=N(\varphi) N ( φ + s 1 ) = N ( φ ) first in s s s and then in φ a \varphi_a φ a . This gives ∑ b ∂ a ∂ b N = 0 \sum_b \partial_a\partial_bN=0 ∑ b ∂ a ∂ b N = 0 . The column sums also vanish, since ∂ a ∂ b N = ∂ b ∂ a N \partial_a\partial_bN=\partial_b\partial_aN ∂ a ∂ b N = ∂ b ∂ a N . In the following expansion, the terms containing u a w a u_aw_a u a w a and u b w b u_bw_b u b w b therefore sum to zero,
∑ a , b ∂ a ∂ b N ( u a − u b ) ( w a − w b ) = − ∑ a , b ∂ a ∂ b N ( u a w b + u b w a ) = − 2 ( u , Hess N w ) . \begin{aligned}
\sum_{a,b}\partial_a\partial_bN(u_a-u_b)(w_a-w_b)=-\sum_{a,b}\partial_a\partial_bN(u_aw_b+u_bw_a) \\
&=-2(u,\operatorname{Hess}Nw).
\end{aligned} a , b ∑ ∂ a ∂ b N ( u a − u b ) ( w a − w b ) = − a , b ∑ ∂ a ∂ b N ( u a w b + u b w a ) = − 2 ( u , Hess N w ) .
Equivalently,
( u , Hess N w ) = − 1 2 ∑ a , b ∂ a ∂ b N ( u a − u b ) ( w a − w b ) . (40) (u,\operatorname{Hess}Nw)=-\frac{1}{2}\sum_{a,b}\partial_a\partial_bN(u_a-u_b)(w_a-w_b). \tag*{(40)} ( u , Hess N w ) = − 2 1 a , b ∑ ∂ a ∂ b N ( u a − u b ) ( w a − w b ) . (40)
The conditional Hessian identity [4 Lemma 3.12] is
∂ i ∂ j [ − log E C T , C T e − N ( Q T φ ′ + ζ ) ] = T N , C T ( Q T e i , Hess N Q T e j ) − Cov N , C T ( ( Q T e i , ∇ N ) , ( Q T e j , ∇ N ) ) . \partial_i\partial_j\left[-\log\mathbb{E}_{C_T,C_T}e^{-N(Q_T\varphi'+\zeta)}\right]=T_{N,C_T}(Q_Te_i,\operatorname{Hess}NQ_Te_j)-\operatorname{Cov}_{N,C_T}((Q_Te_i,\nabla N),(Q_Te_j,\nabla N)). ∂ i ∂ j [ − log E C T , C T e − N ( Q T φ ′ + ζ ) ] = T N , C T ( Q T e i , Hess N Q T e j ) − Cov N , C T (( Q T e i , ∇ N ) , ( Q T e j , ∇ N )) .
The averages have background Q T φ ′ Q_T\varphi' Q T φ ′ . For the first term, the heat-difference bound gives
∣ Q T ( a , i ) − Q T ( b , i ) ∣ ≤ C d ( a , b ) ℓ ′ − 3 e − 2 μ d ( a , i ) / ℓ ′ + 2 μ d ( a , b ) / ℓ ′ . \lvert Q_T(a,i)-Q_T(b,i)\rvert\le Cd(a,b)\ell'^{-3}e^{-2\mu d(a,i)/\ell'+2\mu d(a,b)/\ell'}. ∣ Q T ( a , i ) − Q T ( b , i )∣ ≤ C d ( a , b ) ℓ ′ − 3 e − 2 μ d ( a , i ) / ℓ ′ + 2 μ d ( a , b ) / ℓ ′ .
For M ≥ 16 M\ge16 M ≥ 16 the second spatial moment costs only C μ ℓ 2 C_\mu\ell^2 C μ ℓ 2 , independently of M M M , since
d ( a , b ) 2 e 4 μ d ( a , b ) / ℓ ′ ≤ [ d ( a , b ) 2 e − μ d ( a , b ) / ( 4 ℓ ) ] e μ d ( a , b ) / ( 2 ℓ ) ≤ C μ ℓ 2 e μ d ( a , b ) / ( 2 ℓ ) . d(a,b)^2e^{4\mu d(a,b)/\ell'}\le\left[d(a,b)^2e^{-\mu d(a,b)/(4\ell)}\right]e^{\mu d(a,b)/(2\ell)}\le C_\mu\ell^2e^{\mu d(a,b)/(2\ell)}. d ( a , b ) 2 e 4 μ d ( a , b ) / ℓ ′ ≤ [ d ( a , b ) 2 e − μ d ( a , b ) / ( 4 ℓ ) ] e μ d ( a , b ) / ( 2 ℓ ) ≤ C μ ℓ 2 e μ d ( a , b ) / ( 2 ℓ ) .
Thus (40) and the weighted row bound give
∣ ( Q T e i , Hess N Q T e j ) ∣ ≤ C ℓ ′ − 6 ∑ a e − 2 μ [ d ( a , i ) + d ( a , j ) ] / ℓ ′ ∑ b d ( a , b ) 2 e 4 μ d ( a , b ) / ℓ ′ ∣ ∂ a ∂ b N ∣ ≤ C μ ℓ 2 ℓ ′ − 6 ∑ a e − 2 μ [ d ( a , i ) + d ( a , j ) ] / ℓ ′ ∑ b e μ d ( a , b ) / ( 2 ℓ ) sup φ ∣ ∂ a ∂ b N ∣ ≤ C μ H ζ ℓ ℓ ′ − 6 ∑ a e − 2 μ [ d ( a , i ) + d ( a , j ) ] / ℓ ′ ≤ C μ H ζ ℓ ℓ ′ − 4 e − μ d ( i , j ) / ℓ ′ . \begin{aligned}
\lvert(Q_Te_i,\operatorname{Hess}NQ_Te_j)\rvert
&\le C\ell'^{-6}\sum_a e^{-2\mu[d(a,i)+d(a,j)]/\ell'}\sum_b d(a,b)^2e^{4\mu d(a,b)/\ell'}\lvert\partial_a\partial_bN\rvert\\
&\le C_\mu\ell^2\ell'^{-6}\sum_a e^{-2\mu[d(a,i)+d(a,j)]/\ell'}\sum_b e^{\mu d(a,b)/(2\ell)}\sup_\varphi\lvert\partial_a\partial_bN\rvert\\
&\le C_\mu H_\zeta\ell\ell'^{-6}\sum_a e^{-2\mu[d(a,i)+d(a,j)]/\ell'}\\
&\le C_\mu H_\zeta\ell\ell'^{-4}e^{-\mu d(i,j)/\ell'}.
\end{aligned} ∣( Q T e i , Hess N Q T e j )∣ ≤ C ℓ ′ − 6 a ∑ e − 2 μ [ d ( a , i ) + d ( a , j )] / ℓ ′ b ∑ d ( a , b ) 2 e 4 μ d ( a , b ) / ℓ ′ ∣ ∂ a ∂ b N ∣ ≤ C μ ℓ 2 ℓ ′ − 6 a ∑ e − 2 μ [ d ( a , i ) + d ( a , j )] / ℓ ′ b ∑ e μ d ( a , b ) / ( 2 ℓ ) φ sup ∣ ∂ a ∂ b N ∣ ≤ C μ H ζ ℓ ℓ ′ − 6 a ∑ e − 2 μ [ d ( a , i ) + d ( a , j )] / ℓ ′ ≤ C μ H ζ ℓ ℓ ′ − 4 e − μ d ( i , j ) / ℓ ′ .
The last sum is bounded by
∑ a e − 2 μ [ d ( a , i ) + d ( a , j ) ] / ℓ ′ ≤ e − μ d ( i , j ) / ℓ ′ ∑ a e − μ d ( a , i ) / ℓ ′ ≤ C μ ℓ ′ 2 e − μ d ( i , j ) / ℓ ′ . \sum_a e^{-2\mu[d(a,i)+d(a,j)]/\ell'}\le e^{-\mu d(i,j)/\ell'}\sum_a e^{-\mu d(a,i)/\ell'}\le C_\mu\ell'^2e^{-\mu d(i,j)/\ell'}. a ∑ e − 2 μ [ d ( a , i ) + d ( a , j )] / ℓ ′ ≤ e − μ d ( i , j ) / ℓ ′ a ∑ e − μ d ( a , i ) / ℓ ′ ≤ C μ ℓ ′2 e − μ d ( i , j ) / ℓ ′ .
For the covariance use ( W i ) a a = e μ d ( a , i ) / ℓ ′ (W_i)_{aa}=e^{\mu d(a,i)/\ell'} ( W i ) aa = e μ d ( a , i ) / ℓ ′ . The two observables have gradients Hess N Q T e i \operatorname{Hess}NQ_Te_i Hess N Q T e i and Hess N Q T e j \operatorname{Hess}NQ_Te_j Hess N Q T e j . Thus (31) and (36) with F = N F=N F = N give
∣ Cov N , C T ( ( Q T e i , ∇ N ) , ( Q T e j , ∇ N ) ) ∣ ≤ C ℓ ′ 2 e − μ d ( i , j ) / ℓ ′ ( H ζ ℓ ℓ ′ − 2 ℓ ′ − 1 ) 2 ≤ C β , μ , M H ζ 2 ℓ ′ − 4 e − μ d ( i , j ) / ℓ ′ . \begin{aligned}
\left\lvert\operatorname{Cov}_{N,C_T}((Q_Te_i,\nabla N),(Q_Te_j,\nabla N))\right\rvert
&\le C\ell'^2e^{-\mu d(i,j)/\ell'}(H_\zeta\ell\ell'^{-2}\ell'^{-1})^2\\
&\le C_{\beta,\mu,M}H_\zeta^2\ell'^{-4}e^{-\mu d(i,j)/\ell'}.
\end{aligned} ∣ Cov N , C T (( Q T e i , ∇ N ) , ( Q T e j , ∇ N )) ∣ ≤ C ℓ ′2 e − μ d ( i , j ) / ℓ ′ ( H ζ ℓ ℓ ′ − 2 ℓ ′ − 1 ) 2 ≤ C β , μ , M H ζ 2 ℓ ′ − 4 e − μ d ( i , j ) / ℓ ′ .
Multiply these two bounds by e μ d ( i , j ) / ( 2 ℓ ′ ) e^{\mu d(i,j)/(2\ell')} e μ d ( i , j ) / ( 2 ℓ ′ ) and sum over j j j . Since ∑ j e − μ d ( i , j ) / ( 2 ℓ ′ ) ≤ C μ ℓ ′ 2 \sum_j e^{-\mu d(i,j)/(2\ell')}\le C_\mu\ell'^2 ∑ j e − μ d ( i , j ) / ( 2 ℓ ′ ) ≤ C μ ℓ ′2 , we obtain (39) . Neutrality is preserved by Gaussian convolution and by composition with Q T Q_T Q T , which maps constants to constants.
Gaussian averaging
We compare Gaussian averaging with an independent heat average of the phase. The error involves only derivatives with the phase held fixed and vanishes on the original cosine.
Lemma 3.4. Let f f f be C 1 C^{1} C 1 , periodic under constant shifts with period 2 π / c t 2\pi/c_t 2 π / c t and with zero period average. Suppose the norms below are finite. Fix x ∈ Λ x \in\Lambda x ∈ Λ and M ≥ 2 M \ge2 M ≥ 2 . In the heat regime,
∥ E C T f ( ⋅ + ζ ) ∥ ∗ ≤ e − β 2 ( C t + T − C t ) x x [ ∥ f ∥ ∗ + C β , μ ℓ 2 sup i e μ d ( i , x ) / ℓ ∥ ∂ i ( a ) f ∥ ∗ ] . (41) \left\|\mathbb{E}_{C_T}f(\cdot+\zeta)\right\|_* \le e^{-\frac{\beta}{2}(C_{t+T}-C_t)_{xx}}\left[\|f\|_*+C_{\beta,\mu}\ell^2\sup_i e^{\mu d(i,x)/\ell}\left\|\partial_i^{(a)}f\right\|_*\right]. \tag*{(41)} ∥ E C T f ( ⋅ + ζ ) ∥ ∗ ≤ e − 2 β ( C t + T − C t ) xx [ ∥ f ∥ ∗ + C β , μ ℓ 2 i sup e μ d ( i , x ) / ℓ ∂ i ( a ) f ∗ ] . (41)
The constant is independent of M M M , cutoff and volume.
Proof. Split the covariance as C T = C s + ( C T − C s ) C_T=C_s+(C_T-C_s) C T = C s + ( C T − C s ) , and let ζ , η \zeta,\eta ζ , η be independent centred Gaussian vectors with these covariances. We leave ζ \zeta ζ unchanged. From η \eta η we subtract a constant field which removes its contribution to the phase, and replace that contribution by an independent heat average in the angle. Varying s s s from zero to T T T then restores the original correlation between the phase and the other field variables. Thus put
f s ( φ ) = e β 2 ( C t + T − C t + s ) x x Θ t 2 E f ( φ + ζ + η − β ( Q t η ) x c t − 1 ) . f_s(\varphi)=e^{\frac{\beta}{2}(C_{t+T}-C_{t+s})_{xx}}\Theta_t^2\mathbb{E}f\left(\varphi+\zeta+\eta-\sqrt{\beta}(Q_t\eta)_x c_t^{-1}\right). f s ( φ ) = e 2 β ( C t + T − C t + s ) xx Θ t 2 E f ( φ + ζ + η − β ( Q t η ) x c t − 1 ) .
The subtraction removes the change in the phase at x x x ,
β [ Q t ( η − β ( Q t η ) x c t − 1 ) ] x = β ( Q t η ) x − β ( Q t η ) x c t − 1 c t = 0. \sqrt{\beta}\left[Q_t\left(\eta-\sqrt{\beta}(Q_t\eta)_x c_t^{-1}\right)\right]_x=\sqrt{\beta}(Q_t\eta)_x-\sqrt{\beta}(Q_t\eta)_x c_t^{-1}c_t=0. β [ Q t ( η − β ( Q t η ) x c t − 1 ) ] x = β ( Q t η ) x − β ( Q t η ) x c t − 1 c t = 0.
The heat semigroup on the circle restores the phase variance, β [ Q t ( C T − C s ) Q t ] x x = β ( C t + T − C t + s ) x x \beta[Q_t(C_T-C_s)Q_t]_{xx}=\beta(C_{t+T}-C_{t+s})_{xx} β [ Q t ( C T − C s ) Q t ] xx = β ( C t + T − C t + s ) xx , independently of η \eta η . At the two endpoints we therefore have
f T = E C T f ( ⋅ + ζ ) , ∥ f 0 ∥ ∗ ≤ e − β 2 ( C t + T − C t ) x x ∥ f ∥ ∗ . f_T=\mathbb{E}_{C_T}f(\cdot+\zeta),\qquad\|f_0\|_*\le e^{-\frac{\beta}{2}(C_{t+T}-C_t)_{xx}}\|f\|_*. f T = E C T f ( ⋅ + ζ ) , ∥ f 0 ∥ ∗ ≤ e − 2 β ( C t + T − C t ) xx ∥ f ∥ ∗ .
To estimate f T − f 0 f_T-f_0 f T − f 0 , differentiate the commuting averages. As s s s increases, the covariance C s C_s C s grows at rate e − s A e^{-sA} e − s A . The covariance of η \eta η decreases at the same rate, with derivatives ∂ i ( x ) \partial_i^{(x)} ∂ i ( x ) in place of ∂ i \partial_i ∂ i . Differentiating also the remaining angular variance gives
∂ s f s = [ 1 2 ∑ i , j ( e − s A ) i j ( ∂ i ∂ j − ∂ i ( x ) ∂ j ( x ) ) − β 2 ( e − ( t + s ) A ) x x Θ t 2 ] f s . \partial_s f_s=\left[\frac{1}{2}\sum_{i,j}(e^{-sA})_{ij}\left(\partial_i\partial_j-\partial_i^{(x)}\partial_j^{(x)}\right)-\frac{\beta}{2}(e^{-(t+s)A})_{xx}\Theta_t^2\right]f_s. ∂ s f s = [ 2 1 i , j ∑ ( e − s A ) ij ( ∂ i ∂ j − ∂ i ( x ) ∂ j ( x ) ) − 2 β ( e − ( t + s ) A ) xx Θ t 2 ] f s .
The terms with two angular derivatives cancel. Indeed,
∂ i ∂ j − ∂ i ( x ) ∂ j ( x ) = β Θ t [ Q t ( i , x ) ∂ j ( x ) + Q t ( j , x ) ∂ i ( x ) ] + β Q t ( i , x ) Q t ( j , x ) Θ t 2 , \partial_i\partial_j-\partial_i^{(x)}\partial_j^{(x)}=\sqrt{\beta}\Theta_t\left[Q_t(i,x)\partial_j^{(x)}+Q_t(j,x)\partial_i^{(x)}\right]+\beta Q_t(i,x)Q_t(j,x)\Theta_t^2, ∂ i ∂ j − ∂ i ( x ) ∂ j ( x ) = β Θ t [ Q t ( i , x ) ∂ j ( x ) + Q t ( j , x ) ∂ i ( x ) ] + β Q t ( i , x ) Q t ( j , x ) Θ t 2 ,
∑ i , j ( e − s A ) i j Q t ( i , x ) Q t ( j , x ) = ( Q t e − s A Q t ) x x = ( e − ( s + t ) A ) x x , \sum_{i,j}(e^{-sA})_{ij}Q_t(i,x)Q_t(j,x)=(Q_te^{-sA}Q_t)_{xx}=(e^{-(s+t)A})_{xx}, i , j ∑ ( e − s A ) ij Q t ( i , x ) Q t ( j , x ) = ( Q t e − s A Q t ) xx = ( e − ( s + t ) A ) xx ,
∂ s f s = β Θ t ∑ i ( e − ( s + t / 2 ) A ) i x ∂ i ( x ) f s . \partial_s f_s=\sqrt{\beta}\Theta_t\sum_i(e^{-(s+t/2)A})_{ix}\partial_i^{(x)}f_s. ∂ s f s = β Θ t i ∑ ( e − ( s + t /2 ) A ) i x ∂ i ( x ) f s .
The phase-fixed derivative is controlled by the hypothesis. To bound the angular derivative without differentiating f f f again, we take it on the remaining heat kernel on the circle. The derivative of this kernel has L 1 L^1 L 1 norm at most C a − 1 / 2 Ca^{-1/2} C a − 1/2 for a ≤ 1 a\le1 a ≤ 1 , while for a ≥ 1 a\ge1 a ≥ 1 the Fourier bound ∑ q ≠ 0 ∣ q ∣ e − q 2 a ≤ C e − a \sum_{q\ne0}|q|e^{-q^2a}\le Ce^{-a} ∑ q = 0 ∣ q ∣ e − q 2 a ≤ C e − a applies. Commuting with the semigroup in (24) , we obtain
∥ Θ t e a Θ t 2 h ∥ ∗ ≤ C ( 1 + a − 1 / 2 ) e − a ∥ h ∥ ∗ . \left\|\Theta_t e^{a\Theta_t^2}h\right\|_*\le C(1+a^{-1/2})e^{-a}\|h\|_*. Θ t e a Θ t 2 h ∗ ≤ C ( 1 + a − 1/2 ) e − a ∥ h ∥ ∗ .
Apply this bound to the angular heat semigroup in f s f_s f s and contract the other averages. Integrating from the independent-phase endpoint to the original average gives
∥ f T − f 0 ∥ ∗ ≤ C ∫ 0 T ( 1 + 1 β ( C t + T − C t + s ) x x ) e − β 2 ( C t + T − C t + s ) x x ∥ ∑ i β ( e − ( s + t / 2 ) A ) i x ∂ i ( x ) f ∥ ∗ d s . (42) \|f_T-f_0\|_* \le C\int_0^T\left(1+\frac{1}{\sqrt{\beta(C_{t+T}-C_{t+s})_{xx}}}\right)e^{-\frac{\beta}{2}(C_{t+T}-C_{t+s})_{xx}}\left\|\sum_i\sqrt{\beta}(e^{-(s+t/2)A})_{ix}\partial_i^{(x)}f\right\|_*\,ds. \tag*{(42)} ∥ f T − f 0 ∥ ∗ ≤ C ∫ 0 T ( 1 + β ( C t + T − C t + s ) xx 1 ) e − 2 β ( C t + T − C t + s ) xx i ∑ β ( e − ( s + t /2 ) A ) i x ∂ i ( x ) f ∗ d s . (42)
We can improve the time decay in (42) because the phase-fixed derivatives sum to zero,
∑ i ∂ i ( x ) f = ∑ i ∂ i f − β ∑ i Q t ( i , x ) Θ t f = c t Θ t f − c t Θ t f = 0. \sum_i\partial_i^{(x)}f=\sum_i\partial_i f-\sqrt{\beta}\sum_i Q_t(i,x)\Theta_t f=c_t\Theta_t f-c_t\Theta_t f=0. i ∑ ∂ i ( x ) f = i ∑ ∂ i f − β i ∑ Q t ( i , x ) Θ t f = c t Θ t f − c t Θ t f = 0.
Subtracting the same value ( e − ( s + t / 2 ) A ) x x (e^{-(s+t/2)A})_{xx} ( e − ( s + t /2 ) A ) xx from every coefficient therefore leaves the sum unchanged. By evenness of the kernel about x x x , the resulting difference is half a centred second spatial difference. Telescoping twice along a shortest path and using Lemma A.1 gives
∣ ( e − ( s + t / 2 ) A ) i x − ( e − ( s + t / 2 ) A ) x x ∣ = 1 2 ∣ ( e − ( s + t / 2 ) A ) i x + ( e − ( s + t / 2 ) A ) 2 x − i , x − 2 ( e − ( s + t / 2 ) A ) x x ∣ ≤ C d ( i , x ) 2 ( ℓ 2 + s ) 2 , \left|(e^{-(s+t/2)A})_{ix}-(e^{-(s+t/2)A})_{xx}\right|
=\frac{1}{2}\left|(e^{-(s+t/2)A})_{ix}+(e^{-(s+t/2)A})_{2x-i,x}-2(e^{-(s+t/2)A})_{xx}\right|
\le C\frac{d(i,x)^2}{(\ell^2+s)^2}, ( e − ( s + t /2 ) A ) i x − ( e − ( s + t /2 ) A ) xx = 2 1 ( e − ( s + t /2 ) A ) i x + ( e − ( s + t /2 ) A ) 2 x − i , x − 2 ( e − ( s + t /2 ) A ) xx ≤ C ( ℓ 2 + s ) 2 d ( i , x ) 2 ,
∥ ∑ i β ( e − ( s + t / 2 ) A ) i x ∂ i ( x ) f ∥ ∗ = ∥ ∑ i β [ ( e − ( s + t / 2 ) A ) i x − ( e − ( s + t / 2 ) A ) x x ] ∂ i ( x ) f ∥ ∗ ≤ C β ( ℓ 2 + s ) 2 ∑ i d ( i , x ) 2 ∥ ∂ i ( x ) f ∥ ∗ ≤ C β , μ ℓ 4 ( ℓ 2 + s ) 2 sup i e μ d ( i , x ) / ℓ ∥ ∂ i ( x ) f ∥ ∗ . \begin{aligned}
\left\|\sum_i\sqrt{\beta}(e^{-(s+t/2)A})_{ix}\partial_i^{(x)}f\right\|_*
&=\left\|\sum_i\sqrt{\beta}\left[(e^{-(s+t/2)A})_{ix}-(e^{-(s+t/2)A})_{xx}\right]\partial_i^{(x)}f\right\|_*\\
&\le\frac{C_\beta}{(\ell^2+s)^2}\sum_i d(i,x)^2\|\partial_i^{(x)}f\|_*\\
&\le\frac{C_{\beta,\mu}\ell^4}{(\ell^2+s)^2}\sup_i e^{\mu d(i,x)/\ell}\|\partial_i^{(x)}f\|_*.
\end{aligned} i ∑ β ( e − ( s + t /2 ) A ) i x ∂ i ( x ) f ∗ = i ∑ β [ ( e − ( s + t /2 ) A ) i x − ( e − ( s + t /2 ) A ) xx ] ∂ i ( x ) f ∗ ≤ ( ℓ 2 + s ) 2 C β i ∑ d ( i , x ) 2 ∥ ∂ i ( x ) f ∥ ∗ ≤ ( ℓ 2 + s ) 2 C β , μ ℓ 4 i sup e μ d ( i , x ) / ℓ ∥ ∂ i ( x ) f ∥ ∗ .
Here ∑ i d ( i , x ) 2 e − μ d ( i , x ) / ℓ ≤ 4 ∑ n ≥ 0 ( n + 1 ) n 2 e − μ n / ℓ ≤ C μ ℓ 4 \sum_i d(i,x)^2e^{-\mu d(i,x)/\ell}\le4\sum_{n\ge0}(n+1)n^2e^{-\mu n/\ell}\le C_\mu\ell^4 ∑ i d ( i , x ) 2 e − μ d ( i , x ) / ℓ ≤ 4 ∑ n ≥ 0 ( n + 1 ) n 2 e − μ n / ℓ ≤ C μ ℓ 4 . The factor ( ℓ 2 + s ) − 2 (\ell^2+s)^{-2} ( ℓ 2 + s ) − 2 must compensate for the phase damping lost up to time s s s . The covariance logarithm bounds that loss, and the lower heat diagonal controls the remaining circle variance,
e β 2 ( C t + s − C t ) x x ≤ C β ( 1 + s / ℓ 2 ) β / ( 8 π ) , e^{\frac{\beta}{2}(C_{t+s}-C_t)_{xx}}\le C_\beta(1+s/\ell^2)^{\beta/(8\pi)}, e 2 β ( C t + s − C t ) xx ≤ C β ( 1 + s / ℓ 2 ) β / ( 8 π ) ,
β ( C t + T − C t + s ) x x = β ∫ t + s t + T ( e − u A ) x x d u ≥ c β T − s ℓ 2 . \beta(C_{t+T}-C_{t+s})_{xx}=\beta\int_{t+s}^{t+T}(e^{-uA})_{xx}\,du\ge c_\beta\frac{T-s}{\ell^2}. β ( C t + T − C t + s ) xx = β ∫ t + s t + T ( e − u A ) xx d u ≥ c β ℓ 2 T − s .
Take e − β ( C t + T − C t ) x x / 2 e^{-\beta(C_{t+T}-C_t)_{xx}/2} e − β ( C t + T − C t ) xx /2 out of (42) and set u = 1 + s / ℓ 2 u=1+s/\ell^2 u = 1 + s / ℓ 2 , so d s = ℓ 2 d u ds=\ell^2du d s = ℓ 2 d u and T − s = ℓ 2 ( M 2 − u ) T-s=\ell^2(M^2-u) T − s = ℓ 2 ( M 2 − u ) . Then
∫ 0 T ( 1 + 1 β ( C t + T − C t + s ) x x ) e β 2 ( C t + s − C t ) x x ℓ 4 ( ℓ 2 + s ) 2 d s ≤ C β ℓ 2 ∫ 1 M 2 u β / ( 8 π ) − 2 ( 1 + M 2 M 2 − u ) d u . \begin{aligned}
\int_0^T\left(1+\frac{1}{\sqrt{\beta(C_{t+T}-C_{t+s})_{xx}}}\right)e^{\frac{\beta}{2}(C_{t+s}-C_t)_{xx}}\frac{\ell^4}{(\ell^2+s)^2}\,ds
&\le C_\beta\ell^2\int_1^{M^2}u^{\beta/(8\pi)-2}\left(1+\sqrt{\frac{M^2}{M^2-u}}\right)\,du.
\end{aligned} ∫ 0 T ( 1 + β ( C t + T − C t + s ) xx 1 ) e 2 β ( C t + s − C t ) xx ( ℓ 2 + s ) 2 ℓ 4 d s ≤ C β ℓ 2 ∫ 1 M 2 u β / ( 8 π ) − 2 ( 1 + M 2 − u M 2 ) d u .
The second spatial difference gives the power β / ( 8 π ) − 2 \beta/(8\pi)-2 β / ( 8 π ) − 2 , which is integrable at infinity. Near the upper endpoint u = M 2 u=M^2 u = M 2 the angular variance tends to zero, and differentiation of its heat kernel costs the inverse square root of that variance. This singularity is also integrable. To separate the two estimates, split at M 2 / 2 M^2/2 M 2 /2 and set v = u / M 2 v=u/M^2 v = u / M 2 in the upper half,
∫ 1 M 2 u β / ( 8 π ) − 2 ( 1 + M 2 M 2 − u ) d u ≤ ( 1 + 2 ) ∫ 1 ∞ u β / ( 8 π ) − 2 d u + C β M β / ( 4 π ) − 2 ∫ 1 / 2 1 ( 1 + ( 1 − v ) − 1 / 2 ) d v ≤ 1 + 2 1 − β / ( 8 π ) + C β M β / ( 4 π ) − 2 ≤ C β 1 − β / ( 8 π ) . (43) \begin{aligned}
\int_{1}^{M^2} u^{\beta/(8\pi)-2}\left(1+\sqrt{\frac{M^2}{M^2-u}}\right)\,du
&\le(1+\sqrt{2})\int_{1}^{\infty}u^{\beta/(8\pi)-2}\,du+C_{\beta}M^{\beta/(4\pi)-2}\int_{1/2}^{1}\left(1+(1-v)^{-1/2}\right)\,dv\\
&\le\frac{1+\sqrt{2}}{1-\beta/(8\pi)}+C_{\beta}M^{\beta/(4\pi)-2}\le\frac{C_{\beta}}{1-\beta/(8\pi)}. \tag*{(43)}
\end{aligned} ∫ 1 M 2 u β / ( 8 π ) − 2 ( 1 + M 2 − u M 2 ) d u ≤ ( 1 + 2 ) ∫ 1 ∞ u β / ( 8 π ) − 2 d u + C β M β / ( 4 π ) − 2 ∫ 1/2 1 ( 1 + ( 1 − v ) − 1/2 ) d v ≤ 1 − β / ( 8 π ) 1 + 2 + C β M β / ( 4 π ) − 2 ≤ 1 − β / ( 8 π ) C β . (43)
Substitution in (42) yields
∥ f T − f 0 ∥ ∗ ≤ C β , μ ℓ 2 e − β 2 ( C t + T − C t ) x x sup i e μ d ( i , x ) / ℓ ∥ ∂ i ( x ) f ∥ ∗ . (44) \|f_T-f_0\|_* \le C_{\beta,\mu}\ell^2 e^{-\frac{\beta}{2}(C_{t+T}-C_t)_{xx}}\sup_i e^{\mu d(i,x)/\ell}\|\partial_i^{(x)}f\|_*. \tag*{(44)} ∥ f T − f 0 ∥ ∗ ≤ C β , μ ℓ 2 e − 2 β ( C t + T − C t ) xx i sup e μ d ( i , x ) / ℓ ∥ ∂ i ( x ) f ∥ ∗ . (44)
The bound on f 0 f_0 f 0 then proves (41) . Convolution approximation gives the same conclusion for C 1 C^1 C 1 functions, since the bounds above use only the first derivatives in the statement. □
To pass from the Gaussian average to E 0 g x \mathbf{E}_0g_x E 0 g x , we compare at the actual conditional mean. A small Hessian controls fluctuations, but not the displacement of this mean. The displacement is harmless because the preceding estimate is uniform in the background.
Lemma 3.5. In the heat regime, let F ∈ C 2 F\in C^2 F ∈ C 2 , 0 < s ≤ T 0<s\le T 0 < s ≤ T and s ∥ Hess F ∥ ∞ , o p ≤ 1 / 2 s\|\operatorname{Hess}F\|_{\infty,\mathrm{op}}\le1/2 s ∥ Hess F ∥ ∞ , op ≤ 1/2 . For bounded C 1 C^1 C 1 functions f f f with bounded first derivatives,
∥ T F , C s f ∥ ∞ ≤ ∥ E C s f ( ⋅ + ζ ) ∥ ∞ + C s ∥ Hess F ∥ ∞ , o p ∑ i ∥ ∂ i f ∥ ∞ . (45) \|T_{F,C_s}f\|_\infty\le\|\mathbf{E}_{C_s}f(\mathord\cdot+\zeta)\|_\infty+C_s\|\operatorname{Hess}F\|_{\infty,\mathrm{op}}\sum_i\|\partial_i f\|_\infty. \tag*{(45)} ∥ T F , C s f ∥ ∞ ≤ ∥ E C s f ( ⋅ + ζ ) ∥ ∞ + C s ∥ Hess F ∥ ∞ , op i ∑ ∥ ∂ i f ∥ ∞ . (45) At each background, let E \mathbf{E} E denote its conditional expectation. For each direction w w w , comparison at its actual mean gives
∣ E ∂ w f ( ψ ) − E C s ∂ w f ( E ψ + ζ ) ∣ ≤ C s ∥ Hess F ∥ ∞ , o p ( w , C s − 1 w ) ∑ i ∥ ∂ i f ∥ ∞ . (46) |\mathbf{E}\partial_wf(\psi)-\mathbf{E}_{C_s}\partial_wf(\mathbf{E}\psi+\zeta)|\le C_s\|\operatorname{Hess}F\|_{\infty,\mathrm{op}}\sqrt{(w,C_s^{-1}w)}\sum_i\|\partial_i f\|_\infty. \tag*{(46)} ∣ E ∂ w f ( ψ ) − E C s ∂ w f ( E ψ + ζ ) ∣ ≤ C s ∥ Hess F ∥ ∞ , op ( w , C s − 1 w ) i ∑ ∥ ∂ i f ∥ ∞ . (46) If F F F is neutral and f f f has period 2 π / c t 2\pi/c_t 2 π / c t under common shifts and zero shift average, then
∥ T F , C s f ∥ ∗ ≤ ∥ E C s f ( ⋅ + ζ ) ∥ ∗ + C s ∥ Hess F ∥ ∞ , o p ∑ i ∥ ∂ i f ∥ ∗ . \|T_{F,C_s}f\|_*\le\|\mathbf{E}_{C_s}f(\mathord\cdot+\zeta)\|_*+C_s\|\operatorname{Hess}F\|_{\infty,\mathrm{op}}\sum_i\|\partial_i f\|_*. ∥ T F , C s f ∥ ∗ ≤ ∥ E C s f ( ⋅ + ζ ) ∥ ∗ + C s ∥ Hess F ∥ ∞ , op i ∑ ∥ ∂ i f ∥ ∗ . The constants are independent of M M M , cutoff and volume.
Proof. Fix the background φ \varphi φ . Since C s ≤ sid C_s\le\operatorname{sid} C s ≤ sid and s ∥ Hess F ∥ ∞ , o p ≤ 1 / 2 s\|\operatorname{Hess}F\|_{\infty,\mathrm{op}}\le1/2 s ∥ Hess F ∥ ∞ , op ≤ 1/2 , the conditional Hamiltonian has Hessian bounded below by
C s − 1 + Hess F ≥ C s − 1 − ∥ Hess F ∥ ∞ , o p id ≥ ( 1 − s ∥ Hess F ∥ ∞ , o p ) C s − 1 ≥ 1 2 C s − 1 . C_s^{-1}+\operatorname{Hess}F\ge C_s^{-1}-\|\operatorname{Hess}F\|_{\infty,\mathrm{op}}\operatorname{id}\ge(1-s\|\operatorname{Hess}F\|_{\infty,\mathrm{op}})C_s^{-1}\ge\frac{1}{2}C_s^{-1}. C s − 1 + Hess F ≥ C s − 1 − ∥ Hess F ∥ ∞ , op id ≥ ( 1 − s ∥ Hess F ∥ ∞ , op ) C s − 1 ≥ 2 1 C s − 1 .
Brascamp–Lieb gives Var ( h ) ≤ 2 E ( ∇ h , C s ∇ h ) \operatorname{Var}(h)\le2\mathbf{E}(\nabla h,C_s\nabla h) Var ( h ) ≤ 2 E ( ∇ h , C s ∇ h ) . Integrating the gradient of the conditional density gives the mean,
0 = E [ ∇ F + C s − 1 ( ψ − φ ) ] , E ψ − φ = − C s E ∇ F . 0=\mathbf{E}[\nabla F+C_s^{-1}(\psi-\varphi)],\qquad\mathbf{E}\psi-\varphi=-C_s\mathbf{E}\nabla F. 0 = E [ ∇ F + C s − 1 ( ψ − φ )] , E ψ − φ = − C s E ∇ F .
Use the Ornstein–Uhlenbeck interpolation about this mean, [2 (2.7.3), (2.7.5)] ,
f u ( ψ ) = E C s f ( E ψ + e − u ( ψ − E ψ ) + 1 − e − 2 u ζ ) . f_u(\psi)=\mathbf{E}_{C_s}f\left(\mathbf{E}\psi+e^{-u}(\psi-\mathbf{E}\psi)+\sqrt{1-e^{-2u}}\,\zeta\right). f u ( ψ ) = E C s f ( E ψ + e − u ( ψ − E ψ ) + 1 − e − 2 u ζ ) .
It joins f f f to the constant E C s f ( E ψ + ζ ) \mathbb E_{C_s}f(\mathbb E\psi+\zeta) E C s f ( E ψ + ζ ) , and ∥ ∂ i f u ∥ ∞ ≤ e − u ∥ ∂ i f ∥ ∞ \|\partial_i f_u\|_\infty\le e^{-u}\|\partial_i f\|_\infty ∥ ∂ i f u ∥ ∞ ≤ e − u ∥ ∂ i f ∥ ∞ . Conditional integration by parts replaces the second derivatives of f u f_u f u by the gradient of the conditional Hamiltonian,
E Tr ( C s Hess f u ) = E ( C s ∇ F + ψ − φ , ∇ f u ) . \mathbb E\operatorname{Tr}(C_s\operatorname{Hess}f_u)=\mathbb E(C_s\nabla F+\psi-\varphi,\nabla f_u). E Tr ( C s Hess f u ) = E ( C s ∇ F + ψ − φ , ∇ f u ) .
Substitution into the Ornstein–Uhlenbeck equation therefore gives
d d u E f u = E [ Tr ( C s Hess f u ) − ( ψ − E ψ , ∇ f u ) ] = E ( C s ∇ F + E ψ − φ , ∇ f u ) = ∑ i Cov ( ( C s e i , ∇ F ) , ∂ i f u ) . \begin{aligned}
\frac{d}{du}\mathbb E f_u
&=\mathbb E[\operatorname{Tr}(C_s\operatorname{Hess}f_u)-(\psi-\mathbb E\psi,\nabla f_u)]\\
&=\mathbb E(C_s\nabla F+\mathbb E\psi-\varphi,\nabla f_u)\\
&=\sum_i\operatorname{Cov}((C_se_i,\nabla F),\partial_i f_u).
\end{aligned} d u d E f u = E [ Tr ( C s Hess f u ) − ( ψ − E ψ , ∇ f u )] = E ( C s ∇ F + E ψ − φ , ∇ f u ) = i ∑ Cov (( C s e i , ∇ F ) , ∂ i f u ) .
Centering at the conditional mean has canceled the mean force. Brascamp–Lieb now bounds its fluctuation by Hess F F F . The image of the coordinate vector e i e_i e i under C s C_s C s has a bound independent of the number of sites,
∥ C s e i ∥ 2 = ∥ ∫ 0 s e − v A e i d v ∥ 2 ≤ ∫ 0 s ( e − 2 v A ) i i d v ≤ C ∫ 0 s ( 1 + v ) − 1 / 2 d v ≤ C s . \|C_se_i\|_2=\left\|\int_0^s e^{-vA}e_i\,dv\right\|_2\le\int_0^s\sqrt{(e^{-2vA})_{ii}}\,dv\le C\int_0^s(1+v)^{-1/2}\,dv\le C\sqrt{s}. ∥ C s e i ∥ 2 = ∫ 0 s e − v A e i d v 2 ≤ ∫ 0 s ( e − 2 v A ) ii d v ≤ C ∫ 0 s ( 1 + v ) − 1/2 d v ≤ C s .
Consequently,
Var ( ( C s e i , ∇ F ) ) ≤ 2 E ( Hess F C s e i , C s Hess F C s e i ) ≤ 2 s ∥ Hess F ∥ ∞ , o p 2 ∥ C s e i ∥ 2 2 ≤ C s 2 ∥ Hess F ∥ ∞ , o p 2 . \begin{aligned}
\operatorname{Var}((C_se_i,\nabla F))
&\le2\mathbb E(\operatorname{Hess}F C_se_i,C_s\operatorname{Hess}F C_se_i)\\
&\le2s\|\operatorname{Hess}F\|_{\infty,\mathrm{op}}^2\|C_se_i\|_2^2\\
&\le Cs^2\|\operatorname{Hess}F\|_{\infty,\mathrm{op}}^2.
\end{aligned} Var (( C s e i , ∇ F )) ≤ 2 E ( Hess F C s e i , C s Hess F C s e i ) ≤ 2 s ∥ Hess F ∥ ∞ , op 2 ∥ C s e i ∥ 2 2 ≤ C s 2 ∥ Hess F ∥ ∞ , op 2 .
Integrating the covariance identity and using Cauchy–Schwarz yields
∣ E f − E C s f ( E ψ + ζ ) ∣ ≤ ∫ 0 ∞ ∑ i Var ( ( C s e i , ∇ F ) ) e − u ∥ ∂ i f ∥ ∞ d u ≤ C s ∥ Hess F ∥ ∞ , o p ∑ i ∥ ∂ i f ∥ ∞ . \begin{aligned}
|\mathbb E f-\mathbb E_{C_s}f(\mathbb E\psi+\zeta)|
&\le\int_0^\infty\sum_i\sqrt{\operatorname{Var}((C_se_i,\nabla F))}e^{-u}\|\partial_i f\|_\infty\,du\\
&\le Cs\|\operatorname{Hess}F\|_{\infty,\mathrm{op}}\sum_i\|\partial_i f\|_\infty.
\end{aligned} ∣ E f − E C s f ( E ψ + ζ ) ∣ ≤ ∫ 0 ∞ i ∑ Var (( C s e i , ∇ F )) e − u ∥ ∂ i f ∥ ∞ d u ≤ C s ∥ Hess F ∥ ∞ , op i ∑ ∥ ∂ i f ∥ ∞ .
Taking the supremum over the background proves the scalar comparison. To apply this calculation to ∂ w f \partial_w f ∂ w f , integrate the derivative in direction w w w against the Gaussian density. The observable then still requires only one derivative,
∂ i ( ∂ w f ) u ( ψ ) = e − u 1 − e − 2 u E C s [ ( C s − 1 w , ζ ) ∂ i f ( E ψ + e − u ( ψ − E ψ ) + 1 − e − 2 u ζ ) ] , \partial_i(\partial_w f)_u(\psi)=\frac{e^{-u}}{\sqrt{1-e^{-2u}}}\mathbb E_{C_s}\left[(C_s^{-1}w,\zeta)\partial_i f(\mathbb E\psi+e^{-u}(\psi-\mathbb E\psi)+\sqrt{1-e^{-2u}}\zeta)\right], ∂ i ( ∂ w f ) u ( ψ ) = 1 − e − 2 u e − u E C s [ ( C s − 1 w , ζ ) ∂ i f ( E ψ + e − u ( ψ − E ψ ) + 1 − e − 2 u ζ ) ] ,
∥ ∂ i ( ∂ w f ) u ∥ ∞ ≤ e − u 1 − e − 2 u E C s ( C s − 1 w , ζ ) 2 ∥ ∂ i f ∥ ∞ = e − u 1 − e − 2 u ( w , C s − 1 w ) ∥ ∂ i f ∥ ∞ . \begin{aligned}
\|\partial_i(\partial_w f)_u\|_\infty\le\frac{e^{-u}}{\sqrt{1-e^{-2u}}}\sqrt{\mathbb E_{C_s}(C_s^{-1}w,\zeta)^2}\|\partial_i f\|_\infty\\
&=\frac{e^{-u}}{\sqrt{1-e^{-2u}}}\sqrt{(w,C_s^{-1}w)}\|\partial_i f\|_\infty.
\end{aligned} ∥ ∂ i ( ∂ w f ) u ∥ ∞ ≤ 1 − e − 2 u e − u E C s ( C s − 1 w , ζ ) 2 ∥ ∂ i f ∥ ∞ = 1 − e − 2 u e − u ( w , C s − 1 w ) ∥ ∂ i f ∥ ∞ .
Since e − u / 1 − e − 2 u e^{-u}/\sqrt{1-e^{-2u}} e − u / 1 − e − 2 u is integrable over ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , the variance estimate proves (46) . The same first-derivative bounds justify convolution approximation for C 1 C^1 C 1 functions. If F F F is neutral, a change of variables in its conditional density gives
T F , C s [ f ( ⋅ + θ 1 ) ] ( φ ) = T F , C s f ( φ + θ 1 ) . T_{F,C_s}[f(\cdot+\theta1)](\varphi)=T_{F,C_s}f(\varphi+\theta1). T F , C s [ f ( ⋅ + θ 1 )] ( φ ) = T F , C s f ( φ + θ 1 ) .
Integrating this identity against the heat kernel on the circle shows that T F , C s T_{F,C_s} T F , C s commutes with e a Θ t 2 e^{a\Theta_t^2} e a Θ t 2 . Apply the scalar estimate to e a Θ t 2 f e^{a\Theta_t^2}f e a Θ t 2 f , multiply by e a e^a e a and take the supremum over a ≥ 0 a\ge0 a ≥ 0 . Gaussian convolution and the first derivatives commute with the same semigroup, giving the estimate in ∥ ⋅ ∥ ∗ \|\cdot\|_* ∥ ⋅ ∥ ∗ .
Estimates at the next scale
We first estimate the value by comparing G x G_x G x with the average M 2 E 0 g x M^2\mathbb{E}_0g_x M 2 E 0 g x under N N N . The effect of adding λ R \lambda R λ R is quadratic in the coupling, as differentiation of the numerator and normalization shows,
d d λ E λ g x = − E λ ( g x R ) + ( E λ g x ) ( E λ R ) = − ℓ − 2 ∑ y Cov λ ( g x , g y ) . \frac{d}{d\lambda}\mathbb{E}_\lambda g_x=-\mathbb{E}_\lambda(g_xR)+(\mathbb{E}_\lambda g_x)(\mathbb{E}_\lambda R)=-\ell^{-2}\sum_y\operatorname{Cov}_\lambda(g_x,g_y). d λ d E λ g x = − E λ ( g x R ) + ( E λ g x ) ( E λ R ) = − ℓ − 2 y ∑ Cov λ ( g x , g y ) .
Integrating in λ \lambda λ gives
G x − M 2 E 0 g x = M 2 ∫ 0 1 ( E λ g x − E 0 g x ) d λ = − M 2 ℓ − 2 ∑ y ∫ 0 1 ( 1 − λ ) Cov λ ( g x , g y ) d λ . (47) \begin{aligned}
G_x-M^2\mathbb{E}_0g_x=M^2\int_0^1(\mathbb{E}_\lambda g_x-\mathbb{E}_0g_x)\,d\lambda\\
&=-M^2\ell^{-2}\sum_y\int_0^1(1-\lambda)\operatorname{Cov}_\lambda(g_x,g_y)\,d\lambda. \tag*{(47)}
\end{aligned} G x − M 2 E 0 g x = M 2 ∫ 0 1 ( E λ g x − E 0 g x ) d λ = − M 2 ℓ − 2 y ∑ ∫ 0 1 ( 1 − λ ) Cov λ ( g x , g y ) d λ . (47)
The product covariance bound and local gradients give
∣ Cov λ ( g x , g y ) ∣ ≤ C M ℓ ′ 2 ζ ℓ 2 ℓ − 2 e − 6 μ d ( x , y ) / ℓ ′ ≤ C M ζ ℓ 2 e − 6 μ d ( x , y ) / ℓ ′ , \left|\operatorname{Cov}_\lambda(g_x,g_y)\right|\leq CM\ell^{\prime2}\zeta_\ell^2\ell^{-2}e^{-6\mu d(x,y)/\ell'}\leq CM\zeta_\ell^2e^{-6\mu d(x,y)/\ell'}, ∣ Cov λ ( g x , g y ) ∣ ≤ C M ℓ ′2 ζ ℓ 2 ℓ − 2 e − 6 μ d ( x , y ) / ℓ ′ ≤ C M ζ ℓ 2 e − 6 μ d ( x , y ) / ℓ ′ ,
∥ G x − M 2 E 0 g x ∥ ∞ ≤ C M M 2 ℓ − 2 ζ ℓ 2 ∑ y e − 6 μ d ( x , y ) / ℓ ′ ≤ C M ζ ℓ 2 . \left\|G_x-M^2\mathbb{E}_0g_x\right\|_\infty\leq C_M M^2\ell^{-2}\zeta_\ell^2\sum_y e^{-6\mu d(x,y)/\ell'}\leq C_M\zeta_\ell^2. G x − M 2 E 0 g x ∞ ≤ C M M 2 ℓ − 2 ζ ℓ 2 y ∑ e − 6 μ d ( x , y ) / ℓ ′ ≤ C M ζ ℓ 2 .
For the average under N N N , the two errors in Lemmas 3.4 and 3.5 are, by (26) and (27) ,
ℓ 2 sup i e μ d ( i , x ) / ℓ ∥ ∂ i ( x ) g x ∥ ∗ ≤ C B ζ ℓ 2 , \ell^2\sup_i e^{\mu d(i,x)/\ell}\left\|\partial_i^{(x)}g_x\right\|_*\leq CB\zeta_\ell^2, ℓ 2 i sup e μ d ( i , x ) / ℓ ∂ i ( x ) g x ∗ ≤ C B ζ ℓ 2 ,
T ∥ Hess N ∥ ∞ , o p ∑ i ∥ ∂ i g x ∥ ∗ ≤ C M , K T H ℓ ℓ − 2 ζ ℓ ≤ C M ζ ℓ 2 . T\|\operatorname{Hess}N\|_{\infty,\mathrm{op}}\sum_i\|\partial_i g_x\|_*\leq C_{M,K}TH_\ell\ell^{-2}\zeta_\ell\leq C_M\zeta_\ell^2. T ∥ Hess N ∥ ∞ , op i ∑ ∥ ∂ i g x ∥ ∗ ≤ C M , K T H ℓ ℓ − 2 ζ ℓ ≤ C M ζ ℓ 2 .
Neutrality of N N N makes T N , C T T_N,C_T T N , C T commute with shift averaging. The relation Q T 1 = ( c t + T / c t ) 1 Q_T1=(c_{t+T}/c_t)1 Q T 1 = ( c t + T / c t ) 1 identifies the old and new periods, and invertibility of Q T Q_T Q T preserves the field supremum in ∥ ⋅ ∥ ∗ \|\cdot\|_* ∥ ⋅ ∥ ∗ . Thus Π E 0 g x = 0 \Pi\mathbb{E}_0g_x=0 Π E 0 g x = 0 . We remove the period average only from the quadratic correction, where the constant in norm equivalence is harmless. The leading term in the bound retains the exact Gaussian multiplier,
g x ′ = M 2 E 0 g x + ( id − Π ) ( G x − M 2 E 0 g x ) , g_x'=M^2\mathbb{E}_0g_x+(\operatorname{id}-\Pi)(G_x-M^2\mathbb{E}_0g_x), g x ′ = M 2 E 0 g x + ( id − Π ) ( G x − M 2 E 0 g x ) ,
∥ g x ′ ∥ ∗ ≤ M 2 ∥ E 0 g x ∥ ∗ + C M ζ ℓ 2 ≤ M 2 e − β 2 ( C t + T − C t ) x x [ 2 ℓ ′ 2 ∣ z ℓ ∣ + C B ζ ℓ 2 ] + C M ζ ℓ 2 (48) \begin{aligned}
\|g_x'\|_*\leq M^2\|\mathbb{E}_0g_x\|_*+C_M\zeta_\ell^2 \\
&\leq M^2e^{-\frac{\beta}{2}(C_{t+T}-C_t)_{xx}}\left[2\ell^{\prime2}|z_\ell|+CB\zeta_\ell^2\right]+C_M\zeta_\ell^2 \tag*{(48)}
\end{aligned} ∥ g x ′ ∥ ∗ ≤ M 2 ∥ E 0 g x ∥ ∗ + C M ζ ℓ 2 ≤ M 2 e − 2 β ( C t + T − C t ) xx [ 2 ℓ ′2 ∣ z ℓ ∣ + C B ζ ℓ 2 ] + C M ζ ℓ 2 (48)
≤ 2 ℓ ′ 2 ∣ z ℓ ∣ + [ C M α B + C M ] ζ ℓ 2 . \leq2\ell^{\prime2}|z_\ell|+[CM^\alpha B+C_M]\zeta_\ell^2. ≤ 2 ℓ ′2 ∣ z ℓ ∣ + [ C M α B + C M ] ζ ℓ 2 .
Indeed, the linear term agrees with the Gaussian coefficient at the next scale,
M 2 e − β 2 ( C t + T − C t ) x x ⋅ 2 ℓ 2 ∣ z ℓ ∣ = 2 ℓ ′ 2 ∣ z ℓ ∣ . M^2e^{-\frac{\beta}{2}(C_{t+T}-C_t)_{xx}}\cdot2\ell^2|z_\ell|=2\ell^{\prime2}|z_\ell|. M 2 e − 2 β ( C t + T − C t ) xx ⋅ 2 ℓ 2 ∣ z ℓ ∣ = 2 ℓ ′2 ∣ z ℓ ∣.
For the quadratic term we use e − β ( C t + T − C t ) x x / 2 ≤ C β M − β / ( 4 π ) e^{-\beta(C_{t+T}-C_t)_{xx}/2}\leq C_\beta M^{-\beta/(4\pi)} e − β ( C t + T − C t ) xx /2 ≤ C β M − β / ( 4 π ) .
For the derivative at x x x , period averaging commutes with derivatives and contracts the supremum norm. The density estimate gives
∥ ∂ x g x ′ ∥ ∞ ≤ 2 ∥ ∂ x G x ∥ ∞ ≤ 2 2 M 2 ∥ g x ∥ ∞ ( Q T e x , C T − 1 Q T e x ) ≤ C β M 2 ζ ℓ ℓ ′ − 2 = C β M − α ζ ℓ ℓ ′ − 2 . (49) \begin{aligned}
\|\partial_xg_x'\|_\infty\leq2\|\partial_xG_x\|_\infty\leq2\sqrt{2}M^2\|g_x\|_\infty\sqrt{(Q_Te_x,C_T^{-1}Q_Te_x)} \\
&\leq C_\beta M^2\zeta_\ell\ell^{\prime-2}=C_\beta M^{-\alpha}\zeta_\ell\ell^{\prime-2}. \tag*{(49)}
\end{aligned} ∥ ∂ x g x ′ ∥ ∞ ≤ 2∥ ∂ x G x ∥ ∞ ≤ 2 2 M 2 ∥ g x ∥ ∞ ( Q T e x , C T − 1 Q T e x ) ≤ C β M 2 ζ ℓ ℓ ′ − 2 = C β M − α ζ ℓ ℓ ′ − 2 . (49)
This bound is independent of the previous K K K , H H H , B B B because ∥ g x ∥ ∞ ≤ ( 2 C β + 1 ) ζ ℓ \|g_x\|_\infty\leq(2C_\beta+1)\zeta_\ell ∥ g x ∥ ∞ ≤ ( 2 C β + 1 ) ζ ℓ .
For the full Hessian, differentiate each scalar G x G_x G x before summing over x x x . This retains the distance to x x x in both derivative indices. With F = N + λ R F=N+\lambda R F = N + λ R , the conditional derivative formula gives
∂ j ∂ i E λ g x = ∂ j E λ [ Q T ∇ g x ] i − ∂ j Cov λ ( g x , [ Q T ∇ F ] i ) . \partial_j\partial_i\mathbb{E}_\lambda g_x=\partial_j\mathbb{E}_\lambda[Q_T\nabla g_x]_i-\partial_j\operatorname{Cov}_\lambda(g_x,[Q_T\nabla F]_i). ∂ j ∂ i E λ g x = ∂ j E λ [ Q T ∇ g x ] i − ∂ j Cov λ ( g x , [ Q T ∇ F ] i ) .
The first observable uses only one derivative of g x g_x g x ,
∥ [ Q T ∇ g x ] i ∥ ∞ ≤ ∑ a Q T ( i , a ) ∥ ∂ a g x ∥ ∞ ≤ C ( K + 1 ) ζ ℓ ′ − 2 ∑ a Q T ( i , a ) e − 6 μ d ( a , x ) / ℓ ≤ C ( K + 1 ) ζ ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ . \begin{aligned}
\lVert[Q_T\nabla g_x]_i\rVert_\infty\leq\sum_a Q_T(i,a)\lVert\partial_a g_x\rVert_\infty\\
&\leq C(K+1)\zeta\ell^{\prime-2}\sum_a Q_T(i,a)e^{-6\mu d(a,x)/\ell} \\
&\leq C(K+1)\zeta\ell^{\prime-2}e^{-6\mu d(i,x)/\ell'}.
\end{aligned} ∥[ Q T ∇ g x ] i ∥ ∞ ≤ a ∑ Q T ( i , a ) ∥ ∂ a g x ∥ ∞ ≤ C ( K + 1 ) ζ ℓ ′ − 2 a ∑ Q T ( i , a ) e − 6 μ d ( a , x ) / ℓ ≤ C ( K + 1 ) ζ ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ .
We put its next derivative on the conditional density. For the covariance we use (38) . Thus
∣ ∂ j E λ [ Q T ∇ g x ] i ∣ ≤ 2 ( Q T e j , C T − 1 Q T e j ) ∥ [ Q T ∇ g x ] i ∥ ∞ ≤ C ( K + 1 ) ζ ℓ ′ − 4 e − 6 μ d ( i , x ) / ℓ ′ , \begin{aligned}
\left|\partial_j\mathbb{E}_\lambda[Q_T\nabla g_x]_i\right|\leq\sqrt{2(Q_Te_j,C_T^{-1}Q_Te_j)}\lVert[Q_T\nabla g_x]_i\rVert_\infty\\
&\leq C(K+1)\zeta\ell^{\prime-4}e^{-6\mu d(i,x)/\ell'},
\end{aligned} ∣ ∂ j E λ [ Q T ∇ g x ] i ∣ ≤ 2 ( Q T e j , C T − 1 Q T e j ) ∥[ Q T ∇ g x ] i ∥ ∞ ≤ C ( K + 1 ) ζ ℓ ′ − 4 e − 6 μ d ( i , x ) / ℓ ′ ,
∣ ∂ j Cov λ ( g x , [ Q T ∇ F ] i ) ∣ ≤ C M , K H ζ 2 ℓ ′ − 4 e − 6 μ d ( i , x ) / ℓ ′ . \left|\partial_j\operatorname{Cov}_\lambda(g_x,[Q_T\nabla F]_i)\right|\leq C_{M,K}H\zeta^2\ell^{\prime-4}e^{-6\mu d(i,x)/\ell'}. ∣ ∂ j Cov λ ( g x , [ Q T ∇ F ] i ) ∣ ≤ C M , K H ζ 2 ℓ ′ − 4 e − 6 μ d ( i , x ) / ℓ ′ .
Since H ζ ℓ ≤ 1 H\zeta\ell\leq1 H ζ ℓ ≤ 1 , integration in λ \lambda λ gives ∣ ∂ j ∂ i G x ∣ ≤ C M , K ζ ℓ ′ − 4 e − 6 μ d ( i , x ) / ℓ ′ |\partial_j\partial_iG_x|\leq C_{M,K}\zeta\ell^{\prime-4}e^{-6\mu d(i,x)/\ell'} ∣ ∂ j ∂ i G x ∣ ≤ C M , K ζ ℓ ′ − 4 e − 6 μ d ( i , x ) / ℓ ′ , with M 2 M^2 M 2 absorbed in C M , K C_{M,K} C M , K . Gaussian convolution makes G x G_x G x smooth in its background. Symmetry of this scalar Hessian gives the same bound with i , j i,j i , j interchanged, and hence decay in both indices,
∣ ∂ i ∂ j G x ∣ ≤ C M , K ζ ℓ ′ − 4 min { e − 6 μ d ( i , x ) / ℓ ′ , e − 6 μ d ( j , x ) / ℓ ′ } ≤ C M , K ζ ℓ ′ − 4 e − 3 μ [ d ( i , x ) + d ( j , x ) ] / ℓ ′ . \begin{aligned}
|\partial_i\partial_jG_x|\leq C_{M,K}\zeta\ell^{\prime-4}\min\{e^{-6\mu d(i,x)/\ell'},e^{-6\mu d(j,x)/\ell'}\} \\
&\leq C_{M,K}\zeta\ell^{\prime-4}e^{-3\mu[d(i,x)+d(j,x)]/\ell'}.
\end{aligned} ∣ ∂ i ∂ j G x ∣ ≤ C M , K ζ ℓ ′ − 4 min { e − 6 μ d ( i , x ) / ℓ ′ , e − 6 μ d ( j , x ) / ℓ ′ } ≤ C M , K ζ ℓ ′ − 4 e − 3 μ [ d ( i , x ) + d ( j , x )] / ℓ ′ .
We can now sum directly with the weight needed for the next step. The triangle inequality gives
ℓ ′ − 2 ∑ x , j e μ d ( i , j ) / ( 2 ℓ ′ ) sup φ ′ ∣ ∂ i ∂ j G x ∣ ≤ C M , K ζ ℓ ′ − 6 ∑ x e − 5 μ d ( i , x ) / ( 2 ℓ ′ ) ∑ j e − 5 μ d ( j , x ) / ( 2 ℓ ′ ) ≤ C M , K ζ ℓ ′ − 2 . \begin{aligned}
\ell^{\prime-2}\sum_{x,j}e^{\mu d(i,j)/(2\ell')}\sup_{\varphi'}|\partial_i\partial_jG_x|
\leq C_{M,K}\zeta\ell^{\prime-6}\sum_xe^{-5\mu d(i,x)/(2\ell')}\sum_je^{-5\mu d(j,x)/(2\ell')} \\
&\leq C_{M,K}\zeta\ell^{\prime-2}.
\end{aligned} ℓ ′ − 2 x , j ∑ e μ d ( i , j ) / ( 2 ℓ ′ ) φ ′ sup ∣ ∂ i ∂ j G x ∣ ≤ C M , K ζ ℓ ′ − 6 x ∑ e − 5 μ d ( i , x ) / ( 2 ℓ ′ ) j ∑ e − 5 μ d ( j , x ) / ( 2 ℓ ′ ) ≤ C M , K ζ ℓ ′ − 2 .
Before projection, the full potential is the neutral convolution plus ℓ ′ − 2 ∑ x G x \ell^{\prime-2}\sum_xG_x ℓ ′ − 2 ∑ x G x . Combining the last bound with (39) therefore proves
sup i ∑ j e μ d ( i , j ) / ( 2 ℓ ′ ) sup φ ′ ∣ ∂ i ∂ j ( N ′ + ℓ ′ − 2 ∑ x g x ′ ) ∣ ≤ [ C μ H + C M , K + C β , μ , M H 2 ζ ℓ ] ζ ℓ ′ − 2 . (50) \begin{aligned}
\sup_i\sum_j e^{\mu d(i,j)/(2\ell')}\sup_{\varphi'}|\partial_i\partial_j(N'+\ell^{\prime-2}\sum_xg_x')| \\
&\leq[C_\mu H+C_{M,K}+C_{\beta,\mu,M}H^2\zeta\ell]\zeta\ell^{\prime-2}. \tag*{(50)}
\end{aligned} i sup j ∑ e μ d ( i , j ) / ( 2 ℓ ′ ) φ ′ sup ∣ ∂ i ∂ j ( N ′ + ℓ ′ − 2 x ∑ g x ′ ) ∣ ≤ [ C μ H + C M , K + C β , μ , M H 2 ζ ℓ ] ζ ℓ ′ − 2 . (50)
First derivatives of the local terms
To estimate the phase-fixed derivative of g x ′ g_x' g x ′ , we apply (44) to g x g_x g x over the first half of the time step. Applying it to ∂ i ( x ) g x \partial_i^{(x)}g_x ∂ i ( x ) g x instead would require second derivatives of g x g_x g x . We differentiate the remaining Gaussian density to estimate the derivative of the convolved error.
Lemma 3.6. Assume M ≥ max { 256 , 192 / α } M\geq\max\{256,192/\alpha\} M ≥ max { 256 , 192/ α } , (25) –(26) , and the scale smallness assumptions. Then, uniformly in field, cutoff and volume,
∥ ∂ i ( x ) ( G x − Π G x ) ∥ ∗ ≤ [ C M 3 α / 2 B + C M ( 1 + B α / 4 ) ] ζ 2 ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ . (51) \lVert\partial_i^{(x)}(G_x-\Pi G_x)\rVert_*\leq[CM^{3\alpha/2}B+C_M(1+B^{\alpha/4})]\zeta^2\ell^{\prime-2}e^{-6\mu d(i,x)/\ell'}. \tag*{(51)} ∥ ∂ i ( x ) ( G x − Π G x ) ∥ ∗ ≤ [ C M 3 α /2 B + C M ( 1 + B α /4 )] ζ 2 ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ . (51) The constant C C C is independent of M M M , K K K , H H H .
Proof. We first compensate a change at i i i by a change at x x x . This keeps the phase fixed using a localized direction. After estimating it, (54) will give the derivative ∂ i ( x ) \partial_i^{(x)} ∂ i ( x ) , whose compensation uses a common field shift. Under the convolution the localized direction is
w i = Q T e i − Q t + T ( i , x ) Q t + T ( x , x ) Q T e x , w_i=Q_T e_i-\frac{Q_{t+T}(i,x)}{Q_{t+T}(x,x)}Q_T e_x, w i = Q T e i − Q t + T ( x , x ) Q t + T ( i , x ) Q T e x ,
( Q t e x , w i ) = Q t + T ( i , x ) − Q t + T ( i , x ) Q t + T ( x , x ) Q t + T ( x , x ) = 0. (Q_t e_x,w_i)=Q_{t+T}(i,x)-\frac{Q_{t+T}(i,x)}{Q_{t+T}(x,x)}Q_{t+T}(x,x)=0. ( Q t e x , w i ) = Q t + T ( i , x ) − Q t + T ( x , x ) Q t + T ( i , x ) Q t + T ( x , x ) = 0.
Thus ∂ w i g x = ∑ a ( w i ) a ∂ a ( x ) g x \partial_{w_i}g_x=\sum_a(w_i)_a\partial_a^{(x)}g_x ∂ w i g x = ∑ a ( w i ) a ∂ a ( x ) g x , where the phase-fixed derivatives on the right use the incoming time t t t . With F = N + λ R F=N+\lambda R F = N + λ R , differentiation of the normalized conditional average gives
( ∂ i − Q t + T ( i , x ) Q t + T ( x , x ) ∂ x ) E λ g x = E λ ∂ w i g x − Cov λ ( g x , ∂ w i F ) . (52) \left(\partial_i-\frac{Q_{t+T}(i,x)}{Q_{t+T}(x,x)}\partial_x\right)\mathbb{E}_{\lambda}g_x=\mathbb{E}_{\lambda}\partial_{w_i}g_x-\operatorname{Cov}_{\lambda}(g_x,\partial_{w_i}F). \tag*{(52)} ( ∂ i − Q t + T ( x , x ) Q t + T ( i , x ) ∂ x ) E λ g x = E λ ∂ w i g x − Cov λ ( g x , ∂ w i F ) . (52)
The force covariance was estimated in (37) . The heat-kernel upper bound and diagonal lower bound give Q t + T ( i , x ) / Q t + T ( x , x ) ≤ C e − 6 μ d ( i , x ) / ℓ ′ Q_{t+T}(i,x)/Q_{t+T}(x,x)\le Ce^{-6\mu d(i,x)/\ell'} Q t + T ( i , x ) / Q t + T ( x , x ) ≤ C e − 6 μ d ( i , x ) / ℓ ′ , so
∣ Cov λ ( g x , ∂ w i F ) ∣ ≤ ∣ Cov λ ( g x , [ Q T ∇ F ] i ) ∣ + Q t + T ( i , x ) Q t + T ( x , x ) ∣ Cov λ ( g x , [ Q T ∇ F ] x ) ∣ ≤ C M ζ 2 ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ . \begin{aligned}
\left|\operatorname{Cov}_{\lambda}(g_x,\partial_{w_i}F)\right|\le\left|\operatorname{Cov}_{\lambda}(g_x,[Q_T\nabla F]_i)\right|+\frac{Q_{t+T}(i,x)}{Q_{t+T}(x,x)}\left|\operatorname{Cov}_{\lambda}(g_x,[Q_T\nabla F]_x)\right| \\
&\le CM\zeta^2\ell'^{-2}e^{-6\mu d(i,x)/\ell'}.
\end{aligned} ∣ Cov λ ( g x , ∂ w i F ) ∣ ≤ ∣ Cov λ ( g x , [ Q T ∇ F ] i ) ∣ + Q t + T ( x , x ) Q t + T ( i , x ) ∣ Cov λ ( g x , [ Q T ∇ F ] x ) ∣ ≤ C M ζ 2 ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ .
Only the first term of (52) still needs damping.
The incoming derivative bound gives stronger spatial decay than the conclusion requires. The heat-kernel estimates apply with rate 24 μ / α ≤ ρ 0 24\mu/\alpha\le\rho_0 24 μ / α ≤ ρ 0 , and M ≥ 192 / α M\ge192/\alpha M ≥ 192/ α ensures 24 μ / ( α ℓ ′ ) ≤ μ / ( 8 ℓ ) 24\mu/(\alpha\ell')\le\mu/(8\ell) 24 μ / ( α ℓ ′ ) ≤ μ / ( 8 ℓ ) , so this weight is absorbed by the incoming decay. Consequently,
∥ ∂ w i g x ∥ ∗ ≤ B ζ 2 ℓ ′ − 2 ∑ a [ Q T ( a , i ) + Q t + T ( i , x ) Q t + T ( x , x ) Q T ( a , x ) ] e − 6 μ d ( a , x ) / ℓ ≤ C B ζ 2 ℓ ′ − 2 ℓ ′ − 2 e − 24 μ d ( i , x ) / ( α ℓ ′ ) ∑ a e − 3 μ d ( a , x ) / ℓ ≤ C B ζ 2 ℓ ′ − 2 e − 24 μ d ( i , x ) / ( α ℓ ′ ) . (53) \begin{aligned}
\|\partial_{w_i}g_x\|_*\le B\zeta^2\ell'^{-2}\sum_a\left[Q_T(a,i)+\frac{Q_{t+T}(i,x)}{Q_{t+T}(x,x)}Q_T(a,x)\right]e^{-6\mu d(a,x)/\ell} \\
&\le CB\zeta^2\ell'^{-2}\ell'^{-2}e^{-24\mu d(i,x)/(\alpha\ell')}\sum_a e^{-3\mu d(a,x)/\ell} \\
&\le CB\zeta^2\ell'^{-2}e^{-24\mu d(i,x)/(\alpha\ell')}. \tag*{(53)}
\end{aligned} ∥ ∂ w i g x ∥ ∗ ≤ B ζ 2 ℓ ′ − 2 a ∑ [ Q T ( a , i ) + Q t + T ( x , x ) Q t + T ( i , x ) Q T ( a , x ) ] e − 6 μ d ( a , x ) / ℓ ≤ C B ζ 2 ℓ ′ − 2 ℓ ′ − 2 e − 24 μ d ( i , x ) / ( α ℓ ′ ) a ∑ e − 3 μ d ( a , x ) / ℓ ≤ C B ζ 2 ℓ ′ − 2 e − 24 μ d ( i , x ) / ( α ℓ ′ ) . (53)
Put S = T / 2 S=T/2 S = T /2 and apply (44) to g x g_x g x , with the comparison time T T T replaced by S S S . Let f 0 f_0 f 0 be the initial endpoint of that interpolation, where the phase is averaged independently. The incoming bound gives
∥ E C S g x − f 0 ∥ ∗ ≤ C B ζ 2 ℓ ′ e − β ( C t + S − C t ) x x / 2 . \|\mathbb{E}_{C_S}g_x-f_0\|_*\le CB\zeta^2\ell' e^{-\beta(C_{t+S}-C_t)_{xx}/2}. ∥ E C S g x − f 0 ∥ ∗ ≤ C B ζ 2 ℓ ′ e − β ( C t + S − C t ) xx /2 .
The remaining covariance C T − C S C_T-C_S C T − C S lets us differentiate this error through the Gaussian density. By (35) ,
( w i , ( C T − C S ) − 1 w i ) ≤ 2 ( Q T e i , ( C T − C S ) − 1 Q T e i ) + 2 Q t + T ( i , x ) 2 Q t + T ( x , x ) 2 ( Q T e x , ( C T − C S ) − 1 Q T e x ) ≤ C ℓ ′ − 4 ( 1 + Q t + T ( i , x ) 2 Q t + T ( x , x ) 2 ) ≤ C ℓ ′ − 4 . \begin{aligned}
(w_i,(C_T-C_S)^{-1}w_i)\le2(Q_Te_i,(C_T-C_S)^{-1}Q_Te_i)
+2\frac{Q_{t+T}(i,x)^2}{Q_{t+T}(x,x)^2}(Q_Te_x,(C_T-C_S)^{-1}Q_Te_x) \\
&\le C\ell'^{-4}\left(1+\frac{Q_{t+T}(i,x)^2}{Q_{t+T}(x,x)^2}\right)\le C\ell'^{-4}.
\end{aligned} ( w i , ( C T − C S ) − 1 w i ) ≤ 2 ( Q T e i , ( C T − C S ) − 1 Q T e i ) + 2 Q t + T ( x , x ) 2 Q t + T ( i , x ) 2 ( Q T e x , ( C T − C S ) − 1 Q T e x ) ≤ C ℓ ′ − 4 ( 1 + Q t + T ( x , x ) 2 Q t + T ( i , x ) 2 ) ≤ C ℓ ′ − 4 .
The derivative commutes with the Gaussian and angular averages defining f 0 f_0 f 0 , and the independent phase retains its contraction,
∥ ∂ w i E C T − C S f 0 ∥ ∗ ≤ e − β ( C t + S − C t ) x x / 2 ∥ ∂ w i g x ∥ ∗ . \|\partial_{w_i}\mathbb{E}_{C_T-C_S}f_0\|_*\le e^{-\beta(C_{t+S}-C_t)_{xx}/2}\|\partial_{w_i}g_x\|_*. ∥ ∂ w i E C T − C S f 0 ∥ ∗ ≤ e − β ( C t + S − C t ) xx /2 ∥ ∂ w i g x ∥ ∗ .
For the error, we may either put the derivative on this density or commute it through the two averages. The two bounds give
∥ ∂ w i e C T − C S ( E C S g x − f 0 ) ∥ ∗ ≤ min { ( w i , ( C T − C S ) − 1 w i ) ∥ E C S g x − f 0 ∥ ∗ , 2 ∥ ∂ w i g x ∥ ∗ } ≤ C B ζ ℓ 2 ℓ ′ − 2 min { e − β ( C t + S − C t ) x x / 2 , e − 24 μ d ( i , x ) / ( α ℓ ′ ) } . \begin{aligned}
\left\|\partial_{w_i}e_{C_T-C_S}\left(\mathbb{E}_{C_S}g_x-f_0\right)\right\|_* \leq\min\left\{\sqrt{(w_i,(C_T-C_S)^{-1}w_i)}\left\|\mathbb{E}_{C_S}g_x-f_0\right\|_*,2\left\|\partial_{w_i}g_x\right\|_*\right\} \\
&\leq CB\zeta_\ell^2\ell'^{-2}\min\left\{e^{-\beta(C_{t+S}-C_t)_{xx}/2},e^{-24\mu d(i,x)/(\alpha\ell')}\right\}.
\end{aligned} ∥ ∂ w i e C T − C S ( E C S g x − f 0 ) ∥ ∗ ≤ min { ( w i , ( C T − C S ) − 1 w i ) ∥ E C S g x − f 0 ∥ ∗ , 2 ∥ ∂ w i g x ∥ ∗ } ≤ C B ζ ℓ 2 ℓ ′ − 2 min { e − β ( C t + S − C t ) xx /2 , e − 24 μ d ( i , x ) / ( α ℓ ′ ) } .
The inequality min { a , b } ≤ a 1 − α / 4 b α / 4 \min\{a,b\}\leq a^{1-\alpha/4}b^{\alpha/4} min { a , b } ≤ a 1 − α /4 b α /4 retains the spatial rate 6 μ 6\mu 6 μ while using only a fraction of the damping. Since β ( C t + T − C t + S ) x x ≤ C β \beta(C_{t+T}-C_{t+S})_{xx}\leq C_\beta β ( C t + T − C t + S ) xx ≤ C β , the covariance logarithm gives
∥ E C T ∂ w i g x ∥ ∗ ≤ C B ζ ℓ 2 ℓ ′ − 2 M − β ( 1 − α / 4 ) / ( 4 π ) e − 6 μ d ( i , x ) / ℓ ′ . \left\|\mathbb{E}_{C_T}\partial_{w_i}g_x\right\|_* \leq CB\zeta_\ell^2\ell'^{-2}M^{-\beta(1-\alpha/4)/(4\pi)}e^{-6\mu d(i,x)/\ell'}. ∥ E C T ∂ w i g x ∥ ∗ ≤ C B ζ ℓ 2 ℓ ′ − 2 M − β ( 1 − α /4 ) / ( 4 π ) e − 6 μ d ( i , x ) / ℓ ′ .
Apply (46) to compare the conditional average with a Gaussian average centred at the conditional mean. The latter satisfies the bound just proved, which is uniform in the background. The inverse-covariance estimate also holds with C T − 1 C_T^{-1} C T − 1 because C T − C S ≤ C T C_T-C_S\leq C_T C T − C S ≤ C T . Hence
∣ E λ ∂ w i g x − E C T ∂ w i g x ( E λ ψ + ζ ) ∣ ≤ C T ∥ Hess F ∥ ∞ , op ℓ ′ − 2 ∑ k ∥ ∂ k g x ∥ ∞ ≤ C M , K H ζ ℓ 2 ℓ ′ − 2 ≤ C M ζ ℓ 2 ℓ ′ − 2 . \begin{aligned}
\left|\mathbb{E}_\lambda\partial_{w_i}g_x-\mathbb{E}_{C_T}\partial_{w_i}g_x(\mathbb{E}_\lambda\psi+\zeta)\right|\leq CT\|\operatorname{Hess}F\|_{\infty,\operatorname{op}}\ell'^{-2}\sum_k\|\partial_k g_x\|_\infty\\
&\leq C_{M,K}H\zeta_\ell^2\ell'^{-2} \\
&\leq CM\zeta_\ell^2\ell'^{-2}.
\end{aligned} ∣ E λ ∂ w i g x − E C T ∂ w i g x ( E λ ψ + ζ ) ∣ ≤ C T ∥ Hess F ∥ ∞ , op ℓ ′ − 2 k ∑ ∥ ∂ k g x ∥ ∞ ≤ C M , K H ζ ℓ 2 ℓ ′ − 2 ≤ C M ζ ℓ 2 ℓ ′ − 2 .
Here the constant is independent of B B B . Contraction of each average and (53) give a bound proportional to B B B , with spatial decay. Interpolation keeps only the power B α / 4 B^{\alpha/4} B α /4 ,
∣ E λ ∂ w i g x − E C T ∂ w i g x ( E λ ψ + ζ ) ∣ ≤ C M ζ ℓ 2 ℓ ′ − 2 min { 1 , B e − 24 μ d ( i , x ) / ( α ℓ ′ ) } ≤ C M B α / 4 ζ ℓ 2 ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ . \begin{aligned}
\left|\mathbb{E}_\lambda\partial_{w_i}g_x-\mathbb{E}_{C_T}\partial_{w_i}g_x(\mathbb{E}_\lambda\psi+\zeta)\right|\leq CM\zeta_\ell^2\ell'^{-2}\min\left\{1,Be^{-24\mu d(i,x)/(\alpha\ell')}\right\} \\
&\leq C_MB^{\alpha/4}\zeta_\ell^2\ell'^{-2}e^{-6\mu d(i,x)/\ell'}.
\end{aligned} ∣ E λ ∂ w i g x − E C T ∂ w i g x ( E λ ψ + ζ ) ∣ ≤ C M ζ ℓ 2 ℓ ′ − 2 min { 1 , B e − 24 μ d ( i , x ) / ( α ℓ ′ ) } ≤ C M B α /4 ζ ℓ 2 ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ .
Substitute this comparison, the Gaussian estimate and the force covariance bound into (52) to obtain
∥ ( ∂ i − Q t + T ( i , x ) Q t + T ( x , x ) ∂ x ) E λ g x ∥ ∞ ≤ [ C B M − β ( 1 − α / 4 ) / ( 4 π ) + C M ( 1 + B α / 4 ) ] ζ ℓ 2 ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ . \left\|\left(\partial_i-\frac{Q_{t+T}(i,x)}{Q_{t+T}(x,x)}\partial_x\right)\mathbb{E}_\lambda g_x\right\|_\infty
\leq\left[CBM^{-\beta(1-\alpha/4)/(4\pi)}+C_M(1+B^{\alpha/4})\right]\zeta_\ell^2\ell'^{-2}e^{-6\mu d(i,x)/\ell'}. ( ∂ i − Q t + T ( x , x ) Q t + T ( i , x ) ∂ x ) E λ g x ∞ ≤ [ C B M − β ( 1 − α /4 ) / ( 4 π ) + C M ( 1 + B α /4 ) ] ζ ℓ 2 ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ .
We now express the outgoing derivative ∂ i ( x ) \partial_i^{(x)} ∂ i ( x ) , at time t + T t+T t + T , in terms of the directions just estimated. Since ∑ k Q t + T ( k , x ) = c t + T / β \sum_k Q_{t+T}(k,x)=c_{t+T}/\sqrt{\beta} ∑ k Q t + T ( k , x ) = c t + T / β , the two coefficients of ∂ x \partial_x ∂ x cancel,
( ∂ i − Q t + T ( i , x ) Q t + T ( x , x ) ∂ x ) − β Q t + T ( i , x ) c t + T ∑ k ( ∂ k − Q t + T ( k , x ) Q t + T ( x , x ) ∂ x ) = ∂ i − β Q t + T ( i , x ) c t + T ∑ k ∂ k = ∂ i ( x ) . (54) \begin{aligned}
\left(\partial_i-\frac{Q_{t+T}(i,x)}{Q_{t+T}(x,x)}\partial_x\right)-\frac{\sqrt{\beta}Q_{t+T}(i,x)}{c_{t+T}}\sum_k\left(\partial_k-\frac{Q_{t+T}(k,x)}{Q_{t+T}(x,x)}\partial_x\right) \\
&=\partial_i-\frac{\sqrt{\beta}Q_{t+T}(i,x)}{c_{t+T}}\sum_k\partial_k=\partial_i^{(x)}. \tag*{(54)}
\end{aligned} ( ∂ i − Q t + T ( x , x ) Q t + T ( i , x ) ∂ x ) − c t + T β Q t + T ( i , x ) k ∑ ( ∂ k − Q t + T ( x , x ) Q t + T ( k , x ) ∂ x ) = ∂ i − c t + T β Q t + T ( i , x ) k ∑ ∂ k = ∂ i ( x ) . (54)
The coefficient of this sum compensates for the number of sites within distance ℓ ′ \ell' ℓ ′ of x x x ,
β Q t + T ( i , x ) c t + T ∑ k e − 6 μ d ( k , x ) / ℓ ′ ≤ C ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ ℓ ′ 2 ≤ C e − 6 μ d ( i , x ) / ℓ ′ . (55) \frac{\sqrt{\beta}Q_{t+T}(i,x)}{c_{t+T}}\sum_k e^{-6\mu d(k,x)/\ell'}\leq C\ell'^{-2}e^{-6\mu d(i,x)/\ell'}\ell'^2\leq Ce^{-6\mu d(i,x)/\ell'}. \tag*{(55)} c t + T β Q t + T ( i , x ) k ∑ e − 6 μ d ( k , x ) / ℓ ′ ≤ C ℓ ′ − 2 e − 6 μ d ( i , x ) / ℓ ′ ℓ ′2 ≤ C e − 6 μ d ( i , x ) / ℓ ′ . (55)
Here β / c t + T = e ε 2 m 2 ( t + T ) / 2 ≤ C \sqrt{\beta}/c_{t+T}=e^{\varepsilon^2m^2(t+T)/2}\leq C β / c t + T = e ε 2 m 2 ( t + T ) /2 ≤ C in the heat regime. This proves the same bound for ∂ i ( x ) E λ g x \partial_i^{(x)}\mathbb{E}_\lambda g_x ∂ i ( x ) E λ g x , independently of volume. After multiplication by M 2 M^2 M 2 in G x G_x G x , the exponent of its term proportional to B B B is
2 − β 4 π ( 1 − α / 4 ) = α + α ( 2 − α ) 4 ≤ 3 α 2 < 2 α . 2-\frac{\beta}{4\pi}(1-\alpha/4)=\alpha+\frac{\alpha(2-\alpha)}{4}\leq\frac{3\alpha}{2}<2\alpha. 2 − 4 π β ( 1 − α /4 ) = α + 4 α ( 2 − α ) ≤ 2 3 α < 2 α .
The term proportional to B B B therefore grows more slowly than the induction bound B ζ ℓ 2 = B M 2 α ζ ℓ 2 B\zeta_\ell^2=BM^{2\alpha}\zeta_\ell^2 B ζ ℓ 2 = B M 2 α ζ ℓ 2 . Finally ∂ i ( x ) ( id − Π ) G x = ( id − Π ) ∂ i ( x ) G x \partial_i^{(x)}(\operatorname{id}-\Pi)G_x=(\operatorname{id}-\Pi)\partial_i^{(x)}G_x ∂ i ( x ) ( id − Π ) G x = ( id − Π ) ∂ i ( x ) G x . Projection and norm equivalence act only on this quadratic error, and give (51) . ∧็บ
Completion of the estimates
Proof of Proposition 3.1. Fix μ \mu μ as above. Let C C C cover the propagated quadratic coefficients independently of M M M , K K K , H H H , and let C μ C_{\mu} C μ include the linear coefficient in (50) . Choose
M ≥ max { 256 , 192 / α } , M α / 2 ≥ 4 C , M α ≥ 4 C μ . (56) M \ge\max\{256,192/\alpha\},\qquad M^{\alpha/2} \ge4C,\qquad M^{\alpha} \ge4C_{\mu}. \tag*{(56)} M ≥ max { 256 , 192/ α } , M α /2 ≥ 4 C , M α ≥ 4 C μ . (56)
First choose K ≥ max { 1 , 2 β , C β M 2 − α } K \ge\max\{1,2\sqrt{\beta},C_{\beta}M^{2-\alpha}\} K ≥ max { 1 , 2 β , C β M 2 − α } using (49) , then H ≥ max { 1 , 2 β , 4 C M , K M − α } H \ge\max\{1,2\beta,4C_{M,K}M^{-\alpha}\} H ≥ max { 1 , 2 β , 4 C M , K M − α } using (50) . Let C M ≥ 1 C_{M} \ge1 C M ≥ 1 cover all remaining errors, including the fixed K , H K,H K , H . The choice of M M M controls the term proportional to B B B . Since α / 4 < 1 \alpha/4<1 α /4 < 1 , the remaining terms divided by B B B tend to zero. We can therefore choose B B B large enough that
C M − α / 2 + C M ( 1 + B α / 4 ) B M 2 α ≤ 1 4 + 1 4 = 1 2 . CM^{-\alpha/2}+\frac{C_{M}(1+B^{\alpha/4})}{BM^{2\alpha}}\le\frac{1}{4}+\frac{1}{4}=\frac{1}{2}. C M − α /2 + B M 2 α C M ( 1 + B α /4 ) ≤ 4 1 + 4 1 = 2 1 .
Equivalently,
[ C M 3 α / 2 B + C M ( 1 + B α / 4 ) ] ζ ℓ 2 ≤ 1 2 B ζ ℓ 2 . (57) [CM^{3\alpha/2}B+C_{M}(1+B^{\alpha/4})]\zeta_{\ell}^{2}\le\frac{1}{2}B\zeta_{\ell}^{2}. \tag*{(57)} [ C M 3 α /2 B + C M ( 1 + B α /4 )] ζ ℓ 2 ≤ 2 1 B ζ ℓ 2 . (57)
Finally choose 0 < ζ ∗ ≤ 1 0<\zeta_{*}\le1 0 < ζ ∗ ≤ 1 such that
B ζ ∗ ≤ 1 , C M H ζ ∗ ≤ 1 8 . (58) B\zeta_{*}\le1,\qquad C_{M}H\zeta_{*}\le\frac{1}{8}. \tag*{(58)} B ζ ∗ ≤ 1 , C M H ζ ∗ ≤ 8 1 . (58)
The last choice ensures conditional convexity and bounds the quadratic term in (50) . Substituting our choices in that estimate gives
[ C μ H + C M , K + C β , μ , M H 2 ζ ℓ ] ζ ℓ ℓ ′ − 2 ≤ H ( M α / 4 + M α / 4 + 1 / 8 ) ζ ℓ ℓ ′ − 2 ≤ H ζ ℓ ℓ ′ − 2 . [C_{\mu}H+C_{M,K}+C_{\beta,\mu,M}H^{2}\zeta_{\ell}]\zeta_{\ell}^{\ell^{\prime}-2}
\le H(M^{\alpha/4}+M^{\alpha/4}+1/8)\zeta_{\ell}^{\ell^{\prime}-2}
\le H\zeta_{\ell}^{\ell^{\prime}-2}. [ C μ H + C M , K + C β , μ , M H 2 ζ ℓ ] ζ ℓ ℓ ′ − 2 ≤ H ( M α /4 + M α /4 + 1/8 ) ζ ℓ ℓ ′ − 2 ≤ H ζ ℓ ℓ ′ − 2 .
The derivative at x x x is covered by our choice of K K K , independently of its previous bound. Since M α ≤ M 3 α / 2 M^{\alpha}\le M^{3\alpha/2} M α ≤ M 3 α /2 , the same choice of B B B covers the quadratic term in (48) . Equation (57) gives the phase-fixed derivative bound. Reality, periodicity and zero average are preserved by the construction. Every constant has now been fixed in terms of β \beta β , before cutoff or volume is chosen. □ \square □
Proof of Proposition 2.1. For z = 0 z=0 z = 0 there is nothing to estimate. At ℓ = 1 \ell=1 ℓ = 1 , the cosine has zero phase-fixed derivatives, norm 2 ∣ z ∣ ε α 2|z|\varepsilon^{\alpha} 2∣ z ∣ ε α , derivative at x x x at most 2 β ζ 1 2\sqrt{\beta}\zeta_{1} 2 β ζ 1 , and full Hessian row at most 2 β ζ 1 2\beta\zeta_{1} 2 β ζ 1 . Our choices of K , H K,H K , H cover these bounds. Choose c β c_{\beta} c β small enough for the heat regime and smallness conditions. Proposition 3.1 then gives the decomposition at lengths M j M^{j} M j .
For ℓ ≥ M \ell\ge M ℓ ≥ M , choose j ≥ 0 j\ge0 j ≥ 0 with M ≤ ℓ / M j < M 2 M\le\ell/M^{j}<M^{2} M ≤ ℓ / M j < M 2 . Iterate to M j M^{j} M j , then take the final step with ratio ℓ / M j \ell/M^{j} ℓ / M j . We need only the value and Hessian bounds at this last step. In the proofs of (48) and (50) , the heat bounds are uniform in the ratio and the remaining scale factors are bounded powers of it. In particular, the product ensuring conditional convexity is at most C ( ℓ / M j ) 2 H ζ M j ≤ C M 4 H ζ ℓ C(\ell/M^{j})^{2}H\zeta_{M^{j}}\le CM^{4}H\zeta_{\ell} C ( ℓ / M j ) 2 H ζ M j ≤ C M 4 H ζ ℓ . A single reduction of c β c_{\beta} c β therefore makes the final step admissible for every ratio in [ M , M 2 ] [M,M^{2}] [ M , M 2 ] . These two estimates give
∥ g x ∥ ∗ ≤ C β ζ ℓ , sup i ∑ j e μ d ( i , j ) / ( 2 ℓ ) sup φ ∣ ∂ i ∂ j ( N + R ) ∣ ≤ C β ζ ℓ ℓ − 2 . \lVert g_{x}\rVert_{*}\le C_{\beta}\zeta_{\ell},\qquad
\sup_{i}\sum_{j}e^{\mu d(i,j)/(2\ell)}
\sup_{\varphi}\left|\partial_{i}\partial_{j}(N+R)\right|
\le C_{\beta}\zeta_{\ell}\ell^{-2}. ∥ g x ∥ ∗ ≤ C β ζ ℓ , i sup j ∑ e μ d ( i , j ) / ( 2 ℓ ) φ sup ∣ ∂ i ∂ j ( N + R ) ∣ ≤ C β ζ ℓ ℓ − 2 .
The period average N N N inherits this row bound. On the sine-Gordon trajectory it is also invariant under spatial translations. Indeed, permuting the field values by $a\mapsto a+v leaves V 0 V_0 V 0 and the Gaussian law unchanged and commutes with Q t Q_t Q t and period averaging. Consequently,
sup φ ∣ ∂ a ∂ a + v N ( φ ) ∣ = sup φ ∣ ∂ 0 ∂ v N ( φ ) ∣ . \sup_{\varphi}|\partial_a\partial_{a+v}N(\varphi)|=\sup_{\varphi}|\partial_0\partial_vN(\varphi)|. φ sup ∣ ∂ a ∂ a + v N ( φ ) ∣ = φ sup ∣ ∂ 0 ∂ v N ( φ ) ∣.
This equality concerns field suprema; the Hessian at a fixed field need not be spatially translation invariant.
The neutral difference identity (40) now turns the row bound into the required gradient form. Represent each torus displacement v v v by a shortest lattice vector, ∣ v ∣ 1 = d ( 0 , v ) |v|_1=d(0,v) ∣ v ∣ 1 = d ( 0 , v ) , and choose a shortest path 0 = v 0 , … , v ∣ v ∣ 1 = v 0=v_0,\ldots,v_{|v|_1}=v 0 = v 0 , … , v ∣ v ∣ 1 = v . Then
∑ a ( w a + v − w a ) 2 ≤ ∣ v ∣ 1 ∑ k = 1 ∣ v ∣ 1 ∑ a ( w a + v k − w a + v k − 1 ) 2 ≤ ∣ v ∣ 1 2 ∥ D w ∥ 2 2 . \sum_a(w_{a+v}-w_a)^2\leq|v|_1\sum_{k=1}^{|v|_1}\sum_a(w_{a+v_k}-w_{a+v_{k-1}})^2
\leq|v|_1^2\lVert Dw\rVert_2^2. a ∑ ( w a + v − w a ) 2 ≤ ∣ v ∣ 1 k = 1 ∑ ∣ v ∣ 1 a ∑ ( w a + v k − w a + v k − 1 ) 2 ≤ ∣ v ∣ 1 2 ∥ D w ∥ 2 2 .
The first inequality is Cauchy–Schwarz along the path. In the second, translation of the sum over a a a bounds each nearest-neighbour difference by the gradient energy. Applying this estimate before summing over displacements gives
∣ ( w , Hess N ( φ ) w ) ∣ ≤ 1 2 ∑ v sup ψ ∣ ∂ 0 ∂ v N ( ψ ) ∣ ∑ a ( w a + v − w a ) 2 ≤ 1 2 ∑ v d ( 0 , v ) 2 sup ψ ∣ ∂ 0 ∂ v N ( ψ ) ∣ ∥ D w ∥ 2 2 ≤ C μ ℓ 2 ∑ v e μ d ( 0 , v ) / ( 2 ℓ ) sup ψ ∣ ∂ 0 ∂ v N ( ψ ) ∣ ∥ D w ∥ 2 2 ≤ C β ζ ℓ ∥ D w ∥ 2 2 . (59) \begin{aligned}
|(w,\operatorname{Hess}N(\varphi)w)|&\leq\frac{1}{2}\sum_v\sup_\psi|\partial_0\partial_vN(\psi)|\sum_a(w_{a+v}-w_a)^2\\
&\leq\frac{1}{2}\sum_v d(0,v)^2\sup_\psi|\partial_0\partial_vN(\psi)|\lVert Dw\rVert_2^2\\
&\leq C_\mu\ell^2\sum_v e^{\mu d(0,v)/(2\ell)}\sup_\psi|\partial_0\partial_vN(\psi)|\lVert Dw\rVert_2^2\\
&\leq C_\beta\zeta_\ell\lVert Dw\rVert_2^2. \tag*{(59)}
\end{aligned} ∣ ( w , Hess N ( φ ) w ) ∣ ≤ 2 1 v ∑ ψ sup ∣ ∂ 0 ∂ v N ( ψ ) ∣ a ∑ ( w a + v − w a ) 2 ≤ 2 1 v ∑ d ( 0 , v ) 2 ψ sup ∣ ∂ 0 ∂ v N ( ψ ) ∣ ∥ D w ∥ 2 2 ≤ C μ ℓ 2 v ∑ e μ d ( 0 , v ) / ( 2 ℓ ) ψ sup ∣ ∂ 0 ∂ v N ( ψ ) ∣ ∥ D w ∥ 2 2 ≤ C β ζ ℓ ∥ D w ∥ 2 2 . (59)
The third line uses r 2 e − μ r / ( 2 ℓ ) ≤ C μ ℓ 2 r^2e^{-\mu r/(2\ell)}\leq C_\mu\ell^2 r 2 e − μ r / ( 2 ℓ ) ≤ C μ ℓ 2 . Thus the same weighted row controls both the full Hessian and the neutral gradient form. In physical units the three bounds are
∥ Hess [ V t ( Q t ⋅ ) ] ∥ ∞ , o p ≤ C β ζ ℓ ℓ − 2 = C β ∣ z ∣ ε 2 ( ε ℓ ) − β / ( 4 π ) , \lVert\operatorname{Hess}[V_t(Q_t\cdot)]\rVert_{\infty,\mathrm{op}}\leq C_\beta\zeta_\ell\ell^{-2}=C_\beta|z|\varepsilon^2(\varepsilon\ell)^{-\beta/(4\pi)}, ∥ Hess [ V t ( Q t ⋅ )] ∥ ∞ , op ≤ C β ζ ℓ ℓ − 2 = C β ∣ z ∣ ε 2 ( ε ℓ ) − β / ( 4 π ) ,
osc R ≤ 2 ∣ Λ ∣ ℓ − 2 C β ζ ℓ = C β ∣ z ∣ L 2 ( ε ℓ ) − β / ( 4 π ) , \operatorname{osc}R\leq2|\Lambda|\ell^{-2}C_\beta\zeta_\ell=C_\beta|z|L^2(\varepsilon\ell)^{-\beta/(4\pi)}, osc R ≤ 2∣Λ∣ ℓ − 2 C β ζ ℓ = C β ∣ z ∣ L 2 ( ε ℓ ) − β / ( 4 π ) ,
∣ ( w , Hess N w ) ∣ ≤ C β ζ ℓ ∥ D w ∥ 2 2 = C β ∣ z ∣ ( ε ℓ ) α ∥ D w ∥ 2 2 . |(w,\operatorname{Hess}Nw)|\leq C_\beta\zeta_\ell\lVert Dw\rVert_2^2=C_\beta|z|(\varepsilon\ell)^\alpha\lVert Dw\rVert_2^2. ∣ ( w , Hess N w ) ∣ ≤ C β ζ ℓ ∥ D w ∥ 2 2 = C β ∣ z ∣ ( ε ℓ ) α ∥ D w ∥ 2 2 .
For 1 ≤ ℓ < M 1\leq\ell<M 1 ≤ ℓ < M , take N = 0 N=0 N = 0 and R = V t ( Q t ⋅ ) R=V_t(Q_t\cdot) R = V t ( Q t ⋅ ) . The hypothesis of (32) follows directly from the bare Hessian,
∥ C t ∥ o p ∥ Hess V 0 ∥ ∞ , o p ≤ 2 β t ∣ z ∣ ε α ≤ 2 β M 2 ζ 1 ≤ 1 8 , \lVert C_t\rVert_{\mathrm{op}}\lVert\operatorname{Hess}V_0\rVert_{\infty,\mathrm{op}}\leq2\beta t|z|\varepsilon^\alpha\leq2\beta M^2\zeta_1\leq\frac{1}{8}, ∥ C t ∥ op ∥ Hess V 0 ∥ ∞ , op ≤ 2 β t ∣ z ∣ ε α ≤ 2 β M 2 ζ 1 ≤ 8 1 ,
after reducing c β c_\beta c β . For t > 0 t>0 t > 0 , apply that estimate with Σ = C t \Sigma=C_t Σ = C t and W 1 = W 2 = i d W_1=W_2=\mathrm{id} W 1 = W 2 = id , and use ∥ Q t ∥ o p ≤ 1 \lVert Q_t\rVert_{\mathrm{op}}\leq1 ∥ Q t ∥ op ≤ 1 . At t = 0 t=0 t = 0 the same conclusion is the bare Hessian bound. Also inf V 0 ≤ − log E C t e − V 0 ( φ + ζ ) ≤ sup V 0 \inf V_0\leq-\log\mathbb{E}_{C_t}e^{-V_0(\varphi+\zeta)}\leq\sup V_0 inf V 0 ≤ − log E C t e − V 0 ( φ + ζ ) ≤ sup V 0 . Hence
∥ Hess [ V t ( Q t ⋅ ) ] ∥ ∞ , o p ≤ 4 β ∣ z ∣ ε α ≤ C β ∣ z ∣ ε 2 ( ε ℓ ) − β / ( 4 π ) , \lVert\operatorname{Hess}[V_t(Q_t\cdot)]\rVert_{\infty,\mathrm{op}}\leq4\beta|z|\varepsilon^\alpha\leq C_\beta|z|\varepsilon^2(\varepsilon\ell)^{-\beta/(4\pi)}, ∥ Hess [ V t ( Q t ⋅ )] ∥ ∞ , op ≤ 4 β ∣ z ∣ ε α ≤ C β ∣ z ∣ ε 2 ( ε ℓ ) − β / ( 4 π ) ,
osc R ≤ osc V 0 ≤ 4 ∣ z ∣ L 2 ε − β / ( 4 π ) ≤ C β ∣ z ∣ L 2 ( ε ℓ ) − β / ( 4 π ) . \operatorname{osc}R\leq\operatorname{osc}V_0\leq4|z|L^2\varepsilon^{-\beta/(4\pi)}\leq C_\beta|z|L^2(\varepsilon\ell)^{-\beta/(4\pi)}. osc R ≤ osc V 0 ≤ 4∣ z ∣ L 2 ε − β / ( 4 π ) ≤ C β ∣ z ∣ L 2 ( ε ℓ ) − β / ( 4 π ) .
The factors lost here are at most powers of the fixed M M M . The neutral form bound is zero. This proves the proposition for every real length ℓ ≥ 1 \ell\geq1 ℓ ≥ 1 .
Heat-kernel estimates
The heat estimates of [3 Lemmas A.1 and 3.3] give the following bounds with spatial weights and differences, uniformly before the diffusion length reaches the mass length or the side of the torus. Write D k D^k D k for any k k k nearest-neighbour differences at either endpoint, and d d d for periodic nearest-neighbour distance. Fix 0 < c 0 ≤ 1 0 < c_0 \le1 0 < c 0 ≤ 1 and assume
ℓ ′ ≥ 1 , ℓ ′ ≤ c 0 L / ε , ε m ℓ ′ ≤ c 0 . (60) \ell' \ge1,\qquad\ell' \le c_0L/\varepsilon,\qquad\varepsilon m\ell' \le c_0. \tag*{(60)} ℓ ′ ≥ 1 , ℓ ′ ≤ c 0 L / ε , ε m ℓ ′ ≤ c 0 . (60)
Lemma A.1. There is ρ 0 > 0 \rho_0 > 0 ρ 0 > 0 such that, for 0 < ρ ≤ ρ 0 0 < \rho\le\rho_0 0 < ρ ≤ ρ 0 , 0 ≤ s ≤ 2 ℓ ′ 2 0 \le s \le2\ell'^2 0 ≤ s ≤ 2 ℓ ′2 and k = 0 , 1 , 2 k = 0,1,2 k = 0 , 1 , 2 ,
sup x ∑ y e ρ d ( x , y ) / ℓ ′ ∣ D k e − s A ( x , y ) ∣ ≤ C ρ ( 1 + s ) − k / 2 , (61) \sup_x \sum_y e^{\rho d(x,y)/\ell'}\lvert D^k e^{-sA}(x,y)\rvert\le C_\rho(1+s)^{-k/2}, \tag*{(61)} x sup y ∑ e ρ d ( x , y ) / ℓ ′ ∣ D k e − s A ( x , y )∣ ≤ C ρ ( 1 + s ) − k /2 , (61) ∣ D k e − s A ( x , y ) ∣ ≤ C ρ ( 1 + s ) − 1 − k / 2 e − ρ d ( x , y ) / ℓ ′ . (62) \lvert D^k e^{-sA}(x,y)\rvert\le C_\rho(1+s)^{-1-k/2}e^{-\rho d(x,y)/\ell'}. \tag*{(62)} ∣ D k e − s A ( x , y )∣ ≤ C ρ ( 1 + s ) − 1 − k /2 e − ρ d ( x , y ) / ℓ ′ . (62) The row bound also holds for columns, and its time integral gives
sup x ∑ y e ρ d ( x , y ) / ℓ ′ C r ( x , y ) ≤ C ρ r , 0 ≤ r ≤ 2 ℓ ′ 2 . (63) \sup_x \sum_y e^{\rho d(x,y)/\ell'}C_r(x,y) \le C_\rho r,\qquad0 \le r \le2\ell'^2. \tag*{(63)} x sup y ∑ e ρ d ( x , y ) / ℓ ′ C r ( x , y ) ≤ C ρ r , 0 ≤ r ≤ 2 ℓ ′2 . (63) For 0 ≤ t ≤ ℓ ′ 2 0 \le t \le\ell'^2 0 ≤ t ≤ ℓ ′2 ,
C t ( 0 , 0 ) = 1 4 π log ( 1 + t ) + O ( 1 ) , ( e − t A ) x x ≥ c 1 + t . (64) C_t(0,0)=\frac{1}{4\pi}\log(1+t)+O(1),\qquad(e^{-tA})_{xx}\ge\frac{c}{1+t}. \tag*{(64)} C t ( 0 , 0 ) = 4 π 1 log ( 1 + t ) + O ( 1 ) , ( e − t A ) xx ≥ 1 + t c . (64) All constants are independent of cutoff and volume. If ℓ ′ = M ℓ \ell'=M\ell ℓ ′ = M ℓ , ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 , M ≥ 2 M\ge2 M ≥ 2 , and T = ℓ ′ 2 − ℓ 2 T=\ell'^2-\ell^2 T = ℓ ′2 − ℓ 2 , then Q T , D Q T Q_T,DQ_T Q T , D Q T satisfy pointwise bounds with prefactors C ρ ℓ − 2 C_\rho\ell^{-2} C ρ ℓ − 2 , C ρ ℓ ′ − 3 C_\rho\ell'^{-3} C ρ ℓ ′ − 3 and the same exponential decay. Also, C ℓ ′ 2 − 1 ( 0 , 0 ) − C ℓ 2 − 1 ( 0 , 0 ) = ( 2 π ) − 1 log M + O ( 1 ) C_{\ell'^2-1}(0,0)-C_{\ell^2-1}(0,0)=(2\pi)^{-1}\log M+O(1) C ℓ ′2 − 1 ( 0 , 0 ) − C ℓ 2 − 1 ( 0 , 0 ) = ( 2 π ) − 1 log M + O ( 1 ) , with constants independent of M M M .
Proof. The lattice heat-kernel estimate and torus periodization in [3 Lemma A.1] give, for p s ( x ) = ( e s Δ Z 2 ) ( 0 , x ) p_s(x)=(e^{s\Delta_{\mathbb{Z}^2}})(0,x) p s ( x ) = ( e s Δ Z 2 ) ( 0 , x ) and sufficiently small ρ 0 \rho_0 ρ 0 ,
e 2 ρ ∣ x ∣ 1 / ℓ ′ ∣ D k p s ( x ) ∣ ≤ C ( 1 + s ) − 1 − k / 2 e − c ∣ x ∣ 1 / ( 2 1 + s ) . e^{2\rho\lvert x\rvert_1/\ell'}\lvert D^k p_s(x)\rvert\le C(1+s)^{-1-k/2}e^{-c\lvert x\rvert_1/(2\sqrt{1+s})}. e 2 ρ ∣ x ∣ 1 / ℓ ′ ∣ D k p s ( x )∣ ≤ C ( 1 + s ) − 1 − k /2 e − c ∣ x ∣ 1 / ( 2 1 + s ) .
For s < 1 s<1 s < 1 , the same bound follows from p s ≤ p 1 ( 0 ) − 1 p s + 1 p_s\le p_1(0)^{-1}p_{s+1} p s ≤ p 1 ( 0 ) − 1 p s + 1 and the expression of D k D^k D k as finitely many translates. The exponential sum is at most C ( 1 + s ) C(1+s) C ( 1 + s ) , which proves the weighted row bound on Z 2 \mathbb{Z}^2 Z 2 . On the torus,
( e − s A ) ( x , y ) = e − ε 2 m 2 s ∑ v ∈ ( L / ε ) Z 2 p s ( y − x + v ) . (e^{-sA})(x,y)=e^{-\varepsilon^2m^2s}\sum_{v\in(L/\varepsilon)\mathbb{Z}^2}p_s(y-x+v). ( e − s A ) ( x , y ) = e − ε 2 m 2 s v ∈ ( L / ε ) Z 2 ∑ p s ( y − x + v ) .
The periodic distance d ( x , y ) d(x,y) d ( x , y ) is at most ∣ y − x + v ∣ 1 \lvert y-x+v\rvert_1 ∣ y − x + v ∣ 1 for every lift. Summing over y y y and its lifts therefore bounds the weighted torus row by the infinite-lattice row. For the point estimate, retain one rate ρ \rho ρ for d ( x , y ) d(x,y) d ( x , y ) and sum the remaining exponential over lifts. Their separation is L / ε ≥ ℓ ′ / c 0 L/\varepsilon\ge\ell'/c_0 L / ε ≥ ℓ ′ / c 0 , so this sum is uniformly bounded. The mass factor is at most one. Time integration gives (A.4), and T / 2 ≍ ℓ ′ 2 T/2\asymp\ell'^2 T /2 ≍ ℓ ′2 gives the stated bounds on Q T Q_T Q T and D Q T DQ_T D Q T .
The covariance logarithm follows from [3 Lemma 3.3] . Its condition (3.17) allows either L m ≥ 1 Lm\ge1 L m ≥ 1 or t ≤ ε − 2 ( m − 2 ∧ L 2 ) t\le\varepsilon^{-2}(m^{-2}\wedge L^2) t ≤ ε − 2 ( m − 2 ∧ L 2 ) . The latter follows from (A.1). In this range its length is 1 ∨ t 1\vee\sqrt{t} 1 ∨ t , whose logarithm is 1 2 log ( 1 + t ) + O ( 1 ) \frac12\log(1+t)+O(1) 2 1 log ( 1 + t ) + O ( 1 ) . Subtracting at the two times gives the increment with an O ( 1 ) O(1) O ( 1 ) error independent of M M M . The lower diagonal follows from the diagonal asymptotic in [3 Lemma A.1] , the bound p t ( 0 ) ≥ e − 4 t p_t(0)\ge e^{-4t} p t ( 0 ) ≥ e − 4 t on bounded intervals, positivity of periodization and e − ε 2 m 2 t ≥ e − c 0 2 e^{-\varepsilon^2m^2t}\ge e^{-c_0^2} e − ε 2 m 2 t ≥ e − c 0 2 .
Acknowledgements
The author was partially supported by a grant of the Ministry of Research, Innovation and Digitization, CCCDI - UEFISCDI, project number ROSUA-2024-0001, within PNCDI IV.
Declaration of generative AI use
This paper was written by the author and was not generated by artificial intelligence. OpenAI Codex 5.6 Sol was used to assist the author in a first pass check for correctness of mathematical calculations from handwritten notes, as well as literature checks. All mathematical ideas, arguments, and proofs are those of the author, who takes full responsibility for the contents of the paper.
Data availability statement
No data were used for the research described in the article.
Conflict of interest statement
The author has no conflict of interest to declare.
References