Introduction and main results
The 2DOCP in finite volume. Let N ≥ 1 N \ge1 N ≥ 1 be an integer, and let Λ N ⊂ R 2 \Lambda_N \subset\mathbb{R}^2 Λ N ⊂ R 2 be the disk with centre 0 and radius R N : = N π R_N := \sqrt{\frac{N}{\pi}} R N := π N . Denote by X N : = ( x 1 , … , x N ) X_N := (x_1,\ldots,x_N) X N := ( x 1 , … , x N ) an N N N -tuple of points in Λ N \Lambda_N Λ N and let X N : = ∑ i = 1 N δ x i \mathbf{X}_N := \sum_{i=1}^{N} \delta_{x_i} X N := ∑ i = 1 N δ x i be the associated atomic measure. We define the logarithmic interaction energy F N ( X N ) F_N(X_N) F N ( X N ) as:
F N ( X N ) : = 1 2 ∬ ( x , y ) ∈ Λ N × Λ N x ≠ y − log ∣ x − y ∣ d ( X N − L e b ) ( x ) d ( X N − L e b ) ( y ) , (1) F_N(X_N) := \frac{1}{2} \iint_{\substack{(x,y)\in\Lambda_N\times\Lambda_N\\x\ne y}} -\log|x-y|\,\mathrm{d}(\mathbf{X}_N-\mathrm{Leb})(x)\mathrm{d}(\mathbf{X}_N-\mathrm{Leb})(y), \tag*{(1)} F N ( X N ) := 2 1 ∬ ( x , y ) ∈ Λ N × Λ N x = y − log ∣ x − y ∣ d ( X N − Leb ) ( x ) d ( X N − Leb ) ( y ) , (1)
where Leb denotes the Lebesgue measure on R 2 \mathbb{R}^2 R 2 . For x ≠ 0 x \ne0 x = 0 , we write g ( x ) : = − log ∣ x ∣ g(x) := -\log|x| g ( x ) := − log ∣ x ∣ for the two-dimensional Coulomb kernel. We then define a probability density on the space of N N N -tuples of points in Λ N \Lambda_N Λ N by setting, for β > 0 \beta> 0 β > 0 :
d P N , β ( X N ) : = 1 K N , β exp ( − β F N ( X N ) ) d X N , (2) \mathrm{d}\mathbb{P}_{N,\beta}(X_N) := \frac{1}{K_{N,\beta}}\exp\left(-\beta F_N(X_N)\right)\mathrm{d}X_N, \tag*{(2)} d P N , β ( X N ) := K N , β 1 exp ( − β F N ( X N ) ) d X N , (2)
where the partition function K N , β : = ∫ ( Λ N ) N exp ( − β F N ( X N ) ) d X N K_{N,\beta}:=\int_{(\Lambda_N)^N}\exp(-\beta F_N(X_N))\,\mathrm{d}X_N K N , β := ∫ ( Λ N ) N exp ( − β F N ( X N )) d X N is the normalizing constant. Here and below, d X N \mathrm{d}X_N d X N denotes the product Lebesgue measure d x 1 ⋯ d x N \mathrm{d}x_1\cdots\mathrm{d}x_N d x 1 ⋯ d x N on ( R 2 ) N (\mathbb{R}^2)^N ( R 2 ) N . The probability measure P N , β \mathbb{P}_{N,\beta} P N , β is called the finite-volume canonical Gibbs measure of the two-dimensional one-component plasma (2DOCP) at inverse temperature β \beta β .
Infinite volume: limit points, DLR equations. If X N X_N X N is sampled from P N , β \mathbb{P}_{N,\beta} P N , β , it is known that:
This gives meaning to a 2DOCP in infinite volume at arbitrary temperature (the case β = 2 \beta=2 β = 2 , the Ginibre point process, is very special and was understood long ago [43 ] ). However, existence of several distinct limit points was not ruled out. One consequence of the present paper is that for 0 < β ≤ 2 0<\beta\leq2 0 < β ≤ 2 the law of X N X_N X N does have an infinite-volume limit, which we call Gin β \operatorname{Gin}_\beta Gin β .
The statistical physics picture. The 2DOCP is a classical model of statistical mechanics, where the points are viewed as carrying unit charges and the Lebesgue measure on the disk forms a “uniform neutralizing background”. This electrostatic system interacts through the two-dimensional Coulomb kernel − log ∣ x − y ∣ -\log|x-y| − log ∣ x − y ∣ .
Understanding the phase diagram of the 2DOCP, namely the qualitative and quantitative changes of behaviour as β \beta β varies, is an old open question in mathematical physics, see Subsection 1.2. The case β = 0 \beta=0 β = 0 is trivial, and the case β = 2 \beta=2 β = 2 is explicitly solvable [43 ] . At zero temperature (β = + ∞ \beta=+\infty β = + ∞ ), a recent preprint [75 ] proves that the triangular lattice minimizes the energy.
In this paper, we prove the existence of a fluid phase at high temperature (that is, small β \beta β , here β ≤ 2 \beta\leq2 β ≤ 2 ). This phase presumably extends much further, but the present techniques break down for β > 2 \beta>2 β > 2 .
Main results
Exponential decay of correlations. A characteristic feature of a fluid phase is the absence of long-range order. Here and below, ℓ D : = ( 2 π β ) − 1 / 2 \ell_{\mathrm{D}}:=(2\pi\beta)^{-1/2} ℓ D := ( 2 π β ) − 1/2 is the Debye length , the screening length predicted by the Debye–Hückel theory (see Theorem 5 below). We express it in terms of the two-point correlation function (5) , the cylinder observables of (8) and their Dirichlet energies (9) .
Theorem 1 (Exponential decay of correlations). Let 0 < β ≤ 2 0<\beta\leq2 0 < β ≤ 2 . Let P P P be a limit point , as N → ∞ N\to\infty N → ∞ , of the laws of the local configurations X N X_N X N under P N , β \mathbb{P}_{N,\beta} P N , β .
(i) The two-point correlation function of P P P has a continuous version ρ 2 \rho_2 ρ 2 such that, for every ℓ > ℓ D \ell>\ell_{\mathrm{D}} ℓ > ℓ D , there is C ( P , β , ℓ ) < ∞ C(P,\beta,\ell)<\infty C ( P , β , ℓ ) < ∞ satisfying
∣ ρ 2 ( x , y ) − 1 ∣ ≤ C ( P , β , ℓ ) exp ( − ∣ x − y ∣ ℓ ) for all x , y ∈ R 2 . |\rho_2(x,y)-1|\leq C(P,\beta,\ell)\exp\left(-\frac{|x-y|}{\ell}\right)\qquad\text{for all }x,y\in\mathbb{R}^2. ∣ ρ 2 ( x , y ) − 1∣ ≤ C ( P , β , ℓ ) exp ( − ℓ ∣ x − y ∣ ) for all x , y ∈ R 2 . (ii) Let ℓ > ℓ D \ell>\ell_{\mathrm{D}} ℓ > ℓ D , and let C ( β , ℓ ) : = 2 ( ℓ D − 2 − ℓ − 2 ) − 1 C(\beta,\ell):=2(\ell_{\mathrm{D}}^{-2}-\ell^{-2})^{-1} C ( β , ℓ ) := 2 ( ℓ D − 2 − ℓ − 2 ) − 1 . If F F F and G G G are two cylinder observables which are respectively B ( x , R ) B(x,R) B ( x , R ) -local and B ( y , S ) B(y,S) B ( y , S ) -local, then
∣ Cov P [ F , G ] ∣ ≤ C ( β , ℓ ) exp ( − dist ( B ( x , R ) , B ( y , S ) ) ℓ ) D P ( F ) 1 / 2 D P ( G ) 1 / 2 . |\operatorname{Cov}_P[F,G]|\leq C(\beta,\ell)\exp\left(-\frac{\operatorname{dist}(B(x,R),B(y,S))}{\ell}\right)D_P(F)^{1/2}D_P(G)^{1/2}. ∣ Cov P [ F , G ] ∣ ≤ C ( β , ℓ ) exp ( − ℓ dist ( B ( x , R ) , B ( y , S )) ) D P ( F ) 1/2 D P ( G ) 1/2 .
This is the first result on exponential decay, in fact on any kind of pointwise decay, of correlations for the 2DOCP outside the integrable case β = 2 \beta=2 β = 2 . At β = 2 \beta=2 β = 2 , this is not sharp, indeed, the truncated two-point correlation function of the Ginibre point process, ρ 2 ( x , y ) − 1 = − e − π ∣ x − y ∣ 2 \rho_2(x,y)-1=-e^{-\pi|x-y|^2} ρ 2 ( x , y ) − 1 = − e − π ∣ x − y ∣ 2 , is known exactly ([Gin65]). It exhibits Gaussian, not merely exponential, decay.
Uniqueness of the Gibbs state. Our second characterization of the fluid phase is an absence of phase transition, in the sense that the Gibbs state is unique among good processes.
Theorem 2 (Uniqueness of the Gibbs state). For every 0 < β ≤ 2 0 < \beta\le2 0 < β ≤ 2 :
(i) The infinite-volume free energy of the 2DOCP has a unique minimizer.
(ii) It is the unique good DLR state.
It is rotation-invariant and ergodic. We call it Gin β \operatorname{Gin}_{\beta} Gin β .
The free energy was introduced in [68 ] , see § 9 . The class of good processes is defined in Subsection 2.4, and it contains all the limit points (Corollary 5.11). Uniqueness among this class is part of our result. Since Gin β \operatorname{Gin}_{\beta} Gin β is the only limit point, Theorem 1 applies to it, and we denote its two-point correlation function by ρ 2 , β \rho_{2,\beta} ρ 2 , β .
Consequences. Combining Theorem 2 with the results of [59 ] , we obtain the existence of a limit for the 2DOCP local behaviour at high temperature. The only known limit was for the integrable case β = 2 \beta= 2 β = 2 .
Corollary 3 (Existence of a limit). For every 0 < β ≤ 2 0 < \beta\le2 0 < β ≤ 2 , the local point configuration X N X_N X N converges in distribution to Gin β \operatorname{Gin}_{\beta} Gin β as N → ∞ N \to\infty N → ∞ . The same holds for ∑ i = 1 N δ x i − x N \sum_{i=1}^{N} \delta_{x_i-x_N} ∑ i = 1 N δ x i − x N for every bounded deterministic sequence ( x N ) N ≥ 1 (x_N)_{N\ge1} ( x N ) N ≥ 1 in R 2 \mathbb{R}^{2} R 2 .
In particular, at β = 2 \beta= 2 β = 2 , we obtain a new proof of the existence of the limit, which does not use the determinantal structure. This limit is known as the Ginibre point process [43 ] .
Corollary 4 (Ginibre point process). The Ginibre point process coincides with our Gin 2 \operatorname{Gin}_{2} Gin 2 point process. It is the only minimizer of the free energy F 2 F_2 F 2 , and the only good DLR state at β = 2 \beta= 2 β = 2 .
A DLR variational principle. The proof of Theorem 2 goes through the following proposition, valid at arbitrary temperature. It corresponds to one direction of the Dobrushin–Lanford–Ruelle variational principle for our model, see § 9 .
Proposition 1.1 (DLR variational principle). For all β > 0 \beta> 0 β > 0 , every good DLR state is a minimizer of the free energy F β F_\beta F β .
For β ≤ 2 \beta\le2 β ≤ 2 , all the limit points are good (Corollary 5.11). For β > 2 \beta> 2 β > 2 , we do not know whether they satisfy the decay condition (G3), so the statement might be void in that regime.
Debye–Hückel asymptotics. At high temperature (β ↓ 0 \beta\downarrow0 β ↓ 0 ), the Debye–Hückel theory [DH23] predicts the correlation length and the profile of the correlation functions, see [31 , 71 ] . For the 2DOCP, the correlation length should be the Debye length ℓ D \ell_{\mathrm{D}} ℓ D , which tends to infinity as β ↓ 0 \beta\downarrow0 β ↓ 0 , and the predicted truncated two-point function behaves asymptotically like − β K 0 ( ∣ x ∣ ℓ D ) -\beta K_0\left(\frac{|x|}{\ell_{\mathrm{D}}}\right) − β K 0 ( ℓ D ∣ x ∣ ) , where K 0 K_0 K 0 is the modified Bessel function of the second kind of order zero, see the paragraph on screening in Subsection 1.2. We prove this prediction.
Theorem 5 (Debye–Hückel asymptotics). Locally uniformly on R 2 ∖ { 0 } \mathbb{R}^{2} \setminus\{0\} R 2 ∖ { 0 } and in S ′ ( R 2 ) \mathcal{S}'(\mathbb{R}^{2}) S ′ ( R 2 ) , we have
1 β ( ρ 2 , β ( 0 , ℓ D x ) − 1 ) → β ↓ 0 − K 0 ( ∣ x ∣ ) . \frac{1}{\beta}\left(\rho_{2,\beta}(0,\ell_{\mathrm{D}}x)-1\right) \xrightarrow[\beta\downarrow0]{} -K_0(|x|). β 1 ( ρ 2 , β ( 0 , ℓ D x ) − 1 ) β ↓ 0 − K 0 ( ∣ x ∣ ) .
Moreover, the fluctuations on the Debye scale also have a Gaussian limit. For each 0 < β ≤ 2 0 < \beta\le2 0 < β ≤ 2 , let X X X have law Gin β \operatorname{Gin}_{\beta} Gin β . For f ∈ C c ∞ ( R 2 ; R ) f \in C_{\mathrm{c}}^{\infty}(\mathbb{R}^{2};\mathbb{R}) f ∈ C c ∞ ( R 2 ; R ) , define
Z β ( f ) : = ℓ D − 1 ( ∑ x ∈ X f ( x / ℓ D ) − ℓ D 2 ∫ R 2 f ) . (3) \mathcal{Z}_{\beta}(f) := \ell_{\mathrm{D}}^{-1}\left(\sum_{x\in X} f(x/\ell_{\mathrm{D}})-\ell_{\mathrm{D}}^{2}\int_{\mathbb{R}^{2}} f\right). \tag*{(3)} Z β ( f ) := ℓ D − 1 ( x ∈ X ∑ f ( x / ℓ D ) − ℓ D 2 ∫ R 2 f ) . (3)
Theorem 6 (Central limit theorem on the Debye scale). As β ↓ 0 \beta\downarrow0 β ↓ 0 , the finite-dimensional distributions of ( Z β ( f ) ) f ∈ C c ∞ ( R 2 ; R ) (\mathcal{Z}_{\beta}(f))_{f\in C_{\mathrm{c}}^{\infty}(\mathbb{R}^{2};\mathbb{R})} ( Z β ( f ) ) f ∈ C c ∞ ( R 2 ; R ) converge to those of a centred Gaussian field Z \mathcal{Z} Z . For every f , h ∈ C c ∞ ( R 2 ; R ) f,h \in C_{\mathrm{c}}^{\infty}(\mathbb{R}^{2};\mathbb{R}) f , h ∈ C c ∞ ( R 2 ; R ) , its covariance is
E [ Z ( f ) Z ( h ) ] = ⟨ f , ( − Δ ) ( 1 − Δ ) − 1 h ⟩ L 2 ( R 2 ) . \mathbb{E}[\mathcal{Z}(f)\mathcal{Z}(h)] = \langle f,(-\Delta)(1-\Delta)^{-1}h\rangle_{L^{2}(\mathbb{R}^{2})}. E [ Z ( f ) Z ( h )] = ⟨ f , ( − Δ ) ( 1 − Δ ) − 1 h ⟩ L 2 ( R 2 ) .
The proofs in a nutshell.
Concentration (§ 6 ). Our basic tool is a Poincaré inequality for the finite-volume gas, with a constant independent of N N N . For a probability measure e − Φ d x e^{-\Phi}\mathrm{d}x e − Φ d x on R n \mathbb{R}^{n} R n , when the Hessian of Φ \Phi Φ is bounded below, ∇ 2 Φ ≥ a > 0 \nabla^{2}\Phi\ge a > 0 ∇ 2 Φ ≥ a > 0 , such an inequality is classical: it is the Brascamp–Lieb inequality
Var [ F ] ≤ a − 1 E ∣ ∇ F ∣ 2 . \operatorname{Var}[F] \le a^{-1}\mathbb{E}|\nabla F|^{2}. Var [ F ] ≤ a − 1 E ∣∇ F ∣ 2 .
It does not apply here, because the logarithm is not convex in dimension two: the Hessian of − log ∣ x ∣ -\log|x| − log ∣ x ∣ has the eigenvalues ∣ x ∣ − 2 |x|^{-2} ∣ x ∣ − 2 and − ∣ x ∣ − 2 -|x|^{-2} − ∣ x ∣ − 2 , so the Hessian of the energy is unbounded below when two particles come close. However, these two eigenvalues cancel in the Laplacian (− log ∣ x ∣ -\log|x| − log ∣ x ∣ is harmonic away from the origin). There is a counterpart of the Brascamp–Lieb inequality in which the Hessian is replaced by the so-called Levi matrix (or complex Hessian): it is the L 2 L^{2} L 2 estimate of Hörmander for the Cauchy–Riemann operator ∂ ˉ \bar{\partial} ∂ ˉ , and it is proven by the same integration by parts. In our setting, the positivity of the Levi matrix relies on a trick which is only valid for β ≤ 2 \beta\le2 β ≤ 2 (see the paragraph “Shifting weights to gain positivity” in Subsection 3.1). This is where our assumption on β \beta β comes from.
The estimate of Hörmander only controls the distance to the holomorphic functions, not to the constants. Using the symmetries of the gas, we show that this distance boils down to the standard deviation, and we obtain a “true” Poincaré inequality
Var P N , β [ F ] ≤ 1 2 π β E P N , β ∣ ∇ F ∣ 2 . \operatorname{Var}_{\mathsf{P}_{N,\beta}}[F] \le\frac{1}{2\pi\beta}\mathbb{E}_{\mathsf{P}_{N,\beta}}|\nabla F|^{2}. Var P N , β [ F ] ≤ 2 π β 1 E P N , β ∣∇ F ∣ 2 .
An additional argument gives the same inequality, with a larger constant, for all symmetric observables. Applying the Poincaré inequality to tilted versions of the Gibbs measure (an argument due to Herbst), we also obtain exponential concentration for some extensive observables.
Decay of correlations (§ 8 ). For a Gaussian measure with covariance matrix A − 1 A^{-1} A − 1 , a classical argument of Combes and Thomas gives the exponential decay of the entries ( A − 1 ) k j (A^{-1})_{kj} ( A − 1 ) k j as ∣ k − j ∣ → ∞ |k-j| \to\infty ∣ k − j ∣ → ∞ , when A ≥ a > 0 A \ge a > 0 A ≥ a > 0 is a band matrix, that is, when each site only interacts with the sites at a bounded distance. For non-Gaussian measures, a formula of Helffer and Sjöstrand writes covariances in a similar way, Cov [ F , G ] = ⟨ A 1 − 1 ∇ F , ∇ G ⟩ \operatorname{Cov}[F,G] = \langle A_{1}^{-1}\nabla F,\nabla G\rangle Cov [ F , G ] = ⟨ A 1 − 1 ∇ F , ∇ G ⟩ , where A 1 A_{1} A 1 is an operator acting on vector fields with one component per particle, built with the Hessian of the energy, and equal to A A A in the Gaussian case. Decay of correlations then amounts to the decay of A 1 − 1 A_{1}^{-1} A 1 − 1 between the components of distant particles.
At first sight, this cannot work for the 2DOCP: the interaction has infinite range, and the Hessian of the energy couples all pairs of particles, with entries of order ∣ x i − x j ∣ − 2 |x_i-x_j|^{-2} ∣ x i − x j ∣ − 2 . As for concentration, we use instead the complex version of the formula of Helffer and Sjöstrand, in which the Hessian is replaced by the Levi matrix. Since the logarithm is harmonic, the off-diagonal entries of the Levi matrix vanish as soon as x i ≠ x j x_i \ne x_j x i = x j , and only Dirac masses on the collisions { x i = x j } \{x_i=x_j\} { x i = x j } remain. In this formalism, the 2DOCP thus behaves like a system whose interactions have range zero. The argument of Combes and Thomas then gives the exponential decay of covariances on every length ℓ > ℓ D \ell> \ell_D ℓ > ℓ D , uniformly in N N N , for every limit point (Theorem 1 ). The main technical point is to show that the Dirac masses at the collisions do not produce an error term.
Uniqueness (§ 9 and § 10 ). By Theorem 1 , the limit points satisfy the integrability condition (G3) on ρ 2 − 1 \rho_2-1 ρ 2 − 1 , which was the only missing condition for them to be good processes (Subsection 2.4). The proof of uniqueness then has two steps.
First, we prove that every good DLR state minimizes the free energy F β \mathcal{F}_{\beta} F β of [68 ] (Proposition 1.1 ). Such a statement is called a DLR variational principle. In a disk B R B_R B R , we compare the law of the configuration under a good DLR state P \mathsf{P} P with that of a 2DOCP in B R B_R B R without exterior field, whose number of points has the same law as under P \mathsf{P} P . By the DLR equations, the relative entropy of the former with respect to the latter is bounded by the mean interaction energy between the points in B R B_R B R and the configuration outside B R B_R B R . Thanks to (G3) and to hyperuniformity, we show that this energy is O ( R 3 / 2 ) O(R^{3/2}) O ( R 3/2 ) , which is negligible compared with the area of B R B_R B R . This allows us to identify the free energy of P \mathsf{P} P with the minimal one.
Second, we prove that F β \mathcal{F}_{\beta} F β has only one minimizer. By the large deviation principle of [68 ] , the empirical field of the gas of n n n points (the law of the configuration seen from a uniformly chosen point) comes close to every minimizer of F β \mathcal{F}_{\beta} F β with a probability which is not exponentially small in n n n . Hence, an observable which is a continuous function of the empirical field, and which has exponential concentration, takes the same value at all the minimizers. Applying this to the extensive observables of the first item shows that all the free energy minimizers have the same rotation average. A limit point of the gas is a free energy minimizer, by the first step, and it is rotation-invariant: it is therefore the rotation average of every minimizer. Finally, the Poincaré inequality for general symmetric observables shows that the limit points are ergodic, so they cannot be non-trivial averages of rotated copies of another stationary process. Hence every free energy minimizer is equal to the limit point. This excludes, for instance, a crystal with a uniformly random orientation.
Debye–Hückel asymptotics (§ 11 ). As β ↓ 0 \beta\downarrow0 β ↓ 0 , the Debye length ℓ D \ell_{\mathrm{D}} ℓ D tends to infinity, and we measure distances in units of ℓ D \ell_{\mathrm{D}} ℓ D . We first compute the covariance of two smooth linear statistics on this scale, divided by ℓ D 2 \ell_{\mathrm{D}}^2 ℓ D 2 , which is bounded by the Poincaré inequality. Moving all the points along a vector field ∇ f \nabla f ∇ f and integrating by parts in the Gibbs measure (the Ward identity, or loop equation) gives an exact formula for these covariances. As β ↓ 0 \beta\downarrow0 β ↓ 0 , this formula gives the convergence of Theorem 5 in the sense of distributions.
To obtain the value of the two-point function at a given distance, we fix two particles of the finite gas, and write the two-point function as the Laplace transform of the linear statistic of their potential. Once the total charge of the two particles is compensated, the fluctuations of this statistic are small, and an expansion to second order, combined with the covariances of the first step, replaces the bare interaction − log r -\log r − log r by K 0 ( r ) K_0(r) K 0 ( r ) .
Section 3 presents these four arguments in more detail.
Connection to the literature
Physics literature around the 2DOCP. The phase diagram of the 2DOCP has long been studied in the physics literature. At low temperature (large β \beta β ) one expects a crystalline phase, and a fluid phase at high temperature (small β \beta β ). A transition between the two was investigated in [3 ] . Numerical evidence for a transition around β ≈ 140 \beta\approx140 β ≈ 140 was reported in [32 , 26 , 55 , 22 ] , with further numerical evidence in [47 ] . The existence and nature of such a transition in the thermodynamic limit remain open mathematically, and the physical picture has also been questioned in [84 , 73 , 89 ] . At high temperature, the Debye–Hückel approximation [DH23, DL74] predicts exponential decay of the truncated two-point function ρ 2 , β − 1 \rho_{2,\beta}-1 ρ 2 , β − 1 , at the Debye length and with the profile of Theorem 5 .
The case β = 2 \beta=2 β = 2 is explicitly solvable: it is known as the Ginibre point process [43 ] . Its truncated two-point function has Gaussian decay at large distances. Because of its integrability and its connection with random matrix theory, it has been studied extensively. We refer to [39 ] .
Symmetry breaking. An important question is whether the symmetries of the model, here translations and rotations, can be broken in infinite volume. At positive temperature, the Mermin–Wagner theorem of [40 ] excludes breaking of translation invariance under hypotheses on the interaction which do not cover the logarithmic potential. For the 2DOCP, all limit points of the local configurations X N \mathbf{X}_N X N are stationary at every temperature [59 ] . They are also rotation-invariant, since the finite-volume measures are. For 0 < β ≤ 2 0<\beta\leq2 0 < β ≤ 2 , we prove that the free energy has a unique minimizer and that this minimizer is rotation-invariant and ergodic (Theorem 2 ). In particular, it cannot be a non-trivial mixture of stationary states with different orientations.
Screening and correlations. The rapid decay of correlations in a Coulomb fluid is a collective effect of the long-range interaction. Because charge imbalances interact over large distances, the equilibrium correlations must satisfy strong cancellation conditions [71 ] . In the fluid regime, the truncated particle correlations are expected to decay exponentially, despite the long range of the bare interaction − log ∣ x − y ∣ -\log|x-y| − log ∣ x − y ∣ . This phenomenon is called Debye screening 3 . The exponential form of the decay depends on the Coulomb structure; other long-range repulsive interactions can lead to algebraic correlations [71 Section II.B.3] .
At high temperature, the Debye–Hückel approximation predicts the scale and profile of this decay:
ρ 2 , β ( 0 , x ) − 1 ≈ − β K 0 ( ∣ x ∣ / ℓ D ) \rho_{2,\beta}(0,x)-1 \approx-\beta K_0(|x|/\ell_D) ρ 2 , β ( 0 , x ) − 1 ≈ − β K 0 ( ∣ x ∣/ ℓ D )
Here K 0 K_0 K 0 is the modified Bessel function of the second kind, which behaves like − log t -\log t − log t as t ↓ 0 t \downarrow0 t ↓ 0 and decays exponentially at infinity. The prediction exhibits two effects.
First, at distances ∣ x ∣ = ℓ D r |x| = \ell_D r ∣ x ∣ = ℓ D r with fixed r > 0 r > 0 r > 0 , the truncated correlations have amplitude of order β \beta β .
Second, their range, the Debye length ℓ D \ell_D ℓ D , diverges as β ↓ 0 \beta\downarrow0 β ↓ 0 . In a disk of radius R ≪ ℓ D R \ll\ell_D R ≪ ℓ D , the system can afford discrepancies of the same order as a Poisson process, but for R ≫ ℓ D R \gg\ell_D R ≫ ℓ D , they become much smaller and the system looks neutral. The length scale β − 1 / 2 \beta^{-1/2} β − 1/2 already appears (up to a logarithmic factor) in [5 ] as the minimal scale above which local laws were proven.
The weak amplitude and growing range are consistent with the perfect screening sum rule
∫ R 2 ( ρ 2 , β ( 0 , x ) − 1 ) d x = − 1 , \int_{\mathbb{R}^{2}}(\rho_{2,\beta}(0,x)-1)\,\mathrm{d}x=-1, ∫ R 2 ( ρ 2 , β ( 0 , x ) − 1 ) d x = − 1 ,
the simplest of the sum rules reviewed in [71 ] , which holds for Gin β \operatorname{Gin}_{\beta} Gin β (see the proof of Lemma 6.1 ). Around a particle fixed at the origin, the excess density of the other particles has total mass − 1 -1 − 1 : the surrounding gas compensates the tagged particle’s charge on average.
Order within disorder. The picture above, with weak correlations whose range is large , might suggest that at high temperature the 2DOCP resembles a Poisson process in every respect. It does not, and it is amusing to note that two different notions of order coexist here. In the classical sense of physicists, the 2DOCP at high temperature is very disordered: its correlations decay exponentially fast (Theorem 1 ). In a sense familiar to probabilists, it remains very ordered. For every β > 0 \beta> 0 β > 0 , it is hyperuniform : the variance of the number of points in a disk of radius R R R (in the bulk of a large system) is o ( R 2 ) o(R^2) o ( R 2 ) as R → ∞ R \to\infty R → ∞ [58 ] , and even O ( R ) O(R) O ( R ) for good DLR states (Proposition 2.5 ). Perhaps more anecdotally, its infinite-volume limit points are number-rigid [59 ] in the sense of Ghosh–Peres [44 ] : the number of points in a bounded domain is almost surely determined by the configuration outside of it. There is no contradiction: decay of correlations concerns pairs of local observables, hyperuniformity concerns the number of points in large domains, and number rigidity involves the whole exterior configuration at once, which no local observable sees.
The sum rule also shows how the two notions fit together. The correlations are weak, but they integrate to − 1 -1 − 1 , which is exactly what is needed to cancel the volume term of the number variance. In the Debye–Hückel approximation, the number variance in a disk of radius R R R is of order π R 2 \pi R^2 π R 2 , for a Poisson process, for R ≪ ℓ D R \ll\ell_D R ≪ ℓ D , and of order π R ℓ D \pi R\ell_D π R ℓ D , the perimeter, for R ≫ ℓ D R \gg\ell_D R ≫ ℓ D . Heuristically, hyperuniformity thus only becomes visible beyond the Debye length.
Mathematical treatment. The present paper belongs to the line of works initiated by Sandier and Serfaty on the 2DOCP at the microscopic scale [83 ] , see [81 ] for a presentation. We use in particular the following results.
The large deviation principle and its consequences for the partition function [68 ] .
The fluctuations of smooth linear statistics [69 , 80 ] , see also [8 ] for a different approach.
The existence of limit points of the local configurations, which comes from the local laws of [5 ] .
The DLR equations satisfied by these limit points [59 ] .
Screening has been proved in other Coulomb models:
In dimension two, Yang [92 ] proved Debye screening for a two-component grand-canonical gas with Dirichlet electrostatic boundary conditions, in the high-temperature, small-activity regime.
In dimension three, Brydges and Federbush proved exponential clustering on the Debye scale for dilute Coulomb gases [21 , 13 ] , see [18 Theorem II.1] . Imbrie extended this to the three-dimensional one-component plasma at high temperature and low density [54 ] , Theorem 2.3. These proofs use the co-called sine–Gordon representation .
The one-dimensional log-gas. The 2DOCP has a close one-dimensional relative: the 1d log-gas, a system of particles with logarithmic interaction on the line. For this system, the DLR equations were established in [DHLM21], uniqueness of the minimizer of the free energy was proven in [37 ] , and uniqueness of the stationary solution of the DLR equations under a finite renormalized electric energy assumption in [7 ] . The truncated two-point function decays as a power law for every β > 0 \beta> 0 β > 0 [77 , 6 ] . Polynomial decay bounds for spatially averaged truncated correlations of higher order were proved in [33 ] . A DLR variational principle and another proof of existence of the infinite-volume limit, recovering the breakthrough result of [90 ] by physical methods, are the object of [25 ] .
Recent related works. Poincaré-type inequalities, which are the key to our concentration estimates, were obtained recently by Chafaï [Cha26] and Suzuki [85 ] for β = 2 \beta= 2 β = 2 , and by Rosenzweig [78 ] for β ≤ 2 \beta\le2 β ≤ 2 , with very similar techniques. Around the same time, Becker and Menegaki [19 ] obtain spectral gaps for mean-field Coulomb gases in dimension d ≥ 3 d \ge3 d ≥ 3 .
Open questions
The range of exponential decay. Theorem 1 covers β ≤ 2 \beta\le2 β ≤ 2 , and the present techniques break down for β > 2 \beta> 2 β > 2 , since the collisions contribute to the Levi matrix in Lemma 4.4 with the factor 2 − β 2-\beta 2 − β , which becomes negative. Do correlations decay exponentially up to the transition reported numerically at β ≈ 140 \beta\approx140 β ≈ 140 (see Subsection 1.2)?
Other centres. Let ( x N ) N ≥ 1 (x_N)_{N\ge1} ( x N ) N ≥ 1 be a sequence in R 2 \mathbb{R}^2 R 2 escaping to infinity within the bulk in the sense of [59 ] . Does the configuration ∑ i = 1 N δ x i − x N \sum_{i=1}^{N}\delta_{x_i-x_N} ∑ i = 1 N δ x i − x N converge in distribution? By Theorem 2 , this would follow once (G3) is established for their limit points.
Universality with respect to the confinement. Replace the quadratic confining potential ∣ x i ∣ 2 |x_i|^2 ∣ x i ∣ 2 in (4) by a more general potential V ( x i ) V(x_i) V ( x i ) , or equivalently replace the uniform background measure by a non-uniform one. Let x N x_N x N be a point in the bulk of the corresponding equilibrium measure. After rescaling to intensity 1, does the local configuration seen from x N x_N x N converge to Gin β \operatorname{Gin}_{\beta} Gin β as N → ∞ N \to\infty N → ∞ ? Universality has been proven for one-dimensional log-gases [12 , 11 , 14 ] . The 2DOCP at β = 2 \beta= 2 β = 2 has also attracted a lot of interest see for example [1 , 4 , 53 ] . But as far as we know, there are no universality results concerning the 2DOCP for β ≠ 2 \beta\ne2 β = 2 .
By Theorem 2 , universality would follow if we knew that under a general confining potential, limit points are good DLR states. Our proof of (G3) does not apply as such, since it uses the rotation invariance of the finite-volume measures. For a radial potential V V V , the finite-volume measures are still rotation-invariant. Does our proof of (G3) extend to this case, and give universality for radial potentials?
Connection with the Laughlin wave function. When β = 2 m \beta= 2m β = 2 m for an integer m ≥ 1 m \ge1 m ≥ 1 , the density of the soft 2DOCP is, up to a dilation, the squared modulus of the Laughlin wave function at filling 1 / m 1/m 1/ m [56 ] . In the physics literature, the exponential decay of correlations of the plasma has been related to a spectral gap above this state for its parent Hamiltonian [46 ] . A recent preprint proves this gap at β = 6 \beta= 6 β = 6 [74 Corollaries 1.2 and 7.1] . Does it imply the exponential decay of correlations at β = 6 \beta= 6 β = 6 ?
Breaking of rotation invariance at large β \beta β . Every limit point of the local configurations X N X_N X N is rotation-invariant, since the finite-volume measures are. At large β \beta β , can such a limit point be a non-trivial mixture of processes that are not rotation-invariant, for example a crystal with a uniformly random orientation?
Statement of AI use
The proofs of our main results were found by prompting GPT-6 Astra. We then studied these arguments and turned them into the form given here. We used Codex and Claude Opus 5.5 to edit some parts of our manuscript and to check it for mistakes. We take full responsibility for the contents of the paper.
Preliminaries
Notation. We write C C C and c c c for positive constants that depend only on the parameters of the statement being proved, and that may change from line to line. For a function on R 2 \mathbb{R}^{2} R 2 , ∥ ⋅ ∥ p \|\cdot\|_{p} ∥ ⋅ ∥ p is the norm of L p ( R 2 ) L^{p}(\mathbb{R}^{2}) L p ( R 2 ) , and for a random variable, it is the norm of L p L^{p} L p under the probability measure of the context. We fix a radial function χ ∈ C c ∞ ( B 2 ) \chi\in C^{\infty}_{c}(B_{2}) χ ∈ C c ∞ ( B 2 ) with 0 ≤ χ ≤ 1 0 \le\chi\le1 0 ≤ χ ≤ 1 and χ = 1 \chi=1 χ = 1 on B 1 B_{1} B 1 , and we let χ R : = χ ( ⋅ / R ) \chi_{R}:=\chi(\cdot/R) χ R := χ ( ⋅ / R ) for R > 0 R>0 R > 0 . We identify R 2 \mathbb{R}^{2} R 2 with C \mathbb{C} C .
The model: rewriting and confinement
We start with an identity that replaces the background from the Gibbs measure by a “confining potential”.
Lemma 2.1 (Rewriting the Gibbs measure). Let
H N ( X N ) : = ∑ 1 ≤ i < j ≤ N g ( x i − x j ) + π 2 ∑ i = 1 N ∣ x i ∣ 2 . (4) H_{N}(X_{N}):=\sum_{1\le i<j\le N}g(x_{i}-x_{j})+\frac{\pi}{2}\sum_{i=1}^{N}|x_{i}|^{2}. \tag*{(4)} H N ( X N ) := 1 ≤ i < j ≤ N ∑ g ( x i − x j ) + 2 π i = 1 ∑ N ∣ x i ∣ 2 . (4) The Gibbs measure (2) can be rewritten as follows:
d P N , β ( X N ) = 1 K ~ N , β e − β H N ( X N ) ∏ i = 1 N 1 { x i ∈ Λ N } d X N = 1 K ~ N , β ∏ i < j ∣ x i − x j ∣ β e − π β 2 ∑ i = 1 N ∣ x i ∣ 2 ∏ i = 1 N 1 { x i ∈ Λ N } d X N , \mathrm{d}\mathbb{P}_{N,\beta}(X_{N})=\frac{1}{\widetilde K_{N,\beta}}e^{-\beta H_{N}(X_{N})}\prod_{i=1}^{N}\mathbf{1}_{\{x_{i}\in\Lambda_{N}\}}\mathrm{d}X_{N}
=\frac{1}{\widetilde K_{N,\beta}}\prod_{i<j}|x_{i}-x_{j}|^{\beta}e^{-\frac{\pi\beta}{2}\sum_{i=1}^{N}|x_{i}|^{2}}\prod_{i=1}^{N}\mathbf{1}_{\{x_{i}\in\Lambda_{N}\}}\mathrm{d}X_{N}, d P N , β ( X N ) = K N , β 1 e − β H N ( X N ) i = 1 ∏ N 1 { x i ∈ Λ N } d X N = K N , β 1 i < j ∏ ∣ x i − x j ∣ β e − 2 π β ∑ i = 1 N ∣ x i ∣ 2 i = 1 ∏ N 1 { x i ∈ Λ N } d X N , with some new normalizing constant K ~ N , β \widetilde K_{N,\beta} K N , β . In other words, F N F_{N} F N and H N H_{N} H N differ by a constant on ( Λ N ) N (\Lambda_{N})^{N} ( Λ N ) N .
Proof. The proof is a simple computation. Expanding the double integral in (1) gives
F N ( X N ) = ∑ i < j g ( x i − x j ) − ∑ i = 1 N U N ( x i ) + 1 2 ∬ Λ N × Λ N g ( x − y ) d x d y , U N ( x ) : = ∫ Λ N g ( x − y ) d y . F_{N}(X_{N})=\sum_{i<j}g(x_{i}-x_{j})-\sum_{i=1}^{N}U_{N}(x_{i})+\frac{1}{2}\iint_{\Lambda_{N}\times\Lambda_{N}}g(x-y)\,\mathrm{d}x\mathrm{d}y,\qquad U_{N}(x):=\int_{\Lambda_{N}}g(x-y)\,\mathrm{d}y. F N ( X N ) = i < j ∑ g ( x i − x j ) − i = 1 ∑ N U N ( x i ) + 2 1 ∬ Λ N × Λ N g ( x − y ) d x d y , U N ( x ) := ∫ Λ N g ( x − y ) d y .
The first term is the interaction of the points with each other, the second one is the effect of the background on the points, and the last one, the self-interaction of the background, is a constant which cancels in the partition function. Since − Δ g = 2 π δ 0 -\Delta g=2\pi\delta_{0} − Δ g = 2 π δ 0 , the potential U N U_{N} U N created by the uniform background satisfies Δ U N = − 2 π \Delta U_{N}=-2\pi Δ U N = − 2 π on Λ N \Lambda_{N} Λ N . Since U N U_{N} U N is radial, it is equal to − π 2 ∣ x ∣ 2 -\frac{\pi}{2}|x|^{2} − 2 π ∣ x ∣ 2 plus a constant on Λ N \Lambda_{N} Λ N . Hence F N = H N + c o n s t F_{N}=H_{N}+\mathrm{const} F N = H N + const on ( Λ N ) N (\Lambda_{N})^{N} ( Λ N ) N . Finally, e − β ∑ i < j g ( x i − x j ) = ∏ i < j ∣ x i − x j ∣ β e^{-\beta\sum_{i<j}g(x_{i}-x_{j})}=\prod_{i<j}|x_{i}-x_{j}|^{\beta} e − β ∑ i < j g ( x i − x j ) = ∏ i < j ∣ x i − x j ∣ β is the absolute value of the Vandermonde determinant to the power β \beta β .
Hard versus soft confinement. As is customary in the physics literature, we work with a hard wall : the points are forced to stay in the disk Λ N \Lambda_{N} Λ N , where the neutralizing background lives. By Lemma 2.1, this is the same as a quadratic potential π 2 ∣ x ∣ 2 \frac{\pi}{2}|x|^{2} 2 π ∣ x ∣ 2 acting on each point, restricted to Λ N \Lambda_{N} Λ N . In random matrix theory, one rather considers the soft version, without the constraint x i ∈ Λ N x_{i}\in\Lambda_{N} x i ∈ Λ N : the density is chosen to be proportional to ∏ i < j ∣ x i − x j ∣ β e − π β 2 ∑ i ∣ x i ∣ 2 \prod_{i<j}|x_{i}-x_{j}|^{\beta}e^{-\frac{\pi\beta}{2}\sum_{i}|x_{i}|^{2}} ∏ i < j ∣ x i − x j ∣ β e − 2 π β ∑ i ∣ x i ∣ 2 on ( R 2 ) N (\mathbb{R}^{2})^{N} ( R 2 ) N , sometimes called the two-dimensional β \beta β -ensemble. For β = 2 \beta=2 β = 2 , this gives the law of the (rescaled) eigenvalues of a Ginibre random matrix [43 ] .
The soft-wall and hard-wall models have (after rescaling) the same macroscopic equilibrium measure, uniform on a disk. Our results apply to the soft version as well: the concentration and covariance bounds below are uniform in the wall radius and pass to the soft gas by sending that radius to infinity at fixed N N N . The local laws of [5 ] and the soft-confinement version of the DLR argument described in [59 Section 6.5] then give the same identification of the infinite-volume limit.
Point processes and their correlations
Point configurations
X ( f ) : = ∑ x ∈ X f ( x ) . X(f) := \sum_{x\in X} f(x). X ( f ) := x ∈ X ∑ f ( x ) .
Point processes. A point process is a probability measure on Conf \operatorname{Conf} Conf , that is, the law of a random point configuration. The weak topology is the topology of convergence of E P [ f ] \mathbb{E}_{\mathrm{P}}[f] E P [ f ] for every f : Conf → R f:\operatorname{Conf}\to\mathbb{R} f : Conf → R which is bounded and continuous for the vague topology. We will also use the stronger local and exp-local topologies of [59 Section 4.2] , for which the test functions f f f are not assumed to be continuous, but are local (they only depend on the configuration in a bounded set) and satisfy a growth condition.
For a bounded Borel set Λ \Lambda Λ of positive area and n ≥ 0 n\geq0 n ≥ 0 , Bin Λ , n \operatorname{Bin}_{\Lambda,n} Bin Λ , n is the binomial point process , the law of the configuration made of n n n independent points uniformly distributed in Λ \Lambda Λ . The Poisson point process PPP \operatorname{PPP} PPP of intensity 1 1 1 is the law of a random configuration such that the numbers of points in disjoint bounded Borel sets are independent, and Poisson distributed with mean the area.
A point process P \mathrm{P} P is stationary if it is invariant under the translations τ u X : = ∑ x ∈ X δ x − u \tau_{u}X:=\sum_{x\in X}\delta_{x-u} τ u X := ∑ x ∈ X δ x − u , for u ∈ R 2 u\in\mathbb{R}^{2} u ∈ R 2 . It has intensity 1 1 1 when E P [ N Λ ] = ∣ Λ ∣ \mathbb{E}_{\mathrm{P}}[N_{\Lambda}]=|\Lambda| E P [ N Λ ] = ∣Λ∣ for every bounded Borel set Λ \Lambda Λ . It is locally square-integrable when E P [ N Λ 2 ] < ∞ \mathbb{E}_{\mathrm{P}}[N_{\Lambda}^{2}]<\infty E P [ N Λ 2 ] < ∞ for every bounded Borel set Λ \Lambda Λ . Finally, P \mathrm{P} P is ergodic if it is an extremal point of the convex set of stationary point processes: it cannot be written as a non-trivial mixture of two stationary processes.
Covariance measure and correlation functions. We follow the terminology of [35 Chapters 5 and 8] , see also [36 Section 9.5] . The second factorial moment measure of a locally square-integrable point process P \mathrm{P} P is the measure on R 2 × R 2 \mathbb{R}^{2}\times\mathbb{R}^{2} R 2 × R 2 that gives the mass
E P [ ∑ x , y ∈ X x ≠ y 1 A ( x ) 1 B ( y ) ] \mathbb{E}_{\mathrm{P}}\left[\sum_{\substack{x,y\in X\\x\ne y}} \mathbf{1}_{A}(x)\mathbf{1}_{B}(y)\right] E P x , y ∈ X x = y ∑ 1 A ( x ) 1 B ( y )
to A × B A\times B A × B , for bounded Borel sets A , B ⊂ R 2 A,B\subset\mathbb{R}^{2} A , B ⊂ R 2 . The two-point correlation function of a point process P \mathrm{P} P , if it exists, is the function ρ 2 : ( R 2 ) 2 → [ 0 , ∞ ) \rho_{2}:(\mathbb{R}^{2})^{2}\to[0,\infty) ρ 2 : ( R 2 ) 2 → [ 0 , ∞ ) such that, for every bounded measurable f f f with compact support,
E P [ ∑ x , y ∈ X x ≠ y f ( x , y ) ] = ∬ R 2 × R 2 f ( x , y ) ρ 2 ( x , y ) d x d y . (5) \mathbb{E}_{\mathrm{P}}\left[\sum_{\substack{x,y\in X\\x\ne y}} f(x,y)\right]
=
\iint_{\mathbb{R}^{2}\times\mathbb{R}^{2}} f(x,y)\rho_{2}(x,y)\,\mathrm{d}x\,\mathrm{d}y. \tag*{(5)} E P x , y ∈ X x = y ∑ f ( x , y ) = ∬ R 2 × R 2 f ( x , y ) ρ 2 ( x , y ) d x d y . (5)
In other words, ρ 2 \rho_{2} ρ 2 is the density of this measure with respect to the Lebesgue measure on R 2 × R 2 \mathbb{R}^{2}\times\mathbb{R}^{2} R 2 × R 2 . When this density exists and the process is stationary, the second factorial moment measure has the form ρ 2 ( 0 , y − x ) d x d y \rho_{2}(0,y-x)\,\mathrm{d}x\,\mathrm{d}y ρ 2 ( 0 , y − x ) d x d y , where ρ 2 ( 0 , ⋅ ) \rho_{2}(0,\cdot) ρ 2 ( 0 , ⋅ ) denotes its reduced density, defined almost everywhere on R 2 \mathbb{R}^{2} R 2 .
If P \mathrm{P} P is stationary, locally square-integrable and of intensity 1 1 1 , it has a reduced covariance measure Γ P \Gamma_{\mathrm{P}} Γ P . This is a locally finite signed measure on R 2 \mathbb{R}^{2} R 2 such that, for every bounded measurable F F F with compact support, we have:
E P [ ∬ R 2 × R 2 F ( z , y ) d ( X − Leb ) ( z ) d ( X − Leb ) ( y ) ] = ∫ R 2 ∫ R 2 F ( x , x + u ) d x Γ P ( d u ) . (6) \mathbb{E}_{\mathrm{P}}\left[\iint_{\mathbb{R}^{2}\times\mathbb{R}^{2}}F(z,y)\,\mathrm{d}(X-\operatorname{Leb})(z)\,\mathrm{d}(X-\operatorname{Leb})(y)\right]
=
\int_{\mathbb{R}^{2}}\int_{\mathbb{R}^{2}}F(x,x+u)\,\mathrm{d}x\,\Gamma_{\mathrm{P}}(\mathrm{d}u). \tag*{(6)} E P [ ∬ R 2 × R 2 F ( z , y ) d ( X − Leb ) ( z ) d ( X − Leb ) ( y ) ] = ∫ R 2 ∫ R 2 F ( x , x + u ) d x Γ P ( d u ) . (6)
In words, Γ P ( d u ) \Gamma_{\mathrm{P}}(\mathrm{d}u) Γ P ( d u ) is the covariance between the fluctuations at two points at relative position u u u . When the two-point correlation function exists, we have the identity
Γ P ( d u ) = δ 0 + ( ρ 2 ( 0 , u ) − 1 ) d u . \Gamma_{\mathrm{P}}(\mathrm{d}u)=\delta_{0}+(\rho_{2}(0,u)-1)\,\mathrm{d}u. Γ P ( d u ) = δ 0 + ( ρ 2 ( 0 , u ) − 1 ) d u .
Structure factor. The measure Γ P \Gamma_{\mathrm{P}} Γ P is positive definite, so, by the Bochner–Schwartz theorem, its Fourier transform is a positive tempered measure σ P \sigma_{\mathrm{P}} σ P , called the structure factor of P. With the Fourier transform convention φ ^ ( k ) : = ∫ R 2 e − i k ⋅ x φ ( x ) d x \widehat{\varphi}(k) := \int_{\mathbb{R}^{2}} e^{-ik\cdot x}\varphi(x)\,\mathrm{d}x φ ( k ) := ∫ R 2 e − ik ⋅ x φ ( x ) d x , it satisfies, for φ ∈ C c ∞ ( R 2 ; C ) \varphi\in C_{\mathrm{c}}^{\infty}(\mathbb{R}^{2};\mathbb{C}) φ ∈ C c ∞ ( R 2 ; C ) ,
E P ∣ X ( φ ) − ∫ φ ∣ 2 = 1 ( 2 π ) 2 ∫ ∣ φ ^ ( k ) ∣ 2 σ P ( d k ) . \mathbb{E}_{\mathrm{P}}\left|X(\varphi)-\int\varphi\right|^{2}=\frac{1}{(2\pi)^{2}}\int\left|\widehat{\varphi}(k)\right|^{2}\sigma_{\mathrm{P}}(\mathrm{d}k). E P X ( φ ) − ∫ φ 2 = ( 2 π ) 2 1 ∫ ∣ φ ( k ) ∣ 2 σ P ( d k ) .
More generally, for G G G continuous with compact support and G ^ \widehat{G} G integrable against σ P \sigma_{\mathrm{P}} σ P , we have
∫ G d Γ P = ( 2 π ) − 2 ∫ G ^ d σ P . (7) \int G\,\mathrm{d}\Gamma_{\mathrm{P}}=(2\pi)^{-2}\int\widehat{G}\,\mathrm{d}\sigma_{\mathrm{P}}. \tag*{(7)} ∫ G d Γ P = ( 2 π ) − 2 ∫ G d σ P . (7)
See [35 Chapter 8] , where σ P \sigma_{\mathrm{P}} σ P is called the Bartlett spectrum. It is a general fact that the structure factor is translation-bounded: sup ξ σ P ( B ( ξ , 1 ) ) < ∞ \sup_{\xi}\sigma_{\mathrm{P}}(B(\xi,1))<\infty sup ξ σ P ( B ( ξ , 1 )) < ∞ , hence, for r ≥ 1 r\ge1 r ≥ 1 ,
σ P ( B r ) ≤ C r 2 . \sigma_{\mathrm{P}}(B_r)\le Cr^{2}. σ P ( B r ) ≤ C r 2 .
Observables
We will work with two essentially equivalent types of observables.
Configuration observables. These are the measurable functions F : Conf → R F:\operatorname{Conf}\to\mathbb{R} F : Conf → R .
F ( X ) = φ ( X ( f 1 ) , … , X ( f J ) ) (8) F(X)=\varphi\bigl(X(f_1),\ldots,X(f_J)\bigr) \tag*{(8)} F ( X ) = φ ( X ( f 1 ) , … , X ( f J ) ) (8)
for some J ≥ 1 J\ge1 J ≥ 1 , where φ ∈ C b ∞ ( R J ; R ) \varphi\in C_{\mathrm{b}}^{\infty}(\mathbb{R}^{J};\mathbb{R}) φ ∈ C b ∞ ( R J ; R ) and the test functions ( f j ) j = 1 , … , J (f_j)_{j=1,\ldots,J} ( f j ) j = 1 , … , J are in C c ∞ ( R 2 ; R ) C_{\mathrm{c}}^{\infty}(\mathbb{R}^{2};\mathbb{R}) C c ∞ ( R 2 ; R ) . A cylinder observable only depends on finitely many “smoothed measurements” of the configuration. For Ω ⊂ R 2 \Omega\subset\mathbb{R}^{2} Ω ⊂ R 2 , a cylinder observable is said to be Ω \Omega Ω -local if all the supports of the f j f_j f j are contained in Ω \Omega Ω .
∇ z F ( X ) = ∑ k = 1 J ( ∇ φ ) k ( X ( f 1 ) , … , X ( f J ) ) ∇ f k ( z ) . \nabla_zF(X)=\sum_{k=1}^{J}(\nabla\varphi)_k\bigl(X(f_1),\ldots,X(f_J)\bigr)\nabla f_k(z). ∇ z F ( X ) = k = 1 ∑ J ( ∇ φ ) k ( X ( f 1 ) , … , X ( f J ) ) ∇ f k ( z ) .
D P ( F ) = E P ∑ z ∈ X ∣ ∇ z F ( X ) ∣ 2 . (9) D_{\mathrm{P}}(F)=\mathbb{E}_{\mathrm{P}}\sum_{z\in X}\left|\nabla_zF(X)\right|^{2}. \tag*{(9)} D P ( F ) = E P z ∈ X ∑ ∣ ∇ z F ( X ) ∣ 2 . (9)
It measures the average under P of the sensitivity of F F F to small displacements of the particles.
Functions of N N N positions. We sometimes work with labelled coordinates Z = ( z 1 , … , z N ) ∈ ( R 2 ) N Z=(z_1,\ldots,z_N)\in(\mathbb{R}^{2})^{N} Z = ( z 1 , … , z N ) ∈ ( R 2 ) N .
Such a function is said to be symmetric when F ( z σ ( 1 ) , … , z σ ( N ) ) = F ( Z ) F(z_{\sigma(1)},\ldots,z_{\sigma(N)})=F(Z) F ( z σ ( 1 ) , … , z σ ( N ) ) = F ( Z ) for every permutation σ \sigma σ .
It is said to be rotation-invariant when F ( e i θ z 1 , … , e i θ z N ) = F ( Z ) F(e^{i\theta}z_1,\ldots,e^{i\theta}z_N)=F(Z) F ( e i θ z 1 , … , e i θ z N ) = F ( Z ) for every θ \theta θ , where e i θ z e^{i\theta}z e i θ z is the rotation of z z z by the angle θ \theta θ .
D Q ( F ) = E Q ∑ i = 1 N ∣ ∇ i F ( Z ) ∣ 2 . D_Q(F)=\mathbb{E}_Q\sum_{i=1}^{N}\left|\nabla_iF(Z)\right|^{2}. D Q ( F ) = E Q i = 1 ∑ N ∣ ∇ i F ( Z ) ∣ 2 .
Equivalence of the two definitions. Every configuration observable F F F yields, for each N ≥ 1 N \ge1 N ≥ 1 , a restriction
F ( N ) ( Z ) = F ( ∑ i = 1 N δ z i ) F^{(N)}(Z)=F\left(\sum_{i=1}^{N}\delta_{z_i}\right) F ( N ) ( Z ) = F ( i = 1 ∑ N δ z i )
for Z = ( z 1 , … , z N ) ∈ ( R 2 ) N Z=(z_1,\ldots,z_N)\in(\mathbb{R}^2)^N Z = ( z 1 , … , z N ) ∈ ( R 2 ) N , {#eq:2.7}
which is a symmetric function of N N N positions. Conversely, a symmetric function of N N N positions can be lifted to a function of configurations with N N N points.
F ( N ) ( Z ) = φ ( ∑ i f 1 ( z i ) , … , ∑ i f J ( z i ) ) F^{(N)}(Z)=\varphi\left(\sum_i f_1(z_i),\ldots,\sum_i f_J(z_i)\right) F ( N ) ( Z ) = φ ( i ∑ f 1 ( z i ) , … , i ∑ f J ( z i ) )
is C ∞ C^\infty C ∞ on ( R 2 ) N (\mathbb{R}^2)^N ( R 2 ) N , bounded, with bounded gradient, and, for X = ∑ i δ z i X=\sum_i\delta_{z_i} X = ∑ i δ z i ,
∇ i F ( N ) ( Z ) = ∇ z i F ( X ) . \nabla_iF^{(N)}(Z)=\nabla_{z_i}F(X). ∇ i F ( N ) ( Z ) = ∇ z i F ( X ) .
E Q F ( N ) = E Q c o n f F , D Q ( F ( N ) ) = D Q c o n f ( F ) . \mathbb{E}_QF^{(N)}=\mathbb{E}_{Q^{\mathrm{conf}}}F,\qquad D_Q(F^{(N)})=D_{Q^{\mathrm{conf}}}(F). E Q F ( N ) = E Q conf F , D Q ( F ( N ) ) = D Q conf ( F ) .
Below, we use the same notation for a symmetric law on tuples and its image on configurations, and for an observable and its restriction to tuples. Relative entropy between symmetric laws is unchanged by this identification.
DLR equations and good processes
We recall here the Dobrushin–Lanford–Ruelle (DLR) framework of [59 ] for the 2DOCP.
A brief introduction to DLR equations. By construction, the finite-volume measure P N , β \mathbb{P}_{N,\beta} P N , β has a density d P N , β ∝ e − β F N d Bin Λ N , N \mathrm{d}\mathbb{P}_{N,\beta}\propto e^{-\beta F_N}\,\mathrm{d}\operatorname{Bin}_{\Lambda_N,N} d P N , β ∝ e − β F N d Bin Λ N , N with respect to the binomial point process. It is tempting to ask for a similar description of the infinite-volume limit points P P P , with a global density of the form
d P ∝ e − β ⋅ energy of X d PPP ( X ) . \mathrm{d}P \propto e^{-\beta\cdot\text{energy of }X}\,\mathrm{d}\operatorname{PPP}(X). d P ∝ e − β ⋅ energy of X d PPP ( X ) .
Unfortunately, this does not make sense. The correct framework, due to Dobrushin, Lanford and Ruelle [34 , 65 ] , is local : one describes, for every bounded domain Λ \Lambda Λ , the law of the configuration X Λ X_\Lambda X Λ inside Λ \Lambda Λ conditionally on the configuration X Λ c X_{\Lambda^c} X Λ c outside , as a finite-volume Gibbs measure
d P ( X Λ ∣ X Λ c ) ∝ exp ( − β [ interaction energy of X Λ with itself and with X Λ c ] ) d ( reference measure ) ( X Λ ) . \mathrm{d}P(X_\Lambda\mid X_{\Lambda^c})\propto\exp\left(-\beta\left[\text{interaction energy of }X_\Lambda\text{ with itself and with }X_{\Lambda^c}\right]\right)\,\mathrm{d}(\text{reference measure})(X_\Lambda). d P ( X Λ ∣ X Λ c ) ∝ exp ( − β [ interaction energy of X Λ with itself and with X Λ c ] ) d ( reference measure ) ( X Λ ) .
Gibbs states are the point processes which satisfy these equations for every Λ \Lambda Λ , see [41 Chapter 6] and [42 ] .
Two features are specific to our setting. First, the interaction is long-range, so the energy of X Λ X_\Lambda X Λ with the infinite configuration X Λ c X_{\Lambda^c} X Λ c needs to be properly defined (this was done in [59 ] ). Second, the equations are canonical , which means that the number of points in Λ \Lambda Λ must be specified. In fact, it is determined by the exterior configuration (this phenomenon is called number rigidity).
Energy in a domain. For a bounded Borel set Λ \Lambda Λ and a finite configuration X ~ \widetilde{X} X in Λ \Lambda Λ , we let
F Λ ( X ~ ) : = 1 2 ∬ ( x , y ) ∈ Λ × Λ x ≠ y g ( x − y ) d ( X ~ − Leb ) ( x ) d ( X ~ − Leb ) ( y ) . (10) F_\Lambda(\widetilde{X}):=\frac{1}{2}\iint_{\substack{(x,y)\in\Lambda\times\Lambda\\x\ne y}}g(x-y)\,\mathrm{d}\left(\widetilde{X}-\operatorname{Leb}\right)(x)\,\mathrm{d}\left(\widetilde{X}-\operatorname{Leb}\right)(y). \tag*{(10)} F Λ ( X ) := 2 1 ∬ ( x , y ) ∈ Λ × Λ x = y g ( x − y ) d ( X − Leb ) ( x ) d ( X − Leb ) ( y ) . (10)
It is the energy of the points of X ~ \widetilde{X} X and of the uniform background of density 1 1 1 in Λ \Lambda Λ . For Λ = Λ N \Lambda=\Lambda_N Λ = Λ N , it is the energy F N F_N F N of (1) .
Move functions. Let X ∈ Conf X \in\operatorname{Conf} X ∈ Conf , and let X ~ \widetilde{X} X be a configuration in Λ \Lambda Λ with the same number of points as X Λ X_{\Lambda} X Λ . We define the move function as in [59 (1.7)]
M Λ ( X ~ , X ) : = ∬ Λ × Λ c g ( x − y ) d ( X ~ − X Λ ) ( x ) d ( X Λ c − Leb ) ( y ) . (11) M_{\Lambda}(\widetilde{X},X) := \iint_{\Lambda\times\Lambda^{\mathrm{c}}} g(x-y) \,\mathrm{d}(\widetilde{X}-X_{\Lambda})(x)\,\mathrm{d}(X_{\Lambda^{\mathrm{c}}}-\operatorname{Leb})(y). \tag*{(11)} M Λ ( X , X ) := ∬ Λ × Λ c g ( x − y ) d ( X − X Λ ) ( x ) d ( X Λ c − Leb ) ( y ) . (11)
It is the variation of the energy between Λ \Lambda Λ and its exterior when the configuration X Λ X_{\Lambda} X Λ is replaced by X ~ \widetilde{X} X .
Remark 2.2 (About the definition of M Λ M_{\Lambda} M Λ ). The integral in y y y is not absolutely convergent hence the mere definition of M Λ M_{\Lambda} M Λ requires some care. In [59 ] , the integral in y y y is truncated with a smooth partition of unity ( ϑ j ) j ≥ 0 (\vartheta_j)_{j \ge0} ( ϑ j ) j ≥ 0 adapted to dyadic annuli, and M Λ ( X ~ , X ) M_{\Lambda}(\widetilde{X},X) M Λ ( X , X ) is the limit, as p → ∞ p \to\infty p → ∞ , of the integrals with ∑ j = 0 p ϑ j ( y ) \sum_{j=0}^{p}\vartheta_j(y) ∑ j = 0 p ϑ j ( y ) inserted. For almost every X X X under a limit point of the 2DOCP, this limit exists for every X ~ \widetilde{X} X .
Canonical DLR equations. For X ∈ Conf X \in\operatorname{Conf} X ∈ Conf and a bounded Borel set Λ \Lambda Λ , let n = N Λ ( X ) n=N_{\Lambda}(X) n = N Λ ( X ) , and let γ Λ ( ⋅ ∣ X ) \gamma_{\Lambda}(\cdot\mid X) γ Λ ( ⋅ ∣ X ) be the law of the configuration X ~ + X Λ c \widetilde{X}+X_{\Lambda^{\mathrm{c}}} X + X Λ c , where X ~ \widetilde{X} X is distributed according to
1 K Λ ( X ) e − β ( F Λ ( X ~ ) + M Λ ( X , X ~ ) ) d B i n Λ , n ( X ~ ) , \frac{1}{K_{\Lambda}(X)}e^{-\beta\left(F_{\Lambda}(\widetilde{X})+M_{\Lambda}(X,\widetilde{X})\right)}\,\mathrm{dBin}_{\Lambda,n}(\widetilde{X}), K Λ ( X ) 1 e − β ( F Λ ( X ) + M Λ ( X , X ) ) dBin Λ , n ( X ) ,
with
K Λ ( X ) : = ∫ e − β ( F Λ ( X ~ ) + M Λ ( X , X ~ ) ) d B i n Λ , n ( X ~ ) . (12) K_{\Lambda}(X):=\int e^{-\beta\left(F_{\Lambda}(\widetilde{X})+M_{\Lambda}(X,\widetilde{X})\right)}\,\mathrm{dBin}_{\Lambda,n}(\widetilde{X}). \tag*{(12)} K Λ ( X ) := ∫ e − β ( F Λ ( X ) + M Λ ( X , X ) ) dBin Λ , n ( X ) . (12)
A point process P \mathrm{P} P is a D L R DLR D L R state if, for every bounded Borel set Λ \Lambda Λ and every bounded measurable F : Conf → R F:\operatorname{Conf}\to\mathbb{R} F : Conf → R ,
E P [ F ] = E P [ ∫ F d γ Λ ( ⋅ ∣ X ) ] . (13) \mathbb{E}_{\mathrm{P}}[F]=\mathbb{E}_{\mathrm{P}}\left[\int F\,\mathrm{d}\gamma_{\Lambda}(\cdot\mid X)\right]. \tag*{(13)} E P [ F ] = E P [ ∫ F d γ Λ ( ⋅ ∣ X ) ] . (13)
In words: conditionally on the exterior configuration X Λ c X_{\Lambda^{\mathrm{c}}} X Λ c and on the number n n n of points in Λ \Lambda Λ , the configuration in Λ \Lambda Λ is a finite 2DOCP with n n n points in Λ \Lambda Λ , subject to the field created by the exterior. The equations are canonical : the number of points in Λ \Lambda Λ is not resampled. Here γ Λ ( ⋅ ∣ X ) \gamma_{\Lambda}(\cdot\mid X) γ Λ ( ⋅ ∣ X ) depends on X X X only through X Λ c X_{\Lambda^{\mathrm{c}}} X Λ c and n n n , and (13) says that for P \mathrm{P} P -almost every X X X , with n = N Λ ( X ) n=N_{\Lambda}(X) n = N Λ ( X ) ,
P ( X Λ ∈ ⋅ ∣ X Λ c , N Λ = n ) = law of X ~ under x r e f ‘ ( 2.10 ) ‘ . (14) \mathrm{P}\left(X_{\Lambda}\in\cdot\mid X_{\Lambda^{\mathrm{c}}},N_{\Lambda}=n\right)=\operatorname{law}\text{ of }\widetilde{X}\text{ under }{xref}`(2.10)`. \tag*{(14)} P ( X Λ ∈ ⋅ ∣ X Λ c , N Λ = n ) = law of X under x r e f ‘ ( 2.10 ) ‘. (14)
This is the sense in which a DLR state is described by its local conditional laws, rather than by a global formula such as a Gibbs density.
Good processes. The DLR equations only make sense when the move functions are well-defined. A class of point processes for which this is the case was introduced in [59 Section 8.3] . It includes bounds on the one-point density, on Lipschitz fluctuations and on the energy variations under localized translations, which are all satisfied by the limit points of the 2DOCP, and were used in [59 ] to ensure that the move functions are defined and that the limit points are stationary.
In the present paper, we will use two moment bounds that were included in the requirements of [59 ] (see (G1) and (G2) below), and we place an additional assumption on the decay of correlations.
Definition 2.3 (Good processes). Fix β > 0 \beta>0 β > 0 . We say that P \mathrm{P} P is good if it is good in the sense of [59 Section 8.3] (this includes the bounds (G1) and (G2) below), and satisfies (G3) below. The constants C P C_{\mathrm{P}} C P and s P > 0 s_{\mathrm{P}}>0 s P > 0 may depend on P \mathrm{P} P , but neither on x ∈ R 2 x\in\mathbb{R}^{2} x ∈ R 2 nor on r ≥ 1 r\ge1 r ≥ 1 .
(G1) Number of points. We have log E P exp ( C P − 1 N B ( x , r ) 2 ) ≤ C P r 4 \log\mathbb{E}_{\mathrm{P}}\exp\left(C_{\mathrm{P}}^{-1}N_{B(x,r)}^{2}\right)\le C_{\mathrm{P}}r^{4} log E P exp ( C P − 1 N B ( x , r ) 2 ) ≤ C P r 4 .
(G2) Smooth linear statistics. Let φ \varphi φ be a function of class C 3 C^{3} C 3 , supported in B ( x , r ) B(x,r) B ( x , r ) , with r k ∥ ∇ k φ ∥ ∞ ≤ 1 r^{k}\lVert\nabla^{k}\varphi\rVert_{\infty}\le1 r k ∥ ∇ k φ ∥ ∞ ≤ 1 for 0 ≤ k ≤ 3 0\le k\le3 0 ≤ k ≤ 3 . Then Fluct [ φ ] : = X ( φ ) − ∫ φ \operatorname{Fluct}[\varphi]:=X(\varphi)-\int\varphi Fluct [ φ ] := X ( φ ) − ∫ φ satisfies:
log E P exp ( s Fluct [ φ ] ) ≤ C P for ∣ s ∣ ≤ s P . (15) \log\mathbb{E}_{\mathrm{P}}\exp\left(s\operatorname{Fluct}[\varphi]\right)\le C_{\mathrm{P}}\qquad\text{for }|s|\le s_{\mathrm{P}}. \tag*{(15)} log E P exp ( s Fluct [ φ ] ) ≤ C P for ∣ s ∣ ≤ s P . (15) (G3) Decay of correlations. The process P \mathrm{P} P has a two-point correlation function ρ 2 \rho_{2} ρ 2 , which satisfies
∬ B 1 × R 2 ∣ log ∣ y − x ∣ ∣ ∣ ρ 2 ( x , y ) − 1 ∣ d y d x < ∞ . \iint_{B_{1}\times\mathbb{R}^{2}} \left|\log|y-x|\right|\left|\rho_{2}(x,y)-1\right|\,\mathrm{d}y\,\mathrm{d}x<\infty. ∬ B 1 × R 2 ∣ log ∣ y − x ∣ ∣ ∣ ρ 2 ( x , y ) − 1 ∣ d y d x < ∞. For a stationary process, this is equivalent to
∫ R 2 ∣ log ∣ u ∣ ∣ ∣ ρ 2 ( 0 , u ) − 1 ∣ d u < ∞ . (16) \int_{\mathbb{R}^{2}}|\log|u||\left|\rho_{2}(0,u)-1\right|\,\mathrm{d}u<\infty. \tag*{(16)} ∫ R 2 ∣ log ∣ u ∣∣ ∣ ρ 2 ( 0 , u ) − 1 ∣ d u < ∞. (16)
Every limit point of the 2DOCP belongs to the class of [59 Section 8.3] , for every β > 0 \beta> 0 β > 0 . For the limit points of the configuration seen from the origin, the additional condition (G3) follows when 0 < β ≤ 2 0 < \beta\le2 0 < β ≤ 2 from the exponential decay of correlations (Corollary 5.11). We do not know whether this remains true for large β \beta β .
Some imported results.
Proposition 2.4. For all β > 0 \beta> 0 β > 0 :
(i) The laws of the local configurations X N X_N X N under P N , β P_{N,\beta} P N , β have limit points as N → ∞ N \to\infty N → ∞ in the local topology. Every limit point solves DLR equations, belongs to the class of [59 Section 8.3] , and is invariant under rotations around the origin.
(ii) Every DLR state in the class of [59 Section 8.3] is stationary, with intensity 1 1 1 . In particular, this holds for every good DLR state and every limit point in (i).
(iii) Along a converging subsequence, the expectations of local continuous observables with polynomial growth in a local count converge.
Proof.
Proof of (i). The existence of limit points in the weak topology is [5 Corollary 1.1] , and the convergence has been upgraded to a stronger topology in [59 Section 2.3] . Limit points satisfy the DLR equations by [59 Theorem 1] , and they belong to the class of [59 Section 8.3] , hence satisfy (G1) and (G2). The bounds on the energy under localized translations in this class come from [59 Claim 8.5] , and we correct one step of its proof with Lemma B.5. In the first and last layers, this proof bounds the contribution of κ i , 2 \kappa_{i,2} κ i , 2 by
( X N ( supp χ i ) + ∣ supp χ i ∣ ) sup y ∈ R 2 ∣ Fluct [ κ i , 2 ( ⋅ , y ) ] ∣ . \left(X_N(\operatorname{supp}\chi_i)+\lvert\operatorname{supp}\chi_i\rvert\right)\sup_{y\in\mathbb{R}^2}\lvert\operatorname{Fluct}[\kappa_{i,2}(\cdot,y)]\rvert. ( X N ( supp χ i ) + ∣ supp χ i ∣ ) y ∈ R 2 sup ∣ Fluct [ κ i , 2 ( ⋅ , y )]∣ .
We apply Lemma B.5 to the function 2 2 i κ i , 2 ( x , 2 i + 1 y ) 2^{2i}\kappa_{i,2}(x,2^{i+1}y) 2 2 i κ i , 2 ( x , 2 i + 1 y ) , divided by a large constant. Its derivatives are bounded by [59 (8.15)–(8.17)] and by [59 (8.2)] with k = 3 k=3 k = 3 , which has the same proof. In the first layer, we have 2 i + 1 ≥ 10 L 2^{i+1}\ge10L 2 i + 1 ≥ 10 L , and we take R = 2 i + 1 R=2^{i+1} R = 2 i + 1 . Then B R B_R B R contains the support of χ i \chi_i χ i and the disk B 2 L B_{2L} B 2 L , outside which κ i , 2 ( ⋅ , y ) \kappa_{i,2}(\cdot,y) κ i , 2 ( ⋅ , y ) vanishes, and the product above is at most C ( 1 + N R / R 2 ) C(1+N_R/R^2) C ( 1 + N R / R 2 ) times the supremum in the lemma. In the last layer, the number of points is at most N ≤ C 2 2 i N\le C2^{2i} N ≤ C 2 2 i , so the first factor is at most C 2 2 i C2^{2i} C 2 2 i , and we take R = 3 L R=3L R = 3 L .
Since P N , β P_{N,\beta} P N , β is invariant under rotations, so is the law of X N X_N X N , and this passes to the limit.
Proof of (ii). Stationarity is [59 Theorem 3] . Let ρ \rho ρ be the intensity, which is finite by (G1). We apply (G2) to c φ ( ⋅ / R ) c\varphi(\cdot/R) c φ ( ⋅ / R ) , where φ ∈ C ∞ ( B 1 ) \varphi\in C^\infty(B_1) φ ∈ C ∞ ( B 1 ) is non-negative with positive integral and c > 0 c>0 c > 0 makes the derivative bounds hold. Jensen’s inequality, with both signs of s s s , bounds ∣ c ( ρ − 1 ) R 2 ∫ φ ∣ \lvert c(\rho-1)R^2\int\varphi\rvert ∣ c ( ρ − 1 ) R 2 ∫ φ ∣ independently of R R R . Letting R → ∞ R\to\infty R → ∞ gives ρ = 1 \rho=1 ρ = 1 .
Proof of (iii). See the definition of exp-tame convergence [59 Sections 3.2 and 4.1] .
As a consequence of (15) and stationarity, good DLR states are hyperuniform .
Proposition 2.5 (Hyperuniformity). Let P P P be a good DLR state. There exists C P C_P C P such that:
(i) (spectral hyperuniformity) the structure factor of P P P satisfies
σ P ( B ε ) ≤ C P ε 4 for 0 < ε ≤ 1 ; (17) \sigma_P(B_\varepsilon)\le C_P\varepsilon^4 \qquad\text{for }0<\varepsilon\le1; \tag*{(17)} σ P ( B ε ) ≤ C P ε 4 for 0 < ε ≤ 1 ; (17) (ii) (geometric hyperuniformity) for all x ∈ R 2 x\in\mathbb{R}^2 x ∈ R 2 and R ≥ 1 R\ge1 R ≥ 1 ,
Var P [ N B ( x , R ) ] ≤ C P R . \operatorname{Var}_P[N_{B(x,R)}]\le C_P R. Var P [ N B ( x , R ) ] ≤ C P R .
Proof. For stationary point processes, bounds on smooth linear statistics translate into bounds on the number of points in large disks and into low-frequency estimates for the structure factor. This is the equivalence between spectral and geometric hyperuniformity of Björklund and Hartnick [16 Proposition 3.3 and Theorem 3.6] , in a quantitative form in [66 ] .
For independent points, the number variance would be of the order R 2 R^2 R 2 of the area. A stationary point process whose number variance in disks is o ( R 2 ) o(R^2) o ( R 2 ) is called hyperuniform . If its number variance is O ( R ) O(R) O ( R ) , then it is hyperuniform of class I , in the terminology of [88 ] . The order R R R of the perimeter cannot be improved, by a theorem of Beck [9 ] .
Remark 2.6 (Condition (G3) and hyperuniformity with respect to squares). For a stationary process, condition (G3) says that the integral 1 2 ∫ g ( u ) ( ρ 2 ( 0 , u ) − 1 ) d u \frac{1}{2}\int g(u)(\rho_2(0,u)-1)\,du 2 1 ∫ g ( u ) ( ρ 2 ( 0 , u ) − 1 ) d u , which is the natural candidate for the energy per unit area (see [57 ] ), converges absolutely. It is much weaker than the decay assumptions under which sum rules are usually derived in the physics literature, see [71 ] for a review, and [72 ] for the second moment sum rule of [82 ] .
The same integrability condition appears in [67 Theorem 2] , and, together with hyperuniformity, it implies the ⋆ \star ⋆ -hyperuniformity of [48 ] . One upside for us is that (G3) controls the number variance in squares as well as in disks (see the proof of Lemma 6.1 ).
This condition is not implied by hyperuniformity in disks, even of the strongest form. For instance, the stationary square lattice Z 2 + U \mathbb{Z}^2+U Z 2 + U , with U U U uniform in [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 , has a number variance O ( R ) O(R) O ( R ) in disks, but its number variance in squares can be of order R 2 R^2 R 2 .
Strategy
Subsection 3.1 explains the concentration of observables, the tool behind all our results. Subsection 3.2 presents the Combes–Thomas argument behind the decay of correlations of Theorem 1 . Subsection 3.3 explains how concentration, together with Proposition 1.1 , gives the uniqueness of Theorem 2 . The decay of correlations enters there through the condition (G3). Subsection 3.4 explains how these estimates determine the Debye–Hückel profile of Theorem 5 . The proofs follow in the same order in § 6 to § 11 .
A central tool: concentration of observables
The main tool of this paper is a result about concentration of observables for the 2DOCP at β ≤ 2 \beta\le2 β ≤ 2 .
From the classical to the complex case. A classical case of concentration for a Gibbs measure is that of an energy with uniformly positive Hessian. Let μ = e − Φ d x \mu=e^{-\Phi}\,dx μ = e − Φ d x be a probability measure on R n \mathbb{R}^n R n with ∇ 2 Φ ≥ a > 0 \nabla^2\Phi\ge a>0 ∇ 2 Φ ≥ a > 0 . Then the Brascamp–Lieb inequality [17 ] gives, for every smooth function F : R n → R F:\mathbb{R}^n\to\mathbb{R} F : R n → R ,
Var μ [ F ] ≤ a − 1 E μ [ ∣ ∇ F ∣ 2 ] , (18) \operatorname{Var}_{\mu}[F]\le a^{-1}\mathbb{E}_{\mu}[|\nabla F|^2], \tag*{(18)} Var μ [ F ] ≤ a − 1 E μ [ ∣∇ F ∣ 2 ] , (18)
with a constant that does not depend on the dimension. Unfortunately, the Hessian of g g g at z ∈ R 2 ∖ { 0 } z\in\mathbb{R}^2\setminus\{0\} z ∈ R 2 ∖ { 0 } has the eigenvalues ∣ z ∣ − 2 |z|^{-2} ∣ z ∣ − 2 and − ∣ z ∣ − 2 -|z|^{-2} − ∣ z ∣ − 2 , so the Hessian of H N H_N H N has no lower bound when two particles come close. However, these two eigenvalues cancel in the Laplacian, which suggests using a complex version of the Hessian.
We write Z = ( z 1 , … , z N ) ∈ C N Z=(z_1,\ldots,z_N)\in\mathbb{C}^N Z = ( z 1 , … , z N ) ∈ C N for the positions. For 1 ≤ j ≤ N 1\le j\le N 1 ≤ j ≤ N , let ∂ j = 1 2 ( ∂ Re z j − i ∂ Im z j ) \partial_j=\frac{1}{2}(\partial_{\operatorname{Re}z_j}-i\partial_{\operatorname{Im}z_j}) ∂ j = 2 1 ( ∂ Re z j − i ∂ Im z j ) and ∂ ˉ j = 1 2 ( ∂ Re z j + i ∂ Im z j ) \bar{\partial}_j=\frac{1}{2}(\partial_{\operatorname{Re}z_j}+i\partial_{\operatorname{Im}z_j}) ∂ ˉ j = 2 1 ( ∂ Re z j + i ∂ Im z j ) . The Levi matrix (or complex Hessian ) of a function Φ \Phi Φ of Z Z Z is ( ∂ j ∂ ˉ k Φ ) j , k (\partial_j\bar{\partial}_k\Phi)_{j,k} ( ∂ j ∂ ˉ k Φ ) j , k . For one particle, it is a quarter of the Laplacian of Φ \Phi Φ . Write ∂ ˉ G = ( ∂ ˉ j G ) 1 ≤ j ≤ N \bar{\partial}G=(\bar{\partial}_jG)_{1\le j\le N} ∂ ˉ G = ( ∂ ˉ j G ) 1 ≤ j ≤ N , so that the holomorphic functions are the solutions of ∂ ˉ G = 0 \bar{\partial}G=0 ∂ ˉ G = 0 .
On a product of disks in C N \mathbb{C}^N C N , let ∥ ⋅ ∥ Φ \|\cdot\|_{\Phi} ∥ ⋅ ∥ Φ be the norm of L 2 ( e − Φ d Z ) L^2(e^{-\Phi}\,dZ) L 2 ( e − Φ d Z ) and let Π ker ∂ ˉ \Pi_{\ker\bar{\partial}} Π k e r ∂ ˉ be the orthogonal projection onto the holomorphic functions of this space. If the Levi matrix of Φ \Phi Φ is bounded below by a > 0 a>0 a > 0 , then the L 2 L^2 L 2 estimate of Hörmander [Hör65] gives, for every smooth function G G G with finite norm,
∥ G − Π ker ∂ ˉ G ∥ Φ 2 ≤ a − 1 ∥ ∂ ˉ G ∥ Φ 2 , (19) \|G-\Pi_{\ker\bar{\partial}}G\|_{\Phi}^2\le a^{-1}\|\bar{\partial}G\|_{\Phi}^2, \tag*{(19)} ∥ G − Π k e r ∂ ˉ G ∥ Φ 2 ≤ a − 1 ∥ ∂ ˉ G ∥ Φ 2 , (19)
where ∥ ∂ ˉ G ∥ Φ 2 : = ∑ j ∥ ∂ ˉ j G ∥ Φ 2 \|\bar{\partial}G\|_{\Phi}^2:=\sum_j\|\bar{\partial}_jG\|_{\Phi}^2 ∥ ∂ ˉ G ∥ Φ 2 := ∑ j ∥ ∂ ˉ j G ∥ Φ 2 . In the Brascamp–Lieb inequality, the variance is the squared distance from the observable to the constants, which form the kernel of ∇ \nabla ∇ . Here, we control the distance to the kernel of ∂ ˉ \bar{\partial} ∂ ˉ , that is, to the holomorphic functions. This space is much larger than the constants, and we call (19) a fake Poincaré inequality .
Shifting weights to gain positivity. We would like to apply (19) with Φ = β H N \Phi=\beta H_N Φ = β H N , on the product of disks Λ N N \Lambda_N^N Λ N N (by Lemma 2.1 , e − β H N e^{-\beta H_N} e − β H N is, up to a constant, the density of P N , β \mathbb{P}_{N,\beta} P N , β , hence ∥ F ∥ β H N 2 = K N , β E P N , β [ ∣ F ∣ 2 ] \|F\|_{\beta H_N}^2=K_{N,\beta}\mathbb{E}_{\mathbb{P}_{N,\beta}}[|F|^2] ∥ F ∥ β H N 2 = K N , β E P N , β [ ∣ F ∣ 2 ] : this is the physical L 2 L^2 L 2 space). We thus need a positive lower bound on the Levi matrix of our energy β H N \beta H_N β H N . The quadratic confining potential gives π β / 2 \pi\beta/2 π β /2 times the identity. Away from collisions, that is, when the z i z_i z i 's are distinct, the interaction gives nothing, because the logarithm is harmonic. But at collisions, the identity Δ g = − 2 π δ 0 \Delta g=-2\pi\delta_0 Δ g = − 2 π δ 0 gives a Dirac mass with negative sign .
To get around this problem, we shift things by writing β = 2 + ( β − 2 ) \beta=2+(\beta-2) β = 2 + ( β − 2 ) , which somehow corresponds to “moving the Dirac masses into the test function”. Let us explain. Let V N ( Z ) = ∏ i < j ( z i − z j ) V_N(Z)=\prod_{i<j}(z_i-z_j) V N ( Z ) = ∏ i < j ( z i − z j ) be the Vandermonde polynomial, and define Φ s h ( Z ) = π β 2 ∣ Z ∣ 2 + ( β − 2 ) ∑ i < j g ( z i − z j ) \Phi^{\mathrm{sh}}(Z)=\frac{\pi\beta}{2}|Z|^2+(\beta-2)\sum_{i<j}g(z_i-z_j) Φ sh ( Z ) = 2 π β ∣ Z ∣ 2 + ( β − 2 ) ∑ i < j g ( z i − z j ) as our “shifted weight”. Since
∣ V N ( Z ) ∣ 2 = e − 2 ∑ i < j g ( z i − z j ) , |V_N(Z)|^2=e^{-2\sum_{i<j}g(z_i-z_j)}, ∣ V N ( Z ) ∣ 2 = e − 2 ∑ i < j g ( z i − z j ) ,
the density of P N , β \mathbb{P}_{N,\beta} P N , β can be rewritten as ∣ V N ∣ 2 e − Φ s h |V_N|^2e^{-\Phi^{\mathrm{sh}}} ∣ V N ∣ 2 e − Φ sh (up to some multiplicative constant). The good news is that, in Φ s h \Phi^{\mathrm{sh}} Φ sh , the interaction receives the coefficient β − 2 \beta-2 β − 2 , so when β ≤ 2 \beta\le2 β ≤ 2 the Levi matrix of Φ s h \Phi^{\mathrm{sh}} Φ sh is bounded below (by π β / 2 \pi\beta/2 π β /2 ).
Shifting the observable. We may then use Hörmander’s estimate in the shifted space L 2 ( e − Φ s h ) L^2(e^{-\Phi^{\mathrm{sh}}}) L 2 ( e − Φ sh ) , whose norm we denote by ∥ ⋅ ∥ s h \|\cdot\|_{\mathrm{sh}} ∥ ⋅ ∥ sh . The price to pay is that, if we want to study an observable F F F in the physical space L 2 ( e − β H N ) L^2(e^{-\beta H_N}) L 2 ( e − β H N ) , we need to apply Hörmander’s estimate to a different observable, namely G = V N F G=V_NF G = V N F , in the shifted space. The Dirac masses (implicitly carried by V N V_N V N ) have been “moved” to our observable.
This all may seem far-fetched: in view of (19) , given an observable F F F that we care about,
We need to compute ∥ ∂ ˉ G ∥ s h 2 \|\bar{\partial}G\|_{\mathrm{sh}}^2 ∥ ∂ ˉ G ∥ sh 2 in the shifted space L 2 ( e − Φ s h ) L^2(e^{-\Phi^{\mathrm{sh}}}) L 2 ( e − Φ sh ) for G = V N F G=V_NF G = V N F and relate it to some interesting quantity about F F F .
Worse, we need to understand the distance from G = V N F G=V_NF G = V N F to its orthogonal projection (in the shifted space) onto holomorphic functions...
Two small miracles make this work.
First, V N V_N V N itself is holomorphic, so multiplication by V N V_N V N commutes with ∂ ˉ \bar{\partial} ∂ ˉ and we have ∂ ˉ G = V N ∂ ˉ F \bar{\partial}G=V_N\bar{\partial}F ∂ ˉ G = V N ∂ ˉ F . Hence, since ∣ V N ∣ 2 e − Φ s h = e − β H N |V_N|^2e^{-\Phi^{\mathrm{sh}}}=e^{-\beta H_N} ∣ V N ∣ 2 e − Φ sh = e − β H N , we have ∥ ∂ ˉ G ∥ s h 2 = ∥ ∂ ˉ F ∥ β H N 2 = 1 4 ∥ ∇ F ∥ β H N 2 \|\bar{\partial}G\|_{\mathrm{sh}}^2=\|\bar{\partial}F\|_{\beta H_N}^2=\frac14\|\nabla F\|_{\beta H_N}^2 ∥ ∂ ˉ G ∥ sh 2 = ∥ ∂ ˉ F ∥ β H N 2 = 4 1 ∥∇ F ∥ β H N 2 for a real F F F . This answers our first concern.
Second, in many interesting cases we can actually compute the relevant orthogonal projections and find that they are trivial. Here the logarithmic interaction, through V N V_N V N , plays a very special role.
Obtaining a true Poincaré inequality. Let F F F be an observable which is invariant under permutations of the coordinates and under rotation of all coordinates. Then the projection of G = V N F G=V_NF G = V N F in the shifted space is a constant multiple of V N V_N V N . Indeed, V N V_N V N , and hence G G G , changes sign when two particles are exchanged, and G ( e i θ Z ) = e i θ N ( N − 1 ) / 2 G ( Z ) G(e^{i\theta}Z)=e^{i\theta N(N-1)/2}G(Z) G ( e i θ Z ) = e i θ N ( N − 1 ) /2 G ( Z ) . Since Φ s h \Phi^{\mathrm{sh}} Φ sh itself is invariant under permutations and rotations, the projection of G G G has the same two properties, and the only holomorphic functions with these properties are the multiples of V N V_N V N (Lemma 4.11). If moreover F F F is centred under P N , β \mathbb{P}_{N,\beta} P N , β , then the projection vanishes, and (19) for Φ s h \Phi^{\mathrm{sh}} Φ sh reads ∥ F ∥ β H N 2 ≤ a − 1 ∥ ∂ ˉ F ∥ β H N 2 \|F\|_{\beta H_N}^2\le a^{-1}\|\bar{\partial}F\|_{\beta H_N}^2 ∥ F ∥ β H N 2 ≤ a − 1 ∥ ∂ ˉ F ∥ β H N 2 , with a = π β / 2 a=\pi\beta/2 a = π β /2 . With this, we obtain a true Poincaré inequality for rotation-invariant observables, of the form
Var P N , β [ F ] ≤ 1 2 π β E P N , β [ ∣ ∇ F ∣ 2 ] . \operatorname{Var}_{\mathbb{P}_{N,\beta}}[F]\le\frac{1}{2\pi\beta}\mathbb{E}_{\mathbb{P}_{N,\beta}}[|\nabla F|^2]. Var P N , β [ F ] ≤ 2 π β 1 E P N , β [ ∣∇ F ∣ 2 ] .
This inequality then extends (with the constant 2 / ( π β ) 2/(\pi\beta) 2/ ( π β ) instead of 1 / ( 2 π β ) 1/(2\pi\beta) 1/ ( 2 π β ) ) to all permutation-invariant observables F F F (Subsection 4.8).
Exponential concentration. The Poincaré inequality alone does not give concentration at the exponential level, and we need such tails. We therefore use the argument of Herbst (Subsection 4.7) and tilt the reference measure . We must choose our tilts carefully: let f f f be a non-negative C 2 C^2 C 2 function on the plane with compact support, let B R B_R B R be the disk of radius R R R centred at the origin, and define S R f S_R^f S R f as:
S R f ( z ) = ∫ B R ∫ 0 2 π exp ( − ∑ i f ( e − i θ ( z i − x ) ) ) d θ 2 π d x . S_R^f(z)=\int_{B_R}\int_0^{2\pi}\exp\left(-\sum_i f\left(e^{-\mathrm{i}\theta}(z_i-x)\right)\right)\frac{\mathrm{d}\theta}{2\pi}\,\mathrm{d}x. S R f ( z ) = ∫ B R ∫ 0 2 π exp ( − i ∑ f ( e − i θ ( z i − x ) ) ) 2 π d θ d x .
This observable is symmetric and rotation-invariant. Its Levi matrix is bounded uniformly in N N N and R R R , and ∣ ∇ S R f ∣ 2 \lvert\nabla S_R^f\rvert^2 ∣ ∇ S R f ∣ 2 is at most of order R 2 R^2 R 2 . For small t t t , multiplying e − ϕ f e^{-\phi^f} e − ϕ f by e t S R f e^{tS_R^f} e t S R f , that is, adding − t S R f -tS_R^f − t S R f to the weight, preserves both the symmetries and the positive lower bound on the Levi matrix. The Poincaré inequality therefore still holds.
Integrating those Poincaré inequalities with respect to t t t gives a sub-Gaussian bound on the Laplace transform of S R f − E P N , β S R f S_R^f-\mathbb{E}_{P_{N,\beta}}S_R^f S R f − E P N , β S R f (this is Herbst’s argument, see [60 Section 2.3] ). In particular, for every δ > 0 \delta>0 δ > 0 , we obtain exponential concentration: the probability that ∣ S R f − E P N , β S R f ∣ ≥ δ π R 2 \lvert S_R^f-\mathbb{E}_{P_{N,\beta}}S_R^f\rvert\geq\delta\pi R^2 ∣ S R f − E P N , β S R f ∣ ≥ δ π R 2 is at most 2 e − c R 2 2e^{-cR^2} 2 e − c R 2 , where c > 0 c>0 c > 0 depends only on β \beta β , f f f and δ \delta δ , but neither on N N N nor on R R R (Proposition 4.16 ).
Decay of correlations: the Combes–Thomas argument
We now present the strategy for proving the exponential decay of covariances, uniformly in the number of particles ((47) , in the proof of Proposition 5.1). The decay of the two-point function itself, part (i) of Theorem 1 , is then deduced from it in Subsection 5.7 by a fairly elementary approximation argument.
The Gaussian case
Let us start with the simplest possible situation. Consider a Gaussian field x = ( x k ) 1 ≤ k ≤ N x=(x_k)_{1\leq k\leq N} x = ( x k ) 1 ≤ k ≤ N indexed by the sites k k k of a one-dimensional lattice, with density proportional to e − 1 2 ⟨ x , A x ⟩ e^{-\frac{1}{2}\langle x,Ax\rangle} e − 2 1 ⟨ x , A x ⟩ , where A A A is a symmetric matrix, the interaction matrix . We assume that A ≥ a A\geq a A ≥ a for some a > 0 a>0 a > 0 , and that each site only interacts with nearby sites: A k j = 0 A_{kj}=0 A k j = 0 when ∣ k − j ∣ > r \lvert k-j\rvert>r ∣ k − j ∣ > r , that is, A A A is a band matrix. Decay of correlations concerns observables which depend on the field in two distant sets of sites I I I and J J J , for instance two linear statistics F = ⟨ f , x ⟩ F=\langle f,x\rangle F = ⟨ f , x ⟩ and G = ⟨ g , x ⟩ G=\langle g,x\rangle G = ⟨ g , x ⟩ , where f f f is supported in I I I and g g g in J J J . The Gaussian integration by parts E [ ( A x ) k G ] = E [ ∂ x k G ] \mathbb{E}[(Ax)_kG]=\mathbb{E}[\partial_{x_k}G] E [( A x ) k G ] = E [ ∂ x k G ] gives
Cov [ F , G ] = ⟨ A − 1 ∇ F , ∇ G ⟩ = ∑ k ∈ I , j ∈ J f k ( A − 1 ) k j g j . (20) \operatorname{Cov}[F,G]=\langle A^{-1}\nabla F,\nabla G\rangle=\sum_{k\in I,\;j\in J}f_k(A^{-1})_{kj}g_j. \tag*{(20)} Cov [ F , G ] = ⟨ A − 1 ∇ F , ∇ G ⟩ = k ∈ I , j ∈ J ∑ f k ( A − 1 ) k j g j . (20)
The decay of correlations between I I I and J J J thus amounts to the decay of the entries ( A − 1 ) k j (A^{-1})_{kj} ( A − 1 ) k j of the covariance matrix away from the diagonal, that is, as ∣ k − j ∣ → ∞ \lvert k-j\rvert\to\infty ∣ k − j ∣ → ∞ . The inverse of a band matrix is not a band matrix in general, but, when the band matrix is bounded below by a > 0 a>0 a > 0 , the entries of its inverse decay exponentially away from the diagonal.
Lemma 3.1 (Inverse of a band matrix). Let A A A be a symmetric N × N N\times N N × N matrix with A ≥ a > 0 A\geq a>0 A ≥ a > 0 and A k j = 0 A_{kj}=0 A k j = 0 when ∣ k − j ∣ > r \lvert k-j\rvert>r ∣ k − j ∣ > r , and let χ : = max k ∑ j ≠ k ∣ A k j ∣ \chi:=\max_k\sum_{j\ne k}\lvert A_{kj}\rvert χ := max k ∑ j = k ∣ A k j ∣ . For every ℓ ≥ r \ell\geq r ℓ ≥ r such that χ r 2 / ℓ 2 < a \chi r^2/\ell^2<a χ r 2 / ℓ 2 < a ,
∣ ( A − 1 ) k j ∣ ≤ e − ∣ k − j ∣ / ℓ a − χ r 2 / ℓ 2 for all k , j . \lvert(A^{-1})_{kj}\rvert\leq\frac{e^{-\lvert k-j\rvert/\ell}}{a-\chi r^2/\ell^2}\qquad\text{for all }k,j. ∣( A − 1 ) k j ∣ ≤ a − χ r 2 / ℓ 2 e − ∣ k − j ∣ / ℓ for all k , j .
The proof is due to Combes and Thomas [27 ] . We recall it briefly, as it is the model for what follows.
Proof. Fix j j j , and let M M M be the diagonal matrix with entries M i : = e ∣ i − j ∣ / ℓ M_i:=e^{\lvert i-j\rvert/\ell} M i := e ∣ i − j ∣ / ℓ , an exponential weight which grows with the distance to the site j j j . Since M e j = e j Me_j=e_j M e j = e j and M − 1 e k = e − ∣ k − j ∣ / ℓ e k M^{-1}e_k=e^{-\lvert k-j\rvert/\ell}e_k M − 1 e k = e − ∣ k − j ∣ / ℓ e k , we have
( A − 1 ) k j = ⟨ e k , M − 1 ( M A M − 1 ) − 1 M e j ⟩ = e − ∣ k − j ∣ / ℓ ⟨ e k , ( M A M − 1 ) − 1 e j ⟩ , (A^{-1})_{kj}=\langle e_k,M^{-1}(MAM^{-1})^{-1}Me_j\rangle=e^{-\lvert k-j\rvert/\ell}\langle e_k,(MAM^{-1})^{-1}e_j\rangle, ( A − 1 ) k j = ⟨ e k , M − 1 ( M A M − 1 ) − 1 M e j ⟩ = e − ∣ k − j ∣ / ℓ ⟨ e k , ( M A M − 1 ) − 1 e j ⟩ ,
and it remains to bound the inverse of the conjugated matrix, with a constant independent of N N N . Write M A M − 1 = A + P MAM^{-1}=A+P M A M − 1 = A + P , with P i k = ( M i / M k − 1 ) A i k P_{ik}=(M_i/M_k-1)A_{ik} P ik = ( M i / M k − 1 ) A ik . The entries of P P P are of order 1 / ℓ 1/\ell 1/ ℓ , but its symmetric part is of order 1 / ℓ 2 1/\ell^2 1/ ℓ 2 : for v ∈ R N v \in\mathbb{R}^N v ∈ R N ,
⟨ P v , v ⟩ = ∑ i < k A i k ( M i M k + M k M i − 2 ) v i v k ≥ − κ r 2 ℓ 2 ∣ v ∣ 2 . \langle Pv,v\rangle=\sum_{i<k} A_{ik}\left(\frac{M_i}{M_k}+\frac{M_k}{M_i}-2\right)v_i v_k\geq-\kappa\frac{r^2}{\ell^2}|v|^2. ⟨ P v , v ⟩ = i < k ∑ A ik ( M k M i + M i M k − 2 ) v i v k ≥ − κ ℓ 2 r 2 ∣ v ∣ 2 .
Indeed, A i k = 0 A_{ik}=0 A ik = 0 unless ∣ i − k ∣ ≤ r |i-k|\leq r ∣ i − k ∣ ≤ r , and then e − r / ℓ ≤ M i / M k ≤ e r / ℓ e^{-r/\ell}\leq M_i/M_k\leq e^{r/\ell} e − r / ℓ ≤ M i / M k ≤ e r / ℓ , hence the bracket is at most 2 r 2 / ℓ 2 2r^2/\ell^2 2 r 2 / ℓ 2 , and we use 2 ∣ v i v k ∣ ≤ v i 2 + v k 2 2|v_i v_k|\leq v_i^2+v_k^2 2∣ v i v k ∣ ≤ v i 2 + v k 2 . Hence ⟨ ( A + P ) v , v ⟩ ≥ ( a − κ r 2 / ℓ 2 ) ∣ v ∣ 2 \langle(A+P)v,v\rangle\geq(a-\kappa r^2/\ell^2)|v|^2 ⟨( A + P ) v , v ⟩ ≥ ( a − κ r 2 / ℓ 2 ) ∣ v ∣ 2 for every v v v : the conjugated matrix is still coercive, and the norm of its inverse is at most ( a − κ r 2 / ℓ 2 ) − 1 (a-\kappa r^2/\ell^2)^{-1} ( a − κ r 2 / ℓ 2 ) − 1 .
The three ingredients are the following. The positivity A ≥ a A\geq a A ≥ a gives the coercivity. The weight is equal to 1 where the first observable lives, and it is large where the second one lives. The band structure makes the cost of the conjugation of second order in 1 / ℓ 1/\ell 1/ ℓ .
Beyond the Gaussian case: the Helffer–Sjöstrand representation
Consider now a probability measure μ = e − Φ d x \mu=e^{-\Phi}\,dx μ = e − Φ d x on R N \mathbb{R}^N R N , where x k ∈ R x_k\in\mathbb{R} x k ∈ R is the position of the k k k -th particle, with a non-quadratic energy Φ \Phi Φ whose Hessian is positive: ∇ 2 Φ ≥ a > 0 \nabla^2\Phi\geq a>0 ∇ 2 Φ ≥ a > 0 . The Hessian depends on the configuration, and there is no single matrix to invert. Nevertheless, for smooth observables F , G F,G F , G ,
Cov μ [ F , G ] = ⟨ A 1 − 1 ∇ F , ∇ G ⟩ L 2 ( μ ; R N ) , A 1 : = L R ⊗ I N + ∇ 2 Φ , L R : = − Δ + ∇ Φ ⋅ ∇ . (21) \operatorname{Cov}_{\mu}[F,G]=\langle A_1^{-1}\nabla F,\nabla G\rangle_{L^2(\mu;\mathbb{R}^N)},\qquad A_1:=\mathcal{L}_{\mathbb{R}}\otimes I_N+\nabla^2\Phi,\qquad\mathcal{L}_{\mathbb{R}}:=-\Delta+\nabla\Phi\cdot\nabla. \tag*{(21)} Cov μ [ F , G ] = ⟨ A 1 − 1 ∇ F , ∇ G ⟩ L 2 ( μ ; R N ) , A 1 := L R ⊗ I N + ∇ 2 Φ , L R := − Δ + ∇Φ ⋅ ∇. (21)
This is the representation of Helffer and Sjöstrand [52 ] . Equivalently, u : = A 1 − 1 ∇ F = ∇ ϕ u:=A_1^{-1}\nabla F=\nabla\phi u := A 1 − 1 ∇ F = ∇ ϕ , where ϕ \phi ϕ solves L R ϕ = F − E μ F \mathcal{L}_{\mathbb{R}}\phi=F-\mathbb{E}_{\mu}F L R ϕ = F − E μ F . Its derivation takes a few lines, which we recall in Appendix A.§ 3 . The operator A 1 A_1 A 1 acts on vector fields indexed by the particles, as A A A acts on vectors indexed by the sites, and it is coercive: A 1 ≥ a A_1\geq a A 1 ≥ a , since L R ≥ 0 \mathcal{L}_{\mathbb{R}}\geq0 L R ≥ 0 . The vector field u = A 1 − 1 ∇ F u=A_1^{-1}\nabla F u = A 1 − 1 ∇ F , which replaces A − 1 ∇ F A^{-1}\nabla F A − 1 ∇ F , describes the response of the system to the observable F F F .
The dictionary with the Gaussian case is the following. The site k k k is replaced by the position x k x_k x k of the k k k -th particle: an observable is local to a region if it only depends on the particles located in this region. The weight becomes M k : = e ζ ( x k ) / ℓ M_k:=e^{\zeta(x_k)/\ell} M k := e ζ ( x k ) / ℓ , where ζ \zeta ζ is the distance to the region where F F F lives. The band condition becomes the locality of the off-diagonal entries of the Hessian: ∂ x i ∂ x j Φ ( x ) = 0 \partial_{x_i}\partial_{x_j}\Phi(x)=0 ∂ x i ∂ x j Φ ( x ) = 0 when ∣ x i − x j ∣ ≥ r |x_i-x_j|\geq r ∣ x i − x j ∣ ≥ r . Decay of correlations then amounts to the following statement: the response u k u_k u k is exponentially small when the k k k -th particle is far from the region where F F F lives. The argument of Combes and Thomas proves it as above, with one new feature. The weight does not commute with the differential operator L R \mathcal{L}_{\mathbb{R}} L R , and the conjugation also produces a first-order term, which is antisymmetric and does not contribute, and a zero-order term of size ℓ − 2 \ell^{-2} ℓ − 2 . We carry out this computation only once, for the gas, in Subsections 5.3 and 5.4.
From positive Hessians to the gas
Compared to this situation, the 2D OCP has two drawbacks. First, the Hessian of its energy is not positive. Second, the logarithmic interaction has infinite range, and its Hessian couples all pairs of particles. To solve both issues, we proceed as in Subsection 3.1: we replace the real derivatives by the Cauchy–Riemann derivatives ∂ ˉ k \bar{\partial}_k ∂ ˉ k , and the Hessian by the Levi matrix ( ∂ j ∂ ˉ k ) j , k (\partial_j\bar{\partial}_k)_{j,k} ( ∂ j ∂ ˉ k ) j , k , which we write as a real 2 N × 2 N 2N\times2N 2 N × 2 N matrix (see (27) ). We work in the shifted space L 2 ( e − Φ s h ) L^2(e^{-\Phi^{\mathrm{sh}}}) L 2 ( e − Φ sh ) , with observables multiplied by the Vandermonde polynomial V N V_N V N , which ∂ ˉ \bar{\partial} ∂ ˉ does not see. Table 1 summarizes the correspondence, for this subsection and for Subsection 3.1; the anti-analytic signal is explained below.
Table 1. From positive Hessians to the two-dimensional one-component plasma.
The Levi matrix of Φ s h \Phi^{\mathrm{sh}} Φ sh is bounded below by a = π β / 2 a=\pi\beta/2 a = π β /2 when β ≤ 2 \beta\leq2 β ≤ 2 (Lemma 4.4 ), and it plays the role of the Hessian: the operator A 1 A_1 A 1 becomes L ⊗ I + C Φ s h \mathcal{L}\otimes I+C_{\Phi^{\mathrm{sh}}} L ⊗ I + C Φ sh , where L = ∑ j ∂ ˉ j ∗ ∂ ˉ j \mathcal{L}=\sum_j\bar{\partial}_j^*\bar{\partial}_j L = ∑ j ∂ ˉ j ∗ ∂ ˉ j , and it is coercive (Subsection 5.2). The price to pay is a new issue: the equation L ϕ = V N ( F − E F ) \mathcal{L}\phi=V_N(F-\mathbb{E}F) L ϕ = V N ( F − E F ) , which replaces the equation L R ϕ = F − E μ F \mathcal{L}_{\mathbb{R}}\phi=F-\mathbb{E}_{\mu}F L R ϕ = F − E μ F defining u = ∇ ϕ u=\nabla\phi u = ∇ ϕ in the real case, need not have a solution.
Invertibility: the anti-analytic signal. In the real case, the equation L R ϕ = F − E μ F \mathcal{L}_{\mathbb{R}}\phi=F-\mathbb{E}_{\mu}F L R ϕ = F − E μ F has a solution because L R \mathcal{L}_{\mathbb{R}} L R is invertible on the orthogonal complement of its kernel, the constants, and F − E μ F F-\mathbb{E}_{\mu}F F − E μ F belongs to this complement. Now the kernel of L \mathcal{L} L is much larger: it consists of all the holomorphic functions of the shifted space, and for a real observable F F F , the function V N ( F − E F ) V_N(F-\mathbb{E}F) V N ( F − E F ) need not be orthogonal to them. As for the Poincaré inequality of Subsection 4.8, we use a Fourier-like decomposition into “modes”. Write F = ∑ m F m F=\sum_m F_m F = ∑ m F m , where F m ( e i θ Z ) = e i m θ F m ( Z ) F_m(e^{i\theta}Z)=e^{im\theta}F_m(Z) F m ( e i θ Z ) = e im θ F m ( Z ) is the mode of character m m m of F F F . Since the weight is invariant under rotations, functions with different characters are orthogonal. Since it is invariant under permutations, the antisymmetric function V N F V_NF V N F is orthogonal to the symmetric holomorphic functions, and the antisymmetric holomorphic functions are the products of V N V_N V N by symmetric ones, whose characters are at least the degree d N d_N d N of V N V_N V N , with equality only for the multiples of V N V_N V N . Hence V N F m V_NF_m V N F m , of character d N + m d_N+m d N + m , is orthogonal to all the holomorphic functions when m < 0 m<0 m < 0 , and also when m = 0 m=0 m = 0 if F 0 F_0 F 0 is centred. For a real F F F , F − m = F m ‾ F_{-m}=\overline{F_m} F − m = F m , so the anti-analytic signal H : = F 0 − E F + 2 ∑ m < 0 F m H:=F_0-\mathbb{E}F+2\sum_{m<0}F_m H := F 0 − E F + 2 ∑ m < 0 F m has the same real part as F − E F F-\mathbb{E}F F − E F , hence the same covariance with every real observable, and V N H V_NH V N H is orthogonal to the holomorphic functions (Lemma A.3 ). The function H H H is built from the rotations of F F F around the origin. Since these rotations preserve the disk B R B_R B R , H H H only depends on the particles in B R B_R B R when F F F does: this is why we centre the disk at the origin. This is no loss of generality: the limit points are stationary, so part (ii) of Theorem 1 follows from the case x = 0 x=0 x = 0 .
Finite range: the collisions. The Hessian of the logarithmic interaction couples all pairs of particles, with off-diagonal entries of order ∣ z i − z j ∣ − 2 |z_i-z_j|^{-2} ∣ z i − z j ∣ − 2 . The Levi matrix, on the other hand, is local. Since the logarithm is harmonic away from the origin, the off-diagonal entries of the Levi matrix vanish outside the collisions z i = z j z_i=z_j z i = z j : the long-range part of the interaction is invisible to it, and the band condition holds with r = 0 r=0 r = 0 . At a collision, M i = M j M_i=M_j M i = M j , so the perturbation created by the off-diagonal entries formally vanishes. To make this rigorous, we smooth the logarithm at a scale ε \varepsilon ε (Subsection 5.1). The off-diagonal entries of the Levi matrix are then of size ε − 2 \varepsilon^{-2} ε − 2 and supported in { ∣ z i − z j ∣ < ε } \{|z_i-z_j|<\varepsilon\} { ∣ z i − z j ∣ < ε } , where ( M i − M j ) 2 (M_i-M_j)^2 ( M i − M j ) 2 is of order ε 2 \varepsilon^2 ε 2 . What remains is a collision term , bounded by the mass of the solution near the collisions. The main technical difficulty is to show that the solution does not concentrate on the collisions as ε → 0 \varepsilon\to0 ε → 0 (Lemma 5.5 ), which makes this term disappear when the smoothing is removed.
The weight is now M k = e ζ ( z k ) / ℓ M_k=e^{\zeta(z_k)/\ell} M k = e ζ ( z k ) / ℓ , with ζ ( z ) = min { dist ( z , B R ) , d } \zeta(z)=\min\{\operatorname{dist}(z,B_R),d\} ζ ( z ) = min { dist ( z , B R ) , d } . Apart from the collision term, the only cost of the conjugation is the zero-order term of the differential part. Since ∣ ∂ ˉ k ζ ( z k ) ∣ 2 ≤ 1 / 4 |\bar{\partial}_k\zeta(z_k)|^2\le1/4 ∣ ∂ ˉ k ζ ( z k ) ∣ 2 ≤ 1/4 , the coercivity constant a = π β / 2 = 1 / ( 4 ℓ D 2 ) a=\pi\beta/2=1/(4\ell_D^2) a = π β /2 = 1/ ( 4 ℓ D 2 ) becomes a − 1 / ( 4 ℓ 2 ) a-1/(4\ell^2) a − 1/ ( 4 ℓ 2 ) , which is positive exactly when ℓ > ℓ D \ell>\ell_D ℓ > ℓ D . Covariances therefore decay like e − d / ℓ e^{-d/\ell} e − d / ℓ for every ℓ > ℓ D \ell>\ell_D ℓ > ℓ D , where d d d is the distance between the regions where the two observables live.
Uniqueness from concentration and “goodness”
The proof of Theorem 2 has two ingredients: the exponential concentration estimate of Subsection 3.1, and the “goodness” of good processes.
The role of “goodness”. It is well known that exponential concentration alone does not give uniqueness. The simplest counterexample is the centred massive Gaussian field Ξ \Xi Ξ on Z 2 \mathbb{Z}^2 Z 2 , with covariance ( − Δ + m 2 ) − 1 (-\Delta+m^2)^{-1} ( − Δ + m 2 ) − 1 , where Δ \Delta Δ is the lattice Laplacian and m > 0 m>0 m > 0 . Its energy in a finite box is uniformly convex, and it has exponential concentration at the scale of the volume. Let h h h be a solution of ( − Δ + m 2 ) h = 0 (-\Delta+m^2)h=0 ( − Δ + m 2 ) h = 0 , for instance an exponential in one coordinate. Then the law of Ξ + h \Xi+h Ξ + h solves the same DLR equations as the law of Ξ \Xi Ξ ([42 Theorem 13.24] ), with the same covariance and the same concentration around its mean. What distinguishes the law of Ξ \Xi Ξ is temperedness: a non-zero h h h is not polynomially bounded, and the law of Ξ \Xi Ξ is the only DLR state whose configurations are polynomially bounded. For the 2DOCP, the class of good processes plays a similar role: it bounds the number of points, the smooth linear statistics and the two-point function.
The variational principle. “Goodness” enters through the variational principle: we prove that every good DLR state is a free energy minimizer (Proposition 1.1 ). In a disk B R B_R B R , we compare a good DLR state P with a mixture of gases with no exterior field (a kind of “free boundary condition”), designed to have the same law of the number of points in B R B_R B R as P. A somewhat direct computation, based on the temperedness properties of good processes, shows that the relative entropy of the restriction of P to B R B_R B R with respect to this mixture is at most of order R 3 / 2 R^{3/2} R 3/2 (Proposition 6.5 ), which is small compared with the volume.
Using exponential concentration. Concentration then shows uniqueness of free energy minimizers. We consider the gas of n n n particles in the disk of area n n n . Its empirical field is the law of the configuration seen from a uniform point of this disk. By the large deviation principle of [68 ] , this empirical field visits a neighbourhood of every free energy minimizer at a cost that is subexponential in n n n . A continuous function of the empirical field that concentrates at the exponential scale must therefore take the same value on all the free energy minimizers (Lemma 7.1 ). The observable S R n S_R^n S R n of Subsection 3.1, with π R 2 = n \pi R^2=n π R 2 = n , is of this form: it is the expectation of e − X ( f ) e^{-X(f)} e − X ( f ) under the rotation average of the empirical field, that is, under its law after a rotation by a uniform angle. Since these expectations characterize a point process, all the free energy minimizers have the same rotation average (Proposition 7.2 ). This shows uniqueness among rotation-invariant processes.
Excluding mixtures. It remains to exclude non-trivial mixtures, such as a crystal with a uniformly random orientation. Let P ∞ P_\infty P ∞ be a limit point of the gas, which is stationary and rotation-invariant. The Poincaré inequality for general symmetric observables passes to the limit and makes the spatial averages asymptotically deterministic under P ∞ P_\infty P ∞ , so P ∞ P_\infty P ∞ is extremal among stationary laws (Lemma 7.3 ). If P ∞ P_\infty P ∞ is a free energy minimizer, then it is the rotation average of every free energy minimizer. Since the free energy is defined on stationary processes, every free energy minimizer is stationary, and the extremality of P ∞ P_\infty P ∞ forces it to equal P ∞ P_\infty P ∞ .
This argument uses that the limit points are good, through the condition (G3), which follows from the decay of correlations of Subsection 3.2 and does not use uniqueness (Corollary 5.11).
Debye–Hückel asymptotics
As β ↓ 0 \beta\downarrow0 β ↓ 0 , the Debye length ℓ D \ell_{\mathrm D} ℓ D tends to infinity, and Theorem 5 describes the truncated two-point function at this scale. We therefore measure distances in units of ℓ D \ell_{\mathrm D} ℓ D . Let Y Y Y be the image of the configuration X X X under x ↦ x / ℓ D x \mapsto x/\ell_{\mathrm D} x ↦ x / ℓ D , so that Y ( f ) = X ( f ( ⋅ / ℓ D ) ) Y(f)=X(f(\cdot/\ell_{\mathrm D})) Y ( f ) = X ( f ( ⋅ / ℓ D )) . Under Gin β \operatorname{Gin}_{\beta} Gin β , the configuration Y Y Y has intensity ℓ D 2 \ell_{\mathrm D}^{2} ℓ D 2 . We first identify the averages of the truncated two-point function, through the covariances of linear statistics, and then its value at each distance.
The structure factor on the Debye scale
For smooth functions f , h f,h f , h with compact support, let C β ( f , h ) : = ℓ D − 2 Cov Gin β [ Y ( f ) , Y ( h ) ] \mathcal{C}_{\beta}(f,h):=\ell_{\mathrm D}^{-2}\operatorname{Cov}_{\operatorname{Gin}_{\beta}}[Y(f),Y(h)] C β ( f , h ) := ℓ D − 2 Cov Gin β [ Y ( f ) , Y ( h )] . The constant of the Poincaré inequality for rotation-invariant observables of Subsection 3.1 is 1 / ( 2 π β ) = ℓ D 2 1/(2\pi\beta)=\ell_{\mathrm D}^{2} 1/ ( 2 π β ) = ℓ D 2 . For radial f f f and h h h , it gives ∣ C β ( f , h ) ∣ ≤ ∥ ∇ f ∥ 2 ∥ ∇ h ∥ 2 |\mathcal{C}_{\beta}(f,h)|\leq\|\nabla f\|_2\|\nabla h\|_2 ∣ C β ( f , h ) ∣ ≤ ∥∇ f ∥ 2 ∥∇ h ∥ 2 , which explains the normalization: the covariances divided by ℓ D 2 \ell_{\mathrm D}^{2} ℓ D 2 remain bounded as β ↓ 0 \beta\downarrow0 β ↓ 0 . This is only an upper bound. To identify the limit, we need an identity.
An exact identity. We move the points of the gas along the vector field ∇ f \nabla f ∇ f and integrate by parts in the Gibbs measure. This is the Ward identity of the gas, also called loop equation, see [2 ] for random normal matrices, and [8 ] for related uses in two-dimensional Coulomb gases. In the pair interaction, we replace one point of each pair by the mean density ℓ D 2 \ell_{\mathrm D}^{2} ℓ D 2 of the gas (a mean-field approximation), and call Q f Q_f Q f the error, a centered quadratic statistic of the pairs of points of Y Y Y . We obtain (see Subsection 8.2)
C β ( ( 1 − Δ ) f , h ) = ∫ R 2 ∇ f ⋅ ∇ h − β Cov Gin β [ Q f , Y ( h ) ] . (22) \mathcal{C}_{\beta}\bigl((1-\Delta)f,h\bigr)=\int_{\mathbb{R}^{2}}\nabla f\cdot\nabla h-\beta\operatorname{Cov}_{\operatorname{Gin}_{\beta}}\left[Q_f,Y(h)\right]. \tag*{(22)} C β ( ( 1 − Δ ) f , h ) = ∫ R 2 ∇ f ⋅ ∇ h − β Cov Gin β [ Q f , Y ( h ) ] . (22)
The 1 in 1 − Δ 1-\Delta 1 − Δ is the only contribution of the Coulomb interaction at this order: it comes from the potential created by the displacement of the mean density. Without it, we would get C β ( − Δ f , h ) = ∫ ∇ f ⋅ ∇ h C_\beta(-\Delta f,h)=\int\nabla f\cdot\nabla h C β ( − Δ f , h ) = ∫ ∇ f ⋅ ∇ h , as for a Poisson process. Neglecting Q f Q_f Q f is precisely the Debye–Hückel approximation.
The Debye–Hückel limit. The Poincaré inequality and bounds on the number of points in small disks control Q f Q_f Q f , and the last term of (22) tends to zero as β ↓ 0 \beta\downarrow0 β ↓ 0 . To read off the two-point function, we write the left-hand side with the structure factor. By Theorem 1 , ρ 2 , β ( 0 , ⋅ ) − 1 \rho_{2,\beta}(0,\cdot)-1 ρ 2 , β ( 0 , ⋅ ) − 1 is integrable, hence σ G i n β \sigma_{\mathrm{Gin}_\beta} σ Gin β has a continuous density, which we still denote by σ G i n β \sigma_{\mathrm{Gin}_\beta} σ Gin β . By (7) , after the change of variables k ↦ k / ℓ D k\mapsto k/\ell_D k ↦ k / ℓ D ,
C β ( f , h ) = 1 ( 2 π ) 2 ∫ R 2 σ G i n β ( k / ℓ D ) f ^ ( k ) h ^ ( k ) ‾ d k . C_\beta(f,h)=\frac{1}{(2\pi)^2}\int_{\mathbb{R}^2}\sigma_{\mathrm{Gin}_\beta}(k/\ell_D)\widehat{f}(k)\overline{\widehat{h}(k)}\,dk. C β ( f , h ) = ( 2 π ) 2 1 ∫ R 2 σ Gin β ( k / ℓ D ) f ( k ) h ( k ) d k .
The Fourier transform of ( 1 − Δ ) f (1-\Delta)f ( 1 − Δ ) f is ( 1 + ∣ k ∣ 2 ) f ^ (1+|k|^2)\widehat{f} ( 1 + ∣ k ∣ 2 ) f , and ∫ ∇ f ⋅ ∇ h = ( 2 π ) − 2 ∫ ∣ k ∣ 2 f ^ h ^ ‾ d k \int\nabla f\cdot\nabla h=(2\pi)^{-2}\int|k|^2\widehat{f}\overline{\widehat{h}}\,dk ∫ ∇ f ⋅ ∇ h = ( 2 π ) − 2 ∫ ∣ k ∣ 2 f h d k by Plancherel’s formula. Hence (22) gives
σ G i n β ( k / ℓ D ) → β ↓ 0 ∣ k ∣ 2 1 + ∣ k ∣ 2 . \sigma_{\mathrm{Gin}_\beta}(k/\ell_D)\xrightarrow[\beta\downarrow0]{}\frac{|k|^2}{1+|k|^2}. σ Gin β ( k / ℓ D ) β ↓ 0 1 + ∣ k ∣ 2 ∣ k ∣ 2 .
Without the Coulomb term, we would get 1 1 1 , the structure factor of a Poisson process: the 1 in 1 + ∣ k ∣ 2 1+|k|^2 1 + ∣ k ∣ 2 is the screening. It remains to translate this into a statement on the two-point function. Since the structure factor is the Fourier transform of δ 0 + ( ρ 2 , β ( 0 , u ) − 1 ) d u \delta_0+(\rho_{2,\beta}(0,u)-1)\,du δ 0 + ( ρ 2 , β ( 0 , u ) − 1 ) d u , and since 2 π β ℓ D 2 = 1 2\pi\beta\ell_D^2=1 2 π β ℓ D 2 = 1 ,
σ G i n β ( k / ℓ D ) = 1 + 1 2 π ∫ R 2 e − i k ⋅ u ρ 2 , β ( 0 , ℓ D u ) − 1 β d u . \sigma_{\mathrm{Gin}_\beta}(k/\ell_D)=1+\frac{1}{2\pi}\int_{\mathbb{R}^2}e^{-ik\cdot u}\frac{\rho_{2,\beta}(0,\ell_Du)-1}{\beta}\,du. σ Gin β ( k / ℓ D ) = 1 + 2 π 1 ∫ R 2 e − ik ⋅ u β ρ 2 , β ( 0 , ℓ D u ) − 1 d u .
The Fourier transform of K 0 ( ∣ ⋅ ∣ ) K_0(|\cdot|) K 0 ( ∣ ⋅ ∣ ) is 2 π / ( 1 + ∣ k ∣ 2 ) 2\pi/(1+|k|^2) 2 π / ( 1 + ∣ k ∣ 2 ) , so the limit ∣ k ∣ 2 / ( 1 + ∣ k ∣ 2 ) = 1 − ( 2 π ) − 1 ⋅ 2 π / ( 1 + ∣ k ∣ 2 ) |k|^2/(1+|k|^2)=1-(2\pi)^{-1}\cdot2\pi/(1+|k|^2) ∣ k ∣ 2 / ( 1 + ∣ k ∣ 2 ) = 1 − ( 2 π ) − 1 ⋅ 2 π / ( 1 + ∣ k ∣ 2 ) is the structure factor of a process whose rescaled truncated two-point function is − K 0 -K_0 − K 0 . In other words, ( ρ 2 , β ( 0 , ℓ D ⋅ ) − 1 ) / β (\rho_{2,\beta}(0,\ell_D\cdot)-1)/\beta ( ρ 2 , β ( 0 , ℓ D ⋅ ) − 1 ) / β converges to − K 0 ( ∣ ⋅ ∣ ) -K_0(|\cdot|) − K 0 ( ∣ ⋅ ∣ ) in S ′ ( R 2 ) \mathcal{S}'(\mathbb{R}^2) S ′ ( R 2 ) , which is the first half of Theorem 5 (Proposition 8.2).
Values at fixed distances
The convergence in S ′ ( R 2 ) \mathcal{S}'(\mathbb{R}^2) S ′ ( R 2 ) only concerns averages of the two-point function. To obtain its value at each distance, we compute the effective interaction between two particles of the gas. For r > 0 r>0 r > 0 , let q β ( r ) : = ρ 2 , β ( 0 , ℓ D r e 1 ) q_\beta(r):=\rho_{2,\beta}(0,\ell_Dre_1) q β ( r ) := ρ 2 , β ( 0 , ℓ D r e 1 ) , where e 1 = ( 1 , 0 ) e_1=(1,0) e 1 = ( 1 , 0 ) , and write q β ( r ) = e − β w β ( r ) q_\beta(r)=e^{-\beta w_\beta(r)} q β ( r ) = e − β w β ( r ) : in the language of physics, w β w_\beta w β is the potential of mean force . If the two particles only interacted with each other, we would have w β ( r ) = g ( r e 1 ) = − log r w_\beta(r)=g(re_1)=-\log r w β ( r ) = g ( r e 1 ) = − log r , up to a constant. Theorem 5 says instead that w β ( r ) → β ↓ 0 K 0 ( r ) w_\beta(r)\xrightarrow[\beta\downarrow0]{}K_0(r) w β ( r ) β ↓ 0 K 0 ( r ) : at distances of order ℓ D \ell_D ℓ D , the gas screens the interaction of the two particles, and replaces the logarithm by K 0 K_0 K 0 .
The heuristics: test particles and linear response. The argument follows the classical derivation of the Debye–Hückel theory by test particles and linear response, see for example [49 , 62 ] .
(i) Insertion of two test particles. We fix two particles of the finite gas, at 0 and at r e 1 re_1 r e 1 in the coordinates of Y Y Y , and integrate over the positions of the others. This gives an exact formula (Lemma 8.6 ): up to a constant which does not depend on r r r , the two-point function at ( 0 , r e 1 ) (0,re_1) ( 0 , r e 1 ) equals
e − r 2 / 4 ⏟ confinement × r β ⏟ interaction of the two particles × E [ e − β Y ( g + g ( ⋅ − r e 1 ) ) ] ⏟ interaction with the other particles , \underbrace{e^{-r^2/4}}_{\text{confinement}}\times\underbrace{r^\beta}_{\text{interaction of the two particles}}\times\underbrace{\mathbb{E}\left[e^{-\beta Y(g+g(\cdot-re_1))}\right]}_{\text{interaction with the other particles}}, confinement e − r 2 /4 × interaction of the two particles r β × interaction with the other particles E [ e − β Y ( g + g ( ⋅− r e 1 )) ] ,
where the expectation is taken under the gas of the other particles. This is the test-particle method of Widom [91 ] .
(ii) Linear response. At high temperature, the gas responds linearly to the two test particles, and we expand the logarithm of the expectation to second order in β \beta β : it is − β -\beta − β times a mean, plus β 2 2 \frac{\beta^2}{2} 2 β 2 times a variance. At first order, the gas acts through its mean density, which compensates the confinement: this mean is neutrality. At second order, the only term which depends on r r r is the covariance β 2 Cov [ Y ( g ) , Y ( g ( ⋅ − r e 1 ) ) ] \beta^2\operatorname{Cov}[Y(g),Y(g(\cdot-re_1))] β 2 Cov [ Y ( g ) , Y ( g ( ⋅ − r e 1 ))] . It has a simple meaning: the particle at 0 polarizes the gas, the response of the gas to an external potential being − β -\beta − β times a covariance (fluctuation–dissipation), and the particle at r e 1 r e_1 r e 1 feels this polarization. Formally, we obtain
w β ( r ) ≈ g ( r e 1 ) ⏟ bare − 1 2 π C β ( g , g ( ⋅ − r e 1 ) ) ⏟ induced by the gas . w_\beta(r) \approx\underbrace{g(r e_1)}_{\text{bare}}-\frac{1}{2\pi}\underbrace{\mathcal{C}_\beta\bigl(g,g(\cdot-r e_1)\bigr)}_{\text{induced by the gas}}. w β ( r ) ≈ bare g ( r e 1 ) − 2 π 1 induced by the gas C β ( g , g ( ⋅ − r e 1 ) ) .
(iii) The screening factor. By the spectral representation of C β \mathcal{C}_\beta C β in Subsection 3.4.1, this reads, in Fourier variables,
w ^ β ( k ) ≈ g ^ ( k ) ( 1 − σ G i n β ( k / ℓ D ) ∣ k ∣ 2 ) , g ^ ( k ) = 2 π ∣ k ∣ 2 . \widehat{w}_\beta(k) \approx\widehat{g}(k)\left(1-\frac{\sigma_{\mathrm{Gin}_\beta}(k/\ell_D)}{|k|^2}\right),\qquad\widehat{g}(k)=\frac{2\pi}{|k|^2}. w β ( k ) ≈ g ( k ) ( 1 − ∣ k ∣ 2 σ Gin β ( k / ℓ D ) ) , g ( k ) = ∣ k ∣ 2 2 π .
The effective interaction is the bare one, multiplied by a screening factor which only depends on the structure factor. By the Poincaré inequality, σ G i n β ( k / ℓ D ) ≤ ∣ k ∣ 2 \sigma_{\mathrm{Gin}_\beta}(k/\ell_D)\leq|k|^2 σ Gin β ( k / ℓ D ) ≤ ∣ k ∣ 2 , so this factor lies between 0 and 1: the gas can only screen. With the limit σ G i n β ( k / ℓ D ) → β → 0 ∣ k ∣ 2 / ( 1 + ∣ k ∣ 2 ) \sigma_{\mathrm{Gin}_\beta}(k/\ell_D)\xrightarrow[\beta\to0]{} |k|^2/(1+|k|^2) σ Gin β ( k / ℓ D ) β → 0 ∣ k ∣ 2 / ( 1 + ∣ k ∣ 2 ) of Subsection 3.4.1, we get
w ^ β ( k ) → β → 0 2 π 1 + ∣ k ∣ 2 , \widehat{w}_\beta(k)\xrightarrow[\beta\to0]{} \frac{2\pi}{1+|k|^2}, w β ( k ) β → 0 1 + ∣ k ∣ 2 2 π ,
which is the Fourier transform of K 0 ( ∣ ⋅ ∣ ) K_0(|\cdot|) K 0 ( ∣ ⋅ ∣ ) .
Making the heuristics rigorous. The two test particles carry a total charge 2, to which the gas also responds globally and non-linearly. We neutralize this charge by a fixed charge spread over the unit disk, and remove the latter at the end by an interpolation (see the beginning of § 11 ). Let D r : = g + g ( ⋅ − r e 1 ) − 2 g 1 D_r:=g+g(\cdot-r e_1)-2g_1 D r := g + g ( ⋅ − r e 1 ) − 2 g 1 be the potential of the two particles, neutralized in this way, where g 1 g_1 g 1 is the potential of a unit charge spread over the unit disk (Subsection 8.1). The outcome is the single covariance of the heuristics (Proposition 8.11 ):
1 β log q β ( r ) q β ( s ) = log r s + 1 2 π C β ( g , D r − D s ) + o ( 1 ) . \frac{1}{\beta}\log\frac{q_\beta(r)}{q_\beta(s)}=\log\frac{r}{s}+\frac{1}{2\pi}\mathcal{C}_\beta(g,D_r-D_s)+o(1). β 1 log q β ( s ) q β ( r ) = log s r + 2 π 1 C β ( g , D r − D s ) + o ( 1 ) .
By the screening factor above, the covariance converges to log ( s / r ) − K 0 ( r ) + K 0 ( s ) \log(s/r)-K_0(r)+K_0(s) log ( s / r ) − K 0 ( r ) + K 0 ( s ) (Lemma 8.12 ). The gas cancels the bare interaction log ( r / s ) \log(r/s) log ( r / s ) of the two particles, and replaces it by the screened interaction K 0 K_0 K 0 :
1 β log q β ( r ) q β ( s ) → β → 0 − K 0 ( r ) + K 0 ( s ) , \frac{1}{\beta}\log\frac{q_\beta(r)}{q_\beta(s)}\xrightarrow[\beta\to0]{}-K_0(r)+K_0(s), β 1 log q β ( s ) q β ( r ) β → 0 − K 0 ( r ) + K 0 ( s ) ,
uniformly for r r r and s s s in every compact interval of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) .
Fixing the constant. The ratios determine q β q_\beta q β up to a multiplicative constant, and the convergence in S ′ ( R 2 ) S'(\mathbb{R}^2) S ′ ( R 2 ) of Subsection 3.4.1 fixes this constant: q β ( r ) = 1 − β K 0 ( r ) + o ( β ) q_\beta(r)=1-\beta K_0(r)+o(\beta) q β ( r ) = 1 − β K 0 ( r ) + o ( β ) .
Concentration estimates
The goal of this section is to prove two concentration results for a wide family of observables of the finite-volume system: a Poincaré inequality , then a stronger exponential concentration estimate , both with constants independent of the number of particles.
The mechanism was explained in Subsection 3.1: we factor the squared modulus of the Vandermonde polynomial out of the Gibbs weight (Subsection 4.1), the remaining shifted weight has a positive Levi matrix when β ≤ 2 \beta\leq2 β ≤ 2 (Subsection 4.3), and the L 2 L^2 L 2 estimate of Hörmander gives a fake Poincaré inequality in the shifted space (Subsection 4.4). We then compute the projection on the holomorphic functions (Subsections 4.5, 4.6 and 4.8). This last part, which corresponds to part (ii) of Lemma 4.2 and to Lemma 4.11, is where the logarithmic interaction is truly special.
We do not reprove the estimate of Hörmander: see [Hör65], [Hör90, Chapter IV], or [10 ] for a textbook exposition, and Appendix A for the integration by parts behind it, which is the same as for the Brascamp–Lieb inequality [17 , 23 ] .
Finite-volume measures and the Vandermonde factorization
Generalized Gibbs measures. Fix 0 < β ≤ 2 0 < \beta\le2 0 < β ≤ 2 and N ≥ 1 N \ge1 N ≥ 1 . The two additional parameters a > 0 a > 0 a > 0 and L > 0 L > 0 L > 0 will play a technical role. Define G N , L a G^{a}_{N,L} G N , L a as the probability measure on ( R 2 ) N (\mathbb{R}^{2})^N ( R 2 ) N given, for Z = ( z 1 , … , z N ) ∈ ( R 2 ) N Z=(z_1,\ldots,z_N)\in(\mathbb{R}^{2})^N Z = ( z 1 , … , z N ) ∈ ( R 2 ) N , by
d G N , L a ( Z ) : = 1 K N , L a ∏ i < j ∣ z i − z j ∣ β e − a ∑ i ∣ z i ∣ 2 ∏ i 1 { ∣ z i ∣ < L } d z 1 … d z N . (23) \mathrm{d}G^{a}_{N,L}(Z) := \frac{1}{K^{a}_{N,L}}\prod_{i<j}|z_i-z_j|^\beta e^{-a\sum_i|z_i|^2}\prod_i\mathbf{1}_{\{|z_i|<L\}}\mathrm{d}z_1\ldots\mathrm{d}z_N . \tag*{(23)} d G N , L a ( Z ) := K N , L a 1 i < j ∏ ∣ z i − z j ∣ β e − a ∑ i ∣ z i ∣ 2 i ∏ 1 { ∣ z i ∣ < L } d z 1 … d z N . (23)
The law G N , L a G^{a}_{N,L} G N , L a is symmetric, and supported on B L N : = { Z ∈ ( R 2 ) N : ∣ z i ∣ < L for all i } B^N_L:=\{Z\in(\mathbb{R}^{2})^N:|z_i|<L\text{ for all }i\} B L N := { Z ∈ ( R 2 ) N : ∣ z i ∣ < L for all i } . Equivalently,
d G N , L a ( Z ) = 1 K N , L a e − Φ p h ( Z ) ∏ i 1 { ∣ z i ∣ < L } d Z , Φ p h ( Z ) : = a ∣ Z ∣ 2 + β ∑ i < j g ( z i − z j ) , \mathrm{d}G^{a}_{N,L}(Z)=\frac{1}{K^{a}_{N,L}}e^{-\Phi^{\mathrm{ph}}(Z)}\prod_i\mathbf{1}_{\{|z_i|<L\}}\mathrm{d}Z,\qquad\Phi^{\mathrm{ph}}(Z):=a|Z|^2+\beta\sum_{i<j}g(z_i-z_j), d G N , L a ( Z ) = K N , L a 1 e − Φ ph ( Z ) i ∏ 1 { ∣ z i ∣ < L } d Z , Φ ph ( Z ) := a ∣ Z ∣ 2 + β i < j ∑ g ( z i − z j ) ,
and we call Φ p h \Phi^{\mathrm{ph}} Φ ph the physical weight . For a = π β / 2 a=\pi\beta/2 a = π β /2 , this weight is simply Φ p h = β H N \Phi^{\mathrm{ph}}=\beta H_N Φ ph = β H N , with H N H_N H N as in (4) , and we write G N , L = G N , L π β / 2 G_{N,L}=G^{\pi\beta/2}_{N,L} G N , L = G N , L π β /2 . The finite-volume 2DOCP corresponds to taking a = π β / 2 a=\pi\beta/2 a = π β /2 and L = R N L=R_N L = R N .
The Vandermonde polynomial and the shifted weight. As in Subsection 3.1, a configuration Z = ( z 1 , … , z N ) Z=(z_1,\ldots,z_N) Z = ( z 1 , … , z N ) is a point of C N \mathbb{C}^N C N , and a function from ( R 2 ) N (\mathbb{R}^{2})^N ( R 2 ) N to R 2 \mathbb{R}^2 R 2 is also a function from C N \mathbb{C}^N C N to C \mathbb{C} C . The Vandermonde polynomial V N ( Z ) = ∏ i < j ( z i − z j ) V_N(Z)=\prod_{i<j}(z_i-z_j) V N ( Z ) = ∏ i < j ( z i − z j ) of Subsection 3.1 is antisymmetric, and we have
∣ V N ( Z ) ∣ 2 = e − 2 ∑ i < j g ( z i − z j ) . |V_N(Z)|^2=e^{-2\sum_{i<j}g(z_i-z_j)}. ∣ V N ( Z ) ∣ 2 = e − 2 ∑ i < j g ( z i − z j ) .
Splitting β = 2 + ( β − 2 ) \beta=2+(\beta-2) β = 2 + ( β − 2 ) in the interaction, the physical weight becomes
Φ p h = Φ s h − log ∣ V N ∣ 2 , Φ s h ( Z ) : = a ∣ Z ∣ 2 + ( β − 2 ) ∑ i < j g ( z i − z j ) , e − Φ p h = ∣ V N ∣ 2 e − Φ s h . (24) \Phi^{\mathrm{ph}}=\Phi^{\mathrm{sh}}-\log|V_N|^2,\qquad\Phi^{\mathrm{sh}}(Z):=a|Z|^2+(\beta-2)\sum_{i<j}g(z_i-z_j),\qquad e^{-\Phi^{\mathrm{ph}}}=|V_N|^2e^{-\Phi^{\mathrm{sh}}}. \tag*{(24)} Φ ph = Φ sh − log ∣ V N ∣ 2 , Φ sh ( Z ) := a ∣ Z ∣ 2 + ( β − 2 ) i < j ∑ g ( z i − z j ) , e − Φ ph = ∣ V N ∣ 2 e − Φ sh . (24)
We call Φ s h \Phi^{\mathrm{sh}} Φ sh the shifted weight . The same factorization appears, for β = 2 \beta=2 β = 2 only, in [Cha26, Suz26].
Two L 2 L^2 L 2 spaces. From now on, a a a , L L L , N N N and β \beta β are fixed, and we write G : = G N , L a G:=G^{a}_{N,L} G := G N , L a and K : = K N , L a K:=K^{a}_{N,L} K := K N , L a for the partition function. Functions of N N N positions are complex-valued and defined on the open polydisk B L N B^N_L B L N . We use two L 2 L^2 L 2 spaces:
the physical space L p h 2 : = L 2 ( B L N , e − Φ p h d Z ) L^2_{\mathrm{ph}}:=L^2(B^N_L,e^{-\Phi^{\mathrm{ph}}}\mathrm{d}Z) L ph 2 := L 2 ( B L N , e − Φ ph d Z ) , associated with the physical weight Φ p h \Phi^{\mathrm{ph}} Φ ph ;
the shifted space L s h 2 : = L 2 ( B L N , e − Φ s h d Z ) L^2_{\mathrm{sh}}:=L^2(B^N_L,e^{-\Phi^{\mathrm{sh}}}\mathrm{d}Z) L sh 2 := L 2 ( B L N , e − Φ sh d Z ) , associated with the shifted weight Φ s h \Phi^{\mathrm{sh}} Φ sh .
As in Subsection 3.1, ∥ ⋅ ∥ s h \|\cdot\|_{\mathrm{sh}} ∥ ⋅ ∥ sh is the norm of L s h 2 L^2_{\mathrm{sh}} L sh 2 . More generally, for a measurable function Φ : B L N → [ − ∞ , + ∞ ] \Phi:B^N_L\to[-\infty,+\infty] Φ : B L N → [ − ∞ , + ∞ ] , we let L Φ 2 : = L 2 ( B L N , e − Φ d Z ) L^2_\Phi:=L^2(B^N_L,e^{-\Phi}\mathrm{d}Z) L Φ 2 := L 2 ( B L N , e − Φ d Z ) , with the scalar product ⟨ f , g ⟩ Φ : = ∫ f g ‾ e − Φ d Z \langle f,g\rangle_\Phi:=\int f\overline{g}e^{-\Phi}\mathrm{d}Z ⟨ f , g ⟩ Φ := ∫ f g e − Φ d Z and the norm ∥ ⋅ ∥ Φ \|\cdot\|_\Phi ∥ ⋅ ∥ Φ . Neither measure is normalized. This plays no role, since all the inequalities we prove are invariant under multiplication of the measure by a constant. The normalization only appears when we go back to G G G , through ∥ G ∥ Φ p h 2 = K E G ∣ G ∣ 2 \|G\|_{\Phi^{\mathrm{ph}}}^2=K\mathbb{E}_G|G|^2 ∥ G ∥ Φ ph 2 = K E G ∣ G ∣ 2 . The shifted measure e − Φ s h d Z e^{-\Phi^{\mathrm{sh}}}\mathrm{d}Z e − Φ sh d Z need not be locally finite: near a collision of k k k points, it behaves like ∏ ∣ z i − z j ∣ β − 2 \prod|z_i-z_j|^{\beta-2} ∏ ∣ z i − z j ∣ β − 2 , which is not integrable when k ≥ 4 / ( 2 − β ) k\ge4/(2-\beta) k ≥ 4/ ( 2 − β ) . In particular, the constants need not belong to L s h 2 L^2_{\mathrm{sh}} L sh 2 . This plays no role. The estimate of Hörmander holds for arbitrary plurisubharmonic weights, and we only apply it to functions of the form V N G V_NG V N G , which belong to L s h 2 L^2_{\mathrm{sh}} L sh 2 by Lemma 4.2 .
The Cauchy–Riemann operator
Let F F F be a complex-valued function of N N N positions in R 2 \mathbb{R}^2 R 2 . We use the derivatives ∂ j \partial_j ∂ j and ∂ ‾ j \overline{\partial}_j ∂ j of Subsection 3.1, taken in the sense of distributions, and we write ∂ ‾ F = ( ∂ ‾ j F ) 1 ≤ j ≤ N \overline{\partial}F=(\overline{\partial}_jF)_{1\le j\le N} ∂ F = ( ∂ j F ) 1 ≤ j ≤ N . The operator ∂ ‾ j \overline{\partial}_j ∂ j is the Cauchy–Riemann operator in the variable z j z_j z j , and 4 ∂ j ∂ ‾ j 4\partial_j\overline{\partial}_j 4 ∂ j ∂ j is the Laplacian in z j z_j z j . Each ∂ j F \partial_jF ∂ j F is again complex-valued and ∂ ‾ F \overline{\partial}F ∂ F is a family of N N N such functions. If F F F is real-valued, ∂ ‾ j F \overline{\partial}_jF ∂ j F is the vector 1 2 ∇ j F \frac{1}{2}\nabla_jF 2 1 ∇ j F , hence
∣ ∂ ‾ j F ∣ 2 = 1 4 ∣ ∇ j F ∣ 2 . (25) |\overline{\partial}_jF|^2=\frac{1}{4}|\nabla_jF|^2 . \tag*{(25)} ∣ ∂ j F ∣ 2 = 4 1 ∣ ∇ j F ∣ 2 . (25)
Kernel of ∂ ˉ \bar{\partial} ∂ ˉ . The equation ∂ ˉ h = 0 \bar{\partial}h=0 ∂ ˉ h = 0 does not involve the weight: its solutions are the holomorphic functions on B L N B_L^N B L N (Lemma 4.1 below). In the shifted space, we let
ker ∂ ˉ : = { h ∈ L s h 2 : ∂ ˉ h = 0 } , \ker\bar{\partial}:=\{h\in L_{\mathrm{sh}}^2:\bar{\partial}h=0\}, ker ∂ ˉ := { h ∈ L sh 2 : ∂ ˉ h = 0 } ,
a closed subspace of L s h 2 L_{\mathrm{sh}}^2 L sh 2 , and we denote by Π ker ∂ ˉ \Pi_{\ker\bar{\partial}} Π k e r ∂ ˉ the orthogonal projection of L s h 2 L_{\mathrm{sh}}^2 L sh 2 onto ker ∂ ˉ \ker\bar{\partial} ker ∂ ˉ . For another weight Φ \Phi Φ , we define ker ∂ ˉ \ker\bar{\partial} ker ∂ ˉ and Π ker ∂ ˉ \Pi_{\ker\bar{\partial}} Π k e r ∂ ˉ in the same way in L Φ 2 L_\Phi^2 L Φ 2 .
Lemma 4.1. Let Φ : B L N → [ − ∞ , + ∞ ] \Phi:B_L^N\to[-\infty,+\infty] Φ : B L N → [ − ∞ , + ∞ ] be measurable, with e − Φ e^{-\Phi} e − Φ bounded below by a positive constant on each compact subset of B L N B_L^N B L N (this holds for Φ s h \Phi^{\mathrm{sh}} Φ sh , since β ≤ 2 \beta\le2 β ≤ 2 ). Every element of ker ∂ ˉ ⊂ L Φ 2 \ker\bar{\partial}\subset L_\Phi^2 ker ∂ ˉ ⊂ L Φ 2 has a representative which is a holomorphic function on B L N B_L^N B L N .
Proof. The density e − Φ e^{-\Phi} e − Φ is bounded below on each compact set, so the elements of ker ∂ ˉ \ker\bar{\partial} ker ∂ ˉ are locally in L 2 L^2 L 2 . They are weak solutions of ∂ ˉ j h = 0 \bar{\partial}_j h=0 ∂ ˉ j h = 0 for each j j j . Since 4 ∂ j ∂ ˉ j 4\partial_j\bar{\partial}_j 4 ∂ j ∂ ˉ j is the Laplacian in z j z_j z j , they are harmonic in each position in the sense of distributions, hence smooth by Weyl’s lemma, and holomorphic in each variable: h h h is holomorphic on B L N B_L^N B L N . □ \square □
Effect on the Vandermonde.
Lemma 4.2. The following holds:
(i) Multiplication by V N V_N V N is an isometry from L p h 2 L_{\mathrm{ph}}^2 L ph 2 to L s h 2 L_{\mathrm{sh}}^2 L sh 2 ;
(ii) it commutes with ∂ ˉ \bar{\partial} ∂ ˉ : ∂ ˉ j ( V N G ) = V N ∂ ˉ j G \bar{\partial}_j(V_NG)=V_N\bar{\partial}_jG ∂ ˉ j ( V N G ) = V N ∂ ˉ j G .
Consequently, for every real function F F F of class C 1 C^1 C 1 on a neighbourhood of B L N ‾ \overline{B_L^N} B L N , with F c : = F − E G F F_{\mathrm{c}}:=F-\mathbb{E}_{G}F F c := F − E G F ,
∥ V N F c ∥ s h 2 = K Var G [ F ] , ∥ ∂ ˉ ( V N F c ) ∥ s h 2 = 1 4 K D G ( F ) . (26) \lVert V_NF_{\mathrm{c}}\rVert_{\mathrm{sh}}^2=K\operatorname{Var}_{G}[F],\qquad
\lVert\bar{\partial}(V_NF_{\mathrm{c}})\rVert_{\mathrm{sh}}^2=\tfrac14KD_G(F). \tag*{(26)} ∥ V N F c ∥ sh 2 = K Var G [ F ] , ∥ ∂ ˉ ( V N F c ) ∥ sh 2 = 4 1 K D G ( F ) . (26)
Proof.
Proof of (i). By (4.2), we have ∣ V N G ∣ 2 e − Φ s h = ∣ G ∣ 2 e − Φ p h |V_NG|^2e^{-\Phi^{\mathrm{sh}}}=|G|^2e^{-\Phi^{\mathrm{ph}}} ∣ V N G ∣ 2 e − Φ sh = ∣ G ∣ 2 e − Φ ph .
Proof of (ii). The operator ∂ ˉ j \bar{\partial}_j ∂ ˉ j is a derivation, and ∂ ˉ j ( z i − z k ) = 0 \bar{\partial}_j(z_i-z_k)=0 ∂ ˉ j ( z i − z k ) = 0 for all i , k i,k i , k , hence ∂ ˉ j V N = 0 \bar{\partial}_jV_N=0 ∂ ˉ j V N = 0 .
Proof of (4.4). This identity follows from (i), (ii), ∂ ˉ j F c = ∂ ˉ j F \bar{\partial}_jF_{\mathrm{c}}=\bar{\partial}_jF ∂ ˉ j F c = ∂ ˉ j F and (4.3). □ \square □
Levi matrix of the shifted weight
Let J N : ( R 2 ) N → ( R 2 ) N J_N:(\mathbb{R}^2)^N\to(\mathbb{R}^2)^N J N : ( R 2 ) N → ( R 2 ) N be the multiplication by i i i of each position, J N ( z 1 , … , z N ) : = ( i z 1 , … , i z N ) J_N(z_1,\ldots,z_N):=(iz_1,\ldots,iz_N) J N ( z 1 , … , z N ) := ( i z 1 , … , i z N ) , which is a real linear map. For Φ ∈ C 2 ( ( R 2 ) N ; R ) \Phi\in C^2((\mathbb{R}^2)^N;\mathbb{R}) Φ ∈ C 2 (( R 2 ) N ; R ) and Z ∈ ( R 2 ) N Z\in(\mathbb{R}^2)^N Z ∈ ( R 2 ) N , define the real symmetric 2 N × 2 N 2N\times2N 2 N × 2 N matrix
C Φ ( Z ) : = 1 4 ( ∇ 2 Φ ( Z ) + J N T ∇ 2 Φ ( Z ) J N ) . (27) C_\Phi(Z):=\tfrac14\left(\nabla^2\Phi(Z)+J_N^{T}\nabla^2\Phi(Z)J_N\right). \tag*{(27)} C Φ ( Z ) := 4 1 ( ∇ 2 Φ ( Z ) + J N T ∇ 2 Φ ( Z ) J N ) . (27)
or equivalently, for v = ( v 1 , … , v N ) ∈ ( R 2 ) N v=(v_1,\ldots,v_N)\in(\mathbb{R}^2)^N v = ( v 1 , … , v N ) ∈ ( R 2 ) N ,
v ⋅ C Φ ( Z ) v = 1 4 ( v ⋅ ∇ 2 Φ ( Z ) v + J N v ⋅ ∇ 2 Φ ( Z ) J N v ) . (28) v\cdot C_\Phi(Z)v=\tfrac14\left(v\cdot\nabla^2\Phi(Z)v+J_Nv\cdot\nabla^2\Phi(Z)J_Nv\right). \tag*{(28)} v ⋅ C Φ ( Z ) v = 4 1 ( v ⋅ ∇ 2 Φ ( Z ) v + J N v ⋅ ∇ 2 Φ ( Z ) J N v ) . (28)
The same formulas define C Φ C_\Phi C Φ in the sense of distributions when Φ \Phi Φ is only locally integrable.
Remark 4.3. We have v ⋅ C Φ ( Z ) v = ∑ j , k ∂ j ∂ ˉ k Φ ( Z ) v j v k ‾ v\cdot C_\Phi(Z)v=\sum_{j,k}\partial_j\bar{\partial}_k\Phi(Z)v_j\overline{v_k} v ⋅ C Φ ( Z ) v = ∑ j , k ∂ j ∂ ˉ k Φ ( Z ) v j v k for v ∈ C N v\in\mathbb{C}^N v ∈ C N . Thus C Φ C_\Phi C Φ is the real form of the Levi matrix ( ∂ j ∂ ˉ k Φ ) j , k (\partial_j\bar{\partial}_k\Phi)_{j,k} ( ∂ j ∂ ˉ k Φ ) j , k , and C Φ C_\Phi C Φ being positive is the plurisubharmonicity of Φ \Phi Φ .
The next lemma bounds the Levi matrix of our shifted weight from below.
Lemma 4.4. We have C Φ s h ≥ a I 2 N C_{\Phi^{\mathrm{sh}}}\ge aI_{2N} C Φ sh ≥ a I 2 N in the sense of distributions.
The quadratic term gives exactly a a a . The interaction is harmonic away from collisions, and at the collisions it contributes ( 2 − β ) (2-\beta) ( 2 − β ) times a Dirac mass, since Δ g = − 2 π δ 0 \Delta g=-2\pi\delta_0 Δ g = − 2 π δ 0 ; it is essentially through this sign that the assumption β ≤ 2 \beta\le2 β ≤ 2 enters. We postpone the proof, a direct computation, to Appendix § 13 .§ 1 .
The fake Poincaré inequality
The following statement is the central building block for our concentration estimates. In comparison to the classical Brascamp–Lieb inequality (18) , the constants are replaced by the much larger space ker ∂ ˉ \ker\bar\partial ker ∂ ˉ : we use the name “fake” to emphasize this. We state it for functions of the form V N G V_NG V N G .
Proposition 4.5 (Fake Poincaré inequality). Let G G G be a complex-valued function of class C 1 C^1 C 1 on a neighbourhood of B L N ‾ \overline{B_L^N} B L N . Then V N G V_NG V N G and ∂ ˉ ( V N G ) \bar\partial(V_NG) ∂ ˉ ( V N G ) belong to L s h 2 L^2_{\mathrm{sh}} L sh 2 , and
∥ V N G − Π ker ∂ ˉ ( V N G ) ∥ s h 2 ≤ a − 1 ∥ ∂ ˉ ( V N G ) ∥ s h 2 . \left\|V_NG-\Pi_{\ker\bar\partial}(V_NG)\right\|_{\mathrm{sh}}^2\leq a^{-1}\left\|\bar\partial(V_NG)\right\|_{\mathrm{sh}}^2. ∥ V N G − Π k e r ∂ ˉ ( V N G ) ∥ sh 2 ≤ a − 1 ∂ ˉ ( V N G ) sh 2 .
It is the L 2 L^2 L 2 estimate of Hörmander for ∂ ˉ \bar\partial ∂ ˉ on the pseudoconvex domain B L N B_L^N B L N , with the plurisubharmonic weight Φ s h \Phi^{\mathrm{sh}} Φ sh [Hör65], [Hör90, Chapter IV], see [28 Theorem 5.1] for singular weights. The proof only uses the lower bound of Lemma 4.4 , and applies to weights of the form Φ s h − t S \Phi^{\mathrm{sh}}-tS Φ sh − tS , where S S S is of class C 2 C^2 C 2 on ( R 2 ) N (\mathbb{R}^2)^N ( R 2 ) N with − Θ I 2 N ≤ C S ≤ Θ I 2 N -\Theta I_{2N}\leq C_S\leq\Theta I_{2N} − Θ I 2 N ≤ C S ≤ Θ I 2 N for some Θ \Theta Θ , with a a a replaced by a − ∣ t ∣ Θ a-|t|\Theta a − ∣ t ∣Θ as long as it is positive.
Corollary 4.6 (From fake to true Poincaré). Let F F F be a real function of class C 1 C^1 C 1 on a neighbourhood of B L N ‾ \overline{B_L^N} B L N , and let F c : = F − E G F F_c:=F-\mathbb{E}_GF F c := F − E G F . If
Π ker ∂ ˉ ( V N F c ) = 0 , (29) \Pi_{\ker\bar\partial}(V_NF_c)=0, \tag*{(29)} Π k e r ∂ ˉ ( V N F c ) = 0 , (29) then
Var G [ F ] ≤ 1 4 a D G ( F ) . \operatorname{Var}_G[F]\leq\frac{1}{4a}\mathcal{D}_G(F). Var G [ F ] ≤ 4 a 1 D G ( F ) .
Proof. By Lemma 4.2 , ∥ V N F c ∥ s h 2 = K Var G [ F ] \left\|V_NF_c\right\|_{\mathrm{sh}}^2=\mathrm{K}\operatorname{Var}_G[F] ∥ V N F c ∥ sh 2 = K Var G [ F ] and ∥ ∂ ˉ ( V N F c ) ∥ s h 2 = 1 4 K D G ( F ) \left\|\bar\partial(V_NF_c)\right\|_{\mathrm{sh}}^2=\frac{1}{4}\mathrm{K}\mathcal{D}_G(F) ∂ ˉ ( V N F c ) sh 2 = 4 1 K D G ( F ) . Since the projection vanishes, the fake Poincaré inequality applied to G = F c G=F_c G = F c gives K Var G [ F ] ≤ a − 1 ⋅ 1 4 K D G ( F ) \mathrm{K}\operatorname{Var}_G[F]\leq a^{-1}\cdot\frac{1}{4}\mathrm{K}\mathcal{D}_G(F) K Var G [ F ] ≤ a − 1 ⋅ 4 1 K D G ( F ) .
Remark 4.7 (Connection with the literature). A related inequality is [78 Proposition 4.6] , for real symmetric functions of the gas with soft confinement. After a change of scale, its constant is 1 / ( 2 a ) 1/(2a) 1/ ( 2 a ) with a = π β / 2 a=\pi\beta/2 a = π β /2 . In [78 ] , the pair exponent of the interaction is β / N \beta/N β / N , and the main results concern the diffusive temperature scale, where this exponent is of order 1 / N 1/N 1/ N . Proposition 4.6 only assumes that the pair exponent is at most 2, which is our condition β ≤ 2 \beta\leq2 β ≤ 2 . At β = 2 \beta=2 β = 2 , see also [Cha26, Suz26]. The structure of the argument is the same: a ∂ ˉ \bar\partial ∂ ˉ estimate of Hörmander type, then a control of the holomorphic projection by symmetry. Our contribution at this stage is an explanation of where these inequalities come from, not the inequalities themselves.
Structure of the kernel
In view of Corollary 4.6 , we want to understand the projection of V N F c V_NF_c V N F c onto ker ∂ ˉ \ker\bar\partial ker ∂ ˉ .
Character. Let m ∈ Z m\in\mathbb{Z} m ∈ Z . A function G : ( R 2 ) N → C G:(\mathbb{R}^2)^N\to\mathbb{C} G : ( R 2 ) N → C has character m m m if G ( e i θ Z ) = e i m θ G ( Z ) G(e^{i\theta}Z)=e^{im\theta}G(Z) G ( e i θ Z ) = e im θ G ( Z ) for all θ ∈ R \theta\in\mathbb{R} θ ∈ R and Z ∈ ( R 2 ) N Z\in(\mathbb{R}^2)^N Z ∈ ( R 2 ) N , where e i θ Z : = ( e i θ z 1 , … , e i θ z N ) e^{i\theta}Z:=(e^{i\theta}z_1,\ldots,e^{i\theta}z_N) e i θ Z := ( e i θ z 1 , … , e i θ z N ) : rotating all particles by an angle θ \theta θ rotates the value of G G G by an angle m θ m\theta m θ . A radial real function has character 0.
Lemma 4.8. The function V N V_N V N , which is homogeneous of degree d N : = N ( N − 1 ) / 2 d_N:=N(N-1)/2 d N := N ( N − 1 ) /2 , has character d N d_N d N .
Lemma 4.9. A holomorphic function on B L N B_L^N B L N of character m m m vanishes if m < 0 m<0 m < 0 , and is a homogeneous polynomial of degree m m m if m ≥ 0 m\geq0 m ≥ 0 .
Proof. The Taylor series at 0, h ( Z ) = ∑ α c α Z α h(Z)=\sum_\alpha c_\alpha Z^\alpha h ( Z ) = ∑ α c α Z α , converges on the polydisk B L N B_L^N B L N . The rotation multiplies the monomial Z α Z^\alpha Z α by e i ∣ α ∣ θ e^{i|\alpha|\theta} e i ∣ α ∣ θ . Uniqueness of the coefficients forces c α = 0 c_\alpha=0 c α = 0 as soon as ∣ α ∣ ≠ m |\alpha|\ne m ∣ α ∣ = m .
Lemma 4.10 (Decomposition into characters). Let μ \mu μ be a measure on ( R 2 ) N (\mathbb{R}^2)^N ( R 2 ) N which is invariant under all simultaneous rotations Z ↦ e i θ Z Z\mapsto e^{i\theta}Z Z ↦ e i θ Z . For G ∈ L 2 ( μ ; C ) G\in L^2(\mu;\mathbb{C}) G ∈ L 2 ( μ ; C ) and m ∈ Z m\in\mathbb{Z} m ∈ Z , let
G m ( Z ) = ∫ 0 2 π e − i m θ G ( e i θ Z ) d θ 2 π . (30) G_m(Z)=\int_0^{2\pi}e^{-im\theta}G(e^{i\theta}Z)\,\frac{\mathrm{d}\theta}{2\pi}. \tag*{(30)} G m ( Z ) = ∫ 0 2 π e − im θ G ( e i θ Z ) 2 π d θ . (30)
Then: (i) G m G_m G m is well-defined μ \mu μ -almost everywhere, belongs to L 2 ( μ ; C ) L^2(\mu;\mathbb{C}) L 2 ( μ ; C ) , and has character m m m .
(ii) If U , V ∈ L 2 ( μ ; C ) U,V \in L^2(\mu;\mathbb{C}) U , V ∈ L 2 ( μ ; C ) have characters m 1 ≠ m 2 m_1 \ne m_2 m 1 = m 2 , then ⟨ U , V ⟩ μ = 0 \langle U,V\rangle_\mu=0 ⟨ U , V ⟩ μ = 0 .
(iii) G = ∑ m ∈ Z G m G=\sum_{m\in\mathbb{Z}}G^m G = ∑ m ∈ Z G m , the series converging in L 2 ( μ ; C ) L^2(\mu;\mathbb{C}) L 2 ( μ ; C ) , and ∥ G ∥ μ 2 = ∑ m ∈ Z ∥ G m ∥ μ 2 \lVert G\rVert_\mu^2=\sum_{m\in\mathbb{Z}}\lVert G_m\rVert_\mu^2 ∥ G ∥ μ 2 = ∑ m ∈ Z ∥ G m ∥ μ 2 .
Proof. This is standard Fourier analysis on S 1 S^1 S 1 .
The Vandermonde polynomial reappears in the following lemma, which we use to compute the projection on the kernel in the next subsection.
Lemma 4.11 (Structure of the kernel). An antisymmetric holomorphic function on B L N B_L^N B L N of character m < d N m<d_N m < d N vanishes, and one of character d N d_N d N is a complex multiple of V N V_N V N .
In words: among antisymmetric holomorphic functions, the Vandermonde is, up to a constant factor, the only one of minimal character. We postpone the (elementary) proof to Appendix § 13 .§ 3 .
True Poincaré inequality for rotation-invariant observables
Lemma 4.12 (Symmetries of the projection). Let Φ : B L N → [ − ∞ , + ∞ ] \Phi:B_L^N\to[-\infty,+\infty] Φ : B L N → [ − ∞ , + ∞ ] be measurable, with e − Φ e^{-\Phi} e − Φ bounded below by a positive constant on each compact subset of B L N B_L^N B L N , and invariant under permutations of the positions and under simultaneous rotations. Then the maps
G ↦ G ∘ σ for a permutation σ of the positions , G ↦ G ( e i θ ⋅ ) for θ ∈ R G\mapsto G\circ\sigma\qquad\text{for a permutation }\sigma\text{ of the positions},\qquad G\mapsto G(e^{i\theta}\cdot)\qquad\text{for }\theta\in\mathbb{R} G ↦ G ∘ σ for a permutation σ of the positions , G ↦ G ( e i θ ⋅ ) for θ ∈ R and the multiplications by complex numbers of modulus one are isometries of L Φ 2 L_\Phi^2 L Φ 2 commuting with Π ker ∂ ˉ \Pi_{\ker\bar{\partial}} Π k e r ∂ ˉ and preserving ker ∂ ˉ \ker\bar{\partial} ker ∂ ˉ . Consequently, Π ker ∂ ˉ \Pi_{\ker\bar{\partial}} Π k e r ∂ ˉ maps antisymmetric functions to antisymmetric functions, and functions of character m m m to functions of character m m m .
Proof. These maps are isometries. They preserve holomorphy, hence ker ∂ ˉ \ker\bar{\partial} ker ∂ ˉ . An invertible isometry preserving a closed subspace also preserves its orthogonal complement, hence commutes with the orthogonal projection onto it. For the consequence: G G G is antisymmetric if and only if G ∘ σ = sgn ( σ ) G G\circ\sigma=\operatorname{sgn}(\sigma)G G ∘ σ = sgn ( σ ) G for all σ \sigma σ , and G G G has character m m m if and only if G ( e i θ ⋅ ) = e i m θ G G(e^{i\theta}\cdot)=e^{im\theta}G G ( e i θ ⋅ ) = e im θ G for all θ \theta θ . Both are preserved by Π ker ∂ ˉ \Pi_{\ker\bar{\partial}} Π k e r ∂ ˉ .
Lemma 4.13 (Vanishing of the projection for rotation-invariant observables). Let Φ : B L N → [ − ∞ , + ∞ ] \Phi:B_L^N\to[-\infty,+\infty] Φ : B L N → [ − ∞ , + ∞ ] be measurable, with e − Φ e^{-\Phi} e − Φ bounded below by a positive constant on each compact subset of B L N B_L^N B L N , and invariant under permutations and under simultaneous rotations. Assume that ∥ V N ∥ Φ < ∞ \lVert V_N\rVert_\Phi<\infty ∥ V N ∥ Φ < ∞ , and let μ \mu μ be the probability measure d μ : = ∥ V N ∥ Φ − 2 ∣ V N ∣ 2 e − Φ d Z d\mu:=\lVert V_N\rVert_\Phi^{-2}|V_N|^2e^{-\Phi}dZ d μ := ∥ V N ∥ Φ − 2 ∣ V N ∣ 2 e − Φ d Z . Let F : ( R 2 ) N → R F:(\mathbb{R}^2)^N\to\mathbb{R} F : ( R 2 ) N → R be symmetric and rotation-invariant, with E μ F 2 < ∞ \mathbb{E}_\mu F^2<\infty E μ F 2 < ∞ , and let F c : = F − E μ F F_c:=F-\mathbb{E}_\mu F F c := F − E μ F . Then
Π ker ∂ ˉ ( V N F c ) = 0. \Pi_{\ker\bar{\partial}}(V_NF_c)=0. Π k e r ∂ ˉ ( V N F c ) = 0.
Proof. The function V N F c V_NF_c V N F c is antisymmetric, since F c F_c F c is symmetric and V N V_N V N antisymmetric, and it has character d N d_N d N , since F c F_c F c has character 0 and V N V_N V N character d N d_N d N . By Lemma 4.12, so does Π ker ∂ ˉ ( V N F c ) \Pi_{\ker\bar{\partial}}(V_NF_c) Π k e r ∂ ˉ ( V N F c ) , and by the division property of Lemma 4.11, Π ker ∂ ˉ ( V N F c ) = λ V N \Pi_{\ker\bar{\partial}}(V_NF_c)=\lambda V_N Π k e r ∂ ˉ ( V N F c ) = λ V N for some λ ∈ C \lambda\in\mathbb{C} λ ∈ C . Now V N V_N V N belongs to ker ∂ ˉ \ker\bar{\partial} ker ∂ ˉ . Since ∣ V N ∣ 2 e − Φ d Z = ∥ V N ∥ Φ 2 d μ |V_N|^2e^{-\Phi}dZ=\lVert V_N\rVert_\Phi^2d\mu ∣ V N ∣ 2 e − Φ d Z = ∥ V N ∥ Φ 2 d μ , we get
λ ∥ V N ∥ Φ 2 = ⟨ Π ker ∂ ˉ ( V N F c ) , V N ⟩ Φ = ⟨ V N F c , V N ⟩ Φ = ∥ V N ∥ Φ 2 E μ F c = 0. \lambda\lVert V_N\rVert_\Phi^2=\langle\Pi_{\ker\bar{\partial}}(V_NF_c),V_N\rangle_\Phi=\langle V_NF_c,V_N\rangle_\Phi=\lVert V_N\rVert_\Phi^2\mathbb{E}_\mu F_c=0. λ ∥ V N ∥ Φ 2 = ⟨ Π k e r ∂ ˉ ( V N F c ) , V N ⟩ Φ = ⟨ V N F c , V N ⟩ Φ = ∥ V N ∥ Φ 2 E μ F c = 0.
hence λ = 0 \lambda=0 λ = 0 and Π ker ∂ ˉ ( V N F c ) = 0 \Pi_{\ker\bar{\partial}}(V_NF_c)=0 Π k e r ∂ ˉ ( V N F c ) = 0 .
Proposition 4.14 (Poincaré inequality for rotation-invariant observables). Let a > 0 a>0 a > 0 , 0 < β ≤ 2 0<\beta\leq2 0 < β ≤ 2 , N ≥ 1 N\geq1 N ≥ 1 and L > 0 L>0 L > 0 . Let F : U → R F:\mathcal{U}\to\mathbb{R} F : U → R be a symmetric, rotation-invariant function of N N N positions, of class C ∞ C^\infty C ∞ on an open set U ⊂ ( R 2 ) N \mathcal{U}\subset(\mathbb{R}^2)^N U ⊂ ( R 2 ) N containing B L N B_L^N B L N . Then
Var G a , N , L α [ F ] ≤ 1 4 a D G a , N , L α ( F ) . \operatorname{Var}_{G_{a,N,L}^\alpha}[F]\leq\frac{1}{4a}\mathcal{D}_{G_{a,N,L}^\alpha}(F). Var G a , N , L α [ F ] ≤ 4 a 1 D G a , N , L α ( F ) .
Proof. The weight Φ s h \Phi^{\mathrm{sh}} Φ sh and the polydisk B L N B_L^N B L N are invariant under permutations and rotations. Lemma 4.13 (with Φ = Φ s h \Phi=\Phi^{\mathrm{sh}} Φ = Φ sh and μ = G \mu=\mathbb{G} μ = G ) shows that condition (29) holds, and Corollary 4.6 concludes.
Extension to tilted measures, exponential concentration
We have obtained a Poincaré inequality for our Gibbs measure. We want something stronger: exponential concentration, that is, a Gaussian bound on the Laplace transform of our observables. A classical way to get such bounds is to obtain Poincaré inequalities for a family of tilted measures , and then to integrate them: this is known as Herbst’s argument, see for example [61 , 15 ] .
Tilting a measure by an observable S S S means multiplying its density by e t S e^{tS} e tS , t ∈ R t \in\mathbb{R} t ∈ R . For our measure, this corresponds to replacing the weight Φ s h \Phi^{\mathrm{sh}} Φ sh by Φ t s h : = Φ s h − t S \Phi_t^{\mathrm{sh}} := \Phi^{\mathrm{sh}} - tS Φ t sh := Φ sh − tS .
It is easy to check that everything we did in Subsections 4.4 and 4.6 applies to Φ t s h \Phi_t^{\mathrm{sh}} Φ t sh as well, provided that:
the corresponding physical measure is well-defined, ∫ ( R 2 ) N ∣ V N ∣ 2 e − Φ t s h d Z < ∞ \int_{(\mathbb{R}^2)^N} |V_N|^2 e^{-\Phi_t^{\mathrm{sh}}}\,dZ < \infty ∫ ( R 2 ) N ∣ V N ∣ 2 e − Φ t sh d Z < ∞ ,
Φ t s h \Phi_t^{\mathrm{sh}} Φ t sh is invariant under permutations and rotations,
we have a positive lower bound on C Φ t s h C_{\Phi_t^{\mathrm{sh}}} C Φ t sh .
We ensure these conditions by choosing the observable appropriately.
Choice of the tilt. Let T \mathcal{T} T be the class of non-negative functions of class C 2 C^2 C 2 on R 2 \mathbb{R}^2 R 2 with compact support. Let f ∈ T f \in\mathcal{T} f ∈ T , supported in a ball B b B_b B b . The constants below depend on f f f . For x ∈ R 2 x \in\mathbb{R}^2 x ∈ R 2 , θ ∈ [ 0 , 2 π ] \theta\in[0,2\pi] θ ∈ [ 0 , 2 π ] , and for X ∈ Conf X \in\operatorname{Conf} X ∈ Conf , set
f x , θ ( z ) = f ( e − i θ ( z − x ) ) , Ψ x , θ ( X ) = e − X ( f x , θ ) , f_{x,\theta}(z) = f(e^{-i\theta}(z-x)), \qquad\Psi_{x,\theta}(X) = e^{-X(f_{x,\theta})}, f x , θ ( z ) = f ( e − i θ ( z − x )) , Ψ x , θ ( X ) = e − X ( f x , θ ) ,
and define
S R f ( X ) : = ∫ B R ∫ 0 2 π Ψ x , θ ( X ) d θ 2 π d x . (31) S_R^f(X) := \int_{B_R}\int_0^{2\pi} \Psi_{x,\theta}(X)\frac{d\theta}{2\pi}\,dx . \tag*{(31)} S R f ( X ) := ∫ B R ∫ 0 2 π Ψ x , θ ( X ) 2 π d θ d x . (31)
The map X ↦ S R f ( X ) X \mapsto S_R^f(X) X ↦ S R f ( X ) is a non-negative configuration observable, bounded by ∣ B R ∣ |B_R| ∣ B R ∣ , and B R + b B_{R+b} B R + b -local. Averaging over θ \theta θ makes it rotation-invariant, while averaging over x x x makes it an extensive quantity. Its restriction to configurations with N N N points (as in (2.7)), which we still denote by S R f S_R^f S R f , is a C 2 C^2 C 2 , symmetric, rotation-invariant function of N N N positions. It corresponds to the average, over all positions in B R B_R B R and all orientations, of the Laplace functional of the bump function f f f placed at x x x and rotated by θ \theta θ .
For t ∈ R t \in\mathbb{R} t ∈ R , we define the tilted weight
Φ t s h = Φ s h − t S R f . \Phi_t^{\mathrm{sh}} = \Phi^{\mathrm{sh}} - tS_R^f. Φ t sh = Φ sh − t S R f .
Quantities depending on f f f . The observable S R f S_R^f S R f is a symmetric, rotation-invariant function of the positions. We need bounds on its derivatives, with constants which depend only on f f f .
Lemma 4.15. There are constants D f , Θ f D_f,\Theta_f D f , Θ f , depending only on f f f , such that for all N ≥ 1 N \ge1 N ≥ 1 and all Z ∈ ( R 2 ) N Z \in(\mathbb{R}^2)^N Z ∈ ( R 2 ) N ,
∑ i ∣ ∇ i S R f ∣ 2 ≤ D f R 2 , − Θ f I 2 N ≤ C S R f ( Z ) ≤ Θ f I 2 N . (32) \sum_i \lvert\nabla_i S_R^f\rvert^2 \le D_fR^2, \qquad-\Theta_f I_{2N} \le C_{S_R^f}(Z) \le\Theta_f I_{2N}. \tag*{(32)} i ∑ ∣ ∇ i S R f ∣ 2 ≤ D f R 2 , − Θ f I 2 N ≤ C S R f ( Z ) ≤ Θ f I 2 N . (32)
Proof. In this proof, the constants C C C depend only on f f f . By (31) , S R f S_R^f S R f is the average of the functions Ψ x , θ \Psi_{x,\theta} Ψ x , θ over x ∈ B R x \in B_R x ∈ B R and θ ∈ [ 0 , 2 π ] \theta\in[0,2\pi] θ ∈ [ 0 , 2 π ] .
Fix ( x , θ ) (x,\theta) ( x , θ ) and Z ∈ ( R 2 ) N Z \in(\mathbb{R}^2)^N Z ∈ ( R 2 ) N . Let
T : = ∑ i f x , θ ( z i ) , Ψ : = Ψ x , θ = e − T . T := \sum_i f_{x,\theta}(z_i), \qquad\Psi:= \Psi_{x,\theta} = e^{-T}. T := i ∑ f x , θ ( z i ) , Ψ := Ψ x , θ = e − T .
Since f x , θ f_{x,\theta} f x , θ is supported in B ( x , b ) B(x,b) B ( x , b ) , we have ∇ f x , θ ( z i ) = 0 \nabla f_{x,\theta}(z_i) = 0 ∇ f x , θ ( z i ) = 0 and Δ f x , θ ( z i ) = 0 \Delta f_{x,\theta}(z_i) = 0 Δ f x , θ ( z i ) = 0 for ∣ z i − x ∣ > b |z_i-x| > b ∣ z i − x ∣ > b . Since f f f is non-negative and C 2 C^2 C 2 , it is classical that ∣ ∇ f ∣ 2 ≤ 2 ∥ ∇ 2 f ∥ ∞ f \lvert\nabla f\rvert^2 \le2\lVert\nabla^2 f\rVert_\infty f ∣ ∇ f ∣ 2 ≤ 2 ∥ ∇ 2 f ∥ ∞ f , hence
∑ ∣ z i − x ∣ ≤ b ∣ ∇ f x , θ ( z i ) ∣ 2 ≤ C T . (33) \sum_{\lvert z_i-x\rvert\le b} \lvert\nabla f_{x,\theta}(z_i)\rvert^2 \le C T. \tag*{(33)} ∣ z i − x ∣ ≤ b ∑ ∣ ∇ f x , θ ( z i ) ∣ 2 ≤ C T . (33)
Gradient. We have ∇ i Ψ = − e − T ∇ f x , θ ( z i ) \nabla_i\Psi=-e^{-T}\nabla f_{x,\theta}(z_i) ∇ i Ψ = − e − T ∇ f x , θ ( z i ) , and (33) gives ∑ i ∣ ∇ i Ψ ∣ 2 ≤ C T e − 2 T ≤ C \sum_i|\nabla_i\Psi|^2\le CT e^{-2T}\le C ∑ i ∣ ∇ i Ψ ∣ 2 ≤ C T e − 2 T ≤ C . Moreover, ∇ i Ψ x , θ ( Z ) = 0 \nabla_i\Psi_{x,\theta}(Z)=0 ∇ i Ψ x , θ ( Z ) = 0 unless ∣ z i − x ∣ ≤ b |z_i-x|\le b ∣ z i − x ∣ ≤ b . Differentiating (31) under the integral, and applying Cauchy–Schwarz on the set { x : ∣ z i − x ∣ ≤ b } \{x:|z_i-x|\le b\} { x : ∣ z i − x ∣ ≤ b } , of area π b 2 \pi b^2 π b 2 , we get
∣ ∇ i S R f ∣ 2 ≤ π b 2 ∫ B R ∫ 0 2 π ∣ ∇ i Ψ x , θ ∣ 2 d θ 2 π d x . |\nabla_iS_R^f|^2\le\pi b^2\int_{B_R}\int_0^{2\pi}|\nabla_i\Psi_{x,\theta}|^2\frac{d\theta}{2\pi}\,dx. ∣ ∇ i S R f ∣ 2 ≤ π b 2 ∫ B R ∫ 0 2 π ∣ ∇ i Ψ x , θ ∣ 2 2 π d θ d x .
Summing over i i i gives the first bound in (32) .
Levi matrix. For v ∈ ( R 2 ) N v\in(\mathbb{R}^2)^N v ∈ ( R 2 ) N , (28) and the chain rule give
v ⋅ C Ψ ( Z ) v = e − T 4 [ ( ∑ ∣ z i − x ∣ ≤ b ∇ f x , θ ( z i ) ⋅ v i ) 2 + ( ∑ ∣ z i − x ∣ ≤ b ∇ f x , θ ( z i ) ⋅ i v i ) 2 − ∑ ∣ z i − x ∣ ≤ b Δ f x , θ ( z i ) ∣ v i ∣ 2 ] . v\cdot C_\Psi(Z)v=\frac{e^{-T}}{4}\left[\left(\sum_{|z_i-x|\le b}\nabla f_{x,\theta}(z_i)\cdot v_i\right)^2+\left(\sum_{|z_i-x|\le b}\nabla f_{x,\theta}(z_i)\cdot iv_i\right)^2-\sum_{|z_i-x|\le b}\Delta f_{x,\theta}(z_i)|v_i|^2\right]. v ⋅ C Ψ ( Z ) v = 4 e − T ∣ z i − x ∣ ≤ b ∑ ∇ f x , θ ( z i ) ⋅ v i 2 + ∣ z i − x ∣ ≤ b ∑ ∇ f x , θ ( z i ) ⋅ i v i 2 − ∣ z i − x ∣ ≤ b ∑ Δ f x , θ ( z i ) ∣ v i ∣ 2 .
By Cauchy–Schwarz, since ∣ i v i ∣ = ∣ v i ∣ |iv_i|=|v_i| ∣ i v i ∣ = ∣ v i ∣ , and by (33) , each of the two squares is at most C T ∑ ∣ z i − x ∣ ≤ b ∣ v i ∣ 2 CT\sum_{|z_i-x|\le b}|v_i|^2 C T ∑ ∣ z i − x ∣ ≤ b ∣ v i ∣ 2 , and the last sum is at most C ∑ ∣ z i − x ∣ ≤ b ∣ v i ∣ 2 C\sum_{|z_i-x|\le b}|v_i|^2 C ∑ ∣ z i − x ∣ ≤ b ∣ v i ∣ 2 . Since T e − T ≤ 1 / e Te^{-T}\le1/e T e − T ≤ 1/ e and e − T ≤ 1 e^{-T}\le1 e − T ≤ 1 ,
∣ v ⋅ C Ψ x , θ ( Z ) v ∣ ≤ C ∑ i 1 { ∣ z i − x ∣ ≤ b } ∣ v i ∣ 2 . |v\cdot C_{\Psi_{x,\theta}}(Z)v|\le C\sum_i\mathbf{1}_{\{|z_i-x|\le b\}}|v_i|^2. ∣ v ⋅ C Ψ x , θ ( Z ) v ∣ ≤ C i ∑ 1 { ∣ z i − x ∣ ≤ b } ∣ v i ∣ 2 .
By linearity, C S R f C_{S_R^f} C S R f is the average of the C Ψ x , θ C_{\Psi_{x,\theta}} C Ψ x , θ over x ∈ B R x\in B_R x ∈ B R and θ \theta θ . Integrating the previous bound, with ∫ 1 { ∣ z i − x ∣ ≤ b } d x = π b 2 \int\mathbf{1}_{\{|z_i-x|\le b\}}dx=\pi b^2 ∫ 1 { ∣ z i − x ∣ ≤ b } d x = π b 2 , gives the second bound in (32) .
Using tilted weights to get exponential concentration.
Proposition 4.16 (Exponential concentration). Fix a > 0 a>0 a > 0 , 0 < β ≤ 2 0<\beta\le2 0 < β ≤ 2 , N ≥ 1 N\ge1 N ≥ 1 , L , R > 0 L,R>0 L , R > 0 and f ∈ T f\in\mathcal{T} f ∈ T . There are constants t a , f > 0 t_{a,f}>0 t a , f > 0 and C a , f < ∞ C_{a,f}<\infty C a , f < ∞ , which depend only on a a a and f f f , such that for all ∣ t ∣ ≤ t a , f |t|\le t_{a,f} ∣ t ∣ ≤ t a , f ,
log E G N , L a exp ( t ( S R f − E G N , L a S R f ) ) ≤ C a , f π R 2 t 2 . (34) \log\mathbb{E}_{\mathbb{G}_{N,L}^{a}}\exp\left(t\left(S_R^f-\mathbb{E}_{\mathbb{G}_{N,L}^{a}}S_R^f\right)\right)\le C_{a,f}\pi R^2t^2. \tag*{(34)} log E G N , L a exp ( t ( S R f − E G N , L a S R f ) ) ≤ C a , f π R 2 t 2 . (34)
Proof. Let t a , f : = a / ( 2 Θ f ) t_{a,f}:=a/(2\Theta_f) t a , f := a / ( 2 Θ f ) . For ∣ t ∣ ≤ t a , f |t|\le t_{a,f} ∣ t ∣ ≤ t a , f , let G t \mathbb{G}_t G t be the probability measure with density proportional to ∣ V N ∣ 2 e − Φ t s h + t S R f 1 B N R |V_N|^2e^{-\Phi_t^{\mathrm{sh}}+tS_R^f}\mathbf{1}_{B_N^R} ∣ V N ∣ 2 e − Φ t sh + t S R f 1 B N R . This corresponds to tilting the measure G \mathbb{G} G by e t S R f e^{tS_R^f} e t S R f .
By Lemma 4.4 and Lemma 4.15 , C Φ t s h ≥ ( a − ∣ t ∣ Θ f ) I 2 N ≥ a 2 I 2 N C_{\Phi_t^{\mathrm{sh}}}\ge(a-|t|\Theta_f)I_{2N}\ge\frac{a}{2}I_{2N} C Φ t sh ≥ ( a − ∣ t ∣ Θ f ) I 2 N ≥ 2 a I 2 N in the sense of distributions. The weight is invariant under permutations and rotations, since S R f S_R^f S R f is symmetric and rotation-invariant.
By Lemma 4.13, applied to Φ t s h \Phi_t^{\mathrm{sh}} Φ t sh , we have Π ker ∂ ˉ ( V N ( S R f − E G t S R f ) ) = 0 \Pi_{\ker\bar{\partial}}(V_N(S_R^f-\mathbb{E}_{\mathbb{G}_t}S_R^f))=0 Π k e r ∂ ˉ ( V N ( S R f − E G t S R f )) = 0 . The fake Poincaré inequality for the weight Φ t s h \Phi_t^{\mathrm{sh}} Φ t sh (see the remark after Proposition 4.5 ) and the computation of Lemma 4.2 , with Φ t s h \Phi_t^{\mathrm{sh}} Φ t sh and G t \mathbb{G}_t G t instead of Φ s h \Phi^{\mathrm{sh}} Φ sh and G \mathbb{G} G , give
Var G t [ S R f ] ≤ D G t ( S R f ) 2 a ≤ D f R 2 2 a , (35) \operatorname{Var}_{\mathbb{G}_t}[S_R^f]\le\frac{\mathcal{D}_{\mathbb{G}_t}(S_R^f)}{2a}\le\frac{D_fR^2}{2a}, \tag*{(35)} Var G t [ S R f ] ≤ 2 a D G t ( S R f ) ≤ 2 a D f R 2 , (35)
where we used the gradient bound of Lemma 4.15 .
Finally, we integrate (35) . The function t ↦ log E G e t S R f t\mapsto\log\mathbb{E}_{\mathbb{G}}e^{tS_R^f} t ↦ log E G e t S R f has derivative E G t S R f \mathbb{E}_{\mathbb{G}_t}S_R^f E G t S R f at t = 0 t=0 t = 0 and second derivative Var G t [ S R f ] \operatorname{Var}_{\mathbb{G}_t}[S_R^f] Var G t [ S R f ] at every t t t . By Taylor’s formula,
log E G e t ( S R f − E G S R f ) = t 2 ∫ 0 1 ( 1 − s ) Var G s t [ S R f ] d s ≤ D f R 2 4 a t 2 , \log\mathbb{E}_{\mathbb{G}}e^{t(S_R^f-\mathbb{E}_{\mathbb{G}}S_R^f)}=t^2\int_0^1(1-s)\operatorname{Var}_{\mathbb{G}_{st}}[S_R^f]\,ds\le\frac{D_fR^2}{4a}t^2, log E G e t ( S R f − E G S R f ) = t 2 ∫ 0 1 ( 1 − s ) Var G s t [ S R f ] d s ≤ 4 a D f R 2 t 2 ,
which is (34) with C a , f : = D f / ( 4 π a ) C_{a,f}:=D_f/(4\pi a) C a , f := D f / ( 4 π a ) .
True Poincaré inequality for general observables
We extend the Poincaré inequality of Proposition 4.14 to observables which are not rotation-invariant.
Proposition 4.17 (Poincaré inequality for general observables). Let a > 0 a > 0 a > 0 , 0 < β ≤ 2 0 < \beta\le2 0 < β ≤ 2 , N ≥ 1 N \ge1 N ≥ 1 and L > 0 L > 0 L > 0 . Let F : U → R F : \mathcal{U} \to\mathbb{R} F : U → R be a symmetric function of N N N positions, defined and of class C ∞ C^{\infty} C ∞ on an open set U ⊂ ( R 2 ) N \mathcal{U} \subset(\mathbb{R}^{2})^{N} U ⊂ ( R 2 ) N containing B ‾ L N \overline{B}_{L}^{N} B L N . Then
Var G N , L a [ F ] ≤ 1 a D N , L G a ( F ) . (36) \operatorname{Var}_{G_{N,L}^{a}}[F] \le\frac{1}{a}\mathcal{D}_{N,L}^{G^{a}}(F). \tag*{(36)} Var G N , L a [ F ] ≤ a 1 D N , L G a ( F ) . (36) Consequently, by Cauchy–Schwarz, for two such functions F , G F,G F , G ,
∣ Cov G N , L a [ F , G ] ∣ ≤ 1 a D N , L G a ( F ) 1 / 2 D N , L G a ( G ) 1 / 2 . \left|\operatorname{Cov}_{G_{N,L}^{a}}[F,G]\right| \le\frac{1}{a}\mathcal{D}_{N,L}^{G^{a}}(F)^{1/2}\mathcal{D}_{N,L}^{G^{a}}(G)^{1/2}. Cov G N , L a [ F , G ] ≤ a 1 D N , L G a ( F ) 1/2 D N , L G a ( G ) 1/2 .
The idea: keep only the negative characters. We need again to understand the projection of V N F V_N F V N F onto the kernel ker ∂ ˉ \ker\bar{\partial} ker ∂ ˉ described in Subsection 4.5. The projection is linear, so we can proceed mode by mode.
As observed above, the Vandermonde has character d N d_N d N , and the antisymmetric elements of the kernel have character at least d N d_N d N (division property of Lemma 4.11). Hence, if F m F_m F m is a mode of F F F of character m < 0 m < 0 m < 0 , then the product V N F m V_NF_m V N F m has a character strictly less than d N d_N d N , and is thus already orthogonal to the kernel. This, however, does not work for modes m ≥ 0 m \ge0 m ≥ 0 .
We get rid of those modes using the following simple remark: for a real function F F F , the mode F − m F_{-m} F − m is the complex conjugate of F m F_m F m , so the anti-analytic signal H = F 0 − E G F + 2 ∑ m < 0 F m H = F_0 - \mathbb{E}_G F + 2\sum_{m<0}F_m H = F 0 − E G F + 2 ∑ m < 0 F m of Subsection 3.2 has the same real part as F − E G F F - \mathbb{E}_G F F − E G F , and no modes m > 0 m > 0 m > 0 . Its complex conjugate is the analytic signal of F − E G F F - \mathbb{E}_G F F − E G F from signal processing, see for example [76 Chapter 12] .
Proof of Proposition 4.17.
Step 1. The anti-analytic signal. Fix L 1 > L L_1 > L L 1 > L such that B L 1 N ⊂ U B_{L_1}^{N} \subset\mathcal{U} B L 1 N ⊂ U . For m ∈ Z m \in\mathbb{Z} m ∈ Z , let F m F_m F m be the mode of F F F of character m m m given by (30) . The F m F_m F m are smooth and symmetric on B L 1 N B_{L_1}^{N} B L 1 N , and the series ∑ m F m \sum_m F_m ∑ m F m , together with all its derivatives, converges to F F F uniformly on B ‾ L N \overline{B}_{L}^{N} B L N . Since F F F is real, F − m = F m ‾ F_{-m} = \overline{F_m} F − m = F m .
The anti-analytic signal H H H of F F F is again smooth on a neighbourhood of B ‾ L N \overline{B}_{L}^{N} B L N , and its real part is F c : = F − E G F F_c := F - \mathbb{E}_G F F c := F − E G F . Since F c F_c F c is real, we obtain
Var G [ F ] = E G [ F c Re H ] = Re E G [ F c H ] ≤ Var G [ F ] 1 / 2 ( E G ∣ H ∣ 2 ) 1 / 2 , \operatorname{Var}_{G}[F] = \mathbb{E}_{G}[F_c\operatorname{Re}H] = \operatorname{Re}\mathbb{E}_{G}[F_cH] \le\operatorname{Var}_{G}[F]^{1/2}\left(\mathbb{E}_{G}|H|^2\right)^{1/2}, Var G [ F ] = E G [ F c Re H ] = Re E G [ F c H ] ≤ Var G [ F ] 1/2 ( E G ∣ H ∣ 2 ) 1/2 ,
that is,
Var G [ F ] ≤ E G ∣ H ∣ 2 . (37) \operatorname{Var}_{G}[F] \le\mathbb{E}_{G}|H|^2. \tag*{(37)} Var G [ F ] ≤ E G ∣ H ∣ 2 . (37)
It remains to control the size of H H H . We do it using the fake Poincaré inequality, which requires knowing the projection of V N H V_NH V N H (Step 2) and the Dirichlet energy of H H H (Step 3).
Step 2. The projection of the anti-analytic signal.
Claim 4.18. We have
Π ker ∂ ˉ ( V N H ) = 0. \Pi_{\ker\bar{\partial}}(V_NH) = 0. Π k e r ∂ ˉ ( V N H ) = 0.
Proof. For m < 0 m < 0 m < 0 , V N F m V_NF_m V N F m is antisymmetric and of character d N + m < d N d_N + m < d_N d N + m < d N . By Lemma 4.12, so is its projection, which vanishes by the division property of Lemma 4.11.
It remains to treat the zero mode F 0 − E G F F_0 - \mathbb{E}_G F F 0 − E G F , which is the centred rotation average of F F F , and the projection of V N ( F 0 − E G F ) V_N(F_0 - \mathbb{E}_G F) V N ( F 0 − E G F ) vanishes by Lemma 4.13.
Step 3. The size of the anti-analytic signal. The fake Poincaré inequality (Proposition 4.5 ) with G = H G=H G = H then reads
∥ V N H ∥ s h 2 ≤ a − 1 ∥ ∂ ˉ ( V N H ) ∥ s h 2 . \lVert V_NH\rVert_{\mathrm{sh}}^2 \leq a^{-1}\lVert\bar{\partial}(V_NH)\rVert_{\mathrm{sh}}^2. ∥ V N H ∥ sh 2 ≤ a − 1 ∥ ∂ ˉ ( V N H ) ∥ sh 2 .
By Lemma 4.2 , both sides can be written under G G G :
E G ∣ H ∣ 2 ≤ a − 1 E G ∑ j ∣ ∂ ˉ j H ∣ 2 . (38) \mathbb{E}_G|H|^2 \leq a^{-1}\mathbb{E}_G\sum_j|\bar{\partial}_jH|^2. \tag*{(38)} E G ∣ H ∣ 2 ≤ a − 1 E G j ∑ ∣ ∂ ˉ j H ∣ 2 . (38)
It remains to compare the energy of H H H with that of F F F :
E G ∑ j ∣ ∂ ˉ j H ∣ 2 ≤ 4 E G ∑ j ∣ ∂ ˉ j F ∣ 2 = D G ( F ) . (39) \mathbb{E}_G\sum_j|\bar{\partial}_jH|^2 \leq4\mathbb{E}_G\sum_j|\bar{\partial}_jF|^2 = \mathcal{D}_G(F). \tag*{(39)} E G j ∑ ∣ ∂ ˉ j H ∣ 2 ≤ 4 E G j ∑ ∣ ∂ ˉ j F ∣ 2 = D G ( F ) . (39)
Indeed, by orthogonality of the modes in L 2 ( G ) L^2(G) L 2 ( G ) (Lemma 4.10 ), we have for each j j j :
E G ∣ ∂ ˉ j H ∣ 2 = E G ∣ ( ∂ ˉ j F ) 1 ∣ 2 + 4 ∑ m ≤ 0 E G ∣ ( ∂ ˉ j F ) m ∣ 2 ≤ 4 ∑ m ∈ Z E G ∣ ( ∂ ˉ j F ) m ∣ 2 = 4 E G ∣ ∂ ˉ j F ∣ 2 , \mathbb{E}_G|\bar{\partial}_jH|^2=\mathbb{E}_G|(\bar{\partial}_jF)_1|^2+4\sum_{m\leq0}\mathbb{E}_G|(\bar{\partial}_jF)_m|^2\leq4\sum_{m\in\mathbb{Z}}\mathbb{E}_G|(\bar{\partial}_jF)_m|^2=4\mathbb{E}_G|\bar{\partial}_jF|^2, E G ∣ ∂ ˉ j H ∣ 2 = E G ∣ ( ∂ ˉ j F ) 1 ∣ 2 + 4 m ≤ 0 ∑ E G ∣ ( ∂ ˉ j F ) m ∣ 2 ≤ 4 m ∈ Z ∑ E G ∣ ( ∂ ˉ j F ) m ∣ 2 = 4 E G ∣ ∂ ˉ j F ∣ 2 ,
and 4 ∣ ∂ ˉ j F ∣ 2 = ∣ ∇ j F ∣ 2 4|\bar{\partial}_jF|^2=|\nabla_jF|^2 4∣ ∂ ˉ j F ∣ 2 = ∣ ∇ j F ∣ 2 since F F F is real ((25) ).
Step 4. Conclusion. Combining ((37) ), ((38) ) and ((39) ), we obtain ((36) .
Decay of correlations
In this section, 0 < β ≤ 2 0<\beta\leq2 0 < β ≤ 2 and a = π β / 2 a=\pi\beta/2 a = π β /2 , so that a = 1 / ( 4 ℓ D 2 ) a=1/(4\ell_{\mathrm{D}}^2) a = 1/ ( 4 ℓ D 2 ) . We follow the plan of Subsection 3.2 and Table 1 : we regularize the measure (Subsection 5.1), write covariances with the Helffer–Sjöst rand equation (Subsection 5.2), and conjugate this equation by an exponential weight (Subsections 5.3 to 5.5).
Proposition 5.1 (Decay of correlations for smooth observables). Let ℓ > ℓ D \ell>\ell_{\mathrm{D}} ℓ > ℓ D , and let
C ( β , ℓ ) : = 2 ( ℓ D − 2 − ℓ − 2 ) − 1 = 1 2 ( a − 1 4 ℓ 2 ) . C(\beta,\ell):=2\left(\ell_{\mathrm{D}}^{-2}-\ell^{-2}\right)^{-1}=\frac{1}{2\left(a-\frac{1}{4\ell^2}\right)}. C ( β , ℓ ) := 2 ( ℓ D − 2 − ℓ − 2 ) − 1 = 2 ( a − 4 ℓ 2 1 ) 1 .
Let P \mathbf{P} P be a limit point of the local configurations X N \mathbf{X}_N X N under P N , β \mathbf{P}_{N,\beta} P N , β , let x ∈ R 2 x\in\mathbb{R}^2 x ∈ R 2 , R > 0 R>0 R > 0 , and let Ω ⊂ R 2 \Omega\subset\mathbb{R}^2 Ω ⊂ R 2 be a Borel set at distance d d d from B ( x , R ) B(x,R) B ( x , R ) . For every B ( x , R ) B(x,R) B ( x , R ) -local cylinder observable F F F and every Ω \Omega Ω -local cylinder observable G G G ,
∣ Cov P [ F , G ] ∣ ≤ C ( β , ℓ ) e − d / ℓ D P ( F ) 1 / 2 D P ( G ) 1 / 2 . \left|\operatorname{Cov}_{\mathbf{P}}[F,G]\right|\leq C(\beta,\ell)e^{-d/\ell}\mathcal{D}_{\mathbf{P}}(F)^{1/2}\mathcal{D}_{\mathbf{P}}(G)^{1/2}. ∣ Cov P [ F , G ] ∣ ≤ C ( β , ℓ ) e − d / ℓ D P ( F ) 1/2 D P ( G ) 1/2 .
Regularization
We first regularize the measure. Replacing the hard wall by a wall of finite strength lets us write the equation on C N \mathbb{C}^N C N , and smoothing the logarithm makes the Levi matrix at collisions a function.
Fix N , L N,L N , L . Let ψ ∈ C c ∞ ( B 1 ) \psi\in C^\infty_c(B_1) ψ ∈ C c ∞ ( B 1 ) be radial and non-negative, with ∫ ψ = 1 \int\psi=1 ∫ ψ = 1 , and put
ψ ε ( z ) : = ε − 2 ψ ( z / ε ) and g ε : = g ∗ ψ ε for 0 < ε ≤ 1. \psi_\varepsilon(z):=\varepsilon^{-2}\psi(z/\varepsilon)\qquad\text{and}\qquad g_\varepsilon:=g*\psi_\varepsilon\qquad\text{for }0<\varepsilon\leq1. ψ ε ( z ) := ε − 2 ψ ( z / ε ) and g ε := g ∗ ψ ε for 0 < ε ≤ 1.
For a wall strength λ > 0 \lambda>0 λ > 0 , define
Φ ε s h ( Z ) : = a ∣ Z ∣ 2 + λ ∑ i ( ∣ z i ∣ 2 − L 2 ) + 4 + ( β − 2 ) ∑ i < j g ε ( z i − z j ) . (40) \Phi_\varepsilon^{\mathrm{sh}}(Z):=a|Z|^2+\lambda\sum_i\left(|z_i|^2-L^2\right)_+^4+(\beta-2)\sum_{i<j}g_\varepsilon(z_i-z_j). \tag*{(40)} Φ ε sh ( Z ) := a ∣ Z ∣ 2 + λ i ∑ ( ∣ z i ∣ 2 − L 2 ) + 4 + ( β − 2 ) i < j ∑ g ε ( z i − z j ) . (40)
The corresponding probability measure on C N \mathbb{C}^N C N is
d G ε ( Z ) : = K ε − 1 ∣ V N ( Z ) ∣ 2 e − Φ ε s h ( Z ) d Z , K ε : = ∫ C N ∣ V N ∣ 2 e − Φ ε s h d Z . \mathrm{d}G_\varepsilon(Z):=K_\varepsilon^{-1}|V_N(Z)|^2e^{-\Phi_\varepsilon^{\mathrm{sh}}(Z)}\mathrm{d}Z,\qquad K_\varepsilon:=\int_{\mathbb{C}^N}|V_N|^2e^{-\Phi_\varepsilon^{\mathrm{sh}}}\mathrm{d}Z. d G ε ( Z ) := K ε − 1 ∣ V N ( Z ) ∣ 2 e − Φ ε sh ( Z ) d Z , K ε := ∫ C N ∣ V N ∣ 2 e − Φ ε sh d Z .
For j ≠ k j\neq k j = k , we call
ω k j , ε : = π 2 ( 2 − β ) ψ ε ( z k − z j ) \omega_{kj,\varepsilon}:=\frac{\pi}{2}(2-\beta)\psi_\varepsilon(z_k-z_j) ω k j , ε := 2 π ( 2 − β ) ψ ε ( z k − z j )
the collision coefficients. We keep λ \lambda λ fixed while letting ε → 0 \varepsilon\to0 ε → 0 , then let λ → ∞ \lambda\to\infty λ → ∞ to recover the hard wall. We do not write the dependence on λ \lambda λ . The convergence of covariances and Dirichlet energies is justified in Appendix A.§ 5 .
The Helffer–Sjöstrand equation
We now represent the covariance of two observables under G ε G_{\varepsilon} G ε through the solution of an equation on C N \mathbb{C}^{N} C N . Let F F F be a real cylinder observable, restricted to N N N positions. As in Subsection 4.8, we replace F − E G ε F F-\mathbb{E}_{G_{\varepsilon}}F F − E G ε F by its anti-analytic signal H H H , defined with G ε G_{\varepsilon} G ε in place of G G G , which has the same real part. We work in the space L s h , ε 2 : = L 2 ( C N , e − Φ ε s h d Z ) L_{\mathrm{sh},\varepsilon}^{2}:=L^{2}(\mathbb{C}^{N},e^{-\Phi_{\varepsilon}^{\mathrm{sh}}}\,dZ) L sh , ε 2 := L 2 ( C N , e − Φ ε sh d Z ) , with its norm, its scalar product and the corresponding adjoints. Let
L : = ∑ j ∂ ˉ j ∗ ∂ ˉ j , A 1 : = L ⊗ I N + C Φ ε s h . \mathcal{L}:=\sum_{j}\bar{\partial}_{j}^{*}\bar{\partial}_{j},\qquad A_{1}:=\mathcal{L}\otimes I_{N}+C_{\Phi_{\varepsilon}^{\mathrm{sh}}}. L := j ∑ ∂ ˉ j ∗ ∂ ˉ j , A 1 := L ⊗ I N + C Φ ε sh .
By Lemma A.3 , the function V N H V_{N}H V N H is orthogonal to ker ∂ ˉ = ker L \ker\bar{\partial}=\ker\mathcal{L} ker ∂ ˉ = ker L , so the equation L ϕ = V N H \mathcal{L}\phi=V_{N}H L ϕ = V N H has a solution. When we apply ∂ ˉ \bar{\partial} ∂ ˉ to this equation, commuting ∂ ˉ \bar{\partial} ∂ ˉ with L \mathcal{L} L produces the term C Φ ε s h C_{\Phi_{\varepsilon}^{\mathrm{sh}}} C Φ ε sh of A 1 A_{1} A 1 . By Lemma 4.2 , the right-hand side becomes V N ∂ ˉ H V_{N}\bar{\partial}H V N ∂ ˉ H .
The representation below is the complex version of the computation of Appendix A.§ 3 , and its proof is in Appendix A.§ 5 . Its main feature is that A 1 ≥ a A_{1}\geq a A 1 ≥ a , since L ⊗ I N ≥ 0 \mathcal{L}\otimes I_{N}\geq0 L ⊗ I N ≥ 0 and the Levi matrix is at least a a a . Coercivity is what allows us to bound the solution of an equation: if an operator B B B satisfies Re ⟨ B w , w ⟩ ≥ c ∥ w ∥ 2 \operatorname{Re}\langle Bw,w\rangle\geq c\lVert w\rVert^{2} Re ⟨ B w , w ⟩ ≥ c ∥ w ∥ 2 for every w w w , with c > 0 c>0 c > 0 , then
∥ w ∥ ≤ c − 1 ∥ B w ∥ for every w , (41) \lVert w\rVert\leq c^{-1}\lVert Bw\rVert\qquad\text{for every }w, \tag*{(41)} ∥ w ∥ ≤ c − 1 ∥ B w ∥ for every w , (41)
since c ∥ w ∥ 2 ≤ Re ⟨ B w , w ⟩ ≤ ∥ B w ∥ ∥ w ∥ c\lVert w\rVert^{2}\leq\operatorname{Re}\langle Bw,w\rangle\leq\lVert Bw\rVert\lVert w\rVert c ∥ w ∥ 2 ≤ Re ⟨ B w , w ⟩ ≤ ∥ B w ∥ ∥ w ∥ . Applied to A 1 u = V N ∂ ˉ H A_{1}u=V_{N}\bar{\partial}H A 1 u = V N ∂ ˉ H , it gives a ∥ u ∥ ≤ ∥ V N ∂ ˉ H ∥ a\lVert u\rVert\leq\lVert V_{N}\bar{\partial}H\rVert a ∥ u ∥ ≤ ∥ V N ∂ ˉ H ∥ . We will use this principle again for the weighted solution.
Proposition 5.2 (Helffer–Sjöstrand equation). There exists ϕ ∈ Dom L \phi\in\operatorname{Dom}\mathcal{L} ϕ ∈ Dom L such that L ϕ = V N H \mathcal{L}\phi=V_{N}H L ϕ = V N H . In the sense of quadratic forms, the vector u : = ∂ ˉ ϕ u:=\bar{\partial}\phi u := ∂ ˉ ϕ solves
A 1 u = V N ∂ ˉ H . (42) A_{1}u=V_{N}\bar{\partial}H. \tag*{(42)} A 1 u = V N ∂ ˉ H . (42) For every real cylinder observable G G G , we have
K ε Cov G ε [ F , G ] = Re ⟨ u , V N ∂ ˉ G ⟩ . (43) K_{\varepsilon}\operatorname{Cov}_{G_{\varepsilon}}[F,G]=\operatorname{Re}\langle u,V_{N}\bar{\partial}G\rangle. \tag*{(43)} K ε Cov G ε [ F , G ] = Re ⟨ u , V N ∂ ˉ G ⟩ . (43) Moreover, A 1 ≥ a A_{1}\geq a A 1 ≥ a in the sense of quadratic forms, and
a ∥ u ∥ ≤ ∥ V N ∂ ˉ H ∥ , ∥ V N ∂ ˉ H ∥ 2 ≤ K ε D G ε ( F ) . a\lVert u\rVert\leq\lVert V_{N}\bar{\partial}H\rVert,\qquad\lVert V_{N}\bar{\partial}H\rVert^{2}\leq K_{\varepsilon}D_{G_{\varepsilon}}(F). a ∥ u ∥ ≤ ∥ V N ∂ ˉ H ∥ , ∥ V N ∂ ˉ H ∥ 2 ≤ K ε D G ε ( F ) .
The proof is given in Appendix A.§ 5 .
The deformation of Combes and Thomas
We now take F F F to be B R B_{R} B R -local, with B R B_{R} B R centred at the origin. To bound the covariance (43) with an observable G G G at distance d d d from B R B_{R} B R , we show that each component u k u_{k} u k decays exponentially, in L 2 L^{2} L 2 , as the k k k -th particle moves away from B R B_{R} B R .
The derivative ∂ ˉ k F \bar{\partial}_{k}F ∂ ˉ k F vanishes when the k k k -th particle is outside B R B_{R} B R . Since rotations preserve this disk, the same holds for ∂ ˉ k H \bar{\partial}_{k}H ∂ ˉ k H , and therefore
V N ( Z ) ∂ ˉ k H ( Z ) = 0 when ∣ z k ∣ ≥ R . V_{N}(Z)\bar{\partial}_{k}H(Z)=0\qquad\text{when }|z_{k}|\geq R. V N ( Z ) ∂ ˉ k H ( Z ) = 0 when ∣ z k ∣ ≥ R .
As in Subsection 3.2, we multiply u k u_{k} u k by an exponential growing with the distance from z k z_{k} z k to B R B_{R} B R . For ℓ > 0 \ell>0 ℓ > 0 , set
ζ ( z ) : = min { dist ( z , B R ) , d } , M k ( Z ) : = e ζ ( z k ) / ℓ , M : = diag ( M 1 , … , M N ) , v : = M u . \zeta(z):=\min\{\operatorname{dist}(z,B_{R}),d\},\qquad M_{k}(Z):=e^{\zeta(z_{k})/\ell},\qquad M:=\operatorname{diag}(M_{1},\ldots,M_{N}),\qquad v:=Mu. ζ ( z ) := min { dist ( z , B R ) , d } , M k ( Z ) := e ζ ( z k ) / ℓ , M := diag ( M 1 , … , M N ) , v := M u .
We cap the distance at d d d so that v v v belongs to L 2 L^{2} L 2 before we know anything about the decay of u u u .
The weight has two properties: M k = 1 M_{k}=1 M k = 1 where ∂ ˉ k H ≠ 0 \bar{\partial}_{k}H\neq0 ∂ ˉ k H = 0 , and M k = e d / ℓ M_{k}=e^{d/\ell} M k = e d / ℓ where ∂ ˉ k G ≠ 0 \bar{\partial}_{k}G\neq0 ∂ ˉ k G = 0 , if G G G is local to a set at distance d d d from B R B_{R} B R . Hence, by (43) ,
K ε ∣ Cov G ε [ F , G ] ∣ = ∣ Re ⟨ M u , M − 1 V N ∂ ˉ G ⟩ ∣ ≤ e − d / ℓ ∥ M u ∥ ∥ V N ∂ ˉ G ∥ , K_{\varepsilon}\lvert\operatorname{Cov}_{G_{\varepsilon}}[F,G]\rvert=\lvert\operatorname{Re}\langle Mu,M^{-1}V_{N}\bar{\partial}G\rangle\rvert\leq e^{-d/\ell}\lVert Mu\rVert\lVert V_{N}\bar{\partial}G\rVert, K ε ∣ Cov G ε [ F , G ]∣ = ∣ Re ⟨ M u , M − 1 V N ∂ ˉ G ⟩∣ ≤ e − d / ℓ ∥ M u ∥ ∥ V N ∂ ˉ G ∥ ,
and it remains to bound ∥ M u ∥ \lVert Mu\rVert ∥ M u ∥ by ∥ V N ∂ ˉ H ∥ \lVert V_{N}\bar{\partial}H\rVert ∥ V N ∂ ˉ H ∥ , with a constant independent of N N N . The only information we have on u u u is the Helffer–Sjöstrand equation, and the only bound on its solution comes from the coercivity of A 1 A_{1} A 1 . We therefore proceed in two steps, as for the band matrix of Lemma 3.1.
The operator A 1 + P A_1+P A 1 + P is still coercive, with the constant a − 1 4 ℓ 2 a-\frac{1}{4\ell^2} a − 4 ℓ 2 1 , up to an error term supported near the collisions (Lemma 5.4 ), which vanishes when the regularization is removed (Lemma 5.5 ).
The principle (41) then bounds ∥ M u ∥ \lVert Mu\rVert ∥ M u ∥ .
We now write the equation satisfied by v v v . Each M k M_k M k equals 1 on the support of ∂ ‾ k H \overline{\partial}_kH ∂ k H , so M M M leaves V N ∂ ‾ H V_N\overline{\partial}H V N ∂ H unchanged. Multiplying (42) by M M M and substituting u = M − 1 v u=M^{-1}v u = M − 1 v gives
( A 1 + P ) v = V N ∂ ‾ H with P : = M A 1 M − 1 − A 1 . (44) (A_1+P)v=V_N\overline{\partial}H \qquad\text{with } P:=MA_1M^{-1}-A_1. \tag*{(44)} ( A 1 + P ) v = V N ∂ H with P := M A 1 M − 1 − A 1 . (44)
We have kept A 1 A_1 A 1 as the main operator, and all the terms introduced by the weight are in P P P . The differential part of A 1 A_1 A 1 , which has no analogue in Lemma 3.1, gives two terms. The off-diagonal part of the Levi matrix gives a third one, the collision term, which is the analogue of the matrix P P P of that lemma.
Lemma 5.3 (Decomposition of the perturbation). The perturbation in (44) is
P = M A 1 M − 1 − A 1 = P 1 + P 0 + P c o l l , P=MA_1M^{-1}-A_1=P_1+P_0+P_{\mathrm{coll}}, P = M A 1 M − 1 − A 1 = P 1 + P 0 + P coll , where
( P 1 v ) k = ℓ − 1 ∂ ‾ k ζ ( z k ) ∂ ‾ k v k − ℓ − 1 ∂ ‾ k ∗ ( ( ∂ ‾ k ζ ( z k ) ) v k ) , (P_1v)_k=\ell^{-1}\overline{\partial}_k\zeta(z_k)\overline{\partial}_kv_k-\ell^{-1}\overline{\partial}_k^*((\overline{\partial}_k\zeta(z_k))v_k), ( P 1 v ) k = ℓ − 1 ∂ k ζ ( z k ) ∂ k v k − ℓ − 1 ∂ k ∗ (( ∂ k ζ ( z k )) v k ) , ( P 0 v ) k = − ℓ − 2 ∣ ∂ ‾ k ζ ( z k ) ∣ 2 v k , (P_0v)_k=-\ell^{-2}\lvert\overline{\partial}_k\zeta(z_k)\rvert^2v_k, ( P 0 v ) k = − ℓ − 2 ∣ ∂ k ζ ( z k ) ∣ 2 v k , ( P c o l l v ) k = − ∑ j ≠ k ω k j , ε ( M k M j − 1 ) v j . (P_{\mathrm{coll}}v)_k=-\sum_{j\ne k}\omega_{kj,\varepsilon}\left(\frac{M_k}{M_j}-1\right)v_j. ( P coll v ) k = − j = k ∑ ω k j , ε ( M j M k − 1 ) v j . The identity holds in the sense of forms.
Proof. We first compute with smooth weights. Since M k M_k M k only depends on z k z_k z k , the product rule gives
M k ∂ ‾ k M k − 1 = ∂ ‾ k − ℓ − 1 ∂ ‾ k ζ ( z k ) , M k ∂ ‾ k ∗ M k − 1 = ∂ ‾ k ∗ + ℓ − 1 ∂ ‾ k ζ ( z k ) , M_k\overline{\partial}_kM_k^{-1}=\overline{\partial}_k-\ell^{-1}\overline{\partial}_k\zeta(z_k),\qquad M_k\overline{\partial}_k^*M_k^{-1}=\overline{\partial}_k^*+\ell^{-1}\overline{\partial}_k\zeta(z_k), M k ∂ k M k − 1 = ∂ k − ℓ − 1 ∂ k ζ ( z k ) , M k ∂ k ∗ M k − 1 = ∂ k ∗ + ℓ − 1 ∂ k ζ ( z k ) ,
while ∂ ‾ j ∗ ∂ ‾ j \overline{\partial}_j^*\overline{\partial}_j ∂ j ∗ ∂ j commutes with M k M_k M k for j ≠ k j\ne k j = k . Thus
M k ∂ ‾ k ∗ ∂ ‾ k M k − 1 v k = ∂ ‾ k ∗ ∂ ‾ k v k + ℓ − 1 ∂ ‾ k ζ ( z k ) ∂ ‾ k v k − ℓ − 1 ∂ ‾ k ∗ ( ( ∂ ‾ k ζ ( z k ) ) v k ) − ℓ − 2 ∣ ∂ ‾ k ζ ( z k ) ∣ 2 v k , M_k\overline{\partial}_k^*\overline{\partial}_kM_k^{-1}v_k
=\overline{\partial}_k^*\overline{\partial}_kv_k+\ell^{-1}\overline{\partial}_k\zeta(z_k)\overline{\partial}_kv_k
-\ell^{-1}\overline{\partial}_k^*((\overline{\partial}_k\zeta(z_k))v_k)
-\ell^{-2}\lvert\overline{\partial}_k\zeta(z_k)\rvert^2v_k, M k ∂ k ∗ ∂ k M k − 1 v k = ∂ k ∗ ∂ k v k + ℓ − 1 ∂ k ζ ( z k ) ∂ k v k − ℓ − 1 ∂ k ∗ (( ∂ k ζ ( z k )) v k ) − ℓ − 2 ∣ ∂ k ζ ( z k ) ∣ 2 v k ,
which gives P 1 + P 0 P_1+P_0 P 1 + P 0 .
For the Levi matrix C Φ ε s h C_{\Phi_\varepsilon^{\mathrm{sh}}} C Φ ε sh , the calculation is pointwise:
( M C Φ ε s h M − 1 − C Φ ε s h ) k j = ( M k M j − 1 ) ( C Φ ε s h ) k j . (MC_{\Phi_\varepsilon^{\mathrm{sh}}}M^{-1}-C_{\Phi_\varepsilon^{\mathrm{sh}}})_{kj}
=\left(\frac{M_k}{M_j}-1\right)(C_{\Phi_\varepsilon^{\mathrm{sh}}})_{kj}. ( M C Φ ε sh M − 1 − C Φ ε sh ) k j = ( M j M k − 1 ) ( C Φ ε sh ) k j .
This vanishes for j = k j=k j = k . For j ≠ k j\ne k j = k , (A.6) gives ( C Φ ε s h ) k j = − ω k j , ε (C_{\Phi_\varepsilon^{\mathrm{sh}}})_{kj}=-\omega_{kj,\varepsilon} ( C Φ ε sh ) k j = − ω k j , ε , which yields P c o l l P_{\mathrm{coll}} P coll .
At fixed ε , d \varepsilon,d ε , d , the bounded Lipschitz multipliers M M M and M − 1 M^{-1} M − 1 preserve the form domain. Smooth approximation gives the identity for ζ \zeta ζ .
The weighted energy estimate
We can now try the coercivity estimate again, this time for v v v . As in the proof of Lemma 3.1, only the symmetric part of P P P will contribute. This leaves a zero-order term and the collision term
B ε ( u ) : = ∑ i < j ∫ ω i j , ε ( M i − M j ) 2 Re ( u i u j ‾ ) e − Φ ε s h d Z . B_\varepsilon(u):=\sum_{i<j}\int\omega_{ij,\varepsilon}(M_i-M_j)^2\operatorname{Re}(u_i\overline{u_j})e^{-\Phi_\varepsilon^{\mathrm{sh}}}\,dZ. B ε ( u ) := i < j ∑ ∫ ω ij , ε ( M i − M j ) 2 Re ( u i u j ) e − Φ ε sh d Z .
Lemma 5.4 (Weighted energy estimate). For every ℓ > 0 \ell> 0 ℓ > 0 and d ≥ 0 d \ge0 d ≥ 0 , the solution u u u of (42) satisfies
( a − 1 4 ℓ 2 ) ∥ M u ∥ 2 ≤ ∥ V N ∂ ˉ H ∥ ∥ M u ∥ + ∣ B ε ( u ) ∣ . (45) \left(a-\frac{1}{4\ell^2}\right)\lVert Mu\rVert^2 \le\lVert V_N\bar{\partial}H\rVert\lVert Mu\rVert+\lvert B_\varepsilon(u)\rvert. \tag*{(45)} ( a − 4 ℓ 2 1 ) ∥ M u ∥ 2 ≤ ∥ V N ∂ ˉ H ∥ ∥ M u ∥ + ∣ B ε ( u )∣ . (45) In particular, if ℓ > ℓ D \ell> \ell_D ℓ > ℓ D , then
∥ M u ∥ ≤ ∥ V N ∂ ˉ H ∥ a − 1 4 ℓ 2 + ( ∣ B ε ( u ) ∣ a − 1 4 ℓ 2 ) 1 / 2 . (46) \lVert Mu\rVert\le\frac{\lVert V_N\bar{\partial}H\rVert}{a-\frac{1}{4\ell^2}}+\left(\frac{\lvert B_\varepsilon(u)\rvert}{a-\frac{1}{4\ell^2}}\right)^{1/2}. \tag*{(46)} ∥ M u ∥ ≤ a − 4 ℓ 2 1 ∥ V N ∂ ˉ H ∥ + ( a − 4 ℓ 2 1 ∣ B ε ( u )∣ ) 1/2 . (46)
Proof. We compute the real part of ⟨ P v , v ⟩ \langle Pv,v\rangle ⟨ P v , v ⟩ , using the three terms of Lemma 5.3.
The adjoint of v ↦ ( ℓ − 1 ∂ ˉ k ζ ( z k ) ∂ ˉ k v k ) k v\mapsto(\ell^{-1}\bar{\partial}_k\zeta(z_k)\bar{\partial}_kv_k)_k v ↦ ( ℓ − 1 ∂ ˉ k ζ ( z k ) ∂ ˉ k v k ) k is v ↦ ( ℓ − 1 ∂ ˉ k ∗ ( ( ∂ ˉ k ζ ( z k ) ) v k ) ) k v\mapsto(\ell^{-1}\bar{\partial}_k^*((\bar{\partial}_k\zeta(z_k))v_k))_k v ↦ ( ℓ − 1 ∂ ˉ k ∗ (( ∂ ˉ k ζ ( z k )) v k ) ) k , so P 1 ∗ = − P 1 P_1^*=-P_1 P 1 ∗ = − P 1 and the first-order term does not contribute.
The zero-order term lowers the coercivity constant. Since ∣ ∂ ˉ k ζ ( z k ) ∣ 2 = 1 4 ∣ ∇ ζ ( z k ) ∣ 2 ≤ 1 4 \lvert\bar{\partial}_k\zeta(z_k)\rvert^2=\frac{1}{4}\lvert\nabla\zeta(z_k)\rvert^2\le\frac{1}{4} ∣ ∂ ˉ k ζ ( z k ) ∣ 2 = 4 1 ∣ ∇ ζ ( z k ) ∣ 2 ≤ 4 1 , we have
⟨ P 0 v , v ⟩ = − ℓ − 2 ∑ k ∥ ( ∂ ˉ k ζ ( z k ) ) v k ∥ 2 ≥ − 1 4 ℓ 2 ∥ v ∥ 2 . \langle P_0v,v\rangle=-\ell^{-2}\sum_k\lVert(\bar{\partial}_k\zeta(z_k))v_k\rVert^2\ge-\frac{1}{4\ell^2}\lVert v\rVert^2. ⟨ P 0 v , v ⟩ = − ℓ − 2 k ∑ ∥( ∂ ˉ k ζ ( z k )) v k ∥ 2 ≥ − 4 ℓ 2 1 ∥ v ∥ 2 .
The collision term gives the error. We group the pairs ( i , j ) (i,j) ( i , j ) and ( j , i ) (j,i) ( j , i ) , using ω i j , ε = ω j i , ε \omega_{ij,\varepsilon}=\omega_{ji,\varepsilon} ω ij , ε = ω j i , ε . Since v k = M k u k v_k=M_ku_k v k = M k u k , we have
( M i M j + M j M i − 2 ) Re ( v i v ˉ j ) = ( M i − M j ) 2 Re ( u i u ˉ j ) , \left(\frac{M_i}{M_j}+\frac{M_j}{M_i}-2\right)\operatorname{Re}(v_i\bar{v}_j)=(M_i-M_j)^2\operatorname{Re}(u_i\bar{u}_j), ( M j M i + M i M j − 2 ) Re ( v i v ˉ j ) = ( M i − M j ) 2 Re ( u i u ˉ j ) ,
and therefore Re ⟨ P c o l l v , v ⟩ = − B ε ( u ) \operatorname{Re}\langle P_{\mathrm{coll}}v,v\rangle=-B_\varepsilon(u) Re ⟨ P coll v , v ⟩ = − B ε ( u ) .
Since A 1 ≥ a A_1\ge a A 1 ≥ a , pairing (44) with v v v and taking real parts gives
( a − 1 4 ℓ 2 ) ∥ v ∥ 2 ≤ Re ⟨ V N ∂ ˉ H , v ⟩ + B ε ( u ) . \left(a-\frac{1}{4\ell^2}\right)\lVert v\rVert^2\le\operatorname{Re}\langle V_N\bar{\partial}H,v\rangle+B_\varepsilon(u). ( a − 4 ℓ 2 1 ) ∥ v ∥ 2 ≤ Re ⟨ V N ∂ ˉ H , v ⟩ + B ε ( u ) .
The Cauchy–Schwarz inequality then gives (45) , and solving this quadratic inequality in ∥ M u ∥ \lVert Mu\rVert ∥ M u ∥ gives (46) .
:::
Control of the collision term
At an exact collision, M i = M j M_i=M_j M i = M j . We therefore expect the error in (45) to disappear when the interaction is no longer smoothed. The point to check is that the solutions do not concentrate near collisions as ε → 0 \varepsilon\to0 ε → 0 .
Lemma 5.5 (Control of the collision term). Fix N , L , λ , R , ℓ , d N,L,\lambda,R,\ell,d N , L , λ , R , ℓ , d , with λ , ℓ > 0 \lambda,\ell>0 λ , ℓ > 0 and d ≥ 0 d\ge0 d ≥ 0 , and a B R B_R B R -local cylinder observable F F F . Let u ε u_\varepsilon u ε be the solution of (42) associated with F F F . Then
B ε ( u ε ) ⟶ ε → 0 0. B_\varepsilon(u_\varepsilon)\longrightarrow_{\varepsilon\to0}0. B ε ( u ε ) ⟶ ε → 0 0.
The proof is given in Appendix § 13 .§ 5 . It uses the permutation symmetry of the solution to exclude concentration on the collision diagonals. When β = 2 \beta=2 β = 2 , the collision term is identically zero.
Remark 5.6. In the deformation of Combes and Thomas, the weight produces a commutator with the matrix of the equation, which is large when the entries of this matrix decay slowly. For the circular Riesz gas with repulsion ∣ x ∣ − s \lvert x\rvert^{-s} ∣ x ∣ − s [20 ] , this matrix is the Hessian in gap coordinates. When the gaps are close to 1 1 1 , its entries are of order ( 1 + r ) − s (1+r)^{-s} ( 1 + r ) − s for two gaps r r r labels apart, so an exponential weight would give a commutator with entries that grow with r r r . The weight there is a power ( 1 + r ) α (1+r)^\alpha ( 1 + r ) α of the distance r r r to the observable, with an exponent tuned to s s s . The commutator is small for 1 / 2 < α < s − 1 / 2 1/2<\alpha<s-1/2 1/2 < α < s − 1/2 [20 Lemma 4.3] , and this range is empty when s ≤ 1 s\le1 s ≤ 1 .
Here, the interaction has long range, but its Levi matrix is local. Since g g g is harmonic away from the origin, the off-diagonal entries of the Levi matrix are supported on collisions, where M i = M j M_i=M_j M i = M j . The weight can therefore be exponential, and the only restriction on its rate comes from the zero-order term, as in Lemma 5.4 . After regularization, the square ( M i − M j ) 2 (M_i-M_j)^2 ( M i − M j ) 2 compensates for the size ε − 2 \varepsilon^{-2} ε − 2 of the mollifier, which only makes the coefficient bounded. The non-concentration of u ε u_\varepsilon u ε then makes the integral vanish.
Proof of Proposition 5.1 (smooth observables)
Proof. We first take F F F to be B R B_R B R -local and G G G to be Ω \Omega Ω -local, with dist ( B R , Ω ) = d \operatorname{dist}(B_R,\Omega)=d dist ( B R , Ω ) = d . Fix ℓ > ℓ D \ell>\ell_D ℓ > ℓ D . The anti-analytic-signal estimate and the identity ∣ ∂ ˉ k G ∣ 2 = 1 4 ∣ ∇ k G ∣ 2 \lvert\bar{\partial}_kG\rvert^2=\frac{1}{4}\lvert\nabla_kG\rvert^2 ∣ ∂ ˉ k G ∣ 2 = 4 1 ∣ ∇ k G ∣ 2 give
∥ V N ∂ ˉ H ∥ 2 ≤ K ε D G ε ( F ) , ∥ V N ∂ ˉ G ∥ 2 = K ε 4 D G ε ( G ) . \lVert V_N\bar{\partial}H\rVert^2\le K_\varepsilon\mathcal{D}_{G_\varepsilon}(F),\qquad\lVert V_N\bar{\partial}G\rVert^2=\frac{K_\varepsilon}{4}\mathcal{D}_{G_\varepsilon}(G). ∥ V N ∂ ˉ H ∥ 2 ≤ K ε D G ε ( F ) , ∥ V N ∂ ˉ G ∥ 2 = 4 K ε D G ε ( G ) .
Since ∂ ˉ k G \bar{\partial}_kG ∂ ˉ k G vanishes outside Ω \Omega Ω , the inverse weight is at most e − d / ℓ e^{-d/\ell} e − d / ℓ on its support. The covariance formula gives
K ε ∣ Cov G ε [ F , G ] ∣ = ∣ Re ⟨ M u ε , M − 1 V N ∂ ˉ G ⟩ ∣ ≤ e − d / ℓ ∥ M u ε ∥ ∥ V N ∂ ˉ G ∥ . K_\varepsilon\lvert\operatorname{Cov}_{G_\varepsilon}[F,G]\rvert=\left\lvert\operatorname{Re}\langle Mu_\varepsilon,M^{-1}V_N\bar{\partial}G\rangle\right\rvert\le e^{-d/\ell}\lVert Mu_\varepsilon\rVert\lVert V_N\bar{\partial}G\rVert. K ε ∣ Cov G ε [ F , G ]∣ = Re ⟨ M u ε , M − 1 V N ∂ ˉ G ⟩ ≤ e − d / ℓ ∥ M u ε ∥ ∥ V N ∂ ˉ G ∥ .
We insert (46) , and then let ε → 0 \varepsilon\to0 ε → 0 at fixed N N N , L L L , λ \lambda λ , ℓ \ell ℓ , d d d , using Lemma 5.5 , and let λ → ∞ \lambda\to\infty λ → ∞ , using (A.5). We obtain
∣ Cov G N , L [ F , G ] ∣ ≤ e − d / ℓ 2 ( a − 1 4 ℓ 2 ) D G N , L ( F ) 1 / 2 D G N , L ( G ) 1 / 2 . (47) \lvert\operatorname{Cov}_{G_{N,L}}[F,G]\rvert\le\frac{e^{-d/\ell}}{2\left(a-\frac{1}{4\ell^2}\right)}\mathcal{D}_{G_{N,L}}(F)^{1/2}\mathcal{D}_{G_{N,L}}(G)^{1/2}. \tag*{(47)} ∣ Cov G N , L [ F , G ]∣ ≤ 2 ( a − 4 ℓ 2 1 ) e − d / ℓ D G N , L ( F ) 1/2 D G N , L ( G ) 1/2 . (47)
The constant is independent of N N N and L L L .
We take L = R N L=R_N L = R N and pass along a subsequence whose local configurations converge to P. Covariances converge because cylinder observables are bounded and continuous. Their Dirichlet energies converge by Proposition 2.4 : for a cylinder F F F , the function ∑ z ∈ X ∣ ∇ z F ( X ) ∣ 2 \sum_{z\in X}\lvert\nabla_zF(X)\rvert^2 ∑ z ∈ X ∣ ∇ z F ( X ) ∣ 2 is local and continuous, and is bounded by C ( F ) N K C(F)N_K C ( F ) N K for a fixed compact set K K K . Thus (47) passes to P.
Finally, P is stationary by part (ii) of Proposition 2.4 . Translating both observables gives the estimate for a disk centred at every x ∈ R 2 x\in\mathbb{R}^2 x ∈ R 2 , with the stated constant C ( β , ℓ ) C(\beta,\ell) C ( β , ℓ ) .
From smooth observables to the two-point function
Proposition 5.1 controls covariances of smooth observables through their Dirichlet energies. A pointwise bound on ρ 2 − 1 \rho_2-1 ρ 2 − 1 does not follow automatically: averages of ρ 2 − 1 \rho_2-1 ρ 2 − 1 may be small while ρ 2 − 1 \rho_2-1 ρ 2 − 1 oscillates. This happens in the one-dimensional log-gas, whose truncated two-point function has an oscillating part, with the mean spacing as period, that decays more slowly than its averages for β > 2 \beta>2 β > 2 , as predicted by Forrester [38 ] .
Two facts allow us to pass to pointwise values. First, Proposition 5.1 holds at every scale: concentrating the test functions at scale δ \delta δ only costs a factor δ − 4 \delta^{-4} δ − 4 . Second, ρ 2 \rho_2 ρ 2 is smooth away from the diagonal, with uniform bounds on all its derivatives (Lemma 5.9 ). Averaging ρ 2 − 1 \rho_2-1 ρ 2 − 1 against test functions at scale δ \delta δ whose moments vanish up to order k − 1 k-1 k − 1 then gives its pointwise value up to O ( δ k ) O(\delta^k) O ( δ k ) . Choosing δ \delta δ exponentially small in the distance balances the two errors.
Let P be as in Proposition 5.1. By part (ii) of Proposition 2.4 , the process P is stationary with intensity 1.
Proposition 5.7 (Pointwise decay of the two-point function). The two-point function of P has a continuous version ρ 2 \rho_2 ρ 2 such that, for every ℓ > ℓ D \ell>\ell_D ℓ > ℓ D , there is C ( P , β , ℓ ) < ∞ C(P,\beta,\ell)<\infty C ( P , β , ℓ ) < ∞ satisfying
∣ ρ 2 ( x , y ) − 1 ∣ ≤ C ( P , β , ℓ ) e − ∣ x − y ∣ / ℓ for all x , y ∈ R 2 . \lvert\rho_2(x,y)-1\rvert\le C(P,\beta,\ell)e^{-\lvert x-y\rvert/\ell}\qquad\text{for all }x,y\in\mathbb{R}^2. ∣ ρ 2 ( x , y ) − 1 ∣ ≤ C ( P , β , ℓ ) e − ∣ x − y ∣ / ℓ for all x , y ∈ R 2 .
Lemma 5.8 (Regularity of the two-point function). The two-point function of P has a bounded continuous version, and there is C ( P , β ) < ∞ C(P,\beta)<\infty C ( P , β ) < ∞ such that
∣ ρ 2 ( x 1 , y 1 ) − ρ 2 ( x 2 , y 2 ) ∣ ≤ C ( P , β ) ( ∣ x 1 − x 2 ∣ min ( β , 1 ) + ∣ y 1 − y 2 ∣ min ( β , 1 ) ) \lvert\rho_2(x_1,y_1)-\rho_2(x_2,y_2)\rvert\le C(P,\beta)\left(\lvert x_1-x_2\rvert^{\min(\beta,1)}+\lvert y_1-y_2\rvert^{\min(\beta,1)}\right) ∣ ρ 2 ( x 1 , y 1 ) − ρ 2 ( x 2 , y 2 )∣ ≤ C ( P , β ) ( ∣ x 1 − x 2 ∣ m i n ( β , 1 ) + ∣ y 1 − y 2 ∣ m i n ( β , 1 ) ) whenever x 1 , y 1 , x 2 , y 2 ∈ R 2 x_1,y_1,x_2,y_2\in\mathbb{R}^2 x 1 , y 1 , x 2 , y 2 ∈ R 2 and ∣ x 1 − x 2 ∣ , ∣ y 1 − y 2 ∣ ≤ 1 \lvert x_1-x_2\rvert,\lvert y_1-y_2\rvert\le1 ∣ x 1 − x 2 ∣ , ∣ y 1 − y 2 ∣ ≤ 1 .
The proof of Lemma 5.8 is given in Appendix § 13 .§ 6 .
Lemma 5.9 (Smoothness away from the diagonal). The function ρ 2 \rho_2 ρ 2 is C ∞ C^\infty C ∞ on { ( x , y ) : x ≠ y } \{(x,y):x\ne y\} {( x , y ) : x = y } . For every integer m ≥ 1 m\ge1 m ≥ 1 and d 0 > 0 d_0>0 d 0 > 0 , there is C ( P , β , m , d 0 ) < ∞ C(P,\beta,m,d_0)<\infty C ( P , β , m , d 0 ) < ∞ such that
sup ∣ x − y ∣ ≥ d 0 ∣ ∇ m ρ 2 ( x , y ) ∣ ≤ C ( P , β , m , d 0 ) . \sup_{\lvert x-y\rvert\ge d_0}\lvert\nabla^m\rho_2(x,y)\rvert\le C(P,\beta,m,d_0). ∣ x − y ∣ ≥ d 0 sup ∣ ∇ m ρ 2 ( x , y )∣ ≤ C ( P , β , m , d 0 ) .
The proof, using the DLR equations, (G1) and (G2), is given in Appendix § 13 .§ 6 .
Proof of Proposition 5.7 . Fix ℓ > ℓ D \ell>\ell_D ℓ > ℓ D . Let k ≥ 1 k\ge1 k ≥ 1 be an integer such that ℓ 1 : = k ℓ / ( k + 4 ) > ℓ D \ell_1:=k\ell/(k+4)>\ell_D ℓ 1 := k ℓ / ( k + 4 ) > ℓ D .
Step 1. Linear statistics. Proposition 5.1 is stated for cylinder observables, which are bounded. We extend it to linear statistics X ( f ) X(f) X ( f ) by truncation. For f ∈ C ∞ ( R 2 ; R ) f \in C^\infty(\mathbb{R}^2;\mathbb{R}) f ∈ C ∞ ( R 2 ; R ) , we take truncations T m ∈ C ∞ ( R ) T_m \in C^\infty(\mathbb{R}) T m ∈ C ∞ ( R ) with ∣ ∂ t T m ∣ ≤ 1 |\partial_tT_m| \le1 ∣ ∂ t T m ∣ ≤ 1 , ∣ T m ( t ) ∣ ≤ ∣ t ∣ |T_m(t)| \le|t| ∣ T m ( t ) ∣ ≤ ∣ t ∣ and T m ( t ) → m → ∞ t T_m(t) \xrightarrow[m\to\infty]{} t T m ( t ) m → ∞ t . The local second moments give T m ( X ( f ) ) → m → ∞ X ( f ) T_m(X(f)) \xrightarrow[m\to\infty]{} X(f) T m ( X ( f )) m → ∞ X ( f ) in L 2 ( P ) L^2(\mathbb{P}) L 2 ( P ) . Since ∣ ∂ t T m ∣ ≤ 1 |\partial_tT_m| \le1 ∣ ∂ t T m ∣ ≤ 1 , the Dirichlet energy of T m ( X ( f ) ) T_m(X(f)) T m ( X ( f )) is at most E P ∑ z ∈ X ∣ ∇ f ( z ) ∣ 2 \mathbb{E}_{\mathbb{P}}\sum_{z\in X}|\nabla f(z)|^2 E P ∑ z ∈ X ∣∇ f ( z ) ∣ 2 , which Campbell’s formula computes, P \mathbb{P} P having intensity 1 1 1 :
D P ( T m ( X ( f ) ) ) ≤ E P ∑ z ∈ X ∣ ∇ f ( z ) ∣ 2 = ∥ ∇ f ∥ 2 2 . \mathcal{D}_{\mathbb{P}}(T_m(X(f))) \le\mathbb{E}_{\mathbb{P}}\sum_{z\in X}|\nabla f(z)|^2 = \|\nabla f\|_2^2. D P ( T m ( X ( f ))) ≤ E P z ∈ X ∑ ∣∇ f ( z ) ∣ 2 = ∥∇ f ∥ 2 2 .
Applying Proposition 5.1 to both truncated statistics and letting m → ∞ m \to\infty m → ∞ yields
∣ Cov P [ X ( f ) , X ( g ) ] ∣ ≤ C ( β , ℓ 1 ) e − d / ℓ 1 ∥ ∇ f ∥ 2 ∥ ∇ g ∥ 2 , (48) |\operatorname{Cov}_{\mathbb{P}}[X(f),X(g)]| \le C(\beta,\ell_1)e^{-d/\ell_1}\|\nabla f\|_2\|\nabla g\|_2, \tag*{(48)} ∣ Cov P [ X ( f ) , X ( g )] ∣ ≤ C ( β , ℓ 1 ) e − d / ℓ 1 ∥∇ f ∥ 2 ∥∇ g ∥ 2 , (48)
for f , g ∈ C ∞ ( R 2 ; R ) f,g \in C^\infty(\mathbb{R}^2;\mathbb{R}) f , g ∈ C ∞ ( R 2 ; R ) , when f f f is supported in a disk at distance d d d from the support of g g g .
Step 2. From averages to values. When the supports of f f f and g g g are disjoint, Cov P [ X ( f ) , X ( g ) ] \operatorname{Cov}_{\mathbb{P}}[X(f),X(g)] Cov P [ X ( f ) , X ( g )] is the average of ρ 2 − 1 \rho_2-1 ρ 2 − 1 against f ⊗ g f\otimes g f ⊗ g . We choose f f f and g g g concentrated at scale δ \delta δ around x x x and y y y , and recover the value of ρ 2 − 1 \rho_2-1 ρ 2 − 1 at ( x , y ) (x,y) ( x , y ) by an elementary interpolation: an average at scale δ \delta δ and a bound on the derivatives of order k k k determine the value at the centre up to O ( δ k ) O(\delta^k) O ( δ k ) . Let η ∈ C c ∞ ( B 1 ) \eta\in C^\infty_c(B_1) η ∈ C c ∞ ( B 1 ) be real and such that
∫ η = 1 , ∫ z α η ( z ) d z = 0 for 1 ≤ ∣ α ∣ < k . \int\eta=1,\qquad\int z^\alpha\eta(z)\,\mathrm{d}z=0\quad\text{for }1\le|\alpha|<k. ∫ η = 1 , ∫ z α η ( z ) d z = 0 for 1 ≤ ∣ α ∣ < k .
Put η δ , x ( z ) : = δ − 2 η ( ( z − x ) / δ ) \eta_{\delta,x}(z):=\delta^{-2}\eta((z-x)/\delta) η δ , x ( z ) := δ − 2 η (( z − x ) / δ ) . Then ∥ ∇ η δ , x ∥ 2 = δ − 1 ∥ ∇ η ∥ 2 \|\nabla\eta_{\delta,x}\|_2=\delta^{-1}\|\nabla\eta\|_2 ∥∇ η δ , x ∥ 2 = δ − 1 ∥∇ η ∥ 2 .
Claim 5.10. Let x , y ∈ R 2 x,y\in\mathbb{R}^2 x , y ∈ R 2 , 0 < δ ≤ 1 0<\delta\le1 0 < δ ≤ 1 , and let h h h be a C k C^k C k function on a neighbourhood of B ‾ ( x , δ ) × B ‾ ( y , δ ) \overline{B}(x,\delta)\times\overline{B}(y,\delta) B ( x , δ ) × B ( y , δ ) . Then
∣ h ( x , y ) − ∬ η δ , x ( u ) η δ , y ( v ) h ( u , v ) d u d v ∣ ≤ C ( η , k ) δ k sup B ‾ ( x , δ ) × B ‾ ( y , δ ) ∣ ∇ k h ∣ . \left|h(x,y)-\iint\eta_{\delta,x}(u)\eta_{\delta,y}(v)h(u,v)\,\mathrm{d}u\,\mathrm{d}v\right|
\le C(\eta,k)\delta^k\sup_{\overline{B}(x,\delta)\times\overline{B}(y,\delta)}|\nabla^k h|. h ( x , y ) − ∬ η δ , x ( u ) η δ , y ( v ) h ( u , v ) d u d v ≤ C ( η , k ) δ k B ( x , δ ) × B ( y , δ ) sup ∣ ∇ k h ∣.
Proof. The function η δ , x ⊗ η δ , y \eta_{\delta,x}\otimes\eta_{\delta,y} η δ , x ⊗ η δ , y on R 4 \mathbb{R}^4 R 4 has integral $1$, and its moments ∫ ( w − ( x , y ) ) α ( η δ , x ⊗ η δ , y ) ( w ) d w \int(w-(x,y))^\alpha(\eta_{\delta,x}\otimes\eta_{\delta,y})(w)\,\mathrm{d}w ∫ ( w − ( x , y ) ) α ( η δ , x ⊗ η δ , y ) ( w ) d w vanish for 1 ≤ ∣ α ∣ < k 1\le|\alpha|<k 1 ≤ ∣ α ∣ < k : each such moment is a product of a moment of η δ , x \eta_{\delta,x} η δ , x and a moment of η δ , y \eta_{\delta,y} η δ , y , one of which has order between 1 1 1 and $k-1. We write Taylor’s formula for h h h at ( x , y ) (x,y) ( x , y ) , to order k − 1 k-1 k − 1 with a remainder of order k k k , on the convex set B ‾ ( x , δ ) × B ‾ ( y , δ ) \overline{B}(x,\delta)\times\overline{B}(y,\delta) B ( x , δ ) × B ( y , δ ) , and integrate it against η δ , x ⊗ η δ , y \eta_{\delta,x}\otimes\eta_{\delta,y} η δ , x ⊗ η δ , y . The constant term gives h ( x , y ) h(x,y) h ( x , y ) , the terms of degree between 1 1 1 and k − 1 k-1 k − 1 vanish, and the remainder is at most C δ k sup ∣ ∇ k h ∣ ∥ η ∥ 1 2 C\delta^k\sup|\nabla^k h|\|\eta\|_1^2 C δ k sup ∣ ∇ k h ∣∥ η ∥ 1 2 . □
For r : = ∣ x − y ∣ ≥ 2 r:=|x-y|\ge2 r := ∣ x − y ∣ ≥ 2 and 0 < δ ≤ 1 / 4 0<\delta\le1/4 0 < δ ≤ 1/4 , the supports of η δ , x \eta_{\delta,x} η δ , x and η δ , y \eta_{\delta,y} η δ , y are disjoint, so (48) and the definition of ρ 2 \rho_2 ρ 2 give
∣ ∬ η δ , x ( u ) η δ , y ( v ) ( ρ 2 ( u , v ) − 1 ) d u d v ∣ = ∣ Cov P [ X ( η δ , x ) , X ( η δ , y ) ] ∣ ≤ C ( β , ℓ 1 ) δ − 4 ∥ ∇ η ∥ 2 2 e − ( r − 2 δ ) / ℓ 1 . \left|\iint\eta_{\delta,x}(u)\eta_{\delta,y}(v)(\rho_2(u,v)-1)\,\mathrm{d}u\,\mathrm{d}v\right|
=\left|\operatorname{Cov}_{\mathbb{P}}[X(\eta_{\delta,x}),X(\eta_{\delta,y})]\right|
\le C(\beta,\ell_1)\delta^{-4}\|\nabla\eta\|_2^2e^{-(r-2\delta)/\ell_1}. ∬ η δ , x ( u ) η δ , y ( v ) ( ρ 2 ( u , v ) − 1 ) d u d v = ∣ Cov P [ X ( η δ , x ) , X ( η δ , y )] ∣ ≤ C ( β , ℓ 1 ) δ − 4 ∥∇ η ∥ 2 2 e − ( r − 2 δ ) / ℓ 1 .
We apply Claim 5.10 to h = ρ 2 − 1 h=\rho_2-1 h = ρ 2 − 1 . On B ‾ ( x , δ ) × B ‾ ( y , δ ) \overline{B}(x,\delta)\times\overline{B}(y,\delta) B ( x , δ ) × B ( y , δ ) we have ∣ u − v ∣ ≥ r − 2 δ ≥ 1 |u-v|\ge r-2\delta\ge1 ∣ u − v ∣ ≥ r − 2 δ ≥ 1 , so Lemma 5.9 with d 0 = 1 d_0=1 d 0 = 1 bounds ∇ k ρ 2 \nabla^k\rho_2 ∇ k ρ 2 . This gives
∣ ρ 2 ( x , y ) − 1 ∣ ≤ C ( δ k + δ − 4 e − r / ℓ 1 ) . |\rho_2(x,y)-1|\le C\left(\delta^k+\delta^{-4}e^{-r/\ell_1}\right). ∣ ρ 2 ( x , y ) − 1∣ ≤ C ( δ k + δ − 4 e − r / ℓ 1 ) .
Taking δ = 1 4 e − r / ( k ℓ ) \delta=\frac14e^{-r/(k\ell)} δ = 4 1 e − r / ( k ℓ ) yields
∣ ρ 2 ( x , y ) − 1 ∣ ≤ C e − r / ℓ . |\rho_2(x,y)-1|\le Ce^{-r/\ell}. ∣ ρ 2 ( x , y ) − 1∣ ≤ C e − r / ℓ .
For r < 2 r<2 r < 2 , the bound follows by increasing C C C , since ρ 2 \rho_2 ρ 2 is bounded. □
Corollary 5.11 (The limit points are good). Every limit point of the local configurations X N X_N X N under P N , β \mathbb{P}_{N,\beta} P N , β satisfies (G3). In particular, it is a good DLR state.
Proof. Membership in the class of [[59 ] , Section 8.3], the DLR equations and stationarity follow from Proposition 2.4 . Proposition 5.7 gives (2.14)}, since ∫ R 2 ∣ log ∣ u ∣ ∣ e − c ∣ u ∣ d u < ∞ \int_{\mathbb{R}^2}|\log|u||e^{-c|u|}\,\mathrm{d}u<\infty ∫ R 2 ∣ log ∣ u ∣∣ e − c ∣ u ∣ d u < ∞ for every c > 0 c>0 c > 0 . □
Proof of Theorem 1. Proposition 5.7 gives (i), and Proposition 5.1 with Ω = B ( y , S ) \Omega=B(y,S) Ω = B ( y , S ) gives (ii). Neither uses the uniqueness of the limit point. □
A DLR variational principle: proof of Proposition 1.1
What is a DLR variational principle? There are three main ways to talk about infinite-volume Gibbs measures:
Take a finite-volume Gibbs measure and show that it has a limit (possibly after passing to a subsequence).
Write down DLR equations, as in Subsection 2.4, and ask for solutions.
Derive a free energy functional in infinite volume and ask for minimizers.
In general, (i) implies (ii), in the sense that the limit points solve the DLR equations (for the 2DOCP, this is proven in [59 ] ), but it can happen that there are exotic solutions which cannot be obtained as limits of finite-volume systems.
A DLR variational principle posits, for a given model, the equivalence between (ii) and (iii). Here we prove that every good DLR state minimizes the free energy, for every β > 0 \beta> 0 β > 0 . For 0 < β ≤ 2 0 < \beta\le2 0 < β ≤ 2 , the converse follows from uniqueness of the free energy minimizer and existence of a good DLR state: the latter is a minimizer and therefore coincides with the unique one.
The DLR variational principle has been proven for a large class of lattice spin systems [41 ] but no general result is known for continuous models with long-range interactions. The closely related case of the one-dimensional log-gas was recently addressed in [25 ] , with fairly similar arguments.
The free energy functional
The free energy functional F β \mathcal{F}_{\beta} F β is defined on stationary point processes of intensity 1 1 1 . It is the weighted sum F β : = β W + E \mathcal{F}_{\beta} := \beta\mathcal{W} + \mathcal{E} F β := β W + E of two ingredients:
We normalize W \mathcal{W} W as one half of the electric energy of [68 Section 2.7.4] , with background density 1 1 1 . Thus β W + E \beta\mathcal{W}+\mathcal{E} β W + E is the free energy of [68 (1.8)] at inverse temperature β \beta β . This convention agrees with the factor 1 2 \frac{1}{2} 2 1 in our finite-volume energy (1) . We refer to [81 ] for a presentation of the electric energy.
Intrinsic energy. The free energy F β \mathcal{F}_{\beta} F β involves the electric energy W \mathcal{W} W , defined through electric fields. Its definition is a bit involved, and we do not need it here. Indeed, what naturally comes out of the DLR equations is rather the intrinsic energy , introduced in [57 ] , and whose definition is simply a “large-scale limit of interaction between points”, namely:
W i n t ( P ) : = lim inf R → ∞ 1 ∣ C R ∣ E P [ F C R ( X C R ) ] , (49) \mathcal{W}^{\mathrm{int}}(\mathrm{P}) := \liminf_{R\to\infty} \frac{1}{|C_R|}\mathbb{E}_{\mathrm{P}}\left[F_{C_R}(X_{C_R})\right], \tag*{(49)} W int ( P ) := R → ∞ lim inf ∣ C R ∣ 1 E P [ F C R ( X C R ) ] , (49)
where C R : = [ − R 2 , R 2 ] 2 C_R := \left[-\frac{R}{2},\frac{R}{2}\right]^2 C R := [ − 2 R , 2 R ] 2 is the square of side R R R . Accordingly, W i n t \mathcal{W}^{\mathrm{int}} W int is one half of the intrinsic energy of [57 ] . The associated free energy functional is then defined as:
F β i n t ( P ) : = β W i n t ( P ) + E ( P ) . \mathcal{F}^{\mathrm{int}}_{\beta}(\mathrm{P}) := \beta\mathcal{W}^{\mathrm{int}}(\mathrm{P}) + \mathcal{E}(\mathrm{P}). F β int ( P ) := β W int ( P ) + E ( P ) .
The link between W i n t \mathcal{W}^{\mathrm{int}} W int , F β i n t \mathcal{F}^{\mathrm{int}}_{\beta} F β int and the functionals W , F β \mathcal{W},\mathcal{F}_{\beta} W , F β was studied in [57 ] . Here we only need the electric-intrinsic inequality, see Corollary 6.2 below, but in a slightly different form than in [57 ] , which is possible thanks to the following lemma.
Lemma 6.1 (Squares versus disks). Assume that P \mathrm{P} P is stationary with intensity 1 1 1 and hyperuniform, and that its two-point correlation function satisfies the integrability condition (16) .
(i) We have
lim R → ∞ 1 ∣ C R ∣ E P [ F C R ( X C R ) ] = lim R → ∞ 1 ∣ B R ∣ E P [ F B R ( X B R ) ] = 1 2 ∫ R 2 g ( u ) ( ρ 2 ( 0 , u ) − 1 ) d u . \lim_{R\to\infty} \frac{1}{|C_R|}\mathbb{E}_{\mathrm{P}}\left[F_{C_R}(X_{C_R})\right]
=
\lim_{R\to\infty} \frac{1}{|B_R|}\mathbb{E}_{\mathrm{P}}\left[F_{B_R}(X_{B_R})\right]
=
\frac{1}{2}\int_{\mathbb{R}^2} g(u)\bigl(\rho_2(0,u)-1\bigr)\,du. R → ∞ lim ∣ C R ∣ 1 E P [ F C R ( X C R ) ] = R → ∞ lim ∣ B R ∣ 1 E P [ F B R ( X B R ) ] = 2 1 ∫ R 2 g ( u ) ( ρ 2 ( 0 , u ) − 1 ) d u .
In particular, we can replace squares by disks in the definition (49) of W i n t ( P ) \mathcal{W}^{\mathrm{int}}(\mathrm{P}) W int ( P ) , as the limit is the same:
W i n t ( P ) = lim R → ∞ 1 ∣ B R ∣ E P [ F B R ( X B R ) ] . (50) \mathcal{W}^{\mathrm{int}}(\mathrm{P})=\lim_{R\to\infty}\frac{1}{|B_R|}\mathbb{E}_{\mathrm{P}}\left[F_{B_R}(X_{B_R})\right]. \tag*{(50)} W int ( P ) = R → ∞ lim ∣ B R ∣ 1 E P [ F B R ( X B R ) ] . (50)
(ii) We have, as R → ∞ R\to\infty R → ∞
Var P [ N C R ] log R = o ( ∣ C R ∣ ) . (51) \operatorname{Var}_{\mathrm{P}}[N_{C_R}]\log R=o(|C_R|). \tag*{(51)} Var P [ N C R ] log R = o ( ∣ C R ∣ ) . (51)
The proof is a direct computation, which we postpone to Appendix § 13 .§ 8 .
Corollary 6.2 (Electric-intrinsic inequality). Let P \mathrm{P} P be as in Lemma 6.1 . Then
W ( P ) ≤ lim R → ∞ 1 ∣ B R ∣ E P [ F B R ( X B R ) ] . \mathcal{W}(\mathrm{P})\leq\lim_{R\to\infty}\frac{1}{|B_R|}\mathbb{E}_{\mathrm{P}}\left[F_{B_R}(X_{B_R})\right]. W ( P ) ≤ R → ∞ lim ∣ B R ∣ 1 E P [ F B R ( X B R ) ] .
Proof. In the two-dimensional logarithmic case, [57 Theorem 1] , after division of both energies by 2, gives
W ( P ) ≤ lim inf R → ∞ 1 ∣ C R ∣ E P [ F C R ( X C R ) ] + D log ( P ) , \mathcal{W}(\mathrm{P})\leq\liminf_{R\to\infty}\frac{1}{|C_R|}\mathbb{E}_{\mathrm{P}}\left[F_{C_R}(X_{C_R})\right]+\mathcal{D}_{\log}(\mathrm{P}), W ( P ) ≤ R → ∞ lim inf ∣ C R ∣ 1 E P [ F C R ( X C R ) ] + D l o g ( P ) ,
where the first term is the intrinsic energy (49) , defined with squares, and D log \mathcal{D}_{\log} D l o g is one half of the quantity in [57 (1.8)] . It is a fixed multiple of lim sup R → ∞ log R ⋅ ∣ C R ∣ − 1 Var P [ N C R ] \limsup_{R\to\infty}\log R\cdot|C_R|^{-1}\operatorname{Var}_{\mathrm{P}}[N_{C_R}] lim sup R → ∞ log R ⋅ ∣ C R ∣ − 1 Var P [ N C R ] . By (51) , D log ( P ) = 0 \mathcal{D}_{\log}(\mathrm{P})=0 D l o g ( P ) = 0 , and by part (i) of Lemma 6.1 , the liminf over squares is the same as the limit over disks (50) .
Relative entropy. For two probability measures μ , ν \mu,\nu μ , ν on the same space, the relative entropy of μ \mu μ with respect to ν \nu ν is
Ent ( μ ∣ ν ) : = ∫ log d μ d ν d μ if μ ≪ ν , Ent ( μ ∣ ν ) : = + ∞ otherwise . \operatorname{Ent}(\mu\mid\nu):=\int\log\frac{\mathrm{d}\mu}{\mathrm{d}\nu}\,\mathrm{d}\mu\quad\text{if }\mu\ll\nu,\qquad\operatorname{Ent}(\mu\mid\nu):=+\infty\quad\text{otherwise}. Ent ( μ ∣ ν ) := ∫ log d ν d μ d μ if μ ≪ ν , Ent ( μ ∣ ν ) := + ∞ otherwise .
We use the classical properties of this quantity (non-negativity, convexity, monotonicity under taking marginals), see for example [42 Chapter 15] , and the chain rule : if μ , ν \mu,\nu μ , ν are laws of a pair ( Y , Z ) (Y,Z) ( Y , Z ) , with marginals μ Y , ν Y \mu_Y,\nu_Y μ Y , ν Y for Y Y Y and conditional laws μ ( ⋅ ∣ Y ) , ν ( ⋅ ∣ Y ) \mu(\mathord{\cdot}\mid Y),\nu(\mathord{\cdot}\mid Y) μ ( ⋅ ∣ Y ) , ν ( ⋅ ∣ Y ) for Z Z Z , then
Ent ( μ ∣ ν ) = Ent ( μ Y ∣ ν Y ) + ∫ Ent ( μ ( ⋅ ∣ y ) ∣ ν ( ⋅ ∣ y ) ) d μ Y ( y ) . \operatorname{Ent}(\mu\mid\nu)=\operatorname{Ent}(\mu_Y\mid\nu_Y)+\int\operatorname{Ent}(\mu(\mathord{\cdot}\mid y)\mid\nu(\mathord{\cdot}\mid y))\,\mathrm{d}\mu_Y(y). Ent ( μ ∣ ν ) = Ent ( μ Y ∣ ν Y ) + ∫ Ent ( μ ( ⋅ ∣ y ) ∣ ν ( ⋅ ∣ y )) d μ Y ( y ) .
For a stationary point process P \mathrm{P} P of intensity 1, the specific relative entropy of P \mathrm{P} P is the entropy per unit area with respect to the Poisson point process PPP \operatorname{PPP} PPP of intensity 1:
E ( P ) : = lim R → ∞ 1 R 2 Ent ( P C R ∣ PPP C R ) , \mathcal{E}(\mathrm{P}):=\lim_{R\to\infty}\frac{1}{R^2}\operatorname{Ent}(\mathrm{P}_{C_R}\mid\operatorname{PPP}_{C_R}), E ( P ) := R → ∞ lim R 2 1 Ent ( P C R ∣ PPP C R ) ,
where P C R \mathrm{P}_{C_R} P C R is the restriction of P \mathrm{P} P to the square C R C_R C R .
Outline of the proof
The basic idea is to compare each good DLR state P \mathrm{P} P in some large disk B R B_R B R to a reference ν R \nu_R ν R which is some kind of 2DOCP, the distance being measured through relative entropy.
Using DLR equations, we see that computing the relative entropy in B R B_R B R boils down to estimating the typical contribution of boundary conditions (the effect of the configuration outside B R B_R B R , that is, the move function of Subsection 2.4) to the energy. By a direct computation, we find that for good DLR states, this is negligible compared to the volume. Such a subvolume bound alone does not imply uniqueness. For absolutely summable lattice interactions, a uniform O ( 1 ) O(1) O ( 1 ) bound on the energy change caused by changing the boundary condition gives a sufficient uniqueness criterion [41 Theorem 6.40] .
However, o ( ∣ B R ∣ ) o(|B_R|) o ( ∣ B R ∣ ) is exactly what is needed to identify P \mathrm{P} P as a minimizer of the free energy. This gives us the DLR variational principle. The only technical difficulty is the number of points: our DLR equations are canonical, so the reference must have the same law of the number of points in B R B_R B R as P \mathrm{P} P , and thus mixes gases which are not neutral. We construct this reference in Subsection 6.3.
A reference measure suited to a given point process
In the rest of this section, and until the last step of the proof of Proposition 1.1 , we fix R ≥ 2 R \ge2 R ≥ 2 such that ∣ B R ∣ = π R 2 |B_R| = \pi R^2 ∣ B R ∣ = π R 2 is an integer. The constants C C C and c c c below do not depend on R R R .
For n ≥ 0 n \ge0 n ≥ 0 , we write G n = G n , R π β / 2 G_n = G_{n,R}^{\pi\beta/2} G n = G n , R π β /2 for the law of n n n points in B R B_R B R as defined in Subsection 4.1. This measure is rotation-invariant for every n n n . For n = 0 n = 0 n = 0 , G 0 G_0 G 0 is the Dirac mass at the empty configuration. For n = ∣ B R ∣ n = |B_R| n = ∣ B R ∣ the neutral gas G ∣ B R ∣ G_{|B_R|} G ∣ B R ∣ is the finite-volume 2DOCP.
We write F : = F B R F := F_{B_R} F := F B R and Bin n : = Bin B R , n \operatorname{Bin}_n := \operatorname{Bin}_{B_R,n} Bin n := Bin B R , n . Let U U U be the potential created by the background, given by
U ( y ) : = ∫ B R g ( y − x ) d x for y ∈ R 2 . U(y) := \int_{B_R} g(y-x)\,\mathrm{d}x \qquad\text{for } y \in\mathbb{R}^2 . U ( y ) := ∫ B R g ( y − x ) d x for y ∈ R 2 .
As in the proof of Lemma 2.1 , expanding the double integral in the definition of F F F gives, for Z = ( z 1 , … , z n ) ∈ B R n Z = (z_1,\ldots,z_n) \in B_R^n Z = ( z 1 , … , z n ) ∈ B R n ,
F ( Z ) = ∑ i < j g ( z i − z j ) − ∑ i = 1 n U ( z i ) + E b g with E b g : = 1 2 ∫ B R U . (52) F(Z) = \sum_{i<j} g(z_i-z_j) - \sum_{i=1}^{n} U(z_i) + E_{\mathrm{bg}}
\qquad\text{with } E_{\mathrm{bg}} := \frac{1}{2}\int_{B_R} U . \tag*{(52)} F ( Z ) = i < j ∑ g ( z i − z j ) − i = 1 ∑ n U ( z i ) + E bg with E bg := 2 1 ∫ B R U . (52)
The three terms are the interaction of the points with each other, the effect of the background on the points, and the self-interaction of the background, which does not depend on Z Z Z . By Newton’s theorem, U ( x ) = − ∣ B R ∣ log R + π 2 ( R 2 − ∣ x ∣ 2 ) U(x) = -|B_R|\log R + \frac{\pi}{2}(R^2-|x|^2) U ( x ) = − ∣ B R ∣ log R + 2 π ( R 2 − ∣ x ∣ 2 ) for x ∈ B R x \in B_R x ∈ B R , hence G n G_n G n is indeed the measure of Subsection 4.1, and we can write:
G n = 1 K n e − β F Bin n , K n : = ∫ e − β F d Bin n . (53) G_n = \frac{1}{K_n}e^{-\beta F}\operatorname{Bin}_n,
\qquad
K_n := \int e^{-\beta F}\,\mathrm{d}\operatorname{Bin}_n . \tag*{(53)} G n = K n 1 e − β F Bin n , K n := ∫ e − β F d Bin n . (53)
Here K n K_n K n uses the background energy F F F and the probability measure Bin n \operatorname{Bin}_n Bin n ; it is not the Lebesgue normalizing constant K n , R π β / 2 K_{n,R}^{\pi\beta/2} K n , R π β /2 of Subsection 4.1.
Definition of the reference measure. Let P be a good DLR state, and consider the law of the number of points in B R B_R B R
p P , R ( n ) : = P ( N R = n ) for n ≥ 0. p_{P,R}(n) := \mathbb{P}(N_R=n) \qquad\text{for } n \ge0 . p P , R ( n ) := P ( N R = n ) for n ≥ 0.
We then associate to P its reference measure , defined as the mixture
ν R P : = ∑ n ≥ 0 p P , R ( n ) G n . (54) \nu_R^{P} := \sum_{n\ge0} p_{P,R}(n) G_n . \tag*{(54)} ν R P := n ≥ 0 ∑ p P , R ( n ) G n . (54)
This mixture has exactly the same law as P regarding the number of points in B R B_R B R .
Partition functions for non-neutral systems
In view of our definition of ν R P \nu_R^{P} ν R P , we are led to consider 2DOCPs with a possible non-neutrality (for n ≠ ∣ B R ∣ n \ne|B_R| n = ∣ B R ∣ in (54) ). The discrepancy of a configuration X X X in B R B_R B R is a measure of the non-neutrality, defined as:
Dis R ( X ) : = N R ( X ) − ∣ B R ∣ . \operatorname{Dis}_R(X) := N_R(X)-|B_R| . Dis R ( X ) := N R ( X ) − ∣ B R ∣.
The following proposition gives a lower bound on the partition function of a non-neutral 2DOCP.
Proposition 6.3 (A grand-canonical-like estimate). Let n n n be a random variable with values in N \mathbb{N} N , finite variance and E n = ∣ B R ∣ \mathbb{E}n = |B_R| E n = ∣ B R ∣ , and let Dis : = n − ∣ B R ∣ \operatorname{Dis} := n-|B_R| Dis := n − ∣ B R ∣ . Then there exists a universal constant C C C such that
E [ log K n ] ≥ log K ∣ B R ∣ − C β ( log R ⋅ E ∣ Dis ∣ + E Dis 2 ) . \mathbb{E}\left[\log K_n\right] \ge\log K_{|B_R|} - C\beta\left(\log R \cdot\mathbb{E}|\operatorname{Dis}| + \mathbb{E}\operatorname{Dis}^2\right) . E [ log K n ] ≥ log K ∣ B R ∣ − C β ( log R ⋅ E ∣ Dis ∣ + E Dis 2 ) .
The proof is elementary and postponed to Appendix § 13 .§ 9 : we compare K n K_n K n with K ∣ B R ∣ K_{|B_R|} K ∣ B R ∣ by adding ∣ Dis ∣ |\operatorname{Dis}| ∣ Dis ∣ points to, or removing them from, the neutral gas, and we estimate the resulting change of energy.
The boundary cost
We express the boundary cost through the following kernel.
Lemma 6.4 (The geometric boundary kernel). For u ∈ R 2 ∖ { 0 } u \in\mathbb{R}^{2} \setminus\{0\} u ∈ R 2 ∖ { 0 } , let
J R ( u ) : = ∫ { x ∈ B R , x + u ∉ B R } log ∣ x + u ∣ ∣ u ∣ d x , J R ( 0 ) : = 0. (55) J_R(u):=\int_{\{x\in B_R,\ x+u\notin B_R\}}\log\frac{|x+u|}{|u|}\,\mathrm{d}x,\qquad J_R(0):=0. \tag*{(55)} J R ( u ) := ∫ { x ∈ B R , x + u ∈ / B R } log ∣ u ∣ ∣ x + u ∣ d x , J R ( 0 ) := 0. (55) This is a radial and continuous function, supported on B 2 R B_{2R} B 2 R with
J R = R 2 J 1 ( ⋅ / R ) . J_R=R^2J_1(\cdot/R). J R = R 2 J 1 ( ⋅ / R ) . Moreover, Δ J 1 \Delta J_1 Δ J 1 is an integrable function, such that:
∣ Δ J 1 ^ ( w ) ∣ ≤ C ( 1 + ∣ w ∣ ) − 1 / 2 for w ∈ R 2 . (56) \left|\widehat{\Delta J_1}(w)\right|\leq C(1+|w|)^{-1/2}\qquad\text{for }w\in\mathbb{R}^{2}. \tag*{(56)} Δ J 1 ( w ) ≤ C ( 1 + ∣ w ∣ ) − 1/2 for w ∈ R 2 . (56) We postpone the elementary proof to Appendix B.7.
Recall that P is a good DLR state and that ν R P \nu_R^P ν R P is the reference measure for P defined in (54) .
Proposition 6.5 (Boundary cost). There is C P < ∞ C_P<\infty C P < ∞ , independent of R R R , such that
Ent ( P B R ∣ ν R P ) ≤ C P R 3 / 2 . (57) \operatorname{Ent}\left(P_{B_R}\mid\nu_R^P\right)\leq C_P R^{3/2}. \tag*{(57)} Ent ( P B R ∣ ν R P ) ≤ C P R 3/2 . (57)
Proof.
Step 1. The conditional law as a tilted finite gas. For X ∈ Conf X\in\operatorname{Conf} X ∈ Conf and a configuration X ~ \widetilde X X in B R B_R B R with N R ( X ~ ) = N R ( X ) N_R(\widetilde X)=N_R(X) N R ( X ) = N R ( X ) , define
h R X ( z ) : = ∫ B R c log ∣ y ∣ ∣ y − z ∣ d ( X − Leb ) ( y ) for z ∈ B R , h_R^X(z):=\int_{B_R^c}\log\frac{|y|}{|y-z|}\,\mathrm{d}(X-\operatorname{Leb})(y)\qquad\text{for }z\in B_R, h R X ( z ) := ∫ B R c log ∣ y − z ∣ ∣ y ∣ d ( X − Leb ) ( y ) for z ∈ B R ,
the potential created in B R B_R B R by the exterior charges (points and background), normalized by taking h R X ( 0 ) = 0 h_R^X(0)=0 h R X ( 0 ) = 0 . It depends only on X B R c X_{B_R^c} X B R c . It is proven in [59 ] that this quantity exists for P-almost every X X X and for all X ~ \widetilde X X . It is then harmonic in B R B_R B R . The move function of Subsection 2.4 can be alternatively written as:
M B R ( X ~ , X ) = X ~ ( h R X ) − X B R ( h R X ) , M_{B_R}(\widetilde X,X)=\widetilde X(h_R^X)-X_{B_R}(h_R^X), M B R ( X , X ) = X ( h R X ) − X B R ( h R X ) ,
see [59 Section 6.2] . When writing things this way, the second term does not depend on X ~ \widetilde X X .
For a configuration X X X , let n = N R ( X ) n=N_R(X) n = N R ( X ) and write P ( d X ~ ∣ X ) P(\mathrm{d}\widetilde X\mid X) P ( d X ∣ X ) for the conditional law of X B R X_{B_R} X B R under P given the exterior configuration X B R c X_{B_R^c} X B R c and the number of points n n n in B R B_R B R . By the DLR equations (14) we have, for P-almost every X X X ,
P ( d X ~ ∣ X ) = e − β X ~ ( h R X ) E G n [ e − β X ( h R X ) ] G n ( d X ~ ) . (58) P(\mathrm{d}\widetilde X\mid X)=\frac{e^{-\beta\widetilde X(h_R^X)}}{\mathbb{E}_{G_n}\left[e^{-\beta X(h_R^X)}\right]}G_n(\mathrm{d}\widetilde X). \tag*{(58)} P ( d X ∣ X ) = E G n [ e − β X ( h R X ) ] e − β X ( h R X ) G n ( d X ) . (58)
This is a finite 2DOCP with n n n points in B R B_R B R , tilted by the exterior field. Nothing new, just rewriting.
Step 2. Relative entropy and boundary interaction.
Claim 6.6. If X B R ( h R X ) X_{B_R}(h_R^X) X B R ( h R X ) is P-integrable (which we check in Step 3), then
Ent ( P B R ∣ ν R P ) ≤ − β E P [ X B R ( h R X ) ] . (59) \operatorname{Ent}\left(P_{B_R}\mid\nu_R^P\right)\leq-\beta\mathbb{E}_P\left[X_{B_R}(h_R^X)\right]. \tag*{(59)} Ent ( P B R ∣ ν R P ) ≤ − β E P [ X B R ( h R X ) ] . (59)
The right-hand side is the signed interaction energy between the points in B R B_R B R and the exterior. Proof. The one-point density of G n \mathbb{G}_n G n is radial, since G n \mathbb{G}_n G n is invariant under rotations around the origin. Since h R X h_R^X h R X is harmonic in B R B_R B R , its mean on every circle centred at 0 0 0 is h R X ( 0 ) = 0 h_R^X(0)=0 h R X ( 0 ) = 0 , hence E G n [ X ~ ( h R X ) ] = 0 \mathbb{E}_{\mathbb{G}_n}[\widetilde{X}(h_R^X)]=0 E G n [ X ( h R X )] = 0 . By Jensen’s inequality, log E G n [ e − β X ~ ( h R X ) ] ≥ 0 \log\mathbb{E}_{\mathbb{G}_n}[e^{-\beta\widetilde{X}(h_R^X)}]\ge0 log E G n [ e − β X ( h R X ) ] ≥ 0 , and by (58) ,
Ent ( P ( ⋅ ∣ X ) ∣ G n ) = − β E P ( ⋅ ∣ X ) [ X ~ ( h R X ) ] − log E G n [ e − β X ~ ( h R X ) ] ≤ − β E P ( ⋅ ∣ X ) [ X ~ ( h R X ) ] . (60) \operatorname{Ent}(\mathbb{P}(\cdot\mid X)\mid\mathbb{G}_n)=-\beta\mathbb{E}_{\mathbb{P}(\cdot\mid X)}[\widetilde{X}(h_R^X)]-\log\mathbb{E}_{\mathbb{G}_n}[e^{-\beta\widetilde{X}(h_R^X)}]\le-\beta\mathbb{E}_{\mathbb{P}(\cdot\mid X)}[\widetilde{X}(h_R^X)]. \tag*{(60)} Ent ( P ( ⋅ ∣ X ) ∣ G n ) = − β E P ( ⋅ ∣ X ) [ X ( h R X )] − log E G n [ e − β X ( h R X ) ] ≤ − β E P ( ⋅ ∣ X ) [ X ( h R X )] . (60)
Now, P B R \mathbb{P}_{B_R} P B R and ν R P \nu_R^P ν R P have the same law of the number of points, so the chain rule for relative entropy gives:
Ent ( P B R ∣ ν R P ) = ∑ n ≥ 0 p P , R ( n ) Ent ( P B R ( ⋅ ∣ N R = n ) ∣ G n ) . (61) \operatorname{Ent}(\mathbb{P}_{B_R}\mid\nu_R^P)=\sum_{n\ge0}p_{P,R}(n)\operatorname{Ent}(\mathbb{P}_{B_R}(\cdot\mid N_R=n)\mid\mathbb{G}_n). \tag*{(61)} Ent ( P B R ∣ ν R P ) = n ≥ 0 ∑ p P , R ( n ) Ent ( P B R ( ⋅ ∣ N R = n ) ∣ G n ) . (61)
Averaging (14) over the exterior configuration, P B R ( ⋅ ∣ N R = n ) \mathbb{P}_{B_R}(\cdot\mid N_R=n) P B R ( ⋅ ∣ N R = n ) is the average of P ( ⋅ ∣ X ) \mathbb{P}(\cdot\mid X) P ( ⋅ ∣ X ) over X X X distributed according to P ( ⋅ ∣ N R = n ) \mathbb{P}(\cdot\mid N_R=n) P ( ⋅ ∣ N R = n ) . By convexity of the relative entropy and (60) ,
Ent ( P B R ∣ ν R P ) ≤ E P [ Ent ( P ( ⋅ ∣ X ) ∣ G N R ) ] ≤ − β E P [ E P ( ⋅ ∣ X ) [ X ~ ( h R X ) ] ] . \operatorname{Ent}(\mathbb{P}_{B_R}\mid\nu_R^P)\le\mathbb{E}_P[\operatorname{Ent}(\mathbb{P}(\cdot\mid X)\mid\mathbb{G}_{N_R})]\le-\beta\mathbb{E}_P\left[\mathbb{E}_{\mathbb{P}(\cdot\mid X)}[\widetilde{X}(h_R^X)]\right]. Ent ( P B R ∣ ν R P ) ≤ E P [ Ent ( P ( ⋅ ∣ X ) ∣ G N R )] ≤ − β E P [ E P ( ⋅ ∣ X ) [ X ( h R X )] ] .
Finally, the DLR equations give E P [ E P ( ⋅ ∣ X ) [ X ~ ( h R X ) ] ] = E P [ X B R ( h R X ) ] \mathbb{E}_P\left[\mathbb{E}_{\mathbb{P}(\cdot\mid X)}[\widetilde{X}(h_R^X)]\right]=\mathbb{E}_P[X_{B_R}(h_R^X)] E P [ E P ( ⋅ ∣ X ) [ X ( h R X )] ] = E P [ X B R ( h R X )] .
The boundary interaction.
Claim 6.7. X B R ( h R X ) X_{B_R}(h_R^X) X B R ( h R X ) is P P P -integrable and, with J R J_R J R the geometric kernel of Lemma 6.4 , we have:
E P [ X B R ( h R X ) ] = ∫ R 2 J R d Γ P . (62) \mathbb{E}_P[X_{B_R}(h_R^X)]=\int_{\mathbb{R}^2}J_R\,\mathrm{d}\Gamma_P. \tag*{(62)} E P [ X B R ( h R X )] = ∫ R 2 J R d Γ P . (62)
Proof. Let ( ϑ j ) j ≥ 0 (\vartheta_j)_{j\ge0} ( ϑ j ) j ≥ 0 be the dyadic partition used to define the move function. The difficulty is that h R X h_R^X h R X is only conditionally convergent: for ∣ y ∣ ≫ R |y|\gg R ∣ y ∣ ≫ R , the kernel log ( ∣ y ∣ / ∣ y − z ∣ ) \log(|y|/|y-z|) log ( ∣ y ∣/∣ y − z ∣ ) is of order R / ∣ y ∣ R/|y| R /∣ y ∣ , which is not summable over the points y y y , and the series converges only thanks to the background and to cancellations. To exchange the expectation with this series, we need an L 1 L^1 L 1 bound which is uniform in z ∈ B R z\in B_R z ∈ B R , hence a supremum over z z z , which we control by Sobolev embedding. We first truncate the exterior field:
h R , p X ( z ) : = ∫ B R c ∑ j = 0 p ϑ j ( y ) log ∣ y ∣ ∣ y − z ∣ d ( X − Leb ) ( y ) . h_{R,p}^X(z):=\int_{B_R^{\mathrm{c}}}\sum_{j=0}^{p}\vartheta_j(y)\log\frac{|y|}{|y-z|}\,\mathrm{d}(X-\operatorname{Leb})(y). h R , p X ( z ) := ∫ B R c j = 0 ∑ p ϑ j ( y ) log ∣ y − z ∣ ∣ y ∣ d ( X − Leb ) ( y ) .
Each h R , p X h_{R,p}^X h R , p X is harmonic in B R B_R B R and vanishes at 0 0 0 , hence ∫ B R h R , p X = 0 \int_{B_R}h_{R,p}^X=0 ∫ B R h R , p X = 0 .
We check convergence in L 1 ( P ) L^1(P) L 1 ( P ) after summing over the interior points. For 2 j ≥ 8 R 2^j\ge8R 2 j ≥ 8 R and for multi-indices α \alpha α and α 1 \alpha_1 α 1 with ∣ α ∣ ≤ 2 |\alpha|\le2 ∣ α ∣ ≤ 2 and ∣ α 1 ∣ ≤ 3 |\alpha_1|\le3 ∣ α 1 ∣ ≤ 3 , the kernel satisfies
sup z ∈ B 2 R ∥ ∂ y α 1 ∂ z α ( ϑ j ( y ) log ∣ y ∣ ∣ y − z ∣ ) ∥ ∞ ≤ C ( R ) 2 − j ( 1 + ∣ α 1 ∣ ) . \sup_{z\in B_{2R}}\left\|\partial_y^{\alpha_1}\partial_z^\alpha\left(\vartheta_j(y)\log\frac{|y|}{|y-z|}\right)\right\|_\infty\le C(R)2^{-j(1+|\alpha_1|)}. z ∈ B 2 R sup ∂ y α 1 ∂ z α ( ϑ j ( y ) log ∣ y − z ∣ ∣ y ∣ ) ∞ ≤ C ( R ) 2 − j ( 1 + ∣ α 1 ∣ ) .
Applying (G2) to the rescaled kernels, then Sobolev embedding in z z z , gives
∥ sup z ∈ B R ∣ h R , j X ( z ) − h R , j − 1 X ( z ) ∣ ∥ 2 ≤ C ( R ) 2 − j . \left\|\sup_{z\in B_R}|h_{R,j}^X(z)-h_{R,j-1}^X(z)|\right\|_2\le C(R)2^{-j}. z ∈ B R sup ∣ h R , j X ( z ) − h R , j − 1 X ( z ) ∣ 2 ≤ C ( R ) 2 − j .
By (G1), N B R ∈ L 2 ( P ) N_{B_R}\in L^2(P) N B R ∈ L 2 ( P ) , so Cauchy–Schwarz shows that the corresponding tails of X B R ( h R , p X ) X_{B_R}(h_{R,p}^X) X B R ( h R , p X ) tend to zero in L 1 ( P ) L^1(P) L 1 ( P ) . The finitely many inner terms are integrable by (G3) and local integrability of the two-point density: their only singularity is logarithmic at y = z y=z y = z . Thus X B R ( h R , p X ) → p → ∞ X B R ( h R X ) X_{B_R}(h_{R,p}^X)\xrightarrow[p\to\infty]{}X_{B_R}(h_R^X) X B R ( h R , p X ) p → ∞ X B R ( h R X ) in L 1 ( P ) . L^1(P). L 1 ( P ) . Here R R R is fixed, and the constants C ( R ) C(R) C ( R ) depend on R R R .
For each finite p p p , stationarity, intensity 1 1 1 and ∫ B R h R , p X = 0 \int_{B_R}h_{R,p}^X=0 ∫ B R h R , p X = 0 give
E P [ X B R ( h R , p X ) ] = ∫ B R ∫ B R c ∑ j = 0 p ϑ j ( y ) log ∣ y ∣ ∣ y − z ∣ ( ρ 2 ( 0 , y − z ) − 1 ) d y d z . \mathbb{E}_P[X_{B_R}(h_{R,p}^X)]=\int_{B_R}\int_{B_R^{\mathrm{c}}}\sum_{j=0}^{p}\vartheta_j(y)\log\frac{|y|}{|y-z|}\bigl(\rho_2(0,y-z)-1\bigr)\,\mathrm{d}y\,\mathrm{d}z. E P [ X B R ( h R , p X )] = ∫ B R ∫ B R c j = 0 ∑ p ϑ j ( y ) log ∣ y − z ∣ ∣ y ∣ ( ρ 2 ( 0 , y − z ) − 1 ) d y d z .
After setting u = y − z u=y-z u = y − z , the integral in z z z of the absolute value of the kernel is bounded by C ( R ) ( 1 + ∣ log ∣ u ∣ ∣ ) C(R)(1+|\log|u||) C ( R ) ( 1 + ∣ log ∣ u ∣∣ ) for ∣ u ∣ ≤ 2 R |u|\le2R ∣ u ∣ ≤ 2 R , and by C ( R ) / ∣ u ∣ C(R)/|u| C ( R ) /∣ u ∣ for ∣ u ∣ > 2 R |u|>2R ∣ u ∣ > 2 R . These bounds are integrable against ∣ ρ 2 ( 0 , u ) − 1 ∣ d u |\rho_2(0,u)-1|\,\mathrm{d}u ∣ ρ 2 ( 0 , u ) − 1∣ d u , by (G3) and local integrability. Dominated convergence therefore gives (62) , since Γ P = δ 0 + ( ρ 2 ( 0 , ⋅ ) − 1 ) Leb \Gamma_P=\delta_0+(\rho_2(0,\cdot)-1)\operatorname{Leb} Γ P = δ 0 + ( ρ 2 ( 0 , ⋅ ) − 1 ) Leb and J R ( 0 ) = 0 J_R(0)=0 J R ( 0 ) = 0 . □ \square □
Step 4. Fourier computation. Moving to the Fourier side, using (62) and Lemma 6.4 , we get:
E P [ X B R ( h R X ) ] = 1 ( 2 π ) 2 ∫ J R ^ d σ P = − R 2 ( 2 π ) 2 ∫ Δ J 1 ^ ( R k ) ∣ k ∣ 2 σ P ( d k ) . \mathbb{E}_{P}\left[X_{B_R}(h_R^X)\right]=\frac{1}{(2\pi)^2}\int\widehat{J_R}\,\mathrm{d}\sigma_P=-\frac{R^2}{(2\pi)^2}\int\frac{\Delta\widehat{J_1}(Rk)}{|k|^2}\,\sigma_P(\mathrm{d}k). E P [ X B R ( h R X ) ] = ( 2 π ) 2 1 ∫ J R d σ P = − ( 2 π ) 2 R 2 ∫ ∣ k ∣ 2 Δ J 1 ( R k ) σ P ( d k ) .
where σ P \sigma_P σ P is the structure factor of P P P as in (7) . Using the decay estimate (56) yields:
∣ E P [ X B R ( h R X ) ] ∣ ≤ C R 2 ∫ ( 1 + R ∣ k ∣ ) − 1 / 2 ∣ k ∣ 2 σ P ( d k ) ≤ C R 3 / 2 ∫ ∣ k ∣ − 5 / 2 σ P ( d k ) . \left|\mathbb{E}_{P}\left[X_{B_R}(h_R^X)\right]\right|\le CR^2\int\frac{(1+R|k|)^{-1/2}}{|k|^2}\,\sigma_P(\mathrm{d}k)\le CR^{3/2}\int|k|^{-5/2}\,\sigma_P(\mathrm{d}k). E P [ X B R ( h R X ) ] ≤ C R 2 ∫ ∣ k ∣ 2 ( 1 + R ∣ k ∣ ) − 1/2 σ P ( d k ) ≤ C R 3/2 ∫ ∣ k ∣ − 5/2 σ P ( d k ) .
Splitting into dyadic shells, using (17) near k = 0 k=0 k = 0 and the translation-boundedness σ P ( B r ) ≤ C r 2 \sigma_P(B_r)\le Cr^2 σ P ( B r ) ≤ C r 2 at infinity, gives ∫ ∣ k ∣ − 5 / 2 σ P ( d k ) ≤ C ∑ j ≥ 0 ( 2 − 3 j / 2 + 2 − j / 2 ) < ∞ \int|k|^{-5/2}\sigma_P(\mathrm{d}k)\le C\sum_{j\geq0}(2^{-3j/2}+2^{-j/2})<\infty ∫ ∣ k ∣ − 5/2 σ P ( d k ) ≤ C ∑ j ≥ 0 ( 2 − 3 j /2 + 2 − j /2 ) < ∞ . With (59) , this proves (57) . □ \square □
Remark 6.8 (A softer argument). The variational principle only needs Ent ( P B R ∣ ν R P ) = o ( R 2 ) \operatorname{Ent}(P_{B_R}\mid\nu_R^P)=o(R^2) Ent ( P B R ∣ ν R P ) = o ( R 2 ) . For this weaker bound, we only need the integrability of the truncated two-point function u ↦ ρ 2 ( 0 , u ) − 1 u\mapsto\rho_2(0,u)-1 u ↦ ρ 2 ( 0 , u ) − 1 , which follows from (G1) and (G3). This integrability is not a trivial condition, however. For the limit points of the 2DOCP, it is a consequence of the decay of correlations of Theorem 1 (Corollary 5.11), and we do not know whether it holds for β > 2 \beta>2 β > 2 . By Lemma 6.4 , J R = R 2 J 1 ( ⋅ / R ) J_R=R^2J_1(\mathord{\cdot}/R) J R = R 2 J 1 ( ⋅ / R ) , where J 1 J_1 J 1 is continuous, supported on B 2 B_2 B 2 , and vanishes at 0. Hence
∣ ∫ R 2 J R d Γ P ∣ ≤ R 2 ∫ R 2 ∣ J 1 ( u / R ) ∣ ∣ ρ 2 ( 0 , u ) − 1 ∣ d u = o ( R 2 ) \left|\int_{\mathbb{R}^2}J_R\,\mathrm{d}\Gamma_P\right|\le R^2\int_{\mathbb{R}^2}|J_1(u/R)|\,|\rho_2(0,u)-1|\,\mathrm{d}u=o(R^2) ∫ R 2 J R d Γ P ≤ R 2 ∫ R 2 ∣ J 1 ( u / R ) ∣ ∣ ρ 2 ( 0 , u ) − 1∣ d u = o ( R 2 )
by dominated convergence. The Fourier computation gives the rate R 3 / 2 R^{3/2} R 3/2 .
Remark 6.9. The bound (56) is convenient but not optimal: a numerical computation suggests a decay of order log ∣ k ∣ / ∣ k ∣ \log|k|/|k| log ∣ k ∣/∣ k ∣ , which would give a boundary cost O ( R log R ) O(R\log R) O ( R log R ) .
Proof of Proposition 1.1 (DLR variational principle)
Proof.
Step 1. Chain rule. We start from (61) again, where the chain rule for relative entropy gave:
Ent ( P B R ∣ ν R P ) = ∑ n ≥ 0 p P , R ( n ) Ent ( P B R ( ⋅ ∣ N R = n ) ∣ G n ) . \operatorname{Ent}\left(P_{B_R}\mid\nu_R^P\right)=\sum_{n\geq0}p_{P,R}(n)\operatorname{Ent}\left(P_{B_R}(\mathord{\cdot}\mid N_R=n)\mid G_n\right). Ent ( P B R ∣ ν R P ) = n ≥ 0 ∑ p P , R ( n ) Ent ( P B R ( ⋅ ∣ N R = n ) ∣ G n ) .
Using the very definition of G n G_n G n , we can change the reference measure in the relative entropy:
Ent ( P B R ( ⋅ ∣ N R = n ) ∣ G n ) = Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) + β E P B R ( ⋅ ∣ N R = n ) [ F ] + log K n . \operatorname{Ent}\left(P_{B_R}(\mathord{\cdot}\mid N_R=n)\mid G_n\right)=\operatorname{Ent}\left(P_{B_R}(\mathord{\cdot}\mid N_R=n)\mid\operatorname{Bin}_n\right)+\beta\mathbb{E}_{P_{B_R}(\mathord{\cdot}\mid N_R=n)}[F]+\log K_n. Ent ( P B R ( ⋅ ∣ N R = n ) ∣ G n ) = Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) + β E P B R ( ⋅ ∣ N R = n ) [ F ] + log K n .
Averaging over n n n , we obtain
∑ n p P , R ( n ) [ Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) + β E P B R ( ⋅ ∣ N R = n ) [ F ] ] = Ent ( P B R ∣ ν R P ) − E P [ log K N R ] . (63) \sum_n p_{P,R}(n)\left[\operatorname{Ent}\left(P_{B_R}(\mathord{\cdot}\mid N_R=n)\mid\operatorname{Bin}_n\right)+\beta\mathbb{E}_{P_{B_R}(\mathord{\cdot}\mid N_R=n)}[F]\right]=\operatorname{Ent}\left(P_{B_R}\mid\nu_R^P\right)-\mathbb{E}_P[\log K_{N_R}]. \tag*{(63)} n ∑ p P , R ( n ) [ Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) + β E P B R ( ⋅ ∣ N R = n ) [ F ] ] = Ent ( P B R ∣ ν R P ) − E P [ log K N R ] . (63)
The right-hand side is finite, by Propositions 6.5 and 6.3.
Step 2. Entropy decomposition. The Poisson point process conditioned to have n n n points in B R B_R B R is Bin n \operatorname{Bin}_n Bin n . By the chain rule, we obtain a decomposition similar to (63) , where Poi ∣ B R ∣ \operatorname{Poi}_{|B_R|} Poi ∣ B R ∣ is the Poisson distribution of mean ∣ B R ∣ |B_R| ∣ B R ∣ :
Ent ( P B R ∣ PPP B R ) = Ent ( p P , R ∣ Poi ∣ B R ∣ ) + ∑ n p P , R ( n ) Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) . \operatorname{Ent}\left(P_{B_R}\mid\operatorname{PPP}_{B_R}\right)=\operatorname{Ent}\left(p_{P,R}\mid\operatorname{Poi}_{|B_R|}\right)+\sum_n p_{P,R}(n)\operatorname{Ent}\left(P_{B_R}(\mathord{\cdot}\mid N_R=n)\mid\operatorname{Bin}_n\right). Ent ( P B R ∣ PPP B R ) = Ent ( p P , R ∣ Poi ∣ B R ∣ ) + n ∑ p P , R ( n ) Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) .
The first term is negligible. We have Ent ( p P , R ∣ Poi ∣ B R ∣ ) ≤ − ∑ n p P , R ( n ) log Poi ∣ B R ∣ ( n ) \operatorname{Ent}(p_{P,R}\mid\operatorname{Poi}_{|B_R|})\le-\sum_n p_{P,R}(n)\log\operatorname{Poi}_{|B_R|}(n) Ent ( p P , R ∣ Poi ∣ B R ∣ ) ≤ − ∑ n p P , R ( n ) log Poi ∣ B R ∣ ( n ) . By Stirling’s formula, − log Poi ∣ B R ∣ ( n ) ≤ ( n − ∣ B R ∣ ) 2 / ∣ B R ∣ + O ( log ( n + 2 ) ) -\log\operatorname{Poi}_{|B_R|}(n)\le(n-|B_R|)^2/|B_R|+O(\log(n+2)) − log Poi ∣ B R ∣ ( n ) ≤ ( n − ∣ B R ∣ ) 2 /∣ B R ∣ + O ( log ( n + 2 )) . Therefore:
Ent ( p P , R ∣ Poi ∣ B R ∣ ) ≤ Var P [ N R ] ∣ B R ∣ + C log R . \operatorname{Ent}(p_{P,R}\mid\operatorname{Poi}_{|B_R|})\le\frac{\operatorname{Var}_P[N_R]}{|B_R|}+C\log R. Ent ( p P , R ∣ Poi ∣ B R ∣ ) ≤ ∣ B R ∣ Var P [ N R ] + C log R .
and Proposition 2.5 ensures that this is negligible compared to the volume ∣ B R ∣ |B_R| ∣ B R ∣ as R → ∞ R \to\infty R → ∞ . Thus:
lim inf R → ∞ 1 ∣ B R ∣ Ent ( P B R ∣ PPP B R ) = lim inf R → ∞ 1 ∣ B R ∣ ∑ n p P , R ( n ) Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) . \liminf_{R\to\infty} \frac{1}{|B_R|}\operatorname{Ent}\left(P_{B_R}\mid\operatorname{PPP}_{B_R}\right)
=
\liminf_{R\to\infty} \frac{1}{|B_R|}\sum_n p_{P,R}(n)\operatorname{Ent}\left(P_{B_R}(\cdot\mid N_R=n)\mid\operatorname{Bin}_n\right). R → ∞ lim inf ∣ B R ∣ 1 Ent ( P B R ∣ PPP B R ) = R → ∞ lim inf ∣ B R ∣ 1 n ∑ p P , R ( n ) Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) .
The left-hand side is equal to the specific relative entropy: for a stationary process, the specific relative entropy can be computed along disks as well as along squares, see [45 ] . We obtain:
lim inf R → ∞ 1 ∣ B R ∣ ∑ n p P , R ( n ) Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) = E ( P ) . (64) \liminf_{R\to\infty} \frac{1}{|B_R|}\sum_n p_{P,R}(n)\operatorname{Ent}\left(P_{B_R}(\cdot\mid N_R=n)\mid\operatorname{Bin}_n\right)=\mathcal{E}(P). \tag*{(64)} R → ∞ lim inf ∣ B R ∣ 1 n ∑ p P , R ( n ) Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) = E ( P ) . (64)
This just says that the randomness of the distribution of points in each disk is small enough (hyperuniformity) that it does not contribute to the entropy, which we can simply decompose according to n n n .
Step 3. Energy decomposition. We decompose E P [ F B R ( X B R ) ] \mathbb{E}_P[F_{B_R}(X_{B_R})] E P [ F B R ( X B R )] according to the number of points and write:
1 ∣ B R ∣ E P [ F B R ( X B R ) ] = 1 ∣ B R ∣ ∑ n p P , R ( n ) E P B R ( ⋅ ∣ N R = n ) [ F B R ] . \frac{1}{|B_R|}\mathbb{E}_P[F_{B_R}(X_{B_R})]
=
\frac{1}{|B_R|}\sum_n p_{P,R}(n)\mathbb{E}_{P_{B_R}(\cdot\mid N_R=n)}[F_{B_R}]. ∣ B R ∣ 1 E P [ F B R ( X B R )] = ∣ B R ∣ 1 n ∑ p P , R ( n ) E P B R ( ⋅ ∣ N R = n ) [ F B R ] .
By part (i) of Lemma 6.1, in the form (50) , we get:
lim inf R → ∞ 1 ∣ B R ∣ ∑ n p P , R ( n ) E P B R ( ⋅ ∣ N R = n ) [ F B R ] = W i n t ( P ) . (65) \liminf_{R\to\infty}\frac{1}{|B_R|}\sum_n p_{P,R}(n)\mathbb{E}_{P_{B_R}(\cdot\mid N_R=n)}[F_{B_R}]
=
\mathcal{W}^{\mathrm{int}}(P). \tag*{(65)} R → ∞ lim inf ∣ B R ∣ 1 n ∑ p P , R ( n ) E P B R ( ⋅ ∣ N R = n ) [ F B R ] = W int ( P ) . (65)
Step 4. Conclusion. We go back to the chain rule (63) , divide by ∣ B R ∣ |B_R| ∣ B R ∣ and let R → ∞ R\to\infty R → ∞ . On the left-hand side, we have, by (6.16) and (6.17):
lim inf R → ∞ 1 ∣ B R ∣ ∑ n p P , R ( n ) [ Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) + β E P B R ( ⋅ ∣ N R = n ) [ F ] ] ≥ E ( P ) + β W i n t ( P ) = F β i n t ( P ) . \liminf_{R\to\infty}\frac{1}{|B_R|}\sum_n p_{P,R}(n)\left[\operatorname{Ent}\left(P_{B_R}(\cdot\mid N_R=n)\mid\operatorname{Bin}_n\right)+\beta\mathbb{E}_{P_{B_R}(\cdot\mid N_R=n)}[F]\right]
\geq\mathcal{E}(P)+\beta\mathcal{W}^{\mathrm{int}}(P)=\mathcal{F}^{\mathrm{int}}_\beta(P). R → ∞ lim inf ∣ B R ∣ 1 n ∑ p P , R ( n ) [ Ent ( P B R ( ⋅ ∣ N R = n ) ∣ Bin n ) + β E P B R ( ⋅ ∣ N R = n ) [ F ] ] ≥ E ( P ) + β W int ( P ) = F β int ( P ) .
On the other hand, by Proposition 6.5,
1 ∣ B R ∣ Ent ( P B R ∣ ν R P ) → R → ∞ 0. \frac{1}{|B_R|}\operatorname{Ent}(P_{B_R}\mid\nu_R^P)\xrightarrow[R\to\infty]{}0. ∣ B R ∣ 1 Ent ( P B R ∣ ν R P ) R → ∞ 0.
It thus remains to bound from below the asymptotics of the non-neutral partition function K N R K_{N_R} K N R , that is, the limit inferior of ∣ B R ∣ − 1 E P [ log K N R ] |B_R|^{-1}\mathbb{E}_P[\log K_{N_R}] ∣ B R ∣ − 1 E P [ log K N R ] as R → ∞ R\to\infty R → ∞ . Recall the discrepancy Dis R = N R − ∣ B R ∣ \operatorname{Dis}_R=N_R-|B_R| Dis R = N R − ∣ B R ∣ of Subsection 6.4. We have E P Dis R = 0 \mathbb{E}_P\operatorname{Dis}_R=0 E P Dis R = 0 by stationarity, and E P Dis R 2 = Var P [ N R ] ≤ C P R \mathbb{E}_P\operatorname{Dis}_R^2=\operatorname{Var}_P[N_R]\leq C P_R E P Dis R 2 = Var P [ N R ] ≤ C P R by the hyperuniformity result of Proposition 2.5. The lower bound on the non-neutral partition function from Proposition 6.3, applied with n = N R n=N_R n = N R , gives:
1 ∣ B R ∣ E P [ log K N R ] ≥ 1 ∣ B R ∣ log K ∣ B R ∣ − C β ( log R R 3 / 2 + 1 R ) = 1 ∣ B R ∣ log K ∣ B R ∣ − o ( 1 ) . \frac{1}{|B_R|}\mathbb{E}_P[\log K_{N_R}]
\geq
\frac{1}{|B_R|}\log K_{|B_R|}
-
C\beta\left(\frac{\log R}{R^{3/2}}+\frac{1}{R}\right)
=
\frac{1}{|B_R|}\log K_{|B_R|}-o(1). ∣ B R ∣ 1 E P [ log K N R ] ≥ ∣ B R ∣ 1 log K ∣ B R ∣ − C β ( R 3/2 log R + R 1 ) = ∣ B R ∣ 1 log K ∣ B R ∣ − o ( 1 ) .
Here K ∣ B R ∣ K_{|B_R|} K ∣ B R ∣ is the partition function of the neutral gas in B R B_R B R , whose asymptotics is a consequence of the large deviation principle of [68 ] :
lim R → ∞ 1 ∣ B R ∣ log K ∣ B R ∣ = − min F β . \lim_{R\to\infty}\frac{1}{|B_R|}\log K_{|B_R|}=-\min\mathcal{F}_\beta. R → ∞ lim ∣ B R ∣ 1 log K ∣ B R ∣ = − min F β .
Therefore, we get:
lim inf R → ∞ 1 ∣ B R ∣ E P [ log K N R ] ≥ − min F β . \liminf_{R\to\infty}\frac{1}{|B_R|}\mathbb{E}_P[\log K_{N_R}]\geq-\min\mathcal{F}_\beta. R → ∞ lim inf ∣ B R ∣ 1 E P [ log K N R ] ≥ − min F β .
and thus, putting things together, we must have:
F β i n t ( P ) = E ( P ) + β W i n t ( P ) ≤ min F β . \mathcal{F}^{\mathrm{int}}_\beta(P)=\mathcal{E}(P)+\beta\mathcal{W}^{\mathrm{int}}(P)\leq\min\mathcal{F}_\beta. F β int ( P ) = E ( P ) + β W int ( P ) ≤ min F β .
We conclude using the electric-intrinsic inequality of Corollary 6.2: F β ( P ) ≤ F β i n t ( P ) ≤ min F β \mathcal{F}_\beta(P)\leq\mathcal{F}^{\mathrm{int}}_\beta(P)\leq\min\mathcal{F}_\beta F β ( P ) ≤ F β int ( P ) ≤ min F β . All these quantities are therefore equal, and F β ( P ) = min F β \mathcal{F}_\beta(P)=\min\mathcal{F}_\beta F β ( P ) = min F β .
Uniqueness of the free energy minimizer: proof of Theorem 2
In this section, 0 < β ≤ 2 0 < \beta\le2 0 < β ≤ 2 and we want to prove that there is a unique good DLR state. By Proposition 1.1 , every good DLR state minimizes F β \mathcal{F}_\beta F β , so it suffices to prove that F β \mathcal{F}_\beta F β has a unique minimizer, which is what we will do next. Most likely, there is a direct proof of uniqueness for DLR states along the lines of what was found in [7 ] for the one-dimensional log-gas, but we choose to work with the free energy as it is an important quantity on its own.
The minimizers of F β \mathcal{F}_\beta F β correspond to the zeros of the rate function of the large deviation principle of [68 ] for the empirical field of the neutral gas. We start by briefly recalling this result, as it will be useful for us. We then combine our exponential concentration estimates from § 6 for rotation-invariant observables with the large deviation principle of [68 ] in order to prove that all the minimizers of F β \mathcal{F}_\beta F β have the same rotation average.
An extremality argument (Subsection 7.3), which only uses the Poincaré inequality for general observables, then shows that F β \mathcal{F}_\beta F β has only one minimizer. Physically, this excludes orientational order: a crystal with a uniformly random orientation has an isotropic rotation average, but it is a non-trivial mixture.
Large deviations and exponential concentration
The large deviation principle. For n ≥ 1 n \ge1 n ≥ 1 , recall that R n = n / π R_n = \sqrt{n/\pi} R n = n / π , so that ∣ B R n ∣ = n |B_{R_n}| = n ∣ B R n ∣ = n , and let a : = π β / 2 a := \pi\beta/2 a := π β /2 . The neutral 2DOCP with n n n points in B R n B_{R_n} B R n is P n , β = G n , R n a \mathbb{P}_{n,\beta} = \mathbb{G}^{a}_{n,R_n} P n , β = G n , R n a . If X X X is a point configuration in B R n B_{R_n} B R n , its empirical field is the random probability measure describing the configuration seen from a uniformly chosen point of the disk:
Emp n ( X ) : = 1 ∣ B R n ∣ ∫ B R n δ τ x X d x . \operatorname{Emp}_n(X) := \frac{1}{|B_{R_n}|}\int_{B_{R_n}} \delta_{\tau_x X}\,\mathrm{d}x. Emp n ( X ) := ∣ B R n ∣ 1 ∫ B R n δ τ x X d x .
By [68 Theorem 1] , the law of the empirical field under P n , β \mathbb{P}_{n,\beta} P n , β satisfies a large deviation principle at speed n n n , for the weak topology on point processes, with good rate function F β − min F β \mathcal{F}_\beta- \min\mathcal{F}_\beta F β − min F β .
An abstract lemma. We use the following folklore lemma, for which we have not found a reference.
Lemma 7.1 (Large deviations versus exponential concentration). Let ( X n ) n (X_n)_n ( X n ) n be random variables in a Polish space X \mathcal{X} X whose laws satisfy a large deviation principle at speed n n n with a good rate function I I I . Let G : X → R G : \mathcal{X} \to\mathbb{R} G : X → R be bounded and continuous, and assume that, for every δ > 0 \delta> 0 δ > 0 , there is c ( δ ) > 0 c(\delta) > 0 c ( δ ) > 0 such that, for n n n large enough, an exponential concentration estimate holds:
P ( ∣ G ( X n ) − E G ( X n ) ∣ > δ ) ≤ e − c ( δ ) n . (66) \mathbb{P}\left(\left|G(X_n)-\mathbb{E}G(X_n)\right|>\delta\right)\le e^{-c(\delta)n}. \tag*{(66)} P ( ∣ G ( X n ) − E G ( X n ) ∣ > δ ) ≤ e − c ( δ ) n . (66) Then G G G is constant on the set of minimizers of I I I , and E G ( X n ) \mathbb{E}G(X_n) E G ( X n ) converges to this constant as n → ∞ n \to\infty n → ∞ .
Proof. Without loss of generality, we assume that min I = 0 \min I = 0 min I = 0 . Since I I I is a good rate function, { I = 0 } \{I=0\} { I = 0 } is non-empty. Let x ∗ x^* x ∗ be a minimizer and δ > 0 \delta> 0 δ > 0 , and assume that ∣ G ( X n ) − G ( x ∗ ) ∣ > 2 δ |G(X_n)-G(x^*)| > 2\delta ∣ G ( X n ) − G ( x ∗ ) ∣ > 2 δ along a subsequence. The open set U : = { ∣ G − G ( x ∗ ) ∣ < δ } U := \{|G-G(x^*)| < \delta\} U := { ∣ G − G ( x ∗ ) ∣ < δ } contains x ∗ x^* x ∗ , so the lower bound of the large deviation principle gives lim inf n 1 n log P ( X n ∈ U ) ≥ − inf U I = 0 \liminf_n \frac{1}{n}\log\mathbb{P}(X_n \in U) \ge-\inf_U I = 0 lim inf n n 1 log P ( X n ∈ U ) ≥ − inf U I = 0 . On the other hand, along the subsequence, X n ∈ U X_n \in U X n ∈ U implies ∣ G ( X n ) − E G ( X n ) ∣ > δ |G(X_n)-\mathbb{E}G(X_n)| > \delta ∣ G ( X n ) − E G ( X n ) ∣ > δ , an event of probability at most e − c ( δ ) n e^{-c(\delta)n} e − c ( δ ) n by assumption. Taking the log, dividing by n n n and sending n → ∞ n \to\infty n → ∞ yields a contradiction. Hence every limit point of E G ( X n ) \mathbb{E}G(X_n) E G ( X n ) equals G ( x ∗ ) G(x^*) G ( x ∗ ) , for every minimizer x ∗ x^* x ∗ .
In words: at the exponential scale, the system visits a neighbourhood of every minimizer of the rate function at a subexponential cost, and exponential concentration forbids G G G to take different values on these neighbourhoods.
The neutral gas: convergence of the empirical field
The observable. Let f ∈ T f \in\mathcal{T} f ∈ T be a test function belonging to the class considered in Subsection 4.7, and let Ψ f \Psi_f Ψ f be the rotation-average Ψ f ( X ) : = ∫ 0 2 π e − X ( f ( e − i θ ⋅ ) ) d θ 2 π \Psi_f(X) := \int_0^{2\pi} e^{-X(f(e^{-i\theta}\cdot))}\,\frac{\mathrm{d}\theta}{2\pi} Ψ f ( X ) := ∫ 0 2 π e − X ( f ( e − i θ ⋅ )) 2 π d θ , a bounded continuous local function on Conf \operatorname{Conf} Conf . We make two observations.
A function of the empirical field. Recall the Laplace average S R f S_R^f S R f of (31) . It coincides with the (unnormalized) average of Ψ f \Psi_f Ψ f over the centres x ∈ B R x \in B_R x ∈ B R . We have:
S R f ( X ) = ∫ B R Ψ f ( τ x X ) d x . S_R^f(\mathrm{X}) = \int_{B_R} \Psi_f(\tau_x\mathrm{X})\,\mathrm{d}x. S R f ( X ) = ∫ B R Ψ f ( τ x X ) d x .
This means that if X \mathrm{X} X is sampled from the 2DOCP, the observable S R n f ( X ) S_{R_n}^f(\mathrm{X}) S R n f ( X ) is actually a function of the empirical field Emp n ( X ) \operatorname{Emp}_n(\mathrm{X}) Emp n ( X ) . More precisely, we have (after normalization by the volume)
1 n S R n f ( X ) = ∫ Ψ f d Emp n ( X ) = G f ( Emp n ( X ) ) , where G f ( P ) : = ∫ Ψ f d P . \frac{1}{n}S_{R_n}^f(\mathrm{X}) = \int\Psi_f\mathrm{d}\operatorname{Emp}_n(\mathrm{X}) = G_f(\operatorname{Emp}_n(\mathrm{X})), \qquad\text{where }G_f(\mathrm{P}) := \int\Psi_f\mathrm{d}\mathrm{P}. n 1 S R n f ( X ) = ∫ Ψ f d Emp n ( X ) = G f ( Emp n ( X )) , where G f ( P ) := ∫ Ψ f dP .
Exponential concentration. By Proposition 4.16 , applied with N = n N=n N = n and L = R = R n L=R=R_n L = R = R n , for every n ≥ 1 n\geq1 n ≥ 1 and ∣ t ∣ ≤ t a , f |t|\leq t_{a,f} ∣ t ∣ ≤ t a , f , we have, with C a , f C_{a,f} C a , f independent of n n n
log E P n , β [ e t ( S R n f − E P n , β S R n f ) ] ≤ C a , f n t 2 . \log\mathbb{E}_{\mathrm{P}_{n,\beta}}\left[e^{t(S_{R_n}^f-\mathbb{E}_{\mathrm{P}_{n,\beta}}S_{R_n}^f)}\right]\leq C_{a,f}nt^2. log E P n , β [ e t ( S R n f − E P n , β S R n f ) ] ≤ C a , f n t 2 .
A deviation of S R n f / n S_{R_n}^f/n S R n f / n of order δ \delta δ from its mean thus has a probability of order e − c δ 2 n e^{-c\delta^2n} e − c δ 2 n . In particular the assumption (66) holds for X n = Emp n X_n=\operatorname{Emp}_n X n = Emp n and G = G f G=G_f G = G f , with some c ( δ ) > 0 c(\delta)>0 c ( δ ) > 0 depending on δ \delta δ , a a a and f f f . For a point process P \mathrm{P} P , let Rot ( P ) : = ∫ 0 2 π ( e i θ ) ∗ P d θ 2 π \operatorname{Rot}(\mathrm{P}):=\int_0^{2\pi}(e^{i\theta})_*\mathrm{P}\,\frac{\mathrm{d}\theta}{2\pi} Rot ( P ) := ∫ 0 2 π ( e i θ ) ∗ P 2 π d θ be its rotation average, where ( e i θ ) ∗ P (e^{i\theta})_*\mathrm{P} ( e i θ ) ∗ P is the law of a configuration of law P \mathrm{P} P rotated by the angle θ \theta θ . We can therefore combine the large deviation principle with the exponential concentration through Lemma 7.1 , and get the following result.
Proposition 7.2 (Rotation averages of the free energy minimizers). All the minimizers of F β \mathcal{F}_\beta F β have the same rotation average. In particular, F β \mathcal{F}_\beta F β has at most one rotation-invariant minimizer.
Proof. Fix f ∈ T f\in\mathcal{T} f ∈ T . By the discussion above, Lemma 7.1 applies to X n = Emp n X_n=\operatorname{Emp}_n X n = Emp n and G = G f G=G_f G = G f . It says that G f G_f G f takes the same value at all the minimizers of F β \mathcal{F}_\beta F β , and that E G f ( Emp n ) \mathbb{E}G_f(\operatorname{Emp}_n) E G f ( Emp n ) converges to this value.
Let now P \mathrm{P} P be a free energy minimizer. By definition,
G f ( P ) = E P [ Ψ f ] = ∫ 0 2 π E P [ e − X ( f ( e − i θ ⋅ ) ) ] d θ 2 π = E Rot ( P ) [ e − X ( f ) ] , G_f(\mathrm{P})=\mathbb{E}_{\mathrm{P}}[\Psi_f]=\int_0^{2\pi}\mathbb{E}_{\mathrm{P}}\left[e^{-\mathrm{X}(f(e^{-i\theta}\cdot))}\right]\frac{\mathrm{d}\theta}{2\pi}=\mathbb{E}_{\operatorname{Rot}(\mathrm{P})}\left[e^{-\mathrm{X}(f)}\right], G f ( P ) = E P [ Ψ f ] = ∫ 0 2 π E P [ e − X ( f ( e − i θ ⋅ )) ] 2 π d θ = E Rot ( P ) [ e − X ( f ) ] ,
which is the Laplace functional of Rot ( P ) \operatorname{Rot}(\mathrm{P}) Rot ( P ) evaluated at f f f . For every f ∈ T f\in\mathcal{T} f ∈ T , the Laplace functional of Rot ( P ) \operatorname{Rot}(\mathrm{P}) Rot ( P ) at f f f is therefore the same for all minimizers. Laplace functionals characterize the law of a point process, see for example [36 Chapter 9] , and it is enough to know them on T \mathcal{T} T , by a standard approximation argument. Hence Rot ( P ) \operatorname{Rot}(\mathrm{P}) Rot ( P ) does not depend on the minimizer. If P \mathrm{P} P is rotation-invariant, then Rot ( P ) = P \operatorname{Rot}(\mathrm{P})=\mathrm{P} Rot ( P ) = P , which gives the second assertion.
We also deduce that the rotation-averaged empirical field of the neutral gas converges in probability to a deterministic limit, without extraction.
Extremality
In order to prove that F β \mathcal{F}_\beta F β has only one minimizer, we show that the limit points of the 2DOCP are extremal among stationary point processes. For this, we return to finite volume, use our general Poincaré inequality, and pass to the limit.
Lemma 7.3 (Extremality of the limit of the 2DOCP). Let P \mathrm{P} P be a limit point of the local configurations X N \mathrm{X}_N X N under P N , β \mathbb{P}_{N,\beta} P N , β . For every cylinder observable F F F , we have
Var P [ F ] ≤ 1 a D P ( F ) with a = π β / 2. (67) \operatorname{Var}_{\mathrm{P}}[F]\leq\frac{1}{a}D_{\mathrm{P}}(F) \qquad\text{with }a=\pi\beta/2. \tag*{(67)} Var P [ F ] ≤ a 1 D P ( F ) with a = π β /2. (67) Moreover, P \mathrm{P} P is extremal, namely it cannot be written as a non-trivial mixture of two stationary laws.
Proof. For each N N N , the restriction F ( N ) F^{(N)} F ( N ) of F F F is a smooth symmetric function of N N N positions, so Proposition 4.17 with L = R N L=R_N L = R N gives (67) under P N , β \mathbb{P}_{N,\beta} P N , β . The variance and the Dirichlet energy pass to the limit thanks to the upgraded convergence of [59 ] .
Next, assume that F F F is K K K -local for some compact set K K K , and consider its spatial average
A R F : = ∣ B R ∣ − 1 ∫ B R F ∘ τ x d x . A_R F := \lvert B_R\rvert^{-1} \int_{B_R} F\circ\tau_x\,\mathrm{d}x. A R F := ∣ B R ∣ − 1 ∫ B R F ∘ τ x d x .
Stationarity and Cauchy–Schwarz give D P ( A R F ) ≤ ∣ K ∣ ∣ B R ∣ D P ( F ) \mathcal{D}_{\mathbb{P}}(A_R F)\leq\frac{\lvert K\rvert}{\lvert B_R\rvert}\mathcal{D}_{\mathbb{P}}(F) D P ( A R F ) ≤ ∣ B R ∣ ∣ K ∣ D P ( F ) . The Riemann sums of A R F A_R F A R F are cylinder observables. Since F F F and ∇ z F \nabla_z F ∇ z F are bounded, ∇ z F \nabla_z F ∇ z F vanishes for z z z outside K K K and P \mathbb{P} P has a finite intensity, dominated convergence shows that these sums converge to A R F A_R F A R F in L 2 ( P ) L^2(\mathbb{P}) L 2 ( P ) and that their Dirichlet energies converge to D P ( A R F ) \mathcal{D}_{\mathbb{P}}(A_R F) D P ( A R F ) . Hence (67) holds for A R F A_R F A R F , and Var P [ A R F ] → R → ∞ 0 \operatorname{Var}_{\mathbb{P}}[A_R F]\xrightarrow[R\to\infty]{}0 Var P [ A R F ] R → ∞ 0 , which means that spatial averages are asymptotically deterministic.
If P \mathbb{P} P were written as P = t P 1 + ( 1 − t ) P 2 \mathbb{P}=t\mathbb{P}_1+(1-t)\mathbb{P}_2 P = t P 1 + ( 1 − t ) P 2 with 0 < t < 1 0<t<1 0 < t < 1 and P 1 , P 2 \mathbb{P}_1,\mathbb{P}_2 P 1 , P 2 stationary, then we would have Var P [ A R F ] ≥ t ( 1 − t ) ( P 1 ( F ) − P 2 ( F ) ) 2 \operatorname{Var}_{\mathbb{P}}[A_R F]\geq t(1-t)(\mathbb{P}_1(F)-\mathbb{P}_2(F))^2 Var P [ A R F ] ≥ t ( 1 − t ) ( P 1 ( F ) − P 2 ( F ) ) 2 , so P 1 ( F ) = P 2 ( F ) \mathbb{P}_1(F)=\mathbb{P}_2(F) P 1 ( F ) = P 2 ( F ) for all cylinders, and P 1 = P 2 \mathbb{P}_1=\mathbb{P}_2 P 1 = P 2 .
Proof of Theorem 2 and Corollaries 3 and 4 (uniqueness and convergence)
Proof of Theorem 2 and Corollary 3. We start by choosing a rotation-invariant minimizer of F β \mathcal{F}_\beta F β .
Step 1. A reference minimizer. Let P ∞ \mathbb{P}_\infty P ∞ be a limit point of the local configurations X N \mathbf{X}_N X N under P N , β \mathbb{P}_{N,\beta} P N , β .
By Proposition 2.4 and Corollary 5.11, P ∞ \mathbb{P}_\infty P ∞ is a good DLR state, stationary and rotation-invariant.
By Proposition 1.1 , P ∞ \mathbb{P}_\infty P ∞ minimizes F β \mathcal{F}_\beta F β .
By Lemma 7.3 , P ∞ \mathbb{P}_\infty P ∞ is extremal among stationary laws.
Step 2. Uniqueness of the minimizer. Let P \mathbb{P} P be another stationary minimizer of F β \mathcal{F}_\beta F β .
Both P \mathbb{P} P and P ∞ \mathbb{P}_\infty P ∞ are minimizers, so Proposition 7.2 gives Rot ( P ) = Rot ( P ∞ ) = P ∞ \operatorname{Rot}(\mathbb{P})=\operatorname{Rot}(\mathbb{P}_\infty)=\mathbb{P}_\infty Rot ( P ) = Rot ( P ∞ ) = P ∞ , the last equality because P ∞ \mathbb{P}_\infty P ∞ is rotation-invariant.
Thus P ∞ \mathbb{P}_\infty P ∞ is the average of the stationary laws ( e i θ ) ∗ P (e^{i\theta})_*\mathbb{P} ( e i θ ) ∗ P . Since P ∞ \mathbb{P}_\infty P ∞ is extremal, it means that ( e i θ ) ∗ P = P ∞ (e^{i\theta})_*\mathbb{P}=\mathbb{P}_\infty ( e i θ ) ∗ P = P ∞ for almost every θ \theta θ .
Since P ∞ \mathbb{P}_\infty P ∞ is rotation-invariant, P = ( e − i θ ) ∗ P ∞ = P ∞ \mathbb{P}=(e^{-i\theta})_*\mathbb{P}_\infty=\mathbb{P}_\infty P = ( e − i θ ) ∗ P ∞ = P ∞ .
Hence P ∞ \mathbb{P}_\infty P ∞ is the unique minimizer of F β \mathcal{F}_\beta F β (this is part (i) of Theorem 2 ). We call it Gin β \operatorname{Gin}_\beta Gin β .
Step 3. Uniqueness of the good DLR state. Let P \mathbb{P} P be a good DLR state. By Proposition 2.4 and Proposition 1.1 it is a stationary minimizer of F β \mathcal{F}_\beta F β . By Step 2, P = Gin β \mathbb{P}=\operatorname{Gin}_\beta P = Gin β .
Step 4. The remaining claims. The process Gin β \operatorname{Gin}_\beta Gin β is rotation-invariant (Step 1), and ergodic, since it is extremal among stationary laws. Finally, every limit point of the laws of X N \mathbf{X}_N X N is a minimizer of F β \mathcal{F}_\beta F β (Step 1), hence equal to Gin β \operatorname{Gin}_\beta Gin β , which gives the convergence of X N \mathbf{X}_N X N .
Let ( x N ) N ≥ 1 (x_N)_{N\geq1} ( x N ) N ≥ 1 be a bounded deterministic sequence in R 2 \mathbb{R}^2 R 2 . Then every subsequence admits a further subsequence along which x N x_N x N converges to a point x x x . Continuity of translations then gives convergence of the configuration seen from x N x_N x N to a translate of Gin β \operatorname{Gin}_\beta Gin β , which has the same law by stationarity. This proves the convergence for bounded sequences of observation centres as well.
Proof of Corollary 4. The Ginibre point process is the determinantal point process with kernel K ( z , w ) = exp ( π z w ‾ − π 2 ( ∣ z ∣ 2 + ∣ w ∣ 2 ) ) K(z,w)=\exp(\pi z\overline{w}-\frac{\pi}{2}(\lvert z\rvert^2+\lvert w\rvert^2)) K ( z , w ) = exp ( π z w − 2 π (∣ z ∣ 2 + ∣ w ∣ 2 )) with respect to the Lebesgue measure [43 ] . We show that the laws of X N \mathbf{X}_N X N under P N , β \mathbb{P}_{N,\beta} P N , β with β = 2 \beta=2 β = 2 converge to this process. Corollary 3 then identifies it with Gin 2 \operatorname{Gin}_2 Gin 2 , and the other two assertions are those of Theorem 2 at β = 2 \beta=2 β = 2 .
By Lemma 2.1 , the density of P N , β \mathbb{P}_{N,\beta} P N , β at β = 2 \beta=2 β = 2 is proportional to ∏ i < j ∣ x i − x j ∣ 2 ∏ i e − π ∣ x i ∣ 2 \prod_{i<j}\lvert x_i-x_j\rvert^2\prod_i e^{-\pi\lvert x_i\rvert^2} ∏ i < j ∣ x i − x j ∣ 2 ∏ i e − π ∣ x i ∣ 2 on Λ N \Lambda_N Λ N . The weight e − π ∣ z ∣ 2 e^{-\pi\lvert z\rvert^2} e − π ∣ z ∣ 2 on Λ N \Lambda_N Λ N is radial, so the monomials are orthogonal for it, and the computation of Ginibre applies. In polar coordinates, the squared norm of z k z^k z k for this weight is
π − k ∫ 0 N u k e − u d u = π − k k ! Poi N ( [ k + 1 , ∞ ) ) , \pi^{-k}\int_0^N u^k e^{-u}\,\mathrm{d}u=\pi^{-k}k!\,\operatorname{Poi}_N([k+1,\infty)), π − k ∫ 0 N u k e − u d u = π − k k ! Poi N ([ k + 1 , ∞ )) ,
where Poi N \operatorname{Poi}_N Poi N is the Poisson distribution with mean N N N . Hence X N X_N X N is determinantal on Λ N \Lambda_N Λ N , with kernel
K N ( z , w ) = e − π 2 ( ∣ z ∣ 2 + ∣ w ∣ 2 ) ∑ k = 0 N − 1 ( π z w ‾ ) k k ! Poi N ( [ k + 1 , ∞ ) ) K_N(z,w)=e^{-\frac{\pi}{2}(|z|^2+|w|^2)}\sum_{k=0}^{N-1}\frac{(\pi z\overline{w})^k}{k!\operatorname{Poi}_N([k+1,\infty))} K N ( z , w ) = e − 2 π ( ∣ z ∣ 2 + ∣ w ∣ 2 ) k = 0 ∑ N − 1 k ! Poi N ([ k + 1 , ∞ )) ( π z w ) k
For fixed k k k , the probability Poi N ( [ k + 1 , ∞ ) ) \operatorname{Poi}_N([k+1,\infty)) Poi N ([ k + 1 , ∞ )) tends to 1 1 1 as N → ∞ N\to\infty N → ∞ . For k ≤ N − 1 k\leq N-1 k ≤ N − 1 , it is at least c 0 : = inf n ≥ 1 Poi n ( [ n , ∞ ) ) c_0:=\inf_{n\geq1}\operatorname{Poi}_n([n,\infty)) c 0 := inf n ≥ 1 Poi n ([ n , ∞ )) , which is positive by the central limit theorem. By dominated convergence in k k k , we obtain K N ⟶ K K_N\longrightarrow K K N ⟶ K uniformly on compact sets. The same lower bound gives K N ( z , z ) ≤ c 0 − 1 K ( z , z ) = c 0 − 1 K_N(z,z)\leq c_0^{-1}K(z,z)=c_0^{-1} K N ( z , z ) ≤ c 0 − 1 K ( z , z ) = c 0 − 1 on Λ N \Lambda_N Λ N .
The n n n -point correlation functions of X N X_N X N are the determinants of ( K N ( z i , z j ) ) 1 ≤ i , j ≤ n (K_N(z_i,z_j))_{1\leq i,j\leq n} ( K N ( z i , z j ) ) 1 ≤ i , j ≤ n . They therefore converge locally uniformly to those of the Ginibre point process, and Hadamard’s inequality bounds them by c 0 − n c_0^{-n} c 0 − n . Let f f f be continuous with compact support. The moments of X ( f ) X(f) X ( f ) are integrals of these correlation functions, so they converge to the moments of X ( f ) X(f) X ( f ) under the Ginibre point process. Since the correlation functions of this process are at most 1 1 1 , the variable X ( f ) X(f) X ( f ) has exponential moments under it, and its law is determined by its moments. Therefore X ( f ) X(f) X ( f ) converges in law for every such f f f , which is the weak convergence of the laws of X N X_N X N .
Debye–Hückel asymptotics
In this section we prove the Debye–Hückel asymptotics of Theorem 5 and the Gaussian limit of Theorem 6 . Recall from Subsection 3.4 that q β ( r ) = ρ 2 , β ( 0 , ℓ D r e 1 ) q_\beta(r)=\rho_{2,\beta}(0,\ell_D r e_1) q β ( r ) = ρ 2 , β ( 0 , ℓ D r e 1 ) for r > 0 r>0 r > 0 , where e 1 = ( 1 , 0 ) e_1=(1,0) e 1 = ( 1 , 0 ) . We start by collecting, in Subsection 8.1, the estimates on the gas at the scale ℓ D \ell_D ℓ D that we use throughout.
The proof of the Debye–Hückel asymptotics has two parts.
First part: averages of the two-point function. The covariances of smooth linear statistics only see averages of q β − 1 q_\beta-1 q β − 1 . We compute their limit in Subsection 8.2, with the identity (80) , obtained by an integration by parts in the Gibbs measure. This gives the convergence in S ′ ( R 2 ) S'(\mathbb{R}^2) S ′ ( R 2 ) . The same integration by parts, applied to characteristic functions, gives the Gaussian limit in Subsection 8.3.
Second part: values at fixed distances. To get the value of q β q_\beta q β at each distance, we fix two particles of the finite gas and write the two-point correlation function as a Laplace transform (Subsection 8.4). We expand its logarithm to second order in β \beta β in Subsection 8.5. In Subsection 8.6, this identifies the ratios q β ( r ) / q β ( s ) q_\beta(r)/q_\beta(s) q β ( r ) / q β ( s ) to first order in β \beta β , and the convergence in S ′ ( R 2 ) S'(\mathbb{R}^2) S ′ ( R 2 ) fixes the remaining multiplicative constant.
In this second part, the main issue is that the two test particles carry a total charge 2 2 2 . Their potential g + g ( ⋅ − r e 1 ) g+g(\cdot-r e_1) g + g ( ⋅ − r e 1 ) behaves like − 2 log ∣ z ∣ -2\log|z| − 2 log ∣ z ∣ at infinity, Y ( g + g ( ⋅ − r e 1 ) ) Y(g+g(\cdot-r e_1)) Y ( g + g ( ⋅ − r e 1 )) is not a fluctuation of order one, and the gas responds to this charge also globally, by pushing two units of charge towards the boundary. This global response is neither small nor linear, but it does not depend on r r r , and we only need the ratios q β ( r ) / q β ( s ) q_\beta(r)/q_\beta(s) q β ( r ) / q β ( s ) . We therefore renormalize the inserted charge, in three technical steps.
Neutralization. We split g + g ( ⋅ − r e 1 ) = D r + 2 g 1 g+g(\cdot-r e_1)=D_r+2g_1 g + g ( ⋅ − r e 1 ) = D r + 2 g 1 , where 2 g 1 2g_1 2 g 1 is the potential of a charge 2 2 2 spread over the unit disk, which does not depend on r r r , and D r D_r D r is the potential of a neutral distribution of charges, which decays like r / ∣ z ∣ r/|z| r /∣ z ∣ at infinity. We include the term 2 β Y ( g 1 ) 2\beta Y(g_1) 2 β Y ( g 1 ) in the Gibbs measure, which becomes the measure P 1 \mathbb{P}_1 P 1 of a gas interacting with a fixed charge 2 2 2 spread over the unit disk. This term cancels in the ratios (83) . A similar compensation, of a point charge by a fixed smeared charge that does not depend on the position of the point, is used in [63 Section 3.1] to study the potential of the 2DOCP at a point.
Expansion. Since D r D_r D r is neutral, Y ( D r ) Y(D_r) Y ( D r ) is a fluctuation: by the Poincaré inequality, it is of order ℓ D \ell_D ℓ D , up to a logarithmic factor. Hence β Y ( D r ) \beta Y(D_r) β Y ( D r ) fluctuates by β ℓ D ≈ β \beta\ell_D\approx\sqrt{\beta} β ℓ D ≈ β , and the expansion to second order holds, with an error O ( β 3 / 2 ∣ log β ∣ 3 / 2 ) O(\beta^{3/2}|\log\beta|^{3/2}) O ( β 3/2 ∣ log β ∣ 3/2 ) (Lemma 8.8 ). The logarithmic singularities of D r D_r D r at the two particles are smoothed at the scale ℓ D − 1 \ell_D^{-1} ℓ D − 1 of the distance between neighbouring points, and, since they are non-negative, they do not spoil the exponential moments of − β Y ( D r ) -\beta Y(D_r) − β Y ( D r ) .
Removing the added charge. The mean and the variance are now taken under P 1 \mathbb{P}_1 P 1 rather than under the gas without the added charge. We interpolate: let P t \mathbb{P}_t P t , for 0 ≤ t ≤ 1 0 \le t \le1 0 ≤ t ≤ 1 , be the gas interacting with the charge 2 t 2t 2 t , so that d d t E t [ G ] = − 2 β Cov t [ G , Y ( g 1 ) ] \frac{d}{dt}\mathbb{E}_t[G] = -2\beta\operatorname{Cov}_t[G,Y(g_1)] d t d E t [ G ] = − 2 β Cov t [ G , Y ( g 1 )] for every observable G G G , and the derivative of a covariance is a third cumulant. These derivatives must be controlled for every t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] , and not only at t = 0 t=0 t = 0 . This is possible because each P t \mathbb{P}_t P t is a gas of the same type: the added potential only lowers the Levi matrix of the shifted weight by O ( β ) O(\beta) O ( β ) , so the Poincaré inequality and the exponential decay of covariances of § 8 hold uniformly in t t t , and they bound the third cumulants (Lemmas 8.1, 8.9 and 8.10). Interpolating along a family of perturbed gases, for which the estimates hold uniformly, is a classical device. The outcome is that the added charge changes the mean at first order, by a linear-response term which we keep, and changes the variances only by a negligible amount.
The gas on the Debye scale
As in Subsection 3.4, we measure distances in units of ℓ D \ell_{\mathrm{D}} ℓ D , and Y Y Y is the image of a configuration X X X under x ↦ x / ℓ D x \mapsto x/\ell_{\mathrm{D}} x ↦ x / ℓ D , so that Y ( f ) = X ( f ( ⋅ / ℓ D ) ) Y(f)=X(f(\cdot/\ell_{\mathrm{D}})) Y ( f ) = X ( f ( ⋅ / ℓ D )) . When X X X has the law Gin β \operatorname{Gin}_{\beta} Gin β , the configuration Y Y Y has intensity ℓ D 2 = ( 2 π β ) − 1 \ell_{\mathrm{D}}^2=(2\pi\beta)^{-1} ℓ D 2 = ( 2 π β ) − 1 .
For the finite gas, we use the soft confinement of Subsection 2.1. Let N ≥ 1 N \ge1 N ≥ 1 and 0 ≤ t ≤ 1 0 \le t \le1 0 ≤ t ≤ 1 . Let g 1 g_1 g 1 be the smoothed logarithm of Subsection 5.1, with ε = 1 \varepsilon=1 ε = 1 , and let P t \mathbb{P}_t P t be the probability measure on ( R 2 ) N (\mathbb{R}^2)^N ( R 2 ) N with density proportional to
exp ( − 1 4 ∑ j = 1 N ∣ y j ∣ 2 − β ∑ i < j g ( y i − y j ) − 2 t β ∑ j = 1 N g 1 ( y j ) ) (68) \exp\left(-\frac{1}{4}\sum_{j=1}^{N}|y_j|^2-\beta\sum_{i<j}g(y_i-y_j)-2t\beta\sum_{j=1}^{N}g_1(y_j)\right) \tag*{(68)} exp ( − 4 1 j = 1 ∑ N ∣ y j ∣ 2 − β i < j ∑ g ( y i − y j ) − 2 tβ j = 1 ∑ N g 1 ( y j ) ) (68)
We write Y : = ∑ j = 1 N δ y j Y:=\sum_{j=1}^{N}\delta_{y_j} Y := ∑ j = 1 N δ y j , and E t \mathbb{E}_t E t , Var t \operatorname{Var}_t Var t and Cov t \operatorname{Cov}_t Cov t for the expectation, the variance and the covariance under P t \mathbb{P}_t P t . The measure P t \mathbb{P}_t P t also depends on N N N and β \beta β , which we do not write. If ( x 1 , … , x N ) (x_1,\ldots,x_N) ( x 1 , … , x N ) has the law G N , L G_{N,L} G N , L of (23) without the constraint ∣ x j ∣ < L |x_j|<L ∣ x j ∣ < L , then ( x 1 / ℓ D , … , x N / ℓ D ) (x_1/\ell_{\mathrm{D}},\ldots,x_N/\ell_{\mathrm{D}}) ( x 1 / ℓ D , … , x N / ℓ D ) has the law P 0 \mathbb{P}_0 P 0 , since π β ℓ D 2 / 2 = 1 / 4 \pi\beta\ell_{\mathrm{D}}^2/2=1/4 π β ℓ D 2 /2 = 1/4 . The measure P t \mathbb{P}_t P t is rotation-invariant, because g 1 g_1 g 1 is radial. In Subsection 8.4, the potential 2 t β Y ( g 1 ) 2t\beta Y(g_1) 2 tβ Y ( g 1 ) compensates the logarithmic growth of the potential of two inserted particles. For a function f f f on R 2 \mathbb{R}^2 R 2 , we write
Fluct t [ f ] : = Y ( f ) − E t Y ( f ) . \operatorname{Fluct}_t[f]:=Y(f)-\mathbb{E}_tY(f). Fluct t [ f ] := Y ( f ) − E t Y ( f ) .
The next lemma collects what we need on P t \mathbb{P}_t P t . The Poincaré inequality of § 6 and the decay of covariances of § 8 still hold with the additional potential, and we also bound the number of points and the moments of linear statistics.
Lemma 8.1 (Poincaré inequality and moments on the Debye scale). There exists β 0 > 0 \beta_0>0 β 0 > 0 such that, for every 0 < β < β 0 0<\beta<\beta_0 0 < β < β 0 , every N ≥ 4 / β N\ge4/\beta N ≥ 4/ β , every 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , every p ≥ 2 p\ge2 p ≥ 2 and every smooth symmetric function F F F on ( R 2 ) N (\mathbb{R}^2)^N ( R 2 ) N ,
Var t [ F ] ≤ C D P t ( F ) and ∥ F − E t F ∥ p ≤ C p ∥ ( ∑ i = 1 N ∣ ∇ i F ∣ 2 ) 1 / 2 ∥ p . (69) \operatorname{Var}_t[F]\le C\mathcal{D}_{\mathbb{P}_t}(F)
\quad\text{and}\quad
\lVert F-\mathbb{E}_tF\rVert_p\le Cp\left\lVert\left(\sum_{i=1}^{N}|\nabla_iF|^2\right)^{1/2}\right\rVert_p . \tag*{(69)} Var t [ F ] ≤ C D P t ( F ) and ∥ F − E t F ∥ p ≤ C p ( i = 1 ∑ N ∣ ∇ i F ∣ 2 ) 1/2 p . (69) Moreover, let R , d > 0 R,d>0 R , d > 0 . If F F F is the restriction of a smooth B R B_R B R -local observable, and G G G that of a smooth observable that is local to the complement of B R + d B_{R+d} B R + d , then
∣ Cov t [ F , G ] ∣ ≤ C e − c d D P t ( F ) 1 / 2 D P t ( G ) 1 / 2 . (70) \left|\operatorname{Cov}_t[F,G]\right|\le Ce^{-c d}\mathcal{D}_{\mathbb{P}_t}(F)^{1/2}\mathcal{D}_{\mathbb{P}_t}(G)^{1/2}. \tag*{(70)} ∣ Cov t [ F , G ] ∣ ≤ C e − c d D P t ( F ) 1/2 D P t ( G ) 1/2 . (70) For every x ∈ R 2 x\in\mathbb{R}^2 x ∈ R 2 , every R > 0 R>0 R > 0 , every 0 ≤ γ ≤ 1 0\le\gamma\le1 0 ≤ γ ≤ 1 and every j ≥ 1 j\ge1 j ≥ 1 ,
E t e γ Y ( B ( x , R ) ) ≤ e C γ ℓ D 2 R 2 and ρ j , t ≤ C ( j ) ℓ D 2 j , (71) \mathbb{E}_t e^{\gamma Y(B(x,R))}\le e^{C\gamma\ell_{\mathrm{D}}^2R^2}
\quad\text{and}\quad
\rho_{j,t}\le C(j)\ell_{\mathrm{D}}^{2j}, \tag*{(71)} E t e γ Y ( B ( x , R )) ≤ e C γ ℓ D 2 R 2 and ρ j , t ≤ C ( j ) ℓ D 2 j , (71) where ρ j , t \rho_{j,t} ρ j , t is the j j j -point correlation function of Y Y Y under P t \mathbb{P}_t P t . These two bounds are due to Thoma [86 , 87 ] . For every even p ≥ 2 p\ge2 p ≥ 2 and every bounded Lipschitz function f f f on R 2 \mathbb{R}^2 R 2 with ∇ f ∈ L 2 ( R 2 ) ∩ L p ( R 2 ) \nabla f\in L^2(\mathbb{R}^2)\cap L^p(\mathbb{R}^2) ∇ f ∈ L 2 ( R 2 ) ∩ L p ( R 2 ) ,
∥ Fluct t [ f ] ∥ p ≤ C ( p ) ( ℓ D ∥ ∇ f ∥ 2 + ℓ D 2 / p ∥ ∇ f ∥ p ) . (72) \lVert\operatorname{Fluct}_t[f]\rVert_p\le C(p)\left(\ell_{\mathrm{D}}\lVert\nabla f\rVert_2+\ell_{\mathrm{D}}^{2/p}\lVert\nabla f\rVert_p\right). \tag*{(72)} ∥ Fluct t [ f ] ∥ p ≤ C ( p ) ( ℓ D ∥ ∇ f ∥ 2 + ℓ D 2/ p ∥ ∇ f ∥ p ) . (72) The inequalities (69) , (70) and (72) also hold for local observables of Y Y Y when X ∼ Gin β X\sim\operatorname{Gin}_{\beta} X ∼ Gin β , with Y ( f ) − ℓ D 2 ∫ f Y(f)-\ell_{\mathrm{D}}^2\int f Y ( f ) − ℓ D 2 ∫ f in place of Fluct t [ f ] \operatorname{Fluct}_t[f] Fluct t [ f ] . In these inequalities, locality and gradients refer to the rescaled positions y y y ,
and the Dirichlet energy is E G i n β ∑ y ∈ Y ∣ ∇ y F ( Y ) ∣ 2 E_{\mathrm{Gin}_{\beta}}\sum_{y\in Y}|\nabla_y F(Y)|^2 E Gin β ∑ y ∈ Y ∣ ∇ y F ( Y ) ∣ 2 . The bound on the correlation functions in (71) also holds. When f f f is not compactly supported, this centred statistic is the limit in L p ( G i n β ) L^p(\mathrm{Gin}_{\beta}) L p ( Gin β ) of the centred statistics of χ R f \chi_R f χ R f as R → ∞ R\to\infty R → ∞ , which exists when ∥ ∇ ( ( 1 − χ R ) f ) ∥ 2 \|\nabla((1-\chi_R)f)\|_2 ∥∇ (( 1 − χ R ) f ) ∥ 2 and ∥ ∇ ( ( 1 − χ R ) f ) ∥ p \|\nabla((1-\chi_R)f)\|_p ∥∇ (( 1 − χ R ) f ) ∥ p tend to zero.
By approximation, (69) and (70) also hold for F = Y ( f ) F=Y(f) F = Y ( f ) and G = Y ( h ) G=Y(h) G = Y ( h ) , when f f f and h h h are bounded Lipschitz functions with compact support (and the corresponding locality for (70) ). Indeed, use the mollifier ψ δ \psi_\delta ψ δ of Subsection 5.1. Then Y ( f ∗ ψ δ ) → δ → 0 Y ( f ) Y(f*\psi_\delta)\xrightarrow[\delta\to0]{}Y(f) Y ( f ∗ ψ δ ) δ → 0 Y ( f ) pointwise, with ∣ Y ( f ∗ ψ δ ) ∣ ≤ N ∥ f ∥ ∞ |Y(f*\psi_\delta)|\le N\|f\|_\infty ∣ Y ( f ∗ ψ δ ) ∣ ≤ N ∥ f ∥ ∞ , and ∇ ( f ∗ ψ δ ) = ( ∇ f ) ∗ ψ δ → δ → 0 ∇ f \nabla(f*\psi_\delta)=(\nabla f)*\psi_\delta\xrightarrow[\delta\to0]{}\nabla f ∇ ( f ∗ ψ δ ) = ( ∇ f ) ∗ ψ δ δ → 0 ∇ f almost everywhere, with ∣ ∇ ( f ∗ ψ δ ) ∣ ≤ ∥ ∇ f ∥ ∞ |\nabla(f*\psi_\delta)|\le\|\nabla f\|_\infty ∣∇ ( f ∗ ψ δ ) ∣ ≤ ∥∇ f ∥ ∞ . Since the positions have a density under P t \mathbb{P}_t P t , dominated convergence applies to both sides. We use this for the function A A A of (88) .
The proof is given in Appendix B.8.
Structure factor on the Debye scale
We start with the averages of q β − 1 q_\beta-1 q β − 1 , which we read off the covariances C β ( f , h ) = ℓ D − 2 Cov G i n β [ Y ( f ) , Y ( h ) ] C_\beta(f,h)=\ell_D^{-2}\operatorname{Cov}_{\mathrm{Gin}_\beta}[Y(f),Y(h)] C β ( f , h ) = ℓ D − 2 Cov Gin β [ Y ( f ) , Y ( h )] of Subsection 3.4, for f , h ∈ C c ∞ ( R 2 ; R ) f,h\in C^\infty_c(\mathbb{R}^2;\mathbb{R}) f , h ∈ C c ∞ ( R 2 ; R ) .
Proposition 8.2 (Structure factor on the Debye scale). Let 0 < β ≤ 2 0<\beta\le2 0 < β ≤ 2 . The structure factor σ G i n β \sigma_{\mathrm{Gin}_\beta} σ Gin β of Subsection 2.2 has a continuous radial density, which we also denote by σ G i n β \sigma_{\mathrm{Gin}_\beta} σ Gin β , and
0 ≤ σ G i n β ( k / ℓ D ) ≤ ∣ k ∣ 2 for every k ∈ R 2 . (73) 0\le\sigma_{\mathrm{Gin}_\beta}(k/\ell_D)\le|k|^2\qquad\text{for every }k\in\mathbb{R}^2. \tag*{(73)} 0 ≤ σ Gin β ( k / ℓ D ) ≤ ∣ k ∣ 2 for every k ∈ R 2 . (73) For every f , h ∈ C c ∞ ( R 2 ; R ) f,h\in C^\infty_c(\mathbb{R}^2;\mathbb{R}) f , h ∈ C c ∞ ( R 2 ; R ) ,
C β ( f , h ) = 1 ( 2 π ) 2 ∫ R 2 σ G i n β ( k / ℓ D ) f ^ ( k ) h ^ ( k ) ‾ d k . (74) C_\beta(f,h)=\frac{1}{(2\pi)^2}\int_{\mathbb{R}^2}\sigma_{\mathrm{Gin}_\beta}(k/\ell_D)\widehat{f}(k)\overline{\widehat{h}(k)}\,\mathrm{d}k. \tag*{(74)} C β ( f , h ) = ( 2 π ) 2 1 ∫ R 2 σ Gin β ( k / ℓ D ) f ( k ) h ( k ) d k . (74) Moreover, as β ↓ 0 \beta\downarrow0 β ↓ 0 , the functions σ G i n β ( ⋅ / ℓ D ) \sigma_{\mathrm{Gin}_\beta}(\cdot/\ell_D) σ Gin β ( ⋅ / ℓ D ) converge to k ↦ ∣ k ∣ 2 / ( 1 + ∣ k ∣ 2 ) k\mapsto|k|^2/(1+|k|^2) k ↦ ∣ k ∣ 2 / ( 1 + ∣ k ∣ 2 ) in S ′ ( R 2 ) \mathcal{S}'(\mathbb{R}^2) S ′ ( R 2 ) , and
ρ 2 , β ( 0 , ℓ D ⋅ ) − 1 β → β ↓ 0 − K 0 ( ∣ ⋅ ∣ ) in S ′ ( R 2 ) . (75) \frac{\rho_{2,\beta}(0,\ell_D\cdot)-1}{\beta}\xrightarrow[\beta\downarrow0]{}-K_0(|\cdot|)\qquad\text{in }\mathcal{S}'(\mathbb{R}^2). \tag*{(75)} β ρ 2 , β ( 0 , ℓ D ⋅ ) − 1 β ↓ 0 − K 0 ( ∣ ⋅ ∣ ) in S ′ ( R 2 ) . (75)
The idea of the proof of Proposition 8.2 is to move the points of Y Y Y along a vector field ∇ f \nabla f ∇ f and to integrate by parts, as in (B.12). The interaction of the moved points with the mean density gives back the statistic Y ( f ) Y(f) Y ( f ) , and the interaction between the points produces a quadratic statistic of Y Y Y , which we now define. Let Y [ 2 ] Y^{[2]} Y [ 2 ] be the sum of the Dirac masses δ ( y , z ) \delta_{(y,z)} δ ( y , z ) over the ordered pairs of distinct points y , z y,z y , z of Y Y Y . For f ∈ C c ∞ ( R 2 ; R ) f\in C^\infty_c(\mathbb{R}^2;\mathbb{R}) f ∈ C c ∞ ( R 2 ; R ) and R > 0 R>0 R > 0 , let Q f , R Q_{f,R} Q f , R be the centred version, under G i n β \mathrm{Gin}_\beta Gin β , of
ℓ D − 2 2 ∬ χ R ( x ) χ R ( z ) ( ∇ f ( x ) − ∇ f ( z ) ) ⋅ ∇ g ( x − z ) ( Y [ 2 ] ( d x d z ) − ℓ D 2 Y ( d x ) d z − ℓ D 2 d x Y ( d z ) + ℓ D 4 d x d z ) . (76) \frac{\ell_D^{-2}}{2}\iint\chi_R(x)\chi_R(z)\bigl(\nabla f(x)-\nabla f(z)\bigr)\cdot\nabla g(x-z)\bigl(Y^{[2]}(\mathrm{d}x\,\mathrm{d}z)-\ell_D^2Y(\mathrm{d}x)\,\mathrm{d}z-\ell_D^2\mathrm{d}x\,Y(\mathrm{d}z)+\ell_D^4\mathrm{d}x\,\mathrm{d}z\bigr). \tag*{(76)} 2 ℓ D − 2 ∬ χ R ( x ) χ R ( z ) ( ∇ f ( x ) − ∇ f ( z ) ) ⋅ ∇ g ( x − z ) ( Y [ 2 ] ( d x d z ) − ℓ D 2 Y ( d x ) d z − ℓ D 2 d x Y ( d z ) + ℓ D 4 d x d z ) . (76)
Since ∣ ∇ f ( x ) − ∇ f ( z ) ∣ ≤ C ∣ x − z ∣ |\nabla f(x)-\nabla f(z)|\le C|x-z| ∣∇ f ( x ) − ∇ f ( z ) ∣ ≤ C ∣ x − z ∣ , the kernel of this integral is bounded.
Lemma 8.3 (Bounds for the quadratic statistic). Let f ∈ C c ∞ ( R 2 ; R ) f\in C^\infty_c(\mathbb{R}^2;\mathbb{R}) f ∈ C c ∞ ( R 2 ; R ) . There exist β 0 > 0 \beta_0>0 β 0 > 0 and C > 0 C>0 C > 0 such that, for every 0 < β < β 0 0<\beta<\beta_0 0 < β < β 0 , the statistics Q f , R Q_{f,R} Q f , R converge in L 2 ( G i n β ) L^2(\mathrm{Gin}_\beta) L 2 ( Gin β ) as R → ∞ R\to\infty R → ∞ , and their limit Q f Q_f Q f satisfies ∥ Q f − Q f , R ∥ 2 ≤ C ( f , β ) R − 1 \|Q_f-Q_{f,R}\|_2\le C(f,\beta)R^{-1} ∥ Q f − Q f , R ∥ 2 ≤ C ( f , β ) R − 1 for every R ≥ 1 R\ge1 R ≥ 1 . Let ℓ D − 1 ≤ δ ≤ 1 \ell_D^{-1}\le\delta\le1 ℓ D − 1 ≤ δ ≤ 1 , and let Q f n e a r Q_f^{\mathrm{near}} Q f near and Q f f a r Q_f^{\mathrm{far}} Q f far be defined as Q f Q_f Q f , with the kernel multiplied by χ δ ( x − z ) \chi_\delta(x-z) χ δ ( x − z ) and by 1 − χ δ ( x − z ) 1-\chi_\delta(x-z) 1 − χ δ ( x − z ) , respectively. Then
∥ Q f n e a r ∥ 2 ≤ C ( ℓ D 2 δ 2 + δ ) and ∥ Q f f a r ∥ 2 ≤ C δ − 1 . (77) \|Q_f^{\mathrm{near}}\|_2\le C(\ell_D^2\delta^2+\delta)\qquad\text{and}\qquad\|Q_f^{\mathrm{far}}\|_2\le C\delta^{-1}. \tag*{(77)} ∥ Q f near ∥ 2 ≤ C ( ℓ D 2 δ 2 + δ ) and ∥ Q f far ∥ 2 ≤ C δ − 1 . (77)
The proof is given in Appendix B.8.
Remark 8.4 (The anisotropy and the Poincaré inequality). Up to a constant factor and the truncation by χ R \chi_R χ R , the statistic Q f , R Q_{f,R} Q f , R is the centred anisotropy of Leblé and Serfaty [69 Definition 4.1] for the transport ∇ f \nabla f ∇ f , in the form of [80 Lemma 4.1] . The anisotropy is the first variation of the energy when the background moves with the points. In our integration by parts, the background stays in place, and this produces the term Y ( f ) Y(f) Y ( f ) . The anisotropy is the main error term in the proofs of the central limit theorem for linear statistics of [69 , 80 ] . The proof of Lemma 8.3 treats it in two parts. The pairs of points at distance at least δ \delta δ form a statistic whose gradient in each position is itself a centred linear statistic. The Poincaré inequality, combined with the bounds of Lemma 8.1 on the number of points and on the moments of linear statistics, therefore bounds its variance by C δ − 2 C\delta^{-2} C δ − 2 , uniformly in β \beta β small. The pairs at distance less than δ \delta δ are only controlled by the bounds on the number of points in small disks, which ignore the cancellations. As a result, our bound on ∥ Q f ∥ 2 \lVert Q_f\rVert_2 ∥ Q f ∥ 2 is of order ℓ D 2 / 3 \ell_{\mathrm{D}}^{2/3} ℓ D 2/3 , which is not small: it suffices here only because Q f Q_f Q f comes with the factor β \beta β in (80) .
The identity ((80) ), formally. Before the proof, we derive formally the identity (80) on which it rests, for the finite gas. Recall that Y Y Y has intensity ℓ D 2 \ell_{\mathrm{D}}^2 ℓ D 2 : on the Debye scale, the gas is dense. We compute formally with the finite gas P 0 \mathbb{P}_0 P 0 of (68) , whose density on ( R 2 ) N (\mathbb{R}^2)^N ( R 2 ) N is proportional to e − Φ e^{-\Phi} e − Φ , with
Φ ( y ) : = 1 4 ∑ j ∣ y j ∣ 2 + β ∑ i < j g ( y i − y j ) . \Phi(y):=\frac{1}{4}\sum_j \lvert y_j\rvert^2+\beta\sum_{i<j}g(y_i-y_j). Φ ( y ) := 4 1 j ∑ ∣ y j ∣ 2 + β i < j ∑ g ( y i − y j ) .
Let v v v be a smooth vector field with compact support, and let G G G be a smooth observable. The integral ∫ G e − Φ d y \int G e^{-\Phi}\,dy ∫ G e − Φ d y does not change under the change of variables y j ↦ y j + τ v ( y j ) y_j\mapsto y_j+\tau v(y_j) y j ↦ y j + τ v ( y j ) , which acts on the observable, on the Gibbs weight, and on the Lebesgue measure through its Jacobian. Differentiating at τ = 0 \tau=0 τ = 0 , we get
E 0 [ ∑ j v ( y j ) ⋅ ∇ j G ] = E 0 [ G ( ∑ j v ( y j ) ⋅ ∇ j Φ − Y ( div v ) ) ] . (78) \mathbb{E}_0\left[\sum_j v(y_j)\cdot\nabla_jG\right]=\mathbb{E}_0\left[G\left(\sum_j v(y_j)\cdot\nabla_j\Phi-Y(\operatorname{div}v)\right)\right]. \tag*{(78)} E 0 [ j ∑ v ( y j ) ⋅ ∇ j G ] = E 0 [ G ( j ∑ v ( y j ) ⋅ ∇ j Φ − Y ( div v ) ) ] . (78)
The left-hand side is the variation of the observable, and the right-hand side comes from the variation of the weight and from the Jacobian. Taking G = 1 G=1 G = 1 shows that the factor of G G G on the right-hand side is centred, hence the right-hand side is a covariance. This is the Ward identity of Subsection 3.4.
We apply (78) with v = ∇ f v=\nabla f v = ∇ f and G = Y ( h ) G=Y(h) G = Y ( h ) . The roles of f f f and h h h are not symmetric: f f f defines the transport, and h h h the observable which is transported. Since ∇ j Y ( h ) = ∇ h ( y j ) \nabla_jY(h)=\nabla h(y_j) ∇ j Y ( h ) = ∇ h ( y j ) , the left-hand side is E 0 [ Y ( ∇ f ⋅ ∇ h ) ] = ℓ D 2 ∫ ∇ f ⋅ ∇ h \mathbb{E}_0[Y(\nabla f\cdot\nabla h)]=\ell_{\mathrm{D}}^2\int\nabla f\cdot\nabla h E 0 [ Y ( ∇ f ⋅ ∇ h )] = ℓ D 2 ∫ ∇ f ⋅ ∇ h . On the right-hand side, we have Y ( div v ) = Y ( Δ f ) Y(\operatorname{div}v)=Y(\Delta f) Y ( div v ) = Y ( Δ f ) , and, since ∇ g \nabla g ∇ g is odd,
∑ j ∇ f ( y j ) ⋅ ∇ j Φ = 1 2 Y ( y ⋅ ∇ f ) + β 2 ∑ i ≠ j K f ( y i , y j ) \sum_j\nabla f(y_j)\cdot\nabla_j\Phi=\frac{1}{2}Y(y\cdot\nabla f)+\frac{\beta}{2}\sum_{i\ne j}K_f(y_i,y_j) j ∑ ∇ f ( y j ) ⋅ ∇ j Φ = 2 1 Y ( y ⋅ ∇ f ) + 2 β i = j ∑ K f ( y i , y j )
with K f ( x , z ) : = ( ∇ f ( x ) − ∇ f ( z ) ) ⋅ ∇ g ( x − z ) K_f(x,z):=(\nabla f(x)-\nabla f(z))\cdot\nabla g(x-z) K f ( x , z ) := ( ∇ f ( x ) − ∇ f ( z )) ⋅ ∇ g ( x − z ) .
To understand the pair term, we replace, in each pair, one of the two points by the mean density ℓ D 2 \ell_{\mathrm{D}}^2 ℓ D 2 of the gas (a mean-field approximation). Since − Δ g = 2 π δ 0 -\Delta g=2\pi\delta_0 − Δ g = 2 π δ 0 , and by Newton’s theorem (the integral in z z z being over the disk occupied by the gas),
β ℓ D 2 ∫ K f ( x , z ) d z = β ℓ D 2 ( 2 π f ( x ) − π x ⋅ ∇ f ( x ) ) = f ( x ) − 1 2 x ⋅ ∇ f ( x ) , \beta\ell_{\mathrm{D}}^2\int K_f(x,z)\,dz=\beta\ell_{\mathrm{D}}^2\left(2\pi f(x)-\pi x\cdot\nabla f(x)\right)=f(x)-\frac{1}{2}x\cdot\nabla f(x), β ℓ D 2 ∫ K f ( x , z ) d z = β ℓ D 2 ( 2 π f ( x ) − π x ⋅ ∇ f ( x ) ) = f ( x ) − 2 1 x ⋅ ∇ f ( x ) ,
where we used 2 π β ℓ D 2 = 1 2\pi\beta\ell_{\mathrm{D}}^2=1 2 π β ℓ D 2 = 1 . The first term, 2 π ℓ D 2 f 2\pi\ell_{\mathrm{D}}^2f 2 π ℓ D 2 f , is the potential created by the charge − div ( ℓ D 2 ∇ f ) = − ℓ D 2 Δ f -\operatorname{div}(\ell_{\mathrm{D}}^2\nabla f)=-\ell_{\mathrm{D}}^2\Delta f − div ( ℓ D 2 ∇ f ) = − ℓ D 2 Δ f , that is, by the displacement of the mean density. The second term is the work of the field of the mean density on the moving point, and it cancels the confinement term 1 2 Y ( y ⋅ ∇ f ) \frac{1}{2}Y(y\cdot\nabla f) 2 1 Y ( y ⋅ ∇ f ) . Hence, up to a constant,
∑ j ∇ f ( y j ) ⋅ ∇ j Φ = Y ( f ) + β ℓ D 2 Q f , \sum_j\nabla f(y_j)\cdot\nabla_j\Phi=Y(f)+\beta\ell_{\mathrm{D}}^2Q_f, j ∑ ∇ f ( y j ) ⋅ ∇ j Φ = Y ( f ) + β ℓ D 2 Q f ,
where Q f Q_f Q f , the error of the mean-field approximation, is a centred quadratic statistic of the pairs of points of Y Y Y , see (76) . Inserting this into (78) , where the left-hand side is ℓ D 2 ∫ ∇ f ⋅ ∇ h \ell_{\mathrm{D}}^2\int\nabla f\cdot\nabla h ℓ D 2 ∫ ∇ f ⋅ ∇ h and the right-hand side is the covariance of Y ( h ) Y(h) Y ( h ) with Y ( f ) − Y ( Δ f ) + β ℓ D 2 Q f Y(f)-Y(\Delta f)+\beta\ell_{\mathrm{D}}^2Q_f Y ( f ) − Y ( Δ f ) + β ℓ D 2 Q f , and dividing by ℓ D 2 \ell_{\mathrm{D}}^2 ℓ D 2 , we obtain (80) below. The following lemma makes this computation rigorous under Gin β \operatorname{Gin}_\beta Gin β . Recall from (3) that ℓ D Z β ( h ) = Y ( h ) − ℓ D 2 ∫ h \ell_{\mathrm{D}}Z_\beta(h)=Y(h)-\ell_{\mathrm{D}}^2\int h ℓ D Z β ( h ) = Y ( h ) − ℓ D 2 ∫ h .
Lemma 8.5 (Ward identity for cylinder observables). Let f ∈ C c ∞ ( R 2 ; R ) f\in C^\infty_c(\mathbb{R}^2;\mathbb{R}) f ∈ C c ∞ ( R 2 ; R ) , and let β 0 > 0 \beta_0>0 β 0 > 0 be as in Lemma 8.3. For every 0 < β < min { β 0 , 2 } 0<\beta<\min\{\beta_0,2\} 0 < β < min { β 0 , 2 } and every cylinder observable G G G of the form (8) , we have
E Gin β [ ∑ y ∈ Y ∇ f ( y ) ⋅ ∇ y G ( Y ) ] = ℓ D E Gin β [ Z β ( ( 1 − Δ f ) G ( Y ) ) ] + β ℓ D 2 E Gin β [ Q f G ( Y ) ] . (79) \mathbb{E}_{\operatorname{Gin}_\beta}\left[\sum_{y\in Y}\nabla f(y)\cdot\nabla_yG(Y)\right]=\ell_{\mathrm{D}}\mathbb{E}_{\operatorname{Gin}_\beta}\left[Z_\beta((1-\Delta f)G(Y))\right]+\beta\ell_{\mathrm{D}}^2\mathbb{E}_{\operatorname{Gin}_\beta}\left[Q_fG(Y)\right]. \tag*{(79)} E Gin β y ∈ Y ∑ ∇ f ( y ) ⋅ ∇ y G ( Y ) = ℓ D E Gin β [ Z β (( 1 − Δ f ) G ( Y )) ] + β ℓ D 2 E Gin β [ Q f G ( Y ) ] . (79)
Proof. For the kernel K f ( x , z ) = ( ∇ f ( x ) − ∇ f ( z ) ) ⋅ ∇ g ( x − z ) K_f(x,z)=(\nabla f(x)-\nabla f(z))\cdot\nabla g(x-z) K f ( x , z ) = ( ∇ f ( x ) − ∇ f ( z )) ⋅ ∇ g ( x − z ) , set
R f , R : = β 2 ∬ χ R ( x ) χ R ( z ) K f ( x , z ) Y [ 2 ] ( d x d z ) + Y ( 1 2 x ⋅ ∇ f − Δ f ) . \mathcal{R}_{f,R}:=\frac{\beta}{2}\iint\chi_R(x)\chi_R(z)K_f(x,z)Y^{[2]}(dx\,dz)+Y\left(\frac{1}{2}x\cdot\nabla f-\Delta f\right). R f , R := 2 β ∬ χ R ( x ) χ R ( z ) K f ( x , z ) Y [ 2 ] ( d x d z ) + Y ( 2 1 x ⋅ ∇ f − Δ f ) .
Let b > 0 b>0 b > 0 be such that the disk B b B_b B b contains the support of f f f , and let R ≥ 2 b R\ge2b R ≥ 2 b , so that χ R = 1 \chi_R=1 χ R = 1 on B 2 b B_{2b} B 2 b .
Integration by parts. After rescaling, Y Y Y satisfies the DLR equations (13) under Gin β \operatorname{Gin}_{\beta} Gin β , with the interaction β g \beta g β g and the background ℓ D 2 Leb \ell_{\mathrm{D}}^{2}\operatorname{Leb} ℓ D 2 Leb . We condition on the configuration outside Λ : = B 2 R \Lambda:= B_{2R} Λ := B 2 R and on the number of points in Λ \Lambda Λ , and we integrate by parts in the positions of these points, along the field ∇ f \nabla f ∇ f . Since ∇ f \nabla f ∇ f is supported in B b B_b B b , there is no boundary term on ∂ Λ \partial\Lambda ∂ Λ . For a collision cutoff at ∣ y − z ∣ ≃ δ |y-z|\simeq\delta ∣ y − z ∣ ≃ δ , the factor δ − 1 \delta^{-1} δ − 1 is compensated by ∣ ∇ f ( y ) − ∇ f ( z ) ∣ ≤ C ∣ y − z ∣ |\nabla f(y)-\nabla f(z)|\leq C|y-z| ∣∇ f ( y ) − ∇ f ( z ) ∣ ≤ C ∣ y − z ∣ , and the cutoff error tends to zero by dominated convergence. The derivative of the conditional energy along ∇ f \nabla f ∇ f is
β ∑ y ∈ Y ∇ f ( y ) ⋅ ∫ z ≠ y ∇ g ( y − z ) d ( Y − ℓ D 2 Leb ) ( z ) \beta\sum_{y\in Y}\nabla f(y)\cdot\int_{z\ne y}\nabla g(y-z)\,\mathrm{d}(Y-\ell_{\mathrm{D}}^{2}\operatorname{Leb})(z) β y ∈ Y ∑ ∇ f ( y ) ⋅ ∫ z = y ∇ g ( y − z ) d ( Y − ℓ D 2 Leb ) ( z )
and the Jacobian gives − Y ( Δ f ) -Y(\Delta f) − Y ( Δ f ) . We split the integral with χ R ( z ) + ( 1 − χ R ( z ) ) \chi_R(z)+(1-\chi_R(z)) χ R ( z ) + ( 1 − χ R ( z )) . In the part with χ R ( z ) \chi_R(z) χ R ( z ) , we may insert χ R ( y ) \chi_R(y) χ R ( y ) , which equals 1 on the support of ∇ f \nabla f ∇ f , and we symmetrize the sum over pairs: since ∇ g \nabla g ∇ g is odd, the points give β 2 ∑ y ≠ z χ R ( y ) χ R ( z ) K f ( y , z ) \frac{\beta}{2}\sum_{y\ne z}\chi_R(y)\chi_R(z)K_f(y,z) 2 β ∑ y = z χ R ( y ) χ R ( z ) K f ( y , z ) . For y ∈ B b y\in B_b y ∈ B b , Newton’s theorem gives
∫ χ R ( z ) ∇ g ( y − z ) d z = − π y , \int\chi_R(z)\nabla g(y-z)\,\mathrm{d}z=-\pi y, ∫ χ R ( z ) ∇ g ( y − z ) d z = − π y ,
hence the background gives β π ℓ D 2 Y ( y ⋅ ∇ f ) = 1 2 Y ( y ⋅ ∇ f ) \beta\pi\ell_{\mathrm{D}}^{2}Y(y\cdot\nabla f)=\frac{1}{2}Y(y\cdot\nabla f) β π ℓ D 2 Y ( y ⋅ ∇ f ) = 2 1 Y ( y ⋅ ∇ f ) . We obtain
E Gin β [ ∑ y ∈ Y ∇ f ( y ) ⋅ ∇ y G ( Y ) ] = E Gin β [ G ( Y ) ( R f , R + β E R ) ] , E R : = ∑ y ∈ Y ∇ f ( y ) ⋅ V R ( y ) , \mathbb{E}_{\operatorname{Gin}_{\beta}}\left[\sum_{y\in Y}\nabla f(y)\cdot\nabla_yG(Y)\right]
=
\mathbb{E}_{\operatorname{Gin}_{\beta}}\left[G(Y)\left(\mathcal{R}_{f,R}+\beta\mathcal{E}_R\right)\right],
\qquad
\mathcal{E}_R:=\sum_{y\in Y}\nabla f(y)\cdot V_R(y), E Gin β y ∈ Y ∑ ∇ f ( y ) ⋅ ∇ y G ( Y ) = E Gin β [ G ( Y ) ( R f , R + β E R ) ] , E R := y ∈ Y ∑ ∇ f ( y ) ⋅ V R ( y ) ,
where V R ( y ) : = ∫ ( 1 − χ R ( z ) ) ∇ g ( y − z ) d ( Y − ℓ D 2 Leb ) ( z ) V_R(y):=\int(1-\chi_R(z))\nabla g(y-z)\,\mathrm{d}(Y-\ell_{\mathrm{D}}^{2}\operatorname{Leb})(z) V R ( y ) := ∫ ( 1 − χ R ( z )) ∇ g ( y − z ) d ( Y − ℓ D 2 Leb ) ( z ) is the field created by the points and the background far from the support of f f f . This integral is defined with the dyadic partition ( ϑ j ) j ≥ 0 (\vartheta_j)_{j\geq0} ( ϑ j ) j ≥ 0 of Remark 2.2. Taking G = 1 G=1 G = 1 shows that R f , R + β E R \mathcal{R}_{f,R}+\beta\mathcal{E}_R R f , R + β E R is centred.
The exterior field. Let y ∈ B b y\in B_b y ∈ B b , j ≥ 0 j\geq0 j ≥ 0 , and ∣ α ∣ ≤ 1 |\alpha|\leq1 ∣ α ∣ ≤ 1 . The function z ↦ ( 1 − χ R ( z ) ) ϑ j ( z ) ∂ y α ∇ g ( y − z ) z\mapsto(1-\chi_R(z))\vartheta_j(z)\partial_y^\alpha\nabla g(y-z) z ↦ ( 1 − χ R ( z )) ϑ j ( z ) ∂ y α ∇ g ( y − z ) vanishes unless 2 j ≥ c R 2^j\geq cR 2 j ≥ c R . It is supported in a disk of radius C 2 j C2^j C 2 j , and its derivatives of order k ≤ 3 k\leq3 k ≤ 3 are bounded by C 2 − j ( 1 + ∣ α ∣ + k ) C2^{-j(1+|\alpha|+k)} C 2 − j ( 1 + ∣ α ∣ + k ) . By (G2) for Gin β \operatorname{Gin}_{\beta} Gin β , after rescaling, the L 4 ( Gin β ) L^4(\operatorname{Gin}_{\beta}) L 4 ( Gin β ) norm of its centred statistic is at most C ( β ) 2 − j C(\beta)2^{-j} C ( β ) 2 − j . We sum over j j j , and use the Sobolev embedding W 1 , 4 ( B b ) ↪ L ∞ ( B b ) W^{1,4}(B_b)\hookrightarrow L^\infty(B_b) W 1 , 4 ( B b ) ↪ L ∞ ( B b ) in the variable y y y . This gives
∥ sup y ∈ B b ∣ V R ( y ) ∣ ∥ L 4 ( Gin β ) ≤ C ( b , β ) R − 1 . \left\|\sup_{y\in B_b}|V_R(y)|\right\|_{L^4(\operatorname{Gin}_{\beta})}\leq C(b,\beta)R^{-1}. y ∈ B b sup ∣ V R ( y ) ∣ L 4 ( Gin β ) ≤ C ( b , β ) R − 1 .
With the fourth moment of Y ( B b ) Y(B_b) Y ( B b ) from (G1), we obtain ∥ E R ∥ L 2 ( Gin β ) ≤ C ( f , β ) R − 1 \|\mathcal{E}_R\|_{L^2(\operatorname{Gin}_{\beta})}\leq C(f,\beta)R^{-1} ∥ E R ∥ L 2 ( Gin β ) ≤ C ( f , β ) R − 1 . Therefore
E Gin β [ ∑ y ∈ Y ∇ f ( y ) ⋅ ∇ y G ( Y ) ] = lim R → ∞ Cov Gin β [ G ( Y ) , R f , R ] . \mathbb{E}_{\operatorname{Gin}_{\beta}}\left[\sum_{y\in Y}\nabla f(y)\cdot\nabla_yG(Y)\right]
=
\lim_{R\to\infty}\operatorname{Cov}_{\operatorname{Gin}_{\beta}}[G(Y),\mathcal{R}_{f,R}]. E Gin β y ∈ Y ∑ ∇ f ( y ) ⋅ ∇ y G ( Y ) = R → ∞ lim Cov Gin β [ G ( Y ) , R f , R ] .
Identification. For R R R large enough, the radiality of χ R \chi_R χ R and Newton’s theorem give
∫ χ R ( z ) K f ( x , z ) d z = 2 π f ( x ) − π x ⋅ ∇ f ( x ) . \int\chi_R(z)K_f(x,z)\,\mathrm{d}z=2\pi f(x)-\pi x\cdot\nabla f(x). ∫ χ R ( z ) K f ( x , z ) d z = 2 π f ( x ) − π x ⋅ ∇ f ( x ) .
Expanding (76) and using 2 π β ℓ D 2 = 1 2\pi\beta\ell_{\mathrm{D}}^{2}=1 2 π β ℓ D 2 = 1 , we obtain
R f , R − E Gin β R f , R = ℓ D Z β ( ( 1 − Δ ) f ) + β ℓ D 2 Q f , R . \mathcal{R}_{f,R}-\mathbb{E}_{\operatorname{Gin}_{\beta}}\mathcal{R}_{f,R}
=
\ell_{\mathrm{D}}\mathcal{Z}_{\beta}\bigl((1-\Delta)f\bigr)+\beta\ell_{\mathrm{D}}^{2}Q_{f,R}. R f , R − E Gin β R f , R = ℓ D Z β ( ( 1 − Δ ) f ) + β ℓ D 2 Q f , R .
The convergence of Q f , R Q_{f,R} Q f , R in L 2 L^2 L 2 , from Lemma 8.3, proves (79) .
Proof of Proposition 8.2.
Step 1. Spectral density and its bound. By Proposition 5.7, the function ρ 2 , β ( 0 , ⋅ ) − 1 \rho_{2,\beta}(0,\cdot)-1 ρ 2 , β ( 0 , ⋅ ) − 1 is integrable. The structure factor σ Gin β \sigma_{\operatorname{Gin}_{\beta}} σ Gin β therefore has a continuous radial density. After the change of variables k ↦ ℓ D k k\mapsto\ell_{\mathrm{D}}k k ↦ ℓ D k , the identity that defines σ Gin β \sigma_{\operatorname{Gin}_{\beta}} σ Gin β gives (74) . If f f f is radial, then the statistic Y ( f ) Y(f) Y ( f ) is rotation-invariant. We apply Proposition 4.14 to the finite gas, with a = π β / 2 a=\pi\beta/2 a = π β /2 , for which 1 / ( 4 a ) = ℓ D 2 1/(4a)=\ell_{\mathrm{D}}^{2} 1/ ( 4 a ) = ℓ D 2 , and pass to the limit along the hard-wall gas, by Corollary 3 and part (iii) of Proposition 2.4. This gives C β ( f , f ) ≤ ∫ ∣ ∇ f ∣ 2 C_{\beta}(f,f)\leq\int|\nabla f|^2 C β ( f , f ) ≤ ∫ ∣∇ f ∣ 2 . Since σ Gin β ( ⋅ / ℓ D ) \sigma_{\operatorname{Gin}_{\beta}}(\cdot/\ell_{\mathrm{D}}) σ Gin β ( ⋅ / ℓ D ) is radial and continuous, testing (74) with radial functions gives (73) .
Step 2. The integration by parts. Let f , h ∈ C ∞ ( R 2 ; R ) f,h \in C^\infty(\mathbb{R}^2;\mathbb{R}) f , h ∈ C ∞ ( R 2 ; R ) . For β \beta β sufficiently small, we apply Lemma 8.5 to smooth bounded truncations of Y ( h ) Y(h) Y ( h ) with uniformly bounded first derivatives. The local second moments and the L 2 L^2 L 2 integrability of Q f Q_f Q f justify passage to Y ( h ) Y(h) Y ( h ) . Since E Gin β Y ( ∇ f ⋅ ∇ h ) = ℓ D 2 ∫ R 2 ∇ f ⋅ ∇ h \mathbb{E}_{\operatorname{Gin}_\beta}Y(\nabla f\cdot\nabla h)=\ell_D^2\int_{\mathbb{R}^2}\nabla f\cdot\nabla h E Gin β Y ( ∇ f ⋅ ∇ h ) = ℓ D 2 ∫ R 2 ∇ f ⋅ ∇ h and Q f Q_f Q f is centred, division by ℓ D 2 \ell_D^2 ℓ D 2 gives
c β ( ( 1 − Δ ) f , h ) = ∫ R 2 ∇ f ⋅ ∇ h − β Cov Gin β [ Q f , Y ( h ) ] . (80) c_\beta\bigl((1-\Delta)f,h\bigr)=\int_{\mathbb{R}^2}\nabla f\cdot\nabla h-\beta\operatorname{Cov}_{\operatorname{Gin}_\beta}\left[Q_f,Y(h)\right]. \tag*{(80)} c β ( ( 1 − Δ ) f , h ) = ∫ R 2 ∇ f ⋅ ∇ h − β Cov Gin β [ Q f , Y ( h ) ] . (80)
Step 3. The limit β ↓ 0 \beta\downarrow0 β ↓ 0 . Let ℓ D − 1 ≤ δ ≤ 1 \ell_D^{-1}\leq\delta\leq1 ℓ D − 1 ≤ δ ≤ 1 . By Lemma 8.3, we have ∥ Q f ∥ 2 ≤ C ( f ) ( ℓ D 2 δ 2 + δ + δ − 1 ) \lVert Q_f\rVert_2\leq C(f)(\ell_D^2\delta^2+\delta+\delta^{-1}) ∥ Q f ∥ 2 ≤ C ( f ) ( ℓ D 2 δ 2 + δ + δ − 1 ) . By (8.7) and (8.6), the variance of Y ( h ) Y(h) Y ( h ) under Gin β \operatorname{Gin}_\beta Gin β is at most ℓ D 2 ∫ ∣ ∇ h ∣ 2 \ell_D^2\int|\nabla h|^2 ℓ D 2 ∫ ∣∇ h ∣ 2 . Since β ℓ D 2 = ( 2 π ℓ D ) − 1 \beta\ell_D^2=(2\pi\ell_D)^{-1} β ℓ D 2 = ( 2 π ℓ D ) − 1 , the Cauchy–Schwarz inequality bounds the last term of (8.13) by C ( f , h ) ( ℓ D δ 2 + δ / ℓ D + 1 / ( ℓ D δ ) ) C(f,h)(\ell_D\delta^2+\delta/\ell_D+1/(\ell_D\delta)) C ( f , h ) ( ℓ D δ 2 + δ / ℓ D + 1/ ( ℓ D δ )) . With δ = ℓ D − 2 / 3 \delta=\ell_D^{-2/3} δ = ℓ D − 2/3 , we obtain
∣ c β ( ( 1 − Δ ) f , h ) − ∫ R 2 ∇ f ⋅ ∇ h ∣ ≤ C ( f , h ) ℓ D − 1 / 3 . (81) \left|c_\beta\bigl((1-\Delta)f,h\bigr)-\int_{\mathbb{R}^2}\nabla f\cdot\nabla h\right|\leq C(f,h)\ell_D^{-1/3}. \tag*{(81)} c β ( ( 1 − Δ ) f , h ) − ∫ R 2 ∇ f ⋅ ∇ h ≤ C ( f , h ) ℓ D − 1/3 . (81)
By (8.6), the measures σ Gin β ( k / ℓ D ) d k \sigma_{\operatorname{Gin}_\beta}(k/\ell_D)\,dk σ Gin β ( k / ℓ D ) d k are bounded on compact sets, uniformly in β \beta β . The test functions ( 1 + ∣ k ∣ 2 ) f ^ ( k ) h ^ ( k ) (1+|k|^2)\widehat{f}(k)\widehat{h}(k) ( 1 + ∣ k ∣ 2 ) f ( k ) h ( k ) are Schwartz functions. By (8.6), their integrals against these measures over { ∣ k ∣ > K } \{|k|>K\} { ∣ k ∣ > K } are bounded by ∫ ∣ k ∣ > K ∣ k ∣ 2 ( 1 + ∣ k ∣ 2 ) ∣ f ^ ( k ) h ^ ( k ) ∣ d k \int_{|k|>K}|k|^2(1+|k|^2)|\widehat{f}(k)\widehat{h}(k)|\,dk ∫ ∣ k ∣ > K ∣ k ∣ 2 ( 1 + ∣ k ∣ 2 ) ∣ f ( k ) h ( k ) ∣ d k , uniformly in β \beta β , hence these integrals pass to weak-* limits. By (8.7) and (8.14), every weak-* limit μ \mu μ of these measures as β ↓ 0 \beta\downarrow0 β ↓ 0 satisfies ( 1 + ∣ k ∣ 2 ) μ ( d k ) = ∣ k ∣ 2 d k (1+|k|^2)\mu(dk)=|k|^2\,dk ( 1 + ∣ k ∣ 2 ) μ ( d k ) = ∣ k ∣ 2 d k . Hence these measures converge to ∣ k ∣ 2 ( 1 + ∣ k ∣ 2 ) − 1 d k |k|^2(1+|k|^2)^{-1}dk ∣ k ∣ 2 ( 1 + ∣ k ∣ 2 ) − 1 d k against continuous functions with compact support. By (8.6), this convergence also holds against Schwartz functions, which proves the convergence in S ′ ( R 2 ) \mathcal{S}'(\mathbb{R}^2) S ′ ( R 2 ) of σ Gin β ( ⋅ / ℓ D ) \sigma_{\operatorname{Gin}_\beta}(\cdot/\ell_D) σ Gin β ( ⋅ / ℓ D ) . Finally, the Fourier transform of ( ρ 2 , β ( 0 , ℓ D ⋅ ) − 1 ) / β (\rho_{2,\beta}(0,\ell_D\cdot)-1)/\beta ( ρ 2 , β ( 0 , ℓ D ⋅ ) − 1 ) / β is 2 π ( σ Gin β ( ⋅ / ℓ D ) − 1 ) 2\pi(\sigma_{\operatorname{Gin}_\beta}(\cdot/\ell_D)-1) 2 π ( σ Gin β ( ⋅ / ℓ D ) − 1 ) , and that of K 0 ( ∣ ⋅ ∣ ) K_0(|\cdot|) K 0 ( ∣ ⋅ ∣ ) is 2 π / ( 1 + ∣ k ∣ 2 ) 2\pi/(1+|k|^2) 2 π / ( 1 + ∣ k ∣ 2 ) . This proves (8.8).
Gaussian fluctuations on the Debye scale
The covariance in Theorem 6 is the limit computed in Proposition 8.2. To identify the law, we apply the Ward identity to characteristic functions. This gives a differential equation in the Fourier variable, with an inhomogeneous term containing the anisotropy. The same equation, obtained by the transport method of [69 ] , is the starting point of the local central limit theorem of [70 ] for two-dimensional Coulomb gases at fixed β \beta β . Here, the contribution of the quadratic statistic Q f Q_f Q f is small because it comes with a factor β \beta β .
Proof of Theorem 6 . First, let f ∈ C c ∞ ( R 2 ; R ) f\in C^\infty_c(\mathbb{R}^2;\mathbb{R}) f ∈ C c ∞ ( R 2 ; R ) and h = ( 1 − Δ ) f h=(1-\Delta)f h = ( 1 − Δ ) f , so that
∫ R 2 ∇ f ⋅ ∇ h = ⟨ h , ( − Δ ) ( 1 − Δ ) − 1 h ⟩ L 2 ( R 2 ) . \int_{\mathbb{R}^2}\nabla f\cdot\nabla h=\left\langle h,(-\Delta)(1-\Delta)^{-1}h\right\rangle_{L^2(\mathbb{R}^2)}. ∫ R 2 ∇ f ⋅ ∇ h = ⟨ h , ( − Δ ) ( 1 − Δ ) − 1 h ⟩ L 2 ( R 2 ) .
Let φ β ( t ) : = E Gin β e i t Z β ( h ) \varphi_\beta(t):=\mathbb{E}_{\operatorname{Gin}_\beta}e^{itZ_\beta(h)} φ β ( t ) := E Gin β e i t Z β ( h ) . We apply (8.12) to the real and imaginary parts of e i t Z β ( h ) e^{itZ_\beta(h)} e i t Z β ( h ) . Differentiation of the exponential and the identity Y ( ∇ f ⋅ ∇ h ) = ℓ D 2 ∫ ∇ f ⋅ ∇ h + ℓ D Z β ( ∇ f ⋅ ∇ h ) Y(\nabla f\cdot\nabla h)=\ell_D^2\int\nabla f\cdot\nabla h+\ell_D Z_\beta(\nabla f\cdot\nabla h) Y ( ∇ f ⋅ ∇ h ) = ℓ D 2 ∫ ∇ f ⋅ ∇ h + ℓ D Z β ( ∇ f ⋅ ∇ h ) give
∂ t φ β ( t ) + t φ β ( t ) ∫ ∇ f ⋅ ∇ h = − t ℓ D E Gin β [ Z β ( ∇ f ⋅ ∇ h ) e i t Z β ( h ) ] − i β ℓ D E Gin β [ Q f e i t Z β ( h ) ] . \partial_t\varphi_\beta(t)+t\varphi_\beta(t)\int\nabla f\cdot\nabla h=-\frac{t}{\ell_D}\mathbb{E}_{\operatorname{Gin}_\beta}\left[Z_\beta(\nabla f\cdot\nabla h)e^{itZ_\beta(h)}\right]-i\beta\ell_D\mathbb{E}_{\operatorname{Gin}_\beta}\left[Q_fe^{itZ_\beta(h)}\right]. ∂ t φ β ( t ) + t φ β ( t ) ∫ ∇ f ⋅ ∇ h = − ℓ D t E Gin β [ Z β ( ∇ f ⋅ ∇ h ) e i t Z β ( h ) ] − i β ℓ D E Gin β [ Q f e i t Z β ( h ) ] .
By (8.7) and (8.6), ∥ Z β ( ∇ f ⋅ ∇ h ) ∥ 2 ≤ ∥ ∇ ( ∇ f ⋅ ∇ h ) ∥ 2 \lVert Z_\beta(\nabla f\cdot\nabla h)\rVert_2\leq\lVert\nabla(\nabla f\cdot\nabla h)\rVert_2 ∥ Z β ( ∇ f ⋅ ∇ h ) ∥ 2 ≤ ∥ ∇ ( ∇ f ⋅ ∇ h ) ∥ 2 . Taking δ = ℓ D − 2 / 3 \delta=\ell_D^{-2/3} δ = ℓ D − 2/3 in Lemma 8.3 gives ∥ Q f ∥ 2 ≤ C ( f ) ℓ D 2 / 3 \lVert Q_f\rVert_2\leq C(f)\ell_D^{2/3} ∥ Q f ∥ 2 ≤ C ( f ) ℓ D 2/3 . Since β ℓ D 2 = ( 2 π ) − 1 \beta\ell_D^2=(2\pi)^{-1} β ℓ D 2 = ( 2 π ) − 1 , the Cauchy–Schwarz inequality gives, for every T > 0 T>0 T > 0 ,
sup ∣ t ∣ ≤ T ∣ ∂ t φ β ( t ) + t φ β ( t ) ∫ ∇ f ⋅ ∇ h ∣ ≤ C ( f , T ) ℓ D − 1 / 3 → β ↓ 0 0. \sup_{|t|\leq T}\left|\partial_t\varphi_\beta(t)+t\varphi_\beta(t)\int\nabla f\cdot\nabla h\right|\leq C(f,T)\ell_D^{-1/3}\xrightarrow[\beta\downarrow0]{}0. ∣ t ∣ ≤ T sup ∂ t φ β ( t ) + t φ β ( t ) ∫ ∇ f ⋅ ∇ h ≤ C ( f , T ) ℓ D − 1/3 β ↓ 0 0.
Together with φ β ( 0 ) = 1 \varphi_\beta(0)=1 φ β ( 0 ) = 1 , this implies that φ β ( t ) \varphi_\beta(t) φ β ( t ) converges to exp ( − t 2 2 ∫ ∇ f ⋅ ∇ h ) \exp\left(-\frac{t^2}{2}\int\nabla f\cdot\nabla h\right) exp ( − 2 t 2 ∫ ∇ f ⋅ ∇ h ) as β ↓ 0 \beta\downarrow0 β ↓ 0 , uniformly on bounded intervals.
For a general h ∈ C ∞ ( R 2 ; R ) h\in C^\infty(\mathbb{R}^2;\mathbb{R}) h ∈ C ∞ ( R 2 ; R ) , let f = ( 1 − Δ ) − 1 h f=(1-\Delta)^{-1}h f = ( 1 − Δ ) − 1 h and h R = ( 1 − Δ ) ( χ R f ) h_R=(1-\Delta)(\chi_Rf) h R = ( 1 − Δ ) ( χ R f ) . The function f f f belongs to S ( R 2 ) \mathcal{S}(\mathbb{R}^2) S ( R 2 ) , so h R → R → ∞ h h_R\xrightarrow{R\to\infty}h h R R → ∞ h in H 1 ( R 2 ) H^1(\mathbb{R}^2) H 1 ( R 2 ) . The constant β 0 \beta_0 β 0 of Lemma 8.3 depends on χ R f \chi_Rf χ R f , hence on R R R . This is harmless, since we let β ↓ 0 \beta\downarrow0 β ↓ 0 at fixed R R R before letting R → ∞ R\to\infty R → ∞ . By (8.7) and (8.6), uniformly in β \beta β ,
∣ E Gin β e i t Z β ( h ) − E Gin β e i t Z β ( h R ) ∣ ≤ ∣ t ∣ ∥ ∇ ( h − h R ) ∥ 2 . \left|\mathbb{E}_{\operatorname{Gin}_\beta}e^{itZ_\beta(h)}-\mathbb{E}_{\operatorname{Gin}_\beta}e^{itZ_\beta(h_R)}\right|\leq|t|\lVert\nabla(h-h_R)\rVert_2. E Gin β e i t Z β ( h ) − E Gin β e i t Z β ( h R ) ≤ ∣ t ∣ ∥ ∇ ( h − h R ) ∥ 2 .
The Gaussian variances converge as well. First letting β ↓ 0 \beta\downarrow0 β ↓ 0 at fixed R R R , then R → ∞ R \to\infty R → ∞ , proves the convergence for h h h . Linearity of Z β Z_\beta Z β and the Cramér–Wold theorem give the finite-dimensional convergence.
Insertion of two particles
To compare q β q_\beta q β at two distances, we go back to the finite gas. For N ≥ 1 N \ge1 N ≥ 1 , let ρ 2 , 0 ( N ) \rho^{(N)}_{2,0} ρ 2 , 0 ( N ) be the two-point correlation function of Y Y Y under the measure P 0 \mathbb{P}_0 P 0 on ( R 2 ) N (\mathbb{R}^2)^N ( R 2 ) N . Here the second subscript records the tilt parameter t = 0 t=0 t = 0 . In the insertion identities below, E 0 \mathbb{E}_0 E 0 and E 1 \mathbb{E}_1 E 1 are expectations for the gas with N N N particles.
Lemma 8.6 (Insertion of two particles). Let 0 < β ≤ 2 0 < \beta\le2 0 < β ≤ 2 and N ≥ 1 N \ge1 N ≥ 1 . There exists c N , β > 0 c_{N,\beta} > 0 c N , β > 0 such that, for all distinct u , v ∈ R 2 u,v \in\mathbb{R}^2 u , v ∈ R 2 ,
ρ 2 , 0 ( N + 2 ) ( u , v ) = c N , β e − ( ∣ u ∣ 2 + ∣ v ∣ 2 ) / 4 − β g ( u − v ) E 0 exp [ − β Y ( g ( ⋅ − u ) + g ( ⋅ − v ) ) ] . (82) \rho^{(N+2)}_{2,0}(u,v)=c_{N,\beta}e^{-(|u|^2+|v|^2)/4-\beta g(u-v)}\mathbb{E}_0\exp\left[-\beta Y\left(g(\cdot-u)+g(\cdot-v)\right)\right]. \tag*{(82)} ρ 2 , 0 ( N + 2 ) ( u , v ) = c N , β e − ( ∣ u ∣ 2 + ∣ v ∣ 2 ) /4 − β g ( u − v ) E 0 exp [ − β Y ( g ( ⋅ − u ) + g ( ⋅ − v ) ) ] . (82)
Proof. We fix two of the N + 2 N+2 N + 2 positions at u u u and v v v in the density (68) with t = 0 t=0 t = 0 . The density of the remaining N N N positions is then that of P 0 \mathbb{P}_0 P 0 , multiplied by e − β Y ( g ( ⋅ − u ) + g ( ⋅ − v ) ) e^{-\beta Y(g(\cdot-u)+g(\cdot-v))} e − β Y ( g ( ⋅− u ) + g ( ⋅− v )) and by the factor e − ( ∣ u ∣ 2 + ∣ v ∣ 2 ) / 4 − β g ( u − v ) e^{-(|u|^2+|v|^2)/4-\beta g(u-v)} e − ( ∣ u ∣ 2 + ∣ v ∣ 2 ) /4 − β g ( u − v ) , which does not depend on them. Integrating over these positions gives (82) , where c N , β c_{N,\beta} c N , β is ( N + 2 ) ( N + 1 ) (N+2)(N+1) ( N + 2 ) ( N + 1 ) times the ratio of the normalizing constants of (68) with t = 0 t=0 t = 0 for N N N and for N + 2 N+2 N + 2 particles.
We apply Lemma 8.6 with u = 0 u=0 u = 0 and v = r e 1 v=re_1 v = r e 1 , for r > 0 r>0 r > 0 . The potential g + g ( ⋅ − r e 1 ) g+g(\cdot-re_1) g + g ( ⋅ − r e 1 ) of the two particles grows logarithmically at infinity. We subtract twice the radial function g 1 g_1 g 1 from it, which gives the function D r = g + g ( ⋅ − r e 1 ) − 2 g 1 D_r=g+g(\cdot-re_1)-2g_1 D r = g + g ( ⋅ − r e 1 ) − 2 g 1 of Subsection 3.4. In words, D r D_r D r is the potential of the two particles together with a charge − 2 -2 − 2 , spread over the unit disk with density 2 ψ 2\psi 2 ψ . Since this distribution of charges is neutral, D r D_r D r decays at infinity. More precisely, since g 1 = g g_1=g g 1 = g outside B 1 B_1 B 1 , we have D r = g ( ⋅ − r e 1 ) − g D_r=g(\cdot-re_1)-g D r = g ( ⋅ − r e 1 ) − g outside B 1 B_1 B 1 , hence D r D_r D r is of order r / ∣ z ∣ r/|z| r /∣ z ∣ at infinity. The remaining term 2 β Y ( g 1 ) 2\beta Y(g_1) 2 β Y ( g 1 ) is the interaction of the gas with the opposite charge 2 2 2 . It does not depend on r r r , and we include it in the measure P 1 \mathbb{P}_1 P 1 . For r , s > 0 r,s>0 r , s > 0 , Lemma 8.6 therefore gives
ρ 2 , 0 ( N + 2 ) ( 0 , r e 1 ) ρ 2 , 0 ( N + 2 ) ( 0 , s e 1 ) = e − ( r 2 − s 2 ) / 4 + β log ( r / s ) E 1 e − β Y ( D r ) E 1 e − β Y ( D s ) . (83) \frac{\rho^{(N+2)}_{2,0}(0,re_1)}{\rho^{(N+2)}_{2,0}(0,se_1)}
=e^{-(r^2-s^2)/4+\beta\log(r/s)}
\frac{\mathbb{E}_1e^{-\beta Y(D_r)}}{\mathbb{E}_1e^{-\beta Y(D_s)}}. \tag*{(83)} ρ 2 , 0 ( N + 2 ) ( 0 , s e 1 ) ρ 2 , 0 ( N + 2 ) ( 0 , r e 1 ) = e − ( r 2 − s 2 ) /4 + β l o g ( r / s ) E 1 e − β Y ( D s ) E 1 e − β Y ( D r ) . (83)
Expansion of the Laplace transform
We now expand log E t e − β Y ( D r ) \log\mathbb{E}_t e^{-\beta Y(D_r)} log E t e − β Y ( D r ) to second order in β \beta β . The only difficulty comes from the logarithmic singularities of D r D_r D r at 0 0 0 and r e 1 re_1 r e 1 . We smooth them at the scale ℓ D − 1 \ell_D^{-1} ℓ D − 1 of the distance between neighbouring particles, with the smoothed logarithm of Subsection 5.1. For β ≤ 1 / ( 2 π ) \beta\le1/(2\pi) β ≤ 1/ ( 2 π ) , we have ℓ D − 1 ≤ 1 \ell_D^{-1}\le1 ℓ D − 1 ≤ 1 , and we let
D ~ r : = g 1 / ℓ D + g 1 / ℓ D ( ⋅ − r e 1 ) − 2 g 1 . \widetilde{D}_r:=g_{1/\ell_D}+g_{1/\ell_D}(\cdot-re_1)-2g_1. D r := g 1/ ℓ D + g 1/ ℓ D ( ⋅ − r e 1 ) − 2 g 1 .
For 0 < ε ≤ 1 0<\varepsilon\le1 0 < ε ≤ 1 , the function g − g ε g-g_\varepsilon g − g ε is non-negative and supported in B ε B_\varepsilon B ε , by Newton’s theorem. The difference D r − D ~ r D_r-\widetilde{D}_r D r − D r is the sum of the translates of g − g 1 / ℓ D g-g_{1/\ell_D} g − g 1/ ℓ D by 0 0 0 and by r e 1 re_1 r e 1 .
Lemma 8.7 (Moments of the potential of two particles). Let 0 < r 0 < 1 0<r_0<1 0 < r 0 < 1 . There exist C , β 0 > 0 C,\beta_0>0 C , β 0 > 0 such that, for every 0 < β < β 0 0<\beta<\beta_0 0 < β < β 0 , every N ≥ 4 / β N\ge4/\beta N ≥ 4/ β , every 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , every r 0 ≤ r ≤ 1 / r 0 r_0\le r\le1/r_0 r 0 ≤ r ≤ 1/ r 0 , every 0 < ε ≤ 1 0<\varepsilon\le1 0 < ε ≤ 1 , every x ∈ R 2 x\in\mathbb{R}^2 x ∈ R 2 and every p ∈ { 2 , 4 , 6 } p\in\{2,4,6\} p ∈ { 2 , 4 , 6 } ,
∥ Fluct t [ ( g − g ε ) ( ⋅ − x ) ] ∥ p ≤ C ( ℓ D 2 ε 2 + ℓ D 2 / p ε 2 / p ) . (84) \left\|\operatorname{Fluct}_t\left[(g-g_\varepsilon)(\cdot-x)\right]\right\|_p
\le C\left(\ell_D^2\varepsilon^2+\ell_D^{2/p}\varepsilon^{2/p}\right). \tag*{(84)} ∥ Fluct t [ ( g − g ε ) ( ⋅ − x ) ] ∥ p ≤ C ( ℓ D 2 ε 2 + ℓ D 2/ p ε 2/ p ) . (84) Moreover, for every 2 ≤ p ≤ 6 2\le p\le6 2 ≤ p ≤ 6 ,
∥ Fluct t [ D r ] ∥ p ≤ C ℓ D ( 1 + ∣ log β ∣ ) 1 / 2 and E t e − 2 β Fluct t [ D r ] ≤ C . (85) \left\|\operatorname{Fluct}_t[D_r]\right\|_p\le C\ell_D(1+|\log\beta|)^{1/2}
\qquad\text{and}\qquad
\mathbb{E}_t e^{-2\beta\operatorname{Fluct}_t[D_r]}\le C. \tag*{(85)} ∥ Fluct t [ D r ] ∥ p ≤ C ℓ D ( 1 + ∣ log β ∣ ) 1/2 and E t e − 2 β Fluct t [ D r ] ≤ C . (85)
The proof is given in Appendix § 13 .§ 11 . By (85) , the fluctuations of β Y ( D r ) \beta Y(D_r) β Y ( D r ) are at most of order β ℓ D ( 1 + ∣ log β ∣ ) 1 / 2 \beta\ell_D(1+|\log\beta|)^{1/2} β ℓ D ( 1 + ∣ log β ∣ ) 1/2 , which tends to zero, so we can expand the Laplace transform.
Lemma 8.8 (Laplace expansion). Let 0 < r 0 < 1 0<r_0<1 0 < r 0 < 1 . There exist C , β 0 > 0 C,\beta_0>0 C , β 0 > 0 such that, for every 0 < β < β 0 0<\beta<\beta_0 0 < β < β 0 , every N ≥ 4 / β N\ge4/\beta N ≥ 4/ β , every 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 and every r 0 ≤ r ≤ 1 / r 0 r_0\le r\le1/r_0 r 0 ≤ r ≤ 1/ r 0 ,
∣ log E t e − β Y ( D r ) + β E t Y ( D r ) − β 2 2 Var t [ Y ( D r ) ] ∣ ≤ C β 3 / 2 ( 1 + ∣ log β ∣ ) 3 / 2 . (86) \left|\log\mathbb{E}_t e^{-\beta Y(D_r)}+\beta\mathbb{E}_tY(D_r)-\frac{\beta^2}{2}\operatorname{Var}_t[Y(D_r)]\right|\le C\beta^{3/2}(1+|\log\beta|)^{3/2}. \tag*{(86)} log E t e − β Y ( D r ) + β E t Y ( D r ) − 2 β 2 Var t [ Y ( D r )] ≤ C β 3/2 ( 1 + ∣ log β ∣ ) 3/2 . (86)
Proof. Since ℓ D = ( 2 π β ) − 1 / 2 \ell_D=(2\pi\beta)^{-1/2} ℓ D = ( 2 π β ) − 1/2 , the first bound of (85) gives ∥ Fluct t [ D r ] ∥ 6 ≤ C β − 1 / 2 ( 1 + ∣ log β ∣ ) 1 / 2 \|\operatorname{Fluct}_t[D_r]\|_6\le C\beta^{-1/2}(1+|\log\beta|)^{1/2} ∥ Fluct t [ D r ] ∥ 6 ≤ C β − 1/2 ( 1 + ∣ log β ∣ ) 1/2 . We expand the exponential with the Taylor formula, and bound the remainder with the Cauchy–Schwarz inequality and the second bound of (85) . This gives
∣ E t e − β Fluct t [ D r ] − 1 − β 2 2 E t [ Fluct t [ D r ] 2 ] ∣ ≤ C β 3 ∥ Fluct t [ D r ] ∥ 6 3 ( 1 + ( E t e − 2 β Fluct t [ D r ] ) 1 / 2 ) ≤ C β 3 / 2 ( 1 + ∣ log β ∣ ) 3 / 2 . \begin{aligned}
\left|\mathbb{E}_t e^{-\beta\operatorname{Fluct}_t[D_r]}-1-\frac{\beta^2}{2}\mathbb{E}_t[\operatorname{Fluct}_t[D_r]^2]\right|\le C\beta^3\|\operatorname{Fluct}_t[D_r]\|_6^3\left(1+\left(\mathbb{E}_t e^{-2\beta\operatorname{Fluct}_t[D_r]}\right)^{1/2}\right) \\
&\le C\beta^{3/2}(1+|\log\beta|)^{3/2}.
\end{aligned} E t e − β Fluct t [ D r ] − 1 − 2 β 2 E t [ Fluct t [ D r ] 2 ] ≤ C β 3 ∥ Fluct t [ D r ] ∥ 6 3 ( 1 + ( E t e − 2 β Fluct t [ D r ] ) 1/2 ) ≤ C β 3/2 ( 1 + ∣ log β ∣ ) 3/2 .
Since β 2 E t [ Fluct t [ D r ] 2 ] ≤ C β ( 1 + ∣ log β ∣ ) \beta^2\mathbb{E}_t[\operatorname{Fluct}_t[D_r]^2]\le C\beta(1+|\log\beta|) β 2 E t [ Fluct t [ D r ] 2 ] ≤ C β ( 1 + ∣ log β ∣ ) , taking the logarithm adds an error of at most C β 2 ( 1 + ∣ log β ∣ ) 2 C\beta^2(1+|\log\beta|)^2 C β 2 ( 1 + ∣ log β ∣ ) 2 , which the right-hand side of (86) absorbs. We conclude with the identity log E t e − β Y ( D r ) = − β E t Y ( D r ) + log E t e − β Fluct t [ D r ] \log\mathbb{E}_t e^{-\beta Y(D_r)}=-\beta\mathbb{E}_tY(D_r)+\log\mathbb{E}_t e^{-\beta\operatorname{Fluct}_t[D_r]} log E t e − β Y ( D r ) = − β E t Y ( D r ) + log E t e − β Fluct t [ D r ] . □
We use this expansion at t = 1 t=1 t = 1 . To come back to Gin β \operatorname{Gin}_\beta Gin β , we then have to remove the charge added to the gas, that is, to compare the mean and the variance at t = 1 t=1 t = 1 with those at t = 0 t=0 t = 0 , where the gas converges to Gin β \operatorname{Gin}_\beta Gin β as N → ∞ N\to\infty N → ∞ . For this, we differentiate the density of P t \mathbb{P}_t P t in t t t . For bounded functions f f f and h h h , the derivative of E t Y ( f ) \mathbb{E}_tY(f) E t Y ( f ) is − 2 β Cov t [ Y ( f ) , Y ( g 1 ) ] -2\beta\operatorname{Cov}_t[Y(f),Y(g_1)] − 2 β Cov t [ Y ( f ) , Y ( g 1 )] , and
d d t Cov t [ Y ( f ) , Y ( h ) ] = − 2 β E t [ Fluct t [ f ] Fluct t [ h ] Fluct t [ g 1 ] ] . (87) \frac{\mathrm{d}}{\mathrm{d}t}\operatorname{Cov}_t[Y(f),Y(h)]=-2\beta\mathbb{E}_t[\operatorname{Fluct}_t[f]\operatorname{Fluct}_t[h]\operatorname{Fluct}_t[g_1]]. \tag*{(87)} d t d Cov t [ Y ( f ) , Y ( h )] = − 2 β E t [ Fluct t [ f ] Fluct t [ h ] Fluct t [ g 1 ]] . (87)
Since P t \mathbb{P}_t P t is rotation-invariant and g 1 g_1 g 1 is radial, the covariance of Y ( f ) Y(f) Y ( f ) and Y ( g 1 ) Y(g_1) Y ( g 1 ) under P t \mathbb{P}_t P t only depends on the angular average of f f f . By Newton’s theorem, the angular average of D r − D s = g ( ⋅ − r e 1 ) − g ( ⋅ − s e 1 ) D_r-D_s=g(\cdot-re_1)-g(\cdot-se_1) D r − D s = g ( ⋅ − r e 1 ) − g ( ⋅ − s e 1 ) is
A ( z ) : = − log max ( ∣ z ∣ , r ) + log max ( ∣ z ∣ , s ) . (88) A(z):=-\log\max(|z|,r)+\log\max(|z|,s). \tag*{(88)} A ( z ) := − log max ( ∣ z ∣ , r ) + log max ( ∣ z ∣ , s ) . (88)
It is supported in B max ( r , s ) B_{\max(r,s)} B m a x ( r , s ) and Lipschitz, uniformly for r 0 ≤ r , s ≤ 1 / r 0 r_0\le r,s\le1/r_0 r 0 ≤ r , s ≤ 1/ r 0 . By (87) , the derivatives in t t t of Cov t [ Y ( A ) , Y ( g 1 ) ] \operatorname{Cov}_t[Y(A),Y(g_1)] Cov t [ Y ( A ) , Y ( g 1 )] and of Var t [ Y ( D ~ r ) ] \operatorname{Var}_t[Y(\widetilde D_r)] Var t [ Y ( D r )] are third moments of linear statistics, which we bound in the next lemma.
Lemma 8.9 (Third moments of linear statistics). Let 0 < r 0 < 1 0<r_0<1 0 < r 0 < 1 . There exist C , β 0 > 0 C,\beta_0>0 C , β 0 > 0 such that, for every 0 < β < β 0 0<\beta<\beta_0 0 < β < β 0 , every N ≥ 4 / β N\ge4/\beta N ≥ 4/ β , every 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 and every r 0 ≤ r , s ≤ 1 / r 0 r_0\le r,s\le1/r_0 r 0 ≤ r , s ≤ 1/ r 0 ,
∣ E t [ Fluct t [ A ] Fluct t [ g 1 ] 2 ] ∣ ≤ C ℓ D 3 and ∣ E t [ Fluct t [ D ~ r ] 2 Fluct t [ g 1 ] ] ∣ ≤ C ℓ D 3 ( 1 + ∣ log β ∣ ) . (89) \left|\mathbb{E}_t[\operatorname{Fluct}_t[A]\operatorname{Fluct}_t[g_1]^2]\right|\le C\ell_D^3
\qquad\text{and}\qquad
\left|\mathbb{E}_t[\operatorname{Fluct}_t[\widetilde D_r]^2\operatorname{Fluct}_t[g_1]]\right|\le C\ell_D^3(1+|\log\beta|). \tag*{(89)} E t [ Fluct t [ A ] Fluct t [ g 1 ] 2 ] ≤ C ℓ D 3 and E t [ Fluct t [ D r ] 2 Fluct t [ g 1 ]] ≤ C ℓ D 3 ( 1 + ∣ log β ∣ ) . (89) The proof is given in Appendix § 13 .§ 11 .
Lemma 8.10 (Removal of the added charge). Let 0 < r 0 < 1 0<r_0<1 0 < r 0 < 1 . There exist C , β 0 > 0 C,\beta_0>0 C , β 0 > 0 such that, for every 0 < β < β 0 0<\beta<\beta_0 0 < β < β 0 , every N ≥ 4 / β N\ge4/\beta N ≥ 4/ β and every r 0 ≤ r , s ≤ 1 / r 0 r_0\le r,s\le1/r_0 r 0 ≤ r , s ≤ 1/ r 0 ,
∣ E 1 Y ( D r − D s ) − E 0 Y ( D r − D s ) + 2 β Cov 0 [ Y ( D r − D s ) , Y ( g 1 ) ] ∣ ≤ C β , (90) \left|\mathbb{E}_1Y(D_r-D_s)-\mathbb{E}_0Y(D_r-D_s)+2\beta\operatorname{Cov}_0[Y(D_r-D_s),Y(g_1)]\right|\le C\sqrt{\beta}, \tag*{(90)} ∣ E 1 Y ( D r − D s ) − E 0 Y ( D r − D s ) + 2 β Cov 0 [ Y ( D r − D s ) , Y ( g 1 )] ∣ ≤ C β , (90) ∣ β Var 1 [ Y ( D r ) ] − Var 0 [ Y ( D r ) ] ∣ ≤ C β ( 1 + ∣ log β ∣ ) . (91) \left|\beta\operatorname{Var}_1[Y(D_r)]-\operatorname{Var}_0[Y(D_r)]\right|\le C\sqrt{\beta}(1+|\log\beta|). \tag*{(91)} ∣ β Var 1 [ Y ( D r )] − Var 0 [ Y ( D r )] ∣ ≤ C β ( 1 + ∣ log β ∣ ) . (91)
Proof. By rotation invariance, Cov t [ Y ( D r − D s ) , Y ( g 1 ) ] = Cov t [ Y ( A ) , Y ( g 1 ) ] \operatorname{Cov}_t[Y(D_r-D_s),Y(g_1)]=\operatorname{Cov}_t[Y(A),Y(g_1)] Cov t [ Y ( D r − D s ) , Y ( g 1 )] = Cov t [ Y ( A ) , Y ( g 1 )] . By (87) and the first bound of (89) , the derivative of this covariance in t t t is at most 2 C β ℓ D 3 ≤ C β − 1 / 2 2C\beta\ell_D^3\le C\beta^{-1/2} 2 C β ℓ D 3 ≤ C β − 1/2 in absolute value. We integrate in t t t the derivative − 2 β Cov t [ Y ( D r − D s ) , Y ( g 1 ) ] -2\beta\operatorname{Cov}_t[Y(D_r-D_s),Y(g_1)] − 2 β Cov t [ Y ( D r − D s ) , Y ( g 1 )] of E t Y ( D r − D s ) \mathbb{E}_tY(D_r-D_s) E t Y ( D r − D s ) , and replace this covariance by its value at t = 0 t=0 t = 0 . This proves (90) .
For the variance, (87) and the second bound of (89) give
∣ d d t Var t [ Y ( D ~ r ) ] ∣ ≤ C β − 1 / 2 ( 1 + ∣ log β ∣ ) . \left|\frac{\mathrm{d}}{\mathrm{d}t}\operatorname{Var}_t[Y(\widetilde D_r)]\right|\le C\beta^{-1/2}(1+|\log\beta|). d t d Var t [ Y ( D r )] ≤ C β − 1/2 ( 1 + ∣ log β ∣ ) .
Since D r − D ~ r D_r-\widetilde D_r D r − D r is the sum of two translates of g − g 1 / ℓ D g-g_1/\ell_D g − g 1 / ℓ D , the bound (84) with ε = ℓ D − 1 \varepsilon=\ell_D^{-1} ε = ℓ D − 1 gives ∥ Fluct t [ D r − D ~ r ] ∥ 2 ≤ C \|\operatorname{Fluct}_t[D_r-\widetilde D_r]\|_2\le C ∥ Fluct t [ D r − D r ] ∥ 2 ≤ C . With (85) , we obtain
β ∣ Var t [ Y ( D r ) ] − Var t [ Y ( D ~ r ) ] ∣ ≤ C β ℓ D ( 1 + ∣ log β ∣ ) 1 / 2 ≤ C β ( 1 + ∣ log β ∣ ) 1 / 2 . \beta\left|\operatorname{Var}_t[Y(D_r)]-\operatorname{Var}_t[Y(\widetilde D_r)]\right|\le C\beta\ell_D(1+|\log\beta|)^{1/2}\le C\sqrt{\beta}(1+|\log\beta|)^{1/2}. β Var t [ Y ( D r )] − Var t [ Y ( D r )] ≤ C β ℓ D ( 1 + ∣ log β ∣ ) 1/2 ≤ C β ( 1 + ∣ log β ∣ ) 1/2 .
We integrate in t t t the bound on the derivative of the variance, and use the bound on β ∣ Var t [ Y ( D r ) ] − Var t [ Y ( D ~ r ) ] ∣ \beta|\operatorname{Var}_t[Y(D_r)]-\operatorname{Var}_t[Y(\widetilde D_r)]| β ∣ Var t [ Y ( D r )] − Var t [ Y ( D r )] ∣ at t = 0 t=0 t = 0 and at t = 1 t=1 t = 1 . This proves (91) . □
Identification of the profile
We now let N → ∞ N \to\infty N → ∞ at fixed β \beta β , and then β ↓ 0 \beta\downarrow0 β ↓ 0 . At fixed β \beta β , the soft gas converges to Ginβ _\beta β (Subsection 2.1), and the argument of Appendix B.4, applied to the soft gas, gives ℓ D − 4 ρ 2 , 0 ( N + 2 ) ( 0 , r e 1 ) → N → ∞ q β ( r ) \ell_{\mathrm{D}}^{-4}\rho_{2,0}^{(N+2)}(0,r e_1) \xrightarrow[N\to\infty]{} q_\beta(r) ℓ D − 4 ρ 2 , 0 ( N + 2 ) ( 0 , r e 1 ) N → ∞ q β ( r ) . Indeed, the estimates of Appendix B.4 only use the DLR equations near the two points and the bounds (G1)–(G2). The finite soft gas satisfies the DLR equations in every bounded domain, and the bounds (G1)–(G2) hold for it uniformly in N N N , by the local laws of [5 ] (see also [59 Section 6.5] ). Hence the functions ρ 2 , 0 ( N + 2 ) \rho_{2,0}^{(N+2)} ρ 2 , 0 ( N + 2 ) are equicontinuous on compact subsets of { u ≠ v } \{u \ne v\} { u = v } , uniformly in N N N , and they converge locally uniformly as N → ∞ N \to\infty N → ∞ . In this limit, (8.16) and Lemmas 8.8 and 8.10 only involve covariances under Ginβ _\beta β , of functions that are singular or not compactly supported. We define these covariances by limits. For r r r , s > 0 s > 0 s > 0 , the covariances C β ( A , g 1 ) C_\beta(A,g_1) C β ( A , g 1 ) , C β ( D r , D r ) C_\beta(D_r,D_r) C β ( D r , D r ) and C β ( g , D r − D s ) C_\beta(g,D_r-D_s) C β ( g , D r − D s ) are the limits of C β ( χ R f ε , χ R h ε ) C_\beta(\chi_R f_\varepsilon,\chi_R h_\varepsilon) C β ( χ R f ε , χ R h ε ) as R → ∞ R \to\infty R → ∞ and then ε → 0 \varepsilon\to0 ε → 0 , where ( f , h ) (f,h) ( f , h ) is the corresponding pair of functions, and f ε f_\varepsilon f ε and h ε h_\varepsilon h ε are obtained from f f f and h h h by replacing g g g by g ε g_\varepsilon g ε . For instance,
C β ( g , D r − D s ) = lim ε → 0 lim R → ∞ C β ( χ R g ε , χ R ( g ε ( ⋅ − r e 1 ) − g ε ( ⋅ − s e 1 ) ) ) . (92) C_\beta(g,D_r-D_s)=\lim_{\varepsilon\to0}\lim_{R\to\infty} C_\beta\left(\chi_R g_\varepsilon,\chi_R\left(g_\varepsilon(\mathord{\cdot}-r e_1)-g_\varepsilon(\mathord{\cdot}-s e_1)\right)\right). \tag*{(92)} C β ( g , D r − D s ) = ε → 0 lim R → ∞ lim C β ( χ R g ε , χ R ( g ε ( ⋅ − r e 1 ) − g ε ( ⋅ − s e 1 ) ) ) . (92)
Proposition 8.11 (Ratios of the two-point function). Let 0 < r 0 < 1 0<r_0<1 0 < r 0 < 1 . There exist C , β 0 > 0 C,\beta_0>0 C , β 0 > 0 such that, for every 0 < β < β 0 0<\beta<\beta_0 0 < β < β 0 , the function q β q_\beta q β is positive on [ r 0 , 1 / r 0 ] [r_0,1/r_0] [ r 0 , 1/ r 0 ] , the three limits that define C β ( A , g 1 ) C_\beta(A,g_1) C β ( A , g 1 ) , C β ( D r , D r ) C_\beta(D_r,D_r) C β ( D r , D r ) and C β ( g , D r − D s ) C_\beta(g,D_r-D_s) C β ( g , D r − D s ) exist for r 0 ≤ r , s ≤ 1 / r 0 r_0\le r,s\le1/r_0 r 0 ≤ r , s ≤ 1/ r 0 , and
sup r 0 ≤ r , s ≤ 1 / r 0 ∣ 1 β log q β ( r ) q β ( s ) − log r s − 1 2 π C β ( g , D r − D s ) ∣ ≤ C β ( 1 + ∣ log β ∣ ) 3 / 2 . (93) \sup_{r_0\le r,s\le1/r_0}\left|\frac{1}{\beta}\log\frac{q_\beta(r)}{q_\beta(s)}-\log\frac{r}{s}-\frac{1}{2\pi}C_\beta(g,D_r-D_s)\right|\le C\sqrt{\beta}(1+|\log\beta|)^{3/2}. \tag*{(93)} r 0 ≤ r , s ≤ 1/ r 0 sup β 1 log q β ( s ) q β ( r ) − log s r − 2 π 1 C β ( g , D r − D s ) ≤ C β ( 1 + ∣ log β ∣ ) 3/2 . (93)
Proof.
Step 1. The ratio for the finite gas. Let N ≥ 4 / β N\ge4/\beta N ≥ 4/ β and r 0 ≤ r , s ≤ 1 / r 0 r_0\le r,s\le1/r_0 r 0 ≤ r , s ≤ 1/ r 0 . By rotation invariance, E 0 Y ( D r − D s ) = E 0 Y ( A ) \mathbb{E}_0Y(D_r-D_s)=\mathbb{E}_0Y(A) E 0 Y ( D r − D s ) = E 0 Y ( A ) , and a direct integration gives ∫ A = − π 2 ( r 2 − s 2 ) \int A=-\frac{\pi}{2}(r^2-s^2) ∫ A = − 2 π ( r 2 − s 2 ) . We insert Lemma 8.8 with t = 1 t=1 t = 1 into (8.16), and then Lemma 8.10. Since ℓ D 2 ∫ A = − ( r 2 − s 2 ) / ( 4 β ) \ell_{\mathrm{D}}^2\int A=-(r^2-s^2)/(4\beta) ℓ D 2 ∫ A = − ( r 2 − s 2 ) / ( 4 β ) , the term ℓ D 2 ∫ A \ell_{\mathrm{D}}^2\int A ℓ D 2 ∫ A cancels the confinement factor, and we obtain
1 β log ρ 2 , 0 ( N + 2 ) ( 0 , r e 1 ) ρ 2 , 0 ( N + 2 ) ( 0 , s e 1 ) = log r s − ( E 0 Y ( A ) − ℓ D 2 ∫ A ) + 2 β Cov 0 [ Y ( A ) , Y ( g 1 ) ] + β 2 ( Var 0 [ Y ( D r ) ] − Var 0 [ Y ( D s ) ] ) + O ( β ( 1 + ∣ log β ∣ ) 3 / 2 ) . (94) \frac{1}{\beta}\log\frac{\rho_{2,0}^{(N+2)}(0,r e_1)}{\rho_{2,0}^{(N+2)}(0,s e_1)}
=\log\frac{r}{s}-\left(\mathbb{E}_0Y(A)-\ell_{\mathrm{D}}^2\int A\right)+2\beta\operatorname{Cov}_0[Y(A),Y(g_1)]
+\frac{\beta}{2}\left(\operatorname{Var}_0[Y(D_r)]-\operatorname{Var}_0[Y(D_s)]\right)+O\left(\sqrt{\beta}(1+|\log\beta|)^{3/2}\right). \tag*{(94)} β 1 log ρ 2 , 0 ( N + 2 ) ( 0 , s e 1 ) ρ 2 , 0 ( N + 2 ) ( 0 , r e 1 ) = log s r − ( E 0 Y ( A ) − ℓ D 2 ∫ A ) + 2 β Cov 0 [ Y ( A ) , Y ( g 1 )] + 2 β ( Var 0 [ Y ( D r )] − Var 0 [ Y ( D s )] ) + O ( β ( 1 + ∣ log β ∣ ) 3/2 ) . (94)
where the error is bounded uniformly in N N N , r r r and s s s .
Step 2. The limit N → ∞ N\to\infty N → ∞ . We let N → ∞ N\to\infty N → ∞ at fixed β \beta β . Since A A A is compactly supported and Y Y Y has intensity ℓ D 2 \ell_{\mathrm{D}}^2 ℓ D 2 under Ginβ _\beta β , the centred mean E 0 Y ( A ) − ℓ D 2 ∫ A \mathbb{E}_0Y(A)-\ell_{\mathrm{D}}^2\int A E 0 Y ( A ) − ℓ D 2 ∫ A tends to zero.
For the covariances in (8.27), we first truncate the functions, with errors that are uniform in N N N . By (8.17), replacing g g g by g ε g_\varepsilon g ε in D r D_r D r and D s D_s D s changes these covariances by a quantity that tends to zero with ε \varepsilon ε . By (8.5), the centred statistics of ( 1 − χ R ) D r (1-\chi_R)D_r ( 1 − χ R ) D r and ( 1 − χ R ) D s (1-\chi_R)D_s ( 1 − χ R ) D s have L 2 L^2 L 2 norms at most C ( β ) R − 1 C(\beta)R^{-1} C ( β ) R − 1 . By (8.3), the covariance of Y ( A ) Y(A) Y ( A ) with Y ( ( χ 2 j − χ 2 j − 1 ) g 1 ) Y((\chi_{2j}-\chi_{2j-1})g_1) Y (( χ 2 j − χ 2 j − 1 ) g 1 ) is at most C ( β ) ( 1 + j ) e − c j 2 C(\beta)(1+j)e^{-cj^2} C ( β ) ( 1 + j ) e − c j 2 , which is summable in j j j . After these truncations, the covariances converge by the local convergence of the soft gas to Ginβ _\beta β and by (8.4). This shows that the limits that define C β ( A , g 1 ) C_\beta(A,g_1) C β ( A , g 1 ) , C β ( D r , D r ) C_\beta(D_r,D_r) C β ( D r , D r ) and C β ( D s , D s ) C_\beta(D_s,D_s) C β ( D s , D s ) exist, and that the covariances in (8.27) converge to them, multiplied by ℓ D 2 \ell_{\mathrm{D}}^2 ℓ D 2 .
The same argument applies to the limit (8.25), with g 1 g_1 g 1 in place of g 1 g_1 g 1 . Since Ginβ _\beta β is rotation-invariant and χ R g ε \chi_Rg_\varepsilon χ R g ε is radial, we may replace the second argument of (8.25) by its angular average. Angular averaging commutes with the convolution by ψ ε \psi_\varepsilon ψ ε , so these angular averages converge to A A A in H 1 H^1 H 1 as ε → 0 \varepsilon\to0 ε → 0 , and the bounds on the covariances with the dyadic pieces of g ε g_\varepsilon g ε justify the limit ε → 0 \varepsilon\to0 ε → 0 .
Step 3. Positivity. By (8.8), the integral of q β ( ∣ ⋅ ∣ ) q_\beta(|\cdot|) q β ( ∣ ⋅ ∣ ) against a non-negative radial function in C c ∞ ( R 2 ) C^\infty_c(\mathbb{R}^2) C c ∞ ( R 2 ) , supported in { r 0 < ∣ z ∣ < 1 / r 0 } \{r_0<|z|<1/r_0\} { r 0 < ∣ z ∣ < 1/ r 0 } and with integral one, is positive for β \beta β small. Hence q β q_\beta q β is positive at some distance in [ r 0 , 1 / r 0 ] [r_0,1/r_0] [ r 0 , 1/ r 0 ] . At fixed β \beta β , the right-hand side of (8.27) is bounded uniformly in N ≥ 4 / β N\ge4/\beta N ≥ 4/ β and r 0 ≤ r , s ≤ 1 / r 0 r_0\le r,s\le1/r_0 r 0 ≤ r , s ≤ 1/ r 0 . We take s s s equal to this distance and let N → ∞ N\to\infty N → ∞ . This shows that q β q_\beta q β is positive on [ r 0 , 1 / r 0 ] [r_0,1/r_0] [ r 0 , 1/ r 0 ] , and that the logarithms in (8.27) converge.
Step 4. Cancellation of the covariances. Since 2 β ℓ D 2 = 1 / π 2\beta\ell_D^2=1/\pi 2 β ℓ D 2 = 1/ π , the covariance terms of (94) converge to π − 1 C β ( A , g 1 ) \pi^{-1}C_\beta(A,g_1) π − 1 C β ( A , g 1 ) and ( 4 π ) − 1 ( C β ( D r , D r ) − C β ( D s , D s ) ) (4\pi)^{-1}(C_\beta(D_r,D_r)-C_\beta(D_s,D_s)) ( 4 π ) − 1 ( C β ( D r , D r ) − C β ( D s , D s )) . We now show that
1 π C β ( A , g 1 ) + C β ( D r , D r ) − C β ( D s , D s ) 4 π = 1 2 π C β ( g , D r − D s ) . (95) \frac{1}{\pi}C_\beta(A,g_1)+\frac{C_\beta(D_r,D_r)-C_\beta(D_s,D_s)}{4\pi}=\frac{1}{2\pi}C_\beta(g,D_r-D_s). \tag*{(95)} π 1 C β ( A , g 1 ) + 4 π C β ( D r , D r ) − C β ( D s , D s ) = 2 π 1 C β ( g , D r − D s ) . (95)
We first compute formally with (74) . The Fourier transforms of g g g , g 1 g_1 g 1 and D r D_r D r are
g ^ ( k ) = 2 π ∣ k ∣ 2 , g 1 ^ ( k ) = 2 π ∣ k ∣ 2 ψ ^ ( k ) and D r ^ ( k ) = 2 π ∣ k ∣ 2 ( 1 + e − i r k 1 − 2 ψ ^ ( k ) ) . \widehat{g}(k)=\frac{2\pi}{|k|^2},\qquad\widehat{g_1}(k)=\frac{2\pi}{|k|^2}\widehat{\psi}(k)\qquad\text{and}\qquad\widehat{D_r}(k)=\frac{2\pi}{|k|^2}\left(1+e^{-irk_1}-2\widehat{\psi}(k)\right). g ( k ) = ∣ k ∣ 2 2 π , g 1 ( k ) = ∣ k ∣ 2 2 π ψ ( k ) and D r ( k ) = ∣ k ∣ 2 2 π ( 1 + e − i r k 1 − 2 ψ ( k ) ) .
Since σ G i n β \sigma_{\mathrm{Gin}_\beta} σ Gin β and ψ ^ \widehat{\psi} ψ are real and radial, we may replace A ^ \widehat{A} A by D r ^ − D s ^ \widehat{D_r}-\widehat{D_s} D r − D s , and e − i r k 1 e^{-irk_1} e − i r k 1 by cos ( r k 1 ) \cos(rk_1) cos ( r k 1 ) . By (74) , the three terms of (95) are then the integrals of σ G i n β ( k / ℓ D ) ∣ k ∣ − 4 ( cos ( r k 1 ) − cos ( s k 1 ) ) \sigma_{\mathrm{Gin}_\beta}(k/\ell_D)|k|^{-4}(\cos(rk_1)-\cos(sk_1)) σ Gin β ( k / ℓ D ) ∣ k ∣ − 4 ( cos ( r k 1 ) − cos ( s k 1 )) multiplied by
ψ ^ ( k ) π , 1 − 2 ψ ^ ( k ) 2 π , and 1 2 π , \frac{\widehat{\psi}(k)}{\pi},\qquad\frac{1-2\widehat{\psi}(k)}{2\pi},\qquad\text{and}\qquad\frac{1}{2\pi}, π ψ ( k ) , 2 π 1 − 2 ψ ( k ) , and 2 π 1 ,
respectively, and the first two add up to the third. These computations are formal, since the Fourier transforms above are singular at the origin. At a fixed ε \varepsilon ε , interchanging a bounded translation and the cutoff χ R \chi_R χ R produces an error whose gradient has L 2 L^2 L 2 norm at most C ( ε ) R − 1 log R C(\varepsilon)R^{-1}\log R C ( ε ) R − 1 log R . The gradients of the functions with the cutoff χ R \chi_R χ R have L 2 L^2 L 2 norms at most C ( ε ) log R C(\varepsilon)\log R C ( ε ) log R , so (73) bounds the resulting errors in the covariances by C ( ε ) R − 1 ( log R ) 2 C(\varepsilon)R^{-1}(\log R)^2 C ( ε ) R − 1 ( log R ) 2 . Hence the computation above holds in the limit R → ∞ R\to\infty R → ∞ , at fixed ε \varepsilon ε , and Step 2 justifies the limit ε → 0 \varepsilon\to0 ε → 0 . Combining (94) and (95) gives (93) . □
We now compute the limit of the covariance in (93) .
Lemma 8.12 (Limit of the logarithmic covariance). Let 0 < r 0 < 1 0<r_0<1 0 < r 0 < 1 . Then, uniformly for r 0 ≤ r , s ≤ 1 / r 0 r_0\le r,s\le1/r_0 r 0 ≤ r , s ≤ 1/ r 0 ,
1 2 π C β ( g , D r − D s ) → β ℓ D 2 → ∞ log s r − K 0 ( r ) + K 0 ( s ) . \frac{1}{2\pi}C_\beta(g,D_r-D_s)\xrightarrow[\beta\ell_D^2\to\infty]{}\log\frac{s}{r}-K_0(r)+K_0(s). 2 π 1 C β ( g , D r − D s ) β ℓ D 2 → ∞ log r s − K 0 ( r ) + K 0 ( s ) .
Proof. Let J 0 J_0 J 0 be the Bessel function of the first kind of order zero. We integrate (74) in polar coordinates, with g ^ ( k ) = 2 π ∣ k ∣ − 2 \widehat{g}(k)=2\pi|k|^{-2} g ( k ) = 2 π ∣ k ∣ − 2 , and obtain
1 2 π C β ( g , D r − D s ) = ∫ 0 ∞ σ G i n β ( ϱ e 1 / ℓ D ) J 0 ( r ϱ ) − J 0 ( s ϱ ) ϱ 3 d ϱ . (96) \frac{1}{2\pi}C_\beta(g,D_r-D_s)=\int_0^\infty\sigma_{\mathrm{Gin}_\beta}(\varrho e_1/\ell_D)\frac{J_0(r\varrho)-J_0(s\varrho)}{\varrho^3}\,d\varrho. \tag*{(96)} 2 π 1 C β ( g , D r − D s ) = ∫ 0 ∞ σ Gin β ( ϱ e 1 / ℓ D ) ϱ 3 J 0 ( r ϱ ) − J 0 ( s ϱ ) d ϱ . (96)
The constant Fourier mode of g g g does not contribute, since σ G i n β ( 0 ) = 0 \sigma_{\mathrm{Gin}_\beta}(0)=0 σ Gin β ( 0 ) = 0 . By (73) and the bounds on J 0 J_0 J 0 at zero and at infinity,
∣ σ G i n β ( ϱ e 1 / ℓ D ) J 0 ( r ϱ ) − J 0 ( s ϱ ) ϱ 3 ∣ ≤ { C ϱ , 0 < ϱ ≤ 1 , C ϱ − 3 / 2 , ϱ ≥ 1. \left|\sigma_{\mathrm{Gin}_\beta}(\varrho e_1/\ell_D)\frac{J_0(r\varrho)-J_0(s\varrho)}{\varrho^3}\right|\le\begin{cases}C\varrho,&0<\varrho\le1,\\ C\varrho^{-3/2},&\varrho\ge1.\end{cases} σ Gin β ( ϱ e 1 / ℓ D ) ϱ 3 J 0 ( r ϱ ) − J 0 ( s ϱ ) ≤ { C ϱ , C ϱ − 3/2 , 0 < ϱ ≤ 1 , ϱ ≥ 1.
On every compact subinterval of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , the functions ϱ ↦ ( J 0 ( r ϱ ) − J 0 ( s ϱ ) ) ϱ − 3 \varrho\mapsto(J_0(r\varrho)-J_0(s\varrho))\varrho^{-3} ϱ ↦ ( J 0 ( r ϱ ) − J 0 ( s ϱ )) ϱ − 3 with r 0 ≤ r , s ≤ 1 / r 0 r_0\le r,s\le1/r_0 r 0 ≤ r , s ≤ 1/ r 0 form a compact family of continuous functions. By Proposition 8.2, the non-negative functions σ G i n β ( ⋅ / ℓ D ) \sigma_{\mathrm{Gin}_\beta}(\cdot/\ell_D) σ Gin β ( ⋅ / ℓ D ) converge in S ′ ( R 2 ) S'(\mathbb{R}^2) S ′ ( R 2 ) , hence against continuous functions with compact support. With these uniform tail bounds, (96) converges uniformly to
∫ 0 ∞ J 0 ( r ϱ ) − J 0 ( s ϱ ) ϱ ( 1 + ϱ 2 ) d ϱ = log s r − K 0 ( r ) + K 0 ( s ) . \int_0^\infty\frac{J_0(r\varrho)-J_0(s\varrho)}{\varrho(1+\varrho^2)}\,d\varrho=\log\frac{s}{r}-K_0(r)+K_0(s). ∫ 0 ∞ ϱ ( 1 + ϱ 2 ) J 0 ( r ϱ ) − J 0 ( s ϱ ) d ϱ = log r s − K 0 ( r ) + K 0 ( s ) .
This identity follows from 1 / ( ϱ ( 1 + ϱ 2 ) ) = 1 / ϱ − ϱ / ( 1 + ϱ 2 ) 1/(\varrho(1+\varrho^2))=1/\varrho-\varrho/(1+\varrho^2) 1/ ( ϱ ( 1 + ϱ 2 )) = 1/ ϱ − ϱ / ( 1 + ϱ 2 ) , from ∫ 0 ∞ ( J 0 ( r ϱ ) − J 0 ( s ϱ ) ) ϱ − 1 d ϱ = log ( s / r ) \int_0^\infty(J_0(r\varrho)-J_0(s\varrho))\varrho^{-1}\,d\varrho=\log(s/r) ∫ 0 ∞ ( J 0 ( r ϱ ) − J 0 ( s ϱ )) ϱ − 1 d ϱ = log ( s / r ) and from ∫ 0 ∞ J 0 ( r ϱ ) ϱ ( 1 + ϱ 2 ) − 1 d ϱ = K 0 ( r ) \int_0^\infty J_0(r\varrho)\varrho(1+\varrho^2)^{-1}\,d\varrho=K_0(r) ∫ 0 ∞ J 0 ( r ϱ ) ϱ ( 1 + ϱ 2 ) − 1 d ϱ = K 0 ( r ) . □
In (93) , the limit of the covariance therefore cancels the term log ( r / s ) \log(r/s) log ( r / s ) of the interaction of the two inserted particles, and replaces it by − K 0 ( r ) + K 0 ( s ) -K_0(r)+K_0(s) − K 0 ( r ) + K 0 ( s ) . Physically, the gas screens the interaction of the two particles at distances larger than the Debye length.
Proof of Theorem 5. The convergence in S ′ ( R 2 ) \mathcal{S}'(\mathbb{R}^2) S ′ ( R 2 ) is (75) . Let 0 < r 0 < 1 0<r_0<1 0 < r 0 < 1 . By Proposition 8.11 and Lemma 8.12 , the function q β q_\beta q β is positive on [ r 0 , 1 / r 0 ] [r_0,1/r_0] [ r 0 , 1/ r 0 ] for β \beta β small, and
sup r 0 ≤ r , s ≤ 1 / r 0 ∣ 1 β log q β ( r ) q β ( s ) + K 0 ( r ) − K 0 ( s ) ∣ → β → 0 0. \sup_{r_0\le r,s\le1/r_0}\left|\frac{1}{\beta}\log\frac{q_\beta(r)}{q_\beta(s)}+K_0(r)-K_0(s)\right|\xrightarrow[\beta\to0]{}0. r 0 ≤ r , s ≤ 1/ r 0 sup β 1 log q β ( s ) q β ( r ) + K 0 ( r ) − K 0 ( s ) β → 0 0.
The function K 0 K_0 K 0 is bounded on [ r 0 , 1 / r 0 ] [r_0,1/r_0] [ r 0 , 1/ r 0 ] , so exponentiating this limit gives, uniformly for r 0 ≤ r , s ≤ 1 / r 0 r_0\le r,s\le1/r_0 r 0 ≤ r , s ≤ 1/ r 0 ,
q β ( r ) = q β ( s ) ( 1 − β K 0 ( r ) + β K 0 ( s ) + o ( β ) ) . q_\beta(r)=q_\beta(s)\left(1-\beta K_0(r)+\beta K_0(s)+o(\beta)\right). q β ( r ) = q β ( s ) ( 1 − β K 0 ( r ) + β K 0 ( s ) + o ( β ) ) .
Let φ ∈ C c ∞ ( R 2 ) \varphi\in C^\infty_c(\mathbb{R}^2) φ ∈ C c ∞ ( R 2 ) be non-negative and radial, supported in { r 0 < ∣ z ∣ < 1 / r 0 } \{r_0<|z|<1/r_0\} { r 0 < ∣ z ∣ < 1/ r 0 } , with ∫ φ = 1 \int\varphi=1 ∫ φ = 1 . By rotation invariance of Gin β \operatorname{Gin}_\beta Gin β , we have ρ 2 , β ( 0 , ℓ d z ) = q β ( ∣ z ∣ ) \rho_{2,\beta}(0,\ell\,\mathrm{d}z)=q_\beta(|z|) ρ 2 , β ( 0 , ℓ d z ) = q β ( ∣ z ∣ ) . We integrate the expansion of q β ( r ) q_\beta(r) q β ( r ) above against φ \varphi φ , with r = ∣ z ∣ r=|z| r = ∣ z ∣ , and obtain
∫ R 2 φ ( z ) q β ( ∣ z ∣ ) d z = q β ( s ) ( 1 + β K 0 ( s ) − β ∫ R 2 φ ( z ) K 0 ( ∣ z ∣ ) d z + o ( β ) ) . \int_{\mathbb{R}^2}\varphi(z)q_\beta(|z|)\,\mathrm{d}z=q_\beta(s)\left(1+\beta K_0(s)-\beta\int_{\mathbb{R}^2}\varphi(z)K_0(|z|)\,\mathrm{d}z+o(\beta)\right). ∫ R 2 φ ( z ) q β ( ∣ z ∣ ) d z = q β ( s ) ( 1 + β K 0 ( s ) − β ∫ R 2 φ ( z ) K 0 ( ∣ z ∣ ) d z + o ( β ) ) .
By (75) , the left-hand side is 1 − β ∫ φ ( z ) K 0 ( ∣ z ∣ ) d z + o ( β ) 1-\beta\int\varphi(z)K_0(|z|)\,\mathrm{d}z+o(\beta) 1 − β ∫ φ ( z ) K 0 ( ∣ z ∣ ) d z + o ( β ) . Comparing the two expressions, we obtain q β ( s ) = 1 − β K 0 ( s ) + o ( β ) q_\beta(s)=1-\beta K_0(s)+o(\beta) q β ( s ) = 1 − β K 0 ( s ) + o ( β ) , uniformly for r 0 ≤ s ≤ 1 / r 0 r_0\le s\le1/r_0 r 0 ≤ s ≤ 1/ r 0 . As r 0 r_0 r 0 is arbitrary, this is the locally uniform convergence on R 2 ∖ { 0 } \mathbb{R}^2\setminus\{0\} R 2 ∖ { 0 } .
Appendix A. Integration by parts and the Helffer–Sjöstrand equation
This appendix explains the integration by parts behind the Poincaré inequalities of the paper, explains why the Helffer–Sjöstrand equation is natural, and proves the Helffer–Sjöstrand representation of Proposition 5.2.
Brascamp–Lieb and Hörmander: the same integration by parts
A uniformly positive Hessian gives a Poincaré inequality, and a uniformly positive Levi matrix gives (19) . In both cases the proof is one integration by parts, written with a first-order operator, its adjoint, and a commutator. Uniform convexity enters the real argument through the Hessian lower bound, and the Levi matrix plays the same role in the complex argument.
We do the computations for smooth compactly supported functions, and do not discuss domains of operators. On the polydisk B L N B_L^N B L N , the integration by parts also produces a boundary term, which is non-negative because the polydisk is pseudoconvex [Hör65]. This is the complex analogue of the Brascamp–Lieb inequality on a convex domain. These domains are standard, see [15 Chapters 3–4] for the real case, and [Hör65, Hör90, Ber10] for the complex case. Throughout this subsection, Φ \Phi Φ is smooth and μ = e − Φ d Z \mu=e^{-\Phi}\mathrm{d}Z μ = e − Φ d Z is a probability measure.
We use the following elementary lemma.
Lemma A.1. Let L ≥ 0 \mathcal{L}\ge0 L ≥ 0 be a self-adjoint operator on a Hilbert space, and c > 0 c>0 c > 0 . Assume that ∥ L f ∥ 2 ≥ c ⟨ f , L f ⟩ \lVert\mathcal{L}f\rVert^2\ge c\langle f,\mathcal{L}f\rangle ∥ L f ∥ 2 ≥ c ⟨ f , L f ⟩ for all f ∈ Dom L f\in\operatorname{Dom}\mathcal{L} f ∈ Dom L . Then, denoting by Π \Pi Π the orthogonal projection onto ker L \ker\mathcal{L} ker L ,
∥ f − Π f ∥ 2 ≤ c − 1 ⟨ f , L f ⟩ for all f ∈ Dom L . \lVert f-\Pi f\rVert^2\le c^{-1}\langle f,\mathcal{L}f\rangle\qquad\text{for all }f\in\operatorname{Dom}\mathcal{L}. ∥ f − Π f ∥ 2 ≤ c − 1 ⟨ f , L f ⟩ for all f ∈ Dom L .
Proof. Let ν f \nu_f ν f be the spectral measure of L \mathcal{L} L associated with f f f . The assumption reads ∫ λ 2 d ν f ≥ c ∫ λ d ν f \int\lambda^2\,\mathrm{d}\nu_f\ge c\int\lambda\,\mathrm{d}\nu_f ∫ λ 2 d ν f ≥ c ∫ λ d ν f for all f f f . We apply it to the spectral projection of f f f on ( 0 , c ) (0,c) ( 0 , c ) , and see that L \mathcal{L} L has no spectrum in ( 0 , c ) (0,c) ( 0 , c ) . Hence ∥ f − Π f ∥ 2 = ν f ( ( 0 , ∞ ) ) = ν f ( [ c , ∞ ) ) ≤ c − 1 ∫ λ d ν f = c − 1 ⟨ f , L f ⟩ \lVert f-\Pi f\rVert^2=\nu_f((0,\infty))=\nu_f([c,\infty))\le c^{-1}\int\lambda\,\mathrm{d}\nu_f=c^{-1}\langle f,\mathcal{L}f\rangle ∥ f − Π f ∥ 2 = ν f (( 0 , ∞ )) = ν f ([ c , ∞ )) ≤ c − 1 ∫ λ d ν f = c − 1 ⟨ f , L f ⟩ .
In each case below, L = D ∗ D \mathcal{L}=D^*D L = D ∗ D for a first-order operator D D D (the gradient, or ∂ ˉ \bar\partial ∂ ˉ ), hence ⟨ F , L F ⟩ = ∥ D F ∥ 2 \langle F,\mathcal{L}F\rangle=\lVert DF\rVert^2 ⟨ F , L F ⟩ = ∥ D F ∥ 2 and ker L = ker D \ker\mathcal{L}=\ker D ker L = ker D . It remains to prove the inequality ∥ F L ∥ 2 ≥ c ∥ D F ∥ 2 \lVert F\mathcal{L}\rVert^2\ge c\lVert DF\rVert^2 ∥ F L ∥ 2 ≥ c ∥ D F ∥ 2 , and we do it by an integration by parts.
One real particle
Write Z = ( x , y ) ∈ R 2 Z=(x,y)\in\mathbb{R}^2 Z = ( x , y ) ∈ R 2 , let α , γ ∈ { x , y } \alpha,\gamma\in\{x,y\} α , γ ∈ { x , y } , and let ⟨ f , g ⟩ = ∫ f g d μ \langle f,g\rangle=\int fg\,\mathrm{d}\mu ⟨ f , g ⟩ = ∫ f g d μ be the real scalar product. The adjoint. Integrating by parts, ∫ ( ∂ α f ) g e − Φ = ∫ f ( − ∂ α g + ( ∂ α Φ ) g ) e − Φ \int(\partial_{\alpha}f)g e^{-\Phi} = \int f(-\partial_{\alpha}g + (\partial_{\alpha}\Phi)g)e^{-\Phi} ∫ ( ∂ α f ) g e − Φ = ∫ f ( − ∂ α g + ( ∂ α Φ ) g ) e − Φ , hence, for the operator L R \mathcal{L}_{\mathbb{R}} L R of (21) , we have
∂ α ∗ = − ∂ α + ∂ α Φ , L R = ∂ x ∗ ∂ x + ∂ y ∗ ∂ y = − Δ + ∇ Φ ⋅ ∇ , ⟨ F , L R F ⟩ = ∫ ∣ ∇ F ∣ 2 d μ . \partial_{\alpha}^{*}=-\partial_{\alpha}+\partial_{\alpha}\Phi,\qquad\mathcal{L}_{\mathbb{R}}=\partial_{x}^{*}\partial_{x}+\partial_{y}^{*}\partial_{y}=-\Delta+\nabla\Phi\cdot\nabla,\qquad\langle F,\mathcal{L}_{\mathbb{R}}F\rangle=\int|\nabla F|^{2}d\mu. ∂ α ∗ = − ∂ α + ∂ α Φ , L R = ∂ x ∗ ∂ x + ∂ y ∗ ∂ y = − Δ + ∇Φ ⋅ ∇ , ⟨ F , L R F ⟩ = ∫ ∣∇ F ∣ 2 d μ .
The commutator. For every g g g , [ ∂ γ , ∂ α ∗ ] g = ∂ γ ( − ∂ α g + ( ∂ α Φ ) g ) − ( − ∂ α ∂ γ g + ( ∂ α Φ ) ∂ γ g ) = ( ∂ γ ∂ α Φ ) g [\partial_{\gamma},\partial_{\alpha}^{*}]g=\partial_{\gamma}(-\partial_{\alpha}g+(\partial_{\alpha}\Phi)g)-(-\partial_{\alpha}\partial_{\gamma}g+(\partial_{\alpha}\Phi)\partial_{\gamma}g)=(\partial_{\gamma}\partial_{\alpha}\Phi)g [ ∂ γ , ∂ α ∗ ] g = ∂ γ ( − ∂ α g + ( ∂ α Φ ) g ) − ( − ∂ α ∂ γ g + ( ∂ α Φ ) ∂ γ g ) = ( ∂ γ ∂ α Φ ) g . The commutator is the multiplication by an entry of the Hessian:
[ ∂ γ , ∂ α ∗ ] = ∂ γ ∂ α Φ . [\partial_{\gamma},\partial_{\alpha}^{*}]=\partial_{\gamma}\partial_{\alpha}\Phi. [ ∂ γ , ∂ α ∗ ] = ∂ γ ∂ α Φ.
The integration by parts. Using the definition of the adjoint, then the commutator,
∥ L R F ∥ 2 = ∑ α , γ ⟨ ∂ γ ∗ ∂ γ F , ∂ α ∗ ∂ α F ⟩ = ∑ α , γ ⟨ ∂ γ F , ∂ γ ∂ α ∗ ∂ α F ⟩ = ∑ α , γ ⟨ ∂ γ F , ∂ α ∗ ∂ γ ∂ α F ⟩ + ∑ α , γ ⟨ ∂ γ F , ( ∂ γ ∂ α Φ ) ∂ α F ⟩ = ∑ α , γ ∥ ∂ α ∂ γ F ∥ 2 + ∫ ∇ F ⋅ ∇ 2 Φ ∇ F d μ . \begin{aligned}
\|\mathcal{L}_{\mathbb{R}}F\|^{2}
&=\sum_{\alpha,\gamma}\langle\partial_{\gamma}^{*}\partial_{\gamma}F,\partial_{\alpha}^{*}\partial_{\alpha}F\rangle
=\sum_{\alpha,\gamma}\langle\partial_{\gamma}F,\partial_{\gamma}\partial_{\alpha}^{*}\partial_{\alpha}F\rangle\\
&=\sum_{\alpha,\gamma}\langle\partial_{\gamma}F,\partial_{\alpha}^{*}\partial_{\gamma}\partial_{\alpha}F\rangle+\sum_{\alpha,\gamma}\langle\partial_{\gamma}F,(\partial_{\gamma}\partial_{\alpha}\Phi)\partial_{\alpha}F\rangle\\
&=\sum_{\alpha,\gamma}\|\partial_{\alpha}\partial_{\gamma}F\|^{2}+\int\nabla F\cdot\nabla^{2}\Phi\,\nabla Fd\mu.
\end{aligned} ∥ L R F ∥ 2 = α , γ ∑ ⟨ ∂ γ ∗ ∂ γ F , ∂ α ∗ ∂ α F ⟩ = α , γ ∑ ⟨ ∂ γ F , ∂ γ ∂ α ∗ ∂ α F ⟩ = α , γ ∑ ⟨ ∂ γ F , ∂ α ∗ ∂ γ ∂ α F ⟩ + α , γ ∑ ⟨ ∂ γ F , ( ∂ γ ∂ α Φ ) ∂ α F ⟩ = α , γ ∑ ∥ ∂ α ∂ γ F ∥ 2 + ∫ ∇ F ⋅ ∇ 2 Φ ∇ F d μ .
In the last line, we used the adjoint once more and ∂ γ ∂ α = ∂ α ∂ γ \partial_{\gamma}\partial_{\alpha}=\partial_{\alpha}\partial_{\gamma} ∂ γ ∂ α = ∂ α ∂ γ . If ∇ 2 Φ ≥ a \nabla^{2}\Phi\ge a ∇ 2 Φ ≥ a , the right-hand side is at least a ∫ ∣ ∇ F ∣ 2 d μ = a ⟨ F , L R F ⟩ a\int|\nabla F|^{2}d\mu=a\langle F,\mathcal{L}_{\mathbb{R}}F\rangle a ∫ ∣∇ F ∣ 2 d μ = a ⟨ F , L R F ⟩ . We apply Lemma A.1 with c = a c=a c = a . The kernel of ∇ \nabla ∇ consists of the constants, so Π F = E μ F \Pi F=\mathbb{E}_{\mu}F Π F = E μ F and
Var μ [ F ] ≤ a − 1 E μ ∣ ∇ F ∣ 2 . \operatorname{Var}_{\mu}[F]\le a^{-1}\mathbb{E}_{\mu}|\nabla F|^{2}. Var μ [ F ] ≤ a − 1 E μ ∣∇ F ∣ 2 .
Two real particles
Now Z = ( z 1 , z 2 ) ∈ ( R 2 ) 2 Z=(z_{1},z_{2})\in(\mathbb{R}^{2})^{2} Z = ( z 1 , z 2 ) ∈ ( R 2 ) 2 . There is one derivative ∂ α \partial_{\alpha} ∂ α for each of the four real coordinates α \alpha α of z 1 z_{1} z 1 and z 2 z_{2} z 2 , and the computation is word for word the same, with sums over α \alpha α and γ \gamma γ :
∥ L R F ∥ 2 = ∑ α , γ ∥ ∂ α ∂ γ F ∥ 2 + ∫ ∇ F ⋅ ∇ 2 Φ ∇ F d μ . \|\mathcal{L}_{\mathbb{R}}F\|^{2}=\sum_{\alpha,\gamma}\|\partial_{\alpha}\partial_{\gamma}F\|^{2}+\int\nabla F\cdot\nabla^{2}\Phi\,\nabla Fd\mu. ∥ L R F ∥ 2 = α , γ ∑ ∥ ∂ α ∂ γ F ∥ 2 + ∫ ∇ F ⋅ ∇ 2 Φ ∇ F d μ .
The Hessian is now a 4 × 4 4\times4 4 × 4 matrix made of four 2 × 2 2\times2 2 × 2 blocks ∇ z i ∇ z k Φ \nabla_{z_i}\nabla_{z_k}\Phi ∇ z i ∇ z k Φ . The off-diagonal blocks describe how the force on one particle varies when the other one moves. For the pair interaction − β log ∣ z 1 − z 2 ∣ -\beta\log|z_{1}-z_{2}| − β log ∣ z 1 − z 2 ∣ , the Hessian has the eigenvalue − 2 β ∣ z 1 − z 2 ∣ − 2 -2\beta|z_{1}-z_{2}|^{-2} − 2 β ∣ z 1 − z 2 ∣ − 2 , in the direction ( i ( z 1 − z 2 ) , − i ( z 1 − z 2 ) ) (i(z_{1}-z_{2}),-i(z_{1}-z_{2})) ( i ( z 1 − z 2 ) , − i ( z 1 − z 2 )) . It is unbounded below near the collision, and no confinement a ∣ Z ∣ 2 a|Z|^{2} a ∣ Z ∣ 2 can make the condition ∇ 2 Φ ≥ a \nabla^{2}\Phi\ge a ∇ 2 Φ ≥ a hold.
One complex particle
We write z = x + i y z=x+iy z = x + i y for the point Z = ( x , y ) Z=(x,y) Z = ( x , y ) , and we set
∂ ˉ : = 1 2 ( ∂ x + i ∂ y ) , ∂ : = 1 2 ( ∂ x − i ∂ y ) . \bar{\partial}:=\frac{1}{2}(\partial_{x}+i\partial_{y}),\qquad\partial:=\frac{1}{2}(\partial_{x}-i\partial_{y}). ∂ ˉ := 2 1 ( ∂ x + i ∂ y ) , ∂ := 2 1 ( ∂ x − i ∂ y ) .
We work with complex-valued functions and the Hermitian product ⟨ F , G ⟩ = ∫ F G ‾ d μ \langle F,G\rangle=\int F\overline{G}d\mu ⟨ F , G ⟩ = ∫ F G d μ . The weight Φ \Phi Φ stays real-valued.
The adjoint. Integrating by parts, and using ∂ G ‾ = ∂ ˉ G ‾ \overline{\partial G}=\bar{\partial}\overline{G} ∂ G = ∂ ˉ G and ∂ Φ = ∂ ˉ Φ \partial\Phi=\bar{\partial}\Phi ∂ Φ = ∂ ˉ Φ (Φ \Phi Φ is real), ∫ ( ∂ ˉ F ) G ‾ e − Φ = ∫ F ( − ∂ G + ( ∂ Φ ) G ) e − Φ \int(\bar{\partial}F)\overline{G}e^{-\Phi}=\int F(-\partial G+(\partial\Phi)G)e^{-\Phi} ∫ ( ∂ ˉ F ) G e − Φ = ∫ F ( − ∂ G + ( ∂ Φ ) G ) e − Φ . Hence
∂ ˉ ∗ = − ∂ + ∂ Φ , L = ∂ ˉ ∗ ∂ ˉ , ⟨ F , L F ⟩ = ∫ ∣ ∂ ˉ F ∣ 2 d μ . \bar{\partial}^{*}=-\partial+\partial\Phi,\qquad\mathcal{L}=\bar{\partial}^{*}\bar{\partial},\qquad\langle F,\mathcal{L}F\rangle=\int|\bar{\partial}F|^{2}d\mu. ∂ ˉ ∗ = − ∂ + ∂ Φ , L = ∂ ˉ ∗ ∂ ˉ , ⟨ F , L F ⟩ = ∫ ∣ ∂ ˉ F ∣ 2 d μ .
The commutator. As in the real case, [ ∂ ˉ , ∂ ˉ ∗ ] g = ∂ ˉ ( − ∂ g + ( ∂ Φ ) g ) − ( − ∂ ∂ ˉ g + ( ∂ Φ ) ∂ ˉ g ) = ( ∂ ˉ ∂ Φ ) g [\bar{\partial},\bar{\partial}^{*}]g=\bar{\partial}(-\partial g+(\partial\Phi)g)-(-\partial\bar{\partial}g+(\partial\Phi)\bar{\partial}g)=(\bar{\partial}\partial\Phi)g [ ∂ ˉ , ∂ ˉ ∗ ] g = ∂ ˉ ( − ∂ g + ( ∂ Φ ) g ) − ( − ∂ ∂ ˉ g + ( ∂ Φ ) ∂ ˉ g ) = ( ∂ ˉ ∂ Φ ) g :
[ ∂ ˉ , ∂ ˉ ∗ ] = ∂ ˉ ∂ Φ = 1 4 Δ Φ . [\bar{\partial},\bar{\partial}^{*}]=\bar{\partial}\partial\Phi=\frac{1}{4}\Delta\Phi. [ ∂ ˉ , ∂ ˉ ∗ ] = ∂ ˉ ∂ Φ = 4 1 ΔΦ.
The second derivative which appears is no longer the Hessian, but its complex part ∂ ˉ ∂ Φ \bar{\partial}\partial\Phi ∂ ˉ ∂ Φ , that is, a quarter of the Laplacian. The integration by parts. As before,
∥ L F ∥ 2 = ⟨ ∂ ˉ F , ∂ ˉ ∂ ˉ ∗ ∂ ˉ F ⟩ = ⟨ ∂ ˉ F , ∂ ˉ ∗ ∂ ˉ ∂ ˉ F ⟩ + ⟨ ∂ ˉ F , ( ∂ ∂ ˉ Φ ) ∂ ˉ F ⟩ = ∥ ∂ ˉ ∂ ˉ F ∥ 2 + ∫ ∂ ∂ ˉ Φ ∣ ∂ ˉ F ∣ 2 d μ . \begin{aligned}
\|\mathcal{L}F\|^2=\langle\bar\partial F,\bar\partial\bar\partial^*\bar\partial F\rangle=\langle\bar\partial F,\bar\partial^*\bar\partial\bar\partial F\rangle+\langle\bar\partial F,(\partial\bar\partial\Phi)\bar\partial F\rangle\\
&=\|\bar\partial\bar\partial F\|^2+\int\partial\bar\partial\Phi\,|\bar\partial F|^2d\mu.
\end{aligned} ∥ L F ∥ 2 = ⟨ ∂ ˉ F , ∂ ˉ ∂ ˉ ∗ ∂ ˉ F ⟩ = ⟨ ∂ ˉ F , ∂ ˉ ∗ ∂ ˉ ∂ ˉ F ⟩ + ⟨ ∂ ˉ F , ( ∂ ∂ ˉ Φ ) ∂ ˉ F ⟩ = ∥ ∂ ˉ ∂ ˉ F ∥ 2 + ∫ ∂ ∂ ˉ Φ ∣ ∂ ˉ F ∣ 2 d μ .
If ∂ ∂ ˉ Φ = 1 4 Δ Φ ≥ a \partial\bar\partial\Phi=\frac{1}{4}\Delta\Phi\ge a ∂ ∂ ˉ Φ = 4 1 ΔΦ ≥ a , the right-hand side is at least a ⟨ F , L F ⟩ a\langle F,\mathcal{L}F\rangle a ⟨ F , L F ⟩ . We apply Lemma A.1 with c = a c=a c = a . Here ker L = ker ∂ ˉ \operatorname{ker}\mathcal{L}=\operatorname{ker}\bar\partial ker L = ker ∂ ˉ consists of the holomorphic functions in L 2 ( μ ) L^2(\mu) L 2 ( μ ) , and we get
∥ F − Π ker ∂ ˉ F ∥ 2 ≤ a − 1 E μ ∣ ∂ ˉ F ∣ 2 . \|F-\Pi_{\operatorname{ker}\bar\partial}F\|^2\le a^{-1}E_\mu|\bar\partial F|^2. ∥ F − Π ker ∂ ˉ F ∥ 2 ≤ a − 1 E μ ∣ ∂ ˉ F ∣ 2 .
For a real function F F F , ∣ ∂ ˉ F ∣ 2 = 1 4 ∣ ∇ F ∣ 2 |\bar\partial F|^2=\frac{1}{4}|\nabla F|^2 ∣ ∂ ˉ F ∣ 2 = 4 1 ∣∇ F ∣ 2 . The function Π ker ∂ ˉ F \Pi_{\operatorname{ker}\bar\partial}F Π ker ∂ ˉ F is complex-valued in general, even for real F F F .
Two complex particles
Now Z = ( z 1 , z 2 ) ∈ C 2 Z=(z_1,z_2)\in\mathbb{C}^2 Z = ( z 1 , z 2 ) ∈ C 2 , with the derivatives ∂ ˉ j \bar\partial_j ∂ ˉ j , ∂ j \partial_j ∂ j in the variable z j z_j z j , and ∂ ˉ j ∗ = − ∂ j + ∂ j Φ \bar\partial_j^*=-\partial_j+\partial_j\Phi ∂ ˉ j ∗ = − ∂ j + ∂ j Φ . The commutators are
[ ∂ ˉ k , ∂ ˉ j ∗ ] = ∂ ˉ k ∂ j Φ , [\bar\partial_k,\bar\partial_j^*]=\bar\partial_k\partial_j\Phi, [ ∂ ˉ k , ∂ ˉ j ∗ ] = ∂ ˉ k ∂ j Φ ,
the entries of the Levi matrix (or complex Hessian) ( ∂ j ∂ ˉ k Φ ) j , k = 1 , 2 (\partial_j\bar\partial_k\Phi)_{j,k=1,2} ( ∂ j ∂ ˉ k Φ ) j , k = 1 , 2 , a Hermitian 2 × 2 2\times2 2 × 2 matrix. With L = ∂ ˉ 1 ∗ ∂ ˉ 1 + ∂ ˉ 2 ∗ ∂ ˉ 2 \mathcal{L}=\bar\partial_1^*\bar\partial_1+\bar\partial_2^*\bar\partial_2 L = ∂ ˉ 1 ∗ ∂ ˉ 1 + ∂ ˉ 2 ∗ ∂ ˉ 2 , the same computation gives
∥ L F ∥ 2 = ∑ j , k ∥ ∂ ˉ j ∂ ˉ k F ∥ 2 + ∫ ∑ j , k ( ∂ j ∂ ˉ k Φ ) ∂ ˉ j F ∂ ˉ k F ‾ d μ . \|\mathcal{L}F\|^2=\sum_{j,k}\|\bar\partial_j\bar\partial_kF\|^2+\int\sum_{j,k}(\partial_j\bar\partial_k\Phi)\,\bar\partial_jF\,\overline{\bar\partial_kF}\,d\mu. ∥ L F ∥ 2 = j , k ∑ ∥ ∂ ˉ j ∂ ˉ k F ∥ 2 + ∫ j , k ∑ ( ∂ j ∂ ˉ k Φ ) ∂ ˉ j F ∂ ˉ k F d μ .
If the Levi matrix is ≥ a \ge a ≥ a (as a Hermitian matrix), we get by Lemma A.1 that ∥ F − Π ker ∂ ˉ F ∥ 2 ≤ a − 1 E μ ( ∣ ∂ ˉ 1 F ∣ 2 + ∣ ∂ ˉ 2 F ∣ 2 ) \|F-\Pi_{\operatorname{ker}\bar\partial}F\|^2\le a^{-1}E_\mu(|\bar\partial_1F|^2+|\bar\partial_2F|^2) ∥ F − Π ker ∂ ˉ F ∥ 2 ≤ a − 1 E μ ( ∣ ∂ ˉ 1 F ∣ 2 + ∣ ∂ ˉ 2 F ∣ 2 ) , where ker ∂ ˉ \operatorname{ker}\bar\partial ker ∂ ˉ now consists of the holomorphic functions of ( z 1 , z 2 ) (z_1,z_2) ( z 1 , z 2 ) .
In real coordinates, for v = ( v 1 , v 2 ) ∈ ( R 2 ) 2 v=(v_1,v_2)\in(\mathbb{R}^2)^2 v = ( v 1 , v 2 ) ∈ ( R 2 ) 2 , identified with a vector of C 2 \mathbb{C}^2 C 2 , we have ∑ j , k ( ∂ j ∂ ˉ k Φ ) v j v k = 1 4 ( v ⋅ ∇ 2 Φ v + J 2 v ⋅ ∇ 2 Φ J 2 v ) = v ⋅ C Φ ( Z ) v \sum_{j,k}(\partial_j\bar\partial_k\Phi)v_jv_k=\frac{1}{4}(v\cdot\nabla^2\Phi\,v+J_2v\cdot\nabla^2\Phi\,J_2v)=v\cdot C_\Phi(Z)v ∑ j , k ( ∂ j ∂ ˉ k Φ ) v j v k = 4 1 ( v ⋅ ∇ 2 Φ v + J 2 v ⋅ ∇ 2 Φ J 2 v ) = v ⋅ C Φ ( Z ) v by (28) : the Levi matrix is C Φ C_\Phi C Φ of (27) . For the pair interaction, log ∣ z 1 − z 2 ∣ \log|z_1-z_2| log ∣ z 1 − z 2 ∣ is the real part of the holomorphic function log ( z 1 − z 2 ) \log(z_1-z_2) log ( z 1 − z 2 ) away from the collision, so its Levi matrix vanishes there: the eigenvalue − 2 β ∣ z 1 − z 2 ∣ − 2 -2\beta|z_1-z_2|^{-2} − 2 β ∣ z 1 − z 2 ∣ − 2 of the Hessian is compensated in the average over v v v and J 2 v J_2v J 2 v . At the collision, the Levi matrix of − β log ∣ z 1 − z 2 ∣ -\beta\log|z_1-z_2| − β log ∣ z 1 − z 2 ∣ is a negative Dirac mass (Poincaré–Lelong). The shift by log ∣ V N ∣ 2 \log|V_N|^2 log ∣ V N ∣ 2 of Subsection 4.1 takes care of this.
The two proofs differ only in the first-order operator. When we pass from ∇ \nabla ∇ to ∂ ˉ \bar\partial ∂ ˉ , the Hessian is replaced by the Levi matrix in the commutator, and the constants by the holomorphic functions in the kernel.
The general L 2 L^2 L 2 estimate of Hörmander uses the same identity, applied to general families U = ( U j ) j U=(U_j)_j U = ( U j ) j instead of U = ∂ ˉ F U=\bar\partial F U = ∂ ˉ F .
The Helffer–Sjöstrand equation in the real case
We derive the representation (21) of Subsection 3.2, and explain why it is a natural object. Let μ = e − Φ d x \mu=e^{-\Phi}dx μ = e − Φ d x be a probability measure on R N \mathbb{R}^N R N , where Φ \Phi Φ is smooth and its Hessian is positive: ∇ 2 Φ ≥ a \nabla^2\Phi\ge a ∇ 2 Φ ≥ a for some a > 0 a>0 a > 0 . We compute formally and do not discuss domains of operators.
Two elementary remarks. We write ⟨ ⋅ , ⋅ ⟩ \langle\cdot,\cdot\rangle ⟨ ⋅ , ⋅ ⟩ and ∥ ⋅ ∥ \|\cdot\| ∥ ⋅ ∥ for the scalar product and the norm of L 2 ( μ ) L^2(\mu) L 2 ( μ ) , and also of L 2 ( μ ; R N ) L^2(\mu;\mathbb{R}^N) L 2 ( μ ; R N ) , with ⟨ u , w ⟩ : = ∑ k u k w k d μ \langle u,w\rangle:=\sum_k u_kw_k\,d\mu ⟨ u , w ⟩ := ∑ k u k w k d μ for vector fields. We use two elementary identities.
∂ x j ∗ = − ∂ x j + ( ∂ x j Φ ) , (97) \partial_{x_j}^*=-\partial_{x_j}+(\partial_{x_j}\Phi), \tag*{(97)} ∂ x j ∗ = − ∂ x j + ( ∂ x j Φ ) , (97)
where ( ∂ x j Φ ) (\partial_{x_j}\Phi) ( ∂ x j Φ ) acts by multiplication. This is an integration by parts.
[ ∂ x k , ∂ x j ∗ ] = ( ∂ x k ∂ x j Φ ) . (98) [\partial_{x_k},\partial_{x_j}^*]=(\partial_{x_k}\partial_{x_j}\Phi). \tag*{(98)} [ ∂ x k , ∂ x j ∗ ] = ( ∂ x k ∂ x j Φ ) . (98)
Covariances as a pairing of gradients. Recall the operator L R = − Δ + ∇ Φ ⋅ ∇ \mathcal{L}_{\mathrm{R}}=-\Delta+\nabla\Phi\cdot\nabla L R = − Δ + ∇Φ ⋅ ∇ of (21) . By (A.1), we have L R = ∑ j ∂ x j ∗ ∂ x j \mathcal{L}_{\mathrm{R}}=\sum_j\partial_{x_j}^{*}\partial_{x_j} L R = ∑ j ∂ x j ∗ ∂ x j , hence ⟨ L R ϕ , G ⟩ = ∑ j ⟨ ∂ x j ϕ , ∂ x j G ⟩ = ⟨ ∇ ϕ , ∇ G ⟩ \langle\mathcal{L}_{\mathrm{R}}\phi,G\rangle=\sum_j\langle\partial_{x_j}\phi,\partial_{x_j}G\rangle=\langle\nabla\phi,\nabla G\rangle ⟨ L R ϕ , G ⟩ = ∑ j ⟨ ∂ x j ϕ , ∂ x j G ⟩ = ⟨ ∇ ϕ , ∇ G ⟩ for all smooth ϕ , G \phi,G ϕ , G . In particular, the kernel of L R \mathcal{L}_{\mathrm{R}} L R consists of the constants and, since L R \mathcal{L}_{\mathrm{R}} L R is self-adjoint, it maps the orthogonal complement of the constants to itself. On this space, L R \mathcal{L}_{\mathrm{R}} L R is invertible. To see this, we use the Brascamp–Lieb inequality recalled at the beginning of Subsection 3.1, which holds since ∇ 2 Φ ≥ a \nabla^2\Phi\geq a ∇ 2 Φ ≥ a . With the identity above, it reads
∥ f ∥ 2 ≤ a − 1 ⟨ L R f , f ⟩ whenever E μ f = 0 , \lVert f\rVert^2\leq a^{-1}\langle\mathcal{L}_{\mathrm{R}}f,f\rangle\qquad\text{whenever }\mathbb{E}_{\mu}f=0, ∥ f ∥ 2 ≤ a − 1 ⟨ L R f , f ⟩ whenever E μ f = 0 ,
that is, L R \mathcal{L}_{\mathrm{R}} L R is bounded below by a > 0 a>0 a > 0 on this space (one says that L R \mathcal{L}_{\mathrm{R}} L R has a spectral gap ).
Now let F F F be a smooth observable with bounded derivatives. Since F − E μ F F-\mathbb{E}_{\mu}F F − E μ F is orthogonal to the constants, there is a solution ϕ \phi ϕ of
L R ϕ = F − E μ F . \mathcal{L}_{\mathrm{R}}\phi=F-\mathbb{E}_{\mu}F. L R ϕ = F − E μ F .
Then, for every smooth observable G G G with bounded derivatives,
Cov μ [ F , G ] = ⟨ F − E μ F , G ⟩ = ⟨ L R ϕ , G ⟩ = ⟨ u , ∇ G ⟩ , u = ∇ ϕ . (99) \operatorname{Cov}_{\mu}[F,G]=\langle F-\mathbb{E}_{\mu}F,G\rangle=\langle\mathcal{L}_{\mathrm{R}}\phi,G\rangle=\langle u,\nabla G\rangle,\qquad u=\nabla\phi. \tag*{(99)} Cov μ [ F , G ] = ⟨ F − E μ F , G ⟩ = ⟨ L R ϕ , G ⟩ = ⟨ u , ∇ G ⟩ , u = ∇ ϕ . (99)
The covariance of F F F and G G G is thus the pairing of the gradient of G G G with a vector field u u u , which describes the response of the system to the observable F F F . To bound covariances, we need to understand u u u .
The Helffer–Sjöstrand equation. The operator A 1 = L R ⊗ I N + ∇ 2 Φ A_1=\mathcal{L}_{\mathrm{R}}\otimes I_N+\nabla^2\Phi A 1 = L R ⊗ I N + ∇ 2 Φ of (21) acts on vector fields by ( A 1 w ) k = L R w k + ( ∇ 2 Φ w ) k (A_1w)_k=\mathcal{L}_{\mathrm{R}}w_k+(\nabla^2\Phi\,w)_k ( A 1 w ) k = L R w k + ( ∇ 2 Φ w ) k . Using (A.2) to move ∂ x k \partial_{x_k} ∂ x k to the right of ∂ x j ∗ \partial_{x_j}^{*} ∂ x j ∗ , we get, for every smooth function ϕ \phi ϕ ,
∂ x k L R ϕ = ∑ j ∂ x j ∗ ∂ x j ( ∂ x k ϕ ) + ∑ j ( ∂ x k ∂ x j Φ ) ∂ x j ϕ , that is, ∇ L R = A 1 ∇ . \partial_{x_k}\mathcal{L}_{\mathrm{R}}\phi=\sum_j\partial_{x_j}^{*}\partial_{x_j}(\partial_{x_k}\phi)+\sum_j(\partial_{x_k}\partial_{x_j}\Phi)\,\partial_{x_j}\phi,\qquad\text{that is,}\qquad\nabla\mathcal{L}_{\mathrm{R}}=A_1\nabla. ∂ x k L R ϕ = j ∑ ∂ x j ∗ ∂ x j ( ∂ x k ϕ ) + j ∑ ( ∂ x k ∂ x j Φ ) ∂ x j ϕ , that is, ∇ L R = A 1 ∇.
We apply this to the solution of L R ϕ = F − E μ F \mathcal{L}_{\mathrm{R}}\phi=F-\mathbb{E}_{\mu}F L R ϕ = F − E μ F , and obtain the Helffer–Sjöstrand equation [52 ]
A 1 u = ∇ F . (100) A_1u=\nabla F. \tag*{(100)} A 1 u = ∇ F . (100)
Formally, u = A 1 − 1 ∇ F u=A_1^{-1}\nabla F u = A 1 − 1 ∇ F , hence (A.3) reads
Cov μ [ F , G ] = ⟨ A 1 − 1 ∇ F , ∇ G ⟩ . \operatorname{Cov}_{\mu}[F,G]=\langle A_1^{-1}\nabla F,\nabla G\rangle. Cov μ [ F , G ] = ⟨ A 1 − 1 ∇ F , ∇ G ⟩ .
This is the analogue of the Gaussian formula (20) , with A 1 A_1 A 1 in place of A A A .
Coercivity, and a bound on u u u . The analogue of the condition A ≥ a A\geq a A ≥ a is that the operator A 1 A_1 A 1 is coercive: by the definition of the adjoint and since ∇ 2 Φ ≥ a \nabla^2\Phi\geq a ∇ 2 Φ ≥ a ,
⟨ A 1 w , w ⟩ = ∑ j , k ∥ ∂ x j w k ∥ 2 + ⟨ w , ∇ 2 Φ w ⟩ ≥ a ∥ w ∥ 2 for every vector field w . \langle A_1w,w\rangle=\sum_{j,k}\lVert\partial_{x_j}w_k\rVert^2+\langle w,\nabla^2\Phi\,w\rangle\geq a\lVert w\rVert^2\qquad\text{for every vector field }w. ⟨ A 1 w , w ⟩ = j , k ∑ ∥ ∂ x j w k ∥ 2 + ⟨ w , ∇ 2 Φ w ⟩ ≥ a ∥ w ∥ 2 for every vector field w .
We apply the coercivity principle (41) to the Helffer–Sjöstrand equation (A.4), and obtain
∥ u ∥ ≤ a − 1 ∥ ∇ F ∥ . \lVert u\rVert\leq a^{-1}\lVert\nabla F\rVert. ∥ u ∥ ≤ a − 1 ∥ ∇ F ∥ .
With (A.3), we get ∣ Cov μ [ F , G ] ∣ ≤ a − 1 ∥ ∇ F ∥ ∥ ∇ G ∥ |\operatorname{Cov}_{\mu}[F,G]|\leq a^{-1}\lVert\nabla F\rVert\lVert\nabla G\rVert ∣ Cov μ [ F , G ] ∣ ≤ a − 1 ∥ ∇ F ∥ ∥ ∇ G ∥ , the covariance form of the Brascamp–Lieb inequality. This bound does not depend on where F F F and G G G live. With the argument of Combes and Thomas, we improve it into a bound that decays with the distance between them.
The general case: a Hessian which depends on the configuration. In the Gaussian case of Subsection 3.2, the Hessian is the constant matrix A A A . For a general Φ \Phi Φ , the Hessian ∇ 2 Φ ( x ) \nabla^2\Phi(x) ∇ 2 Φ ( x ) depends on x x x , and there is no single matrix to invert. The operator A 1 = L R ⊗ I N + ∇ 2 Φ A_1=\mathcal{L}_{\mathrm{R}}\otimes I_N+\nabla^2\Phi A 1 = L R ⊗ I N + ∇ 2 Φ plays the role of A A A , in the sense that Cov μ [ F , G ] = ⟨ A 1 − 1 ∇ F , ∇ G ⟩ \operatorname{Cov}_{\mu}[F,G]=\langle A_1^{-1}\nabla F,\nabla G\rangle Cov μ [ F , G ] = ⟨ A 1 − 1 ∇ F , ∇ G ⟩ . The term L R ⊗ I N \mathcal{L}_{\mathrm{R}}\otimes I_N L R ⊗ I N accounts for the dependence of the Hessian on the configuration. It vanishes on constant vector fields, which is why it is invisible in the Gaussian case with linear observables.
Brascamp–Lieb again. Since L R ⊗ I N ≥ 0 \mathcal{L}_{\mathbb{R}} \otimes I_N \ge0 L R ⊗ I N ≥ 0 , we have A 1 ≥ ∇ 2 Φ A_1 \ge\nabla^2\Phi A 1 ≥ ∇ 2 Φ as operators on L 2 ( μ ; R N ) L^2(\mu;\mathbb{R}^N) L 2 ( μ ; R N ) , where ∇ 2 Φ \nabla^2\Phi ∇ 2 Φ acts by multiplication. Hence A 1 − 1 ≤ ( ∇ 2 Φ ) − 1 A_1^{-1} \le(\nabla^2\Phi)^{-1} A 1 − 1 ≤ ( ∇ 2 Φ ) − 1 , and taking G = F G=F G = F gives
Var μ [ F ] = ⟨ A 1 − 1 ∇ F , ∇ F ⟩ ≤ E μ ⟨ ∇ F , ( ∇ 2 Φ ) − 1 ∇ F ⟩ . \operatorname{Var}_{\mu}[F]=\langle A_1^{-1}\nabla F,\nabla F\rangle\le\mathbb{E}_{\mu}\langle\nabla F,(\nabla^2\Phi)^{-1}\nabla F\rangle. Var μ [ F ] = ⟨ A 1 − 1 ∇ F , ∇ F ⟩ ≤ E μ ⟨ ∇ F , ( ∇ 2 Φ ) − 1 ∇ F ⟩ .
This is the original form of the Brascamp–Lieb inequality [17 ] , which implies (18) when ∇ 2 Φ ≥ a \nabla^2\Phi\ge a ∇ 2 Φ ≥ a .
The complex case. For the gas, ∇ \nabla ∇ is replaced by ∂ ˉ \bar{\partial} ∂ ˉ and the Hessian by the Levi matrix (Subsection 5.2). The Hessian of the logarithmic interaction couples all pairs of particles, with entries of order ∣ z i − z j ∣ − 2 |z_i-z_j|^{-2} ∣ z i − z j ∣ − 2 . The Levi matrix, on the other hand, is local: since g g g is harmonic away from the origin, its off-diagonal entries are supported on the collisions (near the collisions, after smoothing). For this reason, the argument of Combes and Thomas works despite the long range of the interaction.
Regularization and the Helffer–Sjöstrand equation
We justify the regularization of Subsection 5.1 and prove Proposition 5.2.
Regularization. The weight {(5.1)} is of class C 3 C^3 C 3 , invariant under permutations and simultaneous rotations. Since ψ ε \psi_\varepsilon ψ ε is radial and supported in B ε B_\varepsilon B ε , we get from the mean value and super-mean value properties of g g g that
g ε ( x ) = g ( x ) for ∣ x ∣ ≥ ε , g ε ≤ g , Δ g ε = − 2 π ψ ε . g_\varepsilon(x)=g(x)\quad\text{for }|x|\ge\varepsilon,\qquad g_\varepsilon\le g,\qquad\Delta g_\varepsilon=-2\pi\psi_\varepsilon. g ε ( x ) = g ( x ) for ∣ x ∣ ≥ ε , g ε ≤ g , Δ g ε = − 2 π ψ ε .
Since β ≤ 2 \beta\le2 β ≤ 2 , since g ε ≤ g g_\varepsilon\le g g ε ≤ g and since the wall term is non-negative, we have
e − Φ ε s h ( Z ) ≤ e − a ∣ Z ∣ 2 ∏ i < j ∣ z i − z j ∣ β − 2 . e^{-\Phi_\varepsilon^{\mathrm{sh}}(Z)}\le e^{-a|Z|^2}\prod_{i<j}|z_i-z_j|^{\beta-2}. e − Φ ε sh ( Z ) ≤ e − a ∣ Z ∣ 2 i < j ∏ ∣ z i − z j ∣ β − 2 .
After multiplication by ∣ V N ∣ 2 |V_N|^2 ∣ V N ∣ 2 , the right-hand side is integrable, so K ε K_\varepsilon K ε is finite. At fixed λ \lambda λ , we let ε → 0 \varepsilon\to0 ε → 0 to remove the smoothing of the interaction. We then let λ → ∞ \lambda\to\infty λ → ∞ to obtain the hard wall. At both steps, by dominated convergence, the partition functions converge, and so do the normalized densities in L 1 L^1 L 1 . Hence, for bounded functions F F F and G G G of N N N positions with bounded gradients,
lim λ → ∞ lim ε → 0 Cov G ε [ F , G ] = Cov G [ F , G ] , lim λ → ∞ lim ε → 0 D G ε ( F ) = D G ( F ) . (101) \lim_{\lambda\to\infty}\lim_{\varepsilon\to0}\operatorname{Cov}_{G_\varepsilon}[F,G]=\operatorname{Cov}_{G}[F,G],\qquad
\lim_{\lambda\to\infty}\lim_{\varepsilon\to0}\mathcal{D}_{G_\varepsilon}(F)=\mathcal{D}_{G}(F). \tag*{(101)} λ → ∞ lim ε → 0 lim Cov G ε [ F , G ] = Cov G [ F , G ] , λ → ∞ lim ε → 0 lim D G ε ( F ) = D G ( F ) . (101)
The Levi matrix of the regularized weight is a function. We apply (B.1) to h = ( β − 2 ) g ε h=(\beta-2)g_\varepsilon h = ( β − 2 ) g ε , for which 1 4 Δ h = π 2 ( 2 − β ) ψ ε \frac{1}{4}\Delta h=\frac{\pi}{2}(2-\beta)\psi_\varepsilon 4 1 Δ h = 2 π ( 2 − β ) ψ ε , and (B.2) to t ↦ λ ( t − L 2 ) + 4 t\mapsto\lambda(t-L^2)_+^4 t ↦ λ ( t − L 2 ) + 4 . We obtain that the matrix C ε Φ s h ( Z ) C_{\varepsilon}^{\Phi^{\mathrm{sh}}}(Z) C ε Φ sh ( Z ) acts on U = ( U k ) 1 ≤ k ≤ N ∈ ( R 2 ) N U=(U_k)_{1\le k\le N}\in(\mathbb{R}^2)^N U = ( U k ) 1 ≤ k ≤ N ∈ ( R 2 ) N by
( C ε Φ s h U ) k = ( a + 4 λ ( ∣ z k ∣ 2 − L 2 ) + 2 ( 4 ∣ z k ∣ 2 − L 2 ) ) U k + ∑ j ≠ k ω k j , ε ( U k − U j ) , (102) \left(C_{\varepsilon}^{\Phi^{\mathrm{sh}}}U\right)_k=
\left(a+4\lambda\left(|z_k|^2-L^2\right)_+^2\left(4|z_k|^2-L^2\right)\right)U_k
+\sum_{j\ne k}\omega_{kj,\varepsilon}(U_k-U_j), \tag*{(102)} ( C ε Φ sh U ) k = ( a + 4 λ ( ∣ z k ∣ 2 − L 2 ) + 2 ( 4∣ z k ∣ 2 − L 2 ) ) U k + j = k ∑ ω k j , ε ( U k − U j ) , (102)
where ω k j , ε \omega_{kj,\varepsilon} ω k j , ε are the collision coefficients of Subsection 5.1. They are non-negative since β ≤ 2 \beta\le2 β ≤ 2 , and so is the coefficient of the wall. Hence U ⋅ C ε Φ s h U ≥ a ∣ U ∣ 2 U\cdot C_{\varepsilon}^{\Phi^{\mathrm{sh}}}U\ge a|U|^2 U ⋅ C ε Φ sh U ≥ a ∣ U ∣ 2 , which is the regularized form of Lemma 4.4 .
The operators L \mathcal{L} L and A 1 A_1 A 1 . We work in the space L s h , ε 2 L^2_{\mathrm{sh},\varepsilon} L sh , ε 2 of Subsection 5.2, with the Hermitian product ⟨ f , g ⟩ = ∫ f g ˉ e − Φ ε s h d Z \langle f,g\rangle=\int f\bar{g}e^{-\Phi_\varepsilon^{\mathrm{sh}}}\,dZ ⟨ f , g ⟩ = ∫ f g ˉ e − Φ ε sh d Z and the norm ∥ ⋅ ∥ \|\cdot\| ∥ ⋅ ∥ . For families U = ( U k ) 1 ≤ k ≤ N U=(U_k)_{1\le k\le N} U = ( U k ) 1 ≤ k ≤ N and U ~ = ( U ~ k ) 1 ≤ k ≤ N \widetilde{U}=(\widetilde{U}_k)_{1\le k\le N} U = ( U k ) 1 ≤ k ≤ N in this space, we let ⟨ U , U ~ ⟩ : = ∑ k ⟨ U k , U ~ k ⟩ \langle U,\widetilde{U}\rangle:=\sum_k\langle U_k,\widetilde{U}_k\rangle ⟨ U , U ⟩ := ∑ k ⟨ U k , U k ⟩ .
The adjoint of ∂ ˉ j \bar{\partial}_j ∂ ˉ j in L s h , ε 2 L^2_{\mathrm{sh},\varepsilon} L sh , ε 2 acts on smooth functions by ∂ ˉ j ∗ f = − ∂ j f + ( ∂ j Φ ε s h ) f \bar{\partial}_j^*f=-\partial_jf+(\partial_j\Phi_\varepsilon^{\mathrm{sh}})f ∂ ˉ j ∗ f = − ∂ j f + ( ∂ j Φ ε sh ) f . We let ∂ ˉ ∗ U : = ∑ j ∂ ˉ j ∗ U j \bar{\partial}^*U:=\sum_j\bar{\partial}_j^*U_j ∂ ˉ ∗ U := ∑ j ∂ ˉ j ∗ U j for a family U U U . The operator L \mathcal{L} L of Subsection 5.2 is then
L = ∂ ˉ ∗ ∂ ˉ = ∑ j = 1 N ∂ ˉ j ∗ ∂ ˉ j = 1 4 ( ∇ Φ ε s h ⋅ ∇ − Δ ) + i 4 ( J N ∇ Φ ε s h ) ⋅ ∇ . \mathcal{L}=\bar{\partial}^*\bar{\partial}=\sum_{j=1}^{N}\bar{\partial}_j^*\bar{\partial}_j=\frac{1}{4}\left(\nabla\Phi_\varepsilon^{\mathrm{sh}}\cdot\nabla-\Delta\right)+\frac{i}{4}\left(J_N\nabla\Phi_\varepsilon^{\mathrm{sh}}\right)\cdot\nabla. L = ∂ ˉ ∗ ∂ ˉ = j = 1 ∑ N ∂ ˉ j ∗ ∂ ˉ j = 4 1 ( ∇ Φ ε sh ⋅ ∇ − Δ ) + 4 i ( J N ∇ Φ ε sh ) ⋅ ∇.
As an operator on L s h , ε 2 L^2_{\mathrm{sh},\varepsilon} L sh , ε 2 , L \mathcal{L} L is the self-adjoint operator associated with the closed quadratic form f ↦ ∥ ∂ ˉ f ∥ 2 f\mapsto\|\bar{\partial}f\|^2 f ↦ ∥ ∂ ˉ f ∥ 2 on the domain of ∂ ˉ \bar{\partial} ∂ ˉ , and its kernel is ker ∂ ˉ \ker\bar{\partial} ker ∂ ˉ . On families, the operator A 1 = L ⊗ I N + C ε Φ s h A_1=\mathcal{L}\otimes I_N+C_{\varepsilon}^{\Phi^{\mathrm{sh}}} A 1 = L ⊗ I N + C ε Φ sh of Subsection 5.2 acts by
( A 1 U ) k = L U k + ( C ε Φ s h U ) k for 1 ≤ k ≤ N . (A_1U)_k=\mathcal{L}U_k+\left(C_{\varepsilon}^{\Phi^{\mathrm{sh}}}U\right)_k\qquad\text{for }1\le k\le N. ( A 1 U ) k = L U k + ( C ε Φ sh U ) k for 1 ≤ k ≤ N .
For a real weight Φ \Phi Φ , Helffer and Sjöstrand [52 ] use the operator − Δ + ∇ Φ ⋅ ∇ -\Delta+\nabla\Phi\cdot\nabla − Δ + ∇Φ ⋅ ∇ on functions and the operator ( − Δ + ∇ Φ ⋅ ∇ ) ⊗ I + ∇ 2 Φ \left(-\Delta+\nabla\Phi\cdot\nabla\right)\otimes I+\nabla^{2}\Phi ( − Δ + ∇Φ ⋅ ∇ ) ⊗ I + ∇ 2 Φ on vector fields, which maps ∇ f \nabla f ∇ f to ∇ ( − Δ f + ∇ Φ ⋅ ∇ f ) \nabla\left(-\Delta f+\nabla\Phi\cdot\nabla f\right) ∇ ( − Δ f + ∇Φ ⋅ ∇ f ) . In L \mathcal{L} L and A 1 A_{1} A 1 , we replace the gradient by ∂ ˉ \bar{\partial} ∂ ˉ and the Hessian by the Levi matrix. The next lemma shows that the same identity holds.
Lemma A.2 (Intertwining and coercivity). Let f f f be a smooth function on ( R 2 ) N (\mathbb{R}^{2})^{N} ( R 2 ) N , and let U U U be a family of N N N smooth functions with compact support. Then
A 1 ∂ ˉ f = ∂ ˉ L f and ∂ ˉ ∗ A 1 U = L ∂ ˉ ∗ U . A_{1}\bar{\partial}f=\bar{\partial}\mathcal{L}f
\qquad\text{and}\qquad
\bar{\partial}^{*}A_{1}U=\mathcal{L}\bar{\partial}^{*}U. A 1 ∂ ˉ f = ∂ ˉ L f and ∂ ˉ ∗ A 1 U = L ∂ ˉ ∗ U . Moreover, in L s h , ε 2 L^{2}_{\mathrm{sh},\varepsilon} L sh , ε 2 ,
⟨ A 1 U , U ⟩ = ∑ j , k = 1 N ∥ ∂ ˉ j U k ∥ 2 + ∫ U ⋅ C Φ ε s h U e − Φ ε s h d Z ≥ a ∥ U ∥ 2 . \langle A_{1}U,U\rangle=\sum_{j,k=1}^{N}\|\bar{\partial}_{j}U_{k}\|^{2}+\int U\cdot C_{\Phi_{\varepsilon}^{\mathrm{sh}}}U\,e^{-\Phi_{\varepsilon}^{\mathrm{sh}}}\,dZ\geq a\|U\|^{2}. ⟨ A 1 U , U ⟩ = j , k = 1 ∑ N ∥ ∂ ˉ j U k ∥ 2 + ∫ U ⋅ C Φ ε sh U e − Φ ε sh d Z ≥ a ∥ U ∥ 2 .
The anti-analytic signal. Since G ε G_{\varepsilon} G ε is invariant under simultaneous rotations, we decompose functions into their modes of character m m m , given by (30) , and we define the anti-analytic signal under G ε G_{\varepsilon} G ε as in Subsection 4.8, with G ε G_{\varepsilon} G ε in place of G G G .
Lemma A.3 (Anti-analytic signal). Let 0 < ε ≤ 1 0<\varepsilon\leq1 0 < ε ≤ 1 , let F F F be a real symmetric function of N N N positions, smooth and bounded with bounded derivatives on ( R 2 ) N (\mathbb{R}^{2})^{N} ( R 2 ) N , and let H H H be its anti-analytic signal under G ε G_{\varepsilon} G ε . Then Re H = F − E G ε F \operatorname{Re}H=F-\mathbb{E}_{G_{\varepsilon}}F Re H = F − E G ε F ,
Π ker ∂ ˉ ( V N H ) = 0 , (103) \Pi_{\ker\bar{\partial}}(V_{N}H)=0, \tag*{(103)} Π k e r ∂ ˉ ( V N H ) = 0 , (103) and
E G ε ∑ j = 1 N ∣ ∂ ˉ j H ∣ 2 ≤ D G ε ( F ) . (104) \mathbb{E}_{G_{\varepsilon}}\sum_{j=1}^{N}|\bar{\partial}_{j}H|^{2}\leq D_{G_{\varepsilon}}(F). \tag*{(104)} E G ε j = 1 ∑ N ∣ ∂ ˉ j H ∣ 2 ≤ D G ε ( F ) . (104)
Proof. Since F F F is real, we have F − m = F m ‾ F_{-m}=\overline{F_{m}} F − m = F m , and therefore Re H = F − E G ε F \operatorname{Re}H=F-\mathbb{E}_{G_{\varepsilon}}F Re H = F − E G ε F .
The projection. The function V N H V_{N}H V N H belongs to L s h , ε 2 L^{2}_{\mathrm{sh},\varepsilon} L sh , ε 2 , since ∥ V N H ∥ 2 = K ε E G ε ∣ H ∣ 2 \|V_{N}H\|^{2}=K_{\varepsilon}\mathbb{E}_{G_{\varepsilon}}|H|^{2} ∥ V N H ∥ 2 = K ε E G ε ∣ H ∣ 2 . Since e − Φ ε s h e^{-\Phi_{\varepsilon}^{\mathrm{sh}}} e − Φ ε sh is continuous and positive, we see as in the proof of Lemma 4.1 that the elements of ker ∂ ˉ \ker\bar{\partial} ker ∂ ˉ are entire functions on C N \mathbb{C}^{N} C N . By Lemma 4.11, applied on each polydisk centred at the origin, an antisymmetric entire function of character m m m vanishes if m < d N m<d_{N} m < d N , and is a complex multiple of V N V_{N} V N if m = d N m=d_{N} m = d N . The weight Φ ε s h \Phi_{\varepsilon}^{\mathrm{sh}} Φ ε sh is invariant under permutations and simultaneous rotations. Therefore Lemmas 4.12 and 4.13 hold in L s h , ε 2 L^{2}_{\mathrm{sh},\varepsilon} L sh , ε 2 , with the same proofs, with C N \mathbb{C}^{N} C N in place of B L N B_{L}^{N} B L N and with μ = G ε \mu=G_{\varepsilon} μ = G ε . For m < 0 m<0 m < 0 , the function V N F m V_{N}F_{m} V N F m is antisymmetric and of character d N + m < d N d_{N}+m<d_{N} d N + m < d N , and so is its projection, which therefore vanishes. The projection of V N ( F 0 − E G ε F ) V_{N}(F_{0}-\mathbb{E}_{G_{\varepsilon}}F) V N ( F 0 − E G ε F ) vanishes by Lemma 4.13. This gives (A.7).
The energy. Since ∂ ˉ j [ F ( e i θ ⋅ ) ] = e − i θ ( ∂ ˉ j F ) ( e i θ ⋅ ) \bar{\partial}_{j}[F(e^{i\theta}\cdot)]=e^{-i\theta}(\bar{\partial}_{j}F)(e^{i\theta}\cdot) ∂ ˉ j [ F ( e i θ ⋅ )] = e − i θ ( ∂ ˉ j F ) ( e i θ ⋅ ) , the operator ∂ ˉ j \bar{\partial}_{j} ∂ ˉ j maps the mode of character m m m of F F F to the mode of character m + 1 m+1 m + 1 of ∂ ˉ j F \bar{\partial}_{j}F ∂ ˉ j F . Hence ∂ ˉ j H = ( ∂ ˉ j F ) 1 + 2 ∑ m ≤ 0 ( ∂ ˉ j F ) m \bar{\partial}_{j}H=(\bar{\partial}_{j}F)_{1}+2\sum_{m\leq0}(\bar{\partial}_{j}F)_{m} ∂ ˉ j H = ( ∂ ˉ j F ) 1 + 2 ∑ m ≤ 0 ( ∂ ˉ j F ) m . The modes are orthogonal in L 2 ( G ε ) L^{2}(G_{\varepsilon}) L 2 ( G ε ) by Lemma 4.10 , hence
E G ε ∣ ∂ ˉ j H ∣ 2 = E G ε ∣ ( ∂ ˉ j F ) 1 ∣ 2 + 4 ∑ m ≤ 0 E G ε ∣ ( ∂ ˉ j F ) m ∣ 2 ≤ 4 E G ε ∣ ∂ ˉ j F ∣ 2 = E G ε ∣ ∇ j F ∣ 2 , \mathbb{E}_{G_{\varepsilon}}|\bar{\partial}_{j}H|^{2}
=\mathbb{E}_{G_{\varepsilon}}|(\bar{\partial}_{j}F)_{1}|^{2}
+4\sum_{m\leq0}\mathbb{E}_{G_{\varepsilon}}|(\bar{\partial}_{j}F)_{m}|^{2}
\leq4\mathbb{E}_{G_{\varepsilon}}|\bar{\partial}_{j}F|^{2}
=\mathbb{E}_{G_{\varepsilon}}|\nabla_{j}F|^{2}, E G ε ∣ ∂ ˉ j H ∣ 2 = E G ε ∣ ( ∂ ˉ j F ) 1 ∣ 2 + 4 m ≤ 0 ∑ E G ε ∣ ( ∂ ˉ j F ) m ∣ 2 ≤ 4 E G ε ∣ ∂ ˉ j F ∣ 2 = E G ε ∣ ∇ j F ∣ 2 ,
by (25) . Summing over j j j gives (A.8). □
The representation. We prove Proposition 5.2. The weak formulation of A 1 ∂ ˉ ϕ ˉ = V N ∂ ˉ H A_{1}\bar{\partial}\bar{\phi}=V_{N}\bar{\partial}H A 1 ∂ ˉ ϕ ˉ = V N ∂ ˉ H is as follows: the functions ∂ ˉ j ∂ ˉ k ϕ \bar{\partial}_{j}\bar{\partial}_{k}\phi ∂ ˉ j ∂ ˉ k ϕ belong to L s h , ε 2 L^{2}_{\mathrm{sh},\varepsilon} L sh , ε 2 , the number ⟨ C Φ ε s h ∂ ˉ ϕ , ∂ ˉ ϕ ⟩ \langle C_{\Phi_{\varepsilon}^{\mathrm{sh}}}\bar{\partial}\phi,\bar{\partial}\phi\rangle ⟨ C Φ ε sh ∂ ˉ ϕ , ∂ ˉ ϕ ⟩ is finite, and
∑ j , k = 1 N ⟨ ∂ ˉ j ∂ ˉ k ϕ , ∂ ˉ j U k ⟩ + ⟨ C Φ ε s h ∂ ˉ ϕ , U ⟩ = ⟨ V N ∂ ˉ H , U ⟩ (105) \sum_{j,k=1}^{N}\langle\bar{\partial}_{j}\bar{\partial}_{k}\phi,\bar{\partial}_{j}U_{k}\rangle+\langle C_{\Phi_{\varepsilon}^{\mathrm{sh}}}\bar{\partial}\phi,U\rangle=\langle V_{N}\bar{\partial}H,U\rangle \tag*{(105)} j , k = 1 ∑ N ⟨ ∂ ˉ j ∂ ˉ k ϕ , ∂ ˉ j U k ⟩ + ⟨ C Φ ε sh ∂ ˉ ϕ , U ⟩ = ⟨ V N ∂ ˉ H , U ⟩ (105)
for every family U U U in L s h , ε 2 L^{2}_{\mathrm{sh},\varepsilon} L sh , ε 2 such that the functions ∂ ˉ j U k \bar{\partial}_{j}U_{k} ∂ ˉ j U k belong to L s h , ε 2 L^{2}_{\mathrm{sh},\varepsilon} L sh , ε 2 and ⟨ C Φ ε s h U , U ⟩ \langle C_{\Phi_{\varepsilon}^{\mathrm{sh}}}U,U\rangle ⟨ C Φ ε sh U , U ⟩ is finite.
Proof of Lemma A.2. For 1 ≤ j , k ≤ N 1 \le j,k \le N 1 ≤ j , k ≤ N , the commutator of ∂ ˉ k \bar{\partial}_k ∂ ˉ k and ∂ ˉ j ∗ \bar{\partial}_j^* ∂ ˉ j ∗ is the multiplication by ∂ ˉ k ∂ j Φ ε s h \bar{\partial}_k\partial_j\Phi^{\mathrm{sh}}_\varepsilon ∂ ˉ k ∂ j Φ ε sh , since
∂ ˉ k ∂ ˉ j ∗ f − ∂ ˉ j ∗ ∂ ˉ k f = ∂ ˉ k ( ( ∂ j Φ ε s h ) f ) − ( ∂ j Φ ε s h ) ∂ ˉ k f = ( ∂ ˉ k ∂ j Φ ε s h ) f . \bar{\partial}_k\bar{\partial}_j^*f-\bar{\partial}_j^*\bar{\partial}_kf=\bar{\partial}_k\bigl((\partial_j\Phi^{\mathrm{sh}}_\varepsilon)f\bigr)-(\partial_j\Phi^{\mathrm{sh}}_\varepsilon)\bar{\partial}_kf=(\bar{\partial}_k\partial_j\Phi^{\mathrm{sh}}_\varepsilon)f. ∂ ˉ k ∂ ˉ j ∗ f − ∂ ˉ j ∗ ∂ ˉ k f = ∂ ˉ k ( ( ∂ j Φ ε sh ) f ) − ( ∂ j Φ ε sh ) ∂ ˉ k f = ( ∂ ˉ k ∂ j Φ ε sh ) f .
By Remark 4.3, these numbers are the entries of the complex form of C Φ ε s h C_{\Phi^{\mathrm{sh}}_\varepsilon} C Φ ε sh . Hence
∂ ˉ k L f = ∑ j = 1 N ( ∂ ˉ j ∗ ∂ ˉ j ∂ ˉ k f + ( ∂ ˉ k ∂ j Φ ε s h ) ∂ ˉ j f ) = L ∂ ˉ k f + ( C Φ ε s h ∂ ˉ f ) k . \bar{\partial}_k\mathcal{L}f=\sum_{j=1}^{N}\left(\bar{\partial}_j^*\bar{\partial}_j\bar{\partial}_kf+(\bar{\partial}_k\partial_j\Phi^{\mathrm{sh}}_\varepsilon)\bar{\partial}_jf\right)=\mathcal{L}\bar{\partial}_kf+\left(C_{\Phi^{\mathrm{sh}}_\varepsilon}\bar{\partial}f\right)_k. ∂ ˉ k L f = j = 1 ∑ N ( ∂ ˉ j ∗ ∂ ˉ j ∂ ˉ k f + ( ∂ ˉ k ∂ j Φ ε sh ) ∂ ˉ j f ) = L ∂ ˉ k f + ( C Φ ε sh ∂ ˉ f ) k .
which is the first identity. Since the matrix C Φ ε s h ( Z ) C_{\Phi^{\mathrm{sh}}_\varepsilon}(Z) C Φ ε sh ( Z ) is symmetric and commutes with J N J_N J N , the operators L \mathcal{L} L and A 1 A_1 A 1 are symmetric on smooth functions and families with compact support. The second identity follows from the first by duality. Finally, ⟨ U k , U k ⟩ = ∑ j ∥ ∂ ˉ j U k ∥ 2 \langle U_k,U_k\rangle=\sum_j\|\bar{\partial}_jU_k\|^2 ⟨ U k , U k ⟩ = ∑ j ∥ ∂ ˉ j U k ∥ 2 for each k k k , and ∑ k ( C Φ ε s h U ) k U k ‾ = U ⋅ C Φ ε s h U ‾ \sum_k(C_{\Phi^{\mathrm{sh}}_\varepsilon}U)_k\overline{U_k}=U\cdot C_{\Phi^{\mathrm{sh}}_\varepsilon}\overline{U} ∑ k ( C Φ ε sh U ) k U k = U ⋅ C Φ ε sh U , since C Φ ε s h C_{\Phi^{\mathrm{sh}}_\varepsilon} C Φ ε sh commutes with J N J_N J N . This gives the identity, and the lower bound follows from (A.6). □
Proof of Proposition 5.2.
Step 1. The solution ϕ \phi ϕ and the covariance. The weight Φ ε s h \Phi^{\mathrm{sh}}_\varepsilon Φ ε sh is plurisubharmonic, with Levi matrix at least a a a by (A.6). By the L 2 L^2 L 2 estimate of Hörmander on the pseudoconvex domain C N \mathbb{C}^N C N [Hör65], [Hör90, Chapter IV], the fake Poincaré inequality of Proposition 4.5 therefore holds in L s h , ε 2 L^2_{\mathrm{sh},\varepsilon} L sh , ε 2 :
∥ h − Π ker ∂ ˉ h ∥ 2 ≤ a − 1 ∥ ∂ ˉ h ∥ 2 = a − 1 ⟨ L h , h ⟩ . \|h-\Pi_{\ker\bar{\partial}}h\|^2\le a^{-1}\|\bar{\partial}h\|^2=a^{-1}\langle\mathcal{L}h,h\rangle. ∥ h − Π k e r ∂ ˉ h ∥ 2 ≤ a − 1 ∥ ∂ ˉ h ∥ 2 = a − 1 ⟨ L h , h ⟩ .
for every h h h in the domain of L \mathcal{L} L . Hence L ≥ a \mathcal{L}\ge a L ≥ a on the orthogonal complement of ker ∂ ˉ = ker L \ker\bar{\partial}=\ker\mathcal{L} ker ∂ ˉ = ker L , and L \mathcal{L} L is invertible on this complement, with an inverse of norm at most a − 1 a^{-1} a − 1 . By (A.7), V N H V_NH V N H belongs to this complement, and there exists ϕ \phi ϕ in the domain of L \mathcal{L} L such that L ϕ = V N H \mathcal{L}\phi=V_NH L ϕ = V N H .
The measure ∣ V N ∣ 2 e − Φ ε s h d Z |V_N|^2e^{-\Phi^{\mathrm{sh}}_\varepsilon}\,dZ ∣ V N ∣ 2 e − Φ ε sh d Z is K ε d G ε K_\varepsilon\,dG_\varepsilon K ε d G ε , the function G G G is real, and Re H = F − E G ε F \operatorname{Re}H=F-\mathbb{E}_{G_\varepsilon}F Re H = F − E G ε F by Lemma A.3 . Moreover, V N ( G − E G ε G ) V_N(G-\mathbb{E}_{G_\varepsilon}G) V N ( G − E G ε G ) belongs to the domain of ∂ ˉ \bar{\partial} ∂ ˉ , and ∂ ˉ \bar{\partial} ∂ ˉ commutes with the multiplication by V N V_N V N by Lemma 4.2 . Hence
K ε Cov G ε [ F , G ] = Re ⟨ V N H , V N ( G − E G ε G ) ⟩ = Re ⟨ L ϕ , V N ( G − E G ε G ) ⟩ = Re ⟨ ∂ ˉ ϕ , V N ∂ ˉ G ⟩ , K_\varepsilon\operatorname{Cov}_{G_\varepsilon}[F,G]=\operatorname{Re}\langle V_NH,V_N(G-\mathbb{E}_{G_\varepsilon}G)\rangle=\operatorname{Re}\langle\mathcal{L}\phi,V_N(G-\mathbb{E}_{G_\varepsilon}G)\rangle=\operatorname{Re}\langle\bar{\partial}\phi,V_N\bar{\partial}G\rangle, K ε Cov G ε [ F , G ] = Re ⟨ V N H , V N ( G − E G ε G )⟩ = Re ⟨ L ϕ , V N ( G − E G ε G )⟩ = Re ⟨ ∂ ˉ ϕ , V N ∂ ˉ G ⟩ ,
which is (43) .
Step 2. The Helffer–Sjöstrand equation. Let U U U be a family of smooth functions with compact support. Since L \mathcal{L} L is self-adjoint and ∂ ˉ ∗ U \bar{\partial}^*U ∂ ˉ ∗ U is a function of class C 2 C^2 C 2 with compact support, we get by Lemma A.2
⟨ ∂ ˉ ϕ , A 1 U ⟩ = ⟨ ϕ , L ∂ ˉ ∗ U ⟩ = ⟨ V N H , ∂ ˉ ∗ U ⟩ = ⟨ V N ∂ ˉ H , U ⟩ . \langle\bar{\partial}\phi,A_1U\rangle=\langle\phi,\mathcal{L}\bar{\partial}^*U\rangle=\langle V_NH,\bar{\partial}^*U\rangle=\langle V_N\bar{\partial}H,U\rangle. ⟨ ∂ ˉ ϕ , A 1 U ⟩ = ⟨ ϕ , L ∂ ˉ ∗ U ⟩ = ⟨ V N H , ∂ ˉ ∗ U ⟩ = ⟨ V N ∂ ˉ H , U ⟩ .
Let u = ∂ ˉ ϕ u=\bar{\partial}\phi u = ∂ ˉ ϕ be as in the statement, and let
q ε ( U ) : = ∑ j , k ∥ ∂ ˉ j U k ∥ 2 + ⟨ C Φ ε s h U , U ⟩ . q_\varepsilon(U):=\sum_{j,k}\|\bar{\partial}_jU_k\|^2+\langle C_{\Phi^{\mathrm{sh}}_\varepsilon}U,U\rangle. q ε ( U ) := j , k ∑ ∥ ∂ ˉ j U k ∥ 2 + ⟨ C Φ ε sh U , U ⟩ .
By local elliptic regularity for L ϕ = V N H \mathcal{L}\phi=V_NH L ϕ = V N H , we have u ∈ H l o c 1 u\in H^1_{\mathrm{loc}} u ∈ H loc 1 . Let χ ~ R ( Z ) : = χ R ( ∣ Z ∣ ) \widetilde{\chi}_R(Z):=\chi_R(|Z|) χ R ( Z ) := χ R ( ∣ Z ∣ ) for Z ∈ C N Z\in\mathbb{C}^N Z ∈ C N , with χ R \chi_R χ R as in § 3 . It equals 1 1 1 on B R ⊂ C N B_R\subset\mathbb{C}^N B R ⊂ C N , it is supported in B 2 R B_{2R} B 2 R , and ∣ ∇ χ ~ R ∣ ≤ C / R |\nabla\widetilde{\chi}_R|\le C/R ∣∇ χ R ∣ ≤ C / R . We test the distributional equation A 1 u = V N ∂ ˉ H A_1u=V_N\bar{\partial}H A 1 u = V N ∂ ˉ H with χ ~ R 2 u \widetilde{\chi}_R^2u χ R 2 u , by local approximation, and get
q ε ( χ ~ R u ) = Re ⟨ V N ∂ ˉ H , χ ~ R 2 u ⟩ + ∑ j , k ∥ ( ∂ ˉ j χ ~ R ) u k ∥ 2 ≤ ∥ V N ∂ ˉ H ∥ ∥ u ∥ + C R 2 ∥ u ∥ 2 . q_\varepsilon(\widetilde{\chi}_Ru)=\operatorname{Re}\langle V_N\bar{\partial}H,\widetilde{\chi}_R^2u\rangle+\sum_{j,k}\|(\bar{\partial}_j\widetilde{\chi}_R)u_k\|^2\le\|V_N\bar{\partial}H\|\,\|u\|+\frac{C}{R^2}\|u\|^2. q ε ( χ R u ) = Re ⟨ V N ∂ ˉ H , χ R 2 u ⟩ + j , k ∑ ∥ ( ∂ ˉ j χ R ) u k ∥ 2 ≤ ∥ V N ∂ ˉ H ∥ ∥ u ∥ + R 2 C ∥ u ∥ 2 .
Since χ ~ R u → R → ∞ u \widetilde{\chi}_Ru\xrightarrow[R\to\infty]{}u χ R u R → ∞ u in L s h , ε 2 L^2_{\mathrm{sh},\varepsilon} L sh , ε 2 , we get q ε ( u ) < ∞ q_\varepsilon(u)<\infty q ε ( u ) < ∞ by lower semicontinuity of the closed form. By cutoff and local convolution, we approximate every element of its domain by smooth families with compact support. Therefore the weak equation extends to all the test families in (A.9). In particular,
q ε ( u ) = ⟨ V N ∂ ˉ H , u ⟩ = ⟨ ∂ ˉ ( V N H ) , ∂ ˉ ϕ ⟩ = ∥ V N H ∥ 2 . q_\varepsilon(u)=\langle V_N\bar{\partial}H,u\rangle=\langle\bar{\partial}(V_NH),\bar{\partial}\phi\rangle=\|V_NH\|^2. q ε ( u ) = ⟨ V N ∂ ˉ H , u ⟩ = ⟨ ∂ ˉ ( V N H ) , ∂ ˉ ϕ ⟩ = ∥ V N H ∥ 2 .
The bound a ∥ u ∥ ≤ ∥ V N ∂ ˉ H ∥ a\|u\|\le\|V_N\bar{\partial}H\| a ∥ u ∥ ≤ ∥ V N ∂ ˉ H ∥ follows from a ∥ u ∥ 2 ≤ q ε ( u ) = ⟨ V N ∂ ˉ H , u ⟩ ≤ ∥ V N ∂ ˉ H ∥ ∥ u ∥ a\|u\|^2\le q_\varepsilon(u)=\langle V_N\bar{\partial}H,u\rangle\le\|V_N\bar{\partial}H\|\,\|u\| a ∥ u ∥ 2 ≤ q ε ( u ) = ⟨ V N ∂ ˉ H , u ⟩ ≤ ∥ V N ∂ ˉ H ∥ ∥ u ∥ . The estimate for ∥ V N ∂ ˉ H ∥ 2 \|V_N\bar{\partial}H\|^2 ∥ V N ∂ ˉ H ∥ 2 follows from (A.8), since ∥ V N ∂ ˉ H ∥ 2 = K ε E G ε ∑ k ∣ ∂ k H ∣ 2 \|V_N\bar{\partial}H\|^2=K_\varepsilon\mathbb{E}_{G_\varepsilon}\sum_k|\partial_kH|^2 ∥ V N ∂ ˉ H ∥ 2 = K ε E G ε ∑ k ∣ ∂ k H ∣ 2 . □
Postponed proofs
Proof of Lemma 4.4 (Levi matrix of the shifted weight)
Proof. Let v ∈ ( R 2 ) N v \in(\mathbb{R}^{2})^{N} v ∈ ( R 2 ) N . For smooth functions, we deduce from (28) the following three identities:
v ⋅ C a ∣ Z ∣ 2 v = a ∣ v ∣ 2 , v \cdot\mathbf{C}_{a|Z|^{2}}v=a|v|^{2}, v ⋅ C a ∣ Z ∣ 2 v = a ∣ v ∣ 2 ,
v ⋅ C ∑ i < j h ( z i − z j ) v = 1 4 ∑ i < j Δ h ( z i − z j ) ∣ v i − v j ∣ 2 (106) v \cdot\mathbf{C}_{\sum_{i<j}h(z_i-z_j)}v=\frac{1}{4}\sum_{i<j}\Delta h(z_i-z_j)|v_i-v_j|^{2} \tag*{(106)} v ⋅ C ∑ i < j h ( z i − z j ) v = 4 1 i < j ∑ Δ h ( z i − z j ) ∣ v i − v j ∣ 2 (106)
v ⋅ C ∑ i η ( ∣ z i ∣ 2 ) v = ∑ i ( ( ∂ u η ) ( ∣ z i ∣ 2 ) + ∣ z i ∣ 2 ( ∂ u 2 η ) ( ∣ z i ∣ 2 ) ) ∣ v i ∣ 2 (107) v \cdot\mathbf{C}_{\sum_i\eta(|z_i|^{2})}v=\sum_i\left((\partial_u\eta)(|z_i|^{2})+|z_i|^{2}(\partial_u^{2}\eta)(|z_i|^{2})\right)|v_i|^{2} \tag*{(107)} v ⋅ C ∑ i η ( ∣ z i ∣ 2 ) v = i ∑ ( ( ∂ u η ) ( ∣ z i ∣ 2 ) + ∣ z i ∣ 2 ( ∂ u 2 η ) ( ∣ z i ∣ 2 ) ) ∣ v i ∣ 2 (107)
for h ∈ C 2 ( R 2 ) h \in C^{2}(\mathbb{R}^{2}) h ∈ C 2 ( R 2 ) and η ∈ C 2 ( [ 0 , ∞ ) ) \eta\in C^{2}([0,\infty)) η ∈ C 2 ([ 0 , ∞ )) . The quantity v ⋅ C Φ ( Z ) v v \cdot\mathbf{C}_{\Phi(Z)}v v ⋅ C Φ ( Z ) v is a quarter of the Laplacian of ( s , t ) ↦ Φ ( Z + s v + t J N v ) (s,t) \mapsto\Phi(Z+sv+tJ_Nv) ( s , t ) ↦ Φ ( Z + s v + t J N v ) at ( 0 , 0 ) (0,0) ( 0 , 0 ) . For each i i i , the map ( s , t ) ↦ z i + s v i + t v i (s,t) \mapsto z_i+sv_i+tv_i ( s , t ) ↦ z i + s v i + t v i is an affine similarity of the plane with dilation factor ∣ v i ∣ |v_i| ∣ v i ∣ , so the Laplacian of ( s , t ) ↦ h ( z i + s v i + t v i ) (s,t) \mapsto h(z_i+sv_i+tv_i) ( s , t ) ↦ h ( z i + s v i + t v i ) at ( 0 , 0 ) (0,0) ( 0 , 0 ) equals ∣ v i ∣ 2 Δ h ( z i ) |v_i|^{2}\Delta h(z_i) ∣ v i ∣ 2 Δ h ( z i ) . The same holds for z i − z j z_i-z_j z i − z j , with dilation factor ∣ v i − v j ∣ |v_i-v_j| ∣ v i − v j ∣ . Using Δ ∣ x ∣ 2 = 4 \Delta|x|^{2}=4 Δ∣ x ∣ 2 = 4 and Δ [ η ( ∣ x ∣ 2 ) ] = 4 ( ∂ u η ) ( ∣ x ∣ 2 ) + 4 ∣ x ∣ 2 ( ∂ u 2 η ) ( ∣ x ∣ 2 ) \Delta[\eta(|x|^{2})]=4(\partial_u\eta)(|x|^{2})+4|x|^{2}(\partial_u^{2}\eta)(|x|^{2}) Δ [ η ( ∣ x ∣ 2 )] = 4 ( ∂ u η ) ( ∣ x ∣ 2 ) + 4∣ x ∣ 2 ( ∂ u 2 η ) ( ∣ x ∣ 2 ) , we obtain the three identities.
Both sides of (B.1) depend continuously on h h h in the sense of distributions, so (B.1) holds for every locally integrable h h h . We apply it to h = ( β − 2 ) g h=(\beta-2)g h = ( β − 2 ) g , for which Δ h = 2 π ( 2 − β ) δ 0 \Delta h=2\pi(2-\beta)\delta_0 Δ h = 2 π ( 2 − β ) δ 0 , and obtain Lemma 4.4 4.4 4.4 .
Proof of Lemma 4.11 (structure of the kernel: division by the Vandermonde)
Proof. Let h h h be an antisymmetric holomorphic function on B L N B_L^N B L N . We first show that h = V N q h=V_Nq h = V N q with q q q holomorphic and symmetric, and then use the characters.
We first prove a local division property. We fix i < j i<j i < j and take as coordinates u = z i − z j u=z_i-z_j u = z i − z j and ξ \xi ξ , made of ( z i + z j ) / 2 (z_i+z_j)/2 ( z i + z j ) /2 and of the other positions. Let h h h be holomorphic on B L N B_L^N B L N and vanish on { u = 0 } \{u=0\} { u = 0 } . For fixed ξ \xi ξ , the set of admissible u u u is convex and contains 0 0 0 (the disk is convex), and Hadamard’s lemma gives
h ( u , ξ ) = h ( u , ξ ) − h ( 0 , ξ ) = ∫ 0 1 d d t h ( t u , ξ ) d t = u q ( u , ξ ) with q ( u , ξ ) : = ∫ 0 1 ∂ u h ( t u , ξ ) d t , h(u,\xi)=h(u,\xi)-h(0,\xi)=\int_0^1\frac{\mathrm{d}}{\mathrm{d}t}h(tu,\xi)\,\mathrm{d}t=uq(u,\xi)
\qquad\text{with}\qquad
q(u,\xi):=\int_0^1\partial_u h(tu,\xi)\,\mathrm{d}t, h ( u , ξ ) = h ( u , ξ ) − h ( 0 , ξ ) = ∫ 0 1 d t d h ( t u , ξ ) d t = u q ( u , ξ ) with q ( u , ξ ) := ∫ 0 1 ∂ u h ( t u , ξ ) d t ,
and q q q is holomorphic on B L N B_L^N B L N , by differentiation under the integral.
By antisymmetry under the transposition of i i i and j j j , h h h vanishes on { z i = z j } \{z_i=z_j\} { z i = z j } , and we can divide it by z i − z j z_i-z_j z i − z j by the local division property. We divide successively by all the factors z i − z j z_i-z_j z i − z j : on a new collision hyperplane, away from the hyperplanes already treated, the factors already divided out do not vanish, so the current quotient vanishes there. This set is dense in the hyperplane, so the vanishing holds everywhere by continuity. We obtain h = V N q h=V_Nq h = V N q with q q q holomorphic on B L N B_L^N B L N , and q q q is symmetric since h h h and V N V_N V N are antisymmetric. If h h h has character m m m , then q q q has character m − d N m-d_N m − d N , first away from the collisions and then everywhere by continuity. We conclude by Lemma 4.9 4.9 4.9 : q = 0 q=0 q = 0 if m < d N m<d_N m < d N , and q q q is a complex constant if m = d N m=d_N m = d N .
Proof of Lemma 5.5 (control of the collision term)
Proof. We may assume that N ≥ 2 N\geq2 N ≥ 2 and β < 2 \beta<2 β < 2 . We fix N , L , λ , R , ℓ N,L,\lambda,\mathcal{R},\ell N , L , λ , R , ℓ and d d d , and we use the norm of the space L s h , ε 2 L_{\mathrm{sh},\varepsilon}^{2} L sh , ε 2 of Subsection 5.2. Let
w ε : = e − Φ ε s h , C : = ⋃ i < j { Z : z i = z j } . w_{\varepsilon}:=e^{-\Phi_{\varepsilon}^{\mathrm{sh}}},
\qquad
\mathcal{C}:=\bigcup_{i<j}\{Z:z_i=z_j\}. w ε := e − Φ ε sh , C := i < j ⋃ { Z : z i = z j } .
We first record the bounds and symmetries of u ε u_{\varepsilon} u ε . By the anti-analytic-signal estimate and the domination of the physical density,
∥ V N ∂ ‾ H ∥ 2 ≤ ∫ ∣ V N ∣ 2 ∑ k ∣ ∇ k F ∣ 2 w ε d Z ≤ C . \|V_N\overline{\partial}H\|^{2}\leq\int|V_N|^{2}\sum_k|\nabla_kF|^{2}w_{\varepsilon}\,\mathrm{d}Z\leq C. ∥ V N ∂ H ∥ 2 ≤ ∫ ∣ V N ∣ 2 k ∑ ∣ ∇ k F ∣ 2 w ε d Z ≤ C .
We test A 1 u ε = V N ∂ ˉ H A_1u_\varepsilon=V_N\bar{\partial}H A 1 u ε = V N ∂ ˉ H with u ε u_\varepsilon u ε and use the positive terms in the form of A 1 A_1 A 1 . This gives
sup 0 < ε ≤ 1 ( ∥ u ε ∥ 2 + ∑ j , k ∥ ∂ ˉ j u ε , k ∥ 2 ) < ∞ . (108) \sup_{0<\varepsilon\leq1}\left(\lVert u_\varepsilon\rVert^2+\sum_{j,k}\lVert\bar{\partial}_j u_{\varepsilon,k}\rVert^2\right)<\infty. \tag*{(108)} 0 < ε ≤ 1 sup ∥ u ε ∥ 2 + j , k ∑ ∥ ∂ ˉ j u ε , k ∥ 2 < ∞. (108)
For a permutation σ \sigma σ , put ( σ Z ) i = z σ − 1 ( i ) (\sigma Z)_i=z_{\sigma^{-1}(i)} ( σ Z ) i = z σ − 1 ( i ) . By the symmetry of F F F , the antisymmetry of V N V_N V N , and uniqueness for the coercive equation, we have
u ε , σ ( i ) ( σ Z ) = sgn ( σ ) u ε , i ( Z ) . (109) u_{\varepsilon,\sigma(i)}(\sigma Z)=\operatorname{sgn}(\sigma)u_{\varepsilon,i}(Z). \tag*{(109)} u ε , σ ( i ) ( σ Z ) = sgn ( σ ) u ε , i ( Z ) . (109)
No concentration on compact sets. We show that, for every compact K ⊂ C N K\subset\mathbb{C}^N K ⊂ C N ,
lim η → 0 lim sup ε → 0 ∫ K ∩ { dist ( Z , C ) < η } ∣ u ε ∣ 2 w ε d Z = 0. (110) \lim_{\eta\to0}\limsup_{\varepsilon\to0}\int_{K\cap\{\operatorname{dist}(Z,\mathcal{C})<\eta\}}|u_\varepsilon|^2w_\varepsilon\,dZ=0. \tag*{(110)} η → 0 lim ε → 0 lim sup ∫ K ∩ { dist ( Z , C ) < η } ∣ u ε ∣ 2 w ε d Z = 0. (110)
Fix Z 0 ∈ C Z^0\in\mathcal{C} Z 0 ∈ C . We partition the indices into groups I 1 , … , I q I_1,\ldots,I_q I 1 , … , I q according to which coordinates coincide at Z 0 Z^0 Z 0 . Let r > 0 r>0 r > 0 be such that B ( Z 0 , 4 r ) B(Z^0,4r) B ( Z 0 , 4 r ) contains no collision between distinct groups. We add a Gaussian to the weight:
Ψ ε : = Φ ε s h + ∣ Z − Z 0 ∣ 2 r 2 . \Psi_\varepsilon:=\Phi^{\mathrm{sh}}_\varepsilon+\frac{|Z-Z^0|^2}{r^2}. Ψ ε := Φ ε sh + r 2 ∣ Z − Z 0 ∣ 2 .
Its Levi matrix is at least a + r − 2 a+r^{-2} a + r − 2 . We project each component of u ε u_\varepsilon u ε onto the holomorphic functions in L 2 ( e − Ψ ε ) L^2(e^{-\Psi_\varepsilon}) L 2 ( e − Ψ ε ) , and we call the resulting family h ε h_\varepsilon h ε . By the L 2 L^2 L 2 estimate of Hörmander, with the weight Ψ ε \Psi_\varepsilon Ψ ε , we have
∫ B ( Z 0 , 2 r ) ∣ u ε − h ε ∣ 2 w ε d Z ≤ e 4 r 2 ∫ e − ∣ Z − Z 0 ∣ 2 / r 2 ∑ j , k ∣ ∂ ˉ j u ε , k ∣ 2 w ε d Z . (111) \int_{B(Z^0,2r)}|u_\varepsilon-h_\varepsilon|^2w_\varepsilon\,dZ\leq e^{4r^2}\int e^{-|Z-Z^0|^2/r^2}\sum_{j,k}|\bar{\partial}_j u_{\varepsilon,k}|^2w_\varepsilon\,dZ. \tag*{(111)} ∫ B ( Z 0 , 2 r ) ∣ u ε − h ε ∣ 2 w ε d Z ≤ e 4 r 2 ∫ e − ∣ Z − Z 0 ∣ 2 / r 2 j , k ∑ ∣ ∂ ˉ j u ε , k ∣ 2 w ε d Z . (111)
The added Gaussian is invariant under permutations within each group. The projection therefore preserves (B.4) for these permutations.
Because the wall is fixed, w ε w_\varepsilon w ε has a positive lower bound on each compact set, uniformly in ε \varepsilon ε . The projection is a contraction, so the holomorphic functions h ε , k h_{\varepsilon,k} h ε , k have uniformly bounded unweighted L 2 L^2 L 2 norms on B ( Z 0 , 4 r ) B(Z^0,4r) B ( Z 0 , 4 r ) . By interior estimates, their derivatives are uniformly bounded on smaller balls.
Let I ( k ) I(k) I ( k ) be the group containing k k k . Exchanging two indices i , j i,j i , j in the same group, both different from k k k , changes the sign of h ε , k h_{\varepsilon,k} h ε , k . Hence this function vanishes on z i = z j z_i=z_j z i = z j and is divisible by z i − z j z_i-z_j z i − z j . We divide successively and obtain
h ε , k = h ~ ε , k ∏ p = 1 q ∏ i < j , i , j ∈ I p i , j ≠ k ( z i − z j ) , h_{\varepsilon,k}=\widetilde{h}_{\varepsilon,k}\prod_{p=1}^{q}\prod_{\substack{i<j,\ i,j\in I_p\\ i,j\neq k}}(z_i-z_j), h ε , k = h ε , k p = 1 ∏ q i < j , i , j ∈ I p i , j = k ∏ ( z i − z j ) ,
where h ~ ε , k \widetilde{h}_{\varepsilon,k} h ε , k is uniformly bounded on B ( Z 0 , 2 r ) B(Z^0,2r) B ( Z 0 , 2 r ) . Each division by a linear factor costs one further derivative on a slightly larger ball.
The squared factors cancel the singularities of the weight for all pairs not involving k k k . Distances between distinct groups are bounded below. Consequently,
∣ h ε , k ( Z ) ∣ 2 w ε ( Z ) ≤ C ( r , Z 0 ) ∏ j ∈ I ( k ) ∖ { k } ∣ z j − z k ∣ β − 2 on B ( Z 0 , 2 r ) . (112) |h_{\varepsilon,k}(Z)|^2w_\varepsilon(Z)\leq C(r,Z^0)\prod_{j\in I(k)\setminus\{k\}}|z_j-z_k|^{\beta-2}\quad\text{on }B(Z^0,2r). \tag*{(112)} ∣ h ε , k ( Z ) ∣ 2 w ε ( Z ) ≤ C ( r , Z 0 ) j ∈ I ( k ) ∖ { k } ∏ ∣ z j − z k ∣ β − 2 on B ( Z 0 , 2 r ) . (112)
This is an integrable majorant, as we see in the coordinates z k z_k z k and the independent differences z j − z k z_j-z_k z j − z k , since β − 2 > − 2 \beta-2>-2 β − 2 > − 2 . The projected part thus puts uniformly vanishing mass in shrinking neighbourhoods of C \mathcal{C} C .
To sum the errors in (B.6), we fix a sequence ε → 0 \varepsilon\to0 ε → 0 . By (B.3), we can extract a subsequence such that
∣ u ε ∣ 2 w ε d Z → ε → 0 μ , ∑ j , k ∣ ∂ ˉ j u ε , k ∣ 2 w ε d Z → ε → 0 ν |u_\varepsilon|^2w_\varepsilon\,dZ\xrightarrow[\varepsilon\to0]{}\mu,\qquad\sum_{j,k}|\bar{\partial}_j u_{\varepsilon,k}|^2w_\varepsilon\,dZ\xrightarrow[\varepsilon\to0]{}\nu ∣ u ε ∣ 2 w ε d Z ε → 0 μ , j , k ∑ ∣ ∂ ˉ j u ε , k ∣ 2 w ε d Z ε → 0 ν
against m a t h c a l C 0 ( C N ) mathcal{C}_0(\mathbb{C}^N) ma t h c a l C 0 ( C N ) , with μ , ν \mu,\nu μ , ν finite measures. We test ∣ u ε ∣ 2 ≤ 2 ∣ h ε ∣ 2 + 2 ∣ u ε − h ε ∣ 2 |u_\varepsilon|^2\leq2|h_\varepsilon|^2+2|u_\varepsilon-h_\varepsilon|^2 ∣ u ε ∣ 2 ≤ 2∣ h ε ∣ 2 + 2∣ u ε − h ε ∣ 2 with continuous cutoffs supported in B ( Z 0 , 2 r ) B(Z^0,2r) B ( Z 0 , 2 r ) and shrinking to C \mathcal{C} C . We use (B.7) to remove the projected part, and obtain
μ ( C ∩ B ( Z 0 , r ) ) ≤ 2 e 4 r 2 ∫ e − ∣ Z − Z 0 ∣ 2 / r 2 d ν ( Z ) . \mu(\mathcal{C}\cap B(Z^0,r))\leq2e^{4r^2}\int e^{-|Z-Z^0|^2/r^2}\,d\nu(Z). μ ( C ∩ B ( Z 0 , r )) ≤ 2 e 4 r 2 ∫ e − ∣ Z − Z 0 ∣ 2 / r 2 d ν ( Z ) .
Fix a partition of the indices and a compact set of collision configurations of this type, with distinct groups separated by at least 1 / m 1/m 1/ m . For r < 1 / ( 16 m ) r < 1/(16m) r < 1/ ( 16 m ) , we cover this set by balls B ( Z ( n ) , r ) B(Z^{(n)},r) B ( Z ( n ) , r ) whose centres are r r r -separated. By the elementary Gaussian packing bound
sup Z ∑ n e − ∣ Z − Z ( n ) ∣ 2 / r 2 ≤ C ( N ) , \sup_Z \sum_n e^{-|Z-Z^{(n)}|^2/r^2} \le C(N), Z sup n ∑ e − ∣ Z − Z ( n ) ∣ 2 / r 2 ≤ C ( N ) ,
its μ \mu μ -mass is at most C ( N ) r 2 ν ( C N ) C(N)r^{2\nu}(\mathbb{C}^N) C ( N ) r 2 ν ( C N ) . We let r → 0 r \to0 r → 0 , and then take the union over m m m , the finitely many partitions, and a compact exhaustion. We obtain μ ( C ) = 0 \mu(\mathcal{C})=0 μ ( C ) = 0 . Since this holds for every subsequential limit, we obtain (B.5) by the Portmanteau theorem.
Control at infinity. For τ > 0 \tau>0 τ > 0 , let
Q τ ( Z ) : = e a min { ∣ Z ∣ , τ } . Q_\tau(Z):=e^{\sqrt{a}\min\{|Z|,\tau\}}. Q τ ( Z ) := e a m i n { ∣ Z ∣ , τ } .
This weight is common to all components, so it commutes with the Levi matrix. We repeat the conjugation calculation of Subsection 5.4, now with ∑ j ∣ ∂ ˉ j log Q τ ∣ 2 ≤ a / 4 \sum_j|\bar{\partial}_j\log Q_\tau|^2\le a/4 ∑ j ∣ ∂ ˉ j log Q τ ∣ 2 ≤ a /4 , and obtain
3 a 4 ∥ Q τ u ε ∥ ≤ ∥ Q τ V N ∂ ˉ H ∥ . \frac{3a}{4}\|Q_\tau u_\varepsilon\|\le\|Q_\tau V_N\bar{\partial}H\|. 4 3 a ∥ Q τ u ε ∥ ≤ ∥ Q τ V N ∂ ˉ H ∥.
The multiplier is bounded and Lipschitz at fixed τ \tau τ , so the calculation is valid in the form domain. Since Q τ Q_\tau Q τ is invariant under simultaneous rotations, we also get, as in the proof of (A.8) with the weight Q τ 2 Q_\tau^2 Q τ 2 ,
∥ Q τ V N ∂ ˉ H ∥ 2 ≤ ∫ e 2 a ∣ Z ∣ ∣ V N ∣ 2 ∑ k ∣ ∇ k F ∣ 2 w ε d Z ≤ C . \|Q_\tau V_N\bar{\partial}H\|^2\le\int e^{2\sqrt{a}|Z|}|V_N|^2\sum_k|\nabla_kF|^2w_\varepsilon\,dZ\le C. ∥ Q τ V N ∂ ˉ H ∥ 2 ≤ ∫ e 2 a ∣ Z ∣ ∣ V N ∣ 2 k ∑ ∣ ∇ k F ∣ 2 w ε d Z ≤ C .
The last bound follows from ∣ V N ∣ 2 w ε ≤ e − a ∣ Z ∣ 2 ∏ i < j ∣ z i − z j ∣ β |V_N|^2w_\varepsilon\le e^{-a|Z|^2}\prod_{i<j}|z_i-z_j|^\beta ∣ V N ∣ 2 w ε ≤ e − a ∣ Z ∣ 2 ∏ i < j ∣ z i − z j ∣ β . Letting τ → ∞ \tau\to\infty τ → ∞ , we obtain
sup 0 < ε ≤ 1 ∫ e 2 a ∣ Z ∣ ∣ u ε ∣ 2 w ε d Z < ∞ . (113) \sup_{0<\varepsilon\le1}\int e^{2\sqrt{a}|Z|}|u_\varepsilon|^2w_\varepsilon\,dZ<\infty. \tag*{(113)} 0 < ε ≤ 1 sup ∫ e 2 a ∣ Z ∣ ∣ u ε ∣ 2 w ε d Z < ∞. (113)
The collision term. Since ζ \zeta ζ is 1-Lipschitz and bounded by d d d , ∣ M i − M j ∣ ≤ ℓ − 1 e d / ℓ ∣ z i − z j ∣ |M_i-M_j|\le\ell^{-1}e^{d/\ell}|z_i-z_j| ∣ M i − M j ∣ ≤ ℓ − 1 e d / ℓ ∣ z i − z j ∣ . On the support of ω i j , ε \omega_{ij,\varepsilon} ω ij , ε , the last distance is at most ε \varepsilon ε . Thus
0 ≤ ω i j , ε ( M i − M j ) 2 ≤ C ( β , ψ ) ℓ − 2 e 2 d / ℓ 1 { ∣ z i − z j ∣ ≤ ε } . 0\le\omega_{ij,\varepsilon}(M_i-M_j)^2\le C(\beta,\psi)\ell^{-2}e^{2d/\ell}\mathbf{1}_{\{|z_i-z_j|\le\varepsilon\}}. 0 ≤ ω ij , ε ( M i − M j ) 2 ≤ C ( β , ψ ) ℓ − 2 e 2 d / ℓ 1 { ∣ z i − z j ∣ ≤ ε } .
Using 2 ∣ Re ( u i u j ‾ ) ∣ ≤ ∣ u i ∣ 2 + ∣ u j ∣ 2 2|\operatorname{Re}(u_i\overline{u_j})|\le|u_i|^2+|u_j|^2 2∣ Re ( u i u j ) ∣ ≤ ∣ u i ∣ 2 + ∣ u j ∣ 2 , we conclude that
∣ B ε ( u ε ) ∣ ≤ C ∫ { dist ( Z , C ) ≤ ε } ∣ u ε ∣ 2 w ε d Z . |\mathcal{B}_\varepsilon(u_\varepsilon)|\le C\int_{\{\operatorname{dist}(Z,\mathcal{C})\le\varepsilon\}}|u_\varepsilon|^2w_\varepsilon\,dZ. ∣ B ε ( u ε ) ∣ ≤ C ∫ { dist ( Z , C ) ≤ ε } ∣ u ε ∣ 2 w ε d Z .
On a fixed compact set, the right-hand side tends to zero by (B.5). Its complement is uniformly small by (B.8). This proves the lemma.
Proof of Lemmas 5.8 and 5.9 (regularity of the two-point function)
Proof of Lemma 5.8. Let h h h be harmonic on B 3 B_3 B 3 , let ξ 1 , … , ξ m ∈ C \xi_1,\ldots,\xi_m\in\mathbb{C} ξ 1 , … , ξ m ∈ C , and let
U ( t ) = e h ( t ) ∏ j = 1 m ∣ t − ξ j ∣ β . U(t)=e^{h(t)}\prod_{j=1}^{m}|t-\xi_j|^\beta. U ( t ) = e h ( t ) j = 1 ∏ m ∣ t − ξ j ∣ β .
If U U U is integrable on B 3 B_3 B 3 , then
∣ U ( t 1 ) − U ( t 2 ) ∣ ≤ C ( β ) ∥ U ∥ L 1 ( B 3 ) ∣ t 1 − t 2 ∣ min ( β , 1 ) (114) |U(t_1)-U(t_2)|\le C(\beta)\|U\|_{L^1(B_3)}|t_1-t_2|^{\min(\beta,1)} \tag*{(114)} ∣ U ( t 1 ) − U ( t 2 ) ∣ ≤ C ( β ) ∥ U ∥ L 1 ( B 3 ) ∣ t 1 − t 2 ∣ m i n ( β , 1 ) (114)
for t 1 , t 2 ∈ B 1 t_1,t_2\in B_1 t 1 , t 2 ∈ B 1 . The constant is independent of m m m , h h h and the ξ j \xi_j ξ j . To prove this, we may assume, adding a constant to h h h , that ∥ U ∥ L 1 ( B 3 ) = 1 \lVert U\rVert_{L^{1}(B_{3})}=1 ∥ U ∥ L 1 ( B 3 ) = 1 . The function U U U is subharmonic, so sup B 2 U ≤ C \sup_{B_{2}}U\le C sup B 2 U ≤ C . Near a zero ξ j ∈ B 2 \xi_j\in B_{2} ξ j ∈ B 2 , the function U ( t ) / ∣ t − ξ j ∣ β U(t)/\lvert t-\xi_j\rvert^{\beta} U ( t ) / ∣ t − ξ j ∣ β is also subharmonic. We integrate its Poisson bound on circles centred at ξ j \xi_j ξ j over radii between 1 / 4 1/4 1/4 and 1 / 2 1/2 1/2 , and obtain
U ( t ) ∣ t − ξ j ∣ β ≤ C ( β ) if ∣ t − ξ j ∣ ≤ 1 / 8. \frac{U(t)}{\lvert t-\xi_j\rvert^{\beta}}\le C(\beta)\qquad\text{if }\lvert t-\xi_j\rvert\le1/8. ∣ t − ξ j ∣ β U ( t ) ≤ C ( β ) if ∣ t − ξ j ∣ ≤ 1/8.
With d ( t ) : = min ( 1 / 8 , dist ( t , { ξ j } ) ) d(t):=\min(1/8,\operatorname{dist}(t,\{\xi_j\})) d ( t ) := min ( 1/8 , dist ( t , { ξ j })) , this yields U ( t ) ≤ C ( β ) d ( t ) β U(t)\le C(\beta)d(t)^{\beta} U ( t ) ≤ C ( β ) d ( t ) β on B 3 / 2 B_{3/2} B 3/2 . Off the zeros, log U \log U log U is harmonic on B ( t , d ( t ) / 2 ) B(t,d(t)/2) B ( t , d ( t ) /2 ) , where U ≤ C ( β ) d ( t ) β U\le C(\beta)d(t)^{\beta} U ≤ C ( β ) d ( t ) β . By the gradient estimate for the non-negative harmonic function log ( C ( β ) d ( t ) β / U ) \log(C(\beta)d(t)^{\beta}/U) log ( C ( β ) d ( t ) β / U ) , we get
∣ ∇ U ( t ) ∣ ≤ C d ( t ) U ( t ) log C ( β ) d ( t ) β U ( t ) ≤ C C ( β ) d ( t ) β d ( t ) ≤ C ( β ) d ( t ) β − 1 for t ∈ B 1 ∖ { ξ j } . \lvert\nabla U(t)\rvert\le\frac{C}{d(t)}U(t)\log\frac{C(\beta)d(t)^{\beta}}{U(t)}
\le\frac{CC(\beta)d(t)^{\beta}}{d(t)}
\le C(\beta)d(t)^{\beta-1}
\qquad\text{for }t\in B_{1}\setminus\{\xi_j\}. ∣ ∇ U ( t )∣ ≤ d ( t ) C U ( t ) log U ( t ) C ( β ) d ( t ) β ≤ d ( t ) C C ( β ) d ( t ) β ≤ C ( β ) d ( t ) β − 1 for t ∈ B 1 ∖ { ξ j } .
If ∣ t 1 − t 2 ∣ < 1 / 64 \lvert t_1-t_2\rvert<1/64 ∣ t 1 − t 2 ∣ < 1/64 , then either a zero is within 2 ∣ t 1 − t 2 ∣ 2\lvert t_1-t_2\rvert 2 ∣ t 1 − t 2 ∣ of t 1 t_1 t 1 , in which case both values are bounded by C ( β ) ∣ t 1 − t 2 ∣ β C(\beta)\lvert t_1-t_2\rvert^{\beta} C ( β ) ∣ t 1 − t 2 ∣ β , or d ≥ ∣ t 1 − t 2 ∣ d\ge\lvert t_1-t_2\rvert d ≥ ∣ t 1 − t 2 ∣ along the segment from t 1 t_1 t 1 to t 2 t_2 t 2 , and we integrate the gradient bound. In both cases, we obtain (B.9). When ∣ t 1 − t 2 ∣ ≥ 1 / 64 \lvert t_1-t_2\rvert\ge1/64 ∣ t 1 − t 2 ∣ ≥ 1/64 , it follows from the bound on U U U .
Let ρ 2 ( N ) \rho^{(N)}_2 ρ 2 ( N ) be the two-point density of the gas of N N N particles under P N , β P_{N,\beta} P N , β , along the subsequence defining P P P . Fix x 0 , y x_0,y x 0 , y , with B ( x 0 , 3 ) B(x_0,3) B ( x 0 , 3 ) inside the wall. From the Gibbs density, we get the positive mixture
e a ∣ t ∣ 2 ρ 2 ( N ) ( x 0 + t , y ) = ∫ e − 2 a x 0 ⋅ t ∣ x 0 + t − y ∣ β ∏ z ∈ ω ∣ x 0 + t − z ∣ β d ν ( ω ) , e^{a\lvert t\rvert^2}\rho^{(N)}_2(x_0+t,y)=\int e^{-2ax_0\cdot t}\lvert x_0+t-y\rvert^\beta\prod_{z\in\omega}\lvert x_0+t-z\rvert^\beta\,d\nu(\omega), e a ∣ t ∣ 2 ρ 2 ( N ) ( x 0 + t , y ) = ∫ e − 2 a x 0 ⋅ t ∣ x 0 + t − y ∣ β z ∈ ω ∏ ∣ x 0 + t − z ∣ β d ν ( ω ) ,
where ω \omega ω contains the N − 2 N-2 N − 2 unmarked positions and ν \nu ν is independent of t t t . Each integrand has the form just considered. We integrate (B.9) and remove the Gaussian factor. This gives
∣ ρ 2 ( N ) ( x 0 + t 1 , y ) − ρ 2 ( N ) ( x 0 + t 2 , y ) ∣ ≤ C ( β ) ∣ t 1 − t 2 ∣ min ( β , 1 ) ∫ B ( x 0 , 3 ) ρ 2 ( N ) ( v , y ) d v for t 1 , t 2 ∈ B 1 . (115) \begin{aligned}
\left\lvert\rho^{(N)}_2(x_0+t_1,y)-\rho^{(N)}_2(x_0+t_2,y)\right\rvert
&\le C(\beta)\lvert t_1-t_2\rvert^{\min(\beta,1)}
\int_{B(x_0,3)}\rho^{(N)}_2(v,y)\,dv
\qquad\text{for }t_1,t_2\in B_1. \tag*{(115)}
\end{aligned} ρ 2 ( N ) ( x 0 + t 1 , y ) − ρ 2 ( N ) ( x 0 + t 2 , y ) ≤ C ( β ) ∣ t 1 − t 2 ∣ m i n ( β , 1 ) ∫ B ( x 0 , 3 ) ρ 2 ( N ) ( v , y ) d v for t 1 , t 2 ∈ B 1 . (115)
The Gaussian factor contributes a Lipschitz error controlled by the same integral, by subharmonicity. By subharmonicity in each variable, we also have, whenever the unit disks lie inside the wall,
ρ 2 ( N ) ( x , y ) ≤ ( e a π ) 2 ∫ B ( x , 1 ) × B ( y , 1 ) ρ 2 ( N ) ( u , v ) d u d v ≤ ( e a π ) 2 E P N , β [ N B ( x , 1 ) N B ( y , 1 ) ] . \rho^{(N)}_2(x,y)\le\left(\frac{e^a}{\pi}\right)^2\int_{B(x,1)\times B(y,1)}\rho^{(N)}_2(u,v)\,du\,dv\le\left(\frac{e^a}{\pi}\right)^2\mathbb{E}_{P_{N,\beta}}\left[N_{B(x,1)}N_{B(y,1)}\right]. ρ 2 ( N ) ( x , y ) ≤ ( π e a ) 2 ∫ B ( x , 1 ) × B ( y , 1 ) ρ 2 ( N ) ( u , v ) d u d v ≤ ( π e a ) 2 E P N , β [ N B ( x , 1 ) N B ( y , 1 ) ] .
By the local moment bounds of Proposition 2.4, ρ 2 ( N ) \rho^{(N)}_2 ρ 2 ( N ) is therefore bounded uniformly on compact sets. With (B.10), applied in each variable, we obtain local equicontinuity. By Arzelà–Ascoli and convergence of the factorial moment measures, there is a continuous limiting density ρ 2 \rho_2 ρ 2 . Since a continuous density is unique, the convergence is locally uniform along the chosen subsequence.
We pass to the limit in the submean bound above, and use stationarity and Cauchy–Schwarz. This gives
0 ≤ ρ 2 ( x , y ) ≤ ( e a π ) 2 E P [ N B ( x , 1 ) N B ( y , 1 ) ] ≤ ( e a π ) 2 E P N B 1 2 < ∞ . 0\le\rho_2(x,y)\le\left(\frac{e^a}{\pi}\right)^2\mathbb{E}_{P}\left[N_{B(x,1)}N_{B(y,1)}\right]\le\left(\frac{e^a}{\pi}\right)^2\mathbb{E}_{P}N_{B_1}^2<\infty. 0 ≤ ρ 2 ( x , y ) ≤ ( π e a ) 2 E P [ N B ( x , 1 ) N B ( y , 1 ) ] ≤ ( π e a ) 2 E P N B 1 2 < ∞.
Finally, we pass to the limit in (B.10) and use this global bound. We vary the two variables one after the other, and this proves the lemma. □
Proof of Lemma 5.9. We use the cutoffs χ 1 / 8 \chi_{1/8} χ 1/8 and χ 1 / 2 \chi_{1/2} χ 1/2 of § 3 . The first equals 1 1 1 on B 1 / 8 B_{1/8} B 1/8 and is supported in B 1 / 4 B_{1/4} B 1/4 , and the second equals 1 1 1 on B 1 / 2 B_{1/2} B 1/2 and is supported in B 1 B_1 B 1 . For ϵ ∈ { ϵ 1 , ϵ 2 } \epsilon\in\{\epsilon_1,\epsilon_2\} ϵ ∈ { ϵ 1 , ϵ 2 } , set v ϵ : = χ 1 / 8 ϵ v_\epsilon:=\chi_{1/8}\epsilon v ϵ := χ 1/8 ϵ . These vector fields translate a marked point together with all sufficiently close particles.
Step 1. Moments of every order of the exterior field in every local smoothness norm. Let ( ν j ) j ≥ 0 (\nu_j)_{j\ge0} ( ν j ) j ≥ 0 be the dyadic partition of Remark 2.2, used to define the move function (11) , and let
V ( t ) : = ∫ ( 1 − χ 1 / 2 ( w ) ) ∇ g ( t − w ) d ( X − Leb ) ( w ) for t ∈ B 1 / 2 , \mathcal{V}(t):=\int(1-\chi_{1/2}(w))\nabla g(t-w)\,d(X-\operatorname{Leb})(w)\qquad\text{for }t\in B_{1/2}, V ( t ) := ∫ ( 1 − χ 1/2 ( w )) ∇ g ( t − w ) d ( X − Leb ) ( w ) for t ∈ B 1/2 ,
where the integral is defined with this partition, as for the move function. For every j j j and multi-indices α \alpha α and α 1 \alpha_1 α 1 with ∣ α 1 ∣ ≤ 3 |\alpha_1| \le3 ∣ α 1 ∣ ≤ 3 ,
sup ∣ t ∣ ≤ 5 / 12 ∥ ∂ w α 1 [ ( 1 − χ 1 / 2 ( w ) ) ϑ j ( w ) ∂ t α ∇ g ( t − w ) ] ∥ ∞ ≤ C ( α ) 2 − j ( 1 + ∣ α ∣ + ∣ α 1 ∣ ) . \sup_{|t|\le5/12}\left\|\partial_w^{\alpha_1}\left[(1-\chi_{1/2}(w))\vartheta_j(w)\partial_t^\alpha\nabla g(t-w)\right]\right\|_\infty\le C(\alpha)2^{-j(1+|\alpha|+|\alpha_1|)}. ∣ t ∣ ≤ 5/12 sup ∂ w α 1 [ ( 1 − χ 1/2 ( w )) ϑ j ( w ) ∂ t α ∇ g ( t − w ) ] ∞ ≤ C ( α ) 2 − j ( 1 + ∣ α ∣ + ∣ α 1 ∣ ) .
The finitely many inner terms satisfy the same estimate after increasing C ( α ) C(\alpha) C ( α ) . By (G2), for every 1 ≤ p < ∞ 1 \le p < \infty 1 ≤ p < ∞ ,
sup ∣ t ∣ ≤ 5 / 12 ∥ ∫ ( 1 − χ 1 / 2 ( w ) ) ϑ j ( w ) ∂ t α ∇ g ( t − w ) d ( X − Leb ) ( w ) ∥ p ≤ C ( P , p , α ) 2 − j ( 1 + ∣ α ∣ ) . \sup_{|t|\le5/12}\left\|\int(1-\chi_{1/2}(w))\vartheta_j(w)\partial_t^\alpha\nabla g(t-w)\,\mathrm{d}(X-\operatorname{Leb})(w)\right\|_p \le C(P,p,\alpha)2^{-j(1+|\alpha|)}. ∣ t ∣ ≤ 5/12 sup ∫ ( 1 − χ 1/2 ( w )) ϑ j ( w ) ∂ t α ∇ g ( t − w ) d ( X − Leb ) ( w ) p ≤ C ( P , p , α ) 2 − j ( 1 + ∣ α ∣ ) .
We sum over j j j and apply Sobolev embedding in t t t . We obtain convergence in L p ( P ; C k ( B ‾ 1 / 3 ) ) L^p(\mathbb{P};C^k(\overline{B}_{1/3})) L p ( P ; C k ( B 1/3 )) and
E P [ ∥ V ∥ C k ( B ‾ 1 / 3 ) p ] < ∞ for every k ≥ 0 and 1 ≤ p < ∞ . (116) \mathbb{E}_{\mathbb{P}}\left[\|V\|_{C^k(\overline{B}_{1/3})}^p\right]<\infty\qquad\text{for every } k\ge0 \text{ and } 1\le p<\infty. \tag*{(116)} E P [ ∥ V ∥ C k ( B 1/3 ) p ] < ∞ for every k ≥ 0 and 1 ≤ p < ∞. (116)
Step 2. Local transports and integration by parts without singular collision terms. For v ∈ { v e 1 , v e 2 } v\in\{v_{e_1},v_{e_2}\} v ∈ { v e 1 , v e 2 } , put
T v F : = ∑ z ∈ X v ( z ) ⋅ ∇ z F , K v ( z , w ) : = ( v ( z ) − v ( w ) ) ⋅ ∇ g ( z − w ) , T_vF:=\sum_{z\in X}v(z)\cdot\nabla_zF,\qquad K_v(z,w):=(v(z)-v(w))\cdot\nabla g(z-w), T v F := z ∈ X ∑ v ( z ) ⋅ ∇ z F , K v ( z , w ) := ( v ( z ) − v ( w )) ⋅ ∇ g ( z − w ) ,
and define
R v : = β 2 ∑ z , w ∈ X z ≠ w χ 1 / 2 ( z ) χ 1 / 2 ( w ) K v ( z , w ) R_v:=\frac{\beta}{2}\sum_{\substack{z,w\in X\\z\ne w}}\chi_{1/2}(z)\chi_{1/2}(w)K_v(z,w) R v := 2 β z , w ∈ X z = w ∑ χ 1/2 ( z ) χ 1/2 ( w ) K v ( z , w )
+ β ∑ z ∈ X ∩ B 1 / 4 v ( z ) ⋅ ( V ( z ) − ∫ χ 1 / 2 ( w ) ∇ g ( z − w ) d w ) − X ( div v ) . \qquad+\beta\sum_{z\in X\cap B_{1/4}}v(z)\cdot\left(\mathcal V(z)-\int\chi_{1/2}(w)\nabla g(z-w)\,\mathrm{d}w\right)-X(\operatorname{div}v). + β z ∈ X ∩ B 1/4 ∑ v ( z ) ⋅ ( V ( z ) − ∫ χ 1/2 ( w ) ∇ g ( z − w ) d w ) − X ( div v ) .
The sum of the first two terms is β \beta β times the derivative of the conditional DLR energy under the flow of v v v . We differentiate the conditional density and obtain
E P [ T v F ] = E P [ F R v ] . (117) \mathbb{E}_{\mathbb{P}}[T_vF]=\mathbb{E}_{\mathbb{P}}[FR_v]. \tag*{(117)} E P [ T v F ] = E P [ F R v ] . (117)
To prove it, we condition in a bounded domain containing the support of v v v , and integrate by parts in its finitely many particle positions. There is no boundary contribution from the domain. For a collision cutoff at ∣ z − w ∣ ≃ δ |z-w|\simeq\delta ∣ z − w ∣ ≃ δ , the factor δ − 1 \delta^{-1} δ − 1 is cancelled by ∣ v ( z ) − v ( w ) ∣ ≤ C ∣ z − w ∣ |v(z)-v(w)|\le C|z-w| ∣ v ( z ) − v ( w ) ∣ ≤ C ∣ z − w ∣ , and the cutoff error tends to zero by dominated convergence.
Every iterated transport derivative of K v K_v K v is bounded. To see this, let z ↦ z s z\mapsto z_s z ↦ z s be a composition of the flows of the v e v_e v e , at small times given by the parameter s s s . All parameter derivatives of
log ∣ z s − w s ∣ ∣ z − w ∣ \log\frac{|z_s-w_s|}{|z-w|} log ∣ z − w ∣ ∣ z s − w s ∣
are bounded uniformly for z ≠ w z\ne w z = w and small s s s , since the flows and their inverses are uniformly Lipschitz, and every parameter derivative of z s z_s z s is Lipschitz in z z z . The exterior configuration in V \mathcal V V is fixed by these flows, and ∫ χ 1 / 2 ( w ) ∇ g ( ⋅ − w ) d w \int\chi_{1/2}(w)\nabla g(\mathord\cdot-w)\,\mathrm{d}w ∫ χ 1/2 ( w ) ∇ g ( ⋅ − w ) d w is smooth. Hence, for u 1 , … , u j , v ∈ { v e 1 , v e 2 } u_1,\ldots,u_j,v\in\{v_{e_1},v_{e_2}\} u 1 , … , u j , v ∈ { v e 1 , v e 2 } ,
∣ T u 1 ⋯ T u j R v ∣ ≤ C ( β , j ) ( 1 + N B 1 ) 2 ( 1 + ∥ v ∥ C j ( B ‾ 1 / 3 ) ) . |T_{u_1}\cdots T_{u_j}R_v|\le C(\beta,j)(1+N_{B_1})^2(1+\|v\|_{C^j(\overline{B}_{1/3})}). ∣ T u 1 ⋯ T u j R v ∣ ≤ C ( β , j ) ( 1 + N B 1 ) 2 ( 1 + ∥ v ∥ C j ( B 1/3 ) ) .
By (G1) and (B.11), these variables and their finite products belong to every L p ( P ) L^p(\mathbb{P}) L p ( P ) with p < ∞ p<\infty p < ∞ . Thus (B.12) also holds for the products used below.
Fix ∣ r ∣ ≥ 1 |r|\ge1 ∣ r ∣ ≥ 1 , a non-negative η ∈ C c ∞ ( B 1 / 16 ) \eta\in C^\infty_c(B_{1/16}) η ∈ C c ∞ ( B 1/16 ) with ∫ η = 1 \int\eta=1 ∫ η = 1 , and f ∈ C c ∞ ( B ( r , 1 / 16 ) ) f\in C^\infty_c(B(r,1/16)) f ∈ C c ∞ ( B ( r , 1/16 )) . Set
φ ( x , y ) : = η ( x ) f ( y − x ) , F φ ( X ) : = ∑ z , w ∈ X z ≠ w φ ( z , w ) . \varphi(x,y):=\eta(x)f(y-x),\qquad F_\varphi(X):=\sum_{\substack{z,w\in X\\z\ne w}}\varphi(z,w). φ ( x , y ) := η ( x ) f ( y − x ) , F φ ( X ) := z , w ∈ X z = w ∑ φ ( z , w ) .
On the support of φ \varphi φ , v e ( x ) = e v_e(x)=e v e ( x ) = e and v e ( y ) = 0 v_e(y)=0 v e ( y ) = 0 , hence
T v e F φ = F ∂ x , e φ . \mathcal{T}_{v_e}F_\varphi=F_{\partial_{x,e}\varphi}. T v e F φ = F ∂ x , e φ .
For a word of coordinate directions e i 1 , … , e i m e_{i_1},\ldots,e_{i_m} e i 1 , … , e i m , start from W 0 = 1 W_0=1 W 0 = 1 and set
W j + 1 : = R v e i j + 1 W j − T v e i j + 1 W j . W_{j+1}:=\mathcal{R}_{v_{e_{i_{j+1}}}}W_j-\mathcal{T}_{v_{e_{i_{j+1}}}}W_j. W j + 1 := R v e i j + 1 W j − T v e i j + 1 W j .
We apply (B.12) repeatedly and get
∬ ∂ x α φ ( x , y ) ρ 2 ( x , y ) d x d y = E P [ F φ W m ] , ∥ W m ∥ p ≤ C ( P , β , m , p ) . \iint\partial_x^\alpha\varphi(x,y)\rho_2(x,y)\,\mathrm{d}x\,\mathrm{d}y=\mathbb{E}_{P}[F_\varphi W_m],\qquad\|W_m\|_p\le C(P,\beta,m,p). ∬ ∂ x α φ ( x , y ) ρ 2 ( x , y ) d x d y = E P [ F φ W m ] , ∥ W m ∥ p ≤ C ( P , β , m , p ) .
where α \alpha α is the multi-index of the word. By stationarity, ρ 2 ( x , y ) = ρ 2 ( 0 , y − x ) \rho_2(x,y)=\rho_2(0,y-x) ρ 2 ( x , y ) = ρ 2 ( 0 , y − x ) . Integrating the derivatives of η \eta η , we get
⟨ ∂ α ρ 2 ( 0 , ⋅ ) , f ⟩ = E P [ ∑ z , w ∈ X z ≠ w η ( z ) f ( w − z ) W m ] . \langle\partial^\alpha\rho_2(0,\cdot),f\rangle=\mathbb{E}_{P}\left[\sum_{\substack{z,w\in X\\z\ne w}}\eta(z)f(w-z)W_m\right]. ⟨ ∂ α ρ 2 ( 0 , ⋅ ) , f ⟩ = E P z , w ∈ X z = w ∑ η ( z ) f ( w − z ) W m .
The right-hand side defines an absolutely continuous signed measure in w − z w-z w − z , since the second factorial moment measure is absolutely continuous. Its total variation on B ( r , 1 / 16 ) B(r,1/16) B ( r , 1/16 ) is at most
∥ η ∥ ∞ E P [ ∣ W m ∣ N B 1 / 16 N B ( r , 1 / 8 ) ] ≤ C ( P , β , m ) , \|\eta\|_\infty\mathbb{E}_{P}\left[|W_m|N_{B_{1/16}}N_{B(r,1/8)}\right]\le C(P,\beta,m), ∥ η ∥ ∞ E P [ ∣ W m ∣ N B 1/16 N B ( r , 1/8 ) ] ≤ C ( P , β , m ) ,
by Hölder’s inequality, (G1) and stationarity. Thus
sup ∣ r ∣ ≥ 1 ∥ ∂ α ρ 2 ( 0 , ⋅ ) ∥ L 1 ( B ( r , 1 / 16 ) ) ≤ C ( P , β , α ) for every multi-index α . \sup_{|r|\ge1}\|\partial^\alpha\rho_2(0,\cdot)\|_{L^1(B(r,1/16))}\le C(P,\beta,\alpha)\qquad\text{for every multi-index }\alpha. ∣ r ∣ ≥ 1 sup ∥ ∂ α ρ 2 ( 0 , ⋅ ) ∥ L 1 ( B ( r , 1/16 )) ≤ C ( P , β , α ) for every multi-index α .
By local Sobolev embedding, every derivative of ρ 2 ( 0 , ⋅ ) \rho_2(0,\cdot) ρ 2 ( 0 , ⋅ ) is uniformly bounded on { ∣ r ∣ ≥ 1 } \{|r|\ge1\} { ∣ r ∣ ≥ 1 } . We choose the cutoff radii smaller than a fixed multiple of min ( d 0 , 1 ) \min(d_0,1) min ( d 0 , 1 ) , and obtain the same assertion on { ∣ r ∣ ≥ d 0 } \{|r|\ge d_0\} { ∣ r ∣ ≥ d 0 } for every d 0 > 0 d_0>0 d 0 > 0 .
Proof of Lemma 6.1 (squares versus disks)
Proof. Let Λ \Lambda Λ be either the square [ − 1 2 , 1 2 ] 2 [-\frac{1}{2},\frac{1}{2}]^2 [ − 2 1 , 2 1 ] 2 or the disk of area 1 1 1 , and let Ω R : = R Λ \Omega_R:=R\Lambda Ω R := R Λ (in this proof only), so that Ω R = C R \Omega_R=C_R Ω R = C R in the first case and B R = Ω π R B_R=\Omega_{\sqrt{\pi}R} B R = Ω π R in the second case. Let g Λ ( u ) : = ∣ Λ ∩ ( Λ + u ) ∣ g_\Lambda(u):=|\Lambda\cap(\Lambda+u)| g Λ ( u ) := ∣Λ ∩ ( Λ + u ) ∣ be the covariogram of Λ \Lambda Λ . It satisfies 0 ≤ g Λ ≤ 1 = g Λ ( 0 ) 0\le g_\Lambda\le1=g_\Lambda(0) 0 ≤ g Λ ≤ 1 = g Λ ( 0 ) , and it is Lipschitz, with constant at most the perimeter of Λ \Lambda Λ .
Proof of (i). Expanding the product measure in (10) , using the definition of ρ 2 \rho_2 ρ 2 , the intensity 1 1 1 and stationarity, we get
E P [ F Ω R ( X Ω R ) ] = 1 2 ∬ Ω R × Ω R g ( x − y ) ( ρ 2 ( x , y ) − 1 ) d x d y = ∣ Ω R ∣ 2 ∫ R 2 g ( u ) g Λ ( u / R ) ( ρ 2 ( 0 , u ) − 1 ) d u . \mathbb{E}_{P}[F_{\Omega_R}(X_{\Omega_R})]=\frac{1}{2}\iint_{\Omega_R\times\Omega_R}g(x-y)\bigl(\rho_2(x,y)-1\bigr)\,\mathrm{d}x\,\mathrm{d}y=\frac{|\Omega_R|}{2}\int_{\mathbb{R}^2}g(u)g_\Lambda(u/R)\bigl(\rho_2(0,u)-1\bigr)\,\mathrm{d}u. E P [ F Ω R ( X Ω R )] = 2 1 ∬ Ω R × Ω R g ( x − y ) ( ρ 2 ( x , y ) − 1 ) d x d y = 2 ∣ Ω R ∣ ∫ R 2 g ( u ) g Λ ( u / R ) ( ρ 2 ( 0 , u ) − 1 ) d u .
all the integrals being absolutely convergent by (16) and the local integrability of ρ 2 \rho_2 ρ 2 . For every u u u , we have g Λ ( u / R ) → R → ∞ 1 g_\Lambda(u/R)\xrightarrow[R\to\infty]{}1 g Λ ( u / R ) R → ∞ 1 . Since 0 ≤ g Λ ≤ 1 0\le g_\Lambda\le1 0 ≤ g Λ ≤ 1 , the integrand is bounded by ∣ log ∣ u ∣ ∣ ∣ ρ 2 ( 0 , u ) − 1 ∣ |\log|u||\,|\rho_2(0,u)-1| ∣ log ∣ u ∣∣ ∣ ρ 2 ( 0 , u ) − 1∣ , and we obtain (i) by dominated convergence.
Proof of (ii). By (16) and the local integrability of ρ 2 \rho_2 ρ 2 , the function ρ 2 ( 0 , ⋅ ) − 1 \rho_2(0,\cdot)-1 ρ 2 ( 0 , ⋅ ) − 1 is integrable on R 2 \mathbb{R}^2 R 2 . By (6) , with Γ P = δ 0 + ( ρ 2 ( 0 , u ) − 1 ) d u \Gamma_{\mathsf{P}}=\delta_0+(\rho_2(0,u)-1)\,\mathrm{d}u Γ P = δ 0 + ( ρ 2 ( 0 , u ) − 1 ) d u ,
∣ Ω R ∣ − 1 Var P [ N Ω R ] = 1 + ∫ R 2 g Λ ( u / R ) ( ρ 2 ( 0 , u ) − 1 ) d u . |\Omega_R|^{-1}\operatorname{Var}_{\mathsf{P}}[N_{\Omega_R}]=1+\int_{\mathbb{R}^2}g_{\Lambda}(u/R)(\rho_2(0,u)-1)\,\mathrm{d}u. ∣ Ω R ∣ − 1 Var P [ N Ω R ] = 1 + ∫ R 2 g Λ ( u / R ) ( ρ 2 ( 0 , u ) − 1 ) d u .
For the disk, the left-hand side tends to 0 0 0 by hyperuniformity, and the right-hand side to 1 + ∫ ( ρ 2 ( 0 , u ) − 1 ) d u 1+\int(\rho_2(0,u)-1)\,\mathrm{d}u 1 + ∫ ( ρ 2 ( 0 , u ) − 1 ) d u by dominated convergence. Hence ∫ ( ρ 2 ( 0 , u ) − 1 ) d u = − 1 \int(\rho_2(0,u)-1)\,\mathrm{d}u=-1 ∫ ( ρ 2 ( 0 , u ) − 1 ) d u = − 1 , and, for both shapes,
∣ Ω R ∣ − 1 Var P [ N Ω R ] = ∫ R 2 ( g Λ ( u / R ) − 1 ) ( ρ 2 ( 0 , u ) − 1 ) d u . |\Omega_R|^{-1}\operatorname{Var}_{\mathsf{P}}[N_{\Omega_R}]=\int_{\mathbb{R}^2}(g_{\Lambda}(u/R)-1)(\rho_2(0,u)-1)\,\mathrm{d}u. ∣ Ω R ∣ − 1 Var P [ N Ω R ] = ∫ R 2 ( g Λ ( u / R ) − 1 ) ( ρ 2 ( 0 , u ) − 1 ) d u .
Since g Λ g_{\Lambda} g Λ is Lipschitz and 0 ≤ g Λ ≤ 1 = g Λ ( 0 ) 0\leq g_{\Lambda}\leq1=g_{\Lambda}(0) 0 ≤ g Λ ≤ 1 = g Λ ( 0 ) , the right-hand side is at most
C R ∫ ∣ u ∣ ≤ R ∣ u ∣ ∣ ρ 2 ( 0 , u ) − 1 ∣ d u + ∫ ∣ u ∣ > R ∣ ρ 2 ( 0 , u ) − 1 ∣ d u . \frac{C}{R}\int_{|u|\leq\sqrt{R}}|u||\rho_2(0,u)-1|\,\mathrm{d}u+\int_{|u|>\sqrt{R}}|\rho_2(0,u)-1|\,\mathrm{d}u. R C ∫ ∣ u ∣ ≤ R ∣ u ∣∣ ρ 2 ( 0 , u ) − 1∣ d u + ∫ ∣ u ∣ > R ∣ ρ 2 ( 0 , u ) − 1∣ d u .
The first term is at most C R − 1 / 2 ∫ ∣ ρ 2 ( 0 , u ) − 1 ∣ d u CR^{-1/2}\int|\rho_2(0,u)-1|\,\mathrm{d}u C R − 1/2 ∫ ∣ ρ 2 ( 0 , u ) − 1∣ d u . The second is at most 2 log R ∫ ∣ u ∣ > R log ∣ u ∣ ∣ ρ 2 ( 0 , u ) − 1 ∣ d u \frac{2}{\log R}\int_{|u|>\sqrt{R}}\log|u||\rho_2(0,u)-1|\,\mathrm{d}u l o g R 2 ∫ ∣ u ∣ > R log ∣ u ∣∣ ρ 2 ( 0 , u ) − 1∣ d u , which is o ( 1 / log R ) o(1/\log R) o ( 1/ log R ) by (16) . This gives (ii). The same computation is done in [48 Proposition 1.8] for disks. ▫
Proof of Proposition 6.3 (a grand-canonical-like estimate)
Proof. For a finite configuration X X X in B R B_R B R , the potential created by the points of X X X and the background is
φ X ( y ) : = ∫ g ( y − x ) d X ( x ) − U ( y ) for y ∈ R 2 . (118) \varphi_X(y):=\int g(y-x)\,\mathrm{d}X(x)-U(y)\qquad\text{for }y\in\mathbb{R}^2. \tag*{(118)} φ X ( y ) := ∫ g ( y − x ) d X ( x ) − U ( y ) for y ∈ R 2 . (118)
Under the neutral gas, let ( x 1 , … , x ∣ B R ∣ ) (x_1,\ldots,x_{|B_R|}) ( x 1 , … , x ∣ B R ∣ ) be distributed according to G ∣ B R ∣ \mathbb{G}_{|B_R|} G ∣ B R ∣ , let X : = ∑ i δ x i X:=\sum_i\delta_{x_i} X := ∑ i δ x i , and let y y y and y ~ \widetilde{y} y be independent and uniform in B R B_R B R , independent of X X X . Let F p p : = ∑ r < s g ( x r − x s ) F_{\mathrm{pp}}:=\sum_{r<s}g(x_r-x_s) F pp := ∑ r < s g ( x r − x s ) be the first term of (52) . We define the following four quantities:
m : = E [ φ X ( y ) ] m:=\mathbb{E}[\varphi_X(y)] m := E [ φ X ( y )]
(potential at a random point),
m ~ : = E [ φ X − δ x 1 ( x 1 ) ] \widetilde{m}:=\mathbb{E}[\varphi_{X-\delta_{x_1}}(x_1)] m := E [ φ X − δ x 1 ( x 1 )]
(potential felt by a particle),
g ‾ R : = E [ g ( y − y ~ ) ] \overline{g}_R:=\mathbb{E}[g(y-\widetilde{y})] g R := E [ g ( y − y )]
(interaction between two random points),
g R ( 2 ) : = E [ g ( x 1 − x 2 ) ] g_R^{(2)}:=\mathbb{E}[g(x_1-x_2)] g R ( 2 ) := E [ g ( x 1 − x 2 )]
(interaction between two particles).
Step 1. The Gibbs variational principle.
Claim B.1. For every n ≥ 0 n\geq0 n ≥ 0 and every probability measure μ \mu μ on B R n B_R^n B R n ,
log K n ≥ − β E μ [ F ] − Ent ( μ ∣ Bin n ) , \log K_n\geq-\beta\mathbb{E}_{\mu}[F]-\operatorname{Ent}(\mu\mid\operatorname{Bin}_n), log K n ≥ − β E μ [ F ] − Ent ( μ ∣ Bin n ) , with equality for μ = G n \mu=\mathbb{G}_n μ = G n .
Proof. By (53) , 0 ≤ Ent ( μ ∣ G n ) = Ent ( μ ∣ Bin n ) + β E μ [ F ] + log K n 0\leq\operatorname{Ent}(\mu\mid\mathbb{G}_n)=\operatorname{Ent}(\mu\mid\operatorname{Bin}_n)+\beta\mathbb{E}_{\mu}[F]+\log K_n 0 ≤ Ent ( μ ∣ G n ) = Ent ( μ ∣ Bin n ) + β E μ [ F ] + log K n . ▫
Step 2. Inserting and removing points.
Claim B.2 (Inserting and removing points). For k ≥ 0 k\geq0 k ≥ 0 and 0 ≤ j ≤ ∣ B R ∣ 0\leq j\leq|B_R| 0 ≤ j ≤ ∣ B R ∣ ,
log K ∣ B R ∣ + k ≥ log K ∣ B R ∣ − β k m − β ( k 2 ) g ‾ R , \log K_{|B_R|+k}\geq\log K_{|B_R|}-\beta k m-\beta\binom{k}{2}\overline{g}_R, log K ∣ B R ∣ + k ≥ log K ∣ B R ∣ − β k m − β ( 2 k ) g R , log K ∣ B R ∣ − j ≥ log K ∣ B R ∣ + β j m ~ − β ( j 2 ) g R ( 2 ) . \log K_{|B_R|-j}\geq\log K_{|B_R|}+\beta j\widetilde{m}-\beta\binom{j}{2}g_R^{(2)}. log K ∣ B R ∣ − j ≥ log K ∣ B R ∣ + β j m − β ( 2 j ) g R ( 2 ) .
Proof.
Insertion. Let μ : = G ∣ B R ∣ ⊗ Bin k \mu:= \mathbb{G}_{|B_R|} \otimes\operatorname{Bin}_k μ := G ∣ B R ∣ ⊗ Bin k : the first ∣ B R ∣ |B_R| ∣ B R ∣ points are distributed according to the neutral gas, and k k k points are added, independently and uniformly. Then Ent ( μ ∣ Bin ∣ B R ∣ + k ) = Ent ( G ∣ B R ∣ ∣ Bin ∣ B R ∣ ) \operatorname{Ent}(\mu\mid\operatorname{Bin}_{|B_R|+k}) = \operatorname{Ent}(\mathbb{G}_{|B_R|} \mid\operatorname{Bin}_{|B_R|}) Ent ( μ ∣ Bin ∣ B R ∣ + k ) = Ent ( G ∣ B R ∣ ∣ Bin ∣ B R ∣ ) . Writing y 1 , … , y k y_1,\ldots,y_k y 1 , … , y k for the added points,
F ( X + δ y 1 + ⋯ + δ y k ) = F ( X ) + ∑ p ≤ k φ X ( y p ) + ∑ p < q ≤ k g ( y p − y q ) . F(X+\delta_{y_1}+\cdots+\delta_{y_k})=F(X)+\sum_{p\leq k}\varphi_X(y_p)+\sum_{p<q\leq k}g(y_p-y_q). F ( X + δ y 1 + ⋯ + δ y k ) = F ( X ) + p ≤ k ∑ φ X ( y p ) + p < q ≤ k ∑ g ( y p − y q ) .
whose expectation under μ \mu μ is E G ∣ B R ∣ [ F ] + k m ‾ + ( k 2 ) g ‾ R \mathbb{E}_{\mathbb{G}_{|B_R|}}[F]+k\overline{m}+\binom{k}{2}\overline{g}_R E G ∣ B R ∣ [ F ] + k m + ( 2 k ) g R . We apply Claim B.1 to μ \mu μ , and use its equality case for G ∣ B R ∣ \mathbb{G}_{|B_R|} G ∣ B R ∣ . This gives the first inequality.
Removal. Let μ \mu μ be the law of ( x j + 1 , … , x ∣ B R ∣ ) (x_{j+1},\ldots,x_{|B_R|}) ( x j + 1 , … , x ∣ B R ∣ ) when ( x 1 , … , x ∣ B R ∣ ) ∼ G ∣ B R ∣ (x_1,\ldots,x_{|B_R|})\sim\mathbb{G}_{|B_R|} ( x 1 , … , x ∣ B R ∣ ) ∼ G ∣ B R ∣ . The relative entropy decreases when taking marginals, hence Ent ( μ ∣ Bin ∣ B R ∣ − j ) ≤ Ent ( G ∣ B R ∣ ∣ Bin ∣ B R ∣ ) \operatorname{Ent}(\mu\mid\operatorname{Bin}_{|B_R|-j})\leq\operatorname{Ent}(\mathbb{G}_{|B_R|}\mid\operatorname{Bin}_{|B_R|}) Ent ( μ ∣ Bin ∣ B R ∣ − j ) ≤ Ent ( G ∣ B R ∣ ∣ Bin ∣ B R ∣ ) . Write X ~ \widetilde{X} X for the remaining configuration. Separating the interactions of the removed points,
F ( X ) = F ( X ~ ) + ∑ r ≤ j φ X − δ x r ( x r ) − ∑ r < s ≤ j g ( x r − x s ) . F(X)=F(\widetilde{X})+\sum_{r\leq j}\varphi_{X-\delta_{x_r}}(x_r)-\sum_{r<s\leq j}g(x_r-x_s). F ( X ) = F ( X ) + r ≤ j ∑ φ X − δ x r ( x r ) − r < s ≤ j ∑ g ( x r − x s ) .
hence E μ [ F ] = E G ∣ B R ∣ [ F ] − j m ~ + ( j 2 ) g ‾ R \mathbb{E}_{\mu}[F]=\mathbb{E}_{\mathbb{G}_{|B_R|}}[F]-j\widetilde{m}+\binom{j}{2}\overline{g}_R E μ [ F ] = E G ∣ B R ∣ [ F ] − j m + ( 2 j ) g R , by exchangeability. By Claim B.1 , we obtain the second inequality.
The potential felt by a particle.
Claim B.3 (Potential felt by a particle and potential at a random point).
m ~ − m = 2 ∣ B R ∣ E G ∣ B R ∣ [ F ] . \widetilde{m}-m=\frac{2}{|B_R|}\mathbb{E}_{\mathbb{G}_{|B_R|}}[F]. m − m = ∣ B R ∣ 2 E G ∣ B R ∣ [ F ] .
Proof. Let ( x 1 , … , x ∣ B R ∣ ) (x_1,\ldots,x_{|B_R|}) ( x 1 , … , x ∣ B R ∣ ) and X X X be as above. The potential felt by the r r r -th particle is, by (B.13),
φ r : = φ X − δ x r ( x r ) = ∑ s ≠ r g ( x r − x s ) − U ( x r ) . \varphi_r:=\varphi_{X-\delta_{x_r}}(x_r)=\sum_{s\ne r}g(x_r-x_s)-U(x_r). φ r := φ X − δ x r ( x r ) = s = r ∑ g ( x r − x s ) − U ( x r ) .
By exchangeability of G ∣ B R ∣ \mathbb{G}_{|B_R|} G ∣ B R ∣ , the definition {(xref)}(B.15) of m ~ \widetilde{m} m reads ∣ B R ∣ m ~ = E [ ∑ r φ r ] |B_R|\widetilde{m}=\mathbb{E}[\sum_r\varphi_r] ∣ B R ∣ m = E [ ∑ r φ r ] . Summing over r r r and using (52) ,
∑ r = 1 ∣ B R ∣ φ r = 2 F p p − ∑ r = 1 ∣ B R ∣ U ( x r ) = 2 F + ∑ r = 1 ∣ B R ∣ U ( x r ) − 2 E b g , \sum_{r=1}^{|B_R|}\varphi_r=2F_{\mathrm{pp}}-\sum_{r=1}^{|B_R|}U(x_r)=2F+\sum_{r=1}^{|B_R|}U(x_r)-2E_{\mathrm{bg}}, r = 1 ∑ ∣ B R ∣ φ r = 2 F pp − r = 1 ∑ ∣ B R ∣ U ( x r ) = 2 F + r = 1 ∑ ∣ B R ∣ U ( x r ) − 2 E bg ,
and thus
∣ B R ∣ m ~ = 2 E [ F ] + E [ ∑ r = 1 ∣ B R ∣ U ( x r ) ] − 2 E b g . |B_R|\widetilde{m}=2\mathbb{E}[F]+\mathbb{E}\left[\sum_{r=1}^{|B_R|}U(x_r)\right]-2E_{\mathrm{bg}}. ∣ B R ∣ m = 2 E [ F ] + E r = 1 ∑ ∣ B R ∣ U ( x r ) − 2 E bg .
On the other hand, by (B.13), ∫ B R φ X ( y ) d y = ∑ r U ( x r ) − ∫ B R U = ∑ r U ( x r ) − 2 E b g \int_{B_R}\varphi_X(y)\,\mathrm{d}y=\sum_rU(x_r)-\int_{B_R}U=\sum_rU(x_r)-2E_{\mathrm{bg}} ∫ B R φ X ( y ) d y = ∑ r U ( x r ) − ∫ B R U = ∑ r U ( x r ) − 2 E bg . We average over y y y first in the definition (B.14) of m m m , and obtain
∣ B R ∣ m = E [ ∑ r = 1 ∣ B R ∣ U ( x r ) ] − 2 E b g . |B_R|m=\mathbb{E}\left[\sum_{r=1}^{|B_R|}U(x_r)\right]-2E_{\mathrm{bg}}. ∣ B R ∣ m = E r = 1 ∑ ∣ B R ∣ U ( x r ) − 2 E bg .
Subtracting the last two displays, ∣ B R ∣ ( m ~ − m ) = 2 E [ F ] |B_R|(\widetilde{m}-m)=2\mathbb{E}[F] ∣ B R ∣ ( m − m ) = 2 E [ F ] .
This is the electrostatic fact that the energy is half the sum of the potentials felt by the charges. Taken separately, m m m and m ~ \widetilde{m} m are not small, but only their difference will appear.
Step 4. Crude bounds.
Claim B.4 (Crude bounds). The following holds:
(i) g ˉ R = − log R + 1 4 < 0 \bar{g}_{R}=-\log R+\frac{1}{4}<0 g ˉ R = − log R + 4 1 < 0 ;
(ii) − π 2 ∣ B R ∣ ≤ E G ∣ B R ∣ [ F ] ≤ ∣ B R ∣ 2 ( log R − 1 4 ) -\frac{\pi}{2}|B_{R}|\leq\mathbb{E}_{G|B_{R}|}[\mathrm{F}]\leq\frac{|B_{R}|}{2}\left(\log R-\frac{1}{4}\right) − 2 π ∣ B R ∣ ≤ E G ∣ B R ∣ [ F ] ≤ 2 ∣ B R ∣ ( log R − 4 1 ) ;
(iii) g R ( 2 ) ≤ − log R + 3 2 g_{R}^{(2)}\leq-\log R+\frac{3}{2} g R ( 2 ) ≤ − log R + 2 3 .
Proof.
Proof of (i). We have g ˉ R = − log R + E [ g ( y − y ~ ) ] \bar{g}_{R}=-\log R+\mathbb{E}[g(y-\tilde{y})] g ˉ R = − log R + E [ g ( y − y ~ )] for y , y ~ y,\tilde{y} y , y ~ uniform in the unit disk, and this expectation is 1 4 \frac{1}{4} 4 1 .
Proof of (ii). We start with the upper bound. By Claim B.1 with μ = Bin ∣ B R ∣ \mu=\operatorname{Bin}_{|B_{R}|} μ = Bin ∣ B R ∣ , and its equality case, we have
β E G ∣ B R ∣ [ F ] ≤ β E Bin ∣ B R ∣ [ F ] − Ent ( G ∣ B R ∣ ∣ Bin ∣ B R ∣ ) ≤ β E Bin ∣ B R ∣ [ F ] . \beta\mathbb{E}_{G|B_{R}|}[\mathrm{F}]\leq\beta\mathbb{E}_{\operatorname{Bin}_{|B_{R}|}}[\mathrm{F}]-\operatorname{Ent}(G|B_{R}|\mid\operatorname{Bin}_{|B_{R}|})\leq\beta\mathbb{E}_{\operatorname{Bin}_{|B_{R}|}}[\mathrm{F}]. β E G ∣ B R ∣ [ F ] ≤ β E Bin ∣ B R ∣ [ F ] − Ent ( G ∣ B R ∣ ∣ Bin ∣ B R ∣ ) ≤ β E Bin ∣ B R ∣ [ F ] .
We compute E Bin n [ F ] = g ˉ R ( ( n − ∣ B R ∣ ) 2 2 − n 2 ) \mathbb{E}_{\operatorname{Bin}_{n}}[\mathrm{F}]=\bar{g}_{R}\left(\frac{(n-|B_{R}|)^{2}}{2}-\frac{n}{2}\right) E Bin n [ F ] = g ˉ R ( 2 ( n − ∣ B R ∣ ) 2 − 2 n ) , which is − ∣ B R ∣ 2 g ˉ R -\frac{|B_{R}|}{2}\bar{g}_{R} − 2 ∣ B R ∣ g ˉ R for n = ∣ B R ∣ n=|B_{R}| n = ∣ B R ∣ . With (i), this gives the upper bound.
For the lower bound, the inequality F ( X ) ≥ − π 2 ∣ B R ∣ \mathrm{F}(X)\geq-\frac{\pi}{2}|B_{R}| F ( X ) ≥ − 2 π ∣ B R ∣ for every configuration X X X of ∣ B R ∣ |B_{R}| ∣ B R ∣ points in B R B_{R} B R is the classical lower bound on the energy of the jellium, see for example [64 ] and [79 ] .
Proof of (iii). By (52) , E F p p = E F + E ∑ i U ( x i ) − E b g \mathbb{E}\mathrm{F}_{\mathrm{pp}}=\mathbb{E}\mathrm{F}+\mathbb{E}\sum_{i}U(x_{i})-E_{\mathrm{bg}} E F pp = E F + E ∑ i U ( x i ) − E bg , with U ≤ − ∣ B R ∣ log R + ∣ B R ∣ 2 U\leq-|B_{R}|\log R+\frac{|B_{R}|}{2} U ≤ − ∣ B R ∣ log R + 2 ∣ B R ∣ on B R B_{R} B R (by the expression of U U U above) and E b g = ∣ B R ∣ 2 2 g ˉ R E_{\mathrm{bg}}=\frac{|B_{R}|^{2}}{2}\bar{g}_{R} E bg = 2 ∣ B R ∣ 2 g ˉ R . With the upper bound on F \mathrm{F} F , this gives E F p p ≤ − ∣ B R ∣ ( ∣ B R ∣ − 1 ) 2 log R + 3 8 ∣ B R ∣ 2 \mathbb{E}\mathrm{F}_{\mathrm{pp}}\leq-\frac{|B_{R}|(|B_{R}|-1)}{2}\log R+\frac{3}{8}|B_{R}|^{2} E F pp ≤ − 2 ∣ B R ∣ ( ∣ B R ∣ − 1 ) log R + 8 3 ∣ B R ∣ 2 , and g R ( 2 ) = 2 ∣ B R ∣ ( ∣ B R ∣ − 1 ) E F p p ≤ − log R + 3 4 ∣ B R ∣ ∣ B R ∣ − 1 ≤ − log R + 3 2 g_{R}^{(2)}=\frac{2}{|B_{R}|(|B_{R}|-1)}\mathbb{E}\mathrm{F}_{\mathrm{pp}}\leq-\log R+\frac{3}{4}\frac{|B_{R}|}{|B_{R}|-1}\leq-\log R+\frac{3}{2} g R ( 2 ) = ∣ B R ∣ ( ∣ B R ∣ − 1 ) 2 E F pp ≤ − log R + 4 3 ∣ B R ∣ − 1 ∣ B R ∣ ≤ − log R + 2 3 .
Step 5. Conclusion. We apply Claim B.2 with k = ( Dis ) + k=(\operatorname{Dis})_{+} k = ( Dis ) + on { Dis ≥ 0 } \{\operatorname{Dis}\geq0\} { Dis ≥ 0 } and j = ( Dis ) − j=(\operatorname{Dis})_{-} j = ( Dis ) − on { Dis < 0 } \{\operatorname{Dis}<0\} { Dis < 0 } , and take the expectation. By Claim B.4 , g ˉ R < 0 \bar{g}_{R}<0 g ˉ R < 0 and g R ( 2 ) ≤ 3 2 g_{R}^{(2)}\leq\frac{3}{2} g R ( 2 ) ≤ 2 3 , hence the terms of second order are bounded below by − 3 2 β Dis 2 -\frac{3}{2}\beta\operatorname{Dis}^{2} − 2 3 β Dis 2 : they can only help. (With the kernel − log -\log − log , the self-energy of an excess charge k k k spread over B R B_{R} B R is negative , of order − k 2 2 log R -\frac{k^{2}}{2}\log R − 2 k 2 log R .) The terms of first order give
− β m E [ ( Dis ) + ] + β m ~ E [ ( Dis ) − ] = β ( m ~ − m ) E [ ( Dis ) + ] , -\beta m\,\mathbb{E}[(\operatorname{Dis})_{+}]+\beta\tilde{m}\,\mathbb{E}[(\operatorname{Dis})_{-}]=\beta(\tilde{m}-m)\,\mathbb{E}[(\operatorname{Dis})_{+}], − β m E [( Dis ) + ] + β m ~ E [( Dis ) − ] = β ( m ~ − m ) E [( Dis ) + ] ,
because E Dis = 0 \mathbb{E}\operatorname{Dis}=0 E Dis = 0 , which means that the positive and negative parts of Dis \operatorname{Dis} Dis have the same mean. This is a chemical potential, and it cancels. By Claims B.3 and B.4, ∣ m ~ − m ∣ ≤ C log R |\tilde{m}-m|\leq C\log R ∣ m ~ − m ∣ ≤ C log R , and E [ ( Dis ) + ] ≤ E ∣ Dis ∣ \mathbb{E}[(\operatorname{Dis})_{+}]\leq\mathbb{E}|\operatorname{Dis}| E [( Dis ) + ] ≤ E ∣ Dis ∣ . We only need a lower bound on the partition functions, and the Gibbs variational principle gives one for every trial measure. We use no penalization of the number of points and no rescaling of non-neutral gases.
Proof of Lemma 6.4 (the geometric boundary kernel)
Proof. Radiality, continuity and the scaling J R ( u ) = R 2 J 1 ( u / R ) \mathcal{J}_{R}(u)=R^{2}\mathcal{J}_{1}(u/R) J R ( u ) = R 2 J 1 ( u / R ) follow from the definition (55) . If ∣ u ∣ ≥ 2 R |u|\geq2R ∣ u ∣ ≥ 2 R , the whole disk B R B_{R} B R is in the domain of integration, and, by the mean value property of the harmonic function x ↦ log ∣ x + u ∣ x\mapsto\log|x+u| x ↦ log ∣ x + u ∣ on B R B_{R} B R , we have ∫ B R log ∣ x + u ∣ d x = ∣ B R ∣ log ∣ u ∣ \int_{B_{R}}\log|x+u|\,dx=|B_{R}|\log|u| ∫ B R log ∣ x + u ∣ d x = ∣ B R ∣ log ∣ u ∣ , hence J R ( u ) = 0 \mathcal{J}_{R}(u)=0 J R ( u ) = 0 .
Write r = ∣ u ∣ r=|u| r = ∣ u ∣ . For 0 < r < 2 0<r<2 0 < r < 2 , the set of x ∈ B 1 x\in B_{1} x ∈ B 1 with ∣ x + u ∣ ≥ 1 |x+u|\geq1 ∣ x + u ∣ ≥ 1 is the unit disk minus a lens, of area a ( r ) = π − 2 arccos ( r / 2 ) + r 2 4 − r 2 a(r)=\pi-2\arccos(r/2)+\frac{r}{2}\sqrt{4-r^{2}} a ( r ) = π − 2 arccos ( r /2 ) + 2 r 4 − r 2 , and J 1 ( u ) = b ( u ) − a ( r ) log r \mathcal{J}_{1}(u)=b(u)-a(r)\log r J 1 ( u ) = b ( u ) − a ( r ) log r , with b ( u ) : = ∫ B 1 log + ∣ x + u ∣ d x b(u):=\int_{B_{1}}\log_{+}|x+u|\,dx b ( u ) := ∫ B 1 log + ∣ x + u ∣ d x . We have ∂ r a ( r ) = 4 − r 2 \partial_{r}a(r)=\sqrt{4-r^{2}} ∂ r a ( r ) = 4 − r 2 . The Laplacian of log + ∣ ⋅ ∣ \log_{+}|\cdot| log + ∣ ⋅ ∣ is the arc-length measure of the unit circle, hence Δ b ( u ) \Delta b(u) Δ b ( u ) is the length of the arc of the unit circle contained in B ( u , 1 ) B(u,1) B ( u , 1 ) , namely 2 arccos ( r / 2 ) 2\arccos(r/2) 2 arccos ( r /2 ) . The radial Laplacian of a ( r ) log r a(r)\log r a ( r ) log r is ( ∂ r 2 a + ∂ r a / r ) log r + 2 ∂ r a / r (\partial_{r}^{2}a+\partial_{r}a/r)\log r+2\partial_{r}a/r ( ∂ r 2 a + ∂ r a / r ) log r + 2 ∂ r a / r . Hence
Δ J 1 ( u ) = 2 arccos ( r / 2 ) − 4 − 2 r 2 r 4 − r 2 log r − 2 4 − r 2 r for 0 < r < 2 , \Delta\mathcal{J}_{1}(u)=2\arccos(r/2)-\frac{4-2r^{2}}{r\sqrt{4-r^{2}}}\log r-\frac{2\sqrt{4-r^{2}}}{r}\qquad\text{for }0<r<2, Δ J 1 ( u ) = 2 arccos ( r /2 ) − r 4 − r 2 4 − 2 r 2 log r − r 2 4 − r 2 for 0 < r < 2 ,
and Δ J 1 = 0 \Delta\mathcal{J}_1=0 Δ J 1 = 0 for r > 2 r>2 r > 2 . There is no singular part at 0 0 0 , since r ∂ r J 1 → r → 0 0 r\partial_r\mathcal{J}_1 \xrightarrow[r\to0]{} 0 r ∂ r J 1 r → 0 0 , nor on the circle r = 2 r=2 r = 2 , since J 1 \mathcal{J}_1 J 1 and ∂ r J 1 \partial_r\mathcal{J}_1 ∂ r J 1 vanish there. The singularities of Δ J 1 \Delta\mathcal{J}_1 Δ J 1 are of order r − 1 ( 1 + ∣ log r ∣ ) r^{-1}(1+|\log r|) r − 1 ( 1 + ∣ log r ∣ ) at the origin and ( 2 − r ) − 1 / 2 (2-r)^{-1/2} ( 2 − r ) − 1/2 inside the circle r = 2 r=2 r = 2 , hence integrable. Its gradient is of order r − 2 ( 1 + ∣ log r ∣ ) r^{-2}(1+|\log r|) r − 2 ( 1 + ∣ log r ∣ ) and ( 2 − r ) − 3 / 2 (2-r)^{-3/2} ( 2 − r ) − 3/2 there. Separating the neighbourhoods of size ∣ h ∣ |h| ∣ h ∣ of these two sets from their complement, we get, for 0 < ∣ h ∣ ≤ 1 / 2 0<|h|\le1/2 0 < ∣ h ∣ ≤ 1/2 ,
∥ Δ J 1 ( ⋅ − h ) − Δ J 1 ∥ 1 ≤ C ( ∣ h ∣ log 2 ( 2 / ∣ h ∣ ) + ∣ h ∣ 1 / 2 ) . \|\Delta\mathcal{J}_1(\cdot-h)-\Delta\mathcal{J}_1\|_1\le C\left(|h|\log^2(2/|h|)+|h|^{1/2}\right). ∥Δ J 1 ( ⋅ − h ) − Δ J 1 ∥ 1 ≤ C ( ∣ h ∣ log 2 ( 2/∣ h ∣ ) + ∣ h ∣ 1/2 ) .
For ∣ k ∣ ≥ 2 π |k|\ge2\pi ∣ k ∣ ≥ 2 π , let h : = π k / ∣ k ∣ 2 h:=\pi k/|k|^2 h := π k /∣ k ∣ 2 , so that e − i k ⋅ h = − 1 e^{-ik\cdot h}=-1 e − ik ⋅ h = − 1 and ∣ h ∣ ≤ 1 / 2 |h|\le1/2 ∣ h ∣ ≤ 1/2 . Then
Δ J ^ 1 ( k ) = 1 2 ∫ ( Δ J 1 ( x ) − Δ J 1 ( x − h ) ) e − i k ⋅ x d x , \widehat{\Delta\mathcal{J}}_1(k)=\frac{1}{2}\int\left(\Delta\mathcal{J}_1(x)-\Delta\mathcal{J}_1(x-h)\right)e^{-ik\cdot x}\,dx, Δ J 1 ( k ) = 2 1 ∫ ( Δ J 1 ( x ) − Δ J 1 ( x − h ) ) e − ik ⋅ x d x ,
and the previous bound gives ∣ Δ J ^ 1 ( k ) ∣ ≤ C ∣ k ∣ − 1 / 2 |\widehat{\Delta\mathcal{J}}_1(k)|\le C|k|^{-1/2} ∣ Δ J 1 ( k ) ∣ ≤ C ∣ k ∣ − 1/2 . For ∣ k ∣ ≤ 2 π |k|\le2\pi ∣ k ∣ ≤ 2 π , we use the L 1 L^1 L 1 norm.
Proof of Lemmas 8.1, 8.3, 8.7 and 8.9 (estimates on the Debye scale)
Lemma 8.1: Poincaré inequality and moments
Proof of Lemma 8.1.
Step 1. Poincaré inequality and decay of covariances. The density {(xref)}(8.1) is that of G N , L 1 / 4 G_{N,L}^{1/4} G N , L 1/4 without the constraint ∣ y j ∣ < L |y_j|<L ∣ y j ∣ < L , with the additional radial potential 2 t β g 1 2t\beta g_1 2 tβ g 1 at each particle. The confinement contributes 1 / 4 1/4 1/4 to the Levi matrix of the shifted weight. Since Δ g 1 = − 2 π ψ \Delta g_1=-2\pi\psi Δ g 1 = − 2 π ψ , the additional potential contributes 1 4 Δ ( 2 t β g 1 ) = − π t β ψ ≥ − π β ∥ ψ ∥ ∞ \frac{1}{4}\Delta(2t\beta g_1)=-\pi t\beta\psi\ge-\pi\beta\|\psi\|_\infty 4 1 Δ ( 2 tβ g 1 ) = − π tβ ψ ≥ − π β ∥ ψ ∥ ∞ . For β ≤ ( 8 π ∥ ψ ∥ ∞ ) − 1 \beta\le(8\pi\|\psi\|_\infty)^{-1} β ≤ ( 8 π ∥ ψ ∥ ∞ ) − 1 , the Levi matrix is therefore at least 1 / 8 1/8 1/8 . The weight remains invariant under permutations and rotations, so we obtain the first bound of {(xref)}(8.2) as in the proof of Proposition {(xref)}4.17. We apply it to the powers of ∣ F − E t F ∣ |F-\mathbb{E}_tF| ∣ F − E t F ∣ and obtain the second bound. The additional term of the Levi matrix is diagonal, and it commutes with the multiplication by the weight of Section 5. Therefore, repeating the proof of {(xref)}(5.8), we obtain {(xref)}(8.3), with a fixed c > 0 c>0 c > 0 .
Step 2. Number of points and correlation functions. In the coordinates x j = ℓ D y j x_j=\ell_Dy_j x j = ℓ D y j , the additional potential is 2 t β g 1 ( x / ℓ D ) 2t\beta g_1(x/\ell_D) 2 tβ g 1 ( x / ℓ D ) , which is superharmonic. The exponential moment bound of Thoma [87 Theorem 2] therefore gives the first bound of {(xref)}(8.4): the methods of [86 , 87 ] are stable under perturbations of the potential which are superharmonic in each variable (see the remark after [87 Theorem 5] , a result of [86 ] ). We then apply the isotropic averaging bound on the correlation functions [87 Proposition 3.1] on balls of fixed radius on the Debye scale, and the first bound. This gives the bound on ρ j , t \rho_{j,t} ρ j , t .
Step 3. Moments. For F = Fluct t [ f ] F=\operatorname{Fluct}_t[f] F = Fluct t [ f ] , we have ∑ i ∣ ∇ i F ∣ 2 = Y ( ∣ ∇ f ∣ 2 ) \sum_i|\nabla_iF|^2=Y(|\nabla f|^2) ∑ i ∣ ∇ i F ∣ 2 = Y ( ∣∇ f ∣ 2 ) . We expand the integer moments of Y ( ∣ ∇ f ∣ 2 ) Y(|\nabla f|^2) Y ( ∣∇ f ∣ 2 ) into factorial moments, and use the bound on the correlation functions in {(xref)}(8.4). This gives
∥ Y ( ∣ ∇ f ∣ 2 ) ∥ p / 2 ≤ C ( p ) ( ℓ D 2 ∥ ∇ f ∥ 2 2 + ℓ D 4 / p ∥ ∇ f ∥ p 2 ) . \|Y(|\nabla f|^2)\|_{p/2}\le C(p)\left(\ell_D^2\|\nabla f\|_2^2+\ell_D^{4/p}\|\nabla f\|_p^2\right). ∥ Y ( ∣∇ f ∣ 2 ) ∥ p /2 ≤ C ( p ) ( ℓ D 2 ∥∇ f ∥ 2 2 + ℓ D 4/ p ∥∇ f ∥ p 2 ) .
With the second bound of {(xref)}(8.2), we then obtain {(xref)}(8.5). By truncating the values of the observables, we extend these inequalities from bounded observables to local polynomials.
Step 4. The bounds under Gin β \operatorname{Gin}_\beta Gin β . The process Gin β \operatorname{Gin}_\beta Gin β is the limit of the hard-wall gas P N , β \mathbb{P}_{N,\beta} P N , β , not of P 0 \mathbb{P}_0 P 0 . We prove the bounds for the hard-wall gas, on the Debye scale and at macroscopic distance from the wall, and pass to the limit. Let Y N Y_N Y N be the image of X N X_N X N under x ↦ x / ℓ D x\mapsto x/\ell_D x ↦ x / ℓ D , with ( x 1 , … , x N ) (x_1,\ldots,x_N) ( x 1 , … , x N ) distributed according to P N , β = G N , R N π β / 2 \mathbb{P}_{N,\beta}=\mathbb{G}_{N,R_N}^{\pi\beta/2} P N , β = G N , R N π β /2 . Let F F F have the form {(xref)}(2.5), with ψ \psi ψ of smooth and of polynomial growth together with its first derivatives. By Proposition {(xref)}4.17, in these coordinates,
Var P N , β [ F ( Y N ) ] ≤ 4 E P N , β ∑ y ∈ Y N ∣ ∇ y F ( Y N ) ∣ 2 . \operatorname{Var}_{\mathbb{P}_{N,\beta}}[F(Y_N)]\le4\mathbb{E}_{\mathbb{P}_{N,\beta}}\sum_{y\in Y_N}|\nabla_yF(Y_N)|^2. Var P N , β [ F ( Y N )] ≤ 4 E P N , β y ∈ Y N ∑ ∣ ∇ y F ( Y N ) ∣ 2 .
Both F 2 F^2 F 2 and ∑ y ∈ Y ∣ ∇ y F ∣ 2 \sum_{y\in Y}|\nabla_yF|^2 ∑ y ∈ Y ∣ ∇ y F ∣ 2 are local and continuous, with polynomial growth in a local count. By Corollary 3 and part (iii) of Proposition {(xref)}2.4, we can therefore pass to the limit in this inequality, and it holds under Gin β \operatorname{Gin}_{\beta} Gin β , at fixed β \beta β . We apply it to smooth approximations of ∣ F − E Gin β F ∣ p / 2 |F-\mathbb{E}_{\operatorname{Gin}_{\beta}}F|^{p/2} ∣ F − E Gin β F ∣ p /2 , and obtain the second inequality of (69) for every p ≥ 2 p \ge2 p ≥ 2 . By Proposition 5.1, with ℓ = 2 ℓ D \ell=2\ell_D ℓ = 2 ℓ D , and after rescaling, we obtain (70) , first for cylinder observables and then for local polynomials by truncation.
The results of Thoma used in Step 2 apply in the same way to the hard-wall gas, at macroscopic distance from the wall. Let ρ ~ j ( N ) \widetilde{\rho}^{(N)}_j ρ j ( N ) be the j j j -point correlation function of Y N Y_N Y N . For every compact set K ⊂ R 2 K \subset\mathbb{R}^2 K ⊂ R 2 , every j ≥ 1 j \ge1 j ≥ 1 , and N N N large enough, we obtain ρ ~ j ( N ) ≤ C ( j ) ℓ D 2 j \widetilde{\rho}^{(N)}_j \le C(j)\ell_D^{2j} ρ j ( N ) ≤ C ( j ) ℓ D 2 j on K j K^j K j , uniformly in β \beta β small. For a non-negative continuous function φ \varphi φ with compact support on ( R 2 ) j (\mathbb{R}^2)^j ( R 2 ) j , the integral of φ \varphi φ against ρ ~ j ( N ) \widetilde{\rho}^{(N)}_j ρ j ( N ) is the expectation of the sum of φ \varphi φ over the j j j -tuples of distinct points of Y N Y_N Y N . This is a local continuous observable, with polynomial growth in a local count. By Corollary 3 and part (iii) of Proposition 2.4 , we can pass to the limit, and the j j j -point correlation function of Y Y Y under Gin β \operatorname{Gin}_{\beta} Gin β is at most C ( j ) ℓ D 2 j C(j)\ell_D^{2j} C ( j ) ℓ D 2 j . This proves the correlation bound of (71) under Gin β \operatorname{Gin}_{\beta} Gin β . With the factorial-moment calculation of Step 3, we then obtain (72) under Gin β \operatorname{Gin}_{\beta} Gin β for compactly supported f f f .
For a function f f f that is not compactly supported, we apply (72) to ( χ R − χ R 1 ) f (\chi_R-\chi_{R_1})f ( χ R − χ R 1 ) f . Under the stated assumptions, the centred statistics of χ R f \chi_R f χ R f form a Cauchy sequence in L p ( Gin β ) L^p(\operatorname{Gin}_{\beta}) L p ( Gin β ) . Finally, a smooth compactly supported kernel of two variables is the limit, together with its derivatives, of finite sums of products f ( x ) h ( z ) f(x)h(z) f ( x ) h ( z ) . By the correlation bounds, the associated quadratic statistics and their particle gradients converge in every finite L p L^p L p . Both inequalities of (69) therefore hold for these quadratic statistics.
Lemma 8.3: the quadratic statistic
Proof of Lemma 8.3.
Step 1. The near part. The kernel of Q f n e a r Q_f^{\mathrm{near}} Q f near is bounded, and it vanishes unless both points are at distance at most 2 δ 2\delta 2 δ from the support of f f f . We expand the square of its sum over pairs. The terms with four distinct points, with three distinct points and with a repeated pair are integrals over sets of volume at most C δ 4 C\delta^4 C δ 4 , C δ 4 C\delta^4 C δ 4 and C δ 2 C\delta^2 C δ 2 , respectively. By (71) under Gin β \operatorname{Gin}_{\beta} Gin β , the correlation function of j j j points is at most C ( j ) ℓ D 2 j C(j)\ell_D^{2j} C ( j ) ℓ D 2 j . With the prefactor ℓ D − 4 \ell_D^{-4} ℓ D − 4 of the square, we obtain the bound C ( ℓ D 4 δ 4 + δ 2 ) C(\ell_D^4\delta^4+\delta^2) C ( ℓ D 4 δ 4 + δ 2 ) on the second moment. The compensating terms of (76) satisfy the same bound, and this proves the first inequality of (77) .
Step 2. The far part. Let Υ ( x , z ) : = 1 2 ( ∇ f ( x ) − ∇ f ( z ) ) ⋅ ∇ g ( x − z ) ( 1 − χ δ ( x − z ) ) \Upsilon(x,z):=\frac{1}{2}(\nabla f(x)-\nabla f(z))\cdot\nabla g(x-z)(1-\chi_\delta(x-z)) Υ ( x , z ) := 2 1 ( ∇ f ( x ) − ∇ f ( z )) ⋅ ∇ g ( x − z ) ( 1 − χ δ ( x − z )) be the kernel of Q f f a r Q_f^{\mathrm{far}} Q f far . Since Υ \Upsilon Υ vanishes near the diagonal, the derivative of Q f f a r Q_f^{\mathrm{far}} Q f far in the position y i y_i y i of a point of Y Y Y is given, first with the cutoffs χ R \chi_R χ R and then in the limit R → ∞ R\to\infty R → ∞ , by
∇ y i Q f f a r = 2 ℓ D − 2 ( Υ ( ∇ x Υ ( y i , ⋅ ) ) − ℓ D 2 ∫ R 2 ∇ x Υ ( y i , z ) d z ) . \nabla_{y_i}Q_f^{\mathrm{far}}=2\ell_D^{-2}\left(\Upsilon(\nabla_x\Upsilon(y_i,\mathord{\cdot}))-\ell_D^2\int_{\mathbb{R}^2}\nabla_x\Upsilon(y_i,z)\,\mathrm{d}z\right). ∇ y i Q f far = 2 ℓ D − 2 ( Υ ( ∇ x Υ ( y i , ⋅ )) − ℓ D 2 ∫ R 2 ∇ x Υ ( y i , z ) d z ) .
We tile the plane by squares of side δ \delta δ . For each square S S S , we get E Gin β Y ( S ) 2 ≤ ℓ D 2 δ 2 + C ℓ D 4 δ 4 \mathbb{E}_{\operatorname{Gin}_{\beta}}Y(S)^2\le\ell_D^2\delta^2+C\ell_D^4\delta^4 E Gin β Y ( S ) 2 ≤ ℓ D 2 δ 2 + C ℓ D 4 δ 4 from the intensity and the two-point bound of (71) . Since δ ≥ ℓ D − 1 \delta\ge\ell_D^{-1} δ ≥ ℓ D − 1 , it follows that ∥ Y ( S ) ∥ 2 ≤ C ℓ D δ \|Y(S)\|_2\le C\ell_D\delta ∥ Y ( S ) ∥ 2 ≤ C ℓ D δ . On each square, we use the Hölder inequality, the embedding H 2 ↪ L ∞ H^2\hookrightarrow L^\infty H 2 ↪ L ∞ at scale δ \delta δ in the first variable of ∇ x Υ \nabla_x\Upsilon ∇ x Υ , and (72) with p = 4 p=4 p = 4 in its second variable. For every multi-index α \alpha α with ∣ α ∣ ≤ 2 |\alpha|\le2 ∣ α ∣ ≤ 2 , let T α : = ∇ z α ∇ x Υ T_\alpha:=\nabla_z^\alpha\nabla_x\Upsilon T α := ∇ z α ∇ x Υ . With the Poincaré inequality (69) , we then bound Var Gin β [ Q f f a r ] \operatorname{Var}_{\operatorname{Gin}_{\beta}}[Q_f^{\mathrm{far}}] Var Gin β [ Q f far ] by
C ∑ ∣ α ∣ ≤ 2 δ 2 ∣ α ∣ ∫ R 2 ( ∥ T α ( x , ⋅ ) ∥ 2 2 + ℓ D − 1 ∥ T α ( x , ⋅ ) ∥ 4 2 ) d x . (119) C\sum_{|\alpha|\le2}\delta^{2|\alpha|}\int_{\mathbb{R}^2}\left(\|T_\alpha(x,\mathord{\cdot})\|_2^2+\ell_D^{-1}\|T_\alpha(x,\mathord{\cdot})\|_4^2\right)\,\mathrm{d}x. \tag*{(119)} C ∣ α ∣ ≤ 2 ∑ δ 2∣ α ∣ ∫ R 2 ( ∥ T α ( x , ⋅ ) ∥ 2 2 + ℓ D − 1 ∥ T α ( x , ⋅ ) ∥ 4 2 ) d x . (119)
For ∣ x − z ∣ ≤ 1 |x-z|\le1 ∣ x − z ∣ ≤ 1 , we have ∣ T α ( x , z ) ∣ ≤ C ∣ x − z ∣ − 2 − ∣ α ∣ |T_\alpha(x,z)|\le C|x-z|^{-2-|\alpha|} ∣ T α ( x , z ) ∣ ≤ C ∣ x − z ∣ − 2 − ∣ α ∣ outside the support of χ δ ( x − z ) \chi_\delta(x-z) χ δ ( x − z ) , and ∣ T α ( x , z ) ∣ ≤ C δ − 2 − ∣ α ∣ |T_\alpha(x,z)|\le C\delta^{-2-|\alpha|} ∣ T α ( x , z ) ∣ ≤ C δ − 2 − ∣ α ∣ on it. Integrated in x x x , the squared L 2 L^2 L 2 and L 4 L^4 L 4 norms of T α ( x , ⋅ ) T_\alpha(x,\mathord{\cdot}) T α ( x , ⋅ ) on B ( x , 1 ) B(x,1) B ( x , 1 ) are therefore at most C δ − 2 − 2 ∣ α ∣ C\delta^{-2-2|\alpha|} C δ − 2 − 2∣ α ∣ and C δ − 3 − 2 ∣ α ∣ C\delta^{-3-2|\alpha|} C δ − 3 − 2∣ α ∣ . When ∣ x − z ∣ ≥ 1 |x-z|\ge1 ∣ x − z ∣ ≥ 1 , one of the two points lies in the support of ∇ f \nabla f ∇ f , and T α T_\alpha T α decays quadratically with the other point. Hence (B.16) is at most C ( δ − 2 + ℓ D − 1 δ − 3 ) ≤ C δ − 2 C(\delta^{-2}+\ell_D^{-1}\delta^{-3})\le C\delta^{-2} C ( δ − 2 + ℓ D − 1 δ − 3 ) ≤ C δ − 2 , which is the second inequality of (77) .
Step 3. The cutoff. For the kernel Υ ( x , z ) ( 1 − χ R ( x ) χ R ( z ) ) \Upsilon(x,z)(1-\chi_R(x)\chi_R(z)) Υ ( x , z ) ( 1 − χ R ( x ) χ R ( z )) , we bound (B.16) by C ( δ ) R − 2 C(\delta)R^{-2} C ( δ ) R − 2 with the same computation, and the derivatives of χ R \chi_R χ R satisfy the same bound. With δ = 1 \delta=1 δ = 1 , this proves the convergence of Q f , R Q_{f,R} Q f , R in L 2 L^2 L 2 and the bound on Q f − Q f , R Q_f-Q_{f,R} Q f − Q f , R .
Lemma 8.7: moments of the potential of two particles
Proof of Lemma 8.7.
Step 1. Moments. Since g − g ε = ( g − g 1 ) ( ⋅ / ε ) g-g_{\varepsilon}=(g-g_{1})(\cdot/\varepsilon) g − g ε = ( g − g 1 ) ( ⋅ / ε ) is non-negative and supported in B ε B_{\varepsilon} B ε , we have ∫ ( g − g ε ) p ≤ C ε 2 \int(g-g_{\varepsilon})^{p}\le C\varepsilon^{2} ∫ ( g − g ε ) p ≤ C ε 2 . For p ∈ { 2 , 4 , 6 } p\in\{2,4,6\} p ∈ { 2 , 4 , 6 } , we expand the p p p -th moment of Fluct t [ ( g − g ε ) ( ⋅ − x ) ] \operatorname{Fluct}_{t}[(g-g_{\varepsilon})(\cdot-x)] Fluct t [( g − g ε ) ( ⋅ − x )] into factorial moments, and use (71) . This gives (84) . With ε = ℓ D − 1 \varepsilon=\ell_{D}^{-1} ε = ℓ D − 1 , it gives ∥ Fluct t [ D r − D ~ r ] ∥ p ≤ C \|\operatorname{Fluct}_{t}[D_{r}-\widetilde D_{r}]\|_{p}\le C ∥ Fluct t [ D r − D r ] ∥ p ≤ C . Moreover,
∥ ∇ D ~ r ∥ 2 ≤ C ( 1 + ∣ log β ∣ ) 1 / 2 and ℓ D 2 / p − 1 ∥ ∇ D ~ r ∥ p ≤ C for p ∈ { 4 , 6 } . \|\nabla\widetilde D_{r}\|_{2}\le C(1+|\log\beta|)^{1/2}
\qquad\text{and}\qquad
\ell_{D}^{2/p-1}\|\nabla\widetilde D_{r}\|_{p}\le C\quad\text{for }p\in\{4,6\}. ∥∇ D r ∥ 2 ≤ C ( 1 + ∣ log β ∣ ) 1/2 and ℓ D 2/ p − 1 ∥∇ D r ∥ p ≤ C for p ∈ { 4 , 6 } .
By (72) , we obtain ∥ Fluct t [ D ~ r ] ∥ p ≤ C ℓ D ( 1 + ∣ log β ∣ ) 1 / 2 \|\operatorname{Fluct}_{t}[\widetilde D_{r}]\|_{p}\le C\ell_{D}(1+|\log\beta|)^{1/2} ∥ Fluct t [ D r ] ∥ p ≤ C ℓ D ( 1 + ∣ log β ∣ ) 1/2 for p ∈ { 2 , 4 , 6 } p\in\{2,4,6\} p ∈ { 2 , 4 , 6 } . The norm ∥ ⋅ ∥ p \|\cdot\|_{p} ∥ ⋅ ∥ p is non-decreasing in p p p , so this proves the first bound of (85) .
Step 2. The exponential bound. We first consider the smooth function w : = g 1 ( ⋅ − r e 1 ) − g 1 w:=g_{1}(\cdot-re_{1})-g_{1} w := g 1 ( ⋅ − r e 1 ) − g 1 . Let R 0 : = 4 / r 0 R_{0}:=4/r_{0} R 0 := 4/ r 0 . We split w w w into χ R 0 w \chi_{R_{0}}w χ R 0 w and the parts w R : = ( χ R − χ R / 2 ) w w_{R}:=(\chi_{R}-\chi_{R/2})w w R := ( χ R − χ R /2 ) w , for R = 2 m R 0 R=2^{m}R_{0} R = 2 m R 0 with m ≥ 1 m\ge1 m ≥ 1 . Each part w R w_{R} w R is supported in B 2 R ∖ B R / 2 B_{2R}\setminus B_{R/2} B 2 R ∖ B R /2 , with ∣ w R ∣ ≤ C / R |w_{R}|\le C/R ∣ w R ∣ ≤ C / R and ∣ ∇ w R ∣ ≤ C / R 2 |\nabla w_{R}|\le C/R^{2} ∣∇ w R ∣ ≤ C / R 2 . By the first bound of (71) with γ = 1 \gamma=1 γ = 1 , we have ∥ Y ( B 2 R ) ∥ p ≤ C ( ℓ D R 2 + p ) \|Y(B_{2R})\|_{p}\le C(\ell_{D}R^{2}+p) ∥ Y ( B 2 R ) ∥ p ≤ C ( ℓ D R 2 + p ) for every p ≥ 2 p\ge2 p ≥ 2 . Therefore, by the second bound of (69) ,
∥ R Fluct t [ w R ] ∥ p ≤ C p ( ℓ D + p R ) . \|R\operatorname{Fluct}_{t}[w_{R}]\|_{p}\le Cp\left(\ell_{D}+\frac{\sqrt{p}}{R}\right). ∥ R Fluct t [ w R ] ∥ p ≤ C p ( ℓ D + R p ) .
By Markov’s inequality, with p p p proportional to u u u , we have P t ( R ∣ Fluct t [ w R ] ∣ > ℓ D u ) ≤ C e − c u \mathbb P_{t}(R|\operatorname{Fluct}_{t}[w_{R}]|>\ell_{D}u)\le Ce^{-cu} P t ( R ∣ Fluct t [ w R ] ∣ > ℓ D u ) ≤ C e − c u for u ≤ c ℓ D 2 R 2 u\le c\ell_{D}^{2}R^{2} u ≤ c ℓ D 2 R 2 . For larger u u u , we use the bound R ∣ Fluct t [ w R ] ∣ ≤ C ( Y ( B 2 R ) + E t Y ( B 2 R ) ) R|\operatorname{Fluct}_{t}[w_{R}]|\le C(Y(B_{2R})+\mathbb E_{t}Y(B_{2R})) R ∣ Fluct t [ w R ] ∣ ≤ C ( Y ( B 2 R ) + E t Y ( B 2 R )) . With (71) , we deduce that P t ( R ∣ Fluct t [ w R ] ∣ > ℓ D u ) ≤ C e − c ℓ D u \mathbb P_{t}(R|\operatorname{Fluct}_{t}[w_{R}]|>\ell_{D}u)\le Ce^{-c\ell_{D}u} P t ( R ∣ Fluct t [ w R ] ∣ > ℓ D u ) ≤ C e − c ℓ D u for u ≥ C ℓ D R 2 u\ge C\ell_{D}R^{2} u ≥ C ℓ D R 2 . For β \beta β small, the two ranges of u u u overlap, hence E t e c R ∣ Fluct t [ w R ] ∣ / ℓ D ≤ C \mathbb E_{t}e^{cR|\operatorname{Fluct}_{t}[w_{R}]|/\ell_{D}}\le C E t e c R ∣ Fluct t [ w R ] ∣/ ℓ D ≤ C . The part χ R 0 w \chi_{R_{0}}w χ R 0 w satisfies the same bound, with R = R 0 R=R_{0} R = R 0 . Summing these bounds with the Jensen inequality (weights proportional to R − 1 R^{-1} R − 1 ), we get
E t e c ∣ Fluct t [ w ] ∣ / ℓ D ≤ C . (120) \mathbb E_{t}e^{c|\operatorname{Fluct}_{t}[w]|/\ell_{D}}\le C. \tag*{(120)} E t e c ∣ Fluct t [ w ] ∣/ ℓ D ≤ C . (120)
By (72) , the sum of the centred statistics of the parts converges to Fluct t [ w ] \operatorname{Fluct}_{t}[w] Fluct t [ w ] in every L p L^{p} L p .
Finally, D r = ( g − g 1 ) + ( g − g 1 ) ( ⋅ − r e 1 ) + w D_{r}=(g-g_{1})+(g-g_{1})(\cdot-re_{1})+w D r = ( g − g 1 ) + ( g − g 1 ) ( ⋅ − r e 1 ) + w . The first two terms are non-negative and supported in B 1 B_{1} B 1 and B ( r e 1 , 1 ) B(re_{1},1) B ( r e 1 , 1 ) , and their expectations under P t \mathbb P_{t} P t are at most C ℓ D 2 C\ell_{D}^{2} C ℓ D 2 , by (71) . Their centred statistics are therefore at least − C ℓ D 2 -C\ell_{D}^{2} − C ℓ D 2 . Multiplying by 2 β = ( π ℓ D 2 ) − 1 2\beta=(\pi\ell_{D}^{2})^{-1} 2 β = ( π ℓ D 2 ) − 1 , we obtain − 2 β Fluct t [ D r ] ≤ C − 2 β Fluct t [ w ] -2\beta\operatorname{Fluct}_{t}[D_{r}]\le C-2\beta\operatorname{Fluct}_{t}[w] − 2 β Fluct t [ D r ] ≤ C − 2 β Fluct t [ w ] . For β \beta β small, 2 β 2\beta 2 β is at most c ℓ D − 1 c\ell_{D}^{-1} c ℓ D − 1 , and we obtain the second bound of (85) from (B.17).
Lemma 8.9: third moments of linear statistics
Proof of Lemma 8.9. We decompose g 1 = ∑ j ≥ 0 φ j g_{1}=\sum_{j\ge0}\varphi_{j} g 1 = ∑ j ≥ 0 φ j , where φ 0 : = χ 1 g 1 \varphi_{0}:=\chi_{1}g_{1} φ 0 := χ 1 g 1 and φ j : = ( χ 2 j − χ 2 j − 1 ) g 1 \varphi_{j}:=(\chi_{2^{j}}-\chi_{2^{j-1}})g_{1} φ j := ( χ 2 j − χ 2 j − 1 ) g 1 for j ≥ 1 j\ge1 j ≥ 1 . The function φ 0 \varphi_{0} φ 0 is supported in B 2 B_{2} B 2 , and φ j \varphi_{j} φ j is supported in B 2 j + 1 ∖ B 2 j − 1 B_{2^{j+1}}\setminus B_{2^{j-1}} B 2 j + 1 ∖ B 2 j − 1 for $j\ge1. By (72) ,
∥ Fluct t [ φ j ] ∥ 4 + D P t ( Fluct t [ φ j ] ) 1 / 2 ≤ C ( 1 + j ) ℓ D . \|\operatorname{Fluct}_{t}[\varphi_{j}]\|_{4}+\mathcal D_{\mathbb P_{t}}\bigl(\operatorname{Fluct}_{t}[\varphi_{j}]\bigr)^{1/2}\le C(1+j)\ell_{D}. ∥ Fluct t [ φ j ] ∥ 4 + D P t ( Fluct t [ φ j ] ) 1/2 ≤ C ( 1 + j ) ℓ D .
For two functions f f f and h h h , we bound D P t ( Fluct t [ f ] Fluct t [ h ] ) 1 / 2 \mathcal D_{\mathbb P_{t}}(\operatorname{Fluct}_{t}[f]\operatorname{Fluct}_{t}[h])^{1/2} D P t ( Fluct t [ f ] Fluct t [ h ] ) 1/2 by C C C times the product of the right-hand sides of (72) for f f f and for h h h , with p = 4 p=4 p = 4 , using the product rule and the Hölder inequality.
Step 1. The first bound. By symmetry, we may assume that j ≥ k j\ge k j ≥ k . Let j j j be larger than a constant that depends on r 0 r_{0} r 0 . If j ≤ k + 2 j\le k+2 j ≤ k + 2 , then we apply (70) to Fluct t [ A ] \operatorname{Fluct}_{t}[A] Fluct t [ A ] and Fluct t [ φ j ] Fluct t [ φ k ] \operatorname{Fluct}_{t}[\varphi_{j}]\operatorname{Fluct}_{t}[\varphi_{k}] Fluct t [ φ j ] Fluct t [ φ k ] . If j ≥ k + 3 j\ge k+3 j ≥ k + 3 , then we apply it to Fluct t [ A ] Fluct t [ φ k ] \operatorname{Fluct}_{t}[A]\operatorname{Fluct}_{t}[\varphi_{k}] Fluct t [ A ] Fluct t [ φ k ] and Fluct t [ φ j ] \operatorname{Fluct}_{t}[\varphi_{j}] Fluct t [ φ j ] . In both cases, the supports of the two observables are at distance at least c 2 j c2^{j} c 2 j . For the finitely many remaining pairs, we bound the left-hand side of (B.18) by C ℓ D 3 C\ell_{D}^{3} C ℓ D 3 , using the Hölder inequality and (72) . With the bounds on the Dirichlet energies above, we obtain, for all j , k ≥ 0 j,k\ge0 j , k ≥ 0 ,
∣ E t [ Fluct t [ A ] Fluct t [ φ j ] Fluct t [ φ k ] ] ∣ ≤ C ℓ D 3 ( 1 + j ) ( 1 + k ) e − c 2 max ( j , k ) . (121) \left|\mathbb E_{t}\left[\operatorname{Fluct}_{t}[A]\operatorname{Fluct}_{t}[\varphi_{j}]\operatorname{Fluct}_{t}[\varphi_{k}]\right]\right|
\le C\ell_{D}^{3}(1+j)(1+k)e^{-c2^{\max(j,k)}}. \tag*{(121)} ∣ E t [ Fluct t [ A ] Fluct t [ φ j ] Fluct t [ φ k ] ] ∣ ≤ C ℓ D 3 ( 1 + j ) ( 1 + k ) e − c 2 m a x ( j , k ) . (121)
The double sum over j j j and k k k converges absolutely, and this proves the first bound of (89) .
Step 2. The second bound. Let R = 2 j R=2^j R = 2 j be large. The function χ R / 8 D ~ r \chi_{R/8}\widetilde{D}_r χ R /8 D r is supported in B R / 4 B_{R/4} B R /4 and equal to D ~ r \widetilde{D}_r D r in B R / 8 B_{R/8} B R /8 . Since ∣ D ~ r ( z ) ∣ ≤ C ∣ z ∣ − 1 |\widetilde{D}_r(z)|\le C|z|^{-1} ∣ D r ( z ) ∣ ≤ C ∣ z ∣ − 1 and ∣ ∇ D ~ r ( z ) ∣ ≤ C ∣ z ∣ − 2 |\nabla\widetilde{D}_r(z)|\le C|z|^{-2} ∣∇ D r ( z ) ∣ ≤ C ∣ z ∣ − 2 for large ∣ z ∣ |z| ∣ z ∣ , we get from (72)
∥ Fluct t [ ( 1 − χ R / 8 ) D ~ r ] ∥ 4 ≤ C ℓ D R − 1 . \left\|\operatorname{Fluct}_t[(1-\chi_{R/8})\widetilde{D}_r]\right\|_4\le C\ell_D R^{-1}. Fluct t [( 1 − χ R /8 ) D r ] 4 ≤ C ℓ D R − 1 .
By (70) and the bound on the Dirichlet energy of products,
∣ Cov t [ Fluct t [ χ R / 8 D ~ r ] 2 , Fluct t [ φ j ] ] ∣ ≤ C ℓ D 3 ( 1 + j ) ( 1 + ∣ log β ∣ ) e − c R . \left|\operatorname{Cov}_t\left[\operatorname{Fluct}_t[\chi_{R/8}\widetilde{D}_r]^2,\operatorname{Fluct}_t[\varphi_j]\right]\right|\le C\ell_D^3(1+j)(1+|\log\beta|)e^{-cR}. Cov t [ Fluct t [ χ R /8 D r ] 2 , Fluct t [ φ j ] ] ≤ C ℓ D 3 ( 1 + j ) ( 1 + ∣ log β ∣ ) e − c R .
By the Hölder inequality, (84) and (85) , replacing χ R / 8 D ~ r \chi_{R/8}\widetilde{D}_r χ R /8 D r by D ~ r \widetilde{D}_r D r in this covariance costs at most C ℓ D 3 ( 1 + j ) R − 1 ( 1 + ∣ log β ∣ ) 1 / 2 C\ell_D^3(1+j)R^{-1}(1+|\log\beta|)^{1/2} C ℓ D 3 ( 1 + j ) R − 1 ( 1 + ∣ log β ∣ ) 1/2 . Consequently,
∣ E t [ Fluct t [ D ~ r ] 2 Fluct t [ φ j ] ] ∣ ≤ C ℓ D 3 ( 1 + j ) 2 − j ( 1 + ∣ log β ∣ ) . (122) \left|\mathbb{E}_t\left[\operatorname{Fluct}_t[\widetilde{D}_r]^2\operatorname{Fluct}_t[\varphi_j]\right]\right|\le C\ell_D^3(1+j)2^{-j}(1+|\log\beta|). \tag*{(122)} E t [ Fluct t [ D r ] 2 Fluct t [ φ j ] ] ≤ C ℓ D 3 ( 1 + j ) 2 − j ( 1 + ∣ log β ∣ ) . (122)
By the Hölder inequality, the finitely many remaining indices j j j contribute at most C ℓ D 3 ( 1 + ∣ log β ∣ ) C\ell_D^3(1+|\log\beta|) C ℓ D 3 ( 1 + ∣ log β ∣ ) . Summing (B.19) over j j j proves the second bound of (89) .
At fixed N N N , the sums over j j j converge in every L p L^p L p because of the confinement, and the estimates are uniform in N N N because the bounds above are summable. These bounds also justify the differentiations in t t t of (87) .
Energy under localized translations
In the proof of [59 Claim 8.5] , the bound on the term κ i , 2 \kappa_{i,2} κ i , 2 of the first layer is slightly incorrect. It requires exponential moments of 2 − 2 i 2^{-2i} 2 − 2 i times the squared number of points in the support of χ i \chi_i χ i . This number of points is typically of order 2 2 i 2^{2i} 2 2 i , so the logarithm of these moments is of order 2 2 i 2^{2i} 2 2 i . We replace this step by the following lemma, which we use in the proof of part (i) of Proposition 2.4.
Lemma B.5. Let β > 0 \beta>0 β > 0 and s > 0 s>0 s > 0 . Then there exist R 0 ( β , s ) R_0(\beta,s) R 0 ( β , s ) and C ( β , s ) C(\beta,s) C ( β , s ) such that, for every R ≥ R 0 R\ge R_0 R ≥ R 0 , every N ≥ 1 N\ge1 N ≥ 1 for which the pair ( 0 , R ) (0,R) ( 0 , R ) satisfies the bulk condition [59 (3.1)] , and every smooth function φ \varphi φ on R 2 × R 2 \mathbb{R}^2\times\mathbb{R}^2 R 2 × R 2 that vanishes unless ∣ x ∣ < R |x|<R ∣ x ∣ < R and ∣ y ∣ < 1 |y|<1 ∣ y ∣ < 1 , and that satisfies
∣ φ ∣ ≤ 1 and R k ∣ ∇ x k ∂ y 1 ∂ y 2 φ ∣ ≤ 1 for 1 ≤ k ≤ 3 , |\varphi|\le1\qquad\text{and}\qquad R^k|\nabla_x^k\partial_{y_1}\partial_{y_2}\varphi|\le1\quad\text{for }1\le k\le3, ∣ φ ∣ ≤ 1 and R k ∣ ∇ x k ∂ y 1 ∂ y 2 φ ∣ ≤ 1 for 1 ≤ k ≤ 3 , we have
log E P N , β exp ( s ( 1 + N R R 2 ) sup y ∈ R 2 ∣ Fluct [ φ ( ⋅ , y ) ] ∣ ) ≤ C ( β , s ) . \log\mathbb{E}_{\mathrm{P}_{N,\beta}}\exp\left(s\left(1+\frac{N_R}{R^2}\right)\sup_{y\in\mathbb{R}^2}|\operatorname{Fluct}[\varphi(\cdot,y)]|\right)\le C(\beta,s). log E P N , β exp ( s ( 1 + R 2 N R ) y ∈ R 2 sup ∣ Fluct [ φ ( ⋅ , y )] ∣ ) ≤ C ( β , s ) .
Proof. Since φ ( x , ⋅ ) \varphi(x,\cdot) φ ( x , ⋅ ) vanishes outside the unit disk, the fundamental theorem of calculus gives, for y ∈ ( − 1 , 1 ) 2 y\in(-1,1)^2 y ∈ ( − 1 , 1 ) 2 ,
Fluct [ φ ( ⋅ , y ) ] = ∫ ( − 1 , y 1 ) × ( − 1 , y 2 ) Fluct [ ∂ y 1 ∂ y 2 φ ( ⋅ , z ) ] d z . \operatorname{Fluct}[\varphi(\cdot,y)]=\int_{(-1,y_1)\times(-1,y_2)}\operatorname{Fluct}[\partial_{y_1}\partial_{y_2}\varphi(\cdot,z)]\,\mathrm{d}z. Fluct [ φ ( ⋅ , y )] = ∫ ( − 1 , y 1 ) × ( − 1 , y 2 ) Fluct [ ∂ y 1 ∂ y 2 φ ( ⋅ , z )] d z .
For every z z z , the function ∂ y 1 ∂ y 2 φ ( ⋅ , z ) \partial_{y_1}\partial_{y_2}\varphi(\cdot,z) ∂ y 1 ∂ y 2 φ ( ⋅ , z ) vanishes outside B R B_R B R , and its derivatives of order k k k are at most R − k R^{-k} R − k for $1\le k\le3. If R R R is large enough, depending on β \beta β and s s s , then [59 Lemma 3.3] at the scale R R R , applied with both signs, gives
E P N , β exp ( 4 s ∣ Fluct [ ∂ y 1 ∂ y 2 φ ( ⋅ , z ) ] ∣ ) ≤ C ( β , s ) . \mathbb{E}_{\mathrm{P}_{N,\beta}}\exp\left(4s|\operatorname{Fluct}[\partial_{y_1}\partial_{y_2}\varphi(\cdot,z)]|\right)\le C(\beta,s). E P N , β exp ( 4 s ∣ Fluct [ ∂ y 1 ∂ y 2 φ ( ⋅ , z )] ∣ ) ≤ C ( β , s ) .
We bound the supremum over y y y by the integral over ( − 1 , 1 ) 2 (-1,1)^2 ( − 1 , 1 ) 2 , apply the Jensen inequality with the uniform measure on this square, and integrate the last bound over z z z . This gives
E P N , β exp ( s sup y ∈ R 2 ∣ Fluct [ φ ( ⋅ , y ) ] ∣ ) ≤ C ( β , s ) . (123) \mathbb{E}_{\mathrm{P}_{N,\beta}}\exp\left(s\sup_{y\in\mathbb{R}^2}|\operatorname{Fluct}[\varphi(\cdot,y)]|\right)\le C(\beta,s). \tag*{(123)} E P N , β exp ( s y ∈ R 2 sup ∣ Fluct [ φ ( ⋅ , y )] ∣ ) ≤ C ( β , s ) . (123)
We now include the factor 1 + N R / R 2 1+N_R/R^2 1 + N R / R 2 . By [59 Lemma 3.1] , there exists c ( β ) > 0 c(\beta)>0 c ( β ) > 0 such that E P N , β e c N R 2 ≤ e R 2 / c \mathbb{E}_{\mathrm{P}_{N,\beta}}e^{cN_R^2}\le e^{R^2/c} E P N , β e c N R 2 ≤ e R 2 / c . On the event { N R ≤ 2 R 2 / c } \{N_R\le2R^2/c\} { N R ≤ 2 R 2 / c } , the factor 1 + N R / R 2 1+N_R/R^2 1 + N R / R 2 is at most C ( β ) C(\beta) C ( β ) , and (B.20), with C ( β ) s C(\beta)s C ( β ) s in place of s s s , bounds the contribution of this event by C ( β , s ) C(\beta,s) C ( β , s ) . On the complementary event, we have R 2 < c N R / 2 R^2<cN_R/2 R 2 < c N R /2 , and ∣ φ ∣ ≤ 1 |\varphi|\le1 ∣ φ ∣ ≤ 1 gives ∣ Fluct [ φ ( ⋅ , y ) ] ∣ ≤ N R + π R 2 ≤ C ( β ) N R |\operatorname{Fluct}[\varphi(\cdot,y)]|\le N_R+\pi R^2\le C(\beta)N_R ∣ Fluct [ φ ( ⋅ , y )] ∣ ≤ N R + π R 2 ≤ C ( β ) N R for every y y y . The exponent in the lemma is therefore at most C ( β ) s N R 2 / R 2 C(\beta)sN_R^2/R^2 C ( β ) s N R 2 / R 2 , which is at most c N R 2 / 2 cN_R^2/2 c N R 2 /2 if R R R is large enough, depending on β \beta β and s s s . Since c N R 2 / 2 > 2 R 4 / c cN_R^2/2 > 2R^4/c c N R 2 /2 > 2 R 4 / c on this event, its contribution is at most
e − 2 R 4 / c E P N , β e c N R 2 ≤ e − R 4 / c . e^{-2R^4/c}\mathbb{E}_{P_{N,\beta}}e^{cN_R^2} \le e^{-R^4/c}. e − 2 R 4 / c E P N , β e c N R 2 ≤ e − R 4 / c .
Together with the bound on the event { N R ≤ 2 R 2 / c } \{N_R \le2R^2/c\} { N R ≤ 2 R 2 / c } , this proves the lemma.
References