Surface energy and boundary layers for a chain of atoms at low temperature

We analyze the surface energy and boundary layers for a chain of atoms at low temperature for an interaction potential of Lennard-Jones type. The pressure (stress) is assumed small but positive and bounded away from zero, while the temperature $\beta^{-1}$ goes to zero. Our main results are: (1) As $\beta \to \infty$ at fixed positive pressure $p>0$, the Gibbs measures $\mu_\beta$ and $\nu_\beta$ for infinite chains and semi-infinite chains satisfy path large deviations principles. The rate functions are bulk and surface energy functionals $\overline{\mathcal{E}}_{\mathrm{bulk}}$ and $\overline{\mathcal{E}}_\mathrm{surf}$. The minimizer of the surface functional corresponds to zero temperature boundary layers. (2) The surface correction to the Gibbs free energy converges to the zero temperature surface energy, characterized with the help of the minimum of $\overline{\mathcal{E}}_\mathrm{surf}$. (3) The bulk Gibbs measure and Gibbs free energy can be approximated by their Gaussian counterparts. (4) Bounds on the decay of correlations are provided, some of them uniform in $\beta$.


Introduction
The purpose of the present article is to analyze the low-temperature behavior for a onedimensional chain of atoms that interact via a Lennard-Jones type potential. The model is atomistic and in terms of the Gibbs measures of classical statistical mechanics. Two limiting procedures are at play: the zero-temperature limit, for which the inverse temperature β goes to infinity, and the thermodynamic limit, where the number of particles N and the system size go to infinity. The order of the limits matters. When the zero-temperature limit is taken before the N → ∞ limit, the analysis of Gibbs measures is replaced by energy minimization, leading to variational models of non-linear elasticity. We perform instead the zero-temperature limit after the thermodynamic limit. The zero-temperature limit for infinite systems is far from trivial, see [vER07,CGU11,CH10] and the discussion in [BRS10].
For the one-dimensional Lennard-Jones interaction, it is known that energy minimizers (ground states) converge to a periodic lattice [GR79] ("crystallization"). For one-dimensional systems with pair potentials that decay faster than 1/r 2 it is well-known that, in contrast, at positive temperature, no matter how small, there is no crystallization [BL15]. Nevertheless, some quantities can be approximated well by their zero-temperature counterpart. For the bulk free energy this is to be expected, for other quantities such as surface corrections this is already more subtle. For the decay of correlations, it is a priori not even clear what the zero-temperature counterpart should be; we propose a natural candidate, see Eqs (2.10) and (2.11).
At zero temperature, surface corrections and boundary layers have been studied, for example, in order to better understand variational models of fracture, see e.g. [BC07,SSZ11] and the references therein. Fracture might be expected for elongated chains, forced to stretch beyond their preferred length. At small positive temperature, large interparticle distances correspond to low pressure (stress) p = p β → 0. We address this regime in a subsequent work and focus here on the elastic regime of positive pressure p > 0, though the case of small pressure p β → 0 is discussed in some comments.
Our main results come in four parts. They are listed in Sections 2.1-2.4 and proven in Sections 3-7. At zero temperature, we extend the result on bulk periodicity from [GR79] to a more general class of potentials and positive pressure, see Theorem 2.1. We prove the existence of bounded surface corrections, and characterize them with the help of an energy functional E surf for semi-infinite chains (Theorem 2.2).
At positive temperature, we prove large deviations principles for the Gibbs measures µ β and ν β on R Z + and R N + (product topology) as β → ∞ at fixed p > 0 (Theorem 2.4). The speed is β and the respective rate functions are energy functionals E bulk and E surf − min E surf whose minimizers are, respectively, the periodic bulk ground state and the zero-temperature boundary layer. The convergence of positive-temperature surface corrections to their zero-temperature counterpart is addressed in Theorem 2.5. These results are intimately related to path large deviations for Markov processes and Hamilton-Jacobi-Bellman equations [FK06], semi-classical analysis [Hel02], and a more direct approach to low-temperature expansions [SL17]. We remark that our results are valid for long range interactions which in particular are not assumed to have superlinear growth at infinity. The large deviations principle is complemented by a result on Gaussian approximations for the bulk Gibbs measure and the Gibbs free energy, valid for finite interaction range m (Theorems 2.7 and 2.8).
Finally we study the temperature-dependence of correlations and informally discuss how correlations connect with effective interactions of defects and the decay of boundary layers. Theorem 2.9 provides a priori estimates that hold for all β, p > 0. In Theorem 2.11 we show that for finite m and small positive pressure p, the decay of correlations is exponential with a rate of decay that stays bounded as β → ∞-the associated Markov chain has a spectral gap bounded away from zero. This uniform estimate is proven with perturbation theory for the transfer operator. For infinite m, we provide instead a uniform estimate for restricted Gibbs measures (Proposition 2.10), which follows from the convexity of the energy (in a neighborhood of the periodic gound state) and techniques from the realm of Brascamp-Lieb inequalities [Hel02]. At vanishing pressure p β → 0 or fixed high pressure p > 0, the spectral gap might become exponentially small because of fracture or metastable wells [BdH15] in non-convex energy landscapes.
Bringing statistical mechanics into atomistic models of crystals and elasticity has a rich tradition [BH98,Wei02,BCF86,Pen02]. Modern developments include: the study of gradient Gibbs measures [FS97] with sophisticated tools such as renormalization groups and cluster expansions [AKM16], random walk representations [BFS82], and Witten Laplacians [Hel02]; scaling limits and gradient Young-Gibbs measures [Pre09,KL14,Run15]; the extension of approximation schemes, e.g., the quasi-continuum method, to positive temperature [BLBLP10,TM11]. In addition, there have been some inroads into the open problem of proving crystallization in the form of orientational order for two-dimensional models [Aum15,HMR14].
To the best of our knowledge, all of the aforementioned mathematical literature, notably on Gibbs gradient measures, is limited to potentials with a superlinear growth at infinity. This is in stark contrast with the decay to zero typically imposed in statistical mechanics of point particles [Rue69]. We work with potentials v(r) → 0, an additional linear term pr enters because we work in the constant pressure ensemble, which is the most convenient ensemble for one-dimensional systems [Rue69, Section 5.6.6]. As a consequence, the by now classical combination of Bakry-Emery estimates and Holley-Stroock perturbation principle, see [Men14] and the references therein, becomes potentially more delicate. We use instead estimates on energy penalties, some aspects of which might generalize to higher-dimensional models.
Another aspect that might generalize to higher dimension concerns the large deviations principle. The existence of a large deviations principle for the Gibbs measure as β → ∞, proven using a exponential tightness and fixed point equation for the measure, amounts to the construction of an infinite volume energy functional that vanishes on ground states only. In higher dimension, the role of the fixed point equation is taken by DLR-conditions named after Dobrushin, Lanford, Ruelle [Geo11] and the proof of a large deviations principle reduces to the investigation of a higher-dimensional analogue of a Bellman equation. The theory of the latter, for non-unique ground states, might mirror possible intricacies of the zero-temperature limit of Gibbs measure described in [vER07].
Finally we remark that the results of this work allow for a detailed analysis of typical atomic configurations at low temperature and low density. In [JKST19] we will in particular prove that, when the density is strictly smaller than the density of the ground state lattice, a system with N particles fills space by alternating approximately crystalline domains ("clusters") with empty domains ("cracks"). The number of domains is of the order of N exp(−βe surf /2) with e surf the surface energy from Theorem 2.2 below.

Main results
2.1. Zero temperature. Let v : (0, ∞) → R be a pair potential, m ∈ N ∪ {∞} a truncation parameter and p ≥ 0 the pressure. At zero temperature we allow for p = 0, at positive temperature we impose p > 0. The Gibbs energy at zero temperature and pressure p for a system of N particles with positions x 1 < . . . < x N and interparticle spacings z j = x j+1 − x j , j = 1, . . . , N − 1, is E N (z 1 , . . . , z N −1 ) = 1≤i<j≤N |i−j|≤m v(z i + · · · + z j−1 ) + p The parameter m restricts the range of the interaction: m = 2 corresponds to a next-nearest neighbor interaction. This section deals with the minimization problem E N = inf z1,...,zN−1>0 E N (z 1 , . . . , z N −1 ) in the limit N → ∞. Throughout we assume that the following assumption holds.
(i) Shape of v: z max is the unique minimizer of v and satisfies v(z max ) < 0. v is decreasing on (0, z max ) and increasing and non-positive on (z max , ∞). (ii) Growth of v: v(z) ≥ −α 1 z −s for all z > 0 and v(z) + v(z max ) − 2α 1 ∞ n=2 (nz) −s > 0 for all z < z min .
(iii) Shape of v ′′ : v ′′ is decreasing on [z min , z max ] and increasing and non-positive on [2z min , ∞).
(iv) Growth of v ′′ : v ′′ (z) ≥ −α 2 z −s−2 for all z > r hc and v ′′ (z max ) + ∞ n=2 n 2 v ′′ (nz min ) > 0. The assumption is satisfied, for example, by the Lennard-Jones potential v(r) = r −12 − r −6 . As we will see, parts (i) and (ii) of the assumption guarantee that energy minimizers at p = 0 have interparticle spacings z j in (z min , z max ), parts (iii) and (iv) ensure that E N is uniformly strictly convex in (z min , z max ) N −1 ; moreover the Hessian D 2 E N is diagonally dominant with positive diagonal entries and negative off-diagonal entries.
Assumption 2. The pressure p satisfies 0 ≤ p < p * with p * := |v(zmax)| zmax . At positive temperature we shall assume in addition that p > 0, r hc > 0, and for some results we need lim rցr hc v(r) = ∞. The next theorem is the adaptation of a similar result by Gardner and Radin [GR79]. It is proven in Section 3.1.
(a) For every N ≥ 2, the map E N : The limit e 0 = lim N →∞ (E N /N ) < 0 exists and is given by Let D 0 ⊂ (r hc , ∞) N be the space of sequences (z j ) j∈N with none or at most finitely many elements different from a. Define When m = ∞, h((z j ) j∈N ) is a function of the whole sequence. E surf is the Gibbs energy of a semi-infinite chain, with additive constant chosen in such a way that at spacings z j ≡ a the Gibbs energy is zero; h(z 1 , z 2 , . . .) represents the interaction of the left-most particle with everybody else.
The theorem is proven in Section 3.2. Note that −pa − ∞ k=1 kv(ka) is the surface energy for a clamped chain with all spacings equal to a and E surf encodes the effect of boundary layers. E surf is multiplied by 2 because finite chains have two ends. We note that min E surf is exactly the boundary layer energy introduced by Braides and Cicalese [BC07]; Braides and Cicalese dealt with the special case m = 2 of next-nearest neighbor interactions but more general potentials. For finite m ≥ 2, see [SS18,Theorem 4.2].
For later purpose we also define a bulk functional It is defined, a priori, on the space D + 0 of positive bi-infinite sequences (z j ) j∈Z ∈ (r hc , ∞) Z that have at most finitely many elements z j = a. Denoting the space of square summable strains D + = {(z j ) j∈Z ∈ (r hc , ∞) Z | j∈Z (z j − a) 2 < ∞}, an analysis similar to the one for the surface functional yields the following result.
2.2. Small positive temperature. Next we analyze infinite volume Gibbs measures on R N + and R Z + in the limit β → ∞. We focus on fixed positive p ∈ (0, |v(z max )|/z max ) but comment on vanishing p = p β → 0 at the end of the section. Let Q (β) N be the probability measure on R N −1 + defined by Standard arguments (see Section 4) show there is a uniquely defined probability measure ν β on the product space R N + such that for every k ∈ N, every bounded continuous test function Similarly, there is a uniquely defined probabilty measure µ β on R Z + such that for all local test functions f as above, and all sequences i N with i N → ∞ and N − i N → ∞, Moreover the measure µ β is shift-invariant and mixing. The measure µ β describes the bulk behavior of a semi-infinite chain, the measure ν β is the equilibrium measure for a semi-infinite chain and encodes the probability distribution of boundary layers.
Our first result is a large deviations principle for the equilibrium measure ν β as β → ∞. The rate function is a suitable extension of E surf : define E surf : In the same way E bulk extends to a map E bulk from R Z + to R ∪ {∞}. Both R N + and R Z + are equipped with the product topology.
The theorem is proven in Section 5.3. The large deviations principle for ν β says that for every closed set A ⊂ R N + and every open set O ⊂ R N + (product topology) (2.5) It is essential that we work in the product topology. Indeed we shall later see that ν β is mixing, therefore for every ε > 0, the measure ν β gives full mass 1 to sequences (z j ) j∈N that have infinitely many spacings |z j − a| > ε. Thus for every ball to be contrasted with the lower bound in Eq. (2.5).
Another consequence concerns the evaluation of the Gibbs energies of localized defects: suppose that because of some impurity, the energy is not E N but E N + V, where V is, say, continuous in the product topology, localized in the bulk, and bounded from below. Then by Varadhan's lemma [DZ98], as β → ∞, the effective Gibbs energy converges to the zero temperature energy of the defect, Surface energies occur as a specific type of defect, when V cancels all interactions between two half-infinite chains (see Proposition 4.9(a)), which leads to the following theorem. Define the Gibbs free energy g(β) per particle in the bulk and the surface correction g surf (β).
Theorem 2.5. Fix p ∈ (0, p * ) and m ∈ N ∪ {∞}. The limits (2.6) exist. If in addition r hc > 0 and lim rցr hc v(r) = ∞, then the bulk and surface Gibbs energy approach their zero-temperature counterparts when β → ∞: This proves that the thermodynamic limit and the zero temperature limit can be exchanged, which is non-trivial (and in fact, fails when the pressure goes to zero too fast, see below). One last consequence of Theorem 2.4 concerns the distribution of spacings and the pressuredensity (or stress-strain) relation. The Gibbs free energy and our partition functions correspond to an ensemble where the overall length of the system is not fixed, but instead may fluctuate with a law that depends on the pressure-high pressures p favor compressed states. In the thermodynamic limit N → ∞, though, the average spacing between particles becomes a well-defined quantity, given by (2.7) By the contraction principle [DZ98, Theorem 4.2.1], the distribution of z 0 under µ β satisfies a large deviations principle with good rate function w(z) = inf{E bulk ((z j ) j∈Z ) | (z j ) j∈Z ∈ R Z + , z 0 = z}. The unique minimizer of w(z) is the ground state spacing a. Lemma 5.1 implies that the distribution of spacings has exponential tails for some β-independent constant C.
In particular, for large β, we have ℓ(β) < a 0 where a 0 is the minimizer of the zero-stress Cauchy-Born energy density k v(kr). Conversely, spacings ℓ(β) > a 0 (elongated chains) imply vanishing pressure p = p β → 0. This is clearly apparent for nearest neighbor interactions (m = 1, Takahashi nearest neighbor gas [Tak42,LM66]), for which (2.8) Comments on vanishing pressure. We add a superscript to indicate that zero-temperature quantities are evaluated at p = 0. When p = p β → 0 slower than any exponential, it is still true that g(β) → e 0 0 . When βp β = exp(−βν) with ν > 0, one can show with [JKM15,Jan12] that lim β→∞ g(β) = min(e 0 0 , −ν). (2.9) At pressures vanishing faster than exp(−β|e 0 0 |), the most likely configurations have very large spacings (dilute gas phase, ℓ(β) → ∞) and the previous results no longer apply. For lim inf 1 β log(βp β ) > e 0 0 , we expect that large deviations principles with rate functions E 0 bulk and E 0 surf − min E 0 surf still hold (in fact our proofs still show weak large deviations principles). However rate functions have non-compact level sets and exponential tightness is lost. Moreover large spacings may contribute to the average (2.7) and Corollary 2.6 need no longer be true, thus allowing for spacings ℓ(β) → ℓ > a 0 .
2.3. Gaussian approximation. Here we complement the large deviations result by a Gaussian approximation. This section deals with finite m and the bulk measure µ β only. Remember d = m − 1. We will see that the Hessian of E bulk at (. . . , a, a, . . .) is associated with a positivedefinite, bounded operator H in ℓ 2 (Z). It is represented by a doubly-infinite matrix (H ij ) i,j∈Z that is diagonally dominant. Write (H −1 ) ij for the matrix elements of the inverse operator and let µ Gauss be the uniquely defined measure on R Z , equipped with the product topology and its associated Borel σ-algebra, such that for all i, j ∈ Z, and every finite-dimensional marginal of µ Gauss is a multi-dimensional Gaussian distribution. Equivalently, µ Gauss is the distribution of a Gaussian process (N j ) j∈Z with mean zero and covariance E[N i N j ] = (H −1 ) ij . More concrete expressions for the probability density functions of nd-dimensional marginals of µ Gauss are provided in Proposition 6.17 below.
In the following we identify the measure µ β on R Z + with the measure 1l R Z + µ β on R Z . We exclude the trivial case m = 1.
Theorem 2.7. Assume 2 ≤ m < ∞, p ∈ (0, p * ), and r hc > 0. Then for every n ∈ N, the n-dimensional marginals of µ β and µ Gauss have probability density functions ρ It follows that the distribution of the spacings, suitably rescaled, converges locally to the Gaussian measure µ Gauss : for every bounded function f : R Z → R that depends on finitely many spacings z j only (bounded cylinder functions), we have For example, in the limit β → ∞, the distribution of a single spacing z j is approximately normal, with mean a and variance β −1 (H −1 ) ii . We expect that Theorem 2.7 stays true for m = ∞ but a proof or disproof is beyond the scope of this article. The next theorem says that the Gibbs free energy is close to the Gibbs free energy of the approximate Gaussian model.
Theorem 2.8. Assume 2 ≤ m < ∞, p ∈ (0, p * ), and r hc > 0. The Gibbs free energy satisfies, as β → ∞, The matrix C is introduced in Eq. (6.18), see also Lemma 6.7, it is a function of the Hessian of the energy.
Remark (Gaussian approximation and semi-classical expansions). If v is smooth and p > 0 is fixed, the Gibbs energy should admit an asymptotic expansion of form to arbitrarily high order n, for some c > 0 and coefficients a j ∈ R. The first correction comes from a Gaussian approximation of the partition function (harmonic crystal ), see Section 6, with the constant c capturing the asymptotic behavior of the determinant of the Hessian around the energy minimum. Higher order corrections correspond to anharmonic effects. A similar expansion holds for g surf (β). Rigorous results for finite m are derived with semi-classical analysis [Hel02, Møl01,BM03] which build on the analogy with the → 0 limit from quantum mechanics. For m = 2 and potentials with superlinear growth at infinity, independent results are given in [SL17].
2.4. Decay of correlations. Suppose that two defects change the energy functional from E bulk to E bulk + V 0 + V k , where we assume for simplicity that V 0 and V k depend on z 0 and z k alone. For large k, we may expect that the Gibbs energies are approximately additive, i.e., should be small when the defects are far apart. I (β) eff (k) represents an effective interaction between the defects. In the study of systems with many defects it is important to understand how fast the effective interaction decreases at large distances. Some intuition is gained from the zerotemperature counterpart however in general the limits β, k → ∞ cannot be interchanged and a full study of (2.10) for large k requires techniques beyond variational calculus. A closely related problem is about the localization of changes induced by a defect: at zero temperature, if (z j ) j∈Z is a minimizer of E bulk + V 0 , how fast does z k converge to the ground state spacing a as k → ±∞? On a similar note, how fast does z k → a for a minimizer of the surface energy E surf (decay of boundary layers)? At positive temperature, the question is about the speed of convergence, for test functions f : , so that f n+i = f i • τ n when τ denotes the left shift on R Z + . These questions naturally lead to the investigation of the decay of correlations. We start with a general result which holds for all β, p > 0.
Theorem 2.9. Assume m ∈ N ∪ {∞} and p > 0. There exist c, C > 0 such that for all β, p > 0, k ∈ N, and bounded f, g : R k + → R, When m is finite and k = m − 1, we have the stronger bound The theorem is proven in Section 4.2. When m is finite, it implies exponential decay of correlations as n → ∞, however the rate − log(1 − e −cβ ) can be exponentially small for large β. When m is infinite, Theorem 2.9 implies algebraic decay of correlations: for q = ⌊n ε ⌋ and sufficiently large n, (1 − e −cβ ) q is negligible compared to β(q/n) s−2 and we find that as n → ∞ (2.12) Better bounds are available for restricted Gibbs measures. Letμ (N ) β be the measure Q (β) N conditioned on [z min , z max ] N −1 andμ β the probability measure on [z min , z max ] Z obtained from the thermodynamic limit ofμ (N ) β . Proposition 2.10. Let m ∈ N ∪ {∞}. There exists c > 0 such that for all β, p > 0, smooth f, g : R + → R, and i = j, Remark. When m is finite, the uniform algebraic decay for the restricted Gibbs measure is replaced with uniform exponential decay exp(−γ|j − i|) with β-independent γ > 0.
The proposition is proven in Section 7. It follows from the uniform convexity of the energy (Lemma 3.3) and known results from the realm of Brascamp-Lieb, Poincaré and Log-Sobolev inequalities. Proposition 2.10 differs from the estimate (2.12) in two ways: there is no exponentially large prefactor exp(cβ), and the rate of algebraic decay is 1/n s instead of 1/n s−2 . Exponentially large prefactors are absent because the energy landscape has no local minimum. The improved algebraic decay 1/n s arises, roughly, because the Gibbs measure is comparable to a Gaussian measure whose covariance is the inverse of the energy's Hessian near the minimum, and instead of the tails of v(r), it is the tails of v ′′ (r) that count.
We suspect that for large β and small pressure, these improvements should carry over to the full Gibbs measure µ β , but we have proofs for interactions involving finitely many neighbors only.
The theorem is proven in Section 6 with perturbation theory for compact integral operators in L 2 (R d ). When m = 2, the relevant operators are self-adjoint and spectral norms and operator norms coincide, leading to improved statements. We conclude with a few comments.
Lagrangian vs. Eulerian point of view. The theorems above formulate decay of correlations in terms of labelled spacings, which in the language of continuum mechanics is a Lagrangian viewpoint. On the other hand, in statistical mechanics of point particles it is more common to deal with unlabelled particles (Eulerian viewpoint) and correlations are between portions of space rather than labelled interparticle distances. The difference between the two approaches becomes quite clear for nearest neighbor interactions (m = 1, see Eq. (2.8)), for which the spacings are i.i.d. with probability density q β (r) proportional to exp(−β[v(r) + p β r]). Because of the independence of spacings, correlations in terms of spacings vanish, µ β (f 0 g n ) − µ β (f 0 )µ β (g n ) = 0. On the other hand, the two-point function ρ 2 (0, x) 1 studied in statistical mechanics of particles is a sum over the number of particles contained in (0, x], with q * k β the n-fold convolution of q β with itself. It is a well-known fact from renewal theory [Fel71, Chapter XI] that but in general the difference is non-zero finite for x-in fact changing q β the convergence as x → ∞ can be arbitrarily slow, even though correlations of labelled interparticle spacings vanish identically. One should keep this difference in mind when browsing the literature. Path-large deviations, non-linear semi-groups, Bellman equation. For m = 2, we may view µ β as the law of a stationary Markov chain with state space R + and transition kernel P β defined in Eq. (6.6). Theorem 2.4 is a path-large deviations result for the Markov chain. Path large deviations are often investigated with the help of non-linear semi-groups and Hamilton-Jacobi-Bellman equations [FK06]. In our context, a natural non-linear semi-group is and for sufficiently smooth f we have a convergence of the form Vanishing pressure. When βp = βp β → 0 faster than exp(−β|e 0 0 |) (see (2.9)), the Gibbs measure should no longer be comparable to a Gaussian. Instead, it should be close to the ideal gas measure, for which spacings are i.i.d. exponentially distributed with parameter βp β , and we may again expect uniform exponential decay of correlations (for finite m). When βp β → 0 at a speed comparable to exp(−β|e 0 0 |), we should instead expect an exponentially small spectral gap: the Markov chain has two metastable wells, one corresponding to the optimal spacing a and another well at infinity. The exponentially small spectral gap is associated with the fracture of the chain of atoms, in the spirit of "fracture as a phase transition" [Tru96].

Energy estimates
In this section we analyze the variational problems arising at zero temperature. Throughout the section we assume that p ∈ [0, p * ) as in Assumption 2.
Proof. Let z 1 , . . . , z N −1 > 0. If z j > z max for some j, define a new configuration by shrinking z j to z max , leaving all other spacings unchanged: z ′ i = z i for i = j and z ′ j = z max . Since z max is a strict minimizer of v and r → v(r) increases on [z max , ∞), shrinking the bonds decreases E N strictly and the original configuration could not have been a minimizer.
If some interparticle spacing is smaller than z min , we remove a particle and reattach it to one end of the chain as follows. Assume b := min(z 1 , . . . , z N −1 ) < z min and let j ∈ {1, . . . , N − 1} with z j = b. Let x 1 = 0 and x i = z 1 + · · · + z i−1 , i = 2, . . . , N be associated particle positions. Thus by Assumption 1(i). Removing the particle x j thus leads to a configuration of N atoms whose energy has decreased by at least (3.1) The last inequality holds because of Assumption 1(ii) and b < z min . We define a new configuration by attaching the removed particle to either end of the chain at a distance r = z max . Since v(z max ) + pz max < 0 by Assumption 2, this decreases E N further, so overall the new configuration has strictly smaller energy, and the original sequence of spacings cannot be a minimizer of E N .
At zero pressure, it is a well-known fact that the N -particle energy is subadditive, E N +M ≤ E N + E M . Indeed placing two N ,M -particle minimizers side by side with large mutual distance, because of v(r) → 0 at r → ∞, yields an N + M -particle configuration with energy ≤ E N + E M . Positive pressure penalizes large mutual distances between two consecutive blocks, so the construction has to be modified. Proof. Let z ∈ (r hc , ∞) N −1 and w ∈ (r hc , ∞) M−1 be minimizers of E N and E M respectively. Define y ∈ (r hc , ∞) M+N −2 by concatenating z and w. By Lemma 3.1, all spacings are in [z min , z max ]. Therefore interactions that involve bonds from both blocks are for spacings ≥ 2z min > z max , hence negative, and As a consquence, a n := E n+1 is subadditive. By Fekete's subadditive lemma, the limit e 0 = lim a n /n = lim E n /n exists and is equal to the infimum of a n /n, hence (In the terminology of statistical mechanics, the energy is stable [Rue69, Chapter 3.2].) The next lemma in particular shows that E N is uniformly convex on [z min , z max ] N −1 . For later purposes, we state and prove this on a slightly larger set.
Note that the Hessian is independent of the pressure p.
For i = j and i, j ∈ L we have j∈L z j ≥ 2z min hence v ′′ ( L z j ) ≤ 0; it follows that the offdiagonal entries of the Hessian are non-positive. Next we show that the row-sums are bounded from below by some constant Assumption 1 guarantees that η > 0 for ε > 0 sufficiently small. Thus row sums are positive, off-diagonal matrix elements non-positive, and consequently diagonal elements positive. Moreover, with C = 2 max{v ′′ (r) | r ∈ [z min , z max + ε]} the diagonal elements are bounded from above by C 2 . The proof of the lemma is then completed with the help of standard arguments, for example every eigenvalue of (∂ i ∂ j E N (z)) N1≤i.j≤N2−1 lies in a Gershgorin circle with center ∂ 2 i E N and radius j =i |∂ i ∂ j E N |. In particular, (∂ i ∂ j E N (z)) N1≤i,j≤N2−1 is an M-matrix and thus monotone.
Proof of Theorem 2.1. (a) By Lemma 3.1 minimizers lie in the compact set [z min , z max ] N −1 . On that set the Hessian of E N is positive definite because of Lemma 3.3, so E N is strictly convex and the minimzer is unique.
, with the help of Lemma 3.3 is a straightforward adaptation of the corresponding proof in [GR79] and will be omitted. By Assumption 1(ii) we even have a > z min . We remark that the proof in [GR79] also shows that max{z (N +1) This in turn implies that the convergence is in fact uniform away from a boundary layer of vanishing volume fraction.
Notice that also a < z max except for the exceptional cases in which only nearest neighbors interact, i.e. m = 1 or v(z) = 0 for z ≥ 2z max , and the pressure vanishes.
Proposition 3.4. Let m ∈ N ∪ {∞} and p ≥ 0. Then Proof. For simplicity we write down the proof for m = ∞; the proof when m ∈ N is completely analogous. Fix k ≥ 2 and ε > 0. Let n 1 , n 2 ∈ N with n 2 ≥ k and N = n 1 + n 2 + 1. Let z = (z −n1 , . . . , z n2−1 ) ∈ [z min , z max ] n1+n2 be the spacings of the N -particle ground state, labelled by j = −n 1 , . . . , n 2 − 1 rather than 1, . . . , N − 1. Choosing n 1 and n 2 large enough we may assume Since the Hessian has matrix norm uniformly bounded from above (Lemma 3.3), changing the spacings z 0 , . . . , z k−1 to a increases the energy by Cε at most thus We decompose the energy of the modified configuration as where W gathers interactions that involve bonds from two consecutive blocks. The term D N represents the interactions between the left and right blocks. It satisfies which goes to zero as k → ∞. Next we subtract N e 0 from E N and distribute it as N e 0 = n 1 e 0 + (k + 1)e 0 + (n 2 − k)e 0 over the first three sums. The middle block contributes as k → ∞. For the first block, we notice that Indeed the only missing piece are negative interactions between the left block and the right tail of a semi-infinite chain. The contribution of the right block C N is estimated in a similar way. We combine the estimates and let first n 1 , n 2 → ∞, then k → ∞, and finally ε → 0 and find lim inf For the upper bound, we take approximate minimizers of E surf and glue them together to an N -particle configuration by assigning them to the left and right boundaries, with spacings a in between. This yields an N -particle configuration with energy , and the required upper bound follows.
The right-hand side is absolutely convergent for all (z j ) j∈N ∈ D.
for some suitable m-independent constant c. In particular, when m = ∞ the sum j β j γ j is absolutely convergent. In order to show that the double sum over k and j in Eq. (3.2) is absolutely convergent, we proceed with estimates analogous to Lemma 3.3. Assume first that all spacings z j = γ j + a are larger than z min . Set sup r≥zmin |v ′′ (r)| = c 1 and note that, by Assumption 1(iii) for all k ≥ 2, sup r≥kzmin |v ′′ (r)| ≤ |v ′′ (kz min )|. Hence More generally, if (γ j ) ∈ ℓ 2 (N) ∩ (r hc − a, ∞) N , then γ j → 0 and because of a > z min , there is an i ∈ N such that z j ≥ z min for all j ≥ i. Let ε = min{|z j | | j = 1, . . . , i}. Summands with j ≥ i can be estimated as before. For j ≤ i and k ≥ i + 2, we proceed as before as well, except that we This leaves a finite sum over j ≤ i, k ≤ i + 2 and overall, the sum is absolutely convergent.
Lemma 3.6. The map D → R, (z j ) → E surf (z j ) j∈N defined by (3.2) is continuous.
Proof. Let z, z (1) , z (2) , . . . be sequences in D such that z (n) −z → 0 in ℓ 2 (N). As lim i→∞ j≥i (γ (n) j ) 2 = 0 uniformly in n, the estimates above show that for every ε > 0, we can find i ∈ N such that the sum over {(j, k) | j ≥ i or k ≥ i} contributes to E surf (γ (n) ) and E surf (γ) an amount bounded by ε. In the remaining finite sum the continuity of v(r) allows us to pass to the limit. The proof is easily concluded with an ε/3 argument.
Lemma 3.7. The restriction of E surf to D ∩ [z min , z max + ε] N is strictly convex and satisfies Proof. The proof of the convexity is similar to Lemma 3.3 and therefore omitted. For the coercivity, Next we cut and paste (z 1 , . . . , z n ) into the middle of a large ground state chain: let . A Taylor expansion of E N around the minmizer z (N ) together with Lemma 3.3 and Theorem 2.1 yields On the other hand, let C 1 = ∞ ℓ=2 ℓ|v(ℓz min )| be a bound for interactions between blocks and remember E k ≥ ke 0 by Lemma 3.2 and e 0 ≤ 0. Then We combine with Eq. (3.3) and let first k 1 , k 2 → ∞, then n → ∞, and conclude that η 2 ∞ j=1 γ 2 j ≤ E surf (z) + C 2 with the help of Lemma 3.6. This proves the coercivity in the case m = ∞. The proof for finite m is similar.
Proof. We proceed as in Section 3.1. Let (z j ) j∈N ∈ D. If one of the z j 's is larger than z max , we can define a new configuration by shrinking this spacing to z max , leaving all other configurations unchanged. This decreases E surf . If one of the z j 's is smaller than z min , let b be the smallest among them, and j ∈ N with b = z j . Then we can define a new configuration by removing a participating particle and possibly shrinking a bond, i.e., (z 1 , z 2 , . . .) → (z 1 , z 2 , . . . , z j−1 , min(z j + z j+1 , z max ), z j+2 , . . .). Since e 0 ≤ 0, just as in Lemma 3.1, we see that this decreases the energy. Repeating these steps if necessary, the initial configuration is mapped to a new one that has strictly lower energy and all spacings in [z min , z max ].
The existence of a minimizer now follows from the coercivity proven in Lemma 3.7, the compactness of [z min , z max ] N ∩ D with respect to the weak ℓ 2 -convergence (shifted by (a, a, . . .)) and the weak lower semicontinuity of E surf on that set due to Lemmas 3.6 and 3.7. The minimizer is unique because of the strict convexity from Lemma 3.7.
Proof of Proposition 2.3. In complete analogy to Lemma 3.5 we obtain for all (z j ) j∈Z ∈ D + 0 , and as in Lemma 3.6, we see that (3.4) defines a continuous map D + → R. The proof of strict convexity, even on [z min , z max + ε] Z ∩ D + for some ε > 0, is again similar to Lemma 3.3. As in Lemma 3.8 we have that E bulk has a unique minimizer in D, which lies in D ∩ [z min , z max ] N . Since a ∈ (z min , z max ] and ∂ i E bulk ((z j ) j∈Z ) = 0 for every i ∈ Z by (3.4), the minimizer of E bulk is (. . . , a, a, . . .). Clearly, E bulk (. . . , a, a, . . .) = 0. Finally, the formula connecting E bulk and E surf is clear on D + 0 and follows on D + by approximation.

A fixed point equation.
In the following we assume that v has a hard core: (3.5) Our main aim in this subsection is to obtain the following characterisation of E surf , cf. (2.4).
Proposition 3.9. Let I = E surf − min E surf . Then I is the unique lower semi-continuous solution (product topology) of the equation such that min I = 0 and I = ∞ if z j ≤ r hc for one of the z j 's.
Note that, by induction, (3.6) is equivalent to for all k ∈ N and z = (z j ) j∈N ∈ R N + . (Observe that h(z) > −∞ for all z ∈ R N + by the decay assumption on v and r hc > 0. ) We begin with a technical auxiliary result.
The new configuration satisfies for some suitable constant C that depends on r hc ,c and v only. Let z (N ) be the N -particle ground state with spacings labelled by j = −n 1 + 1, . . . , k + n 2 rather than 1, . . . , N − 1. Since E N (z (N ) ) = E N ≥ N e 0 by Lemma 3.2 and e 0 ≤ 0, we get Suppose that all spacings z j are in [z min , z max ]. We use a Taylor approximation around the minimizer z (N ) , apply Lemma 3.3 and Theorem 2.1, and obtain.
Letting k → ∞ we obtain an upper bound for the ℓ 2 -norm of (z j − a) j∈N . If there are z j with z j < z min or z j > z max , we modify the configuration z 1 , . . . , z k without increasing its energy as in the proof of Lemma 3.1 to obtain z ′ 1 , . . . , z ′ k . When we shrink bonds z j > z max to z ′ j = z max , leaving all other spacings unchanged, both z ′ j and z j are strictly larger than ε 0 so the truncated ℓ 2 -norm k j=1 min (z j − a) 2 , ε 2 0 ) is unaffected. On the other hand suppose z i = min(z j ) < z min . Then we remove the particle x i , reattach it a distance z max to the left of the k-particle block. This effects the change on the ℓ 2 -norm. Both |z i − a| and |z max − a| are larger than ε 0 , moreover min((z i−1 + z i − a) 2 , ε 2 0 ) − min((z i−1 − a) 2 , ε 2 0 ) ≤ ε 2 0 . So the truncated ℓ 2 -norm increases by at most ε 2 0 . Let n be the number of times this step has to be performed. Iterating we arrive at a configuration z ′′ and E k+1 (z ′′ ) ≤ E k+1 (z) − nδ for some δ > 0, cf. (3.1). Making ε 0 smaller if necessary we may assume ε 2 0 < δ. We combine with Eq. (3.8) forẑ ′′ and C ′′ = C − nδ and obtain k j=1 min((z j − a) 2 , ε 2 0 ) ≤ C − nδ + nε 2 0 ≤ C.
We let k → ∞ and find that the truncated ℓ 2 -norm of (z j ) j∈N is finite. It follows in particular that there are only finitely many spacings |z j − a| ≥ ε 0 , and (z j − a) j∈N is square summable. This establishes the first assertion. In order to show the convergence of the partial sums to E surf , first observe that E surf satisfies (3.7) for I = E surf . This is clear for z ∈ D 0 and follows for general z ∈ D by continuity. If z ∈ D, the sequence of shifts ((z j ) j≥k ) k∈N converges to (. . . , a, a, . . .) strongly and thus . , a, a, . . .) = E surf (z).
We have actually proven the following: for sufficiently small ε 0 > 0, suitable c 1 , c 2 > 0, and all (z j ) j∈N ∈ R N + , Proof of Proposition 3.9. Let I = E surf − min E surf . Observe that I satisfies (3.6). This is clear for z ∈ D 0 and for z / ∈ D. For the remaining z it follows from Lemma 3.6. We now show that I is lower semi-continuous with respect to pointwise convergence. Without loss we suppose that z (n) ∈ D converges to z ∈ [r hc , ∞) N pointwise with I(z (n) ) ≤c < ∞ for some constantc > 0. Passing to a subsequence (not relabelled) we may furthermore assume that lim inf n→∞ I(z (n) ) = lim n→∞ I(z (n) ). Fix an ε > 0 such that the estimate in Lemma 3.7 is satisfied. By (3.9) Passing to a further subsequence (not relabelled) and choosing N sufficiently large we may achieve that either such indices do not exist or that j n → ∞ as n → ∞. In both cases we get that z j ∈ [z min , z max + ε] for j ≥ N . In particular, z j > r hc for j ≥ N .
In the second case we define new configurationsz (n) by applying the procedure detailed in the proof of Lemma 3.8 to the tails (z (n) j ) j≥N shrinking the bonds z (n) j > z max + ε, j ≥ N , and deleting particles . In the first case we simply setz (n) = z (n) . Since j n → ∞ in the second case, we still havez (n) → z pointwise.
By (3.7) with k = N − 1 we have From the decay properties of v and z (n) j ≥ r hc > 0 it is easy to see that, for any j ∈ N, h(z (n) j , z (n) j+1 , . . .) converges to h(z j , z j+1 , . . .). Since I(z (n) ) ≤c and I ≥ 0, from Assumption 3 we also get z j > r hc for j = 1, . . . , N − 1. So In particular, I((z (n) j ) j≥N ) ≤ C and so Lemma 3.7 implies that z ∈ D andz (n) − z ⇀ 0 in ℓ 2 by coercivity and hence that lim inf Note that, as I ≡ ∞, this inequality also shows that I(a, a, . . .) < ∞.
For the reverse inequality, by choosing k large enough in (3.7) we first see that (3.10) holds true for all z ∈ D 0 . We denote by z (N ) the truncation with z (N ) j = z j for j ≤ N and z (N ) j = a for j ≥ N + 1. Since z (N ) → z pointwise and z (N ) − z → 0 in ℓ 2 as N → ∞, lower semi-continuity of I and strong continuity of E surf (see Lemma 3.6) give where we have used that z (N ) ∈ D 0 for all N .
We now restrict to the case m < ∞. Let d = m − 1. By (3.7) with k = d we have Taking the infimum over (z j ) j∈N ∈ D 0 , with fixed z 1 , . . . , z d and setting In Chapter 6 we will need the following estimate.
A simpler proof gives the following estimate that will also be needed in Chapter 6.
Proof. By continuity we may assume that z = (z j ) j∈Z ∈ D + 0 \ [z min , z max + ε] Z . If z i > z max + ε, we define z ′ = (z ′ j ) j∈Z by setting z ′ j = z j for j = i and z ′ i = z max . Then If b = min{z j : j ∈ Z} < z min . We choose the smallest i with z i = b and define z = (z ′ j ) j∈N by setting z ′ j = z j for j < i, z ′ i = min{z i + z i+1 , z max } and z ′ j = z j+1 for j > i. As in (3.13) we get This concludes the proof.

Gibbs measures for the infinite and semi-infinite chains
Here we prove the existence of ν β , µ β , g(β), g surf (β) and check that µ β is shift-invariant and mixing, hence ergodic; the results and methods are fairly standard. In addition, we provide an a priori estimate on the decay of correlations with explicit analysis of the β-dependence (Theorem 4.4) which to the best of our knowledge is new. The results from this section need only very little on the pair potential: we only use that v has a hard core and that v(r) = O(1/r s ), for large r, with s > 2. The technical assumption of a hard core frees us from superstability estimates [LP76,Rue76]. The decay of the potential ensures that the infinite volume Gibbs measure is unique, see e.g. [Geo11, Chapter 8.3] and [Pap84a,Pap84b,Kle85].
We follow the classical treatment of one-dimensional systems with transfer operators. For compactly supported pair potentials with a hard core (or, in our case, when m is chosen finite), the transfer operators are integral operators in L 2 (R m−1 + , dx) [Rue69, Chapter 5.6], see Section 6. For long-range interactions, the transfer operator (also known as Ruelle operator or Ruelle-Perron-Frobenius operator ) acts instead from the left on functions of infinitely many variables, and from the right on measures [Rue68, GMS70,Rue78]. The formalism of transfer operators keeps being developed in the context of dynamical systems and ergodic theory [Bal00b,Bal00a].
For the decay of correlations, we adapt [Pol00] to the present context of continuous unbounded spins and carefully track the β-dependence in the bounds. In Section 5.3, transfer operators will also help us investigate the large deviations behavior of the Gibbs measures; notably the eigenvalue equation from Lemma 4.1 translates into a fixed point equation for the rate function (see Lemma 5.4).
We will show in Proposition 4.9 that ν β is the measure satisfying (2.2). The non-compactness of (r hc , ∞) N forms an obstacle to the application of a Schauder-Tychonoff fixed point theorem for the map ν → L * β ν/ν(L β 1), see e.g. [Rue68, Proposition 2]. It might be possible to remove the obstacle using tightness estimates, but we prefer to follow a different route and exploit the known uniqueness of infinite volume Gibbs measures [Geo11, Chapter 8.3] instead.
The previous identities hold for all non-negative test functions f , consequentlyνγ 0 k =ν for all k ∈ N 0 andν is a Gibbs measure as well.
Define the operator We will show in Proposition 4.9 that µ β is the measure satisfying (2.3). Notice that for every bounded measurable function f that depends on right-chain variables z 1 , z 2 , . . . only, (a) µ β is shift-invariant.
(b) For all f, g : R N + → R + and all n ∈ N, we have µ β (f (g • τ n )) = µ β ((S n β f )g). The proof is standard [Rue78] and therefore omitted. The lemma can be rephrased as follows: let (Z n ) n∈Z be a stochastic process with law µ β , defined on some probability space (Ω, F , P). Then (Z n ) n∈Z is stationary, and . . a.s. Our next task is to show that the process is not only stationary but in fact ergodic and to estimate the decay of correlations.
It follows that the variation is summable, Notice that for all q ∈ N 0 , C q is independent of β and p. In fact the pressure only enters the oscillation var 0 (h). By a slight abuse of notation we identify a function f : R N + → R with the function f 1 : R Z + → R + , (z j ) j∈Z → f ((z j ) j∈N ) and write µ β (f ) instead of µ β (f 1 ). The results of this subsection hold for all p > 0.
Theorem 4.4. Let m ∈ N ∪ {∞} and p > 0. The measure µ β is mixing with respect to shifts, i.e., µ β (f (g • τ n )) → µ β (f )µ β (g) as n → ∞, for all f, g ∈ L 1 (R Z + , µ β ). Moreover for γ(β) = exp(−3βC 0 ) and all bounded f, g : R N + → R, q, n ∈ N, N ≥ qn, We prove Theorem 4.4 with Pollicott's method of conditional expectations [Pol00]. For alternative approaches, see [Sar02] and the references therein. The principal idea is the following: for n ∈ N, f ∈ L 1 (R N + , ϕ β ν β ) let Π n f be the projection onto the subspace of functions that depend on the first n coordinates only, i.e., var n (f ) = 0. In terms of the stationary process (Z n ) n∈Z with law µ β , The difference enclosed in parentheses represents a truncation error; it is made small by choosing n large. On the subspace of mean-zero functions, the truncated operator S n Π n satisfies a contraction property uniformly in n (Lemma 4.7), and (S n β Π n ) q goes to zero exponentially fast as q → ∞. Lemma 4.5. We have var q (log ϕ β ) ≤ βC q for all q ∈ N 0 and β, p > 0.
). The claim then follows from the definition (4.1) of the invariant function.
Proof of Theorem 4.4. Let f, g : R N + → R be bounded functions and q, n ∈ N, N ≥ qn. Using Eq. (4.2) and Lemmas 4.7 and 4.8, we get The explicit estimate on the decay of correlations follows. That µ β is mixing then follows from standard approximation arguments.
Proof of Theorem 2.9. The estimate for infinite m is an immediate consequence of Theorem 2.9. For finite m and n = m − 1, the truncation error in Lemma 4.8 for a function f : R n + → R actually vanishes since var n (f ) = 0 and C n = 0. The bound simplifies accordingly.
(a) The Gibbs free energy and its surface correction defined by the limits (2.6) exist and are given by (b) Eqs. (2.2) and (2.3) hold true.

Large deviations as β → ∞
Here we analyze the behavior of the bulk and surface Gibbs measures µ β and ν β and of the energies g(β) and g surf (β). The large deviations result for the surface measure ν β is a consequence of the eigenvalue equation from Lemma 4.1, exponential tightness, and the uniqueness of the solution to the fixed point equation in Proposition 3.9. Since the bulk measure is absolutely continuous with respect to the product measure of two independent half-infinite chains (Eq. (4.2) and Proposition 4.9(b)), we may go from the surface to the bulk measure with the help of Varadhan's integral lemma [DZ98,Chapter 4.3]. The asymptotic behavior of e surf (β) is based on the representation from Proposition 4.9(a).

5.1.
A tightness estimate. The following estimate will help us prove that the infinite-volume measure ν β is exponentially tight (see the proof of Lemma 5.3) which enters the proof of Theorem 2.4.
Lemma 5.1. For all β, p > 0, N ∈ N, k ∈ {1, . . . , N − 1}, and r ≥ 0, we have Proof. Fix k ∈ N and r ≥ 0. For z = (z 1 , . . . , z N −1 ) ∈ R N −1 + with z k ≥ z max + r we define a new configuration z ′ by setting z ′ k = z k − r and leaving all other spacings unchanged. This decreases the Gibbs energy by an amount at least and the proof of the lemma is easily concluded.

Gibbs free energy in the bulk.
Lemma 5.2. Let β → ∞ at fixed p. Then Proof of Lemma 5.2. The relation between g(β) and λ 0 (β) has been proven in Proposition 4.9.
Lemma 5.3. Every sequence β j → ∞ has a subsequence along which (ν βj ) j∈N satisfies a large deviations principle with speed β j and some good rate function.
Remark. If p = p β → 0, we lose exponential tightness and only know that every sequence (ν βj ) has a subsequence along which it satisfies a weak large deviations principle [DZ98, Lemma 4.1.23], which means that the upper bound in (2.5) is required to hold for compact sets rather than closed sets.
It follows that the family of measures (ν β ) β≥1 is exponentially tight, i.e., for every M > 0, we can find a compact subset K ⊂ R N + such that lim sup β→∞ 1 β log ν β (K c ) ≤ −M . R N + endowed with the product topology is separable and metrizable and therefore has a countable base. Lemma 4.1.23 in [DZ98] applies and yields the claim.
Since n was arbitrary we have shown that I ≡ ∞ on R N + \ (r hc , ∞) N . In particular, as ν β satisfies a large deviations principle on R N + with rate function I, the same large deviations principle holds on (r hc , ∞) N .
We now establish another (weak) large deviations principle on (r hc , ∞) N . Let K ⊂ (r hc , ∞) N be a (relatively) closed set and [α, b] ⊂ (r hc , ∞) a compact interval. Then (5.1) with n = 1 yields Write f β (z 1 ; K) for the inner integral. As h is bounded from below and for every fixed z 1 > r hc , (z 2 , z 3 , . . for all z 1 ∈ [α, b]. Next we note that for all (z j ) j∈N ∈ (r hc , ∞) N , z ′ 1 > r hc , and suitable C > 0, |h(z 1 , z 2 , . . .) − h(z ′ 1 , z 2 , . . .)| ≤ |v(z 1 ) − v(z ′ 1 )| + C|z 1 − z ′ 1 |. For z 1 , z ′ 1 bounded away from r hc we may exploit that the derivative of v is bounded and drop the first term, making C larger if need be. Plugging these estimates into the definition of f β (z 1 , K), we find that for some C α > 0 and all β > 0, It follows that the upper bound (5.2) is uniform on compact subsets of (r hc , ∞) and  (x 1 , x 2 (z 2 , z 3 , . . .) . This is possible since I is lower semicontinuous. With the help of (5.3) we can now deduce that (5.3) holds for V instead of [α, b] × K.) Since (r hc , ∞) N is a Polish space, the rate function in a weak large deviations principle is uniquely defined [DZ98, Chapter 4.1], hence J = I on (r hc , ∞) N . To finish the proof it remains to observe that also J = I on R N + \ (r hc , ∞) N because both I and h are equal to ∞ on that set.
Proof of Theorem 2.4. The large deviations principle for ν β with good rate function E surf −min E surf is an immediate consequence of Lemmas 5.3 and 5.4 and Proposition 3.9. As a consequence, ν − β ⊗ ν + β satisfies a deviations principle with good rate function (z j ) j∈Z → E surf (z 1 , z 2 , . . .) + E surf (z 0 , z −1 , . . .) − 2 min E surf on R Z + and on [r hc , ∞) Z , The large deviations principle for µ β thus follows from Eq. (4.2), Lemmas 4.3.4 and 4.3.6 in [DZ98], min E bulk = 0 and by Proposition 2.3, and the observation that W 0 is continuous on [r hc , ∞) Z .

Surface corrections to the Gibbs free energy.
Proof of Theorem 2.5. The statements about g(β) have already been proven in Lemma 5.2. For g surf (β), we start from the formula in Proposition 4.9(a), to which we apply Lemma 5.2, Theorem 2.4 and Varadhan's lemma. This yields

Gaussian approximation
Here we prove Theorems 2.7 and 2.8 on the Gaussian approximation to the bulk measure µ β when m is finite. We start from a standard idea, namely perturbation theory for transfer operators [Hel02], however we need to put some work into a good choice of transfer operator as the standard symmetrized choice (6.2) does not work well. This aspect is explained in more detail in Section 6.1. Throughout this section m satisfies 2 ≤ m < ∞. Remember d = m − 1.
6.1. Decomposition of the energy. Choice of transfer operator. For finite m, the treatment with transfer operators from Section 4.1 can be considerably simplified: instead of an operator that acts on functions of infinitely many variables, the transfer operator becomes an integral operator in L 2 (R d ) (L 2 space with respect to Lebesgue measure). There are several possible choices, corresponding each to an additive decomposition of the energy. Let V (z 1 , . . . , z d ) := E m (z 1 , . . . , z d ) and W (z 1 , . . . , z d ; z d+1 , . . . , z 2d ) = 1≤i≤d<j≤2d |i−j|≤d v(z i + · · · + z j ).
Let us block variables as x j = (z dj+1 , . . . , z dj+d ). Then for (z j ) j∈Z ∈ D + 0 we have with only finitely many non-zero summands. By Proposition 2.3 the sum extends to D + by continuity. The transfer operator associated with the representation (6.1) is the integral operator with kernel exp(−β[V (x)+W (x; y)]); it is clearly related to the d-th power of the transfer operator L β from Section 4.1. The analysis is simpler for a symmetrized operator with kernel which has the advantage of being Hilbert-Schmidt: The pressure term present in V (x) and V (y) ensures that T β (x, y) decays exponentially fast when |x|+|y| → ∞ so that R 2d T β (x, y) 2 dxdy < ∞.
The transfer operator T β corresponds to a rewriting of (6.1), For the analysis of the limit β → ∞, we would like to have a transfer operator that concentrates in some sense around the optimal spacings so that we may approximate it with a Gaussian operator. When m ≥ 3, unfortunately, the function (x, y) → 1 2 V (x) + W (x; y) + 1 2 V (y) need not have its minimum at (x, y) = (a, a), with a = (a, . . . , a) ∈ R d . Therefore we introduce yet another variant of the transfer operator: we look for a function H(x, y) such that and H(x, y) ≥ H(a, a) = 0, and work with the kernel By a slight abuse of notation we use the same letter for the integral operator The function H is defined as follows. Set .
the formula for H follows. Because of The roles of x and σx can be exchanged, hence u(x) − u(σx) is bounded.
6.2. Some properties of the transfer operator.
The lemma follows from Lemma 6.1, the elementary proofs are omitted. By the Krein-Rutman theorem [KR48], [Dei85,Chapter 6], the operator norm ||K β || =: Λ 0 (β) is a simple eigenvalue of K β , the associated eigenfunction φ β can be chosen strictly positive on (r hc , ∞) d , and the other eigenvalues of K β have absolute value strictly smaller than Λ 0 (β), i.e., Let Π β be the rank-one projection in L 2 (R d ) given by Then K β Π β = Λ 0 (β)Π β = Π β K β and an induction over n ∈ N shows (6.4) Since Λ 1 (β) is nothing else but the spectral radius of K β − Λ 0 (β)Π β , it follows that lim sup The spectral properties of K β are related to the Gibbs free energy and the Gibbs measure as follows.
6.3. Gaussian transfer operator. Here we introduce the Gaussian counterpart to the transfer operator K β and study its spectral properties. We start from the quadratic approximation to the bulk energy E bulk . The differentiability of E bulk in a neighborhood of the constant sequence z j ≡ a is checked in Lemma 6.11 below, for the definition of the Gaussian transfer operator we only need the infinite matrix of partial derivatives at (. . . , a, a, . . .).
In the following we block variables as x j = (z dj , . . . , z dj+d−1 ) for z = (z j ) j∈Z and ξ j = (ζ dj , . . . , ζ dj+d−1 ) for ζ = (ζ j ) j∈Z . Remember the decomposition (6.1). Set a = (a, . . . , a) ∈ R d and define the d × d matrices A := W yy (a, a) + V xx (a) + W xx (a, a), B := −W xy (a, a). (6.7) We note the following relations: The Hessian D 2 E bulk at (. . . , a, a, . . .) is a doubly infinite, band-diagonal matrix with block form  Note that Lemma 3.3 implies that D 2 E bulk (. . . , a, a, . . .) is positive definite. We look for a quadratic form Q(x, y) on R 2d that is positive-definite and satisfies One candidate choice could be Q(x, y) := 1 2 x, Ax − 2 x, By ′ + 1 2 y, Ay (x ′ , y ′ ∈ R d ), but it is not easily related to H(x, y). We make a different choice which mimicks the definition of H(x, y) and show later that this amounts to picking the Hessian of H(x, y) (see Lemma 6.12 below).
We introduce the quadratic counterparts to the functions H(x, y), w(x), and H(x, y) from Section 6.2. Remember the bulk Hessian from (6.9). Since it is positive-definite, there exist uniquely defined positive-definite matrices M ∈ R 2d×2d and N ∈ R d×d such that We will see in the proof of Lemma 6.12 that M , N and M are the Hessians of H at (a, a), w at a and H at (a, a), respectively. The relation between Q and Q(x, y) is clarified in Lemma 6.7 below. We are going to work with the kernel and the associated integral operator (G β f )(x) = R d G β (x, y)f (y)dy. In Section 6.4 we show that G β is a good approximation for K β , here we study the operator G β on its own. Clearly it is enough to understand the integral operator G with kernel G(x, y) := exp(− 1 2 Q(x, y)), since G and G β are related by the change of variables x → √ β(x − a), see Eq. (6.21) below.
Proof. First we show that M is positive semi-definite, by an argument similar to Lemma 6.2(a). Define Clearly for all (x, y) ∈ R d × R d and M is positive semi-definite. Next let (x 0 , y 0 ) ∈ R d × R d be a zero of the quadratic form associated with M . Then by (6.13), the function y → F (x 0 , y) must be minimal at y = y 0 , hence ∇ y F (x 0 , y) = 0. Similarly, the function y → F (x, y 0 ) must be minimal at x = x 0 , hence ∇ x F (x 0 , y 0 ) = 0. Thus (x 0 , y 0 ) is a critical point of F . But F is strictly convex because M is positive-definite, therefore the critical point (x 0 , y 0 ) is a global minimizer of F which yields (x 0 , y 0 ) = 0. It follows that M is positive-definite.
It follows from Lemma 6.4 that R 2d G(x, y) 2 dxdy < ∞, hence G is Hilbert-Schmidt with strictly positive integral kernel and Krein-Rutman theorem is applicable. So we may ask for its principal eigenvalue and eigenvector and its spectral gap. It is natural to look for a Gaussian eigenfunction.
Lemma 6.5. Let F be a positive-definite, symmetric d × d matrix. Then the following two statements are equivalent: The function x → x, F x satisfies the quadratic Bellman equation (6.14) Proof. The proof is by a straightforward completion of squares: write  Lemma 6.6. The principal eigenvalue of G is (2π) d / det C and the principal eigenfunction is exp(− 1 2 x, 1 2 N x ) (up to scalar multiples).
Proof. A close look at our definitions shows that F := 1 2 N solves (6.14) (it is positive-definite because N is). Indeed, by the definition of Q, M , we have Therefore, by Lemma 6.5, the function φ(x) = exp(− 1 4 x, N x ) is an eigenfunction of G. The matrix M 3 + F in (6.15) is equal to (C − 1 2 N ) + F = C, and we find that the principal eigenvalue of G is (2π) d / det C.
In order to identify the block C in (6.16), we introduce the quadratic analogue to the function u(x). Let A and B be the d × d matrices from (6.7) and A 1 := V xx (a) + W xx (a, a). The infinite matrix (∂ i ∂ j E surf (a, a, . . .)) i,j∈N is band-diagonal with block structure The matrix differs from the bulk Hessian (6.9) by the upper left corner A 1 : we have Proof. Clearly by a completion of squares similar to the proof of Lemma 6.5. We add W yy (a, a) to both sides, remember (6.17), and obtain the equation for C. It is easy to see that which proves (6.16). Furthermore, Let us check that the two expressions for N are indeed identical, and that σN σ = N . Combining with (6.17) and (6.19), the two expressions for N become The two expressions are indeed equal, and from the end formula and (6.8) we read off that σN σ = N . Actually which is the analogue of w(x) = u(x) + u(σx) − V (x). Now we compute M . The off-diagonal blocks of M are the same as those of M . The upper left diagonal block is Proof. Let U β : L 2 (R d ) → L 2 (R d ) be the unitary operator given by We have and the principal eigenvalue and eigenfunction of G β are obtained from those of G in Lemma 6.6 by straightforward transformations.
Remark. When m = 2, all eigenvalues and eigenfunctions of G (hence G β ) can be computed explicitly, and the eigenfunctions are expressed with Hermite polynomials. See [Hel02, Section 5.2] on the harmonic Kac operator.
6.4. Perturbation theory. Remember the unitary operator U β from (6.20) and the relation G β = β −d/2 U * β GU β . The main technical result of this section is the following.
Lemma 6.11. The mapping E bulk is C 2 in some open neighborhood in D + of the constant sequence (. . . , a, a, . . .).
For all z ∈ D + 0 the derivative of E bulk at z is given by for all ζ ∈ ℓ 2 (Z) with ζ j = 0 for all but finitely many j. So for z, z ′ ∈ D + in a neighborhood of (. . . , a, a, . . .) with a uniform constant C, the right hand side of (6.22) extends to a uniformly continuous function there. Writing for z, z ′ ∈ D + 0 , a standard approximation argument shows that indeed E bulk is C 1 in a neighborhood of (. . . , a, a, . . .) also in D + with DE bulk given by (6.22). In fact, E bulk is even C 2 on a neighborhood of (. . . , a, a, . . .) in D + and (6.23) This follows similarly as above by extending the derivative of DE bulk , where we now use that the . . , a) and so D 2 E bulk extends to a continuous mapping from a neighborhood of (. . . , a, a, . . .) to L(ℓ 2 (Z)) (the space of bounded linear operators on ℓ 2 (Z)) given by (6.23).
Next we show that M is in fact the Hessian of H.
Lemma 6.12. Assume 2 ≤ m < ∞, p ∈ [0, p * ), and r hc > 0. We have H(x, y) ≥ H(a, a) = 0 for all x, y ∈ R d + , moreover as x, y → a, The lemma leaves open whether (a, a) is the unique global minimizer of H.
Proof. The first part of the lemma has already been proven in Lemma 6.2(a). With M ∈ R 2d×2d , N ∈ R d×d as in (6.10) and (6.11) we let M as in (6.12). It remains to show that D 2 H(a, a) = M . Since, for a suitable ε > 0, E bulk is convex on D + ∩ [z min , z max + ε] Z , see (the proof of) Proposition 2.3, Lemma 3.12 shows that there is a unique function on a neighborhood of (a, a) in R d × R d with values in R −N × R N , (x, y) →z = (z − , z + ) = (z − (x, y), z + (x, y)) such that H(x, y) = E bulk (z − (x, y), x, y, z + (x, y)).
The latter identity implies D (x,y) H = D (x,y) E bulk (z − , ·, ·, z + ), Proof. Since the pair potential v is bounded from below, we have for some constant c > 0 W (x; y) ≥ −c.
In combination with Lemma 6.1 this yields the claim.
In order to estimate ||K β − G β ||, we split the configuration space into a neighborhood A ⊃ B δ (a) of a and its complement B = R d \ A and treat blocks separately. For U ⊂ R d , we write 1 U for the multiplication operator with the indicator function 1l U .
This holds true for every ε > 0, so the left-hand side converges to zero. Since operator norms are bounded by Hilbert-Schmidt norms, the lemma follows. Proof. We may view K B β = 1 B K β 1 B as an operator in L 2 (B, dx). The Krein-Rutman theorem is applicable and shows that λ = ||K B β || is a simple eigenvalue and there exists an eigenfunction ψ that is strictly positive on B ∩ (r hc , ∞) d . Because of the symmetry H(σy, σx) = H(x, y), the function ψ • σ is a left eigenfunction. Moreover for all f, g ∈ L 2 (B, dx), we have W (x i , x i+1 ).
Proof of Proposition 6.9. Let ε > 0, A ε := [z min , z max + ε] d , and B = R d \ A. The sets A and B are clearly invariant under reversals, moreover z min < a ≤ z max by Theorem 2.1(b), so a is in the interior of A and bounded away from B. Thus A and B satisfy the assumptions of Lemmas 6.14 and 6.15. By Lemma 3.11, they also satisfy the condition (6.24) from Lemma 6.16. By the triangle inequality, The first term on the right-hand side, multiplied by β d/2 , goes to zero by Lemma 6.14. For the second term, we estimate ||K β − 1 A K β 1 A || ≤ ||1 B K β 1 B || + ||1 A K β 1 B || + ||1 B K β 1 A || and conclude from Lemmas 6.15 and 6.16 that d β d/2 ||K β − 1 A K β 1 A || → 0. Bounding Hilbert-Schmidt norms, it is straightforward to check that ||β d/2 (G β − 1 A G β 1 A )|| → 0 as well, and the proof is complete.
Proof of Theorem 2.11. The theorem is an immediate consequence of Lemma 6.3(c) and Corollary 6.10.
For the proof of Theorem 2.7, we first express the marginals of µ Gauss in terms of the matrices A and B from Eq. (6.7) and the matrix C from (6.18). We group variables in blocks x j ∈ R d as usual and view µ Gauss as a measure on (R d ) Z . This proves part (b) of the lemma. The proof of (c) is similar. Part (a) follows from (b) and a relation similar to (6.26).

A Brascamp-Lieb type covariance estimate for m = ∞
Here we prove Proposition 2.10. Key to the proof is a matrix lower bound A for the Hessian of E N . For Gaussian measures with probability density proportional to exp(− β 2 z, Az ) and test functions f i = z i , g j = z j , we end up estimating the covariance C ij = ([βA] −1 ) ij . We follow [Men14], see also [OR07]. For 1 ≤ i ≤ N − 1 we also have n v ′′ (nz min ) =: ρ > 0 by Assumption 1(iv). Moreover n 2 |v ′′ (nz min )| > 0 again by Assumption 1(iv). Let A N be the (N −1)×(N −1)-matrix with diagonal ρ and off-diagonal entries −κ |j−i| ; notice that η, κ j−i , ρ do not depend on N . A N is symmetric and positive-definite.
Similar estimates apply to other r. Combining with (7.3) we find It follows that Notice that the series is convergent. The bound is plugged into the estimate (7.2) and the proposition follows by passing to the limit N → ∞.