Introduction
The success of thermodynamics in describing the macroscopic world is based on the fact that most laboratory systems achieve equilibrium states in a reasonably short time. When the microscopic laws of motion are known and trusted, time averages of observables in or about equilibrium states are described by the appropriate distributions dictated by the statistical mechanics of the corresponding microscopic problem. Such substitution is based on the ergodic hypothesis , which allows one to replace time averages of observables (to wash out microscopic fluctuations) to ensemble averages. This intuition, due to Boltzmann [1 ] for classical dynamics, and ported by Von Neumann for quantum dynamics [2 ] , served so well to understand the physical world, that it became every physicist’s intuition when approaching the study of a non-exactly solvable (non-integrable) model made of many particles. It is then surprising when the ergodic hypothesis fails and, as everything which is counterintuitive, it is worth a better look because it could imply the necessity for a paradigm shift in the theory [3 ] .
The failure modes of the ergodic hypothesis almost always involve a timescale which diverge very fast with the system size. Therefore, one can build a macroscopic timescale from a microscopic timescale and the inversion of the infinite-time and the infinite-volume limits (also known as thermodynamic limit) implicitly assumed in the ergodic hypothesis, is impossible.
Sometimes the divergence is understood as spontaneous symmetry breaking: a ferromagnet magnetises in one direction although its Hamiltonian is rotation invariant. Sometimes the broken symmetry is not apparent in the microscopic Hamiltonian, as in mean-field spin glasses [4 , 5 ] or in Anderson [6 , 7 ] and many-body localization [8 , 10 ] [9], where it is revealed only after a rewriting of the quenched average of the partition function [11 , 12 ] .
Ergodicity breaking therefore, for the purpose of our discussion, arises from the birth of a timescale for equilibration, τ e q \tau_{\mathrm{eq}} τ eq , which diverges fast, say faster than any polynomial, in the thermodynamic limit. Typically, a polynomial divergence is confined to scale-invariant critical points (critical slowing down) or is due to hydrodynamic modes (Thouless time) . Ergodicity breaking timescales are, instead, fatser than those. However, they are difficult to control quantitatively and/or analytically. For example, a divergence of the form τ e q ∼ e c N \tau_{\mathrm{eq}} \sim e^{cN} τ eq ∼ e c N can be easily mistaken, in numerical experiments, for a large polynomial τ e q ∼ N a \tau_{\mathrm{eq}} \sim N^{a} τ eq ∼ N a , with a ∼ 10 a \sim10 a ∼ 10 or for a stretched exponential τ e q ∼ e N δ \tau_{\mathrm{eq}} \sim e^{N^{\delta}} τ eq ∼ e N δ (with δ < 1 \delta< 1 δ < 1 ), although these laws might have different physical mechanisms at their origins. A time N a N^{a} N a with a ≫ 1 a \gg1 a ≫ 1 is usually due, for example, to subdiffusion [13 ] and we do not call a subdiffusive system non-ergodic.
It is then of utmost importance to have firm, analytical results on these timescales. In this paper, we plan to do exactly that, leveraging on a series of results on the breakdown of ergodicity occurring in the discrete, non-linear Schrödinger equation (DNLSE) [14 , 15 , 16 ] and on a series of results of the eigenvalues of the Laplacian operator on large-dimensional manifolds.
The paper is organized as follows: We first present the model, the approximations and the theorems characterizing the geometry of the microcanonical surface. These are known from previous works [14 , 15 , 17 ] , and are collected in Section III. The transition at ε = ε c \varepsilon= \varepsilon_{\mathrm{c}} ε = ε c is a dynamical transition [14 , 18 ] , where the manifold gets separated into N N N distinct symmetry sectors, and the junctions between them have small volume and sectional area. Traversing from one sector, where a certain charge q i q_i q i is large, to another, takes time which one can link to the inverse of the smallest, non-zero eigenvalue of the Laplacian on the manifold, call it λ 1 \lambda_1 λ 1 . λ 0 = 0 \lambda_0 = 0 λ 0 = 0 is the “trivial” equilibrium state eigenvalue.
This leads to the second part of the paper, where we estimate λ 1 \lambda_1 λ 1 from above and below. To do this we need two different techniques. Upper bounds are simpler, since they are variational in nature. We can find a set of functions on the manifold on which to calculate λ 1 \lambda_1 λ 1 and this gives a proper upper bound. The optimization in a class of problems is connected to the solution of a problem of supersymmetric quantum mechanics a la Witten [19 ] . We do this in Section IV.
Lower bounds are more complicated, since one has to find an inequality valid for all functions on the manifold . This is obtained by a completely different method which takes the name of Bochner [20 ] and Reilly [21 ] , and is explained in detail in Sections V–VII.
The model and the results
The DNLSE is a classical problem, where the Hamiltonian of a complex field is taken as
H = − κ ∑ ⟨ i , j ⟩ ∈ G ( ψ i ∗ ψ j + ψ j ∗ ψ i ) + ∑ i = 1 N ∣ ψ i ∣ 4 (1) H=-\kappa\sum_{\langle i,j\rangle\in G}\left(\psi_i^*\psi_j+\psi_j^*\psi_i\right)+\sum_{i=1}^{N}|\psi_i|^4 \tag*{(1)} H = − κ ⟨ i , j ⟩ ∈ G ∑ ( ψ i ∗ ψ j + ψ j ∗ ψ i ) + i = 1 ∑ N ∣ ψ i ∣ 4 (1)
where ⟨ i , j ⟩ ∈ G \langle i,j\rangle\in G ⟨ i , j ⟩ ∈ G is a graph over N N N sites, and the canonical Poisson relations { ψ i ∗ , ψ j } = i δ i j \{\psi_i^*,\psi_j\}=i\delta_{ij} { ψ i ∗ , ψ j } = i δ ij dictate the dynamics. Equivalently, separating ψ j = ( x j + i p j ) / 2 \psi_j=(x_j+ip_j)/\sqrt{2} ψ j = ( x j + i p j ) / 2 one could write the Hamiltonian in position-momentum variables ( x , p ) (x,p) ( x , p ) , while writing in the “charge-phase/action-angle” variables ψ j = q j e i ϕ j \psi_j=\sqrt{q_j}e^{i\phi_j} ψ j = q j e i ϕ j we connect to the problem of Josephson junction arrays.
The dynamics conserves both the energy and the total charge Q = ∑ i ∣ ψ i ∣ 2 = ∑ i q i Q=\sum_i|\psi_i|^2=\sum_iq_i Q = ∑ i ∣ ψ i ∣ 2 = ∑ i q i . Localization in the DNLSE consists in the following fact: For particular initial conditions, when the energy density E / N = ε E/N=\varepsilon E / N = ε is higher than a critical threshold ε c \varepsilon_c ε c , the energy tends to concentrate in a single oscillator for a very long time [22 , 23 ] . Numerical calculations often fail to see the decay of such soliton [18 , 24 ] , the question arises then, what dictates the timescale of such decay.
If the kinetic energy term gets substituted by a random walk on the resulting microcanonical surface M N , ε M_{N,\varepsilon} M N , ε of constant Q = ∑ i q i Q=\sum_iq_i Q = ∑ i q i and E = ε N = ∑ i ∣ ψ i ∣ 4 = ∑ i q i 2 E=\varepsilon N=\sum_i|\psi_i|^4=\sum_iq_i^2 E = εN = ∑ i ∣ ψ i ∣ 4 = ∑ i q i 2 and q i ≥ 0 q_i\geq0 q i ≥ 0 (with reflecting boundary conditions and natural measure μ \mu μ ) most of the statistical properties of the original NLSE are conserved [25 , 26 ] . In particular, the Brownian motion on the microcanonical surface shows the phenomenon of localization present in the original model [15 ] in all its qualitative and quantitative aspects, as explained in statistical mechanics terms [14 , 16 , 27 ] . Notice that, in this Brownian motion dynamics, the graph G G G disappears from the problem.
Two comments on the choice of the Brownian dynamics, which we will expand a bit in the Conclusion. The metric on M N , ε M_{N,\varepsilon} M N , ε is a modelling choice: other reversible dynamics with the same equilibrium measure share all the static properties discussed below, but not necessarily the subleading terms of the time scales (see the Conclusions). All the results of this paper concern this Brownian motion. Their transfer to the Hamiltonian dynamics of the DNLSE rests on this substitution, which is yet to be proven.
In this paper we prove that the Brownian motion equilibration timescales behave in two different ways, depending we are in the normal or condensed phase, namely ε < 2 \varepsilon< 2 ε < 2 or ε > 2 \varepsilon> 2 ε > 2 . First of all, however, let us define the equilibration time.
Let P t = e t Δ M / 2 P_t = e^{t\Delta_M/2} P t = e t Δ M /2 be the semigroup of the Brownian motion on M N , ε M_{N,\varepsilon} M N , ε with reflecting conditions. − 1 2 Δ M -\frac{1}{2}\Delta_M − 2 1 Δ M is self-adjoint and non-negative on L 2 ( μ ) L^2(\mu) L 2 ( μ ) , with spectral gap λ 1 / 2 \lambda_1/2 λ 1 /2 , where λ 1 \lambda_1 λ 1 is the first non-zero Neumann eigenvalue of − Δ M -\Delta_M − Δ M , eq. (9) below.
The relaxation (or equilibration) time of the Brownian motion on M N , ε M_{N,\varepsilon} M N , ε is
τ e q = 2 λ 1 . (2) \tau_{\mathrm{eq}} = \frac{2}{\lambda_1}. \tag*{(2)} τ eq = λ 1 2 . (2)
The ground state is the constant function ρ e q ≡ 1 \rho_{\mathrm{eq}} \equiv1 ρ eq ≡ 1 on M M M , with λ 0 = 0 \lambda_0 = 0 λ 0 = 0 .
The relaxation time is characterised by the following, which follows from the spectral theorem by expanding in eigenfunctions of − Δ M -\Delta_M − Δ M . For every f ∈ L 2 ( μ ) f \in L^2(\mu) f ∈ L 2 ( μ ) and every t ≥ 0 t \ge0 t ≥ 0 ,
∥ P t f − μ ( f ) ∥ L 2 ( μ ) ≤ e − t / τ e q ∥ f − μ ( f ) ∥ L 2 ( μ ) , (3) \left\|P_t f-\mu(f)\right\|_{L^2(\mu)} \le e^{-t/\tau_{\mathrm{eq}}}\left\|f-\mu(f)\right\|_{L^2(\mu)}, \tag*{(3)} ∥ P t f − μ ( f ) ∥ L 2 ( μ ) ≤ e − t / τ eq ∥ f − μ ( f ) ∥ L 2 ( μ ) , (3)
with equality when f f f is an eigenfunction associated with λ 1 \lambda_1 λ 1 . Equivalently, if the initial law has density ρ 0 \rho_0 ρ 0 with respect to μ \mu μ , the density ρ t \rho_t ρ t at time t t t satisfies
∥ ρ t − ρ e q ∥ L 2 ( μ ) ≤ e − t / τ e q ∥ ρ 0 − ρ e q ∥ L 2 ( μ ) . (4) \left\|\rho_t-\rho_{\mathrm{eq}}\right\|_{L^2(\mu)} \le e^{-t/\tau_{\mathrm{eq}}}\left\|\rho_0-\rho_{\mathrm{eq}}\right\|_{L^2(\mu)}. \tag*{(4)} ∥ ρ t − ρ eq ∥ L 2 ( μ ) ≤ e − t / τ eq ∥ ρ 0 − ρ eq ∥ L 2 ( μ ) . (4)
In particular, the equilibrium autocorrelation of any observable, C f ( t ) = E μ [ ( f ( q 0 ) − μ ( f ) ) ( f ( q t ) − μ ( f ) ) ] C_f(t) = \mathbb{E}_{\mu}\left[(f(q_0)-\mu(f))(f(q_t)-\mu(f))\right] C f ( t ) = E μ [ ( f ( q 0 ) − μ ( f )) ( f ( q t ) − μ ( f )) ] , obeys ∣ C f ( t ) ∣ ≤ e − t / τ e q Var μ ( f ) \left|C_f(t)\right| \le e^{-t/\tau_{\mathrm{eq}}}\operatorname{Var}_{\mu}(f) ∣ C f ( t ) ∣ ≤ e − t / τ eq Var μ ( f ) . The decay rate of C f C_f C f equals λ 1 / 2 \lambda_1/2 λ 1 /2 only if f f f has a non-zero component along the eigenspace of λ 1 \lambda_1 λ 1 , if not C f C_f C f decays faster.
The equilibration time τ e q \tau_{\mathrm{eq}} τ eq is the slowest time observed in the dynamics. Some correlation functions or operator expectations will decay faster, if they are not coupled to the slowest sector, and this is observed in numerics. However, on timescales t ≫ τ e q t \gg\tau_{\mathrm{eq}} t ≫ τ eq any signal of the initial condition will be reduced by a factor e − t / τ e q e^{-t/\tau_{\mathrm{eq}}} e − t / τ eq .
In this paper we prove that, for the Brownian motion defined before, there exists an ε c \varepsilon_c ε c (ε c = 2 \varepsilon_c = 2 ε c = 2 in our units) such that for ε < ε c \varepsilon< \varepsilon_c ε < ε c the equilibration time is τ e q = O ( 1 ) \tau_{\mathrm{eq}} = O(1) τ eq = O ( 1 ) and for ε > ε c \varepsilon> \varepsilon_c ε > ε c , τ e q = e Θ ( N ) \tau_{\mathrm{eq}} = e^{\Theta(\sqrt{N})} τ eq = e Θ ( N ) . We prove an upper and a lower bound on λ 1 \lambda_1 λ 1 in each phase: the upper bounds in Section IV, the lower bounds in § 0.6 –VII.
Without discussing the strategy, we present here the results. For ε < 2 \varepsilon< 2 ε < 2 , and N N N sufficiently large (i.e. N > 2000 / ( 2 − ε ) 2 N > 2000/(2-\varepsilon)^2 N > 2000/ ( 2 − ε ) 2 ), one has
( 2 − ε ) 2 100 ( ε − 1 ) < λ 1 ≲ λ M F , (5) \frac{(2-\varepsilon)^2}{100(\varepsilon-1)} < \lambda_1 \lesssim\lambda_{\mathrm{MF}}, \tag*{(5)} 100 ( ε − 1 ) ( 2 − ε ) 2 < λ 1 ≲ λ MF , (5)
where λ M F \lambda_{\mathrm{MF}} λ MF is the lowest non-zero eigenvalue of a second-order differential equation in one variable which can be easily tabulated to arbitrary accuracy. In the limit ε → 2 \varepsilon\to2 ε → 2 , λ M F ( 2 ) = 1 / 4 \lambda_{\mathrm{MF}}(2) = 1/4 λ MF ( 2 ) = 1/4 while for ε → 1 \varepsilon\to1 ε → 1 , λ M F ≃ 1 / ( ε − 1 ) \lambda_{\mathrm{MF}} \simeq1/(\varepsilon-1) λ MF ≃ 1/ ( ε − 1 ) . The upper bound holds asymptotically, lim sup N → ∞ λ 1 ≤ λ M F \limsup_{N\to\infty} \lambda_1 \leq\lambda_{\mathrm{MF}} lim sup N → ∞ λ 1 ≤ λ MF (Section IV A), hence “≲ \lesssim ≲ ”. The numerical experiments find a decay time which is much closer to the upper bound 2 / λ M F 2/\lambda_{\mathrm{MF}} 2/ λ MF than to the lower bound, which suggests that the lower bound is not strict. We could not, however, find a better lower bound.
For ε > 2 \varepsilon> 2 ε > 2 we have (assuming the limit exists)
( 2 − 1 ) ε − 2 ≤ lim N → ∞ log τ e q N ≤ ε − 2 . (6) (\sqrt{2}-1)\sqrt{\varepsilon-2} \leq\lim_{N\to\infty}\frac{\log\tau_{\mathrm{eq}}}{\sqrt{N}} \leq\sqrt{\varepsilon-2}. \tag*{(6)} ( 2 − 1 ) ε − 2 ≤ N → ∞ lim N log τ eq ≤ ε − 2 . (6)
(Theorems IV.1 and V.1). The stretched-exponential time scale is a consequence of the second conservation law. In condensed zero-range processes, which conserve only the mass, the condensate moves between sites on a polynomial time scale [28 ] . Here it cannot move without passing through configurations in which two sites share the excess energy, whose area is smaller by a factor e − ( 2 − 1 ) N ( ε − 2 ) e^{-(\sqrt{2}-1)\sqrt{N(\varepsilon-2)}} e − ( 2 − 1 ) N ( ε − 2 ) (Section III C).
In the next Section, we introduce some results about the geometry of the microcanonical surface that will be used in Section IV. More specialized results to be used in § 0.6 will be introduced at the beginning of that Section.
Geometry of the microcanonical surface
The surface and its regions
Throughout, N ≥ 5 N \geq5 N ≥ 5 and 1 < ε < N 1 < \varepsilon< N 1 < ε < N , and we use units in which the total charge is Q = N Q = N Q = N . The microcanonical surface is the intersection of the simplex { q i ≥ 0 , ∑ i q i = N } \{q_i \geq0,\ \sum_i q_i = N\} { q i ≥ 0 , ∑ i q i = N } with the sphere ∑ i q i 2 = E = N ε \sum_i q_i^2 = E = N\varepsilon ∑ i q i 2 = E = N ε :
M = M N , ε = { q ∈ [ 0 , ∞ ) N : ∑ i q i = N , ∑ i q i 2 = N ε } , R 2 : = N ( ε − 1 ) . M = M_{N,\varepsilon} = \left\{q \in[0,\infty)^N : \sum_i q_i = N,\ \sum_i q_i^2 = N\varepsilon\right\}, \qquad R^2 := N(\varepsilon-1). M = M N , ε = { q ∈ [ 0 , ∞ ) N : i ∑ q i = N , i ∑ q i 2 = N ε } , R 2 := N ( ε − 1 ) .
Since ∣ q − 1 ∣ 2 = R 2 \lvert q-1\rvert^2 = R^2 ∣ q − 1 ∣ 2 = R 2 , M M M is the part of a round sphere of radius R R R (centre 1 1 1 , inside the hyperplane ∑ i q i = N \sum_i q_i = N ∑ i q i = N ) cut out by the N N N half-spaces q j ≥ 0 q_j \geq0 q j ≥ 0 : a spherical polytope of dimension d = N − 2 ≥ 3 d = N - 2 \ge3 d = N − 2 ≥ 3 , whose boundary ∂ M = { q ∈ M : ∃ j , q j = 0 } \partial M = \{q \in M : \exists j,\ q_j = 0\} ∂ M = { q ∈ M : ∃ j , q j = 0 } consists of the faces { q j = 0 } \{q_j = 0\} { q j = 0 } . We write σ = H N − 2 \sigma= \mathcal{H}^{N-2} σ = H N − 2 for its Riemannian measure, μ = σ / σ ( M ) \mu= \sigma/\sigma(M) μ = σ / σ ( M ) for the normalised measure, and M ∘ M^\circ M ∘ for its interior. The largest value of a coordinate on M M M is
t ∗ : = 1 + ( N − 1 ) ( ε − 1 ) t^* := 1 + \sqrt{(N-1)(\varepsilon-1)} t ∗ := 1 + ( N − 1 ) ( ε − 1 ) , attained only at the pole P i P_i P i , where q i = t ∗ q_i=t^* q i = t ∗ and all the other coordinates are equal. We write ⟨ ψ ⟩ : = 1 N ∑ j ψ ( q j ) \langle\psi\rangle:= \frac{1}{N}\sum_j\psi(q_j) ⟨ ψ ⟩ := N 1 ∑ j ψ ( q j ) for a function ψ \psi ψ of one variable, and I I ( v , w ) = ⟨ ∇ v ν , w ⟩ \mathrm{II}(v,w)=\langle\nabla_v\nu,w\rangle II ( v , w ) = ⟨ ∇ v ν , w ⟩ for the second fundamental form of a boundary face with outward normal ν \nu ν , so that convex faces have I I ≥ 0 \mathrm{II}\ge0 II ≥ 0 . For ε > 2 \varepsilon>2 ε > 2 we put a : = ε − 2 a:=\sqrt{\varepsilon-2} a := ε − 2 .
The following are elementary standard computations which will be used throughout the paper:
Lemma III.1. Let γ ( x ) : = 1 − 1 N − ( x − 1 ) 2 R 2 \gamma(x):=1-\frac{1}{N}-\frac{(x-1)^2}{R^2} γ ( x ) := 1 − N 1 − R 2 ( x − 1 ) 2 . On M M M , for all i , j i,j i , j :
⟨ ∇ q i , ∇ q j ⟩ = δ i j − 1 N − ( q i − 1 ) ( q j − 1 ) R 2 \langle\nabla q_i,\nabla q_j\rangle=\delta_{ij}-\frac{1}{N}-\frac{(q_i-1)(q_j-1)}{R^2} ⟨ ∇ q i , ∇ q j ⟩ = δ ij − N 1 − R 2 ( q i − 1 ) ( q j − 1 ) , so that ∣ ∇ q i ∣ 2 = γ ( q i ) |\nabla q_i|^2=\gamma(q_i) ∣∇ q i ∣ 2 = γ ( q i ) ; moreover ∑ j ∇ q j = ∑ j q j ∇ q j = 0 \sum_j\nabla q_j=\sum_jq_j\nabla q_j=0 ∑ j ∇ q j = ∑ j q j ∇ q j = 0 .
∇ 2 q i = − q i − 1 R 2 g \nabla^2q_i=-\frac{q_i-1}{R^2}g ∇ 2 q i = − R 2 q i − 1 g , Δ q i = − N − 2 R 2 ( q i − 1 ) \Delta q_i=-\frac{N-2}{R^2}(q_i-1) Δ q i = − R 2 N − 2 ( q i − 1 ) and Ric = N − 3 R 2 g \operatorname{Ric}=\frac{N-3}{R^2}g Ric = R 2 N − 3 g .
If γ ( K ) > 0 \gamma(K)>0 γ ( K ) > 0 , the face { q j = K } \{q_j=K\} { q j = K } of { q j ≥ K } \{q_j\ge K\} { q j ≥ K } has outward normal ν = − ∇ q j / ∣ ∇ q j ∣ \nu=-\nabla q_j/|\nabla q_j| ν = − ∇ q j /∣∇ q j ∣ and I I = K − 1 R 2 ∣ ∇ q j ∣ g \mathrm{II}=\frac{K-1}{R^2|\nabla q_j|}g II = R 2 ∣∇ q j ∣ K − 1 g . Since q j = 1 q_j=1 q j = 1 at the centre 1 1 1 , { q j ≥ K } \{q_j\ge K\} { q j ≥ K } is larger than a hemisphere, with concave boundary, for K < 1 K<1 K < 1 , and is a cap with convex boundary for K > 1 K>1 K > 1 . In particular the faces { q j = 0 } \{q_j=0\} { q j = 0 } of M M M are concave.
For F = ⟨ ψ ⟩ F=\langle\psi\rangle F = ⟨ ψ ⟩ and all α , β ∈ R \alpha,\beta\in\mathbb{R} α , β ∈ R , ∣ ∇ F ∣ 2 ≤ 1 N ⟨ ( ψ ′ − α − β x ) 2 ⟩ |\nabla F|^2\le\frac{1}{N}\langle(\psi'-\alpha-\beta x)^2\rangle ∣∇ F ∣ 2 ≤ N 1 ⟨( ψ ′ − α − β x ) 2 ⟩ ; on { q i = 0 } \{q_i=0\} { q i = 0 } , ⟨ ∇ F , ∇ q i ⟩ = 1 N ( ψ ′ ( 0 ) + ⟨ ( x − ε ) ψ ′ ⟩ ε − 1 ) \langle\nabla F,\nabla q_i\rangle=\frac{1}{N}\left(\psi'(0)+\frac{\langle(x-\varepsilon)\psi'\rangle}{\varepsilon-1}\right) ⟨ ∇ F , ∇ q i ⟩ = N 1 ( ψ ′ ( 0 ) + ε − 1 ⟨( x − ε ) ψ ′ ⟩ ) .
( ε − 1 ) Δ F = ( ε − 1 ) ⟨ ψ ′ ′ ⟩ − ⟨ ( x − 1 ) ψ ′ ⟩ + R ψ (\varepsilon-1)\Delta F=(\varepsilon-1)\langle\psi''\rangle-\langle(x-1)\psi'\rangle+R_\psi ( ε − 1 ) Δ F = ( ε − 1 ) ⟨ ψ ′′ ⟩ − ⟨( x − 1 ) ψ ′ ⟩ + R ψ , with ∣ R ψ ∣ ≤ 4 ( ε − 1 ) sup ∣ ψ ′ ′ ∣ / N |R_\psi|\le4(\varepsilon-1)\sup|\psi''|/N ∣ R ψ ∣ ≤ 4 ( ε − 1 ) sup ∣ ψ ′′ ∣/ N .
Slices and meridians
Many geometric properties of M M M will be recovered by slicing it at a given coordinate, and reducing it to a lower dimensional microcanonical surface of lower dimension and different energy. To make this idea quantitative we fix i i i , put n = N − 1 n=N-1 n = N − 1 and
c ( t ) = N − t N − 1 , ρ ( t ) 2 = R 2 − N ( t − 1 ) 2 N − 1 , c(t)=\frac{N-t}{N-1},\qquad\rho(t)^2=R^2-\frac{N(t-1)^2}{N-1}, c ( t ) = N − 1 N − t , ρ ( t ) 2 = R 2 − N − 1 N ( t − 1 ) 2 ,
e ( t ) = ( N − 1 ) ( N ε − t 2 ) ( N − t ) 2 = 1 + ρ 2 n c 2 , θ ( t ) = 2 − e ( t ) . e(t)=\frac{(N-1)(N\varepsilon-t^2)}{(N-t)^2}=1+\frac{\rho^2}{nc^2},\qquad\theta(t)=2-e(t). e ( t ) = ( N − t ) 2 ( N − 1 ) ( N ε − t 2 ) = 1 + n c 2 ρ 2 , θ ( t ) = 2 − e ( t ) .
For q ∈ M q\in M q ∈ M with q i = t q_i=t q i = t , the other n n n coordinates have sum N − t N-t N − t and sum of squares N ε − t 2 N\varepsilon-t^2 N ε − t 2 , so that y j = q j / c ( t ) y_j=q_j/c(t) y j = q j / c ( t ) , j ≠ i j\ne i j = i , satisfy 1 n ∑ j ≠ i y j = 1 \frac{1}{n}\sum_{j\ne i}y_j=1 n 1 ∑ j = i y j = 1 and 1 n ∑ j ≠ i y j 2 = e ( t ) \frac{1}{n}\sum_{j\ne i}y_j^2=e(t) n 1 ∑ j = i y j 2 = e ( t ) . The slice { q i = t } \{q_i=t\} { q i = t } is thus a rescaled microcanonical surface of energy e ( t ) e(t) e ( t ) , and θ ( t ) \theta(t) θ ( t ) measures how far it lies below the critical energy 2 2 2 (to leading order θ ( t ) ≈ ( t 2 − ( ε − 2 ) N ) / N \theta(t)\approx(t^2-(\varepsilon-2)N)/N θ ( t ) ≈ ( t 2 − ( ε − 2 ) N ) / N ).
Lemma III.2.
The slice Σ i ( t ) = M ∩ { q i = t } \Sigma_i(t)=M\cap\{q_i=t\} Σ i ( t ) = M ∩ { q i = t } is c ( t ) ⋅ M n , e ( t ) c(t)\cdot M_{n,e(t)} c ( t ) ⋅ M n , e ( t ) , a homothetic copy of radius ρ ( t ) \rho(t) ρ ( t ) .
In the meridian coordinates ( q j ) j ≠ i = c ( t ) 1 + ρ ( t ) ω (q_j)_{j\ne i}=c(t)\mathbf{1}+\rho(t)\omega ( q j ) j = i = c ( t ) 1 + ρ ( t ) ω , with ω \omega ω in the unit sphere of { v ∈ R n : ∑ j v j = 0 } \{v\in\mathbb{R}^n:\sum_jv_j=0\} { v ∈ R n : ∑ j v j = 0 } , g = d t 2 / γ ( t ) + ρ 2 g S g=\mathrm{d}t^2/\gamma(t)+\rho^2g_S g = d t 2 / γ ( t ) + ρ 2 g S and d σ = ρ N − 3 γ − 1 / 2 d t d ω \mathrm{d}\sigma=\rho^{N-3}\gamma^{-1/2}\mathrm{d}t\,\mathrm{d}\omega d σ = ρ N − 3 γ − 1/2 d t d ω .
For t > ε t>\varepsilon t > ε , e ′ ( t ) < 0 e'(t)<0 e ′ ( t ) < 0 and ( c / ρ ) ′ > 0 (c/\rho)'>0 ( c / ρ ) ′ > 0 : the admissible directions ω \omega ω increase with t t t , and every meridian through { q i > ε } \{q_i>\varepsilon\} { q i > ε } stays in M ∘ M^\circ M ∘ and reaches the pole P i P_i P i .
The regions. For K ≥ ε K\ge\varepsilon K ≥ ε , the cap { q ∈ M : q i > K } \{q\in M:q_i>K\} { q ∈ M : q i > K } is connected, since its meridians reach P i P_i P i . Its boundary consists of the face { q i = K } \{q_i=K\} { q i = K } , which is convex because K > 1 K>1 K > 1 , and of pieces of the concave faces { q j = 0 } \{q_j=0\} { q j = 0 } , j ≠ i j\ne i j = i . We will use the cells A i A_i A i and three such caps,
A i = { q ∈ M ∘ : q i > q j ∀ j ≠ i } , U i = { q ∈ M ∘ : q i > ε } , A_i=\{q\in M^\circ:q_i>q_j\ \forall j\ne i\},\qquad U_i=\{q\in M^\circ:q_i>\varepsilon\}, A i = { q ∈ M ∘ : q i > q j ∀ j = i } , U i = { q ∈ M ∘ : q i > ε } ,
H i = { q ∈ M ∘ : q i > 2 ε } , G i = { q ∈ M : q i ≥ T } , H_i=\{q\in M^\circ:q_i>2\varepsilon\},\qquad G_i=\{q\in M:q_i\ge T\}, H i = { q ∈ M ∘ : q i > 2 ε } , G i = { q ∈ M : q i ≥ T } ,
where the level of the deep cap G i G_i G i , used only for ε ≥ 2 \varepsilon\ge2 ε ≥ 2 , is
T ∈ [ T 0 , T 0 + 1 ] , T 0 = { ε − 2 N + 16 ( ε − 2 ) − 1 / 4 N 1 / 4 , ε > 2 , 16 N 1 / 3 , ε = 2 , (8) T\in[T_0,T_0+1],\qquad
T_0=
\begin{cases}
\sqrt{\varepsilon-2}\sqrt{N}+16(\varepsilon-2)^{-1/4}N^{1/4}, & \varepsilon>2,\\
16N^{1/3}, & \varepsilon=2,
\end{cases} \tag*{(8)} T ∈ [ T 0 , T 0 + 1 ] , T 0 = { ε − 2 N + 16 ( ε − 2 ) − 1/4 N 1/4 , 16 N 1/3 , ε > 2 , ε = 2 , (8)
chosen outside the finitely many levels at which the face { q i = T } \{q_i=T\} { q i = T } does not meet the faces { q j = 0 } \{q_j=0\} { q j = 0 } transversally (Proposition V.3 (b)). For N N N large, 2 ε < T < T + 1 < t ∗ ≤ N / 2 2\varepsilon<T<T+1<t^*\le N/2 2 ε < T < T + 1 < t ∗ ≤ N /2 (the last inequality holds, for example, when N ≥ 4 ε N\ge4\varepsilon N ≥ 4 ε ), so that ∅ ≠ G i ⊂ H i ⊂ U i \varnothing\ne G_i\subset H_i\subset U_i ∅ = G i ⊂ H i ⊂ U i . The regions are shown in Figure 3 .
Figure 3. The regions G i ⊂ H i ⊂ U i G_i \subset H_i \subset U_i G i ⊂ H i ⊂ U i and the cells A i A_i A i of M M M , in a planar view and on the sphere.
Lemma III.4. Let 1 < ε < N / 2 1<\varepsilon<N/2 1 < ε < N /2 . The largest coordinate of every q ∈ M ∘ q\in M^\circ q ∈ M ∘ exceeds ε \varepsilon ε , because ε = ∑ j ( q j / N ) q j \varepsilon=\sum_j(q_j/N)q_j ε = ∑ j ( q j / N ) q j is a weighted average of the coordinates. Hence M ∘ = ⋃ i U i M^\circ=\bigcup_iU_i M ∘ = ⋃ i U i and A i ⊂ U i A_i\subset U_i A i ⊂ U i ;
the cells A i A_i A i partition M M M up to null sets, and μ ( A i ) = 1 / N \mu(A_i)=1/N μ ( A i ) = 1/ N . Moreover U i ∩ U j ≠ ∅ U_i \cap U_j \ne\varnothing U i ∩ U j = ∅ , since it contains the configurations with q i = q j > ε q_i=q_j>\varepsilon q i = q j > ε and all other coordinates equal, so that M M M is connected.
The connectedness of M M M for 1 < ε < N / 2 1<\varepsilon<N/2 1 < ε < N /2 is due to [15 Sec. 6 and App. A] .
Neumann eigenvalues. A probability density ρ t \rho_t ρ t on M M M evolves by ∂ t ρ t = 1 2 Δ M ρ t \partial_t\rho_t=\frac{1}{2}\Delta_M\rho_t ∂ t ρ t = 2 1 Δ M ρ t , with Neumann (reflecting) conditions on ∂ M \partial M ∂ M . The measure μ \mu μ is reversible, and the relaxation time (2) is τ e q = 2 / λ 1 \tau_{\rm eq}=2/\lambda_1 τ eq = 2/ λ 1 , where
λ 1 = λ 1 ( N , ε ) = inf { ∫ M ∣ ∇ f ∣ 2 d μ Var μ f : f ∈ H 1 ( M ) , Var μ f > 0 } (9) \lambda_1=\lambda_1(N,\varepsilon)=\inf\left\{\frac{\int_M |\nabla f|^2\,d\mu}{\operatorname{Var}_{\mu}f}: f\in H^1(M),\ \operatorname{Var}_{\mu}f>0\right\} \tag*{(9)} λ 1 = λ 1 ( N , ε ) = inf { Var μ f ∫ M ∣∇ f ∣ 2 d μ : f ∈ H 1 ( M ) , Var μ f > 0 } (9)
is the first nonzero Neumann eigenvalue of − Δ M -\Delta_M − Δ M . We write E ( f , f ) = ∫ M ∣ ∇ f ∣ 2 d μ \mathcal{E}(f,f)=\int_M|\nabla f|^2d\mu E ( f , f ) = ∫ M ∣∇ f ∣ 2 d μ for the Dirichlet form. For a subdomain D ⊂ M D\subset M D ⊂ M , λ 1 ( D ) \lambda_1(D) λ 1 ( D ) is the first nonzero Neumann eigenvalue of D D D , and P ( D ) = 1 / λ 1 ( D ) P(D)=1/\lambda_1(D) P ( D ) = 1/ λ 1 ( D ) is its Poincaré constant, the smallest constant with min c ∈ R ∫ D ∣ f − c ∣ 2 d σ ≤ P ( D ) ∫ D ∣ ∇ f ∣ 2 d σ \min_{c\in\mathbb{R}}\int_D|f-c|^2\,d\sigma\le P(D)\int_D|\nabla f|^2\,d\sigma min c ∈ R ∫ D ∣ f − c ∣ 2 d σ ≤ P ( D ) ∫ D ∣∇ f ∣ 2 d σ for all f ∈ H 1 ( D ) f\in H^1(D) f ∈ H 1 ( D ) . We abbreviate P G = P ( G i ) P_G=P(G_i) P G = P ( G i ) , P H = P ( H i ) P_H=P(H_i) P H = P ( H i ) , P U = P ( U i ) P_U=P(U_i) P U = P ( U i ) .
Areas
The areas of M M M and of its coordinate-defined subsets are expressed through the microcanonical partition function
Ω n ( S , E ) = ∫ [ 0 , ∞ ) n d n q δ ( S − ∑ i q i ) δ ( E − ∑ i q i 2 ) . (10) \Omega_n(S,E)=\int_{[0,\infty)^n}d^nq\,\delta\left(S-\sum_i q_i\right)\delta\left(E-\sum_i q_i^2\right). \tag*{(10)} Ω n ( S , E ) = ∫ [ 0 , ∞ ) n d n q δ ( S − i ∑ q i ) δ ( E − i ∑ q i 2 ) . (10)
The constraint gradients ∇ ∑ i q i = 1 \nabla\sum_i q_i=1 ∇ ∑ i q i = 1 and ∇ ∑ i q i 2 = 2 q \nabla\sum_i q_i^2=2q ∇ ∑ i q i 2 = 2 q have the Gram matrix
G = ( N 2 N 2 N 4 N ε ) , det G = 2 N ε − 1 , (11) G=\begin{pmatrix}N&2N\\2N&4N\varepsilon\end{pmatrix},\qquad\sqrt{\det G}=2N\sqrt{\varepsilon-1}, \tag*{(11)} G = ( N 2 N 2 N 4 N ε ) , det G = 2 N ε − 1 , (11)
which is constant on M M M . Moreover, by the coarea formula:
d μ ( q ) = 1 Ω N ( N , N ε ) δ ( N − ∑ i q i ) δ ( N ε − ∑ i q i 2 ) ∏ i = 1 N 1 { q i ≥ 0 } d q i , (12) d\mu(q)=\frac{1}{\Omega_N(N,N\varepsilon)}\delta\left(N-\sum_iq_i\right)\delta\left(N\varepsilon-\sum_iq_i^2\right)\prod_{i=1}^{N}\mathbf{1}_{\{q_i\geq0\}}\,dq_i, \tag*{(12)} d μ ( q ) = Ω N ( N , N ε ) 1 δ ( N − i ∑ q i ) δ ( N ε − i ∑ q i 2 ) i = 1 ∏ N 1 { q i ≥ 0 } d q i , (12)
and, since e − ∑ i q i = e − N e^{-\sum_iq_i}=e^{-N} e − ∑ i q i = e − N on M M M , the conditioned product of exponentials
μ = Law ( X | ∑ i X i = N , ∑ i X i 2 = N ε ) , X i i . i . d . E x p ( 1 ) . (13) \mu=\operatorname{Law}\left(X\ \middle|\ \sum_iX_i=N,\ \sum_iX_i^2=N\varepsilon\right),\qquad X_i\ {\rm i.i.d.}\ {\rm Exp}(1). \tag*{(13)} μ = Law ( X i ∑ X i = N , i ∑ X i 2 = N ε ) , X i i.i.d. Exp ( 1 ) . (13)
For any number n n n of coordinates, let X 1 , … , X n X_1,\ldots,X_n X 1 , … , X n be i.i.d. Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) . For L ∈ ( 0 , ∞ ] L\in(0,\infty] L ∈ ( 0 , ∞ ] let f n < L ( S , E ) f_n^{<L}(S,E) f n < L ( S , E ) be the density of ( ∑ j X j , ∑ j X j 2 ) (\sum_jX_j,\sum_jX_j^2) ( ∑ j X j , ∑ j X j 2 ) restricted to the event { max j X j < L } \{\max_jX_j<L\} { max j X j < L } , and f n : = f n < ∞ f_n:=f_n^{<\infty} f n := f n < ∞ ; let Ω n < L \Omega_n^{<L} Ω n < L be defined as Ω n \Omega_n Ω n , with all n n n coordinates restricted to [ 0 , L ) [0,L) [ 0 , L ) .
Lemma III.5 (areas as densities).
(1) f n < L ( S , E ) = e − S Ω n < L ( S , E ) f_n^{<L}(S,E)=e^{-S}\Omega_n^{<L}(S,E) f n < L ( S , E ) = e − S Ω n < L ( S , E ) . More generally, if X 1 , … , X n X_1,\ldots,X_n X 1 , … , X n are i.i.d. with a density p p p , the density of ( ∑ j X j , ∑ j X j 2 ) (\sum_jX_j,\sum_jX_j^2) ( ∑ j X j , ∑ j X j 2 ) is bounded by π 3 ( sup p ) 3 \frac{\pi}{\sqrt{3}}(\sup p)^3 3 π ( sup p ) 3 for n ≥ 3 n\geq3 n ≥ 3 ; in particular f n < L ≤ π / 3 f_n^{<L}\leq\pi/\sqrt{3} f n < L ≤ π / 3 .
(2) σ ( M N , ε ) = 2 N ε − 1 Ω N ( N , N ε ) = 2 N ε − 1 e N f N ( N , N ε ) \sigma(M_{N,\varepsilon})=2N\sqrt{\varepsilon-1}\,\Omega_N(N,N\varepsilon)=2N\sqrt{\varepsilon-1}\,e^Nf_N(N,N\varepsilon) σ ( M N , ε ) = 2 N ε − 1 Ω N ( N , N ε ) = 2 N ε − 1 e N f N ( N , N ε ) .
(3) For sets defined by conditions on k ≤ N − 3 k\leq N-3 k ≤ N − 3 coordinates, areas are obtained by integrating out those coordinates. For instance the joint density of two coordinates is
μ ( q i ∈ d u , q j ∈ d v ) = Ω N − 2 ( N − u − v , N ε − u 2 − v 2 ) Ω N ( N , N ε ) d u d v , \mu(q_i\in du,q_j\in dv)=\frac{\Omega_{N-2}\left(N-u-v,N\varepsilon-u^2-v^2\right)}{\Omega_N(N,N\varepsilon)}\,du\,dv, μ ( q i ∈ d u , q j ∈ d v ) = Ω N ( N , N ε ) Ω N − 2 ( N − u − v , N ε − u 2 − v 2 ) d u d v , so that
μ ( U i ∩ U j ) = 1 f N ( N , N ε ) ∫ ε ∞ ∫ ε ∞ e − u − v f N − 2 ( N − u − v , N ε − u 2 − v 2 ) d u d v . \mu(U_i\cap U_j)=\frac{1}{f_N(N,N\varepsilon)}\int_{\varepsilon}^{\infty}\int_{\varepsilon}^{\infty}e^{-u-v}f_{N-2}\left(N-u-v,N\varepsilon-u^2-v^2\right)\,du\,dv. μ ( U i ∩ U j ) = f N ( N , N ε ) 1 ∫ ε ∞ ∫ ε ∞ e − u − v f N − 2 ( N − u − v , N ε − u 2 − v 2 ) d u d v .
Lemma III.6 (local limit theorem [29 §19] ). Uniformly in z ∈ R 2 z \in\mathbb{R}^{2} z ∈ R 2 , n f n ( n + n z 1 , 2 n + n z 2 ) n f_n(n+\sqrt{n}z_1,2n+\sqrt{n}z_2) n f n ( n + n z 1 , 2 n + n z 2 ) converges to a nondegenerate Gaussian density in z z z . In particular f n ≥ c / n f_n \ge c/n f n ≥ c / n within O ( n ) O(\sqrt{n}) O ( n ) of ( n , 2 n ) (n,2n) ( n , 2 n ) .
Lemma III.6 controls the partition functions Ω n ( S , E ) = e S f n ( S , E ) \Omega_n(S,E)=e^S f_n(S,E) Ω n ( S , E ) = e S f n ( S , E ) only within O ( n ) O(\sqrt{n}) O ( n ) of ( n , 2 n ) (n,2n) ( n , 2 n ) . Far from it, at energy densities n E / S 2 > 2 nE/S^2>2 n E / S 2 > 2 , where the residual system is supercritical, we use a Chernoff bound with a cutoff: when all coordinates lie below a cutoff, the energy fugacity may be taken negative.
Lemma III.7 (tilted bound). Fix ε > 2 \varepsilon>2 ε > 2 , a = ε − 2 a=\sqrt{\varepsilon-2} a = ε − 2 and k ∈ { 1 , 2 } k\in\{1,2\} k ∈ { 1 , 2 } , and let n = N − k n=N-k n = N − k . Uniformly in ς ≥ 0 \varsigma\ge0 ς ≥ 0 , r ≥ 0 r\ge0 r ≥ 0 and ζ ∈ [ 1 / ( 2 N ) , ε ] \zeta\in[1/(2\sqrt{N}),\sqrt{\varepsilon}] ζ ∈ [ 1/ ( 2 N ) , ε ] ,
ln Ω n ζ N ( N − ς N , N ( ε − r 2 ) ) ≤ N − N [ ς + ( a 2 − r 2 ) + max ( ζ , a / 2 ) ] + C ln N , (14) \ln\Omega_n^{\zeta\sqrt{N}}\left(N-\varsigma\sqrt{N},N(\varepsilon-r^2)\right)\le N-\sqrt{N}\left[\varsigma+\frac{(a^2-r^2)_+}{\max(\zeta,a/2)}\right]+C\ln N, \tag*{(14)} ln Ω n ζ N ( N − ς N , N ( ε − r 2 ) ) ≤ N − N [ ς + max ( ζ , a /2 ) ( a 2 − r 2 ) + ] + C ln N , (14) with C C C depending only on ε \varepsilon ε , and ( x ) + = max ( x , 0 ) (x)_+=\max(x,0) ( x ) + = max ( x , 0 ) .
The proof is in Appendix A.
For ε > 2 \varepsilon>2 ε > 2 the area of M M M is dominated by the configurations in which one coordinate, of size about a N a\sqrt{N} a N , carries the whole excess energy a 2 N a^2N a 2 N (Section III C). The following two Propositions give the fundamental estimates on volumes needed throughout our paper and will be proved in Appendix B:
Proposition III.8 (total area). For fixed ε ≥ 2 \varepsilon\ge2 ε ≥ 2 and N N N large ,
exp { N − a N − O ( ln N ) } ≤ Ω N ( N , N ε ) ≤ exp { N − a N + O ( ln N ) } , (15) \exp\{N-a\sqrt{N}-O(\ln N)\}\le\Omega_N(N,N\varepsilon)\le\exp\{N-a\sqrt{N}+O(\ln N)\}, \tag*{(15)} exp { N − a N − O ( ln N )} ≤ Ω N ( N , N ε ) ≤ exp { N − a N + O ( ln N )} , (15) that is, e − a N − O ( ln N ) ≤ f N ( N , N ε ) ≤ e − a N + O ( ln N ) e^{-a\sqrt{N}-O(\ln N)}\le f_N(N,N\varepsilon)\le e^{-a\sqrt{N}+O(\ln N)} e − a N − O ( l n N ) ≤ f N ( N , N ε ) ≤ e − a N + O ( l n N ) .
The gluing in the lower bound needs the caps U i U_i U i to overlap substantially. By Lemma III.5(3) the overlap is a ratio of values of the densities f n f_n f n ; the underlying phenomenon, condensation on the intersection of a simplex and a sphere, is described in Section III C and in [17 , 30 ] .
Combined with the local limit theorem, the upper bound in Proposition III.8 has the following consequence, which is crucial for the proof of the lower bound when ε ≥ 2 \varepsilon\ge2 ε ≥ 2 ; part (a), at ε = 2 \varepsilon=2 ε = 2 , needs only the local limit theorem:
Proposition III.9. Let i ≠ j i\ne j i = j .
(a) At ε = 2 \varepsilon=2 ε = 2 , μ ( U i ∩ U j ) ≥ α 2 : = 1 2 ( e − 3 − e − 4 ) 2 \mu(U_i\cap U_j)\ge\alpha_2:=\frac{1}{2}(e^{-3}-e^{-4})^2 μ ( U i ∩ U j ) ≥ α 2 := 2 1 ( e − 3 − e − 4 ) 2 for N ≥ N 0 N\ge N_0 N ≥ N 0 .
(b) For fixed ε > 2 \varepsilon> 2 ε > 2 there is C ε C_{\varepsilon} C ε such that μ ( U i ∩ U j ) ≥ N − C ε \mu(U_i \cap U_j) \ge N^{-C_{\varepsilon}} μ ( U i ∩ U j ) ≥ N − C ε for N ≥ N ε N \ge N_{\varepsilon} N ≥ N ε ; any C ε > 2 C_{\varepsilon} > 2 C ε > 2 is admissible.
Finally, along the meridians of a cap: by Lemma III.2(2), on { q i > ε } \{q_i > \varepsilon\} { q i > ε } one has d σ = v ( t ) d t d ω \mathrm{d}\sigma= v(t)\,\mathrm{d}t\,\mathrm{d}\omega d σ = v ( t ) d t d ω , with v ( t ) = ρ ( t ) N − 3 γ ( t ) − 1 / 2 v(t) = \rho(t)N^{-3\gamma(t)-1/2} v ( t ) = ρ ( t ) N − 3 γ ( t ) − 1/2 and ω \omega ω ranging over the admissible directions. For t , s ∈ ( ε , 2 ε + 1 ] t,s \in(\varepsilon,2\varepsilon+1] t , s ∈ ( ε , 2 ε + 1 ] and N N N large,
γ ( t ) ≥ 1 2 , v ( t ) v ( s ) ≤ C ε . (16) \gamma(t) \ge\frac{1}{2}, \qquad\frac{v(t)}{v(s)} \le C_{\varepsilon}. \tag*{(16)} γ ( t ) ≥ 2 1 , v ( s ) v ( t ) ≤ C ε . (16)
Deeper in the cap the area decays like e − q i e^{-q_i} e − q i ; in the chart of H i H_i H i this is quantified by Lemma VI.1(a).
The landscape: chambers and the tube
This subsection explains where the two proofs come from. For ε > 2 \varepsilon> 2 ε > 2 the surface consists of N N N chambers, one for each site that can carry the condensate. They are joined pairwise by tubes whose narrowest section, the neck , is smaller than a chamber by a factor e − b ( ε ) N e^{-b(\varepsilon)\sqrt{N}} e − b ( ε ) N , and the spectral gap is governed by this neck.
Upper bound (Section IV). The landscape dictates the test function: a function of two occupations, + 1 +1 + 1 when the condensate sits on site 1, − 1 -1 − 1 when it sits on site 2 and 0 otherwise, which changes at the neck level q ≃ a N / 2 q \simeq a\sqrt{N}/2 q ≃ a N /2 , so that its Dirichlet energy is set by the area of the necks. Proposition IV.2 makes the neck estimate rigorous.
Lower bound (§ 0.6 –VII). The landscape explains why the Bochner argument must be localised to the deep caps G i G_i G i , which lie inside a single chamber, and why the lower bound loses against the neck (Remark IV.6).
The subsection itself is heuristic. Apart from the exact identities (19) and (G1), it uses the asymptotics of Ω n \Omega_n Ω n from [14 ] , and its displays hold to the indicated orders without error control. It describes the landscape behind both proofs, but none of the proofs uses it: the neck exponent b ( ε ) b(\varepsilon) b ( ε ) of (27) is proved as an upper bound in Proposition IV.2, and the total area as Proposition III.8.
Entropy. The computation of Ω N \Omega_N Ω N has been done in [14 ] . By homogeneity it suffices to know Ω n ( n , n x ) \Omega_n(n,nx) Ω n ( n , n x ) as a function of the energy density x x x . With fugacities η \eta η and β \beta β conjugate to mass and energy, and the single-site integral
Z ( η , β ) = ∫ 0 ∞ d q e − η q − β q 2 , (17) Z(\eta,\beta)=\int_{0}^{\infty}dq\,e^{-\eta q-\beta q^{2}}, \tag*{(17)} Z ( η , β ) = ∫ 0 ∞ d q e − η q − β q 2 , (17)
finite for β ≥ 0 \beta\ge0 β ≥ 0 , the saddle-point equations E p [ q ] = 1 \mathbb{E}_{p}[q]=1 E p [ q ] = 1 , E p [ q 2 ] = x \mathbb{E}_{p}[q^{2}]=x E p [ q 2 ] = x for the tilted density p ( q ) = e − η q − β q 2 / Z ( η , β ) p(q)=e^{-\eta q-\beta q^{2}}/Z(\eta,\beta) p ( q ) = e − η q − β q 2 / Z ( η , β ) have a solution with β > 0 \beta>0 β > 0 when 1 < x < 2 1<x<2 1 < x < 2 , and
ln Ω n ( n , n x ) = n S ( x ) + O ( ln n ) , S ( x ) = η + β x + ln Z ( η , β ) . (18) \ln\Omega_{n}(n,nx)=nS(x)+O(\ln n),\qquad S(x)=\eta+\beta x+\ln Z(\eta,\beta). \tag*{(18)} ln Ω n ( n , n x ) = n S ( x ) + O ( ln n ) , S ( x ) = η + β x + ln Z ( η , β ) . (18)
S S S is concave, with maximum S ( 2 ) = 1 S(2)=1 S ( 2 ) = 1 at the infinite-temperature point x = 2 x=2 x = 2 (β = 0 \beta=0 β = 0 , η = 1 \eta=1 η = 1 , p ( q ) = e − q p(q)=e^{-q} p ( q ) = e − q ), and S ( x ) ≃ 1 − ( x − 2 ) 2 / 8 S(x)\simeq1-(x-2)^{2}/8 S ( x ) ≃ 1 − ( x − 2 ) 2 /8 there. As x → 1 + x\to1^{+} x → 1 + , β → + ∞ \beta\to+\infty β → + ∞ ; for 1 < ϵ ≤ N / ( N − 1 ) 1<\epsilon\le N/(N-1) 1 < ϵ ≤ N / ( N − 1 ) the surface is a round sphere of radius R R R inside the simplex, and λ 1 = ( N − 2 ) / R 2 \lambda_{1}=(N-2)/R^{2} λ 1 = ( N − 2 ) / R 2 exactly (compare Theorem V.1 (i)).
The fluid phase 1 < ϵ < 2 1<\epsilon<2 1 < ϵ < 2 . Locally μ \mu μ is a product of the densities p ( q ) p(q) p ( q ) , which decay faster than e − q e^{-q} e − q : the largest occupation is O ( ln N ) O(\ln N) O ( ln N ) and there are no chambers. Since S S S is increasing for x < 2 x<2 x < 2 , an excursion q i = z i N q_i=z_i\sqrt{N} q i = z i N of one or two sites costs a factor e − N [ S ( ϵ ) − S ( ϵ − z 1 2 − z 2 2 ) ] e^{-N[S(\epsilon)-S(\epsilon-z_1^2-z_2^2)]} e − N [ S ( ϵ ) − S ( ϵ − z 1 2 − z 2 2 )] . There is no tunnel to cross, consistent with Theorem V.1 (i). For ϵ > 2 \epsilon>2 ϵ > 2 the saddle-point equations have no solution with β ≥ 0 \beta\ge0 β ≥ 0 : the saddle sticks at the branch point β = 0 \beta=0 β = 0 , the residual system stays critical, and the excess energy N ( ϵ − 2 ) N(\epsilon-2) N ( ϵ − 2 ) is carried by a single site, the condensate .
Two-site sections for $\epsilon>2. A passage from one chamber to another involves two coordinates, the one losing the excess energy and the one gaining it. We use the two-site density of Lemma III.5(3) with q i = z i N q_i=z_i\sqrt{N} q i = z i N . The residual system has n = N − 2 n=N-2 n = N − 2 sites, mass S ′ = N − q 1 − q 2 S'=N-q_1-q_2 S ′ = N − q 1 − q 2 and energy E ′ = N ϵ − q 1 2 − q 2 2 E'=N\epsilon-q_1^2-q_2^2 E ′ = N ϵ − q 1 2 − q 2 2 ; the exact homogeneity Ω n ( S , E ) = m n − 3 Ω n ( S / m , E / m 2 ) \Omega_n(S,E)=m^{n-3}\Omega_n(S/m,E/m^2) Ω n ( S , E ) = m n − 3 Ω n ( S / m , E / m 2 ) with m = S ′ / n m=S'/n m = S ′ / n gives
ln Ω n ( S ′ , E ′ ) = ( n − 3 ) ln m + ln Ω n ( n , n x ) , x = E ′ n S ′ 2 , (19) \ln\Omega_n(S',E')=(n-3)\ln m+\ln\Omega_n(n,nx),\qquad x=\frac{E'n}{S'^{2}}, \tag*{(19)} ln Ω n ( S ′ , E ′ ) = ( n − 3 ) ln m + ln Ω n ( n , n x ) , x = S ′ 2 E ′ n , (19)
where
( n − 3 ) ln m = − ( z 1 + z 2 ) N − ( z 1 + z 2 ) 2 2 + 2 + O ( N − 1 / 2 ) , (20) (n-3)\ln m=-(z_1+z_2)\sqrt{N}-\frac{(z_1+z_2)^2}{2}+2+O(N^{-1/2}), \tag*{(20)} ( n − 3 ) ln m = − ( z 1 + z 2 ) N − 2 ( z 1 + z 2 ) 2 + 2 + O ( N − 1/2 ) , (20)
x = ( ϵ − z 1 2 − z 2 2 ) [ 1 + 2 ( z 1 + z 2 ) N ] + O ( N − 1 ) . (21) x=\left(\epsilon-z_1^2-z_2^2\right)\left[1+\frac{2(z_1+z_2)}{\sqrt{N}}\right]+O(N^{-1}). \tag*{(21)} x = ( ϵ − z 1 2 − z 2 2 ) [ 1 + N 2 ( z 1 + z 2 ) ] + O ( N − 1 ) . (21)
The two spikes remove energy q i 2 q_i^2 q i 2 but also mass q i q_i q i ; the mass removal is the O ( N − 1 / 2 ) O(N^{-1/2}) O ( N − 1/2 ) correction in (21) , which displaces the tunnel by O ( 1 ) O(1) O ( 1 ) . Since ln Ω n ( n , n x ) = n − n ( x − 2 ) + O ( log n ) \ln\Omega_n(n,nx)=n-\sqrt{n(x-2)}+O(\log n) ln Ω n ( n , n x ) = n − n ( x − 2 ) + O ( log n ) for x > 2 x>2 x > 2 [14 ] ,
ln Ω N − 2 ( N − q 1 − q 2 , N ϵ − q 1 2 − q 2 2 ) = N − N Φ ( z 1 , z 2 ) + O ( ln N ) , (22) \ln\Omega_{N-2}\left(N-q_1-q_2,N\epsilon-q_1^2-q_2^2\right)=N-\sqrt{N}\,\Phi(z_1,z_2)+O(\ln N), \tag*{(22)} ln Ω N − 2 ( N − q 1 − q 2 , N ϵ − q 1 2 − q 2 2 ) = N − N Φ ( z 1 , z 2 ) + O ( ln N ) , (22)
with Φ ( z 1 , z 2 ) = z 1 + z 2 + a 2 − z 1 2 − z 2 2 \Phi(z_1,z_2)=z_1+z_2+\sqrt{a^2-z_1^2-z_2^2} Φ ( z 1 , z 2 ) = z 1 + z 2 + a 2 − z 1 2 − z 2 2 for z 1 2 + z 2 2 ≤ a 2 z_1^2+z_2^2\le a^2 z 1 2 + z 2 2 ≤ a 2 . Φ \Phi Φ is symmetric, since transposing two coordinates is an isometry of M M M that preserves μ \mu μ . At z = 0 z=0 z = 0 it gives ln Ω N ( N , N ε ) = N − a N ≡ ln V c \ln\Omega_N(N,N\varepsilon)=N-a\sqrt{N}\equiv\ln V_c ln Ω N ( N , N ε ) = N − a N ≡ ln V c , the volume of the surface and, to exponential order, of a single chamber. The transverse section at fixed ( q 1 , q 2 ) (q_1,q_2) ( q 1 , q 2 ) is therefore
A ( z 1 , z 2 ) V c = exp { − N [ Φ ( z 1 , z 2 ) − a ] } , V c = e N − a N . (23) \frac{\mathcal{A}(z_1,z_2)}{V_c}=\exp\left\{-\sqrt{N}\left[\Phi(z_1,z_2)-a\right]\right\},\qquad V_c=e^{N-a\sqrt{N}}. \tag*{(23)} V c A ( z 1 , z 2 ) = exp { − N [ Φ ( z 1 , z 2 ) − a ] } , V c = e N − a N . (23)
The tube. The square root in (22) vanishes on the circle
z 1 2 + z 2 2 = a 2 , (24) z_1^2+z_2^2=a^2, \tag*{(24)} z 1 2 + z 2 2 = a 2 , (24)
where the two peeled coordinates carry the whole excess energy and the other N − 2 N-2 N − 2 are exactly critical. Inside the disk a third coordinate hosts a spike of size ( a 2 − ∣ z ∣ 2 ) N \sqrt{(a^2-|z|^2)N} ( a 2 − ∣ z ∣ 2 ) N , so a path through the interior is a concatenation of two passages and is not cheaper. The direct passage between two chambers is therefore confined to (24) , where Φ = ∥ z ∥ 1 \Phi=\|z\|_1 Φ = ∥ z ∥ 1 at fixed ∥ z ∥ 2 = a \|z\|_2=a ∥ z ∥ 2 = a . With z = a ( cos θ , sin θ ) z=a(\cos\theta,\sin\theta) z = a ( cos θ , sin θ ) ,
Φ − a = a [ 2 cos ( θ − π 4 ) − 1 ] , \Phi-a=a\left[\sqrt{2}\cos\left(\theta-\frac{\pi}{4}\right)-1\right], Φ − a = a [ 2 cos ( θ − 4 π ) − 1 ] ,
which vanishes at the ends θ = 0 , π / 2 \theta=0,\pi/2 θ = 0 , π /2 (the chambers) and is maximal at θ = π / 4 \theta=\pi/4 θ = π /4 . At finite N N N the mass removed by the peeled coordinates pushes the tube outwards (Figure 1, panel b). Balancing the Gaussian fluctuation of the residual system, ln Ω n ( n , n x ) ≃ ln Ω n ( n , 2 n ) − n ( x − 2 ) 2 / 8 \ln\Omega_n(n,nx)\simeq\ln\Omega_n(n,2n)-n(x-2)^2/8 ln Ω n ( n , n x ) ≃ ln Ω n ( n , 2 n ) − n ( x − 2 ) 2 /8 for ∣ x − 2 ∣ ≪ n − 1 / 3 |x-2|\ll n^{-1/3} ∣ x − 2∣ ≪ n − 1/3 , against the linear cost in (20) places the tube at
∥ z ∥ 2 = a + ( 2 − 1 a 2 ) 2 cos ( θ − π 4 ) N + O ( N − 1 ) , (25) \|z\|_2=a+\left(2-\frac{1}{a^2}\right)\frac{\sqrt{2}\cos\left(\theta-\frac{\pi}{4}\right)}{\sqrt{N}}+O(N^{-1}), \tag*{(25)} ∥ z ∥ 2 = a + ( 2 − a 2 1 ) N 2 cos ( θ − 4 π ) + O ( N − 1 ) , (25)
each large coordinate being shifted by 2 − 1 / a 2 2-1/a^2 2 − 1/ a 2 . Collecting the O ( 1 ) O(1) O ( 1 ) terms of (20) , (25) and of the Gaussian fluctuation, the section along the tube is
ln A ( θ ) V c = − N a [ 2 cos ( θ − π 4 ) − 1 ] − c ∗ ( ε ) sin 2 θ + O ( N − 1 / 2 ) , c ∗ ( ε ) = 2 + a 2 2 − 1 2 a 2 , (26) \ln\frac{\mathcal{A}(\theta)}{V_c}=-\sqrt{N}a\left[\sqrt{2}\cos\left(\theta-\frac{\pi}{4}\right)-1\right]-c_*(\varepsilon)\sin2\theta+O(N^{-1/2}),\qquad c_*(\varepsilon)=2+\frac{a^2}{2}-\frac{1}{2a^2}, \tag*{(26)} ln V c A ( θ ) = − N a [ 2 cos ( θ − 4 π ) − 1 ] − c ∗ ( ε ) sin 2 θ + O ( N − 1/2 ) , c ∗ ( ε ) = 2 + 2 a 2 − 2 a 2 1 , (26)
valid for a ≫ N − 1 / 6 a\gg N^{-1/6} a ≫ N − 1/6 ; a numerical evaluation of Ω n \Omega_n Ω n up to N = 12800 N=12800 N = 12800 confirms (25) and the value of c ∗ ( ε ) c_*(\varepsilon) c ∗ ( ε ) . In the coordinates q ˉ = ( q 1 + q 2 ) / 2 \bar q=(q_1+q_2)/2 q ˉ = ( q 1 + q 2 ) /2 , τ = q 1 − q 2 \tau=q_1-q_2 τ = q 1 − q 2 one has sin 2 θ = 1 − τ 2 / ( a 2 N ) \sin2\theta=1-\tau^2/(a^2N) sin 2 θ = 1 − τ 2 / ( a 2 N ) on the tube, and the section of the tunnel joining two chambers is
A ( τ ) = V c exp { − N [ 2 a 2 − τ 2 N − a ] − c ∗ ( ε ) ( 1 − τ 2 a 2 N ) } , ∣ τ ∣ ≤ a N , (27) \mathcal{A}(\tau)=V_c\exp\left\{-\sqrt{N}\left[\sqrt{2a^2-\frac{\tau^2}{N}}-a\right]-c_*(\varepsilon)\left(1-\frac{\tau^2}{a^2N}\right)\right\},\qquad|\tau|\le a\sqrt{N}, \tag*{(27)} A ( τ ) = V c exp { − N [ 2 a 2 − N τ 2 − a ] − c ∗ ( ε ) ( 1 − a 2 N τ 2 ) } , ∣ τ ∣ ≤ a N , (27)
The neck. The narrowest point is τ = 0 \tau=0 τ = 0 , where q 1 = q 2 = a N / 2 + 2 − 1 / a 2 q_1=q_2=a\sqrt{N/2+2-1/a^2} q 1 = q 2 = a N /2 + 2 − 1/ a 2 and
A ( 0 ) = V c e − N b ( ε ) − c ∗ ( ε ) , b ( ε ) = ( 2 − 1 ) a . \mathcal{A}(0)=V_c e^{-\sqrt{N}b(\varepsilon)-c_*(\varepsilon)},\qquad b(\varepsilon)=(\sqrt{2}-1)a. A ( 0 ) = V c e − N b ( ε ) − c ∗ ( ε ) , b ( ε ) = ( 2 − 1 ) a .
It is exactly the midpoint of the tube: at q 1 = a N / 2 q_1=a\sqrt{N/2} q 1 = a N /2 the excess N a 2 / 2 Na^2/2 N a 2 /2 left to the other sites must be carried by a single coordinate, so q 2 = q 1 q_2=q_1 q 2 = q 1 . Expanding the square root,
ln A ( τ ) A ( 0 ) = τ 2 2 2 a N + c ∗ ( ε ) τ 2 a 2 N + O ( τ 4 N 3 / 2 ) (28) \ln\frac{\mathcal{A}(\tau)}{\mathcal{A}(0)}=\frac{\tau^2}{2\sqrt{2}a\sqrt{N}}+c_*(\varepsilon)\frac{\tau^2}{a^2N}+O\left(\frac{\tau^4}{N^{3/2}}\right) \tag*{(28)} ln A ( 0 ) A ( τ ) = 2 2 a N τ 2 + c ∗ ( ε ) a 2 N τ 2 + O ( N 3/2 τ 4 ) (28)
so the section stays within a constant factor of its minimum over ∣ τ ∣ ≲ ( 2 2 a ) 1 / 2 N 1 / 4 |\tau|\lesssim(2\sqrt{2}a)^{1/2}N^{1/4} ∣ τ ∣ ≲ ( 2 2 a ) 1/2 N 1/4 . The bottleneck is a window of width N 1 / 4 N^{1/4} N 1/4 in a channel of length 2 a N 2a\sqrt{N} 2 a N .
Figure 1 and Figure 2 are obtained from a numerically exact evaluation of Ω n ( S , E ) \Omega_n(S,E) Ω n ( S , E ) , by exponential tilting and a decomposition into fluid and condensate (Appendix G); it agrees with an exact-constraint Montecarlo at N = 100 N=100 N = 100 .
Figure 1. Several figures to understand the geometry of the high-dimensional space obtained by a numerically exact evaluation of the densities (exponential tilting plus the fluid–condensate decomposition (G1)), at N = 100 N = 100 N = 100 . a) A two-dimensional projection (higher volumes are brighter colors) for ε = 1.5 < 2 \varepsilon= 1.5 < 2 ε = 1.5 < 2 ; b) The same with ε = 4 > 2 \varepsilon= 4 > 2 ε = 4 > 2 where one can identify the two chambers q 1 ≃ a N q_1 \simeq a\sqrt{N} q 1 ≃ a N and q 2 ≃ a N q_2 \simeq a\sqrt{N} q 2 ≃ a N and, around the origin the projection of the remaining N − 2 N - 2 N − 2 chambers; c) Projection on a single coordinate which shows the appearance of the chambers and the tunnel and the agreement with Montecarlo results.
Figure 2. The numerical computed areas at N = 100 , … , 12800 N = 100, \ldots, 12800 N = 100 , … , 12800 to show convergence to the asymptotic formula. Leading order in left and O ( 1 ) O(1) O ( 1 ) corrections in right panel.
Upper bound on λ 1 \lambda_1 λ 1
One can obtain upper bounds easily by variational methods. By (9) , every f f f with E μ f = 0 \mathbb{E}_\mu f=0 E μ f = 0 gives λ 1 ≤ E μ ∣ ∇ M f ∣ 2 / E μ f 2 \lambda_1\leq\mathbb{E}_\mu|\nabla_Mf|^2/\mathbb{E}_\mu f^2 λ 1 ≤ E μ ∣ ∇ M f ∣ 2 / E μ f 2 , and the closer f f f is to the eigenfunction, the better the bound. The natural trial functions are different in the two phases, in particular the number of variables they must depend non-trivially upon changes from one for ε ≤ 2 \varepsilon\leq2 ε ≤ 2 to two for ε ≥ 2 \varepsilon\geq2 ε ≥ 2 .
The fluid phase: ε < 2 \varepsilon<2 ε < 2
In the fluid phase, ε < 2 \varepsilon<2 ε < 2 , all the variables have the same order of magnitude q i ∼ O ( 1 ) q_i\sim O(1) q i ∼ O ( 1 ) and the natural trial functions are functions of a single variable only, f ( q ) = φ ( q 1 ) − E μ φ ( q 1 ) f(q)=\varphi(q_1)-\mathbb{E}_\mu\varphi(q_1) f ( q ) = φ ( q 1 ) − E μ φ ( q 1 ) . In the limit N → ∞ N\to\infty N → ∞ the sites become independent, and the slowest mode of a product of independent processes is a function of one coordinate. For such functions both averages reduce to one-dimensional integrals. Since ∇ M f = φ ′ ( q 1 ) ∇ M q 1 \nabla_Mf=\varphi'(q_1)\nabla_Mq_1 ∇ M f = φ ′ ( q 1 ) ∇ M q 1 , and ∣ ∇ M q 1 ∣ 2 = γ ( q 1 ) = 1 − 1 N − ( q 1 − 1 ) 2 N ( ε − 1 ) |\nabla_Mq_1|^2=\gamma(q_1)=1-\frac{1}{N}-\frac{(q_1-1)^2}{N(\varepsilon-1)} ∣ ∇ M q 1 ∣ 2 = γ ( q 1 ) = 1 − N 1 − N ( ε − 1 ) ( q 1 − 1 ) 2 is a function of q 1 q_1 q 1 alone (Lemma III.1), one finds exactly
λ 1 ≤ ∫ 0 ∞ d q 1 γ ( q 1 ) φ ′ ( q 1 ) 2 p N ( q 1 ) ∫ 0 ∞ d q 1 ( φ ( q 1 ) − φ ˉ ) 2 p N ( q 1 ) , p N ( q 1 ) = Ω N − 1 ( N − q 1 , N ε − q 1 2 ) Ω N ( N , N ε ) , (29) \lambda_1\leq\frac{\int_0^\infty dq_1\,\gamma(q_1)\varphi'(q_1)^2p_N(q_1)}{\int_0^\infty dq_1\,(\varphi(q_1)-\bar{\varphi})^2p_N(q_1)},\qquad p_N(q_1)=\frac{\Omega_{N-1}(N-q_1,N\varepsilon-q_1^2)}{\Omega_N(N,N\varepsilon)}, \tag*{(29)} λ 1 ≤ ∫ 0 ∞ d q 1 ( φ ( q 1 ) − φ ˉ ) 2 p N ( q 1 ) ∫ 0 ∞ d q 1 γ ( q 1 ) φ ′ ( q 1 ) 2 p N ( q 1 ) , p N ( q 1 ) = Ω N ( N , N ε ) Ω N − 1 ( N − q 1 , N ε − q 1 2 ) , (29)
where p N p_N p N is the marginal of one occupation and φ ˉ = ∫ 0 ∞ φ ( q ) p N ( q ) d q \bar{\varphi}=\int_0^\infty\varphi(q)p_N(q)\,dq φ ˉ = ∫ 0 ∞ φ ( q ) p N ( q ) d q . First of all, since γ ( q 1 ) = 1 − O ( 1 / N ) < 1 \gamma(q_1)=1-O(1/N)<1 γ ( q 1 ) = 1 − O ( 1/ N ) < 1 (and goes to 1 when N → ∞ N\to\infty N → ∞ , for any ε > 1 \varepsilon>1 ε > 1 ) we can replace this upper bound with another, simpler one in which γ → 1 \gamma\to1 γ → 1 :
λ 1 ≤ ∫ 0 ∞ d q 1 φ ′ ( q 1 ) 2 p N ( q 1 ) ∫ 0 ∞ d q 1 ( φ ( q 1 ) − φ ˉ ) 2 p N ( q 1 ) . (30) \lambda_1 \le\frac{\int_0^\infty\mathrm{d}q_1\,\varphi'(q_1)^2 p_N(q_1)}{\int_0^\infty\mathrm{d}q_1\,\left(\varphi(q_1)-\bar{\varphi}\right)^2 p_N(q_1)}. \tag*{(30)} λ 1 ≤ ∫ 0 ∞ d q 1 ( φ ( q 1 ) − φ ˉ ) 2 p N ( q 1 ) ∫ 0 ∞ d q 1 φ ′ ( q 1 ) 2 p N ( q 1 ) . (30)
The best choice of φ \varphi φ solves the Euler–Lagrange equation of (30) , a one-dimensional Sturm–Liouville problem,
− 1 p N d d q ( p N d φ d q ) = Λ N φ , φ ′ = 0 at the ends of the support , (31) -\frac{1}{p_N}\frac{\mathrm{d}}{\mathrm{d}q}\left(p_N\frac{\mathrm{d}\varphi}{\mathrm{d}q}\right)=\Lambda_N\varphi,\qquad\varphi'=0\ \text{at the ends of the support}, \tag*{(31)} − p N 1 d q d ( p N d q d φ ) = Λ N φ , φ ′ = 0 at the ends of the support , (31)
and its first non-zero eigenvalue Λ N ( ε ) \Lambda_N(\varepsilon) Λ N ( ε ) is an upper bound on λ 1 \lambda_1 λ 1 at every finite N N N .
By the equivalence of ensembles (Section III C, see also [17 ] ) p N ( q ) p_N(q) p N ( q ) tends to the tilted density p ( q ) = e − η q − β q 2 / Z ( η , β ) p(q)=e^{-\eta q-\beta q^2}/Z(\eta,\beta) p ( q ) = e − η q − β q 2 / Z ( η , β ) of (17) , with ( η , β ) (\eta,\beta) ( η , β ) fixed by E p [ q ] = 1 \mathbb{E}_p[q]=1 E p [ q ] = 1 and E p [ q 2 ] = ε \mathbb{E}_p[q^2]=\varepsilon E p [ q 2 ] = ε . Equation (31) then becomes the eigenvalue problem of a single reflected diffusion in the potential W ( q ) = η q + β q 2 W(q)=\eta q+\beta q^2 W ( q ) = η q + β q 2 ,
− φ ′ ′ + ( η + 2 β q ) φ ′ = λ φ , q ≥ 0 , φ ′ ( 0 ) = 0. (32) -\varphi''+(\eta+2\beta q)\varphi'=\lambda\varphi,\qquad q\ge0,\qquad\varphi'(0)=0. \tag*{(32)} − φ ′′ + ( η + 2 β q ) φ ′ = λ φ , q ≥ 0 , φ ′ ( 0 ) = 0. (32)
This equation is solved exactly. In the variable ξ = 2 β q + ξ 0 \xi=\sqrt{2\beta}\,q+\xi_0 ξ = 2 β q + ξ 0 , with ξ 0 = η / 2 β \xi_0=\eta/\sqrt{2\beta} ξ 0 = η / 2 β , it becomes the Hermite equation φ ′ ′ − ξ φ ′ + ν φ = 0 \varphi''-\xi\varphi'+\nu\varphi=0 φ ′′ − ξ φ ′ + ν φ = 0 with ν = λ / 2 β \nu=\lambda/2\beta ν = λ /2 β . Its solution with polynomial growth is φ = H e ν ( ξ ) \varphi=He_\nu(\xi) φ = H e ν ( ξ ) , and the Neumann condition at ξ = ξ 0 \xi=\xi_0 ξ = ξ 0 selects D ν − 1 ( ξ 0 ) = 0 D_{\nu-1}(\xi_0)=0 D ν − 1 ( ξ 0 ) = 0 , where D ν D_\nu D ν is the parabolic cylinder function. The first non-zero eigenvalue
λ M F ( ε ) = 2 β ν ( ξ 0 ) (33) \lambda_{\mathrm{MF}}(\varepsilon)=2\beta\,\nu(\xi_0) \tag*{(33)} λ MF ( ε ) = 2 β ν ( ξ 0 ) (33)
is the gap of one independent site. Inserting this φ \varphi φ in (30) and letting N → ∞ N\to\infty N → ∞ gives
lim sup N → ∞ λ 1 ( N , ε ) ≤ λ M F ( ε ) , 1 < ε < 2. (34) \limsup_{N\to\infty}\lambda_1(N,\varepsilon)\le\lambda_{\mathrm{MF}}(\varepsilon),\qquad1<\varepsilon<2. \tag*{(34)} N → ∞ lim sup λ 1 ( N , ε ) ≤ λ MF ( ε ) , 1 < ε < 2. (34)
The function λ M F \lambda_{\mathrm{MF}} λ MF decreases from the sphere value 1 / ( ε − 1 ) 1/(\varepsilon-1) 1/ ( ε − 1 ) at ε → 1 \varepsilon\to1 ε → 1 , where φ \varphi φ is linear, to 1 / 4 1/4 1/4 at ε → 2 − \varepsilon\to2^- ε → 2 − . Figure 4 compares λ M F ( ε ) \lambda_{\mathrm{MF}}(\varepsilon) λ MF ( ε ) with the relaxation rates measured by a Montecarlo numerical algorithm.
Figure 4. The analytic value of the upper bound λ M F \lambda_{\mathrm{MF}} λ MF compared with the numerical values extracted from the correlation functions of the Montecarlo procedure discussed in the text. The deviations at ε = 1.9 \varepsilon= 1.9 ε = 1.9 are indicative of the critical region.
We measured the relaxation rates of the Brownian motion directly, with an exactly reversible Montecarlo dynamics: each move rotates three occupations along the circle that preserves their mass and energy, and time is calibrated so that the generator is Δ M \Delta_M Δ M , the normalisation of λ M F \lambda_{\mathrm{MF}} λ MF . Rates were measured separately in the symmetry sectors of S N S_N S N , extrapolated to zero step size and then to N → ∞ N\to\infty N → ∞ ; details are in Appendix G.
The results are reproduced in Figure 4 .
The condensate phase ε > 2 \varepsilon> 2 ε > 2
For ε > 2 \varepsilon> 2 ε > 2 the result of the previous subsection is no longer useful. In the limit N → ∞ N \to\infty N → ∞ the one-site marginal tends to p ( q ) = e − q p(q) = e^{-q} p ( q ) = e − q , the bulk at β = 0 \beta= 0 β = 0 , and the condensate q i ≃ a N q_i \simeq a\sqrt{N} q i ≃ a N , a = ε − 2 a = \sqrt{\varepsilon- 2} a = ε − 2 , disappears from it: it carries weight 1 / N 1/N 1/ N and escapes to infinity. The solutions of
− φ ′ ′ ( q ) + φ ′ ( q ) = λ φ ( q ) , φ ′ ( 0 ) = 0 , (35) -\varphi^{\prime\prime}(q) + \varphi^{\prime}(q) = \lambda\varphi(q), \qquad\varphi^{\prime}(0) = 0, \tag*{(35)} − φ ′′ ( q ) + φ ′ ( q ) = λ φ ( q ) , φ ′ ( 0 ) = 0 , (35)
that are square integrable with respect to e − q d q e^{-q}dq e − q d q are the constant, with λ 0 = 0 \lambda_0 = 0 λ 0 = 0 , and a continuum starting at λ = 1 / 4 \lambda= 1/4 λ = 1/4 . Formally λ M F ( ε > 2 ) = 1 / 4 \lambda_{\mathrm{MF}}(\varepsilon> 2) = 1/4 λ MF ( ε > 2 ) = 1/4 , which is a correct but very loose bound.
The slow process in the condensed phase is the transfer of the condensate from one site to another. Its simplest description uses two sites, say 1 and 2, and test functions of q 1 q_1 q 1 and q 2 q_2 q 2 only. Such a function can be + 1 +1 + 1 when the condensate sits on site 1, − 1 -1 − 1 when it sits on site 2, and 0 when it sits anywhere else. During the transfer the configuration must cross the tunnel of Section III C. This leads to:
Theorem IV.1 (upper bound). Fix ε > 2 \varepsilon> 2 ε > 2 and let a = ε − 2 a = \sqrt{\varepsilon- 2} a = ε − 2 . Then
λ 1 ( N , ε ) ≤ exp { − b ( ε ) N + O ( ln N ) } , b ( ε ) = ( 2 − 1 ) a . \lambda_1(N,\varepsilon) \le\exp\left\{-b(\varepsilon)\sqrt{N} + O(\ln N)\right\}, \qquad b(\varepsilon) = (\sqrt{2} - 1)a. λ 1 ( N , ε ) ≤ exp { − b ( ε ) N + O ( ln N ) } , b ( ε ) = ( 2 − 1 ) a .
We prove this theorem by going to two-site test functions. The exponential bound is not improved by going to 3 and higher variables test functions.
Two-site test functions
The reduction of Section IV A extends to functions of two occupations, f ( q ) = φ ( q 1 , q 2 ) − E μ φ f(q)=\varphi(q_1,q_2)-\mathbb{E}_{\mu}\varphi f ( q ) = φ ( q 1 , q 2 ) − E μ φ . The subtraction − E μ φ -\mathbb{E}_{\mu}\varphi − E μ φ is necessary to orthogonalize with respect to the ground state, constant function. The gradient is ∇ M f = ∑ k = 1 , 2 ∂ k φ ∇ M q k \nabla_M f=\sum_{k=1,2}\partial_k\varphi\,\nabla_M q_k ∇ M f = ∑ k = 1 , 2 ∂ k φ ∇ M q k , and the matrix Π k l = ⟨ ∇ M q k , ∇ M q l ⟩ \Pi_{kl}=\langle\nabla_Mq_k,\nabla_Mq_l\rangle Π k l = ⟨ ∇ M q k , ∇ M q l ⟩ , k , l ∈ { 1 , 2 } k,l\in\{1,2\} k , l ∈ { 1 , 2 } , given in Lemma III.1, depends on q 1 q_1 q 1 and q 2 q_2 q 2 only. Both averages in the variational principle therefore reduce exactly to integrals over the quadrant ( q 1 , q 2 ) (q_1,q_2) ( q 1 , q 2 ) , with the two-site density of Lemma III.5(3),
p 2 ( q 1 , q 2 ) = Ω N − 2 ( N − q 1 − q 2 , N ε − q 1 2 − q 2 2 ) Ω N ( N , N ε ) . (36) p_2(q_1,q_2)=\frac{\Omega_{N-2}\left(N-q_1-q_2,\;N\varepsilon-q_1^2-q_2^2\right)}{\Omega_N(N,N\varepsilon)}. \tag*{(36)} p 2 ( q 1 , q 2 ) = Ω N ( N , N ε ) Ω N − 2 ( N − q 1 − q 2 , N ε − q 1 2 − q 2 2 ) . (36)
This is the density drawn in Figure 1 . One finds
λ 1 ≤ ∬ d q 1 d q 2 p 2 ( q 1 , q 2 ) ∇ φ ⋅ Π ∇ φ ∬ d q 1 d q 2 p 2 ( q 1 , q 2 ) ( φ − E μ φ ) 2 , (37) \lambda_1\leq\frac{\iint dq_1\,dq_2\,p_2(q_1,q_2)\,\nabla\varphi\cdot\Pi\nabla\varphi}{\iint dq_1\,dq_2\,p_2(q_1,q_2)\left(\varphi-\mathbb{E}_{\mu}\varphi\right)^2}, \tag*{(37)} λ 1 ≤ ∬ d q 1 d q 2 p 2 ( q 1 , q 2 ) ( φ − E μ φ ) 2 ∬ d q 1 d q 2 p 2 ( q 1 , q 2 ) ∇ φ ⋅ Π∇ φ , (37)
for every φ \varphi φ , at every N N N . In the same way as we replaced γ → 1 \gamma\to1 γ → 1 in the previous subsection, since Π \Pi Π is the restriction of an orthogonal projector, ∇ φ ⋅ Π ∇ φ ≤ ∣ ∇ φ ∣ 2 \nabla\varphi\cdot\Pi\nabla\varphi\leq|\nabla\varphi|^2 ∇ φ ⋅ Π∇ φ ≤ ∣∇ φ ∣ 2 , and Π → δ k l \Pi\to\delta_{kl} Π → δ k l at the price of a weaker but still valid bound. The loss is O ( 1 / N ) O(1/N) O ( 1/ N ) where q 1 , q 2 = O ( 1 ) q_1,q_2=O(1) q 1 , q 2 = O ( 1 ) , and a factor of order one near the condensate; at the neck, Π \Pi Π acts as the identity along the tube direction ( 1 , − 1 ) (1,-1) ( 1 , − 1 ) .
The optimal φ \varphi φ solves a two-dimensional Sturm–Liouville problem (with Π \Pi Π replaced by δ k l \delta_{kl} δ k l ), − ∇ ⋅ ( p 2 ∇ φ ) = Λ N ( 2 ) p 2 φ -\nabla\cdot(p_2\nabla\varphi)=\Lambda_N^{(2)}p_2\varphi − ∇ ⋅ ( p 2 ∇ φ ) = Λ N ( 2 ) p 2 φ with natural boundary conditions. Its first non-zero eigenvalue Λ N ( 2 ) \Lambda_N^{(2)} Λ N ( 2 ) is an upper bound on λ 1 \lambda_1 λ 1 and lies below the one-site value Λ N \Lambda_N Λ N , since one-site functions are a special case.
The two-site problem has a transparent meaning. In the potential W = − ln p 2 W=-\ln p_2 W = − ln p 2 the quadrant looks like panel (b) of Figure 1 :
The lowest antisymmetric mode is a two-well Kramers problem: φ ≃ + 1 \varphi\simeq+1 φ ≃ + 1 in one well, ≃ − 1 \simeq-1 ≃ − 1 in the other, ≃ 0 \simeq0 ≃ 0 in the bulk.
To prove Theorem IV.1 it is enough to write down one explicit function of this type. We do not solve the two-dimensional problem but a connection with Witten’s supersymmetric quantum mechanics [19 ] is mentioned at the end of this section.
The trial function
Let q ± = a N / 2 ± ℓ / 2 q_{\pm}=a\sqrt{N/2}\pm\ell/2 q ± = a N /2 ± ℓ /2 , with a fixed width ℓ > 0 \ell>0 ℓ > 0 , and let χ ( q ) \chi(q) χ ( q ) be the ramp equal to 0 for q ≤ q − q\leq q_{-} q ≤ q − , to 1 for q ≥ q + q\geq q_{+} q ≥ q + , and linear in between. We take
φ ( q 1 , q 2 ) = χ ( q 1 ) − χ ( q 2 ) . (38) \varphi(q_{1},q_{2})=\chi(q_{1})-\chi(q_{2}). \tag*{(38)} φ ( q 1 , q 2 ) = χ ( q 1 ) − χ ( q 2 ) . (38)
Then φ = + 1 \varphi=+1 φ = + 1 when site 1 is above the neck level a N / 2 a\sqrt{N/2} a N /2 and site 2 is below it, φ = − 1 \varphi=-1 φ = − 1 in the opposite case, and φ = 0 \varphi=0 φ = 0 when both are below, which is the case when the condensate is elsewhere. The function changes only when one of the two occupations crosses the neck level. The three ingredients of the Rayleigh quotient are as follows.
Mean. φ \varphi φ is antisymmetric under q 1 ↔ q 2 q_{1}\leftrightarrow q_{2} q 1 ↔ q 2 , which preserves μ \mu μ . Hence E μ φ = 0 \mathbb{E}_{\mu}\varphi=0 E μ φ = 0 and φ ⊥ 1 \varphi\perp1 φ ⊥ 1 .
Dirichlet energy. The intrinsic gradient is the projection of the ambient one, so ∣ ∇ M φ ∣ 2 ≤ χ ′ ( q 1 ) 2 + χ ′ ( q 2 ) 2 \lvert\nabla_{M}\varphi\rvert^{2}\leq\chi'(q_{1})^{2}+\chi'(q_{2})^{2} ∣ ∇ M φ ∣ 2 ≤ χ ′ ( q 1 ) 2 + χ ′ ( q 2 ) 2 . With I = [ q − , q + ] I=[q_{-},q_{+}] I = [ q − , q + ] and p N p_{N} p N the one-site marginal of (30) ,
E μ ∣ ∇ M φ ∣ 2 ≤ 2 ℓ 2 ∫ I p N ( q ) d q ≤ 2 ℓ sup q ∈ I p N ( q ) . (39) \mathbb{E}_{\mu}\lvert\nabla_{M}\varphi\rvert^{2}\leq\frac{2}{\ell^{2}}\int_{I}p_{N}(q)\,\mathrm{d}q\leq\frac{2}{\ell}\sup_{q\in I}p_{N}(q). \tag*{(39)} E μ ∣ ∇ M φ ∣ 2 ≤ ℓ 2 2 ∫ I p N ( q ) d q ≤ ℓ 2 q ∈ I sup p N ( q ) . (39)
The energy is thus set by the probability that one given site sits exactly at the neck level.
Variance. φ 2 = 1 \varphi^{2}=1 φ 2 = 1 when one of the two sites is above q + q_{+} q + and the other below q − q_{-} q − . Since ∑ i q i 2 = N ε \sum_{i}q_{i}^{2}=N\varepsilon ∑ i q i 2 = N ε , at most n ∗ = N ε / q − 2 ≃ 2 ε / a 2 n_{*}=N\varepsilon/q_{-}^{2}\simeq2\varepsilon/a^{2} n ∗ = N ε / q − 2 ≃ 2 ε / a 2 sites can exceed q − q_{-} q − . So if a condensate above q + q_{+} q + exists, it can be paired with at least N − n ∗ N-n_{*} N − n ∗ sites below q − q_{-} q − . By exchangeability,
E μ φ 2 ≥ 2 μ ( q 1 ≥ q + , q 2 ≤ q − ) ≥ 2 ( N − n ∗ ) N ( N − 1 ) μ ( max i q i ≥ q + ) . (40) \mathbb{E}_{\mu}\varphi^{2}\geq2\mu\left(q_{1}\geq q_{+},q_{2}\leq q_{-}\right)\geq\frac{2(N-n_{*})}{N(N-1)}\mu\left(\max_{i}q_{i}\geq q_{+}\right). \tag*{(40)} E μ φ 2 ≥ 2 μ ( q 1 ≥ q + , q 2 ≤ q − ) ≥ N ( N − 1 ) 2 ( N − n ∗ ) μ ( i max q i ≥ q + ) . (40)
For ε > 2 \varepsilon>2 ε > 2 the last probability tends to one (Lemma IV.3 below), and E μ φ 2 ≥ ( 2 / N ) ( 1 − o ( 1 ) ) \mathbb{E}_{\mu}\varphi^{2}\geq(2/N)(1-o(1)) E μ φ 2 ≥ ( 2/ N ) ( 1 − o ( 1 )) .
Inserting (39) and (40) in the variational principle gives the two-site Rayleigh bound
λ 1 ≤ N ℓ ( 1 + o ( 1 ) ) sup ∣ q − a N / 2 ∣ ≤ ℓ / 2 p N ( q ) . (41) \lambda_{1}\leq\frac{N}{\ell}(1+o(1))\sup_{\lvert q-a\sqrt{N/2}\rvert\leq\ell/2}p_{N}(q). \tag*{(41)} λ 1 ≤ ℓ N ( 1 + o ( 1 )) ∣ q − a N /2 ∣ ≤ ℓ /2 sup p N ( q ) . (41)
Everything is now reduced to the value of the one-site marginal at the neck level.
The single-site density at the neck level
The heuristic picture of Section III C predicts p N p_N p N at once. Fix q 1 = z N q_1=z\sqrt{N} q 1 = z N with z < a z<a z < a . The remaining sites must carry the excess energy ( a 2 − z 2 ) N (a^2-z^2)N ( a 2 − z 2 ) N , and the cheapest way to do it is a second spike of height ( a 2 − z 2 ) N \sqrt{(a^2-z^2)N} ( a 2 − z 2 ) N . Then
p N ( z N ) ≈ exp { − N [ Φ 1 ( z ) − a ] } , Φ 1 ( z ) = z + a 2 − z 2 . (42) p_N(z\sqrt{N}) \approx\exp\left\{-\sqrt{N}\left[\Phi_1(z)-a\right]\right\}, \qquad\Phi_1(z)=z+\sqrt{a^2-z^2}. \tag*{(42)} p N ( z N ) ≈ exp { − N [ Φ 1 ( z ) − a ] } , Φ 1 ( z ) = z + a 2 − z 2 . (42)
Φ 1 \Phi_1 Φ 1 is the ℓ 1 \ell^1 ℓ 1 norm of the two spikes at fixed ℓ 2 \ell^2 ℓ 2 norm a a a . It is maximal at z = a / 2 z=a/\sqrt{2} z = a / 2 , where the two spikes are equal: the configuration is at the neck of the tube joining site 1 to the site of the second spike. There Φ 1 = 2 a \Phi_1=\sqrt{2}a Φ 1 = 2 a , and p N ( a N / 2 ) ≈ e − b ( ε ) N p_N(a\sqrt{N/2})\approx e^{-b(\varepsilon)\sqrt{N}} p N ( a N /2 ) ≈ e − b ( ε ) N . Geometrically, the slice { q 1 = a N / 2 } \{q_1=a\sqrt{N/2}\} { q 1 = a N /2 } cuts all the N − 1 N-1 N − 1 tubes leaving the chamber of site 1, each at its neck. This is the valley between the bulk and the condensate peak visible in panel (c) of Figure 1 .
The rigorous version is:
Proposition IV.2 (the neck valley). Fix ε > 2 \varepsilon>2 ε > 2 and K > 0 K>0 K > 0 . Uniformly for ∣ q − a N / 2 ∣ ≤ K \lvert q-a\sqrt{N/2}\rvert\le K ∣ q − a N /2 ∣ ≤ K ,
p N ( q ) ≤ exp { − b ( ε ) N + O ( ln N ) } . (43) p_N(q)\le\exp\left\{-b(\varepsilon)\sqrt{N}+O(\ln N)\right\}. \tag*{(43)} p N ( q ) ≤ exp { − b ( ε ) N + O ( ln N ) } . (43)
Lemma IV.3 (a condensate exists). Fix ε > 2 \varepsilon>2 ε > 2 and K > 0 K>0 K > 0 . Then μ ( max i q i < a N / 2 + K ) ≤ exp { − b ( ε ) N + O ( ln N ) } \mu\left(\max_i q_i<a\sqrt{N/2}+K\right)\le\exp\left\{-b(\varepsilon)\sqrt{N}+O(\ln N)\right\} μ ( max i q i < a N /2 + K ) ≤ exp { − b ( ε ) N + O ( ln N ) } .
The proofs (Appendix C) use the same tools as Proposition III.8, plus one observation. Condition on the largest of the occupations q 2 , … , q N q_2,\ldots,q_N q 2 , … , q N , say q 2 = v q_2=v q 2 = v . All the other sites are then bounded by v v v , and on a bounded support the energy fugacity may be taken negative: this is the precise sense in which the saddle point sits on the branch cut β < 0 \beta<0 β < 0 of (17) . Lemma III.7 then bounds the residual partition function by exp { N − N F ( z 1 , z 2 ) + O ( ln N ) } \exp\{N-\sqrt{N}F(z_1,z_2)+O(\ln N)\} exp { N − N F ( z 1 , z 2 ) + O ( ln N )} , with
F ( z 1 , z 2 ) = z 1 + z 2 + ( a 2 − z 1 2 − z 2 2 ) + max ( z 2 , a / 2 ) , z i = q i / N , (44) F(z_1,z_2)=z_1+z_2+\frac{(a^2-z_1^2-z_2^2)_+}{\max(z_2,a/2)}, \qquad z_i=q_i/\sqrt{N}, \tag*{(44)} F ( z 1 , z 2 ) = z 1 + z 2 + max ( z 2 , a /2 ) ( a 2 − z 1 2 − z 2 2 ) + , z i = q i / N , (44)
with ( x ) + = max ( x , 0 ) (x)_+=\max(x,0) ( x ) + = max ( x , 0 ) as in Lemma III.7.
The exponent of (43) follows from an elementary inequality: at z 1 = a / 2 z_1=a/\sqrt{2} z 1 = a / 2 ,
min z 2 ≥ 0 F ( a 2 , z 2 ) = 2 a , (45) \min_{z_2\ge0}F\left(\frac{a}{\sqrt{2}},z_2\right)=\sqrt{2}a, \tag*{(45)} z 2 ≥ 0 min F ( 2 a , z 2 ) = 2 a , (45)
attained at z 2 = a / 2 z_2=a/\sqrt{2} z 2 = a / 2 , i.e. at the neck.
Proof of Theorem IV.1. Insert (43) with K = ℓ / 2 K=\ell/2 K = ℓ /2 in (41) , for any fixed ℓ \ell ℓ .
The exponent does not depend on ℓ \ell ℓ : it is a property of the neck. A wider ramp improves only the prefactor. Expanding Φ 1 \Phi_1 Φ 1 around its maximum, Φ 1 ′ ′ ( a / 2 ) = − 2 2 / a \Phi_1''(a/\sqrt{2})=-2\sqrt{2}/a Φ 1 ′′ ( a / 2 ) = − 2 2 / a , the valley of p N p_N p N is flat within a constant factor over ∣ q − a N / 2 ∣ ≲ N 1 / 4 |q-a\sqrt{N/2}|\lesssim N^{1/4} ∣ q − a N /2 ∣ ≲ N 1/4 , the window already found in (28) . Taking ℓ ≃ N 1 / 4 \ell\simeq N^{1/4} ℓ ≃ N 1/4 gains a factor N − 1 / 4 N^{-1/4} N − 1/4 .
and changing it between various “natural” choices changes the prefactor of the bound but not the exponent. So one should not take the prefactor seriously when comparing, for example, with Hamiltonian dynamics.
Remark IV.6 (the exponent). Theorems V.1 and IV.1 give
( 2 − 1 ) a ≤ lim inf N → ∞ − log λ 1 N ≤ lim sup N → ∞ − log λ 1 N ≤ a . (\sqrt{2}-1)a \le\liminf_{N\to\infty}\frac{-\log\lambda_1}{\sqrt{N}} \le\limsup_{N\to\infty}\frac{-\log\lambda_1}{\sqrt{N}} \le a. ( 2 − 1 ) a ≤ N → ∞ lim inf N − log λ 1 ≤ N → ∞ lim sup N − log λ 1 ≤ a .
Above ε c \varepsilon_c ε c the spectral gap therefore decays as e − Θ ( N ) e^{-\Theta(\sqrt{N})} e − Θ ( N ) : faster than any power of N N N , but slower than e − c N e^{-cN} e − c N for every c > 0 c>0 c > 0 . This should be compared with the transition in the lowest eigenvalue of the Laplacian on M M M described in [15 ] . The loss in the lower bound does not come from the neck. It comes from the transport of Steps 2–3 of § 0.7 , which must reach the configurations whose largest coordinate is close to ε \varepsilon ε (Lemma III.4): the density oscillation along a cap between q i ≈ 2 ε q_i \approx2\varepsilon q i ≈ 2 ε and q i ≈ T q_i \approx T q i ≈ T is e T e^{T} e T , while the neck costs only e b ( ε ) N e^{b(\varepsilon)\sqrt{N}} e b ( ε ) N . The identification of the exponent, presumably b ( ε ) b(\varepsilon) b ( ε ) from the upper bound, is left open.
Supersymmetric formulation. During our investigations, we have found that the whole discussion of the upper bound can be recast in the language of Witten’s supersymmetric quantum mechanics [19 ] . The substitution ψ = p φ \psi=\sqrt{p}\,\varphi ψ = p φ turns the Sturm–Liouville problems of this section into Schrödinger operators H = A † A H=A^\dagger A H = A † A with superpotential W = − ln p W=-\ln p W = − ln p , whose zero-energy state is the equilibrium. In the condensed phase, with ℏ = N − 1 / 2 \hbar=N^{-1/2} ℏ = N − 1/2 , the exponentially small eigenvalues arise from tunnelling between the wells through the neck, as in the semiclassical analysis of the Witten complex [35 , 36 , 37 ] would follow from boundary and cusp versions of the Eyring–Kramers formula [38 , 39 ] . We will present this point of view elsewhere.
Lower bounds and a unified Bochner criterion
The remaining Sections of the paper are devoted to the proof of the following:
Theorem V.1 (lower bounds).
(i) If 1 < ε < 2 1<\varepsilon<2 1 < ε < 2 and N ( 2 − ε ) 2 ≥ 2000 N(2-\varepsilon)^2\ge2000 N ( 2 − ε ) 2 ≥ 2000 , then λ 1 ( N , ε ) ≥ ( 2 − ε ) 2 / [ 100 ( ε − 1 ) ] \lambda_1(N,\varepsilon)\ge(2-\varepsilon)^2/[100(\varepsilon-1)] λ 1 ( N , ε ) ≥ ( 2 − ε ) 2 / [ 100 ( ε − 1 )] .
(ii) There are c > 0 c>0 c > 0 and N 0 N_0 N 0 such that λ 1 ( N , 2 ) ≥ c N − 4 e − 16 N 1 / 3 \lambda_1(N,2)\ge cN^{-4}e^{-16N^{1/3}} λ 1 ( N , 2 ) ≥ c N − 4 e − 16 N 1/3 for N ≥ N 0 N\ge N_0 N ≥ N 0 .
(iii) For fixed ε > 2 \varepsilon> 2 ε > 2 there are c ε , C ε > 0 c_{\varepsilon}, C_{\varepsilon} > 0 c ε , C ε > 0 and N ε N_{\varepsilon} N ε such that, for N ≥ N ε N \ge N_{\varepsilon} N ≥ N ε ,
λ 1 ( N , ε ) ≥ c ε N − C ε exp { − N ( ε − 2 ) − 16 ( N ε − 2 ) 1 / 4 } . (46) \lambda_1(N,\varepsilon) \ge c_{\varepsilon}N^{-C_{\varepsilon}}\exp\left\{-\sqrt{N(\varepsilon-2)}-16\left(\frac{N}{\varepsilon-2}\right)^{1/4}\right\}. \tag*{(46)} λ 1 ( N , ε ) ≥ c ε N − C ε exp { − N ( ε − 2 ) − 16 ( ε − 2 N ) 1/4 } . (46)
The statements (ii)–(iii) are asymptotic, because the caps G i G_i G i must be nonempty: at ε = 2 \varepsilon= 2 ε = 2 the condition T < t ∗ T < t^* T < t ∗ , i.e. 16 N 1 / 3 < 1 + N − 1 16N^{1/3} < 1+\sqrt{N-1} 16 N 1/3 < 1 + N − 1 , already requires N N N of order 1.7 ⋅ 10 7 1.7 \cdot10^7 1.7 ⋅ 1 0 7 .
Part (i) follows from a Bochner–Reilly criterion (Proposition V.2 ) applied to M M M with a single vector field (§ 0.6 C). For ε ≥ 2 \varepsilon\ge2 ε ≥ 2 the criterion is applied to the caps G i G_i G i , and the resulting Poincaré inequalities are transported to the sets U i U_i U i and glued (§ 0.6 D).
The Bochner–Reilly criterion
We will use the following version of Reilly’s application of the (integrated) Bochner’s Formula for domain with corners.
Proposition V.2 (Bochner–Reilly criterion with corners). Let D D D be a compact connected d d d -dimensional Riemannian domain with smooth corners, d ≥ 3 d \ge3 d ≥ 3 , and let F ⊂ D ‾ F \subset\overline{D} F ⊂ D be finite (possibly empty). Suppose that X X X is smooth on D ‾ ∖ F \overline{D} \setminus F D ∖ F , that ∣ X ∣ ≤ C dist ( ⋅ , F ) − 1 |X| \le C \operatorname{dist}(\cdot,F)^{-1} ∣ X ∣ ≤ C dist ( ⋅ , F ) − 1 near F F F , and, for some λ ∗ > 0 , \lambda_* > 0, λ ∗ > 0 ,
Ric + ( div X − ∣ X ∣ 2 ) g ≥ λ ∗ g on D ∖ F , (47) \operatorname{Ric}+(\operatorname{div}X-|X|^2)g \ge\lambda_*g \qquad\text{on } D \setminus F, \tag*{(47)} Ric + ( div X − ∣ X ∣ 2 ) g ≥ λ ∗ g on D ∖ F , (47) I I ≥ ⟨ X , ν ⟩ g on each open boundary face outside F . (48) \mathrm{II} \ge\langle X,\nu\rangle g \qquad\text{on each open boundary face outside } F. \tag*{(48)} II ≥ ⟨ X , ν ⟩ g on each open boundary face outside F . (48) Here ν \nu ν is outward, I I ( v , v ) = ⟨ ∇ v ν , v ⟩ \mathrm{II}(v,v)=\langle\nabla_v\nu,v\rangle II ( v , v ) = ⟨ ∇ v ν , v ⟩ , and (48) is an inequality on vectors tangent to the face. Then λ 1 ( D ) ≥ λ ∗ \lambda_1(D) \ge\lambda_* λ 1 ( D ) ≥ λ ∗ . If F = ∅ F=\varnothing F = ∅ , the restriction d ≥ 3 d \ge3 d ≥ 3 is unnecessary.
Proof. Integrals over D D D are taken with respect to its Riemannian measure σ = H d \sigma=\mathcal{H}^d σ = H d , and integrals over a boundary face with respect to H d − 1 \mathcal{H}^{d-1} H d − 1 . Let u u u be a first nonconstant variational Neumann eigenfunction, write λ 1 = λ 1 ( D ) \lambda_1=\lambda_1(D) λ 1 = λ 1 ( D ) , and set V = ∇ u V=\nabla u V = ∇ u . The regularity theorem of Mitrea–Taylor–Vasy [40 Proposition 3.2 and Corollary 4.2] , applied to the absolute one-form d u du d u , gives V ∈ H 1 ( D ) V \in H^1(D) V ∈ H 1 ( D ) : indeed d ( d u ) = 0 d(du)=0 d ( d u ) = 0 , δ d u = λ 1 u ∈ L 2 \delta du=\lambda_1u \in L^2 δ d u = λ 1 u ∈ L 2 , and ν ⌟ d u = 0 \nu\mathbin{\lrcorner}du=0 ν ┘ d u = 0 . The facewise Gaffney identity (the integrated Bochner–Weitzenböck formula; see [41 Theorem 3] for one-forms with zero normal component) gives
λ 1 ∫ D ∣ V ∣ 2 d σ = ∫ D ( ∣ ∇ V ∣ 2 + Ric ( V , V ) ) d σ + ∑ S ∫ S I I ( V , V ) d H d − 1 , (49) \lambda_1\int_D |V|^2\,d\sigma
=
\int_D\left(|\nabla V|^2+\operatorname{Ric}(V,V)\right)\,d\sigma
+\sum_S\int_S \mathrm{II}(V,V)\,d\mathcal{H}^{d-1}, \tag*{(49)} λ 1 ∫ D ∣ V ∣ 2 d σ = ∫ D ( ∣∇ V ∣ 2 + Ric ( V , V ) ) d σ + S ∑ ∫ S II ( V , V ) d H d − 1 , (49)
where the sum is over the open boundary faces S S S . For a smooth field tangent to every face, the same identity with left-hand side ∫ D ( ( div V ) 2 + ∣ d V ♭ ∣ 2 ) d σ \int_D \left((\operatorname{div} V)^2 + \lvert dV^\flat\rvert^2\right)\,d\sigma ∫ D ( ( div V ) 2 + ∣ d V ♭ ∣ 2 ) d σ follows by applying the divergence theorem to ( div V ) V − ∇ V V (\operatorname{div} V)V-\nabla_VV ( div V ) V − ∇ V V . This identity extends to H 1 H^1 H 1 tangential fields by density and the trace theorem: in a corner chart, extend each coordinate component oddly across its own face and evenly across the other faces, then mollify. Substitution of V = ∇ u V=\nabla u V = ∇ u yields (49) .
Integrating ∣ ∇ V + Z ⊗ V ∣ 2 ≥ 0 \lvert\nabla V+Z\otimes V\rvert^2\ge0 ∣ ∇ V + Z ⊗ V ∣ 2 ≥ 0 , for any Lipschitz field Z Z Z , gives
∫ D ∣ ∇ V ∣ 2 d σ ≥ ∫ D ( div Z − ∣ Z ∣ 2 ) ∣ V ∣ 2 d σ − ∑ S ∫ S ⟨ Z , ν ⟩ ∣ V ∣ 2 d H d − 1 . (50) \int_D \lvert\nabla V\rvert^2\,d\sigma\ge\int_D \left(\operatorname{div} Z-\lvert Z\rvert^2\right)\lvert V\rvert^2\,d\sigma-\sum_S\int_S \langle Z,\nu\rangle\lvert V\rvert^2\,d\mathcal{H}^{d-1}. \tag*{(50)} ∫ D ∣ ∇ V ∣ 2 d σ ≥ ∫ D ( div Z − ∣ Z ∣ 2 ) ∣ V ∣ 2 d σ − S ∑ ∫ S ⟨ Z , ν ⟩ ∣ V ∣ 2 d H d − 1 . (50)
If F = ∅ F=\varnothing F = ∅ , take Z = X Z=X Z = X in (50) and combine (49) with (47) –(48) . Otherwise put r = dist ( ⋅ , F ) r=\operatorname{dist}(\mathord{\cdot},F) r = dist ( ⋅ , F ) and Z = ζ δ X Z=\zeta_\delta X Z = ζ δ X , where ζ δ = 0 \zeta_\delta=0 ζ δ = 0 on r < δ r<\delta r < δ , ζ δ = 1 \zeta_\delta=1 ζ δ = 1 on r > 2 δ r>2\delta r > 2 δ , 0 ≤ ζ δ ≤ 1 0\le\zeta_\delta\le1 0 ≤ ζ δ ≤ 1 , and ∣ ∇ ζ δ ∣ ≤ C / δ \lvert\nabla\zeta_\delta\rvert\le C/\delta ∣ ∇ ζ δ ∣ ≤ C / δ . Since Ricci curvature and the face second fundamental forms are bounded, (47) –(50) imply
λ 1 ∫ D ∣ V ∣ 2 d σ ≥ λ ∗ ∫ D ∣ V ∣ 2 d σ − C ∫ D ∩ { r < 2 δ } ∣ V ∣ 2 d σ − C ∫ D ∩ { δ < r < 2 δ } r − 2 ∣ V ∣ 2 d σ − C ∑ S ∫ S ∩ { r < 2 δ } ∣ V ∣ 2 d H d − 1 . (51) \lambda_1\int_D\lvert V\rvert^2\,d\sigma\ge\lambda_*\int_D\lvert V\rvert^2\,d\sigma-C\int_{D\cap\{r<2\delta\}}\lvert V\rvert^2\,d\sigma-C\int_{D\cap\{\delta<r<2\delta\}}r^{-2}\lvert V\rvert^2\,d\sigma-C\sum_S\int_{S\cap\{r<2\delta\}}\lvert V\rvert^2\,d\mathcal{H}^{d-1}. \tag*{(51)} λ 1 ∫ D ∣ V ∣ 2 d σ ≥ λ ∗ ∫ D ∣ V ∣ 2 d σ − C ∫ D ∩ { r < 2 δ } ∣ V ∣ 2 d σ − C ∫ D ∩ { δ < r < 2 δ } r − 2 ∣ V ∣ 2 d σ − C S ∑ ∫ S ∩ { r < 2 δ } ∣ V ∣ 2 d H d − 1 . (51)
The unweighted volume and face error integrals in (51) tend to zero by integrability and the trace theorem. For the integral with weight r − 2 r^{-2} r − 2 , use the local point-singularity Hardy inequality: for p ∈ D ‾ p\in\overline D p ∈ D and sufficiently small fixed a > 0 a>0 a > 0 ,
∫ D ∩ B a ( p ) ∣ V ∣ 2 dist ( q , p ) 2 d σ ( q ) ≤ C a ∥ V ∥ H 1 ( D ) 2 . (52) \int_{D\cap B_a(p)}\frac{\lvert V\rvert^2}{\operatorname{dist}(q,p)^2}\,d\sigma(q)\le Ca\lVert V\rVert_{H^1(D)}^2. \tag*{(52)} ∫ D ∩ B a ( p ) dist ( q , p ) 2 ∣ V ∣ 2 d σ ( q ) ≤ C a ∥ V ∥ H 1 ( D ) 2 . (52)
Estimate (52) follows from the Euclidean Hardy inequality [42 Section 1.2.3] , after a cutoff and a bounded H 1 H^1 H 1 extension in a corner chart [43 Chapter VI, Section 3] . It requires d ≥ 3 d\ge3 d ≥ 3 but no vanishing boundary trace; the H 1 H^1 H 1 norm includes the necessary L 2 L^2 L 2 term. Since F F F is finite, (52) makes r − 2 ∣ V ∣ 2 r^{-2}\lvert V\rvert^2 r − 2 ∣ V ∣ 2 locally integrable near F F F . The weighted integral in (51) therefore tends to zero as δ ↓ 0 \delta\downarrow0 δ ↓ 0 , proving λ 1 ≥ λ ∗ \lambda_1\ge\lambda_* λ 1 ≥ λ ∗ .
On the microcanonical surface. Proposition V.2 is applied to D = M D=M D = M for 1 < ε < 2 1<\varepsilon<2 1 < ε < 2 , and to D = G i D=G_i D = G i for ε ≥ 2 \varepsilon\ge2 ε ≥ 2 . By Lemma III.1(2)–(3) its two hypotheses become scalar conditions:
N − 3 R 2 + div X − ∣ X ∣ 2 ≥ λ ∗ , (53) \frac{N-3}{R^2}+\operatorname{div}X-\lvert X\rvert^2\ge\lambda_*, \tag*{(53)} R 2 N − 3 + div X − ∣ X ∣ 2 ≥ λ ∗ , (53)
⟨ X , ∇ q j ⟩ ≥ 1 − K R 2 on a face { q j = K } of { q j ≥ K } . (54) \langle X,\nabla q_j\rangle\ge\frac{1-K}{R^2}\quad\text{on a face }\{q_j=K\}\text{ of }\{q_j\ge K\}. \tag*{(54)} ⟨ X , ∇ q j ⟩ ≥ R 2 1 − K on a face { q j = K } of { q j ≥ K } . (54)
On the faces { q j = 0 } \{q_j = 0\} { q j = 0 } (K = 0 K = 0 K = 0 ), (54) reads ⟨ X , ∇ q j ⟩ ≥ 1 / R 2 \langle X, \nabla q_j\rangle\ge1/R^2 ⟨ X , ∇ q j ⟩ ≥ 1/ R 2 : since ∇ q j \nabla q_j ∇ q j points into the domain, X X X must enter through each of these faces with normal component at least 1 / ( R 2 ∣ ∇ q j ∣ ) 1/(R^2|\nabla q_j|) 1/ ( R 2 ∣∇ q j ∣ ) , which compensates their concavity. On the convex face { q i = T } \{q_i = T\} { q i = T } of G i G_i G i (K = T > 1 K = T > 1 K = T > 1 ), (54) holds for every field tangent to the face.
The cost of the concavity. Integrating (53) over M M M , with the divergence theorem, (54) on the faces { q j = 0 } \{q_j = 0\} { q j = 0 } , and the coarea identity ∫ { q j = 0 } ∣ ∇ q j ∣ − 1 d H N − 3 = σ ( M ) p N ( 0 ) \int_{\{q_j=0\}}|\nabla q_j|^{-1}d\mathcal{H}^{N-3}=\sigma(M)p_N(0) ∫ { q j = 0 } ∣∇ q j ∣ − 1 d H N − 3 = σ ( M ) p N ( 0 ) , where p N p_N p N is the one-site marginal of (30) , gives for every field allowed in Proposition V.2
λ ∗ ≤ N − 3 R 2 − N p N ( 0 ) R 2 − ∫ M ∣ X ∣ 2 d μ ≤ 1 − p N ( 0 ) ε − 1 . (55) \lambda_* \le\frac{N-3}{R^2}-\frac{Np_N(0)}{R^2}-\int_M |X|^2\,d\mu\le\frac{1-p_N(0)}{\varepsilon-1}. \tag*{(55)} λ ∗ ≤ R 2 N − 3 − R 2 N p N ( 0 ) − ∫ M ∣ X ∣ 2 d μ ≤ ε − 1 1 − p N ( 0 ) . (55)
As N → ∞ N \to\infty N → ∞ , p N ( 0 ) p_N(0) p N ( 0 ) tends to the density at 0 0 0 of the one-site law of Section III C: this limit is smaller than 1 1 1 for ε < 2 \varepsilon< 2 ε < 2 , with 1 − lim N p N ( 0 ) = ( 2 − ε ) / 2 + O ( ( 2 − ε ) 2 ) 1-\lim_N p_N(0)=(2-\varepsilon)/2+O((2-\varepsilon)^2) 1 − lim N p N ( 0 ) = ( 2 − ε ) /2 + O (( 2 − ε ) 2 ) , and equals 1 1 1 for ε ≥ 2 \varepsilon\ge2 ε ≥ 2 . A single field on M M M can therefore give an N N N -uniform bound only for ε < 2 \varepsilon< 2 ε < 2 . For ε ≥ 2 \varepsilon\ge2 ε ≥ 2 the criterion is applied to the caps G i G_i G i , whose slices are below the critical energy (Remark III.3).
Corners
Proposition V.2 is stated for compact domains with smooth corners, i.e. points with a neighbourhood diffeomorphic to a neighbourhood of 0 0 0 in [ 0 , ∞ ) m × R d − m [0,\infty)^m \times\mathbb{R}^{d-m} [ 0 , ∞ ) m × R d − m . The following observation proves that smooth corners can fail only where the faces through a point have linearly dependent gradients and that this happens only at the points of F ε : = { q ∈ M : q j ∈ { 0 , ε } ∀ j } F_\varepsilon:= \{q \in M : q_j \in\{0,\varepsilon\}\ \forall j\} F ε := { q ∈ M : q j ∈ { 0 , ε } ∀ j } , the configurations with k k k coordinates equal to ε \varepsilon ε and the others equal to 0 0 0 , which exist only at the exceptional energies N / ε = k ∈ Z N/\varepsilon= k \in\mathbb{Z} N / ε = k ∈ Z , and on the face { q i = K } \{q_i = K\} { q i = K } of a cap for finitely many levels K K K .
Proposition V.3 (corners of the domains). (a) Every q ∈ M ∖ F ε q \in M \setminus F_\varepsilon q ∈ M ∖ F ε has a neighbourhood in M M M diffeomorphic to a neighbourhood of 0 0 0 in [ 0 , ∞ ) m × R d − m [0,\infty)^m \times\mathbb{R}^{d-m} [ 0 , ∞ ) m × R d − m , m m m being the number of coordinates vanishing at q q q . In particular, if N / ε ∉ Z N/\varepsilon\notin\mathbb{Z} N / ε ∈ / Z , M M M is a compact connected domain with smooth corners.
(b) For K ∈ ( ε , t ∗ ) K \in(\varepsilon,t^*) K ∈ ( ε , t ∗ ) outside a set of at most 2 N 2N 2 N values, the cap { q ∈ M : q i ≥ K } \{q \in M:q_i \ge K\} { q ∈ M : q i ≥ K } is a compact connected domain with smooth corners, and it contains no point of F ε F_\varepsilon F ε . This holds for G i G_i G i , by the choice of T T T in (8) .
(c) If N / ε = k ∈ Z N/\varepsilon= k \in\mathbb{Z} N / ε = k ∈ Z with k ≤ N − 2 k \le N-2 k ≤ N − 2 , M M M is not a domain with corners at the points p ∈ F ε p \in F_{\varepsilon} p ∈ F ε : on M M M
∑ j : p j = 0 q j = ∣ q − p ∣ 2 2 ε , (56) \sum_{j:p_j=0} q_j = \frac{|q-p|^2}{2\varepsilon}, \tag*{(56)} j : p j = 0 ∑ q j = 2 ε ∣ q − p ∣ 2 , (56)
so every curve in M M M starting at p p p is tangent to the ( k − 1 ) (k-1) ( k − 1 ) -dimensional subspace { v ∈ T p : v j = 0 if p j = 0 } \{v \in T_p: v_j=0\text{ if }p_j=0\} { v ∈ T p : v j = 0 if p j = 0 } , and k − 1 < d k-1<d k − 1 < d . (For k = N − 1 k=N-1 k = N − 1 , M = S R M=S_R M = S R is a round sphere.)
The proof is in Appendix D.
Part (b) also covers the slices of a cap at exceptional energies: at a level t t t with n / e ( t ) ∈ Z n/e(t)\in\mathbb{Z} n / e ( t ) ∈ Z , the faces of the slice Σ i ( t ) ≃ M n , e ( t ) \Sigma_i(t)\simeq M_{n,e(t)} Σ i ( t ) ≃ M n , e ( t ) meet non-transversally at the points of c ( t ) F e ( t ) c(t)F_{e(t)} c ( t ) F e ( t ) , but these points are not in F ε F_{\varepsilon} F ε , since q i = t > ε q_i=t>\varepsilon q i = t > ε , so they are regular corner points of the cap. In Step 1 they are only isolated singular points of the vector field, which Proposition V.2 allows. At the exceptional energies N / ε ∈ Z N/\varepsilon\in\mathbb{Z} N / ε ∈ Z of M M M itself, part (c) shows that Proposition V.2 cannot be applied to M M M ; Theorem V.1 (i) is obtained there by approximating the energy from below (Section VII).
The global field for 1 < ε < 2 1<\varepsilon<2 1 < ε < 2
The vector field X X X is built from the profile
φ ( x ) = ( 2 − ε ) x ( x − ε ) 2 ( 2 − ε ) x + 8 ε , x ≥ 0 , (57) \varphi(x)=\frac{(2-\varepsilon)x(x-\varepsilon)^2}{(2-\varepsilon)x+8\varepsilon},\qquad x\ge0, \tag*{(57)} φ ( x ) = ( 2 − ε ) x + 8 ε ( 2 − ε ) x ( x − ε ) 2 , x ≥ 0 , (57)
through the averages
H = ⟨ φ ⟩ , J = ( ε − 1 ) φ ′ ( 0 ) + ⟨ ( x − ε ) φ ′ ⟩ , X = ∇ H J . (58) H=\langle\varphi\rangle,\qquad J=(\varepsilon-1)\varphi'(0)+\langle(x-\varepsilon)\varphi'\rangle,\qquad X=\frac{\nabla H}{J}. \tag*{(58)} H = ⟨ φ ⟩ , J = ( ε − 1 ) φ ′ ( 0 ) + ⟨( x − ε ) φ ′ ⟩ , X = J ∇ H . (58)
Since φ ≥ 0 \varphi\ge0 φ ≥ 0 vanishes only at 0 0 0 and ε \varepsilon ε , H > 0 H>0 H > 0 on M ∖ F ε M\setminus F_{\varepsilon} M ∖ F ε .
Proposition V.4 (the global field). Let 1 < ε < 2 1<\varepsilon<2 1 < ε < 2 and N ( 2 − ε ) 2 ≥ 2000 N(2-\varepsilon)^2\ge2000 N ( 2 − ε ) 2 ≥ 2000 . The field X = ∇ H / J X=\nabla H/J X = ∇ H / J of (58) is smooth on M ∖ F ε M\setminus F_{\varepsilon} M ∖ F ε , and there
⟨ X , ∇ q j ⟩ = 1 R 2 on each face { q j = 0 } , (59) \langle X,\nabla q_j\rangle=\frac{1}{R^2}\qquad\text{on each face }\{q_j=0\}, \tag*{(59)} ⟨ X , ∇ q j ⟩ = R 2 1 on each face { q j = 0 } , (59)
N − 3 R 2 + div X − ∣ X ∣ 2 ≥ ( 2 − ε ) 2 100 ( ε − 1 ) , (60) \frac{N-3}{R^2}+\operatorname{div}X-|X|^2\ge\frac{(2-\varepsilon)^2}{100(\varepsilon-1)}, \tag*{(60)} R 2 N − 3 + div X − ∣ X ∣ 2 ≥ 100 ( ε − 1 ) ( 2 − ε ) 2 , (60)
∣ X ∣ ≤ 2 ( N H ) − 1 / 2 . (61) |X|\le2(NH)^{-1/2}. \tag*{(61)} ∣ X ∣ ≤ 2 ( N H ) − 1/2 . (61)
In particular, if F ε = ∅ F_{\varepsilon}=\varnothing F ε = ∅ , Proposition V.2 applies with D = M D=M D = M , F = ∅ F=\varnothing F = ∅ and λ ∗ = ( 2 − ε ) 2 / [ 100 ( ε − 1 ) ] \lambda_{*}=(2-\varepsilon)^2/[100(\varepsilon-1)] λ ∗ = ( 2 − ε ) 2 / [ 100 ( ε − 1 )] , and gives λ 1 ( N , ε ) ≥ ( 2 − ε ) 2 / [ 100 ( ε − 1 ) ] \lambda_{1}(N,\varepsilon)\ge(2-\varepsilon)^2/[100(\varepsilon-1)] λ 1 ( N , ε ) ≥ ( 2 − ε ) 2 / [ 100 ( ε − 1 )] .
Thus (59) is (54) with equality, (60) is (53) , and (61) controls X X X near the zeros of H H H .
To prove the Proposition we will need the following elementary properties of the function φ \varphi φ , which will be proved in Appendix § 0.5.5 . For x ≥ 0 x\ge0 x ≥ 0 :
( φ 1 ) 0 ≤ φ ( x ) ≤ ( x − ε ) 2 , φ ′ ( 0 ) = ε ( 2 − ε ) / 8 and ∣ φ ′ ′ ( x ) ∣ ≤ 2 ; (\varphi1)\quad0\le\varphi(x)\le(x-\varepsilon)^2,\quad\varphi'(0)=\varepsilon(2-\varepsilon)/8\quad\text{and}\quad|\varphi''(x)|\le2; ( φ 1 ) 0 ≤ φ ( x ) ≤ ( x − ε ) 2 , φ ′ ( 0 ) = ε ( 2 − ε ) /8 and ∣ φ ′′ ( x ) ∣ ≤ 2 ;
( φ 2 ) 2 φ ( x ) ≤ ( x − ε ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − x ε ) ) ≤ 13 4 φ ( x ) ; (\varphi2)\quad2\varphi(x)\le(x-\varepsilon)\left(\varphi'(x)-\varphi'(0)\left(1-\frac{x}{\varepsilon}\right)\right)\le\frac{13}{4}\varphi(x); ( φ 2 ) 2 φ ( x ) ≤ ( x − ε ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − ε x ) ) ≤ 4 13 φ ( x ) ;
( φ 3 ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − x ε ) ) 2 ≤ 13 4 ( x − ε ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − x ε ) ) ; (\varphi3)\quad\left(\varphi'(x)-\varphi'(0)\left(1-\frac{x}{\varepsilon}\right)\right)^2\le\frac{13}{4}(x-\varepsilon)\left(\varphi'(x)-\varphi'(0)\left(1-\frac{x}{\varepsilon}\right)\right); ( φ 3 ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − ε x ) ) 2 ≤ 4 13 ( x − ε ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − ε x ) ) ;
( φ 4 ) ( φ ′ ( x ) + ( x − ε ) φ ′ ′ ( x ) − ( φ ′ ( 0 ) − ε φ ′ ′ ( 0 ) ) ( 1 − x ε ) ) 2 ≤ 144 φ ( x ) ; (\varphi4)\quad\left(\varphi'(x)+(x-\varepsilon)\varphi''(x)-\left(\varphi'(0)-\varepsilon\varphi''(0)\right)\left(1-\frac{x}{\varepsilon}\right)\right)^2\le144\varphi(x); ( φ 4 ) ( φ ′ ( x ) + ( x − ε ) φ ′′ ( x ) − ( φ ′ ( 0 ) − ε φ ′′ ( 0 ) ) ( 1 − ε x ) ) 2 ≤ 144 φ ( x ) ;
( φ 5 ) for every probability measure d ν on [ 0 , ∞ ) with ∫ x d ν = 1 and ∫ x 2 d ν = ε , φ ′ ( 0 ) + ∫ ( φ ′ ′ − φ ′ ) d ν ≥ 3 ( 2 − ε ) 2 16 ε . \begin{aligned}
(\varphi5)\quad&\text{for every probability measure }d\nu\text{ on }[0,\infty)\text{ with }\int x\,d\nu=1\text{ and }\int x^2\,d\nu=\varepsilon,\\
&\varphi'(0)+\int(\varphi''-\varphi')\,d\nu\ge\frac{3(2-\varepsilon)^2}{16\varepsilon}.
\end{aligned} ( φ 5 ) for every probability measure d ν on [ 0 , ∞ ) with ∫ x d ν = 1 and ∫ x 2 d ν = ε , φ ′ ( 0 ) + ∫ ( φ ′′ − φ ′ ) d ν ≥ 16 ε 3 ( 2 − ε ) 2 .
Proof. All statements are on M ∖ F ε M\setminus F_{\varepsilon} M ∖ F ε , where H > 0 H>0 H > 0 .
The normalisation. Since ⟨ x ⟩ = 1 \langle x\rangle=1 ⟨ x ⟩ = 1 and ⟨ x 2 ⟩ = ε \langle x^2\rangle=\varepsilon ⟨ x 2 ⟩ = ε , we have ⟨ ( x − ε ) ( 1 − x / ε ) ⟩ = 1 − ε \langle(x-\varepsilon)(1-x/\varepsilon)\rangle=1-\varepsilon ⟨( x − ε ) ( 1 − x / ε )⟩ = 1 − ε , so that J = ⟨ ( x − ε ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − x / ε ) ) ⟩ J=\left\langle(x-\varepsilon)\left(\varphi'(x)-\varphi'(0)(1-x/\varepsilon)\right)\right\rangle J = ⟨ ( x − ε ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − x / ε ) ) ⟩ . By (φ2), and by H ≤ ⟨ ( x − ε ) 2 ⟩ = ε ( ε − 1 ) < 2 ( ε − 1 ) H\le\langle(x-\varepsilon)^2\rangle=\varepsilon(\varepsilon-1)<2(\varepsilon-1) H ≤ ⟨( x − ε ) 2 ⟩ = ε ( ε − 1 ) < 2 ( ε − 1 ) , which follows from (φ1),
2 H ≤ J ≤ 13 4 H < 13 2 ( ε − 1 ) . (62) 2H\le J\le\frac{13}{4}H<\frac{13}{2}(\varepsilon-1). \tag*{(62)} 2 H ≤ J ≤ 4 13 H < 2 13 ( ε − 1 ) . (62)
In particular J > 0 J>0 J > 0 , so X X X is smooth on M ∖ F ε M\setminus F_{\varepsilon} M ∖ F ε .
The flux (59) . On { q j = 0 } \{q_j=0\} { q j = 0 } , Lemma III.1(4) with ψ = φ \psi=\varphi ψ = φ gives ⟨ ∇ H , ∇ q j ⟩ = 1 N ( φ ′ ( 0 ) + ⟨ ( x − ε ) φ ′ ⟩ / ( ε − 1 ) ) = J / R 2 \langle\nabla H,\nabla q_j\rangle=\frac{1}{N}\left(\varphi'(0)+\langle(x-\varepsilon)\varphi'\rangle/(\varepsilon-1)\right)=J/R^2 ⟨ ∇ H , ∇ q j ⟩ = N 1 ( φ ′ ( 0 ) + ⟨( x − ε ) φ ′ ⟩ / ( ε − 1 ) ) = J / R 2 , because R 2 = N ( ε − 1 ) R^2=N(\varepsilon-1) R 2 = N ( ε − 1 ) . Dividing by J J J gives (59) .
Gradients and the size (61) . Apply Lemma III.1(4) to H = ⟨ φ ⟩ H=\langle\varphi\rangle H = ⟨ φ ⟩ with α + β x = φ ′ ( 0 ) ( 1 − x / ε ) \alpha+\beta x=\varphi'(0)(1-x/\varepsilon) α + β x = φ ′ ( 0 ) ( 1 − x / ε ) , and to J = ⟨ ψ ⟩ J=\langle\psi\rangle J = ⟨ ψ ⟩ , ψ ( x ) = ( x − ε ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − x / ε ) ) \psi(x)=(x-\varepsilon)\left(\varphi'(x)-\varphi'(0)(1-x/\varepsilon)\right) ψ ( x ) = ( x − ε ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − x / ε ) ) , with α + β x = ψ ′ ( 0 ) ( 1 − x / ε ) \alpha+\beta x=\psi'(0)(1-x/\varepsilon) α + β x = ψ ′ ( 0 ) ( 1 − x / ε ) . Since ψ ′ ( x ) − ψ ′ ( 0 ) ( 1 − x / ε ) = φ ′ ( x ) + ( x − ε ) φ ′ ′ ( x ) − ( φ ′ ( 0 ) − ε φ ′ ′ ( 0 ) ) ( 1 − x / ε ) \psi'(x)-\psi'(0)(1-x/\varepsilon)=\varphi'(x)+(x-\varepsilon)\varphi''(x)-(\varphi'(0)-\varepsilon\varphi''(0))(1-x/\varepsilon) ψ ′ ( x ) − ψ ′ ( 0 ) ( 1 − x / ε ) = φ ′ ( x ) + ( x − ε ) φ ′′ ( x ) − ( φ ′ ( 0 ) − ε φ ′′ ( 0 )) ( 1 − x / ε ) , properties (φ3) and (φ4) give, using H ≤ J / 2 H\le J/2 H ≤ J /2 ,
∣ ∇ H ∣ 2 ≤ 13 4 J N , ∣ ∇ J ∣ 2 ≤ 144 H N ≤ 72 J N . (63) |\nabla H|^2\le\frac{13}{4}\frac{J}{N},\qquad|\nabla J|^2\le144\frac{H}{N}\le72\frac{J}{N}. \tag*{(63)} ∣∇ H ∣ 2 ≤ 4 13 N J , ∣∇ J ∣ 2 ≤ 144 N H ≤ 72 N J . (63)
With J ≥ 2 H J\ge2H J ≥ 2 H , the first bound gives ∣ X ∣ 2 = ∣ ∇ H ∣ 2 / J 2 ≤ 13 / ( 8 N H ) < 4 / ( N H ) |X|^2=|\nabla H|^2/J^2\le13/(8NH)<4/(NH) ∣ X ∣ 2 = ∣∇ H ∣ 2 / J 2 ≤ 13/ ( 8 N H ) < 4/ ( N H ) , which is (61) .
The interior inequality (60) . Lemma III.1(5) gives J + ( ε − 1 ) Δ H = ( ε − 1 ) { φ ′ ( 0 ) + ⟨ φ ′ ′ − φ ′ ⟩ } + R φ J+(\varepsilon-1)\Delta H=(\varepsilon-1)\{\varphi'(0)+\langle\varphi''-\varphi'\rangle\}+\mathcal{R}_{\varphi} J + ( ε − 1 ) Δ H = ( ε − 1 ) { φ ′ ( 0 ) + ⟨ φ ′′ − φ ′ ⟩} + R φ , where ∣ R φ ∣ ≤ 8 ( ε − 1 ) / N |\mathcal{R}_{\varphi}|\le8(\varepsilon-1)/N ∣ R φ ∣ ≤ 8 ( ε − 1 ) / N by (\varphi1). Applying (\varphi5) to the measure d ν ( x ) = 1 N ∑ j δ ( x − q j ) d x \mathrm{d}\nu(x)=\frac{1}{N}\sum_j\delta(x-q_j)\,\mathrm{d}x d ν ( x ) = N 1 ∑ j δ ( x − q j ) d x , whose first two moments are 1 1 1 and ε \varepsilon ε , and using ε < 2 \varepsilon<2 ε < 2 ,
J + ( ε − 1 ) Δ H ≥ ( ε − 1 ) ( 3 ( 2 − ε ) 2 32 − 8 N ) . (64) J+(\varepsilon-1)\Delta H\ge(\varepsilon-1)\left(\frac{3(2-\varepsilon)^2}{32}-\frac{8}{N}\right). \tag*{(64)} J + ( ε − 1 ) Δ H ≥ ( ε − 1 ) ( 32 3 ( 2 − ε ) 2 − N 8 ) . (64)
Since X = ∇ H / J X=\nabla H/J X = ∇ H / J and R 2 = N ( ε − 1 ) R^2=N(\varepsilon-1) R 2 = N ( ε − 1 ) ,
N − 3 R 2 + div X − ∣ X ∣ 2 = J + ( ε − 1 ) Δ H ( ε − 1 ) J − 3 N ( ε − 1 ) − ⟨ ∇ H , ∇ J ⟩ + ∣ ∇ H ∣ 2 J 2 . (65) \frac{N-3}{R^2}+\operatorname{div}X-|X|^2=\frac{J+(\varepsilon-1)\Delta H}{(\varepsilon-1)J}-\frac{3}{N(\varepsilon-1)}-\frac{\langle\nabla H,\nabla J\rangle+|\nabla H|^2}{J^2}. \tag*{(65)} R 2 N − 3 + div X − ∣ X ∣ 2 = ( ε − 1 ) J J + ( ε − 1 ) Δ H − N ( ε − 1 ) 3 − J 2 ⟨ ∇ H , ∇ J ⟩ + ∣∇ H ∣ 2 . (65)
By (62) , 1 / ( ε − 1 ) < 13 / ( 2 J ) 1/(\varepsilon-1)<13/(2J) 1/ ( ε − 1 ) < 13/ ( 2 J ) ; by (63) , ∣ ∇ H ∣ ∣ ∇ J ∣ ≤ ( 13 / 4 ) ⋅ 72 J / N < 16 J / N |\nabla H||\nabla J|\le\sqrt{(13/4)\cdot72}\,J/N<16J/N ∣∇ H ∣∣∇ J ∣ ≤ ( 13/4 ) ⋅ 72 J / N < 16 J / N , since ( 13 / 4 ) ⋅ 72 = 234 < 256 (13/4)\cdot72=234<256 ( 13/4 ) ⋅ 72 = 234 < 256 . Hence the last two terms in (65) satisfy
3 N ( ε − 1 ) < 20 N J , ⟨ ∇ H , ∇ J ⟩ + ∣ ∇ H ∣ 2 J 2 < 16 + 4 N J = 20 N J . (66) \frac{3}{N(\varepsilon-1)}<\frac{20}{NJ},\qquad\frac{\langle\nabla H,\nabla J\rangle+|\nabla H|^2}{J^2}<\frac{16+4}{NJ}=\frac{20}{NJ}. \tag*{(66)} N ( ε − 1 ) 3 < N J 20 , J 2 ⟨ ∇ H , ∇ J ⟩ + ∣∇ H ∣ 2 < N J 16 + 4 = N J 20 . (66)
Inserting (64) and (66) in (65) ,
N − 3 R 2 + div X − ∣ X ∣ 2 ≥ 1 J ( 3 ( 2 − ε ) 2 32 − 48 N ) . \frac{N-3}{R^2}+\operatorname{div}X-|X|^2\ge\frac{1}{J}\left(\frac{3(2-\varepsilon)^2}{32}-\frac{48}{N}\right). R 2 N − 3 + div X − ∣ X ∣ 2 ≥ J 1 ( 32 3 ( 2 − ε ) 2 − N 48 ) .
Since N ( 2 − ε ) 2 ≥ 2000 N(2-\varepsilon)^2\ge2000 N ( 2 − ε ) 2 ≥ 2000 , 48 / N < 50 / N ≤ ( 2 − ε ) 2 / 40 48/N<50/N\le(2-\varepsilon)^2/40 48/ N < 50/ N ≤ ( 2 − ε ) 2 /40 , and 3 / 32 − 1 / 40 = 11 / 160 > 13 / 200 3/32-1/40=11/160>13/200 3/32 − 1/40 = 11/160 > 13/200 . With J < 13 2 ( ε − 1 ) J<\frac{13}{2}(\varepsilon-1) J < 2 13 ( ε − 1 ) this gives
N − 3 R 2 + div X − ∣ X ∣ 2 ≥ 13 200 ( 2 − ε ) 2 J ≥ ( 2 − ε ) 2 100 ( ε − 1 ) . \frac{N-3}{R^2}+\operatorname{div}X-|X|^2\ge\frac{13}{200}\frac{(2-\varepsilon)^2}{J}\ge\frac{(2-\varepsilon)^2}{100(\varepsilon-1)}. R 2 N − 3 + div X − ∣ X ∣ 2 ≥ 200 13 J ( 2 − ε ) 2 ≥ 100 ( ε − 1 ) ( 2 − ε ) 2 .
The last claim. If F ε = ∅ F_\varepsilon=\varnothing F ε = ∅ , X X X is smooth on M M M , and M M M is a domain with smooth corners by Proposition V.3 (a); Proposition V.2 applies with F = ∅ F=\varnothing F = ∅ . □
Proposition V.4 proves Theorem V.1 (i) when N / ε ∉ Z N/\varepsilon\notin\mathbb{Z} N / ε ∈ / Z ; Section VII handles the exceptional energies. Figure 5 shows the field X X X on the two-dimensional surface M 4 , 1.7 M_{4,1.7} M 4 , 1.7 .
Figure 5. The global field X = ∇ H / J X=\nabla H/J X = ∇ H / J of Proposition V.4 on the two-dimensional surface M 4 , 1.7 \mathcal{M}_{4,1.7} M 4 , 1.7 . The field crosses the concave faces { q j = 0 } \{q_j=0\} { q j = 0 } inwards, with the flux (59) that compensates their concavity.
The route from local estimates to the global gap
Recall the cells and the caps of Section III A,
A i = { q ∈ M ∘ : q i > q j ∀ j ≠ i } , U i = { q ∈ M ∘ : q i > ε } , A_i=\{q\in M^\circ:q_i>q_j\ \forall j\ne i\},\qquad U_i=\{q\in M^\circ:q_i>\varepsilon\}, A i = { q ∈ M ∘ : q i > q j ∀ j = i } , U i = { q ∈ M ∘ : q i > ε } ,
H i = { q ∈ M ∘ : q i > 2 ε } , G i = { q ∈ M : q i ≥ T } , H_i=\{q\in M^\circ:q_i>2\varepsilon\},\qquad G_i=\{q\in M:q_i\ge T\}, H i = { q ∈ M ∘ : q i > 2 ε } , G i = { q ∈ M : q i ≥ T } ,
where T T T , given by (8) , lies slightly above N ( ε − 2 ) \sqrt{N(\varepsilon-2)} N ( ε − 2 ) for ε > 2 \varepsilon>2 ε > 2 and equals 16 N 1 / 3 16N^{1/3} 16 N 1/3 up to 1 1 1 for ε = 2 \varepsilon=2 ε = 2 ; thus A i ⊂ U i A_i\subset U_i A i ⊂ U i and G i ⊂ H i ⊂ U i G_i\subset H_i\subset U_i G i ⊂ H i ⊂ U i . For ε ≥ 2 \varepsilon\ge2 ε ≥ 2 the proof bounds the Poincaré constants P G = P ( G i ) P_G=P(G_i) P G = P ( G i ) , P H = P ( H i ) P_H=P(H_i) P H = P ( H i ) , P U = P ( U i ) P_U=P(U_i) P U = P ( U i ) in four steps:
Step 1. Lift the vector field X X X in (58) to the slices of G i G_i G i . The cap estimate (74) gives P G ≤ C ε N 1 / 2 P_G \le C_{\varepsilon}N^{1/2} P G ≤ C ε N 1/2 for ε > 2 \varepsilon>2 ε > 2 and P G ≤ C N 2 / 3 P_G \le CN^{2/3} P G ≤ C N 2/3 for ε = 2 \varepsilon=2 ε = 2 .
Step 2. Replace f f f by the function g g g in (75) , which is constant for q i ≥ T + 1 q_i \ge T+1 q i ≥ T + 1 . The convex chart of H i H_i H i then gives P H ≤ C ε N 3 e T P_H \le C_{\varepsilon}N^3e^T P H ≤ C ε N 3 e T in (80) .
Step 3. Integrate along meridians across ε < q i ≤ 2 ε \varepsilon<q_i\le2\varepsilon ε < q i ≤ 2 ε to obtain P U ≤ C ε ( 1 + P H ) P_U \le C_{\varepsilon}(1+P_H) P U ≤ C ε ( 1 + P H ) in (83) .
Step 4. Apply Lemma VII.1 to the cells A i ⊂ U i A_i\subset U_i A i ⊂ U i and the overlap bound μ ( U i ∩ U j ) ≥ α \mu(U_i\cap U_j)\ge\alpha μ ( U i ∩ U j ) ≥ α . Equation (87) gives λ 1 ≥ c ε α N − 4 e − T \lambda_1\ge c_{\varepsilon}\alpha N^{-4}e^{-T} λ 1 ≥ c ε α N − 4 e − T .
Only Step 1 uses Proposition V.2 . The factor e T e^T e T enters through the density ratio bounded in (79) ; Proposition III.9 bounds α − 1 \alpha^{-1} α − 1 by a power of N N N .
From deep caps to the covering sets, ε ≥ 2 \varepsilon\ge2 ε ≥ 2
Constants in this section may depend on the fixed energy ε \varepsilon ε , but not on N N N , except for the local singularity estimates explicitly identified in § 0.7 A. Domains carry the induced metric and the area measure σ \sigma σ , and T T T is the level (8) .
Step 1: the lifted field on G i G_i G i
On M n , e ( t ) M_{n,e(t)} M n , e ( t ) , write y j = q j / c ( t ) y_j=q_j/c(t) y j = q j / c ( t ) for j ≠ i j\ne i j = i and set
φ t ( y ) = θ ( t ) y ( y − e ( t ) ) 2 θ ( t ) y + 8 e ( t ) , H t = 1 n ∑ j ≠ i φ t ( y j ) , J t = ( e ( t ) − 1 ) φ t ′ ( 0 ) + 1 n ∑ j ≠ i ( y j − e ( t ) ) φ t ′ ( y j ) , Z t = ∇ M n , e ( t ) H t J t . (67) \begin{aligned}
\varphi_t(y)&=\frac{\theta(t)y(y-e(t))^2}{\theta(t)y+8e(t)},\qquad H_t=\frac{1}{n}\sum_{j\ne i}\varphi_t(y_j),\\
J_t&=(e(t)-1)\varphi_t'(0)+\frac{1}{n}\sum_{j\ne i}(y_j-e(t))\varphi_t'(y_j),\qquad Z_t=\frac{\nabla_{M_{n,e(t)}}H_t}{J_t}. \tag*{(67)}
\end{aligned} φ t ( y ) J t = θ ( t ) y + 8 e ( t ) θ ( t ) y ( y − e ( t ) ) 2 , H t = n 1 j = i ∑ φ t ( y j ) , = ( e ( t ) − 1 ) φ t ′ ( 0 ) + n 1 j = i ∑ ( y j − e ( t )) φ t ′ ( y j ) , Z t = J t ∇ M n , e ( t ) H t . (67)
Thus Z t Z_t Z t is the vector field X X X in (58) with ( N , ε ) (N,\varepsilon) ( N , ε ) replaced by ( n , e ( t ) ) (n,e(t)) ( n , e ( t )) , so that 2 − ε 2-\varepsilon 2 − ε becomes θ ( t ) \theta(t) θ ( t ) . If D t : M n , e ( t ) → Σ i ( t ) D_t:M_{n,e(t)}\to\Sigma_i(t) D t : M n , e ( t ) → Σ i ( t ) is the homothety of ratio c ( t ) c(t) c ( t ) , define
Y = c − 2 ( D t ) ∗ Z t , X = ϖ Y , ϖ = ρ 2 c R 2 ∈ ( 0 , 1 ] . (68) Y=c^{-2}(D_t)_*Z_t,\qquad X=\varpi Y,\qquad\varpi=\frac{\rho^2}{cR^2}\in(0,1]. \tag*{(68)} Y = c − 2 ( D t ) ∗ Z t , X = ϖ Y , ϖ = c R 2 ρ 2 ∈ ( 0 , 1 ] . (68)
The bound ϖ ≤ 1 \varpi\le1 ϖ ≤ 1 follows from c R 2 − ρ 2 = N ( t − 1 ) ( t − ε ) / ( N − 1 ) ≥ 0 cR^2-\rho^2=N(t-1)(t-\varepsilon)/(N-1)\ge0 c R 2 − ρ 2 = N ( t − 1 ) ( t − ε ) / ( N − 1 ) ≥ 0 . By Remark III.3, Proposition V.4 applies on M n , e ( t ) M_{n,e(t)} M n , e ( t ) with ( N , ε ) (N,\varepsilon) ( N , ε ) replaced by ( n , e ( t ) ) (n,e(t)) ( n , e ( t )) , away from the finite set F e ( t ) F_{e(t)} F e ( t ) . Scaling (60) by D t D_t D t gives
n − 3 ρ 2 + div Σ Y − ∣ Y ∣ 2 ≥ n θ 2 100 ρ 2 . (69) \frac{n-3}{\rho^2}+\operatorname{div}_{\Sigma}Y-\lvert Y\rvert^2\ge\frac{n\theta^2}{100\rho^2}. \tag*{(69)} ρ 2 n − 3 + div Σ Y − ∣ Y ∣ 2 ≥ 100 ρ 2 n θ 2 . (69)
Since X X X is tangent to the slices and ϖ \varpi ϖ depends only on t t t , the volume element in Lemma III.2(2) gives div X = ϖ div Σ Y \operatorname{div}X=\varpi\operatorname{div}_{\Sigma}Y div X = ϖ div Σ Y . Combining (68) –(69) yields
N − 3 R 2 + div X − ∣ X ∣ 2 ≥ 1 R 2 ( n θ 2 100 c − D ( t ) ) ≥ n θ 2 200 R 2 ≥ κ N 400 ( ε − 1 ) , D ( t ) = N − 4 c − ( N − 3 ) = ( t − 2 ) N − 3 t + 4 N − t ≤ 2 t . (70) \begin{aligned}
\frac{N-3}{R^2}+\operatorname{div}X-\lvert X\rvert^2
&\ge\frac{1}{R^2}\left(\frac{n\theta^2}{100c}-D(t)\right)
\ge\frac{n\theta^2}{200R^2}
\ge\frac{\kappa_N}{400(\varepsilon-1)},\\
D(t)&=\frac{N-4}{c}-(N-3)=\frac{(t-2)N-3t+4}{N-t}\le2t. \tag*{(70)}
\end{aligned} R 2 N − 3 + div X − ∣ X ∣ 2 D ( t ) ≥ R 2 1 ( 100 c n θ 2 − D ( t ) ) ≥ 200 R 2 n θ 2 ≥ 400 ( ε − 1 ) κ N , = c N − 4 − ( N − 3 ) = N − t ( t − 2 ) N − 3 t + 4 ≤ 2 t . (70)
The second inequality in (70) uses c ≤ 1 c\le1 c ≤ 1 and n θ 2 ≥ 400 t n\theta^2\ge400t n θ 2 ≥ 400 t from (7) ; the third uses n ≥ N / 2 n\ge N/2 n ≥ N /2 and R 2 = N ( ε − 1 ) R^2=N(\varepsilon-1) R 2 = N ( ε − 1 ) . On each face { q j = 0 } \{q_j=0\} { q j = 0 } , j ≠ i j\ne i j = i , scaling (59) gives
⟨ X , ∇ q j ⟩ = ϖ ⟨ Y , ∇ Σ q j ⟩ = ϖ c n ( e − 1 ) = 1 R 2 . (71) \langle X,\nabla q_j\rangle=\varpi\langle Y,\nabla_{\Sigma}q_j\rangle=\frac{\varpi}{cn(e-1)}=\frac{1}{R^2}. \tag*{(71)} ⟨ X , ∇ q j ⟩ = ϖ ⟨ Y , ∇ Σ q j ⟩ = c n ( e − 1 ) ϖ = R 2 1 . (71)
On q i = T q_i=T q i = T , tangency gives ⟨ X , ∇ q i ⟩ = 0 ≥ ( 1 − T ) / R 2 \langle X,\nabla q_i\rangle=0\ge(1-T)/R^2 ⟨ X , ∇ q i ⟩ = 0 ≥ ( 1 − T ) / R 2 . Thus (71) and tangency verify (54) on every face of G i G_i G i .
To check the singularity hypothesis of Proposition V.2, let F F F consist of the pole P i P_i P i and the configurations at the finitely many levels t k ∈ [ T , t ∗ ) t_k\in[T,t^*) t k ∈ [ T , t ∗ ) with n / e ( t k ) = k ∈ Z n/e(t_k)=k\in\mathbb{Z} n / e ( t k ) = k ∈ Z , k < n k<n k < n , at which k k k of the coordinates y j y_j y j , j ≠ i j\ne i j = i , equal e ( t k ) e(t_k) e ( t k ) and the others vanish. Near such a configuration, let A ⊂ { 1 , … , N } ∖ { i } A\subset\{1,\ldots,N\}\setminus\{i\} A ⊂ { 1 , … , N } ∖ { i } be the set of its positive coordinates. The function H t H_t H t in (67) satisfies
H t ≥ c 0 ( ∑ j ∉ A j ≠ i y j + ∑ j ∈ A ( y j − e ( t ) ) 2 ) . (72) H_t\ge c_0\left(\sum_{\substack{j\notin A\\j\ne i}}y_j+\sum_{j\in A}(y_j-e(t))^2\right). \tag*{(72)} H t ≥ c 0 j ∈ / A j = i ∑ y j + j ∈ A ∑ ( y j − e ( t ) ) 2 . (72)
The mass constraint gives k ( e ( t ) − e ( t k ) ) = − ∑ j ∉ A , j ≠ i y j − ∑ j ∈ A ( y j − e ( t ) ) k(e(t)-e(t_k))=-\sum_{j\notin A,j\ne i}y_j-\sum_{j\in A}(y_j-e(t)) k ( e ( t ) − e ( t k )) = − ∑ j ∈ / A , j = i y j − ∑ j ∈ A ( y j − e ( t )) . Since e ′ ( t k ) ≠ 0 e'(t_k)\ne0 e ′ ( t k ) = 0 , this identity and (72) imply H t ≥ c 1 dist ( q , F ) 2 H_t\ge c_1\operatorname{dist}(q,F)^2 H t ≥ c 1 dist ( q , F ) 2 . Estimate (61) , with the slice parameters, gives ∣ Z t ∣ ≤ 2 ( n H t ) − 1 / 2 |Z_t|\le2(nH_t)^{-1/2} ∣ Z t ∣ ≤ 2 ( n H t ) − 1/2 ; hence (68) implies
∣ X ( q ) ∣ ≤ C / dist ( q , F ) . (73) |X(q)|\le C/\operatorname{dist}(q,F). \tag*{(73)} ∣ X ( q ) ∣ ≤ C / dist ( q , F ) . (73)
Near P i P_i P i , all y j → 1 y_j\to1 y j → 1 , so (67) gives H t ≥ c 2 e ( t ) ( e ( t ) − 1 ) H_t\ge c_2e(t)(e(t)-1) H t ≥ c 2 e ( t ) ( e ( t ) − 1 ) and ∣ X ∣ ≤ C ρ / ( c R 2 ) → 0 |X|\le C\rho/(cR^2)\to0 ∣ X ∣ ≤ C ρ / ( c R 2 ) → 0 . The constants in this paragraph may depend on N N N : Proposition V.2 requires the inverse-distance bound for each fixed N N N .
The cap G i G_i G i is a compact connected domain with smooth corners (Proposition V.3 (b)). Proposition V.2 , applied using (70) , (71) , and (73) , gives
λ 1 ( G i ) ≥ κ N 400 ( ε − 1 ) , P G ≤ { ( ε − 1 ) N / ε − 2 , ε > 2 , N 2 / 3 / 90 , ε = 2. (74) \lambda_1(G_i)\ge\frac{\kappa_N}{400(\varepsilon-1)},\qquad P_G\le
\begin{cases}
(\varepsilon-1)\sqrt{N}/\sqrt{\varepsilon-2}, & \varepsilon>2,\\
N^{2/3}/90, & \varepsilon=2.
\end{cases} \tag*{(74)} λ 1 ( G i ) ≥ 400 ( ε − 1 ) κ N , P G ≤ { ( ε − 1 ) N / ε − 2 , N 2/3 /90 , ε > 2 , ε = 2. (74)
Figure 6 shows the lifted field on an invariant two-sphere of M 108 2 , 5 M_{108^{2},5} M 10 8 2 , 5 , where the cap G 1 G_1 G 1 is nonempty, and marks its corner, its singular points and the pole.
Figure 6. The lifted field of Step 1 on an invariant two-sphere of M 108 2 , 5 M_{108^{2},5} M 10 8 2 , 5 , where the cap G 1 G_1 G 1 is nonempty. The corner of G 1 G_1 G 1 , the singular points of the field and the pole are marked.
Step 2: freezing and transport from G i G_i G i to H i H_i H i
For K > ε K>\varepsilon K > ε the coordinates q j q_j q j , j ≠ i j\ne i j = i , measured as fractions of their total mass, give a convex chart of the cap { q i > K } \{q_i>K\} { q i > K } . We state it for H i H_i H i .
Lemma VI.1. Fix i i i and write t : = q i t:=q_i t := q i . On H i H_i H i , t t t ranges over ( 2 ε , t ∗ ] (2\varepsilon,t^*] ( 2 ε , t ∗ ] , the value t ∗ t^* t ∗ being attained only at the pole P i P_i P i . Let T = { x ∈ [ 0 , ∞ ) N − 1 : ∑ j x j = 1 } \mathcal{T}=\{x\in[0,\infty)^{N-1}:\sum_jx_j=1\} T = { x ∈ [ 0 , ∞ ) N − 1 : ∑ j x j = 1 } and k ( t ) = ( N ε − t 2 ) / ( N − t ) 2 k(t)=(N\varepsilon-t^2)/(N-t)^2 k ( t ) = ( N ε − t 2 ) / ( N − t ) 2 ; k k k is strictly decreasing on ( ε , N ) (\varepsilon,N) ( ε , N ) . The weighted coordinates x j = q j / ( N − t ) x_j=q_j/(N-t) x j = q j / ( N − t ) , j ≠ i j\ne i j = i , map H i H_i H i diffeomorphically onto the convex set
Ω = { x ∈ T ∘ : ∣ x ∣ 2 < k ( 2 ε ) } , \Omega=\{x\in\mathcal{T}^{\circ}:|x|^2<k(2\varepsilon)\}, Ω = { x ∈ T ∘ : ∣ x ∣ 2 < k ( 2 ε )} , a simplex intersected with a ball. On Ω \Omega Ω , t t t is the function t ( x ) = k − 1 ( ∣ x ∣ 2 ) t(x)=k^{-1}(|x|^2) t ( x ) = k − 1 ( ∣ x ∣ 2 ) , and the inverse chart is F ( x ) = ( q i = t ( x ) , q j = ( N − t ( x ) ) x j ) F(x)=(q_i=t(x),\ q_j=(N-t(x))x_j) F ( x ) = ( q i = t ( x ) , q j = ( N − t ( x )) x j ) . Moreover, for N ≥ 7 ε N\ge7\varepsilon N ≥ 7 ε :
(a) d σ = w ( t ( x ) ) d x d\sigma=w(t(x))\,dx d σ = w ( t ( x )) d x , where d x dx d x is the ( N − 2 ) (N-2) ( N − 2 ) -dimensional Lebesgue measure on the hyperplane { ∑ j x j = 1 } \{\sum_jx_j=1\} { ∑ j x j = 1 } and
w ( t ) = ε − 1 N − 1 ( N − t ) N − 1 t − ε ; w(t)=\sqrt{\frac{\varepsilon^{-1}}{N-1}}\frac{(N-t)^{N-1}}{t-\varepsilon}; w ( t ) = N − 1 ε − 1 t − ε ( N − t ) N − 1 ; w w w is decreasing in t t t , and N N − 1 e − t − t 2 / N ≤ ( N − t ) N − 1 ≤ N N − 1 e − t + t / N N^{N-1}e^{-t-t^2/N}\le(N-t)^{N-1}\le N^{N-1}e^{-t+t/N} N N − 1 e − t − t 2 / N ≤ ( N − t ) N − 1 ≤ N N − 1 e − t + t / N for 0 ≤ t ≤ N / 2 0\le t\le N/2 0 ≤ t ≤ N /2 ;
(b) diam ( Ω ) 2 ≤ 8 ε / N \operatorname{diam}(\Omega)^2\le8\varepsilon/N diam ( Ω ) 2 ≤ 8 ε / N ;
(c) ∥ d F ∥ 2 ≤ ( 2 + 6 / ε ) N 3 \lVert dF\rVert^2\le(2+6/\varepsilon)N^3 ∥ d F ∥ 2 ≤ ( 2 + 6/ ε ) N 3 .
The main factor ( N − t ) N − 1 ≈ N N − 1 e − t (N-t)^{N-1}\approx N^{N-1}e^{-t} ( N − t ) N − 1 ≈ N N − 1 e − t of the density says that taking mass t t t away from the other N − 1 N-1 N − 1 coordinates shrinks their volume by ( 1 − t / N ) N − 1 ≈ e − t (1-t/N)^{N-1}\approx e^{-t} ( 1 − t / N ) N − 1 ≈ e − t ; it is the origin of the factor e T e^T e T below. The proof is in Appendix § 0.16 .
Let m m m be the mean of f ∈ H 1 ( H i ) f\in H^1(H_i) f ∈ H 1 ( H i ) on G i G_i G i . Choose 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 with χ = 0 \chi=0 χ = 0 on ( − ∞ , T ] (-\infty,T] ( − ∞ , T ] , χ = 1 \chi=1 χ = 1 on [ T + 1 , ∞ ) [T+1,\infty) [ T + 1 , ∞ ) and ∣ χ ′ ∣ ≤ 2 |\chi'|\le2 ∣ χ ′ ∣ ≤ 2 , and set
g = f − χ ( q i ) ( f − m ) . (75) g=f-\chi(q_i)(f-m). \tag*{(75)} g = f − χ ( q i ) ( f − m ) . (75)
Since ∣ ∇ q i ∣ ≤ 1 |\nabla q_i|\le1 ∣∇ q i ∣ ≤ 1 , the Poincaré bound (74) gives
∫ H i ∣ f − g ∣ 2 d σ ≤ P G ∫ H i ∣ ∇ f ∣ 2 d σ , ∫ H i ∣ ∇ g ∣ 2 d σ ≤ ( 2 + 8 P G ) ∫ H i ∣ ∇ f ∣ 2 d σ . (76) \begin{aligned}
\int_{H_i}|f-g|^2\,d\sigma&\le P_G\int_{H_i}|\nabla f|^2\,d\sigma,\\
\int_{H_i}|\nabla g|^2\,d\sigma&\le(2+8P_G)\int_{H_i}|\nabla f|^2\,d\sigma. \tag*{(76)}
\end{aligned} ∫ H i ∣ f − g ∣ 2 d σ ∫ H i ∣∇ g ∣ 2 d σ ≤ P G ∫ H i ∣∇ f ∣ 2 d σ , ≤ ( 2 + 8 P G ) ∫ H i ∣∇ f ∣ 2 d σ . (76)
In the convex chart F : Ω → H i F:\Omega\to H_i F : Ω → H i of Lemma VI.1, g ∘ F = m g \circ F=m g ∘ F = m for t ≥ T + 1 t\ge T+1 t ≥ T + 1 . For the density w ( t ) w(t) w ( t ) of Lemma VI.1(a), define
w ~ ( t ) = max { w ( t ) , w ( T + 1 ) } , sup Ω w ~ inf Ω w ~ ≤ Ξ : = w ( 2 ε ) w ( T + 1 ) = ( N − 2 ε N − T − 1 ) N − 1 T + 1 − ε ε . (77) \widetilde{w}(t)=\max\{w(t),w(T+1)\},\qquad\frac{\sup_{\Omega}\widetilde{w}}{\inf_{\Omega}\widetilde{w}}\leq\Xi:=\frac{w(2\varepsilon)}{w(T+1)}=\left(\frac{N-2\varepsilon}{N-T-1}\right)^{N-1}\frac{T+1-\varepsilon}{\varepsilon}. \tag*{(77)} w ( t ) = max { w ( t ) , w ( T + 1 )} , inf Ω w sup Ω w ≤ Ξ := w ( T + 1 ) w ( 2 ε ) = ( N − T − 1 N − 2 ε ) N − 1 ε T + 1 − ε . (77)
The Payne–Weinberger inequality [44 , 45 ] on Ω \Omega Ω and (77) give the weight w ~ \widetilde{w} w a Poincaré constant at most Ξ diam ( Ω ) 2 / π 2 \Xi\operatorname{diam}(\Omega)^2/\pi^2 Ξ diam ( Ω ) 2 / π 2 . The same constant bounds the variance of g ∘ F g\circ F g ∘ F with weight w w w : use w ≤ w ~ w\leq\widetilde{w} w ≤ w in the variance integral and ∇ ( g ∘ F ) = 0 \nabla(g\circ F)=0 ∇ ( g ∘ F ) = 0 wherever w ≠ w ~ w\neq\widetilde{w} w = w in the energy integral. Lemma VI.1(b)–(c) and (76) therefore give
min c ∈ R ∫ H i ∣ g − c ∣ 2 d σ ≤ C ε N 2 Ξ ∫ H i ∣ ∇ g ∣ 2 d σ , (78) \min_{c\in\mathbb{R}}\int_{H_i}|g-c|^2\,d\sigma\leq C_\varepsilon N^2\Xi\int_{H_i}|\nabla g|^2\,d\sigma, \tag*{(78)} c ∈ R min ∫ H i ∣ g − c ∣ 2 d σ ≤ C ε N 2 Ξ ∫ H i ∣∇ g ∣ 2 d σ , (78)
P H ≤ 2 P G + C ε N 2 Ξ ( 1 + P G ) . P_H\leq2P_G+C_\varepsilon N^2\Xi(1+P_G). P H ≤ 2 P G + C ε N 2 Ξ ( 1 + P G ) .
The logarithmic bounds in Lemma VI.1(a), applied to (77) , imply
Ξ ≤ T + 1 ε exp ( T + 1 − 2 ε + 2 ε + ( T + 1 ) 2 N ) ≤ { C ε N 1 / 2 e T , ε > 2 , C N 1 / 3 e T , ε = 2. (79) \Xi\leq\frac{T+1}{\varepsilon}\exp\left(T+1-2\varepsilon+\frac{2\varepsilon+(T+1)^2}{N}\right)\leq
\begin{cases}
C_\varepsilon N^{1/2}e^T, & \varepsilon>2,\\
CN^{1/3}e^T, & \varepsilon=2.
\end{cases} \tag*{(79)} Ξ ≤ ε T + 1 exp ( T + 1 − 2 ε + N 2 ε + ( T + 1 ) 2 ) ≤ { C ε N 1/2 e T , C N 1/3 e T , ε > 2 , ε = 2. (79)
Substituting (74) and (79) into (78) proves
P H ≤ C ε N 3 e T . (80) P_H\leq C_\varepsilon N^3e^T. \tag*{(80)} P H ≤ C ε N 3 e T . (80)
Step 3: meridian transport from H i H_i H i to U i U_i U i
Write d σ = υ ( t ) d t d ω d\sigma=\upsilon(t)\,dt\,d\omega d σ = υ ( t ) d t d ω on U i U_i U i , as in (16) . By Lemma III.2(3), the admissible meridian directions increase with t t t . Thus every ( t , ω ) ∈ L : = U i ∖ H i (t,\omega)\in L:=U_i\setminus H_i ( t , ω ) ∈ L := U i ∖ H i is joined within U i U_i U i to ( s , ω ) (s,\omega) ( s , ω ) for 2 ε < s < 2 ε + 1 2\varepsilon<s<2\varepsilon+1 2 ε < s < 2 ε + 1 , and
∣ f ( t , ω ) − f ( s , ω ) ∣ 2 ≤ 2 ( ε + 1 ) ∫ t s ∣ ∇ f ( r , ω ) ∣ 2 d r . (81) |f(t,\omega)-f(s,\omega)|^2\leq2(\varepsilon+1)\int_t^s|\nabla f(r,\omega)|^2\,dr. \tag*{(81)} ∣ f ( t , ω ) − f ( s , ω ) ∣ 2 ≤ 2 ( ε + 1 ) ∫ t s ∣∇ f ( r , ω ) ∣ 2 d r . (81)
Let m m m be the mean of f f f on H i H_i H i . Average ∣ f ( t , ω ) − m ∣ 2 ≤ 2 ∣ f ( s , ω ) − m ∣ 2 + 2 ∣ f ( t , ω ) − f ( s , ω ) ∣ 2 |f(t,\omega)-m|^2\leq2|f(s,\omega)-m|^2+2|f(t,\omega)-f(s,\omega)|^2 ∣ f ( t , ω ) − m ∣ 2 ≤ 2∣ f ( s , ω ) − m ∣ 2 + 2∣ f ( t , ω ) − f ( s , ω ) ∣ 2 over s ∈ ( 2 ε , 2 ε + 1 ) s\in(2\varepsilon,2\varepsilon+1) s ∈ ( 2 ε , 2 ε + 1 ) and integrate over L L L . Equations (81) and (16) give
∫ L ∣ f − m ∣ 2 d σ ≤ C ε ∫ H i ∣ f − m ∣ 2 d σ + C ε ∫ U i ∣ ∇ f ∣ 2 d σ . (82) \int_L|f-m|^2\,d\sigma\leq C_\varepsilon\int_{H_i}|f-m|^2\,d\sigma+C_\varepsilon\int_{U_i}|\nabla f|^2\,d\sigma. \tag*{(82)} ∫ L ∣ f − m ∣ 2 d σ ≤ C ε ∫ H i ∣ f − m ∣ 2 d σ + C ε ∫ U i ∣∇ f ∣ 2 d σ . (82)
Adding ∫ H i ∣ f − m ∣ 2 d σ \int_{H_i}|f-m|^2\,d\sigma ∫ H i ∣ f − m ∣ 2 d σ to (82) and applying (80) on H i H_i H i proves
P U ≤ C ε ( 1 + P H ) ≤ C ε N 3 e T . (83) P_U\leq C_\varepsilon(1+P_H)\leq C_\varepsilon N^3e^T. \tag*{(83)} P U ≤ C ε ( 1 + P H ) ≤ C ε N 3 e T . (83)
For f ∈ H 1 ( U i ) f\in H^1(U_i) f ∈ H 1 ( U i ) , equations (81) –(82) follow by local Sobolev approximation and absolute continuity along almost every meridian.
Completion of the lower bounds
Exceptional energies in Theorem V.1 (i). Suppose ε 0 = N / k ∈ ( 1 , 2 ) \varepsilon_0=N/k\in(1,2) ε 0 = N / k ∈ ( 1 , 2 ) and N ( 2 − ε 0 ) 2 ≥ 2000 N(2-\varepsilon_0)^2\ge2000 N ( 2 − ε 0 ) 2 ≥ 2000 . On the unit sphere S S S in { ∑ j y j = 0 } \{\sum_j y_j=0\} { ∑ j y j = 0 } put K τ = { u ∈ S : u j ≥ − τ ∀ j } K_\tau=\{u\in S:u_j\ge-\tau\ \forall j\} K τ = { u ∈ S : u j ≥ − τ ∀ j } . The map u ↦ 1 + R ( ε ) u u\mapsto1+R(\varepsilon)u u ↦ 1 + R ( ε ) u identifies M N , ε M_{N,\varepsilon} M N , ε with ( K 1 / R ( ε ) , R ( ε ) 2 g S ) (K_{1/R(\varepsilon)},R(\varepsilon)^2g_S) ( K 1/ R ( ε ) , R ( ε ) 2 g S ) . As nonexceptional ε ↑ ε 0 \varepsilon\uparrow\varepsilon_0 ε ↑ ε 0 , the sets K 1 / R ( ε ) K_{1/R(\varepsilon)} K 1/ R ( ε ) decrease to K 1 / R ( ε 0 ) K_{1/R(\varepsilon_0)} K 1/ R ( ε 0 ) , and their difference has vanishing measure. For restrictions of smooth functions on S S S , the Poincaré inequality obtained from (59) and (60) therefore passes to the limit. Its spectral-gap lower bound in the metric g S g_S g S is
R ( ε ) 2 ( 2 − ε ) 2 100 ( ε − 1 ) = N ( 2 − ε ) 2 100 . (84) R(\varepsilon)^2\frac{(2-\varepsilon)^2}{100(\varepsilon-1)}=\frac{N(2-\varepsilon)^2}{100}. \tag*{(84)} R ( ε ) 2 100 ( ε − 1 ) ( 2 − ε ) 2 = 100 N ( 2 − ε ) 2 . (84)
Restrictions of smooth functions on S S S are dense in H 1 ( K 1 / R ( ε 0 ) H^1(K_{1/R(\varepsilon_0)} H 1 ( K 1/ R ( ε 0 ) ): truncate the function, cut it off near the finite set corresponding to F ε 0 F_{\varepsilon_0} F ε 0 at squared-gradient cost O ( δ d − 2 ) O(\delta^{d-2}) O ( δ d − 2 ) , where d = N − 2 > 2 d=N-2>2 d = N − 2 > 2 , and use corner-chart density away from that set. The limit of (84) , rescaled by R ( ε 0 ) − 2 R(\varepsilon_0)^{-2} R ( ε 0 ) − 2 , proves Theorem V.1 (i) at ε 0 \varepsilon_0 ε 0 . This limiting argument treats the points of F ε 0 F_{\varepsilon_0} F ε 0 , at which M M M is not a domain with corners (Proposition V.3 (c)).
Step 4: Theorem V.1 (ii)–(iii). The local Poincaré inequalities on the sets U i U_i U i are glued by the following lemma.
Lemma VII.1. Let A i ⊂ U i A_i\subset U_i A i ⊂ U i , 1 ≤ i ≤ N 1\le i\le N 1 ≤ i ≤ N , where the A i A_i A i partition M M M up to null sets and μ ( A i ) = 1 / N \mu(A_i)=1/N μ ( A i ) = 1/ N . If P ( U i ) ≤ P P(U_i)\le P P ( U i ) ≤ P and μ ( U i ∩ U j ) ≥ α > 0 \mu(U_i\cap U_j)\ge\alpha>0 μ ( U i ∩ U j ) ≥ α > 0 for i ≠ j i\ne j i = j , then
Var μ f ≤ N ( 1 + 2 α ) 2 P ∫ M ∣ ∇ f ∣ 2 d μ . (85) \operatorname{Var}_{\mu} f\le N\left(1+\sqrt{\frac{2}{\alpha}}\right)^2P\int_M|\nabla f|^2\,d\mu. \tag*{(85)} Var μ f ≤ N ( 1 + α 2 ) 2 P ∫ M ∣∇ f ∣ 2 d μ . (85)
Proof. Let m i m_i m i be the mean of f f f on U i U_i U i , v = m i v=m_i v = m i on A i A_i A i , and E i = ∫ U i ∣ ∇ f ∣ 2 d μ E_i=\int_{U_i}|\nabla f|^2d\mu E i = ∫ U i ∣∇ f ∣ 2 d μ . The Poincaré inequalities on U i U_i U i imply
∥ f − v ∥ 2 2 ≤ P ∑ i E i ≤ N P E ( f , f ) , α ∣ m i − m j ∣ 2 ≤ 2 P ( E i + E j ) . (86) \|f-v\|_2^2\le P\sum_iE_i\le NP\mathcal{E}(f,f),\qquad\alpha|m_i-m_j|^2\le2P(E_i+E_j). \tag*{(86)} ∥ f − v ∥ 2 2 ≤ P i ∑ E i ≤ N P E ( f , f ) , α ∣ m i − m j ∣ 2 ≤ 2 P ( E i + E j ) . (86)
Summing the second inequality in (86) gives Var μ v = N − 2 ∑ i < j ( m i − m j ) 2 ≤ 2 P E ( f , f ) / α \operatorname{Var}_{\mu}v=N^{-2}\sum_{i<j}(m_i-m_j)^2\le2P\mathcal{E}(f,f)/\alpha Var μ v = N − 2 ∑ i < j ( m i − m j ) 2 ≤ 2 P E ( f , f ) / α . The triangle inequality for the centred L 2 L^2 L 2 norm, together with the first inequality in (86) , gives ( N + 2 / α ) 2 P (\sqrt{N}+\sqrt{2/\alpha})^2P ( N + 2/ α ) 2 P as a Poincaré constant, which is no larger than the constant in (85) .
□ \square □
Lemma III.4 supplies the cover U i U_i U i and the cells A i A_i A i required by Lemma VII.1. Proposition III.9 gives α ≥ α 2 > 0 \alpha\ge\alpha_2>0 α ≥ α 2 > 0 at ε = 2 \varepsilon=2 ε = 2 and α ≥ N − C ε \alpha\ge N^{-C\varepsilon} α ≥ N − C ε at fixed ε > 2 \varepsilon>2 ε > 2 . Since α ≤ 1 \alpha\le1 α ≤ 1 and ( 1 + 2 / α ) 2 ≤ 8 / α (1+\sqrt{2/\alpha})^2\leq8/\alpha ( 1 + 2/ α ) 2 ≤ 8/ α , substituting (83) into (85) yields
Var μ f ≤ C ε N 4 α − 1 e T ∫ M ∣ ∇ f ∣ 2 d μ , λ 1 ( N , ε ) ≥ c ε α N − 4 e − T . (87) \begin{aligned}
\operatorname{Var}_{\mu} f &\leq C_{\varepsilon}N^4\alpha^{-1}e^T\int_M|\nabla f|^2d\mu,\\
\lambda_1(N,\varepsilon)&\geq c_{\varepsilon}\alpha N^{-4}e^{-T}. \tag*{(87)}
\end{aligned} Var μ f λ 1 ( N , ε ) ≤ C ε N 4 α − 1 e T ∫ M ∣∇ f ∣ 2 d μ , ≥ c ε α N − 4 e − T . (87)
Finally, (8) gives e − T ≥ e − 1 e − T 0 e^{-T}\geq e^{-1}e^{-T_0} e − T ≥ e − 1 e − T 0 . Substitution of the two values of T 0 T_0 T 0 and the corresponding overlap bounds into (87) proves Theorem V.1(ii)–(iii).
Concluding summary
We have shown that substituting the Hamiltonian dynamics with a Brownian dynamics on the microcanonical surface, we can put rigorous bounds on the equilibration time. The fluid phase, for ε < 2 \varepsilon<2 ε < 2 equilibrates in O ( 1 ) O(1) O ( 1 ) time. The condensate phase for ε > 2 \varepsilon>2 ε > 2 equilibrates in time which diverges with N N N ,
τ e q ≍ e b ∗ ( ε ) N + o ( N ) (88) \tau_{\mathrm{eq}}\asymp e^{b_*(\varepsilon)\sqrt{N}+o(\sqrt{N})} \tag*{(88)} τ eq ≍ e b ∗ ( ε ) N + o ( N ) (88)
and we have the exact bounds
( 2 − 1 ) ε − 2 ≤ b ∗ ( ε ) ≤ ε − 2 . (89) (\sqrt{2}-1)\sqrt{\varepsilon-2}\leq b_*(\varepsilon)\leq\sqrt{\varepsilon-2}. \tag*{(89)} ( 2 − 1 ) ε − 2 ≤ b ∗ ( ε ) ≤ ε − 2 . (89)
To obtain the lower bound on λ 1 \lambda_1 λ 1 , which is an upper bound to τ e q \tau_{\mathrm{eq}} τ eq , we have used a geometric technique which we dubbed Bochner method because it rests on Bochner’s formula: integrated for the gradient of an eigenfunction, it turns a lower bound on the Ricci curvature into a lower bound on λ 1 \lambda_1 λ 1 , and a vector field that crosses the concave faces of M M M inwards turns the wrong-sign boundary term into an interior cost that the curvature of the sphere can pay (Proposition V.2 ). For ε < 2 \varepsilon<2 ε < 2 one such field on the whole surface gives a bound uniform in N N N ; for ε ≥ 2 \varepsilon\geq2 ε ≥ 2 it works only on the deep caps around a condensate, and transporting and gluing these local inequalities across M M M gives λ 1 ≥ e − N ( ε − 2 ) − O ( N 1 / 4 ) \lambda_1\geq e^{-\sqrt{N}(\varepsilon-2)-O(N^{1/4})} λ 1 ≥ e − N ( ε − 2 ) − O ( N 1/4 ) (Theorem V.1 ). The proof of the lower bound to λ 1 \lambda_1 λ 1 is involved (although conceptually simple) but it is necessary so. In fact, one must prove that for any square integrable, zero average, function f f f one cannot get a ratio ∫ d μ ∣ ∇ M f ∣ 2 / ∫ d μ ∣ f ∣ 2 \int d\mu|\nabla_M f|^2/\int d\mu|f|^2 ∫ d μ ∣ ∇ M f ∣ 2 / ∫ d μ ∣ f ∣ 2 lower than something prescribed (the putative bound), which is exponentially small in N \sqrt{N} N . Physically, this means that the only long time dynamics occurs because of the entropic effects at the neck. In other words, that the chambers equilibrate quickly. This is the most difficult part of the proof.
To obtain the upper bound to λ 1 \lambda_1 λ 1 , which is a lower bound on τ e q \tau_{\mathrm{eq}} τ eq , instead it is sufficient to present an appropriate variational trial functions on the microcanonical surface. We found that for ε ≤ 2 \varepsilon\le2 ε ≤ 2 it is sufficient (as supported by the numerics) to have trial functions of a single variable q 1 q_1 q 1 while for ε > 2 \varepsilon> 2 ε > 2 one must switch to two functions ( q 1 , q 2 ) (q_1, q_2) ( q 1 , q 2 ) in order to capture the event of a localized mass to shift from one lattice site to another. When we cannot solve the 2 2 2 -d d d partial differential equation, we can present a simple linear trial function of ( q 1 , q 2 ) (q_1, q_2) ( q 1 , q 2 ) which does the job.
Our upper bound also admits a natural formulation in terms of Witten’s supersymmetric quantum mechanics [19 ] : the equilibrium measure is the unique zero-energy state, and in the condensed phase the N − 1 N - 1 N − 1 exponentially small eigenvalues are the ground states of the N N N chambers, lifted from zero by tunnelling through the necks. This point of view will be developed in a separate work.
The choice of the Brownian motion also deserves a comment. Any reversible dynamics with the microcanonical measure as equilibrium has the same geometric bottleneck, so the exponent b ( ε ) N b(\varepsilon)\sqrt{N} b ( ε ) N in the condensed phase is a property of the measure and not of the dynamics. The prefactors and the O ( 1 ) O(1) O ( 1 ) corrections, instead, depend on the metric used to define the Brownian motion, and for ε ≤ 2 \varepsilon\le2 ε ≤ 2 the dependence affects the gap itself. For instance, the Brownian motion in the variables ψ i \psi_i ψ i , on the surface { ∑ i ∣ ψ i ∣ 2 = N , ∑ i ∣ ψ i ∣ 4 = N ε } ⊂ C N \{\sum_i |\psi_i|^2 = N,\ \sum_i |\psi_i|^4 = N\varepsilon\} \subset\mathbb{C}^N { ∑ i ∣ ψ i ∣ 2 = N , ∑ i ∣ ψ i ∣ 4 = N ε } ⊂ C N , has the same equilibrium measure but a diffusion coefficient proportional to q i q_i q i in the occupations, closer to the Hamiltonian mass transfer between sites. At ε = 2 \varepsilon= 2 ε = 2 this surface is, to leading order, an Einstein manifold with Ric ≃ 2 \operatorname{Ric} \simeq2 Ric ≃ 2 , and its gap is close to 2, whereas for the Brownian motion in the variables q i q_i q i studied here λ 1 ≲ 1 / 4 \lambda_1 \lesssim1/4 λ 1 ≲ 1/4 (with the equality most likely being correct, as shown by the numerics).
The N \sqrt{N} N dependence of the (log of) the equilibration time is an entropy effect as measured by the reduction of the volume of the region (the tunnel ) that the phase space point has to pass to go from one condensate configuration to another one. Such entropic barriers are common to other statistical mechanics problems [46 , 47 ] , and in recent years have been argued to hinder the performance of both classical and quantum algorithms [48 ] .
Going back to look at the Hamiltonian dynamics, we must notice a curious “coincidence” with the work [24 ] . In [24 ] the Hamiltonian dynamics of the DNLSE coupled to a bath, at positive but high temperature has been studied and a breather has been found to melt on timescales τ b ∼ e α ε t b ( 0 ) \tau_b \sim e^{\alpha\varepsilon t b(0)} τ b ∼ e α εt b ( 0 ) where b ( 0 ) = ∣ ψ i ( 0 ) ∣ 2 b(0) = |\psi_i(0)|^2 b ( 0 ) = ∣ ψ i ( 0 ) ∣ 2 is the mass of the breather at position i i i and α m e l t = 0.91 \alpha_{\mathrm{melt}} = 0.91 α melt = 0.91 . If we bootstrap this numerical law to our configurations at negative temperatures, where ε > 2 \varepsilon> 2 ε > 2 , b ( 0 ) = q i ( 0 ) = ε − 2 N b(0) = q_i(0) = \sqrt{\varepsilon- 2}\sqrt{N} b ( 0 ) = q i ( 0 ) = ε − 2 N and ln τ e q = α ε − 2 N \ln\tau_{\mathrm{eq}} = \alpha\sqrt{\varepsilon- 2}\sqrt{N} ln τ eq = α ε − 2 N with α ∈ [ 2 − 1 , 1 ] \alpha\in[\sqrt{2}-1,1] α ∈ [ 2 − 1 , 1 ] in this paper so we have the same N \sqrt{N} N dependence and a compatible prefactor.
The comparison is meaningful because the background of [24 ] has mean mass density close to one, so that b ( 0 ) b(0) b ( 0 ) and q i q_i q i are measured in the same units, and their exponent α m e l t \alpha_{\mathrm{melt}} α melt is in the interval we found. The mechanisms are prima facie different: in [24 ] the slowness comes from an adiabatic invariant associated with the fast rotation of the breather, which our Brownian approximation discards, while our barrier is purely entropic and has no counterpart at positive temperature. However, the results suggest they must be connected somehow.
The slowness of the Hamiltonian dynamics above ε c = 2 \varepsilon_{\mathrm{c}} = 2 ε c = 2 is also visible in the numerical study [18 ] , where the infinite-temperature line ε = 2 \varepsilon= 2 ε = 2 of the Gross–Pitaevskii lattice was found to mark the onset of long-lived localized excitations, whose lifetimes have power-law distributions and make the dynamics weakly nonergodic at slightly larger energy densities.
To our knowledge, the one presented in this paper is the first rigorous result in which the relaxation time of a dynamics on the DNLS microcanonical surface is bounded, from above and from below, by quantities that diverge with the system size: in the condensed phase both bounds are stretched exponentials, e Θ ( N ) e^{\Theta(\sqrt{N})} e Θ ( N ) . This should be contrasted with the rigorous Hamiltonian results based on adiabatic invariants [49 , 50 ] , which are uniform in N N N and in which the small parameter is the specific energy, and with the time scales of condensate motion in stochastic particle systems [28 ] , which grow polynomially in the system size.
We believe the techniques exposed in this paper can be used to find rigorous bounds of this kind, for many stochastic problems in the future.
Acknowledgements
Claude Opus was used in this paper heavily to run the numerics, to set up and bring to closure the cumbersome calculations needed in the proof of the lower bound, and for checking overall consistency of the methods and results. We like to believe, however, that the ideas which have not appeared in the literature before are our own.
The tilted bound (Lemma III.7)
Proof of Lemma III.7. Write s = ζ N ≥ 1 2 s=\zeta\sqrt{N}\ge\frac{1}{2} s = ζ N ≥ 2 1 , S = N − ζ N S=N-\zeta\sqrt{N} S = N − ζ N and E = N ( ε − r 2 ) E=N(\varepsilon-r^2) E = N ( ε − r 2 ) . For real β \beta β let p s , β ( q ) = e − q − β q 2 1 [ 0 , s ] ( q ) / Z s ( 1 , β ) p_{s,\beta}(q)=e^{-q-\beta q^2}\mathbf{1}_{[0,s]}(q)/\mathcal{Z}_s(1,\beta) p s , β ( q ) = e − q − β q 2 1 [ 0 , s ] ( q ) / Z s ( 1 , β ) , with Z s ( 1 , β ) = ∫ 0 s e − q − β q 2 d q \mathcal{Z}_s(1,\beta)=\int_0^s e^{-q-\beta q^2}\,dq Z s ( 1 , β ) = ∫ 0 s e − q − β q 2 d q . Since ∏ i e − q i − β q i 2 = e − S − β E \prod_i e^{-q_i-\beta q_i^2}=e^{-S-\beta E} ∏ i e − q i − β q i 2 = e − S − β E on the support of the δ \delta δ -functions,
Ω n < s ( S , E ) = e S + β E Z s ( 1 , β ) n f n s , β ( S , E ) (90) \Omega_n^{<s}(S,E)=e^{S+\beta E}\mathcal{Z}_s(1,\beta)^n f_n^{s,\beta}(S,E) \tag*{(90)} Ω n < s ( S , E ) = e S + β E Z s ( 1 , β ) n f n s , β ( S , E ) (90)
exactly, where f n s , β f_n^{s,\beta} f n s , β is the density of ( ∑ i q i , ∑ i q i 2 ) (\sum_i q_i,\sum_i q_i^2) ( ∑ i q i , ∑ i q i 2 ) for n n n i.i.d. variables with density p s , β p_{s,\beta} p s , β . By Lemma III.5(1),
sup f n s , β ≤ π 3 ( sup q p s , β ( q ) ) 3 . (91) \sup f_n^{s,\beta}\le\frac{\pi}{\sqrt{3}}\left(\sup_q p_{s,\beta}(q)\right)^3. \tag*{(91)} sup f n s , β ≤ 3 π ( q sup p s , β ( q ) ) 3 . (91)
Case r ≥ a r\ge a r ≥ a . Take β = 0 \beta=0 β = 0 . Then Z s ( 1 , 0 ) ≤ 1 \mathcal{Z}_s(1,0)\le1 Z s ( 1 , 0 ) ≤ 1 , and sup p s , 0 ≤ 1 / ( 1 − e − 1 / 2 ) \sup p_{s,0}\le1/(1-e^{-1/2}) sup p s , 0 ≤ 1/ ( 1 − e − 1/2 ) since s ≥ 1 2 s\ge\frac{1}{2} s ≥ 2 1 ; so ln Ω n < s ≤ S + O ( 1 ) = N − ζ N + O ( 1 ) \ln\Omega_n^{<s}\le S+O(1)=N-\zeta\sqrt{N}+O(1) ln Ω n < s ≤ S + O ( 1 ) = N − ζ N + O ( 1 ) .
Case r < a r<a r < a . Take β = − κ / N < 0 \beta=-\kappa/\sqrt{N}<0 β = − κ / N < 0 with
κ = 1 − δ c , c = max ( ζ , a 2 ) , δ = 6 ln N a N . \kappa=\frac{1-\delta}{c},\qquad c=\max\left(\zeta,\frac{a}{2}\right),\qquad\delta=\frac{6\ln N}{a\sqrt{N}}. κ = c 1 − δ , c = max ( ζ , 2 a ) , δ = a N 6 ln N .
The exponent L ( q ) = − q + κ q 2 / N \mathcal{L}(q)=-q+\kappa q^2/\sqrt{N} L ( q ) = − q + κ q 2 / N is convex, and at the cutoff L ( s ) = − s ( 1 − κ ζ ) ≤ 0 \mathcal{L}(s)=-s(1-\kappa\zeta)\le0 L ( s ) = − s ( 1 − κ ζ ) ≤ 0 because κ ζ ≤ 1 − δ \kappa\zeta\le1-\delta κ ζ ≤ 1 − δ . So e L ≤ 1 e^{\mathcal{L}}\le1 e L ≤ 1 on [ 0 , s ] [0,s] [ 0 , s ] , Z s ( 1 , β ) ≥ 1 − e − 1 / 2 \mathcal{Z}_s(1,\beta)\ge1-e^{-1/2} Z s ( 1 , β ) ≥ 1 − e − 1/2 , and (A2) is bounded. Split Z s ( 1 , β ) = ∫ 0 s e L ( q ) d q \mathcal{Z}_s(1,\beta)=\int_0^s e^{\mathcal{L}(q)}\,dq Z s ( 1 , β ) = ∫ 0 s e L ( q ) d q at L = N 1 / 8 L=N^{1/8} L = N 1/8 :
on [ 0 , L ] [0,L] [ 0 , L ] , e κ q 2 / N = 1 + κ q 2 / N + O ( q 4 / N ) e^{\kappa q^2/\sqrt{N}}=1+\kappa q^2/\sqrt{N}+O(q^4/N) e κ q 2 / N = 1 + κ q 2 / N + O ( q 4 / N ) gives ∫ 0 L e L ( q ) d q ≤ 1 + 2 κ / N + O ( N − 1 ) \int_0^L e^{\mathcal{L}(q)}\,dq\le1+2\kappa/\sqrt{N}+O(N^{-1}) ∫ 0 L e L ( q ) d q ≤ 1 + 2 κ / N + O ( N − 1 ) ;
on [ L , s ] [L,s] [ L , s ] , convexity bounds L \mathcal{L} L by its endpoint values: L ( L ) ≤ − L / 2 \mathcal{L}(L)\le-L/2 L ( L ) ≤ − L /2 , while L ( s ) ≤ − s / 2 \mathcal{L}(s)\le-s/2 L ( s ) ≤ − s /2 if ζ ≤ a / 4 \zeta\le a/4 ζ ≤ a /4 (where κ ζ ≤ 1 2 \kappa\zeta\le\frac{1}{2} κ ζ ≤ 2 1 ) and L ( s ) ≤ − δ s ≤ − 3 2 ln N \mathcal{L}(s)\le-\delta s\le-\frac{3}{2}\ln N L ( s ) ≤ − δ s ≤ − 2 3 ln N if ζ > a / 4 \zeta>a/4 ζ > a /4 ; so this part is O ( N − 1 ) O(N^{-1}) O ( N − 1 ) .
Hence n ln Z s ( 1 , β ) ≤ 2 κ N + O ( 1 ) n\ln\mathcal{Z}_s(1,\beta)\le2\kappa\sqrt{N}+O(1) n ln Z s ( 1 , β ) ≤ 2 κ N + O ( 1 ) . Inserting S S S and β E = − κ N ( ε − r 2 ) \beta E=-\kappa\sqrt{N}(\varepsilon-r^2) β E = − κ N ( ε − r 2 ) in (A1), and using ε − 2 = a 2 \varepsilon-2=a^2 ε − 2 = a 2 ,
ln Ω n < s ≤ N − N [ ζ + κ ( a 2 − r 2 ) ] + O ( 1 ) = N − N [ ζ + a 2 − r 2 c ] + δ N ( a 2 − r 2 ) c + O ( 1 ) , \ln\Omega_n^{<s}\le N-\sqrt{N}\left[\zeta+\kappa(a^2-r^2)\right]+O(1)=N-\sqrt{N}\left[\zeta+\frac{a^2-r^2}{c}\right]+\frac{\delta\sqrt{N}(a^2-r^2)}{c}+O(1), ln Ω n < s ≤ N − N [ ζ + κ ( a 2 − r 2 ) ] + O ( 1 ) = N − N [ ζ + c a 2 − r 2 ] + c δ N ( a 2 − r 2 ) + O ( 1 ) ,
and the middle term is at most 2 a δ N = 12 ln N 2a\delta\sqrt{N}=12\ln N 2 a δ N = 12 ln N .
Proofs of Propositions III.8 and III.9
Proof of Proposition III.8. Lower bound. Integrating out one coordinate in the definition of Ω N \Omega_N Ω N gives Ω N ( N , N ε ) = ∫ 0 ∞ Ω N − 1 ( N − x , N ε − x 2 ) d x \Omega_N(N,N\varepsilon)=\int_0^\infty\Omega_{N-1}(N-x,N\varepsilon-x^2)\,dx Ω N ( N , N ε ) = ∫ 0 ∞ Ω N − 1 ( N − x , N ε − x 2 ) d x ; restrict x x x to [ a N , a N + 1 ] [a\sqrt{N},a\sqrt{N}+1] [ a N , a N + 1 ] . The other n = N − 1 n = N - 1 n = N − 1 coordinates have mass S ′ = N − x S' = N - x S ′ = N − x and energy E ′ = N ε − x 2 E' = N\varepsilon- x^2 E ′ = N ε − x 2 , so that S ′ − n = − a N + O ( 1 ) S' - n = -a\sqrt{N} + O(1) S ′ − n = − a N + O ( 1 ) and E ′ − 2 n = O ( N ) E' - 2n = O(\sqrt{N}) E ′ − 2 n = O ( N ) . Lemma III.6 gives f n ( S ′ , E ′ ) ≥ c / N f_n(S', E') \ge c/N f n ( S ′ , E ′ ) ≥ c / N , hence Ω n ( S ′ , E ′ ) = e S ′ f n ( S ′ , E ′ ) ≥ c e N − a N / N \Omega_n(S', E') = e^{S'}f_n(S', E') \ge ce^{N-a\sqrt{N}}/N Ω n ( S ′ , E ′ ) = e S ′ f n ( S ′ , E ′ ) ≥ c e N − a N / N . Integrating over x x x ,
Ω N ( N , N ε ) ≥ c e N − a N / N . \Omega_N(N,N\varepsilon) \ge ce^{N-a\sqrt{N}}/N. Ω N ( N , N ε ) ≥ c e N − a N / N .
Upper bound. The largest coordinate exceeds ε \varepsilon ε (Lemma III.4) and is almost surely attained at a single site. Conditioning on its position and value u u u ,
Ω N ( N , N ε ) = N ∫ ε t ∗ Ω N − 1 < u ( N − u , N ε − u 2 ) d u . \Omega_N(N,N\varepsilon) = N\int_{\varepsilon}^{t^*}\Omega_{N-1}^{<u}(N-u,N\varepsilon-u^2)\,du. Ω N ( N , N ε ) = N ∫ ε t ∗ Ω N − 1 < u ( N − u , N ε − u 2 ) d u .
Apply Lemma III.7 with k = 1 k = 1 k = 1 and ς = r = ζ = u / N \varsigma= r = \zeta= u/\sqrt{N} ς = r = ζ = u / N , which is admissible since ε ≤ u ≤ t ∗ ≤ N ε \varepsilon\le u \le t^* \le\sqrt{N\varepsilon} ε ≤ u ≤ t ∗ ≤ N ε for N N N large. Its proof (Appendix A) gives in fact the sharper bound
ln Ω N − 1 < u ( N − u , N ε − u 2 ) ≤ N − N [ ζ + ( 1 − δ ) a 2 − ζ 2 max ( ζ , a / 2 ) ] + O ( 1 ) , δ = 6 ln N a N . \ln\Omega_{N-1}^{<u}(N-u,N\varepsilon-u^2) \le N-\sqrt{N}\left[\zeta+(1-\delta)\frac{a^2-\zeta^2}{\max(\zeta,a/2)}\right]+O(1),\qquad\delta=\frac{6\ln N}{a\sqrt{N}}. ln Ω N − 1 < u ( N − u , N ε − u 2 ) ≤ N − N [ ζ + ( 1 − δ ) max ( ζ , a /2 ) a 2 − ζ 2 ] + O ( 1 ) , δ = a N 6 ln N .
This exponent lies below N − a N N-a\sqrt{N} N − a N by at least:
a N / 4 a\sqrt{N}/4 a N /4 for ζ < a / 2 \zeta<a/2 ζ < a /2 , because there the bracket is concave in ζ \zeta ζ and at least min ( 2 ( 1 − δ ) a , a / 2 + 3 2 ( 1 − δ ) a ) ≥ 5 4 a \min\left(2(1-\delta)a,a/2+\frac{3}{2}(1-\delta)a\right)\ge\frac{5}{4}a min ( 2 ( 1 − δ ) a , a /2 + 2 3 ( 1 − δ ) a ) ≥ 4 5 a for N N N large.
Integrating over u u u gives the stronger upper bound
Ω N ( N , N ε ) ≤ C ε ′ N e N − a N . \Omega_N(N,N\varepsilon)\le C'_{\varepsilon}N e^{N-a\sqrt{N}}. Ω N ( N , N ε ) ≤ C ε ′ N e N − a N .
The two bounds give (15) , and the bounds on f N ( N , N ε ) = e − N Ω N ( N , N ε ) f_N(N,N\varepsilon)=e^{-N}\Omega_N(N,N\varepsilon) f N ( N , N ε ) = e − N Ω N ( N , N ε ) follow.
Proof of Proposition III.9. By Lemma III.5(3),
Ω N ( N , N ε ) μ ( U i ∩ U j ) = ∫ ε ∞ ∫ ε ∞ Ω N − 2 ( N − u − v , N ε − u 2 − v 2 ) d u d v . \Omega_N(N,N\varepsilon)\,\mu(U_i\cap U_j)=\int_{\varepsilon}^{\infty}\int_{\varepsilon}^{\infty}\Omega_{N-2}(N-u-v,N\varepsilon-u^2-v^2)\,du\,dv. Ω N ( N , N ε ) μ ( U i ∩ U j ) = ∫ ε ∞ ∫ ε ∞ Ω N − 2 ( N − u − v , N ε − u 2 − v 2 ) d u d v .
(a) Let ε = 2 \varepsilon=2 ε = 2 and restrict u , v u,v u , v to [ 3 , 4 ] ⊂ ( ε , ∞ ) [3,4]\subset(\varepsilon,\infty) [ 3 , 4 ] ⊂ ( ε , ∞ ) . Since Ω n = e S ′ f n \Omega_n=e^{S'}f_n Ω n = e S ′ f n ,
μ ( U i ∩ U j ) ≥ 1 f N ( N , 2 N ) ∫ 3 4 ∫ 3 4 e − u − v f N − 2 ( N − u − v , 2 N − u 2 − v 2 ) d u d v . \mu(U_i\cap U_j)\ge\frac{1}{f_N(N,2N)}\int_3^4\int_3^4e^{-u-v}f_{N-2}(N-u-v,2N-u^2-v^2)\,du\,dv. μ ( U i ∩ U j ) ≥ f N ( N , 2 N ) 1 ∫ 3 4 ∫ 3 4 e − u − v f N − 2 ( N − u − v , 2 N − u 2 − v 2 ) d u d v .
Both densities are evaluated within O ( 1 ) O(1) O ( 1 ) of their barycentres ( n , 2 n ) (n,2n) ( n , 2 n ) , where Lemma III.6 gives n f n → G ( 0 ) > 0 nf_n \to G(0)>0 n f n → G ( 0 ) > 0 , G G G the limiting Gaussian density. Hence the right side tends to ( e − 3 − e − 4 ) 2 = 2 α 2 (e^{-3}-e^{-4})^2=2\alpha_2 ( e − 3 − e − 4 ) 2 = 2 α 2 , and it exceeds α 2 \alpha_2 α 2 for N ≥ N 0 N\ge N_0 N ≥ N 0 .
(b) Let ε > 2 \varepsilon>2 ε > 2 and a = ε − 2 a=\sqrt{\varepsilon-2} a = ε − 2 . Restrict u , v u,v u , v to ( ε , ε + 1 ) (\varepsilon,\varepsilon+1) ( ε , ε + 1 ) , and write Ω N − 2 ( S , E ) = ∫ 0 ∞ Ω N − 3 ( S − x , E − x 2 ) d x \Omega_{N-2}(S,E)=\int_0^\infty\Omega_{N-3}(S-x,E-x^2)\,dx Ω N − 2 ( S , E ) = ∫ 0 ∞ Ω N − 3 ( S − x , E − x 2 ) d x , restricting x x x to [ a N , a N + 1 ] [a\sqrt{N},a\sqrt{N}+1] [ a N , a N + 1 ] . The remaining n = N − 3 n=N-3 n = N − 3 sites have mass S ′ = N − u − v − x S'=N-u-v-x S ′ = N − u − v − x and energy E ′ = N ε − u 2 − v 2 − x 2 E'=N\varepsilon-u^2-v^2-x^2 E ′ = N ε − u 2 − v 2 − x 2 , so that S ′ − n = − a N + O ( 1 ) S'-n=-a\sqrt{N}+O(1) S ′ − n = − a N + O ( 1 ) and E ′ − 2 n = O ( N ) E'-2n=O(\sqrt{N}) E ′ − 2 n = O ( N ) . Lemma III.6 gives f n ( S ′ , E ′ ) ≥ c / N f_n(S',E')\ge c/N f n ( S ′ , E ′ ) ≥ c / N , hence Ω n ( S ′ , E ′ ) = e S ′ f n ( S ′ , E ′ ) ≥ c e N − a N / N \Omega_n(S',E')=e^{S'f_n(S',E')}\ge ce^{N-a\sqrt{N}}/N Ω n ( S ′ , E ′ ) = e S ′ f n ( S ′ , E ′ ) ≥ c e N − a N / N , and
Ω N ( N , N ε ) μ ( U i ∩ U j ) ≥ c e N − a N / N . \Omega_N(N,N\varepsilon)\,\mu(U_i\cap U_j)\ge ce^{N-a\sqrt{N}}/N. Ω N ( N , N ε ) μ ( U i ∩ U j ) ≥ c e N − a N / N .
Together with the stronger upper bound Ω N ( N , N ε ) ≤ C ε ′ N e N − a N \Omega_N(N,N\varepsilon)\le C'_\varepsilon N e^{N-a\sqrt{N}} Ω N ( N , N ε ) ≤ C ε ′ N e N − a N proved above,
μ ( U i ∩ U j ) ≥ c ε N − 2 , \mu(U_i\cap U_j)\ge c_\varepsilon N^{-2}, μ ( U i ∩ U j ) ≥ c ε N − 2 ,
so (b) holds with any C ε > 2 C_\varepsilon>2 C ε > 2 .
Consequently Theorem V.1(iii) holds with N − 6 N^{-6} N − 6 in place of N − C ε N^{-C_\varepsilon} N − C ε . We do not know whether the power N − 2 N^{-2} N − 2 is sharp; at ε = 2 \varepsilon=2 ε = 2 the overlap is bounded below by a constant.
Proofs of the neck estimates
Proof of Proposition IV.2. Step 1: exact representation. Almost surely the largest of q 2 , … , q N q_2,\ldots,q_N q 2 , … , q N is attained at a unique site. By permutation symmetry and (12) , conditioning on its position and value v v v gives
p N ( u ) = N − 1 Ω N ( N , N ε ) ∫ d v Ω N − 2 < v ( N − u − v , N ε − u 2 − v 2 ) , (92) p_N(u)=\frac{N-1}{\Omega_N(N,N\varepsilon)}\int dv\,\Omega_{N-2}^{<v}(N-u-v,N\varepsilon-u^2-v^2), \tag*{(92)} p N ( u ) = Ω N ( N , N ε ) N − 1 ∫ d v Ω N − 2 < v ( N − u − v , N ε − u 2 − v 2 ) , (92)
with Ω N − 2 < v \Omega_{N-2}^{<v} Ω N − 2 < v as in Lemma III.5. The range of v v v is bounded. On one side v ≤ N ε v\le\sqrt{N\varepsilon} v ≤ N ε . On the other side the N − 2 N-2 N − 2 remaining sites carry mass N − u − v N-u-v N − u − v and are all below v v v , so v ≥ ( N − u ) / ( N − 1 ) ≥ 1 2 v\ge(N-u)/(N-1)\ge\frac12 v ≥ ( N − u ) / ( N − 1 ) ≥ 2 1 for u ≤ N ε u\le\sqrt{N\varepsilon} u ≤ N ε and N N N large. Note that v v v may be larger or smaller than u u u : no ordering between sites 1 and 2 is assumed.
Step 2: the tilted bound. Write z 1 = u / N z_1=u/\sqrt{N} z 1 = u / N , z 2 = v / N z_2=v/\sqrt{N} z 2 = v / N . Lemma III.7 with k = 2 k=2 k = 2 , ς = z 1 + z 2 \varsigma=z_1+z_2 ς = z 1 + z 2 , r 2 = z 1 2 + z 2 2 r^2=z_1^2+z_2^2 r 2 = z 1 2 + z 2 2 and cutoff ζ = z 2 ≥ 1 / ( 2 N ) \zeta=z_2\ge1/(2\sqrt{N}) ζ = z 2 ≥ 1/ ( 2 N ) gives
ln Ω N − 2 < v ( N − u − v , N ε − u 2 − v 2 ) ≤ N − N F ( z 1 , z 2 ) + C ln N , (93) \ln\Omega_{N-2}^{<v}(N-u-v,N\varepsilon-u^2-v^2)\le N-\sqrt{N}F(z_1,z_2)+C\ln N, \tag*{(93)} ln Ω N − 2 < v ( N − u − v , N ε − u 2 − v 2 ) ≤ N − N F ( z 1 , z 2 ) + C ln N , (93)
with F F F as in (44) .
Step 3: the variational inequality. Let z 1 = a / 2 + δ z_1=a/\sqrt{2}+\delta z 1 = a / 2 + δ with ∣ δ ∣ ≤ K / N |\delta|\le K/\sqrt{N} ∣ δ ∣ ≤ K / N . Since max ( z 2 , a / 2 ) ≥ a / 2 \max(z_2,a/2)\ge a/2 max ( z 2 , a /2 ) ≥ a /2 , one has ∣ ∂ F / ∂ z 1 ∣ ≤ 1 + 4 z 1 / a |\partial F/\partial z_1|\le1+4z_1/a ∣ ∂ F / ∂ z 1 ∣ ≤ 1 + 4 z 1 / a on both sides of the circle ∣ z ∣ = a |z|=a ∣ z ∣ = a , and F F F is continuous across it. Hence F ( z 1 , z 2 ) ≥ F ( a / 2 , z 2 ) − C 1 K / N F(z_1,z_2)\ge F(a/\sqrt{2},z_2)-C_1K/\sqrt{N} F ( z 1 , z 2 ) ≥ F ( a / 2 , z 2 ) − C 1 K / N , with C 1 = 1 + 4 ε / a C_1=1+4\sqrt{\varepsilon}/a C 1 = 1 + 4 ε / a .
At z 1 = a / 2 z_1=a/\sqrt{2} z 1 = a / 2 there are three cases.
0 ≤ z 2 < a / 2 0\le z_2<a/2 0 ≤ z 2 < a /2 : F = a / 2 + a + z 2 − 2 z 2 2 / a F=a/\sqrt{2}+a+z_2-2z_2^2/a F = a / 2 + a + z 2 − 2 z 2 2 / a . This is concave in z 2 z_2 z 2 , so it is at least its value at the endpoints, a / 2 + a > 2 a a/\sqrt{2}+a>\sqrt{2}a a / 2 + a > 2 a .
Hence min z 2 F ( a / 2 , z 2 ) = 2 a \min_{z_2}F(a/\sqrt{2},z_2)=\sqrt{2}a min z 2 F ( a / 2 , z 2 ) = 2 a , attained at the neck z 2 = a / 2 z_2=a/\sqrt{2} z 2 = a / 2 , which is (45) . The integrand of (C1) is therefore at most exp { N − 2 a N + C 1 K + C ln N } \exp\{N-\sqrt{2}a\sqrt{N}+C_1K+C\ln N\} exp { N − 2 a N + C 1 K + C ln N } .
Step 4: normalisation. By Proposition III.8, Ω N ( N , N ε ) ≥ exp { N − a N − O ( ln N ) } \Omega_N(N,N\varepsilon)\ge\exp\{N-a\sqrt{N}-O(\ln N)\} Ω N ( N , N ε ) ≥ exp { N − a N − O ( ln N )} . The prefactor N − 1 N-1 N − 1 and the length N ε \sqrt{N\varepsilon} N ε of the range of v v v are polynomial. Hence p N ( u ) ≤ exp { − ( 2 − 1 ) a N + O ( ln N ) } p_N(u)\le\exp\{-(\sqrt{2}-1)a\sqrt{N}+O(\ln N)\} p N ( u ) ≤ exp { − ( 2 − 1 ) a N + O ( ln N )} , uniformly for ∣ u − a N / 2 ∣ ≤ K |u-a\sqrt{N}/2|\le K ∣ u − a N /2∣ ≤ K .
Proof of Lemma IV.3. Let u K = a N / 2 + K u_K=a\sqrt{N}/2+K u K = a N /2 + K . As in the proof of the upper bound in Proposition III.8, condition on the position and value u u u of the largest coordinate, which exceeds ε \varepsilon ε by Lemma III.4:
Ω N ( N , N ε ) μ ( max i q i < u K ) = N ∫ ε u K Ω N − 1 ≤ u ( N − u , N ε − u 2 ) d u . \Omega_N(N,N\varepsilon)\,\mu\left(\max_i q_i<u_K\right)=N\int_{\varepsilon}^{u_K}\Omega_{N-1}^{\le u}(N-u,N\varepsilon-u^2)\,du. Ω N ( N , N ε ) μ ( i max q i < u K ) = N ∫ ε u K Ω N − 1 ≤ u ( N − u , N ε − u 2 ) d u .
Lemma III.7 with k = 1 k=1 k = 1 and ζ = r = ς = u / N \zeta=r=\varsigma=u/\sqrt{N} ζ = r = ς = u / N bounds the integrand by exp { N − N F ( ζ ) + C ln N } \exp\{N-\sqrt{N}F(\zeta)+C\ln N\} exp { N − N F ( ζ ) + C ln N } , with F ( ζ ) = ζ + ( a 2 − ζ 2 ) + / max ( ζ , a / 2 ) F(\zeta)=\zeta+(a^2-\zeta^2)_+/\max(\zeta,a/2) F ( ζ ) = ζ + ( a 2 − ζ 2 ) + / max ( ζ , a /2 ) . On the range ζ ≤ a / 2 + K / N \zeta\le a/\sqrt{2}+K/\sqrt{N} ζ ≤ a / 2 + K / N there are two cases.
Hence the integrand is at most exp { N − 2 a N + O ( ln N ) } \exp\{N-\sqrt{2}a\sqrt{N}+O(\ln N)\} exp { N − 2 a N + O ( ln N )} . Dividing by the lower bound of Proposition III.8 proves the lemma.
The corners of the domains
Proof of Proposition V.3. Linear independence of the gradients. The tangent space of S R S_R S R at q q q is the orthogonal complement of 1 \mathbf{1} 1 and q q q , and ∇ q j \nabla q_j ∇ q j is the orthogonal projection of e j e_j e j onto it. Hence, for a set I I I of indices, a relation ∑ j ∈ I c j ∇ q j = 0 \sum_{j\in I}c_j\nabla q_j=0 ∑ j ∈ I c j ∇ q j = 0 holds if and only if ∑ j ∈ I c j e j = α 1 + β q \sum_{j\in I}c_je_j=\alpha\mathbf{1}+\beta q ∑ j ∈ I c j e j = α 1 + β q for some α , β ∈ R \alpha,\beta\in\mathbb{R} α , β ∈ R , that is,
c j = α + β q j ( j ∈ I ) , 0 = α + β q l ( l ∉ I ) . (94) c_j=\alpha+\beta q_j\quad(j\in I),\qquad0=\alpha+\beta q_l\quad(l\notin I). \tag*{(94)} c j = α + β q j ( j ∈ I ) , 0 = α + β q l ( l ∈ / I ) . (94)
In the two cases below some l ∉ I l\notin I l ∈ / I exists. If β = 0 \beta=0 β = 0 , then (D1) gives α = 0 \alpha=0 α = 0 and all c j = 0 c_j=0 c j = 0 . A nontrivial relation therefore has β ≠ 0 \beta\ne0 β = 0 , and then all the coordinates q l q_l q l , l ∉ I l\notin I l ∈ / I , are equal to the same m = − α / β m=-\alpha/\beta m = − α / β , which is positive.
(i) Let q ∈ M ∖ F ε q\in M\setminus F_\varepsilon q ∈ M ∖ F ε and I = { j : q j = 0 } I=\{j:q_j=0\} I = { j : q j = 0 } ; some l ∉ I l\notin I l ∈ / I exists because ∑ j q j = N \sum_jq_j=N ∑ j q j = N . If the k k k coordinates outside I I I were all equal to m m m , the constraints would give k m = N km=N k m = N and k m 2 = N ε km^2=N\varepsilon k m 2 = N ε , so m = ε m=\varepsilon m = ε and q ∈ F ε q\in F_\varepsilon q ∈ F ε . Hence the gradients of the coordinates vanishing at q q q are linearly independent.
(ii) Let K ∈ ( ε , t ∗ ) K\in(\varepsilon,t^*) K ∈ ( ε , t ∗ ) , q ∈ M q\in M q ∈ M with q i = K q_i=K q i = K , and I = { i } ∪ { j : q j = 0 } I=\{i\}\cup\{j:q_j=0\} I = { i } ∪ { j : q j = 0 } ; some l ∉ I l\notin I l ∈ / I exists, since otherwise q i = N q_i=N q i = N and N 2 = N ε N^2=N\varepsilon N 2 = N ε . If the k ∈ { 1 , … , N − 1 } k\in\{1,\ldots,N-1\} k ∈ { 1 , … , N − 1 } coordinates outside I I I were all equal to m m m , the constraints would give K + k m = N K+km=N K + k m = N and K 2 + k m 2 = N ε K^2+km^2=N\varepsilon K 2 + k m 2 = N ε , that is, k K 2 + ( N − K ) 2 = k N ε kK^2+(N-K)^2=kN\varepsilon k K 2 + ( N − K ) 2 = k N ε . For each k k k this quadratic equation has at most two roots K K K . Hence, for K K K outside a set of at most 2 N 2N 2 N values, at every q ∈ M q\in M q ∈ M with q i = K q_i=K q i = K the gradients of q i q_i q i and of the coordinates vanishing at q q q are linearly independent.
(a) and (b). Let D D D be M M M or the cap, written as D = { q ∈ S R : f 1 ( q ) ≥ 0 , … , f s ( q ) ≥ 0 } D=\{q\in S_R:f_1(q)\ge0,\ldots,f_s(q)\ge0\} D = { q ∈ S R : f 1 ( q ) ≥ 0 , … , f s ( q ) ≥ 0 } , where the f f f ’s are the coordinates and, for the cap, q i − K q_i-K q i − K . D D D is closed in the compact manifold S R S_R S R , hence compact; it is connected by Lemma III.4 and Lemma III.2(3). Let q ∈ D q\in D q ∈ D , q ∉ F ε q\notin F_\varepsilon q ∈ / F ε , and let f 1 , … , f m f_1,\ldots,f_m f 1 , … , f m be the functions vanishing at q q q . Their differentials at q q q are linearly independent: by (i) if q i > K q_i>K q i > K or D = M D=M D = M , by (ii) if q i = K q_i=K q i = K . (The cap contains no point of F ε F_\varepsilon F ε , whose coordinates are at most ε < K \varepsilon<K ε < K .) By the inverse function theorem, f 1 , … , f m f_1,\ldots,f_m f 1 , … , f m together with d − m d-m d − m suitable linear functions are coordinates of S R S_R S R near q q q . The other f f f ’s are positive near q q q , so in these coordinates D D D is, near q q q , the set { f 1 ≥ 0 , … , f m ≥ 0 } \{f_1\ge0,\ldots,f_m\ge0\} { f 1 ≥ 0 , … , f m ≥ 0 } : a neighbourhood of 0 0 0 in [ 0 , ∞ ) m × R d − m [0,\infty)^m\times\mathbb{R}^{d-m} [ 0 , ∞ ) m × R d − m .
(c). Let Z = { j : p j = 0 } Z=\{j:p_j=0\} Z = { j : p j = 0 } and write q = p + v q=p+v q = p + v with q ∈ M q\in M q ∈ M . Then ∑ j v j = 0 \sum_jv_j=0 ∑ j v j = 0 , and ∑ j ( p j + v j ) 2 = N ε = ∑ j p j 2 \sum_j(p_j+v_j)^2=N\varepsilon=\sum_jp_j^2 ∑ j ( p j + v j ) 2 = N ε = ∑ j p j 2 gives 2 ε ∑ j ∉ Z v j + ∣ v ∣ 2 = 0 2\varepsilon\sum_{j\notin Z}v_j+|v|^2=0 2 ε ∑ j ∈ / Z v j + ∣ v ∣ 2 = 0 . Since ∑ j ∉ Z v j = − ∑ j ∈ Z v j \sum_{j\notin Z}v_j=-\sum_{j\in Z}v_j ∑ j ∈ / Z v j = − ∑ j ∈ Z v j and v j = q j v_j=q_j v j = q j for j ∈ Z j\in Z j ∈ Z , this is (56) . Along a curve q ( s ) q(s) q ( s ) in M M M with q ( 0 ) = p q(0)=p q ( 0 ) = p , the left side of (56) is a sum of nonnegative terms of order s 2 s^2 s 2 , so q ˙ j ( 0 ) = 0 \dot{q}_j(0)=0 q ˙ j ( 0 ) = 0 for j ∈ Z j\in Z j ∈ Z . The vectors of T p T_p T p with v j = 0 v_j=0 v j = 0 for j ∈ Z j\in Z j ∈ Z form a subspace of dimension k − 1 < d k-1<d k − 1 < d , while the tangent vectors at 0 0 0 of curves in [ 0 , ∞ ) m × R d − m [0,\infty)^m\times\mathbb{R}^{d-m} [ 0 , ∞ ) m × R d − m span R d \mathbb{R}^d R d .
The properties of the profile φ \varphi φ
Proof of ( φ 1 ) (\varphi1) ( φ 1 ) –( φ 5 ) (\varphi5) ( φ 5 ) . Put
u ( x ) = ε ( 10 − ε ) ( 2 − ε ) x + 8 ε , 0 < u ( x ) ≤ u ( 0 ) = 10 − ε 8 < 9 8 . u(x)=\frac{\varepsilon(10-\varepsilon)}{(2-\varepsilon)x+8\varepsilon},\qquad0<u(x)\le u(0)=\frac{10-\varepsilon}{8}<\frac{9}{8}. u ( x ) = ( 2 − ε ) x + 8 ε ε ( 10 − ε ) , 0 < u ( x ) ≤ u ( 0 ) = 8 10 − ε < 8 9 .
and note that φ ( x ) / ( x − ε ) = ( 2 − ε ) x ( x − ε ) / ( ( 2 − ε ) x + 8 ε ) \varphi(x)/(x-\varepsilon)=(2-\varepsilon)x(x-\varepsilon)/((2-\varepsilon)x+8\varepsilon) φ ( x ) / ( x − ε ) = ( 2 − ε ) x ( x − ε ) / (( 2 − ε ) x + 8 ε ) is a rational function, with
( φ ( x ) x − ε ) 2 = φ ( x ) ( 2 − ε ) x ( 2 − ε ) x + 8 ε ≤ φ ( x ) . \left(\frac{\varphi(x)}{x-\varepsilon}\right)^2=\varphi(x)\frac{(2-\varepsilon)x}{(2-\varepsilon)x+8\varepsilon}\le\varphi(x). ( x − ε φ ( x ) ) 2 = φ ( x ) ( 2 − ε ) x + 8 ε ( 2 − ε ) x ≤ φ ( x ) .
Let ξ ( x ) = ( x − ε ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − x / ε ) ) \xi(x)=(x-\varepsilon)(\varphi'(x)-\varphi'(0)(1-x/\varepsilon)) ξ ( x ) = ( x − ε ) ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − x / ε )) ; then ξ ′ ( x ) − ξ ′ ( 0 ) ( 1 − x / ε ) \xi'(x)-\xi'(0)(1-x/\varepsilon) ξ ′ ( x ) − ξ ′ ( 0 ) ( 1 − x / ε ) is the function in ( φ 4 ) (\varphi4) ( φ 4 ) . Differentiating (57) gives
φ ′ ( 0 ) = ε ( 2 − ε ) 8 , φ ′ ′ = 2 − 16 u 3 10 − ε , \varphi'(0)=\frac{\varepsilon(2-\varepsilon)}{8},\qquad\varphi''=2-\frac{16u^3}{10-\varepsilon}, φ ′ ( 0 ) = 8 ε ( 2 − ε ) , φ ′′ = 2 − 10 − ε 16 u 3 ,
φ ′ ( x ) − φ ′ ( 0 ) ( 1 − x ε ) = φ ( x ) x − ε ⋅ 18 − ε + 8 u ( x ) 8 , hence ξ = φ 18 − ε + 8 u 8 , \varphi'(x)-\varphi'(0)\left(1-\frac{x}{\varepsilon}\right)=\frac{\varphi(x)}{x-\varepsilon}\cdot\frac{18-\varepsilon+8u(x)}{8},\qquad\text{hence}\qquad\xi=\varphi\frac{18-\varepsilon+8u}{8}, φ ′ ( x ) − φ ′ ( 0 ) ( 1 − ε x ) = x − ε φ ( x ) ⋅ 8 18 − ε + 8 u ( x ) , hence ξ = φ 8 18 − ε + 8 u ,
ξ ′ ( x ) − ξ ′ ( 0 ) ( 1 − x ε ) = φ ( x ) x − ε B ( x ) , B ( x ) = ε 2 − 24 ε + 172 32 + ( 14 − ε ) u ( x ) 4 + 2 u ( x ) 2 . \xi'(x)-\xi'(0)\left(1-\frac{x}{\varepsilon}\right)=\frac{\varphi(x)}{x-\varepsilon}B(x),\qquad B(x)=\frac{\varepsilon^2-24\varepsilon+172}{32}+\frac{(14-\varepsilon)u(x)}{4}+2u(x)^2. ξ ′ ( x ) − ξ ′ ( 0 ) ( 1 − ε x ) = x − ε φ ( x ) B ( x ) , B ( x ) = 32 ε 2 − 24 ε + 172 + 4 ( 14 − ε ) u ( x ) + 2 u ( x ) 2 .
The bounds follow from 1 < ε < 2 1<\varepsilon<2 1 < ε < 2 and 0 < 8 u ≤ 10 − ε < 9 0<8u\le10-\varepsilon<9 0 < 8 u ≤ 10 − ε < 9 .
( φ 1 ) (\varphi1) ( φ 1 ) . Since 0 ≤ ( 2 − ε ) x / ( ( 2 − ε ) x + 8 ε ) < 1 0\le(2-\varepsilon)x/((2-\varepsilon)x+8\varepsilon)<1 0 ≤ ( 2 − ε ) x / (( 2 − ε ) x + 8 ε ) < 1 , 0 ≤ φ ≤ ( x − ε ) 2 0\le\varphi\le(x-\varepsilon)^2 0 ≤ φ ≤ ( x − ε ) 2 . Moreover 2 > φ ′ ′ ≥ 2 − ( 10 − ε ) 2 / 32 > − 1 2>\varphi''\ge2-(10-\varepsilon)^2/32>-1 2 > φ ′′ ≥ 2 − ( 10 − ε ) 2 /32 > − 1 .
( φ 2 ) (\varphi2) ( φ 2 ) . Since 16 < 18 − ε + 8 u < 26 16<18-\varepsilon+8u<26 16 < 18 − ε + 8 u < 26 , 2 φ ≤ ξ ≤ 13 4 φ 2\varphi\le\xi\le\frac{13}{4}\varphi 2 φ ≤ ξ ≤ 4 13 φ .
( φ 3 ) (\varphi3) ( φ 3 ) . By the bound on ( φ / ( x − ε ) ) 2 (\varphi/(x-\varepsilon))^2 ( φ / ( x − ε ) ) 2 ,
( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − x ε ) ) 2 ≤ φ ( 18 − ε + 8 u 8 ) 2 = ξ 18 − ε + 8 u 8 ≤ 13 4 ξ . \left(\varphi'(x)-\varphi'(0)\left(1-\frac{x}{\varepsilon}\right)\right)^2\le\varphi\left(\frac{18-\varepsilon+8u}{8}\right)^2=\xi\frac{18-\varepsilon+8u}{8}\le\frac{13}{4}\xi. ( φ ′ ( x ) − φ ′ ( 0 ) ( 1 − ε x ) ) 2 ≤ φ ( 8 18 − ε + 8 u ) 2 = ξ 8 18 − ε + 8 u ≤ 4 13 ξ .
( φ 4 ) (\varphi4) ( φ 4 ) . The three terms of B B B are positive and smaller than 5, 4 and 3, so 0 < B < 12 0<B<12 0 < B < 12 and
( ξ ′ ( x ) − ξ ′ ( 0 ) ( 1 − x ε ) ) 2 ≤ 144 φ ( x ) . \left(\xi'(x)-\xi'(0)\left(1-\frac{x}{\varepsilon}\right)\right)^2\le144\varphi(x). ( ξ ′ ( x ) − ξ ′ ( 0 ) ( 1 − ε x ) ) 2 ≤ 144 φ ( x ) .
( φ 5 ) (\varphi5) ( φ 5 ) . Let Ψ = φ ′ ( 0 ) + φ ′ ′ − φ ′ \Psi=\varphi'(0)+\varphi''-\varphi' Ψ = φ ′ ( 0 ) + φ ′′ − φ ′ . Since φ ′ ′ − 2 \varphi''-2 φ ′′ − 2 is a negative multiple of u 3 u^3 u 3 , whose derivatives alternate in sign, φ ( 4 ) < 0 < φ ( 5 ) \varphi^{(4)}<0<\varphi^{(5)} φ ( 4 ) < 0 < φ ( 5 ) and Ψ ′ ′ ′ = φ ( 5 ) − φ ( 4 ) > 0 \Psi'''=\varphi^{(5)}-\varphi^{(4)}>0 Ψ ′′′ = φ ( 5 ) − φ ( 4 ) > 0 . Therefore
Ψ ( x ) ≥ Ψ ( 0 ) + Ψ ′ ( 0 ) x + 1 2 Ψ ′ ′ ( 0 ) x 2 \Psi(x)\ge\Psi(0)+\Psi'(0)x+\frac{1}{2}\Psi''(0)x^2 Ψ ( x ) ≥ Ψ ( 0 ) + Ψ ′ ( 0 ) x + 2 1 Ψ ′′ ( 0 ) x 2
for x ≥ 0 x\ge0 x ≥ 0 , and integrating against d ν d\nu d ν ,
∫ Ψ d ν ≥ Ψ ( 0 ) + Ψ ′ ( 0 ) + ε 2 Ψ ′ ′ ( 0 ) = 3 ( 10 − ε ) 2 ( 2 − ε ) 2 1024 ε ≥ 3 ( 2 − ε ) 2 16 ε . \int\Psi\,d\nu\ge\Psi(0)+\Psi'(0)+\frac{\varepsilon}{2}\Psi''(0)=\frac{3(10-\varepsilon)^2(2-\varepsilon)^2}{1024\varepsilon}\ge\frac{3(2-\varepsilon)^2}{16\varepsilon}. ∫ Ψ d ν ≥ Ψ ( 0 ) + Ψ ′ ( 0 ) + 2 ε Ψ ′′ ( 0 ) = 1024 ε 3 ( 10 − ε ) 2 ( 2 − ε ) 2 ≥ 16 ε 3 ( 2 − ε ) 2 .
The chart of H i H_i H i
Proof of Lemma VI.1. The chart. For q ∈ H i q \in H_i q ∈ H i the other coordinates are positive with sum N − t N-t N − t , so x ∈ T ∘ x \in T^\circ x ∈ T ∘ ; and ∑ j ≠ i q j 2 = N ε − t 2 \sum_{j\ne i}q_j^2=N\varepsilon-t^2 ∑ j = i q j 2 = N ε − t 2 gives ∣ x ∣ 2 = k ( t ) |x|^2=k(t) ∣ x ∣ 2 = k ( t ) . Since
k ′ ( t ) = − 2 N ( t − ε ) ( N − t ) 3 < 0 ( ε < t < N ) , k'(t)=-\frac{2N(t-\varepsilon)}{(N-t)^3}<0 \qquad(\varepsilon<t<N), k ′ ( t ) = − ( N − t ) 3 2 N ( t − ε ) < 0 ( ε < t < N ) ,
the condition t > 2 ε t>2\varepsilon t > 2 ε is equivalent to ∣ x ∣ 2 < k ( 2 ε ) |x|^2<k(2\varepsilon) ∣ x ∣ 2 < k ( 2 ε ) . Conversely, let x ∈ Ω x\in\Omega x ∈ Ω . By Cauchy–Schwarz ∣ x ∣ 2 ≥ 1 / ( N − 1 ) = k ( t ∗ ) |x|^2\ge1/(N-1)=k(t^*) ∣ x ∣ 2 ≥ 1/ ( N − 1 ) = k ( t ∗ ) ; the last equality holds because at the pole the other coordinates all equal ( N − t ∗ ) / ( N − 1 ) (N-t^*)/(N-1) ( N − t ∗ ) / ( N − 1 ) . So there is a unique t ( x ) ∈ ( 2 ε , t ∗ ] t(x)\in(2\varepsilon,t^*] t ( x ) ∈ ( 2 ε , t ∗ ] with k ( t ( x ) ) = ∣ x ∣ 2 k(t(x))=|x|^2 k ( t ( x )) = ∣ x ∣ 2 , smooth in x x x because k ′ ≠ 0 k'\ne0 k ′ = 0 , and F ( x ) ∈ H i F(x)\in H_i F ( x ) ∈ H i : its coordinates are positive, sum to t + ( N − t ) = N t+(N-t)=N t + ( N − t ) = N , and have sum of squares t 2 + ( N − t ) 2 k ( t ) = N ε t^2+(N-t)^2k(t)=N\varepsilon t 2 + ( N − t ) 2 k ( t ) = N ε . The two maps are inverse to each other.
(a) The density. In the meridian coordinates of Lemma III.2(2), ( q j ) j ≠ i = c 1 + ρ ω (q_j)_{j\ne i}=c\mathbf{1}+\rho\omega ( q j ) j = i = c 1 + ρ ω with ω \omega ω in the unit sphere of { v ∈ R N − 1 : ∑ j v j = 0 } \{v\in\mathbb{R}^{N-1}:\sum_jv_j=0\} { v ∈ R N − 1 : ∑ j v j = 0 } , and d σ = ρ N − 3 γ − 1 / 2 d t d ω d\sigma=\rho^{N-3}\gamma^{-1/2}\,dt\,d\omega d σ = ρ N − 3 γ − 1/2 d t d ω . Dividing by N − t = ( N − 1 ) c N-t=(N-1)c N − t = ( N − 1 ) c gives x = x ˉ + r ω x=\bar{x}+r\omega x = x ˉ + r ω , with x ˉ = 1 / ( N − 1 ) \bar{x}=1/(N-1) x ˉ = 1/ ( N − 1 ) and r = ρ / ( N − t ) r=\rho/(N-t) r = ρ / ( N − t ) . These are polar coordinates around x ˉ \bar{x} x ˉ in the hyperplane, so d x = r N − 3 d r d ω dx=r^{N-3}\,dr\,d\omega d x = r N − 3 d r d ω . Since r 2 = ∣ x ∣ 2 − 1 N − 1 = k ( t ) − 1 N − 1 r^2=|x|^2-\frac{1}{N-1}=k(t)-\frac{1}{N-1} r 2 = ∣ x ∣ 2 − N − 1 1 = k ( t ) − N − 1 1 , we have 2 r d r = k ′ ( t ) d t 2r\,dr=k'(t)\,dt 2 r d r = k ′ ( t ) d t , i.e. ∣ d t / d r ∣ = r ( N − t ) 3 / ( N ( t − ε ) ) |dt/dr|=r(N-t)^3/(N(t-\varepsilon)) ∣ d t / d r ∣ = r ( N − t ) 3 / ( N ( t − ε )) . Hence
d σ = ρ N − 3 r N − 3 γ − 1 / 2 ∣ d t d r ∣ d x = ( N − t ) N − 1 N ( t − ε ) ρ γ − 1 / 2 d x . d\sigma=\frac{\rho^{N-3}}{r^{N-3}}\gamma^{-1/2}\left|\frac{dt}{dr}\right|dx=\frac{(N-t)^{N-1}}{N(t-\varepsilon)}\rho\gamma^{-1/2}dx. d σ = r N − 3 ρ N − 3 γ − 1/2 d r d t d x = N ( t − ε ) ( N − t ) N − 1 ρ γ − 1/2 d x .
From the definitions of ρ \rho ρ and γ \gamma γ (Lemmas III.1 and III.2), ρ 2 = N N − 1 R 2 γ \rho^2=\frac{N}{N-1}R^2\gamma ρ 2 = N − 1 N R 2 γ , so ρ γ − 1 / 2 = R N / ( N − 1 ) = N ( ε − 1 ) / ( N − 1 ) \rho\gamma^{-1/2}=R\sqrt{N/(N-1)}=N\sqrt{(\varepsilon-1)/(N-1)} ρ γ − 1/2 = R N / ( N − 1 ) = N ( ε − 1 ) / ( N − 1 ) ; this gives w w w . Both factors ( N − t ) N − 1 (N-t)^{N-1} ( N − t ) N − 1 and 1 / ( t − ε ) 1/(t-\varepsilon) 1/ ( t − ε ) decrease in t t t . The two-sided bound follows from − u − u 2 ≤ log ( 1 − u ) ≤ − u -u-u^2\le\log(1-u)\le-u − u − u 2 ≤ log ( 1 − u ) ≤ − u for 0 ≤ u ≤ 1 2 0\le u\le\frac12 0 ≤ u ≤ 2 1 , applied with u = t / N u=t/N u = t / N and multiplied by N − 1 N-1 N − 1 .
(b) The diameter. Ω \Omega Ω lies in the ball of centre x ˉ \bar{x} x ˉ and radius r 0 r_0 r 0 , where r 0 2 = k ( 2 ε ) − 1 N − 1 < N ε ( N − 2 ε ) 2 ≤ 2 ε N r_0^2=k(2\varepsilon)-\frac{1}{N-1}<\frac{N\varepsilon}{(N-2\varepsilon)^2}\le\frac{2\varepsilon}{N} r 0 2 = k ( 2 ε ) − N − 1 1 < ( N − 2 ε ) 2 N ε ≤ N 2 ε , because ( N − 2 ε ) 2 ≥ N 2 / 2 (N-2\varepsilon)^2\ge N^2/2 ( N − 2 ε ) 2 ≥ N 2 /2 for N ≥ 7 ε N\ge7\varepsilon N ≥ 7 ε . So diam ( Ω ) 2 ≤ 4 r 0 2 ≤ 8 ε / N \operatorname{diam}(\Omega)^2\le4r_0^2\le8\varepsilon/N diam ( Ω ) 2 ≤ 4 r 0 2 ≤ 8 ε / N .
(c) The Lipschitz bound. Let v v v be tangent to Ω \Omega Ω , i.e. ∑ j v j = 0 \sum_jv_j=0 ∑ j v j = 0 . Differentiating k ( t ( x ) ) = ∣ x ∣ 2 k(t(x))=|x|^2 k ( t ( x )) = ∣ x ∣ 2 gives k ′ ( t ) d t ( v ) = 2 ⟨ x , v ⟩ = 2 ⟨ x − x ˉ , v ⟩ k'(t)\,dt(v)=2\langle x,v\rangle=2\langle x-\bar{x},v\rangle k ′ ( t ) d t ( v ) = 2 ⟨ x , v ⟩ = 2 ⟨ x − x ˉ , v ⟩ , so, using t − ε > ε t-\varepsilon>\varepsilon t − ε > ε , N − t < N N-t<N N − t < N and r 0 ≤ 2 ε / N r_0\le\sqrt{2\varepsilon/N} r 0 ≤ 2 ε / N ,
∣ d t ( v ) ∣ ≤ 2 r 0 ∣ v ∣ ∣ k ′ ( t ) ∣ = r 0 ( N − t ) 3 N ( t − ε ) ∣ v ∣ ≤ r 0 N 2 ε ∣ v ∣ ≤ 2 ε N 3 / 2 ∣ v ∣ . |dt(v)|\le\frac{2r_0|v|}{|k'(t)|}=\frac{r_0(N-t)^3}{N(t-\varepsilon)}|v|\le\frac{r_0N^2}{\varepsilon}|v|\le\sqrt{\frac{2}{\varepsilon}}N^{3/2}|v|. ∣ d t ( v ) ∣ ≤ ∣ k ′ ( t ) ∣ 2 r 0 ∣ v ∣ = N ( t − ε ) r 0 ( N − t ) 3 ∣ v ∣ ≤ ε r 0 N 2 ∣ v ∣ ≤ ε 2 N 3/2 ∣ v ∣.
Since d F ( v ) = ( d t ( v ) , ( ( N − t ) v j − x j d t ( v ) ) j ≠ i ) dF(v)=\left(dt(v),\left((N-t)v_j-x_jdt(v)\right)_{j\ne i}\right) d F ( v ) = ( d t ( v ) , ( ( N − t ) v j − x j d t ( v ) ) j = i ) and ∣ x ∣ ≤ 1 |x|\le1 ∣ x ∣ ≤ 1 ,
∣ d F ( v ) ∣ 2 ≤ d t ( v ) 2 + 2 ( N − t ) 2 ∣ v ∣ 2 + 2 ∣ x ∣ 2 d t ( v ) 2 ≤ 2 N 2 ∣ v ∣ 2 + 3 d t ( v ) 2 ≤ ( 2 + 6 ε ) N 3 ∣ v ∣ 2 . |dF(v)|^2\le dt(v)^2+2(N-t)^2|v|^2+2|x|^2dt(v)^2\le2N^2|v|^2+3\,dt(v)^2\le\left(2+\frac{6}{\varepsilon}\right)N^3|v|^2. ∣ d F ( v ) ∣ 2 ≤ d t ( v ) 2 + 2 ( N − t ) 2 ∣ v ∣ 2 + 2∣ x ∣ 2 d t ( v ) 2 ≤ 2 N 2 ∣ v ∣ 2 + 3 d t ( v ) 2 ≤ ( 2 + ε 6 ) N 3 ∣ v ∣ 2 .
Numerical methods
This appendix describes the numerical methods behind Figure 1 , Figure 2 and Figure 4 . None of the proofs uses them.
Evaluation of the partition functions
Figure 1 evaluates Ω n ( S , E ) \Omega_n(S,E) Ω n ( S , E ) by exponential tilting and an exact decomposition into fluid and condensate. For a cutoff ln N ≪ L ≪ a N \ln N \ll L \ll a\sqrt{N} ln N ≪ L ≪ a N , classifying configurations by the number k k k of sites with q i > L q_i>L q i > L ,
Ω n ( S , E ) = ∑ k ≥ 0 ( n k ) ∫ ( L , ∞ ) k d k q Ω n − k < L ( S − ∑ j q j , E − ∑ j q j 2 ) (95) \Omega_n(S,E)=\sum_{k\geq0}\binom{n}{k}\int_{(L,\infty)^k}\mathrm{d}^kq\ \Omega_{n-k}^{<L}\left(S-\sum_jq_j,E-\sum_jq_j^2\right) \tag*{(95)} Ω n ( S , E ) = k ≥ 0 ∑ ( k n ) ∫ ( L , ∞ ) k d k q Ω n − k < L ( S − j ∑ q j , E − j ∑ q j 2 ) (95)
with Ω m < L \Omega_m^{<L} Ω m < L as in Lemma III.5. On a bounded support Z L ( η , β ) = ∫ 0 L e − η q − β q 2 d q \mathcal{Z}_L(\eta,\beta)=\int_0^L e^{-\eta q-\beta q^2}\,\mathrm{d}q Z L ( η , β ) = ∫ 0 L e − η q − β q 2 d q is finite for every real β \beta β , including β < 0 \beta<0 β < 0 , so the saddle point always exists, and the local limit theorem gives ln Ω m < L ( S ′ , E ′ ) = η S ′ + β E ′ + m ln Z L − ln ( 2 π m det C ) + O ( 1 / m ) \ln\Omega_m^{<L}(S',E')=\eta S'+\beta E'+m\ln\mathcal{Z}_L-\ln(2\pi m\sqrt{\det C})+O(1/m) ln Ω m < L ( S ′ , E ′ ) = η S ′ + β E ′ + m ln Z L − ln ( 2 π m det C ) + O ( 1/ m ) , with C C C the tilted covariance of ( q , q 2 ) (q,q^2) ( q , q 2 ) . The terms k = 1 , 2 k=1,2 k = 1 , 2 are computed by quadrature and k ≥ 3 k\geq3 k ≥ 3 is negligible. Independence of L L L , the normalisation of the one- and two-site marginals, and agreement with an exact-constraint Montecarlo at N = 100 N=100 N = 100 check the accuracy.
Montecarlo dynamics
We simulated Brownian motion on the microcanonical surface M N , ε = { q ∈ R ≥ 0 N : ∑ i q i = N , ∑ i q i 2 = N ε } M_{N,\varepsilon}=\{q\in\mathbb{R}_{\geq0}^N:\sum_iq_i=N,\ \sum_iq_i^2=N\varepsilon\} M N , ε = { q ∈ R ≥ 0 N : ∑ i q i = N , ∑ i q i 2 = N ε } and measured its relaxation rates directly. We use the technique in [26 , 51 ] . The elementary move picks three distinct sites ( i , j , k ) (i,j,k) ( i , j , k ) at random and rotates ( q i , q j , q k ) (q_i,q_j,q_k) ( q i , q j , q k ) by an angle φ ∼ U [ − d , d ] \varphi\sim U[-d,d] φ ∼ U [ − d , d ] along the circle on which both q i + q j + q k q_i+q_j+q_k q i + q j + q k and q i 2 + q j 2 + q k 2 q_i^2+q_j^2+q_k^2 q i 2 + q j 2 + q k 2 are fixed. A move that would make a coordinate negative is rejected. The microcanonical measure restricted to each such circle is uniform in the angle, so the move satisfies detailed balance exactly for any d d d . Time is calibrated through the exact identity ∑ t X t 2 = N R 2 Δ M \sum_tX_t^2=NR^2\Delta_M ∑ t X t 2 = N R 2 Δ M , with X t X_t X t the rotation generator of triple t t t and R 2 = N ( ε − 1 ) R^2=N(\varepsilon-1) R 2 = N ( ε − 1 ) . Assigning each move the time step δ t = d 2 R 2 / [ 3 ( N − 1 ) ( N − 2 ) ] \delta t=d^2R^2/[3(N-1)(N-2)] δ t = d 2 R 2 / [ 3 ( N − 1 ) ( N − 2 )] makes the generator of the simulated process Δ M \Delta_M Δ M , so the measured rates are eigenvalues of − Δ M -\Delta_M − Δ M , in the same normalisation as λ M F \lambda_{\mathrm{MF}} λ MF . The dynamics commutes with permutations of the sites, so L 2 ( M ) L^2(M) L 2 ( M ) splits into S N S_N S N sectors, each with its own gap.
We probed these sectors separately, using time autocorrelations of (i) single-site functions f ( q i ) f(q_i) f ( q i ) , f ∈ { q , q 2 , q 3 , e − q } f \in\{q,q^2,q^3,e^{-q}\} f ∈ { q , q 2 , q 3 , e − q } , from which the combination C s a m e − C c r o s s C_{\mathrm{same}}-C_{\mathrm{cross}} C same − C cross isolates the standard sector; (ii) two-site observables ( φ a ( q i ) − φ a ( q l ) ) ( φ b ( q j ) − φ b ( q k ) ) (\varphi_a(q_i)-\varphi_a(q_l))(\varphi_b(q_j)-\varphi_b(q_k)) ( φ a ( q i ) − φ a ( q l )) ( φ b ( q j ) − φ b ( q k )) on disjoint quadruples of sites, which have no trivial or standard component and therefore probe the sectors ( N − 2 , 2 ) (N-2,2) ( N − 2 , 2 ) , ( N − 2 , 1 , 1 ) (N-2,1,1) ( N − 2 , 1 , 1 ) , … \ldots … ; and (iii) collective observables ( ∑ i q i 3 , ∑ i e − q i , ∑ i q i , max i q i ) (\sum_i q_i^3,\sum_i e^{-q_i},\sum_i \sqrt{q_i},\max_i q_i) ( ∑ i q i 3 , ∑ i e − q i , ∑ i q i , max i q i ) , which probe the trivial sector.
In each sector the slowest rate was obtained from a generalised eigenvalue problem on the correlation matrix of the observables, followed by a single-exponential fit on a window selected from the data alone, with no reference to λ M F \lambda_{\mathrm{MF}} λ MF . Pointwise errors on the correlations come from Bartlett’s formula, and errors on the rates from a leave-one-seed-out jackknife over 8 independent seeds. We used ε ∈ { 1.3 , 1.5 , 1.7 , 1.9 } \varepsilon\in\{1.3,1.5,1.7,1.9\} ε ∈ { 1.3 , 1.5 , 1.7 , 1.9 } and N ∈ { 100 , … , 1600 } N \in\{100,\ldots,1600\} N ∈ { 100 , … , 1600 } , with a simulated time per seed such that T λ M F ≈ 3 × 10 4 T\lambda_{\mathrm{MF}} \approx3 \times10^4 T λ MF ≈ 3 × 1 0 4 . Rejection at the faces q i = 0 q_i=0 q i = 0 induces a bias in the rates that is linear in the step size, so every point was run at d = 0.4 d=0.4 d = 0.4 , 0.2 0.2 0.2 and 0.1 0.1 0.1 .
The rates were then extrapolated first linearly to d → 0 d \to0 d → 0 and then as λ M C ( N ) = λ M C + c / N \lambda_{\mathrm{MC}}(N)=\lambda_{\mathrm{MC}}+c/N λ MC ( N ) = λ MC + c / N to N → ∞ N \to\infty N → ∞ . At ε = 1.9 \varepsilon=1.9 ε = 1.9 we added runs at N = 3200 N=3200 N = 3200 and 6400 6400 6400 with d = 0.05 d=0.05 d = 0.05 and 0.1 0.1 0.1 . These use reflection of the angle at the ends of the allowed arc instead of rejection. Reflection is also exactly reversible and reduces the slope of the O ( d ) O(d) O ( d ) bias by about one third.
References