Annales de l’Institut Henri Poincaré - Probabilités et Statistiques

2018, Vol. 54, No. 3, 1731–1757

https://doi.org/10.1214/17-AIHP853
© Association des Publications de l’Institut Henri Poincaré, 2018

www.imstat.org/aihp

Interpolation process between standard diffusion

and fractional diffusion1
Cédric Bernardina,2 , Patrícia Gonçalvesb,3 , Milton Jarac,4 and Marielle Simond,e,5
a Université Côte d’Azur, CNRS, LJAD, Parc Valrose, 06108 NICE Cedex 02, France. E-mail: cbernard@unice.fr
b Center for Mathematical Analysis, Geometry and Dynamical Systems, Instituto Superior Técnico, Universidade de Lisboa, Av. Rovisco Pais,
1049-001 Lisboa, Portugal. E-mail: patricia.goncalves@math.tecnico.ulisboa.pt
c Instituto de Matemática Pura e Aplicada, Estrada Dona Castorina 110, 22460-320 Rio De Janeiro, Brazil. E-mail: mjara@impa.br
d Inria Lille – Nord Europe, 40 avenue du Halley, 59650 Villeneuve d’Ascq, France. E-mail: marielle.simon@inria.fr
e Laboratoire Paul Painlevé, UMR CNRS 8524, Cité Scientifique, 59655 Villeneuve d’Ascq, France

Received 25 July 2016; revised 29 March 2017; accepted 10 July 2017

Abstract. We consider a Hamiltonian lattice field model with two conserved quantities, energy and volume, perturbed by stochastic
noise preserving the two previous quantities. It is known that this model displays anomalous diffusion of energy of fractional type
due to the conservation of the volume (Nonlinearity 25 (4) (2012) 1099–1133; Arch. Ration. Mech. Anal. 220 (2) (2016) 505–542).
We superpose to this system a second stochastic noise conserving energy but not volume. If the intensity of this noise is of order
one, normal diffusion of energy is restored while it is without effect if intensity is sufficiently small. In this paper we investigate the
nature of the energy fluctuations for a critical value of the intensity. We show that the latter are described by an Ornstein–Uhlenbeck
process driven by a Lévy process which interpolates between Brownian motion and the maximally asymmetric 3/2-stable Lévy
process. This result extends and solves a problem left open in (J. Stat. Phys. 159 (6) (2015) 1327–1368).

Résumé. Nous considérons un modèle de champs sur réseau Hamiltonien avec deux quantités conservées, l’énergie et le volume,
perturbé par un bruit stochastique conservant les deux quantités précédentes. Il est connu que ce modèle produit une diffusion
anormale de l’énergie de type fractionnaire en raison de la conservation du volume (Nonlinearity 25 (4) (2012) 1099–1133; Arch.
Ration. Mech. Anal. 220 (2) (2016) 505–542). Nous superposons à cette dynamique un second bruit stochastique conservant
l’énergie mais pas le volume. Si l’intensité de ce bruit est d’ordre 1, la diffusion normale de l’énergie est restaurée tandis qu’elle
est sans effet si l’intensité est suffisamment faible. Dans ce papier nous étudions la nature des fluctuations d’énergie pour une
valeur critique de l’intensité. Nous montrons que ces dernières sont décrites par un processus d’Ornstein–Uhlenbeck dirigé par
un processus de Lévy qui interpole entre le mouvement Brownien et le processus de Lévy stable 3/2 totalement asymétrique. Ce
résultat étend et résout un problème laissé ouvert dans (J. Stat. Phys. 159 (6) (2015) 1327–1368).
MSC: 60K35; 82C22; 82C44; 60G22; 74A25

Keywords: Anomalous diffusion; Chain of oscillators; Equilibrium fluctuations; Lévy process

1 This work benefited from the support of the project EDNHS ANR-14-CE25-0011 of the French National Research Agency (ANR) and of the PHC
Pessoa Project 37854WM. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon
2020 research and innovative programme (Grant agreement No 715734).
2 Supported by the French National Research Agency (ANR) Grant ANR-15-CE40-0020-01 (LSD).
3 Supported by the FCT/Portugal project UID/MAT/04459/2013.
4 Supported by the CNPq Grant 401628/2012-4 and by the FAPERJ Grant JCNE E17/2012. Also supported in part by the NWO Gravitation Grant
024.002.003-NETWORKS and by MathAmSud Grant LSBS-2014.
5 Supported in part by Labex CEMPI (ANR-11-LABX-0007-01).
1732 C. Bernardin et al.

1. Introduction

Since the seminal work of Fermi–Pasta–Ulam (FPU) [6], heat conduction in chains of oscillators has attracted a lot of
attention. In one-dimensional chains, superdiffusion of energy has been observed numerically in unpinned FPU chains,
which corresponds to anomalous thermal conductivity. This anomalous thermal conductivity is generally attributed
to a small scattering rate for low modes, which is due to momentum conservation. When the system has a pinning
potential, destroying the conservation of momentum, normal diffusion of energy is expected. In [1,2], it was proposed
to perturb the Hamiltonian dynamics with stochastic interactions that conserve energy and momentum, like random
exchanges of velocity between nearest neighbours. These models have the advantage to be studied rigorously keeping
at the same time the features of deterministic models. For linear interactions, in dimension d ≥ 3, energy follows
normal diffusion, while in dimensions d = 1, 2 energy is superdiffusive [2]. If a pinning potential is added to the
dynamics, normal diffusivity can be proved regardless of the dimension.
In [9] it was proved that in dimension d = 1, energy fluctuations follow the fractional heat equation ∂t u =
−c(−)3/4 u, with c > 0. As mentioned above, in the presence of a pinning potential, energy fluctuations follow
the usual heat equation ∂t u = Du, where D > 0 is the diffusion coefficient. Our goal is to provide a crossover
between these two universality classes, aiming for a better understanding of the origin of the superdiffusivity of the
energy in one-dimensional chains. In particular, we aim to clarify the role of the conservation of momentum.
The stochastic chains considered in [2] have three conserved quantities: the energy, the momentum and the stretch
of the chain. Since we are interested in the role of the conservation of momentum, for simplicity we will consider a
Hamiltonian lattice field model introduced by Bernardin and Stoltz [5], which has only two conserved quantities, see
Section 2.1, but which displays similar superdiffusion features. We call these conserved quantities energy and volume.
In [5] the authors add to the deterministic dynamics an energy and volume conservative Poissonian noise, which is
discrete in nature. Here we consider instead a conservative Brownian noise, for a reason that will be explained ahead.
In [3] a similar result to [9] has been obtained by different techniques for these models.
Let n ∈ N be a scaling parameter, which represents the inverse mesh of the stochastic chain. We add to the dynamics
a second stochastic interaction that conserves only the energy and we scale down the strength of this second interaction
by an , with a > 0. We prove that energy fluctuations follow an evolution equation of the form

∂t u = La u,

where La has the Fourier representation

2 2
a (k) = − √ 4π k
L , k ∈ R.
a + 2iπk
√
In particular, we note that La → −c{(−)3/4 − ∇(−)1/4 } as a → 0 and aLa →  as a → ∞, providing in this
way a crossover between anomalous and normal diffusion of energy in the model. Note as well that the interpolation
between the fractional and normal Laplacians can be understood as an ultraviolet cut-off at modes of order O(a): low
modes behave diffusively, while high modes behave superdiffusively.
In [4] another version of the model of Bernardin and Stoltz of [5] was considered. An almost complete phase
diagram was obtained, although the interpolating part of the diagram described here was missing there. The interested
reader may verify that the methods presented in this article allow to complete the phase diagram in [4] as well as to
prove the results stated there to the model considered here.

Energy fluctuations

Let us describe in a more precise way the main result proved in [3] for the model considered here. Let {ωxn (t)}x∈Z ∈ RZ
be the infinite dimensional
 diffusion process defined
 in Section 2.1. The (formal) conserved quantities of the model
are the energy x∈Z [ωxn (t)]2 and the volume x∈Z ωxn (t). Let {μβ ; β > 0} be the family of Gibbs homogeneous
product measures which are invariant by the dynamics. Under μβ , the random variables {ωx }x∈Z are independent
centered Gaussian variables with variance β −1 . The probability measure on the space of trajectories which is induced
 Interpolation process between standard diffusion and fractional diffusion 1733

by the initial law μβ and the Markov process {ωxn (t)}x∈Z is denoted by Pβ and its corresponding expectation by Eβ .
Define the energy correlation function as

β 2  n 2  2 
Sn (t, x) = Eβ ωx (t) − β −1 ω0n (0) − β −1 .
2
We prove here the following scaling limit for Sn (t, x): for any test functions ϕ, ψ : R → R in the usual Schwartz
space S(R),

1   3/2  x y
lim Sn tn , y − x ϕ ψ = Pt (v − s)ϕ(s)ψ(v) ds dv, (1.1)
n→∞ n n n R2
x,y∈Z

where Pt (·) has the Fourier representation


Pt (k) = e−t La (k) , k ∈ R.

In other words, Sn (tn3/2 , nx) converges, in a weak sense, to the fundamental solution of the evolution equation
∂t u = La u. The case a = 0 is the case considered in [3] for the model with a Poissonian noise. That result is a simple
consequence of a stronger scaling limit, which is the main result of this article. To state it properly let us define the
energy fluctuation field as

1  n 2  x
Etn (ϕ) = √ ωx (t) − β −1 ϕ (1.2)
n n
x∈Z

for test functions ϕ : R → R in S(R). We will prove that this field converges in law to the Gaussian process which is
the stationary solution of the equation

∂t Et = La Et + 2β −2 (−Sa )∇Wt , (1.3)

where Wt is a space-time white noise, La is the adjoint of La in L2 (R) and Sa is its symmetric part given by
Sa = 12 (La + La ). This convergence implies the limit
   
lim Eβ Etn (ϕ)E0n (ψ) = Eβ Et (ϕ)E0 (ψ) ,
n→∞

which is exactly the limit stated in (1.1).

We point out that with respect to [3] and [4], the model considered in this article has a Brownian noise instead of
a Poissonian noise. At the level of the correlation function Sn (t, x), the choice of a Poissonian or a Brownian noise
does not make a sensitive difference. In particular, the method of proof in this article allows to prove (1.1) also for
Poissonian noises chosen in a proper way. However, at the level of the Gaussian fluctuations, key tightness estimates
do not hold for Poissonian noises due to rare events that may introduce huge discontinuities on the observables we are
interested in. We believe that at the level of finite-dimensional distributions the process (1.3) still describes the scaling
limit of energy fluctuations in the model with the Poissonian noise considered in [4]. However, it is not clear whether
the obstructions in order to prove tightness are technical or intrinsic to those kind of noises.

A sketch of the proof

Our proof of the convergence of the energy fluctuation field (1.2) follows the usual scheme of convergence in law of
stochastic processes: we show tightness of the processes Etn in a suitable topology, then we prove that any limit point
of the sequence {Etn }n∈N satisfies a weak formulation of the equation (1.3) and then we rely on a uniqueness result for
the solutions of (1.3).
One technical difficulty comes from what is known in the literature by the replacement lemma: it is not very
difficult to write down a martingale decomposition for Etn that should heuristically converge to the martingale problem
1734 C. Bernardin et al.

associated to Et . But the drift term of this martingale decomposition involves the energy current ωxn (t)ωx+1
n (t). This

current is not a function of the energy and therefore we say that the martingale problem for Et is not closed. To
n

overcome that, we need to replace the current ωxn (t)ωx+1 n (t) by a function of the energy. This is accomplished by

studying the relation between the energy fluctuations and the fluctuations of the correlation field given by

1   x + y |y − x|
ωxn (t)ωyn (t) − δx,y β −1 f , √ , (1.4)
n3/4 2n n
x,y∈Z

on some regular two-dimensional test function f . Above, δx,y is the usual indicator function that equals 1 if x = y
and 0 otherwise. Note that, at least heuristically, the energy current is given by the correlation field evaluated at the
diagonal y = x + 1. The introduction of this field is one of the main conceptual innovations in [3]. This field can be
interpreted as the tensor product of the volume fluctuation field with itself. It turns out that volume fluctuations have
two characteristic time scales. First, the speed of sound associated to the volume is equal to 2, and therefore, volume
fluctuations evolve in the hyperbolic time scale tn following a linear transport equation. If the volume fluctuation field
is modified by a Galilean transformation that drives out the transport dynamics, then it evolves in a diffusive time
scale tn2 , following an equation of the form (1.3) with the operator L replaced by the usual Laplacian operator . In
the definition of the correlation field (1.4), we introduced two different spatial scales. This non-homogeneous spatial
scaling allows to observe both natural time scales at once. In fact, the correlation field (1.4) has a scaling limit in the
hyperbolic time scale tn given by the stationary solution of
 
dZt = −∂x + ∂yy
2
− a Zt dt + dMt ,

where Mt is an infinite-dimensional martingale (see also Section 2.3 for more details). We point out that although
we do not prove neither this result6 nor anything related to it, this limiting equation was used as a guideline for the
computations below. Since the energy fluctuations evolve in the superdiffusive time scale tn3/2 , the correlation field
acts as a fast variable for the evolution of the energy.
The structure of the paper is described as follows. Below we introduce the model with notations, and we state the
main result of this work, namely Theorem 2.5. Section 3 is devoted to the decomposition of the energy field into a
martingale problem, using both the energy field and the correlation field. In Sections 4 and 5, we prove, respectively,
tightness of the processes and characterization of their limit points, for establishing the convergence. Appendix A
collects some results on the Lévy operator L, while in Appendices B and C we gather all technical details used along
the proof.

2. Preliminaries

2.1. The model

In this section we define the BS model (as introduced in [5]) with continuous noises. For that purpose we need to
introduce two real parameters: λ > 0 and γn > 0, the latter depending on a scale parameter n ∈ N. Let us consider
a system of diffusions evolving on the state space := RZ , in the time scale n3/2 , and generated by the operator
n3/2 Ln , where Ln is decomposed as the sum Ln = A + λS1 + γn S2 , where
 ∂
A= (ωx+1 − ωx−1 ) ,
∂ωx
x∈Z
 
S1 = (Xx ◦ Xx ), S2 = (Yx ◦ Yx ),
x∈Z x∈Z

6 This result can be guessed by using the computations in Appendix B.2 but its rigorous proof is not trivial and would require a paper by itself. The
interested reader is invited to consult [8] for a similar result in the context of the symmetric simple exclusion process.
 Interpolation process between standard diffusion and fractional diffusion 1735

and the family of operators {Xx , Yx }x∈Z is given by

∂ ∂ ∂
Xx = (ωx+1 − ωx ) + (ωx − ωx−1 ) + (ωx−1 − ωx+1 ) ,
∂ωx−1 ∂ωx+1 ∂ωx
∂ ∂
Yx = ωx+1 − ωx .
∂ωx ∂ωx+1
The generator A is the generator corresponding to the infinite system of coupled ODE’s dωx (t) = (ωx+1 (t) −
ωx−1 (t)) dt , x ∈ Z. A simple change of variables [5] shows that it is equivalent to the dynamics generated by an
infinite system of coupled harmonic oscillators. With this change of variables, ωx represents either the momentum of
a particle or the interdistance between two nearest neighbor particles. The diffusion operator Xx is nothing but the
generator of a Brownian motion on the circle {(ωx−1 , ωx , ωx+1 ) ∈ R3 ; ωx−1
2 + ωx2 + ωx+1
2 = 1, ωx−1 + ωx + ωx+1 = 0}
while Yx is the generator of a Brownian motion on the circle {(ωx , ωx+1 ) ∈ R ; ωx + ωx+1 = 1}.
2 2 2
 
We call energy the formal quantity x [ωx ]2 and volume the formal quantity x ωx . The Liouville operator A as
well as the noise S1 conserves both energy and volume, while the operator S2 conserves only energy. We assume that
the strength of the second noise scales as
a
γn = (2.1)
n
for some a > 0. We emphasize that one could easily treat the general case γn = nab , b ≥ 0, as in [4], using the same
methods as in this paper, but we chose here to focus on the most interesting case b = 1 where the interpolation
happens.
The Markov process generated by the accelerated operator n3/2 Ln is denoted by ωn (t) = {ωxn (t)}x∈Z . This diffusion
has a family {μβ ; β > 0} of invariant measures given by the Gibbs homogeneous product measures

 β βω2
μβ (dω) = exp − x dωx .
2π 2
x∈Z

Here β represents the inverse temperature, and we denote by ϕ β the average of ϕ : → R with respect to μβ .
The law of the process {ωxn (t); t ≥ 0}x∈Z starting from the invariant measure μβ is denoted by Pβ , and the expec-
tation with respect to Pβ is denoted by Eβ . Note that under μβ , the averaged energy per site equals ωx2 β = β −1 , and
the averaged volume per site equals ωx β = 0.

2.2. Fluctuation fields

From now on, the Markov process {ωxn (t); t ≥ 0}x∈Z is considered starting from μβ . The energy fluctuation field is
defined as the distribution-valued process Etn given by
1  n 2  x
Etn (ϕ) = √ ωx (t) − β −1 ϕ (2.2)
n n
x∈Z

for any ϕ : R → R in the usual Schwartz space S(R) of test functions. For fixed t and ϕ, the random variables Etn (ϕ)
satisfy a central limit theorem: they converge to a centered normal random variable of variance 2β −2 ϕ 22 , where
· 2 denotes the usual norm of the Hilbert space L2 (R).
Our main goal is to obtain a convergence result for the S (R)-valued process {Etn ; t ≥ 0}. It turns out that the
analysis of the correlation field
1   x + y |y − x|
ωxn (t)ωyn (t) − δx,y β −1 f , √ (2.3)
n3/4 2n n
x,y∈Z

will play a fundamental role on the derivation of the scaling limit of Etn . Recall that δx,y is the indicator function that
equals 1 if x = y and 0 otherwise, and f : R × R+ → R is a smooth function. The non-isotropic scaling is crucial in
order to see the scaling limit of Etn .
1736 C. Bernardin et al.

2.3. Generalized Ornstein–Uhlenbeck equation associated to a Lévy process

√
First of all, let us introduce some notations: for any complex number z ∈ C, we denote by z its principal
√ square
√ root,
which has positive real part: if z = re with r ≥ 0 and θ ∈ (−π, π], then its principal square root is z = reiθ/2 .
iθ

Let also ψ : R → C be the Fourier transform of a function ψ ∈ L1 (R), which is defined by

ψ(k) := e−2iπuk ψ(u) du, k ∈ R. (2.4)

R

For any ϕ ∈ S(R), we define Lϕ via the action of the operator L on Schwartz spaces: precisely, the operator L acts
on the Fourier transform of ϕ as:

 1 (2iπk)2
Lϕ(k) = √ √ ϕ(k), k ∈ R. (2.5)
2 3λ a + iπk
This operator has nice properties, stated in the next proposition:

Proposition 2.1. The operator L is the generator of a Lévy process. It leaves the space S(R) invariant, and its
Lévy–Khintchine representation is given by
 
(Lϕ)(u) = ϕ(u − y) − ϕ(u) + yϕ (u) a (dy), (2.6)
R

where a is the measure on R defined by

 
4a 5/2 −2ay 3 1 1
a (dy) = − √ e + + 1(0,+∞) (y). (2.7)
6λπ 16(ay)5/2 2(ay)3/2 (ay)1/2

Proof. For the sake of readability, we postpone this proof to Appendix A.1. 

Let us give here an alternative definition of Lϕ, which will turn out to be more tractable in the forthcoming
computations. We claim that Lϕ can equivalently be defined as follows: for any u ∈ R,

(Lϕ)(u) = −2∂u f (u, 0), (2.8)

where f : R × R+ → R is the function such that its Fourier transform with respect to its first variable:

Fk (v) := e−2iπuk f (u, v) du, k ∈ R, v ≥ 0,

R

is given by

1 (2iπk)ϕ(k) a + iπk
Fk (v) = − √ √ exp − v , v ≥ 0. (2.9)
4 3λ a + iπk 3λ
The function f defined in this way satisfies the integrability conditions

f 2 (u, v) du dv < ∞ and ∂v f 2 (u, v) du dv < ∞. (2.10)

R×R+ R×R+

Moreover the function f is solution of the Laplace equation


2 f − ∂ f − 2af )(u, v) = 0, for u ∈ R, v > 0,
(6λ∂vv u
(2.11)
12λ∂v f (u, 0) = ϕ (u), for u ∈ R.

This last claim is proved in Appendix A.2.

 Interpolation process between standard diffusion and fractional diffusion 1737

Let L be the adjoint of L in L2 (R) and S := 12 (L + L ) be its symmetric part. Let us fix a time horizon T > 0.
We are going to explain the meaning of a stationary solution of the infinite dimensional Ornstein–Uhlenbeck equation
driven by L, written as follows:

∂t Et = L Et + 2β −2 (−S)Wt , (2.12)

where {Wt ; t ∈ [0, T ]} is a S (R)-valued space-time white noise.

Definition 2.2. We say that an S (R)-valued process {Et ; t ∈ [0, T ]} is β-stationary if, for any t ∈ [0, T ], the S (R)-
valued random variable Et is a white noise (in space) of variance 2β −2 , namely: for any ϕ ∈ S(R), the real-valued
random variable Et (ϕ) has a normal distribution of mean zero and variance 2β −2 ϕ 22 .

Definition 2.3. We say that the S (R)-valued process {Et ; t ∈ [0, T ]} is a stationary solution of (2.12) if:
(1) {Et ; t ∈ [0, T ]} is β-stationary;
(2) for any time differentiable function ϕ : [0, T ] × R → R, such that for each t ∈ [0, T ] both ϕt and ∂t ϕt belong to
S(R), the process
t  
Et (ϕt ) − E0 (ϕ0 ) − Es (∂s + L)ϕs ds
0

is a continuous martingale of quadratic variation

t
2β −2 ϕs (u)(−Sϕs )(u) du ds.
0 R

Thanks to the fact that L is the generator of a Lévy process, the same argument used in [7, Appendix B] can be
worked out here to prove the uniqueness of such solutions:

Proposition 2.4 ([7]). Two stationary solutions of (2.12) have the same distribution.

Let us denote by C([0, T ], S (R)) the space of continuous functions from [0, T ] to S (R). Roughly speaking, the
main result of this work states that the energy fluctuations described by Etn (defined in (2.2)) satisfy an approximate
martingale problem, which, in the limit n → ∞ becomes the martingale characterization of the limiting process
described in Definition 2.3. It can be precisely formulated as follows:

Theorem 2.5. The sequence of processes {Etn ; t ∈ [0, T ]}n∈N converges in law, as n → ∞, with respect to the weak
topology of C([0, T ], S (R)), to the stationary solution of the infinite-dimensional Ornstein–Uhlenbeck process given
by (2.12).

The proof of Theorem 2.5 follows from two steps:

(1) We prove in Section 4 that the sequence {Etn ; t ∈ [0, T ]}n∈N is tight.
(2) We characterize all its limit points in Section 5 by means of a martingale problem.
First, we need to do an investigation of the fluctuation field Etn , and of the discrete martingale problem that it satisfies.

3. Martingale decompositions

In this section we fix ϕ ∈ S(R). Let f : R × R+ → R be as in Section 2.3. Let us introduce the time dependent
bidimensional field, defined as Ctn (f ) := C(f )(ωn (t)) with

1   n
C(f )(ω) := ωx ωy − δx,y β −1 fx,y ,
n
x,y∈Z
1738 C. Bernardin et al.

where, for any x, y ∈ Z,

x + y |y − x|
n
fx,y := f , √ . (3.1)
2n n
Note that, for any sufficiently regular square-integrable function f , since under μβ the variables {ωx }x∈N are inde-
pendent and centered Gaussian, we have, by an application of the Cauchy–Schwarz inequality, that
 2  C(β)   n 2
Eβ Ctn (f ) ≤ 2 fx,y −−−→ 0. (3.2)
n n→∞
x,y∈Z

3.1. Martingale decomposition for the energy

We first need to define the two discrete operators ∇n and n , acting on ϕ ∈ S(R) as follows: for any x ∈ Z let
   
x x +1 x x x +1 x −1 x
∇n ϕ := n ϕ −ϕ , n ϕ := n ϕ
2
+ϕ − 2ϕ .
n n n n n n n
From Dynkin’s formula, see for example [10], for any ϕ ∈ S(R), the process
t  
MEt,n (ϕ) := Etn (ϕ) − E0n (ϕ) − n3/2 Ln Esn (ϕ) ds (3.3)
0

is a martingale. A straightforward computation shows that

   x

n3/2 Ln Esn (ϕ) = −2 ωxn (s)ωx+1
n
(s)∇n ϕ + Rns (ϕ), (3.4)
n
x∈Z

where
1  n 2  x
Rns (ϕ) = (2γn + 4λ) ωx (s) − β −1 n ϕ
n n
x∈Z
  
1  n 2  x+2 x −2 x
+ (2λ) ωx (s) − β −1 n2 ϕ +ϕ − 2ϕ
n n n n
x∈Z

1 x +1
+ (2λ) ωxn (s)ωx+2
n
(s)n ϕ
n n
x∈Z

1 x x+1
− (4λ) ωxn (s)ωx+1
n
(s) n ϕ + n ϕ .
n n n
x∈Z
t
The second term in the right hand side of (3.4), when integrated in time between 0 and t – namely 0 Rns (ϕ) ds –
is negligible in L2 (Pβ ) as a consequence of the Cauchy–Schwarz inequality (recall that ωx ωx+2 ωy ωy+2 β = 0 for
x = y). Analogously, the first term in the right hand side (3.4), integrated in time, can be replaced thanks to Cauchy–
Schwarz inequality, up to a vanishing error in L2 (Pβ ), by
t  x t  x
−2 ωxn (s)ωx+1
n
(s)ϕ ds = − ωxn (s)ωx+1
n
(s)24λ∂v f , 0 ds
0 x∈Z n 0 x∈Z n

the last equality being a consequence of (2.11). Therefore, we have

t  x t
Etn (ϕ) − E0n (ϕ) = − ωxn (s)ωx+1
n
(s)24λ∂v f , 0 ds + MEt,n (ϕ) + εn (s) ds, (3.5)
0 x∈Z n 0
 Interpolation process between standard diffusion and fractional diffusion 1739

where MEt,n (ϕ) is a martingale, whose quadratic variation will be computed in Section 3.4. Moreover, εn (t) satisfies
two estimates: first, for any t > 0 fixed,
 t 2
lim Eβ εn (s) ds =0 (3.6)
n→∞ 0

and second,
 2 
lim sup Eβ εn (t) < +∞. (3.7)
n→∞ t∈[0,T ]

3.2. Martingale decomposition for the correlation field

Now let us turn to the bidimensional field Ctn (f ). From Dynkin’s formula, for any f : R2 → R, the process
t  
MCt,n (f ) := Ctn (f ) − C0n (f ) − n3/2 Ln Csn (f ) ds (3.8)
0

is a martingale. The computations of Appendix B allow us to write

 
  2  2  x 1 1
n3/2 Ln C(f ) = − √ ωx − β −1 ∂u f ,0 + O √ +O √ (3.9)
n n n n
x∈Z
  
x 1
+ ωx ωx+1 24λ∂v f ,0 + O (3.10)
n n
x∈Z
    
4  x x 1
+√ ωx ωx+1 af , 0 − ωx+1 ωx−1 λ∂vv f2
,0 + O √ , (3.11)
n n n n
x∈Z

where O(εn ) denotes a sequence of functions in Z bounded by cεn for some finite constant c that does not depend
on n. Note that (3.10) contains the same term that we made appear above in (3.5).
Observe that w.r.t. the computations of Appendix B an extra term has been introduced (precisely in the first display
(3.9)): this term is

2β −1  x β −1  x
√ ∂u f ,0 = − √ (Lϕ) ,
n n n n
x∈Z x∈Z

where the last equality follows from (2.8). We claim that this new quantity is at most of order n−1/2 . To
 justify this,
recall that by Proposition 2.1 the function h = Lϕ is in the Schwartz space and that its integral equals R h(u) du =
ĥ(0) = 0. Moreover we have
   
1  x   
x+1

n x
n h − h(u) du =  h − h(u) du
n R x n
x∈Z x∈Z n

 
 
x+1
n x+1
=  h (u) − u du
x n
x∈Z n

1    
x+1
h (u) du = 1 h (u) du = O 1 .
n
≤
n x n R n
x∈Z n


x∈Z h( n ) = O(1)
x
Therefore and the claim is proved.
1740 C. Bernardin et al.

Let us go one step further, and replace the local function ωx−1 ωx+1 that appears in (3.11) with the local function
ωx ωx+1 . This is the purpose of Lemma 3.1 below: from that result we can rewrite the time integral as
t   2 t  2  x
n3/2 Ln Csn (f ) ds = − √ ωxn (s) − β −1 ∂u f , 0 ds
0 n 0 x∈Z n
t  x
+ 24λ ωxn (s)ωx+1
n
(s)∂v f , 0 ds
0 x∈Z n

4 t    x t
+√ ωxn (s)ωx+1
n
(s) af − λ∂vv
2
f , 0 ds + εn (s) ds,
n 0 x∈Z n 0

where εn (t) satisfies the same estimates as εn (t), namely (3.6) and (3.7).

Lemma 3.1. Let {ψn (x)}x∈Z be a real-valued sequence such that

1  2
ψn (x) < +∞. (3.12)
n
x∈Z

Then,
 2
t 1   
lim Eβ √ ψn (x) ωxn − ωx−1
n n
(s)ωx+1 (s) ds = 0. (3.13)
n→∞ 0 n
x∈Z

Proof. To prove the lemma we use a general inequality for the variance of additive functionals of Markov processes:
we have
 t 1  2
 n  t
Eβ √ ψn (x) ωx − ωx−1 (s)ωx+1 (s) ds
n n
≤ C(β) 3/2  2[tn3/2 ]−1 ,−1 , (3.14)
0 n n
x∈Z

where
1 
(ω) := √ ψn (x)(ωx − ωx−1 )ωx+1 ,
n
x∈Z

and, for any z > 0,

 
 2z,−1 := , (z − λS1 − γn S2 )−1  β
     
= sup 2 g β − z g 2 β − g(−λS1 − γn S2 )g β , (3.15)
g

where the supremum is restricted over functions g in the domain of S. In order to prove (3.14), we first apply
Lemma 3.9 of [12], with the operator −t −1 Id + n3/2 L and we get:
 t  2
   −1 
Eβ n
 ω (s) ds ≤ C(β)t , t −1 − n3/2 L  β
0
C(β)t   3/2 −1 −1 
= 3/2
, tn −L  β
n
C(β)t   3/2 −1 −1 
≤ 3/2 , tn −S β
n
C(β)t
= 3/2  2[tn3/2 ]−1 ,−1 .
n
 Interpolation process between standard diffusion and fractional diffusion 1741

We can forget about the positive operator (z − λS1 ), and bound the norm (3.15) as follows:
 
 2z,−1 ≤ , (−γn S2 )−1  β .

One can easily check that

1 1
S2 ωx−1 ωx+1 − ωx ωx+1 = (ωx − ωx−1 )ωx+1 ,
4 6

which implies that (−γn S2 )−1  is explicit and given by

 
−1 1  1 1
(−γn S2 ) (ω) = √ ψn (x) ωx ωx+1 − ωx−1 ωx+1 ,
γn n 6 4
x∈Z

so that, finally,
C(β)  2
 2
z,−1 ≤ ψn (x) .
γn n
x∈Z

Recall γn = an , and then after replacing the previous bound in (3.14) we get
 2
t 1    t n 1  2 1
Eβ √ ψn (x) ωxn − ωx−1
n n
(s)ωx+1 (s) ds ≤ C(β) ψn (x) = O √ ,
0 n n3/2 an n
x∈Z x∈Z

which vanishes as n → ∞. 

3.3. Sum of the two decompositions

Combining the two decompositions (3.3) and (3.8) we get

t 2  n 2  x
Etn (ϕ) − E0n (ϕ) = − √ ωx (s) − β −1 ∂u f , 0 ds (3.16)
0 n n
x∈Z
t 4  n   x
+ √ n
ωx (s)ωx+1 (s) af − λ∂vv
2
f , 0 ds (3.17)
0 n n
x∈Z
  t
+ MCt,n (f ) − Ctn (f ) − C0n (f ) + MEt,n (ϕ) + εn (s) ds, (3.18)
0

where εn (t) = εn (t) + ε (t). Note that

t 2  n 2  x t  
√ ωx (s) − β −1 ∂u f , 0 ds = 2 Esn ∂u f (·, 0) ds.
0 n n 0
x∈Z

Since the terms in (3.17) and (3.18) will be proved to vanish, as n → ∞, this will permit to close the martingale
equation in terms of the energy field. From (3.2), the term (Ctn (f ) − C0n (f )) vanishes
√ in L (Pβ ). Finally, the term
2

(3.17), which is in the same form as (3.4) (but of smaller order, since it is divided by n), is treated by repeating the
same procedure: let g : R × R+ → R be solution of the equation

2 g − ∂ g − 2ag)(u, v) = 0,
(6λ∂vv for u ∈ R, v > 0,
u
(3.19)
24λ∂v g(u, 0) = 4(af − λ∂vv f )(u, 0), for u ∈ R,
2

where f is given in Section 2.3. The function g is defined by its Fourier transform w.r.t. the first variable as it has been
done to define f . Then, using the same computations as before, but with ϕ (u) replaced by 2(af − λ∂vv 2 f )(u, 0), we
1742 C. Bernardin et al.

get that
t 4  n   x
√ n
ωx (s)ωx+1 (s) af − λ∂vv
2
f , 0 ds
0 n n
x∈Z
t 24λ  n x
= √ n
ωx (s)ωx+1 (s)∂v g , 0 ds
0 n n
x∈Z
1  
= √ Ctn (g) − C0n (g) − MCt,n (g) (3.20)
n
t 2  2  x
+ ωxn (s) − β −1 ∂u g , 0 ds (3.21)
0 n n
x∈Z
t 4   x t
− ωxn (s)ωx+1
n
(s) ag − λ∂vv
2
g , 0 ds + εn (s) ds. (3.22)
0 n n 0
x∈Z

Note that in (3.21) we introduced the extra term

2β −1  x
∂u g ,0
n n
x∈Z

as we did above
 for f . The same argument works here: one can prove that this additional quantity is of order at most
O( n1 ) since R ∂u g(u, 0) du = 0.
From the Cauchy–Schwarz inequality, both terms (3.21) and (3.22) vanish in L2 (Pβ ), as n → ∞, and give a
contribution εn (t) which also satisfies the same conditions as (3.6) and (3.7) (note that this is the same argument
used in Section 3.1). Besides, from (3.2), Ctn (g) − C0n (g) also vanishes in L2 (Pβ ), as n → ∞. Summarizing, the
approximate discrete martingale equation can be written as
t  
Etn (ϕ) − E0n (ϕ) = −2 Esn ∂u f (·, 0) ds
0
1 t
+ MEt,n (ϕ) + MCt,n (f ) − √ MCt,n (g) + ε n (s) ds, (3.23)
n 0

where ε n (t) satisfies (3.6) and (3.7). In the following paragraph, by computing quadratic variations we prove that the
only martingale term that will give a non-zero contribution to the limit is the one coming from the correlation field,
namely MCt,n (f ).

3.4. Convergence of quadratic variations

We start by showing that the quadratic variations of the martingales ME·,n (ϕ), MC·,n (f ) and MC·,n (g) converge in
mean, as n → ∞.

Lemma 3.2. For any ϕ ∈ S(R) and t > 0,

 
lim Eβ ME·,n (ϕ) t = 0.
n→∞

Proof. We have
  n3/2 t     
ME·,n (ϕ) t = Ln F 2 ωn (s) − 2F (Ln F ) ωn (s) ds
n 0
√ t   2  2  n 
= n 2λ Xz (F ) + 2γn Yz (F ) ω (s) ds, (3.24)
0 z∈Z
 Interpolation process between standard diffusion and fractional diffusion 1743


where F (ω) := 2 n
x∈Z ωx ϕx and ϕxn := ϕ( xn ). Note that (3.24) can also be written as

√ t  
n λQ1 (F, F ) + γn Q2 (F, F ) ωn (s) ,
0

where the bilinear operators Qi (i = 1, 2) are given by

Qi (f, g) = Si (f g) − f Si g − gSi f.

In some contexts, the bilinear form Qi is called the carré du champ. A long but simple computation (using Ap-
pendix B.1) gives that
 E  √ t   n   
M·,n (ϕ) t = n 4λ ωxn (s)ωx+1
n
(s) ϕx+1 − ϕxn + ωxn (s)ωx−1
n
(s) ϕxn − ϕx−1
n
0 x∈Z
 2
+ ωx−1
n n
(s)ωx+1 (s) n
ϕx−1 − ϕx+1
n

 
 2
+ 4γn ωxn (s)ωx+1
n
(s) n
ϕx+1 − ϕxn ds. (3.25)
x∈Z

Therefore, taking the expectation, since ϕ ∈ S(R) we get

  1  z 2
1
Eβ ME·,n (ϕ) t ≤ tC(β)(λ + γn ) 3/2 ∇n ϕ =O √ ,
n n n
z∈Z

which proves the lemma. 

Lemma 3.3. Let f : R × R+ → R be as in Section 2.3. Then, for t > 0

   
lim Eβ MC·,n (f ) t = 2tβ −2 8af 2 + 24λ(∂v f )2 (u, v) du dv.
n→∞ R×R+

Moreover the term on the right hand side of last expression equals to

2tβ −2 ϕ(u)(−Sϕ)(u) du.

R

Proof. As before, we have

  n3/2 t   2  2  n 
MC·,n (f ) t = 2 2λ Xz (F ) + 2γn Yz (F ) ω (s) ds, (3.26)
n 0 z∈Z

where F (ω) := x,y∈Z ωx ωy fx,y n with f n defined in (3.1). Since the computations are a bit longer, we decompose
x,y
them as follows: first, note that
 n 
(Xz )(F ) = 2ωz+1 ωz−1 −fz+1,z+1 + fz−1,z−1
n
− fz−1,z
n
+ fz,z+1
n
(3.27)
 n 
+ 2ωz ωz−1 fz,z − fz−1,z−1
n
− fz,z+1
n
+ fz−1,z+1
n
(3.28)
 n 
+ 2ωz ωz+1 −fz,z + fz+1,z+1
n
− fz−1,z+1
n
+ fz−1,z
n
(3.29)
 2 n     
+ 2 ωz fz,z+1 − fz−1,z
n
+ ωz+1
2 n
fz−1,z+1 − fz,z+1
n
+ ωz−1
2 n
fz−1,z − fz−1,z+1
n
(3.30)
   n   
+2 ωy ωz fz+1,y − fz−1,y n
+ ωz+1 fz−1,yn
− fz,y
n
(3.31)
y ∈{z−1,z,z+1}
/
 n 
+ ωz−1 fz,y − fz+1,y
n
. (3.32)
1744 C. Bernardin et al.

In the last expression we consider separately two terms: the first expression involving only the coordinates ωz−1 , ωz
and ωz+1 (from (3.27) to (3.30)) that we denote by (I), and the last remaining sum over y ∈ / {z − 1, z, z + 1} (namely
(3.31)–(3.32)) that we denote by (II). In order to compute Eβ [ MC·,n (f ) t ], we first estimate the Pβ -average of

n3/2 t   2  
2λ Xz (F ) ωn (s) ds (3.33)
n2 0 z∈Z

to which (I) contributes as

 2
1  z 2
1  z 1
λtC(β) 3/2 ∂v f ,0 + ∂u f ,0 +O ,
n n n5/2 n n3/2
z∈Z z∈Z

therefore it vanishes as n → ∞. The second term (II) is the only contributor to the limit. By using a Taylor expansion
(see also (B.5) below), one has

2 z + y |y − z| 1
n
fz+1,y − fz−1,y
n
= − √ ∂v f , √ +O , (3.34)
n 2n n n
1 z + y |y − z| 1
n
fz−1,y − fz,y
n
= √ ∂v f , √ +O . (3.35)
n 2n n n

Therefore, in the estimate of (3.33) the second term (II) will contribute as

    24 z + y |y − z| 2
1
2λt ω02 ω12 β 3/2
∂v f , √ +O ,
n 2n n n
z∈Z y ∈{z−1,z,z+1}
/

which converges, as n → ∞, to

48λtβ −2 (∂v f )2 (u, v) du dv. (3.36)

R×R+

Let us now take care of the second stochastic noise that appears with Yz . We have:
 n    n
(Yz )(F ) = 2ωz ωz+1 fz,z − fz+1,z+1
n
− 2 ωz2 − ωz+1
2
fz,z+1
  
−2 n
ωy ωz fz+1,y − ωz+1 fz,y
n
.
y ∈{z,z+1}
/

Recall that γn = an . One can check that

  
n3/2 t  2  
Eβ 2γn Yz (F ) ωn (s) ds (3.37)
n2 0 z∈Z

can be rewritten by the translation invariance of μβ as

  1    2 1
16at ω02 ω12 β 3/2 n
fy,z +O ,
n n
z∈Z y ∈{z−1,z,z+1}
/

and it converges, as n → ∞, to

16atβ −2 f 2 (u, v) du dv. (3.38)

R×R+
 Interpolation process between standard diffusion and fractional diffusion 1745

As a consequence of (3.36) and (3.38) , we have

   
Eβ MC·,n (f ) t −−−→ 2tβ −2 8af 2 + 24λ(∂v f )2 (u, v) du dv.
n→∞ R×R+

An explicit resolution of (2.11) via Fourier transforms given in Appendix A.2 easily gives

 
8af 2 + 24λ(∂v f )2 (u, v) du dv = ϕ(u)(−Sϕ)(u) du
R×R+ R

which is enough to conclude. 

Remark 3.4. We note that by, similar computations to the ones of the previous lemma, we can prove that

   
Eβ MC·,n (g) t −−−→ 2tβ −2 ag 2 + 3λ(∂v g)2 (u, v) du dv,
n→∞ R×R+

where g has been defined before as the solution to (3.19).

Lemma 3.5 (L2 (Pβ ) convergence of quadratic variations). For ϕ ∈ S(R) and f : R × R+ → R as in Section 2.3,
we have
    2 
lim Eβ ME·,n (ϕ) t − Eβ ME·,n (ϕ) t = 0, (3.39)
n→∞
    2 
lim Eβ MC·,n (f ) t − Eβ MC·,n (f ) t = 0, (3.40)
n→∞

Proof. The proof of this lemma is postponed to Appendix C. 

3.5. Conclusion

From Lemma 3.2 and the remark above, we know that, for each fixed t > 0, the martingales MEt,n (ϕ) and √1 MC
n t,n (g)
vanish, as n → ∞, in L2 (Pβ ). Therefore the non vanishing terms remaining in the right hand side of the decomposition
(3.23) are
t  
−2 Esn ∂u f (·, 0) ds + MCt,n (f ). (3.41)
0

4. Tightness

The tightness of the sequence {Etn ; t ∈ [0, T ]}n∈N in the space C([0, T ], S (R)) is proved by standard arguments.
First, Mitoma’s criterion [11] reduces the proof of tightness of distribution-valued processes to the proof of tight-
ness for real-valued processes. Indeed, it is enough to show tightness of the sequence {Etn (ϕ); t ∈ [0, T ]} for any
ϕ ∈ S(R). According to (3.23), we are reduced to prove that the processes
 
 n  t  
E0 (ϕ) n∈N , Esn ∂u f (·, 0) ds; t ∈ [0, T ]
0 n∈N

are tight, where f : R × R+ → R is solution to (2.11). We will also prove that the martingales
 
 E   C  1 C
Mt,n (ϕ); t ∈ [0, T ] n∈N , Mt,n (f ); t ∈ [0, T ] n∈N , √ Mt,n (g); t ∈ [0, T ] (4.1)
n n∈N
1746 C. Bernardin et al.

are convergent and, in particular, they are tight, and finally that the process
 t 
εn (s) ds; t ∈ [0, T ]
0 n∈N

is tight.

4.1. Tightness for {E0n (ϕ)}n∈N

As mentioned at the beginning of Section 2.2, {E0n (ϕ)}n∈N converges in distribution, as n → ∞, towards a centered
normal random variable of variance 2β −2 ϕ 2L2 (R) , and in particular the sequence is tight.

t t
4.2. Tightness for { 0 Esn (∂u f (·, 0)) ds; t ∈ [0, T ]}n∈N and for { 0 ε n (s) ds; t ∈ [0, T ]}n∈N

For these two integral terms we use the following tightness criterion:
t
Proposition 4.1 ([7, Proposition 3.4]). A sequence of processes of the form { 0 Xn (s) ds; t ∈ [0, T ]}n∈N is tight with
respect to the uniform topology in C([0, T ], R) if
 
lim sup E Xn2 (t) < +∞.
n→∞ t∈[0,T ]

One can easily check from the Cauchy–Schwarz inequality that

  2  C(β)  x 2  2
Eβ Esn ∂u f (·, 0) ≤ ∂u f ,0 −−−→ C(β)t 2 ∂u f (u, 0) du,
n n n→∞ R
x∈Z

and recall that ε n (t) satisfies (3.7). Therefore, the criterion of Proposition 4.1 holds for both processes, and tightness
follows.

4.3. Convergence of martingales

By definition, and more precisely (3.3) and (3.8), for any n ∈ N and ϕ ∈ S(R), f as in Section 2.3 and g solution of
(3.19), the martingales
 
 E   C  1 C
Mt,n (ϕ); t ∈ [0, T ] , Mt,n (f ); t ∈ [0, T ] , √ Mt,n (g); t ∈ [0, T ]
n

are continuous in time. In order to prove that the sequences of martingales written in (4.1) are convergent as n → ∞,
we use the following criterion, adapted from [13, Theorem 2.1] to the case of continuous processes:

Proposition 4.2. A sequence {Mnt ; t ∈ [0, T ]}n∈N of square-integrable martingales converges in distribution with
respect to the uniform topology of C([0, T ]; R), as n → ∞, to a Brownian motion of variance σ 2 if for any t ∈ [0, T ],
the quadratic variation Mn t converges in distribution, as n → ∞, towards σ 2 t .

From Section 3.4 we conclude that the martingales

 
 E  1 C
Mt,n (ϕ); t ∈ [0, T ] n∈N , √ Mt,n (g); t ∈ [0, T ]
n n∈N

vanish in distribution, as n → ∞, and from Proposition 4.2 we conclude that the martingales
 C 
Mt,n (f ); t ∈ [0, T ] n∈N
 Interpolation process between standard diffusion and fractional diffusion 1747

converge in distribution as n → ∞ to a Brownian motion of variance

2tβ −2 ϕ(u)(−Sϕ)(u) du.

R

From this, we conclude that all the martingales are tight.

5. Characterization of limit points

From the previous section, we know that the sequence {Etn ; t ∈ [0, T ]}n∈N is tight. Let {Et ; t ∈ [0, T ]} be one limit
point in C([0, T ], S (R)). For simplicity, we still index the convergent subsequence by n.
We already know that {E0n (ϕ)}n∈N converges in distribution, as n → ∞, towards a centered Gaussian random
variable of variance 2β −2 ϕ 2L2 (R) .
For the integral term it is easy to see that the convergence in law
t t
Esn (Lϕ) ds −−−→ Es (Lϕ) ds
0 n→∞ 0

holds. The convergence for the martingale term has already been proved in Section 4.3. Putting all these elements
together, we conclude that, for any ϕ ∈ S(R), we have
t
Et (ϕ) = E0 (ϕ) + Es (Lϕ) ds + Mt (ϕ),
0

where Mt (ϕ) is a Brownian motion of quadratic variation

2tβ −2 ϕ(u)(−Sϕ)(u) du.

R

By Proposition 2.4, the distribution of {Et ; t ∈ [0, T ]} is uniquely determined. We conclude that the sequence {Etn ; t ∈
[0, T ]}n∈N has a unique limit point, and since it is tight, it converges to this limit point. This ends the proof of
Theorem 2.5.

Appendix A: Fourier transforms and Lévy–Khintchine decomposition

A.1. Lévy–Khintchine decomposition

Let us first prove that L lets S(R) invariant. Since the Fourier transform is a bijection from S(R) into itself, it is
 ∈ S(R). Since a > 0, the function
sufficient to prove that if ϕ ∈ S(R) then Lϕ

(2iπk)2
θ :k∈R→ √ ∈C
a + iπk
is a smooth function and we have that for any p ≥ 0, there exist constants Cp , αp > 0 such that
 (p)   
∀k ∈ R, θ (k) ≤ Cp 1 + |k| αp . (A.1)

 ∈ S(R).
Therefore, we have that Lϕ
Let X be a random variable distributed according to the Gamma distribution ( 12 , 1). More precisely, its density
fX with respect to the Lebesgue measure is given by

e−x
fX (x) := 1(0,+∞) (x) √ , x ∈ R,
πx
1748 C. Bernardin et al.

and its characteristic function is

  1 +∞ e−x
X (t) = E eitX = √ = lim √ eitx dx, t ∈ R. (A.2)
1 − it ε→0 ε πx

Lemma A.1. For any t ∈ R,

t2 +∞  
H (t) := √ = eitx − 1 − itx (dx),
1 − it 0

where (dx) := fX (x) dx.

Proof. Note that, for ε > 0, an integration by parts gives

+∞ 1 − eitε 1 +∞  
fX (x)eitx dx = fX (ε) − eitx − 1 fX (x) dx
ε it it ε
1 +∞  
−−→ − eitx − 1 fX (x) dx,
ε→0 it 0

the last convergence holds since fX (ε)  √1

πε
as ε → 0. Therefore, we have the following identity

+∞  
−itX (t) = eitx − 1 fX (x) dx. (A.3)
0

A second integration by parts can now be done in the same way, and one can check that
+∞   1 − eitε 1 +∞  
eitx − 1 fX (x) dx = + ε fX (ε) − eitx − 1 − itx fX (x) dx.
ε it it ε

e−ε
Since fX (ε) = − √πε (1 + 1
2ε ), by taking the limit as ε → 0 in the previous identity, using (A.3) and recalling (A.2),
Lemma A.1 follows. 

The function we are interested in is the one that appears in (2.5), namely:

1 (2iπt)2 2a 3/2 πt
a (t) := √ √ =− √ H − , a > 0,
2 3λ a + iπt 3λ a
where H is given in Lemma A.1. From that lemma we get

2a 3/2 +∞ πtx iπtx

a (t) = − √ e−i a −1+ fX (x) dx
3λ 0 a
4a 5/2 +∞  
=− √ e−2iπty − 1 + 2iπty fX (2ay) dy.
3λ 0

A simple computation gives

e−x 1 3
fX (x) = √ 1+ + 2 .
πx x 4x
Therefore
+∞  
a (t) = e−2iπty − 1 + 2iπty a (dy),
0

where a has been defined in (2.7). Proposition 2.1 easily follows.

 Interpolation process between standard diffusion and fractional diffusion 1749

A.2. Aternative definition: Fourier transformation and resolution

Recall that f : R × R+ → R is such that its Fourier transform with respect to the first variable is given by (2.9). For
any fixed k ∈ R, the function Fk (·) is solution to

6λFk (v) − (2a + 2iπk)Fk (v) = 0, v ≥ 0,
(A.4)
12λFk (0) = 2iπk ϕ(k).

If we assume (2.8), one can easily check that

 1 (2iπk)2
Lϕ(k) = −4iπkFk (0) = √ √ ϕ(k), k ∈ R,
2 3λ a + iπk
and therefore it coincides with (2.5). Moreover, by inverting in Fourier space the system (A.4), one can easily recover
the partial differential equation satisfied by f and given in (2.11). Finally, the integrability conditions (2.10) follow
from the Parseval identity:
   2  2
8af 2 + 24λ(∂v f )2 (u, v) du dv = 8a Fk (v) + 24λFk (v) dk dv
R×R+ R×R+
√
|2iπk|2 a + |a + iπk|  2
= √ ϕ(k) dk
R 2 6λ |a + iπk|

= 
ϕ(−k)(−Sϕ)(k) dk
R

= ϕ(u)(−Sϕ)(u) du.
R

Appendix B: Aside computations

B.1. The carré du champ

Let f, g : → R be local smooth functions. Since the operator A is a first-order operator, we have the Leibniz rule

A(f g) = f Ag + gAf.

The operators S1 , S2 are second-order differential operators. Therefore, the relation above does not hold. Recall that
the bilinear operators Qi (i = 1, 2) are given by

Qi (f, g) = Si (f g) − f Si g − gSi f.

In our situation, these carrés des champs have simple expressions:


Q1 (f, g) = 2 (Xx f )(Xx g),
x∈Z

Q2 (f, g) = 2 (Yx f )(Yx g).
x∈Z

We will only evaluate the carré du champ on pairs of functions of the form (ωx , ωy ). In the case of Q1 , we have four
cases. First, Q1 (ωx , ωy ) = 0 if |y − x| ≥ 3. We have that

Q1 (ωx−1 , ωx+1 ) = 2(Xx ωx−1 )(Xx ωx+1 )

= 2(ωx+1 − ωx )(ωx − ωx−1 ).
1750 C. Bernardin et al.

Using the identity 2(a − b)(b − c) = (a − c)2 − (a − b)2 − (b − c)2 we can rewrite

Q1 (ωx−1 , ωx+1 ) = (ωx+1 − ωx−1 )2 − (ωx+1 − ωx )2 − (ωx − ωx−1 )2 .

In a similar way,

Q1 (ωx , ωx+1 ) = 2(ωx+1 − ωx )2 − (ωx+2 − ωx+1 )2 − (ωx+2 − ωx )2

− (ωx+1 − ωx−1 )2 − (ωx − ωx−1 )2 ,
Q1 (ωx , ωx ) = 2(ωx+2 − ωx+1 )2 + 2(ωx+1 − ωx−1 )2 + 2(ωx−1 − ωx−2 )2 .

In the case of Q2 we have three different cases:

Q2 (ωx , ωy ) = 0, |y − x| ≥ 2,
Q2 (ωx , ωx+1 ) = −2ωx ωx+1 ,
Q2 (ωx , ωx ) = 2ωx−1
2
+ 2ωx+1
2
.

B.2. The generator applied to quadratic functions

As mentioned before, the correlation field plays a fundamental role in the derivation of energy fluctuations. In order
to see this, we need to make a very detailed study of the action of the generator Ln over functions of the form

ωx ωy qx,y ,
x,y∈Z

where q : Z2 → R will be chosen within a few lines and is supposed to be symmetric: qx,y = qy,x . We have

Ln (ωx ωy ) = ωx Ln ωy + ωy Ln ωx + λQ1 (ωx , ωy ) + γn Q2 (ωx , ωy ). (B.1)

˜ u
Let us introduce some notation that will be useful later on. For u : Z → R we define ∇u, ˜ : Z → R as

˜ x = 1 (ux+1 − ux−1 ),
∇u
1
˜ x = (ux−2 + 2ux−1 − 6ux + 2ux+1 + ux+2 ).
u
2 6
One can check that
˜ x + 6λω
Ln ωx = 2∇ω ˜ x − 2γn ωx .

For q : Z2 → R define Aq : Z2 → R as

Aqx,y = qx+1,y − qx−1,y + qx,y+1 − qx,y−1 .

In other words,
˜ x ,y + 2∇q
Aqx,y = 2∇q ˜ x,y↑ ,
↑

where the arrows indicate on which variable the ∇˜ operator acts. Define as well Sq : Z2 → R as
˜ x ,y + 6q
Sqx,y = 6q ˜ x,y↑ .
↑

Performing an integration by parts and using (B.1) we have that

 
Ln ωx ωy qx,y = ωx ωy (−A + λS − 4γn )qx,y
x,y∈Z x,y∈Z
 
+ λQ1 (ωx , ωy ) + γn Q2 (ωx , ωy ) qx,y .
x,y∈Z
 Interpolation process between standard diffusion and fractional diffusion 1751

The second sum on the right hand side of the last identity is what we call the stochastic interaction term, since it only
appears due to the stochastic nature of the dynamics. Although the first sum also depends on the stochastic noise, it
can be constructed from deterministic dynamics as well.
The computations of Section B.1 show that

Q1 (ωx , ωy )qx,y
x,y∈Z

= 2(ωx+1 − ωx−1 )2 {qx,x + qx−1,x+1 − qx−1,x − qx,x+1 }
x∈Z

+ 2(ωx+1 − ωx )2 {qx−1,x−1 + 2qx,x+1 + qx+2,x+2 }
x∈Z

− 2(ωx+1 − ωx )2 {qx−1,x + qx−1,x+1 + qx,x+2 + qx+1,x+2 },
x∈Z

= 2ωx2 {qx−2,x−2 + 2qx−1,x−1 + 2qx+1,x+1 + qx+2,x+2 }
x∈Z

− 4ωx2 {qx−1,x+1 + qx+1,x+2 + qx−2,x−1 }
x∈Z

− 4ωx ωx+1 {qx−1,x−1 + 2qx,x+1 + qx+2,x+2 − qx−1,x − qx−1,x+1 − qx,x+2 − qx+1,x+2 }
x∈Z

− 4ωx+1 ωx−1 {qx,x + qx−1,x+1 − qx−1,x − qx,x+1 }, (B.2)
x∈Z

and we also have

    
Q2 (ωx , ωy )qx,y = 2
2 ωx−1 + ωx+1
2
qx,x − 4ωx ωx+1 qx,x+1 . (B.3)
x,y∈Z x∈Z

Let us go on and consider now the particular choice qx,y := fx,y

n given in (3.1) where f : R × R → R is a smooth
+
function with enough decay at infinity. The computations are pretty involved; we consider in this section only the
linear part

ωx ωy (−A + λS − 4γn )fx,y
n
.
x,y∈Z

To simplify the notation we define f (u, v) := f (u, −v) for v < 0. We call this definition symmetrization. Extending
f in this way, the resulting function may be no longer differentiable at v = 0 (but it is smooth in u and has left and
n , we have:
right derivatives in ν at 0). Moreover, with this extension, and recalling the definition (3.1) of fx,y
n
fx,y = fy,x
n
, for any x, y ∈ Z. (B.4)
n and Sf n . Consider (x, y) situated on the upper half-plane delimited by the diagonal
We start by computing Afx,y x,y
{x = y}, namely: y ≥ x. Then, for any i ∈ Z and for any j ≥ 0, we have
x +y i y −x j x +y y −x
f + , √ +√ −f , √
2n 2n n n 2n n
j x +y y −x 1 i j2 2 x +y y −x
= √ ∂v f , √ + ∂u + ∂vv f , √
n 2n n n 2 2 2n n
1 ij 2 j3 3 x +y y −x 1
+ ∂uv + ∂vvv f , √ + Oi,j , (B.5)
n3/2 2 6 2n n n2
1752 C. Bernardin et al.

where Oi,j ( n12 ) represents a sequence of functions in Z2 bounded by c(i,j

n2
)
for some finite constant c(i, j ) and for any
n ∈ N. In the following, we denote O( n12 ) when the sequence of functions is bounded by nc2 and c does not depend on
any index.
From now on we denote

∂f ( x+y y−x
√
2n , n ) if y > x,
∂fx,y =
n
x +
∂f ( n , 0 ) if y = x,

where ∂ can be any differentiate operator involving the variable v. For x = y we have from (B.5) that
2 1
n
Afx,y = ∂u fx,y
n
+O 2 .
n n
For x = y, the expression is different due to the symmetrization of f . We have that
2 1
n
Afx,x = ∂u fx,x
n
+ O 3/2 .
n n

Note that the term of order O( n1 ) is the same in both expressions, the difference appears only at order O( n3/2
1
).
Now let us compute Sfx,y . The lack of regularity of f at v = 0 affects the computations if |x − y| ≤ 1. In particular,
n
we can ensure that all the differences of the form fx+k,y+ − fx,y
n appear in such a way that x + k ≤ y +  and x ≤ y.

With this precaution, we avoid to cross the axis {x = y} where derivatives can have jumps due to the irregularity of f .
For |y − x| ≥ 2 we have
12 2 n 1
n
Sfx,y = ∂ f +O 2 .
n vv x,y n

For y = x + 1 we write the Taylor expansion centered at ( xn , 0) as follows

4 12 2 1
n
Sfx,x+1 = √ ∂v + ∂vv n
fx,x + O 3/2 .
n n n
For y = x we have
16 12 2 1
n
Sfx,x = √ ∂v + ∂vv n
fx,x + O 3/2 .
n n n
Putting together all the expressions computed above, and recalling (2.1), we see that
   
2  n 1
ωx ωy (−A + λS − 4γn )fx,y =
n
ωx ωy −∂u + 6λ∂vv − 2a fx,y + O 2
2
n n
x,y∈Z x,y∈Z
  
8λ 1
+ ωx ωx+1 √ ∂v fx,x
n
+ O 3/2
n n
x∈Z
  
2 16λ 1
+ ωx √ ∂v fx,x + O 3/2 .
n
n n
x∈Z

B.3. The carré du champ revisited

In this section we perform the same computations for both carrés des champs. It is quite easy to see from (B.3) that
      
4a n 1 4a n 1
γn Q2 (ωx , ωy )fx,y
n
= ωx ωx+1 − fx,x + O 3/2 + ωx2 fx,x + O 2 .
n n n n
x,y∈Z x∈Z x∈Z
 Interpolation process between standard diffusion and fractional diffusion 1753

We now deal with Q1 (see (B.2)). First, we consider the term with ωx2 , and we write the Taylor expansion at ( xn , 0) as

1 n 1 n
f + fx−1,x−1
n
+ fx+1,x+1
n
+ fx+2,x+2 − fx−1,x+1
n
− fx+1,x+2
n
− fx−2,x−1
n
2 x−2,x−2 2
4 3 2 1
= − √ ∂v − ∂vv n
fx,x + O 3/2 .
n n n
Then, we have the term with ωx ωx+1 , and we write the Taylor expansions at ( xn , 0):
n
fx−1,x−1 + 2fx,x+1
n
+ fx+2,x+2
n
− fx−1,x
n
− fx−1,x+1
n
− fx,x+2
n
− fx+1,x+2
n

4 4 2 1
= − √ ∂v − ∂vv n
fx,x + O 3/2 .
n n n
Finally, the term with ωx+1 ωx−1 gives the Taylor expansion centered at ( xn , 0) as:

1 2 n 1
n
fx,x + fx−1,x+1
n
− fx−1,x
n
− fx,x+1
n
= ∂vv fx,x + O 3/2 .
n n
Therefore,
   
16λ 12λ 2 1
n
λQ1 (ωx , ωy )fx,y = ωx2 − √ ∂v − n
∂vv fx,x + O 3/2
n n n
x,y∈Z x∈Z
  
16λ 16λ 2 1
+ ωx ωx+1 √ ∂v + n
∂vv fx,x + O 3/2
n n n
x∈Z
  
4λ 2 n 1
+ ωx+1 ωx−1 − ∂vv fx,x + O 3/2 .
n n
x∈Z

Putting every computation together, we obtain

  
2    n 1
Ln ωx ωy fx,y =
n
ωx ωy −∂u + 6λ∂vv − 2a fx,y + O
2
n n
x,y∈Z x,y∈Z
 
4  1   n 1
+√ ωx ωx+1 6λ∂v − √ a − 4λ∂vv 2
fx,x + O
n n n
x∈Z
 
4 2   n 1
+ ωx −3λ∂vv 2
+ a fx,x +O √
n n
x∈Z
 
4λ  1
− ωx+1 ωx−1 ∂vv fx,x + O √
2 n
.
n n
x∈Z

and after simplifications

  
2   n 1
Ln n
ωx ωy fx,y = ωx ωy −∂u + 6λ∂vv − 2a fx,y + O
2
n n
x,y∈Z x=y
 
2 2 1
− n
ωx ∂u fx,x +O √
n n
x∈Z
 
24λ  1
+√ ωx ωx+1 ∂v fx,x + O
n
n n
x∈Z
1754 C. Bernardin et al.

4   n 
+ 2
ωx ωx+1 4λ∂vv − a fx,x
n
x∈Z
  
4 1
− fx,x + O √
2 n
ωx+1 ωx−1 λ∂vv .
n n
x∈Z

Appendix C: L2 convergence of quadratic variations

In this section we prove Lemma 3.5. We start by showing the L2 (Pβ ) convergence for ME·,n (ϕ) t , namely (3.39).
Recall the explicit formula for the quadratic variation given in (3.25). By using the inequality (x + y)2 ≤ 2x 2 + 2y 2
several times, we split the four terms appearing in (3.25) and we control each one separately by using exactly the same
approach. We only give the proof of the control for one of them. We start by computing the variance of
 2
√ t  n 
n 4λ ωxn (s)ωx+1
n
(s) ϕx+1 − ϕxn ds,
0 x∈Z

where ϕxn = ϕ( xn ). Last expression can be written as

√ t   n  n 
n 4λ ωxn (s)ωx+1
n
(s)ωyn (s)ωy+1
n
(s) ϕx+1 − ϕxn ϕy+1 − ϕyn ds. (C.1)
0 x,y∈Z

Note that under the equilibrium probability measure μβ the expectation of [ωxn ωx+1
n ωn ωn ](s) is non-zero only for
y y+1
diagonal terms y = x, so that the expectation of (C.1) is equal to
√   2 2  n 2
4λt n ω0 ω1 β ϕx+1 − ϕxn .
x,y∈Z

Define χx,x+1 := ωx2 ωx+1

2 − ω02 ω12 β which are centered random variables. By stationarity and the Cauchy–Schwarz
inequality, the variance of (C.1) is bounded by

  n 2 2
Ct 2 n χx,x+1 ϕx+1 − ϕxn μβ (dω) (C.2)
x∈Z

  n  n  2
+ Ct 2 n ωx ωx+1 ωy ωy+1 ϕx+1 − ϕxn ϕy+1 − ϕyn μβ (dω) (C.3)
x=y∈Z

for some constant C > 0. First we look at the diagonal terms. Developing the square of the sum, since the variables
χx,x+1 and χy,y+1 are correlated only if |y − x| ≤ 1, by the Cauchy–Schwarz inequality, the term (C.2) can be bounded
from above by
 4  
t 2 C(β)n n
ϕx+1 − ϕxn = O n−2 .
x∈Z

For the remaining term, by developing the square of the sum and using the fact that the variables {ωx }x∈Z have mean
zero and are i.i.d. under μβ we bound it from above by
 2  2  
t 2 C(β)n n
ϕx+1 − ϕxn n
ϕy+1 − ϕyn = O n−1 .
x,y∈Z

We let the reader work out the same argument in order to finish the proof of (3.39).
 Interpolation process between standard diffusion and fractional diffusion 1755

Now we turn to MC·,n (f ) t and we prove (3.40). Recall the explicit expression (3.26) in the proof of Lemma 3.3.
We use again the inequality (x + y)2 ≤ 2x 2 + 2y 2 several times and we control each term separately by using exactly
the same approach. We present the proof for the contribution of the term with Xz but we note that for the term with
Yz the estimates are analogous. Recall (3.27)–(3.32). We note that the most demanding terms are those coming from
(3.31) and (3.32). To make the exposition as simple as possible, we look only at one of these terms, which is of the
form

1 t    n  2
√ 2λ 2 ωyn (s)ωzn (s) fz+1,y − fz−1,y
n
ds
n 0 z∈Z y ∈{z−1,z,z+1}
/

and can be written as

8λ t  2   n  2
√ ωzn (s) ωyn (s) fz+1,y − fz−1,y
n
ds.
n 0 z∈Z y ∈{z−1,z,z+1}
/

We sum and subtract the mean of (ωzn (s))2 to write last term as

8λ t  2    n  2
√ ωzn (s) − β −1 ωyn (s) fz+1,y − fz−1,y
n
ds (C.4)
n 0 z∈Z y ∈{z−1,z,z+1}
/

8λ t    n  2
+√ β −1 ωyn (s) fz+1,y − fz−1,y
n
ds. (C.5)
n 0 z∈Z y ∈{z−1,z,z+1}
/

Now we estimate the variance of each term separately. First, we note that the mean of (C.4) is zero so that its variance
is given by

Ct 2     n  2
ωz2 − β −1 ωy fz+1,y − fz−1,y
n
n
z∈Z y ∈{z−1,z,z+1}
/

    n  2
× ωz̄2 − β −1 ωu fz̄+1,u − fz̄−1,u
n
μβ (dω).
z̄∈Z u∈{z̄−1,z̄,z̄+1}
/

To bound from above this last expression, we expand the squares and use the independence of the centered random
variables {ωx }x∈Z . Therefore last expectation is bounded from above by the sum of two terms, according to z = z̄ and
z = z̄. The first one is

t 2 C(β)    2  2
n
fz+1,y − fz−1,y
n n
fz+1,u − fz−1,u
n
,
n
z∈Z y,u∈{z−1,z,z+1}
/

which, by (3.34), can be bounded from above by

t 2 C(β)    2 2
C
n
∂v fz,y ≤
n3 n
z∈Z y ∈{z−1,z,z+1}
/

and vanishes as n → ∞. The second is

Ct 2     n 2  2   n 2
ωz2 − β −1 ωz̄2 fz+1,z̄ − fz−1,z̄
n
ωz̄ − β −1 ωz2 fz̄+1,z − fz̄−1,z
n
μβ (dω).
n
z=z̄∈Z
1756 C. Bernardin et al.

Last expectation is bounded from above by

t 2 C(β)   4 C
n
∂v fz+1,z̄ ≤ 3/2 ,
n3 n
z=z̄∈Z

and vanishes as n → ∞. Now we compute the variance of (C.5) which, by developing the square in the sum, can be
written as
8λ t    n  n 
√ β −1 ωyn (s)ωȳn (s) fz+1,y − fz−1,y
n
fz+1,ȳ − fz−1,
n
ȳ ds.
n 0 z∈Z y,ȳ ∈{z−1,z,z+1}
/

First note that its mean is given by

8λt  −2   2
√ β n
fz+1,y − fz−1,y
n
,
n
z∈Z y ∈{z−1,z,z+1}
/

and therefore, its variance can be bounded from above by

Ct 2     n 2 2
β −1 ωy2 − β −1 fz+1,y − fz−1,y
n
μβ (dω)
n
z∈Z y ∈{z−1,z,z+1}
/

Ct 2    n  n  2
+ β −1 ωy ωȳ fz+1,y − fz−1,y
n
fz+1,ȳ − fz−1,
n
ȳ μβ (dω).
n
z∈Z y=ȳ
y,ȳ ∈{z−1,z,z+1}
/

Now, the first expectation in the previous display can be bounded from above by

C(β)t 2    2  2 C
n
∂v fz,y n
∂v fz̄,y ≤
n3 n
z,z̄∈Z y ∈{z−1,z,z+1}
/

and vanishes as n → ∞; while the second one can be bounded from above by

C(β)t 2   n  n  n  n 
fz+1,y − fz−1,y
n
fz+1,ȳ − fz−1,
n
ȳ fz̄+1,ȳ − fz̄−1,ȳ fz̄+1,y − fz̄−1,y
n n
n
y=ȳ z,z̄

which is equal to

C(β)t 2     2
C
n
∂v fz,y ∂v fz,nȳ ≤
n3 n
y=ȳ z
y,ȳ ∈{z−1,z,z+1}
/

and vanishes as n → ∞.

Acknowledgement

M.S. thanks CAPES (Brazil) and IMPA (Instituto de Matematica Pura e Aplicada, Rio de Janeiro) for the post-doctoral
fellowship.

References

[1] G. Basile, C. Bernardin and S. Olla. Momentum conserving model with anomalous thermal conductivity in low dimensional systems. Phys.
Rev. Lett. 96 (2006) 204303.
 Interpolation process between standard diffusion and fractional diffusion 1757

[2] G. Basile, C. Bernardin and S. Olla. Thermal conductivity for a momentum conserving model. Comm. Math. Phys. 287 (1) (2009) 67–98.
MR2480742
[3] C. Bernardin, P. Gonçalves and M. Jara. 3/4-Fractional superdiffusion in a system of harmonic oscillators perturbed by a conservative noise.
Arch. Ration. Mech. Anal. 220 (2) (2016) 505–542. MR3461356
[4] C. Bernardin, P. Gonçalves, M. Jara, M. Sasada and M. Simon. From normal diffusion to superdiffusion of energy in the evanescent flip noise
limit. J. Stat. Phys. 159 (6) (2015) 1327–1368. MR3350374
[5] C. Bernardin and G. Stoltz. Anomalous diffusion for a class of systems with two conserved quantities. Nonlinearity 25 (4) (2012) 1099–1133.
MR2904271
[6] E. Fermi, J. Pasta and S. Ulam. Studies of nonlinear problems. I. Los Alamos Report LA-1940, 1955. Published later in Collected Papers of
Enrico Fermi, E. Segré (Ed.), University of Chicago Press, 1965.
[7] P. Gonçalves and M. Jara. Density fluctuations for exclusion processes with long jumps. Probab. Theory Related Fields 170 (1–2) (2018)
311–362. MR3748326
[8] M. Jara. Quadratic fluctuations of the simple exclusion process. Preprint, 2014. Available at arXiv:1401.2609.
[9] M. Jara, T. Komorowski and S. Olla. Superdiffusion of energy in a chain of harmonic oscillators with noise. Comm. Math. Phys. 339 (2)
(2015) 407–453. MR3370610
[10] C. Kipnis and C. Landim. Scaling Limits of Interacting Particle Systems. Springer-Verlag, Berlin, 1999. MR1707314
[11] I. Mitoma. Tightness of probabilities on C([0, 1]; S ) and D([0, 1]; S ). Ann. Probab. 11 (4) (1983) 989–999. MR0714961
[12] S. Sethuraman. Central limit theorems for additive functionals of the simple exclusion process. Ann. Probab. 28 (2000) 277–302. MR1756006
[13] W. Whitt. Proofs of the martingale FCLT. Probab. Surv. 4 (2007) 268–302. MR2368952