Journal of Physics A: Mathematical and Theoretical

J. Phys. A: Math. Theor. 57 (2024) 435001 (16pp) https://doi.org/10.1088/1751-8121/ad7faa

General properties of the response

function in a class of solvable
non-equilibrium models
F Corberi1,2 and L Smaldone1,2,∗
1
Physics Department ‘E.R. Caianiello’, University of Salerno, Via Giovanni Paolo
II, 132, 84084 Fisciano, Salerno, Italy
2
INFN Sezione di Napoli, Gruppo collegato di Salerno, Fisciano, Italy

E-mail: [email protected] and [email protected]

Received 11 July 2024; revised 2 September 2024

Accepted for publication 25 September 2024
Published 7 October 2024

Abstract
We study the non-equilibrium response function Rij (t, t ′ ), namely the variation
of the local magnetization ⟨Si (t)⟩ on site i at time t as an effect of a perturb-
ation applied at the earlier time t′ on site j, in a class of solvable spin models
characterized by the vanishing of the so-called asymmetry. This class encom-
passes both systems brought out of equilibrium by the variation of a thermo-
dynamic control parameter, as after a temperature quench, or intrinsically out
of equilibrium models with violation of detailed balance. The one-dimensional
Ising model and the voter model (on an arbitrary graph) are prototypical
examples of these two situations which are used here as guiding examples.
Defining the fluctuation-dissipation ratio Xij (t, t ′ ) = βRij /(∂Gij /∂t ′ ), where
Gij (t, t ′ ) = ⟨Si (t)Sj (t ′ )⟩ is the spin–spin correlation function and β is a para-
meter regulating the strength of the perturbation (corresponding to the inverse
temperature when detailed balance holds), we show that, in the quite general
case of a kinetics obeying dynamical scaling, on equal sites this quantity has
a universal form Xii (t, t ′ ) = (t + t ′ )/(2t), whereas limt→∞ Xij (t, t ′ ) = 1/2 for
any ij couple. The specific case of voter models with long-range interactions is
thoroughly discussed.

Keywords: voter model, response function, fluctuation-dissipation theorem,

nonequilibrium dynamics, spin systems

∗
Author to whom any correspondence should be addressed.

Original Content from this work may be used under the terms of the Creative Commons Attribution
4.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the
title of the work, journal citation and DOI.

© 2024 The Author(s). Published by IOP Publishing Ltd 1

J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

1. Introduction

While statistical mechanics provides a rather complete and detailed theory of equilibrium sys-
tems, understanding out of equilibrium states is still an open challenge. Depending on the
system at hand, in some cases such states can be treated as weak perturbations around a
given equilibrium, allowing sometimes to build some kind of perturbation approach. However,
there are in nature a wealth of phenomena that cannot be considered near to equilibrium
by any means. One example is provided by systems building an ordered state out of an ini-
tially disordered one, as in the case of phase-ordering [1–3] in quenched magnets or phase-
separation in composite materials as, for instance, binary liquids or alloys. Glassy materi-
als [4, 5] and other disordered systems [6] also fall into the same category. These systems
remain out of equilibrium for very long times because the ergodic time diverges in the thermo-
dynamic limit and may therefore become larger than any accessible experimental timescale.
Moreover, although the kinetics becomes slower and slower as time elapses, a phenomenon
dubbed as aging, the probability distribution in phase-space keeps changing at any time as it
can be revealed by inspection of suitable observables as, for instance, two-time correlations.
This slowly relaxing state can be viewed as gradual exploration of larger and larger portion
of phase-space in a progressive effort to reinstate ergodicity, a scenario referred to as weak
ergodicity breaking. The problem, therefore, cannot be handled in a statistical mechanical
approach by restricting over a subset of phase space, as it would be appropriate for the study of
metastability [7].
A different, but related, class of systems are those kept away from equilibrium by a small
external power input done by stationary non-conservative forces violating detailed balance, as
in active matter or in sheared or stirred materials, and/or periodically time-dependent forces.
In this context, a wealth of studies have been concerned with the behavior of two-time
quantities and, in particular, on the relation between the response function R(t, t ′ ), with t ⩾ t ′ ,
and the associated correlation function A(t, t ′ ), the fluctuation-dissipation relation. Here R(t, t ′ )
is the so-called autoresponse, namely the local reaction of the system at position ⃗x at time t,
due to a perturbation acting in the same place at the earlier time t′ . Similarly, A(t, t ′ ) is the
autocorrelation function between a local observable measured at ⃗x at two different times t′
and t. In equilibrium, the fluctuation-dissipation theorem (FDT) [8] provides a linear relation
between these two quantities involving the inverse thermodynamic temperature β = 1/kB T

∂
β −1 R (t, t ′ ) = A (t, t ′ ) . (1)
∂t ′
In a canonical setting, this guarantees that a thermometer immersed in the system under study
displays a temperature equal to the one of the bath the sample is in contact with [9]. Out of
equilibrium, the FDT does not hold and the quantity
∂ A(t,t ′ )
βeff = βX−1 (t, t ′ ) = ∂t ′
(2)
R (t, t ′ )

has been interpreted, in some cases, as an effective inverse temperature [10], different from
that of the reservoir. For instance, in a class of aging systems as, e.g. the p-spin model of a glass
[11, 12], the so called fluctuation-dissipation ratio (FDR) X(t, t ′ ) takes, for large times t′ , only
a couple of values, X(t, t ′ ) = 1 for t − t ′ ≪ t ′ and X(t, t ′ ) = X∞ ̸= 1 for t − t ′ ≫ t ′ , showing
the presence, next to the bath temperature, of an additional effective (inverse) temperature
βeff = βX− 1
∞ . In addition, the FDR has been also used as an indicator of the out-of-equilibrium

2
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

nature of the system, and has been related to the rate of energy dissipation [13] or the entropy
production rate [14].
Interestingly, besides evidencing the nature of the deviation from equilibrium, in slowly
relaxing systems X(t, t ′ )—a quantity compute in a far-from-equilibrium state—encodes also
information on the equilibrium state the system is approaching. Indeed, a theorem [15, 16]
relates the FDR to the overlap probability distribution of the target equilibrium.
In this paper we discuss the form of the FDR in a class of models where the so-called
asymmetry (see section 2, equation (12) in particular) vanishes, a fact that makes an analytical
approach possible. Brownian diffusion is perhaps the simplest example. The problem is defined
in terms of spin variables Si = ±1 located on the vertices i of a completely general graph. For
concreteness, after discussing general properties, we will consider the voter model and perform
a quantitative analysis on regular D-dimensional lattices, considering interactions depending
on the distance r between spins [17–20]. The one-dimensional case with nearest neighbors
(NN) interactions can be mapped to the kinetic Ising chain, whose FDR during phase-ordering
was derived long ago in [21, 22]. Besides that, we will consider here cases with long-range
interactions between spins decaying as r−α , for any α. In the scaling regime of the model,
where the time evolution can be reparametrized in terms of a single growing length L(t) (see
section 2, especially around equation (17) for a precise definition) the FDR is a non-trivial
universal function of its arguments decaying from X ≡ 1 at equal times t = t ′ , to X ≡ 1/2
for large time separations t − t ′ → ∞. We also discuss the space-dependent FDR obtained
through a generalization of equations (1) and (2) where the response describes the reaction of
the system at ⃗x due to a perturbation acting earlier on a different site ⃗x +⃗r. This quantity has a
more complicated behavior than its equal site counterpart but, similarly to it, it also converges
to X = 1/2 for large t − t ′ .
The paper is organized as follows: in section 2 we present some generalities on fluctuation-
dissipation theorem and we derive some results on FDR which will be tested in the voter
model. In section 3 we review the voter model with long-range interactions in its general form
and then we provide explicit calculations of both response function and FDR in both one-
and two-dimensions. Finally in section 4 we present conclusions and discussion. For reader’s
convenience we have devoted appendix to derive some auxiliary computations which were
employed in the main text.

2. Fluctuation-dissipation relations: generalities

In this section we review some generalities on linear response theory and fluctuation-
dissipation relations in and out of equilibrium. We will use the paradigm of spin systems,
but most of the properties reviewed here are rather general.
Let us consider a collection of Ising variables Si = ±1 evolving according to a Markovian
′
stochastic process in discrete time. In general, the linear response function RAB
ij (t, t ) describes
the modification of an observable Ai at site i and time t due to a small perturbation b coupling
linearly with another variable Bj at previous times t′ . The simplest instance, in a spin system,
is when Ai ≡ Si , Bj ≡ Sj and b = hj is a magnetic field. When a kick perturbation acting on a
very short time ∆t is applied

hj (t) = hθ (t − t ′ ) θ (t ′ + ∆t − t) δi,j , (3)

where θ is the Heaviside function, the impulsive response is defined as [23, 24]

3
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

1 δ⟨Si (t)⟩
Rij(t,t ′ ) = lim . (4)
∆t→0 ∆t δhj (t ′ ) h= 0

Here ⟨. . . ⟩ indicates a non-equilibrium statistical average, namely taken over initial configur-
ations and stochastic realizations of the evolution, and

δ⟨Si (t)⟩ = ⟨Si (t)⟩h − ⟨Si (t)⟩ (5)

where ⟨. . . ⟩h means an average over trajectories in the presence of the perturbation.

The response function can be alternatively written as [21, 25, 26]
 
−1 ′ 1 ∂ ′ ′
β Rij (t, t ) = Gij (t, t ) − ⟨Si (t) Bj (t )⟩ , (6)
2 ∂t ′

where

Gij (t, t ′ ) = ⟨Si (t) Sj (t ′ )⟩ (7)

is the spin–spin correlation function, and

Bj = −2Sj w (Sj ) , (8)

w(Si ) being the transition rate for flipping a single spin Si in the absence of the perturbation.
Equation (6) holds true for any kind of unperturbed dynamic rules, also if detailed balance
is violated, for any perturbation modifying the unperturbed transition rates as w(Si ) → wh (Si )
with

wh (Si ) = w (Si ) (1 − βhj ) , (9)

where β is a parameter regulating the effect of the kick. If the unperturbed evolution
obeys detailed balance, β can be identified with the inverse thermodynamic temperature
β = (kB T)−1 , kB being the Boltzmann constant. Equation (9) can be interpreted as the require-
ment that the perturbed Markov chain obeys detailed balance at least with respect to the effect
of a perturbation adding an energy −hj Sj to the unperturbed system. Clearly, if the latter obeys
detailed balance with respect to a certain Hamiltonian, equation (9) amounts to the natural
request that detailed balance is still valid after switching on the perturbation. Notice that on
the r.h.s. of equation (6) there are only correlation functions computed in the unperturbed
dynamics. Therefore equation (6) can be qualified as a non-equilibrium generalization of the
FDT. Indeed, it is simple to show [21, 25, 26] that it reduces to the usual equilibrium relation

∂
β −1 Rij (t, t ′ ) = Gij (t, t ′ ) (10)
∂t ′
in equilibrium conditions. This will be further discussed below. Let us also mention that
equation (6) also holds, with a suitable definition of Bj , for systems of continuous variables
described, e.g. by Langevin or Fokker–Planck equations [27].
It is useful to re-write equation (6) in the alternative form
 
−1 ′ 1 ∂Gij (t, t ′ ) ∂Gij (t, t ′ ) ′
β Rij (t, t ) = − − Aij (t, t ) , (11)
2 ∂t ′ ∂t

4
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

where
∂Gij (t, t ′ )
Aij (t, t ′ ) = ⟨Si (t) Bj (t ′ )⟩ − (12)
∂t
∂ Gij (t,t ′ )
is the so called asymmetry. Since it can be shown [27] that ∂t = ⟨Bi (t)Sj (t ′ )⟩, this quantity
amounts to a sort of commutator,

Aij (t, t ′ ) = ⟨Si (t) Bj (t ′ )⟩ − ⟨Bi (t) Sj (t ′ )⟩ , (13)

whereby the name comes from. In stationary states with time-inversion invariance, Aij vanishes
and one recovers the equilibrium FDT, equation (10). However, equilibrium is not a necessary
condition for having Aij ≡ 0, as we will discuss below.
The generalized FDT (6) can be also written in the form [21]

∂Gij (t, t ′ )
βRij (t, t ′ ) = Xij (t, t ′ ) , (14)
∂t ′
where
 
∂ Gij (t,t ′ )
1 Aij (t, t ′ ) 
Xij (t, t ′ ) = 1 − ∂t
∂ Gij (t,t ′ )
− ∂ Gij (t,t ′ )
(15)
2
∂t ′ ∂t ′

is the FDR. This quantity can, in some cases, be associated to a non-equilibrium effective
temperature [10]. Clearly, in equilibrium one has Xij (t, t ′ ) ≡ 1. In the following we will denote
the equal-site FDR Xii (t, t ′ ) as X(t, t ′ ) tout court. Similarly, also the autocorrelation function
Gii (t, t ′ ) will be denoted as A(t, t ′ ).
The considerations above are general. Now we restrict our discussion to systems where the
asymmetry vanishes. Indeed, as we mentioned before, equilibrium states are not a necessary
condition for having Aij ≡ 0 as there are plenty of examples of models without asymmetry
also away from equilibrium. For instance, this happens in the aging stage of the 1D Ising
model [21, 22, 25], in some instances of trap models [28–30] and, as we will see soon, in
the voter model in arbitrary space dimension and with arbitrary spacial interactions. However,
the physical significance of having vanishing asymmetry has not been fully elucidated today,
because this quantity cannot be straightforwardly associated, in general, to an interpretable
physical observable [31, 32]. In general, the vanishing of the asymmetry depends on both
the type of model and the nature of the responding variable (the spin Si , in this article). In
particular, a class of variables has been identified in [31], denoted as there as randomizing, for
which the asymmetry can be generally proven to vanishes. However, this is not a necessary
condition for having Aij ≡ 0. For instance, it is easily seen that the spin variable in the voter
model, that will be thoroughly considered in this paper, is not randomizing in the sense of [31],
but is nevertheless associated to a vanishing asymmetry.
The FDR reads, in this case,
 
∂ Gij (t,t ′ )
1
Xij (t, t ′ ) = 1 − ∂ G ∂(tt,t ′ )  . (16)
2 ij
′
∂t

The first observation is that, in stationary states where Gij is a function of the time difference t −
t ′ only, one has Xij (t, t ′ ) ≡ 1, as in equilibrium. In systems where detailed balance is violated,

5
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

such states are non-equilibrium ones and occur, for instance in the voter model, as we will
discuss. In addition, it can be easily proved that limt′ →t X(t, t ′ ) = 1, even in cases in which
detailed balance does not hold (see appendix). This can be clearly seen in figures 1 and 3,
where X(t, t ′ ) is plotted for the d = 1 and d = 2 voter model with long-range interactions (see
section 3), respectively. In this figures, following established practice, X(t, t ′ ) is plotted against
A(t, t ′ ). Given that A(t, t ′ ) is a function decreasing monotonously from A(t, t) = 1 to zero (for
t ≫ t ′ ), the region t ′ → t corresponds to the right part of the plot.
Instead, if i ̸= j the behavior of Xij (t, t ′ ) in the above equal times limit is radically different,
since one has limt→t ′ Xi̸=j (t, t ′ ) = 0. This is shown in appendix. The physical intuition is that
flipping Si at time t or Sj at time t′ occurs with the same probability, hence ∂ tij = ∂ t ′ij in
∂G ∂G

equation (16), as t ′ → t. Alternatively, one arrives at the same conclusion upon arguing that
a perturbation exerted somewhere else at j, cannot reach i in a vanishing time interval. As a
concrete example, this can be observed for the long-range d = 1 voter model in figure 2.
Now we focus on aging systems. Assuming spatial homogeneity it is Gij (t, t ′ ) = G(r; t, t ′ ),
where r is the distance between i and j. Generally the late-time dynamics obeys a dynamical
scaling symmetry [1–3, 33–41] expressed by the form
 
r L (t)
G (r; t, t ′ ) = g ; , (17)
L (t ′ ) L (t ′ )

where L(t) is a dynamical correlation length. From equation (16) one has
" dg ∂ L(t)
#
′ 1
X (r; t, t ) = 1 + dg dz dg ∂t
∂ L(t ′ )
, (18)
2 y dy + z dz ∂t ′

where y = r/L(t ′ ) and z = L(t)/L(t ′ ) are the arguments of g in equation (17). If, as it is usually
the case, L(t) grows algebraically in time, L(t) ∝ ta , then it is
" dg
#
′ 1 dz a−1
X (r; t, t ) = x (y; z) = 1 + dg z a , (19)
2 y dy + z dg
dz

showing that X is also a function x of the scaling variables y and z alone. Focusing on the case
r = 0, which is the situation most frequently considered, the above expression gives

1 + z− a t + t′
1

X (t, t ′ ) = = . (20)
2 2t

Then X(t, t ′ ) is a universal quantity, independent on the specific model considered among those
with vanishing asymmetry, in a regime characterized by dynamical scaling. Let us just stress
that, when plotting against A(t, t ′ ) as it is done in figures 1 and 3, curves inherit the model
dependence of A(t, t ′ ), as it can be visually observed.
On the contrary, for r ̸= 0, equation (19) shows that the model-dependent scaling function
g informs X. The limiting value [22]

X∞ (r; t ′ ) = lim X (r; t, t ′ ) , (21)

t→∞

(taken with fixed r and t′ ), however, retains a universal character similar to the one pointed out
for X(t, t ′ ). In fact, the second term in square brackets in equation (19) generally vanishes as

6
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

z → ∞ (because z dg
dz → 0 and a ⩽ 1), and hence

1
X∞ (r; t ′ ) ≡ , (22)
2

for any r and t′ . This is clearly seen in figure 2.

The discussion above is not limited to any specific model, though the voter model (which is
examined in detail in the next section) has occasionally been used to illustrate our reasoning.
The concepts presented have a general applicability, with some restrictions (such as the van-
ishing of asymmetry or the occurrence of aging states) noted above. In the following section,
we will focus specifically on the voter model.

3. The voter model

The voter model [42–54] is described by a set of binary variables located on the nodes i of a
graph, which can assume the values Si = ±1. Two spins Si and Sk interact with a probability
Pik , so that the probability to flip Si is

1X
w (Si ) = Pi k (1 − Si Sk ) . (23)
2
k

With standard techniques of stochastic calculus [55] it is easy to show (see appendix) that
′
= −2⟨Si (t)Sj (t ′ )w(Si (t))⟩ which, provides
∂ t ⟨Si (t)Sj (t )⟩
∂

∂Gij (t, t ′ ) X
= − Gij (t, t ′ ) + Pi k Gkj (t, t ′ ) . (24)
∂t
k

Regarding the response function, using the transition rate w of equations (23) and (6)
amounts to
" #
1 ∂G (t, t ′
) X
β −1 Rij (t, t ′ ) = + Gij (t, t ′ ) − Pj k Gik (t, t ′ ) .
i j
(25)
2 ∂t ′
k

Comparing with equation (24), and using the fact that Gij and Rij are symmetric in i, j, it follows
 
−1 ′ 1 ∂Gij (t, t ′ ) ∂Gij (t, t ′ )
β Rij (t, t ) = − , (26)
2 ∂t ′ ∂t

showing that the asymmetry is vanishing in this model.

3.1. Voter model on a regular lattice

We now focus on the case where the graph is a D-dimensional lattice and the probability
Pik that spin Si interacts with Sk depends only on their distance. Considering the correlation
Gij (t, t ′ ) = G(r; t, t ′ ), where r is the distance between i and j, and indicating with dk the distance
between k and j, the evolution equation for this quantity reads [17–20, 56]

∂G (r; t, t ′ ) X X n(ℓ)
= −G (r; t, t ′ ) + P (ℓ) G (dk (r, ℓ) , t) , (27)
∂t
ℓ k=1

7
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

where ℓ runs over all the possible distances on the lattice and n(ℓ) is the number of lattice
sites k at distance ℓ from i. Equation (27) must be solved with initial condition G(r; t, t) =
C(r; t), where the equal-time correlation C obeys [17–19] a differential equation analogous
to equation (27). We will consider the case of the ordering process originating from a fully
disordered state at t = 0. Further details on this problem can be found in [19].
In the following we will focus on the case of interactions decaying algebraically with dis-
tance, i.e.

P (ℓ) ∝ ℓ−α , (28)

with α > 0. In the two opposite limits α → ∞ and α → 0 one recovers, respectively, the original
voter model with NN interactions only, and the mean field version where the agents are fully
connected.
We start by considering the one-dimensional case where equation (27) simplifies to

∂G (r; t, t ′ ) X
N/2
= −G (r; t, t ′ ) + P (ℓ) [G (r − ℓ; t, t ′ ) + G (r + ℓ; t, t ′ )] . (29)
∂t
ℓ=1

Before starting the discussion of the behavior of the model in the different sectors of α-
values let us remark that, for any value of α, the system obeys dynamical scaling (with some
caveats discussed in [17, 19] in the aging states.)
In the NN limit α → ∞ the model can be mapped exactly onto the Ising model. Therefore, in
this case (but only in this case) the model obeys detailed balance. More in general, the above
equation was studied with periodic boundary conditions in [19], where different behaviors
were found as α is changed. In the following, building on this knowledge, we discuss the
behavior of the response function. We will always assume the thermodynamic limit N → ∞.
For α > 3, G(r; t, t ′ ) has the form (17), with L(t) ∝ t1/2 , for sufficiently small values y <
y of y = r/L(t ′ ) (this will be further detailed later on). Then equation (20) holds true. The
∗

autocorrelation function A(z) reads [19]

√ !
2 2
A (z) = arctan √ , (30)
π z2 − 1

which is the same result as in the NN model [21, 22] and is independent on α. Inverting
equation (30) to obtain z(A) and plugging into equation (16) one has [21, 22]

1
X (A) =
, α > 3. (31)
2 − sin2 π
2A

This analytic determination is plotted with a black heavy-dotted curve in figure 1. In the same
figure, continuous curves are obtained by numerically solving equation (29), thus obtaining
A(t, t ′ ) by letting r = 0 in G(r; t, t ′ ) and in X(r; t, t ′ ), the latter quantity being obtained, starting
from G(r; t, t ′ ), by means of equation (16). One sees that the curves for α > 3 are α-independent
and fall onto the analytic form (31).
We now move to study X(r; t, t ′ ) for finite values of r. To the best of our knowledge this
quantity has not been thoroughly studied before. Its behavior, for α = 5, is shown in figure 2,
where it is plotted against A(t, t ′ ), as we already did for X(t, t ′ ) in figure 1. As we mentioned
in section 2, X(r; t, t ′ ) is bound to take the values X(r; t, t) = 0 for equal times (right part of
the figure, A = 1), and X(r; t → ∞, t ′ ) = 1/2 in the large time separation sector (left part of the

8
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

Figure 1. X(t, t ′ ) is plotted against G(t, t ′ ), for the 1D voter model with long-range
algebraic interactions. different curves are for different values of α, indicated in the
legend, for t ′ = 10 (t ′ = 102 for α = 3). System size is N = 103 for α ⩾ 3, N = 104 for
2 ⩽ α < 3, N = 105 for α < 2. The black heavy-dotted curve is the analytical form for
the 1d-Ising model with NN interactions obtained in [21, 22]. Curves for α = 0, 0.5, 1
superimpose.

plot, A = 0). This is clearly observed in figure 2. It can also be noticed that the curves approach
the one for r = 0 (black line) below a characteristic value A∗ . Looking at equation (19), this
means that the term z dg/dz in the denominator on the r.h.s. eventually prevails over y dg/dy.
Clearly, this occurs at values of z (i.e. of t) larger as y (hence r) are taken larger, as indeed
we see in figure 2. Moreover, we notice a rather different form of such approach for small
(r < r∗ , with r∗ ≃ 25) of large (r > r∗ ) values of r (to better display this we draw data for
r < r∗ with continuous lines and those for r > r∗ with dashed ones). For r < r∗ the curves
rise to a maximum and then approach the curve for X(t, t ′ ) (i.e. black) smoothly from above.
For r > r∗ , instead, there is an additional wiggle. The origin of r∗ and its impact on the form
of X(r; t, t ′ ) can be understood as follows. How we anticipated earlier, a unique scaling form
as in equation (17) is obeyed by G(r; t, t ′ ) only for sufficiently small values y < y∗ of y =
r/L(t ′ ). Indeed, for larger values of y, both the growth law of the scaling length and the form
of the scaling function g change significantly [19]. In particular, while for y < y∗ both g and
L(t) behave as in the NN case, for y > y∗ they become α-dependent. The value y∗ separating
these different behaviors is an increasing function y∗ (t ′ , z) of z and t′ and diverges as α →
∞, meaning that for NN interactions a unique scaling form is recovered. Plotting X(r; t, t ′ )
for fixed t′ and a small value of r, such that r < y∗ (t ′ , z = 1) = r∗ , the condition r < y∗ (t ′ , z)
will remain true for all the subsequent evolution, meaning that equation (16) only probes the
y < y∗ (t ′ , z) behavior of G(r; t, t ′ ). This produces the (continuous) lines for r < r∗ shown in
figure 2. On the other hand, considering X(r; t, t ′ ) for a value of r as large as to have r >
y∗ (t ′ , z = 1) = r∗ , equation (16) initially probes the large-y α-dependent scaling of G(r; t, t ′ ).
As time elapses (and z increases), since y∗ (t ′ , z) is an increasing function of z, one crosses
over to the universal scaling form valid for y < y∗ (t ′ , z). This produces the wiggle observed in
figure 2. Clearly, the larger is the value of r considered, the later the wiggle will be produced,
which is indeed observed in figure 2. It is interesting to notice the effectiveness of X(r; t, t ′ )
in providing such detailed information on the spatial structure of correlations in the systems.
Indeed, the existence of y∗ can be hardly inferred by the direct study of correlation functions,
particularly if numerical, because correlations are already very small at y∗ and masked by any
source of noise. Instead, X(r; t, t ′ ) displays a very clear (additional) non-monotonicity. Let us

9
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

Figure 2. X(r; t, t ′ ) is plotted against A(t, t ′ ), for α = 5 and different values of r, indic-
ated in the legend. System size is N = 103 and t ′ = 10.

also comment that such an effect is a peculiar feature of the extended interactions present in
the model, since there is no such a feature in the NN case (because y∗ = ∞).
The quantitative discussion above holds true for α > 3. Using the results contained in [19],
and proceeding similarly, one can extend the previous analysis to the range with α ⩽ 3.
Specifically, for 2 < α ⩽ 3 and large t′ and z, the autocorrelation function reads

A (z) = a1 z−1 , (32)

where a1 is a constant, with L(t) ∝ t1/(α−1) . Then one has

α−1
1 + (A/a1 )
X (A) = . (33)
2
This equation shows that X(A) decreases significantly towards the asymptotic value X(0) =
1/2 when (A/a1 )α−1 gets smaller than a certain number ϵ ≪ 1. Solving we get A ≪
a1 ϵ1/(α−1) , thus implying that the descent occurs at smaller and smaller values of A as α
decreases, which is indeed seen in figure 1.
In the range 1 < α ⩽ 2 the system undergoes a coarsening phase with

A (z) = a2 z− α−1 ,
1
(34)

with another constant a2 and with L(t) ∝ t, while approaching to a non-trivial stationary state,
characterized by
α−2
A (t, t ′ ) = Astat (τ ) = A0 τ α−1 , τ ≡ t − t′ . (35)

Then, one has

α−1
1 + (A/a2 )
X (A) = . (36)
2
during such coarsening stage. Notice that the same form holds (apart from the value of the con-
stants a1 , a2 ) both for 2 < α ⩽ 3 (equation (33)) and for 1 < α ⩽ 2 (equation (36)) despite the
fact that both the growth law L(t) and the autocorrelation behave differently. As a consequence,

10
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

Figure 3. X(t, t ′ ) is plotted against A(t, t ′ ), for the 2D voter model with long-range
algebraic interactions. Different curves are for different values of α, indicated in the
legend, for t ′ = 10. System size is N = 1012 . Curves for α ⩽ 2 superimpose.

the curves for X(t, t ′ ) in figure 1 behave similarly to those of the cases with 2 < α ⩽ 3 and keep
staying closer to X = 1 down to smaller and smaller A-values as α is progressively lowered,
for the reasons previously discussed.
Later, when the system enters the stationary state, one has
1
(α − 2) τ 1−α
R (t, t ′ ) = Rstat (τ ) = β A0 , (37)
α−1
and X(r; t, t ′ ) ≡ 1 trivially, as in any stationary state.
For α ⩽ 1, the situation is different because the stationary state is approached in a micro-
scopic time, without any previous coarsening stage [17]. At stationarity the system behaves
similarly to the mean-field model, with

A (t, t ′ ) = Astat (τ ) = a3 e−τ , (38)

a3 being a constant. Then

Rstat (τ ) = β A (τ ) , (39)

and X(r; t, t ′ ) ≡ 1. This can be seen in figure 1.

We complete our discussion of the voter model on a lattice with interaction probability
decaying algebraically as in equation (28) by briefly considering the case of a two-dimensional
square lattice. This model has been studied in [18] but two-time quantities, in particular
G(r; t, t ′ ) have never been analytically determined. Besides probing the effect of dimensional-
ity, this case is interesting also because, at variance with the D = 1 model, here dynamical scal-
ing is violated, at least for α > 4 [18], so that the analysis below equation (17) does not apply.
Our numerical results, obtained by numerically solving equation (27) with periodic boundary
conditions, are displayed in figure 3. The overall behavior is similar to the one-dimensional
case, with X(t, t ′ ) decaying from X(t, t) = 1 to X(t → ∞, t ′ ) = 1/2 on smaller and smaller val-
ues of A as α is decreased. Also, the mean-field behavior with X(t, t ′ ) ≡ 1 is observed for α ⩽ 2
since, in any dimension, this is expected for α ⩽ D. At variance with the D = 1 case, however,
X depends on α in the whole range explored and a short-range universal behavior where X is
α independent, corresponding to α > 3 for D = 1, is not observed.

11
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

4. Conclusions

In this work, we studied the impulsive response function Rij (t, t ′ ) for a class of spin sys-
tems characterized by vanishing asymmetry and subjected to a small magnetic perturba-
tion. The voter and the Ising model can be viewed as prototype examples. We derived
some general properties of the FDR Xij (t, t ′ ) for this class of models, particularly concern-
ing the equal time limit t → t ′ . Furthermore, for systems exhibiting dynamical scaling in
+

their kinetics, we found Xii (t, t ′ ) = (t + t ′ )/(2t) and limt→∞ Xij (t, t ′ ) = 1/2 for any ij pair.
Then, restricting to the one-dimensional voter model with algebraic interactions, and build-
ing on the findings from [19], which provided the analytical form of two-time autocor-
relation functions, we have computed the response function and the FDR using the FDT
in its general form for spin systems (6). This was achieved for any value of the interac-
tion exponent α, both in the aging states and in the stationary ones. Besides these ana-
lytical results, we also provided numerical results for the aforementioned model, both in
one and 2D.
The behavior of the response function and the associated FDR has been a widely studied
in aging system in the last years [10, 21–26, 37–39, 57–67]. One important direction in these
studies is the possibility to infer static properties of hard-to-equilibrate systems, as glasses and
spin glasses, from their dynamical response, enforcing the static-dynamic connection [15, 16]
mentioned in section 1 which, under certain hypotheses, relates the out-of-equilibrium FDR
to the equilibrium structure of phase space. However, it was found that in the one-dimensional
Ising model with NN interactions one of those hypotheses is violated. Indeed, the FDR in
that case (discussed in section 3.1) has an highly non-trivial form, completely independent of
the trivial properties of the equilibrium low-temperature states. The origin of such violation is
in the nature of the domain walls motion in D = 1, which is purely diffusive, since there is no
surface tension on the line. This produces an anomalously large response [68], thus invalidating
the static-dynamic connection. Such connection, however, is restored when considering the
same Ising model in D > 1, because, in that case, surface tension regularizes the motion of the
interfaces. The voter model considered in this paper is special because surface tension is absent
in any dimension and, for this reason, it displays a non-trivial FDR qualitatively similar to the
one of the NN Ising model, in any dimension (see section 3.1, figures 1 and 3, in particular).
This is presumably true for the whole class of models with vanishing asymmetry considered
in this paper.

Data availability statement

No new data were created or analyzed in this study.

Acknowledgments

F C acknowledges financial support by MUR PRIN PNRR 2022B3WCFS.

12
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

Appendix. Time derivatives of correlation functions

In this appendix we evaluate some time derivatives of Gij (t, t ′ ) for a general system of discrete
variables (e.g. spins) denoted by S ≡ (S1 , . . . SN ), subjected to a Markov process. By definition

X


Gi j (t, t ′ ) = Si Sj′ p S, t|S ′ , t ′ p S ′ , t ′ (A.1)
S,S ′


where p(S, t) is the probability of state S at time t and p S, t|S ′ , t ′ is the conditional probability
of having the state S at time t if we had the state S ′ at time t′ .
The increment of Gij when the largest time t goes from t to t + δ, is
∆t Gij (t, t ′ ) = ⟨Si (t + δ) Sj (t ′ )⟩ − ⟨Si (t) Sj (t ′ )⟩, (A.2)
which explicitly reads
X



∆t Gi j (t, t ′ ) = (Si′ ′ − Si ) Sj′ p S ′ ′ , t + δ|S, t p S, t|S ′ , t ′ p S ′ , t ′ . (A.3)
S,S ′ ,S ′ ′

In the limit of small δ a single spin-flip attempt is only possible. Moreover, it is clear that only
flip of Si contributes to the r.h.s. of equation (A.3), due to the Si′ ′ − Si factor. Therefore we can
make the replacement p(S ′ ′ , t + δ|S, t) = w(Si )δ, leading to

∂Gi j (t, t ′ ) ∆t Gij (t, t ′ ) X

= lim = (Si′ ′ − Si ) Sj′ w (Si ) p S, t|S ′ , t ′ p S ′ , t ′ . (A.4)
∂t δ→0 δ
S,S ′ ,S ′ ′

For Ising spins (Si = ±1) the only contribution is for Si′ ′ = −Si hence, performing the sum
over S ′ ′ one arrives at

∂Gi j (t, t ′ ) X


= −2 Si Sj′ w (Si ) p S, t|S ′ , t ′ p S ′ , t ′
∂t ′ S,S

= −2 ⟨w (Si (t)) Si (t) Sj (t ′ )⟩ . (A.5)

This shows that the derivative of a correlator with respect of the largest time is a (different)
correlator itself.

Let us now consider the limit t ′ → t, where we can make the substitution p S, t|S ′ , t ′ →
δS,S ′ and use this δ-function to perform the S ′ summation. One has
∂Gi j (t, t ′ ) X
lim
′
= −2 Si Sj w (Si ) p (S, t) = −2 ⟨Si (t) Sj (t) w (Si (t))⟩ . (A.6)
t →t ∂t
S

In particular, one has

∂Gii (t, t ′ )
lim
′
= −2⟨w (Si (t))⟩, (A.7)
t →t ∂t
which simply expresses that for each spin flip, which occurs with probability w(Si ), the cor-
relation decreases by two.

13
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

Let us now consider the increment with respect to the smallest time t′ , which is given by
X


∆t′ Gi j (t, t ′ ) = Si Sj′ ′ − Sj′ p S, t|S ′ ′ , t ′ + δ
S,S ′ ,S ′ ′



× p S ′ ′ , t ′ + δ|S ′ , t ′ p S ′ , t ′ . (A.8)
∆Gi j (t,t ′ )
Taking the incremental ratio δ , proceeding as before and performing the limit δ → 0
we get
∂Gi j (t, t ′ ) X



= −2 Si Sj′ w Sj′ p S, t|Fj S ′ , t ′ p S ′ , t ′ , (A.9)
∂t ′ ′ S,S

where we have defined the operator Fj which flips the jth spin, i.e. Fj S ≡
(S1 , S2 , . . . , −Sj , . . . , SN ). This expression shows that the derivative of a correlator with respect
to the smallest time cannot in general be written as a correlator, because of the Fj in the
conditional probability.

Let us now consider the limit t ′ → t, where we can make the substitution p S, t|Fj S ′ , t ′ →
′
δS,Fj S ′ . Using this δ function to perform the S summation in equation (A.9) one has

∂Gi i (t, t ′ )
lim
′
t →t ∂t ′
 P
=
2 P S w (Si ) p(S, t) = 2⟨w(Si (t)⟩, for i = j,
(A.10)
−2 S Si Sj w (Si ) p (S, t) = −2 ⟨Si (t)Sj (t)w(Si (t))⟩, for i ̸= j,

showing that, only for equal times, also this derivative can be expressed in the form of an
average quantity. Comparing with equation (A.6) one has

∂Gij (t, t ′ )  ∂ Gi i (t,t ′ )
− limt→t ′ ∂t , for i = j,
lim′ = ′ (A.11)
t→t ∂t ′  lim ′ ∂ G ij ( t,t )
, ̸ j.
for i =
t→t ∂t

Notice that this result is independent on the detailed balance holding true or not and applies
to intrinsically non-equilibrium models as well. Plugging into equation (16) this shows that
limt′ →t X(t, t ′ ) = 1, while limt′ →t Xij (t, t ′ ) = 0, ∀i ̸= j.

ORCID iDs

F Corberi  https://orcid.org/0000-0002-4763-425X
L Smaldone  https://orcid.org/0000-0002-6738-5537

References

[1] Bray A 1994 Adv. Phys. 43 357–459

[2] Puri S and Wadhaven V 2009 Kinetics of Phase Transitions (Taylor and Francis)
[3] Cugliandolo L F 2015 C. R. Phys. 16 257–66
[4] Bouchaud J P, Cugliandolo L F, Kurchan J and Mézard M 1997 Disordered systems and neural
networks (available at: https://api.semanticscholar.org/CorpusID:117096862)
[5] Corberi F, Cugliandolo L F and Yoshino H 2011 Growing Length Scales in Aging Systems, in:
Dynamical Heterogeneities in Glasses, Colloids and Granular Media (Oxford University Press)
[6] Corberi F 2015 C. R. Phys. 16 332–42
[7] Langer J 1980 Kinetics of metastable states Systems Far From Equilibrium ed L Garrido (Springer)
pp 12–47

14
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

[8] Kubo R, Toda M and Hashitsume N 1991 Statistical Mechanics of Linear Response (Springer) pp
146–202
[9] Cugliandolo L F, Kurchan J and Peliti L 1997 Phys. Rev. E 55 3898–914
[10] Corberi F, Lippiello E and Zannetti M 2004 J. Stat. Mech. 12007
[11] Cugliandolo L F and Kurchan J 1993 Phys. Rev. Lett. 71 173–6
[12] Cugliandolo L F and Kurchan J 1995 Phil. Mag. B 71 501–14
[13] Harada T and Sasa S i 2005 Phys. Rev. Lett. 95 130602
[14] Cugliandolo L F, Dean D S and Kurchan J 1997 Phys. Rev. Lett. 79 2168–71
[15] Franz S, Mézard M, Parisi G and Peliti L 1998 Phys. Rev. Lett. 81 1758–61
[16] Franz S, Mézard M, Parisi G and Peliti L 1999 J. Stat. Phys. 97 459–88
[17] Corberi F and Castellano C 2024 J. Phys. Complex. 5 025021
[18] Corberi F and Smaldone L 2024 Phys. Rev. E 109 034133
[19] Corberi F and Smaldone L 2024 J. Stat. Mech. 053204
[20] Corberi F, dello Russo S and Smaldone L 2024 Phys. Rev. E 110 024143
[21] Lippiello E and Zannetti M 2000 Phys. Rev. E 61 3369–74
[22] Godréche C and Luck J M 2000 J. Phys. A: Math. Gen. 33 1151
[23] Chatelain C 2003 J. Phys. A: Math. Gen. 36 10739
[24] Crisanti A and Ritort F 2003 J. Phys. A: Math. Gen. 36 R181
[25] Lippiello E, Corberi F and Zannetti M 2005 Phys. Rev. E 71 036104
[26] Corberi F, Lippiello E, Sarracino A and Zannetti M 2010 Phys. Rev. E 81 011124
[27] Lippiello E, Corberi F, Sarracino A and Zannetti M 2008 Phys. Rev. E 78 041120
[28] Monthus C and Bouchaud J P 1996 J. Phys. A: Math. Gen. 29 3847
[29] Sollich P 2003 J. Phys. A: Math. Gen. 36 10807
[30] Ritort F 2003 J. Phys. A: Math. Gen. 36 10791
[31] Diezemann G 2005 Phys. Rev. E 72 011104
[32] Diezemann G 2012 Phys. Rev. E 85 051502
[33] Furukawa H 1989 J. Stat. Soc. Japan 58 216
[34] Furukawa H 1989 Phys. Rev. B 40 2341–7
[35] Newman T J and Bray A J 1990 J. Phys. A: Math. Gen. 23 4491
[36] Filipe J A N, Bray A J and Puri S 1995 Phys. Rev. E 52 6082–100
[37] Corberi F and Cugliandolo L F J. Stat. Mech. 11019
[38] Chamon C, Corberi F and Cugliandolo L F 2011 J. Stat. Mech. 08015
[39] Corberi F, Lippiello E and Zannetti M 2002 Phys. Rev. E 65 046136
[40] Corberi F, Nicodemi M, Piccioni M and Coniglio A 1999 Phys. Rev. Lett. 83 5054–7
[41] Emmott C L and Bray A J 1999 Phys. Rev. E 59 213–7
[42] Kimura M and Weiss G H 1964 Genetics 49 561–76
[43] Malécot G, Yermanos D and Collection K M R 1970 The Mathematics of Heredity ed W H Freeman
[44] Clifford P and Sudbury A 1973 Biometrika 60 581–8
[45] Holley R A and Liggett T M 1975 Ann. Probab. 3 643–63
[46] Liggett T 2004 Interacting Particle Systems Classics in Mathematics (Springer)
[47] Cox J T and Griffeath D 1986 Ann. Probab. 14 347–70
[48] Scheucher M and Spohn H 1988 J. Stat. Phys. 53 279–94
[49] Krapivsky P L 1992 Phys. Rev. A 45 1067–72
[50] Frachebourg L and Krapivsky P L 1996 Phys. Rev. E 53 R3009–12
[51] Ben-Naim E, Frachebourg L and Krapivsky P L 1996 Phys. Rev. E 53 3078–87
[52] Sood V and Redner S 2005 Phys. Rev. Lett. 94 178701
[53] Sood V, Antal T and Redner S 2008 Phys. Rev. E 77 041121
[54] Castellano C, Fortunato S and Loreto V 2009 Rev. Mod. Phys. 81 591–646
[55] Glauber R J 1963 J. Math. Phys. 4 294–307
[56] Corberi F and dello Russo S 2024 Ordering kinetics with long-range interactions: interpolating
between voter and ising models (arXiv:2404.06917)
[57] Corberi F, Lippiello E and Zannetti M 2006 Phys. Rev. E 74 041106
[58] Corberi F, Castellano C, Lippiello E and Zannetti M 2004 Phys. Rev. E 70 017103
[59] Corberi F, Lippiello E and Zannetti M 2003 Phys. Rev. Lett. 90 099601
[60] Corberi F, Lippiello E and Zannetti M 2005 Phys. Rev. E 72 056103
[61] Corberi F and Cugliandolo L F 2009 J. Stat. Mech. 09015

15
J. Phys. A: Math. Theor. 57 (2024) 435001 F Corberi and L Smaldone

[62] Lippiello E, Corberi F and Zannetti M 2006 Phys. Rev. E 74 041113

[63] Baity-Jesi M et al 2017 Proc. Natl Acad. Sci. USA 114 1838–43
[64] Franz S and Rieger H 1995 J. Stat. Phys. 79 749
[65] Marinari E, Parisi G, Ricci-Tersenghi F and Ruiz-Lorenzo J J 1998 J. Phys. A: Math. Gen. 31 2611
[66] Marinari E, Parisi G, Ricci-Tersenghi F and Ruiz-Lorenzo J J 2000 J. Phys. A: Math. Gen. 33 2373
[67] Cruz A, Fernández L A, Jiménez S, Ruiz-Lorenzo J J and Tarancón A 2003 Phys. Rev. B 67 214425
[68] Corberi F, Candia A d, Lippiello E and Zannetti M 2002 Phys. Rev. E 65 046114

16