
2
tures q↔βbijectively thanks to positivity of the static
covariance matrix Cij =−∂qi/∂βj.
Euler hydrodynamics asserts that the slow variation in
space of the state (1) induces a corresponding slow vari-
ation in time, such that the state keeps the local equi-
librium form; one may propose that at all times (1) cor-
rectly describes the states, with βi
ini(x)→βi(x, t) (see,
e.g., [6,25,27]). In (1), correlations vanish exponentially
with the distance, under very broad conditions including
non-zero temperature (finite βi’s) and the lack of phase
transition [28,29] (which we assume here). Thus, spa-
tial correlations should vanish exponentially, even during
time evolution; indeed each fluid cell is at “local equilib-
rium”, and entropy maximisation should occur indepen-
dently in every fluid cell [19,23,30,31].
In fact, certain long-range, algebraic correlations are
known to emerge in non-equilibrium situations when con-
servation laws are present. This is well studied for dif-
fusive systems in non-equilibrium steady states (NESS)
[32–34]: unbalanced thermostats at the system’s bound-
aries lead to nonzero gradients, and correlations between
conserved densities at macroscopic distances decay as
(system size)−1. This is due to the breaking of detailed
balance at the diffusive scale and determined by viscous
coefficients, and may be quantitatively described by fluc-
tuating hydrodynamics and macroscopic fluctuation the-
ory (see, e.g., [33–35]). But what happens at the Euler
scale, where viscous effects are scaled down to zero size?
In NESS emerging from the partitioning protocol in
systems of infinite size [36], gradients vanish and corre-
lations are weaker. The strongest are found in integrable
systems, including free particles, where conserved density
correlations decay as (distance)−2because of discontinu-
ities in the occupation function of hydrodynamic modes
[37,38]. But this decay is too quick to correlate Euler-
scale fluid cells (see below).
We note that a similar situation occurs at zero temper-
ature, under the different physics of quantum fluctuations
at Fermi points, and that a theory for the transport of
such weak algebraic correlations on top of moving fluids
is proposed in [30,39] (in GHD). Very far from equilib-
rium, stronger long-range correlations may develop: for
instance, global quantum quenches generate finite den-
sities of entangled particles that may propagate (diffu-
sively or ballistically) and carry nontrivial entanglement
[40–42] and correlations [43,44]. But entangled particle
production is not expected to occur in long-wavelength
states, Eq. (1).
Up to now, there has been no prediction, observa-
tion or theory for eventual long-range correlations emerg-
ing under ballistic scaling from (1). The assumption of
uncorrelated Euler-scale fluid cells, and that the form
(1) stays valid in time, has remained, and appears to
play an important role in recent studies of the evolu-
tion of correlations and fluctuations under inhomoge-
neous conditions and nonlinear hydrodynamic response
theory [19,23,30,39,45].
In this manuscript, we show that the assumption of
uncorrelated Euler-scale fluid cells is generically incor-
rect. We show that correlations of strength ∝ℓ−1develop
dynamically from (1), at macroscopic (∝ℓ) times and
distances, under generic conditions for systems exhibit-
ing ballistic transport. In particular, if QR
i(ℓt), QR′
j(ℓt)
are total charges lying on finite but macroscopically
large regions R, R′that are separated by a macro-
scopic distance, |R|,|R′| ∝ ℓ, dist(R, R′)∝ℓ, evaluated
at macroscopic time ℓt, then their covariance is large,
⟨QR
i(ℓt)QR′
j(ℓt)⟩c∝ℓ. This shows strong correlations
between separated cells. The appearance of ballistically
scaled long-range correlations at all macroscopic times is
a general phenomenon which, to our knowledge, has not
been discussed before. It holds no matter the nature of
the system, quantum or classical, integrable or not, and
is solely controlled by its Euler hydrodynamics.
This phenomenon is not explained by the theories for
diffusive long-range correlations recalled above, as it does
not depend on viscous coefficients or phenomenological
noise, and occurs in ballistic times t∝x. It gives the
dominant correlations on large distances, beyond diffu-
sive broadening and of larger strength than the 1/x2cor-
relations due to occupation discontinuities. It is not due
to quasi-particle excitations, as it is a universal hydrody-
namic effect. By contrast, we show that the phenomenon
occurs in long-wavelength inhomogeneous state (as in
(1)), only if the Euler hydrodynamic theory is interact-
ing, and only if it admits at least two different fluid modes
(with different velocities). Euler-scale long-range correla-
tions invalidate the assumption that on every time-slice a
state such as (1) is found. This thus calls for a new under-
standing of the principles of Euler hydrodynamics, and a
re-think of recent studies of hydrodynamic nonlinear re-
sponse and the evolution of correlations and fluctuations.
We quantify this phenomenon by proposing that the
principle replacing independent local entropy maximisa-
tion of fluid cells is that of relaxation of fluctuations: lo-
cal observables relax to fixed, non-fluctuating functions
of conserved densities, which themselves fluctuate. This
is developed into a universal theory, the ballistic macro-
scopic fluctuation theory (BMFT). The BMFT is a hy-
drodynamic large-deviation theory, solely based on the
emergent Euler hydrodynamic data of the model, which
characterises all fluctuations and correlations at the bal-
listic hydrodynamic scale, including under fluid motion.
For illustration, we study the paradigmatic hard-rod
model of statistical physics, which is simple enough to
be amenable to high-accuracy numerical simulations, yet
truly interacting. We find that the model does indeed
develop long-range correlations, which are quantitatively
in excellent agreement with our theory.
Ballistic long-range correlations.— We show that
correlations in the initial state (1), between macroscopi-
cally separated observables ˆo1(ℓx1, ℓt) and ˆo2(ℓx2, ℓt) at
macroscopic times, generically has strength ℓ−1. That
is, the connected correlation function ⟨ˆo1ˆo2⟩c:= ⟨ˆo1ˆo2⟩ −