Annurev Conmatphys 031720 032419

Annual Review of Condensed Matter Physics
Irreversibility and Biased

Ensembles in Active Matter:
Insights from Stochastic
Access provided by 2a01:e0a:840:f510:6066:683d:4413:fca5 on 03/17/23. For personal use only.
Annu. Rev. Condens. Matter Phys. 2022.13:215-238. Downloaded from www.annualreviews.org
Thermodynamics
Étienne Fodor,1 Robert L. Jack,2,3
and Michael E. Cates2
1
Department of Physics and Materials Science, University of Luxembourg, Luxembourg;
email: [email protected]
2
Department of Applied Mathematics and Theoretical Physics, University of Cambridge,
Cambridge, United Kingdom; email: [email protected], [email protected]
3
Yusuf Hamied Department of Chemistry, University of Cambridge, Cambridge, United
Kingdom
Annu. Rev. Condens. Matter Phys. 2022. 13:215–38 Keywords

First published as a Review in Advance on
self-propelled particles, nonequilibrium field theories, dissipation, entropy
November 10, 2021
production, large deviations, phase transitions
The Annual Review of Condensed Matter Physics is
online at conmatphys.annualreviews.org Abstract
https://doi.org/10.1146/annurev-conmatphys-
Active systems evade the rules of equilibrium thermodynamics by constantly
031720-032419
dissipating energy at the level of their microscopic components. This energy
Copyright © 2022 by Annual Reviews.
flux stems from the conversion of a fuel, present in the environment, into sus-
All rights reserved
tained individual motion. It can lead to collective effects without any equi-
librium equivalent, some of which can be rationalized by using equilibrium
tools to recapitulate nonequilibrium transitions. An important challenge is
then to delineate systematically to what extent the character of these active
transitions is genuinely distinct from equilibrium analogs. We review recent
works that use stochastic thermodynamics tools to identify, for active sys-
tems, a measure of irreversibility comprising a coarse-grained or informatic
entropy production. We describe how this relates to the underlying energy
dissipation or thermodynamic entropy production, and how it is influenced
by collective behavior. Then, we review the possibility of constructing ther-
modynamic ensembles out of equilibrium, where trajectories are biased to-
ward atypical values of nonequilibrium observables. We show that this is a
generic route to discovering unexpected phase transitions in active matter
systems, which can also inform their design.
215
1. INTRODUCTION
Active matter is a class of nonequilibrium systems whose components extract energy from the envi-
ronment to produce an autonomous motion (1–3). Examples are found in biological systems, such
as swarms of bacteria (4) and assemblies of cells (5); social systems, such as groups of animals (6)
and human crowds (7); and synthetic systems, such as vibrated polar particles (8) and catalytic
colloids (9). The combination of individual self-propulsion and interactions between individuals
can lead to collective effects without any equivalent in equilibrium. Examples include collective
directed motion, as observed in bird flocks (6), and the spontaneous formation of clusters made
of purely repulsive particles without even a depletion interaction, as reported for Janus colloids
in a fuel bath (9). To study these effects, minimal models have been proposed based on simple
dynamical rules. Some are formulated at particle level, for instance with automaton rules (10) or
by extending Langevin dynamics (11). Others work at the coarse-grained level in terms of hydro-
dynamic fields (12, 13). The latter can be obtained systematically by coarse-graining the particle
dynamics or postulated phenomenologically. Both the microscopic and hydrodynamic approaches
have successfully reproduced experimental behavior, such as the emergence of a long-ranged polar
order, known as the flocking transition (14), and phase separation that occurs without any micro-
scopic attraction, known as motility-induced phase separation (MIPS) (15).
Although active systems evade the rules of equilibrium statistical mechanics, some works have
built a framework to predict their properties based on the partial applicability of thermodynamic
concepts beyond equilibrium. A first approach was to map some active systems onto equilibrium
ones with a similar steady state, allowing one to define effective free energies for active matter (16,
17). Other studies have extended the definitions of standard observables such as pressure (18–20),
surface tension (21, 22), and chemical potential (23, 24), hoping to establish equations of state
relating, for instance, pressure and density. Interestingly, one finds that the existence of such state
functions cannot generically be relied upon (e.g., the pressure on a wall can depend on the type
of wall) (20), highlighting the limitations of equilibrium analogies. In trying to build a thermody-
namic framework for active matter, an important challenge is then to identify regimes in which it
is possible to deploy equilibrium tools and to clearly distinguish these from genuine nonequilib-
rium regimes. In other words, how should we delineate where and when activity really matters in
the emerging phenomenology? And, most importantly, can we define a systematic, unambiguous
measure of the departure from equilibrium?
In passive systems, the steady state (Boltzmann) distribution involves the Hamiltonian that also
drives the microdynamics. This ensures thermodynamic consistency of the dynamics, which fur-
ther implies that the mechanical and thermodynamic definitions of pressure are equivalent, and
precludes (real space) steady-state currents. Although such currents offer a clear nonequilibrium
signature of activity, as observed, for instance, in collective motion (14), it is more challenging to
distinguish, say, MIPS from standard phase separation without tracking individual particle mo-
tion (15). In particular, it is not helpful to define departure from equilibrium via deviation from an
effective Boltzmann distribution in systems that are not thermodynamically consistent. Instead,
the cornerstone of modern statistical mechanics, which allows one to dissociate fundamentally
active and passive systems, is the reversibility of equilibrium dynamics (25). In a steady equilib-
rium state, forward and backward dynamics are indistinguishable, because all fluctuations exhibit
time-reversal symmetry (TRS). This constraint entails other important properties, such as the
fluctuation–dissipation theorem (FDT) (26), and the absence of dissipated heat at equilibrium.
For active systems, the irreversibility of the dynamics then stands out as the key differentiating
property that causes the violation of these and other equilibrium laws.
To quantify the breakdown of TRS, we use stochastic thermodynamics (27, 28). This frame-
work extends standard notions of thermodynamics, such as the first and second laws, to fluctuating
216 Fodor • Jack • Cates

trajectories. It was first developed for systems in contact with one or more reservoirs, such as en-
ergy and/or volume reservoirs, provided that each reservoir satisfies some equilibrium constraints.
Such constraints do not preclude the system from operating out of equilibrium, for instance, under
external fields and/or thermostats at different temperatures: In this context, the thermodynamic
consistency of the dynamics yields some explicit connections among irreversibility, dissipated heat,
and entropy production. In contrast, the dynamics of active systems are often formulated via phe-
nomenological arguments, which do not follow a priori the requirements of thermodynamic con-
sistency. A natural question is, then, Is it legitimate to extend the methods of stochastic thermo-
dynamics to active matter? And, what can we learn from their irreversibility measures when their
connection with other thermodynamic observables becomes blurred?
Another interesting direction in active matter theory is to generalize familiar thermodynamic
ensembles to nonequilibrium settings. Here, the important roles of steady-state currents and other
types of irreversible dynamics make it insufficient to study ensembles of configurations (charac-

terized by a stationary measure, of which the Boltzmann distribution is an example). Instead, we
must address ensembles of trajectories, as used previously to analyze fluctuation theorems (29), and
other dynamical effects (30–32). Such ensembles are built similarly to the canonical ensemble of
statistical mechanics, with relevant (extensive) observables coupled to conjugate (intensive) fields.
In practice, one selects a dynamical observable of interest and focuses on dynamical trajectories
where it has some atypical target value. The resulting biased ensembles of rare trajectories are
intrinsically connected to large deviation theory (33, 34). As such, they provide insight into mech-
anisms for unusual and interesting fluctuations. In the passive context, they have proven useful for
unveiling dynamical transitions, e.g., in kinetically constrained models of glasses (32). This leads
one to ask the following question: How do the transitions in biased ensembles of active matter
differ from their passive counterparts? And then, can we identify settings in which bias-induced
transitions emerge in active matter that have no passive equivalent?
In what follows (Section 2), we describe how the irreversibility of active systems can be quan-
tified, for both particle-based and hydrodynamic theories, using generalized forms of entropy
production. We discuss how to relate these, in certain cases, to energy dissipation and how they
allow one to identify phases and/or spatial regions in which activity comes to the fore. In Section 3,
we then describe how biased ensembles provide novel insights into the emergence of collective
effects in active matter. We review some useful tools of large deviation theory, such as representing
dynamical bias in terms of control forces, show the utility of these methods in an active context,
and discuss how they can reveal unexpected transitions in generic active systems. In Section 4, we
give a brief conclusion that includes the implications of these methods for the rational design of
new materials.
2. IRREVERSIBILITY AND DISSIPATION

2.1. Modeling Active Matter: From Particles to Hydrodynamic Fields
Typical particle-based dynamics start from the seminal Langevin equation. Provided that inertia
is negligible, this balances the forces stemming from the heat bath (damping and thermal noise),
the force deriving from a potential U, and nonconservative forces fi :

ṙi = μ fi − ∇iU ) + 2μT ξ i , 1.
where μ is the mobility, T the temperature of the bath, and ξ i a set of zero-mean, unit-variance
white Gaussian noises: ξ iα (t)ξ jβ (t ) = δ ij δ αβ δ(t − t ), where Latin and Greek indices, respectively,
refer to particle labels and spatial coordinates. Hereafter, we refer to such noise as unit white
noise. The potential U describes particle interactions and/or an external perturbation applied by
www.annualreviews.org • Stochastic Thermodynamics: Active Matter 217

the operator. The nonconservative forces fi model self-propulsion, which converts a source of
energy, present in the environment, into directed motion. In principle, fi can include external
perturbations that do not derive from a potential. When fi = 0, the system is at equilibrium with
Boltzmann statistics (∝e−U/T ).
The self-propulsion force fi is often modeled as, effectively, an additional noise. In contrast with
the thermal noise ξ i , it has some persistence, which captures the propensity of active particles to
sustain directed motion. Typically, a persistence time τ sets the exponential decay of the two-point
correlations: fiα (t ) f jβ (0) = δi j δαβ f02 e−|t|/τ . In recent years, two variants have emerged as popu-
lar models of self-propelled particles. First, for active Ornstein–Uhlenbeck particles (AOUPs),
the statistics of fi is Gaussian, so that it can be viewed as obeying an (autonomous) Ornstein–
Uhlenbeck process (17, 35):
√
AOUP: τ ḟi = −fi + f0 2τ ζ i , 2.

where the ζ i are unit white noises. Scaling the amplitude as μ f0 = μTa /τ , the self-propulsion
itself converges to a white noise source in the limit of vanishing persistence, in which regime the
system reduces to passive particles at temperature T + Ta . Second, for active Brownian particles
(ABPs), the magnitude of fi is now fixed, which leads to non-Gaussian statistics, described by the
following process in two dimensions (11):

ABP: fi = f0 (cos θi , sin θi ), θ˙i = 2/τ ηi , 3.
where ηi is a unit white noise. In Equations 2 and 3, the self-propulsion is independent for each
particle. This corresponds to isotropic particles, which undergo MIPS for sufficiently large per-
sistence and density (15). In contrast, many models consider alignment interactions, yielding a
collective (oriented) motion at high density and enhanced persistence (14).
To study the emergence of nonequilibrium collective effects, it is helpful to introduce the rele-
vant hydrodynamic fields. These can be identified, and their dynamical equations found, by coarse-
graining the microscopic equations of motion using standard tools of stochastic calculus (36).
Another approach consists of postulating such hydrodynamic theories from phenomenological
arguments, respecting any conservation laws and spontaneously broken symmetries (1). For in-
stance, a theory of MIPS can be found by considering the dynamics of a scalar field φ(r, t) that
encodes the local concentration of particles (13, 37):
√
δF
φ̇ = −∇ · J = −∇ · − λ∇ + Jφ + 2λDφ , 4.
δφ
where λ is collective mobility, D is a noise temperature, and φ is (spatiotemporal) unit white noise.
The term Jφ embodies contributions driving the current J of φ that cannot be derived from any
free energy functional F as −λ∇(δF/δφ). For instance, choosing a standard φ 4 , square-gradient
functional for F captures phase separation in equilibrium. On introducing activity, symmetry ar-
guments require that Jφ takes the form (φ)2 and/or (φ)(2 φ) to lowest order in φ and its
gradients (13, 37). The former term shifts the coexisting densities with respect to the equilibrium
model (13), whereas the latter can lead to microphase separation, with either the vapor phase
decorated with liquid droplets or the liquid phase decorated with vapor bubbles (37). Note that,
although the conservative term −λ∇(δF/δφ) captures all the contributions to the coarse-grained
dynamics that respect TRS, it may depend on nonequilibrium parameters of the microscopic dy-
namics (37).
The dynamics in Equation 4 can also be coupled to a polar field p(r, t), representing the local
mean orientation of particles, to study, for instance, collective flocking motion (14):
δF √
ṗ = − + J p + 2D p , 5.
δp

where the rotational mobility is taken as unity, and p is another spatiotemporal unit white noise.
As with Jφ , the term Jp represents active relaxations that cannot be written as free-energy deriva-
tives. Note that, for the passive limit of the coupled model to respect detailed balance, the free
energy F in Equation 5 must be the same as that in Equation 4, and the value of D in each equation
must also coincide. In the absence of activity ( Jφ = Jp = 0), the system has Boltzmann statistics
∝e−F/D . To capture the emergence of polar order, a minimal choice is to take F as a p4 functional,
with Jφ proportional to p, and Jp a linear combination of φ, φp, and (p ·)p (1, 14). This can
yield polar bands traveling across an apolar background, which is a key signature of aligning ac-
tive particles models (10). In fact, much of the physics of flocks is retained by setting λ = 0 in
Equation 4 (12). By contrast more complicated theories, describing the emergence of nematic
order (14) and/or the coupling to a momentum-conserving fluid (38, 39), introduce additional
hydrodynamic fields beyond φ and p as considered above.
2.2. How Far from Equilibrium Is Active Matter?

We present various approaches to define time reversal, and to quantify the corresponding TRS
breakdown, in popular models of active matter. This leads us to propose several measures of the
departure from equilibrium.
2.2.1. Forward and time-reversed dynamics: path-probability representation. To quantify

the irreversibility of active dynamics, we compare the path probability P to realize a dynamical
trajectory (across a time interval [0, t]) with that for its time-reversed counterpart (27, 28, 40).
Choosing the time-reversed dynamics is a matter of subtlety (35, 39, 41–49). For the particle-based
dynamics in Equation 1, exploiting the known statistics of the (unit white) noise, trajectories of
position ri and self-propulsion fi have probability P ∼ e−A (50), with action
t
1
2
A= ṙi + μ ∇iU − fi dt + A f . 6.
4μT 0 i
We use Stratonovich discretization, omitting a time-symmetric contribution that is irrelevant in
what follows. A f specifies the statistics of the self-propulsion force fi . For AOUPs and ABPs, fi has
autonomous dynamics, independent of position ri . One can then in principle integrate out that
dynamics to obtain a reduced action Ar . So far, the explicit expression has been derived only for
AOUPs (35, 48, 51, 52):
t t
Ar = dt dt ṙi + μ∇iU t · ṙi + μ∇iU t (t − t ). 7.
0 0 i
The kernel is an even function of time t, which depends on μ, T, f0 , and τ . It reduces to

(t) = δ(t)/[4μ(T + Ta )] for vanishing persistence (τ → 0), when the self-propulsion amplitude
scales as μ f0 = μTa /τ (48), as expected for passive particles. The actions A and Ar describe the
same dynamics from different viewpoints. The former tracks the realizations of position and self-
propulsion, whereas the latter follows the time evolution of positions only. Thus, Ar treats the
self-propulsion as a source of noise, without resolving its time evolution. In contrast, A regards fi
as a configurational coordinate, which might in principle be read out from the particle shape. For
instance, Janus particles, which are spheres with distinct surface chemistry on two hemispheres (2,
9), generally self-propel along a body-fixed heading vector pointing from one to the other (see
also Figure 1).
To compare forward and backward trajectories, one needs to define a time-reversed version
of the dynamics. This necessarily transforms the time variable as t → t − t , and it can also in-
volve flipping some of the variables or fields (as detailed below). For particle-based dynamics,

Orientation reversal
a b
Time reversal
c d
Figure 1
TR for active particles (fish). Each has an orientation (head–tail axis) and moves (swims) as indicated by the
wake lines behind it. The TR operation may or may not reverse orientations. (a) Natural dynamics.
(b) Reversing orientation but not motion: fish swim tail first along their original directions. (c) Reversing
motion but not orientation: fish swim tail first, opposite to their original direction. (d) Reversing both
motion and orientation: fish swim headfirst, opposite to their original direction. The situations in panels b
and c occur with extremely low probability in the natural dynamics (requiring exceptional noise realizations).
In dilute regimes (not shown) the case in panel d is equiprobable to that shown in panel a: Individual fish
swim headfirst in both cases. For the shoals shown here, however, the case in panel d is less probable than
that in panel a, because the natural dynamics has more fish at the front of the shoal: time-reversal symmetry
is broken at a collective level even if orientations are reversed. Abbreviation: TR, time reversal.
time reversal of the positions ri is unambiguous. Writing P in terms of the reduced action Ar in
Equation 7, the time-reversed counterpart ARr reads as
Time reversal: several

t t
backward dynamics ARr = dt
dt ṙi − μ∇iU t · ṙi − μ∇iU t (t − t ) . 8.
that can be defined 0 0 i
depending on the time
signature of the Here, we have changed variables as {t , t } → {t − t , t − t }, and used the fact that is even. For the
degrees of freedom
full action A, defining the time-reversed dynamics requires that we fix the time signature of the
self-propulsion fi . Indeed, there are two possible reversed actions, AR,± , where + and − indices

refer, respectively, to even (fi → fi ) and odd (fi → −fi ) self-propulsion:
t
1
2
AR,± = ṙi − μ ∇iU ∓ fi dt + A f . 9.
4μT 0 i Dissipated heat:
the amount of energy
We assumed that A f is the same for forward and time-reversed dynamics, which holds for ABPs transferred from the
and AOUPs (but not for systems with aligning interactions; 53). ARr and AR,± correspond to three system to the
different definitional choices of the time-reversed dynamics. The first of these deliberately dis- surrounding
thermostat due to
cards the realizations of the self-propulsion forces in order to compare the forward and backward internal active
motions of particle positions. This choice is widely used for AOUPs, where particles have no head- processes
ing vector. It is also the only possible choice when T = 0 in Equation 1 (see 35). In contrast, AR,± ,
Informatic entropy
which for each we assume that both position and self-propulsion are tracked, are usually chosen production:
for particles with a heading vector (45, 53, 54).
a quantification of the
For the field theories of Section 2.1, defining the time-reversed dynamics requires us to choose breakdown of
the time signature of each field. This choice depends on the physical system and the phases under time-reversal
scrutiny (Section 2.4). Scalar fields associated with a local density are clearly even (13, 39), whereas symmetry via the
difference between
scalar fields schematically representing polarization (55), or stream functions for fluid flow (56), forward and backward
should generally be odd. Similarly, the vector field p can be chosen even (57) or odd (47), de- path probabilities
pending on whether it is viewed as the local orientation of particles (mean heading vector; see
Figure 1) or directly as a local velocity (12). One also has to decide which fields to retain or ig-
nore when comparing forward and backward paths. This choice is the counterpart at field level
of retaining (A) or ignoring (Ar ) the propulsive forces in a system of AOUPs. For example, in a
system with φ and p variables, one can choose whether or not to keep separate track of the density
current J alongside φ and p.
In general, there are four different versions of time-reversed dynamics of Equations 4 and 5,
provided that each one of the fields φ and p can be either odd or even. Often, though, constraints
on the time signatures of the fields restrict these choices. To ensure that the dynamics is invariant
under time reversal in the passive limit, the free energy F must be the same in forward and time-
reversed dynamics. This ensures that odd fields can only appear as even powers in F. Hence, if p
appears linearly in F, which for liquid crystal models often includes a p · φ or anchoring term,
it must be chosen even. Further constraints arise if some of the possible noise terms are set to zero
(e.g., 12) so that certain fields are deterministically enslaved to others. For instance, if φ̇ = −∇ · p
(without noise), for consistency φ and p necessarily have different signatures.
2.2.2. Distance from equilibrium: breakdown of time-reversal symmetry. For a given dy-
namics, once its time-reversed counterpart has been chosen, we can define systematically an irre-
versibility measure S as (27, 28, 40)

1 P 1
S = lim ln R = lim AR − A . 10.
t→∞ t P t→∞ t
The average · is taken over noise realizations. The limit of long trajectories gets rid of any tran-
sient relaxations to focus on steady-state fluctuations. Equation 10 is a cornerstone of stochastic
thermodynamics (27, 28, 40). It was first established in thermodynamically consistent models (40),
where S can be shown to be the entropy production rate (EPR) governing the dissipated heat (27,
28). In models in which S has this meaning, thermodynamics constrains the choice of time rever-
sal. In Section 2.3, we return to the question of how far the thermodynamic interpretation extends
to active systems. Meanwhile, S already offers an unambiguous measure of TRS breakdown in ac-
tive matter, which we refer to as informatic entropy production rate (IEPR). Because S changes

when the dynamics is coarse-grained, by eliminating unwanted variables, it differs when retaining
(A) or ignoring (Ar ) the propulsive forces.
In particle-based dynamics, choosing the action Ar so that only particle positions are tracked,
the corresponding IEPR Sr follows from Equations 7, 8, and 10 as
∞
Sr = −4μ (t ) ṙi (t ) · ∇iU (0) dt, 11.
−∞ i
where we have again used the fact that is even. The IEPR Sr vanishes in the absence of any poten-
tial, showing that the dynamics satisfies TRS for free active particles and also for an external har-

monic potential U ∼ ri2 (35, 52). When T = 0, it reduces to Sr = τ ( i ṙi · ∇i )3U /[2(μ f0 )2 ] (35).
When (by using the full action A) one follows the dynamics of both position and self-propulsion,
the two possible IEPRs found from Equations 6, 9, and 10 are
1 μ
S+ = ṙi · fi , S− = ∇iU · fi . 12.
T i T i
t
We have used the fact that 1t 0 i ṙi · ∇iU dt = U (t )−U
t
(0)
vanishes at large t. Substituting the ex-
pression for ṙi from Equation 1 into Equation 12 yields S + +S − = N μ f02 /T , where N is the par-
ticle number, using that self-propulsion fi and thermal noise 2μT ξ i are uncorrelated. Therefore,
in the absence of any potential U, S − vanishes identically, whereas S + remains nonzero. Indeed,
S − compares trajectories whose velocity and self-propulsion both flip on time reversal, retaining
alignment (up to thermal noise) between the two: The forward and backward dynamics are indis-
tinguishable unless potential forces intervene. In contrast, S + quantifies how different trajectories
are when particles move either along with or opposite to their self-propulsion (Figure 1), yielding
the contribution N μ f02 /T even when U = 0.
The IEPR S + is lowest (and S − highest) when the propulsive force fi balances the interaction
force −i U so that particles are almost arrested. Hence, both IEPRs are sensitive to the formation
of particle clusters, which is associated with dynamical slowing-down for isotropic particles (15),
and to formation of a polarized state for aligning particles (14). Note that S + is proportional to
the contribution of self-propulsion to pressure, known as swim pressure (19, 20). Finally, in the
presence of alignment, any dynamical interaction torques appear explicitly in S (53). We defer
further discussion on how interactions shape irreversibility, for both particle-based dynamics and
field theories, to Section 2.4.
2.3. Energetics Far from Equilibrium: Extracted Work and Dissipated Heat
A major success of stochastic thermodynamics is to extend the definition of observables from
classical thermodynamics to cases in which fluctuations cannot be neglected (27, 28, 40). Although
first proposed for thermodynamically consistent models, this approach can be extended to some
(not all) types of active matter.
2.3.1. Particle-based approach: energy transfers in microscopic dynamics. For particle dy-
namics, assuming that the potential U depends on the set of control parameters α n , the work W
produced by varying α n during a time t is (27, 28)
t
∂U
W= α̇n dt . 13.
0 n
∂α n
Note that W is stochastic due to the thermal noise and the self-propulsion, even though the proto-
col α n (t) is deterministic. Equation 13 relies on the precept that some external operator perturbs

the system though U, without prior knowledge of the detailed dynamics; it applies equally for
active and passive systems.
The heat Q is now defined as the energy delivered by the system to the surrounding thermostat.
For the dynamics of Equation 1, the effect of the thermostat is encoded in the damping force and
the thermal noise; the fluctuating observable Q then follows as (27, 28)
t
ṙi
Q= · ṙi − 2μT ξ i dt . 14.
0 i
μ

Substituting Equation 1 into Equation 14, and using the chain rule U̇ = n α̇n (∂U/∂αn ) + i ṙi ·
∇iU , gives a relation among energy U, work W, and heat Q:
t
U (t ) − U (0) = W − Q + ṙi · fi dt . 15.

0 i
In passive systems, the nonconservative force fi represents some intervention by the external op-
t
erator (beyond changes in U), so that the term W̄ ≡ 0 i ṙi · fi dt can be absorbed into the work
W. Then, Equation 15 is the first law of thermodynamics (FLT) (27, 28). The only possible time
reversal for passive dynamics is to choose S + in Equation 12, so that S + = Q̇/T . In this context,
S + coincides with the thermodynamic EPR due to contact of the system with the thermostat,
which connects explicitly irreversibility, entropy production, and dissipation (28). In contrast, for
active dynamics, the contribution W̄ captures the energy cost to sustain the self-propulsion of
particles (58, 59), which is generally distinct from both W and Q. When the potential is static
(α̇n = 0), S + = Q̇/T still holds, yet S + should not be interpreted as a thermodynamic EPR in
general.
Although these considerations offer one consistent approach to the question of energy trans-
fers for active particles, several other approaches are possible. Interestingly, one alternative def-
inition of heat relies on replacing ṙi in Equation 14 by ṙi − μfi . This amounts to regarding the
self-propulsion as being caused by a locally imposed flow, with particle displacements evaluated
in the flow frame (43, 48). It yields vanishing heat in the absence of interactions, for the same
reasons as led us to zero S − in Equation
12. Also, some works addressing
heat engines (60–62)
have redefined heat by replacing 2μT ξ i in Equation14 with 2μT ξ i + μfi , thus considering
the self-propulsion fi as a noise with similar status to 2μT ξ i . This yields vanishing heat when
the potential is static (α̇n = 0) by discarding all the energy dissipated by self-propulsion, allow-
ing one to reinstate the FLT, U (t ) − U (0) = W − Q. Finally, the angular diffusion of ABPs is
as thermal. This gives an angular contribution to Q, which is proportional
often itself regarded
t
to 0 i θ˙i (θ˙i − 2/τ ηi )dt , which vanishes for Equation 3 but is generally nonzero when align-
ing interactions are present (59). For AOUPs, interpreting the self-propulsion dynamics in terms
of thermal damping and noise is less straightforward, although some studies have taken such a
path (41, 42, 63).
Importantly, Equation 1 does not resolve how particles convert fuel into motion. Accordingly,
Equation 14 only captures the energy dissipated by the propulsion itself, ignoring contributions
from underlying, metabolic degrees of freedom. Various schematic models describe the underlying
chemical reactions in thermodynamically consistent terms, maintaining them out of equilibrium
by holding constant the chemical potential difference between products and reactants (43, 46, 64).
For some of these models (43, 46), the time evolution of ri can be mapped into Equation 1 when
a chemical noise parameter is small (58). In this case, the difference between the partial and total
heat, found by discarding or including reactions, is a constant that is independent of the potential
U. In a more refined model, the dynamics tracks the time evolution of chemical concentration,

and the heat features explicitly the chemical current (64). Related features are discussed next, for
continuum fields.
Linear irreversible
thermodynamics 2.3.2. Coarse-grained perspective: energy transfers at hydrodynamic level. For field theo-
(LIT): forces and ries, assuming that the free energy F depends on some set of control parameters α n , the work W
currents that are is defined by analogy with Equation 13 as
linearly coupled, and t
∂F
the dissipated heat is W= α̇n dt . 16.
given in terms of their 0 n
∂α n
product This holds regardless of any nonequilibrium terms in Equations 4 and 5 from the case of pas-
sive to active fields. In contrast, to define heat at a hydrodynamic level, one cannot depend a
priori on a physical interpretation of the various dynamical terms as we made for particles. A
minimal requirement, for emergence of the FLT along the lines of Equation 15, is that F in
Equation 16 is a genuine free energy, which stems from coarse-graining only the passive con-
tributions of the microscopic dynamics. When theories are based solely on phenomenological
arguments, this thermodynamic interpretation is absent.
In some cases, an active field theory ought to allow a thermodynamic interpretation, e.g., for the
formation of membraneless organelles (65), but not when the same theory describes social systems,
e.g., the demixing of animal groups. In the latter, not only active terms but F itself originate in
behavioral rules, which do not produce any mechanical work, so that there is no meaningful FLT.
In the former, to reinstate a thermodynamic framework, one can include chemical fields describing
the fuel consumption underlying activity (57). This relies on linear irreversible thermodynamics
(LIT) (66), as previously used to illuminate particle-based dynamics (64). LIT postulates linear
relations between thermodynamic fluxes and forces, whose product determines the heat Q. A class
of active gel models was indeed first formulated in this way (67, 68). In contrast, the field theories
in Equations 4 and 5 were not derived from LIT a priori, yet they can be embedded within it in a
consistent manner.
To illustrate this, consider a scalar field φ obeying Equation 4. The conservation law φ̇ = −∇ · J
relates this to the thermodynamic flux J, whose conjugate force is −∇(δF/δφ). Likewise, a ther-
modynamic flux ṅ describes some metabolic chemical process, with conjugate force a chemical
potential difference μ. Out-of-equilibrium dynamics is maintained by holding either ṅ or μ
constant (47). Considering here the latter case, LIT requires that the active current Jφ be linear in
μ (although, like J, it can be nonlinear in φ), so that we identify Jφ / μ C as an off-diagonal
Onsager coefficient (57), yielding

δF
J, ṅ = L − ∇ , μ + noise terms, 17.
δφ
with an Onsager matrix obeying LJJ = λI (with I the d-dimensional identity); Lṅṅ = γ , a chemical
mobility (such that ṅ = γ μ in the absence of φ dynamics); and LJṅ = LṅJ = C. This last result
encodes the famous Onsager symmetry, which stems from the underlying reversibility of LIT (25),
so that the form of the active current Jφ also controls the φ coupling in the equation for ṅ. For
Jφ = μ[(φ)2 ] (13), then ṅ = γ μ − ∇(δF/δφ) · ∇[(∇φ)2 ]. The covariance of the noise terms
in Equation 17 is set directly by L. The off-diagonal noise depends on φ through C and thus is
multiplicative. The noise terms accordingly include so-called spurious drift contributions, which
ensure that the LIT dynamics reaches Boltzmann equilibrium when neither ṅ nor μ is held
constant (57).
A crucial assumption of Equation 17 is that φ and n are the only relevant hydrodynamic fields,
so that, on the time and length scales at which they evolve, all other degrees of freedom are ther-
mally equilibrated. Then, the heat Q is given in terms of the entropy production of these fields:

Q = D ln(P[J, ṅ]/P R [J, ṅ]). The resulting thermodynamic EPR differs in form from the IEPR as-
sociated with the dynamics of φ alone (Equation 10). Therefore, it cannot be equated with any of
the various IEPRs for differently time-reversed pure φ dynamics given in Section 2.2. However,
because Equation 17 follows LIT, the heat can be expressed in terms of currents and forces:
t
δF
Q = dr dt − J · ∇ + ṅ μ . 18.
0 δφ
Similar expressions arise for more elaborate field theories, such as polar fields (57). Note that Q is
stochastic, as is W in Equation 16, even though both are defined at hydrodynamic level. Using the

chain rule Ḟ = n α̇n (∂F/∂αn ) + φ̇(δF/δφ)dr and the conservation law φ̇ = −∇ · J, it follows
that free energy F, work W, and heat Q are related as
t
F (t ) − F (0) = W − Q + dr dt ṅ μ. 19.
0
The energy balance in Equation 19 offers the hydrodynamic equivalent of the relation among
the particle-based work W, heat Q, and energy U in Equation 15. When neither ṅ nor μ is
maintained constant, and μ derives from a free energy ( μ = −δFch /δn), the system achieves
equilibrium, so that Equation 19 reduces to the FLT with respect to the total free energy F +

Fch . When the free-energy F does not change (α̇n = 0), the heat rate equals dr ṅ μ, which
illustrates that all the activity ultimately stems from the work done by chemical processes. At fixed

μ (say), ṅ depends on φ, so that dr ṅ μ contains information similar to, but distinct from,
Equation 10 (see discussion in Section 2.4).
2.4. Where and When Activity Matters

The various IEPRs introduced in Section 2 enable one to delineate regimes in which activity most
affects the dynamics compared to equilibrium. Identifying such regimes can quantify the length
and timescales where activity primarily matters and sometimes pinpoint specific locations where
self-propulsion plays an enhanced role. Within this perspective, the most appropriate choice of
IEPR is usually that which best reflects the dynamical symmetries of the emergent order (see
Figure 1). By eliminating gross contributions (e.g., from propulsion pointing opposite to velocity),
the right choice of IEPR can help identify dynamical features that break TRS more subtly and
help unravel how activity can create effects with no equilibrium counterpart.
2.4.1. Spatial and spectral decompositions of irreversibility. If thermal noise is omitted from
Equation 1, only one IEPR can be defined. This is Sr (see Equation 11), tracking positions only. For

AOUPs with pairwise interaction U = i < j V(ri − rj ), one finds a particle-based decomposition

Sr = i σi (69), where
τ 3
σi = (ṙi − ṙ j ) · ∇i V (ri − r j ) . 20.
4μ f02 j
This shows that particle i contributes most to Sr when its neighbors j have a large relative veloc-
ity ṙi − ṙ j . Thus, collisions between slow and fast particles are the main source of irreversibility.
Interestingly, such collisions lie at the basis of the formation of particle clusters, which can poten-
tially lead to nonequilibrium (or motility-induced) phase separation (15). For a phase-separated
density profile, Equation 20 allows one to distinguish the relative contributions from particles in
different spatial zones. In the dilute phase, collisions are rare so that σ i stays small, whereas in
a dense enough phase, particles barely move so that σ i is again modest. At interfaces, collisions

1.6 35
a 1.4 b 30
1.2 25
1.0 20
σ(r)
ρ(r)
0.8 15
0.6 10
0.4 5
0.2 0
0
0 10 20 30 40 50 60 70
r/a
1.0
c
0.004
0.5 0.002
〈ϕ(x)〉
σϕ (x)
0 0
〈ϕ(x)〉
D = 0.001
–0.5 –0.002
D = 0.002
D = 0.02
–1.0 –0.004 D = 0.04
–5 5 0 40 80 120 160 200 240
σ(r) x
Figure 2
Spatial decomposition of the informatic entropy production rate. (a) Phase separation of repulsive active particles, with individual σ i
(color-coded via Equation 20) enhanced at the liquid–vapor interface. (b) Time-averaged profiles of particle density ρ and informatic
entropy production rate density σ show that σ is small in both bulks and peaked at the interface. (c) Similar behavior for a conserved
active scalar field φ. Panels a and b adapted from Reference 39; copyright 2017 American Physical Society. Panel c adapted with
permission from Reference 69; copyright 2021 American Physical Society.
between fast and slow particles, entering respectively from dilute and dense zones, lead to locally
high IEPR (see Figure 2).
In the presence of thermal noise, one can also study S ± (45). These IEPRs track the dy-
namics of both position and self-propulsion. The corresponding particle-based quantities, σi+ =
(1/T )ṙi · fi and σi− = (μ/T )∇iU · fi , which take fi to be even and odd, respectively, are minimal
and maximal for particles at rest. This happens when −i U is equal and opposite to fi . In practice,
σi+ is low in the dense phase and high at the interface (like σ i above), whereas σi− decreases grad-
ually from the bulk of the dense phase to the interface. Considering instead polar clusters, which
emerge when aligning interactions are present, σi− is now low in the dense and dilute phases, with a
higher value at the interface (53). Thus, for isotropic particles it is σi+ that exposes the character of
the irreversibility of self-propulsion at interfaces, whereas for aligning particles σi− does so. Each
separately elucidates the role of interfacial self-propulsion to promote and stabilize clustering.
Turning now to field theories, we consider first the simplest scalar model of Equation 4, with
φ a particle density that is even under time reversal. In the spirit of Landau–Ginzburg theory, the
nonequilibrium forcing term Jφ is generally taken as a local function of φ and its gradients and,

hence, is also even (13, 37). The relevant measure of irreversibility reads as S = σ (r)dr, where
σ = J · Jφ /D can be identified as a local IEPR density (39). To lowest order, Jφ contains three
gradients and two fields, giving leading order (in noise) contributions where φ is large, as applies
near the interface in a phase-separated profile (see Figure 2). This corroborates the particle-
based results for active phase separation given above. (Similar principles govern various local IEPR

densities in field theories of polar active matter; 49.) In contrast to S, the local chemical heat
ṅ μ, found from Equation 18, is high throughout the bulk phases with a (bimodal) dip across
the interface (57). This reflects the fact that the underlying chemical reactions fruitlessly dissipate
Harada–Sasa
energy throughout uniform bulk phases (where C vanishes, modulo noise, so they barely affect relation: spectral
the φ dynamics), whereas at interfaces such reactions locally do work, against the thermodynamic decomposition of
force −∇(δF/δφ), to sustain the counter-diffusive current Jφ . informatic entropy
Another challenge is to identify the timescales on which activity dominates over thermal ef- production across
temporal frequencies
fects. To partially quantify this, one can define a frequency-dependent energy scale Teff (ω) =
and spatial modes
ωCi (ω)/[2Ri (ω)]. The autocorrelation of position reads as Ci (ω) = eiωt ri (t)·ri (0)dt. The response
Ri for small perturbation of the potential (U → U − f û · ri , where û is an arbitrary unit vector)

reads as Ri (ω) = sin(ωt )[δri (t ) · û/δ f (0)] f =0 dt. In the absence of self-propulsion, Teff reduces
to the bath temperature T, restoring the FDT (26). The deviation Teff − T at small ω identi-
fies regimes in which activity dominates, and this has been measured in living systems (70–72).
Interestingly, measuring the FDT violation offers a generic route to quantify irreversibility and
dissipation. In passive systems, the Harada–Sasa relation (73) explicitly connects the position au-
tocorrelation Ci , the response function Ri , and the thermodynamic EPR, S. For active particles, it
can be generalized as (35, 74)

ω
dω ω
dω
S+ = ωCi (ω) − 2T Ri (ω) , Sr = 4μω(ω)Ci (ω) − 2Ri (ω) . 21.
i
μT 2π i
μ 2π
The integrands in Equation 21 provide spectral decompositions of S + and Sr . Using models in

which S + is directly proportional to heat, this decomposition has been measured experimentally
to provide insights into the energetics of living systems (75, 76). Notably the integrand for the
coordinate-only IEPR, Sr , is no longer directly the FDT violation. Instead the irreversibility,
when evaluated from fluctuations of position only, is quantified by the violation of a modified
relation between correlation and response (35). For field theories, the FDT violation can again be
expressed via correlation and response functions in the Fourier domain, depending on both tem-
poral frequencies and spatial modes. It provides a spectral decomposition of the IEPR that usefully
extends the Harada–Sasa relation (39, 77). This helps identify the length and timescales primarily
involved in the breakdown of TRS. When considering theories with several fields, the quantifi-
cation of irreversibility typically involves FDT violations associated with each one of them (39,
47).
2.4.2. Scaling of irreversibility with dynamical parameters. For generic active dynamics,
one can identify as dynamical parameters the coefficients of active terms of the dynamics. Ex-
amples include the strength of microscopic self-propulsion, fi , for particle-based formulations as
in Equation 1, and the coefficients of nonintegrable terms in field theories, such as the leading-
order contributions [(φ)2 ] and (φ)2 φ within Jφ in Equation 4. Estimating how the various
IEPRs depend on these activity parameters, and on temperature or density, allows one to delineate
equilibrium-like regimes in which TRS is restored either exactly or asymptotically. Thus, obtain-
ing precise scalings for IEPRs is an important step toward understanding how far active dynamics
deviates from equilibrium, e.g., with a view to building a thermodynamic framework, starting with
near-equilibrium cases.
Considering AOUPs in Equation 2, the persistence time is a natural parameter that controls
the distance from the equilibrium limit at τ = 0. This approach allows a systematic perturbative
expansion in τ (at fixed Ta ) for the steady state (35, 52, 69). Expanding the IEPR (here Sr ) shows
the irreversibility to vanish at linear order in τ , even as the statistics differ from the Boltzmann

distribution ∼e−U/Ta (35). In other words, there exists a regime of small persistence in which the
steady state is distinct from that of an equilibrium system at temperature Ta , yet TRS is restored
asymptotically. The existence of such a regime lays the groundwork for extending standard rela-
tions of equilibrium thermodynamics.
Various works have studied the behavior of Sr beyond the small τ regime. They have shed
light on a nonmonotonicity of Sr (τ ) for dense systems (78) and also for particles confined in an
external potential (79). Note that, in the latter case, the dependence of Sr on bath temperature T
is also nonmonotonic in general (52). These results suggest that departures from equilibrium can
decrease with persistence and increase with thermal fluctuations. However, note that the IEPR
per particle has units of inverse time (see Equation 10), so that a saturating entropy production
per persistence time gives decreasing Sr (τ ). Turning to repulsive particles, the IEPRs S + and S − ,
respectively, decrease and increase with τ , up to the onset of MIPS (53). This is consistent with
their being, respectively, low and high in clustered regions, and this trend is unaffected by the
rescaling S ± → S ± τ (80).
Another parameter controlling nonequilibrium effects is the density ρ. In a homogeneous state,
because S − increases as the dynamics slows down, S − increases (decreases) with ρ for isotropic
(aligning) particles. In practice, some detailed scalings can be obtained for isotopic pairwise inter-

actions of the form U = i < j V(ri − rj ). For weak interactions, namely when the amplitude of V
is small compared to those of thermal noise and self-propulsion, S − becomes linear in ρ (80). In
general, it can be written in terms of density correlations, as follows:

T S−
=ρ (∇V )2 − T ∇ 2V g2 (r)dr + ρ 2 (∇V (r)) · (∇V (r ))g3 (r, r )drdr , 22.
μ

where g2 = ij δ(r − ri + rj )/(Nρ) and g3 = ijk δ(r − ri + rj )δ(r − ri + rk )/(Nρ)2 are two-
point and three-point density correlators (81). The integrand in Equation 22 can be regarded as
an integral form of the Yvon–Born–Green relation, which constrains g2 and g3 for passive parti-
cles (82), so that the violation of this relation provides access to S − for active ones. A related form
of Equation 22 has been proposed in terms of g2 − g2,eq , where g2,eq is the two-point correlator for
passive particles with the same potential U (81, 83).
In field theories, IEPRs are generically either linear or quadratic in the activity parameters (for
small parameters) depending on whether they break symmetries of the passive theory (39). The
IEPR scalings with D are mechanistically revealing for both scalar (39) and polar (49) models.
As previously noted, the spatial dependence of IEPR density is informative: Independent scalings
can be seen in different bulk phases and at their interfaces. In general, one finds the following
leading order behavior. (a) σ ∼ D−1 : the deterministic part of the dynamics already breaks TRS,
and very unlikely noise realizations are needed to recreate, for the reversed paths, the opposite of a
deterministic forward motion. (b) σ ∼ D0 : TRS is unbroken deterministically but violated at lowest
order in fluctuations. This may require some spatial symmetry breaking at the deterministic level,
so that the background fields act as a ratchet or rectifier for the fluctuations. (c) TRS is broken
only at higher order, which gives σ ∼ D1 in all cases so far studied, although higher powers are
not ruled out.
For scalar fields exhibiting bulk phase separation, the deterministic dynamics gives a stationary
mean-field profile, so that there is no D−1 contribution. TRS is broken at leading order in fluctu-
ations because of the interface between phases (Figure 2). In uniform bulk phases, irreversibility
emerges only at next order, σ ∼ D1 , suggesting that it arises from interactions between fluctua-
tions. Note that active scalar models can also predict entirely new phases, whose σ scalings remain
subject to investigation (37). The behavior of the thermodynamic EPR, ṅ μ(r)/D, found by em-
bedding the same model within LIT and including chemical processes, is quite different (57). This

scales as D−1 but with a reduced amplitude in interfacial regions (see Section 2.4.1). Furthermore,
for one scalar model with both conserved and nonconserved dynamics, TRS violations are gross
(D−1 ) if these two contributions to φ̇ are separately monitored, but perfect reversibility holds if
Stealth entropy
not (84). production: entropy
For dynamics with a scalar density field (even under time reversal) and a polarization (even or production that shows
odd) (47, 49), the IEPR scalings depend on whether the density current J is retained or ignored a nontrivial exponent
(see Equations 4 and 5, and Section 2.2); we assume the latter here. There remain two IEPRs, even when active and
passive models may
S ± , according to whether p is even (+) or odd (−). In the phase comprising traveling bands or
share a universality
clusters (polarized high-density regions that propagate along p), both of S ± scale as D−1 , because class
such deterministic dynamics breaks TRS. Furthermore, because the density wave has a steep front
and a shallow back, flipping p does not restore deterministic-level TRS (Figure 1). In the phase
of uniform p, one finds instead S ± ∼ D0 ; similar arguments might apply now to fluctuations
instead of deterministic motion.

As emphasized already, IEPRs give varying information depending on which degrees of free-
dom are retained and ignored. In particular, the universal properties of the critical point for active
phase separation (including MIPS) can be studied by progressive elimination of degrees of free-
dom via the renormalization group, which could potentially be expected to give rise to emergent
reversibility. Remarkably, Reference 85 has established that (a) the active critical point is in the
same universality class as passive phase separation with TRS, but (b) there is no emergent re-
versibility, in the sense that the IEPR per space–time correlation volume does not scale toward
zero at criticality (it remains constant above four dimensions, and it can even diverge below). This
scenario, referred to as stealth entropy production, can be rationalized by arguing that the scaling
of the IEPR has a nontrivial critical exponent in the universality class shared by active and passive
systems. Because no coarse-graining can make a passive system become active, the IEPR has zero
amplitude in the passive cases previously thought to define the universality class.
3. BIASED ENSEMBLES OF TRAJECTORIES

3.1. Dynamical Bias and Optimal Control
Biased ensembles generalize the canonical ensemble of equilibrium statistical mechanics from
microstates to trajectories (29, 31, 86). They have proven useful in simple models of interacting
particles (87–90) as well as glassy systems (32, 91) and beyond (92–94). They are constructed by
biasing the value of a physical observable, which we denote here by B. The choice of this observable
depends on the physical system of interest, e.g., the heat dissipated during a dynamical trajectory
or the displacement of a tagged particle.
Let X denote a dynamical trajectory of an active system, over a time period [0, t]. Following
Section 2.2, the probability of this trajectory can be represented as P[X ] = p0 [X ]e−A[X ] , where A
is the action and p0 is the probability of the initial condition. A biased ensemble is defined by a
probability distribution over these trajectories:
1
Ps [X ] = p0 [X ] exp (−A[X ] − sB[X ]), 23.
Z (s, t )
where s is the biasing field and Z (s, t ) = e−sB for normalization. Comparing with the canonical
ensemble, one may identify B and s as an (extensive) physical observable and its conjugate (in-
tensive) field, respectively. The field s, according to its sign, biases the distribution toward higher
or lower B. Importantly, the bias has no prejudice about the dynamical mechanism by which B
changes: This allows the system to access fluctuation mechanisms for this quantity that might not

have been anticipated a priori (91, 93, 94). Consistent with this observation, Equation 23 can also
be derived by a maximum entropy computation for trajectories with nontypical B (95, 96).
The next step is to consider very large times t, which is analogous to the thermodynamic limit
Optimal control:
some control forces of large system size in equilibrium statistical mechanics. Because Z (s) is analogous to a partition
that can be added to function, one may define a dynamical free energy density as
the system to mimic
1
biased ensembles as ψ (s) = lim log Z (s, t ). 24.
closely as possible t→∞ t
This construction is natural if the observable B scales extensively with the time t, which is assumed
below; ψ is then a scaled cumulant generating function for B (33). In contrast to the standard en-
sembles of equilibrium statistical mechanics, Equation 23 does not describe practical experimental
systems. However, there are several theoretical contexts in which such distributions are relevant.
Examples include fluctuation theorems, in which B is a measure of irreversibility, whose proba-

bility distributions have symmetries related to TRS, yielding ψ(s) = ψ(1 − s) (28, 29, 40). Biased
ensembles are also deeply connected with large deviation theory (97), applied to time-averaged
quantities (33, 34), which describes rare events in which B/t differs significantly from its typical
value. Viewing ψ(s) as a dynamical free energy, its singularities can be interpreted as dynamical
phase transitions. In most cases, such transitions require a limit where both t and the system size
go to infinity.
Given that biased ensembles do not mimic typical experimental situations, a natural question
is whether the response to the bias s can be connected with its response to some physical pertur-
bation. Such a connection is provided by optimal control theory (98), which is also intrinsically
related to large deviation theory (89, 99, 100). In fact, biased ensembles show how B can be reduced
or increased, and the corresponding mechanisms are as close as possible to the original dynamics.
We now discuss a method for identifying a physical system that mimics a given biased ensem-
ble (34, 99–101). To reduce the level of technicality, we exclude transient behavior that happens
for times close to the beginning or end of a trajectory. Biased ensembles can then be accurately
reproduced by the optimally controlled system (OCS), which is also called auxiliary process (102)
or driven process (101). To find the OCS, consider a wider class of controlled systems, obtained
by modifying the equations of motion of the original active model. For any controlled system, the
con
Kullback–Leibler (KL) divergence measures how its path probability P con [X ] = p0 [X ]e−A [X ]
differs from Ps :

P con [X ]
DKL (P con ||Ps ) = P con [X ] log DX , 25.
Ps [X ]
where DX denotes a path integral. The OCS has limt→∞ t −1 DKL (P con ||Ps ) = 0, so its differences
from Ps are small. From Equations 23–25, one can deduce (34)
1 con
lim A [X ] − A[X ] − sB[X ] con ≤ ψ (s), 26.
t→∞ t
where the average ·con is computed for the controlled system. Equation 26 provides a lower
bound on ψ, which becomes an equality for the OCS. Among all the processes with any given
Bcon , the OCS minimizes A − Acon con . Hence, it provides a mechanism for achieving an atypical
value of B while remaining as close as possible to the original dynamics.
Knowing A, Acon , and B, it suffices to maximize the bound in Equation 26 over the controlled
system to obtain the OCS. The details of this computation depend on the unbiased dynamics of
interest, as discussed extensively in References 34, 99, 100, and 101. In practice, the OCS usually
differs from the original system in two ways: (a) The forces in the original model get modified

by terms that depend in simple and explicit ways on s and B, and (b) other control forces are
added as gradients of an optimal control potential Uopt , which may be computed by solving either
an eigenvalue problem (following Doob, as discussed in Reference 101) or a nonlinear partial
differential equation (98). We emphasize that the optimal control forces do not generically appear
as external forces in the dynamics; they can correspond to complicated interactions that may be
long-ranged. Although exact results for Uopt are rare (103–105), Equation 26 is still useful. For
example, using a controlled system with small KL divergence dramatically speeds up the numerical
sampling of biased ensembles (106–109). Also, minimization of the KL divergence over simple
classes of a controlled system provides useful insight into relevant fluctuation mechanisms, even
without an exact solution.
3.2. Phase Transitions and Symmetry Breaking

We now examine phase transitions in biased ensembles of active particles. Interestingly, they can
lead to the emergence of symmetry breaking without any equivalent in the unbiased dynamics.
Furthermore, biased ensembles also open the door to proposing innovative strategies for designing
active particles to achieve some target collective states.
3.2.1. Biased ensembles for active Brownian particles. To illustrate the behavior of biased
ensembles, we discuss a guiding example for a system of ABPs (94, 105). Section 3.2.2 below
presents biased ensembles of active systems in a wider context. Figure 3 displays the behavior of
biased ensembles for a two-dimensional system of ABPs. Trajectories are biased according to their
active work; that is, B = W̄ in Equation 23, with
t
1
W̄[X ] = W̄i [X ], W̄i [X ] = ṙi · fi dt , 27.
i
w0 0
where the normalization w 0 is chosen so that W̄i = t for noninteracting (or sufficiently dilute)
particles. Physically, W̄ measures how effectively the particles’ propulsive forces are converted into
motion: Freely swimming particles have (W̄i /t ) ≈ 1, whereas those in crowded environments are
impeded by their neighbors, resulting in (W̄i /t ) ≈ 0. Negative W̄i means that a particle’s motion
is opposite to its self-propulsive force, which occurs rarely. Because W̄ is closely related to the
IEPR (Section 2.2), it follows a fluctuation theorem (28): The dynamical free energy of the biased
ensemble obeys ψ(w 0 − s) = ψ(s).
Figure 3 shows how the biasing field modifies the behavior of this system. The following dis-
cussion applies for s < w0 /2; the behavior for larger s can then be obtained via the fluctuation
theorem. For s > 0, the system enters a state that is phase-separated and dynamically arrested (de-
noted PSA). Consistent with the arguments of Section 3.1, this happens because phase separation
is the most natural (or least unlikely) mechanism for particles to collectively reduce their active
work, via local crowding effects. Alternative mechanisms do exist for reduced active work; for ex-
ample, dilute ABPs would have W̄ ≈ 0 during rare fluctuations in which the noise term 2μT ξ i
in Equation 1 is opposite to the propulsion force fi . Collectively, though, these trajectories have
much larger action A than does the PSA state; and for a given value of W̄, the biased ensemble
is dominated by the trajectories of least action. Hence, one observes the PSA state in the biased
ensemble.
This behavior is closely related to that of biased ensembles for equilibrium systems, which can
be analyzed at field-theoretic (or fluctuating hydrodynamic) level using macroscopic fluctuation
theory (MFT) (110). The particle density is a locally conserved field, so large-scale density fluc-
tuations relax on a timescale proportional to L2 (where L is the system size). For large systems,
this slow timescale decouples from particles’ rapid microscopic motion, leading to a theory for the

a b
Bias s < 0 Bias s > 0
c d
Control: aligning Control: turning

self-propulsion toward density
gradient
Figure 3
Behavior in biased ensembles of active Brownian particles, and comparison with controlled systems. The
central panel shows a snapshot of the unbiased dynamics (W̄ W̄). Panels a and b show similar snapshots
of collective motion and phase-separated arrested states (respectively, biased ensembles with s < 0 and s > 0)
associated with atypical W̄. Particles are colored according to their orientations. Panels c and d show how
biased ensembles can be mimicked by control forces (or torques). In panel c, an infinite-ranged interaction
favors orientational alignment; particles with similar colors are aligned with each other, as in panel a. In
panel d, control torques tend to align orientations along density gradients. Particles are colored according to
their control torque strength, which tends to be largest for boundary particles, facilitating phase separation.
Panels a, b, and d adapted with permission from Reference 94; copyright 2019 American Physical Society.
Panel c adapted from Reference 105; copyright 2021 American Physical Society.
density alone. Construction of this theory involves scaling the space–time coordinates, so that the
biasing field s enters via a scaling variable λ = sL2 . For equilibrium steady states, MFT predicts
dynamical phase transitions at finite λ, corresponding to a bias s of order L−2 (88). This argument
applies also in the active case: The result is that sufficiently large systems should phase separate
as soon as s is positive.
To understand this result physically, recall the connection of biased ensembles to optimal con-
trol theory (Equation 26). As well as biased systems, Figure 3 shows those to which control forces
have been applied. To mimic PSA states, a suitable control strategy is to create a macroscopic clus-
ter, whose boundary particles have their self-propulsion forces pointing inward (similar to MIPS
states). This can be achieved by applying torques to the ABP orientations, so that they point along
the local density gradient, stabilizing the phase-separated cluster and increasing its density toward
dynamical arrest (94). Such control forces (or torques) act mostly on the boundary particles, so
that the difference between the action terms in Equation 26 scales as L (i.e., subextensive in the
volume L2 of the system). This is consistent with the irreversibility measure (1/T )ṙi · fi being
high at the interface for a phase-separated profile in the unbiased dynamics (Section 2.4). Because
the subextensive control forces lead to an extensive change in W̄, one can confirm that a PSA state

is expected for all s > 0 (94). Note that control using torques is efficient for ABPs, because it har-
nesses the inherent self-propulsion to stabilize the cluster. Alternatively, applying control forces to
particle positions, as in Reference 111, would result in a larger (though still subextensive) action.
Returning to Figure 3, one sees for s < 0 that the particles align their orientations to create a
state of collective motion. Because ABPs move with fixed speed, this alignment suppresses particle
collisions, which is a natural way to increase the active work W̄. Spontaneous breaking of rotational
symmetry may be surprising, because aligning interactions are completely absent from the ABP
model. However, numerical results (94, 105) show that this biased ensemble can be described
rather accurately by a controlled model with mean-field (infinite-ranged) interactions among the
ABP orientations.
It is clear that any field-level description of collective motion requires a polarization field in
order to capture such alignment. Hence, collective motion cannot be described by a fluctuating
hydrodynamic theory for the density alone, contrary to the PSA state. Also, the collective motion
phase is found for biasing fields beyond a threshold of s∗ −1/τ , rather than at infinitesimal s
as for PSA. These observations are related: In the PSA state, the response to the bias takes place
by a slow (hydrodynamic) field, for which the natural scale is s ∼ L−2 ; for the collective motion
state, the response occurs by a fast field, which requires s of order unity. In fact, responses to small
biasing fields s are generically dominated by slow hydrodynamic modes (34). If the quantity B
couples to these, it responds strongly at small s (87, 88). By contrast, for s of order unity, an accu-
rate description of the large-scale, bias-induced atypical dynamics may require analysis of fields
(like polarization in this example) that are negligible for hydrodynamic descriptions of typical
trajectories.
We now consider a coarse-grained (hydrodynamic) theory that captures the collective motion
in biased ensembles with low active work. This reinforces the correspondence between particle
models and field-theoretic descriptions. As noted above, describing the onset of collective motion
at s ∼ s∗ requires consideration of a polarization field p alongside the density φ, as in Equations 4
and 5:
√
J = v(φ)p − Dc (φ)∇φ + 2M(φ)φ , ṗ = F(φ, p) + 2D p , 28.
where v is self-propulsion, Dc is a (collective) diffusion constant, M is the mobility, and F is an

effective thermodynamic force (105). To analyze biased ensembles, we need the dependence of
the active work on φ, p. At hydrodynamic level, it is consistent to approximate this as an integral
t
of a local work rate w as W̄ ≈ 0 w(φ, p) dr dt . Here, w is averaged over a mesoscopic region
and a time interval where density and polarization have prescribed values.
To understand collective motion, the key point is that the active work tends to be
reduced by alignment (because collisions are reduced); hence, for small polarization p:
w(φ, p) w(φ, 0) − cw (φ)|p|2 , where cw > 0 is the coupling between active work and polarization.
Construction of the coarse-grained action for this minimal theory confirms that p is a fast field,
whereas φ is slow (hydrodynamic). It is consistent to assume that F(φ, p) −cF (φ)p, where cF > 0
is the strength of the restoring force toward p = 0. Hence, for trajectories X with homogeneous
density and small (constant) polarization p, one has Ps [X ] ∼ e−Apol [X ] (see Equation 23), where
c2
Apol ∝ 4F + scw |p|2 + O(|p|4 ). Minimizing the Landau-like action Apol reveals that spontaneous
symmetry breaking occurs for s < sc , where sc = −cF2 /(4cw ) < 0. This connects to particle-level re-
sults via the scaling estimates cw ∼ 1 and γ cF2 ∼ τ −1 , yielding symmetry breaking at sc ∼ −1/τ (105).
3.2.2. Generic features of biased ensembles in active systems. We now discuss the insights
from the previous example in a broader context. First, note that biasing the active work can create

both orientational order and density modulations; this stems from the couplings between fluctua-
tions of the active work and those of local alignment and density. The first of these can already be
deduced via exact biased-ensemble results in a system with just two active particles (104, 105): Pairs
with parallel alignment tend to have fewer collisions, so these states are promoted when biasing to
higher active work (or propulsion efficiency). Broader insight stems from the optimal control in-
terpretation of biased ensembles (Section 3.1). For ABPs biased by the active work, the collective
motion at large W̄ shows that local alignment is an effective route toward efficient swimming (94,
104), suggesting a design strategy to achieve active particles with desired properties.
Similar arguments explain the coupling of active work to local density and, hence, to phase
separation, because crowding reduces the particles’ swimming speed, as found by fluctuating hy-
drodynamic arguments (94, 105, 112), perturbative calculations (80, 81), and variational com-
putations (111, 112). Many other active systems couple alignment, density, and work by similar
mechanisms; so this behavior under bias should be generic. Indeed, the results of References 80,
81, 94, 104, 105, and 111 cover a range of different active systems, including lattice models as
well as AOUPs and ABPs. Dynamical arrest in biased ensembles also has a counterpart in passive
glassy systems (32, 34, 91), and can be explained by general principles that couple biasing fields to
metastable states (34).
All these results concern systems in which the observable B is defined as a sum over all particles.
Instead, B may refer to a tagged particle within a large system (113–115). Perhaps surprisingly,
results for equilibrium systems show that applying such bias to a single particle can generate a
macroscopic response (116). This is because of coupling of the biased quantity to the slow (hy-
drodynamic) density fluctuations, as described by MFT. In active systems, the results of Refer-
ences 113 and 115 indicate a rich behavior for large deviations of the single-particle active work,
which appears related to a coupling with collective (long-ranged) density fluctuations. However,
fluctuations of the single-particle current were explained in Reference 114 by a local argument
based on a density-dependent velocity.
4. CONCLUSION
The tools of stochastic thermodynamics, first introduced for thermodynamically consistent mod-
els (27, 28, 40), provide fruitful insights when analyzing the consequences of irreversibility in active
dynamics. The framework has to be carefully adapted, because various thermodynamic relations
do not carry over a priori. For instance, the usual connections among irreversibility, entropy pro-
duction, and dissipation of energy need to be revisited (Sections 2.2 and 2.3). The FLT no longer
holds in its standard form, which is a fact that strongly influences the behavior of engines using
an active working substance (62). However, the IEPR can be defined without reference to the first
law directly in terms of forward and backward path probabilities. This makes it a useful measure
of irreversibility in active systems. In contrast with equilibrium, care is now needed in choosing
which coarse-grained quantities change sign on time reversal: Different choices yield different
information on how TRS is broken.
Importantly, measuring irreversibility via IEPR can provide a systematic approach for evalu-
ating the distance of active dynamics from an equilibrium reference. This allows one to pinpoint
regimes (such as specific bulk phases, or interfaces between these) where activity plays a major role,
and rationalize the mechanisms whereby self-propulsion at the particle scale stabilizes collective
effects with no equilibrium equivalent (Section 2.4). As reviewed above, this approach has been
validated in several settings, involving either particle-based or field-level dynamics for scalar and
polar materials, but much remains to be explored beyond these examples. In particular, it would
be interesting to consider active nematics (14).

Recent progress in stochastic thermodynamics has opened new avenues whose implications
for active matter remain largely unexplored. In particular, thermodynamic uncertainty relations
(TURs) place bounds on irreversibility in the presence of currents (117). For active systems, it is
not yet known how different types of emergent order affect the tightness of these bounds. Un-
derstanding this could allow delineation of regimes in which TURs are most accurate, with direct
relevance for experimental studies of living systems.
Finally, biased ensembles (Section 3) offer promise for the engineering of active materials (81,
94, 104). Within this perspective, one predefines a desired macroscopic property in terms of a
specific biasing observable. For a given starting material, the sampling of biased trajectories deter-
mines a strategy to tune microscopic interactions and achieve the target property. This approach
is potentially also relevant when developing strategies to control the collective behavior of social
and living systems, such as pedestrian crowds or cars on a freeway. Further use of biased ensembles
in active matter models could therefore help establish ground rules for their wider application in
systems that are far from equilibrium.
DISCLOSURE STATEMENT
The authors are not aware of any affiliations, memberships, funding, or financial holdings that
might be perceived as affecting the objectivity of this review.
ACKNOWLEDGMENTS
The authors acknowledge insightful discussions with Øyvind L. Borthne, Fernando Caballero,
Timothy Ekeh, Yann-Edwin Keta, Yuting Irene Li, Tomer Markovich, Cesare Nardini, Takahiro
Nemoto, Patrick Pietzonka, Sriram Ramaswamy, Udo Seifert, Thomas Speck, Julien Tailleur,
Elsen Tjhung, Laura Tociu, Suriyanarayanan Vaikuntanathan, and Frédéric van Wijland. É.F. ac-
knowledges support from an ATTRACT Grant of the Luxembourg National Research Fund.
M.E.C. is funded by the Royal Society. This work was funded in part by the National Science
Foundation, Grant No. NSF PHY-1748958, and the European Research Council under the
Horizon 2020 Programme, Grant No. 740269.
LITERATURE CITED
1. Marchetti MC, Joanny JF, Ramaswamy S, Liverpool TB, Prost J, et al. 2013. Rev. Mod. Phys. 85:1143–89
2. Bechinger C, Di Leonardo R, Löwen H, Reichhardt C, Volpe G, Volpe G. 2016. Rev. Mod. Phys.
88:045006
3. Fodor É, Marchetti MC. 2018. Physica A 504:106–20
4. Elgeti J, Winkler RG, Gompper G. 2015. Rep. Prog. Phys. 78(5):056601
5. Saw TB, Doostmohammadi A, Nier V, Kocgozlu L, Thampi S, et al. 2017. Nature 544:212–16
6. Cavagna A, Giardina I. 2014. Annu. Rev. Condens. Matter Phys. 5:183–207
7. Bain N, Bartolo D. 2019. Science 363(6422):46–49
8. Deseigne J, Dauchot O, Chaté H. 2010. Phys. Rev. Lett. 105:098001
9. Palacci J, Sacanna S, Steinberg AP, Pine DJ, Chaikin PM. 2013. Science 339(6122):936–40
10. Vicsek T, Czirók A, Ben-Jacob E, Cohen I, Shochet O. 1995. Phys. Rev. Lett. 75:1226–29
11. Fily Y, Marchetti MC. 2012. Phys. Rev. Lett. 108:235702
12. Toner J, Tu Y. 1995. Phys. Rev. Lett. 75:4326–29
13. Wittkowski R, Tiribocchi A, Stenhammar J, Allen RJ, Marenduzzo D, Cates ME. 2014. Nat. Commun.
5:4351
14. Chaté H. 2020. Annu. Rev. Condens. Matter Phys. 11:189–212

15. Cates ME, Tailleur J. 2015. Annu. Rev. Condens. Matter Phys. 6:219–44
16. Tailleur J, Cates ME. 2008. Phys. Rev. Lett. 100:218103
17. Maggi C, Marconi UMB, Gnan N, Di Leonardo R. 2015. Sci. Rep. 5:10742
18. Yang X, Manning ML, Marchetti MC. 2014. Soft Matter 10:6477–84
19. Takatori SC, Yan W, Brady JF. 2014. Phys. Rev. Lett. 113:028103
20. Solon AP, Fily Y, Baskaran A, Cates ME, Kafri Y, et al. 2015. Nat. Phys. 11(8):673–78
21. Bialké J, Siebert JT, Löwen H, Speck T. 2015. Phys. Rev. Lett. 115:098301
22. Zakine R, Zhao Y, Knežević M, Daerr A, Kafri Y, et al. 2020. Phys. Rev. Lett. 124:248003
23. Paliwal S, Rodenburg J, van Roij R, Dijkstra M. 2018. New J. Phys. 20:015003
24. Guioth J, Bertin E. 2019. J. Chem. Phys. 150:094108
25. Onsager L. 1931. Phys. Rev. 37:405–26
26. Kubo R. 1966. Rep. Prog. Phys. 29(1):255–84
27. Sekimoto K. 1998. Prog. Theor. Phys. Suppl. 130:17–27
28. Seifert U. 2012. Rep. Prog. Phys. 75(12):126001

29. Maes C. 1999. J. Stat. Phys. 95(1):367–92
30. Derrida B. 2007. J. Stat. Mech. 2007(07):P07023
31. Lecomte V, Appert-Rolland C, van Wijland F. 2007. J. Stat. Phys. 127(1):51–106
32. Garrahan JP, Jack RL, Lecomte V, Pitard E, van Duijvendijk K, van Wijland F. 2007. Phys. Rev. Lett.
98:195702
33. Touchette H. 2009. Phys. Rep. 478(1):1–69
34. Jack RL. 2020. Eur. Phys. J. B 93:74
35. Fodor É, Nardini C, Cates ME, Tailleur J, Visco P, et al. 2016. Phys. Rev. Lett. 117:038103
36. Dean DS. 1996. J. Phys. A Math. Gen. 29(24):L613
37. Tjhung E, Nardini C, Cates ME. 2018. Phys. Rev. X 8:031080
38. Tiribocchi A, Wittkowski R, Marenduzzo D, Cates ME. 2015. Phys. Rev. Lett. 115:188302
39. Nardini C, Fodor É, Tjhung E, van Wijland F, Tailleur J, et al. 2017. Phys. Rev. X 7:021007. https://doi.
org/10.1103/PhysRevX.7.021007
40. Lebowitz JL, Spohn H. 1999. J. Stat. Phys. 95(1):333–65
41. Mandal D, Klymko K, DeWeese MR. 2017. Phys. Rev. Lett. 119:258001
42. Puglisi A, Marconi UMB. 2017. Entropy 19(7):356
43. Speck T. 2018. Europhys. Lett. 123(2):20007
44. Caprini L, Marconi UMB, Puglisi A, Vulpiani A. 2018. Phys. Rev. Lett. 121:139801
45. Shankar S, Marchetti MC. 2018. Phys. Rev. E 98:020604
46. Pietzonka P, Seifert U. 2018. J. Phys. A Math. Theor. 51(1):01LT01
47. Dadhichi LP, Maitra A, Ramaswamy S. 2018. J. Stat. Mech. 2018(12):123201
48. Dabelow L, Bo S, Eichhorn R. 2019. Phys. Rev. X 9:021009
49. Borthne ØL, Fodor É, Cates ME. 2020. New J. Phys. 22(12):123012
50. Onsager L, Machlup S. 1953. Phys. Rev. 91:1505–12
51. Caprini L, Marconi UMB, Puglisi A, Vulpiani A. 2019. J. Stat. Mech. 2019(5):053203
52. Martin D, de Pirey TA. 2021. J. Stat. Mech. 2021(4):043205
53. Crosato E, Prokopenko M, Spinney RE. 2019. Phys. Rev. E 100:042613
54. Speck T. 2016. Europhys. Lett. 114(3):30006
55. Solon AP, Tailleur J. 2013. Phys. Rev. Lett. 111:078101
56. Alert R, Joanny JF, Casademunt J. 2020. Nat. Phys. 16(6):682–88
57. Markovich T, Fodor É, Tjhung E, Cates ME. 2021. Phys. Rev. X 11:021057
58. Pietzonka P, Fodor É, Lohrmann C, Cates ME, Seifert U. 2019. Phys. Rev. X 9:041032
59. Ekeh T, Cates ME, Fodor É. 2020. Phys. Rev. E 102:010101
60. Zakine R, Solon A, Gingrich T, van Wijland F. 2017. Entropy 19(5):193
61. Holubec V, Steffenoni S, Falasco G, Kroy K. 2020. Phys. Rev. Res. 2:043262
62. Fodor É, Cates ME. 2021. Europhys. Lett. 134(1):10003
63. Loos SAM, Klapp SHL. 2020. New J. Phys. 22(12):123051
64. Gaspard P, Kapral R. 2018. J. Chem. Phys. 148(13):134104

65. Weber CA, Zwicker D, Jülicher F, Lee CF. 2019. Rep. Prog. Phys. 82(6):064601
66. Groot SRD, Mazur P. 1962. Non-Equilibrium Thermodynamics. Amsterdam: North-Holland
67. Kruse K, Joanny JF, Jülicher F, Prost J, Sekimoto K. 2004. Phys. Rev. Lett. 92:078101
68. Prost J, Jülicher F, Joanny JF. 2015. Nat. Phys. 11:111–17
69. Martin D, O’Byrne J, Cates ME, Fodor É, Nardini C, et al. 2021. Phys. Rev. E 103:032607. https://doi.
org/10.1103/PhysRevE.103.032607
70. Fodor É, Guo M, Gov NS, Visco P, Weitz DA, van Wijland F. 2015. Europhys. Lett. 110(4):48005
71. Ahmed WW, Fodor É, Almonacid M, Bussonnier M, Verlhac M-H, et al. 2018. Biophys. J. 114(7):1667–
79
72. Gnesotto FS, Mura F, Gladrow J, Broedersz CP. 2018. Rep. Prog. Phys. 81(6):066601
73. Harada T, Sasa Si. 2005. Phys. Rev. Lett. 95:130602
74. Szamel G. 2019. Phys. Rev. E 100:050603
75. Toyabe S, Okamoto T, Watanabe-Nakayama T, Taketani H, Kudo S, Muneyuki E. 2010. Phys. Rev. Lett.
104:198103
76. Fodor É, Ahmed WW, Almonacid M, Bussonnier M, Gov NS, et al. 2016. Europhys. Lett. 116:30008
77. Seara DS, Machta BB, Murrell MP. 2021. Nat. Commun. 12:392
78. Flenner E, Szamel G. 2020. Phys. Rev. E 102:022607
79. Dabelow L, Bo S, Eichhorn R. 2021. J. Stat. Mech. 2021(3):033216
80. Fodor É, Nemoto T, Vaikuntanathan S. 2020. New J. Phys. 22:013052
81. Tociu L, Fodor É, Nemoto T, Vaikuntanathan S. 2019. Phys. Rev. X 9:041026
82. Hansen JP, McDonald IR. 2013. Theory of Simple Liquids. Oxford: Academic
83. Tociu L, Rassolov G, Fodor É, Vaikuntanathan S. 2020. arXiv:2012.10441
84. Li YI, Cates ME. 2021. J. Stat. Mech. 2021(1):013211
85. Caballero F, Cates ME. 2020. Phys. Rev. Lett. 124:240604
86. Garrahan JP, Jack RL, Lecomte V, Pitard E, van Duijvendijk K, van Wijland F. 2009. J. Phys. A
42(7):075007
87. Bodineau T, Derrida B. 2004. Phys. Rev. Lett. 92(18):180601
88. Appert-Rolland C, Derrida B, Lecomte V, van Wijland F. 2008. Phys. Rev. E 78(2):021122
89. Jack RL, Sollich P. 2015. Eur. Phys. J. Spec. Top. 224(12):2351–67
90. Dolezal J, Jack RL. 2019. J. Stat. Mech. 2019(12):123208
91. Hedges LO, Jack RL, Garrahan JP, Chandler D. 2009. Science 323(5919):1309–13
92. Garrahan JP, Lesanovsky I. 2010. Phys. Rev. Lett. 104(16):160601
93. Weber JK, Jack RL, Schwantes CR, Pande VS. 2014. Biophys. J. 107(4):974–82
94. Nemoto T, Fodor É, Cates ME, Jack RL, Tailleur J. 2019. Phys. Rev. E 99:022605. https://doi.org/10.
1103/PhysRevE.99.022605
95. Simha A, Evans RML, Baule A. 2008. Phys. Rev. E 77(3):031117
96. Pressé S, Ghosh K, Lee J, Dill KA. 2013. Rev. Mod. Phys. 85(3):1115–41
97. den Hollander F. 2000. Large Deviations. Providence, RI: Am. Math. Soc.
98. Bertsekas DP. 2005. Dynamic Programming and Optimal Control, Vol. 1. Belmont, MA: Athena Sci.
99. Dupuis P, Ellis RS. 1997. A Weak Convergence Approach to the Theory of Large Deviations. New York: Wiley
100. Chétrite R, Touchette H. 2015. J. Stat. Mech. 2015(12):P12001
101. Chetrite R, Touchette H. 2015. Ann. Henri Poincaré 16(9):2005–57
102. Jack RL, Sollich P. 2010. Prog. Theor. Phys. Suppl. 184:304–17
103. Tsobgni Nyawo P, Touchette H. 2016. Phys. Rev. E 94:032101
104. Cagnetta F, Mallmin E. 2020. Phys. Rev. E 101:022130
105. Keta YE, Fodor É, van Wijland F, Cates ME, Jack RL. 2021. Phys. Rev. E 103:022603. https://doi.org/
10.1103/PhysRevE.103.022603
106. Nemoto T, Bouchet F, Jack RL, Lecomte V. 2016. Phys. Rev. E 93(6):062123
107. Nemoto T, Jack RL, Lecomte V. 2017. Phys. Rev. Lett. 118(11):115702
108. Ray U, Chan GKL, Limmer DT. 2018. Phys. Rev. Lett. 120:210602
109. Jacobson D, Whitelam S. 2019. Phys. Rev. E 100:052139
110. Bertini L, De Sole A, Gabrielli D, Jona-Lasinio G, Landim C. 2015. Rev. Mod. Phys. 87(2):593–636

111. Whitelam S, Klymko K, Mandal D. 2018. J. Chem. Phys. 148(15):154902
112. GrandPre T, Klymko K, Mandadapu KK, Limmer DT. 2021. Phys. Rev. E 103:012613
113. Cagnetta F, Corberi F, Gonnella G, Suma A. 2017. Phys. Rev. Lett. 119:158002
114. GrandPre T, Limmer DT. 2018. Phys. Rev. E 98:060601
115. Chiarantoni P, Cagnetta F, Corberi F, Gonnella G, Suma A. 2020. J. Phys. A Math. Theor. 53(36):36LT02
116. Dolezal J, Jack RL. 2021. Phys. Rev. E 103:052132
117. Horowitz JM, Gingrich TR. 2020. Nat. Phys. 16(1):15–20

Annual Review of
Condensed Matter
Contents Physics
Volume 13, 2022
Reflections on 65 Years of Helium Research

John D. Reppy p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 1
My Life and Science
Valery L. Pokrovsky p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p15

Russell Donnelly and His Leaks
J.J. Niemela and K.R. Sreenivasan p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p33
Director Deformations, Geometric Frustration, and Modulated Phases
in Liquid Crystals
Jonathan V. Selinger p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p49
Thin Film Skyrmionics
Takaaki Dohi, Robert M. Reeve, and Mathias Kläui p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p73
The Physics of Dense Suspensions
Christopher Ness, Ryohei Seto, and Romain Mari p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p97
Topological Magnets: Functions Based on Berry Phase and Multipoles
Satoru Nakatsuji and Ryotaro Arita p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 119
Active Turbulence
Ricard Alert, Jaume Casademunt, and Jean-François Joanny p p p p p p p p p p p p p p p p p p p p p p p p p p p p 143
Topological Magnons: A Review
Paul A. McClarty p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 171
Olfactory Sensing and Navigation in Turbulent Environments
Gautam Reddy, Venkatesh N. Murthy, and Massimo Vergassola p p p p p p p p p p p p p p p p p p p p p p p p p 191
Irreversibility and Biased Ensembles in Active Matter: Insights from
Stochastic Thermodynamics
Étienne Fodor, Robert L. Jack, and Michael E. Cates p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 215
The Hubbard Model
Daniel P. Arovas, Erez Berg, Steven A. Kivelson, and Srinivas Raghu p p p p p p p p p p p p p p p p p p 239
The Hubbard Model: A Computational Perspective
Mingpu Qin, Thomas Schäfer, Sabine Andergassen, Philippe Corboz,
and Emanuel Gull p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 275
Understanding Hydrophobic Effects: Insights from Water
Density Fluctuations
Nicholas B. Rego and Amish J. Patel p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 303
Modeling of Ferroelectric Oxide Perovskites: From First
to Second Principles
Philippe Ghosez and Javier Junquera p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 325
How Cross-Link Numbers Shape the Large-Scale Physics of
Cytoskeletal Materials
Sebastian Fürthauer and Michael J. Shelley p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 365
Studying Quantum Materials with Scanning SQUID Microscopy

Eylon Persky, Ilya Sochnikov, and Beena Kalisky p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 385
Coherently Coupled Mixtures of Ultracold Atomic Gases
Alessio Recati and Sandro Stringari p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 407
Errata
An online log of corrections to Annual Review of Condensed Matter Physics articles may
be found at http://www.annualreviews.org/errata/conmatphys

Annurev Conmatphys 031720 032419

Uploaded by

Copyright:

Available Formats

Annurev Conmatphys 031720 032419

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Annurev Conmatphys 031720 032419

Uploaded by

Copyright:

Available Formats

Annual Review of Condensed Matter Physics

Irreversibility and Biased

Annu. Rev. Condens. Matter Phys. 2022. 13:215–38 Keywords

216 Fodor • Jack • Cates

types of irreversible dynamics make it insufficient to study ensembles of configurations (charac-

2. IRREVERSIBILITY AND DISSIPATION

www.annualreviews.org • Stochastic Thermodynamics: Active Matter 217

AOUP: τ ḟi = −fi + f0 2τ ζ i , 2.

218 Fodor • Jack • Cates

2.2. How Far from Equilibrium Is Active Matter?

2.2.1. Forward and time-reversed dynamics: path-probability representation. To quantify

The kernel is an even function of time t, which depends on μ, T, f0 , and τ . It reduces to

www.annualreviews.org • Stochastic Thermodynamics: Active Matter 219

Time reversal: several   

220 Fodor • Jack • Cates

www.annualreviews.org • Stochastic Thermodynamics: Active Matter 221

222 Fodor • Jack • Cates

U (t ) − U (0) = W − Q + ṙi · fi dt . 15.

www.annualreviews.org • Stochastic Thermodynamics: Active Matter 223

224 Fodor • Jack • Cates

2.4. Where and When Activity Matters

www.annualreviews.org • Stochastic Thermodynamics: Active Matter 225

226 Fodor • Jack • Cates

The integrands in Equation 21 provide spectral decompositions of S + and Sr . Using models in

www.annualreviews.org • Stochastic Thermodynamics: Active Matter 227

228 Fodor • Jack • Cates

instead of deterministic motion.

3. BIASED ENSEMBLES OF TRAJECTORIES

www.annualreviews.org • Stochastic Thermodynamics: Active Matter 229

Examples include fluctuation theorems, in which B is a measure of irreversibility, whose proba-

230 Fodor • Jack • Cates

3.2. Phase Transitions and Symmetry Breaking

www.annualreviews.org • Stochastic Thermodynamics: Active Matter 231

Control: aligning Control: turning

232 Fodor • Jack • Cates

where v is self-propulsion, Dc is a (collective) diffusion constant, M is the mobility, and F is an

www.annualreviews.org • Stochastic Thermodynamics: Active Matter 233

234 Fodor • Jack • Cates

www.annualreviews.org • Stochastic Thermodynamics: Active Matter 235

28. Seifert U. 2012. Rep. Prog. Phys. 75(12):126001

236 Fodor • Jack • Cates

www.annualreviews.org • Stochastic Thermodynamics: Active Matter 237

238 Fodor • Jack • Cates

Volume 13, 2022

Reflections on 65 Years of Helium Research

Valery L. Pokrovsky p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p15

Studying Quantum Materials with Scanning SQUID Microscopy

You might also like

Time reversal: several