Kinetic theory of collisionless relaxation for systems with long-range interactions

Pierre-Henri Chavanis1
1
Laboratoire de Physique Théorique, Université de Toulouse, CNRS, UPS, Toulouse, France
We develop the kinetic theory of collisionless relaxation for systems with long-range interactions
in relation to the statistical theory of Lynden-Bell. We treat the multi-level case. We make the
connection between the kinetic equation obtained from the quasilinear theory of the Vlasov equation
arXiv:2112.13664v1 [cond-mat.stat-mech] 21 Dec 2021

and the relaxation equation obtained from a maximum entropy production principle. We propose a
method to close the infinite hierarchy of kinetic equations for the phase level moments and obtain
a kinetic equation for the coarse-grained distribution function in the form of a generalized Landau,
Lenard-Balescu or Kramers equation associated with a generalized form of entropy [P.H. Chavanis,
Physica A 332, 89 (2004)]. This allows us to go beyond the two-level case associated with a Fermi-
Dirac-type entropy. We discuss the numerous analogies with two-dimensional turbulence. We also
mention possible applications of the present formalism to fermionic and bosonic dark matter halos.

I. INTRODUCTION

Recently, the dynamics and thermodynamics of systems with long-range interactions has been a subject of consider-
able interest in statistical mechanics [1–5]. For systems with long-range interactions, the relaxation towards statistical
equilibrium (Boltzmann distribution) is governed by the homogeneous [6, 7] or inhomogeneous [8, 9] Lenard-Balescu
equation, which is a generalization of the homogeneous [10] or inhomogeneous [11] Landau equation taking into account
collective effects. The collisional relaxation time generically scales with the number of particles N as trelax ∼ N tD ,
where tD is the dynamical time.1 When N → +∞, the relaxation time diverges and the Boltzmann statistical
equilibrium state is never reached. The system is then governed by the Vlasov equation [16, 17] which describes a
collisionless evolution driven only by mean field effects. The Vlasov equation is reversible and conserves the Boltz-
mann entropy among an infinite number of Casimir invariants. This seems to preclude the relaxation towards an
equilibrium state. However, systems governed by the Vlasov equation can experience a process of violent relaxation
on the coarse-grained scale towards a metaequilibrium state. This process of collisionless relaxation was first evi-
denced by King [18] and Hénon [19] in the case of stellar systems described by the Vlasov-Poisson equations. It can
account for the structure and regularity of galaxies whose collisional relaxation time exceeds the age of the Universe
by many orders of magnitude [13]. Similarly, incompressible and inviscid flows in 2D hydrodynamics are described
by the Euler-Poisson equations. Systems governed by these equations also experience a process of violent relaxation
leading to the formation of large-scale vortices like Jupiter’s Great Red spot [20, 21].
A statistical theory of violent relaxation has been developed by Lynden-Bell [22] for stellar systems. The metae-
quilibrium state is obtained by maximizing a mixing entropy while accounting for all the constraints of the dynamics.
This leads to the Lynden-Bell distribution function (DF) which can be viewed as a Fermi-Dirac-type DF in the
two-level case, or a superposition of Fermi-Dirac-type DFs in the multi-level case. A similar statistical theory has
been developed by Miller [23] and Robert and Sommeria [24] for 2D incompressible and inviscid flows. The numerous
analogies between stellar systems and 2D vortices are described by Chavanis [20, 25–28]. The Lynden-Bell statistical
theory has also been applied to other systems with long-range interactions such as the HMF model [29–36]. How-
ever, the power of prediction of the Lynden-Bell statistical theory is limited by the problem of incomplete relaxation
[22, 37–39]. There are cases where the statistical prediction works well and cases where it fails. The Lynden-Bell
statistical theory relies on an assumption of ergodicity which is not always fulfilled in practice. It is difficult to know
a priori if the system will mix efficiently, as required by the ergodicity assumption.
The Lynden-Bell theory is an equilibrium theory. It is then important to develop a kinetic theory of (violent
or quiescent) collisionless relaxation and obtain an evolution equation for the coarse-grained DF.2 This project was

1 The “collisional” evolution of systems with long-range interactions described by the Lenard-Balescu equation is induced by two-body
correlations among the particles. The Lenard-Balescu equation is valid at the order 1/N in a proper thermodynamic limit where
N → +∞ with m ∼ 1/N . Accordingly, the collisional relaxation time generically scales as trelax ∼ N tD [12]. For self-gravitating
systems, the Chandrasekhar relaxation time scales as trelax ∼ (N/ ln N ) tD because of logarithmic corrections [13]. A similar scaling is
obtained in plasma physics where the particle number N is replaced by the plasma parameter Λ giving the number of electrons in the
Debye sphere [10]. For spatially homogeneous 1D systems with long-range interactions, the Lenard-Balescu collision term vanishes [12]
and one has to account for higher order correlations among the particles. A kinetic equation, taking into account three-body correlations,
has recently been derived in Refs. [14, 15]. This equation is valid at the order 1/N 2 , leading to a collisional relaxation time scaling as
trelax ∼ N 2 tD .
2 Lynden-Bell [22] heuristically described the evolution of the coarse-grained DF by a simple Fokker-Planck equation.
 2

first considered by Kadomtsev and Pogutse (KP) [40] for the Vlasov equation of plasma physics. They developed a
quasilinear theory of collisionless relaxation and, in the two-level case, obtained a fermionic-like Landau or Lenard-
Balescu equation which relaxes towards the Lynden-Bell DF. Because of their assumptions, their kinetic theory can
only describe the late quiescent stages of the relaxation process (gentle relaxation). Their approach was extended by
Severne and Luwel (SL) [41] in the multi-level case in an astrophysical context. The quasilinear theory of collisionless
relaxation was further discussed by Chavanis [42–44]. He used it to derive a truncated model (a sort of fermionic King
model) taking into account the evaporation of high energy stars. This DF has a finite mass in contrast to the original
Lynden-Bell DF which is not normalizable when coupled to gravity. The quasilinear theory of the Vlasov-Poisson
equations has been exported to the case of 2D incompressible and inviscid flows described by the Euler-Poisson
equations [45, 46].
A completely different approach has been developed by Chavanis, Sommeria and Robert (CSR) [26] by using a
maximum entropy production principle (MEPP) previously introduced in 2D hydrodynamics [47–49]. A relaxation
equation is constructed heuristically by maximizing the rate of production of Lynden-Bell’s entropy while accounting
for all the constraints of the dynamics. This leads to a generalized Fokker-Planck equation with a time-dependent
temperature that evolves so as to conserve the energy. In the two-level case, this relaxation equation reduces to a
fermionic-like Kramers equation which generalizes the Fokker-Planck equation for the coarse-grained DF introduced
heuristically by Lynden-Bell [22].
It is possible to obtain the CSR equation from the KP and SL equations by making a sort of thermal bath
approximation. This procedure provides the explicit expression of the diffusion coefficient in the CSR equation which
is not given by the MEPP [26]. The connection between the KP and the CSR equations was made in Refs. [42–
44] in the two-level case and is generalized in the present paper to the multi-level case. Then, there remains a
complicated closure problem: The SL and CSR equations yield an infinite hierarchy of equations for the moments
of the distribution. Following the suggestion of Ref. [44], we introduce a simple method to close this hierarchy of
equations. We first remark that the equilibrium coarse-grained DF extremizes a “generalized entropy” at fixed mass
and energy. We also show that the variance of the distribution at equilibrium is related to the coarse-grained DF
through this generalized entropy. We then propose to close the hierarchy of kinetic equations at the level of the
coarse-grained DF by extending this relation out-of-equilibrium (this can be justified by a local thermodynamical
equilibrium assumption). This leads to a form of generalized Landau or Lenard-Balescu equation associated with
a generalized entropy. This equation has been studied in detail in [44]. It conserves mass and energy and satisfies
an H-theorem for the generalized entropy. In the thermal bath approximation, it reduces to a generalized Kramers
equation and the diffusion coefficient can be calculated explicitly.
We show how the kinetic theory is able to account for the problem of incomplete relaxation through a space and
time-dependent diffusion coefficient.
We also discuss the nonlinear dynamical stability of stationary solutions of the Vlasov equation. We introduce an
energy principle which provides the most refined condition of dynamical stability. A stationary solution of the Vlasov
equation is dynamically stable if and only if it is a minimum of energy with respect to symplectic perturbations (i.e.
perturbations that conserve all the Casimirs). We then propose a relaxation equation that minimizes the energy
while conserving all the Casimirs. This relaxation equation can serve as a numerical algorithm to construct stable
steady states of the Vlasov equation. We also show that the maximization of a pseudo-entropy at fixed mass and
energy (thermodynamic-looking microcanonical principle) provides a sufficient condition of dynamical stability. As a
result, the generalized Landau, Lenard-Balescu and Kramers equations of Refs. [44, 50, 51] can be used as numerical
algorithms to construct stable steady states of the Vlasov equation.
Finally, we discuss the notion of H-functions and “selective decay” following the work of Tremaine et al. [52]. We
propose an evolution equation that monotonically increases all the H-functions while conserving the mass and the
energy. It differs from the generalized Landau, Lenard-Balescu and Kramers equations of Refs. [44, 50, 51] which
monotonically increase only one particular H-function (the generalized entropy) at fixed mass and energy.

II. LYNDEN-BELL’S STATISTICAL THEORY

A. Vlasov equation

We consider a Hamiltonian system of N particles interacting via a long-range binary potential u(|r − r′ |) which
decays at large distances as r−γ with γ ≤ d, where d is the dimension of space. Let f (r, v, t) denotes the DF defined
such that f drdv gives the total mass of particles with position r and velocity v at time t. In the collisionless (or
mean field) regime, the evolution of the DF is governed by the Vlasov equation
∂f ∂f ∂f
+v· − ∇Φ · = 0, (1)
∂t ∂r ∂v
 3

where
Z
Φ(r, t) = u(|r − r′ |)f (r′ , v′ , t) dr′ dv′ (2)

is the mean potential produced self-consistently by the particles. The Vlasov equation can be obtained from the
N -body Liouville equation by making a mean-field approximation, i.e., by writing the N -body DF as a product of N
one-body DFs. The Vlasov equation is valid when t ≪ trelax , where trelax is the collisional relaxation time. Since trelax
grows algebraically with N , the Vlasov equation becomes exact when N → +∞ in a proper thermodynamic limit
where the mass of the particles scales as m ∼ 1/N [5]. We note that the individual mass of the particles does not
appear in the Vlasov equation. This implies that the collisionless dynamics does not lead to a segregation by mass.
The Vlasov (or collisionless Boltzmann) equation simply states that, in the absence of encounters, the DF f is
conserved by the flow in phase space. This can be written as Df /Dt = 0 where D/Dt = ∂/∂t + v · ∇r − ∇Φ · ∇v
is the material derivative (Stokes operator). The conservation of the DF together with the incompressibility of the
flow in phase space imply that the total mass (or hypervolume) of all phase elements with phase density between f
and f + δf is conserved.
R This is equivalent to the conservation of an infinite number of invariants called the Casimir
integrals Ih = h(f ) drdv for any continuous function h(f ).R The conservation of the Casimirs is equivalent R to the
conservation of all the moments of the DF, denoted Mn = f n drdv, which include the total mass M = f drdv
1 1
R 2 R
as a special case. The Vlasov equation
R also conserves the total energy E = 2 f v drdv
R + 2 f Φ drdv (kinetic +
potential), the total impulse P = f v drdv, and the total angular momentum L = f r × v drdv (see Appendices
A and B). The Vlasov equation admits an infinite number of stationary solutions whose general form is given by the
Jeans theorem [53]. For example, any DF of the form f = f (ǫ), where ǫ = v 2 /2 + Φ(r) is the individual energy of the
particles by unit of mass, is a steady state of the Vlasov equation. In astrophysics, it describes spherically symmetric
and isotropic stellar systems [54].

B. Metaequilibrium state

The Vlasov-Poisson equations develop very complex filaments as a result of a mixing process in phase space (col-
lisionless mixing). In this sense, the fine-grained DF f (r, v, t) will never reach a stationary state but will rather
produce intermingled filaments at smaller and smaller scales. However, if we introduce a coarse-graining procedure,
the coarse-grained DF f (r, v, t) is expected to reach a metaequilibrium state f (r, v) on a very short timescale, of the
order of the dynamical time tD . This is because the evolution continues at scales smaller than the scale of observation
(coarse-grained). This process is known as “phase mixing” and “violent relaxation” (or collisionless relaxation) [54].
Lynden-Bell [22] has tried to describe this metaequilibrium state in terms of statistical mechanics. In the following,
we summarize his theory and provide some complements (see also [26, 39, 55]).
Let f0 (r, v) denote the initial (fine-grained) DF. We discretize f0 (r, v) in a series of levels η over which f0 (r, v) ≃ η
is approximately constant. Thus, the levels {η} represent all the values taken by the fine-grained DF. If the initial
condition is unsteady or unstable, the DF f (r, v, t) will Rbe stirred in phase space (phase mixing) but will conserve
its values η and the corresponding hypervolumes γ(η) = δ(f (r, v, t) − η) drdv as a property of the Vlasov equation
(this is equivalent to the conservation of all the Casimirs).
Let us introduce the probability density ρ(r, v, η) of finding the level of phase density η in a small neighborhood of
the position (r, v) in phase space. This probability density satisfies at each point the normalization condition
Z
ρ(r, v, η) dη = 1. (3)

The locally averaged (coarse-grained) DF is then expressed in terms of the probability density as
Z
f (r, v) = ρ(r, v, η)η dη, (4)

and the associated potential satisfies Φ(r, t) = u(|r − r′ |)f (r′ , v′ , t) dr′ dv′ . The conserved quantities of the Vlasov
R
equation can be decomposed into two groups.3 The mass and energy are called robust integrals because they are
(approximately) conserved by the coarse-grained DF: M [f ] = M [f ] and E[f ] ≃ E[f ]. Hence
Z
M = f drdv, (5)

3 This distinction was first made in [37, 44].

 4

1 2 1
Z Z
E= f v drdv + f Φ drdv. (6)
2 2
The potential is smooth since it is expressed as an integral of the DF, so we can express the energy in terms of the
coarse-grained fields f and Φ neglecting the internal energy of the fluctuations f˜Φ̃. Therefore, the mass and the
energy can be calculated at any time of the evolution from the coarse-grained field f . By contrast, the moments
n
R nn ≥ 2 are called fragile integrals because they are altered on the
Mn with coarse-grained scale since f n 6= f , where
f n = ρη dη. Therefore, only the moments of the fine-grained DF, Mnf.g. = Mn [f ] = f n drdv, are conserved, i.e.
R

Z
Mn = ρ(r, v, η)η n drdvdη.
f.g.
(7)

R n
The moments of the coarse-grained DF, Mnc.g. [f ] = f drdv, are not conserved along the evolution since Mn [f ] 6=
Mn [f ]. Instead of conserving the fine-grained moments, we can equivalently conserve the total hypervolume
Z
γ(η) = ρ(r, v, η) drdv (8)

of each phase level η. We note that Mnf.g. = γ(η)η n dη.

R
After a complex evolution, we may expect the system to be in the most probable, i.e. most mixed state, consistent
with all the constraints imposed by the dynamics.4 Specifically, the metaequilibrium state is obtained by maximizing
the Lynden-Bell mixing entropy5
Z
SLB [ρ] = − ρ(r, v, η) ln ρ(r, v, η) drdvdη, (9)

while conserving the mass M , the energy E and all the Casimirs (or the fine-grained moments Mnf.g. ). We also need
to account for the local normalization condition (3). The Lynden-Bell entropy (9) can be obtained from a standard
combinatorial analysis taking into account the specificities of the Vlasov equation (see Refs. [22, 39, 55] for details).
Introducing Lagrange multipliers, the first variations satisfy
X Z Z 
f.g.
δSLB − βδE − αδM − αn δMn − ζ(r, v)δ ρ(r, v, η)dη drdv = 0, (10)
n>1

where β = 1/T is the inverse temperature associated with the conservation of energy and αn are the “chemical
potentials” associated with the conservation of the fine-grained moments Mnf.g. (the function ζ(r, v) accounts for the
local normalization condition).6 This variational principle leads to the Gibbs state
1
ρ(r, v, η) = χ(η)e−η(βǫ+α) , (11)
Z(ǫ)

where ǫ = v 2 /2 + Φ(r) is the energy of a particle by unit of mass.7 In writing Eq. (11), we have distinguished the
Lagrange multipliers α and β associated with the robust integrals M and E from the Lagrange multipliers αn>1 ,
f.g.
associated with the conservation of the fragile moments Mn>1 , which have been regrouped in the function
X
χ(η) ≡ exp(− αn η n ). (12)
n>1

4 This statement relies on an assumption of ergodicity which may not always be realized in practice (see Refs. [22, 37–39] for a discussion
of the concept of incomplete relaxation).
5 We take kB = 1 throughout the paper.
6 For brevity, we shall not consider here the conservation of linear impulse P and angular momentum L. Actually, if we work in the
barycentric frame of reference, the Lagrange multiplier U associated with the conservation of P vanishes. On the other hand, the
conservation of angular momentum implies that the statistical equilibrium state has a solid rotation. When L = 0, it can be shown that
the Lagrange multiplier Ω associated with the conservation of L also vanishes [22, 26]. We will assume that we are in this situation,
even though the general case can be treated straightforwardly.
7 We note that the Lynden-Bell distribution (11) does not lead to a segregation by mass since the individual mass of the particles does
not appear in the Vlasov equation on which the whole theory is based. However, it leads to a segregation by phase levels η.
 5

This distinction will make sense in the following (see also Appendix C). Under this form, we see that the equilibrium
distribution of phase levels is a product of a universal Boltzmann factor e−(βǫ+α)η by a non-universal function χ(η)
which depends on the initial condition. The partition function Z is determined by the local normalization condition
(3) yielding
Z
Z(ǫ) = χ(η)e−η(βǫ+α) dη. (13)

The partition function Z(ǫ) can be used as a generating function for constructing the moments of the fine-grained
distribution (see Appendix D). We note that the Lynden-Bell statistics (11) has a form similar to a superstatistics
(see Ref. [55] for the development of this analogy). The equilibrium coarse-grained DF defined by Eq. (4) can be
written as
χ(η)ηe−η(βǫ+α) dη
R
1
Z
−η(βǫ+α)
f= χ(η)ηe dη = R . (14)
Z(ǫ) χ(η)e−η(βǫ+α) dη
One can easily check that
1 ∂ ln Z
f =− = F (βǫ + α) = f (ǫ). (15)
β ∂ǫ
We note that the coarse-grained DF predicted by Lynden-Bell depends only on the individual energy ǫ of the particles.
As such, it is a particular stationary solution of the Vlasov equation. We also note that f (ǫ) is a monotonically
decreasing function of energy. Indeed, from Eqs. (11) and (15), it is easy to establish that (see Appendix D)
′
f (ǫ) = −βf2 , (16)
where
Z
2
f2 ≡ ρ(η − f )2 dη = f 2 − f ≥ 0 (17)

is the local centered variance of the distribution ρ(r, v, η). Equation (16) is a form of fluctuation-dissipation theorem.
′
We note that f (ǫ) ≤ 0 since β ≥ 0 is required to make the velocity profile normalizable. We can also easily show [55]
that f (r, v) ≤ f0max , where f0max is the maximum value of the initial (fine-grained) DF. The inequality 0 ≤ f ≤ f0max
is clear from physical considerations since the coarse-grained DF can only decrease by mixing. Finally, one can show
that the coarse-grained DF predicted by Lynden-Bell is nonlinearly dynamically Vlasov stable (see Sec. VIII).
For a given initial condition, the statistical theory of Lynden-Bell selects a particular stationary solution of the
Vlasov equation (the most probable – most mixed – one) among an infinity of stationary solutions. The Lynden-Bell
equilibrium state is obtained by solving the integral equation
" #
2
v′
Z
′
Φ(r) = u(|r − r |)f αn ,β + Φ(r ) dr′ dv′
′
(18)
2

and relating the Lagrange multipliers (αn , β) to the constraints (Mnf.g. , E). We also have to make sure that the
equilibrium state is an entropy maximum not a minimum or a saddle point (see Appendix C). We note that the
coarse-grained DF f (ǫ) can take different forms depending on the function χ(η) determined by the fragile moments.
In the present context, the function χ(η) is determined from the constraints a posteriori. Indeed, we have to solve
the full problem in order to get the αn ’s and obtain the expression of χ(η).8 We emphasize that the Lynden-Bell
statistical equilibrium state f (ǫ) resulting from a violent collisionless relaxation depends on the details of the initial
condition. This is different from the Boltzmann statistical equilibrium state resulting from a collisional relaxation
which depends only on the value of the mass M and the energy E. In the present case, we need to know the value of
the fine-grained moments Mnf.g. which are accessible only in the initial condition (or from the fine-grained field) since
the observed moments Mnc.g. are altered for t > 0 by the coarse-graining as the system undergoes a mixing process
(Mnc.g. 6= Mnf.g. ).
Remark: Similar results have been derived by Miller [23] and Robert and Sommeria [24] in 2D turbulence. The
analogy between the Lynden-Bell theory and the Miller-Robert-Sommeria theory is discussed in [20, 26].

8 In this sense, the constraints associated with the conservation of the fine-grained moments are treated microcanonically. Following the
approach of [56] in 2D turbulence, we have suggested in [55] that, when the system is forced by an external medium, the fine-grained
constraints may be treated canonically. In that case, the function χ(η) should be considered as given a priori (it is determined by the
external forcing). Treating the Casimirs canonically also allows us to derive a sufficient condition of thermodynamical stability in the
sense of Lynden-Bell (see Appendix C).
 6

C. Two-level case

If the initial DF takes only two values f0 = η0 and f0 = 0 (vacuum), the Lynden-Bell entropy reduces to
Z     
f f f f
S=− ln + 1− ln 1 − drdv, (19)
η0 η0 η0 η0
which is similar to the Fermi-Dirac entropy. Furthermore, the constraints reduce to the conservation of mass M and
f.g.
= f n drdv = η0n−1 × f drdv = η0n−1 M . The metaequilibrium state is then given by
R R
energy E since Mn>1
η0
f= , (20)
1 + eη0 (βǫ+α)
which is similar to the Fermi-Dirac DF [22, 57]. Morphologically, the Lynden-Bell statistics corresponds to a 4th
type of statistics since the particles are distinguishable but subject to an exclusion principle f ≤ η0 due to the
incompressibility of the flow in phase space [22]. This constraint plays a role similar to the Pauli exclusion principle
in quantum mechanics. In the dilute (nondegenerate) limit of the Lynden-Bell theory f ≪ η0 , the entropy (19) and
the DF (20) reduce to
 
f f
Z
S=− ln − 1 drdv, f = η0 e−η0 (βǫ+α) , (21)
η0 η0
which are similar to the Boltzmann entropy and to the Boltzmann distribution.
Remark: We note that the effective temperature T = 1/β in the DF (21) has not the dimension of a temperature.
Indeed, the mass m of the particles does not appear in the Lynden-Bell theory since it is based on the Vlasov equation
for collisionless systems which is independent of the mass of the particles. However, T /η0 can be interpreted as a
velocity dispersion (in the nondegenerate limit). In this sense, one can say that the temperature in Lynden-Bell’s
theory is proportional to the mass of the particles (or more precisely to the ratio m/η0 ) [22].

D. Generalized entropy

Since the coarse-grained DF f (ǫ) predicted by the statistical theory of Lynden-Bell depends only on the individual
energy and is monotonically decreasing, it extremizes a “generalized entropy” of the form [37, 39, 44, 55]
Z
S[f ] = − C(f ) drdv (22)

at fixed mass M and energy E, where C(f ) is a convex function (i.e. C ′′ > 0). Indeed, introducing Lagrange
multipliers α and β, and writing the variational principle under the form

δS − βδE − αδM = 0, (23)

we find that

C ′ (f ) = −βǫ − α. (24)

Since C ′ (f ) is a monotonically increasing function of f , we can inverse this relation to obtain

f = F (βǫ + α) = f (ǫ), (25)

where the function F (x) = (C ′ )−1 (−x) is determined by the generalized entropy C(f ). Inversely, for a given F (x),
the generalized entropy is given by
Z f
C(f ) = − F −1 (x) dx. (26)

From the identity

′ β
f (ǫ) = − , (27)
C ′′ (f )
 7
′
obtained from Eq. (24), we find that f (ǫ) is a monotonically decreasing function, i.e., f (ǫ) < 0 (since β > 0 as
explained in [58]).
Therefore, for any Gibbs state of the form (11), there exists a generalized entropy of the form (22) that the coarse-
grained DF f , given by Eq. (15), extremizes at fixed mass M and energy E. From the statistical theory of Lynden-Bell,
we have F (x) = −(ln Z)′ (x) [see Eq. (15)], where Z(x) depends only on χ(η) [see Eq. (13)]. Substituting this relation
into Eq. (26), we find that the generalized entropy is given by [39, 55]
Z f
C(f ) = − [(ln Z)′ ]−1 (−x) dx. (28)

We expect that in many cases the coarse-grained distribution (15) maximizes the generalized entropy S at fixed mass
M and energy E (robust constraints) although this is not necessarily the case (see Appendix C). We emphasize
that the generalized entropy (28) is a non-universal function which depends on the initial condition. Indeed, it is
determined by the function χ(η) which depends indirectly on the initial condition through the complicated procedure
9
discussed at the end
R of Sec. II B. In general, the generalized entropy (22) with (28) is not the ordinary Boltzmann
entropy SB [f ] = − f ln f drdv because of the existence of fine-grained constraints (Casimirs) that modify the form
of the entropy that we would naively expect. These constraints are sometimes refered to as hidden constraints since
they are not accessible from the coarse-grained dynamics [38, 39, 44, 55].
Remark: Similar results have been derived by Chavanis [59–62] in 2D turbulence.

E. H-functions and selective decay principle

In order to quantify the importance of mixing during the process of violent relaxation, Tremaine et al. [52] have
introduced the notion of H-functions. They are defined by
Z
H[f ] = − C(f ) drdv, (29)

where C is any convex function (i.e. C ′′ > 0). It can be shown that the H-functions H[f ] calculated with the
coarse-grained DF increase during violent relaxation in the sense that H[f (r, v, t)] ≥ H[f (r, v, 0)] for t > 0 where it
is assumed that, initially, the system is not mixed so that f (r, v, 0) = f (r, v, 0). This is similar to the H-theorem
in kinetic theory. However, contrary to the Boltzmann equation, the Vlasov equation does not single out a unique
functional (the above inequality is true for all H-functions) and the time evolution of the H-functions is not necessarily
monotonic (nothing is implied concerning the relative values of H(t) and H(t′ ) for t, t′ > 0). Yet, this observation
suggests a notion of generalized selective decay principle (for −H):10 Among all the invariants of the collisionless
dynamics, the H-functions (fragile constraints) tend to increase (−H tend to decrease) on the coarse-grained scale
while the mass and the energy (robust constraints) are approximately conserved. According to this phenomenological
principle, we may expect that the metaequilibrium state reached by the system as a result of violent relaxation will
maximize a certain H-function (non-universal), denoted H ∗ [f ], at fixed mass and energy.11 This would guarantee
that the metaequilibrium state is nonlinearly dynamically stable (see Sec. VIII). If the evolution is ergodic, the above-
mentioned statement is presumably correct. The H-function H ∗ [f ] that is effectively maximized at metaequilibrium

9 In the case where the system experiences an external forcing (footnote 8), the function χ(η) and, consequently, the generalized entropy
C(f ) should be considered to be given a priori, being determined by the forcing.
10 A similar selective decay principle has been advocated in 2D turbulence and magnetohydrodynamics (see, e.g., Refs. [61, 63–66] and
references therein). It is either due to a small
R dissipation or to a coarse-graining (for dissipationless
R systems). In 2D Rturbulence,
it has been argued that the enstrophy Γ2 = ω 2 dr decreases while the circulation Γ = ω dr and the energy E = 12 ωψ dr are
conserved so that the system reaches a minimum enstrophy state [66–69]. The minimization of enstrophy at fixed circulation and energy
(which is mathematically equivalent to the minimization of energy at fixed circulation and enstrophy) leads to a linear relationship
ω = λψ + µ between vorticity and stream function. The minimization of “generalized R enstrophies” has also been considered.
R In
magnetohydrodynamics [70–74], it has been argued that the magnetic energy E = B2 dr decreases while the helicity H = A · B dr
is conserved so that the system reaches a minimum energy state. The minimization of magnetic energy at fixed helicity (which is
mathematically equivalent to the maximization or minimization of helicity at fixed energy) leads to a linear relationship ∇ × B = λB
which characterizes a force-free configuration. These variational principles ensure the nonlinear dynamical stability of the system
with respect to a dissipationless evolution (we note that all the above functionals are conserved for a purely inviscid and fine-grained
evolution).
11 This is, however, not necessary: All the H-functions could increase at fixed mass and energy without necessarily implying that the
coarse-grained DF reaches a steady state that maximizes one of them (see Sec. IX).
 8

is the generalized entropy S[f ] defined by Eqs. (22) and (28), as obtained from Lynden-Bell’s theory (according to
the comment that follows Eq. (28) this is expected to be true in many cases but not in all cases). Furthermore, under
the assumption of Appendix C 4, it can be shown that the generalized entropy monotonically increases during violent
relaxation (see Sec. IV D). In case of incomplete relaxation, it is possible (but not necessary) that the metaequilibrium
state maximizes a certain H-function (sometimes also called a generalized entropy) which is different from the one
associated with the Lynden-Bell theory (H ∗ [f ] 6= S[f ]). This discussion shows that the notion of selective decay for
collisionless systems with long-range interactions is quite subtle.

F. Incomplete relaxation

The statistical theory of Lynden-Bell relies on the assumption that the evolution is ergodic so that the equilibrium
state maximizes the mixing entropy (9) under the constaints of the dynamics. In reality, this is not always the case.
It has been understood since the beginning [22] that violent relaxation may be incomplete so that the mixing entropy
(9) is not maximized in the whole available phase space. This is obvious in the case of 3D self-gravitating systems
since there is no maximum entropy state, even in theory (the Lynden-Bell DF has an infinite mass). However, even
for simpler systems for which a maximum entropy state (in the sense of Lynden-Bell) exists, there are cases where this
maximum entropy state is not reached (see the discussion in [38] and in Sec. 6 of [39]). For example, in the context of
the HMF model [29–36, 75, 76] and in the context of 2D turbulence [77–80], situations have been reported where the
Lynden-Bell prediction works well and situations have been reported where the Lynden-Bell does not work well (!).
Some authors have proposed to account for incomplete relaxation by changing the form of entropy and by using for
example the Tsallis entropy [81]. Sometimes, the Tsallis distribution provides a good fit of the metaequilibrium state
reached by the system (see the above-mentioned references). However, this is not general. Furthermore, this type of
appoach leads to some arbitrariness since the generalized entropy depends on unknown parameters (like, e.g., Tsallis’
q parameter) that are not predicted by the theory. In practice, these parameters have to be fitted to the observed
distribution. In Sec. V, we shall discuss an alternative approach to take into account incomplete relaxation, based on
kinetic theory, where there is no such indetermination.

III. KINETIC THEORY OF QUIESCENT COLLISIONLESS RELAXATION

In this section, we recall the kinetic theory of quiescent collisionless relaxation, based on a quasilinear theory of the
Vlasov equation, initially developed by Kadomtsev and Pogutse [40], Severne and Luwel [41], and Chavanis [42–44].
We give some details of derivation and complements.

A. Quasilinear theory

Basically, a collisionless system with long-range interactions is described in a self-consistent mean field approxima-
tion by the Vlasov equation [see Eqs. (1) and (2)]. In principle, this equation completely determines the evolution of
the DF f (r, v, t). However, as discussed in Sec. II, we are not interested in practice by the finely striated structure of
the flow in phase space but only by its macroscopic, i.e. smoothed-out, structure. Indeed, the observations and the
numerical simulations are always realized with a finite resolution. Moreover, the coarse-grained DF f (r, v, t) is likely
to converge towards a steady state f (r, v) (metaequilibrium state) contrary to the exact distribution f (r, v, t) which
develops smaller and smaller scales for all times.
If we decompose the DF and the potential in a mean and fluctuating part (f = f + δf , Φ = Φ + δΦ) and take the
local average of the Vlasov equation (1), we readily obtain an equation for the coarse-grained DF of the form

∂f ∂f ∂f ∂
+v· − ∇Φ · =− · Jf (30)
∂t ∂r ∂v ∂v
with a diffusion current Jf = −δf ∇δΦ related to the correlations of the fine-grained fluctuations. The right hand
side of Eq. (30) can be interpreted as an effective “collision” term. Any systematic calculation of the diffusion current
starting from the Vlasov equation must necessarily introduce an evolution equation for the fluctuation δf . This
equation is simply obtained by subtracting Eq. (30) from Eq. (1). This yields

∂δf ∂δf ∂δf ∂f ∂δf ∂δf

+v· − ∇Φ · = ∇δΦ · + ∇δΦ · − ∇δΦ · . (31)
∂t ∂r ∂v ∂v ∂v ∂v
 9

Equations (30) and (31) are exact since no approximation has been made for the moment. To go further, we need
to implement some approximations. In the sequel, we shall develop a quasilinear theory which was introduced by
Kadomtsev and Pogutse [40] by analogy with the quasilinear theory of collisional relaxation based on the Klimontovich
equation. This will provide a precise theoretical framework to analyze the process of collisionless relaxation in systems
with long-range interactions. The essence of the quasilinear theory is to assume that the fluctuations are weak and
neglect the nonlinear terms in Eq. (31) altogether. In that case, Eqs. (30) and (31) reduce to the coupled system

∂f ∂f ∂f ∂
+v· − ∇Φ · = · δf ∇δΦ, (32)
∂t ∂r ∂v ∂v

∂δf ∂δf ∂δf ∂f

+v· − ∇Φ · − ∇δΦ · = 0. (33)
∂t ∂r ∂v ∂v
Physically, these equations describe the coupling between a subdynamics (played here by the small scale fluctuations
δf ) and a macrodynamics (played by the coarse-grained DF f ). Due to the strong simplifications implied by the
neglect of nonlinear terms in Eq. (31), the quasilinear theory only describes the late quiescent stages of the violent
relaxation process, when the fluctuations have weaken (gentle relaxation). Although this is essentially an asymptotic
theory, it is of importance to develop this theory in detail since it provides an explicit expression of the effective
“collision operator” which appears on the coarse-grained scale.
If we restrict ourselves to spatially homogeneous distributions, the field −∇Φ vanishes. In that case, Eqs. (32) and
(33) reduce to the coupled equations

∂f ∂
= · δf ∇δΦ, (34)
∂t ∂v

∂δf ∂δf ∂f
+v· − ∇δΦ · = 0. (35)
∂t ∂r ∂v
We shall assume that the fluctuations evolve rapidly compared to the evolution of the coarse-grained fields, so that the
time variation of f and Φ can be neglected in the calculation of the collision term. This is similar to the Bogoliubov
ansatz in the collisional theory. Therefore, for the purpose of solving Eq. (35) and obtaining the correlation function
δf δΦ, we shall regard f (v) as constant in time. With this approximation, Eqs. (34) and (35) can be solved with
the aid of Fourier-Laplace transforms and the collision term can be explicitly calculated. The derivation proceeds
similarly to the derivation of the Lenard-Balescu equation from the Klimontovich equation in the collisional theory
(see, e.g., [12]). The collisional and collisionless kinetic theories are analogous because the Klimontovich equation
is formally similar to the Vlasov equation. However, the Klimontovich equation involves a DF which is a sum of
δ-functions while the Vlasov equation involves a continuous DF.

B. Dielectric function

The Fourier-Laplace transform of the fluctuations of the DF δf is defined by

Z +∞
dr
Z
˜
δ f (k, v, ω) = dt e−i(k·r−ωt) δf (r, v, t). (36)
(2π)d 0

The inverse transform is

dω i(k·r−ωt) ˜
Z Z
δf (r, v, t) = dk e δ f (k, v, ω), (37)
C 2π

where the Laplace contour C in the complex ω plane must pass above all poles of the integrand. Similar expressions
hold for the fluctuations of the potential δΦ. We note that, for periodic potentials, the integral over k is replaced by
a discrete summation over the different modes. If we take the Fourier-Laplace transform of Eq. (35), we find that

∂f
− δ fˆ(k, v, 0) − iω δ f˜(k, v, ω) + ik · v δ f˜(k, v, ω) − ik · δ Φ̃(k, ω) = 0, (38)
∂v
 10

where the first term is the spatial Fourier transform of the initial value of the fluctuations
dr −ik·r
Z
ˆ
δ f (k, v, 0) = e δf (r, v, 0). (39)
(2π)d

Equation (38) can be rewritten as

∂f
k · ∂v δ fˆ(k, v, 0)
δ f˜(k, v, ω) = δ Φ̃(k, ω) + , (40)
k·v−ω i(k · v − ω)

where the first term takes into account “collective effects”. The fluctuations of the potential are related to the
fluctuations of the DF by a convolution
Z
δΦ(r, t) = u(|r − r′ |)δf (r′ , v′ , t) dr′ dv′ . (41)

Taking the Fourier-Laplace transform of this equation, we obtain

Z
δ Φ̃(k, ω) = (2π)d û(k) δ f˜(k, v, ω) dv. (42)

Substituting Eq. (40) into Eq. (42), we find that the Fourier-Laplace transform of the fluctuations of the potential is
given by

û(k) δ fˆ(k, v, 0)
Z
δ Φ̃(k, ω) = (2π)d dv , (43)
ǫ(k, ω) i(k · v − ω)

where
∂f
k · ∂v
Z
ǫ(k, ω) = 1 − (2π)d û(k) dv (44)
k·v−ω

is the dielectric function. The Fourier-Laplace transform of the fluctuations of the DF is then given by Eq. (40)
with Eq. (43). The dispersion relation associated with the linearized Vlasov equation corresponds to ǫ(k, ω) = 0. It
determines the proper pulsations of the system. If the DF is Vlasov stable,12 then Im[ω] < 0 for all modes ω. In
particular, ǫ(k, ω) does not vanish when ω is real so that Eq. (43) is well-defined. If collective effects were neglected
in Eq. (40), we would obtain Eq. (43) with ǫ(k, ω) = 1. This shows that, because of collective effects, the bare
potential of interaction û(k) is replaced by a “dressed” potential û(k)/ǫ(k, ω) taking into account the polarization of
the medium.
We can use the foregoing equations to compute the effective collision term appearing on the right hand side of Eq.
(34). One has

dω dω ′ ′ i(k·r−ωt) i(k′ ·r−ω′ t) ˜

Z Z Z Z
′
δf ∇δΦ = dk dk ik e e δ f (k, v, ω)δ Φ̃(k′ , ω ′ ). (45)
C 2π C 2π

Using Eq. (40), we find that

∂f
k · ∂v δ fˆ(k, v, 0)δ Φ̃(k′ , ω ′ )
δ f˜(k, v, ω)δ Φ̃(k′ , ω ′ ) = δ Φ̃(k, ω)δ Φ̃(k′ , ω ′ ) + . (46)
k·v−ω i(k · v − ω)

As we shall see, the first term accounts for a diffusion and the second term accounts for a friction. Let us consider
these two terms separately.

12 The initial DF may be dynamically unstable or even unsteady but we focus on a regime of quiescent relaxation where the DF is
dynamically stable.
 11

C. Diffusion

From Eq. (43), we obtain

û(k)û(k ′ ) δ fˆ(k, v, 0)δ fˆ(k′ , v′ , 0)

Z
δ Φ̃(k, ω)δ Φ̃(k′ , ω ′ ) = −(2π)2d dvdv′ . (47)
ǫ(k, ω)ǫ(k′ , ω ′ ) (k · v − ω)(k′ · v′ − ω ′ )

To proceed further, we have to evaluate the correlation function δf (r, v, t)δf (r′ , v′ , t). Following Kadomtsev and
Pogutse [40] we shall assume that the mixing in phase space is sufficiently efficient that the scale of the kinematic
correlations is small with respect to the coarse-graining mesh size. In that case, we can write

δf (r, v, 0)δf (r′ , v′ , 0) = ǫdr ǫdv δ(r − r′ )δ(v − v′ )f2 (v), (48)

where ǫr and ǫv are the resolution scales in position and velocity respectively and
2
f2 ≡ (δf )2 = (f − f )2 = f 2 − f (49)

is the local variance of the fine-grained fluctuations. Note that ǫdr ǫdv can be interpreted as the hypervolume of a
macrocell in Lynden-Bell’s statistical theory.13 Taking the Fourier transform of Eq. (48), we get
1 d d
δ fˆ(k, v, 0)δ fˆ(k′ , v′ , 0) = ǫ ǫ δ(k + k′ )δ(v − v′ )f2 (v). (50)
(2π)d r v

Substituting Eq. (50) into Eq. (47), we find that

û(k)2 f2 (v)
Z
δ Φ̃(k, ω)δ Φ̃(k′ , ω ′ ) = (2π)d ǫdr ǫdv δ(k + k′ ) dv . (51)
ǫ(k, ω)ǫ(−k, ω ′ ) (k · v − ω)(k · v + ω ′ )

Considering only the contributions that do not decay in time, it can be shown [82] that [(k · v − ω)(k · v + ω ′ )]−1 can
be substituted by (2π)2 δ(ω + ω ′ )δ(k · v − ω).14 Then, using the property ǫ(−k, −ω) = ǫ(k, ω)∗ , one finds that the
correlations of the fluctuations of the potential are given by
2
d+2 d d û(k)
Z
′ ′ ′ ′
δ Φ̃(k, ω)δ Φ̃(k , ω ) = (2π) ǫr ǫv δ(k + k )δ(ω + ω ) δ(k · v − ω)f2 (v) dv. (52)
|ǫ(k, ω)|2

From Eq. (52), we get the contribution to Eq. (45) of the first term of Eq. (46). This yields the diffusion term

∂f
dω k · ∂v û(k)2
Z Z Z
(δf ∇δΦ)Diff
i = −i(2π)d+1 ǫdr ǫdv dk dv′ ki f2 (v′ )δ(k · v′ − ω). (53)
C 2π k · v − ω |ǫ(k, ω)|2

Using the Landau prescription ω → ω + i0+ and the Plemelj formula,

 
1 1
=P ∓ iπδ(x), (54)
x ± i0+ x

where P denotes the principal value, we can replace 1/(k · v − ω − i0+ ) by +iπδ(k · v − ω). Then, integrating over
ω, we obtain

û(k)2 ∂f
Z
(δf ∇δΦ)Diff
i = π(2π)d ǫdr ǫdv dk dv′ ki kj δ[k · (v − v′ )]f2 (v′ ) (v). (55)
|ǫ(k, k · v)|2 ∂vj

13 The Lynden-Bell entropy can be obtained from a combinatorial analysis by dividing the phase space into macrocells and microcells and
by counting the number of microstates associated with a given macrostate. The mixing entropy (9) is equal to the logarithm of this
number [22, 39, 55].
14 See also Appendix A of [83] for a more precise justification of this procedure through a detailed calculation of the integral (Laplace
transform) obtained by substituting Eq. (46) with Eq. (51) into Eq. (45).
 12

D. Friction

Proceeding similarly, we obtain

δ fˆ(k, v, 0)δ Φ̃(k′ , ω ′ ) û(k ′ ) 1 δ fˆ(k, v, 0)δ fˆ(k′ , v′ , 0)

Z
= (2π)d dv′ (56)
i(k · v − ω) ǫ(k′ , ω ′ ) i(k · v − ω) i(k′ · v′ − ω ′ )

and

δ fˆ(k, v, 0)δ Φ̃(k′ , ω ′ ) û(k ′ ) 1 f2 (v)

= ǫdr ǫdv δ(k + k′ ) . (57)
i(k · v − ω) ǫ(k′ , ω ′ ) i(k · v − ω) i(k′ · v − ω ′ )

Considering only the contributions that do not decay in time, it can be shown [82] (see footnote 14 with Eq. (51)
replaced by Eq. (57)) that Eq. (57) can be substituted by

δ fˆ(k, v, 0)δ Φ̃(k′ , ω ′ ) û(k ′ )

= (2π)2 ǫdr ǫdv δ(k + k′ )δ(ω + ω ′ )δ(k · v − ω)f2 (v). (58)
i(k · v − ω) ǫ(k′ , ω ′ )

From Eq. (58), we get the contribution to Eq. (45) of the second term of Eq. (46). This yields the friction term

û(k)
Z
(δf ∇δΦ)Fric
i = ǫ d d
ǫ
r v dk ki Im ǫ(k, k · v)f2 (v). (59)
|ǫ(k, k · v)|2

Using the Landau prescription ω → ω + i0+ and the Plemelj formula (54), the imaginary part of the dielectric function
(44) reads

∂f
Z
Im ǫ(k, ω) = −π(2π)d û(k) k· δ(k · v − ω) dv. (60)
∂v

Substituting this expression into Eq. (59), we obtain

û(k)2 ∂f
Z
(δf ∇δΦ)Fric
i = −π(2π)d ǫdr ǫdv dk dv′ ki kj 2
δ[k · (v − v′ )]f2 (v) ′ (v′ ). (61)
|ǫ(k, k · v)| ∂v j

E. Collision term: Kadomtsev-Pogutse (1970) equation

Regrouping Eqs. (34), (55) and (61), we end up with the kinetic equation
′
!
∂f ∂ û(k)2 ∂f ∂f
Z
= π(2π)d ǫdr ǫdv ′
dk dv ki kj δ[k · (v − v′ )] f2′ − f2 ′ , (62)
∂t ∂vi |ǫ(k, k · v)|2 ∂vj ∂v j

′
where f = f (v, t), f = f (v′ , t), f2 = f2 (v, t), and f2′ = f2 (v′ , t). This equation, which takes collective effects into
account, was first derived by Kadomtsev and Pogutse [40].15 It is formally similar to the Lenard-Balescu equation of
the collisional theory (see, e.g., [12]), except that mf in the Lenard-Balescu equation is replaced by ǫdr ǫdv f2 in the KP
equation. We also note that, contrary to the Lenard-Balescu equation, the KP equation is not closed in the general
case since it involves the variance f2 of the fine-grained distribution. It can be closed exactly only in the two-level
case (see Sec. IV C).
If we neglect collective effects and take |ǫ(k, k · v)| = 1, Eq. (62) reduces to
′
!
∂f ∂ ∂f ∂f
Z
= π(2π)d ǫdr ǫdv dk dv′ ki kj û(k)2 δ[k · (v − v′ )] f2′ − f2 ′ . (63)
∂t ∂vi ∂vj ∂v j

15 It is closely related to an equation previously derived by Dupree [84].

 13

The integral over k can be performed explicitly (see, e.g., [28]) and we obtain
′
!
∂f ∂ w2 δij − wi wj ∂f ∂f
Z
= Kd dv′ f2′ − f2 ′ , (64)
∂t ∂vi w3 ∂vj ∂v j
R +∞
where w = v − v′ is the relative velocity and Kd is a constant with value K3 = 8π 5 ǫ3r ǫ3v 0 k 3 û(k)2 dk in d = 3 and
R +∞
K2 = 8π 3 ǫ2r ǫ2v 0 k 2 û(k)2 dk in d = 2. This kinetic equation, that neglects collective effects, is formally similar to the
Landau equation [10] in the collisional theory (see, e.g., [12]) with the substitution mf → ǫdr ǫdv f2 . For a 3D plasma,
R +∞
using (2π)3 û(k) = 4πe2 /m2 k 2 , we get K3 = (2πe4 /m3 ) ln Λ where ln Λ = 0 dk/k is the Coulomb logarithm that
has to be regularized with appropriate cut-offs. The large-scale cut-off is the Debye length λD (the Debye length
appears naturally when we take into account collective effects) and the small-scale cut-off is the spatial resolution
scale ǫr which replaces the Landau length λL in the collisional theory. This yields ln Λ = ln(λD /ǫr ). For a 2D plasma,
using (2π)2 û(k) = 2πe2 /m2 k 2 and introducing a large-scale cut-off at the Debye length, we obtain K2 = 2πe4 /m3 kD .
There is no need to introduce a small-scale cut-off in that case but the integration should be stoped at ǫr in principle.
Returning to Eqs. (62) and (63), we note that collective effects can be taken into account simply by replacing the
bare potential û(k) in the Landau equation by a “dressed” potential ûd (k) = û(k)/|ǫ(k, k · v)|, including the dielectric
function, without changing the overall structure of the kinetic equation. Physically, this means that the particles are
“dressed” by their polarization cloud. In plasma physics, collective effects are important because they account for
screening effects and regularize, at the scale of the Debye length, the logarithmic divergence that occurs in the Landau
equation. This avoids the introduction of ad hoc cut-offs at large scales.
We can write the kinetic equation (62) in the compact form
′
!
∂f ∂ ′ ∂f ∂f
Z
′
= dv Kij f2 − f2 ′ (65)
∂t ∂vi ∂vj ∂v j

by introducing the tensor

û(k)2
Z
Kij = π(2π)d ǫdr ǫdv dk ki kj δ[k · (v − v′ )]. (66)
|ǫ(k, k · v)|2

We note that it satisfies the identity Kij wj = 0. If we neglect collective effects, the tensor Kij is explicitly given by
(see, e.g., [28])
 
bare 1 wi wj
Kij = Kd δij − . (67)
w w2

We also note that the kinetic equation (65) has the structure of a generalized Fokker-Planck equation involving a
diffusion term and a friction term.
Remark: An equation similar to the KP equation has been introduced in 2D turbulence [45, 46] (see also Eq. (148)
of [85] and Eq. (88) of [86]). It reads

1 ∂ X +∞ ′ ′ 1 ∂ω′
 
∂ω 1 ∂ω
Z
= 2π 2 ǫ2r r dr |n||G(n, r, r′ , nΩ)|2 δ(Ω − Ω′ ) ω2′ − ω2 ′ ′ . (68)
∂t r ∂r n 0 r ∂r r ∂r

F. 3D self-gravitating systems

Systems with attractive long-range interactions are generically spatially inhomogeneous. It is important to develop
a kinetic theory of collisionless relaxation for such systems. If we implement a quasilinear approximation and neglect
collective effects, we can derive a generalized kinetic equation of the form [42–44]
′
Z t !
∂f ∂f ∂f ∂ ∂f ∂f
Z
+v· − ∇Φ · = ǫdr ǫdv ds dr′ dv′ Fi (r′ → r)t Fj (r′ → r)t−s f2′ − f2 ′ (69)
∂t ∂r ∂v ∂vi 0 ∂vj ∂v j
t−s

that is valid for systems that are not necessarily spatially homogeneous and not necessarily Markovian. Here, f =
′
f (r, v, t), f = f (r′ , v′ , t), f2 = f2 (r, v, t) and f2′ = f2 (r′ , v′ , t). On the other hand, F(r′ → r)t denotes the force by
 14

unit of mass exerted by a particle located in r′ on a particle located in r at time t. In certain cases, e.g. for 3D
self-gravitating systems, we can make a local approximation and proceed as if the system were spatially homogeneous.
If we also implement a Markovian approximation, the kinetic equation (69) is replaced by
′
Z +∞ !
∂f ∂f ∂f 3 3 ∂ ′ ∂f ∂f
Z
′ ′ ′ ′
+v· − ∇Φ · = ǫr ǫv ds dr dv Fi (r → r)t Fj (r → r)t−s f2 − f2 ′ , (70)
∂t ∂r ∂v ∂vi 0 ∂vj ∂v j
t

′
where now f = f (r, v, t), f = f (r, v′ , t), f2 = f2 (r, v, t) and f2′ = f2 (r, v′ , t). Passing in Fourier space, we obtain [28]
′
!
∂f ∂f ∂f ∂ ∂f ∂f
Z
+v· − ∇Φ · = π(2π)3 ǫ3r ǫ3v dk dv′ ki kj û(k)2 δ[k · (v − v′ )] f2′ − f2 ′ . (71)
∂t ∂r ∂v ∂vi ∂vj ∂v j

In this equation, the effects of spatial inhomogeneity are kept only in the advection (Vlasov) term, while the collision
term is calculated as if the system were spatially homogeneous. The integral over k can be performed explicitly [28]
and the foregoing equation can be rewritten as
′
!
2
∂f ∂f ∂f ∂ ′ w δij − wi wj ′ ∂f ∂f
Z
+v· − ∇Φ · = K3 dv f2 − f2 ′ . (72)
∂t ∂r ∂v ∂vi w3 ∂vj ∂v j
R +∞
For a 3D self-gravitating system, using (2π)3 û(k) = −4πG/k 2 , we get K3 = 2πǫ3r ǫ3v G2 ln Λ where ln Λ = 0 dk/k is
the Coulomb factor that has to be regularized with appropriate cut-offs. The large-scale cut-off is the Jeans length λJ
(which is of the order of the system size R) and the small-scale cut-off is the spatial resolution scale ǫr which replaces
the gravitational Landau length λL in the collisional theory. This yields ln Λ = ln(λJ /ǫr ).
Remark: We can also extend the KP equation to spatially inhomogeneous systems by making a local approximation.
In this manner, we can heuristically take collective effects into account. In Sec. VII, we introduce more general
kinetic equations written with angle-action variables that take into account spatial inhomogeneity and collective
effects without relying on a local approximation.

IV. MULTI-LEVEL CASE

In the general case, the kinetic equation (62) for the coarse-grained DF f (v, t) is not closed because it depends on
the local centered variance f2 (v, t) of the distribution ρ(v, η, t). It is therefore necessary to extend the kinetic theory
to the multi-level case and work in term of the DF ρ(v, η, t) for each level.

A. Severne-Luwel (1980) equation

A kinetic equation for ρ(v, η, t) has been derived by Severne and Luwel [41] by generalizing the quasilinear theory
of Kadomtsev and Pogutse [40] to the multi-level case.16 They obtained an equation of the form17
′
" #
∂ρ ∂ ′ ∂ρ ∂f
Z
′
= dv Kij f2 − ρ(η − f ) ′ . (74)
∂t ∂vi ∂vj ∂v j
R
We can check that the Rnormalization condition ( ρ dη = 1) is preserved with time and that the equation for the
coarse-grained DF f = ρη dη returns Eq. (65). We can also show that the SL equation conserves the energy and all

16 The derivation is similar to the one detailed in Sec. III. In the multi-level case, one has to work in terms of ρe (r, v, η, t) = δ(f (r, v, t)− η)
and ρ(v, η, t) = hδ(f (r, v, t) − η)i which generalize f (r, v, t) and f (v, t), respectively. Equation (48) is generalized into

δρ(r, v, η, t)δρ(r′ , v′ , η′ , t′ ) = ǫdr ǫdv ρ(r, v, η′ , t)[δ(η − η′ ) − ρ(r, v, η, t)]δ(r − r′ )δ(v − v′ ), (73)
corresponding to Eq. (3.13) of [41].
17 Severne and Luwel [41] developed their theory for spatially inhomogeneous stellar systems. However, at the end of their calculations,
in order to obtain an explicit kinetic equation, they made a local approximation which amounts to proceeding as if the system were
infinite and homogeneous (see Sec. III F).
 15

the Casimirs, that it satisfies and H-theorem for the Lynden-Bell entropy (9), and that it relaxes towards the Gibbs
state (11) (see Appendix E).
From the SL equation, we can derive a hierarchy of equations for the moments f n = ρη n dη of the distribution.
R
The general term of this hierarchy is
′
" #
∂f n ∂ ′ ∂f
n ∂f
Z
′ n+1 n
= dv Kij f2 − (f − f f) ′ . (75)
∂t ∂vi ∂vj ∂v j

For n = 1 we recover the KP equation [Eq. (65)].

Remark: An equation similar to the SL equation can be derived in 2D turbulence [46]. It reads

∂ρ X Z +∞ 
1 ∂ω′

2 21 ∂ ′ ′ ′ 2 ′ ′ 1 ∂ρ
= 2π ǫr r dr |n||G(n, r, r , nΩ)| δ(Ω − Ω ) ω2 − ρ(σ − ω) ′ ′ . (76)
∂t r ∂r n 0 r ∂r r ∂r

B. Chavanis-Sommeria-Robert (1996) equation

By using a different approach based on a Maximum Entropy Production Principle (MEPP), Chavanis, Sommeria
and Robert [26] have proposed an equation for ρ(v, η, t) that relaxes towards the Lynden-Bell distribution.18 This
phenomenological equation is expected to describe the whole process of violent relaxation, including the very nonlinear
early regime. In the late regime of quiescent relaxation, it is possible to connect the CSR equation to the KP and SL
equations as follows.19
The kinetic equation (74) obtained from the quasilinear theory of the Vlasov equation is an integrodifferential
equation. The current in v depends on the value of f and f2 in v′ through an integral over v′ . We can transform
′
the integrodifferential equation (74) into a differential equation by replacing f = f (v′ , t) and f2′ = f2 (v′ , t) by their
equilibrium values obtained from the Gibbs state (11). This amounts to making a thermal bath approximation. Using
the identity from Eq. (16), we find that
′
∂f
= −βf2′ v′ . (77)
∂v′
Substituting Eq. (77) into Eq. (74) we get
 
∂ρ ∂ ∂ρ
Z
= dv′ Kij f2′ + βf2′ ρ(η − f )vj′ . (78)
∂t ∂vi ∂vj

In principle, f2′ should be calculated at equilibrium. However, in order to be more general, we shall evaluate f2 in
Eq. (78) at time t, not at equilibrium. Using the identity Kij (vj − vj′ ) = 0, we can replace vj′ by vj in the last term
of Eq. (78). In this manner, we obtain
  
∂ρ ∂ ∂ρ
= Dij + βρ(η − f )vj (79)
∂t ∂vi ∂vj

with
Z
Dij = dv′ Kij f2′ . (80)

This equation can be viewed as a generalized Fokker-Planck equation of the Kramers type involving a diffusion term
and a friction term. The friction term is the counterpart of Chandrasekhar’s dynamical friction [87] in the kinetic
theory of collisional stellar systems. The friction coefficient is given by a form of Einstein relation ξij = Dij βη.
By making the thermal bath approximation from Eq. (77), we have lost the conservation of energy. Indeed, we have
passed from a microcanonical description (conservation of energy E and H-theorem for the entropy S) to a canonical

18 The MEPP can be formulated in the inhomogeneous case, as discussed in Appendix F, but in this section we consider spatially
homogeneous systems.
19 This connection was made in [42, 43] in the two-level case and is extended here to the multi-level case.
 16

description (fixed temperature T and H-theorem for the free energy F = E − T S). However, the conservation of
energy can be artificially restored by letting β(t) depend on time
R in a suitable manner. To that purpose, using Eq.
(79), we first write the equation for the coarse-grained DF f = ρη dη which reads
  
∂f ∂ ∂f
= Dij + βf2 vj . (81)
∂t ∂vi ∂vj

This equation can also be obtained by applying the thermal bath approximation (77) to Eq. (65). Then, we compute

∂f v 2 ∂Jf v 2
 
∂f
Z Z Z Z
Ė = dv = − dv = Jf · v dv = − Dij + βf2 vj vi dv. (82)
∂t 2 ∂v 2 ∂vj

We can enforce the conservation of energy (Ė = 0) by taking

∂f
R
Dij vi ∂vj
dv
β(t) = − R . (83)
Dij f2 vi vj dv

This approach returns the CSR equations formed by Eqs. (79) and (83) above. It also provides the explicit expression
(80) of the diffusion coefficient – actually a tensor Dij – which was not given by the MEPP [26]. Equation (79) with
Eqs. (80) and (83) is an integrodifferential equation since β and Dij are expressed as integrals of f and f2 , but it
is simpler than the KP and SL equations. We can show (see Ref. [26] and Appendix G) that the CSR equations
conserve the energy and all the Casimirs, and that they monotonically increase the Lynden-Bell entropy (H-theorem).
They usually relax towards the Lynden-Bell DF except in the cases reported in Sec. V where the diffusion tensor Dij
vanishes.
From the CSR equation (79), we can derive a hierarchy of equations for the moments f n = ρη n dη of the
R
distribution. The general term of this hierarchy is

∂f n
  n 
∂ ∂f
= Dij + β(f n+1 − f n f )vj . (84)
∂t ∂vi ∂vj

For n = 1 we recover Eq. (81). This hierarchy of equations can also be obtained by applying the thermal bath
approximation (77) to Eq. (75).
Remark: In the context of 2D turbulence, the MEPP has been introduced by Robert and Sommeria [47]. It can
be viewed as a variational formulation of the linear thermodynamics of Onsager [88–90] (see [51, 91]). The relaxation
equations for the coarse-grained vorticity involves a diffusion term and a drift term [26, 47, 49]. The drift is the
counterpart of Chandrasekhar’s dynamical friction [20, 85, 92]. The moments of the relaxation equation have been
derived in [48, 61, 62, 93]. The connection between the quasilinear theory of the 2D Euler equation and the MEPP
has been discussed by Chavanis [45]. This connection allowed us to compute the diffusion coefficient of 2D vortices
(not given by the MEPP) and recover the heuristic expression given in Ref. [48]. A more detailed discussion of the
quasilinear theory of the 2D Euler equation is given in [46].

C. Two-level case: Fermionic-like equations

If the initial condition in phase space consists of patches of uniform DF f = η0 surrounded by vacuum f = 0
(two-level approximation), we can write f 2 = η0 ×f = η0 f yielding

f2 = f (η0 − f ). (85)

In that case, the local centered variance f2 of the distribution can be trivially related to the coarse-grained DF f . We
see how the self-correlations and the incompressibility of the flow in phase space give rise to an effective “exclusion
principle”. Substituting Eq. (85) into Eq. (65), we obtain a kinetic equation of the form
′
" #
∂f ∂ ′ ∂f ∂f
Z
′ ′
= dv Kij f (η0 − f ) − f (η0 − f ) ′ . (86)
∂t ∂vi ∂vj ∂v j

This equation is similar to the fermionic Landau (or Lenard-Balescu) equation [40–44, 94]. It conserves the mass
and the energy and monotonically increases the Lynden-Bell (or Fermi-Dirac-like) entropy [94]. In the nondegenerate
 17

limit f ≪ η0 , it becomes similar to the classical Landau (or Lenard-Balescu) equation associated with the Boltzmann
entropy.
On the other hand, the CSR equations (79), (80) and (83) reduce to
  
∂f ∂ ∂f
= Dij + β(t)f (η0 − f )vj (87)
∂t ∂vi ∂vj

with
Z
′ ′
Dij = dv′ Kij f (η0 − f ) (88)

and
∂f
R
Dij vi ∂vj
dv
β(t) = − R . (89)
Dij f (η0 − f )vi vj dv

Equation (87) is similar to the fermionic Kramers equation. It can be obtained from Eq. (86) by using the thermal
bath approximation (77). This equation has been derived and studied in [42–44, 94]. It conserves the mass and the
energy and monotonically increases the Lynden-Bell (or Fermi-Dirac-like) entropy. In the nondegenerate limit f ≪ η0 ,
it becomes similar to the classical Kramers equation [22].

D. Chavanis (2004) equation

For more complicated initial conditions (multi-level case), we have to solve the equation for ρ(r, v, η, t), the proba-
bility density of finding the phase level η in (r, v) at time t [see Eq. (74) or Eq. (79)]. The strategy is to discretize the
initial condition into N levels η. This approach then leads us to a closed system of N coupled equations (one for each
level η). However, for generic initial conditions, we have to deal with a great number of levels and these equations are
not convenient to solve when N ≫ 1. We can alternatively try to solve the hierarchy of equations for the moments
f n [see Eq. (75) or Eq. (84)] but we then encounter a difficult closure problem. In practice, we are mainly interested
in the evolution of the first moment, namely the coarse-grained DF f . As we have seen, the equation for f [see Eq.
(65) or Eq. (81)] depends on the variance f2 . In order to obtain a self-consistent kinetic equation for f , we need
to relate the variance f2 to the coarse-grained DF f . In Ref. [44], we have proposed a closure approximation that
leads to a simple kinetic equation. While not being exact, this equation preserves the robust features of the process
of violent relaxation and is amenable to an easier numerical implementation. Its main interest is to go beyond the
two-level approximation while leaving the problem tractable. The idea is to observe that Eqs. (16) and (27) lead to
the important relation [44, 50]
1
f2 = . (90)
C ′′ (f )

This relation is valid at equilibrium but we propose to use it as a closure approximation in Eqs. (75) and (84). This is
expected to be a reasonable approximation if we are close to equilibrium, which is in fact dictated by the quasilinear
approximation. Of course, this procedure assumes that we know the function C(f ) in advance. This is the case if
we have already determined the equilibrium state by the procedure discussed in Sec. II D and we want to describe
the dynamics close to equilibrium.20 A justification of the closure relation (90) is given in Appendix C 4 following an
argument first given in Ref. [59].
If we close the hierarchy of equations (75) with Eq. (90), we obtain a self-consistent kinetic equation of the form
[44]
′
" #
∂f ∂ 1 ∂f 1 ∂f
Z
′
= dv Kij ′ − ′′ . (91)
∂t ∂vi C ′′ (f ) ∂vj C (f ) ∂v ′ j

20 This is also the case when the generalized entropy C(f ) is determined by the external forcing as discussed in footnotes 8 and 9.
 18

It can be viewed as a generalized Landau (or Lenard-Balescu) equation. It conserves the mass and the energy and
monotonically increases the generalized entropy (22) (H-theorem) [44]. In the two-level case, using f2 = 1/C ′′ (f ) =
f (η0 − f ), we recover Eq. (86). The generalized Landau equation (91) has been studied in detail in Ref. [44].
If we close the hierarchy of equations (84) with Eq. (90), we obtain a self-consistent kinetic equation of the form
[50]
  
∂f ∂ ∂f β(t)
= Dij + vj (92)
∂t ∂vi ∂vj C ′′ (f )

with
1
Z
Dij = dv′ Kij ′ (93)
C ′′ (f )

and
∂f
R
Dij vi ∂vj
dv
β(t) = − R . (94)
Dij C ′′1(f ) vi vj dv

Equation (92) can also be derived from Eq. (91) by making the thermal bath approximation (77). It can be viewed
as a generalized Kramers equation. It conserves the mass and the energy and monotonically increases the generalized
entropy (22) (H-theorem) [50]. In two-level case, using f2 = 1/C ′′ (f ) = f (η0 − f ), we recover Eqs. (87)-(89). The
generalized Kramers equations (92)-(94) have been studied in detail in Ref. [50].
Remark: Similar equations have been introduced in the context of 2D turbulence in Refs. [50, 55, 59–62].

E. One-dimensional systems

For one-dimensional (1D) systems, recalling the expression of Kij from Eq. (66), the SL equation (74) becomes
′
" #
∂ρ ∂ û(k)2 ′ ∂ρ ∂f
Z
2 ′ 2 ′
= 2π ǫr ǫv dv dk k δ[k(v − v )] f2 − ρ(η − f ) ′ . (95)
∂t ∂v |ǫ(k, kv)|2 ∂v ∂v

1
Using the identity δ[k(v − v ′ )] = |k| δ(v − v ′ ), it can be rewritten as

û(k)2
 
∂ρ ∂ ∂ρ ∂f
Z
= 2π 2 ǫr ǫv dk |k| f2 − ρ(η − f ) . (96)
∂t ∂v |ǫ(k, kv)|2 ∂v ∂v

The corresponding hierarchy of moment equations takes the form

∂f n û(k)2 ∂f n
 
∂ ∂f
Z
2 n+1 n
= 2π ǫr ǫv dk |k| f2 − (f − f f) . (97)
∂t ∂v |ǫ(k, kv)|2 ∂v ∂v

For n = 1 we get

∂f
= 0. (98)
∂t
This result can be directly obtained from the KP equation (62) applied to 1D systems. It implies that the coarse-
grained DF does not change with time. However, the higher moments evolve in time.
Remark: We note that the “kinetic blocking” of the coarse-grained DF in 1D occurs only for spatially homogeneous
systems. For 1D inhomogeneous systems, the coarse-grained DF evolves in time according to the inhomogeneous KP
equation (108). On the other hand, for 1D homogeneous systems, the CSR equation (81) does not show such a “kinetic
blocking”. This may be related to the fact that this equation is more justified in the phase of violent relaxation than
in the phase of quiescent relaxation.
 19

V. INCOMPLETE VIOLENT RELAXATION

We have seen that the kinetic equation derived from the quasilinear theory relaxes towards the Lynden-Bell distri-
bution. In this sense, it provides a justification of the maximum entropy principle and implies that the evolution is
ergodic. However, direct numerical simulations of the Vlasov equation (or direct simulations of the N -body problem
performed in the collisionless regime) show that violent relaxation is in general incomplete [22, 37–39]. The fluctua-
tions of the potential that are the engine of the collisionless relaxation can die out before the system has reached the
statistical equilibrium equilibrium state. How can we reconcile these apparently contradictory results?
First, we have to recall that the quasilinear theory, which is based on the assumption that the nonlinear terms in
the equation for the fluctuations can be neglected, describes only a regime of late quiescent relaxation. Therefore,
the relaxation toward the Lynden-Bell DF may be limited to this “gentle” situation. In addition, we have assumed
that the correlation function is given by Eq. (48) in which the resolution scales ǫr and ǫv in position and velocity are
constant in time. This is also a strong assumption. At the end of their paper, Kadomtsev and Pogutse [40] argue
that the scale of correlations may decrease in time as the variations of the potential Φ decay. In that case, the kinetic
equation becomes
′
" #
2
∂ρ ∂ û(k) ∂ρ ∂f
Z
= π(2π)d ǫr (t)d ǫv (t)d dv′ dk ki kj δ[k · (v − v′ )] f2′ − ρ(η − f ) ′ , (99)
∂t ∂vi |ǫ(k, k · v)|2 ∂vj ∂v j

where A(t) ≡ ǫr (t)d ǫv (t)d tends to zero for t → +∞. If the scale of correlations decreases rapidly in time,21 the
relaxation towards the Lynden-Bell distribution may be inhibited. This effect may account for incomplete relaxation.
The heuristic CSR approach [26] aims at describing the very nonlinear regime of violent relaxation. It leads to a
relaxation equation of the form
  
∂ρ ∂ρ ∂ρ ∂ ∂ρ
+v· − ∇Φ · = · D(r, v, t) + β(t)(η − f )ρv (100)
∂t ∂r ∂v ∂v ∂v
with a diffusion coefficient D(r, v, t) given by Eq. (80). The diffusion coefficient depends on the local centered variance
f2 of the distribution. Therefore, it vanishes in the regions of phase space where there are no fluctuations. Said
differently, the fluctuations δΦ of the potential must be strong enough to provide an efficient mixing. The vanishing of
the diffusion coefficient can “freeze” the system in a subdomain of phase space and account for incomplete relaxation
and non-ergodicity [26].22 For the same reason, the KP equation may also experience a process of incomplete relaxation
when f2 → 0.
This kinetic justification of incomplete relaxation is interesting because it is not based on a generalized entropy
such as the Tsallis entropy (see Sec. II F), so it does not involve any free parameter like q [81]. However, it demands
to solve a dynamical equation [Eq. (99) or Eq. (100) with Eqs. (80) and (83)] in order to predict the incompletely
mixed equilibrium state reached by the system. The idea is that, in case of incomplete relaxation (non-ergodicity),
the prediction of the equilibrium state is impossible without considering the dynamics [38].
Finally, the kinetic theory allows us to take into account the evaporation of high energy particles, which also prevents
the relaxation of the system towards a true (Lynden-Bell) statistical equilibrium state.23 Truncated models such as
the fermionic King model have been derived in Refs. [42, 44]. These truncated DFs differ from the Lynden-Bell DF
and have a finite mass contrary to the Lynden-Bell DF for 3D self-gravitating systems.
Remark: Similar arguments have been developed in 2D turbulence to explain the process of incomplete violent
relaxation [26, 48, 86].

VI. COLLISIONLESS VERSUS COLLISIONAL RELAXATION

In this section, we discuss the collisionless relaxation time associated with the KP equation (62) and compare it
with the collisional relaxation time associated with the ordinary Lenard-Balescu equation [12]. Our discussion follows

21 Kadomtsev and Pogutse [40] argue that A(t) decreases like t−1/2 or even more rapidly.
22 This “freezing” has been observed numerically in 2D turbulence [48] and led to the concept of maximum entropy “bubbles” [95]. It is
not clear if the vanishing of the diffusion coefficient in certain regions of phase space completely stops the relaxation or simply slows it
down. However, if the relaxation is strongly slowed down (as observed in [48]) the result is essentially the same on a practical point of
view.
23 Evaporation is particularly important for 3D self-gravitating systems since there is no statistical equilibrium state in a strict sense. The
Lynden-Bell DF coupled to the Poisson equation has an infinite mass.
 20

and completes the discussion given in Refs. [40–42].

The Lenard-Balescu equation describing the evolution of the system sourced by finite N effects (collisions) is (see,
e.g., [12])
û(k)2 ∂f ′
 
∂f ∂ ′ ∂f
Z
= π(2π)d m dk dv′ ki kj δ[k · (v − v ′
)] f − f . (101)
∂t ∂vi |ǫ(k, k · v)|2 ∂vj ∂v ′ j
As discussed previously, we can obtain the KP equation (62) describing the collisionless relaxation of the system by
making the substitution mf → ǫdr ǫdv f2 in the right hand side of Eq. (101). In the two-level case, this substitution
becomes mf → ǫdr ǫdv (η0 − f )f . In the nondegenerate limit f ≪ η0 , it reduces to mf → ǫdr ǫdv η0 f , i.e., m → ǫdr ǫdv η0 . In
other words, we have to replace the mass m of the particles by the effective mass
meff = ǫdr ǫdv η0 , (102)
which is the mass of a completely filled macrocell. In the terminology of Dupree [84] and Kadomtsev and Pogutse
[40], this can be viewed as the effective mass of “macroparticles” or “clumps”, i.e., correlated regions. Using the fact
that tR ∝ 1/m [12], the ratio between the collisionless relaxation time t∗R and the collisional relaxation time tR is
t∗R m
= . (103)
tR meff
The quantity m/meff is the so-called reduction factor. In general meff ≫ m so that the reduction of the relaxation
time can be quite large. As a result, the collisionless relaxation takes place on a timescale which is much smaller than
the collisional relaxation time in agreement with the Lynden-Bell concept of violent relaxation.
In d ≥ 2, the collisional relaxation time scales as24
tR ∼ N tD , (104)
where tD is the dynamical time. Using Eq. (103), we find that the collisionless relaxation time scales as
Nm
t∗R ∼ tD ∼ N tD , (105)
meff
where N = N m/meff represents the number of completely filled macrocells or, equivalently, the number of macropar-
ticles. This number is usually much smaller than the number of particles (N ≪ N ). This is another manner to
understand why the collisionless relaxation is much shorter than the collisional relaxation. In general, the collisionless
relaxation time is equal to a few dynamical times. We recall, however, that we need t∗R ≫ tD for the validity of the
Markovian description. This is typically the case when N > ∼ 10 − 100.
The fact that the kinetic equation (62) relaxes towards the Lynden-Bell DF on a few dynamical times is interpreted
by Kadomtsev and Pogutse [40] in terms of “collisions” between macroparticles with a large effective mass meff ∼
ǫdr ǫdv η0 . For partially degenerate systems, the effective mass of the macroparticles scales as meff = ǫdr ǫdv (η0 − f ) or,
more generally, as meff = ǫdr ǫdv f2 /f . It is therefore substantially reduced. Accordingly, the collisionless relaxation time
increases.25 This is because the fluctuations that drive the collisionless relaxation are less effective in establishing a
statistical equilibrium state. As a result, the system can be frozen for a long time in a metaequilibrium state which
is not the most mixed state. This kinetic blocking can account for the process of incomplete relaxation as discussed
in Sec. V.
Remark: In the case of 3D plasmas and 3D stellar systems we have to account for logarithmic corrections. In that
case, we get
 
λ
ln D,Jǫ r
meff = ǫdr ǫdv η0  . (106)
λD,J
ln λL

Detailed estimates of the relaxation time are given in [12, 94, 96, 97]. Similar results are obtained in 2D turbulence
(see Sec. 4.2 of Ref. [86]).

24 In the case of 3D plasmas and 3D stellar systems we have to account for logarithmic corrections yielding tR ∼ [N/ ln(λD,J /λL )] tD . In
d = 1, the KP and Lenard-Balescu operators vanish for spatially homogeneous systems, implying that the relaxation is longer (see Sec.
IV E). In that case, the collisional relaxation time scales as N 2 tD [12, 14, 15].
25 Similarly, the collisional relaxation time of self-gravitating fermions is larger than the collisional relaxation time of self-gravitating
classical particles [94].
 21

VII. KINETIC THEORY OF QUIESCENT COLLISIONLESS RELAXATION FOR SPATIALLY

INHOMOGENEOUS SYSTEMS

We can easily extend the quasilinear theory of quiescent collisionless relaxation to the case of spatially inhomoge-
neous systems by introducing angle-action variables and using the formalism developed in Ref. [9]. The inhomogeneous
Lenard-Balescu equation describing the evolution of the system sourced by finite N effects (collisions) is [8, 9]
∂f ′
 
∂f ∂ X ∂f
Z
= π(2π)d m · dJ′ k |Adk,k′ (J, J′ , k · Ω)|2 δ(k · Ω − k′ · Ω′ ) f ′ k · − f k′ · , (107)
∂t ∂J ∂J ∂J′
′ k,k

where Adk,k′ (J, J′ , ω) is the dressed potential of interaction (it is written as −1/Dk,k′ (J, J′ , ω) in Refs. [8, 9]) and
Ω(J) is the pulsation of the orbit of a particle with action J. This equation can be obtained from a quasilinear
theory based on the Klimontovich equation [9]. As we have previously explained, the quasilinear theory based on
the Vlasov equation is similar to the quasilinear theory based on the Klimontovich equation provided that we make
the substitution mf → ǫdr ǫdv f2 . As a result, the inhomogeneous KP equation describing the quiescent collisionless
relaxation of the system is
′
!
∂f d d d ∂
XZ
′ d ′ 2 ′ ′ ′ ∂f ′ ∂f
= π(2π) ǫr ǫv · dJ k |Ak,k′ (J, J , k · Ω)| δ(k · Ω − k · Ω ) f2 k · − f2 k · . (108)
∂t ∂J ∂J ∂J′
′ k,k

In the multilevel case, we obtain the inhomogeneous SL equation

′
!
∂ρ d d d ∂
XZ
′ d ′ 2 ′ ′ ′ ∂ρ ′ ∂f
= π(2π) ǫr ǫv · dJ k |Ak,k′ (J, J , k · Ω)| δ(k · Ω − k · Ω ) f2 k · − ρ(η − f )k · . (109)
∂t ∂J ∂J ∂J′
′ k,k

If we close the hierarchy of moment equations with the ansatz from Eq. (90), we obtain the inhomogeneous Chavanis
equation
′
" #
∂f ∂ XZ 1 ∂f 1 ∂f
= π(2π)d ǫdr ǫdv · dJ′ k |Adk,k′ (J, J′ , k · Ω)|2 δ(k · Ω − k′ · Ω′ ) ′ k· − k′ · . (110)
∂t ∂J C ′′ (f ) ∂J C ′′ (f ) ∂J′
k,k ′

In the two-level case, the foregoing equations reduce to

′
" #
∂f ∂ X ∂f ∂f
Z
′ ′
= π(2π)d ǫdr ǫdv · dJ′ k |Adk,k′ (J, J′ , k · Ω)|2 δ(k · Ω − k′ · Ω′ ) f (η0 − f )k · − f (η0 − f )k′ · , (111)
∂t ∂J ∂J ∂J′
k,k
′

which can be viewed as a form of inhomogeneous fermionic Lenard-Balescu equation.

If we make a thermal bath approximation, using the identity from Eq. (16) and the relation Ω(J) = ∂ǫ/∂J, we
obtain
′
∂f
= −βf2′ Ω(J′ ). (112)
∂J′
Substituting Eq. (112) into Eq. (109) we get
 
∂ρ d d d ∂
XZ
′ d ′ 2 ′ ′ ′ ∂ρ ′ ′ ′
= π(2π) ǫr ǫv · dJ k |Ak,k′ (J, J , k · Ω)| δ(k · Ω − k · Ω ) f2 k · + βf2 ρ(η − f )k · Ω . (113)
∂t ∂J ∂J
′ k,k

Using the properties of the δ-function (resonance condition), we can replace k′ ·Ω′ by k·Ω in the last term in brackets.
We can then rewrite the foregoing equation as
  
∂ρ ∂ ∂ρ
= Dij + βρ(η − f )Ωj (114)
∂t ∂Ji ∂Jj
with
XZ
Dij = π(2π)d ǫdr ǫdv dJ′ ki kj |Adk,k′ (J, J′ , k · Ω)|2 δ(k · Ω − k′ · Ω′ )f2 (J′ ). (115)
k,k′
 22

The equation for the coarse-grained DF reads

  
∂f ∂f ∂
Dij = + βf2 Ωj . (116)
∂t ∂Ji
∂Jj
R
As in Sec. IV B we can enforce the conservation of energy E = f (J)ǫ(J) dJ by letting the inverse temperature
evolve in time according to
∂f
R
Dij Ωi ∂J j
dJ
β(t) = − R . (117)
Dij f2 Ωi Ωj dJ
In this manner, we obtain another type of CSR equations for inhomogeneous systems written with angle-action
variables. Note that the diffusion coefficient from Eq. (115) does not display a logarithmic divergence at large scales
for self-gravitating systems, contrary to the case where a local approximation is made (see Sec. III F), since spatial
inhomogeneity has been properly accounted for.
The above kinetic equations conserve the energy and all the Casimirs and monotonically increase the Lynden-Bell
entropy (H-theorem).

VIII. NONLINEAR DYNAMICAL STABILITY AND NUMERICAL ALGORITHMS

In this section, we consider the nonlinear dynamical stability of steady states of the Vlasov equation based on
variational principles (see [39, 58, 98] and references therein for additional discussions). We also introduce relaxation
equations that can serve as numerical algorithms to compute stable steady states of the Vlasov equation. Similar
results obtained for the 2D Euler equation are given in [61, 62].

A. Energy principle

The Vlasov equation conserves the energy and an infinite class of Casimirs. It can be shown that a DF which is
an extremum of energy (δE = 0) with respect to symplectic perturbations (i.e. perturbations that conserve all the
Casimirs) is a stationary solution of the Vlasov equation. Furthermore, this DF is dynamically stable if and only if
it is a minimum of energy (δ 2 E > 0) with respect to symplectic perturbations (see, e.g., [58] for a brief presentation
of these results). This energy principle is the most refined stability criterion because it takes into account all the
constraints of the Vlasov equation (an infinity of Casimirs). This stability criterion has been introduced in astrophysics
by Bartholomew [99] and Kandrup [100] for the Vlasov-Poisson equations. It is similar to the Kelvin-Arnol’d energy
principle for 2D inviscid incompressible hydrodynamical flows governed by the Euler-Poisson equations [61]. We are
led therefore to considering the minimization problem

min {E[f ] | symplectic perturbations} (118)

f

or, equivalently,

min {E[f ] | Mn≥1 [f ] = Mn≥1 } . (119)

f

Here, the perturbations must conserve all the Casimirs, which is equivalent to the conservation of all the moments of
the DF.26 If we restrict ourselves to DFs of the form f = f (ǫ) with f ′ (ǫ) < 0, it can be shown [58] that f is a local
minimum of E for isovortical perturbations if and only if
1 (δf )2 1
Z Z
δ 2 E[δf ] ≡ − drdv + δf δΦ drdv > 0,
2 f ′ (ǫ) 2
∀ δf | δE = δMn≥1 = 0. (120)

26 For the Newtonian gravitational interaction, it can be shown that all the DFs of the form f = f (ǫ) with f ′ (ǫ) < 0 are minima of
energy with respect to symplectic perturbations so, according to the stability criterion (118), they are dynamically Vlasov stable (see
Refs. [100–105] for linear stability and Ref. [106] for nonlinear stability). This is, however, no more true in general relativity (see the
discussion in [107]) nor for other potentials of interaction.
 23

For the Coulombian potential of interaction in plasma physics, the second term in Eq. (120) is positive implying that
all the DFs of the form f = f (ǫ) with f ′ (ǫ) < 0 are stable (ǫ = v 2 /2 for homogeneous plasmas). For the Newtonian
potential of interaction in astrophysics, the second term in Eq. (120) is negative. Still, it can be shown that all the
DFs of the form f = f (ǫ) with f ′ (ǫ) < 0 are stable (see footnote 26).
Numerical algorithm: We can easily construct a modified dynamics for the DF that conserves all the Casimirs and
that monotonically dissipates the energy. Let us consider the equation27
∂f
+ {f, ǫ} = α{f, {f, ǫ}}, (121)
∂t
where {f, g} is the Poisson bracket defined by Eq. (A1). When α = 0, we recover the Vlasov equation which conserves
the energy and all the Casimirs (see Appendix B). When α > 0, we show below that Eq. (121) conserves all the
Casimirs while the energy decreases monotonically. Therefore, it relaxes towards a minimum of energy with respect to
symplectic perturbations. By construction, this is a dynamically stable steady state of the Vlasov equation. Therefore,
Eq. (121) can be used as a numerical algorithm to construct stable steady states of the Vlasov equation. This is
interesting because it is generally difficult to construct steady states of the Vlasov equation and be sure that they are
dynamically stable.
Proof: We first show that Eq. (121) conserves all the Casimirs. We have
∂f
Z Z
I˙h = h′ (f ) drdv = α h′ (f ){f, {f, ǫ}} drdv. (122)
∂t
Using the identity from Eq. (A2), we get
Z
I˙h = α {f, ǫ}{h′ (f ), f } drdv. (123)

Then, using the identity from Eq. (A3) and the fact that {f, f } = 0, we obtain
Z
I˙h = α h′′ (f ){f, ǫ}{f, f } drdv = 0. (124)

We now show that the energy decreases monotonically. We have

∂f
Z Z
Ė = ǫ drdv = α ǫ{f, {f, ǫ}} drdv. (125)
∂t
Using the identity from Eq. (A2), we get
Z
Ė = α {f, ǫ}{ǫ, f } drdv. (126)

Then, using the identity from Eq. (A4), we obtain

Z
Ė = −α {f, ǫ}2 drdv ≤ 0. (127)

Therefore, the energy is non increasing. At equilibrium (Ė = 0), we have {f, ǫ} = 0 implying that f is a stationary
solution of the Vlasov equation.

B. Sufficient conditions of dynamical stability

We have seen that a DF is a dynamically stable steady state of the Vlasov equation if and only if it is a minimum
of energy for perturbations that conserve all the Casimirs. Therefore, a sufficient condition of dynamical stability is
that f is a minimum of energy for perturbations that conserve the mass M and one Casimir of the form
Z
S[f ] = − C(f )drdv, (128)

27 This equation was suggested in footnote 6 of [61] based on similar results obtained in 2D hydrodynamics.
 24

where C(f ) is a convex function, i.e. C ′′ > 0 [108, 109]. In that case, it is a fortiori a minimum of energy for
perturbations that conserve all the Casimirs (i.e. for symplectic perturbations). We are therefore led to considering
the two-constraint minimization problem

min {E[f ] | M [f ] = M, S[f ] = S} . (129)

f

It is shown in [58] that this minimization problem is equivalent to the maximization problem28

max {S[f ] | M [f ] = M, E[f ] = E} . (130)

f

The first variations can be treated like in Sec. II D leading to the DF from Eq. (25). It can be shown [58] that f is a
local minimum of E at fixed M and S or a local maximum of S at fixed E and M if and only if

1 (δf )2 1
Z Z
δ 2 E[δf ] ≡ − drdv + δf δΦ drdv > 0,
2 f ′ (ǫ) 2
∀ δf | δE = δM = 0. (131)

Clearly, Eq. (131) implies Eq. (120). Indeed if δ 2 E is positive for all perturbations that conserve mass and energy
at first order, it is a fortiori positive for all perturbations that conserve mass, energy and all the Casimirs at first
order. If we view the functional (128) as a “pseudo (or effective) entropy” [39, 110, 111] the maximization problem
(130) is similar to a condition of microcanonical stability in thermodynamics, i.e., to the maximization of the entropy
at fixed mass and energy.29 Therefore, a maximum of pseudo entropy at fixed mass and energy isRa dynamically
stable steady state of the Vlasov equation. In particular, considering the Boltzmann entropy S = − f ln f drdv of
statistical mechanics leading to the Boltzmann distribution f = e−βǫ−α , we conclude that microcanonical stability
implies (Vlasov) dynamical stability. However, the reciprocal is wrong: a dynamically stable steady state of the
Vlasov equation is not necessarily a maximum of pseudo entropy at fixed mass and energy. For example, we have
indicated in footnote 26 that, in the case of Newtonian self-gravitating systems, all the DFs of the form f = f (ǫ)
with f ′ (ǫ) < 0 are dynamically (Vlasov) stable, even those that do not maximize a pseudo entropy at fixed mass and
energy. The stability criteria (129) and (130) are less refined than the stability criterion (118) because they do not
take into account all the constraints of the Vlasov equation. This is similar to a notion of ensemble inequivalence in
thermodynamics (see below).
An even less refined condition of dynamical stability is that f maximizes J = S − βE at fixed mass or, equivalently,
minimizes F = E − T S at fixed mass, where J or F is the Legendre transform of the pseudo entropy with respect to
the energy. We are therefore led to considering the one-constraint minimization problem

min {F [f ] = E[f ] − T S[f ] | M [f ] = M } . (132)

f

The first variations return the results of Sec. II D so that (130) and (132) have the same critical points. It can be
shown [58] that f is a local minimum of F at fixed M if and only if

1 (δf )2 1
Z Z
2
δ E[δf ] ≡ − drdv + δf δΦ drdv > 0,
2 f ′ (ǫ) 2
∀ δf | δM = 0. (133)

Clearly, Eq. (133) implies Eq. (131). Indeed if δ 2 E is positive for all perturbations that conserve mass, it is a fortiori
positive for all perturbations that conserve mass and energy at first order. If we view the functional F as a “pseudo
(or effective) free energy” [39, 110, 111] the minimization problem (132) is similar to a condition of canonical stability

28 If we view f as the coarse-grained DF f , this maximization problem can be related to the selective decay principle (for −S) of Sec. II E.
29 We stress that we are just making a “thermodynamical analogy” [39, 110, 111]. There is no thermodynamics involved in the dynamical
stability problem of the Vlasov equation.
R q This thermodynamical analogy (or effective thermodynamics) may provide an interpretation
1
of the Tsallis entropy Sq = − q−1 (f − f ) drdv, leading to the Tsallis distribution f = (1/q)1/(q−1) [1 − (q − 1)(βǫ + α)]1/(q−1) , in
terms of a “pseudo entropy” [39, 110, 111] in the sense given above. This “Tsallis pseudo entropy” may be useful for dynamical (not
thermodynamical) stability problems. The maximization of the Tsallis (pseudo) entropy at fixed mass and energy ensures the dynamical
stability of a particular class of stationary solutions of the Vlasov equation known as polytropic DFs (see [36, 39, 76, 107, 110] for a
more detailed discussion).
 25

in thermodynamics, i.e., to the minimization of the free energy at fixed mass. Therefore, a minimum of pseudo free
energy at fixed mass is a dynamically stable steady state of the Vlasov equation. The fact that (132) implies (130)
means that a minimum of free energy at fixed mass is necessarily a maximum of entropy at fixed mass and energy.
However, the reciprocal is wrong: A maximum of entropy at fixed mass and energy is not necessarily a minimum of
free energy at fixed mass. Therefore, canonical stability implies microcanonical stability but not the converse [58].
This corresponds the notion of ensemble inequivalence in thermodynamics for systems with long-range interactions
[5, 58, 112–114]. Transposed to the present (dynamical) context, the minimization of pseudo free energy at fixed
mass (one-constraint problem) provides a sufficient condition of dynamical stability which is less refined than the
maximization of pseudo entropy at fixed mass and energy (two-constraint problem) which is itself less refined than
the minimization of energy under symplectic perturbations (infinite-constraint problem).30 In summary, we have the
correspondances

(132) ⇒ (130) ⇒ (118) ⇔ Vlasov stability (134)

The connection of these results with the so-called nonlinear Antonov first law is discussed in detail in [39, 107, 111].
Numerical algorithms: Let us consider the relaxation equation
1 ∂f ′
 
∂f ∂f ∂f ∂ 1 ∂f
Z
+v· − ∇Φ · = dv′ Kij − , (135)
∂t ∂r ∂v ∂vi C ′′ (f ′ ) ∂vj C ′′ (f ) ∂v ′ j
or the relaxation equation
  
∂f ∂f ∂f ∂ ∂f β(t)
+v· − ∇Φ · = · D + ′′ v , (136)
∂t ∂r ∂v ∂v ∂v C (f )
where D is a strictly positive constant and
∂f
R
∂v · v drdv 3M
β(t) = − R v2
= R v2
(137)
C ′′ (f ) drdv drdv
C ′′ (f )

is a time-dependent inverse temperature.31 These equations conserve the mass M and the energy E and monotonically
increase the pseudo entropy S (H-theorem). They relax towards a maximum entropy state at fixed mass and energy.
This corresponds to a microcanonical description. By construction, this equilibrium state is a stable steady state of
the Vlasov equation. Therefore, Eq. (135) or Eqs. (136) and (137) can be used as numerical algorithms to construct
stable steady states of the Vlasov equation. If we fix β, Eq. (136) can be viewed as a generalized Kramers equation. It
conserves the mass M and monotonically decreases the pseudo free energy F = E − T S [50, 51]. It relaxes towards a
minimum of free energy at fixed mass. This corresponds to a canonical description. By construction, the equilibrium
state of the generalized Kramers equation is a stable steady state of the Vlasov equation. Therefore, Eq. (136) with
fixed β can be used as a numerical algorithm to construct stable steady states of the Vlasov equation. We can also
obtain simpler numerical algorithms by taking the hydrodynamic moments of the generalized Landau and Kramers
equations (135) and (136), and closing the hierarchy of equations with a local thermodynamic equilibrium assumption,
leading to generalized Navier-Stokes, Euler and Smoluchowski equations [50, 51]. Similar numerical algorithms have
been introduced in 2D turbulence [50, 61, 62].

C. Dynamical and thermodynamical stability

It can be shown that a thermodynamical equilibrium state in the sense of Lynden-Bell is nonlinearly dynamically
stable. Indeed, the coarse-grained DF f obtained from the Gibbs state (11) which maximizes the Lynden-Bell entropy

30 We can also introduce a no-constraint problem by considering the maximization of the grand potential G = S − βE − αM . This is the
least refined stability criterion (see Ref. [61]).
31 Here, we use the kinetic equations (135) and (136) as numerical algorithms to compute stable steady states of the Vlasov equation,
not as parametrizations of the coarse-grained dynamics. As a result, we can make the following simplifications: (i) we can write these
equations for spatially inhomogeneous systems and take f ′ = f (r, v′ , t) in Eq. (135) even if the local approximation is not justified for
the true evolution of the system; (ii) we can ignore collective effects in Eq. (135) and define Kij by Eq. (67) instead of Eq. (66); (iii)
we can replace Dij by Dij = Dδij in Eq. (136), where D is a strictly positive constant, in order to make the equation simpler and make
sure that it relaxes towards a maximum entropy state at fixed mass and energy without experiencing a situation of kinetic blocking (see
Sec. V).
 26

(9) at fixed mass, energy and Casimir constraints is a minimum of energy E[f ] with respect to perturbations that
c.g.
conserve the coarse-grained moments Mn≥1 [f ] (see Sec. 7.8 of [61]). Therefore, according to Eq. (119), this is a
nonlinearly dynamically stable steady state of the Vlasov equation. By contrast, the initial condition f0 , even though
it has the same energy as the metaequilibrium state, is generically not a minimum of energy E[f0 ] with respect to
f.g.
perturbations that conserve the fine-grained moments Mn≥1 [f0 ], so it is dynamically unstable and relaxes towards
the metaequilibrium state.
It can be shown that a DF which maximizes the generalized entropy defined by Eqs. (22) and (28) at fixed mass and
energy is (i) nonlinearly dynamically stable (see VIII B) and (ii) thermodynamically stable in the sense of Lynden-
Bell (see Appendix C). We stress, however, that this is just a sufficient condition of dynamical and thermodynamical
stability. In particular, the DF f associated with a thermodynamical equilibrium state in the sense of Lynden-Bell
does not necessarily maximizes the generalized entropy defined by Eqs. (22) and (28) at fixed mass and energy (see
Appendix C).
We would be tempted to believe that Vlasov nonlinear dynamical stability implies Lynden-Bell’s thermodynamical
stability. More precisely, we would be tempted to believe that a DF which is a monotonically decreasing function of ǫ
and which is a minimum of energy with respect to symplectic perturbations is a thermodynamically equilibrium state
in the sense of Lynden-Bell. However, this is not true as shown by the following counter-example. For collisionless
self-gravitating systems, all the DFs of the form f = f (ǫ) with f ′ (ǫ) < 0 are dynamically stable (see footnote 26) even
those that are not thermodynamically stable in the sense of Lynden-Bell. In particular, all the extrema – including
saddle points – of Lynden-Bell’s entropy at fixed mass, energy and Casimir constraints are dynamically stable (since
′
they are of the form f = f (ǫ) with f (ǫ) < 0) even if they are not maxima of Lynden-Bell’s entropy at fixed mass,
energy and Casimir constraints. This is because the dynamical stability criterion involves the coarse-grained moments
c.g. f.g.
Mn≥1 [f ] while the Lynden-Bell thermodynamical criterion involves the fine-grained moments Mn≥1 [ρ]. This can be
easily understood in the two-level case. In that case, the Lynden-Bell statistical equilibrium state is obtained by
f.g.
maximizing the Fermi-Dirac-like entropy (19) at fixed mass and energy (the fine-grained moments Mn≥1 [ρ] are all
proportional to the mass) or, equivalently, by minimizing the energy at fixed mass and Fermi-Dirac-like entropy.
However, we have seen that this optimization problem is just a sufficient condition of dynamical stability. A more
refined condition of dynamical stability is that f is a minimum of energy E[f ] with respect to perturbations that
c.g.
conserve all the coarse-grained moments Mn≥1 [f ] (not just the Fermi-Dirac-like entropy). For self-gravitating systems
this is the case for all DFs of the form f = f (ǫ) with f (ǫ) < 0. Thus, there exist DFs which are dynamically Vlasov
stable while they do not maximize the Fermi-Dirac-like entropy at fixed mass and energy. Such DFs are dynamically
stable but not thermodynamically stable in the sense of Lynden-Bell.

IX. AN EQUATION THAT CONSERVES THE MASS AND THE ENERGY AND THAT
MONOTONICALLY INCREASES ALL THE H-FUNCTIONS

Using the same method as the one developed in Sec. VIII A, we can easily construct a modified dynamics for the
DF that conserves the mass and the energy and that monotonically increases all the H-functions (see also Appendix
H). Let us consider the equation

∂f
+ {f , ǫ} = α{ǫ, {f , ǫ}}. (138)
∂t
When α = 0, we recover the Vlasov equation which conserves the energy and all the Casimirs (see Appendix B).
When α > 0, we show below that Eq. (138) conserves the energy while it increases all the H-functions monotonically.
We note that all the stationary solutions of the Vlasov equation (satisfying {f , ǫ} = 0) are stationary solutions of Eq.
(138). Equation (138) may admit other stationary solutions (satisfying {f , ǫ} = α{ǫ, {f , ǫ}}) but, according to the
result derived below Eq. (143), they are necessarily unstable.
Proof: We first show that Eq. (138) conserves the energy. We have

∂f
Z Z
Ė = ǫ drdv = α ǫ{ǫ, {f, ǫ}} drdv. (139)
∂t

Using the identity from Eq. (A2) and the fact that {ǫ, ǫ} = 0, we get
Z
Ė = α {f , ǫ}{ǫ, ǫ} drdv = 0. (140)
 27

We now show that Eq. (138) monotonically increases all the generalized H-functions. We have

∂f
Z Z
′
Ḣ = C (f ) drdv = α C ′ (f ){ǫ, {f, ǫ}} drdv. (141)
∂t
Using the identity from Eq. (A2), we get
Z
Ḣ = α {f , ǫ}{C ′ (f ), ǫ} drdv. (142)

Then, using the identity from Eq. (A3), we obtain

Z
Ḣ = α C ′′ (f ){f , ǫ}2 drdv ≥ 0. (143)

Therefore, the H-functions are non decreasing. At equilibrium (Ḣ = 0), Eq. (143) implies {f , ǫ} = 0. Therefore,
Eq. (138) relaxes towards a stationary solution of the Vlasov equation. Note that this stationary solution does not
necessarily maximize a particular H-function at fixed mass and energy. It cannot be predicted a priori. One has to
solve the kinetic equation (138) numerically to determine its equilibrium state.

X. CONCLUSION

In this paper, we have discussed the kinetic theory of collisionless relaxation for systems with long-range interactions.
We have recalled the basics of the quasilinear theory of the Vlasov equation developed by Kadomtsev and Pogutse
[40], Severne and Luwel [41], and Chavanis [42–44]. We have established a connection between the kinetic equations
derived from the quasilinear theory and the CSR relaxation equations obtained from a phenomenological MEPP [26].
We have proposed a method to close the hierarchy of moment equations leading to a self-consistent kinetic equation
for the coarse-grained DF which is valid beyond the two-level case [44]. This equation [see Eq. (91)] depends on
a generalized entropy C(f ) which can be obtained from the equilibrium state and then used out-of-equilibrium, or
which can be obtained at any time of the dynamics by using the procedure explained in Appendix C 4. We have also
discussed the nonlinear dynamical stability of steady states of the Vlasov equation and proposed numerical algorithms
in the form of kinetic (relaxation) equations that can be used to construct nonlinearly stable steady states. Similar
results can be obtained in 2D turbulence and vortex dynamics by exploiting the analogy between the Vlasov and the
2D Euler equations. This will be discussed in a specific paper [46].
The statistical mechanics of violent relaxation was initiated by Lynden-Bell [22] in the context of collisionless
stellar systems. However, the present paper was motivated by the possibility to apply these ideas to the context of
fermionic or bosonic dark matter [115–119]. Indeed, these systems also exhibit a process of violent relaxation (known
as gravitational cooling [120] in the case of boson stars). For these systems, we have to take into account the quantum
nature of the particles. In the case of fermionic dark matter, the quantum potential arising from the Heisenberg
uncertainty principle is negligible and we can use the classical Vlasov equation (Thomas-Fermi approximation). The
Lynden-Bell theory of violent relaxation can justify the establishment of a Fermi-Dirac-like DF on a timescale shorter
than the age of the universe [115, 116, 119].32 This leads to dark matter halos with a “core-halo” structure. The
quantum core (fermion ball) solves the core-cusp problem of classical cold dark matter and the isothermal halo leads
to flat rotation curves in agreement with the observations [119]. In the case of bosonic dark matter the quantum
potential is important and we must replace the Vlasov equation by the Wigner equation. A generalization of the
Lynden-Bell theory of violent relaxation taking into account the specificities of the Wigner equation has been recently
proposed in [118]. This theory also leads to dark matter halos with a “core-halo” structure where the quantum core
is a self-gravitating Bose-Einstein condensate (soliton) surrounded by a halo made of quantum interferences. The
collisional kinetic theory of fermions and bosons has been studied in [94, 121, 122]. Fermions and bosons behave
antisymmetrically regarding their collisional relaxation. The Pauli blocking f (η0 − f ) for fermions has the tendency
to slow down the relaxation and the Bose enhancement f (η0 + f ) for bosons, leading to the formation of “granules” or
“quasiparticles”, has the tendency to accelerate the relaxation. Gravitational encounters (“collisions”) are completely
negligible in fermionic dark matter halos. In bosonic dark matter halos, they manifest themselves on a (secular)
timescale of the order of the age of the universe (see [94] and references therein).

32 For self-gravitating fermions, gravitational encounters are completely negligible and cannot establish a statistical equilibrium state on a
relevant timescale. However, a collisional relaxation may be relevant if the fermions are self-interacting [119].
 28

Appendix A: Basic properties of the Poisson bracket

The Poisson brackets are defined by

{f, g} = ∇r f · ∇v g − ∇v f · ∇r g. (A1)

We recall below some basic properties of the Poisson brackets that can be established straightforwardly:
Z Z
f {g, h} drdv = h{f, g} drdv, (A2)

Z Z
{h(f ), g} drdv = h′ (f ){f, g} drdv, (A3)

{f, g} = −{g, f }. (A4)

Appendix B: Basic properties of the Vlasov equation

In this appendix, we establish some basic properties of the Vlasov equation (1).
(i) The conservation of the Casimirs can be established as follows:
  Z  
∂f ∂f ∂f ∂h(f ) ∂h(f )
Z Z
I˙h = h′ (f ) drdv = − h′ (f ) v · − ∇Φ · drdv = − v· − ∇Φ · drdv
∂t ∂r ∂v ∂r ∂v
Z  
∂ ∂
= − · [h(f )v] − · [h(f )∇Φ] drdv = 0. (B1)
∂r ∂v

(ii) The conservation of the energy can be established as follows:

Z   Z  
∂f ∂f ∂f ∂ ∂
Z
Ė = ǫ drdv = − ǫ v · − ∇Φ · drdv = − ǫ · (f v) − · (f ∇Φ) drdv
∂t ∂r ∂v ∂r ∂v
Z Z
= f ∇Φ · v drdv − f v · ∇Φ drdv = 0. (B2)

(iii) The conservation of the impulse can be established as follows:

Z   Z  
∂f ∂f ∂f ∂ ∂
Z
Ṗ = v drdv = − v v · − ∇Φ · drdv = − v · (f v) − · (f ∇Φ) drdv
∂t ∂r ∂v ∂r ∂v
∂
Z Z Z Z
= vi · (f ∂j Φ) drdv = − δij f ∂j Φdrdv = − f ∇Φ drdv = − ρ∇Φ dr = 0. (B3)
∂vj

The last equality results from the fact that the sum of the forces acting on the system vanishes. Indeed, using Eq.
(2), we get
Z Z Z
− ρ∇Φ dr = − drdr ρ(r)ρ(r )∇u(|r − r |) = drdr′ ρ(r)ρ(r′ )∇u(|r − r′ |) = 0.
′ ′ ′
(B4)

To get the second equality, we have interchanged the dummy variables r and r′ , and to get the last equality we have
added the half sum of the two preceding expressions.
(iv) The conservation of the angular momentum can be established as follows:
   
∂f ∂f ∂f ∂ ∂
Z Z Z
L̇ = (r × v) drdv = − (r × v) v · − ∇Φ · drdv = − (r × v) · (f v) − · (f ∇Φ) drdv
∂t ∂r ∂v ∂r ∂v
 
∂ ∂
Z Z Z
= − ǫijk xj vk · (f vl ) − · (f ∂l Φ) drdv = ǫijk δjl vk f vl drdv − ǫijk xj δkl f ∂l Φ drdv
∂xl ∂vl
Z Z Z Z Z
= ǫijk vk f vj drdv − ǫijk xj f ∂k Φ drdv = f v × v drdv − f r × ∇Φ drdv = − ρr × ∇Φ dr = 0. (B5)
 29

The last equality results from the fact that the sum of torques acting on the system vanishes. Indeed, using Eq. (2),
we get
Z Z Z
− ρr × ∇Φ dr = − drdr′ ρ(r)ρ(r′ )r × ∇u(|r − r′ |) = drdr′ ρ(r)ρ(r′ )r′ × ∇u(|r − r′ |)

1 1 r − r′
Z Z
= − drdr′ ρ(r)ρ(r′ )(r − r′ ) × ∇u(|r − r′ |) = − drdr′ ρ(r)ρ(r′ )u′ (|r − r′ |)(r − r′ ) × = 0. (B6)
2 2 |r − r′ |

To get the second equality, we have interchanged the dummy variables r and r′ , and to get the third equality we have
added the half sum of the two preceding expressions.
We can establish these results in a slightly different manner, by using the properties of the Poisson brackets (see
Appendix A). The Vlasov equation can be written as

∂f
+ {f, ǫ} = 0. (B7)
∂t
The steady states of the Vlasov equation satisfy {f, ǫ} = 0. The conservation of the Casimirs can be be established
as follows:
∂f
Z Z Z Z
˙ ′
Ih = h (f ) drdv = − h (f ){f, ǫ} drdv = − ǫ{h (f ), f } drdv = − ǫh′′ (f ){f, f } drdv = 0.
′ ′
(B8)
∂t
The conservation of the energy can be established as follows:
∂f
Z Z Z
Ė = ǫ drdv = − ǫ{f, ǫ} drdv = − f {ǫ, ǫ} drdv = 0. (B9)
∂t

Appendix C: Canonical treatment of the Casimir constraints

In the statistical theory of Lynden-Bell [22], the Casimir constraints are treated microcanonically. This is the
correct approach of the problem for an isolated system since these quantities are conserved by the Vlasov equation.
However, it makes the problem quite complicated to solve because we have to relate a large number of Lagrange
multipliers αn (chemical potentials) to the moments Mnf.g. of the fine-grained DF. For that reason, we may consider
a simpler problem where the Casimir constraints are treated canonically (note that the energy and the mass are still
treated microcanonically). In that case, we assume that the Lagrange multipliers αn for n > 1 are prescribed instead
of the moments Mnf.g. .
There are several justifications for treating the Casimir constraints canonically:
1. If the system is not isolated, we may assume that forcing and dissipation will destroy the conservation of the fine-
grained moments Mnf.g. and fix the Lagrange multipliers αn (chemical potentials) instead. While this is an interesting
and convenient suggestion, it does not rest on a firm solid basis.
2. Treating the Casimir constraints canonically provides a simpler maximization problem which determines a
sufficient condition of thermodynamical stability in the sense of Lynden-Bell (see Appendices C 1-C 3).
3. In the kinetic theory of collisionless relaxation, a canonical description of the Casimir constraints is justified to
close the hierarchy of moments equations if the Lagrange multilpliers αn do not differ too much from their equilibrium
value (see Appendix C 4).
Remark: Similar results have been obtained for the Euler equation in 2D hydrodynamics [56, 59–62, 123] and their
adaptation to the Vlasov equation has been discussed in [39].

1. Sufficient condition of Lynden-Bell’s thermodynamical stability

In the Lynden-Bell theory, the statistical equilibrium state is obtained by maximizing the mixing entropy SLB [ρ]
f.g.
at fixed mass M , energy E, Casimirs Mn>1 , and normalization condition (see Sec. II B). This is a necessary and
sufficient condition of thermodynamical stability in the sense of Lynden-Bell. It determines the most probable state
of the system. We thus have to solve the maximization problem
 Z 
f.g. f.g.
max SLB [ρ] | M [f ] = M, E[f ] = E, Mn>1 [ρ] = Mn>1 , ρ dη = 1 . (C1)
ρ
 30

f.g.
The variational problem determining the extrema of SLB at fixed M , E, Mn>1 and normalization condition is given
f.g.
by Eq. (10), leading to the Gibbs state (11). This equilibrium state is a local maximum of SLB at fixed M , E, Mn>1
and normalization condition if and only if

1 (δρ)2 β
Z Z
δ 2 J[δρ] ≡ − drdvdη − δf δΦ drdv < 0,
2 ρ 2
Z
f.g.
∀ δρ | δE = δM = δMn>1 = δρ dη = 0. (C2)

Let us now consider the maximization of the relative entropy

X
Sχ = SLB − αn Mnf.g. (C3)
n>1

at fixed mass M , energy E and normalization condition. Sχ is the Legendre transform of SLB with respect to the
fine-grained moments. As compared to the original maximization problem, this amounts to treating the Casimir
constraints canonically instead of microcanonically. We thus have to solve the maximization problem
 Z 
max Sχ [ρ] | M [f ] = M, E[f ] = E, ρ dη = 1 . (C4)
ρ

The variational problem determining the extrema of Sχ at fixed M , E and normalization is again given by Eq. (10),
leading to the same Gibbs state (11) as in the original problem. This equilibrium state is a local maximum of Sχ at
fixed M , E and normalization if and only if

1 (δρ)2 β
Z Z
δ 2 J[δρ] ≡ − drdvdη − δf δΦ drdv < 0,
2 ρ 2
Z
∀ δρ | δE = δM = δρ dη = 0. (C5)

The critical points (first variations) of (C1) and (C4) are the same but the condition of stability (second variations)
is different. A maximum of Sχ at fixed M , E and normalization condition is always a maximum of S at fixed M , E,
f.g.
Mn>1 and normalization condition, but the converse is wrong. Indeed if inequality (C5) is satisfied for all variations
that satisfy the conservation at first order of mass, energy and normalization condition, it is a fortiori satisfied for all
variations that satisfy the conservation at first order of mass, energy, normalization condition and Casimirs. Therefore
(C5) implies (C2) but this is not reciprocal. As a result, (C4) provides just a sufficient condition of thermodynamical
stability (in the sense of Lynden-Bell). Making the relative entropy explicit, we get
Z X Z
Sχ = − ρ(r, v, η) ln ρ(r, v, η) drdvdη − αn ρ(r, v, η)η n drdvdη
n>1
Z " #
X
= − ρ(r, v, η) ln ρ(r, v, η) + αn η n drdvdη
n>1
 
ρ(r, v, η)
Z
= − ρ(r, v, η) ln drdvdη, (C6)
χ(η)

where we have used Eq. (12) to get the last equality.

Let us finally consider the maximization of the generalized entropy S[f ] at fixed mass M and energy E (see Sec.
II D). We have to solve the maximization problem

max {S[f ] | M [f ] = M, E[f ] = E}. (C7)

f

The variational problem determining the extrema of S at fixed M and E is given by Eq. (23), leading to the equilibrium
state (25) corresponding to the Lynden-Bell coarse-grained DF. This equilibrium state is a local maximum of S at
 31

fixed M and E if and only if33

1 β
Z Z
2 ′′ 2
δ J[δf ] ≡ − C (f )(δf ) drdv − δf δΦ drdv < 0,
2 2
∀ δf | δE = δM = 0. (C8)

Below we show that the maximization of the relative entropy Sχ [ρ] at fixed mass M , energy E and normalization
condition is equivalent to the maximization of the generalized entropy S[f ] defined by (28) at fixed mass M and
energy E. As a result, (C7) provides a sufficient condition of thermodynamical stability (in the sense of Lynden-Bell).
In summary

(C7) ⇔ (C4) ⇒ (C1) (C9)

Remark: We may miss important solutions by maximizing the relative entropy Sχ at fixed mass and energy instead
of maximizing the Lynden-Bell entropy SLB at fixed mass, energy and Casimirs. This is similar to the notion of
ensemble inequivalence for systems with long-range interactions [5, 58, 112–114]. For example, for systems with
long-range interactions, equilibrium states with negative specific heats are forbidden in the canonical ensemble (fixed
T ) while they are allowed in the microcanonical ensemble (fixed E). Similarly, we may miss important solutions by
treating the Casimirs canonically instead of microcanonically.

2. Equivalence for global maximization

We first show the equivalence of (C4) and R (C7) for global maximization. To maximize Sχ [ρ] at fixed mass M [f ],
energy E[f ] and normalization condition ρ(r, v, η) dη = 1 we can proceed in two steps: R
(i) In a first step, we maximize Sχ [ρ] at fixed mass M [f ], energy E[f ] and normalization condition ρ(r, v, η) dη = 1
for a given DF f (r, v). Since the specification
R of f (r, v) determines M [f ] and E[f R], this is equivalent to maximizing
Sχ [ρ] at fixed normalization condition ρ(r, v, η) dη = 1 and with the constraint ρ(r, v, η)η dη = f (r, v). Writing
the variational problem as
Z Z  Z Z 
δSχ − ζ(r, v)δ ρ(r, v, η)dη drdv − Ψ(r, v)δ ρ(r, v, η)ηdη drdv = 0, (C10)

where ζ(r, v) and Ψ(r, v) are Lagrange multipliers, we get

1
ρ∗ (r, v, η) = χ(η)e−ηΨ(r,v) . (C11)
Z[Ψ(r, v)]

This is the global maximum of entropy with the previous constraints since δ 2 Sχ = − 21 [(δρ)2 /ρ∗ ] drdvdη < 0 (the
R
constraints are linear in ρ so their second variations vanish). The functions Z(Ψ) and Ψ are determined by

1
Z Z
−ηΨ(r,v)
Z[Ψ(r, v)] = χ(η)e dη, f (r, v) = χ(η)ηe−ηΨ(r,v) dη (C12)
Z[Ψ(r, v)]

expressing the normalization condition and the specification of the DF f (r, v). These results are similar to those of
Sec. II B provided that we replace βǫ + α by Ψ. Then, we have
′
f = F (Ψ) = −(ln Z)′ (Ψ), f (Ψ) = −f2 (Ψ), (C13)

where F and f2 are defined in Sec. II B.

We can then determine S[f ] ≡ Sχ [ρ∗ ]. Substituting Eq. (C11) into Eq. (C6) we get
Z Z
S[f ] = f Ψ drdv + ln Z drdv. (C14)

33 Using the identity from Eq. (27), we can check that Eq. (C8) is equivalent to Eq. (131).
 32

This is of the form of Eq. (22) with

C(f ) = −f Ψ − ln Z. (C15)

Using Eq. (C13) we find that

∂Ψ ∂ ln Z ∂Ψ ∂Ψ ∂Ψ
C ′ (f ) = −Ψ − f − = −Ψ − f +f = −Ψ = −[(ln Z)′ ]−1 (−f ). (C16)
∂f ∂Ψ ∂f ∂f ∂f
Therefore,
Z f
C(f ) = − [(ln Z)′ ]−1 (−x) dx. (C17)

This returns the result from Eq. (28) establishing the fact that S[f ] is the generalized entropy from Sec. II D.
Therefore, the generalized entropy S[f ] is equal to the relative entropy Sχ [ρ] calculated at ρ∗ when the Casimir
constraints are treated canonically (this is also true for the Lynden-Bell entropy SLB [ρ] calculated at ρ∗ when the
Casimir constraints are treated microcanonically).
(ii) In a second step, we maximize S[f ] ≡ Sχ [ρ∗ ] at fixed mass M [f ] and energy E[f ]. Proceeding as in Sec. II D,
the cancellation of the first variations yields

C ′ (f ) = −βǫ(r, v) − α. (C18)

Comparing Eqs. (C16) and (C18) we find (at equilibrium) that

Ψ(r, v) = βǫ(r, v) + α. (C19)

Substituting this relation into Eq. (C11) we recover the Gibbs state (11). However, we have proven more than that.
The present approach shows that ρ(r, v, η) is the global maximum of Sχ [ρ] at fixed M , E and normalization condition
if and only if f (r, v) is the global maximum of S[f ] at fixed M and E (this is where we need to treat the Casimir
constraints canonically in order to have a fixed shape of the generalized entropy).
Remark: Equation (C16) implies
1
C ′′ (f ) = − ′ . (C20)
f (Ψ)

Comparing this relation with Eq. (C13) we obtain the important relation
1
f2 = . (C21)
C ′′ (f )

We stress that this relation is valid even before maximizing S[f ] ≡ Sχ [ρ∗ ] at fixed mass M [f ] and energy E[f ]. In this
sense, it is expected to remain valid (or approximately valid) when the coarse-grained DF f (r, v) is out-of-equilibrium
(see Appendix C 4).

3. Equivalence for local maximization

We now show the equivalence of (C4) and (C7) for local maximization, i.e. ρ(r, v, η) is a (local) maximum of Sχ [ρ]
at fixed E, M and normalization condition if and only if the corresponding coarse-grained DF f (r, v) is a (local)
maximum of S[f ] at fixed E and M . To that purpose, we show the equivalence between the stability criteria (C5)
and (C8).
R Let us determine
R the perturbation δρ∗ (r, v, η) that maximizes δ 2 J[δρ] given by (C5) with the constraints δf =
δρη dη and δρ dη = 0, where δf (r, v) is prescribed (assumed to conserve energy and mass at first order). Since
the specification of δf determines δΦ, hence the second integral in Eq. (C5), we can write the variational problem
under the form
(δρ)2
  Z Z  Z 
1
Z Z
δ − drdvdη − λ(r, v)δ δρη dη drdv − ζ(r, v)δ δρ dη drdv = 0, (C22)
2 ρ
 33

where λ(r, v) and ζ(r, v) are Lagrange multipliers. This gives

δρ∗ (r, v, η) = −ρ(r, v, η)[λ(r, v)η + ζ(r, v)], (C23)

which is the global maximum of δ 2 J[δρ] with the previous constraints since δ 2 (δ 2 J) = − {[δ(δρ)]2 /2ρ} drdvdη < 0
R
(the constraints are linear
R in δρ soRtheir second variations vanish). The Lagrange multipliers are determined from
the constraints δf = δρη dη and δρ dη = 0 yielding δf = −λf 2 − ζf and 0 = −λf − ζ. Therefore, the optimal
perturbation (C23) can finally be written

δf
δρ∗ = ρ(η − f ). (C24)
f2

Since it maximizes δ 2 J[δρ], we have δ 2 J[δρ] ≤ δ 2 J[δρ∗ ]. Explicating δ 2 J[δρ∗ ] using Eqs. (C5) and (C24), we obtain

1 (δf )2 1
Z Z
δ 2 J[δρ] ≤ − drdv − β δf δΦ drdv. (C25)
2 f2 2
Finally, using Eq. (90), which is rigorously valid at equilibrium, the foregoing inequality can be rewritten as
1 1
Z Z
2
δ J[δρ] ≤ − ′′
C (f )(δf ) drdv − β δf δΦ drdv ≡ δ 2 J[δf ],
2
(C26)
2 2
where the r.h.s. is precisely the functional appearing in Eq. (C8). Furthermore, there is equality in Eq. (C26) if
and only if δρ = δρ∗ . This proves that the stability criteria (C5) and (C8) are equivalent. Indeed: (i) if inequality
(C8) is fulfilled for all perturbations δf that conserve mass and energy at first order, then according to Eq. (C26), we
know that inequality (C5) is fulfilled for all perturbations δρ that conserve mass, energy, and normalization condition
at first order; (ii) if there exists a perturbation δf c that makes δ 2 J[δf c ] > 0, then the perturbation δρc given by
Eq. (C24) with δf = δf c makes δ 2 J[δρc ] = δ 2 J[δf c ] > 0 (this is where we need to treat the Casimir constraints
canonically otherwise this perturbation might not be allowed by the Casimir constraints). In conclusion, the stability
criteria (C5) and (C8) are equivalent.
Remark: We can also derive this result by using the method of orthogonal perturbations [124] developed in the
Appendix of [39].

4. Out-of-equilibrium distribution and justification of the closure relation from Eq. (90)

We can use the strategy developed above to propose a closure of the hierarchy of moment equations (75) describing
the collisionless relaxation of systems with long-range
R interactions.34 The idea is to maximize, out-of-equilibrium,
R the
relative entropy Sχ [ρ] at fixed normalization ρ(r, v, η, t) dη = 1 and coarse-grained DF f (r, v, t) = ρ(r, v, η, t)η dη.
This amounts to constructing a thermodynamical equilibrium distribution ρ∗ (r, v, η, t) corresponding to an out-of-
equilibrium coarse-grained DF f (r, v, t), just like in the first step of Appendix C 2. This returns, at each time t, the
equations of the first step of Appendix C 2. In particular, one has
1
ρ∗ (r, v, η, t) = χ(η)e−ηΨ(r,v,t) , (C27)
Z[Ψ(r, v, t)]

where Z[Ψ(r, v, t)] and Ψ(r, v, t) are determined in terms of χ(η) and f (r, v, t) by Eq. (C12). As a result, Eq. (C21)
is valid at any time (under the previous assumption) yielding
1
f2 (r, v, t) = . (C28)
C ′′ [f (r, v, t)]
As discussed in Sec. IV D, this important relation allows us to close the hierarchy of kinetic equations. This leads
to Eqs. (91) and (92)-(94). We have already indicated in Sec. IV D that these equations conserve mass and energy
and satisfy an H-theorem for the generalized entropy S[f ]. Since Sχ [ρ∗ ] = S[f ], we conclude that the entropy Sχ (t)
increases monotonically with time until the Gibbs state is reached.

34 This method was first introduced in Appendix C of [59] and in [62] in the context of 2D turbulence.
 34

Remark: If we treat the Casimir constraints microcanonically, we find the same results as above except that, at
f.g.
each time t, we have to relate χ(η) to the Casimirs Mn>1 and to the coarse-grained DF f (r, v, t). As a result, χt (η)
and Ct (f ) become functions of time. Therefore, the shape of the generalized entropy changes with time. The kinetic
equation (91) remains valid except that we have to replace C(f ) by Ct (f ). A manner to justify treating the Casimir
constraints canonically is to assume that the function χt (η) is always close to its equilibrium value so that it does
not change substantially. Actually, maximizing out-of-equilibrium the relative entropy Sχ [ρ] at fixed normalization
and coarse-grained DF to get Eq. (C27) is only valid close to equilibrium so the two assumptions are conditioned to
each other. In the canonical closure approach, we just have to solve the equilibrium problem to get χ(η) and C(f )
once for all. Then, Eq. (91) determines the dynamical evolution of the system for all times t provided that we are
sufficiently close to equilibrium for the above assumptions to be valid. Alternatively, in the microcanonical closure
approach, we have to determine χt (η) and Ct (f ) at each time in order to obtain Eq. (91). Since SLB [ρ∗ ] = St [f ],
we conclude that the Lynden-Bell entropy SLB (t) increases monotonically with time until the Gibbs state is reached.
This microcanonical closure approach is more precise, but it is also much more complicated.

5. The equation for the distribution of phase levels

In the approach developed in the previous section, the coarse-grained DF f (v, t) evolves according to Eq. (91)
or Eqs. (92)-(94). The distribution ρ∗ (v, η, t) is then given by Eq. (C27). It may be of interest to determine the
relaxation equation satisfied by ρ∗ (v, η, t) explicitly. According to Eq. (C27), we have

ln ρ∗ = −ηΨ + ln χ(η) − ln Z(Ψ), (C29)

where Ψ(v, t) is related to f (v, t) according to Eq. (C12). Differentiating Eq. (C29) with respect to t and using Eqs.
(C13) and (C20), we obtain

∂ρ∗ ∂Ψ ∂f
= −ρ∗ (η − f ) = ρ∗ (η − f )C ′′ (f ) . (C30)
∂t ∂t ∂t
Similarly, we have

∂ρ∗ ∂Ψ ∂f
= −ρ∗ (η − f ) = ρ∗ (η − f )C ′′ (f ) . (C31)
∂v ∂v ∂v
Combining Eq. (C30) with Eqs. (91) and (92), we get
′
" #
∂ρ∗ ∂ 1 ∂f 1 ∂f
Z
= ρ∗ (η − f )C ′′ (f ) ′
dv Kij ′ − ′′ (C32)
∂t ∂vi ′′
C (f ) ∂vj C (f ) ∂v ′ j

and
  
∂ρ∗ ∂ ∂f β(t)
= ρ∗ (η − f )C ′′ (f ) Dij + ′′ vj . (C33)
∂t ∂vi ∂vj C (f )

Using Eqs. (C21) and (C31), the foregoing equations can be rewritten as
′
" #
∂ρ∗ ρ∗ (η − f ) ∂ f2 ′ ∂ρ∗ ∂f
Z
′
= dv Kij f2 − ρ∗ (η − f ) ′ (C34)
∂t f2 ∂vi ρ∗ (η − f ) ∂vj ∂v j

and
ρ∗ (η − f ) ∂
  
∂ρ∗ f2 ∂ρ∗
= Dij + β(t)ρ∗ (η − f )vj . (C35)
∂t f2 ∂vi ρ∗ (η − f ) ∂vj

Under that form, we see some analogies (but also crucial differences) with the SL and CSR equations (74) and (79).
Remark: Similar equations have been obtained in the context of 2D turbulence [62]. By proceeding similarly to
Sec. 4.2 of [62], it is also possible to derive a relaxation for ρ(r, v, t) associated with the maximization problem (C4)
where the Casimir constraints are treated canonically. This equation can be used as a numerical algorithm to solve
the maximization problem (C4).
 35

6. Log-entropy

f.g.
In the previous sections, we have treated the fine-grained moments Mn>1 canonically. If we do not take into account
f.g.
at all the contribution of the fine-grained moments Mn>1 in the variational principle, the Gibbs state reduces to

1
ρ∗ (r, v, η) = e−ηΨ(r,v) . (C36)
Z[Ψ(r, v)]

This amounts to writing χ(η) = 1 in Eq. (C11). Using Eqs. (C12) and (C13) it is easy to establish that
1 1 1 2
Z= , f= , f2 = 2
=f . (C37)
Ψ Ψ Ψ
We can then rewrite Eq. (C36) as
1
ρ∗ (r, v, η) = e−η/f(r,v) . (C38)
f (r, v)

The generalized entropy associated with this distribution can be obtained from the relation [see Eq. (C28)]
1 2
= f2 = f , (C39)
C ′′ (f )

leading to the functional

Z
S= ln f drdv. (C40)

This is what we have called the log-entropy in Ref. [55]. The kinetic equation (91) associated with the log-entropy
has been studied in [94, 125]. Using Eq. (C19), the equilibrium DF is given by
1
f= . (C41)
βǫ + α
This is the Lorentzian DF. Note that this DF is not normalizable in d = 3, so there is no equilibrium state in that
case.

Appendix D: Cumulant generating function

In the multi-level case, the equilibrium distribution of the statistical theory of Lynden-Bell is the Gibbs state
1
ρ(r, v, η) = χ(η)e−η(βǫ+α) , (D1)
Z(ǫ)

where
Z
Z(ǫ) = χ(η)e−η(βǫ+α) dη (D2)

R
is the partition function. The coarse-grained DF f = ρη dη (first moment) is given by

1
Z
f = ηχ(η)e−η(βǫ+α) dη
Z(ǫ)
1 1 ∂
Z
= − χ(η)e−η(βǫ+α) dη
β Z(ǫ) ∂ǫ
1 1 ∂Z
= −
β Z(ǫ) ∂ǫ
1
= − (ln Z)′ (ǫ). (D3)
β
 36

To the probability density (D1) we associate the cumulant generating function

κ(λ, ǫ) = ln e−λβη . (D4)
It satisfies
1
Z
e−λβη = e−λβη χ(η)e−η(βǫ+α) dη
Z(ǫ)
1
Z
= χ(η)e−η[β(ǫ+λ)+α] dη
Z(ǫ)
Z(ǫ + λ)
= . (D5)
Z(ǫ)
Therefore,
κ(λ, ǫ) = ln Z(ǫ + λ) − ln Z(ǫ). (D6)
Taking the partial derivative of this expression with respect to λ, we get
∂κ(λ, ǫ)
= (ln Z)′ (ǫ + λ). (D7)
∂λ
Combined with Eq. (D3), we obtain
∂κ(λ, ǫ)
= −βf (ǫ + λ). (D8)
∂λ
Expanding both sides of Eq. (D8) in powers of λ, we find that the cumulants κn (ǫ) = κ(n) (0, ǫ) are related to the
derivatives of f (ǫ) by
dn f
κn+1 (ǫ) = −β . (D9)
dǫn
For example,
2 df
β 2 (f 2 − f ) = −β , (D10)
dǫ

3 d2 f
β 3 (f 3 − 3f f 2 + 2f ) = β , (D11)
dǫ2

2 2 4 d3 f
β 4 (f 4 − 3f 2 − 4f 3 f + 12f 2 f − f ) = −β . (D12)
dǫ3

Appendix E: Properties of the KP and SL equations

1. Conservation of energy, linear impulse and Casimirs

It is easy to show that the KP equation (62) conserves the energy and the linear impulse. Indeed,
′ ′
! Z !
v 2 ∂f ′ ∂f ∂f ′ ∂f ∂f
Z Z
′ ′ ′
Ė = dv = − dvdv vi Kij f2 − f2 ′ = dvdv vi Kij f2 − f2 ′
2 ∂t ∂vj ∂v j ∂vj ∂v j
′
!
1 ∂f ∂f
Z
= − dvdv′ wi Kij f2′ − f2 ′ = 0, (E1)
2 ∂vj ∂v j
where we have interchanged the dummy variables v and v′ to obtain the third equality and used the identity Kij wj = 0
to obtain the last equality. Similarly,
′ ′
! Z !
∂f ∂f ∂f ∂f ∂f
Z Z
Ṗi = dv vi = − dvdv′ Kij f2′ − f2 ′ = dvdv′ Kij f2′ − f2 ′ = 0. (E2)
∂t ∂vj ∂v j ∂vj ∂v j
Since the KP equation (62) is the first moment of the SL equation (74), the SL equation
R conserves the energy and
the linear impulse. The SL equation also (trivially) conserves the hypersurface γ(η) = ρ dv of each level. This is
equivalent to the conservation of all the Casimirs.
 37

2. H-theorem

We can also show that the SL equation (74) satisfies an H-theorem for the Lynden-Bell entropy (9).35 The SL
equation (74) can be rewritten in a more symmetric form as
∂ρ′
 
∂ρ ∂ ′ ∂ρ
Z
′ ′ ′ ′ ′
= dv dη Kij η ρ (η − f ) − ρ(η − f ) ′ . (E3)
∂t ∂vi ∂vj ∂v j
The rate of change of the Lynden-Bell entropy (9) is
∂ρ
Z
ṠLB = − dvdη (1 + ln ρ) . (E4)
∂t
Substituting Eq. (E3) into Eq. (E4), we get
" #
η ′ ∂ρ ′ ∂ρ ∂ρ′
Z
′ ′ ′ ′
ṠLB = dvdv dηdη Kij ρ (η − f ) − ρ(η − f ) ′
ρ ∂vi ∂vj ∂vj
" #
′
′ η ∂ρ ′ ∂ρ ∂ρ′
Z
′ ′ ′
= − dvdv dηdη ′ ′ Kij ρ (η − f ) − ρ(η − f ) ′
ρ ∂vi ∂vj ∂vj
" #
∂ρ′ ∂ρ′
 
1 ′ 1 ′ ′ ∂ρ ′ ∂ρ
Z
′ ′ ′
= dvdv dηdη ρη − ρη ′ Kij ρ (η − f ) − ρ(η − f ) ′ . (E5)
2 ρρ′ ∂vi ∂vi ∂vj ∂vj

To obtain the first line we have integrated by parts, to obtain the second line we have interchanged the primed and
unprimed variables, and to obtain the third line we have taken the half-sum of the first and second lines. Equation
(E5) can be rewritten as
" #
∂ρ′ ∂ρ′
 
1 ′ 1 ′ ∂ρ ′ ∂ρ
Z
′ ′ ′ ′ ′
ṠLB = dvdv dηdη ρ (η − f ) − ρ(η − f ) ′ Kij ρ (η − f ) − ρ(η − f ) ′ + I, (E6)
2 ρρ′ ∂vi ∂vi ∂vj ∂vj

where I is the integral

" #
′ ′
 
1 1 ∂ρ ∂ρ ∂ρ ∂ρ
Z
′ ′
I= dvdv′ dηdη ′ ′ ρ′ f − ρf ′ Kij ρ′ (η ′ − f ) − ρ(η − f ) ′ . (E7)
2 ρρ ∂vi ∂vi ∂vj ∂vj

Expanding the terms in brackets, it can written as the sum of four integrals. The first integral
   
1 1 ′ ∂ρ ′ ∂ρ
Z
I1 = dvdv′ dηdη ′ f Kij ρ′ (η ′ − f ) (E8)
2 ρ ∂vi ∂vj
′ ′ ′
vanishes because dη ′ ρ′ (η ′ − f ) = f − f = 0. The second integral
R

" #
∂ρ′
 
1 ′ ∂ρ
Z
′ ′
I2 = dvdv dηdη f Kij (η − f ) ′ = 0 (E9)
2 ∂vi ∂vj

vanishes because dη ′ (∂ρ′ /∂vj′ ) = 0 (recall that dη ′ ρ′ = 1). The two other integrals I3 and I4 vanish for the same
R R
reasons. As a result, we find that I = 0. The rate of change of the Lynden-Bell entropy (E6) can therefore be written
as
1 1
Z
ṠLB = dvdv′ dηdη ′ ′ Xi Kij Xj (E10)
2 ρρ
with
′ ∂ρ ∂ρ′
X = ρ′ (η ′ − f ) − ρ(η − f ) ′ . (E11)
∂v ∂v

35 This H-theorem was not derived in [41].

 38

Since
û(k)2
Z
Xi Kij Xj = π(2π)d ǫdr ǫdv dk (k · X)2 δ[k · (v − v′ )], (E12)
|ǫ(k, k · v)|2

we conclude that ṠLB ≥ 0 with equality if and only if X is parallel to v′ − v. Therefore, the Lynden-Bell entropy
increases monotonically (H-theorem).

3. Gibbs state

Let us check that the Gibbs state (11) is a stationary solution of the SL equation (74). From Eq. (11) we have

ln ρ = −η(βǫ + α) + ln χ(η) − ln Z. (E13)

Taking the derivative of Eq. (E13) and using Eq. (15) we get

∂ρ
= −βρ(η − f )v. (E14)
∂v
On the other hand, according to Eq. (16), we have

∂f ′
= f (ǫ)v = −βf2 v. (E15)
∂v
Therefore, at statistical equilibrium,
′
∂ρ ∂f
f2′ − ρ(η − f ) ′ = −f2′ βρ(η − f )w. (E16)
∂v ∂v
Since Kij wj = 0, we find that the current in Eq. (74) vanishes implying that ∂ρ/∂t = 0.
Inversely, the condition that X must be parallel to v′ − v at equilibrium (this condition results from the H-theorem
as shown above) can be written as

1 ∂ ln ρ 1 ∂ ln ρ′
− ′ ′
= −A(η, η ′ , v, v′ )(v − v′ ). (E17)
η − f ∂v η ′ − f ∂v

From the symmetry of the left hand side of Eq. (E17) it can be shown [6] that A(η, η ′ , v, v′ ) is a constant that we
shall denote β. This then implies that
∂ ln ρ
+ β(η − f )(v − u) = 0, (E18)
∂v
where u is another constant. At that stage, we can repeat the argument of [26] (see also Appendix G) to show that
Eq. (E18) leads to the Gibbs state (11). In conclusion, the SL equation relaxes towards the Lynden-Bell distribution.

Appendix F: Interpretation of the global temperature in the CSR equations

In this Appendix, we provide a physical interpretation of the inverse temperature β(t) in the CSR equation which
was introduced in [26] as a Lagrange multiplier associated with the conservation of energy.

1. Spatially inhomogeneous systems

For spatially inhomogeneous systems, the CSR equations can be written as

  
∂ρ ∂ρ ∂ρ ∂ ∂ρ
+v· − ∇Φ · = Dij + β(t)ρ(η − f )vj (F1)
∂t ∂r ∂v ∂vi ∂vj
 39

with
Z
Dij = dv′ Kij f2′ (F2)

and
R ∂f
Dij vi ∂vj
drdv
β(t) = − R , (F3)
Dij f2 vi vj drdv
where we have made a local approximation f2′ = f2 (r, v′ , t) in Eq. (F2).36 If we consider a simplified model where
Dij = Dδij with D constant, we obtain after an integration by parts
R
d f drdv
β(t) = R . (F4)
f2 v 2 drdv
In the two-level case, and in the nondegenerate limit, the CSR equations reduce to
  
∂f ∂f ∂f ∂ ∂f
+v· − ∇Φ · = Dij + β(t)η0 f vj (F5)
∂t ∂r ∂v ∂vi ∂vj
with
Z
′
Dij = dv′ Kij η0 f (F6)

and
R ∂f
Dij vi ∂vj
drdv
β(t) = − R . (F7)
Dij η0 f vi vj drdv
Equation (F5) is similar to the classical Kramers equation, except that it involves a time-dependent temperature. If
we consider a simplified model where Dij = Dδij with D constant, we obtain after an integration by parts
R
d f drdv dM
β(t) = R = , (F8)
η0 f v 2 drdv 2η0 K(t)

where M = f drdv is the total mass and K(t) = 21 f v 2 drdv is the total kinetic energy. Writing β = 1/T , we get
R R

2η0 K(t) dM
T (t) = ⇔ K(t) = T (t). (F9)
dM 2 η0
This relation shows that T (t) can be interpreted as a global kinetic temperature. It is, however, different from the
spatial average value of the local kinetic temperature. The local kinetic temperature is defined by
η0 f [v − u(r, t)]2 dv
R
Tkin (r, t) = R , (F10)
d f dv

where u(r, t) = ρ1 f v dv is the local velocity. The spatial average of the kinetic temperature is
R

R Z 
ρTkin (r, t) dr η0
Z
2 2
hTkin i(t) = R = f v drdv − ρu dr . (F11)
ρ dr dM
We have the following relation
η0
Z
T (t) = hTkin i(t) + ρu2 dr (F12)
dM
between the global kinetic temperature (F9) and the spatial average value of the local kinetic temperature (F11).

36 See another possible expression of Dij in Appendix B of [26]. More generally, we can leave Dij unspecified provided that the quadratic
form Dij Xi Xj ≥ 0 for any X is definite positive.
 40

2. Spatially homogeneous systems

For spatially homogeneous systems, the energy reduces to the kinetic energy (K = E) implying that the inverse
temperature defined by Eq. (F8) is constant
1 dM
β= = . (F13)
T 2η0 E
In that case, the CSR equation (F5) becomes
 
∂f ∂ ∂f
=D · + βη0 f v , (F14)
∂t ∂v ∂v
which is similar to the usual Kramers (or Klein-Kramers-Chandrasekhar) equation [87, 126, 127]. For the initial
condition f 0 (v) = M δ(v − v0 ) it has the analytical solution
 d/2 βη (v−e−Dβη0 t v )2
βη0 − 0 0
2(1−e−2Dβη0 t )
f (v, t) = M −2Dβη t
e . (F15)
2π(1 − e 0 )

2
We can check that this solution relaxes towards the Boltzmann DF f (v) = M (βη0 /2π)d/2 e−βη0 v /2 . The solution
(F15) was first found by Lord Rayleigh [128] long before the Rseminal paper of Einstein [129] on Brownian motion (see
[130] for more details). Taking the time derivative of E = 12 f v 2 dv and using Eq. (F14), we get37

Ė + 2Dβη0 E = dDm. (F16)

This equation can be integrated into
 
dM dM
E(t) = E0 − e−2Dβη0 t + . (F17)
2βη0 2βη0
This result can also be directly obtained
R from Eq. (F15). The Kramers equation (F14) satisfies an H-theorem for the
free energy F = E − T S where S = − (f /η0 ) ln(f /η0 ) dv is the Boltzmann entropy. Indeed,
 2
DT ∂f
Z
Ḟ = − + βη0 f v ≤ 0. (F18)
η0 f ∂v
Equations (F15)-(F18) are valid for arbitrary values of β. In general, the energy is not conserved since the Kramers
equation is associated with the canonical ensemble (thermal bath). However, when β is exactly given by Eq. (F13) it
turns out that E(t) = E0 is constant. In that case, Eq. (F14) satisfies an H-theorem for the Boltzmann entropy S.

Appendix G: Generalized CSR equations

In the CSR equations [26] the energy, the linear impulse and the angular momentum are conserved globally thanks
to uniform time-dependent Lagrange multipliers (inverse temperature β(t), linear velocity U(t) and angular velocity
Ω(t)). It is possible to introduce more general relaxation equations that conserve the energy, the linear impulse and
the angular momentum locally. The equation for ρ(r, v, η, t) reads38
  
∂ρ ∂ρ ∂ρ ∂ ∂ρ
+v· − ∇Φ · = Dij + β(r, t)ρ(η − f )(v − u(r, t))j . (G2)
∂t ∂r ∂v ∂vi ∂vj

37 If we make the correspondance v ↔ r, the Kramers equation (F14) is equivalent to the Smoluchowski equation [131] for a Brownian
particle in a harmonic potential. In that case, the kinetic energy is equivalent to the moment of inertia and Eq. (F16) can be interpreted
as a form of virial theorem.
38 This equation can be obtained from the SL equation (73) by first extending it to spatially inhomogeneous systems, making a local
′
approximation (see Sec. III F), then by computing the term ∂f /∂v′ with the distribution
(v−u(r,t))2
 
1 −η β(r,t) 2
+α(r,t)
ρ(r, v′ , η, t) = χ(η)e , (G1)
Z(r, v′ , t)
which relies on a local thermodynamic equilibrium approximation. In that case, the diffusion tensor in Eq. (G2) is given by Eq. (80)
with f2′ = f2 (r, v′ , t). The usual CSR equations [26] are recovered for β(r, t) = β(t) and u(r, t) = U(t) − Ω(t) × r.
 41

Multiplying Eq. (G2) by η and integrating over the phase levels, we get
  
∂f ∂f ∂f ∂ ∂f
+v· − ∇Φ · = Dij + β(r, t)f2 (v − u(r, t))j . (G3)
∂t ∂r ∂v ∂vi ∂vj

These relaxation equation can be written as

∂ρ ∂ρ ∂ρ ∂ ∂f ∂f ∂f ∂
+v· − ∇Φ · =− ·J and +v· − ∇Φ · =− · Jf , (G4)
∂t ∂r ∂v ∂v ∂t ∂r ∂v ∂v
where J is the current of the phase levels η and Jf is the current of the coarse-grained DF given by
   
∂ρ ∂f
Ji = −Dij + β(r, t)ρ(η − f )(v − u(r, t))j
and = −Dij Jfi
+ β(r, t)f2 (v − u(r, t))j . (G5)
∂vj ∂vj
R R
We note that J dη = 0 (according to the normalization condition) and Jf = Jη dη. The local conservation of
linear impulse and energy imposes that
Z Z
Jf dv = 0 and Jf · v dv = 0. (G6)

Substituting the current Jf from Eq. (G5) into the constraints from Eq. (G6), we obtain a set of two linear equations

∂f
Z Z
Dij dv + β(r, t) Dij f2 (v − u(r, t))j dv = 0, (G7)
∂vj

∂f
Z Z
Dij vi dv + β(r, t) Dij vi f2 (v − u(r, t))j dv = 0, (G8)
∂vj

which determine β(r, t) and u(r, t).

The H-theorem can be derived as follows. First we note that the Lynden-Bell entropy, and more generally all the
functionals of ρ, are conserved by the advection term of Eq. (G2). The proof is similar to the one given in Appendix
B for the Vlasov equation:
  Z  
∂ρ ∂ρ ∂ρ ∂h(ρ) ∂h(ρ)
Z Z
˙ ′
Ih = h (ρ) ′
drdvdη = − h (ρ) v · − ∇Φ · drdvdη = − v· − ∇Φ · drdvdη
∂t ∂r ∂v ∂r ∂v
Z  
∂ ∂
= − · [h(ρ)v] − · [h(ρ)∇Φ] drdvdη = 0. (G9)
∂r ∂v

Therefore, the change of entropy is only due to the current J. It is given by

∂ρ ∂ ∂ ln ρ J ∂ρ
Z Z Z Z
ṠLB = − (ln ρ + 1) drdvdη = (ln ρ + 1) · J drdvdη = − · J drdvdη = − · drdvdη. (G10)
∂t ∂v ∂v ρ ∂v
The last term of this equation can be rewritten as
 
∂ρ
Z Z
J J
ṠLB = − · + βρ(η − f )(v − u) drdvdη + βρ(η − f ) · (v − u) drdvdη. (G11)
ρ ∂v ρ

Integrating over η and using the normalization condition and the local conservation of impulse and energy from Eq.
(G6) we see that the second term in Eq. (G11) vanishes:
Z Z
β(η − f )J · (v − u) drdvdη = βJf · (v − u) drdv = 0. (G12)

As a result, there remains

 
∂ρ
Z
J
ṠLB = − · + βρ(η − f )(v − u) drdvdη, (G13)
ρ ∂v
 42

which, using Eq. (G5), can be written as

   
1 ∂ρ ∂ρ
Z
ṠLB = + βρ(η − f )(v − u) Dij + βρ(η − f )(v − u) drdvdη. (G14)
ρ ∂v i ∂v j

Assuming that the quadratic form Xi Dij Xj ≥ 0 for any X is positive definite (we can check that this is the case
with the expression of Dij from Eqs. (66), (67) and (80)) we conclude that ṠLB ≥ 0. At equilibrium, the current J
vanishes leading to the Gibbs state (11). This can be proven as follows. The condition J = 0 can be written as
∂ ln ρ
+ β(η − f )(v − u) = 0. (G15)
∂v
Applying this relation to a reference level η0 , we get
∂ ln ρ0
+ β(η0 − f )(v − u) = 0, (G16)
∂v
where ρ0 = ρ(r, v, η0 ). Subtracting Eqs. (G15) and (G16), we obtain
 
∂ ρ
ln + β(η − η0 )(v − u) = 0. (G17)
∂v ρ0
This equation can be integrated into
 
ρ 1
ln + β(η − η0 )(v − u)2 = A(r, η), (G18)
ρ0 2

where A(r, η) is a constant of integration. At equilibrium, the advection term in Eq. (G2) must also vanish yielding
∂ρ ∂ρ
v· − ∇Φ · = 0. (G19)
∂r ∂v
Repeating the same procedure as above, we get
   
∂ ρ ∂ ρ
v· ln − ∇Φ · ln = 0. (G20)
∂r ρ0 ∂v ρ0

One can show from the combination of Eqs. (G15) and (G19) that, at equilibrium, β must be uniform and u
must vanish (this can be viewed as a consequence of the Jeans theorem [53]). Therefore, limt→+∞ β(r, t) = β and
limt→+∞ u(r, t) = 0.39 Then, Eq. (G18) reduces to
 
ρ 1
ln + β(η − η0 )v 2 = A(r, η). (G21)
ρ0 2
Taking its gradient with respect to r, we get
 
∂ ρ
ln = ∇A(r, η). (G22)
∂r ρ0

Substituting Eq. (G17) and Eq. (G22) into Eq. (G20) we get v · [∇A + β(η − η0 )∇Φ] = 0. This equality must be
true for all v, implying that ∇A + β(η − η0 )∇Φ = 0, which can be integrated into A(r, η) = −β(η − η0 )Φ(r) − B(η),
where B(η) is a constant of integration. Finally, Eq. (G21) can be rewritten as
 
ρ
ln = −β(η − η0 )ǫ − B(η), (G23)
ρ0

which is equivalent to the Gibbs state (11) with 1/Z(r, v) = ρ(r, v, η0 )eβη0 ǫ(r,v) and χ(η)e−ηα = e−B(η) . Inversely,
starting from the Gibbs state (11) and using Eqs. (E13) and (E14), we get J = 0.

39 Note that in the CSR approach [26], at each time t, the inverse temperature β(t) is uniform and u = 0. One then have limt→+∞ β(t) = β.
 43

Remark: If we assume that Dij = Dδij with D constant, the linear equations (G7) and (G8) reduce to

f2 (v − u(r, t))2 dv
R R
f2 v dv
u(r, t) = R , T (r, t) = . (G24)
f2 dv
R
d f dv
In the two-level case and in the nondegenerate limit, we obtain
 
∂f ∂f ∂f ∂ ∂f
+v· − ∇Φ · =D · + β(r, t)f η0 (v − u(r, t)) (G25)
∂t ∂r ∂v ∂v ∂v
with
f (v − u(r, t))2 dv
R R
f v dv
u(r, t) = R , T (r, t) = R . (G26)
f dv d f dv

In that case, we recover the usual expressions of the local velocity and local kinetic temperature. Equations (G25)
and (G26) are similar to the kinetic equations introduced by Dougherty [132] for collisional systems.

Appendix H: Another equation that conserves the mass and the energy and that monotonically increases all
the H-functions

In Sec. IX we have introduced an equation that conserves the mass and the energy and that monotonically increases
all the H-functions. In this Appendix, we introduce another equation that satisfies the same properties.

1. Anisotropic diffusion equation

The CSR equation is given by Eq. (79) with Eq. (83). If we assume for simplicity that the diffusion tensor is
isotropic and constant, so that Dij = Dδij , this equation reduces to
 
∂ρ ∂ ∂ρ
=D · + β(t)ρ(η − f )v (H1)
∂t ∂v ∂v
with
∂f
R
v · ∂v dv
β(t) = − R . (H2)
f2 v 2 dv

If we get rid of the integrals in Eq. (H2), we get

∂f
v · ∂v
β(v, t) = − . (H3)
f2 v 2
Substituting this relation into Eq. (H1) we obtain
" #
∂f
∂ρ ∂ ∂ρ v · ∂v
=D · − ρ(η − f )v . (H4)
∂t ∂v ∂v f2 v 2

This equation can also be obtained by applying the MEPP with a local conservation of energy Jf · v = 0 in velocity
space.40 It conserves the normalization condition, the Casimirs (or the total hypervolume of each phase level η) and
the energy. It also increases the mixing entropy (9) monotonically (H-theorem). The proof is essentially the same as
for the CSR equations (see Ref. [26] and Appendix G). However, it does not relax towards the Gibbs state (11).41

40 An equation similar to Eq. (H4), but acting in position space instead of velocity space, has been obtained in the context of 2D turbulence
in Ref. [65].
41 It is not clear if this property is a drawback of this equation or if it can account for situations of incomplete relaxation where the
quasistationary state is different from the Lynden-Bell statistical equilibrium state.
 44

To see that, let us consider the equation for the coarse-grained DF (81) which, for an isotropic and constant diffusion
tensor, can be written as
 
∂f ∂ ∂f
=D · + β(t)f2 v . (H5)
∂t ∂v ∂v

Replacing β(t) by Eq. (H3) we obtain

!
∂f
∂f ∂ ∂f v · ∂v
=D · − v . (H6)
∂t ∂v ∂v v2

This equation can also be obtained by multiplying Eq. (H4) by η and integrating over η. We note that, unlike Eq.
(H5), this equation is closed since the second moment f2 has cancelled out. Equation (H6) can be rewritten as
 
∂f ∂  vi vj  ∂f
=D δij − 2 . (H7)
∂t ∂vi v ∂vj

This is an anisotropic diffusion equation of the form

 
∂f ∂ ∂f
= Dij (H8)
∂t ∂vi ∂vj

with a diffusion tensor

v 2 δij − vi vj
Dij = D . (H9)
v2
The diffusion tensor Dij has the property that Dij vj = 0. As a result, all isotropic DFs are stationary solutions of
′
Eq. (H6). Indeed, for a DF of the form f = f (v) with v = |v|, we have ∂f /∂vj = f (v)vj /v. Since Dij vj = 0, we
obtain Dij ∂f /∂vj = 0, hence ∂f /∂t = 0. When the initial DF f 0 (v) is anisotropic, the system evolves until f (v, t)
becomes isotropic. Therefore, the effect of the diffusion equation (H6) is to “isotropize” an initially anisotropic DF.42
Remark: For 1D systems, like the HMF model, Eq. (H6) reduces to

∂f
=0 (H10)
∂t
so there is no evolution in that case.

2. Properties of Eq. (H6)

Let us write Eq. (H6) under the conservative form

∂f ∂
=− · Jf (H11)
∂t ∂v
with the diffusion current
     
∂f ∂f ∂f v ∂f ∂f v
Jf = −Dij = −D − v· = −D − . (H12)
∂vj ∂v ∂v v 2 ∂v ∂v v

We note that the diffusion current is normal to the velocity:

Jf · v = 0. (H13)

42 For simplicity, we have considered spatially homogeneous systems. However, Eq. (H6) remains valid for spatially inhomogeneous systems
provided that we introduce an advection term in the left hand side. In that case, it relaxes towards an isotropic DF of the form f (ǫ)
where ǫ = v2 /2 + Φ(r) which cancels both the advection term and the “collision” term.
 45

As a result, Eq. (H6) trivially conserves the energy (6). Indeed

Z 2 Z 2
v ∂f v ∂
Z
Ė = dv = − · Jf dv = Jf · v dv = 0. (H14)
2 ∂t 2 ∂v
We can also show that Eq. (H6) monotonically increases all the H-functions (29). We have

∂f ∂ ∂f
Z Z Z
Ḣ = − C ′ (f ) dv = C ′ (f ) · Jf dv = − C ′′ (f )Jf · dv. (H15)
∂t ∂v ∂v
Using Eq. (H12), the last equality of this equation can be rewritten as
   
1 ∂f v
Z
′′
Ḣ = C (f ) Jf · Jf − D v · dv. (H16)
D ∂v v 2
Using Eq. (H13), we get

J2f
Z
Ḣ = C ′′ (f ) dv ≥ 0. (H17)
D
Therefore, all the H-functions increase monotonically. At equilibrium, we have Jf = 0. This determines an isotropic
DF of the form f = f (v).
Remark: We note that Eqs. (138) and (H6) share similar general properties (conservation of energy and monotonic
increase of all the H-functions). However, these two equations are very different. In particular, Eq. (138) reduces to
∂t f = 0 for spatially homogeneous systems contrary to Eq. (H6), and Eq. (H6) reduces to ∂t f = 0 for 1D systems
contrary to Eq. (138).

3. Analytical solution of Eq. (H6)

It turns out that Eq. (H6) can be solved analytically. Taking the divergence of the current from Eq. (H12), we can
rewrite Eq. (H6) as

∂f
D
= D ∆v f − ∆v f = 2 ∆S f , (H18)
∂t v
2
∂ d−1 ∂
where ∆v is the Laplacian operator in velocity space, ∆v = ∂v 2 + v ∂v is the part of the Laplacian operator which
involves derivatives with respect to the modulus of v, and ∆S is the part of the Laplacian operator which involves
derivatives with respect to the orientation of the vector v (on the unit sphere).
In d = 2, introducing a polar system of coordinates, we have
∂2
∆S = . (H19)
∂θ2
The solution of Eq. (H18) is then
+∞
2
t/v 2
X
f (v, θ, t) = cn (v)einθ e−Dn (H20)
n=−∞

with
2π
1
Z
cn (v) = f 0 (v, θ)e−inθ dθ. (H21)
2π 0

For t → +∞, we get

2π
1
Z
f (v, θ, t) → c0 (v) = f 0 (v, θ) dθ = hf 0 (v, θ)iθ . (H22)
2π 0

Therefore, f (v, t) tends to an isotropic DF which is equal to the average over the angle θ of the initial DF f 0 (v) =
f0 (v, θ).
 46

In d = 3, introducing a spherical system of coordinates, we have

1 ∂2
 
1 ∂ ∂
∆S = + sin θ . (H23)
sin2 θ ∂φ2 sin θ ∂θ ∂θ

In that case, the solution of Eq. (H18) is

+∞ X
+l
X 2
f (v, θ, φ, t) = clm (v)Ylm (θ, φ)e−Dl(l+1)t/v (H24)
l=0 m=−l

with
Z 2π Z π
∗
clm (v) = dφ dθ sin θf 0 (v, θ, φ)Ylm (θ, φ), (H25)
0 0

where Ylm (θ, φ) are the spherical harmonics. For t → +∞, we get
2π π
1
Z Z
f (v, θ, φ, t) → c00 (v)Y00 (θ, φ) = dφ dθ sin θf 0 (v, θ, φ) = hf 0 (v, θ, φ)iθ,φ . (H26)
4π 0 0

Therefore, f (v, t) tends to an isotropic DF which is equal to the average over the angles θ and φ of the initial DF
f 0 (v) = f0 (θ, φ, v).

4. Another type of equation

If we integrate Eq. (H2) by parts, we find that

R
d f dv
β(t) = R . (H27)
f2 v 2 dv

If we get rid of the integrals, we get

df
β(v, t) = . (H28)
f2 v 2

Substituting this relation into Eq. (H5) we obtain

 
∂f ∂ ∂f v
=D · + df 2 . (H29)
∂t ∂v ∂v v

Again, this is a closed equation. Equation (H29) conserves the energy (6). Indeed:

v 2 ∂f v2 ∂
     
∂f ∂f ∂f
Z Z Z Z
v v
Ė = dv = D · + df 2 dv = − D + df 2 · v dv = − D · v + df dv = 0,
2 ∂t 2 ∂v ∂v v ∂v v ∂v
(H30)
where the last equality is obtained after performing an integration by parts. By contrast, nothing general can be said
about the sign of Ḣ.
Remark: Equation (H30) can be interpreted as a Smoluchowski equation describing the evolution of a Brownian
particle coupled to a thermal bath of unit temperature Tc = 1 and submitted to an attractive logarithmic potential
U (v) = d ln |v|, where v plays the role of the position r. The stationary solution f e = Ae−U(v) = A/v d is not
normalizable (the normalization factor diverges logarithmically at both small and large velocities). From the general
theory
R of Fokker-PlanckR equations, we know that Eq. (H30) satisfies an H-theorem (Ḟ ≤ 0) for the free energy
F = f U (v) dv +RTc f ln f dv. At the critical temperature Tc = 1, Eq. (H29) has the particularity to conserve the
energy E = (1/2) f v 2 dv which is analogous to the moment of inertia if we make the correspondance r ↔ v (see
above). In this respect, Eq. (H30) can be interpreted as a form of virial theorem. The study of Eq. (H29) at T ≤ Tc
is subtle because it displays a form of “collapse” or a form of Bose-Einstein condensation in the state v = 0 leading
to a Dirac peak δ(v). This is an example of Bessel process that has been studied in, e.g., Ref. [133].
 47

[1] Dynamics and Thermodynamics of Systems with Long-Range Interactions, edited by T. Dauxois, S. Ruffo, E. Arimondo
and M. Wilkens, Lectures Notes in Physics 602 (Berlin: Springer, 2002)
[2] Dynamics and Thermodynamics of Systems with Long-Range Interactions: Theory and Experiments, edited by A. Campa,
A. Giansanti, G. Morigi and F. Sylos Labini, AIP Conf. Proc. 965 122 (2008)
[3] Long-Range Interacting Systems, edited by T. Dauxois, S. Ruffo and L. Cugliandolo, Les Houches Summer School 2008,
(Oxford: Oxford University Press, 2009)
[4] A. Campa, T. Dauxois, S. Ruffo, Physics Reports 480, 57 (2009)
[5] A. Campa, T. Dauxois, D. Fanelli, S. Ruffo, Physics of Long-Range Interacting Systems (Oxford University Press, 2014)
[6] A. Lenard, Ann. Phys. (N.Y.) 10, 390 (1960)
[7] R. Balescu, Phys. Fluids 3, 52 (1960)
[8] J. Heyvaerts, Mon. Not. R. Astron. Soc. 407, 355 (2010)
[9] P.H. Chavanis, Physica A 391, 3680 (2012)
[10] L.D. Landau, Phys. Z. Sowj. Union 10, 154 (1936)
[11] P.H. Chavanis, Astron. Astrophys. 556, A93 (2013)
[12] P.H. Chavanis, Eur. Phys. J. Plus 127, 19 (2012)
[13] S. Chandrasekhar, Principles of Stellar Dynamics (University of Chicago Press, 1942)
[14] J.B. Fouvry, B. Bar-Or, P.H. Chavanis, Phys. Rev. E 100, 052142 (2019)
[15] J.B. Fouvry, P.H. Chavanis, C. Pichon, Phys. Rev. E 102, 052110 (2020)
[16] J.H. Jeans, Mon. Not. R. Astron. Soc. 76, 70 (1915)
[17] A.A. Vlasov, Zh. Eksp. i Teor. Fiz. 8, 291 (1938)
[18] I.R. King, Astron. J. 67, 471 (1962)
[19] M. Hénon, Ann. Astrophys. 27, 83 (1964)
[20] P.H. Chavanis, Statistical mechanics of two-dimensional vortices and stellar systems, in: Dynamics and thermodynamics
of systems with long range interactions, edited by Dauxois, T, Ruffo, S., Arimondo, E. and Wilkens, M. Lecture Notes in
Physics, Springer (2002)
[21] F. Bouchet, A. Venaille, Phys. Rep. 515, 227 (2012)
[22] D. Lynden-Bell, Mon. Not. R. Astron. Soc. 136, 101 (1967)
[23] J. Miller, Phys. Rev. Lett. 65, 2137 (1990)
[24] R. Robert, J. Sommeria, J. Fluid Mech. 229, 291 (1991)
[25] P.H. Chavanis, Mécanique statistique des tourbillons bidimensionnels. Analogie avec la relaxation violente des systèmes
stellaires, PhD thesis, Ecole Normale Supérieure de Lyon (1996)
[26] P.H. Chavanis, J. Sommeria, R. Robert, Astrophys. J. 471, 385 (1996)
[27] P.H. Chavanis, Theor. Comput. Fluid Dyn. 24, 217 (2010)
[28] P.H. Chavanis, J. Stat. Mech P05019 (2010)
[29] P.H. Chavanis, Eur. Phys. J. B 53, 487 (2006)
[30] A. Antoniazzi, D. Fanelli, J. Barré, P.H. Chavanis, T. Dauxois, S. Ruffo, Phys. Rev. E 75, 011112 (2007)
[31] A. Antoniazzi, F. Califano, D. Fanelli, S. Ruffo, Phys. Rev. Lett. 98, 150602 (2007)
[32] A. Antoniazzi, D. Fanelli, S. Ruffo, Y. Yamaguchi, Phys. Rev. Lett. 99, 040601 (2007)
[33] P.H. Chavanis, G. De Ninno, D. Fanelli, S. Ruffo Out of equilibrium phase transitions in mean field Hamiltonian dynamics
in Proceedings of the conference ”Chaos, Complexity and Transport” (Marseille, 5-9 June 2007); Chandre, Leoncini,
Zaslavsky Eds., Chaos, Complexity and Transport: Theory and Applications, World Scientific (2008) p. 3
[34] F. Staniscia, P.H. Chavanis, G. De Ninno, D. Fanelli, Phys. Rev. E 80, 021138 (2009)
[35] F. Staniscia, P.H. Chavanis, G. De Ninno, Phys. Rev. E 83, 051111 (2011)
[36] A. Campa, P.H. Chavanis, Eur. Phys. J. B 86, 170 (2013)
[37] P.H. Chavanis, Astron. Astrophys. 401, 15 (2003)
[38] P.H. Chavanis, Physica A 365, 102 (2006)
[39] P.H. Chavanis, AIP Conf. Proc. 970, 39 (2008)
[40] B.B. Kadomtsev, O.P. Pogutse, Phys. Rev. Lett. 25, 17 (1970)
[41] G. Severne, M. Luwel, Astrophys. Space Sci. 72, 293 (1980)
[42] P.H. Chavanis, Mon. Not. R. Astron. Soc. 300, 981 (1998)
[43] P.H. Chavanis, Statistical mechanics of violent relaxation in stellar systems, in Multiscale Problems in Science and Tech-
nology, edited by N. Antonić, C.J. van Duijn, W. Jäger, and A. Mikelić (Springer, 2002)
[44] P.H. Chavanis, Physica A 332, 89 (2004)
[45] P.H. Chavanis, Phys. Rev. Lett. 84, 5512 (2000)
[46] P.H. Chavanis, in preparation
[47] R. Robert, J. Sommeria, Phys. Rev. Lett. 69, 2776 (1992)
[48] R. Robert, C. Rosier, J. Stat. Phys. 86, 481 (1997)
[49] P.H. Chavanis, J. Sommeria, Phys. Rev. Lett. 78, 3302 (1997)
[50] P.H. Chavanis, Phys. Rev. E 68, 036108 (2003)
[51] P.H. Chavanis, Eur. Phys. J. B 62, 179 (2008)
[52] S. Tremaine, M. Hénon, D. Lynden-Bell, Mon. Not. R. Astron. Soc. 227, 543 (1987)
 48

[53] J.H. Jeans, MNRAS 76, 70 (1915)

[54] J. Binney, S. Tremaine, Galactic Dynamics (Princeton Series in Astrophysics, 1987)
[55] P.H. Chavanis, Physica A 359, 177 (2006)
[56] R. Ellis, K. Haven, B. Turkington, Nonlinearity 15, 239 (2002)
[57] P.H. Chavanis, J. Sommeria, Mon. Not. R. Astron. Soc. 296, 569 (1998)
[58] A. Campa, P.H. Chavanis, J. Stat. Mech. 06, 06001 (2010)
[59] P.H. Chavanis, Physica D 200, 257 (2005)
[60] P.H. Chavanis, Physica D 237, 1998 (2008)
[61] P.H. Chavanis, Eur. Phys. J. B 70, 73 (2009)
[62] P.H. Chavanis, A. Naso, B. Dubrulle, Eur. Phys. J. B 77, 167 (2010)
[63] R.H. Kraichnan, D. Montgomery, Rep. Prog. Phys. 43, 547 (1980)
[64] W.H. Matthaeus, D. Montgomery, N.Y. Acad. Sci. 357, 203 (1980)
[65] P.H. Chavanis, Fluid Dyn. Res. 46, 061409 (2014)
[66] A. Naso, P.H. Chavanis, B. Dubrulle, Eur. Phys. J. B 77, 187 (2010)
[67] F.P. Bretherton, D.B. Haidvogel, J. Fluid Mech. 78, 129 (1976)
[68] C.E. Leith, Phys. Fluids 27, 1388 (1984)
[69] P.H. Chavanis, J. Sommeria, J. Fluid Mech. 314, 267 (1996)
[70] L. Woltjer, Proc. Nat. Acad. Sci. 44, 489 (1958)
[71] J.B. Taylor, Phys. Rev. Lett. 33, 1139 (1974)
[72] D. Montgomery, L. Turner, G. Vahala, Phys. Fluids 21, 757 (1978)
[73] N. Leprovost, B. Dubrulle, P.H. Chavanis, Phys. Rev. E 71, 036311 (2005)
[74] A. Naso, R. Monchaux, P.H. Chavanis, B. Dubrulle, Phys. Rev. E 81, 066318 (2010)
[75] V. Latora, A. Rapisarda, C. Tsallis, Physica A 305 129 (2002)
[76] P.H. Chavanis, A. Campa, Eur. Phys. J. B 76, 581 (2010)
[77] J. Sommeria, C. Staquet, R. Robert, J. Fluid Mech. 233 661 (1991)
[78] X.P. Huang, C.F. Driscoll, Phys. Rev. Lett. 72 2187 (1994)
[79] B.M. Boghosian, Phys. Rev. E 53 4754 (1996)
[80] H. Brands, P.H. Chavanis, R. Pasmanter, J. Sommeria, Phys. Fluids 11 3465 (1999)
[81] C. Tsallis, J. Stat. Phys. 52 479 (1988)
[82] E.M. Lifshitz, L.P. Pitaevskii, Physical Kinetics (Pergamon Press, Oxford, 1981)
[83] B. Bar-Or, J.B. Fouvry, S. Tremaine, arXiv:2010.10212
[84] T.H. Dupree, Phys. Rev. Lett. 25, 789 (1970)
[85] P.H. Chavanis, Phys. Rev. E 64, 026309 (2001)
[86] P.H. Chavanis, Physica A 387, 1123 (2008)
[87] S. Chandrasekhar, Astrophys. J. 97, 255 (1943)
[88] L. Onsager, Phys. Rev. 37, 405 (1931)
[89] L. Onsager, Phys. Rev. 38, 2265 (1931)
[90] L. Onsager, S. Machlup, Phys. Rev. 91, 1505 (1953)
[91] P.H. Chavanis, Entropy 21, 1006 (2019)
[92] P.H. Chavanis, Phys. Rev. E 58, R1199 (1998)
[93] E. Kazantsev, J. Sommeria, J. Verron, J. Phys. Oceanogr. 28, 1017 (1998)
[94] P.H. Chavanis, Eur. Phys. J. Plus 136, 703 (2021)
[95] P.H. Chavanis, J. Sommeria, J. Fluid Mech. 356, 259 (1998)
[96] P.H. Chavanis, Eur. Phys. J. Plus 128, 126 (2013)
[97] P.H. Chavanis, Eur. Phys. J. Plus 128, 128 (2013)
[98] P.H. Chavanis, L. Delfini, Eur. Phys. J. B 69, 389 (2009)
[99] P. Bartholomew, Mon. Not. Roy. Astr. Soc. 151, 333 (1971)
[100] H. Kandrup, Astrophys. J. 370, 312 (1991)
[101] J.P. Doremus, M.R. Feix, G. Baumann, Phys. Rev. Lett. 26, 725 (1971)
[102] J.P. Doremus, M.R. Feix, G. Baumann, Astron. Astrophys. 29, 401 (1973)
[103] D. Gillon, M. Cantus, J.P. Doremus, G. Baumann, Astron. Astrophys. 50, 467 (1976)
[104] J.F. Sygnet, G. Des Forets, M. Lachieze-Rey, R. Pellat, Astrophys. J. 276, 737 (1984)
[105] H. Kandrup, J.F. Sygnet, Astrophys. J. 298, 27 (1985)
[106] M. Lemou, F. Méhats, P. Raphaël, Commun. Math. Phys. 302, 161 (2011)
[107] P.H. Chavanis, Eur. Phys. J. Plus 135, 290 (2020)
[108] J.R. Ipser, Astrophys. J. 193, 463 (1974)
[109] J.R. Ipser, G. Horwitz, Astrophys. J. 232, 863 (1979)
[110] P.H. Chavanis, C. Sire, Physica A 356, 419 (2005)
[111] P.H. Chavanis, Astron. Astrophys. 451, 109 (2006)
[112] T. Padmanabhan, Phys. Rep. 188, 285 (1990)
[113] J. Katz, Found. Phys. 33, 223 (2003)
[114] P.H. Chavanis, Int. J. Mod. Phys. B 20, 3113 (2006)
[115] P.H. Chavanis, M. Lemou, F. Méhats, Phys. Rev. D 91, 063531 (2015)
[116] P.H. Chavanis, M. Lemou, F. Méhats, Phys. Rev. D 92, 123527 (2015)
 49

[117] P.H. Chavanis, Phys. Rev. D 100, 083022 (2019)

[118] P.H. Chavanis, A heuristic wave equation parameterizing BEC dark matter halos with a quantum core and an isothermal
atmosphere [arXiv:2104.09244]
[119] P.H. Chavanis, Predictive model of fermionic dark matter halos with a quantum core and an isothermal atmosphere
[arXiv:2112.07726]
[120] E. Seidel, W.M. Suen, Phys. Rev. Lett. 72, 2516 (1994)
[121] D.G. Levkov, A.G. Panin, I.I. Tkachev, Phys. Rev. Lett. 121, 051301 (2018)
[122] B. Bar-Or, J.B. Fouvry, S. Tremaine, Astrophys. J. 871, 28 (2019)
[123] F. Bouchet, Physica D 237, 1976 (2008)
[124] T.D. Frank, Nonlinear Fokker-Planck Equations: Fundamentals and Applications (Springer-Verlag, 2005)
[125] P.H. Chavanis, C. Sire, Physica A 375, 140 (2007)
[126] O. Klein, Arkiv för Matematik, Astronomi, och Fysik 16, 1 (1921)
[127] H.A. Kramers, Physica A 7, 284 (1940)
[128] Lord Rayleigh, Phil. Mag. 32, 424 (1891)
[129] A. Einstein, Ann. Physik 17, 549 (1905)
[130] P.H. Chavanis, Eur. Phys. J. Plus 134, 353 (2019)
[131] M. von Smoluchowski, Ann. Physik 48, 1103 (1915)
[132] J.P. Dougherty, Phys. Fluids 7, 1788 (1964)
[133] P.H. Chavanis, Eur. Phys. J. B 78, 139 (2010)