The vmf90 program for the numerical resolution of the Vlasov equation for mean-field

systems
Pierre de Buyl

arXiv:1310.0805v2 [physics.comp-ph] 25 Mar 2014

Center for Nonlinear Phenomena and Complex Systems, Universite libre de Bruxelles, CP231, Campus Plaine, 1050 Brussels, Belgium

Abstract
The numerical resolution of the Vlasov equation provides complementary information with respect to analytical studies and forms
an important research tool in domains such as plasma physics. The study of mean-field models for systems with long-range
interactions is another field in which the Vlasov equation plays an important role. We present the vmf90 program that performs
numerical simulations of the Vlasov equation for this class of mean-field models with the semi-Lagrangian method.
Keywords: Vlasov equation ; Hamiltonian Mean-Field model ; Single Wave model

PROGRAM SUMMARY
Manuscript Title: The vmf90 program for the numerical resolution of
the Vlasov equation for mean-field systems
Author: Pierre de Buyl
Program Title: vmf90
Journal Reference:
Catalogue identifier:
Licensing provisions: vmf90 is available under the GNU General
Public License
Programming language: Fortran 95
Computer: Single CPU computer
Operating system: No specific operating system, the program is tested
under Linux and OS X
RAM: about 5 M bytes
Keywords: Vlasov equation, semi-Lagrangian method, Hamiltonian
Mean-Field model, Single-Wave model
Classification: 1.5 Relativity and Gravitation, 19.8 Kinetic Models,
19.13 Wave-plasma Interactions, 23 Statistical Physics and Thermodynamics.
External routines/libraries: HDF5 for the code (tested with HDF5
v1.8.8 and above). Python, NumPy, h5py and matplotlib for analysis.
Nature of problem:
Numerical resolution of the Vlasov equation for mean-field models
(Hamiltonian Mean-Field model and Single Wave model).
Solution method:
The equation is solved with the semi-Lagrangian method and cubic
spline interpolation.
Running time:
The examples provided with the program take 1m30 for the HamiltonianMean Field model and 10m for the Single Wave model, on an Intel
Core i7 CPU @ 3.33GHz. Increasing the number of grid points or the
number of time steps increases the running time.

Present address: Department of Chemistry, Katholieke Universiteit Leuven, Celestijnenlaan 200F, B-3001 Heverlee, Belgium

Preprint submitted to Computer Physics Communications

1. Introduction
The Vlasov equation provides an appropriate description
of collisionless plasmas, beam propagation, collisionless gravitational dynamics, among other systems of interest. Another
field of application is the study of mean-field models for longrange interactions [14] In these systems, thermodynamical
and dynamical properties can differ strongly with respect to
short-range interacting systems. The understanding of the
kinetic properties in these systems relies importantly on their
Vlasov description. Theoretical considerations and N-body
simulations form the majority of the literature but numerical
simulations of the Vlasov equation are found in a small number
of references [513]. These simulations provide a complementary point of view to N-body simulations by removing
the dependence on the number of simulated particles. The
number N of particles present in a simulation has a significant
impact on the dynamics of these long-range systems, especially on the timescale of the relaxation to equilibrium. Many
studies on long-range interacting systems are performed on
one-dimensional models. This makes the numerical resolution of the Vlasov equation by Eulerian methods, instead of
particle-in-cell methods, computationally accessible.
Numerical resolutions of the Vlasov equation are being
performed since decades, following the work of Cheng and
Knorr [14] on the time discretization of the problem. Most
of the research in this field is related to the simulation of the
Vlasov-Poisson system in plasma physics. As a consequence,
no code is available for the simulation of simpler mean-field
models. Altough existing codes for the Vlasov-Poisson system
could be easily converted for use with mean-field systems, we
propose here a code that is aimed specifically at these systems.
The vmf90 program presented in this article has already
been used to perform the numerical resolution of the Vlasov
equation for the Hamiltonian Mean-Field or free electron laser
models in the references [6] (at that time, a previous C++
March 26, 2014

version was used) and [712]. The results in the present

manuscript have been obtained with version 1.0.0 of the vmf90
program.

2.2. The single-wave model

The so-called single-wave model is a model consisting of
many particles interacting with a single wave. This model originates from the study of beam-plasma interactions [16] but also
proves adequate to study wave-particle interactions, free electron lasers [17] and collective atomic recoil lasing [18].
The Vlasov equation for the single wave model reads

2. Physical models
2.1. The Hamiltonian Mean-Field model
The Hamiltonian Mean-Field (HMF) model, introduced in
Ref. [15], consists of particles interacting via a cosine potential
in a periodic box. It is intended as a simplified description of
systems of charged or gravitational particles in which collective
behavior, caused by the all-to-all coupling of the particles, plays
an important role.
The HMF model serves as a reference model in the field of
systems with long-range interactions [3]. It exhibits peculiar
dynamical properties among which is the existence of the socalled quasi-stationary states (QSS). These states occur, starting from an out-of-equilibrium initial condition, after a O(1)
timescale but do not correspond to the equilibrium predicted by
statistical mechanics. The lifetime of the QSS increases algebraically with the number N of interacting particles in the system. In the thermodynamic limit, the dynamics remains stuck
in the QSS regime.
The Vlasov equation for the HMF model reads
+

V[ f ]()

mx [ f ]

1Z m x [ f ] cos my [ f ] sin ,
dd p f cos ,
Z
dd p f sin ,

=
=
=

f
f
+ 2(A x cos Ay sin )
,
Z
p
Ay + dd p f cos ,
Z
A x dd p f sin ,

(5)

where f = f (, p; t) is the one-particle distribution function and

A x and Ay are the two components of the wave. Specifying
A x and Ay is equivalent to specifying a phase and an intensity.
is a parameter that accounts for the mismatch between the
resonant frequency of the device and the energy of the particle,
it is commonly called the detuning parameter.
The time evolution of Eqs. (1) conserves the following
quantities: the L1 norm
Z
|| f ||1 =
d d p f (, p; t) ,
(6)

f dV[ f ] f

=0,

d p

f
t

my [ f ]

f
t
A x
t
Ay
t

the total momentum


 Z
P(t)[ f ] = A2x + A2y + d d p f (, p; t) p ,

(7)



where A2x + A2y is the momentum of the wave, and the energy
U
Z

!
p2
+ 2 Ay cos + A x sin .
U(t)[ f ] =
d d p f (, p; t)
2
(8)

(1)

where f = f (, p; t) is the one-particle distribution function, V

is the potential, depending self-consistently on f and m x and my
are the two components of the magnetization vector.
The time evolution of Eqs. (1) conserves the following
quantities: the L1 norm
Z
|| f ||1 =
d d p f (, p; t) ,
(2)

2.3. Other models

Other related models can be implemented easily in the code
by taking as an example the HMF model or the single-wave
model files.

the total momentum

P(t)[ f ] =

3. Description of the vmf90 program

d d p f (, p; t) p ,

(3)
vmf90 is written in Fortran 90 and takes advantage of userdefined types, dynamic memory allocation and modules. Fortran 90 provides, with respect to older revisions of the Fortran programming language, many improvements for the structured coding of scientific programs. The reader is referred to
Refs. [1921] for more information on this topic. In addition to
these ideas, vmf90 takes profit of revision control with the git
software [22] and extensive in-code documentation.

and the energy U

Z
U(t)[ f ] = d d p f (, p; t)
!
p2 1 
+ 1 m x [ f ] cos my [ f ] sin .
2
2

(4)

These quantities are discussed in Ref. [7].

3.1. General structure

At the core of the program lies the module Vlasov module
that provides the user-defined type (UDT) grid for phasespace numerical grids. The grid UDT contains information
on:

The program is under revision control with the git software [22]. The information on the status of the source code
repository is included automatically during the build process
of the software and is updated at each rebuild so that the executable program always reflect this status. The programs display revision information: the commit reference (the SHA1) of
the code as well as a status that is one of the following

1. The number of grid points.

2. The spatial and velocity limits of the grid.
3. Storage for:

Clean: the code corresponds to the displayed git commit

reference.

The grid.

Unclean: the code is based on the displayed git commit

reference but has been modified.

The position and velocity marginal distributions.

The second derivative required for the cubic spline
interpolation.

Out of repository: the code has been extracted from a

tarball whose version is given, but is not tracked by git.

The force to apply when evolving the velocities.

4. The periodic or finite character of the spatial dimension.
5. The time step to evolve the system.

In this manner, the results can be linked in a strict manner to

the exact version of the code. It is necessary to use a git-tracked
repository to take profit of this feature.
Due to the use of Fortran UDTs and modules, the full status
of a simulated system is held in a variable. There is thus no need
for global variables (i.e. the common block feature of Fortran
is not used, nor are module-global variables).
All modules are commented in the style of the Doxygen [23] program. A full documentation for the API is thus
automatically generated.
The program is distributed with the ParseText Fortran module that allows an easy parsing of human-readable text files for
Fortran programs. This provides the advantage that the user
is less likely to set the wrong parameters in a configuration file
that would contain only numbers without a close by description.
For instance, reading the time step is done with the following
line of code:

The Fortran declaration of the grid UDT is shown explicitly

in Fig. 1.
This UDT is then used in the description of the HMF and
FEL models. The module Vlasov module also provides the
routines to perform the basic steps for solving the Vlasov equation:
1.
2.
3.
4.

Obtention of the grid coordinates for a given grid point.

Advection in the x- and p-direction.
Computation of the distribution marginals in x and v.
Initialization of the grid with Gaussian or water bag distributions.

For each model, a module provides a UDT containing a

grid variable and additional information on the model, such
as the mean-field variables. A program using this UDT is written for each model, i.e. vmf90 hmf and vmf90 fel, that
sets the simulation up and performs the steps required for the
simulation. As the code that performs the actual steps in the
simulation algorithm is found in Vlasov module, the implementation of other models can benefit from a verified code base.
A separate module, spline module, contains the routines
for the cubic spline interpolation for natural splines (for the velocity dimension and the non-periodic spatial dimension) and
periodic splines (for the spatial dimension in periodic systems).
The vmf90 module provides information on the code version (see more in section 3.2).

DT = PTread_d(PT_variable, DT)
where DT is a double precision variable, PT variable holds
the information from the configuration file and the argument
DT indicates which variable should be looked upon in the
configuration file that is parsed. The routine PTread d is suited
for double precision variables and similar routines exist for the
other common Fortran data types. A configuration file is given
as an example in the Appendix Appendix A.
3.3. Data storage model and analysis tools
The vmf90 program stores data in the HDF5 format that
allows to store data to machine precision in structured files that
simplify the storage and analysis of the data. A derivative of the
H5MD [24, 25] file format is used for that purpose, allowing
a consistent storage of parameters and observables across the
file. Also, the version information for vmf90 described in the
previous section is part of the H5MD file.
To allow users who are not familiar with HDF5 or H5MD
to analyze the data, a small program displaying common results for the simulation is provided. It is a command-line tool
that takes as arguments the filename of the data and a command: snaps to display snapshots of phase space and plot
to display observables such as mass and energy.

3.2. Scientific coding guidelines

With the aim of providing a rigorous scientific computing
tool, vmf90 follows the following principles: use of revision
control, i.e. the revision information is provided in the output file of a vmf90 simulation, no use of global variables for
the computation routines, use of human readable configuration
files instead of hardcoding computation parameters in the program, use of self-describing data files, code documentation with
Doxygen [23]. The combination of all these principles allows
the scientist to track finely the code and parameter sets that produce a given result and the scientist-programmer to maintain a
robust codebase for future developments.
3

type grid
integer :: Nx, Nv
double precision :: xmin, xmax, vmin, vmax
double precision :: dx, dv
double precision :: DT
double precision :: en_int, en_kin, energie, masse, momentum
logical :: is_periodic
double precision, allocatable :: f(:,:)
double precision, allocatable :: d2(:), copy(:)
double precision, allocatable :: rho(:), phi(:), force(:)
end type grid
Figure 1: The Fortran declaration of the grid UDT that serves as the building block for physical models.

This program is written in Python and uses h5py [26],

NumPy [27] and Matplotlib [28]. Its purpose is at the same
time to provide a tool to check on a simulation file as well
as an example program that shows how to access and use the
resulting datafile.

by following the trajectory of the transported quantity backward

in time. The semi-Lagrangian methodology does not impose a
CFL condition on the time step [30]. The semi-Lagrangian algorithm for the Vlasov equation is described in Ref. [30]. We
reproduce here the steps of the algorithm for clarity.
f s (xi , vm ) is the distribution function at time step s and grid
point i, m. i is the spatial index and m the velocity index. The
time discretization reads

3.4. Obtaining, building and running vmf90

vmf90 is available under the GNU General Public License,
on the GitHub code hosting website 1 . In order to have a working version of vmf90 on your computer, here are the steps to
execute in a terminal. Those instructions are provided for the
sake of completeness and to facilitate the reader first steps with
vmf90. The following programs are required: git [22], a Fortran
compiler [e.g. gfortran [29]] and the HDF5 library.

1
f s+1 (x, v) = f s (x t(v + F ( x)t), v tF ( x)) ,
2

(9)

where x = x vt/2 and F is the force field computed at half

a time step.
A practical implementation to compute Eq. (9) on a numerical mesh is the following :

# obtain vmf90 from its main repository

git clone https://github.com/pdebuyl/vmf90.git
# enter vmf90 directory
cd vmf90
# create the directory in which vmf90 is built
mkdir build
# enter the directory in which vmf90 is built
cd build
# build vmf90 for the hmf model
make -f ../scripts/Makefile hmf

1. Advection in the -direction, 1/2 time step

f (i , pm ) = f s (i pm t/2, pm ).
2. Computation of the force field for f .
3. Advection in the p-direction, 1 time step
f (i , pm ) = f (i , pm F (i )t).
4. Advection in the -direction, 1/2 time step
f s+1 (i , pm ) = f (i pm t/2, pm ).
The evaluation of f is done with cubic spline interpolation.
4.2. Boundaries of the domain
The interpolation is made to return the following quantities
when a point outside of the computational box is queried for
interpolation:

At this point, the program vmf90 hmf is created within

the build directory. Running the program is done by typing ./vmf90 hmf and requires to have a configuration file
HMF in in the directory. Section 5 presents an example run
for the HMF model.

In a non-periodic box, f is evaluated with zero value

whenever
v vmax or v vmin or x xmax or x xmin

4. Algorithms

(10)

In a periodic system, x values outside of the box are evaluated at their periodic image inside the box. v values
outside the box are still evaluated to zero.

4.1. Semi-Lagrangian algorithm

The semi-Lagrangian algorithm provides a way to solve numerically an advection equation for data stored on a mesh. Instead of using finite difference equations, the solution is found

The evaluation as zero outside of the box corresponds to a zerovalue boundary condition. The box needs to be taken large
enough so that the normalization (conservation of the L1 norm)
is satisfied.

1 https://github.com/pdebuyl/vmf90

5. Simulation of the Hamiltonian Mean-Field model

6. Simulations for the single-wave model

For reference, we reproduce the first Vlasov simulations for

the HMF model [5] with a waterbag initial condition M0 = 0.5,
U = 0.69. The configuration file is found in the scripts directory of the vmf90 software as HMF in.resonances.
Assuming that you are in the directory build, as in
Sec. 3.4, the instructions are

The single-wave model [16] represents an ensemble of particles in which the interaction is well represented by a single
component of the electromagnetic field. In addition to the particles, represented by the distribution function f , one needs to
consider also the field A x , Ay in the simulation as evidenced by
the coupled evolution equations (5). As isqtypical, we track the

evolution of the intensity of the field I = A2x + A2y .

We consider a simulation that was presented in Ref. [6] with
a homogeneous water bag initial condition of energy U = 0.2
and a seed intensity given by A x = 0.0001, Ay = 0. Without
the seed, even unstable initial conditions may remain stuck because of the lack of granularity of the distribution. The panels
of Fig. 3 are the direct output of the program show vmf90.py.
Mass conservation is respected with a relative accuracy of about
1012 , as it was for the HMF model. The other panels display
the evolution of the energy and of the intensity of the wave.
The snapshot displays the typical structure of a single bump
in which the majority of the distribution function is found.

# copy example configuration file

cp ../scripts/HMF_in.resonances ./HMF_in
# run simulation
./vmf90_hmf
The program creates the file hmf.h5 that contains the trajectory of the system. One may use any HDF5-capable software
to extract the information. We illustrate here the use of the program show vmf90.py that is found in the scripts directory
of vmf90.
# plots Mx and My as a function time
../scripts/show_vmf90.py hmf.h5 plot Mx My
# displays phase space snapshots
../scripts/show_vmf90.py hmf.h5 snaps

7. Discussion
We have presented an open-source implementation of the
semi-Lagrangian algorithm for Vlasov simulations tailored for
mean-field models. This code has served as the basis for several
publications and its modular nature allows for an easy addition
of other models. The Fortran code uses modern additions to
the language that allows for structured programming. Other
technical features include the use of HDF5 for the storage of
simulation data, the inclusion of a simulation analysis program,
revision control of the source code and an extensive in code
documentation.

The two above examples display information that can be saved

in a graphics file format (as supported by matplotlib, e.g. eps,
pdf, png and more). If one wishes to dump the data for analysis
with a tool that lacks HDF5 support, show vmf90.py features
a dump command.
../scripts/show_vmf90.py hmf.h5 dump Mx My
dumps t, M x and My as text into the terminal. This text can be
piped to another program or dumped into a file by appending
> t Mx My.txt to the command line above.
The figures generated by show vmf90.py are reproduced
in Fig. 2 without further processing. One can already observe
the major features of the simulation: the normalization (mass,
here) is respected to about 1012 and the energy is conserved.
The behavior of the magnetization, starting from M x = M0 =
0.5, is to decrease and then display sustained oscillations. Snapshots of the phase space are also displayed, at four regularly
spaced times of the simulation, and the clustering behavior observed in Refs. [5, 15] is well reproduced.

Acknowledgments
The author acknowledges fruitful collaboration with all of
his co-authors listed in the references. The author also acknowledges Peter H. Colberg for many useful technical discussions,
Claire Revillet and Jakub Krajniak for testing the compilation.
Appendix A. Example configuration file

5.1. Performing a parametric study

vmf90 comes with an example program, parametric run.py,
in the scripts directory to perform a parametric study of the
HMF magnetization. Such a study represents a typical use of
the program in the existing literature.
The example reproduces Fig. 3 of Ref. [11], albeit with a
smaller number of grid points and a shorter sampling time, for a
faster execution. The appearance of the figure is also changed to
a scatter plot to make the implementation of the data collection
more readable.

An example configuration file is provided, for the HMF

model, in Fig. A.4. This file is available as HMF in.stable wb
in the program repository. Each parameter appears with an explicit denomination in the file.
References
[1] P.-H. Chavanis, Lynden-bell and tsallis distributions for the hmf model,
Eur. Phys. J. B 53 (2006) 487. doi:10.1140/epjb/e2006-00405-5.
[2] P.-H. Chavanis, Hamiltonian and brownian systems with long-range interactions: II. kinetic equations and stability analysis, Physica A 361 (2006)
81. doi:10.1016/j.physa.2005.06.088.
[3] A. Campa, T. Dauxois, S. Ruffo, Statistical mechanics and dynamics of
solvable models with long-range interactions, Phys. Rep. 480 (2009) 57.
doi:10.1016/j.physrep.2009.07.001.

1e12+1

1.59994240

0.7

mass

energy
en_int
en_kin

0.6

1.39988021

0.5

1.20004007

0.4
0.99997788

0.3
0.80002671

0.2

0.59996452
0

50

time

100

0.1
0

150

0.5

Mx
My

2
1
v

0.4
0.3
0.2

20

40

2
1

0
1

1
v

0.1
0.20

20

40

60

80
time

100

120

140

1
0

120

140

160

t =38.40

2
3 2 1
2

t =80.00

2
3 2 1

160

100

0.0

80
time

t =0.00

2
3 2 1

0.1

60

t =121.60

2
3 2 1

Figure 2: Simulations of the HMF in.resonances example file. The mass evolution is zoomed to 1 with a detail of 1012 . These figures are produced, with no
modification, by the script show vmf90.py that accompanies vmf90.

1e12+1

1.5

mass

9.00002295
7.99993405

energy
en_int
en_kin

1.0

7.00006719

0.5

5.99997829
5.00011144

0.0

4.00002254
3.00004466

0.5

2.00006678
0.99997788

20

40

60
time

80

100

1.0
0

120

0.8

2
v

0.7

0.6

40

60
time

0.3

2
v

0.2

0.1

40

60

80

100

t =60.00

3 2 1

120

time

120

t =30.00

3 2 1
4

4
20

100

3 2 1

0.4

80

t =0.00

0.5

0.0
0

20

t =90.00

4
0

3 2 1

Figure 3: Simulations of the FEL in example file. The mass evolution is zoomed to 1 with a detail of 1012 . These figures are produced, with no modification, by
the script show vmf90.py that accompanies vmf90.

model = HMF
Nx = 256
Nv = 512
vmax = 2.5
DT = 0.1
n_steps = 5
n_top = 300
n_images = 50
IC = wb_eps
width = 4.
bag = 0.72
epsilon = 0.001
Nedf = 0

!
!
!
!
!
!
!
!
!
!
!
!
!

the model name. is checked in the program

number of points in x-direction
number of points in v-direction
size of the box
timestep
number of steps in inner loop
number of executions for inner loop
number of dumps of the phase space DF
initial condition
\Delta\theta for initial condition
\Delta p for initial condition
initial perturbation of the profile
number of points for energy DF
Figure A.4: Example configuration file HMF in.

[20] V. K. Decyk, H. J. Gardner, Object-oriented design patterns in fortran

90/95: mazev1, mazev2 and mazev3, Comput. Phys. Commun. 178
(2008) 611. doi:10.1016/j.cpc.2007.11.013.
[21] D. McCormack, Scientific Software Development in Fortran, Lulu.com,
2009.
[22] L. Torvalds, Contributors. [link].
http://git-scm.com/
[23] D. van Heesch, Doxygen (1997).
http://www.stack.nl/~ dimitri/doxygen/
[24] P. de Buyl, P. Colberg, F. Hofling, H5MD : HDF5 for molecular data.
http://nongnu.org/h5md/
[25] P. de Buyl, P. H. Colberg, F. Hofling, H5md: A structured, efficient, and portable file format for molecular data, Comp. Phys. Commun.doi:10.1016/j.cpc.2014.01.018.
[26] A. Collette, HDF5 for Python (2008).
http://h5py.alfven.org/
[27] T. E. Oliphant, Python for scientific computing, Computing in Science &
Engineering 9 (2007) 1020. doi:10.1109/MCSE.2007.58.
[28] J. D. Hunter, matplotlib (2002).
http://matplotlib.org/
[29] A. Vaught, P. Brook, Gnu fortran (2002).
http://gcc.gnu.org/fortran/
[30] E. Sonnendrucker, J. Roche, P. Bertrand, A. Ghizzo, The semi-lagrangian
method for the numerical resolution of the vlasov equation, J. Comp.
Phys. 149 (1999) 201. doi:10.1006/jcph.1998.6148.

[4] F. Bouchet, S. Gupta, D. Mukamel, Thermodynamics and dynamics of

systems with long-range interactions, Physica A 389 (2010) 43894405.
doi:10.1016/j.physa.2010.02.024.
[5] A. Antoniazzi, F. Califano, D. Fanelli, S. Ruffo, Exploring
the thermodynamic limit of hamiltonian models:
Convergence
to the vlasov equation, Phys. Rev. Lett. 98 (2007) 150602.
doi:10.1103/PhysRevLett.98.150602.
[6] P. de Buyl, D. Fanelli, R. Bachelard, G. D. Ninno, Out-ofequilibrium mean-field dynamics of a model for wave-particle
interaction, Phys. Rev. ST Accel. Beams 12 (2009) 060704.
doi:10.1103/PhysRevSTAB.12.060704.
[7] P. de Buyl, Numerical resolution of the vlasov equation for the hamiltonian mean-field model, Commun. Nonlinear Sci. Numer. Simulat. 15
(2010) 21332139. doi:10.1016/j.cnsns.2009.08.020.
[8] P. de Buyl, D. Mukamel, S. Ruffo, Self-consistent inhomogeneous steady
states in hamiltonian mean-field dynamics, Phys. Rev. E 84 (2011)
061151. doi:10.1103/PhysRevE.84.061151.
[9] P. de Buyl, D. Mukamel, S. Ruffo, Statistical mechanics of collisionless
relaxation in a non-interacting system, Phil. Trans. R. Soc. A 369 (2011)
439452. doi:10.1098/rsta.2010.0251.
[10] P. de Buyl, P. Gaspard, Effectiveness of mixing in violent relaxation, Phys.
Rev. E 84 (2011) 061139. doi:10.1103/PhysRevE.84.061139.
[11] P. de Buyl, D. Fanelli, S. Ruffo, Phase transitions of quasistationary states
in the hamiltonian mean field model, Central European Journal of Physics
10 (2012) 652659. doi:10.2478/s11534-012-0010-6.
[12] P. de Buyl, G. De Ninno, D. Fanelli, C. Nardini, A. Patelli, F. Piazza,
Y. Y. Yamaguchi, Absence of thermalization for systems with long-range
interactions coupled to a thermal bath, Phys. Rev. E 87 (2013) 042110.
doi:10.1103/PhysRevE.87.042110.
[13] S. Ogawa, Y. Y. Yamaguchi, Linear response theory in the vlasov equation
for homogeneous and for inhomogeneous quasistationary states, Phys.
Rev. E 85 (2012) 061115. doi:10.1103/PhysRevE.85.061115.
[14] C. Z. Cheng, G. Knorr, The integration of the vlasov equation in configuration space, J. Comp. Phys. 22 (1976) 330.
doi:10.1016/0021-9991(76)90053-X.
[15] M. Antoni, S. Ruffo, Clustering and relaxation in hamiltonian long-range dynamics, Phys. Rev. E 52 (1995) 2361.
doi:10.1103/PhysRevE.52.2361.
[16] T. ONeil, J. H. Winfrey, J. H. Malmberg, Nonlinear interaction of
a small cold beam and a plasma, Phys. Fluids 14 (1971) 1204.
doi:10.1063/1.1693587.
[17] Y. Elskens, D. Escande, Microscopic Dynamics of Plasmas and Chaos,
Institute of Physics Publishing, 2003.
[18] R. Bachelard, T. Manos, P. de Buyl, F. Staniscia, F. S. Cataliotti, G. D.
Ninno, D. Fanelli, N. Piovella, Experimental perspectives for systems
based on long-range interactions, J. Stat. Mech. 2010 (2010) P06009.
doi:10.1088/1742-5468/2010/06/P06009.
[19] V. K. Decyk, C. D. Norton, B. K. Szymanski, How to support inheritance
and run-time polymorphism in fortran 90, Comput. Phys. Commun. 115
(1998) 9. doi:10.1016/S0010-4655(98)00101-5.