Professional Documents
Culture Documents
Trails in Kinetic
Theory
Foundational Aspects and Numerical
Methods
SEMA SIMAI Springer Series
Volume 25
Editors-in-Chief
Luca Formaggia, MOX–Department of Mathematics, Politecnico di Milano,
Milano, Italy
Pablo Pedregal, ETSI Industriales, University of Castilla–La Mancha, Ciudad Real,
Spain
Series Editors
Mats G. Larson, Department of Mathematics, Umeå University, Umeå, Sweden
Tere Martínez-Seara Alonso, Departament de Matemàtiques, Universitat Politècnica
de Catalunya, Barcelona, Spain
Carlos Parés, Facultad de Ciencias, Universidad de Málaga, Málaga, Spain
Lorenzo Pareschi, Dipartimento di Matematica e Informatica, Università degli Studi
di Ferrara, Ferrara, Italy
Andrea Tosin, Dipartimento di Scienze Matematiche “G. L. Lagrange”, Politecnico
di Torino, Torino, Italy
Elena Vázquez-Cendón, Departamento de Matemática Aplicada, Universidade de
Santiago de Compostela, A Coruña, Spain
Paolo Zunino, Dipartimento di Matemática, Politecnico di Milano, Milano, Italy
As of 2013, the SIMAI Springer Series opens to SEMA in order to publish a
joint series aiming to publish advanced textbooks, research-level monographs and
collected works that focus on applications of mathematics to social and industrial
problems, including biology, medicine, engineering, environment and finance.
Mathematical and numerical modeling is playing a crucial role in the solution of
the complex and interrelated problems faced nowadays not only by researchers
operating in the field of basic sciences, but also in more directly applied and
industrial sectors. This series is meant to host selected contributions focusing on the
relevance of mathematics in real life applications and to provide useful reference
material to students, academic and industrial researchers at an international level.
Interdisciplinary contributions, showing a fruitful collaboration of mathematicians
with researchers of other fields to address complex applications, are welcomed in
this series.
THE SERIES IS INDEXED IN SCOPUS
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2021
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
In the last decades, kinetic theory has emerged as one of the most prominent
fields of modern mathematics. It was originally developed as a field of Mathe-
matical Physics to investigate interacting particle systems and their corresponding
continuum descriptions. Yet, in the recent years, there has been an explosion of
applications of kinetic theory to other areas of research such as biology and social
sciences. Kinetic-type equations currently represent a common ground for the cross-
fertilization between a heterogeneity of communities that include both pure and
applied disciplines.
The source for the broad applicability of kinetic theory lies on the omnipresence
of emergent phenomena in real-life applications. Emergent phenomena correspond
to the appearance of large-scale (observable) structures from the underlying micro-
scopic, discrete dynamics. Kinetic theory provides precisely the mathematical
framework to link these discrete dynamics with their corresponding continuum
equations at the macro-scale. Nowadays, countless applications, ranging from
plasma physics to socio-economic and soft matter, have roots in kinetic theory.
At the same time, the investigation of emerging phenomena in these new fields
of applications presents many mathematical challenges at the level of modelling,
mathematical analysis and numerics.
All these aspects have been presented in the School Trails in Kinetic Theory:
Foundational Aspects and Numerical Methods organized by the editors of the
present book during the Trimester Program on Kinetic Theory at the Hausdorff
Institute for Mathematics of Bonn in June 2019. During this event, four eminent
lecturers and eleven invited speakers of the highest profile presented research
advances and cutting-edge results to broadest scientific community. In the following,
we briefly describe the main topics of this event, which will constitute the backbone
for the book.
Theoretical Aspects
The problem of deriving macroscopic evolution equations from the microscopic
description based on the fundamental laws of mechanics, through suitable scaling
limits, is a central problem of non-equilibrium statistical mechanics. The resulting
v
vi Preface
kinetic equations are essential to describe the relevant physical properties of the
system and their time evolution. Classical examples of kinetic equations are the
Boltzmann equation, the Landau equation and the Vlasov equation. During the
last couple of decades, several important results have been achieved, both in the
derivation and in the analysis of the kinetic equations, which have increased the
current understanding of kinetic theory. Most of the rigorous results available in this
direction have been obtained for dilute gases or for systems with weak interactions.
The rigorous analysis of the mean-field limit and the Vlasov equation is a well-
understood subject in the case of smooth potentials, but the interesting case of
the Coulomb interaction is still open. Much more subtle are the limiting physical
situations leading to the Boltzmann equation and the Landau equation. After the
fundamental Lanford’s result on the rigorous derivation of the Boltzmann equation
from a gas of hard spheres even though only for a short time interval, the case
of smooth short-range interaction potentials has been studied, but the validity (or
non-validity) of the Boltzmann equation in the case of long-range potentials is still
open and challenging, as well as the short-time validity limitation. Concerning the
derivation of the Landau equation, a rigorous proof is still missing, even for short
time. Recently, some partial results on the extension to long-range interactions have
been proposed in the simplified case of the Lorentz gas, which consists of a single
particle moving through infinitely heavy, randomly distributed scatterers. Significant
progress has been made in the derivation of kinetic equations for quantum particles,
despite that the understanding of interacting particle systems is much more partial
than the classical ones.
Another well-established research direction in kinetic theory is the qualitative
behaviour of the solutions of kinetic equations, and the analysis of their long-
time asymptotics, closely related to the problem of hydrodynamics. It is worth
to mention that also in this direction several hard problems still need further
investigation, as, for instance, the proof of global well-posedness of classical
solutions of Boltzmann equation, which is still elusive. Kinetic equations are also
used in various other settings. One example is, for instance, the Smoluchowski
equation arising from problems of polymerization, particle aggregation in aerosols
or drop formation in rain. The analysis and validation from particle systems of
the Smoluchowski equation is much less developed, but there has recently been
progress in understanding the properties of self-similar solutions and a first step
in understanding the derivation of coagulation equations from mechanical particle
systems.
Applications in Socio-Economic and Life Sciences
As mentioned before, emergent phenomena are ubiquitous in nature. Being able to
link phenomena at the different scales is crucial to giving answers to questions in the
experimental sciences: How do we explain the self-organization of a tissue from its
underlying constituents? How crowds of pedestrian self-organize into lanes? How
does opinion dynamics evolve over time from local and macroscopic interactions?
Kinetic theory has found in the recent years multiple applications in biology and
social sciences, especially in the fields of collective motion and opinion dynamics.
Preface vii
This poses many new challenges: first, at the level of the modelling, we need
to consider simple models that are tractable enough but that, at the same time,
capture the phenomena under investigation; second, new coarse-graining tools need
to be developed to obtain macroscopic equations since the classical tools cannot
be applied (this can be due, for example, to the lack of conserved quantities or
to the appearance of phase transitions); third, new numerical methods need to be
developed (for example, for hyperbolic non-conservative equations); finally, some
systems exhibit violation of the propagation of chaos, making it impossible to derive
kinetic equations with classical methods. In conclusion, the applications of kinetic
theory to emergent phenomena in biology and social sciences open new fascinating
questions that will push further the borders of our mathematical understanding and
methodologies.
Numerical Methods and Uncertainty Quantification
The development of numerical methods for kinetic equations has been the subject
of extremely active investigations in the past decades, especially in relation to
efficient approximations of equilibrium states and in connection to the multi-scale
limits of collisional equations. Significant progress has been obtained in a variety
of problems in kinetic theory, e.g. granular gases, kinetic methods for soft matter
physics, optimal control of kinetic equations.
Furthermore, in recent years, significant efforts have been dedicated to incor-
porate possible deviations from the systems’ prescribed deterministic behaviour.
Our analytical understanding of the structural randomness seems to be crucial
to provide reliable descriptions of real-world models. A step towards realistic
modelling demands a quantification of the possible deviations of a model-driven
approach measuring errors and uncertainties. In a kinetic setting, the general strategy
to take into consideration the realistic lack in information due, for example, to
empirical assumptions or incomplete knowledge of boundary terms or initial data
relies on an increased dimensionality of the particles’ distribution. In the recent
literature, authors deal with this challenge through uncertainty quantification (UQ)
methods whose approach provides accurate algorithms for the a priori estimation of
the impact of uncertainties in terms of statistical moments. All these methods seem
to be particularly appropriate also in connection to socio-economic and life sciences
phenomena since most of the models are generally not derived from first principles.
The editors of this book wish to express their gratitude to the Hausdorff Institute for
Mathematics (HIM) of Bonn, Germany, for their financial support in organizing the
School “Trails in Kinetic Theory: Foundational Aspects and Numerical Methods”,
May 20–24, 2019. We would like to thank all the speakers of this exciting event
for their inspiring presentations that nourished fruitful discussions among the
participants.
ix
Contents
xi
Contributors
xiii
xiv Contributors
José Antonio Carrillo, Jingwei Hu, Zheng Ma, and Thomas Rey
Abstract Over the past decades, kinetic description of granular materials has
received a lot of attention in mathematical community and applied fields such as
physics and engineering. This article aims to review recent mathematical results in
kinetic granular materials, especially for those which arose since the last review
Villani (J Stat Phys 124(2):781–822, 2006) by Villani on the same subject. We
will discuss both theoretical and numerical developments. We will finally showcase
some important open problems and conjectures by means of numerical experiments
based on spectral methods.
J. A. Carrillo ()
Mathematical Institute, University of Oxford, Oxford, UK
e-mail: carrillo@maths.ox.ac.uk; carrillo@imperial.ac.uk
J. Hu · Z. Ma
Department of Mathematics, Purdue University, West Lafayette, IN, USA
e-mail: jingweihu@purdue.edu; ma531@purdue.edu
T. Rey
Univ. Lille, CNRS, UMR 8524, Inria – Laboratoire Paul Painlevé, Lille, France
e-mail: thomas.rey@univ-lille.fr
description of rapid granular flows [57, 65, 66] has been able to compute transport
coefficients for hydrodynamic descriptions successfully used in situations that are
a long way from their supposed limits of validity, to describe, for instance, shock
waves in granular gases [28, 82], clustering [32, 37, 61], and the Faraday instability
for vibrating thin granular layers [27, 28, 36, 46, 73, 88–90]. A large amount of
practical systems can be described as a granular gas, such as for example spaceship
reentry in a dusty atmosphere (Mars for instance), planetary rings [7, 70] and sorting
behavior in vibrating layers of mixtures. A lot of other examples can be found in the
thesis manuscript [40], and in the seminal book of Brilliantov and Pöschel [31].
Usually, a granular gas is composed of 106–1016 particles. The study of such
a system will then be impossible with a direct approach, and we shall adopt a
kinetic point of view, studying the behavior of a one-particle distribution function
f , depending on time t ≥ 0, space x ∈ ⊂ Rdx and velocity v ∈ Rd , for dx ≤ d ∈
{1, 2, 3}. The statistical mechanics description of the system has been then admitted
in the physical community as the tool to connect the microscopic description to
macroscopic system of balance laws in rapid granular flows [30, 57, 59, 65, 66] as
in the classical rarefied gases [39]. In this first section, we shall review some basics
on the inelastic Boltzmann equation, and present the mathematical state of the art
since the previous review paper on the subject [91].
Microscopic Dynamics The microscopic dynamics can be summarized with the
following hypotheses:
1. The particles interact via binary collisions. More precisely, the gas is rarefied
enough so that collisions between 3 or more particles can be neglected.
2. These binary collisions are localized in space and time. In particular, all the
particles are considered as point particles, even if they describe macroscopic
objects.
3. Collisions preserve mass and momentum, but dissipate a fraction 1 − e of the
kinetic energy in the impact direction, where the inelasticity parameter e ∈ [0, 1]
is called restitution coefficient:
⎧
⎨ v + v∗ = v + v∗ ,
⎪
1−e 2 (1)
⎪
⎩ |v |2 + |v∗ |2 − |v|2 − |v∗ |2 = − |(v − v∗ ) · ω|2 ≤ 0,
2
with ω ∈ Sd−1 being the impact direction. Using these conservation, one has the
following two possible parametrizations (see also Fig. 2) of the post-collisional
velocities, as a function of the pre-collisional ones:
• The ω-representation or reflection map, given for ω ∈ Sd−1 by
1+e
v = v − ((v − v∗ ) · ω) ω,
2
1+e
v∗ = v∗ + ((v − v∗ ) · ω) ω. (2)
2
Granular Materials 3
v∗ v
v + v∗ 1−e 1+e
v = + (v − v∗ ) + |v − v∗ |σ,
2 4 4
v + v∗ 1−e 1+e
v∗ = − (v − v∗ ) − |v − v∗ |σ. (3)
2 4 4
Remark 1 Taking e = 1 in both (2) and (3) yields the classical energy-conservative
elastic collision dynamics, as illustrated in Fig. 1.
The geometry of collisions is more complex than the classical elastic one. Indeed,
fixing v, v∗ ∈ Rd , denote by
v + v∗ 1−e v + v∗ v + v∗
± := ± (v∗ − v), O := = .
2 4 2 2
Then if u := v − v∗ is the relative velocity, one has
1+e
|+ − v | = |− − v∗ | = |u|,
4
namely v ∈ S (+ , |u|(1 + e)/4) and v∗ ∈ S (− , |u|(1 + e)/4), where S(x, r) is
the sphere centered in x and of radius r (see also Fig. 2).
Restitution Coefficient The physics literature is quite divided on the question of
whether the restitution coefficient e should be a constant or not [31]. Although most
of the early mathematical results on the topic consider a constant e [91], it seems
that this case is only realistic in dimension 1 of velocity (the so-called “collisional
cannon” described in the chapter 4 of [31] is a famous counter-example). The true
realistic case considers that e depends on the relative velocity |v−v∗ | of the colliding
particles. Even more precisely, it must be close to the elastic case 1 for small relative
velocities (namely no dissipation, elastic case), and decay towards 0 when this
relative velocity is large. The first mathematical result on this direction can be found
in [85], where
1
e(|v − v∗ |) = , (4)
1 + c |v − v∗ |γ
4 J. A. Carrillo et al.
v v
ω
σ
θ O −
v v∗
+
v∗
v∗
Fig. 2 Geometry of the inelastic collision in the phase space (dashed lines represent the elastic
case)
for a > 0. More details on the derivation of this expression can be found in [31,
Chapter 3].
Another quite rigorous study has been made in [81], with a threshold-dependent
restitution coefficient:
1 if r < r∗ ,
e(r) =
ē if r ≥ r∗ ,
Remark 2 This model is meaningful even in dimension 1, which is not the case for
elastic collisions. Indeed, such monodimensional collisions are only
{v , v∗ } = {v, v∗ },
meaning that the particle velocities are either swapped or preserved. The particles
being indistinguishable, nothing happens.1 In the 1d inelastic case, the collisions are
given using (3) by
v + v∗ e
{v , v∗ } = {v, v∗ } or ± (v − v∗ )
2 2
1
QI (f, f )(v) ψ(v) dv = f∗ f ψ + ψ∗ − ψ − ψ∗
Rd 2 Rd ×Rd ×Sd−1
· B(|v − v∗ |, cos θ, E(f )) dσ dv dv∗ ,
(6)
where the collision kernel is typically of the form B(|u|, cos θ, E(f )) =
(|u|)b(cos θ, E(f )), and E(f ) is the kinetic energy of f , namely its second
moment in velocity, the postcollisional velocities are computed by (3), and θ is the
angle between σ and u. We shall assume in all the following of this section that the
collision kernel is of generalized hard sphere type, namely
Remark 3 Note that we assumed that the collision kernel B in (6) depends on the
relative velocity, the angle of collision, and on E(f ). These former dependencies are
quite classical, but the latter is not. Nevertheless, it makes a lot of sense physically
speaking, as one can see in [83].
1 Because of that, the elastic collision operator is simply equal to 0 for a one-dimensional velocity
space, the Boltzmann equation reducing only to the free transport equation.
6 J. A. Carrillo et al.
1
QI (f, f )(v) ψ(v) dv = f∗ f ψ + ψ∗ − ψ − ψ∗
Rd 2 Rd ×Rd ×Sd−1
where the postcollisional velocities are computed by (2), θ is the angle between ω
and u, and
1+e
v=v− ((v − v∗ ) · ω) ω,
2e
1+e
v∗ = v∗ + ((v − v∗ ) · ω) ω. (10)
2e
The final strong from of the operator is QI (f, f )(v) = Q+I (f, f )(v) −
f (v) L(f )(v) with the loss part of the operator given by
Eγ
Q+
I (f, f )(v) = + +
e (|u|, cos θ )be (cos θ )
f f∗ dω dv∗ ,
Rd ×Sd−1 J (|u|, cos θ )
(11)
Granular Materials 7
with + +
e (r, s) and be (s) given by
s
be+ (s) =b , (12)
e2 + (1 − e2 )s 2
and
r r λ
+
e (r, s) = e2 + (1 − e2 )s 2 = e2 + (1 − e2 )s 2 . (13)
e e
We can derive now the following strong form of the collision operator also in the
σ -representation by just changing variable in the operator from ω to σ , see [34], to
find the expressions of the loss and the gain terms in the σ -representation:
and
Eγ
Q+
I (f, f )(v) = + +
e (|u|, cos θ )be (cos θ )
f f∗ dσ dv∗ ,
Rd ×Sd−1 J (|u|, cos θ )
(14)
with + +
e (r, s) and be (s) given by
√
(1 + e2 )s − (1 − e2 ) 2
be+ (s) =b , (15)
(1 + e2 ) − (1 − e2 )s (1 + e2 ) − (1 − e2 )s
and
λ
r r
+
e (r, s) = √ (1 + e 2 ) − (1 − e 2 )s = √ (1 + e 2 ) − (1 − e 2 )s .
2e 2e
(16)
v + v∗ 1−e 1+e
v= + (v − v∗ ) + |v − v∗ |σ,
2 4e 4e
v + v∗ 1−e 1+e
v∗ = − (v − v∗ ) − |v − v∗ |σ. (17)
2 4e 4e
8 J. A. Carrillo et al.
The granular gases collision operator has then the same structure of the elastic
Boltzmann operator under Grad’s cutoff assumption, namely it can be seen as the
difference between the inelastic gain term Q+ I (f, f ) and the loss term f L(f ),
which depends only on the chosen collision kernel, but not on the inelasticity.
We shall call the following granular gases equation, or the inelastic Boltzmann
equation:
∂f
+ v · ∇x f = QI (f, f ). (18)
∂t
We shall denote its first fluid moments (resp. density, mean momentum, and kinetic
energy) by
(ρ(f ), ρ(f )u(f ), E(f )) := 1, v, |v|2 /2 f (v) dv.
Rd
G(ρ+ )
QE (f, f )(x, v) = δ d−1 ( + +
e (|u|, cos θ )be (cos θ ) f+ f∗
Rd ×Sd−1 e
−G(ρ− )f− f∗ ) dω dv∗ ,
(19)
where ρ is the local density of f , ± denotes for a given function g = g(x) the
shorthand notation
and G is the local collision rate (also known as the correlation rate, see [91]). The
global existence of renormalized solutions for the full granular gases equation (18)
with the collision operator (19) has been established for both elastic and inelastic
collisions in [45]. Existence and L1 (dx dv) stability of such solutions has been
proved in [95], for close to vacuum initial datum.
2 Note that using a BBGKY approach [50] to derive (11) is not expected to succeed, because among
other problems the macroscopic size of the particles composing a granular gas is incompatible with
the Boltzmann-Grad scaling assumption.
Granular Materials 9
QI (F, F ) + v F = 0
10 J. A. Carrillo et al.
with α = 1 in the Maxwellian molecules case and α = 3/2 in the hard spheres case.
Indeed, the thermal bath gives an input of kinetic energy, preventing the appearance
of trivial Dirac delta equilibria. The propagation of the Sobolev norms of the space
dependent version of this equation was then established in [51]. It uses a careful
estimate of the inelastic entropy production (23), and a fixed point argument for the
existence and uniqueness of solutions.
Finally, the work [77] establishes the global well posedness of the granular
gases equation without a thermal bath, for a general case of collision kernel (which
contains (7)) and velocity dependent restitution coefficients:
Theorem 2 (Theorem 1.4 of [77]) Let 0 ≤ fin ∈ L13 ∩ ∈ BV4 . Then for any T ∈
(0, Tc ), where Tc := sup {T > 0 : E(f )(t) > 0, ∀ t < T } is the so-called blowup
time, there exists an unique nonnegative solution f ∈ C(0, T ; L12 ) ∩ L∞ (0, T ; L13 )
of (20). It preserves mass and momentum, and converges in the weak-* topology of
measures towards a Dirac delta.
Their proof relies on careful estimates of the collision operator QI in Orlicz space
(specially the L log L space of finite entropy measures).
Remark 5 The related (but still mostly open) problem of the propagation of chaos
was considered in [78] for a very simplified inelastic collision operator with a
thermal bath.
Cauchy Problem in the Space Dependent Setting The case of the space inhomo-
geneous setting3 has been much less investigated.
The first result can be found in [9] for the model introduced in [85] with a
restitution coefficient given by (4), in one dimension of space and velocity. This
work establishes the existence and uniqueness of mild (perturbative) solutions, first
for small L1 (dx dv) initial data, and then for compactly supported initial data. Their
main argument is reminiscent from a work due to Bony in [23] concerning discrete
velocity approximation of the Boltzmann equation in dimension 1.
The global existence of mild solutions in the general R3x × R3v setting, for a
large class of velocity-dependent restitution coefficient, but for initial data close
to vacuum, was obtained in [6]. The proof is based on a Kaniel-Shinbrot iteration
on a very small functional space. The stability in L1 (R3x × R3v ) under the same
assumptions was established in [94]. Finally the existence and convergence to
equilibrium in T3x × R3v for a weakly inhomogeneous granular gas4 with a thermal
bath was proved in [87], using a perturbative approach.
3 Physicallymore realistic, in part because of the spontaneous loss of space homogeneity that has
been observed in [58].
4 Namely, the initial condition is chosen with a lot of exponential moments in velocity, and close to
where D(f ) ≥ 0 is the energy dissipation functional, which depends only on the
collision kernel:
The quantity (|u|, E) is the so-called energy dissipation rate, given using (1) by
1 − e2
(|u|, E) := |u · ω|2 B(|u|, cos θ, E) dω ≥ 0, ∀e ∈ [0, 1].
4 Sd−1
(22)
This dissipation of kinetic energy has a major consequence on the behavior of the
solutions to the granular gases equation. Indeed, combined with the conservation of
mass and momentum, it implies (at least formally) an explosive behavior, namely
convergence in the weak-* topology of solutions to (18) towards Dirac deltas,
centered in the mean momentum u:
f (t, ·) δv=u , t → ∞.
As for the entropy, it is not possible to obtain any entropy dissipation for this
equation, in order to precise this large time behavior. Indeed, as noticed in [51],
taking ψ(v) = log f (v) in (6) yields
1 f f∗
QI (f, f )(v) log f (v) dv = f∗ f log B dσ dv dv∗
Rd 2 Rd ×Rd ×Sd−1 f f∗
1 f f∗ f f∗
= f∗ f log − + 1 B dσ dv dv∗
2 Rd ×Rd ×Sd−1 f f∗ f f∗
1
+ f∗ f − f∗ f B dσ dv dv∗ . (23)
2 Rd ×Rd ×Sd−1
12 J. A. Carrillo et al.
The first term in (23), the elastic contribution, is nonpositive because log λ−λ+1 ≤
0 (this is Boltzmann’s celebrated H Theorem). Nevertheless, the second term has no
sign a priori: it is 0 only in the elastic case (because of the involutive collisional
transformation (v, v∗ , σ ) → (v , v∗ , σ )). Boltzmann’s entropy
is then not dissipated by the solution of the granular gases equation if e < 1. Some
work has been done on that direction in the numerical side. Indeed, adding a drift
term or a thermal bath in velocity can yield numerical entropy dissipation, as noticed
in [53].
Kinetic Energy Dissipation and the Haff’s Cooling Law Let us assume in this
subsection that the granular gas considered is space homogeneous, namely f is
solution to (20). Having no known entropy, one has then to use other macroscopic
quantities to study the large time behavior of solutions to (18). Because of its explicit
dissipation functional, kinetic energy is a good candidate for this. Moreover, being
related to the variance, it allows to measure the concentration in velocity of the
solution.
In order to have an explicit bound for the energy dissipation, let us assume
that the collision kernel is of the general type (7). Using polar coordinates, it is
straightforward to compute the dissipation rate (22):
1 − e2 λ+2 γ
(|u|, E) = b1 |u| E , (24)
4
where thanks to (8)
π
b1 = Sd−2 cos2 (θ ) sind−3 (θ ) b(cos(θ )) dθ < ∞.
0
Using the conservation of mass and momentum, one can always assume that the
initial condition is of unit mass and zero momentum. Plugging (24) into (21) then
yields using Hölder and Jensen inequalities
d 1 − e2
E(f )(t) = −b1 E(f )γ (t) f f∗ |v − v∗ |λ+2 dv dv∗
dt 4 Rd ×Rd
1 − e2
≤ −b1 E(f )γ (t) f (v) |v|λ+2 dv
4 Rd
1 − e2
≤ −b1 E(f )1+γ +λ/2 (t). (25)
4
In particular, one will have the following large time behaviors: Setting Ce =
b1 ρ (1 − e2 )/4 and α := γ + 1/2,
Granular Materials 13
All of these formal results have been proven to be rigorous and sharp, with explicit
lower bounds, in [13, 15] for the Maxwellian and hard sphere cases [74], and in [83]
for the anomalous cases. Extension to the viscoelastic case can be found in [5, 6],
where the energy is shown to behave as
All these papers share a common approach of proof, using the fact that the
space homogeneous granular gases equation admits a self-similar behavior. Hence,
introducing some well chosen time-dependent scaling function ω and τ , the
distribution f is written as
to take into account the concentration in the velocity variables.5 The rescaled
function g is then solution to the granular gases equation, with an anti-drift term
in velocity:
∂t g + ∇v · (v g) = QI (g, g).
5 One can see the velocity scaling function ω as the inverse of the variance of the distribution f .
This scaling is then a continuous “zoom” on the blowup, and can be used to develop numerical
methods for solving the full granular gases equation, see [49].
14 J. A. Carrillo et al.
Let us consider in this subsection the following hyperbolic scaling of the granular
gases equation:
∂fε 1
+ v · ∇x fε = QI (fε , fε ). (27)
∂t ε
Determining the precise hyperbolic limit ε → 0 of Eq. (27) is a fundamental, yet
very difficult question.
Indeed, for the elastic case e = 1, one simply has to use the fact that the
equilibria of the collision operator are at the thermodynamical equilibrium (gaussian
distributions) and the conservation of mass, momentum and kinetic energy to obtain
the classical compressible Euler–Fourier system [39]. Because of the trivial Dirac
equilibria, this question is more intricate for the true inelastic case.
Pressureless Euler Dynamics Adopting the same approach as in the elastic case,
one can formally plug the “equilibrium” Dirac deltas in the pressure to obtain the
following pressureless Euler system:
⎧
⎪ ∂ρ
⎪
⎨ + ∇ · (ρ u) = 0,
∂t
(28)
⎪
⎪
⎩ ∂(ρu) + ∇ · (ρ u ⊗ u) = 0.
∂t
Granular Materials 15
This system can describe various interesting physical situations, such as galactic
clusters, but is notoriously difficult to study mathematically. Its solution are in
general ill-posed, as classical solutions cannot exists for large times and weak
solutions are not unique.
In the unidimensional case, it is however possible to recover a well posed theory
by imposing a semi-Lipschitz condition on u. This theory was introduced in [25],
and later extended in [26] and [63]. We cite below the main result of [63], where
M 1 (R) denotes the space of Radon measures on R and L2 (ρ) for ρ ≥ 0 in M 1 (R)
denotes the space of functions which are square integrable against the density ρ.
Theorem 3 (From [63]) For any ρ 0 ≥ 0 in M 1 (R) and any u0 ∈ L2 (ρ 0 ), there
exists ρ ∈ L∞ (R+ , M 1 (R)) and u ∈ L∞ (R+ , L2 (ρ)) solution to (28) in the sense
of distribution and satisfying the semi-Lipschitz Oleinik-type bound
x−y
u(t, x) − u(t, y) ≤ , for a.e. x > y. (29)
t
The proof of Theorem 3 is quite delicate, relying on duality solutions. For this
reason, we only explain the rational behind the bound (29), which can be seen very
simply from the discrete sticky particles dynamics. We refer in particular to [29] for
the limit of this sticky particles dynamics as N → ∞.
Consider N particles on the real line. We describe the ith particle at time t > 0 by
its position xi (t) and its velocity vi (t). Since we are dealing with a one dimensional
dynamics, we can always assume the particles to be initially ordered
Note in particular that particles having the same position at a given time will then
move together at the same velocity. Hence, only a finite number of collisions can
occur, as the particles aggregates.
This property also leads to the Oleinik regularity. Consider any two particles i
and j with xi (t) > xj (t). Because they occupy different positions, they have never
collided and hence xi (s) > xj (s) for any s ≤ t. If neither had undergone any
collision then xi (0) = xi (t) − vi (t) t > xj (0) = xj (t) − vj (t) t or
vi − vj + 1
< , (30)
xi − xj +
t
where x+ := max(x, 0). It is straightforward to check that (30) still holds if particles
i and j had some collisions between time 0 and t.
As one can see this bound is a purely dispersive estimate based on free
transport and the exact equivalent of the traditional Oleinik regularization for Scalar
Conservation Laws, see [80]. It obviously leads to the semi-Lipschitz bound (29) as
N → ∞.
This result was extended to the one dimensional (in space and velocity) granular
gases equation (27) in [64]. The basic idea of the proof of this work is to compare
the granular gases dynamics to the pressureless gas system (28). The main difficulty
is to show that fε becomes monokinetic at the limit (see also the very recent paper
[68]). This is intimately connected to the Oleinik property (29), just as this property
is critical to pass to the limit from the discrete sticky particles dynamics.
Unfortunately it is not possible to obtain (29) directly. Contrary to the sticky
particles dynamics, this bound cannot hold for any finite ε (or for any distribution
that is not monokinetic). This is the reason why it is very delicate to obtain the
pressureless gas system from kinetic equations (no matter how natural it may seem).
Indeed we are only aware of one other such example in [69].
One of the main contributions of [64] is a complete reworking of the Oleinik
estimate, still based on dispersive properties of the free transport operator v ∂x
but compatible with kinetic distributions that are not monokinetic, through the
introduction of a new, global nonlinear energy. The main result in this paper is the
following:
Theorem 4 (from [64]) Consider a sequence of weak solutions fε ∈ L∞ ([0, T ],
Lp (R2 )) for some p > 2 and with total mass 1 to the granular gases Eq. (27)
such that all initial v-moments are uniformly bounded in ε, some moment in x is
uniformly bounded, and fε0 is, uniformly in ε, in some Lp for p > 1. Then any
weak-* limit f of fε is monokinetic, f = ρ(t, x) δ(v − u(t, x)) for a.e. t, where
ρ, u are a solution in the sense of distributions to the pressureless system (28) while
u has the Oleinik property (29).
Quasi-Elastic Limit The physical community usually considers another approach,
that is assuming that the granular gas is in a quasi-elastic 1 − e2 ∼ ε → 0
setting. This was first proposed in [65, 66], using an approach very similar to the
Granular Materials 17
seminal Grad’s 13 moments closure for rarefied gas dynamics. The difficulty of a
hydrodynamic description of granular materials has been addressed in well reasoned
terms in [55, 56], and as already discussed in the introduction, the hydrodynamic
equations obtained with the kinetic theory of granular gases have been shown to be
insightful well beyond their supposed limit of validity, i.e., away from the quasi-
elastic limit assumption with external sources of energy. In fact, assuming that
solutions of the kinetic problem do not deviate from being Gaussians, one can then
obtain in the hard sphere case the following quasi-elastic compressible Euler system
⎧
⎪
⎪ ∂ρ
⎪
⎪ + ∇ · (ρ u) = 0,
⎪
⎪ ∂t
⎪
⎨
∂(ρu)
+ ∇ · (ρ (u ⊗ u) + ρ T I) = 0, (31)
⎪ ∂t
⎪
⎪
⎪
⎪
⎪ ∂W
⎪
⎩ + ∇ · (u (W + ρ T )) = −K ρ T 3/2,
∂t
Other types of fluid limits (such as viscous limits) of the granular gases equation
has been described in the review paper [41] and in the recent survey [54] for many
different physical regimes, but none has been rigorously established. To illustrate
the kind of equations obtained through these procedure, we write the generalized
Navier–Stokes compressible equations for granular media in conservative form, see
[36], as
∂ρ
+ ∇ · (ρu) = 0
∂t
∂(ρu)
+ ∇ · [ρ (u ⊗ u)] = ∇ · P + ρ F (32)
∂t
∂W
+ ∇ · [uW ] = −∇ · q + P : E + u · (∇ · P ) − γ + ρu · F ,
∂t
representing the evolution of number density of particles ρ(t, x), velocity
field u(t, x) and the total energy W , which is given by W = 32 ρT + 12 ρ|u|2 =
ρ + 12 ρ|u|2 , with T being the granular temperature (3D). The symbol F stands for
external forces applied to the system. The constitutive relations for the momentum
and heat fluxes write, as usual,
2
Pij = −p + λ − μ Eii δij + 2μEij
3
i
∂Uj
for the stress tensor, with Eij = 12 ∂U ∂xj
i
+ ∂xi . The thermal conductivity relates
linearly the heat flux q to the temperature gradient, q = −χ∇T .
The equation of state is relevant here: we use the expression G(ν) = [1 −
4
(ν/νmax ) 3 νmax ]−1 for the contact value of the pair correlation function G(ν), which
accounts for high densities, and the equation of state p = ρT (1 + 2(1 + e)νG(ν)),
where ν is the packing fraction and e the constant restitution coefficient. Random
close-packing is achieved in 3D at νmax = 0.65; we do not allow any packing
fraction higher than 99.99% of this value. The transport coefficients for constant
restitution given in [14, 65] write,
12
γ = √ (1 − e2 )ρT 3/2 G(ν),
σ π
for the cooling coefficient γ , which models energy dissipation through collisions.
Other kinetic coefficients are the shear viscosity,
√
πT 5 4 12
μ= ρσ +1+ 1+ G(ν) ,
6 16G(ν) 5 π
Granular Materials 19
8 √
λ = √ ρσ T G(ν)
3 π
where σ · (v− v∗ ) = cos θ (û denotes the unit vector along u), and for simplicity
we assume B does not depend on the kinetic energy E (compare with (7)). From the
numerical point of view, this does not impose any limitation since the E part can
always be factored out from the integral sign.
where
1+e
v = v − (g − |g|σ ).
4
We then assume that f has a compact support: Suppv (f ) ≈ BS , where BS is a ball
centered at the origin with radius S. Hence it suffices to truncate the infinite integral
in g to a larger ball BR with radius R = 2S:
−1
N
2
1
fˆk ei L k·v , fˆk = f (v)e−i L k·v dv.
π π
f (v) ≈ fN (v) =
(2L)d DL
k=− N2
N/2−1
Here k = (k1 , . . . , kd ) is a multidimensional index, and := k=−N/2
N/2−1 √
k1 ,...,kd =−N/2 . The choice of L should be chosen at least as L ≥ (3 + 2)S/2
to avoid aliasing, see [48] for more details. Now substituting fN into (34)
and choosing ψ(v) = e−i L k·v , we can obtain the k-th mode of the collision
π
operator as
N
−1
2
Q̂k = G(l, m)fˆl fˆm , (35)
l,m=− N2
l+m=k
In the original spectral method [48], the weight G(l, m) is precomputed and
stored since it is independent of the solution f which leads to a memory requirement
of O(N 2d ). During the online computation, the weighted sum (35) is directly
evaluated whose complexity is O(N 2d ).
To reduce the complexity of the direct spectral method as well as to alleviate its
memory requirement, the key idea introduced in [62] is to render the weighted
convolution (35) into a pure convolution so that it can be computed efficiently by
the FFT. One way to achieve this is through a low-rank approximation of G(l, m),
namely,
Np
G(l, m) ≈ αp (l + m)βp (m), (36)
p=1
22 J. A. Carrillo et al.
where αp and βp are some functions to be determined and the number of terms Np
in the expansion is small. Then (35) becomes
2 −1
N
Np
Q̂k ≈ αp (k) fˆl βp (m)fˆm , (37)
p=1 l,m=− N2
l+m=k
where the inner summation is a pure convolution of two functions fˆl and
βp (m)fˆm . Hence the total complexity to evaluate Q̂k (for all k) is brought down to
O(Np N d log N), i.e., a few number of FFTs.
Specifically, we first split G(l, m) into a gain term and a loss term:
where
−i πL m·g i πL 1+e
Ggain (l, m) := e B(|g|, σ · ĝ)e 4 (l+m)·(g−|g|σ ) dσ dg,
BR Sd−1
and
e−i L m·g
π
Gloss(m) := B(|g|, σ · ĝ)dσ dg.
BR Sd−1
N
2 −1
Q̂−
k = fˆl G(m)fˆm
l,m=− N2
l+m=k
by FFT. For the gain term, to get a decomposition of form (36), we introduce a
quadrature rule to discretize g, then Ggain (l, m) can be approximated as
wρ wĝ ρ d−1 e−i L ρ m·ĝ F (l + m, ρ, ĝ),
π
Ggain(l, m) ≈ (38)
ρ,ĝ
where ρ := |g| ∈ [0, R] is the radial part of g and ĝ ∈ Sd−1 is the angular part, and
wρ and wĝ are the corresponding quadrature weights. The function F is given by
π 1+e
F (l + m, ρ, ĝ) := B(ρ, σ · ĝ)ei L ρ 4 (l+m)·(ĝ−σ ) dσ. (39)
Sd−1
Granular Materials 23
With this approximation, the gain term of the collision operator can be evaluated as
2 −1
N
!
Q̂+ fˆl e−i L ρ m·ĝ fˆm ,
π
k ≈ wρ wĝ ρ d−1
F (k, ρ, ĝ)
ρ,ĝ l,m=− N2
l+m=k
The accuracy and efficiency of the fast spectral method has been validated in [62].
In this section, we perform some additional tests to demonstrate the potential of
the method in predicting some mathematical theories. We also introduce a GPU
implementation of the method that significantly improves the CPU version in [62].
This is critical especially for long time simulation.
We consider the following spatially homogeneous equation with a heat bath:
∂t f = QI (f, f ) + τ v f, (40)
where τ is the parameter describing the strength of the heat bath. Notice that it is not
necessarily related to the inelasticity parameter e, contrarily to e.g. [76]. The heat
bath v f will also be discretized using the Fourier spectral method and Runge–
Kutta method is used for time marching.
For the collision operator, we consider the simplified variable hard sphere kernel
For Maxwell molecules, given the initial condition f0 (v) whose macroscopic
quantities are ρ0 , u0 and T0 , the density and velocity are conserved so ρ(t) = ρ0 ,
u(t) = u0 and the temperature will evolve as
8τ ρ0 (1 − e2 ) 8τ
T (t) = T0 − exp − t + , (42)
1 − e2 4 1 − e2
We could see
8τ
lim T (t) = .
t →∞ 1 − e2
Table 1 Average running time per evaluation of the collision operator in 3D. Comparison between
the CPU and GPU-parallelized implementation for various Ns (# of Fourier basis in each velocity
dimension) and fixed Nρ = 30, Msph = 32 (# of quadrature points used in radial and spherical
direction, respectively) with 2 Xeon® Silver 4110 2.10 GHz CPUs with 4 NVIDIA Geforce GTX
2080 Ti (Turing) GPUs
N CPU GPU
8 7.68 ms 5.89 ms
16 61.2 ms 5.97 ms
32 546 ms 12.1 ms
64 5.38 s 109 ms
Granular Materials 25
In order to confirm this result, we first choose the following physical parameters
e = 0.95, τ = 1 − e = 0.05 .
For time integrator, a 4th order Runge–Kutta method is used with t = 0.01 and
we compute sufficient long time to get f∞ and in both cases the 2 difference of
solutions between the last 2 time steps are of O(10−12). The results are shown in
Fig. 3 where one can observe the exponential convergence of relative entropy (45)
from the left figure. In the right figure we find a perfect match of the temperature
evolution between numerical solution and the analytical solution (42). We also plot
the profile of the equilibrium solution f∞ in Fig. 4 which shows a Gaussian-like
density function.
26 J. A. Carrillo et al.
Fig. 3 Test 1. Convergence to equilibrium. e = 0.95 with heat bath τ = 0.05. Initial data is
the Maxwellian (44). Left: Semi-log plot of the relative entropy of f and f∞ = f (t = 100, v).
Right: numerical temperature (orange dots) with exact temperature (blue
√ line) as (42). Numerical
parameters: Nv2 = 64 × 64, Nρ = 32, Mĝ = 16, R = 20, L = 5(3 + 2) and t = 0.01
Fig. 4 Test 1. The equilibrium profile of e = 0.95 with heat bath τ = 0.05, initial data is the
Maxwellian (44). Numerical parameters: Nv2 = 64 × 64, Nρ = 32, Mĝ = 16, R = 20, L =
√
5(3 + 2) and t = 0.01
Although the exponential convergence result is only available for the case that
e = 1 − τ , we expect something similar happens even when e and τ are not related.
In order to investigate this, we choose
e = 0.5, τ = 0.1,
and perform the test using the same initial data (44). The results are shown in Figs. 5
and 6 where we indeed observe the same exponential convergence to equilibrium.
Granular Materials 27
Fig. 5 Test 1. Convergence to equilibrium. e = 0.5 with heat bath τ = 0.1, initial data is the
Maxwellian (44). Left: Semi-log plot of the relative entropy of f and f∞ = f (t = 55, v).
Right: numerical temperature (orange dots) with exact temperature (blue
√ line) as (42). Numerical
parameters: Nv2 = 64 × 64, Nρ = 32, Mĝ = 16, R = 20, L = 5(3 + 2) and t = 0.01
Fig. 6 Test 1. The equilibrium profile of e = 0.5 with heat bath τ = 0.1, initial data is the
Maxwellian (44). Numerical parameters: Nv2 = 64 × 64, Nρ = 32, Mĝ = 16, R = 20, L =
√
5(3 + 2) and t = 0.01
Fig. 7 Test 1. Convergence to equilibrium. e = 0.5 with heat bath τ = 0.1, initial data is the flat
function (46). Left: Semi-log plot of the relative entropy of f and f∞ = f (55, v). Right: numerical
√ line) as (42). Numerical parameters: Nv =
temperature (orange dots) with exact temperature (blue 2
Fig. 8 Test 1. The equilibrium profile of e = 0.5 with heat bath τ = 0.1, initial data is the√flat
function (46). Numerical parameters: Nv2 = 64 × 64, Nρ = 32, Mĝ = 16, R = 20, L = 5(3 + 2)
and t = 0.01.
√
with w0 = 2 6 such that ρ0 = 1, u0 = (0, 0) and T0 = 8. With restitution
coefficient e = 0.5 and heat bath τ = 0.1, the results are shown in Figs. 7 and 8.
Test 2: Investigation of Tail Behavior of the Equilibrium in 2D We now compare
the different tail behaviors of the equilibrium solutions for the Maxwell molecules
collision kernel (43) and for the hard spheres collision kernel
Fig. 9 Test 2. The equilibrium profile of e = 0.3, 0.5, 0.7 with heat bath τ = 0.1, initial data
is the flat function (46). Left: Semi-log plot of f∞ (v1 , 0.17) = f (t = 55, v1 , 0.17) for Maxwell
molecules. Right: Semi-log plot of f∞ (v1 , 0.17) = f (t = 55, v1 , 0.17) for hard spheres. The red
lines are the reference profiles. Numerical parameters: Nv2 = 128 × 128, Nρ = 32, Mĝ = 16,
√
R = 20, L = 5(3 + 2) and t = 0.01
in 2D. To see the tail we need higher resolution in velocity space so the velocity
mesh is increased to Nv2 = 128 × 128. We plot the profile in vx (v1 ) by choosing a
fixed vy (v2 ) for different es (0.3, 0.5 and 0.7). From Fig. 9, we see that the numerical
scheme generates overpopulated equilibrium tails: the Maxwell molecules case
behaves like
f (v, t = ∞) ∼ e−α|v| ,
f (v, t = ∞) ∼ e−α|v|
3/2
.
These results corresponds accurately to what was predicted theoretically in [20, 44]
(summarized in Theorem 1).
Test 3: 3D Hard Sphere The last test is more related to physics in real world, by
simulating the so-called “Haff’s cooling Law” (26). We use the following initial
data which is a Maxwellian with nonzero bulk velocity:
ρ0
e−(v−u0) ,
2
f0 (v) = 3/2
(47)
(2πT0 )
1
B= |g|.
4π
30 J. A. Carrillo et al.
In the first two tests, we consider a realistic set-up where the restitution coefficient
e depends on the distance of the relative velocity, i.e., e is a function of ρ = |g|
instead of a constant,
e0 − 1 e0 + 1
e(ρ) = tanh(ρ − 4) + ,
2 2
where 0 < e0 < 1. This allows to mimics the physically relevant visco-elastic hard
spheres case (see also (5)). We numerically evaluate the temperature and the results
for e0 = 0.8 and e0 = 0.2 are shown in Figs. 10 and 11. Compared with the cases
where e is constant, we observe a slight slower decay of the temperature.
Another parameter that may affect the decay rate of temperature is the variable
hard spheres exponent λ from (7). In Fig. 12 we show that, in the presence of heat
bath, for e = 0.5 but with λ = 1 (hard spheres), λ = 0.5 and λ = 0 (Maxwellian
Fig. 10 Test 3. Haff’s cooling law with Maxwellian initial data (47). Left: plot of inhomogeneous
e. Right: comparison of temperature between constant e = 0.8 (dash line) and e(|g| = ρ) =
−0.1 tanh(ρ − 4) + 0.9. Numerical parameters: Nv3 = 32 × 32 × 32, Nρ = 30, Mĝ = 32, R = 8,
√
L = 5(3 + 2) and t = 0.01
Fig. 11 Test 3. Haff’s cooling law with Maxwellian initial data (47). Left: plot of inhomogeneous
e. Right: comparison of temperature between constant e = 0.2 (dash line) and e(|g| = ρ) =
−0.4 tanh(ρ − 4) + 0.6. Numerical parameters: Nv3 = 32 × 32 × 32, Nρ = 30, Mĝ = 32, R = 8,
√
L = 5(3 + 2) and t = 0.01
Granular Materials 31
Fig. 12 Test 3. Haff’s cooling law with heat bath for different variable hard spheres exponent λs
√ (τ = 0.1). Numerical parameters: Nv = 32 × 32 ×
and Maxwellian initial data (47). The heat bath 3
Fig. 13 Test 3. Heated Haff’s cooling law with Maxwellian initial data (47). Left: regular plot.
Right: log-log plot. Numerical parameters: Nv3 = 32 × 32 × 32, Nρ = 30, Mĝ = 32, R = 8,
√
L = 5(3 + 2) and t = 0.01
molecules), the decay rate of temperature will decrease after certain time (notice the
slopes after t = 5).
Finally, with the heat bath τ = 0.1, we numerically evaluate the temperature up
to time tfinal = 20 for various values of restitution coefficients. The time evolution of
T is shown in Fig. 13 where one can observe the transition of decays from e = 0.5
to e = 0.95 (near elastic case).
4 Conclusion
ical part, we discussed the Cauchy theory and large time behavior of the equation,
as well as the hydrodynamic limits. In the numerical part, we discussed the Fourier
Galerkin spectral method to approximate the inelastic collision operator and the
fast version of the method. Leveraging on the fast algorithm and GPU parallelized
architecture, we are able to conduct long time simulations. In the last section, the
exponential convergence to equilibrium and tail behavior were numerically verified.
We also simulated a physically relevant case with visco-elastic hard spheres, where
the “near” Haff’s cooling law has been observed.
References
1. Almazán, L., Carrillo, J.A., Garzó, V., Salueña, C., Pöschel, T.: Supplementary online material
(2012). https://iopscience.iop.org/article/10.1088/1367-2630/15/4/043044
2. Almazán, L., Carrillo, J.A., Salueña, C., Garzó, V., Pöschel, T.: A numerical study of the
Navier–Stokes transport coefficients for two-dimensional granular hydrodynamics. New J.
Phys. 15(4), 043044 (2013)
3. Alonso, R., Cañizo, J.A., Gamba, I., Mouhot, C.: A new approach to the creation and
propagation of exponential moments in the Boltzmann equation. Commun. Partial. Differ. Equ.
38(1), 155–169 (2013)
4. Alonso, R., Lods, B.: Uniqueness and regularity of steady states of the Boltzmann equation
for viscoelastic hard-spheres driven by a thermal bath. Commun. Math. Sci. 11(4), 851–906
(2013)
5. Alonso, R., Lods, B.: Boltzmann model for viscoelastic particles: asymptotic behavior,
pointwise lower bounds and regularity. Commun. Math. Phys. 331 2, 545–591 (2014)
6. Alonso, R.J., Lods, B.: Free cooling and high-energy tails of granular gases with variable
restitution coefficient. SIAM J. Math. Anal. 42(6), 2499–2538 (2010)
7. Araki, S., Tremaine, S.: The dynamics of dense particle disks. Icarus 65(1), 83–109 (1986)
8. Benedetto, D., Caglioti, E., Carrillo, J.A., Pulvirenti, M.: A non-Maxwellian steady distribution
for one-dimensional granular media. J. Statist. Phys. 91(5/6), 979–990 (1998)
9. Benedetto, D., Pulvirenti, M.: On the one-dimensional Boltzmann equation for granular flows.
M2AN Math. Model. Numer. Anal. 35(5), 899–905 (2002)
10. Bird, G.A.: Molecular Gas Dynamics and the Direct Simulation of Gas Flows. Clarendon Press,
Oxford (1994)
11. Bisi, M., Cañizo, J.A., Lods, B.: Uniqueness in the weakly inelastic regime of the equilibrium
state to the Boltzmann equation driven by a particle bath. SIAM J. Math. Anal. 43(6), 2640–
2674 (2011)
12. Bisi, M., Carrillo, J.A., Toscani, G.: Contractive metrics for a Boltzmann equation for granular
gases: diffusive equilibria. J. Stat. Phys. 118(1–2), 301–331 (2005)
13. Bisi, M., Carrillo, J.A., Toscani, G.: Decay rates in probability metrics towards homogeneous
cooling states for the inelastic Maxwell model. J. Stat. Phys. 124(2–4), 625–653 (2006)
Granular Materials 33
14. Bizon, C., Shattuck, M., Swift, J.B., Swinney, H.: Transport coefficients for granular media
from molecular dynamics simulations. Phys. Rev. E 60, 4340 (1999)
15. Bobylev, A.V., Carrillo, J.A., Gamba, I.: On some properties of kinetic and hydrodynamic
equations for inelastic interactions. J. Statist. Phys. 98(3), 743–773 (2000)
16. Bobylev, A.V., Carrillo, J.A., Gamba, I.: Erratum on: “On some properties of kinetic and
hydrodynamic equations for inelastic interactions”. J. Statist. Phys. 103(5–6), 1137–1138
(2001)
17. Bobylev, A.V., Cercignani, C.: Moment equations for a granular material in a thermal bath. J.
Statist. Phys. 1063–4, 547–567 (2002)
18. Bobylev, A.V., Cercignani, C., Gamba, I.: Generalized kinetic Maxwell type models of granular
gases. In: Mathematical Models of Granular Matter. Lecture Notes in Math., vol. 1937, pp. 23–
57. Springer, Berlin (2008)
19. Bobylev, A.V., Cercignani, C., Toscani, G.: Proof of an asymptotic property of self-similar
solutions of the Boltzmann equation for granular materials. J. Statist. Phys. 111(1–2), 403–417
(2003)
20. Bobylev, A.V., Gamba, I., Panferov, V.: Moment inequalities and high-energy tails for
Boltzmann equations with inelastic interactions. J. Stat. Phys. 116(5), 1651–1682 (2004)
21. Bolley, F., Carrillo, J.A.: Tanaka theorem for inelastic Maxwell models. Commun. Math. Phys.
276(2), 287–314 (2007)
22. Bolley, F., Gentil, I., Guillin, A.: Uniform convergence to equilibrium for granular media. Arch.
Ration. Mech. Anal. 208(2), 429–445 (2013)
23. Bony, J.-M.: Solutions globales bornées pour les modèles discrets de l’équation de Boltzmann,
en dimension 1 d’espace. In: Journées “Équations aux derivées partielles” (Saint Jean de
Monts, 1987). École Polytechnique, Palaiseau, 1987. Exp. No. XVI, 10 pp
24. Bouchut, F., Desvillettes, L.: A proof of the smoothing properties of the positive part of
Boltzmann’s kernel. Rev. Mat. Iberoam. 14(1), 47–61 (1998)
25. Bouchut, F., James, F.: Duality solutions for pressureless gases, monotone scalar conservation
laws, and uniqueness. Commun. Partial Differ. Equ. 24(11–12), 2173–2189 (1999)
26. Boudin, L.: A solution with bounded expansion rate to the model of viscous pressureless gases.
SIAM J. Math. Anal. 32(1), 172–193 (2000)
27. Bougie, J., Kreft, J., Swift, J.B., Swinney, H.: Onset of patterns in an oscillated granular layer:
continuum and molecular dynamics simulations. Phys. Rev. E 71, 021301 (2005)
28. Bougie, J., Moon, S.J., Swift, J., Swinney, H.: Shocks in vertically oscillated granular layers.
Phys. Rev. E 66, 051301 (2002)
29. Brenier, Y., Grenier, E.: Sticky particles and scalar conservation laws. SIAM J. Numer. Anal.
35(6), 2317–2328 (1998) (electronic)
30. Brey, J.J., Dufty, J.W., Santos, A.: Kinetic models for granular flow. J. Stat. Phys. 97, 281–322
(1999)
31. Brilliantov, N.V., Pöschel, T.: Kinetic Theory of Granular Gases. Oxford University Press,
Oxford (2004)
32. Brilliantov, N.V., Salueña, C., Schwager, T., Pöschel, T.: Transient structures in a granular gas.
Phys. Rev. Lett. 93, 134301 (2004)
33. Carlen, E., Chow, S.-N., Grigo, A.: Dynamics and hydrodynamic limits of the inelastic
Boltzmann equation. Nonlinearity 23(8), 1807–1849 (2010)
34. Carlen, E.A., Carrillo, J.A., Carvalho, M.C.: Strong convergence towards homogeneous
cooling states for dissipative Maxwell models. Ann. Inst. H. Poincaré Anal. Non Linéaire
26(5), 1675–1700 (2009)
35. Carrillo, J.A., McCann, R.J., Villani, C.: Kinetic equilibration rates for granular media and
related equations: entropy dissipation and mass transportation estimates. Rev. Mat. Iberoam.
19(3), 971–1018 (2004)
36. Carrillo, J.A., Poëschel, T., Salueña, C.: Granular hydrodynamics and pattern formation in
vertically oscillated granular disk layers. J. Fluid Mech. 597, 119–144 (2008)
34 J. A. Carrillo et al.
37. Carrillo, J.A., Salueña, C.: Modelling of shock waves and clustering in hydrodynamic
simulations of granular gases. In: Modelling and Numerics of Kinetic Dissipative Systems,
pp. 163–176. Nova Sci. Publ., Hauppauge (2006)
38. Carrillo, J.A., Toscani, G.: Contractive probability metrics and asymptotic behavior of
dissipative kinetic equations. Riv. Mat. Univ. Parma (7) 6, 75–198 (2007)
39. Cercignani, C., Illner, R., Pulvirenti, M.: The Mathematical Theory of Dilute Gases. Applied
Mathematical Sciences, vol. 106. Springer, New York (1994)
40. Daerr, A.: Dynamique des Avalanches. PhD thesis, Université Denis Diderot Paris §7, 2000
41. Dufty, J.W.: Nonequilibrium Statistical Mechanics and Hydrodynamics for a Granular Fluid.
In: Cichocki, E., Napiorkowski, M., Piasekcki, J. (eds.) Second Warsaw School on Statistical
Physics, no. June 2007, p. 64. Warsaw University Press, Warsaw (2008)
42. Weinan, E., Rykov, Y.G., Sinai, Y.G.: Generalized variational principles, global weak solutions
and behavior with random initial data for systems of conservation laws arising in adhesion
particle dynamics. Commun. Math. Phys. 177(2), 349–380 (1996)
43. Ellis, R.S., Pinsky, M.A.: The first and second fluid approximations to the linearized Boltzmann
equation. J. Math. Pures Appl. 54(9), 125–156 (1975)
44. Ernst, M.H., Brito, R.: Driven inelastic Maxwell models with high energy tails. Phys. Rev. E
65, 040301 (2002)
45. Esteban, M., Perthame, B.: On the modified Enskog equation for elastic and inelastic collisions.
Models with spin. Ann. Inst. H. Poincaré Anal. Non Linéaire 8(3–4), 289–308 (1991)
46. Faraday, M.: On a peculiar class of acoustical figure and on certain forms assumed by groups
of particles upon vibrating elastic surfaces. Philos. Trans. R. Soc. Lond. 121, 299 (1831)
47. Fenichel, N.: Geometric singular perturbation theory for ordinary differential equations. J.
Differ. Equ. 31(1), 53–98 (1979)
48. Filbet, F., Pareschi, L., Toscani, G.: Accurate numerical methods for the collisional motion of
(heated) granular flows. J. Comput. Phys. 202(1), 216–235 (2005)
49. Filbet, F., Rey, T.: A rescaling velocity method for dissipative kinetic equations. Applications
to granular media. J. Comput. Phys. 248, 177–199 (2013)
50. Gallagher, I., Saint-Raymond, L., Texier, B.: From Newton to Boltzmann: Hard Spheres and
Short-Range Potentials. European Mathematical Society, Zürich (2014)
51. Gamba, I., Panferov, V., Villani, C.: On the Boltzmann equation for diffusively excited granular
media. Commun. Math. Phys. 246(3), 503–541 (2004)
52. Gamba, I.M., Tharkabhushanam, S.H.: Spectral-Lagrangian methods for collisional models of
non-equilibrium statistical states. J. Comput. Phys. 228, 2012–2036 (2009)
53. García de Soria, M.I., Maynar, P., Mischler, S., Mouhot, C., Rey, T., Trizac, E.: Towards an
H-theorem for granular gases. J. Stat. Mech: Theory Exp. 2015(11), P11009 (2015)
54. Garzó, V.: Granular Gaseous Flows: A Kinetic Theory Approach to Granular Gaseous Flows.
Springer, Berlin (2019)
55. Goldhirsch, I.: Scales and kinetics of granular flows. Chaos 9(3), 659–672 (1999)
56. Goldhirsch, I.: Probing the boundaries of hydrodynamics. In: Pöschel, T., Luding, S. (eds.)
Granular Gases. Lecture Notes in Physics, vol. 564, pp. 79–99. Springer, Berlin (2001)
57. Goldhirsch, I.: Rapid granular flows. Ann. Rev. Fluid Mech. 35, 267 (2003)
58. Goldhirsch, I., Zanetti, G.: Clustering instability in dissipative gases. Phys. Rev. Lett. 70(11),
1619–1622 (1993)
59. Goldshtein, A., Shapiro, M.: Mechanics of collisional motion of granular materials. Part 1:
general hydrodynamic equations. J. Fluid Mech. 282, 75 (1995)
60. Haff, P.: Grain flow as a fluid-mechanical phenomenon. J. Fluid Mech. 134, 401–30 (1983)
61. Hill, S.A., Mazenko, G.F.: Granular clustering in a hydrodynamic simulation. Phys. Rev. E 67,
061302 (2003)
62. Hu, J., Ma, Z.: A fast spectral method for the inelastic Boltzmann collision operator and
application to heated granular gases. J. Comput. Phys. 385, 119–134 (2019)
63. Huang, F., Wang, Z.: Well posedness for pressureless flow. Commun. Math. Phys. 222(1),
117–146 (2001)
Granular Materials 35
64. Jabin, P.-E., Rey, T.: Hydrodynamic limit of granular gases to pressureless Euler in dimension
1. Q. Appl. Math. 75, 155–179 (2017)
65. Jenkins, J., Richman, M.W.: Grad’s 13-moment system for a dense gas of inelastic spheres.
Arch. Ration. Mech. Anal. 87, 355 (1985)
66. Jenkins, J., Richman, M.W.: Kinetic theory for plane flows of a dense gas of identical, rough,
inelastic, circular disks. Phys. Fluids 28, 3485 (1985)
67. Johnson, C.G., Gray, J. M.N.T.: Granular jets and hydraulic jumps on an inclined plane. J.
Fluid Mech. 675, 87–116 (2011)
68. Kang, M.-J., Kim, J.: Propagation of the mono-kinetic solution in the Cucker–Smale-type
kinetic equations (2019). arXiv preprint 1909.07525
69. Kang, M.-J., Vasseur, A.F.: Asymptotic analysis of Vlasov-type equations under strong local
alignment regime. Math. Models Methods Appl. Sci. 25(11), 2153–2173 (2015)
70. Kawai, T., Shida, K.: An inelastic collision model for the evolution of “planetary rings”. J.
Phys. Soc. Jpn 59(1), 381–388 (1990)
71. Li, H., Toscani, G.: Long-time asymptotics of kinetic models of granular flows. Arch. Ration.
Mech. Anal. 172(3), 407–428 (2004)
72. McNamara, S., Young, W.R.: Inelastic collapse in two dimensions. Phys. Rev. E 50, R28–R31
(1993)
73. Melo, F., Umbanhowar, P., Swinney, H.L.: Hexagons, kinks, and disorder in oscillated granular
layers. Phys. Rev. Lett. 75, 3838–3841 (1995)
74. Mischler, S., Mouhot, C.: Cooling process for inelastic Boltzmann equations for hard spheres,
Part II: Self-similar solutions and tail behavior. J. Statist. Phys. 124(2), 703–746 (2006)
75. Mischler, S., Mouhot, C.: Stability, convergence to self-similarity and elastic limit for the
Boltzmann equation for inelastic hard spheres. Commun. Math. Phys. 288(2), 431–502 (2009)
76. Mischler, S., Mouhot, C.: Stability, convergence to the steady state and elastic limit for the
Boltzmann equation for diffusively excited granular media. Discrete Contin. Dyn. Syst. 24(1),
159–185 (2009)
77. Mischler, S., Mouhot, C., Ricard, M.: Cooling process for inelastic Boltzmann equations for
hard spheres, Part I: the Cauchy problem. J. Statist. Phys. 124(2), 655–702 (2006)
78. Mischler, S., Mouhot, C., Wennberg, B.: A new approach to quantitative propagation of chaos
for drift, diffusion and jump processes. Probab. Theory Relat. Fields 161(1–2), 1–59 (2015)
79. Naldi, G., Pareschi, L., Toscani, G.: Spectral methods for one-dimensional kinetic models of
granular flows and numerical quasi elastic limit. ESAIM: Math. Model. Numer. Anal. 37, 73–
90 (2003)
80. Oleinik, O.: On Cauchy’s problem for nonlinear equations in a class of discontinuous functions.
Doklady Akad. Nauk SSSR (N.S.) 95, 451–454 (1954)
81. Pöschel, T., Brilliantov, N., Schwager, T.: Transient clusters in granular gases. J. Phys.
Condens. Matter 17, 2705–2713 (2005)
82. Rericha, E., Bizon, C., Shattuck, M., Swinney, H.: Shocks in supersonic sand. Phys. Rev. Lett.
88, 1 (2002)
83. Rey, T.: Blow up analysis for anomalous granular gases. SIAM J. Math. Anal. 44(3), 1544–
1561 (2012)
84. Rey, T.: A spectral study of the linearized Boltzmann equation for diffusively excited granular
media (2013). Preprint arXiv 1310.7234
85. Toscani, G.: One-dimensional kinetic models of granular flows. M2AN Math. Model. Numer.
Anal. 34(6), 1277–1291 (2000)
86. Toscani, G.: Kinetic and hydrodynamic models of nearly elastic granular flows. Monatsh.
Math. 142(1), 179–192 (2004)
87. Tristani, I.: Boltzmann equation for granular media with thermal forces in a weakly inhomo-
geneous setting. J. Funct. Anal. 270(5), 1922–1970 (2016)
88. Tsimring, L.S., Aranson, I.S.: Localized and cellular patterns in a vibrated granular layer. Phys.
Rev. Lett. 79, 213–216 (1997)
89. Umbanhowar, P.B., Melo, F., Swinney, H.L.: Localized excitations in a vertically vibrated
granular layer. Nature 382, 793–796 (1996)
36 J. A. Carrillo et al.
90. Umbanhowar, P.B., Melo, F., Swinney, H.L.: Periodic, aperiodic, and transient patterns in
vibrated granular layers. Phys. A 249, 1–9 (1998)
91. Villani, C.: Mathematics of granular materials. J. Stat. Phys. 124(2), 781–822 (2006)
92. Villani, C.: Hypocoercivity. Mem. Am. Math. Soc. 202, 184 (2009)
93. Womersley, R.S.: Efficient spherical designs with good geometric properties. In: Contemporary
Computational Mathematics—a Celebration of the 80th Birthday of Ian Sloan, pp. 1243–1285.
Springer, Cham (2018)
94. Wu, Z.: L1 and BV-type stability of the inelastic Boltzmann equation near vacuum. Continuum
Mech. Thermodyn. 22(3), 239–249 (2009)
95. Wu, Z.: On the inelastic Enskog equation near vacuum. J. Math. Phys. 51(3), 033508 (2010)
Asymptotic Methods for Kinetic
and Hyperbolic Evolution Equations
on Networks
1 Introduction
There have been many attempts to define coupling conditions for macroscopic
partial differential equations on networks including, for example, drift-diffusion
equations, scalar hyperbolic equations, or hyperbolic systems like the wave equation
or Euler type models, see for example [4, 5, 5, 9, 10, 14, 16, 18–22, 26, 29, 33, 35].
In [18, 33] coupling conditions for scalar hyperbolic equations on networks are
discussed and investigated. [26, 46] treat the wave equation and general nonlinear
hyperbolic systems are considered in [4, 5, 9, 14, 19]. Nevertheless, for hyperbolic
systems on networks there are still many unsolved problems, like finding suitable
coupling conditions without restricting to subsonic situations.
On the other hand, coupling conditions for kinetic equations on networks have
been discussed in a much smaller number of publications, see [11, 27, 34]. In [11]
a first attempt to derive a coupling condition for a macroscopic equation from the
underlying kinetic model has been presented for the case of a kinetic equations for
chemotaxis.
Here, a more general and more accurate procedure to derive coupling conditions
for macroscopic equations from the underlying kinetic ones is reviewed, see [12]
and in [13] for more details. We consider linear and nonlinear systems and use
an asymptotic analysis of the situation near the nodes. The procedure has been
motivated by the classical procedure to find kinetic slip boundary conditions for
macroscopic equations. They are derived from the underlying kinetic equations via
an analysis of the kinetic layer. We refer to [7, 8, 30, 45] for a kinetic layer analysis
or [39, 47–49] for the case of hyperbolic relaxation systems.
To explain the procedure in more detail, we consider a scaled kinetic problem
in one spacial dimension with spatial and velocity variables x ∈ R, v ∈ R and the
distribution function f solving
1
∂t f + v∂x f = Q(f ) (1)
with a scaling parameter . One obtains for → 0 associated macroscopic equations
for the moments of f as, for example density and flux
∞ ∞
ρ= f (v)dv , q= vf (v)dv .
−∞ −∞
Considering additionally boundary conditions on an interval [0, b] one uses for the
kinetic model typically
For the macroscopic equations boundary conditions are usually given via conditions
on the ingoing characteristic variables [6, 44].
If such equations are considered on a network, it is sufficient to study a single
coupling point or node, where coupling conditions are required. We consider a node
connecting n edges, which are, for simplicity, oriented away from the node. Each
edge i = 1, . . . , n is parametrized by the interval [0, bi ] and the kinetic quantities
are denoted by f i . A typical choice of coupling conditions for the kinetic problem
is given by
n
f i (0, v) = cij f j (0, −v), v > 0 , (2)
j =1
n
compare [11]. The total mass in the system is conserved, if i=1 cij = 1 holds.
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 39
where the notation A is used for the so called Albedo operator of the half space
problem.
On the network, we proceed as follows. Starting from the kinetic coupling
conditions (2) we determine the coupling conditions for the macroscopic equations
in the following way. We use the kinetic coupling conditions to obtain conditions on
the in- and outgoing solutions of the half space problems on the different arcs. That
means
n
ϕ i (0, v) = cij ϕ j (0, −v), v > 0
j =1
or, if we denote the ingoing function of the half-space problem on arc i by k i (v), v >
0 and the outgoing solution by Ai [k i ](v), v < 0
n
k i (v) = cij Aj [k j ](−v), v > 0 . (4)
j =1
conditions have been derived for the linear wave equation from an underlying linear
kinetic model in [12] and for a simple nonlinear situation in [13].
The survey is organized in the following way. We start with the linear case.
In Sect. 2.1 we discuss the kinetic and macroscopic equations and the boundary
and coupling conditions for the linear case. In Sect. 2.2 kinetic boundary layers
are discussed, as well as an asymptotic analysis of the kinetic equations near the
nodes. This leads to an abstract formulation of the coupling conditions for the
macroscopic equations at the nodes based on a fix-point problem involving kinetic
half-space equations. A refined method to determine the solution of the half space
problems is derived, compared to previous approximate solution methods for half-
space problems and applied to the problem of finding accurate coupling conditions
for the macroscopic equations. Moreover, the macroscopic equations on the network
with the different coupling conditions are numerically compared to each other and
to the full solutions of the kinetic equations on the network in Sect. 2.5.
Then, the nonlinear case is investigated for an equation with a scalar limit prob-
lem. In Sect. 3.1 we state a relaxation model and the associated scalar conservation
law. In Sect. 3.2 kinetic boundary layers are discussed, as well as the combination of
these layer solutions with suitable Riemann solvers. This leads to classical boundary
conditions for the Burgers problem depending on the kinetic boundary condition.
In the following Sect. 3.3 a short outline is given, how coupling conditions for
the scalar hyperbolic problem are derived from the kinetic coupling conditions.
For the case of a node with three edges we state explicit coupling conditions for
the macroscopic equation based on the kinetic coupling conditions. Finally, again,
the solution of the nonlinear macroscopic equation on the network is numerically
compared to the full solution of the kinetic equation in Sect. 3.4.
An outline of further work is included at the end of the paper.
2.1 Equations
We consider a linear kinetic BGK model in 1D with two collision invariants and
x ∈ R, v ∈ [−1, 1]
1 1 v
∂t f + v∂x f = Q(f ) = − f − (ρ + 2 q) (5)
2 a
with a 2 = 1
3 and
∞ ∞
ρ= f (v)dv, q= vf (v)dv .
−∞ −∞
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 41
∂t ρ + ∂x q = 0
(6)
∂t q + ∂x (a 2 ρ) = 0 .
For (6) the boundary conditions are given in characteristic variables [44]. The
corresponding Riemann Invariants are
r1/2 = q ∓ aρ . (7)
As boundary data the value of q + aρ at the left boundary and q − aρ at the right
boundary are prescribed.
If these equations are considered on a network, it is sufficient to study a single
coupling point. At each node coupling conditions are required. In the following we
consider a node connecting n edges, which are oriented away from the node, as
in Fig. 1. Each edge i is parametrized by the interval [0, bi ] and the kinetic and
macroscopic quantities are denoted by f i and ρ i , q i respectively. A natural choice
of coupling conditions for the kinetic problem is given by
n
f i (0, v) = cij f j (0, −v), v > 0 , (8)
j =1
n
cij = 1 (9)
i=1
f + = Cf − , v > 0 ,
where f + = (f 1 (0, v), . . . , f n (0, v))T and f − = (f 1 (0, −v), . . . , f n (0, −v))T
and C is a n × n matrix. The coupling conditions for the macroscopic quantities are
2
42 R. Borsche and A. Klar
1 1
∂t f + v∂x f = Q(f ) .
This yields to first order in the following stationary kinetic half space problem for
x ∈ [0, ∞]
1 v
v∂x ϕ = − ϕ − ρ + 2q , (10)
2 a
where ρ and q are here the zeroth and first moments of ϕ. At x = 0 the boundary
conditions for the half space problem are
The boundary condition for (6) is obtained by determining r2 from the asymptotic
solution of the half space problem setting r2 = q∞ + aρ∞ . The values ρ∞ and q∞
are the macroscopic quantities associated to the solution of the half-space problem
at infinity, which has the form
1 v
ϕ(∞, v) = ρ∞ + 2 q∞ .
2 a
The solution of the half space problem is also used to determine the outgoing
distribution
Coupling conditions for the macroscopic equations on the network are now found
by the following procedure.
We start from the kinetic coupling conditions
n
f i (0, v) = cij f j (0, −v), v > 0.
j =1
n
ϕ (0, v) =
i
cij ϕ j (0, −v), v > 0
j =1
or, if we denote the ingoing function of the half-space problem on arc i by k i (v), v >
0 and the outgoing solution by Ai [k i ](v), v < 0
n
k i (v) = cij Aj [k j ](−v), v > 0 . (11)
j =1
with r1i = q i − aρ i . The coupling conditions for the wave equation are conditions
on the outgoing characteristic variables at x = 0. We define
r2i (0) = q∞
i
[k i ] + aρ∞
i
[k i ] .
44 R. Borsche and A. Klar
One obtains
1 1
ρ − = a − − b− , ρ + = a + + b+ ,
2 2
1 1 1 + 1 +
q − = − a − + b− , q+ = a + b
2 3 2 3
and
1 1 0 1
v 2 f (v)dv = − ρ + + q + , v 2 f (v)dv = − ρ − − q − .
0 6 −1 6
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 45
Finally, integrating the kinetic equation with respect to the corresponding half-
spaces, we get the half-moment approximation of the kinetic equation as
⎧
⎪ 1 ρ 3q
⎪ +
⎪ ∂t ρ + ∂x q
⎪
+ =− ρ+ − +
⎪
⎪ 2 4
⎪
⎪
⎪
⎪ + 1 + + 1 + ρ q
⎪ ∂t q + ∂x − ρ + q
⎨ =− q − +
6 4 2
(13)
⎪
⎪ ∂ ρ− + ∂ q − 1 − ρ 3q
⎪
⎪ =− ρ − −
⎪
⎪
t x
2 4
⎪
⎪
⎪
⎪
⎪ 1 1 − ρ q
⎩ ∂t q − + ∂x − ρ − − q − =− q − − + .
6 4 2
ρ = ρ+ + ρ− , ρ̂ = ρ + − ρ − ,
q = q+ + q− , q̂ = q + − q − ,
Obviously, the half-moment model has again the wave equation (6) as macroscopic
limit as goes to 0.
Rescaling the spatial variable in the half-moment problem with and rewriting
the equations in terms of the even-odd variables, one obtains to zeroth order the
following half-space problem for x ∈ R+
⎧
⎪
⎪ ∂x q + +
=− ρ −
ρ
+
3q
⎪
⎪
⎪
⎪ 2 4
⎪
⎪ ρ
⎪
⎪ 1 + + q
⎪
⎨ x
∂ − ρ + q = − q+ − +
6 4 2
⎪
⎪ ρ 3q
⎪
⎪ ∂x q − −
=− ρ − −
⎪
⎪ 2 4
⎪
⎪ ρ
⎪
⎪
⎪ 1 q
⎩ ∂x − ρ − − q − = − q− − − +
6 4 2
46 R. Borsche and A. Klar
or
⎧
⎪
⎪ ∂x q =0
⎪
⎪
⎪
⎪
⎪
⎪ 1
⎪
⎪ ∂ − ρ + q̂ =0
⎨ x
6
(14)
⎪ 3
⎪
⎪ ∂x q̂ = − ρ̂ − q
⎪
⎪ 2
⎪
⎪
⎪
⎪
⎪ 1 ρ
⎩ ∂x − ρ̂ + q = − q̂ − .
6 2
We have to provide boundary conditions for ρ + (0) and q + (0), as well as a condition
at x = ∞
q∞ − aρ∞ = r1 (0) = C.
Then, the half space problem can be solved explicitly. We determine a solution up
to 3 constants which will be fixed with the above 3 conditions. First, we observe,
that we have 2 invariants
q = C1
ρ
− + q̂ = C2 (15)
6
with constants C1 and C2 . From the last equation in (14) we can deduce that at
x=∞
ρ∞
q̂∞ = .
2
Combining this with (15) gives ρ∞ = 3C2 or q̂∞ = 3C2 2 . From the third equation
of (14) we obtain ρ̂∞ = 3q 3C1
2 = 2 . This simplifies (14) to
⎧
⎪
⎪ q = C1
⎪
⎪
⎪
⎪ ρ = 6q̂− 6C2
⎨ 3
∂x q̂ = − ρ̂ − C1
⎪
⎪ 2
⎪
⎪
⎪
⎪ 1
⎩ ∂x − ρ̂ = − −2q̂ + 3C2 .
6
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 47
2x 2x 3C1
ρ̂ = γ exp(− ) + γ̂ exp( ) +
a a 2
a 2x a 2x 3C2
q̂ = γ exp(− ) − γ̂ exp( ) + .
2 a 2 a 2
Since we are looking only for bounded solutions we are left with
2x 3C1
ρ̂ = γ exp(− )+
a 2
a 2x 3C2
q̂ = γ exp(− ) +
2 a 2
q = C1
2x
ρ = 3aγ exp(− ) + 3C2 .
a
The three parameters are fixed with the 3 conditions mentioned above. At x = 0
inflow data is given
1
q(0) + q̂(0) = q+ (0)
2
1
ρ(0) + ρ̂(0) = ρ+ (0)
2
and the Riemann Invariant at x = ∞ gives
q∞ − aρ∞ = C . (16)
Together with the condition at infinity (16), this determines the asymptotic values
q∞ , ρ∞ and γ . The outgoing quantities ρ− (0), q− (0) are then determined by
q∞ ρ∞ a
− − γ = q− (0)
2 4 4
3q∞ 1 3a − 1
− + ρ∞ + γ = ρ− (0) .
4 2 2
1 q∞ ρ∞
q+ (0) = q+ (∞) = (q∞ + q̂∞ ) = +
2 2 4
and the condition at infinity (16). The outgoing quantities are
ρ∞ q∞
q− (0) = q− (∞) = − +
4 2
ρ∞ 3q∞
ρ− (0) = ρ− (∞) = − .
2 4
To estimate the accuracy of our method, we consider the classical problem of
determining the so-called extrapolation length [31, 40, 43].
For x ∈ R+ , v ∈ [−1, 1] we consider the half space equation
ρ 3
v∂x f = − f − ( + vq)
2 2
"1
with −1 vf dv = q = 0 and f (0, v) = v, v > 0. Thus we obtain
ρ
v∂x f = − f − .
2
ρ∞ a 1
+ γ = q+ (0) =
4 4 3
ρ∞ 3a + 1 1
+( )γ = ρ+ (0) = .
2 2 2
This leads to
9a + 4
ρ∞ =
6a + 3
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 49
and with a 2 = 1
3 we obtain
√
3 3+4
ρ∞ = √ .
2 3+3
n
j
i
q+ (0) = − cij q− (0) . (18)
j =1
j
n−2 i ρi a n−2 j ρ∞ a
q∞ + ∞ + γ i = q∞ + + γj .
2n 4 4 2n 4 4
which gives the invariance of
2(n − 2)
ρ∞ + q∞ + aγ . (20)
n
50 R. Borsche and A. Klar
Further, integrating the kinetic coupling conditions (8) with respect to the positive
and negative half moments, we obtain
n
j
i
ρ+ (0) = cij ρ− (0) .
j =1
n
γi = 0 . (21)
i=1
ρ + Cq .
∂t e + ∂x (ρq) = 0 .
Along one edge this entropy is conserved, but the total entropy in the network can
change according to the entropy-fluxes at the nodes. Note that for all above models
with the coupling invariant ρ + Cq = C̃ and C > 0 the total entropy decays, since
n n
n
n
n
ρq = C̃ − Cq q = C̃ q −C q 2 = −C q2 < 0 .
i=1 i=1 i=1 i=1 i=1
ρi = ρj i = j, i, j = 1, 2, 3
3
qi = 0 .
i=1
52 R. Borsche and A. Klar
3
1 1 i 2
(ρ ) + 2 (q ) dx .
i 2
2 a
i=1
1 0
0.9 −0.05
−0.1
0.8
ρ
0.7
0.6
Fig. 2 Kinetic equation, half-moment equation and wave equation with coupling conditions given
by the Maxwell, the half-moment and the full-moment approach and the assumption of equal
density at time T = 1
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 53
kinetic half-moment
Maxwell wave-half
0.688 −0.178
0.686
−0.18
0.684
ρ
q
−0.182
0.682
0.68 −0.184
0.678
Table 1 Total entropy and entropy loss at time T = 0.1 for different coupling conditions
Coupling conditions Total entropy Entropy loss
Wave equal density 7.2222139e − 01 −8.3353857e − 07
Wave full moment 7.1701785e − 01 −5.2043681e − 03
Wave Maxwell 7.1621400e − 01 −6.0082256e − 03
Wave half moment 7.1597274e − 01 −6.2494790e − 03
Half moment = 10−6 7.1586300e − 01 −6.3592218e − 03
Kinetic = 10−6 7.1574793e − 01 −6.4742970e − 03
Initially, the total entropy at t = 0 is equal to 0.722222. In this case we use a very
fine grid with 30000 spatial cells for all models and 400 cells in velocity space
for the kinetic equation. In Table 1 the value of the total entropy at time T = 0.1
is shown for the different coupling conditions together with the half moment and
the kinetic solution for comparison. One observes the very accurate approximation
given by the wave equation with half moment coupling conditions.
As a second example we consider a more complicated network, see Fig. 4, as, for
example, studied in [26] for the wave equation.
As initial conditions for the kinetic equation we choose f 1 (x, v) = 1,f 2 (x, v) =
6 and f (x, v) = 2 for j = 3, . . . , 7, which corresponds to macroscopic densities
5 j 1
kinetic half-moment
wave - Maxwell wave - full
wave - half wave - equal
1.55
1.5
1.4 1.5
ρ
ρ
1.3 1.45
1.2 1.4
1.1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x on edge 4 x on edge 4
Boundary conditions for the wave equation with full moment, Maxwell and half
moment conditions are derived as detailed above. In case of the equal density
conditions, we determine the ingoing characteristic using ρ = 1, q = 0 at the
E1 -boundary and ρ = 12 , q = 0 at the E7 -boundary.
In Fig. 5 the density ρ 4 on edge 4 is displayed at time t = 3 and t = 10. As
before, we observe a good agreement of the half moment coupling with the kinetic
and half moment model. Also the Maxwell approximation is relatively close to the
kinetic results. The states of the full moment coupling and the equal density coupling
deviate remarkably from the kinetic results.
3.1 Equations
vu − û û + vu
f1 = , f2 = .
2v 2v
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 55
∂t u + ∂x F (u) = 0 . (25)
∂t u + ∂x û = 0
1 (26)
∂t û + ∂x P (u, û) = − û − F (u) .
with
−v ≤ F (u) ≤ v . (28)
We have to prescribe f2 at the left boundary and f1 at the right boundary. For
the nonlinear hyperbolic limit problem, boundary conditions have been considered
in many works [1, 2, 6, 15, 42].
For boundary conditions and layers of hyperbolic problems with stiff relaxation
terms and for the derivation of conditions for the corresponding limit equations we
refer for example to [39, 47–49].
In order to determine the boundary conditions for the limit equations we combine
an analysis of the kinetic layer with the solution of a half-Riemann problem for the
limit equation, compare for example [3, 47]. A similar procedure will then be used
to find kinetic based coupling conditions for the Burgers equation on a network.
To proceed, we first state the layer equations. We consider the left boundary of
the domain located at x = 0. A rescaling of the spacial coordinate near the boundary
with x̃ = x gives the layer problem on [0, ∞) as
vu − F (u)
−v∂x̃ f1 = − f1
2v
(29)
F (u) + vu
v∂x̃ f2 = − f2 .
2v
∂x̃ û = 0
∂x̃ P (u, û) = F (u) − û ,
56 R. Borsche and A. Klar
∂x̃ û = 0
v 2 ∂x̃ u = F (u) − û
and therefore
û = C = const
(30)
a∂x̃ u = F (u) − C
with a = v 2 > 0.
Remark 3 For a right boundary we obtain the layer problem as
−a∂x̃ u = F (u) − C .
In a first step we determine the boundary conditions for the Burgers equation from
the kinetic boundary condition. We use the boundary layer equations and couple
them with half-Riemann solvers. First we discuss the solution in the boundary layer.
The boundary layer equation near a left boundary is given by
a∂x u = u2 − C .
√ √
For C √> 0 this problem has two fixpoints u = ± C, where C is instable
and − √C is a stable fixpoint. The domain of attraction of the stable fix-point is
(−∞, C).
The explicit solution is given by
√ √ √
u(x) = Ctanh(− C(x + C2 )/a) for |u(0)| < C
and
√ √ √
u(x) = Ccoth(− C(x + C2 )/a) for |u(0)| > C.
We determine C2 from
√ √ √
u(0) = Ctanh(− CC2 /a) for |u(0)| < C
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 57
and
√ √ √
u(0) = Ccoth(− CC2 /a) for |u(0)| > C.
√ √
One observes
√ that for u(0) < C the limit x → ∞ leads to u(x) → − C and for
u(0) > C the layer solution diverges at x = −C2 = √a arcoth( u(0)
√ ).
C C
For C = 0 we obtain
a a
u(x) = − =−
x + C2 x − u(0)
a
and convergence to 0 for u(0) < 0 and divergence for u(0) > 0. The solutions are
sketched in Fig. 6. √
In the following we use the notation (U ) for the unstable solution u(x) = C
and the notation (S) for the (partially) stable solutions. The asymptotic states as x →
∞ are denoted by uK . The layer solution for the right boundary can be discussed
analogously.
Since the layer solution can not cover the full range of possible states at a
boundary, we have to consider additionally a Riemann Problem for the Burgers
equation connecting the state in the domain with the layer. In particular, for the left
boundary we need to know, which asymptotic states uK from the kinetic layer can
be connected to a given right side state from the Burgers equation uB using only
waves with non-negative speeds. For the Burgers equation we have the following
cases:
RP1 uB ≥ 0 ⇒ uK ∈ [0, ∞), since there is either an arbitrary wave with positive
speed, if uK > 0, or a rarefaction wave starting at u = 0.
u
√ u
u(0) C
u(0)
u (0)
C =0
u=0 u =0
u(0) √ u (0)
− C u (0)
u (x )
u(0) u (0)
x=∞ x =0 x =∞
x=0
(b)
(a)
Fig. 6 Possible solutions to the layer equation. (a) Layer solutions for C > 0. (b) Layer solutions
for C = 0
58 R. Borsche and A. Klar
RP2 uB < 0 ⇒ uK ∈ {uB } ∪ (−uB , ∞), since there is either no wave or a shock
wave moving to the right.
Thus, for a given uB we can select uK only from the above subsets. For a boundary
on the right hand, we study the analogous cases.
Finally, to find the macroscopic boundary conditions at a left boundary from the
underlying kinetic problem, we combine the solution of the half space problem (30)
on [0, ∞] with asymptotic solution uK and of a Riemann Problem with left state
uK and right state uB .
The details of the procedure are as follows. We consider a domain x ≥ 0 and
determine macroscopic boundary conditions at x = 0 in the following way. For
the kinetic problem we prescribe f2 (x = 0). Moreover, the actual macroscopic
value at x = 0+ is denoted by uB . From these two values we have to determine
a (potentially new) boundary value for the macroscopic solution uK and a value
for f1 (0), the outgoing kinetic value. We consider different cases coupling stable
or unstable layer solutions, denoted by (U ) or (S), and Riemann problem solutions
RP 1 or RP 2.
Case 1, RP1-U The flow is ingoing with uB > 0 and f2 (0) > 0. The layer solution
is
√
u(x) = C >0.
u u u u
u(0) u(x) uK uB
uB
uK
u=0 u =0
u (x )
u (0)
x=0 x=∞ x =0 x =∞
(a) (b)
Fig. 7 Boundary layer and Riemann problem solution for positive uB . (a) uB > 0 and u(0) > 0.
(b) uB > 0 and u(0) < 0
f1 (0) = f2 (0) .
In this case u(0) and uB cannot be connected by an outgoing Burgers wave. The
kinetic layer solution takes care for a part (from u(0) to 0) of the full jump from
u(0) to uB , see Fig. 7b.
u2B −vuB
Case 3, RP2-S The flow is outgoing with uB < 0 and f2 (0) ≤ 2v . In this
case the value of the layer solution at infinity is given by uB . Thus,
√
u(∞) = − C = uB .
u2B √
u(0) = − + 2f2 (0) ≤ −uB = C .
v
Then, the layer solution is given
√ by the formulas in the last subsection and
converges to the stable fixpoint − C. Moreover,
u2B
f1 (0) = u(0) − f2 (0) = − + f2 (0) .
v
In this case u(0) and uB cannot be connected by an outgoing Burgers wave. The
kinetic layer solution handles the full jump, see Fig. 8a.
60 R. Borsche and A. Klar
u u u u
u (0) u (x ) uK
u(0)
u=0 u K uB u =0 uB
u(x)
x=0 x=∞ x =0 x =∞
(a) (b)
Fig. 8 Boundary layer and Riemann problem solution for negative uB . (a) uB < 0 and u(0) ≤
−uB . (b) uB < 0 and u(0) ≥ −uB
u2B −vuB
Case 4, RP2-U The flow is ingoing with uB < 0 and f2 (0) ≥ 2v . As in the
first case the layer solution is given by
√
u(x) = C.
We consider the kinetic and Burgers problems on a network. In the present case
the orientation of the edges is important. In the kinetic case we have to prescribe at
the node for each ingoing edge the quantity f1 and for the outgoing edges f2 . The
macroscopic coupling conditions are then derived from the kinetic conditions using
the above procedures.
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 61
We concentrate on the case of a node with 3 edges and 1 ingoing and 2 outgoing
edges. For other cases we refer to [13].
Arc i is oriented into the node and arc j and k are oriented out of the node
(Fig. 9).
We choose the symmetric kinetic coupling conditions similar to [11]
j
f1i = 12 (f1 + f1k ) (31)
j
f2 = 12 (f2i + f1k ) (32)
j
f2k = 12 (f2i + f1 ) . (33)
In these coupling conditions we assume that the value of v is identical on all edges.
If the velocities are not symmetric the fluxes have to be weighted such that the
conservation of mass is maintained. We reformulate the above conditions in terms
of u and û and obtain
1 j
vui − ûi = vu − ûj + vuk − ûk (34)
2
1
ûj + vuj = ûi + vui + vuk − ûk (35)
2
1
ûk + vuk = ûi + vui + vuj − ûj . (36)
2
Remark 5 In general linear mass conserving kinetic coupling conditions in the 1-2
case have the form
j
f1i = αf1 + (1 − β)f1k
j
f2 = (1 − γ )f2i + βf1k
j
f2k = γf2i + (1 − α)f1
we have eight possible combinations of stable and unstable layer solutions, which
are combined via the coupling conditions.
Step 2: Combine Kinetic Layer and Riemann Problems Assuming the states
j j
uiB , uB , ukB to be given, we have to determine the new states uiK , uK and ukK at
the node. We have to consider eight different configurations of 1- and 2-Riemann
problems. For each of them all possible combinations with stable or unstable layer
solutions have to be discussed.
Going through all possible cases one obtains the following cases for the
macroscopic coupling conditions, we refer for details to [13].
j
Case 1, RP1-1-1 uiB < 0, uB > 0, ukB > 0. Then Ci = Cj = Ck = 0 and
j
uiK = 0 uK = 0 ukK = 0 .
j
Case 2, RP1-1-2 uiB < 0, uB > 0, ukB < 0. Then Ci = Ck , Cj = 0 and
j
uiK = ukB < 0 uK = 0 ukK = ukB < 0 .
j
Case 3, RP1-2-1 uiB < 0, uB < 0, ukB > 0. Then Ci = Cj , Ck = 0 and
j j j
uiK = uB < 0 uK = uB < 0 ukK = 0 .
j Ci Ci
Case 4, RP2-1-1 uiB > 0, uB > 0, ukB > 0. Then Cj = 2 , Ck = 2 and
j 1 1
uiK = uiB > 0 uK = √ uiB > 0 ukK = √ uiB > 0 .
2 2
j
Case 5, RP1-2-2 uiB < 0, uB < 0, ukB < 0. Then Ci = Cj + Ck and
#
j j
uiK = − (u2B )2 + (ukB )2 < 0 uK = uB < 0 ukK = ukB < 0 .
j
Case 6, RP2-1-2 uiB > 0, uB > 0, ukB < 0. Then we have 3 subcases.
√
If uiB ≥ − 2ukB , then Cj = C2i , Ck = C2i and
j 1 1
uiK = uiB > 0 uK = √ uiB > 0 ukK = √ uiB > 0 .
2 2
√
If − 2ukB > uiB ≥ −ukB , then Cj = Ci − Ck and
#
j
uiK = uiB > 0 uK = (uiB )2 − (ukB )2 > 0 ukK = ukB < 0 .
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 63
j
uiK = ukB < 0 uK = 0 ukK = ukB < 0 .
j
Case 7, RP2-2-1 uiB > 0, uB < 0, ukB > 0. We have again 3 subcases.
√ j
If uiB ≥ − 2uB , then Cj = C2i , Ck = C2i and
j 1 1
uiK = uiB > 0 uK = √ uiB > 0 ukK = √ uiB > 0 .
2 2
√ j j
If − 2uB > uiB ≥ −uB then Ck = Ci − Cj and
#
j j j
uiK = uiB > 0 uK = uB < 0 ukK = (uiB )2 − (uB )2 > 0 .
j
If uiB < −uB , then Ci = Cj , Ck = 0 and
j j j
uiK = uB < 0 uK = uB < 0 ukK = 0 .
j
Case 8, RP2-2-2 uiB > 0, uB < 0, ukB < 0. We have 4 subcases.
√ j √
If uiB ≥ − 2uB , uiB ≥ − 2ukB then Cj = C2i , Ck = C2i and
j 1 1
uiK = uiB > 0 uK = √ uiB > 0 ukK = √ uiB > 0 .
2 2
√ j #
j
If − 2uB > uiB ≥ (uB )2 + (ukB )2 , then Ck = Ci − Cj and
#
j j j
uiK = uiB > 0 uK = uB < 0 ukK = (uiB )2 − (uB )2 > 0 .
√ #
j
If − 2ukB > uiB ≥ (uB )2 + (ukB )2 , then Cj = Ci − Ck and
#
j
uiK = uiB > 0 uK = (uiB )2 − (ukB )2 > 0 ukK = ukB < 0 .
#
j
If uiB ≤ (uB )2 + (ukB )2 , then Ci = Cj + Ck and
#
j j j
uiK = − (uB )2 + (ukB )2 < 0 uK = uB < 0 ukK = ukB > 0 .
j
This yields the desired expressions for the new states uiK , uK and ukK at the node
j
for given states uiB , uB , ukB .
64 R. Borsche and A. Klar
In this section the derived coupling conditions are investigated numerically. The
numerical solutions of the macroscopic equation (25) with F (u) = u2 are compared
to those obtained for the kinetic model (24).
As numerical scheme for the kinetic equations the Upwind method for the linear
advective part is combined with an implicit Euler scheme for the source term. The
solution of the macroscopic equation is approximated with a Godunov scheme.
For all computations 1000 cells are used per edge, i.e. a gridsize of Δx = 10−3 ,
as spacial resolution. The time steps Δt are chosen according to the respective
CFL conditions, Δt = Δx/v for the kinetic problem and Δt = Δx/(2|umax |)
for the macroscopic equation. The simulations are computed up to time T = 0.5
and the relaxation parameter is chosen as = 0.0005. The initial conditions are
formulated in macroscopic states, the remaining values in the kinetic model are
chosen according to the relaxed state, i.e. û = F (u). The kinetic speeds are chosen
in agreement with the subcharacteristic condition as v = 2. Larger values of v would
lead to similar results, as the coupling conditions are independent of v.
Here we consider a 1-2-junction. For a 2-1-junction we refer to [13]. The
coupling conditions are tested with six different Riemann problems at the junction.
The first edge is connected to the junction at x = 1, while the other two edges
are connected to the node at x = 0. Thus in the following figures waves move to the
left in the first edge but to the right in edge 2 and 3.
In Fig. 10 the solutions to the initial conditions (u10 , u20 , u30 )(x) = (−1, 0.75, 0.5)
and (u10 , u20 , u30 )(x) = (−1, 0.75, −0.5) are shown. These correspond to the
cases RP1-1-1 and RP1-1-2 respectively. In RP1-1-1 the initial states only have
−0.5 −0.5
u1
u1
−1 −1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
1
0.5
0.5
u2
u2
0
0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
1 0
0.5 −0.5
u3
u3
0 −1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x
Fig. 10 Left: RP1-1-1 with u1 = −1, u2 = 0.75, u3 = 0.5, Right: RP1-1-2 with u1 = −1,
u2 = 0.75, u3 = −0.5
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 65
characteristic speeds away from the junction and the coupling conditions enforce
zero states in all edges. This leads to three rarefaction waves.
On the right hand side RP1-1-2 is considered. Flow is entering from edge 3
but only exiting in edge 1. In edge 1 a rarefaction wave forms and moves to the
left. On the slower end of the rarefaction wave a bump in the kinetic solution
is present. As the initial states at t = 0 do not satisfy the coupling conditions,
this small disturbance arises due to the transition in the new state at the junction.
For smaller values of and when refining the spacial and the temporal grid, this
disturbance becomes narrower and more peaked. Such temporal layers due to the
initial conditions will also occur in other test cases. On edge 2 a boundary layer
connects the junction state and a rarefaction wave. The ingoing flow from edge 3
leads to a small boundary layer.
In Fig. 11 on the left, the flow from the first edge is split to the outgoing edges.
On the right hand side the flow from edge 2 and 3 is merged into edge 1. Two
layers connect the backward going flows with the junction states. In the first edge a
small rarefaction wave travels to the left, followed by a small temporal layer. This
corresponds to the case1-2-2.
Figure 12 shows the last two test-cases for this junction. On the left in the case
RP2-1-2 the flow enters from the first and the third edge and exits into the second
one, where a rarefaction wave moves to the right. The test on the right considers the
case RP2-2-2.
In all tests of the 1-2 junction the kinetic and the macroscopic solutions are very
close. Especially the states at the junction are correctly represented by the derived
coupling conditions. Since the value of is small, also the boundary layers in the
kinetic solution have a small spacial width.
0.5 −0.5
u1
u1
0 −1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
1 0
0.5 −0.5
u2
u2
0 −1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
1 0
0.5 −0.5
u3
u3
0 −1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x
Fig. 11 Left: RP2-1-1 with u1 = 1, u2 = 0.75, u3 = 0.5, Right: RP1-2-2 with u1 = −1,
u2 = −0.75, u3 = −0.5
66 R. Borsche and A. Klar
0.5
0.5
u1
u1
0
0 0.2 0.4 0.6 0.8 1
−0.5
1
−1
0.5
u2
u2
0 0.2 0.4 0.6 0.8 1 −0.8
−1
0 0 0.2 0.4 0.6 0.8 1
−0.4
u3
−0.6
−0.5
u3
−0.8
−1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x
Fig. 12 Left: RP2-1-2 with u1 = 0.6, u2 = 0.75, u3 = −0.5, Right: RP2-2-2 with u1 = 0.8,
u2 = −0.75, u3 = −0.5
4 Outlook
Coupling conditions for macroscopic equations on networks have been derived from
underlying kinetic models via a kinetic layer analysis at the nodes of the network.
We have treated simple model equations like linear BGK-type kinetic equations and
nonlinear relaxation approximations for scalar limit problems.
The approach presented here will be extended to more general situations in future
work. In particular, an investigation of linear models with more collision invariants
leading, in the limit, for example, to the full linearized Euler model is under way.
Moreover, discrete velocity relaxation systems leading in the limit to nonlinear
macroscopic systems of equations similar to the isothermal Euler equations are
currently investigated.
References
1. Andreianov, B., Sbihi, K.: Well-posedness of general boundary-value problems for scalar
conservation laws. Trans. Am. Math. Soc. 367(6), 3763–3806 (2015)
2. Andreianov, B.P., Coclite, G.M., Donadello, C.: Well-posedness for vanishing viscosity
solutions of scalar conservation laws on a network. Discrete Contin. Dyn. Syst. 37(11), 5913–
5942 (2017)
3. Aregba-Driollet, D., Milisic, V.: Kinetic approximation of a boundary value problem for
conservation laws. Numer. Math. 97, 595–633 (2004)
4. Banda, M., Herty, M., Klar, A.: Coupling conditions for gas networks governed by the
isothermal Euler equations. NHM 1(2), 295–314 (2006)
5. Banda, M., Herty, M., Klar, A. : Gas flow in pipeline networks. NHM 1(1), 41–56 (2006)
Asymptotic Methods for Kinetic and Hyperbolic Evolution Equations on Networks 67
6. Bardos, C., le Roux, A.Y., Nédélec, J.-C.: First order quasilinear equations with boundary
conditions. Commun. Partial Differ. Equ. 4(9), 1017–1034 (1979)
7. Bardos, C., Santos, R., Sentis, R.: Diffusion approximation and computation of the critical size.
Trans. Am. Math. Soc. 284(2), 617–649 (1984)
8. Bensoussan, A., Lions, J.L., Papanicolaou, G.C.: Boundary-layers and homogenization of
transport processes. J. Publ. RIMS Kyoto Univ. 15, 53–157 (1979)
9. Borsche, R., Colombo, R., Garavello, M.: On the coupling of systems of hyperbolic conserva-
tion laws with ordinary differential equations. Nonlinearity 23, 112749 (2010)
10. Borsche, R., Göttlich, S., Klar, A., Schillen, P.: The scalar Keller-Segel model on networks.
Math. Models Methods Appl. Sci. 24(2), 221–247 (2014)
11. Borsche, R., Kall, J., Klar, A., Pham, T.N.H.: Kinetic and related macroscopic models for
chemotaxis on networks. Math. Models Methods Appl. Sci. 26(6), 1219–1242 (2016)
12. Borsche, R., Klar, A.: Kinetic layers and coupling conditions for macroscopic equations on
networks. SIAM Sci. Comput. 40(3), 1784–1808 (2018)
13. Borsche, R., Klar, A.: Kinetic layers and coupling conditions for nonlinear scalar equations on
networks. Nonlinearity 31, 351 (2018)
14. Bretti, G., Natalini, R., Ribot, M.: A hyperbolic model of chemotaxis on a network: a numerical
study. ESAIM: M2AN 48(1), 231–258 (2014)
15. Bürger, R., Frid, H., Karlsen, K.H.: On the well-posedness of entropy solutions to conservation
laws with a zero-flux boundary condition. J. Math. Anal. Appl. 326(1), 108–120 (2007)
16. Camilli, F., Corrias, L.: Parabolic models for chemotaxis on weighted networks. Math. Pures
Appl. 108(4), 459–480 (2017)
17. Cercignani, C.: A variational principle for boundary value problems. J. Stat. Phys. 1(2), 297–
311 (1969)
18. Coclite, G.M., Garavello, M., Piccoli, B.: Traffic flow on a road network. SIAM J. Math. Anal.
36, 1862–1886 (2005)
19. Colombo, R., Herty, M., Sachers, V.: On 2 × 2 conservation laws at a junction. SIAM J. Math.
Anal. 40(2), 605–622 (2008)
20. Colombo, R.M., Garavello, M.: On the Cauchy problem for the p-system at a junction. SIAM
J. Math. Anal. 39, 1456–1471 (2008)
21. Colombo, R.M., Mauri, C.: Euler system for compressible fluids at a junction. J. Hyperbolic
Differ. Equ. 5(3), 547–568 (2008)
22. Corli, A., di Ruvo, L., Malaguti, L., Rosini, M.D.: Traveling waves for degenerate diffusive
equations on networks. NHM 12(3), 339–370 (2017)
23. Coron, F.: Computation of the asymptotic states for linear halfspace problems. TTSP 19(2), 89
(1990)
24. Coron, F., Golse, F., Sulem, C.: A classification of well-posed kinetic layer problems. CPAM
41, 409 (1988)
25. Dager, R., Zuazua, E.: Controllability of tree-shaped networks of vibrating strings. C. R. Acad.
Sci. Paris 332, 1087–1092 (2001)
26. Egger, H. Kugler, T.: Damped wave systems on networks: exponential stability and uniform
approximations (2016). https://arxiv.org/abs/1605.03066
27. Fermo, L., Tosin, A.: A fully-discrete-state kinetic theory approach to traffic flow on road
networks. Math. Models Methods Appl. Sci. 25(3), 423–461 (2015)
28. Garavello, M.: A review of conservation laws on networks. NHM 5(3), 565–581 (2010)
29. Garavello, M., Piccoli, B.: Traffic Flow on Networks. AIMS (2006)
30. Golse, F.: Analysis of the boundary layer equation in the kinetic theory of gases. Bull. Inst.
Math. Acad. Sin. 3(1), 211–242 (2008)
31. Golse, F., Klar, A.: Numerical method for computing asymptotic states and outgoing distribu-
tions for a kinetic linear half space problem. J. Stat. Phys. 80(5–6), 1033–1061 (1995)
32. Gugat, M., Herty, M., Klar, A., Leugering, G., Schleper, V.: Well–posedness of networked
hyperbolic systems of balance laws. In: International Series of Numerical Mathematics, vol.
160, 175–198. Springer, Berlin (2011)
68 R. Borsche and A. Klar
33. Herty, M., Klar, A., Piccoli, B.: Existence of solutions for supply chain models based on partial
differential equations. SIAM J. Math. Anal. 39(1), 160–173 (2007)
34. Herty, M., Moutari, S.: A macro-kinetic hybrid model for traffic flow on road networks.
Comput. Methods Appl. Math. 9(3), 238–252 (2009)
35. Leugering, G., Schmidt, E.J.P.G.: On the modelling and stabilization of flows in networks of
open canals. SIAM J. Control Optim. 41(1), 164–180 (2002)
36. Li, Q., Lu, J., Sun, W.: Half-space kinetic equations with general boundary conditions. Math.
Comput. 86, 1269–1301 (2017)
37. Li, Q., Lu, J., Sun, W.: A convergent method for linear half-space kinetic equations. ESAIM:
M2AN 51(5), 1583–1615 (2017)
38. Liu, T.P.: Hyperbolic conservation laws with relaxation. Commun. Math. Phys. 108, 153–175
(1987)
39. Liu, J.-G., Xin, Z.: Boundary-layer behavior in the fluid-dynamic limit for a nonlinear model
Boltzmann equation. Arch. Ration. Mech. Anal. 135, 61–105 (1996)
40. Loyalka, S.K., Ferziger, J.H.: Model dependence of the slip coefficient. Phys. Fluids 108, 1833
(1967)
41. Loyalka, S.K.: Approximate method in the kinetic theory. Phys. Fluids 11(14), 2291 (1971)
42. Otto, F.: Initial-boundary value problem for a scalar conservation law. C. R. Acad. Sci. Paris
Sér. I Math. 322(8), 729–734 (1996)
43. Siewert, C.E., Thomas, J.R.: Strong evaporation into a half space I. Z. Angew. Math. Phys. 32,
421 (1981)
44. Toro, E.F.: Riemann Solvers and Numerical Methods for Fluid Dynamics. Springer, Berlin
(2009)
45. Ukai, S., Yang, T., Yu, S.-H.: Nonlinear boundary layers of the Boltzmann equation. I.
Existence. Commun. Math. Phys. 236(3), 373–393 (2003)
46. Valein, J., Zuazua, E.: Stabilization of the wave equation on 1-d networks. SIAM J. Control
Optim. 48(4), 2771–2797 (2009)
47. Wang, W.-C., Xin, Z.: Asymptotic limit of initial boundary value problems for conservation
laws with relaxational extensions. Commun. Pure Appl. Math. 51(5), 505–535 (1998)
48. Xu, W.-Q.: Boundary conditions and boundary layers for a multi-dimensional relaxation
model. J. Differ. Equ. 197(1), 85–117 (2004)
49. Yong, W.-A.: Boundary conditions for hyperbolic systems with stiff relaxation. Indiana Univ.
Math. J. 48(1), 115–137 (1999)
Coagulation Equations for Aerosol
Dynamics
Marina A. Ferreira
M. A. Ferreira ()
Department of Mathematics and Statistics, Gustaf Hällströmin katu 2b, Helsingin yliopisto,
Finland
e-mail: marina.ferreira@helsinki.fi
1 Introduction
1 x ∞
∂t f (x, t) = K(x − y, y)f (x − y, t)f (y, t)dy − K(x, y)f (x, t)f (y, t)dy
2 0 0
(1.1)
where K(x, y) is the coagulation rate between particles of size x and y. The first
term on the right hand-side is the gain term due to the coagulation between particles
of size x − y and particles of size y to create a particle of size x. The second term
is the loss term which describes the loss of particles of size x by merging with
any other particle in the system. Equation (1.1) is an integrodifferential equation
belonging to the class of kinetic equations.
We also consider more general systems where a constant input of particles may
be present. The number density in this case satisfies the coagulation equation with
an extra source term η ≥ 0,
1 x
∂t f (x, t) = K(x − y, y)f (x − y, t)f (y, t)dy
2 0
∞
− K(x, y)f (x, t)f (y, t)dy + η(x). (1.2)
0
Complementary research lines have expanded over the last decades on exper-
imental, numerical and theoretical aspects of Eqs. (1.1) and (1.2). Algorithms to
simulate these equations have been developed to test hypotheses drawn from
atmospheric data [25, 34] (see [26] for a survey on numerical methods). On the
other hand, theoretical results have clarified issues mainly related to Eq. (1.1), such
as existence and uniqueness of solutions for general classes of kernels [18, 30] or
the behaviour of solutions for explicitly solvable kernels [28] and general kernels
[4, 5, 12, 17].
A particle may be characterized not only by its size but also by its composition,
leading to multi-component coagulation equations where the size is described by
Coagulation Equations 71
a vector x ∈ Rd+ \{0} representing the size of each of the chemical components
of a particle. An application of multi-component equations to aerosol dynamics is
described in Sect. 2.2.
In this paper we review analytic results related to the one-component equa-
tions (1.1) and (1.2) as well as to the corresponding discrete multi-component
equations with x, y ∈ Nd \{0}. We start in Sect. 2 with a short overview on
various topics related to properties of the solutions, applications and derivation
from particle systems. We also introduce some notation that is used throughout
the chapter. In Sect. 3 we study the one-component equations (1.1) and (1.2).
Section 3.2 contains the main steps of the proof of one of the first well-posedness
results for Eq. (1.1) with unbounded coagulation kernels obtained in 1999 by Norris
[30]. Section 3.3 contains a review of the proofs of existence and non-existence of
stationary solutions to coagulation equations to (1.2), obtained recently in [13]. In
Sect. 4 we consider the discrete multi-component equation with constant kernel.
Following the computations presented in [21], we compute in Sect. 4.1 explicit
time-dependent solutions and in Sect. 4.2 we compute stationary solutions when
an additional source at the monomers is present. We also obtain approximations
of both solutions showing explicitly that mass localizes along a straight line in the
multi-dimensional size space for large times and large sizes. Finally in Sect. 5 we
mention some recent results in the literature and open questions.
2 Preliminaries
Interestingly, if the coagulation rate behaves like a power law and if the power
is sufficiently large, then mass-conservation is lost. Such phenomenon is called
gelation and it corresponds to the formation of infinitely large clusters that are not
seen any more by the equation. Therefore these big clusters leave the system and
the total mass decreases. Gelation may be interpreted as a change in state from gas
to gel. Mathematically, this phenomenon poses interesting challenges [11]. Since
gelation has not been observed in atmospheric aerosols we do not discuss it further
here.
We note that, contrarily to the Boltzmann equation, the coagulation equation does
not preserve number of particles, due to the sticky collisions.
Atmospheric aerosols are suspensions of small particles in the air, whose diameter
ranges approximately between 1 nanometre, in the case of molecular particles, to
100 micrometres, in the case of cloud droplets and dust particles [19]. Aerosols
influence sunlight scattering by reflecting and absorbing radiation, and they consti-
tute the seeds that originate the clouds. Therefore, they play an important role in
weather and climate forecast [6].
Aerosols are subject to complex processes that influence their size distribution
over time. One important process is the coagulation of particles to produce larger
ones. Other processes include the formation of new small particles, or monomers,
due to certain physical and chemical processes, the removal of particles due to
gravity or diffusion, and the growth/shrinkage due to condensation/evaporation [25].
Atmospheric aerosols are typically constituted by different chemicals, leading to
multi-component systems, which may alter the rate of the processes mentioned
before and consequently, the particle size distribution [34].
We consider the regime in which the particles are uniformly distributed in space.
Moreover, we assume that removal and growth of particles due to condensation
is not important, which in practice may correspond roughly to sizes between 10
nanometers and 10 micrometers [19]. We are then led to the study of multi-
component systems where particles undergo binary coagulation in the presence of a
source of small particles.
Coagulation kernels K have been derived for atmospheric aerosols using kinetic
theory under several assumptions on the shape and motion of particles [19]. Aerosol
particles are commonly assumed to be spherical and to undergo elastic collisions
with background air particles. The number of such collision events is assumed to
be much larger than the number of collisions between two coalescing particles.
This drives the system towards an equilibrium where the particle velocities follow a
Maxwell-Boltzmann distribution.
Moreover, any collision between coalescing particles yields a coalescing particle.
Two different coagulation kernels have been derived under the previous conditions
for two different regimes. Each regime is defined based on the relation between
particle size and the average distance travelled by a particle between two collisions
Coagulation Equations 73
in air, called mean free path. Under normal pressure and temperature conditions,
the mean free path in air, , is of the order of 10 nanometres. If the size (diameter)
of a spherical particle, x, is much smaller than the mean free path x , the
particle is more likely to travel in straight lines before meeting another coalescing
particle. In this case the rate of coagulation has been estimated by the free molecular
coagulation kernel:
Otherwise, if the size of a particle is much larger than the mean free path, x , the
coalescing particle will meet many background air particles before meeting another
coalescing particle. In this case, the air behaves like a fluid and the coalescing
particle is more likely to diffuse. The coagulation rate has been estimated by the
diffusive coagulation kernel:
This kernel was first derived in the original work by Smoluchowski [32]. Other ker-
nels have been derived under different assumptions on the underlying background
gas and particles, such as particles moving in a laminar shear or turbulent flow [19],
and particles having electric charges [33, 34].
The behaviour and even the existence of solutions to Eq. (1.2) strongly depends
on the coagulation kernel. In Sects. 3.3.1 and 3.3.2 we review the existence of
stationary solutions for a large class of kernels which includes in particular the free
molecular (2.3) and the diffusive kernels (2.4).
The Smoluchowski’s coagulation equation has been rigorously derived using differ-
ent approaches that consider different types of particle systems. In one approach, a
purely stochastic particle system is considered, where pairs of particles are randomly
picked to originate a new particle. The associated stochastic process is usually called
Marcus-Lushnikov process. A different approach considers deterministic particle
systems, where particles move and when they collide they merge with a certain
probability.
The first approach is inspired in Kac-models for the derivation of the Boltzmann
equation [16]. A common strategy is to start from an infinite stochastic particle
system where particles of size x and y coalesce at a rate K(x, y) and to prove that
the number density, after being conveniently rescaled, converges, as the unit volume
tends to infinity, to a measure that solves the Smoluchowski’s coagulation equation
with kernel K. This has been obtained for the additive kernel, product kernel as
well as for a class of sub-multiplicative kernels using combinatorial techniques and
random graphs. See [2] (Chapter 5.2) for an accessible exposition and [1] for a
review on existing results and open problems.
74 M. A. Ferreira
In the second approach, there are fewer rigorous results. The first result to the best
of our knowledge is due to Lang and Xanh [22]. They consider Brownian particles
moving in the three-dimensional Euclidean space according to Brownian motion
with a diffusion coefficient D. The particles are assumed to move independently
on each other provided they are at a distance greater than the sum of their radius
2R. Once they come closer than 2R they coalesce with probability 1/2, forming
one Brownian particle with the same radius R and the same diffusion coefficient
D. In the limit when the number of particles N goes to infinity and the radius
R goes to zero, such that RN remains constant, the authors prove propagation of
chaos and that the density function converges in probability to the solution to the
Smoluchowski’s coagulation equation with constant coagulation kernel. The limit
where RN remains constant is the so-called Boltzmann-Grad limit and is the limit of
constant mean free time. A more general case of coalescing Brownian particles with
diffusion coefficients changing after coalescence, but not the size R, has been treated
in [20]. More recently, the change in size after coalescence has been considered in
[31] (see also [29]) in the case of a tracer particle moving in a straight line and
coalescing with randomly distributed fixed particles of different sizes. In this case,
a linear coagulation equation with a simple shear kernel was derived in the kinetic
limit where the volume fraction filled by the background of particles tends to zero.
2.4 Notation
3 One-Component Equation
for some given constants c1 , c2 , λ and γ and for all (x, y) ∈ (0, ∞)2 . This
class includes in particular the physical kernels (2.3) and (2.4). The parameter γ
represents the homogeneity of the kernel, while λ represents the “off-diagonal” rate.
The parameter γ yields the behaviour under the scaling of the particle size, while
λ measures the importance of collisions between particles of different sizes. Note
that the bounds in (3.2) and (3.3) are homogeneous, i.e., they satisfy for any k > 0,
h(kx, ky) = k γ h(x, y), but the kernels are not necessarily homogeneous.
We assume the following condition on the source η ∈ M+ (R∗ )
3. for all bounded measurable functions ϕ of compact support and t < T it holds
t
ϕ, ft = ϕ, f0 + ϕ, L(fs )ds (3.7)
0
1
ϕ, L(f ) = K(x, y)[ϕ(x + y) − ϕ(x) − ϕ(y)]f (dx)f (dy),
2 R∗ R∗
"
4. R∗ x1x≤1f0 (dx) < ∞ and (3.7) holds with ϕ(x) = x1x≤1.
If T = ∞ we call time-dependent solution to (1.1).
One can easily check that condition 2 is the minimal one to have well-defined
integrals. Condition 3 is the weak formulation commonly used in the literature and
76 M. A. Ferreira
is a stationary injection solution of (1.2) if the following identity holds for any test
function ϕ ∈ Cc (R∗ ):
1
K (x, y) [ϕ (x + y) − ϕ (x) − ϕ (y)] f (dx) f (dy) + ϕ (x) η (dx) = 0 .
2 R∗ R∗ R∗
(3.9)
Condition (3.8) is the minimal one for the integrals in (3.9) to be well-defined.
Stationary injection solutions have a constant in time flux of mass from small to
large sizes, due to the source, therefore they are non-equilibrium solutions. Note
that to be able to be stationary, the volume of particles entering the system has to
balance the volume of particles leaving the system. Interestingly, there is an implicit
removal of particles from the system at infinite sizes that allows the existence
of these solutions. As we will see in the next two Theorems, for some class of
coagulation rates, including the diffusive kernel (2.4), such balance exists, while for
other class of kernels, including the free molecular kernel (2.3), such balance does
not exist.
Theorem 3.4 Assume that K satisfies (3.1)–(3.5) and |γ + 2λ| < 1. Let η = 0
satisfy (3.6). Then, there exists a stationary injection solution f ∈ M+ (R∗ ), f = 0,
to (1.2) in the sense of Definition 3.3.
Theorem 3.5 Suppose that K satisfies (3.1)–(3.5) as well as |γ + 2λ| ≥ 1. Let us
assume also that η = 0 satisfies (3.6). Then, there is not any solution of (1.2) in the
sense of Definition 3.3.
Coagulation Equations 77
Note that the diffusive kernel (2.4) satisfies the growth conditions (3.2)–(3.3)
with γ = 0 and λ = 1/3, while the free molecular kernel (2.3) satisfies the growth
conditions with γ = 1/6 and λ = 1/2. Therefore there exists a stationary solution
for the diffusive but not for the free molecular kernel.
The mass flux from small to large sizes associated to a stationary injection
solutions is given in the next Lemma.
Lemma 3.6 Suppose that the assumptions of Theorem 3.4 hold. Let f be a
stationary injection solution in the sense of Definition 3.3. Then f satisfies for any
R>0
Remark
" 3.7 If R ≥ Lη , the right-hand side of (3.10) is constant equal to Jη =
[1,Lη ] xη(dx) > 0. Therefore, J (R) = Jη for R > Lη , i.e., the mass flux is constant
in the regions that include large sizes.
The main ideas to prove Lemma 3.6 are the following. For each ε > 0, we
define the test function ϕ(x) = xχε (x) ∈ Cc (R∗ ) where χε ∈ Cc∞ (R∗ ) is such that
0 ≤ χε ≤ 1, χε (x) = 1, for 1 ≤ x ≤ R, and χε (x) = 0, for x ≥ R + ε. Using this
test function in (3.9), the result can be obtained after letting ε → 0.
We describe the main ideas of the proof of Theorem 3.2 obtained in [30] (Section 2).
The first step is to prove well-posedness for a truncated problem. The second
step is to obtain estimates that allow us to remove the truncation and to obtain well-
posedness for the original problem.
Let B ⊂ R∗ be a compact set. Note that all measures in M (B) are bounded.
Note that from the hypotheses of Theorem 3.2 on the kernel we have that
for all bounded measurable functions ϕ on R∗ and all a ∈ R, where (ϕ, a), (f, ξ )
denotes ϕ, f + aξ . The truncated equation reads
t
(ϕ, a), (ft , ξt ) = (ϕ, a), (f0 , ξ0 ) + (ϕ, a), LB (f, ξ )ds. (3.12)
0
and proving that there exists a T > 0 such that (f n , ξ n ) converges in M (B) × R
uniformly in t ≤ T to the desired local solution, which is also unique.
The second step is to prove that ft ≥ 0, t ∈ [0, T ], which is obtained using
again an iterative argument similar to the one used in the first step.
Coagulation Equations 79
Finally, the third step is to show that the solution exists for all times t ∈ [0, ∞).
Choosing ϕ = w and a = 1 we obtain that
d 1
(w, ft + ξt ) = {w(x + y) − w(x) − w(y)}K(x, y)ft (dx)ft (dy) ≤ 0,
dt 2 R∗ ×R∗
Using a scaling argument, we may assume without loss of generality that w, f0 +
ξ0 ≤ 1, consequently fT + |ξT | ≤ 1. We can start again from (fT , ξT ) at time T
to extend the solution to [0, 2T ] and so on. Moreover, choosing ϕ = 0 and a = 1
in (3.12), we obtain
d 1
ξt = {w(x + y)1B̄ (x + y)}K(x, y)f (dx)f (dy) + ξt w2 (x)f (dx),
dt 2 R∗ ×R∗ R∗
which implies that ξt ≥ 0, for all t ≥ 0, due to ft ≥ 0, which ends the proof of the
proposition.
Proof of Theorem 3.2 Fix f0 ∈ M+ , such " that w, f0 < ∞. For each compact set
B ⊂ R∗ define f0B = 1B f0 and ξ0B = B̄ w(x)f0 (dx). From Proposition 3.9 there
is a unique solution (f0B , ξtB )t ≥0 to (3.12) starting from (f0B , ξ0B ). We now set ft =
lim ftB and ξt = lim ξtB . Using (3.11), we obtain by dominated convergence,
B→R∗ B→R∗
d 1
ϕ, ft = {ϕ(x +y)−ϕ(x)−ϕ(y)}K(x, y)ft (dx)ft (dy)−ξt ϕw, ft ,
dt 2 R∗ ×R∗
for all bounded measurable functions ϕ. One can prove that for all t < T and for
any local solution (gt )t <T ,
ξt = 0, t > 0. (3.14)
Then (3.13) and (3.14) imply that (ft )t ≥0 is a time-dependent solution to (1.1) and
moreover, it is the only solution.
80 M. A. Ferreira
We present here the main ideas of the proof of the existence Theorem 3.4 obtained
in [13]. The general strategy is similar to the strategy used in the proof of well-
posedness presented in the previous section. First, we prove existence of a stationary
solution for a truncated problem and second, we obtain estimates that allow to
remove the truncation and hence the existence result for the original problem.
Unfortunately the method used to prove existence does not give uniqueness, that
problem needs a separate treatment (see [23] for a simple explanation of the
available techniques).
Let ε > 0 and R∗ ≥ Lη , where Lη is the upper bound of the support of the source
η defined in (3.6). We will eventually make ε → 0 and R∗ → ∞. We consider
kernels Kε,R∗ that are continuous, bounded and have compact support, such that
and
where Kε is continuous and satisfies Kε (x, y) ∈ [a1 (ε), a2 (ε)], for all (x, y) ∈ R2+
and
Additionally, in the evolution equation, we consider a cut-off of the gain term due
to the coagulation that ensures that the measure solutions are supported in [1, 2R∗ ]
and bounded at all times. To this end, we choose ζR∗ ∈ C (R∗ ) such that 0 ≤ ζR∗ ≤
1, ζR∗ (x) = 1 for 0 ≤ x ≤ R∗ , and ζR∗ (x) = 0 for x ≥ 2R∗ . The regularized time
evolution equation then reads as
ζR∗(x)
∂t f (x, t) = Kε,R∗ (x − y, y)f (x − y, t)f (y, t)dy
2 (0,x]
Definition 3.10 Let ε > 0 and R∗ ≥ Lη . Suppose that Kε,R∗ satisfies (3.15)–
(3.17) and η ∈ M+ (R+ ) satisfies (3.6). Consider some initial data f0 ∈ M+ (R∗ )
for which f0 ((0, 1) ∪ (2R∗ , ∞)) = 0. Then f0 ∈ M+,b (R∗ ). We will say
that f ∈ C 1 ([0, T ] , M+,b (R∗ )) satisfying f (·, 0) = f0 (·) is a time-dependent
solution of (3.20) if the following identity holds for any test function ϕ ∈
C 1 ([0, T ] , Cc (R∗ )) and all 0 < t < T ,
d
ϕ (x, t) f (dx, t) − ϕ̇ (x, t) f (dx, t)
dt R∗ R∗
1 $ %
= Kε,R∗ (x, y) ϕ (x + y, t) ζR∗ (x + y) − ϕ (x, t) − ϕ (y, t)
2 R∗ R∗
×f (dx, t) f (dy, t)
f (dx, t) ≤ f0 (dx) + Ct , t ≥ 0,
R∗ R∗
"
for C = R∗ η(dx) ≥ 0 which is independent of f0 , t, and T .
To prove Prop. 3.11 we observe that since the kernel is bounded, the result may
be obtained using Banach fixed-point theorem.
Definition 3.12 Let ε > 0 and R∗ ≥ Lη . Suppose that Kε,R∗ satisfies (3.15)–
(3.17) and η ∈ M+ (R+ ) satisfies (3.6). We will say that f ∈ M+ (R∗ ), satisfying
f ((0, 1) ∪ (2R∗ , ∞)) = 0 is a stationary injection solution of (3.20) if the following
identity holds for any test function ϕ ∈ Cc (R∗ ):
1 $ %
0= Kε,R∗ (x, y) ϕ (x + y) ζR∗ (x + y) − ϕ (x) − ϕ (y) f (dx) f (dy)
2 R∗ R∗
+ ϕ (x) η (dx) .
R∗
82 M. A. Ferreira
S(t)f0 = f (·, t)
is invariant under the time evolution (3.20). Moreover, UM is compact in the ∗-weak
topology due to Banach–Alaoglu’s Theorem (cf.[3]), since it is an intersection of a
∗-weak closed set XR∗ and the closed ball f ≤ M.
The second step is to prove that for each t > 0, both maps S(t) : UM → UM
and t → S(t)f0 are continuous in the ∗-weak topology.
Finally, the third step of the proof reads as follows. Since for each t, the operator
S(t) is continuous and UM is compact and convex when endowed with the ∗−weak
topology, we can apply Schauder fixed point theorem to conclude that for all δ > 0
there is a fixed point fδ of S(δ) in UM . Moreover, since UM is metrizable and hence
sequentially compact, there is a convergent sequence {fδn }n∈N , i.e., there exists fˆ ∈
UM such that fδn → fˆ when δn → 0 in the ∗−weak topology. For each t we choose
δn = t/n. Using the semigroup property we obtain that S(t)fδn = S(nδn )fδn =
S(δn )fδn . Using the continuity of t → S(t)f0 and the fact that S(0)fˆ = fˆ, we
Coagulation Equations 83
obtain S(δn )fδn → fˆ. On the other hand using the continuity of S(t) we obtain
that S(t)fδn → S(t)fˆ. Therefore S(t)fˆ = fˆ and thus fˆ is a stationary solution
to (3.20), which concludes the proof.
The next Lemma provides uniform estimates for integrals.
Lemma 3.14 Let a > 0, R ≥ a and b ∈ (0, 1) be such that bR > a. Suppose
f ∈ M+ (R∗ ), ϕ ∈ C(R∗ ), g ∈ L1 (R∗ ), and g, ϕ ≥ 0. If
1
ϕ(x)f (dx) ≤ g(z) , for z ∈ [a, R] ,
z [bz,z]
then
"
[a,∞) g(z)dz
ϕ(x)f (dx) ≤ + Rg(R) .
[a,R] ln(b −1 )
1
Kε (x, y) [ϕ(x+y)−ϕ(x)−ϕ(y)]fε (dx) fε (dy)+ ϕ(x)η (dx) = 0.
2 [0,∞)2 [0,∞)
and
1 C̃
! fε (dx) ≤ √
3
z 2z
3 ,z z ε
2
for some subsequence {εn }n∈N with limn→∞ εn = 0. Using (3.19) and Lemma 3.14,
one can prove that fε satisfies (3.9) for any ϕ ∈ Cc (R+ ). In particular, f = 0 due
to η = 0.
It only remains to prove (3.8). Taking the limit of (3.22) as ε → 0 we arrive at:
1 C
f (dx) ≤ for all z ∈ (0, ∞),
z [2z/3,z] z3/2+γ /2
which implies
1 zμ
x μ f (dx) ≤ C for all z ∈ (0, ∞),
z [2z/3,z] z3/2+γ /2
for any μ ∈ R. From Lemma 3.14 we obtain the boundedness of the moment of
order μ:
x μ f (dx) < ∞.
[0,∞)
We present the main ideas of the proof of Theorem 3.5 obtained in [13]. The proof
is done by contradiction. Let the kernel K satisfy the power law bounds (3.2)–(3.3)
with |γ + 2λ| ≥ 1. Suppose that f ∈ M+ (R∗ ) is a stationary injection solution
of (1.1) in the sense of Definition 3.3. Then f satisfies the weak formulation (3.9)
as well as the condition on the moments (3.8).
Coagulation Equations 85
The first step is to rewrite (3.9) using the flux formulation. Consider the function
J : R∗ → R+ defined by
where
R = {x ≥ 1, y ≥ 1 : x + y > R, x ≤ R} .
J (R) = xη(dx), R ≥ 1.
[1,R]
J (R) = J (Lη ), R ≥ Lη .
The second step is to prove that the main contribution to the integral (3.23) as R →
∞ is due to collisions between particles of size close to R and particles of size of
order 1. To that end, for a given δ > 0 small, we consider a partition of R =
(1) (2)
Dδ ∩ Dδ such that
(1)
Dδ = {x ≥ 1, y ≥ 1 : y ≤ δx} ,
(2)
Dδ = {x ≥ 1, y ≥ 1 : y > δx}
and we define
Therefore
Using the upper bound for the kernel (3.3), the moment condition (3.8) and the fact
that R ∩ Dδ2 ⊂ [1, R] × [ 1+δ
δR
, ∞) one concludes after some computations that the
contribution of J2 vanishes as R → ∞, i.e.,
lim J2 (R) = 0
R→∞
86 M. A. Ferreira
The third step of the proof consists in obtaining a lower bound for the fluxes that
implies a lower bound for the number of particles in some region of the size space.
Using the upper bound for the kernel (3.3) we obtain after some computations
J Lη
lim inf R a+1
y f (dx) f (dy) ≥
b
. (3.25)
R→∞ (1)
R ∩Dδ c3 1 + δ |a−b|
(1)
R ∩ Dδ ⊂ {(x, y) : 1 ≤ y ≤ δR, R < x + y, x ≤ R}
J Lη 1
y b f (dy) f (dx) ≥ (3.26)
[1,δR] (R−y,R] 2c3 1 + δ |a−b| R a+1
for R ≥ R0 with R0 large enough. We now consider two cases separately a ≥ 0 and
a < 0. Let first a ≥ 0. Due to (3.24) we may define
J Lη 1
− [F (R − y) − F (R)] y b f (dy) ≤ − for R ≥ R0 .
[1,δR] 2c3 1 + δ |a−b| R a+1
Then, using a comparison argument (see Lemma 4.1 in [13]), for some constant
B > 0, it follows that
B
F (R) ≥ if R ≥ R0 , for a > 0, (3.28)
Ra
Coagulation Equations 87
and
B ≤ R a F (R) ≤ x a f (dx)
(R,∞)
Taking the limit when R → ∞ and using (3.24) it follows that B ≤ 0, which leads
to a contradiction. In the case a = 0, the contradiction follows from (3.29) in a
similar way using (3.24).
Let now a < 0. We define the function F by
J Lη 1
− [F (R) − F (R − y)] y b f (dy) ≤ − for R ≥ R0 .
[1,δR] 2c3 1 + δ |a−b| R a+1
As in the previous case, using a comparison argument (see Lemma 4.2 in [13]), it
follows that there is B > 0 such that
B
F (R) ≥ if R ≥ R0 .
Ra
For a small ε > 0 satisfying ε < B there exists M such that
x a f (dx) = ε .
[M,∞)
In this section we consider discrete coagulation equations where the particle size is
a discrete variable representing the number of monomers. In addition, we consider
that particles may be composed of different types of monomers. A particle with
d components is described by a vector α = (αi )i=1,...,d ∈ Nd+ \{0} where αi
represents the size of the ith component. The number density nα (t) of particles
with composition α at time t ≥ 0 satisfies the discrete multi-component coagulation
equation
1
∂t nα (t) = Kα−β,β nα−β (t)nα (t) − Kα,β nα (t)nβ (t) (4.1)
2
0<β<α β>0
where Kα,β represents the coagulation rate between particles with compositions α
and β. We use α < β to denote αi ≤ βi , i = 1, . . . , d, and α = β.
The continuous counterpart of the discrete multi-component coagulation equa-
tion may be written as (4.1) by substituting the sums by multidimensional integrals
and by taking the discrete vector α ∈ Nd+ \{0} as a continuous variable α ∈ Rd+ \{0}.
In the following sections, we consider discrete multi-component equations with
the constant kernel Kα,β = 2. We obtain an explicit time-dependent solution to (4.1)
and an explicit stationary solution in the presence of a source, by following the
computations presented in [21]. Similar results could also be obtained in the case of
continuous multi-component equations.
dnα
(t) = nβ (t)nα−β (t) − 2nα (t) nβ (t) (4.2)
dt
β<α β>0
d
where δα,β = 1 if α = β and δα,β = 0 otherwise. We recall that |β| = i=1 βi .
Coagulation Equations 89
Remark 4.1 Note that the initial condition (4.3) is supported at monomers and its
mass is uniformly distributed by the types of particles. The initial mass of each type
of particle is d1 .
Existence and uniqueness of a solution to (4.2)–(4.3) in the one-component case
d = 1 is proven in [28] for any initial condition satisfying ∞ α=1 nα (0) < ∞ using
Laplace transforms. An explicit solution to the multi-component problem has been
obtained in [21] and [27] using a generating function.
In this section we review the computations described in [21] to obtain an explicit
solution to (4.20). We then study the long-time behaviour using an approximation
of the solution for large times and large sizes as in [21]. In particular, we observe
the phenomenon of mass localization along a straight line in the size space.
Multiplying (4.20) formally by a test function ψα and summing in α we obtain
the weak formulation
∞
∞
∂t ψα nα (t) = [ψα+β − ψα − ψβ ]nα (t)nβ (t). (4.4)
α=1 α,β=1
The solution may be obtained using the generating function defined next.
The generating function F : D × R+ → R associated to a sequence
{nα (t)}α∈Nd \{0} is defined by
F (z, t) = zα nα (t) (4.5)
α>0
where D ⊂ Rd is the domain of convergence of the series and zα = z1α1 z2α2 . . . zdαd .
Using ψα = zα in (4.4) we obtain an equation for the generating function F ,
1
∂t N(t) = −N 2 (t), N(0) = 1 ⇐⇒ N(t) = . (4.7)
1+t
If we subtract Eqs. (4.6) and (4.7) we obtain an equation for F − N. More precisely,
we get ∂t (F − N) = (F − N)2 . Solving this equation and using (4.7) yields an
expression for F
F0 (z)
F (z, t) = (4.8)
(1 + t)(1 + t − tF0 (z))
90 M. A. Ferreira
where F0 (z) = F (z, 0) is given by F0 (z) = d1 di=1 zi after substituting (4.3)
in (4.5). The expression for F will be used in the following to determine the solution
to (4.2).
We note that if {nα }α>0 is a solution to dthe multi-component coagulation
equation (4.2), then {n }
|α| α>0 , where |α| = i=1 αi is the sum variable and n|α|
is defined by n|α| = β>0 nβ δ|α|,|β| , is a solution to the one-component equation
with constant kernel K|α|,|β| = 2 and initial condition n|α| (0) = δ|α|,1 . This result
may be obtained using the weak formulation (4.4) with a test function of the form
ψα = ϕ|α| . We first solve the one-component equation to find an expression for
{n|α| }α>0 .
We consider the generating function f : D × R+ → R, D ⊂ R, associated to the
one-component problem
∞
f (z, t) = z|α| n|α| (t), (4.9)
|α|=1
∞ |α| n (0)
which may be expressed by (4.8) with f0 (z, t) = |α|=1 z |α| = z, i.e.
z
f (z, t) = (4.10)
(1 + t)(1 + t − tz)
Comparing each term of the two series (4.11) and (4.9) we conclude that the solution
to the one-component equation is
t |α|−1
n|α| (t) = . (4.12)
(1 + t)|α|+1
Comparing
with (4.9) and using (4.12) and the fact that (z1 + . . . + zd )k =
k! α1 α2 αd
α1 !α2 !...αd ! z1 z2 . . . zd we finally obtain the solution to the multi-component
|α|=k
coagulation equation (4.2) expressed in terms of n|α| ,
1 |α|!
nα (t) = n|α| (t)g(α) with g(α) = . (4.14)
d |α| α1 !α2 ! . . . αd !
t ) = e to obtain an
To study the long time behaviour, we use the fact that lim( 1+t t
approximation for n|α| (t) for large |α| and large time t
|α|
n|α| (t) ≈ t −2 exp(− ). (4.15)
t
Remark 4.2 In [28] it is shown that the solution to the continuous one-component
equation with constant kernel does approach the form f (x, t) = t −2 exp − xt for
large times, provided the initial mass is either finite, which includes the case treated
in this section, or its mass distribution function diverges sufficiently weakly.
We also consider an approximation of the function g
|α|2−
g(α) ≈ |α|−(d−1)/2 exp(− ) (4.16)
2|α|
d
where |α|2− = 1
d (αi − αj )2 denotes the generalized mass difference variable.
i,j =1
Using (4.15) and (4.16) in (4.14) we obtain for large t and |α| the approximation
|α| |α|2
nα (t) ≈ t −2 |α|−(d−1)/2 exp(− ) exp(− − ). (4.17)
t 2|α|
We observe that, besides the mass scale |α| ∼ t imported from the solution√to
the one-component equation, there is a second mass scale given by |α|− ∼ t.
|α|
Introducing the variables ξ = |α|
t and ρ =
√− we may then write the solution in a
t
scaling form
where
ρ2
φ(ξ, ρ) = ξ −(d−1)/2 exp(−ξ ) exp(− ). (4.19)
2ξ
Finally we note from (4.17) that for any fixed time, nα (t) reaches maximum values
when |α|2− = 0. This condition defines a straight line in the size space given by
92 M. A. Ferreira
where sα is the source term. In analogy to the initial conditions in the time-
dependent case (4.3), the source term is given by
h
sα = δα,β , (4.21)
d
|β|=1
satisfies
√
where S(z) = dh |α|=1 zα = hd di=1 zi and N = h is obtained using an
appropriate test function in (4.22). The solution to (4.24) reads
&
√ |z|
F (z) = h[1 − 1 − ]. (4.25)
d
The solution to (4.20) is obtained by expanding F in powers of the variables zi and
comparing with (4.23), yielding
The existence and uniqueness of a time-dependent solution have also been estab-
lished in [30] for coagulation kernels satisfying (3.2)–(3.3) with γ + λ = −λ and
λ > −1/2 using a similar reasoning as the one we presented in Sect. 3.3.1 [30].
Moreover, in [10] existence is obtained using a functional framework, for a class of
kernels satisfying (3.2)–(3.3) with c1 = c2 = 1, λ ∈ [−1, 1], γ ∈ [0, 2], γ ≤ −2λ,
γ + λ ∈ [−1, 1] and (γ , λ) = (−λ, −1). In [18] uniqueness is proved globally
in time for a class of kernels satisfying some regularity conditions as well as the
bounds (3.2)–(3.3) with γ ≤ 1, λ = 0, and for a different class of kernels such that
γ ∈ (1, 2], λ = 1 up to a gelation time T . For more general classes of kernels,
both existence and uniqueness remain open problems. We refer to the survey [24]
for further references.
In the presence of a constant source of small particles, the existence and non-
existence of stationary solutions presented in Sect. 3 are the most recent existence
results to the best of our knowledge. Previous results [8] were obtained for particular
classes of kernels that are included in the more general setting presented here. In
the case of multi-component equations with d components, source and kernel K
satisfying
d
c1 w(x, y) ≤ K(x, y) ≤ c2 w(x, y) with w(x, y) = x γi −λi y λi + y γi −λi x λi ,
i=1
(5.1)
we expect the existence result (Theorem 3.4) to remain valid for a class of kernels
satisfying |γi +2λi | < 1 for all i = 1, . . . , d [14]. In the same line, the nonexistence
result (Theorem 3.5) should hold true if |γi + 2λi | ≥ 1 for some i. Moreover,
stationary solutions are expected to exhibit mass localization along a straight line
for a class of kernels satisfying growth bounds that are invariant under permutations
of the components [15].
To the best of our knowledge, nothing is known about rigorous results for
multi-component coagulation equations with general kernels. However, the well-
posedness results for the one-component case, are expected to remain true in the
multi-component case provided the kernel satisfies the bounds (5.1) with γi and λi
satisfying the same conditions for well-posedness in dimension d = 1 for all i. Mass
localization for large times is expected to hold for a class of kernels that satisfy the
same bounds with the additional condition that w is invariant under any permutation
of the components. Without this invariance condition, the mass may not localize, the
mass may not localize in a straight line, due to the different rates of coagulation of
each component. Multiscale behaviour is then expected to emerge that could break
down the nice localization structure.
Coagulation Equations 95
References
1. Aldous, A.: Deterministic and stochastic models for coalescence (aggregation and coagula-
tion): a review of the mean-field theory for probabilists. Bernoulli 5(1), 3–48 (1999)
2. Bertoin, J.: Random Fragmentation and Coagulation Processes, vol. 102. Cambridge Univer-
sity Press, Cambridge (2006)
3. Brezis, H.: Functional Analysis, Sobolev Spaces and Partial Differential Equations. Springer,
Berlin (2010)
4. Bonacini, M., Niethammer, B., Velázquez, J.J.L.: Self-similar solutions to coagulation equa-
tions with time-dependent tails: the case of homogeneity smaller than one. Commun. Part.
Differ. Equ. 43(1), 82–117 (2018)
5. Bonacini, M., Niethammer, B., Velázquez, J.J.L.: Self-similar solutions to coagulation equa-
tions with time-dependent tails: the case of homogeneity one. Arch. Rat. Mech. Anal. 233(1),
1–43 (2019)
6. Carslaw, K.S., Lee, L.A., Reddington, C.L., Pringle, K.J., Rap, A., Forster, P.M., Mann, G.W.,
Spracklen, D.V., Woodhouse, M.T., Regayre, L.A., Pierce, J.R.: Large contribution of natural
aerosols to uncertainty in indirect forcing. Nature 503(7474), 67–71 (2013)
7. da Costa, F.P.: Mathematical aspects of coagulation-fragmentation equations. In Bourguignon,
J.P., Jeltsch, R., Pinto, A., Viana, M. (eds.) Mathematics of Energy and Climate Change; CIM
Series in Mathematical Sciences, vol. 2, pp. 83–162. Springer, Cham (2015)
8. Dubovski, P.B.: Mathematical Theory of Coagulation. Lecture Notes Series, vol. 23. Seoul
National University, Seoul (1994)
9. Elimelech, M., Gregory, J., Jia, X., Williams, R.A.: Particle Deposition and Aggregation—
Measurement, Modelling and Simulation. Elsevier, Amsterdam (1995)
10. Escobedo, M., Mischler, S.: Dust and self-similarity for the Smoluchowski coagulation
equation. Ann. I. H. Poincarè Anal. Non Linéaire 23(3), 331–362 (2006)
11. Escobedo, M., Mischler, S., Perthame, B.: Gelation in coagulation and fragmentation models.
Commun. Math. Phys. 231(1), 157–188 (2002)
12. Escobedo, M., Mischler, S., Rodriguez Ricard, M.: On self-similarity and stationary problem
for fragmentation and coagulation models. Ann. Inst. H. Poincaré Anal. Non Linéaire 22(1),
99–125 (2005)
96 M. A. Ferreira
13. Ferreira, M.A., Lukkarinen, J., Nota, A., Velázquez, J.J.L.: Stationary non-equilibrium solu-
tions for coagulation systems. Arch. Ration. Mech. Anal. 1–67 (2019) https://doi.org/10.1007/
s00205-021-01623-w
14. Ferreira, M.A., Lukkarinen, J., Nota, A., Velázquez, J.J.L.: Multicomponent coagulation
systems: existence and non-existence of stationary non-equilibrium solutions (2021). arXiv
preprint arXiv:2103.12763
15. Ferreira, M.A., Lukkarinen, J., Nota, A., Velázquez, J.J.L.: Localization in stationary
non-equilibrium solutions for multicomponent coagulation systems (2020). arXiv preprint
arXiv:2006.14840
16. Fournier, N., Giet, J.S.: Convergence of the Marcus–Lushnikov process. Methodol. Comput.
Appl. Probab. 6, 219–231 (2004)
17. Fournier, N., Laurençot, P.: Existence of self-similar solutions to Smoluchowski’s coagulation
equation. Commun. Math. Phys. 256(3), 589–609 (2005)
18. Fournier, N., Laurençot, P.: Well-posedness of Smoluchowski’s coagulation equation for a class
of homogeneous kernels. J. Funct. Anal. 233, 351–379 (2006)
19. Friedlander, S.K.: Smoke, Dust, and Haze. Oxford University Press, Oxford (2000)
20. Hammond, A., Rezakhanlou, F.: The kinetic limit of a system of coagulating Brownian
particles. Arch. Ration. Mech. Anal. 185(1), 1–67 (2007)
21. Krapivsky, P., Ben-Naim, E.: Aggregation with multiple conservation laws. Phys. Rev. E
53(10), 1103 (1995)
22. Lang, R., Xanh, N.X.: Smoluchowski’s theory of coagulation in colloids holds rigorously in
the Boltzmann-Grad-limit. Z. Wahrscheinlichkeitstheor. verw. Geb. 54(3), 227–280 (1980)
23. Laurençot, P.: Weak compactness techniques and coagulation equations. In: Evolutionary
Equations with Applications in Natural Sciences, pp. 199–253. Springer, Cham (2015)
24. Laurençot, P., Mischler, S.: On coalescence equations and related models. In: Degond, P.,
Pareschi, L., Russo, G. (eds.) Modeling and Computational Methods for Kinetic Equations.
Modeling and Simulation in Science, Engineering and Technology. Birkhäuser, Boston (2004)
25. Lehtinen, K.E.J., Kulmala, M.: A model for particle formation and growth in the atmosphere
with molecular resolution in size. Atmos. Chem. Phys. 3, 251–257 (2003)
26. Lee, M.H.: A survey of numerical solutions to the coagulation equation. J. Phys. A 34(47),
10219 (2001)
27. Lushnikov, A.A.: Evolution of coagulating systems. III. Coagulating mixtures. J. Colloid.
Interf. Sci. 54(1), 94–101 (1976)
28. Menon, G., Pego, R.: Approach to self-similarity in Smoluchowski’s coagulation equation.
Commun. Pure and Appl. Math. 57(9), 1197–1232 (2004)
29. Niethammer, B., Nota, A., Throm, S., Velázquez, J.J.L.: Self-similar asymptotic behavior for
the solutions of a linear coagulation equation. J. Differ. Equ. 266(1), 653–715 (2019)
30. Norris, J.R.: Smoluchowski’s coagulation equation: uniqueness, nonuniqueness and a hydro-
dynamic limit for the stochastic coalescent. Ann. App. Probab. 9, 78–109 (1999)
31. Nota, A., Velázquez, J.J.L.: On the growth of a particle coalescing in a Poisson distribution of
obstacles. Commun Math Phys 354(3), 957–1013 (2017)
32. Smoluchowski, M.: Drei Vorträge über diffusion, Brownsche Molekularbewegung und Koag-
ulation von Kolloidteilchen. Phys. Z. 17, 557–599 (1916)
33. Su, T., Bowers, M.T.: Theory of ion-polar molecule collisions. Comparison with experimental
charge transfer reactions of rare gas ions to geometric isomers of difluorobenzene and
dichloroethylene. J. Chem. Phys. 58(7), 3027–3037 (1973)
34. Olenius, T., Kupiainen-Määttä, O., Ortega, I.K., Kurtén, T., Vehkamäki, H.: Free energy barrier
in the growth of sulfuric acid–ammonia and sulfuric acid–dimethylamine clusters. J. Chem.
Phys. 139(17), 084312 (2013)
Multibody and Macroscopic Impact
Laws: A Convex Analysis Standpoint
Abstract These lecture notes address mathematical issues related to the modeling
of impact laws for systems of rigid spheres and their macroscopic counterpart. We
analyze the so-called Moreau’s approach to define multibody impact laws at the
mircroscopic level, and we analyze the formal macroscopic extensions of these
laws, where the non-overlapping constraint is replaced by a barrier-type constraint
on the local density. We detail the formal analogies between the two settings,
and also their deep discrepancies, detailing how the macroscopic impact laws,
natural ingredient in the so-called Pressureless Euler Equations with a Maximal
Density Constraint, are in some way irrelevant to describe the global motion of
a collection of inertial hard spheres. We propose some preliminary steps in the
direction of designing macroscopic impact models more respectful of the underlying
microscopic structure, in particular we establish micro-macro convergence results
under strong assumptions on the microscopic structure.
1 Introduction
The modeling of particle systems spreads over a wide range of approaches, which
rely on various levels of description of the particles. At one end of this range, the
microscopic/Lagrangian setting is based on an individual description of particles,
which “simply” obey Newton’s Laws. At the other end, macroscopic models rely on
F. Bourdin
Laboratoire de Mathématiques d’Orsay, Université Paris-Sud, Orsay, France
e-mail: felicien.bourdin@ens.fr
B. Maury ()
Département de Mathématiques et Applications, Ecole Normale Supérieure, PSL University,
Paris, France
e-mail: bertrand.maury@ens.psl.eu; bertrand.maury@math.u-psud.fr
expected model takes the form of the so called pressureless equations with maximal
density constraint, which essentially fits into a sound framework in the one-
dimensional setting only [4, 6]. For higher dimension, little is known on this
equation. Let us add that the system is commonly written without any collision law,
the actual choice being usually made in an implicit way, depending on the approach
which is followed. For instance, in [6], particular solutions are built by means of
sticky blocks with a purely inelastic collision law, whereas in [9, 10], the approach
is based on compressible Euler equation with a barrier-like pressure with respect to
the density, natively leading to a purely elastic behavior.
The largest part of this text is meant to be accessible to graduate students, so
we tried to preserve self-consistency as far as possible, writing at some points full
proofs of elementary results, in particular in the appendix.
From Single Collision to Multibody Impact Laws
We introduce here Moreau’s approach of impact laws, which fits in the general class
of Contact Dynamics Methods (see [24, 29]). Let us start with a point particle subject
to remain in the upper half plane R × R+ , with a purely inelastic collision law on the
boundary. We denote by r = (x, y) its position, and by u its velocity. If this particle
is not subject to any force, its motion follows
2
R if y > 0
u = PCr u , with Cr =
+ −
(1)
R × R+ if y = 0
where u− (resp. u+ ) is the pre- (resp. post-) collisional velocity, and PCr is the
Euclidian projection on Cr . When the particle does not touch the wall, the velocity
is constant. When a collision occurs, with pre-collisional velocity u− = (ux , uy )
(with uy < 0), the post-collisional velocity is u+ = (ux , 0).
In the case of an elastic collision, we introduce a restitution coefficient e ∈ (0, 1].
The post-collisional velocity is now u+ = (ux , −euy ). This behavior can be written
in a way which can generalized to the multi-collisional situation. We introduce the
outward normal cone to K, defined as
' (
Nr = Cr◦ = v ∈ R2 , ß < v, w >≤ 0 ∀w ∈ Cr = {0} × R− .
u+ = u− − (1 + e)PNr u− .
rj − ri
Denoting eij = , the set of admissible velocities is
|rj − ri |
) *
Cr = v , Dij (r) = rj − ri − 2R = 0 ⇒ ß < eij , (vj − vi ) >≥ 0 . (2)
(N.B.: we use the notation ß < a, b > to denote the dot product of vectors in the
physical space Rd , while ·|· shall be used for generalized velocity vectors in Rnd ,
or elements in abstract Hilbert spaces.)
Let r = (r1 , . . . , rn ) ∈ K be given. As previously, the outward normal cone to
K at r is defined as the polar set to the cone of feasible velocities:
' (
Nr = Cr◦ = v ∈ Rdn , v|w ≤ 0 ∀w ∈ Cr .
u = U − (1 + e)PNr U, (3)
where e ∈ [0, 1] is the restitution coefficient. Since Nr and Cr are mutually polar, it
holds that
I = PNr + PCr ,
where I is the identity operator in Rdn (see [23], or the proof of Proposition 11 in
the appendix). As a consequence, the post-collisional velocity can be expressed in
terms of PCr U , for any e ∈ [0, 1],
Cr = {v , Bv ≤ 0} = B −1 o+ , + = Rm
+, (7)
Proof Let us start by a preliminary remark: the fact that the image of B is closed (it
is a finite dimensional linear space) is not sufficient to ensure that B (+ ) is closed
(see Remark 11 in the appendix). This property is nevertheless true here, because
B (+ ) is spanned by a finite number of vectors
⎧ ⎫
⎨ ⎬
B (+ ) = − pij Gij , pij ≥ 0 (9)
⎩ ⎭
ij
of constraints is asymptotically 3n, which mean that the dimension of the kernel of
B is asymptotically n.
The problem nevertheless presents some sort of uniqueness property, restricted to
the homogeneous problem (U = 0). The following proposition essentially states a
very intuitive fact: if one considers any static configuration of a finite number of hard
spheres in the open space (i.e. with no walls), under the assumption that interaction
contact forces are only repulsive, then all forces are actually zero. This property
will be used to show that the solution set for the pressure field (Proposition 1) is
bounded.
Lemma 1 We consider an admissible configuration r ∈ K, and the associated
matrix B ∈ Mm,n (R) (the raws of which are given by (6)). The set
) *
+ , B q = 0 = ker B ∩ +
S = q ∈ Rm
is reduced to {0}.
Proof Let us first establish the uniqueness for the homogeneous problem. We
consider q = (qij ) ∈ Rm
+ such that
B q= qij Gij = 0.
i∼j
where i ∼ j means that the particles i and j are in contact. Let i0 denote the index
of an extremal vertex of the convex hull conv(qi , 1 ≤ i ≤ n). By Hahn-Banach’s
theorem, the compact {qi0 } and the set conv{qi , i = i0 }, which is closed and convex,
can be separated in a strict sense by a plane in Rd . We denote by x an element of
this plane, and by v a normal vector to it. One has
so that ß < (qi0 −qj ), v >> 0 for j = i0 . Now the balance of contact forces exerted
upon sphere i0 in the direction v reads
qj i0 ß < ej i0 , v >
j =i0
where ß < ej i0 , v >> 0, and qj i0 ≥ 0 for all j . This quantity is positive unless
qj i0 = 0 for all j = i0 . Therefore all multipliers associated to a contact with sphere
i0 are equal to 0, and this approach can be iterated for the reduced family (qj , j =
i0 ). By downward induction on the number of active spheres, we prove that S is
reduced to {0}.
An important consequence of this expected property is the boundedness of the
solution set for (8).
Multibody and Macroscopic Impact Laws: A Convex Analysis Standpoint 105
Proposition 2 Under the assumptions of Proposition 1, the solution set for the dual
component p
) *
S = q ∈ Rm
+ , B q = U − u = p + ker B ∩ + ,
Note that, since we assumed that the support of ρ is strongly included in , it holds
that ρ ⊂ H01 ().
106 F. Bourdin and B. Maury
It can be easily checked that, for a smooth velocity field v and a regular saturated
zone, belonging to Cρ is equivalent to verifying ∇ · v ≥ 0 on [ρ = 1].
The non-elastic collision law writes
u = PCρ U,
We introduce
where Bv is defined by
. /
Bv, p = v · ∇p dρ. (13)
Note that and are not identified here, and that B maps V onto , so that the
adjoint operator B is defined in L(, V ), the set of continuous linear mappings
from into V .
The saddle-point formulation of the problem can be written
u + ∇p = U ρ-a.e. in ,
“ −∇·u ≤0 in [ρ = 1]”,
≥ 0 ρ-a.e. in , (14)
p
u · ∇p dρ = 0,
where the second equation (between quotation marks) is meant in a weak sense, i.e.
u · ∇q dρ ≤ 0 ∀q ∈ + .
which writes ∇qL2ρ ≥ βqH 1 in the present context. Due to Poincaré Inequality,
which holds true because ⊂ H01 (), it is sufficient to establish that the inequality
∇qL2ρ ≥ β∇qL2 holds for any q ∈ .
For q ∈ , by Theorem 1.56 in [33] one has (1 − ρ) ∇q = (1 − ρ) 1q =0 ∇q = 0,
so that ∇qL2ρ = ∇qL2 . As a consequence, B has a closed range, and so does
B by Banach-Steinhaus Theorem. The range of B is also dense thanks to the same
inequality, thus B is onto.
3 Micro-Macro Issues
We detailed in the previous section impact laws for a collection of rigid spheres, in
the Moreau’s spirit, and we proposed a natural instantiation of the same principles at
the macroscopic level. The macroscopic version may appear as a natural candidate
to handle collision between clusters of infinitely many hard spheres represented
by a diffuse density. We shall see here that some considerations may comfort this
standpoint in the one-dimensional setting. Yet, for dimensions d ≥ 2, we shall prove
that the macroscopic law presented in the previous section is not a relevant model
for describing the impact between large collections of hard spheres.
One Dimensional Setting
In the one-dimensional setting (hard spheres move on a fixed line) the two
approaches are mutually consistent, as we shall see here.
First, the notion of maximal density is well defined at the microscopic level: a
cluster of spheres (represented by segments in 1d) is saturated if the solid phase
covers some zone of the real line, which corresponds to ρ = 1 in the macroscopic
setting.
Now consider such a cluster of n segments covering an interval I ∈ R, and
the corresponding macroscopic density ρ = 1I (characteristic function of I ). We
consider a pre-collisional velocity field U that pushes the configuration against the
boundary of the feasible set, i.e. such that ∂x U ≤ 0. In this case the constraint will be
108 F. Bourdin and B. Maury
BB p = BU, (18)
with
⎛ ⎞
⎛ ⎞ 2 −1 0 · · 0
1 −1 0 . . . ... ⎜ ⎟
⎜ −1 2 −1 0 · · ⎟
⎜0 1 −1 . . . ⎟ ⎜ ⎟
⎜ ... ⎟ ⎜ ·⎟
B=⎜ ⎟ ∈ Mn−1,n (R) , BB = ⎜ 0 −1 · · ⎟ ∈ Mn−1 (R),
⎜ . . ⎟ ⎜ · · · · ·⎟
⎝0 0 .. .. ... ⎠ ⎜
⎜
⎟
⎟
0 0 ... 1 −1 ⎝ · · 2 −1 ⎠
0 · · 0 −1 2
that is the discrete Laplacian matrix. The two formulations are mutually consistent
in the sense that the linear system (18) is a standard finite difference discretization
of the Poisson problem (16), which is covered by rigorous convergence results (see
e.g. [2]).
Multibody and Macroscopic Impact Laws: A Convex Analysis Standpoint 109
Case d ≥ 2
In higher dimensions the situation is fully different. First, the notion of maximal
density is not clearly defined at the microscopic level. Let us consider
√ collections
of identical discs. The maximal packing density ρmax = π/2 3 ≈ 0.9069 . . . ,
and corresponds to the triangular lattice (see Fig. 3, right). Yet the actual density of
moving collections of rigid disks is generally strictly less than this maximal value,
which does not mean that the flow is unconstrained (as the macroscopic setting
would suggest). These considerations call for a clear identification of configurations
which saturate the constraint. It is tempting to consider as maximal in some sense
any density corresponding to such configurations, for which there are no free disks,
so that constraints are activated everywhere. The triangular lattice is clearly jammed,
but so is the Cartesian lattice (ρ = π/4 ≈ 0.79), and it is possible to build looser
jammed configurations, for example by removing some non neighboring discs from
the triangular lattice. We refer to [32] for a general review on the notion of maximal
random packing.
Beyond this difficulty to properly define the notion of maximal density, the
microscopic and macroscopic projections exhibit deep discrepancies in dimensions
higher than 1. We propose here to enlighten these discrepancies by considering
the underlying Poisson problems for the pressure in both settings. Like in the
one-dimensional setting, we first consider the macroscopic setting, which is in
some manner simpler than the microscopic one, in spite of its infinite dimensional
character.
The pressure can be shown, under some assumptions, to verify a Poisson like
problem in the saturated zone, The first step consists in proving that the problem
verifies the abstract maximum principle (see Definition 3), that is
U · ∇q ≥ 0 ∀q ∈ , q ≥ 0 a.e., (19)
110 F. Bourdin and B. Maury
∇p · ∇q = U · ∇q ∀q ∈
takes nonnegative values almost everywhere, i.e. it lies in + . This property takes
the form of a maximum principle for the Laplace operator, in an extended sense:
the saturated zone [ρ = 1] may be not be the closure of an open domain, it may in
particular have an empty interior, while having a positive measure (see Remark 1).
This property is obtained by a standard procedure, which consists in taking a test
function equal to the negative part of p, i.e. q = p− = − min(0, p). We have that
∇q = −∇p1p≤0 (see Theorem 1.56 in [33]), and q ≥ 0, so that
− 2
− ∇p = U · ∇p− ≥ 0,
which implies that ∇p− vanishes almost everywhere, i.e. p− is constant on . Since
it is 0 in the neighborhood of the boundary, it vanishes on i.e. p ≥ 0 a.e. in .
In other words, if the pre-collisional field is non-expansive, i.e. ∇ · U ≤ 0, then
the pressure field p is a weak solution to the Poisson problem
− p = −∇ · U, (20)
N N
n ) ∈2 that is dense in , and a sequence (Rn ) ∈ (0, +∞] such that
(c
πRn ≤ ||/2. For a given rn < Rn , we denote by γn the circle of radius rn ,
centered at cn , by !n the cocentric circle of radius Rn , and by n the ring domain
between these circles. We denote by gn the solution to the following Dirichlet
problem in n
−g = 0 in n ,
g = 0 on γn , (21)
g = 1 on !n ,
extended by 0 inside the small disc, and by 1 outside the large one. Since the capacity
of a point is 0 in R2 (see e.g. [22]), one can choose rn , with 0 < rn < Rn , sufficiently
small to ensure that
1
|∇gn |2 ≤ .
22n
We denote by ω the union of the small discs (centered at cn , with radius rn ), which
is open and dense by construction. Now consider the function Gn = g1 g2 . . . gn . It
holds that
n
∇Gn = ∇gn gj ,
k=0 j =k
so that (all the gj take values between 0 and 1 by construction), by the triangular
inequality in L2 (),
n
n
1
∇Gn L2 () ≤ ∇gk L2 () ≤ ≤ 2.
2k
k=0 k=1
Fig. 4 In the configuration represented here, considering that the distances are subject to remain
0 (constraint Bu = 0), the pre-collisional velocity tends to push any two grains in contact toward
overlapping, and yet the pressure between the two grains in the center will be negative
BB p = BU.
Since the horizontal velocities have a much larger magnitude, in spite of the fact
that the pre-collisional velocity pushes the grains against each other (i.e. BU >
0), it is clear that the pressure associated with the contact between the two central
grains will be negative, which rules out the maximum principle (in the sense of
Definition 3). As a consequence, the solution to the impact problem, with a unilateral
constraint, will not be the same: the grains at the center will be pushed apart during
the collision, which implies (thanks to the complementarity constraint Bu|p = 0)
that the corresponding pressure is 0. Note that some sort of Poisson Problem can be
recovered for the pressure associated with the impact law, by removing the raws of
B which correspond to non activated contacts, i.e. with −Gij · u < 0. If one denote
by B the corresponding matrix, and by p the corresponding pressure, it holds that
B B p = B U,
with a reduced matrix B which may also not verify the maximum principle.
This violation of the maximum principle for BB is generic in the hard-sphere,
microscopic, setting, as soon a dense collections of grains are concerned. It can be
checked for simple situation that the matrix BB , unlike in the one-dimensional
setting, has positive off-diagonal entries.
Square and Triangular Lattices
As an illustration of the previous considerations, and as an introduction to the next
section, let us make some remarks on very specific situations, where the overall
Multibody and Macroscopic Impact Laws: A Convex Analysis Standpoint 113
behavior of a collection of rigid discs can be seen to significantly differ from the
behavior given by the macroscopic impact law.
Consider at first a jammed configuration structured according to a square lattice
(see Fig. 3 (left)). On each row, the non-overlapping constraints impose horizontal
velocities to be non-decreasing. Similarly, on each column, the vertical velocities
must be non-decreasing also. Two fields of Lagrange multipliers can therefore be
associated to the constraints in the main directions x and y, which act on the
system independently from each other. As a consequence, two constraints must be
verified, to be compared to the single scalar constraint of the macroscopic constraint
∇ · u ≥ 0.
In the case of a triangular lattice (see Fig. 3 (right)), the monotonicity of the
velocity is imposed in each of the 3 principal directions.
In both cases, the microscopic constraints are much stronger than the macro-
scopic one, which is therefore obviously irrelevant to model at the macroscopic scale
the collections of hard discs. The next section is dedicated to designing macroscopic
models more respectful of the underlying microscopic structure, in the case of
crystal-like configurations.
1
1 2
3 0 4 3 0 4
5 6
2
Fig. 5 The two structured jammed configuration. On both side, the spheres in contact with the
sphere 0 are labelled from 1 to 4 or 6
114 F. Bourdin and B. Maury
four, each one corresponding to a contact with an adjacent sphere. The microscopic
constraint for spheres in contact described by (2) writes here
(u1 − u0 ) · ey ≥0
(u0 − u2 ) · ey ≥0
(22)
(u4 − u0 ) · ex ≥0
(u − u ) · e ≥0
0 3 x
∂p
− ux ≥0 (23)
∂x
for every nonnegative test function p such that its weak partial derivative in x can be
defined. In order to clarify this last condition, we will introduce anisotropic Sobolev
spaces, naturally defined to formalize the notion of “weakly derivable along one
direction”. The following description of these spaces is extracted from [15]. In what
follows, is a strictly convex bounded open set, with regular boundary. We refer
to [15] for the study of more general domains.
Definition 1 The anisotropic Sobolev space in the direction x on is defined by:
∂f
Hx1 () 2
= f ∈ L () , 2
weakly exists in L () (24)
∂x
∂f
where “ weakly exists in L2 () ” means that
∂x
∂f ∂g
∀g ∈ C 1 () , g=− f. (25)
∂x ∂x
0 02
0 ∂f 0
This space is endowed with the norm f 2H 1 = f 22 +0 0
0 ∂x 0
x
2
Since C x and C y are closed convex cones, the projection problem 1 admits a
unique solution. We can write the saddle-point formulation of the problem, that is
the instantiation of the abstract formulation (50) to the present situation.
Proposition 6 There exists a unique pair of nonnegative Lagrange multipliers (or
pressures) px , py ∈ x × y such that
∂px ∂py
u+ ex + ey = U ρ-a.e. (29)
∂x ∂y
where
⎧ 2 −1 −1
⎨ L ()2 −→ H
x () × Hy ()
B: ∂ux ∂uy
⎩ u −→ − ,−
∂x ∂y
) *
and + = qx , qy ∈ x × y , qx ≥ 0 , qy ≥ 0 a.e. . By Poincaré inequal-
ity, there exists a constant β > 0 such that |B μ| ≥ β|μ|, with
⎧
1 () × H 1 () −→
⎨ H0,x 0,y L2 ()2
B : ∂qx ∂qy
⎩ qx , qy →
− ex + ey .
∂x ∂y
solutions to
∂px
∂x + ux = Ux ρ-a.e.
(30)
∂ux
− ≤ 0 where ρ = 1,
∂x
∂py
∂y + uy = Uy ρ-a.e.
(31)
− ∂uy ≤ 0 where ρ = 1.
∂y
This is the macroscopic counterpart of what we had seen at the end of Sect. 3:
two independant pressure fields appear, acting separately on each component of
the velocity in order to correct the compressions in x and y.
Remark 2 This model introduces anisotropy, so that the collision is no longer
rotationally invariant: Fig. 6 shows the situation of two colliding blocks, under three
angles of impact. In the case of an impact along one of the two principal directions,
no perturbation occurs in the tangential direction whereas in the case of an impact
involving both directions (second case in Fig. 6), a transverse velocity appear. Note
that in the third case of two blocks colliding along a direction that is very close to
one of the two axes, the post-impact velocity is mainly directed along the transverse
direction.
Multibody and Macroscopic Impact Laws: A Convex Analysis Standpoint 117
Fig. 6 Three impacts between opposing blocks, varying the angle of incidence. On the left, the
velocity fields before the impact; on the right, the velocity fields after the impacts
118 F. Bourdin and B. Maury
We shall now build a more pathological macroscopic model derived from the
microscopic configuration of a triangular (or hexagonal) stack of particles (see
Fig. 5, right). The well-posedness of the saddle-point formulation (i.e. existence and
uniqueness of pressure fields) is more delicate than before: uniqueness is lost, as
we shall see later on, and the existence is still an open problem. We introduce unit
vectors along the principal directions:
√ √
1 3 1 3
e0 = (1, 0) , e1 = − , , e2 = − , − .
2 2 2 2
In this case, any sphere has 6 neighbors, two along each axis directed by the ei .
As for the previous configuration, one can write the microscopic constraints on the
sphere 0
(u4 − u0 ) · e0 ≥ 0 , (u0 − u3 ) · e0 ≥ 0 , (u1 − u0 ) · e1 ≥ 0
(32)
(u0 − u6 ) · e1 ≥ 0 , (u5 − u0 ) · e2 ≥ 0 , (u0 − u2 ) · e2 ≥ 0
Fig. 7 Counterexample to
the injectivity of B
−1 1
1 −1
−1 1
Due to the symmetries of f , this equation is compatible with the limit condition
p = 0 on ∂H : every line directed by any ei has an intersection of the same length
2 2
with zones labelled by 1 or −1. Moreover, we have ∂i pi ei = f ei = 0, so
i=0 i=0
p lies in ker(B ).
5 Homogenization Issues
This section deals with issues pertaining to the convergence of microscopic models
towards macroscopic ones. Let us make it clear that such convergence is out of
reach in general. We shall rather describe a general framework to address these
issues, and establish some convergence results in very particular situations, in the
case when the microscopic situation is structured. The idea is the following: we
start from a macroscopic velocity field, and we span the domain with a sequence of
saturated configurations of spheres of radius tending to 0. At each scale, we project
the field on the feasible set, which contains all those fields which comply with the
non-overlapping constraint.
1
Ũi = U (x)dρ(x) ∀1 ≤ i ≤ n. (37)
|Di | Di
n
1B(xin,δn ) (39)
i=1
Multibody and Macroscopic Impact Laws: A Convex Analysis Standpoint 121
weakly converges to some limit density when n goes to +∞, what are the possible
limits of φn (U ) ? Is it possible to prescribe constraints on the microscopic structure
so that φn converges to some projection operator at the macroscopic level, which
would encode the characteristics of the microscopic structure ?
These questions should be seen as a wide research program which is way beyond
the scope of these notes. We shall restrict ourselves to some short comments, and to
providing a detailed answer in very specific situations (see Sect. 5.2).
First, various sorts of constraints can be expected: isotropic ones like in Sect. 2,
anisotropic ones according to some principal directions like in Sect. 4, or possibly
not linked to any underlying regular structure in the grain configuration. The notion
of local maximal value, already discussed in Sect. 2, is also an issue: consider e.g. a
configuration where a part of the saturated domain is spanned by a square lattice, and
another part is spanned by a triangular mesh. As we shall see below, the projection
operators will actually converge√ towards an operator activated respectively when
ρ = (1 − π/4) and ρ = π/2 3, depending on the local microscopic structure, so
that the maximal density is not defined uniformly over the saturated zone.
In all generality, when there is no reason to assume any regularity/periodicity in
the microscopic structure, one may expect some sort of averaging in the direction
of contacts, with a local constraint on the density based on the so-called Random
Maximal Packing, that is around 0.64 for three-dimensional collections of identical
hard spheres [31]. This may legitimate an isotropic approach like the one presented
in Sect. 2.2, based on a uniform maximal density, and an isotropic constraint on
the velocities. Yet, as extensively described in the literature on granular media (see
Fig. 8 On the left, the construction of a microscopic vector field Ũ from the macroscopic one U ,
in the square configuration. The green lines delimit the Voronoï cells associated to the spheres, the
red arrows are the mean of the vector field on each cell (expanded for a sake of clarity). On the
right, construction of a macroscopic vector field v (black) from a microscopic field ũ (red)
122 F. Bourdin and B. Maury
e.g. [30]), complex force networks are observed within collections of grains, over
scales that go way beyond the microscopic size of the grains. Such observations
advocate for the need to develop macroscopic models which would reflect some
anisotropy at the mesoscopic scale, in the spirit of what is done in the next section
for highly structured configurations.
Remark 5 One could question the choice of using Voronoï Cells instead of defining
the field v to be constant on every sphere, and null elsewhere. The reason is
that we aim at showing strong L2 convergence results, which will not hold for
velocities supported on spheres. For instance, consider the constant field U = ex
for the squared configuration with radius tending to 0. As no constraint is activated,
v is automatically equal to U everywhere, and similarly, ũni = ex for every
1 ≤ i ≤ kn (kn being the number of spheres needed to span the domain). If we
define wn piecewise constant on every sphere equal to 0 elsewhere, there subsists
an irreducible gap
0 0
0U − wn 02 = (1 − π/4) λ() + o(1)
2
π/4 being the proportion of spanned by the spheres for a square lattice.
0 02 kn 2
0 0
0f − fnint 0 ≤ f (x) − fnint (x) dx
2 Di
i=1
≤ 2 = 2.
i∈I Di
124 F. Bourdin and B. Maury
0 02 kn 2
0 int 0 int
0un − fnint 0 = un (x) − fnint (x) dx
2 Di
i=1
kn 2
1
= λ(Di ) (u(y) − f (y)) dy
λ(Di ) Di
i=1
kn
≤ |u(y) − f (y)|2 dy
i=1 Di
= u − f 22
Thus uint 2 2
n converges in L (ω) towards u, and so does vn .
Remark 6 In the previous proof, two ingredients can be identified as essential in the
process of elaborating general homogenization results:
• uint
n ∈ Cn : a field that respects the macroscopic constraint must check the
microscopic constraints once integrated on the Voronoï cells; and reciprocally
the piecewise constant approximation vn of the corrected microscopic field must
satisfy the macroscopic constraint. Thus the macro/micro constraints must be
compatible under the mapping that we defined above.
• uint 2
n needs to converge for the L norm toward u: this is in particular true if
the spheres span the whole saturation area, in the sense that the diameter of the
Voronoï cells tends to 0.
6 Evolution Models
We describe here the evolution problems which are associated to the impact laws
that have been described in the first sections of these notes. Let us first make it
clear that writing an evolution problem associated to the impact laws studied in
Sects. 4 and 5 is irrelevant a priori. Indeed, the assumptions which can be made on
the microscopic structure of a granular medium are instantaneously ruled out as soon
as the medium undergoes any deformation. A macroscopic model respectful of the
current state of the medium in terms of microscopic structure should rely on some
parameters to reflect the local organization of grains, which strongly conditions the
impact law as we detailed in the previous sections. We shall rather present evolution
problems for the microscopic setting, which takes the form of a second-order in time
differential inclusion, and for the macroscopic scale we shall consider the isotropic
setting only (the divergence is nonnegative on the saturated zone).
Multibody and Macroscopic Impact Laws: A Convex Analysis Standpoint 125
d 2r
M + Nr ' f. (43)
dt 2
where Gij is defined by (6). It guarantees that contact forces verify the Law of
Action-Reaction, and that only repulsive forces are exerted (grains do not glue to
each other).
Yet, Inclusion (43) is essentially compatible with all impact laws which do not
violate the Law of Action-Reaction, including some laws which would lead to an
increase of kinetic energy. An impact law of the type (3) has to be prescribed. We
shall now write the full evolution system, in the purely inelastic setting, and with an
explicit involvement of interaction forces. In the dynamic setting, these forces are
generically singular in order to instantaneously change the velocities of the grains,
and we shall represent them by positive measures in time, denoted by M+ (0, T ) .
In the purely inelastic setting, the system writes
d 2r
M =f + pij Gij
dt 2
ij
pij ∈ M+ (0, T ) (45)
) *
supp(pij ) ⊂ t , Dij (r(t)) = 0
u+ = PCr u− .
1 This convention is consistent with the definition of Nr as the Fréchet subdifferential of the
indicatrix function IK of K, which is indeed ∅ outside of K.
126 F. Bourdin and B. Maury
where Cρ is the cone of feasible velocities defined by (10). These equation must
be understood in a weak sense. In particular the pressure p is likely to be very
singular in time, like in the microscopic setting, and the momentum equation is
meant in a distributional sense. Little is known concerning this system, which is
usually written without the impact law (last equation of the system). Note that this
law can be replaced, at least formally, by any law of the type
u+ = u− − (1 + e) u− − PCρ u− ,
with e = 1 for the elastic case. This equation is well-understood in the one-
dimensional setting, see e.g.[6] where particular “sticky-blocks” solutions are built,
Multibody and Macroscopic Impact Laws: A Convex Analysis Standpoint 127
0.5
-0.5
-1
-1.5
-2
1/4 1/2 1 2 4
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1/4 1/2 1 2 4
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1/4 1/2 1 2 4
and can be used to build solutions of the system. This class of solutions corresponds
to the situation where the initial density is the sum of characteristic functions of
segments, each one initially moving at a uniform velocity. Since no forcing term is
involved, segments remains segments, possibly merging to form larger segments,
and the model can be treated exactly according to the microscopic model (45).
Note that this approach, presented in the purely inelastic setting, could be extended
to various impact laws (e ∈ (0, 1]). Note also that, since sticky blocks reproduce
the microscopic setting, the non-uniqueness result which we mentioned obviously
extends to the macroscopic problem, if one accounts for a time-dependent forcing
term.
128 F. Bourdin and B. Maury
u−
PC r u −
PN r u−
u+
Fig. 10 Newton’s craddle: computation of u+ with Moreau’s approach, with initially touching
discs (left), and slightly pulled apart discs (right)
apparent post-collisional velocity affects the sphere on the opposite side only, while
the other spheres (including the hitting one) stay still. A straight application of the
approach we presented (Moreau’s approach) in the elastic setting leads to a fully
different picture, presented in Fig. 10 (left): the hitting sphere is pushed backward
(i.e. rightward), almost as if it had hit a wall (the speed is slightly reduced), while
the rest of the spheres are pushed leftward at a small velocity, in such a way
that total momentum and kinetic energy are conserved. Yet, by considering an
initial situation where grains are slightly pulled apart (initial distances set at an
arbitrary small value), the experimentally observed behavior is recovered, after a
series of quasi-simultaneous binary collisions as illustrated again in Fig. 10 (right).
Similar examples of the high sensitivity of the impact law to the configuration,
possibly inducing significant changes in the future behavior of the system, can
straightforwardly be built for the macroscopic one-dimensional problem, in the
elastic setting.
Appendix
We gather here some well-known theoretical results, and some less classical ones,
on the saddle-point formulation of cone-constrained minimization problems.
Let V be a Hilbert space, and J : V −→ R a continuously differentiable
functional. We denote by DJ (u) ∈ V its differential at u, and by ∇J (u) its
gradient:
Linear Constraints
∇J (u) + ξ = 0 , ξ ∈ K ⊥ .
1
v −→ J (v) = |v − U |2 , (47)
2
for a given U ∈ V .
Proposition 9 Let K = ker B be a linear subspace of V , and u a local minimizer
of J (defined by (47)) over K, the linear functional ξ defined in the previous
proposition lies in B ().
If we assume that B has a closed range, then ξ ∈ B (). If we identify V with
its dual space, considering accordingly that B maps to V , it means that there
exists p ∈ such that
u+B p = U
(48)
Bu =0
then (u, p) verifies (48) if and only if it is a saddle point for L in V × , i.e.
Unilateral Constraints
We now consider the projection of an element on a closed convex cone C. This cone,
like all the cones we shall consider in this section, admits the origin as a pole, i.e.
R+ C ⊂ C. More precisely, U ∈ V being given, we aim at minimizing
1
v −→ J (v) = |v − U |2
2
over C. We denote by N the polar cone to C:
N = C ◦ = {v ∈ V , v|w ≤ 0 ∀w ∈ C} .
u + ξ = U , u = PC U , ξ = PN U , u|ξ = 0.
U − u|w − u ≤ 0 ∀w ∈ K.
Applying this to the closed convex set C, and using the fact that C is a cone, we
have that
The first term is nonnegative by polarity, and so is the second one because u is the
projection of U on C.
Finally, u|ξ ≤ 0 and, since 0 ∈ N,
0 ≥ U − ξ |0 − ξ = − u|ξ ,
so that u|ξ = 0.
Conversely, if U = u + ξ with u ∈ C, ξ ∈ N, u|ξ = 0, then for any w ∈ C
U − u|w − u = ξ |w − u = ξ |w ≤ 0,
C = {v ∈ V , Bv|μ ≤ 0 , ∀μ ∈ + } , (49)
where B ∈ L(V , ), is a Hilbert space identified to its dual, and + is a closed
convex cone in (see Fig. 11).
Remark 8 One may interrogate the motivation for defining a convex cone by means
of another convex cone. This approach will be proven fruitful in many situations
where K is natively described in an implicit way, i.e. as the collection of elements
which verify certain unilateral constraints, whereas + is defined in a explicit way,
like Rd+ in the finite dimensional setting, or as a subset of real functions taking
nonnegative values, so that projecting on + can be computed straightforwardly.
u
U
C Λ− = Λ°+
N h =B p
B Λ+
N = C ◦ = {w ∈ V , w|v ≤ 0 , ∀v ∈ C} = B + .
Since μ goes over a cone, the left hand side inequality implies that h|B μ ≤ 0 for
all μ ∈ + , so that α ≥ 0 and h ∈ C by definition of C. We then have h|w > 0,
which contradicts the fact that h ∈ C, w ∈ N = C o .
Let us now introduce the so-called saddle-point formulation of the projection
problem
u + B p =U
Bu ∈ −
(50)
p ∈ +
Bu|p = 0.
1
L(v, q) = |v − U |2 + Bv|q ,
2
i.e.
Now consider the closed convex cone spanned by this parabola, i.e.
3
+ = conv R+ P R + ey ,
where ey is the unit vector in the direction y. The projection of + on the (x, y)
plane is R×]0, +∞[∪ {(0, 0)}, which is not closed.
Yet, an important family of cones enjoys the property of being linearly mapped
onto a closed set, those are the cones spanned by a finite number of vectors.
Lemma 3 Let V be a Hilbert space, and N a convex cone spanned by a finite
number of vectors:
n 4
N= αi Gi , (α1 , . . . , αn ) ∈ Rn+ .
i=1
Proof We give a full proof of this classical result to enlighten the importance of the
fact that N is spanned by a finite number of vectors. We prove the result by induction
on the number of vectors. For n = 1, the result is obvious. Assume that the property
is true for n ≥ 1, and consider the cone N associated to n + 1 vectors. If the Gi s are
independent, we call W the space spanned by these vectors, and we introduce
n+1
G : α ∈ Rn+1 −→ αi Gi ∈ W.
i=1
This map is invertible, and its reciprocal G−1 is linear and continuous from W to
R . Now consider v = α k Gi converging to v ∈ W . then G−1 v k converges to
n+1 k
n+1
μi Gi = 0. (51)
i=1
n+1
α k Gi −→ v.
i=1
which lies in the cone spanned by the n vectors (Gi )i =i0 (by the induction
hypothesis), so it is in N.
We now address some theoretical issues related to the description of solution
sets for the pressure p ∈ for equations of the type (50). Like in the case of
equality constraints (Proposition 10), the solution p is unique a soon as B is onto,
and uniqueness is lost whenever the range of B is not dense in . Yet, in the
finite dimensional setting, the solution set can be proven to be bounded under some
conditions which are typically met for impact laws in granular media. The approach
is based on the notion of asymptotic cone (see e.g. [7]):
136 F. Bourdin and B. Maury
which does not depend on the choice of u ∈ K, is called the asymptotic cone of K
(see e.g. [18]).
Proposition 13 Let V be a Hibert space, and K ⊂ V a closed convex subset. For
→
−
any u ∈ K, the asymptotic cone K is the set of directions h such that the half line
u + R+ h is contained in K.
→
− →
−
Proof If u + R+ h ⊂ K, then h is in K by definition. Conversely, if h ∈ K , h
writes t (v − u) for some t > 0, with v = u + h/t ∈ K, so that u + R+ h ⊂ K.
This notion provides a criterium to identify bounded convex sets (in the finite
dimensional setting).
Proposition 14 Let V be a finite dimensional Hibert space, and K ⊂ V a closed
convex subset which contains 0. Then
→
−
K is bounded ⇐⇒ K = {0} .
→
−
Proof If K is bounded by M, then tK is bounded by tM, so K contains only
0. Conversely, if K is not bounded, there exists a sequence (un ) in K, with
|un | → +∞. Let u be any element of K. Since V is finite-dimensional the unit
sphere is compact, and one can extract a subsequence from (un −u)/|un − u|, which
converges to some v ∈ H , with |v| = 1. Now consider t > 0, and θn = t/|un − u|.
By convexity of K, it holds that
un − u
(1 − μn )u + μn un = u + μn (un − u) = u + t ∈ K.
|un − u|
→
−
Since K is closed, having n go to infinity yields u + tv ∈ K. As a consequence K
contains the nonzero vector v.
Note that the finite dimension is crucial in the previous proposition. Con-
sider for example the case where V = 2 and K is the hypercube
{x = (xn ) ∈ V , 0 ≤ xn ≤ 1}. The closed convex set K does not contain any
half-line, while being not bounded.
We may now establish the main property
Proposition 15 Let V be a finite dimensional Hilbert space, C ⊂ V a closed
convex cone defined by (49), U ∈ V , and u the projection of U on C. We assume
that B (+ ) is closed, so that (by Proposition 12) there exists p ∈ + such that
Multibody and Macroscopic Impact Laws: A Convex Analysis Standpoint 137
is bounded.
Proof The solution set can be written
S = (p + ker B ) ∩ + ,
→
−
it is a closed convex set. Consider h ∈ S . By Proposition 13, the half line p + R+ h
is contained in S ⊂ p + ker B , which implies h ∈ ker B . Since S is also contained
in the cone + is a cone, it also implies that p/t + h ∈ + , for any t > 0, which
yields, by having t go to 0, h ∈ + . To sum up, h ∈ ker B ∩ + = {0}. We proved
→
−
that S = {0}, therefore (by Proposition 14), S is bounded.
We end this appendix by defining a notion which is relevant to classify problems
according to some sort of abstract maximum principle. In the context of collisions,
the issue can be formulated as follows: if the pre-collisional velocity fields tends to
violate all the constraints, it can be expected that all contacts will be active, i.e. that
all interaction forces will be positive, and the unilateral constraints turn out to be
equalities. It is an essential tool to exhibit a Poisson like problem for the pressure in
impact laws (see the end of Sect. 3). We shall see that this intuitive fact is sometimes
ruled out, when a general property is not verified.
Definition 3 (Abstract Maximum Principle) Let C be a closed convex cone in V ,
associated to B ∈ L(V , ) through Eq. (49). Like in proposition 12, we assume that
B () and B (+ ) are closed, so that, for any U ∈ V , the system (50) admits at
least a solution (u, p) ∈ V × + , where u is the projection of U on C. We say that
the couple (B, + ) (which encodes the structure of the projection problem) verifies
the maximum principle if
BB p = BU.
Proof Let us consider the problem with an equality constraint, i.e. u ∈ ker B. We
denote by u the projection of U on C. From the maximum principle there exists
p ∈ λ+ such that BB p = BU , which implies that u = U − B p is in K = ker B,
so that u = PK U by Proposition 9. Since Bu = 0 ∈ − and p ∈ + , the couple
(u, p) is also a solution to the problem with unilateral constraints (50), which ends
the proof.
138 F. Bourdin and B. Maury
References
1. Alder, B.J., Wainwright, T.E.: Studies in molecular dynamics. I. General method. J. Chem.
Phys. 31(2), 459 (1959)
2. Ames, W.F.: Numerical Methods for Partial Differential Equations. Academic Press, New York
(2014)
3. Ballard, P.: The dynamics of discrete mechanical systems with perfect unilateral constraints.
Arch. Ration. Mech. Anal. 154, 199–274 (2000)
4. Berthelin, F.: Existence and weak stability for a pressureless model with unilateral constraint.
Math. Models Methods Appl. Sci. 12(2), 249–272 (2002)
5. Berthelin, F.: Theoretical study of a multi-dimensional pressureless model with unilateral
constraint. SIAM J. Math. Anal. 49(3), 2287–2320 (2017)
6. Bouchut, F., Brenier, Y., Cortes, J., Ripoll, J.-F.: A hierarchy of models for two-phase flows. J.
Nonlinear Sci. 10(6), 639–660 (2000)
7. Bourbaki, N.: Espaces Vectoriels Topologiques. Masson, Paris (1981)
8. Brezis, H.: Functional Analysis, Sobolev Spaces and Partial Differential Equations. Springer,
Berlin (2011)
9. Degond, P., Hua, J., Navoret, L.: Numerical simulations of the Euler system with congestion
constraint. J. Comput. Phys. 230(22), 8057–8088 (2011)
10. Degond, P., Minakowski, P., Navoret, L., Zatorska, E.: Finite volume approximations of the
Euler system with variable congestion. Comput. Fluids 169, 23–39 (2017). https://doi.org/10.
1016/j.compfluid.2017.09.007
11. Ekeland, I., Temam, R.: Analyse convexe et problèmes variationnels. Dunod
12. Falk, R.S.: Error estimates for the approximation of a class of variational inequalities. Math.
Comput. 28(128), 963–971 1974
13. Frémond, M.: Non-Smooth Thermomechanics. Springer, Berlin (2002)
14. Golse, F., Saint-Raymond, L.: The Navier–Stokes limit of the Boltzmann equation for bounded
collision kernels. Invent. Math. 155, 81–161 (2004)
15. Joly, P.: Some trace theorems in anisotripic Sobolev spaces. SIAM J. Math Anal 23(3), 799–
819 (1994)
16. Liu, C., Zhao, Z., Brogliato, B.: Frictionless multiple impacts in multibody systems: part I.
Theoretical framework. Proc. R. Soc. A Math. Phys. Eng. Sci. 464(2100), 3193–3211 (2008)
17. Maury, B., A time-stepping scheme for inelastic collisions. Numer. Math. 102, 649–679 (2006)
18. Maury, B., Analyse fonctionnelle, exercices et problèmes corrigés. Ellipses, Paris (2004)
19. Maury, B., Preux, A.: Pressureless Euler equations with maximal density constraint: a time-
splitting scheme. Topol. Optim. Optimal Transp. Appl. Sci. 17, 333 (2017)
20. Maury, B., Roudneff-Chupin, A., Santambrogio, F.: A macroscopic Crowd motion model of
the gradient-flow type. Math. Models Methods Appl. Sci. 20(10), 1787–1821 (2010)
21. Maury, B., Roudneff-Chupin, A., Santambrogio, F., Venel, J.: Handling congestion in Crowd
motion modeling. Netw. Heterog. Media 6(3), 485–519 2011
22. Maz’ya, V.: Sobolev Spaces: With Applications to Elliptic Partial Differential Equations,
Grundlehren der mathematischen Wissenschaften, vol. 342. Springer, Berlin (2011)
23. Moreau, J.-J.: Décomposition orthogonale d’un espace hilbertien selon deux cônes mutuelle-
ment polaires. C. R. Acad. Sci. Paris, 255, 238–240 (1962)
24. Moreau, J.J.: Some numerical methods in multibody dynamics: application to granular
materials. Eur. J. Mech. A Solids 13(4), 93–114 (1994)
25. Nguyen, N.S., Brogliato, B.: Comparisons of multiple-impact laws for multibody systems:
Moreau’s law, binary impacts, and the LZB approach. In: Leine R., Acary V., Brüls O. (eds.)
Advanced Topics in Nonsmooth Dynamics. Springer, Cham (2018)
26. Perrin, C., Westdickenberg, M.: One-dimensional granular system with memory effects. SIAM
J. Math. Anal. 50(6), 5921–5946 (2018)
27. Ristow, G.: Simulating granular flow with molecular dynamics. J. Phys. I EDP Sci. 2(5), 649–
662 (1992)
Multibody and Macroscopic Impact Laws: A Convex Analysis Standpoint 139
28. Schatzmann, M.: A class of nonlinear differential equations of second order in time. Nonlinear
Anal. Theory Methods Appl. 2, 355–373 (1978)
29. Radjaï, F., Dubois, F., (eds.): Discrete-element Modeling of Granular Materials. Wiley, London
(2011)
30. Radjai, F., Roux, S., Moreau, J.-J.: Contact forces in a granular packing. Chaos: Interdiscip. J.
Nonlinear Sci. 9, 544–544 (1999)
31. Torquato, S., Stillinger, F.H.: Jammed Hard-particle packings: from Kepler to Bernal and
beyond. Rev. Mod. Phys. 82, 2633 (2010)
32. Torquato, S., Truskett, T.M., Debenedetti, P.G.: Is random close packing of spheres well
defined? Phys. Rev. Lett. 84, 2064–2067 (2000)
33. Troianiello, G.M.: Elliptic Differential Equations and Obstacle Problems, University Series in
Mathematics. Springer, Berlin (1987)
34. Villani, C.: Limites hydrodynamiques de l’équation de Boltzmann [d’après C. Bardos, F. Golse,
D. Levermore, P.-L. Lions, N. Masmoudi, N., L. Saint-Raymond]. Séminaire Bourbaki, vol.
2000–2001, Exp. 893
An Introduction to Uncertainty
Quantification for Kinetic Equations
and Related Problems
Lorenzo Pareschi
1 Introduction
L. Pareschi ()
Department of Mathematics and Computer Science, University of Ferrara, Ferrara, Italy
e-mail: lorenzo.pareschi@unife.it
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 141
G. Albi et al. (eds.), Trails in Kinetic Theory, SEMA SIMAI Springer Series 25,
https://doi.org/10.1007/978-3-030-67104-4_5
142 L. Pareschi
The recent growth of interest in UQ for PDEs can be traced back mainly to three
factors: widespread availability of data resulting from advances in technology, the
increased development of high-performance computing and the construction and
Uncertainty Quantification for Kinetic Equation and Related Problems 143
Uncertain Statistics
Statistics
solution of about
about
PDE the PDE uncertain
uncertain
and post- outputs of
inputs
processing interest
analysis of new algorithms for solving differential equations with random inputs.
In presence of uncertainties it becomes necessary to quantify these effects on the
solution of the PDE, or on any quantity of interest (a quantity that depends on
the solution of the PDE for which we want to know some statistical information),
derived from the solution. The complete UQ task then consists of determining
information about the uncertainty in an output of interest that depends on the
solution of a PDE, given information about the uncertainty in the inputs of the PDE
(see Fig. 1).
given the (eventually uncertain) initial data u(z, x, 0) = u0 (z, x), x ∈ Rdx , and
where the terms F, B and D depend on the random parameters.
A realization of a solution of the PDE is a solution obtained for a specific
choice of the random parameters. One instead wants to obtain statistical information
on a quantity of interest e.g., expected values, variances, standard deviations,
covariances, higher statistical moments, etc. Therefore, multiple solutions of the
PDE are necessary in order to achieve such information.
144 L. Pareschi
where p(z) is the probability density function (PDF) of the input parameters.
Below we give some examples of quantities of interests.
Example 1
(i)
1/p
F (u) = uLp (Rdx ) = |u(z, x, t)|p dx , "(w) = w
Rdx
will give as quantity of interest E[uLp (Rdx ) ] the expected value of the Lp -
norm.
(ii)
applied to F (u1 ) and F (u2 ), where u1 and u2 are two solutions of the PDE
gives Cov(F (u1 ), F (u2 )) the covariance.
One of the main challenges for numerical methods, is that the computational
cost associated with UQ increases with the number of parameters used to model the
uncertainty (curse of dimensionality). This is a general problem but it is particularly
relevant for kinetic equations where the dimension of the phase-space is very high.
We can follow two main strategies to alleviate this problem:
• one can try to use relatively few solutions of the PDE and replace the PDE
with a surrogate, low-fidelity, model which is much cheaper to solve. Correlation
between the two models may then be used in a control variate setting.
• for smooth solutions one can design methods which permit an accurate evaluation
of E["(F (u))] using few quadrature points obtained from stochastic orthogonal
polynomials with respect to the PDF.
Uncertainty Quantification for Kinetic Equation and Related Problems 145
We have tacitly assumed that we know p(z), the PDF of the input parameters. In
practice, one usually does not know much about the statistics of the input variables
and need to deal with the corresponding stochastic inverse problem [59].
Many methods have been devised in the literature for approximating statistics of
quantities of interest. We summarize shortly some of the main methods below (see
[4, 22–24, 32, 37, 55, 65] for recent monographs and surveys)
• Monte Carlo sampling: one generates independent realizations of random inputs
based on their PDF (which may be known or not and not necessarily smooth).
For each realization the problem is deterministic and can be solved by standard
methods in a non intrusive way. The advantage is its simplicity but on the other
hand it implies a slow convergence and fluctuations in the solution statistics [6,
25].
• Multi-fidelity, multi-level methods: Accelerate Monte Carlo sampling methods
by using multiple surrogate models with different levels of fidelity in a control
variate setting [15, 16, 41, 53, 54]. Low-fidelity models may also be obtained
from a multi-level hierarchy of numerical discretizations in the phase space [20,
22, 30, 44].
• Stochastic-Galerkin: solutions are expressed as orthogonal polynomials of the
random inputs accordingly to their PDF. Spectral convergence for smooth
solutions in the random space [40, 65]. They require smoothness and knowledge
of the PDF. The intrusive nature may lead to the loss of physical properties and
suffers of the curse of dimensionality [27, 61].
• Other methods: moment methods where the unknowns are the moments of the
solution, stochastic collocation methods based on orthogonal polynomials but
selecting the quadrature points [47, 67].
Let us quickly describe the simple Monte Carlo sampling method. Assume
u(z, x, t), x ∈ R, solution of a PDE with uncertainty only in the initial data
u0 (z, x), z ∈ ⊂ Rdz . The method does not depend on the particular solver used
for the PDE or the dimension dz , and consists of three main steps.
Algorithm (Simple Monte Carlo Method)
1. Sampling: Sample M independent identically distributed (i.i.d.) initial data
uk,0 , k = 1, . . . , M
from the random initial data u0 and approximate on a grid x to get uk,0
x .
146 L. Pareschi
uk,n
x , k = 1, . . . , M.
1
M
E["(F (u(·, t n )))] ≈ EM ["(F (unx ))] := "(F (uk,n
x )).
M
k=1
In the sequel we will consider F (u) = u and "(w) = w, namely the quantity
of interest is E[u]. Let su recall that, from the central limit theorem, the root mean
square error satisfies [6, 43]
!1/2
E (E[u] − EM [u])2 = Var(u)1/2 M −1/2 . (4)
where p ≥ 1 and to keep notations simple we ignored the time discretization error.
Let us define the following norms
0 02 !1/2
(i) E[u(·, t n )] − EM [unx ]Lp (R;L2 ()) := E 0E[u(·, t n )] − EM [unx ]0Lp (R)
0 ! 0
0 n )] − E [un ] 2 1/2 0
0
(ii) E[u(·, t n )] − EM [unx ]L2 (;Lp (R)) := 0
0E E[u(·, t M x 0
Lp (R)
Note that, by the Jensen inequality for any convex function φ we have
M = O(x −2p ).
Therefore, for a method of order p changing the grid from x to x/2 requires to
multiply the number of samples by a factor 22p .
To describe the method, let us assume that the solution of the PDE, u(z, x, t), x ∈ R,
has an uncertain initial data which depends on a one-dimensional random variable
z ∈ ⊂ R.
The method is based on the construction of a set of orthogonal polynomials
{m (z)}M
m=0 , of degree less or equal to M, orthonormal with respect to the
probability density function p(z) [55, 65]
M
uM (z, x, t) = ûm (x, t)m (z), (8)
m=0
which can be evaluated by the same quadrature (Gaussian) used to compute ûm .
In case the quantity of interest is the expectation of the solution we have
M
E[uM ] = uM (z, x, t)p(z) dz = ûm (x, t)E[m (·)] = û0 ,
m=0
M
M
Var(uM ) = E[u2M ] − E[uM ]2 = ûm ûn E[m n ] − û20 = û2m − û20 .
m,n=0 m=0
Stochastic Galerkin approximation in the field of random PDEs are better known
under the name of generalized polynomial chaos (gPC). The solution of the PDE,
is obtained by standard Galerkin approach, first replacing u with uM and then
projecting the PDE to the space generated by {m (z)}M
m=0 .
Let us consider a general PDE in the form
M
∂t ûh (x, t) = ûm ûn E[Q(m , n ) h ], h = 0, . . . , M.
m,n=0
A general problem, is the loss of physical properties (like positivity of the solution
or other invariants) due to the approximation in the orthogonal polynomial space.
Uncertainty Quantification for Kinetic Equation and Related Problems 149
E[(uM (·, x, t) − u(·, x, t))2 ]1/2 = uM (·, x, t) − u(·, x, t)L2 ()
u(·, x, t)H r () (10)
≤C .
Mr
For analytic functions, spectral convergence becomes exponential convergence.
Therefore, we must equilibrate an error relation of the type
and then very small values of M are sufficient to balance the errors in the method.
For multi-dimensional random spaces, assuming the same degree M in each
dimension, the number of degrees of freedom of the polynomial space is
(dz + M)!
K= .
dz !M!
Let us focus our attention on the specific case of kinetic equations of Boltzmann and
mean-field type. More precisely, we consider kinetic equations of the general form
[10, 14, 32, 64]
1
∂t f (z, x, v, t) + v · ∇x f (z, x, v, t) = Q(f, f )(z, x, v, t), (x, v) ∈ Rdx × Rdv
ε
(11)
where ε > 0 is the Knudsen number and z ∈ ⊆ Rdz is a random vector. The
particular structure of the interaction term Q(f, f ) depends on the kinetic model
considered.
Well know examples are given by the Boltzmann equation
v + v∗ |v − v∗ | v + v∗ |v − v∗ |
v = + ω, v∗ = − ω, (13)
2 2 2 2
or by mean-field Vlasov-Fokker–Planck type models
In the classical case of rarefied gas dynamic, we have the collision invariants
1 1
ρ= f dv, u= vf dv, T = (v − u)2 f dv, (19)
Rdv ρ Rdv dv ρ Rdv
are the density, mean velocity and temperature of the gas depending on (z, x, t).
Integrating the Boltzmann equation against the collision invariants φ(v) yields
∂t f (z, x, v, t)φ(v) dv + ∇x · vf (z, x, v, t)φ(v) dv = 0, φ(v) = 1, v, |v|2 .
R3 R3
Uncertainty Quantification for Kinetic Equation and Related Problems 151
These equations descrive the balance of mass, momentum and energy. However, the
system is not closed since it involves higher order moments of f .
The simplest way to find an approximate closure is to assume f ≈ M to obtain
the compressible Euler equations with random inputs
∂t ρ(z, x, t) + ∇x · (ρu)(z, x, t) = 0
∂t (ρu)(z, x, t) + ∇x · (ρu ⊗ u + p)(z, x, t) = 0 (20)
1
∂t E(z, x, t) + ∇x · (Eu + pu)(z, x, t) = 0, p = ρT = (2E − ρu2 ).
dv
Other closure strategies, like the Navier–Stokes approach, lead to more accurate
macroscopic approximations of the moment system.
Two peculiar aspects of kinetic equations are the high dimensionality and the struc-
tural properties (nonnegativity of the solution, conservation of physical quantities,
. . .) which represent a challenge for numerical methods. These difficulties are even
more striking in the context of UQ. We summarize below the main advantages and
drawbacks of MC and SG methods.
MC Methods for UQ
1. easy non intrusive application as they rely on existing numerical solvers.
Efficiency and structural properties are inherited from the existing solvers.
2. lower impact on the curse of dimensionality. Easy to parallelize and convergence
is independent of the dimension of the random space.
3. can be applied even if the PDF of the random vector is not known or lacks of
regularity.
4. convergence behavior is slow.
SG Methods for UQ
1. application is intrusive and problem dependent. Hard to combine with stochastic
methods (phase space) and structural properties often are lost.
2. suffer the curse of dimensionality, in particular for nonlinear problems, and
special techniques are required to reduce the computational cost.
3. require knowledge and smoothness of the PDF.
4. can achieve high accuracy, spectral accuracy for smooth solutions, in the random
space.
In the next sections we will focus on some of the recent progress on MC methods
based on multi-fidelity techniques and on stochastic Galerkin methods using micro-
macro decomposition and hybrid approaches.
152 L. Pareschi
We assume that the equation has been discretized by a deterministic solver on a grid
v and x, which satisfies [17, 60]
f (·, t n ) − fx,v
n
B(R) ≤ C x p + v q , (21)
p
f (z, ·, t)Lp (D×Rdv ) = |f (z, x, v, t)|p (1 + |v|s ) dv dx. (22)
s D×Rdv
For the Monte Carlo method therefore we have the error estimate
∂f
= Q(f, f ), (24)
∂t
where f = f (z, v, t) with initial data f (z, v, 0) = f0 (z, v).
Under suitable assumptions [62, 63], f (z, v, t) → f ∞ (z, v) exponentially as
t → ∞, where f ∞ (z, v) is the Maxwellian equilibrium state s.t. Q(f ∞ , f ∞ ) = 0.
We denote the moments as
Let us recall that a Monte Carlo estimator for E[f ] based on M samples gives
Since f ∞ (z, v) is known, we can assume E[f ∞ ](v) is evaluated with a negligible
error and use the Monte Carlo estimator only on E[g] to get
For a time dependent low-fidelity model f˜(z, v, t) given M samples f k (v, t), k =
1, . . . , M we define
E[f ](v, t) ≈ EM λ
[f ](v, t) := EM [f ](v, t) − λ EM [f˜](v, t) − E[f˜](v, t)
1 k 1 ˜k
M M
= f (v, t) − λ f (v, t) − f̃(v, t) ,
M M
k=1 k=1
(29)
154 L. Pareschi
∂Var(f λ )
= 2λVar(f˜) − 2Cov(f, f˜) = 0.
∂λ
As a consequence we have the following result.
Cov(f, f˜)
Proposition 1 The quantity λ∗ = minimizes Var(f λ ) at (v, t) and gives
Var(f˜)
λ∗ Cov(f, f˜)2
Var(f )= 1− Var(f ) = (1 − ρf,
2
f˜
)Var(f ), (31)
Var(f˜)Var(f )
∗
lim λ∗ (v, t) = 1, lim Var(f λ )(v, t) = 0 ∀ v ∈ Rdv . (32)
t →∞ t →∞
data f0 .
3. Solving the control variate: Compute the control variate f˜v n at time t n by a
4. Solving: For each realization f0k , k = 1, . . . , M the kinetic equation and the
control variate are solved by the deterministic schemes. Denote the solutions at
k,n
time t n by fv , and f˜v
k,n
, k = 1, . . . , M.
n
CovM (fv , f˜v
n
)
5. Estimating: Estimate λ∗ using the M samples as λ∗,n M = .
˜n
VarM (fv ))
Compute the expected value of the random solution as
1 k,n 1 ˜k,n
M M
λ∗ n
ẼM [fv ] = fv − λ∗,n
M fv − f̃v .
n
M M
k=1 k=1
∗ ∗
(33)
+EMλ
[f ](·, t n ) − EM
λ
[fv
n
]B(R;L2 ())
≤ C σf λ∗ M −1/2 + v q ,
where σf λ∗ = (1 − ρ 2 ˜ )1/2Var(f )1/2 B(R) . The statistical error depends on the
f,f
correlation between f and f˜. Since ρf,f˜ → 1 as t → ∞ the statistical error will
vanish for large times.
In Fig. 2 we report the results for the space homogeneous Boltzmann equation
with uncertain initial data using various control variates and values of λ. In Fig. 3
we report the time evolution of the optimal value function λ in the case of a control
variate approach based on the BGK approximation. The initial condition is a two
bumps problem with uncertainty
ρ0 |v − (2 + sz)|2 |v + (1 + sz)|2
f0 (z, v) = exp − + exp − (34)
2π σ σ
with s = 0.2, ρ0 = 0.125, σ = 0.5 and z uniform in [0, 1]. The deterministic
solver adopted for the Boltzmann equation is the fast spectral method [17, 45]
and the discretization parameters are such that the stochastic error dominates the
computation (see [15] for more details). We can see from the computations that
156 L. Pareschi
Fig. 2 Homogeneous relaxation. Error in E[f ] using various control variate strategies. Left M =
10, right M = 100
1
1.1
1
0.9999
0.9
0.8 0.9998
0.7
0.9997
0.6
0.5
0.9996
0.4
0.3 0.9995
Fig. 3 Optimal λ∗ (v, t) for the MSCV method based on a BGK control variate at t = 10 (left)
and t = 50 (right)
with the optimal method based on the BGK model we can gain almost two digits of
precision for the same computational cost. Note that, to divide the MC error by a
factor 100 we need to multiply the number of samples by 10,000!
Consider now, a general space non homogeneous kinetic equation with random
inputs
1
∂t f (z, x, v, t) + v · ∇x f (z, x, v, t) = Q(f, f )(z, x, v, t). (35)
ε
Uncertainty Quantification for Kinetic Equation and Related Problems 157
The fundamental difference between the space homogeneous and the space non
homogeneous case, is that now the variance of
will not vanish asymptotically in time, unless the kinetic equation is close to the
surrogate model (fluid regime), namely for small values of the Knudsen number.
Cov(f, f˜)
Proposition 2 The quantity λ∗ = minimizes Var(f λ ) at (x, v, t) and
Var(f˜)
gives
∗ 2
Var(f λ ) = (1 − ρf, f˜
)Var(f ), (37)
where ρf,f˜ ∈ [−1, 1] is the correlation coefficient between f and f˜. In addition,
we have
∗
lim λ∗ (x, v, t) = 1, lim Var(f λ )(x, v, t) = 0 ∀ (x, v) ∈ Rdx × Rdv .
ε→0 ε→0
(38)
Contrary to the space homogeneous case, one cannot ignore the computational
cost of solving the macroscopic fluid equations or the BGK model, although
considerably smaller than that of the Boltzmann collision operator. Using ME M
samples for the control variate, we get the error estimate
∗
E[f ](·, t n ) − EM,M
λ
E
[fx,v
n
]B(R;L2 ())
(39)
' (
−1/2
≤ C σf λ∗ M −1/2 + τf λ∗ ME + x p + v q
where σf λ∗ = (1 − ρ 2 ˜ )1/2 Var(f )1/2 B(R) , τf λ∗ = ρf,f˜ Var(f )1/2 B(R) .
f,f
158 L. Pareschi
Again the statistical error depends on the correlation between f and f˜. In this
case, ρf,f˜ → 1 as ε → 0, therefore the statistical error will depend only on the fine
scale sampling in the fluid limit.
We point out that, the optimal value of λ depends on the quantity of interest and
in practice does not depend on (x, v, t) unless one is interested in the details of the
distribution function. For a general moment mφ (f ), the optimal value depends on
(x, t) and is given by
Cov(mφ (f ), mφ (f˜))
λ∗ =
Var(mφ (f˜))
Var(EM [f ]) = M −1 Var(f ).
Therefore, taking into account the number of effective samples in the minimization
process and using the independence of the estimators EM [·] and EME [·] we get
λ
Var(EM,ME
[f ]) = M −1 Var(f − λf˜) + ME−1 Var(λf˜)
= M −1 Var(f ) − 2λCov(f, f˜) + (M −1 + ME−1 )λ2 Var(f˜).
Minimizing with respect to λ yields the effective optimal value λ̃∗ which reads
ME Cov(f, f˜)
λ̃∗ = λ∗ , λ∗ = . (40)
M + ME Var(f˜)
Let Cost(·) denote computational cost to compute the solution of a given model for
a fixed value of the random parameter. The total cost is MCost(f ) + ME Cost(f˜).
Fixing a given cost for both models MCost(f ) = ME Cost(f˜), we obtain
Cost(f )
λ̃∗ = λ∗ .
Cost(f ) + Cost(f˜)
Fig. 4 Sod test with uncertain initial data. E[T ] and confidence bands at t = 0.875 (top). Left
ε = 10−2 , right ε = 10−3 . Error in E[T ] with M = 10 and ε = 10−2 (bottom). Left ME = 103 ,
right ME = 104 . Here Nx = 100, Nv = 32
Since we are interested only in the accuracy in the random variable, the numerical
parameters of the deterministic discretization Nv , Nx and t have been selected
such that the deterministic error is smaller than the stochastic one (see [15] for
further details). In Fig. 4 (top), we report the expectation of the solution at the
final time together with the confidence bands. In the same figure (bottom) we also
report the various errors using different control variates for the expected value of
the temperature as a function of time. The optimal values of λ∗ (x, t) have been
computed with respect to the temperature. The improvements obtained by the
various control variates are evident and, as expected, becomes particularly striking
close to fluid regimes.
Next we consider in the same setting the sudden heating problem with uncertain
boundary condition. Initial condition is a local equilibrium with ρ0 = 1, u0 = 0,
160 L. Pareschi
Fig. 5 Sudden heating with uncertain boundary condition. E[T ] and confidence bands at t = 0.9
(top). Left ε = 10−2 , right ε = 10−3 . Error in E[T ] with M = 10 and ε = 10−3 (bottom). Left
ME = 103 , right ME = 104 . Here Nx = 100, Nv = 32
with s = 0.2, z uniform in [0, 1]. The results are summarized in Fig. 5, where we
report the expectation of the temperature at the final time and the various errors
using different control variates. In this case, due to the source of uncertainty at the
boundary there is no relevant difference between the Euler and BGK control variates
and the results is less sensitive to the choice of the Knudsen number.
The bi-fidelity approach developed in the previous sections is fully general and
accordingly to the particular kinetic model studied one can select a suitable
approximated solution as control variate which acts at a given scale. In this section
we extend the methodology to the use of several approximated solutions as control
Uncertainty Quantification for Kinetic Equation and Related Problems 161
variates with the aim to further improve the variance reduction properties of MSCV
methods (see Fig. 6).
Given f˜1 , . . . , f˜L approximations of f (z, v, t) we can consider the random
variable
L
f λ1 ,...,λL (z, v, t) = f (z, v, t) − λh (f˜h (z, v, t) − f̃h (v, t)), (43)
h=1
L
Var(f λ1 ,...,λL ) = Var(f ) + λ2h Var(f˜h )
h=1
⎛ ⎞
L
⎜
L
⎟
+2 λh ⎝ λk Cov(f˜h , f˜k ) − Cov(f, f˜h )⎠ ,
h=1 k=1
k =h
or in vector form
∗ = C −1 b, (45)
In fact, the optimal values λ∗h , h = 1, . . . , L are found by equating to zero the
partial derivatives with respect to λh . This corresponds to the linear system
L
Cov(f, f˜h ) = λk Cov(f˜h , f˜k ), h = 1, . . . , L, (47)
k=1
or equivalently C = b.
Example 2 Let us consider the case L = 2, where f˜1 = f0 and f˜2 = f ∞ .
The optimal values λ∗1 and λ∗2 are readily found and are given by
λ∗ ,λ∗2
EM1 (v, t) =EM [f ](v, t) − λ∗1 (EM [f0 ](v) − f0 (v))
(49)
− λ∗2 EM [f ∞ ](v) − f∞ (v) .
lim λ∗ = 0, lim λ∗ = 1,
t →∞ 1 t →∞ 2
λ∗ ,λ∗2
lim EM1 (v, t) = f∞ (v).
t →∞
In Fig. 7 we report the results obtained for the homogeneous relaxation problem with
uncertain initial data. Compared to the optimal BGK control variate, at the same
computational cost, using the estimator based on two control variates described
above we can gain one additional digit of accuracy.
10-2
10-3
10-4
L2 error E[f]
10-5
10-6
10-7
MC
10-8
MSCV
MSCV2
10-9
0 1 2 3 4 5 6 7 8 9 10
time
Fig. 7 Homogeneous relaxation. Error for E[f ] over time with M = 10 for the MSCV method
based on BGK and the MSCV2 method based on the two control variates f0 and f ∞
Next, to estimate E[fL ] we use ML−1 ML samples with fL−1 as control variate
ˆ
EL [f ] := EML [f ] − λ̂L EML [fL ] − EML−1 [fL ]
+ λ̂L−1 EML−1 [fL−1 ] − EML−2 [fL−1 ]
(50)
...
+ λ̂1 EM1 [f1 ] − EM0 [f1 ] . . . .
ˆ
L
EL [f ] = EML [fL+1 ] − λh (EMh [fh ] − EMh−1 [fh ])
h=1
(51)
L
= λ1 EM0 [f1 ] + (λh+1 EMh [fh+1 ] − λh EMh [fh ]),
h=1
:
L
λh = λ̂j , h = 1, . . . , L, λL+1 = 1, fL+1 = f. (52)
j =h
ˆ
Var(EL [f ]) = λ21 M0−1 Var(f1 )
(53)
L ' (
+ Mh−1 λ2h+1 Var(fh+1 ) + λ2h Var(fh ) − 2λh+1 λh Cov(fh+1 , fh ) .
h=1
−1
Mh−1 {λh Var(fh ) − λh−1 Cov(fh , fh−1 )}
(54)
+ Mh−1 {λh Var(fh ) − λh+1 Cov(fh+1 , fh )} = 0,
:
L
Cov(fj +1 , fj )
λ∗h = λ̂∗j , λ̂∗j = . (55)
Var(fj )
j =h
Uncertainty Quantification for Kinetic Equation and Related Problems 165
ˆ∗
E[f ](·, t n ) − EL [fv
n
]B(R;L2 ())
(56)
L
−1/2 −1/2
≤C ξh σh Mh + ξ0 M0 + v q
h=1
0 1/2 0
0 0
where σh = 0 0 1 − ρ 2
fh ,fh−1 Var(fh ) 1/2 0
0 , τh = ρfh ,fh−1 Var(fh )1/2 B(R)
; B(R)
and ξh = L j =h+1 τj .
If the control variates share the same behavior as t → ∞, namely fh → f ∞
for h = 1, . . . , L, we get ρf2h ,fh−1 → 1 as t → ∞ the statistical error depends
only on the finest level of samples M0 . Similar considerations hold in the space non
homogeneous case as ε → 0.
In Fig. 8 the results obtained in the case of the sudden heating problem (42)
using a three models hierarchy based on the Euler system, the BGK model and the
full Boltzmann equation are reported.
There is a close link between multi-fidelity methods and multi-level Monte Carlo
methods. Let us consider as control variates a hierarchy of discretizations of the
Fig. 8 Sudden heating with uncertain boundary condition. Error for E[T ] over time for ε = 10−2
(left) and ε = 10−3 (right). MSCV method based on BGK and MSCVH2 based on BGK and Euler.
M2 = 10 for Boltzmann, M1 = 102 for BGK and M0 = 104 for Euler
166 L. Pareschi
kinetic equation. For example, in the homogeneous case, with a cartesian grid we
take
where v1 is the mesh width for the coarsest resolution, which corresponds to the
solution with the lowest level of fidelity. Our full model is, therefore, represented
by the fine scale solution obtained for vL . The hierarchy of numerical solutions
fh (z, v, t), h = 1, . . . , L, at time t with mesh vh represents the setting for the
multi-level control variate estimators.
In particular, fixing all λh = 1, h = 1, . . . , L, we get the classical Multi-level
MC estimator [22]
L
EL1 [f ](v, t) = EM0 [f1 ] + (EMh [fh+1 − fh ]), (57)
h=1
1.1 7 10-3
MLMC MLMC
1 quasi-MLMC quasi-MLMC
optimal-MLMC 6 optimal-MLMC
0.9 Ref
0.8 5
0.7
4
0.6
3
0.5
0.4 2
0.3
1
0.2
0.1 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fig. 9 Multilevel MC for BGK. E[ρ] (left) and error in space (right) for the various MLMC using
L = 3, M0 = 320, M1 = 80, M2 = 20 with xh = 21−h (x1 ), h = 1, 2, 3, x1 = 0.1
Uncertainty Quantification for Kinetic Equation and Related Problems 167
M
fM (z, v, t) = fˆm (v, t)m (z), (58)
m=0
In (58) the coefficients fˆm are the projection of the solution with respect to m
fˆm (v, t) = f (z, v, t)m (z)p(z) dz = E[f (·, v, t)m (·)]. (59)
Stochastic Galerkin (SG) methods for kinetic equations based on the use of deter-
ministic methods in the phase space have demonstrated numerical and theoretical
evidence of spectral accuracy[11, 27–29, 32, 40, 61]. However, their practical
application presents some drawbacks.
• SG methods lead to the loss of the physical properties, like positivity
" and
conservation of moments. For example, if we denote by mφ (f ) = f φ(v) dv
the moments of f , we have
M
mφ (f ) = mφ (fM ) = mφ (fˆm )m .
m=0
These properties are essential to characterize the long time behavior of the system
and the Maxwellian equilibrium states [17, 51].
• One possibility, is to modify the coefficients fˆm in the gPC expansion or the
polynomial basis in such a way that the macroscopic moments of f (or positivity)
are preserved. This approach, however, is rather difficult in general and typically
leads to the loss of spectral accuracy [7, 49].
• Additionally, for nonlinear hyperbolic conservation laws, like the Euler system
in the fluid-dynamic limit, the generalized polynomial chaos expansion may lead
to the loss of hyperbolicity of the resulting approximated system (see [12, 13, 28,
35, 56]).
168 L. Pareschi
M
Q̂h (fˆ, fˆ) = fˆm fˆn E [Q(m , n )h ] .
m,n=1
M
L̂h (fˆ∞ , ĝ) = fˆm∞ ĝn E [L(m , n )h ] .
m,n=1
state.
By substituting to ĝh = fˆh − fˆh∞ the SG scheme can be rewritten as
or equivalently
∞ ∞
∂t fM (z, v, t) = QM (fM , fM )(z, v, t) − QM (fM , fM ),
M
QM (fM , fM )(z, v, t) = Q̂m (fˆ, fˆ)m (z).
m=0
C
Q(f, f ) − QM (fM , fM )L2 () ≤ f H r () + Q(fM , fM )H r ()
Mr
∞ ∞ C
QM (fM , fM )L2 () ≤ f ∞ H r () + Q(fM
∞ ∞
, fM )H r ()
Mr
which provide a spectral estimate for the equilibrium preserving SG method.
The approach just described applies to a large variety of kinetic equations where
the equilibrium state is know. In the case of Fokker–Planck equations, the method
can be generalized to the situation where the steady state is not known in advance
[19]. The idea is based on the notion of quasi-equilibrium state. To this aim given a
one-dimensional Fokker–Planck equation characterized by
which gives
The above problem can be solved analytically for f = f ∞ (v) only in some special
cases. More in general we can represent a quasi-stationary solution in the form
v P[f ](z, v∗ , t) + D (z, v∗ )
f q (z, v, t) = C exp − dv∗ (61)
−∞ D(z, v∗ )
it is clear that the formulation presented in Sect. 6.1 applies and we obtain a steady
state preserving method for large times. We refer to [19] for more details.
As an example, let us consider the swarming model with self-propulsion defined
by
10 1 10 0
10 0 10-2
10 -1 10 -4
10 -2 10 -6
-3 -8
10 10
10 -4 0 2 4 6 8 10 10 -10
0 5 10 15 20
Fig. 10 Evolution of the L2 error for the swarming model (62) with standard SG scheme (left) and
with the micro-macro SG scheme (right). The error has been computed with respect to a reference
solution obtained with M = 40, N = 321, t = 10−1 and final time T = 20
The idea is to combine SG methods in the random space with particle Monte
Carlo methods for the approximation of f in the phase space. This novel hybrid
formulation makes it possible to construct efficient methods that preserve the main
physical properties of the solution along with spectral accuracy in the random space
[8, 9, 52].
where
The VFP equation can be derived from the following system of stochastic
differential equations for (Xi (z, t), Vi (z, t)) ∈ Rdv × Rdx , i = 1, . . . , N with
random inputs
⎧
⎪
⎨dXi (z, t) = Vi (z, t)dt
⎪
1
N
⎪
⎪dV (z, t) = P (z, Xi , Xj )(Vj − Vi )dt + 2D(z)dWi ,
⎩ i N
j =1
being {Wi }N
i=1 independent Brownian motions.
We consider the empirical measure associated to the particle system
1
N
f (N) (z, x, v, t) = δ(x − Xi (z, t)) ⊗ δ(v − Vi (z, t))
N
i=1
M
M
XiM (z, t) = X̂i,m (t)m (z), ViM (z, t) = V̂i,m (t)m (z), i = 1, . . . , N
m=0 m=0
where X̂i,m , V̂i,m are the projections of the solution with respect to m
X̂i,m (t) = E[Xi (·, t)m (·)], V̂i,m (t) = E[Vi (·, t)m (·)].
ij √
and Phk = E[P (·, XiM , XjM )h (·)k (·)], Dh = E[ 2D(·)h (·)].
Moments are recovered from the empirical measure as
1
N
(N)
fM (z, x, v, t) = δ(x − XiM (z, t)) ⊗ δ(v − ViM (z, t))
N
i=1
1
N
(N)
mφ (fM )= δ(x − XiM (z, t))φ(ViM (z, t))
N
i=1
Uncertainty Quantification for Kinetic Equation and Related Problems 173
The method just described has the usual quadratic cost O(N 2 ) of a mean field
problem, where each particle at each time step modifies its velocity interacting with
all other particles. In addition, this cost has to be multiplied by the quadratic cost
O(M 2 ) of the SG method. Therefore the overall computational cost is O(M 2 N 2 ).
A reduction of the cost is obtained using a suitable Monte Carlo evaluation of the
interaction dynamics to mitigate the curse of dimensionality [1]
⎧
⎪
⎪ d X̂i,h = V̂i,h dt
⎪
⎨
1 ij
M
⎪
⎪ = Phk (V̂j,k − V̂i,k )dt + Dh dWi
⎪ d V̂
⎩ i,h S
j ∈Si k=0
fix S ≤N.
1. Consider N samples (Xi , Vi ) from f0 (x, v) and
2. Perform gPC representation on the particles : X̂i,h , V̂i,h , for h = 0, . . . , M
3. For n = 0, . . . , T − 1
• Generate N random variables {ηi }N
i=1 ∼ N(0, 1)
• For i = 1, . . . , N
– Sample S particles {j1 , . . . , jS } := Si uniformly without repetition
– Compute the space and velocity change
n+1
X̂i,h = X̂i,h
n
+ V̂i,h
n
t
t ij n √
M
n+1
V̂i,h = n
V̂i,h + Pkh (V̂j,k − V̂i,k
n
) + t Dh ηi
S
j ∈Si k=0
3 10-1
2 10-2
1 10-3
0 10-4
-1 -0.5 0 0.5 1 0 1 2 3 4 5
Fig. 11 Left: expected density at time T = 1 obtained through a standard SG method (gPC) and
the particle gPC scheme (MCgPC) with S = 5 at each time step. The gPC expansion has been
performed up to order M = 5. Right: convergence for the expected temperature of the MCgPC
method with fixed S = 50 and an increasing number of particles. The reference temperature has
been computed with a standard SG method for the mean-field problem
• In the case S = N we obtain the typical convergence rate O(N −1/2 ) due to
Monte Carlo sampling in the phase
√ space. The fast evaluation of the interactions
induces an additional error O 1/S − 1/N with S < N.
We report in Fig. 11 the result of a simulation concerning the simple space
homogeneous one-dimensional alignment process corresponding to P (z, x, x∗ ) =
1 + sz, z ∼ U (0, 1), s = 0.5 and D = 0. The initial data is given by a bimodal
density
(v−μ)2 (v+μ)2
f0 (v) = β e 2σ 2 + e 2σ 2
"
with σ 2 = 0.1, μ = 0.25 and β such that R f0 (v) dv = 1. It is clear that a very
small value M suffices to match the accuracy in the random and in the phase space.
To this aim, we will focus on the space homogeneous Boltzmann equation, and
observe that, in the case of Maxwell molecules B ≡ 1, the collision operator can be
rewritten as
M
viM (z, t) = v̂i,m (t)m (z).
m=0
1 1
vi (z, t) = (vi (z, t) + vj (z, t)) + |vi (z, t) − vj (z, t)|ω,
2 2
1 1
vj (z, t) = (vi (z, t) + vj (z, t)) − |vi (z, t) − vj (z, t)|ω.
2 2
Let us observe that
|vi (z, t) − vj (z, t)| = |vi (z, t) − vj (z, t)|, (65)
1 M 1 M
viM (z, t) = (v (z, t) + vjM (z, t)) + |v (z, t) − vjM (z, t)| ω,
2 i 2 i
1 1 M
viM (z, t) = (viM (z, t) + vjM (z, t)) − |v (z, t) − vjM (z, t)| ω
2 2 i
and then project by integrating against m (z) p(z) on to get for m = 0, . . . , M
1 1 m
v̂i,m (t) = (v̂i,m (t) + v̂j,m (t)) + V̂ ω, (66)
2 2 ij
1 1 m
v̂j,m (t) = (v̂i,m (t) + v̂j,m (t)) − V̂ ω (67)
2 2 ij
176 L. Pareschi
where
V̂ijm = |viM (z, t) − vjM (z, t)|m (z) p(z) dz, (68)
end for
As a numerical example we consider the 2D case with uncertain initial data
corresponding to the exact solution[5, 52]
1 1 − α(z)s(z, t) v2 v2
f (z, v, t) = 1− 1− e− 2s(z,t) , (69)
2πs(z, t) α(z)s(z, t) 2s(z, t)
2 − e−t /8
where s(z, t) = . We will consider α(z) = 2 + κz, with z ∼ U (−1, 1).
2α(z)
To emphasize the good agreement of the computed approximation for all times,
we depict in Fig. 12 the evolution at times t = 0, 1, 5 of the marginal of E[f ] and
Var(f ). Next, in Fig. 13 we present spectral convergence of the scheme computed
through the fourth order moment of the 2D model with α(z) = 2 + κz, κ = 0.25
and κ = 0.75 with z ∼ U (−1, 1). As reference solution we considered the fourth
order moment at time T = 5 obtained with N = 106 particles and M = 25
Galerkin projections and the evolution is computed with t = 10−1 . In the right
plot we present the decay of the L2 () error for increasing M = 0, . . . , 14 in
semilogarithmic scale. In the left plot we represent also the whole evolution of M4
(a) (b) (c)
Uncertainty Quantification for Kinetic Equation and Related Problems
step has been performed in [−5, 5]2 through 1002 gridpoints. (a) t = 0. (b) t = 1. (c) t = 5. (d) t = 0. (e) t = 1. (f) t = 5
178 L. Pareschi
0
10
10 -5
10 -10
-15
10
0 2 4 6 8 10 12 14
Fig. 13 Left: Convergence of the L2 () error with respect to the fourth order moment obtained
from a reference solution computed with N = 106 particles and M = 25 from the DSMC-SG
methods. Right: evolution of the fourth order moment in the interval [0, 5] for exact and DSMC-
SG approximation with N = 106 and M = 5
H
V̂ijm (t) ≈ wh |viM (zh , t) − vjM (zh , t)|m (zh ). (70)
h=0
The resulting scheme requires O(MH ) operations to compute viM (zh , t) and
vjM (zh , t) for all h’s and O(MH ) operations to evaluate V̂ijm (t) for all m’s. Taking
H = M the total cost of a Monte Carlo collision at each time step is therefore
O(M 2 ).
Acknowledgments This work has been supported by the Italian Ministry of Instruction, Univer-
sity and Research (MIUR) under the PRIN Project 2017, No. 2017KKJP4X, “Innovative numerical
methods for evolutionary partial differential equations and applications”.
Uncertainty Quantification for Kinetic Equation and Related Problems 179
References
1. Albi, G., Pareschi, L.: Binary interaction algorithms for the simulation of flocking and
swarming dynamics. Multiscale Model. Simul. 11(1), 1–29 (2013)
2. Albi, G., Pareschi, L., Zanella, M.: Uncertainty quantification in control problems for flocking
models. Math. Probl. Eng. 2015, 850124 (2015)
3. Ayi, N., Faou, E.: Analysis of an asymptotic preserving scheme for stochastic linear kinetic
equations in the diffusion limit. SIAM/ASA J. Uncertain. Quantif. 7(2), 760–785 (2019)
4. Bijl, H., Lucor, D., Mishra, S., Schwab, C. (eds.): Uncertainty Quantification in Computational
Fluid Dynamics. Lecture Notes in Computational Science and Engineering. Springer, Berlin
(2013)
5. Bobylev, A.V.: Exact solutions of the Boltzmann equation. Dokl. Akad. Nauk SSSR 225, 1296–
1299 (1975) (in Russian)
6. Caflisch, R.E.: Monte Carlo and Quasi Monte Carlo methods. Acta Numer. 7, 1–49 (1998)
7. Cai, Z., Fan, Y., Ying, L.: An entropic Fourier method for the Boltzmann equation. SIAM J.
Sci. Comput. 40, A2858–A2882 (2018)
8. Carrillo, J.A., Pareschi, L., Zanella, M.: Particle based gPC methods for mean-field models of
swarming with uncertainty. Commun. Comput. Phys. 25, 508–531 (2019)
9. Carrillo, J.A., Zanella, M.: Monte Carlo gPC methods for diffusive kinetic flocking models
with uncertainties. Vietnam J. Math. 47, 931–954 (2019)
10. Cercignani, C.: The Boltzmann Equation and Its Applications. Springer, New York (1988)
11. Daus, E.S., Jin, S., Liu, L.: Spectral convergence of the stochastic Galerkin approximation to
the Boltzmann equation with multiple scales and large random perturbation in the collision
kernel. Kinetic Related Models 12, 909–922 (2019)
12. Després, B., Perthame, B.: Uncertainty propagation; Intrusive kinetic formulations of scalar
conservation laws. SIAM/ASA J. Uncertain. Quantif. 4, 980–1013 (2016)
13. Després, B., Poëtte, G., Lucor, D.: Robust uncertainty propagation in systems of conservation
laws with the entropy closure method. In: Uncertainty Quantification in Computational Fluid
Dynamics. Lecture Notes in Computational Science and Engineering, vol. 92, pp. 105–149
(2010)
14. Degond, P., Pareschi, L., Russo, G. (eds.): Modeling and computational methods for kinetic
equations. In: Modeling and Simulation in Science, Engineering and Technology. Birkhäuser,
Boston (2004)
15. Dimarco, G., Pareschi, L.: Multi-scale control variate methods for uncertainty quantification
of kinetic equations. J. Comput. Phys. 388, 63–89 (2019)
16. Dimarco, G., Pareschi, L.: Multi-scale variance reduction methods based on multiple control
variates for kinetic equations with uncertainties. Multiscale Model. Simul. 18(1), 351–382
(2020)
17. Dimarco, G., Pareschi, L.: Numerical methods for kinetic equations. Acta Numer. 23, 369–520
(2014)
18. Dimarco, G., Pareschi, L., Zanella, M.: Uncertainty quantification for kinetic models in
socio-economic and life sciences. In: Uncertainty Quantification for Kinetic and Hyperbolic
Equations. SEMA-SIMAI Springer Series (2018)
19. Dimarco, G., Pareschi, L., Zanella, M.: Micro-macro stochastic Galerkin methods for Fokker–
Planck equations (2020). Preprint
20. Fairbanks, H.R., Doostan, A., Ketelsen, C., Iaccarino, G.: A low-rank control variate for
multilevel Monte Carlo simulation of high-dimensional uncertain systems. J. Comput. Phys.
341, 121–139 (2017)
21. Gamba, I.M., Jin, S., Liu, L.: Error estimate of a bi-fidelity method for kinetic equations with
random parameters and multiple scales. to appear in Int. J. Uncertain. Quantif. (2019)
22. Giles, M.B.: Multilevel Monte Carlo methods. Acta Numer. 24, 259–328 (2015)
23. Gittelson, C.J., Schwab, C.: Sparse tensor discretization of high-dimensional parametric and
stochastic PDEs. Acta Numer. 20, 291–467 (2011)
180 L. Pareschi
24. Gunzburger, M.D., Webster, C.G., Zhang, G.: Stochastic finite element methods for partial
differential equations with random input data. Acta Numer. 23, 521–650 (2014)
25. Hammersley, J.M., Handscomb, D.C.: Monte Carlo Methods. Methuen, London, and Wiley,
New York (1964)
26. Heitzinger, C., Leumüller, M., Pammer, G., Rigger, S.: Existence, uniqueness, and a com-
parison of nonintrusive methods for the stochastic nonlinear Poisson-Boltzmann equation.
SIAM/ASA J. Uncertain. Quantif. 6(3), 1019–1042 (2018)
27. Hu, J., Jin, S.: A stochastic Galerkin method for the Boltzmann equation with uncertainty. J.
Comput. Phys. 315, 150–168 (2016)
28. Hu, J., Jin, S., Shu, R.: On stochastic Galerkin approximation of the nonlinear Boltzmann
equation with uncertainty in the fluid regime. J. Comput. Phys. 397, 108838 (2019)
29. Hu, J., Jin, S., Shu, R.: A stochastic Galerkin method for the Fokker–Planck-Landau equation
with random uncertainties. In: Theory, Numerics and Applications of Hyperbolic Problems. II,
pp. 1–19. Springer Proc. Math. Stat., vol. 237. Springer, Cham (2018)
30. Hu, J., Pareschi, L., Wang, Y.: Uncertainty quantification for the kinetic BGK equation using
variance reduction multilevel Monte Carlo methods. to appear in SIAM/ASA J. Uncertain.
Quantif. (2020)
31. Jin, S., Liu, L.: An asymptotic-preserving stochastic Galerkin method for the semiconductor
Boltzmann equation with random inputs and diffusive scalings. Multiscale Model. Simul.
15(1), 157–183 (2017)
32. Jin, S., Pareschi, L. (eds.): Uncertainty Quantification for Kinetic and Hyperbolic Equations.
SEMA-SIMAI Springer Series, vol. 14. Springer, Berlin (2017)
33. Jin, S., Lu, H., Pareschi, L.: Efficient stochastic asymptotic-preserving implicit-explicit
methods for transport equations with diffusive scalings and random inputs. SIAM J. Sci.
Comput. 40(2), A671–A696 (2018)
34. Jin, S., Lu, H., Pareschi, L.: A high order stochastic Asymptotic-Preserving scheme for
chemotaxis kinetic models with random inputs. Mult. Model. Simul. 6, 1884–1915 (2018)
35. Jin, S., Shu, R.: A study of hyperbolicity of kinetic stochastic Galerkin system for the isentropic
Euler equations with uncertainty. Chin. Ann. Math. Ser. B 40(5), 765–780 (2019)
36. Jin, S., Zhu, Y.: Hypocoercivity and uniform regularity for the Vlasov–Poisson–Fokker–Planck
system with uncertainty and multiple scales. SIAM J. Math. Anal. 50(2), 1790–1816 (2018)
37. Le Maitre, O., Knio, O.M.: Spectral Methods for Uncertainty Quantification: With Applica-
tions to Computational Fluid Dynamics. Scientific Computation. Springer Netherlands (2010)
38. Li, Q., Wang, L.: Uniform regularity for linear kinetic equations with random input based on
hypocoercivity. SIAM/ASA J. Uncertain. Quantif. 5(1), 1193–1219 (2017)
39. Liu, L.: A stochastic asymptotic-preserving scheme for the bipolar semiconductor Boltzmann-
Poisson system with random inputs and diffusive scalings. J. Comput. Phys. 376, 634–659
(2019)
40. Liu, L., Jin, S.: Hypocoercivity based sensitivity analysis and spectral convergence of the
stochastic Galerkin approximation to collisional kinetic equations with multiple scales and
random inputs. SIAM Multiscale Model. and Simul. 16, 1085–1114 (2018)
41. Liu, L., Zhu, X.: A bi-fidelity method for the multiscale Boltzmann equation with random
parameters. J. Comput. Phys. 402, 108914 (2020)
42. Liu, T.-P., Yu, S.-H.: Boltzmann equation: micro–macro decomposition and positivity of shock
profiles. Commun. Math. Phys. 246(1), 133–179 (2004)
43. Loève, M.: Probability Theory I, 4th edn. Springer, Berlin (1977)
44. Mishra, S., Schwab, Ch., Šukys, J.: Multi-level Monte Carlo finite volume methods for
nonlinear systems of conservation laws in multi-dimensions. J. Comput. Phys. 231, 3365–3388
(2012)
45. Mouhot, C., Pareschi, L.: Fast algorithms for computing the Boltzmann collision operator.
Math. Comput. 75, 1833–1852 (2006)
46. Naldi, G., Pareschi, L., Toscani, G. (eds.): Mathematical Modeling of Collective Behavior in
Socio-Economic and Life Sciences. Birkhäuser, Boston (2010)
Uncertainty Quantification for Kinetic Equation and Related Problems 181
47. Nobile, F., Tempone, R., Webster, C.G.: A sparse grid stochastic collocation method for partial
differential equations with random input data. SIAM J. Num. Anal. 46(5), 2309–2345 (2008)
48. Pareschi, L., Russo, G.: An introduction to Monte Carlo methods for the Boltzmann equation.
ESAIM Proc. 10, 35–75 (2001)
49. Pareschi, L., Russo, G.: On the stability of spectral methods for the homogeneous Boltzmann
equation. Transp. Theory Stat. Phys. 29, 431–447 (2000)
50. Pareschi, L., Toscani, G.: Interacting Multiagent Systems: Kinetic Equations and Monte Carlo
Methods. Oxford University Press, Oxford (2013)
51. Pareschi, L., Zanella, M.: Structure–preserving schemes for nonlinear Fokker–Planck equa-
tions and applications. J. Sci. Comput. 74, 1575–1600 (2018)
52. Pareschi, L., Zanella, M.: Monte Carlo stochastic Galerkin methods for the Boltzmann equation
with uncertainties: space homogeneous case. J. Comput. Phys. 423, 109822 (2020)
53. Peherstorfer, B., Gunzburger, M., Willcox, K.: Convergence analysis of multifidelity Monte
Carlo estimation. Numer. Math. 139, 683–707 (2018)
54. Peherstorfer, B., Willcox, K., Gunzburger, M.: Optimal model management for multifidelity
Monte Carlo estimation. SIAM J. Sci. Comput. 38(5), A3163–A3194 (2016)
55. Pettersson, P., Iaccarino, G., Nordström, J.: Polynomial Chaos Methods for Hyperbolic Partial
Differential Equations: Numerical Techniques for Fluid Dynamics Problems in the Presence of
Uncertainties. Mathematical Engineering. Springer, Berlin (2015)
56. Poëtte, G., Després, B., Lucor, D.: Uncertainty quantification for systems of conservation laws.
J. Comput. Phys. 228(7), 2443–2467 (2009)
57. Poëtte, G.: A gPC-intrusive Monte-Carlo scheme for the resolution of the uncertain linear
Boltzmann equation. J. Comput. Phys. 385, 135–162 (2019)
58. Poëtte, G.: Spectral convergence of the generalized Polynomial Chaos reduced model obtained
from the uncertain linear Boltzmann equation. Math. Comput. Simul. 177(C), 24–45 (2020)
59. Roosta-Khorasani, F., van den Doel, K., Ascher, U.: Stochastic algorithms for inverse problems
involving PDEs and many measurements. SIAM J. Sci. Comput. 36(5), S3–S22 (2014)
60. Russo, G., Santagati, P., Yun, S-B.: Convergence of a semi-Lagrangian scheme for the BGK
model of the Boltzmann equation. SIAM J. Numer. Anal. 50(3), 1111–1135 (2012)
61. Shu, R., Hu, J., Jin, S.: A stochastic Galerkin method for the Boltzmann equation with
multi-dimensional random inputs using sparse wavelet bases. Numer. Math. Theory Methods
Applications 10, 465–488 (2017)
62. Toscani, G.: Entropy production and the rate of convergence to equilibrium for the Fokker–
Planck equation. Q. Appl. Math. LVII(3), 521–541 (1999)
63. Toscani, G., Villani, C.: Sharp entropy dissipation bounds and explicit rate of trend to
equilibrium for the spatially homogeneous Boltzmann equation. Commun. Math. Phys. 203(3),
667–706 (1999)
64. Villani, C.: A review of mathematical topics in collisional kinetic theory. In: Friedlander, S.,
Serre, D. (eds.) Handbook of Mathematical Fluid Mechanics, vol. I, pp.71–305. North-Holland,
Amsterdam (2002)
65. Xiu, D.: Numerical Methods for Stochastic Computations. Princeton University Press, Prince-
ton (2010)
66. Zhu, Y., Jin, S.: The Vlasov–Poisson–Fokker–Planck system with uncertainty and a one-
dimensional asymptotic-preserving method. SIAM Multiscale Model. Simul. 15(4), 1502–
1529 (2017)
67. Zhu, X., Linebarger, E.M., Xiu, D.: Multi-fidelity stochastic collocation method for computa-
tion of statistical moments. J. Comput. Phys. 341, 386–396 (2017)
A Brief Introduction to the Scaling
Limits and Effective Equations in Kinetic
Theory
Many interesting systems in physics and applied sciences consist of a large number
of identical components so that they are difficult to analyze from a mathematical
point of view. On the other hand, quite often, we are not interested in a detailed
description of the system but rather in its collective behaviour. Therefore, it is
necessary to look for all procedures leading to simplified models, retaining the
interesting features of the original system, cutting away redundant information. This
is the methodology of statistical mechanics and of kinetic theory. Here we want to
M. Pulvirenti ()
Dipartimento di Matematica, Università di Roma La Sapienza, Roma, Italy
International Research Center M&MOCS, Università dell’Aquila, Cisterna di Latina (LT), Italy
e-mail: pulviren@mat.uniroma1.it
S. Simonella
ENS de Lyon, UMPA UMR 5669 CNRS, Lyon Cedex 07, France
e-mail: sergio.simonella@ens-lyon.fr
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 183
G. Albi et al. (eds.), Trails in Kinetic Theory, SEMA SIMAI Springer Series 25,
https://doi.org/10.1007/978-3-030-67104-4_6
184 M. Pulvirenti and S. Simonella
outline the limiting procedure leading from the microscopic description of a large
particle system (based on the fundamental laws like the Newton or Schrödinger
equations) to the more practical picture dictated by kinetic theory.
Although the methods of kinetic theory are frequently applied to a large variety of
complex systems (consisting of a huge number of individuals), we will discuss only
models arising in physics and more precisely in classical mechanics. The starting
point is a system of N identical particles in the physical space. A microscopic state
of the system is a sequence z1 , · · · , zN where zi = (xi , vi ) denotes position and
velocity of the i-th particle. The equations of motion are given by Newton’s laws of
dynamics.
We are interested in a situation where N is very large (for instance, a cubic
centimeter of a rarefied gas contains approximately 1019 molecules). The knowledge
of the microscopic states becomes useless, and we turn to a statistical description.
We introduce a probability measure W0N (ZN )dZN (absolutely continuous with
respect to the Lebesgue measure), defined on the phase space R3N × R3N , where
ZN = (z1 , · · · , zN ) = (xi , vi , · · · , xN , vN ) .
W0N assigns the same statistical weight to two different vectors ZN and ZN differing
only for the order of particles, i.e., identifying the same physical configuration.
The time-evolved measure is defined by
and we look for an equation describing the evolution of f1N . We deduce, in most of
the physically relevant situations, an evolution equation of the form
The first term in the right-hand side is due to the free transport of particles, while
the term Q should describe the interaction of particle 1 with the rest of the system.
A Brief Introduction to the Scaling Limits and Effective Equations in Kinetic Theory 185
We face a big difficulty. Since the interaction is binary, Q will depend on f2N ,
namely the two-particle marginal. In other words, (3) is still useless: to know f1N
we need to know f2N , and to know f2N we need to know f3N , and so on. We handle a
hierarchy of equations, called BBGKY hierarchy (from the names of the physicists
Bogolyubov, Born, Green, Kirkwood, Yvon).
Here enters the property called propagation of chaos, that is,
Accepting (4), Q becomes an operator acting on f1N and (3) is a closed equation.
We have thus replaced a huge ordinary differential system by a single PDE. The
price we pay is that (3) is nonlinear.
The equality in Eq. (4) is certainly false, since it expresses the statistical
independence of particle 1 and particle 2 which, even if assumed at time 0, cannot
hold at later times. Indeed, the dynamics creates correlations. Nevertheless, one
can hope to recover this property in some asymptotic situation described by a
suitable scaling limit. This is what happens in two different physical contexts: the
low-density and the weak-coupling limits, yielding two different kinetic equations,
namely the Boltzmann and the Landau equations, respectively. The passage from
hamiltonian mechanics to this kinetic description is actually very delicate. As
we shall see later on, we go from a deterministic time-reversible system to an
irreversible equation.
A different scaling procedure is the so-called mean-field limit. This leads to the
Vlasov equation, which has still a time-reversible, hamiltonian nature. It is a sort of
continuum limit and hence much simpler than the previous two. Some challenging
and interesting problems concerning the mean-field limit are anyway still open, but
we shall not discuss them in this note.
with
v = v − n[n · (v − v1 )]
v v1
v1
v
As a fundamental feature of (5), one has the formal conservation (in time) of the
five quantities
and proved the famous H theorem asserting the decrease of H (f (t)) along the
solutions of (5).
Finally, in the case of bounded domains or homogeneous solutions (f = f (v, t)
independent of x), the distribution defined for some β > 0, ρ > 0 and u ∈ R3 by
ρ
e−β/2|v−u| ,
2
M(v) = (10)
(2π/β)3/2
called Maxwellian distribution, is stationary for the evolution given by (5). In addi-
tion, M minimizes H among all distributions with given total mass ρ, mean velocity
u and mean energy. The parameter β is interpreted as the inverse temperature.
In conclusion, Boltzmann was able to introduce an evolution equation with the
remarkable properties of expressing mass, momentum and energy conservation and
also the tendency to thermal equilibrium. In this way, he tried to conciliate Newton’s
laws with the second principle of thermodynamics.
The H Theorem is apparently in contrast with the laws of mechanics, which
are time-reversible. This fact caused skepticism among the scientific community,
and the work of Boltzmann was attacked repeatedly. We refer the reader to
the monograph by C. Cercignani [5], which is a beautiful compromise between
historical account and scientific divulgation, to have a faithful idea of the debate
at the time.
To formally derive (5), let us consider a system of N identical hard spheres of
diameter ε and unitary mass, interacting by means of the collision law (7). We
denote by ε the diameter of the particles which, for the moment, is fixed and not
necessarily small.
The phase space !N of the system is the subset of R3N × R3N fulfilling the hard-
core condition, namely |xi − xj | ≥ ε for i = j . The dynamical flow ZN → t (ZN )
is defined as the free flow, i.e., ZN → t (ZN ) = (x1 + v1 t, v1 , · · · , xN + vN t, vN )
up to the first impact time (when |xi − xj | = ε); then an instantaneous collision
takes place according to the law (7), and the flow goes on up to the next collision
instant.
The well-posedness of the hard-sphere dynamics is not obvious, due to the
occurrence of multiple collisions or to the a priori possibility that collision times
accumulate at a finite limiting time. However, such pathologies cannot occur outside
a set of initial conditions ZN of vanishing measure. Indeed following [1] (see also
[6]), the flow ZN → t (ZN ) can be defined for all t ∈ R almost everywhere with
respect to the Lebesgue measure, which is enough for what will follow (even the
proof of this result is not relevant in the following, so that we omit further details).
188 M. Pulvirenti and S. Simonella
where Coll denotes the variation of f1N due to the collisions, which takes the form
In the next section, we will comment on the justification of this equation. Here, let
us accept it and argue on its consequences.
Two given particles should be (almost) uncorrelated if the gas is rarefied enough.
This leads to the propagation of chaos
which might seem contradictory at first sight. In fact, if two particles collide,
correlations are created. Even assuming (13) at some given time, if particle 1
collides with particle 2, such an equation cannot be satisfied at any time after the
collision.
Before discussing the propagation of chaos further, we notice that, in practical
situations, for a rarefied gas, Nε3 (total volume occupied by the particles) is very
small, while Nε2 = O(1). This implies that the collision operator given by (12)
is O(1). Therefore, since we are dealing with a huge number of particles, we are
tempted to perform the limit N → ∞ and ε → 0 in such a way that ε2 = O(N −1 ).
As a consequence, the probability that two tagged particles collide (which is of the
order of the surface of a ball, that is O(ε2 )), is negligible. Instead, the probability
that a given particle collides with any of the remaining N − 1 particles (which is
O(Nε2 ) = O(1)) is not negligible. On the other hand, condition (13) refers to two
preselected particles (say 1 and 2) and it is not unreasonable to conceive that it holds
in the limiting situation in which we work.
A Brief Introduction to the Scaling Limits and Effective Equations in Kinetic Theory 189
Nevertheless, we cannot simply insert (13) into (12), as the integral operator
refers to times both before and after the collision. Let us assume (13) only when
the pair of velocities (v, v2 ) are incoming ((v − v2 ) · n > 0). If the two particles
are initially uncorrelated, it is unlikely that they have collided before a given time t,
hence we assume their statistical independence.
This is a standard argument in textbooks of kinetic theory, but some extra care is
needed. If particles 1 and 2 have not collided directly before a given time t, this does
not imply that they are uncorrelated. Indeed there may exist a chain of collisions
involving a group i1 , i2 , · · · of particles
1 → i1 → i2 → · · · → 2 ,
correlating particles 1 and 2. As we shall see later, this is excluded (at least for a
short time) by a more rigorous analysis. The two clusters of particles influencing
the dynamics of particles 1 and 2 are disjoint with large probability.
Coming back to (12), for the outgoing pair velocities (v, v2 ) (satisfying (v2 − v) ·
n > 0), we shall make use of the continuity property
(∂t + v · ∇x )f = λ−1 dv2 dn (v − v2 ) · n [f (x, v )f (x, v2 ) − f (x, v)f (x, v2 )].
S+
(16)
The parameter λ represents, roughly, the typical length a particle can cover without
undergoing any collision (mean free path). (In (6), we just chose λ = 1.)
It may be worth remarking that, after having taken the limit N → ∞ and ε → 0,
there is no way to distinguish between incoming and outgoing pair velocities. This
is because no trace of the parameter ε is left in (16) and n plays the role of a random
variable. However, keeping in mind the way the Boltzmann equation was derived,
one shall conventionally maintain the name incoming for velocities satisfying the
condition (v − v2 ) · n > 0 (and consequently the pair (v , v2 ) would be outgoing
in (16)).
190 M. Pulvirenti and S. Simonella
Equation (16) (or equivalently (5)–(6)) is the Boltzmann equation for hard
spheres. Such an equation has a statistical nature, and it is not equivalent to the
hamiltonian dynamics from which it has been derived. Indeed the H theorem shows
that it is not reversible in time in contrast with the laws of mechanics.
By the analysis on the order of magnitude of the quantities in the game, we
deduced that the Boltzmann equation works in special situations only. The condition
Nε2 = O(1) means that we consider a rarefied gas, with almost vanishing volume
density. After Boltzmann established the equation, Harold Grad [8, 9] postulated its
validity in the limit N → ∞ and ε → 0 with Nε2 → O(1) as discussed above (this
is often called, indeed, the Boltzmann-Grad limit).
There is no contradiction in the irreversibility or in the trend to equilibrium
obtained after the limit, when they are strictly speaking false for mechanical
systems. However, the arguments above are delicate and require a rigorous, deeper
analysis. If the Boltzmann equation is not a purely phenomenological model
derived by assumptions ad hoc and justified by its practical relevance, but rather
a consequence of a mechanical model, we must derive it rigorously. In particular,
the propagation of chaos should not be a hypothesis but the statement of a theorem.
After the formulation of the mathematical validity problem by Grad, Cercignani
[4] obtained the evolution equation (hierarchy) for the marginals of a hard-sphere
system, and this was the starting point to rigorously derive the Boltzmann equation,
as accomplished by Lanford in his famous paper [14], even though only for a short
time interval.
Lanford’s theorem is probably the most relevant result regarding the mathemat-
ical foundations of kinetic theory. In fact, it dispelled the many previous doubts on
the validity of the Boltzmann equation (although some authors refuse a priori the
problem of deriving the equation starting from mechanical systems [22]).
Unfortunately, the short-time limitation is a serious one. Only for special
systems, as is the case of a very rarefied gas expanding in a vacuum, can we obtain a
global validity [11, 12]. The possibility of deriving the Boltzmann equation globally
in time, at least in cases when we have a global existence of good solutions, is still
an open, challenging problem.
We conclude this section with a few historical remarks. Before Boltzmann,
Maxwell proposed a kinetic equation that is just the Boltzmann equation integrated
against test functions [16, 17]. He considered also more general potentials, in partic-
ular, inverse-power-law potentials, motivated essentially by the special properties of
their cross-section. After Lanford’s result, the case of smooth short-range potentials
has been studied by other authors [7, 13, 19]. It is a nontrivial extension, in particular
when the interacting potential is not “close enough” to a hard-sphere potential. The
validity (or nonvalidity) of the Boltzmann equation in the case of genuine long-
range interactions is open, in absence of techniques suited to deal with collisional
and mean-field terms simultaneously.
A Brief Introduction to the Scaling Limits and Effective Equations in Kinetic Theory 191
In this and in the following section we give more details on the derivation of (5) from
N hard spheres of diameter ε, discussed above heuristically. We remind the reader
that we are interested in the behaviour of the system in the limit N → ∞, ε → 0
fixing ε2 N = 1 (1 chosen for simplicity), according to the Boltzmann-Grad limit.
Namely we have a single scaling parameter ε (or N), and we study the asymptotics
ε → 0 (N → ∞).
We start with the justification of (1). Let A be a measurable set in R3N × R3N .
Then the probability of finding the system in A at time t > 0 is given by
where
W N (ZN , t)χA (ZN ) = W0N (ZN )χ−t (A) (ZN ) = W0N (ZN )χA (t (ZN )) ,
for any bounded Borel function u. Here the integral is extended over all the phase
space !N . By using the Liouville theorem on the transformation ZN → t (ZN ), it
follows that
Here Lεj is the generator of the dynamics of j hard spheres of diameter ε (Liouville
operator of a j −particle system), while
j
Cjε+1 = ε
Ck,j +1 , (19)
k=1
To derive Eq. (18) formally, we would like to give some description of Lεj as
j
differential operator. This poses a difficulty, in fact Lεj = i=1 vi · ∇xi on functions
vanishing on ∂!j and the interacting dynamics is completely coded on the boundary.
In [4, 6], boundary conditions are imposed using (14), and its higher order versions,
and Eq. (18) is derived integrating by parts over !N . However if one is not afraid
of working with delta functions, it is more convenient to use the following compact
description:
j
Lεj = vi · ∇xi − Tεj (22)
i=1
where
Tεj = Tε;
j
i,k
(23)
i<k
i,k=1,··· ,j
and
$ %
Tε;
j
i,k N
fj (Zj ) = ε2 dn Ui,k · n +
δ(Ri,k − εn)bni,k − δ(Ri,k + εn) fjN (Zj ) ,
S2
(24)
The last operator transforms the incoming pair (vi , vk ) into the outgoing (vi , vk )
after a scattering with impact parameter n. Note finally that the operator (20) can be
as well expressed in terms of Tεj :
ε; k,j +1 N
ε2 Ck,j
ε
+1 fj +1 =
N
dzj +1 Tj +1 fj +1 . (25)
We should remind here that the marginals are supported on the space of non-
overlapping hard spheres (|xi − xk | ≥ ε for i = k). Therefore, when we think
of (18) (and (21)) as equations over the whole space R3j × R3j , we should always
complement them with the condition |xi − xk | < ε ⇒ fjN = 0.
Let us now check the expression given for the Liouville equation, based on TεN .
Consider a point particle hitting a sphere of diameter ε of infinite mass, centred at
the origin. Let g(X, V , t) be the probability distribution of the point particle, with
initial datum g(t = 0) = g0 . Let V and V denote the incoming and outgoing
velocity, respectively. It is V = V − 2(V · n)n, where n ∈ S 2 is the impact vector.
V
n V
We denote by Z = (X, V ) → Z(t) = (X(t), V (t)) the dynamical flow. For any test
function ϕ = ϕ(X, V ) we have that
d
ϕ(Z(t)) = V · ∇X ϕ(Z(t)) + δ(t − τ )[ϕ(X + V τ, V ) − ϕ(X + V τ, V )]
dt
where τ is the hitting time. The term [· · · ] describes the jump in velocity. Proceeding
as in (17) we deduce that
d
g(Z, t) ϕ(Z) = − V · ∇X g(Z, t) ϕ(Z) + g0 (Z) δ(t − τ ) [· · · ].
dt Acoll
(26)
which has jacobian determinant of modulus ε2 |V · n|, the last term in Eq. (26)
becomes
∞
ε2 dτ dV dn |V · n| g0 (Z(nε, τ, V )) δ(t − τ ) [ϕ(nε, V ) − ϕ(nε, V )]
0 V ·n<0
Note that in the last step we changed again variables, V → V , in the positive term.
" identify the time derivative in strong form, we now write g(nε, ·)ϕ(nε, ·) =
To
dXδ(X − nε)g(X, ·)ϕ(X, ·), exchange the integrals. and make a last change of
variables n → −n in the negative term. We conclude that
d
g(Z, t) = −V ·∇X g(Z, t)+ε2 dn |V ·n| [δ(x−εn) bn −δ(x+εn)] g(Z, t) ,
dt V ·n>0
where bn flips V into V . The general form of TεN follows easily from this
computation.
In order to arrive to (18), it is enough to proceed as in the standard derivation
of the BBGKY hierarchy for smooth potentials. We split the sums in TεN given by
Eq. (23), as
= + + ,
i<k i<k≤j i≤j,k>j j <i<k
i,k=1,··· ,N
and integrate in dzj +1 · · · dzN . The first sum produces Lεj immediately. The second
sum gives the collision operator ε2 Cjε+1 , multiplied by a symmetry factor (N − j ).
The last sum vanishes by exact compensation of gain and loss in (24).
Equation (18) is the starting point for Lanford’s validity theorem, as we shall see
in the following section.
A Brief Introduction to the Scaling Limits and Effective Equations in Kinetic Theory 195
The iteration of the Duhamel formula for Eq. (18) leads to express fjN (t) as a sum:
N−j t t1 tn−1
fjN (t) = αn (N, j ) dt1 dt2 . . . dtn (27)
n=0 0 0 0
αn (N, j ) = (N − j ) · · · (N − j − n + 1) ε2n .
The main ingredient for the theorem of Lanford stated below, is just this explicit
representation for the solution of the N-particle hierarchy. Actually, this identity
can be rigorously proved directly, without making use of (18) [18, 20].
On the other hand, a similar formula can be established for the tensor product of
solutions to the Boltzmann equation fj (t) := f (t)⊗j . Namely we have that
∞
t t1 tn−1
fj (t) = dt1 dt2 . . . dtn (28)
n=0 0 0 0
f0 (x, v) ≤ h(x)e−βv
2
=
Hypothesis 2 Let !j be the subset of R3j × R3j fulfilling the condition
xi = xk for i = k, 1 ≤ i, k ≤ j.
⊗j
lim fjN (t = 0) = f0 , (29)
ε→0
=
uniformly on compact subsets of !j and the bound
:
j
h(xi )e−βvi .
2
fjN (x1 , v1 , · · · xj , vj , t = 0) ≤
i=1
Theorem ([14]) Under the hypotheses 1 and 2, there exists t0 > 0 (depending only
on z0 , β) such that, for t < t0 we have, for all j ≥ 1,
where f (t) is the unique solution to the Boltzmann equation. The convergence holds
almost everywhere.
Following Lanford, the proof can be organized in two steps.
We first give an a priori bound on the series expansions (27) (uniform in ε) and
(28), using that the time t is small enough. To give a rough idea of this step, let us
cutoff large velocities. In particular, we ignore the factors |vj +1 − vk | in (20). Then,
the string of operators can be estimated brutally by
for some C > 0, where the factorial growth comes from the sum in (19). On the
other hand, the ordered time-integration yields t n /n!, so that the series expansion is
j
bounded by a geometric series n C1 (C2 t)n , for positive C1 , C2 .
In the second step, one shows the term by term convergence of (27)–(28). Here
the short time restriction does not enter anymore.
For more details on Lanford’s proof, we refer to [6, 14, 21].
We conclude with some remarks.
1. The time t0 is explicitly computable. It turns out to be a fraction of the mean free
time between collisions. This time limitation is purely technical.
2. Lanford’s original proof was qualitative: it does not make explicit the rate
of convergence. This can be obtained with some extra care, along the same
arguments [7, 19].
A Brief Introduction to the Scaling Limits and Effective Equations in Kinetic Theory 197
where
:
ZN := dZN f0⊗N (ZN ) 1{|xi −xk |>ε} (ZN )
R3N ×R3N 1≤i<k≤N
The Boltzmann equation is suited to the description of rarefied gases, and one can
ask whether a useful kinetic analysis can be applied also to the case of a dense gas.
To introduce the problem, let us revisit the Boltzmann-Grad limit in an alternative
way. Let ε be a small scale parameter denoting the ratio between the microscopic
and the macroscopic scale, for instance the inverse number of atomic diameters
necessary to cover 1 meter, or the inverse number of atomic characteristic times
necessary to cover 1 s. Then, scale space and time by ε in the equations of motion
(in our case, the hard-sphere hierarchy). We need to specify the number of particles
N. In a box of side 1, there should be N ≈ ε−3 particles if one assumes that the
intermolecular distance is of the same order of the molecular diameter. The number
of collisions of a given particle per macroscopic unit time would be ε−1 . As we have
seen, in a low-density regime, N scales differently, namely N ≈ ε−2 , the number
of collisions per unit time is finite and the one-particle distribution function satisfies
the Boltzmann equation.
A variety of possible scalings describes different physical situations (see the next
section). For instance, the gas may be dense, N = O(ε−3 ) and the particles are
weakly interacting via a smooth two-body potential √ φ. To express the weakness of
the interaction, we assume that φ is rescaled by ε. Since φ varies on a scale ε
(in macroscopic unities), the force will be O( √1ε ) and act on a time interval O(ε).
√
The variation of momentum due to the single scattering is O( ε), and the number
of particles met by a typical particle is O( 1ε ). Hence, the total momentum variation
for unit time is O( √1ε ). However, in the case of a homogeneous gas and symmetric
forces, this variation should
√ be zero in the average. The computation of the variance
leads to a result 1ε O( ε)2 = O(1). Therefore, based on a central-limit type of
198 M. Pulvirenti and S. Simonella
argument, we expect that in the kinetic limit a diffusion equation in the velocity
variable holds.
At the level of kinetic equations, consider a collision operator of Boltzmann type,
for a spherically symmetric, smooth potential φ = φ(x). We assume for simplicity
the potential to be short-range, namely φ(x) = 0 if |x| > 1. The collision operator
QB is given by (6), with (7) replaced by
v = v − ω[ω · (v − v1 )]
where ω is the unit vector in the direction of the transferred momentum, while n is
the impact parameter.1 The potential φ enters in the determination of ω.
v − v1
θ ω
n v − v1
1
QεB (f, f )(x, v) = dv1 dn (v − v1 ) · n
ε 2
S+
1 ' (
= dvdv1 dn U · n u(v + p) + u(v1 − p) − u(v) − u(v1 ) f (v)f (v1 )
2ε 2
S+
1 Note that this is not the conventional form for the Boltzmann equation and usually the factor
1 '
≈ dvdv1 dn U · n p · (∇v u(v) − ∇v1 u(v1 )) + (33)
2ε 2
S+
1 1 2
(
u(v)pα pα +
2
∂α,α ∂α,α u(v1 )pα pα f (v)f (v1 ) ,
2 2
α,α
1
Tα,α (U ) := dn U · n pα pα ,
2ε 2
S+
we write
+∞ 3
1 x(s) 1 1 x(s)
pα = − ds √ ∇xα φ =− √ √ ds dk i kα ei k· ε φ̂(k) .
−∞ ε ε 2π ε R3
Then
3
1 1
Tα,α = − dn U · n
2π 2ε2 2
S+
x(s1 ) x(s2 )
ds1 ds2 dk1 dk2 (k1 )α (k2 )α eik1 · ε eik2 · ε φ̂(k1 )φ̂(k2 ).
But x(s)
ε ≈ n+
Us
ε . Therefore, setting y(s) = n + U s (after rescaling times) we
have that
3
1 1
T α,α =− dn U · n
2π 2 2
S+
ds1 ds2 dk1 dk2 (k1 )α (k2 )α eik1 ·y(s1) eik2 ·y(s2) φ̂(k1 )φ̂(k2 ).
Next we write
eik1 ·y(s1) eik2 ·y(s2) = ei(k1 +k2 )·y(s1) eik2 ·U (s2 −s1 )
200 M. Pulvirenti and S. Simonella
hence we arrive to
(2π)
Tα,α ≈ − dk φ̂ 2 (k) δ(k · U ) kα kα =: aα,α (U ) . (35)
2
(2π)4 1
aα,α (U ) = − dρ φ̂ 2 (ρ) ρ 3 d k̂ δ(Û · k̂) , (36)
2 |U |
where Û is the versor of U . Here we are using that, due to the spherical symmetry,
φ̂ depends on k through |k| only. Setting
∞
B=π dρ φ̂ 2 (ρ) ρ 3 (37)
0
"
and computing d k̂ δ(Û · k̂), we conclude that
B
aα,α (U ) = δα,α − Ûα Ûα . (38)
|U |
B is the kinetic constant coding all the information on the microscopic potential.
We turn now to the evaluation of the first order terms in (33), i.e.
1 1 ⊥
T (U ) := dn U · n pU , T⊥ (U ) := dn U · n pU
2ε 2
S+ 2ε 2
S+
where p = (pU , pU⊥ ) and p > 0 is the projection of p over −U/|U |. Note that T
U ⊥
is vanishing by symmetry. On the other hand, pU = (ω · U )2 /|U | = p2 /|U | so that
1 1 2B
T (U ) = Tα,α (U ) ≈ aα,α (U ) = .
|U | α |U | α |U |2
A Brief Introduction to the Scaling Limits and Effective Equations in Kinetic Theory 201
In conclusion,
where
B
Lu(v, v1 ) := −2 (v − v1 ) · (∇v u(v) − ∇v1 u(v1 )) + T r(a ⊗ D 2 u)(v, v1 )
|v − v1 |3
where T r(a ⊗ D 2 u)(v, v1 ) = α,α aα,α (v − v1 )∂v2α ,vα u(v) and a = (aα,α )α,α is
given by (38).
This leads to introduce the Landau operator, defined by
QL (f, f )(x, v) := dv1 ∇v a(v − v1 ) (∇v − ∇v1 )f (x, v)f (x, v1 ) . (40)
The collision operator QL has been introduced by Landau in 1936 for the study of
a weakly interacting dense plasma [15] and
(∂t + v · ∇x )f = QL (f, f )
2 The Landau equation was obtained from the Boltzmann equation for cutoffed Coulomb potential
(truncated both at short and large distances). Actually the word “Coulomb” is frequently used for
the Landau equation with kernel singularity |U1 | (see (38)), which is somehow misleading. In fact
as we have seen, this singularity is always present.
202 M. Pulvirenti and S. Simonella
Let us give a unified picture of the different regimes discussed so far, leading to the
Boltzmann and the Landau equation.
The starting point is always a classical system of N identical particles of
unit mass. Microscopic positions and velocities are denoted by q1 , · · · , qN and
v1 , · · · , vN . Let τ be the microscopic time. The Newton’s equations read:
d d
qi = vi , vi = F (qi − qj ) (42)
dτ dτ
j =1,··· ,N
j =i
x = ε q, t = ετ ,
and ε has to be sent to zero to extract the essential macroscopic features. Note that
the velocity remains unscaled. In these new variables, the system reads:
d d 1 xi − xj
xi = vi , vi = F . (43)
dt dt ε ε
j =1,··· ,N
j =i
d d 1 xi − xj
xi = vi , vi = − √ ∇φ , (45)
dt dt ε ε
j =1,··· ,N
j =i
where
TNε W N = ε
Tk, WN , (47)
1≤k<≤N
xk − x
ε
Tk, WN = ∇φ · (∇vk − ∇v )W N . (48)
ε
N
The marginals fjN (t) satisfy the BBGKY hierarchy
j =1
⎛ ⎞
j
1 N −j
⎝∂t + vk · ∇k ⎠ fjN = √ Tjε fjN + √ ε3 Cjε+1 fjN+1 , (49)
ε ε
k=1
204 M. Pulvirenti and S. Simonella
j
Cjε+1 = ε
Ck,j +1 ,
k=1
ε N −3 xk − xj +1
Ck,j +1 fj +1 (Zj ) = −ε dvj +1 dxj +1 F · ∇vk fjN+1 (Zj , xj +1 , vj +1 )
ε
One should note that this hierarchy has the same structure of (18), but now
we are considering a smooth and weakly rescaled potential φ. In fact Ck,j ε
+1
describes the “collision” of particle k, belonging to the j -particle subsystem, with a
particle outside the subsystem, conventionally j + 1. The dynamics of the j -particle
subsystem is governed by three effects: the free-streaming operator, the collisions
“inside” the subsystem (the Tjε term), and the collisions with particles “outside” the
subsystem (the Cjε+1 term).
We can complement the above equations with the initial condition
⊗j
fjN (t = 0) = f0 , (51)
⊗j N −j t
fjN (t) =S(t)f0 + √ ε3 S(t − t1 )Cjε+1 fjN+1 (t1 )dt1 + (52)
ε 0
1 t
√ S(t − t1 )Tjε fjN (t1 )dt1 .
ε 0
dX F (X) = 0 ,
A Brief Introduction to the Scaling Limits and Effective Equations in Kinetic Theory 205
which implies Cjε+1 fjN+1 = O(ε) , provided that the second derivatives Dv2 fjN+1 (t)
√ = O(ε − 2 ) , we see that the second term in
7
are bounded uniformly in ε. Since N−jε
the right-hand side of (52) does not give any contribution in the limit. In the same
assumptions,
t
S(t − t1 )Tjε fjN (t1 )dt1 =
0
t (xi − xk ) − (vi − vk )(t − t1 )
F · g(Zj , t1 )dt1
0 ε
i =k
where g is a smooth j -particle function, which is again O(ε) so that the last term in
the right-hand side of (52) is also vanishing in the limit. We are therefore facing the
alternative: either the limit is trivial, or the time evolved marginals are not smooth.
This is indeed bad news: a rigorous derivation of the (expected) Landau equation
seems problematic.
The above difficulty suggests to split fjN (t) into two parts, namely we conjecture
that:
where gjN is the main part of fjN and is smooth, while γjN is small, but strongly
oscillating (hence with large derivatives). The two parts satisfy, by definition,
⎛ ⎞
j
N −j 3 ε N −j
⎝∂t + vk · ∇xk ⎠ gjN = √ ε Cj +1 gjN+1 + √ ε3 Cjε+1 γjN+1 ,
ε ε
k=1
⎛ ⎞
j
1 1
⎝∂t + vk · ∇xk ⎠ γjN = √ Tjε γjN + √ Tjε gjN ,
ε ε
k=1
The remarkable feature of this decomposition is that the singular part can be
eliminated. In fact we have that
N −1 t
f1N (t) =S(t)f0 + √ ε3 S(t − t1 ) C2ε g2N (t1 ) + γ2N (t1 ) dt1 ,
ε 0
206 M. Pulvirenti and S. Simonella
where
1 t
γ2N (t) = √ ds U2ε (s) T2ε g2N (t − s)
ε 0
and U2ε is just the two-particle interacting flow. Indicating by Z2ε (−s) s∈(0,t )
this
flow with final condition Z2ε (0) = Z2 , we have that
1 t x1ε (−s) − x2ε (−s)
γ2N (Z2 , t) = √ ds ∇φ · ∇v1 − ∇v2 g2N (Z2ε (−s), t − s).
ε 0 ε
N −1 3 ε N N −1 3 ε t
∂t + v1 · ∇x1 f1N (t) = √ ε C2 g2 (t) + ε C2 ds U2ε (s) T2ε g2N (t − s) .
ε ε 0
Let u ∈ D be a test function. As already mentioned the first term on the right-hand
side is negligible:
N −1 3 √
√ ε u, C2ε g2N (t) = O ε .
ε
N −1 t
− dz1 dz2 ds ∇v1 u(z1 )
ε 0
x1 − x2 x1 (−s) − x2 (−s)
F F · (∇v1 − ∇v2 )g2N (Z2ε (−s), t − s)
ε ε
∞
≈− dz1 dr dv2 ds ∇v1 u(z1 )
0
x1ε (−εs) − x2ε (−εs)
F (r) F · (∇v1 − ∇v2 )g2N (x1 , v1 , x1 , v2 , t) ,
ε
x1 −x2
after having changed to variables r = ε and s → εs . Here, setting U = v1 − v2 ,
where in the last step we invoked propagation of chaos (g2N ≈ (g1N )⊗2 ) and used
definition (40). Indeed it is not hard to show that
∞ 1 ∞
dr dsF (r)F (r + U s) = dr dsF (r)F (r + U s) = a(U )
0 2 −∞
where a(U ) is the matrix given by (38). Indeed expressing the above identity in
terms of the Fourier transforms we readily arrive to the right-hand side of (34).
Unfortunately, very little is known about the mathematical derivation. We
mention here the only result we are aware of.
Consider the first order (in time) approximation g̃jN of gjN given by
⊗j N −j t
g̃jN (t) = S(t)f0 + √ ε3 S(t − τ )Cjε+1 S(τ )gjN+1 dτ
ε 0
N −j 3 t τ
⊗(j +1)
+ ε dτ dσ S(t − τ )Cjε+1 Ujε+1 (τ − σ )Tjε+1 S(σ )f0 .
ε 0 0
(53)
where D r is any derivative of order r and b > 0. Assume φ ∈ C 2 (R3 ) and φ(x) = 0
if |x| > 1. Assume that the marginals factorize exactly at time zero. Then
t
lim g̃1N (t) = S(t)f0 + dτ S(t − τ )QL (S(τ )f0 , S(τ )f0 ) (55)
ε→0 0
References
1. Alexander, R.K.: The infinite hard sphere system. Ph.D.Thesis, Department of Mathematics,
University of California at Berkeley (1975)
2. Bobylev, A.V., Pulvirenti, M., Saffirio, C.: From particle systems to the landau equation: a
consistency result. Commun. Math. Phys. 319(3), 683–721 (2013)
3. Boltzmann, L.: Lectures on Gas Theory (English edition annotated by S. Brush). University of
California Press, Berkeley (1964, reprint)
4. Cercignani, C.: On the Boltzmann equation for rigid spheres. Transport Theory Stat. Phys.
2(3), 211–225 (1972)
5. Cercignani, C.: Ludwig Boltzmann: The Man Who Trusted Atoms. Oxford University, Oxford
(1998)
6. Cercignani, C., Illner, R., Pulvirenti, M.: The Mathematical Theory of Dilute Gases. Applied
Mathematical Sciences, vol. 106. Springer, New York (1994)
7. Gallagher, I., Saint Raymond, L., Texier, B.: From Newton to Boltzmann: hard spheres and
short–range potentials. Zurich Advanced Lectures in Mathematics Series, vol. 18. EMS, Lewes
(2014)
8. Grad, H.: On the kinetic theory of rarefied gases. Commun. Pure Appl. Math. 2(4), 331–407
(1949)
9. Grad, H.: Principles of the kinetic theory of gases. In: Flügge, S. (ed.) Handbuch der Physik,
vol. 12, pp. 205–294. Springer, Berlin (1958)
10. He, L.: Asymptotic analysis of the spatially homogeneous Boltzmann equation: grazing
collisions limit. J. Stat. Phys. 155, 151–210 (2014)
11. Illner, R., Pulvirenti, M.: Global validity of the Boltzmann equation for a two–dimensional rare
gas in the vacuum. Commun. Math. Phys. 105, 189–203 (1986)
12. Illner, R., Pulvirenti, M.: Global validity of the Boltzmann equation for a two– and three–
dimensional rare gas in vacuum: erratum and improved result. Commun. Math. Phys. 121,
143–146 (1989)
13. King, F.: BBGKY hierarchy for positive potentials. Ph.D. Thesis, Department of Mathematics,
University of California, Berkeley (1975)
14. Lanford, O.E.: Time evolution of large classical systems. In: Moser, J. (ed.) Dynamical
Systems, Theory and Applications (Seattle, 1974). Lecture Notes in Physics, vol. 38, pp. 205–
294. Springer, Berlin (1975)
15. Lifshitz, E.M., Pitaevskii, L.P.: Course of Theoretical Physics “Landau-Lifshits”, vol. 10.
Pergamon Press, Oxford (1981)
16. Maxwell, J.C.: On the dynamical theory of gases. Philos. Trans. Roy. Soc. London Ser. A 157
49–88 (1867)
17. Maxwell, J.C.: The Scientific Letters and Papers of James Clerk Maxwell, vol. 2, pp. 1862–
1873. Cambridge University Press, Cambridge (1995)
18. Pulvirenti, M., Simonella, S.: On the evolution of the empirical measure for the Hard-Sphere
dynamics. Bull. Inst. Math. Acad. Sin. 10(2), 171–204 (2015)
19. Pulvirenti, M., Saffirio, C., Simonella, S.: On the validity of the Boltzmann equation for short–
range potentials. Rev. Math. Phys. 26(2), 1450001 (2014)
20. Simonella, S.: Evolution of correlation functions in the hard sphere dynamics. J. Stat. Phys.
155(6), 1191–1221 (2014)
21. Spohn, H.: Large Scale Dynamics of Interacting Particles. Texts and Monographs in Physics.
Springer, Heidelberg (1991)
22. Truesdell, C., Muncaster, R.G.: Fundamentals of Maxwell’s Kinetic Theory of a Simple
Monatomic Gas: Treated as a Branch of Rational Mechanics. Pure and Applied Mathematics,
vol. 83. Academic, New York (1980)
Statistical Description of Human
Addiction Phenomena
Giuseppe Toscani
1 Introduction
G. Toscani ()
Department of Mathematics, IMATI Institute of CNR, Pavia, Italy
e-mail: giuseppe.toscani@unipv.it
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 209
G. Albi et al. (eds.), Trails in Kinetic Theory, SEMA SIMAI Springer Series 25,
https://doi.org/10.1007/978-3-030-67104-4_7
210 G. Toscani
societies, like conflicts, criminality, and city size formation have been considered
[2–4, 32, 35].
Since social phenomena are deeply based on behavioral aspects of agents,
the microscopic kinetic interactions have been often modeled to reproduce these
features. To our knowledge, the first kinetic model in which psychological and
behavioral components of the agents have been explicitly considered has been
proposed in [47] to model the price formation of a good in a multi-agent market,
consisting of two different trader populations. The kinetic description in [47] was
inspired by the microscopic Lux–Marchesi model [45, 46] (cf. also [41, 42]). The
microscopic trading rules of agents were assumed to depend both on the opinion
of traders like in [55], and on the way they interact with each other and perceive
risks. This last aspect has been done by resorting, in agreement with the pioneering
prospect theory by Kahneman and Twersky [36, 37], to interactions involving a
suitable value function.
Analogous microscopic mechanism has been considered in [33], where the
choice of a particularly adapted value function justified the statistical shape of the
service time distribution in a call center, and, in more generality, the formation of
a number of social phenomena which can be described by a lognormal distribution
[34].
Starting from [19], where the formation of Pareto tails in the wealth distribution
in a western society was studied at different scales, the analytical description of the
stationary distribution of the social phenomenon under investigation was obtained
resorting to a particular asymptotic limit of the kinetic equation of Boltzmann
type, which results in a Fokker–Planck type equation [51, 59], still reminiscent of
the microscopic interaction mechanism. This is particularly evident in the kinetic
description of [33, 34], where the shape of the value function to insert in the
microscopic interaction is deeply connected to the steady state of the Fokker–Planck
asymptotic equation.
The leading idea in [33] was recently applied to the study of the phenomenon of
alcohol consumption in [21]. There, the choice of a new class of value functions,
suitable to model the possibility of addiction in the microscopic interaction, led to a
class of Fokker–Planck equations with a steady state given by a generalized Gamma
distribution [43, 53].
The findings of [21] are in agreement with the exhaustive fitting analysis
presented in [38, 52]. In these papers the analysis of the fitting of real data about
alcohol consumption in a huge number of countries, pushed the authors to conclude
that, among various probability distributions often used in this context, Gamma and
Weibull distributions (particular cases of the generalized Gamma [53]), appeared to
furnish a better fitting with respect to the Log-normal distribution, first proposed by
Ledermann [40] as a reasonable model for the consumption problem.
Statistical Description of Human Addiction Phenomena 211
In Eq. (1) θ, κ and δ are positive constants related to the relevant characteristics of
the phenomenon under study. Moreover, in all cases considered in [21], in agreement
with the fitting analysis in [38, 52], the constant δ belongs to the interval (0, 1]. The
equilibrium state of the Fokker–Planck equation is given the generalized Gamma
density [43, 53]
δ 1 ) *
f∞ (x; θ, κ, δ) = κ
x κ−1 exp − (x/θ )δ . (2)
θ ! (κ/δ)
Similarly to Eq. (1), θ , κ and δ are positive constants related to the relevant
characteristics of the phenomenon under study, and δ ≤ 1. The main difference
between the Fokker–Planck equation (1) and the present one, is that both the
coefficient of diffusion and the drift term are scaled by a factor x δ . This scaling
has no effect on the steady state, so that the new equation (3) has the same steady
state (2) of Eq. (1).
However, this new Fokker–Planck equation seems to be better adapted to
describe the addiction phenomena, since it allows to obtain an explicit rate of
relaxation of the solution towards the equilibrium. Given an initial probability
density f0 (x) with a bounded variance, it can be proven by classical entropy
methods that the (unique) solution to the initial-boundary value problem for
the Fokker–Planck equation (3) converges towards the equilibrium density (2)
exponentially fast in time with explicit rate [56], a result that seems not available
for the solution to (1).
212 G. Toscani
The analysis of the present paper underlines the importance of the generalized
Gamma density in the statistical description of social phenomena. Previous findings
in this direction were concerned with event history and survival analysis [13].
Addiction phenomena which can be described by the Fokker–Planck equation (3)
include alcohol consumption [21], on line gambling [58], as well as the abuse of
the insights of social networking sites [39]. This new form of addiction is very
recent, since online social networking sites reached a very high popularity only
in the last decade, involving more and more individuals of the society to connect
with others who share similar interests. The perceived need to be online was
noticed to often result in compulsive use of these sites, which in extreme cases
may produce symptoms and consequences traditionally associated with substance-
related addictions [39].
In more details, in Sect. 2 we will briefly describe the modeling assumptions
of [21] and [58], relative to the addiction phenomena of alcohol consumption and,
respectively, to web gambling activity. In particular, we will outline the importance
to resort to a variable collision kernel in the underlying linear Boltzmann equation.
A grazing collision limit [59] procedure finally allows to recover the Fokker–Planck
equation (3). This will be the argument of Sect. 3.
A short review of the qualitative analysis of the Fokker–Planck equation (3),
recently obtained in [56], will be done in Sect. 4.
f (x, t) dx = 1.
R+
The density changes in time since individuals connect (and disconnect) many times
in the given period of a day, thus continuously upgrading the time x spent in web
activities. In agreement with the classical kinetic theory of rarefied gases, we will
always denote a single upgrade of the quantity x as a microscopic interaction.
In the phenomena under study, we will focus on two aspects, which appear to be
common and essential in the eventual formation of addiction.
• Assumption A: There is an entry level (represented by values of the variable x
below a certain value x̄) that is accepted by most societies. The assumption of a
moderate quantity of alcohol, an occasional gambling activity or a limited use of
the mobile phone are indeed seen as completely normal.
• Assumption B: There is an objective pleasure in spending time in these
activities. Consequently, it is normally easier to increase the value of the quantity
x than to decrease it. To prevent addiction, it is usual to fix an alarm level
(represented by a suitable value x̄L of the variable x, with x̄L > x̄), that
individuals should not exceed, and to continuously advertise about the dangers
associated with addiction values x > x̄L .
Assumption B, strongly related to human behavior, has been fully considered in
the kinetic modeling, at the level of individual microscopic interactions, in various
papers [21, 33, 34], taking inspiration from the pioneering prospect theory by
Kahneman and Twersky [36, 37] and their representation of value functions.
On the contrary, in [21, 33, 34] Assumption A, mostly related to the collective
behavior of the system of individuals, was not taken into account. The mathematical
translation of the entry level corresponds to assign a different value (frequency)
to the elementary interactions in terms of the value x. A reasonable hypothesis is
214 G. Toscani
In a single update (interaction) the value x of time can be modified for two reasons,
expressed by two terms, both proportional to the value x. In the first one the
coefficient (·), which can assume both positive and negative values, characterizes
the predictable behavior of agents. The second term takes into account a certain
amount of human unpredictability. The usual choice is to assume that the random
variable η is of zero mean and bounded variance, expressed by η = 0, η2 = λ,
with λ > 0. Small random variations of the interaction
√ (4) will be expressed simply
by multiplying η by a small positive constant , with 1, which produces the
new (small) variance λ.
The function plays the role of the value function in the prospect theory of
Kahneman and Twersky [36, 37], and contains the mathematical details of the
expected human behavior in the phenomenon under consideration, namely the fact
that it is normally easier to increase the value of x than to decrease it, in relationship
with the alarm value x̄L . In terms of the variable s = x/x̄L the value functions
considered in [21] to describe alcohol consumption are given by
δ −1)/δ
e(s −1
δ (s) = μ , s ≥ 0, (5)
e(s −1)/δ +1
δ
where 0 < δ ≤ 1 and 0 < μ < 1 are suitable constants characterizing the intensity
of the individual behavior, while the constant > 0 is related to the intensity of the
interaction. Hence, the choice 1 corresponds to small variations of the mean
difference x∗ − x. In (9), the value μ denotes the maximal amount of variation of
Statistical Description of Human Addiction Phenomena 215
x that agents will be able to obtain in a single interaction. Note indeed that the value
function δ (s) is such that
Clearly, the choice μ < 1 implies that, in absence of randomness, the value of x ∗
remains positive if x is positive. As proven in [21], the value function satisfies
and
d d
(1 + s) < (1 − s) . (8)
ds δ ds δ
These properties are in agreement with the expected behavior of agents, since
deviations from the reference point (s = 1 in our case), are bigger below it than
above. Letting δ → 0 in (5) allows to recover the value function
s − 1
0 (s) = μ , s ≥ 0. (9)
s + 1
d < =
ϕ(x) f (x, t) dx = χ(x) ϕ(x∗ ) − ϕ(x) f (x, t) dx . (10)
dt R+ R+
Here expectation · takes into account the presence of the random parameter η in the
microscopic interaction (4). The function χ(x) measures the interaction frequency.
The right-hand side of Eq. (10) measures the variation in density between
individuals that modify their value from x to x∗ (loss term with negative sign) and
agents that change their value from x∗ to x (gain term with positive sign).
In [21], the simplification of the Maxwell molecules, leading to a constant inter-
action kernel χ, has been assumed. This simplification, maybe not so well justified
from a modeling point of view, is the common assumption in the Boltzmann-type
description of socio-economic phenomena [26, 51].
216 G. Toscani
In [27], the Maxwellian assumption has been analyzed in its critical aspects.
There, starting from a careful analysis of the microscopic economic transactions of
the kinetic model, allowed to conclude that the choice of a constant collision kernel
included as possible also interactions which human agents would exclude a priori.
This was evident for example in the case of interactions in which an agent that trades
with a certain amount of wealth, does not receive (excluding the risk) some wealth
back from the market. In strong analogy with the rarefied gas dynamics [9], where
the analysis of the Boltzmann equation for Maxwell pseudo-molecules leads to the
possibility to make use of the Fourier transformed version, this makes clear that, in
the socio-economic modeling, the main advantages of the Maxwellian assumption
are linked to the possibility to obtain analytical results.
Following the approach in [27], we express the mathematical form of the kernel
χ(x) by taking into account Assumption A, which implies that the frequency of
changes which leads to increase the amount of time x is inversely proportional to x.
Hence, in the addiction setting, it seems natural to consider collision kernels in the
form
χ(x) = αx −β , (11)
for some constants α > 0 and β > 0. This kernel assigns a low probability to
happen to interactions in which individuals are subject to a high degree of addiction,
and assigns a high probability to happen to interactions in which the value of the
addiction variable x is close to zero.
The values of the constants α and β can be suitably chosen by resorting to the
following argument. For small values of the x variable, the rate of growth of the
value function (9) is given by
d x μ δ−1
≈ x . (12)
dx δ x̄L x̄Lδ
This shows that, for small values of x, the mean individual growth predicted by the
value function is proportional to x δ−1 . Then, the choice β = δ would correspond to
a collective growth independent of the parameter δ characterizing the value function.
A second important fact is that the individual rate of growth (12) depends linearly
on the positive constant , and it is such that the intensity of the variation decreases
as decreases. Then, the choice
ν
α=
is such that the collective growth remains bounded even in presence of very small
values of the constant . With these assumptions, the weak form of the Boltzmann-
type equation (10), suitable to describe addiction phenomena, is given by
d ν < =
ϕ(x) f (x, t) dx = x −δ ϕ(x∗ ) − ϕ(x) f (x, t) dx . (13)
dt R+ R+
Statistical Description of Human Addiction Phenomena 217
Note that, in consequence of the choice made on the interaction kernel χ, the
evolution of the density f (x, t) is tuned by the parameter , which characterizes
both the intensity of interactions and the interaction frequency.
For any choice of the value function (9), the linear kinetic equation (13) describes
the evolution of the density consequent to interactions of type (4). The parameter
is closely related to the intensity of interactions. In particular, values 1
describe the situation in which a single interaction determines only an extremely
small change of the value x. This situation is well-known in kinetic theory of rarefied
gases, where interactions of this type are called grazing collisions [51, 59]. At the
same time [26], the balance of this smallness with the random part is achieved by
setting
√
η→ η. (14)
In this way the scaling assumptions allow to retain the effect of all parameters
appearing in (4) in the limit procedure. An exhaustive discussion on these scaling
assumptions can be found in [26] (cf. also [33] for analogous computations in the
case of the Log-normal case). For these reasons, we address the interested reader to
these review papers for details.
Letting → 0, the weak solution f (x, t) to the kinetic model (13) converges
towards f (x, t), solution of a Fokker–Planck type equation [26]. Indeed, the limit
density f (x, t) is such that the time variation of the (smooth) observable quantity
ϕ(x) satisfies
d
ϕ(x) f (x, t) dx =
dt R+
δ (15)
μ x λ
ν −ϕ (x) x 1−δ
− 1 + ϕ (x)x 2−δ f (x, t) dx
R+ 2δ x̄L 2
Hence, provided the boundary terms produced by the integration by parts vanish,
Eq. (15) coincides with the weak form of the Fokker–Planck equation
∂f (x, t) λ ∂ 2 2−δ μ ∂ 1 x δ
=ν x f (x, t) + −1 x 1−δ
f (x, t) .
∂t 2 ∂x 2 2 ∂x δ x̄L
(16)
Equation (16) describes the evolution of the distribution density f (x, t) of the
weekly time x ∈ R+ spent on social networks related activities in the limit of the
218 G. Toscani
grazing interactions. The steady state density can be explicitly evaluated [21], and,
by setting γ = μ/λ, it results to be the function
δ 4
γ x
f∞ (x) = f∞ (x̄L )x̄L2−δ x γ /δ+δ−2 exp − 2 −1 . (17)
δ x̄L
By fixing the mass of the steady state (17) equal to one, the consequent probability
density is the generalized Gamma f∞ (x; θ, κ, δ) defined by (2), characterized in
terms of the shape κ > 0, the scale parameter θ > 0, and the exponent δ > 0 that in
the present situation are given by
1/δ
γ δ2
κ= + δ − 1, θ = x̄L . (18)
δ γ
It has to be remarked that the shape κ is positive, only if the constant γ = μ/λ
satisfies the bound
Note that condition (19) holds, independently of δ, when μ ≥ 4λ, namely when the
variance of the random variation in (4) is small with respect to the maximal variation
of the value function. Note that the smallness assumption (19) is typical of addiction
phenomena, where individuals live their addiction without large unpredictable
variations.
The limit δ → 0 in the Fokker–Planck equation (16) corresponds to the drift term
induced by the value function (9). In this case, the equilibrium distribution (17) takes
the form of a lognormal density [33].
Note that for all values δ > 0 the moments are expressed in terms of the
parameters denoting respectively the alarm level x̄L , the variance λ of the random
effects and the values δ and μ characterizing the value function φδ defined in (5).
Going back to the fitting analysis presented in [38, 52], that lead to identify as
correct statistical distributions for alcohol consumption the Gamma and Weibull
ones, we recall that these cases are obtained by choosing δ = 1 and δ = κ,
respectively. In particular, the case of Gamma distribution leads to a mean value of
the addiction equal to the alarm level x̄L , while for the Weibull distribution, where
γ = δ, the mean value of the addiction is given by
1
x f∞ (x) dx = x̄L δ 1/δ−1
! . (20)
R+ δ
Note that, since δ < 1, in this case the mean value is strictly less than the alarm level
x̄L . If for example δ = 1/2, the mean value is equal to x̄L /2. Hence, the Weibull
case corresponds to the situation in which the addiction phenomenon is sensible to
Statistical Description of Human Addiction Phenomena 219
the advertisements about possible dangers. In the general case, the mean takes the
value
!
2 1/δ ! 1 γ + δ
δ δ δ
x f∞ (x) dx = x̄L !. (21)
R+ γ ! 1 γ
+δ−1
δ δ
Exact computation of the mean can be done by choosing, for 1/2 < δ < 1, the value
γ = δ 2 . This choice is such that condition (19) is satisfied. In this case
1
x f∞ (x) dx = x̄L > x̄L .
R+ ! 2− 1
δ
This shows that the alarm level can be exceeded in the presence of a small variance
of the random variation (with respect to the maximal variation of the value function),
which corresponds to a strong addiction phenomenon.
4 Relaxation to Equilibrium
Scaling time in the Fokker–Planck equation (16) allows to write it in the clean
form (3), which outlines the dependence of both the diffusion and drift terms
on the shape κ > 0, the scale parameter θ > 0, and the exponent δ > 0.
Also, Eq. (3) allows to directly recover the generalized Gamma density (2) in
terms of the same parameters. It has to be remarked once more that the limit
procedure described in Sect. 3 leads to Eq. (15), namely to a weak version of
the Fokker–Planck equation (16). Then, suitable boundary conditions have to be
considered, to guarantee the equivalence of the two equations, and consequently
the correct evolution of the main macroscopic quantities, and among them the
mass conservation. The most used are the so-called no–flux boundary conditions,
expressed by
δ
∂ 2−δ
x f (x, t) + x − (κ + 1 − δ)x 1−δ
f (x, t) = 0, t > 0.
∂x θδ x=0
(22)
In presence of the no-flux boundary conditions (22) one can study, without loss of
generality, the initial-boundary value problem for Eq. (3) with a probability density
function, say f0 (x), as initial datum. Then, mass conservation implies that the
solution f (x, t) remains a probability density for all subsequent times t > 0.
The qualitative analysis of the Fokker–Planck equations (3) has been done in the
recent paper [56]. There, the analysis was extended to values of δ in the interval
0 < δ ≤ 2, thus covering generalized Gamma densities that range from the Log-
normal density, corresponding to δ → 0, to Chi-densities, obtained for δ = 2.
220 G. Toscani
In the following, for the sake of completeness, we will briefly present the main
results obtained in [56], as well as the main properties of this class of Fokker–
Planck equations. As extensively discussed in [26], Fokker–Planck type equations
of type (3) can be rewritten in different equivalent forms, each one useful for various
purposes. For given t > 0, let
x
F (x, t) = f (y, t) dy (23)
0
denote the probability distribution induced by the probability density f (x, t),
solution of the Fokker–Planck equation (3). In [56], the writing the Fokker–Planck
equations (3) in terms of the distribution F (x, t), highlighted an interesting feature
of their solutions.
Integrating both sides of Eq. (3) on the interval (0, x), and applying condi-
tion (22) on the boundary x = 0, it is immediate to verify that F (x, t) satisfies
the equation
∂F (x, t) ∂2 δ ∂
= x 2−δ 2 F (x, t) + x − (κ − 1)x 1−δ F (x, t), (24)
∂t ∂x θδ ∂x
The no-flux boundary conditions (22) then guarantee that, for any t ≥ 0
Then it holds
∂ ∂
F (x, t) = m2 G(y, τ ),
∂t ∂τ
while
∂ ∂
F (x, t) = mx m−1 G(y, τ ),
∂x ∂y
and
∂2 2 2m−2 ∂
2 ∂
F (x, t) = m x G(y, τ ) + m(m − 1)x m−2 G(y, τ ).
∂x 2 ∂y 2 ∂y
Statistical Description of Human Addiction Phenomena 221
Hence, substituting into (24) the above identities and using the inverse relation x =
y 1/m , it follows that G(y, τ ) satisfies the Fokker–Planck equation
κ
∂G(y, τ ) ∂2 δ/m ∂
= y 2−δ/m 2 G(y, τ ) + y − −1 y 1−δ/m
G(y, τ ).
∂τ ∂y (θ m )δ/m m ∂y
(27)
Moreover, if F (x, t) satisfies conditions (25) for any t ≥ 0, G(y, τ ) still satisfies
the same conditions for any τ ≥ 0.
Note that Eq. (27) has the same structure of Eq. (24), with the constants κ, θ ,
and δ substituted by θ m , κ/m and δ/m. Consequently, its equilibrium distribution is
given by the generalized Gamma density
κ δ
f∞ y; θ m , , . (28)
m m
∂f (x, t) ∂2 ∂ x κ !
= (xf (x, t)) + − f (x, t) . (29)
∂t ∂x 2 ∂x θ δ δ
The steady state of Eq. (29) is the standard Gamma distribution of shape κ̄ = κ/δ
and scale θ̄ = θ δ
1 1 ) *
f∞ (x; θ δ , κ/δ, 1) = f∞ (x; θ̄, κ̄, 1) κ̄
= x κ̄−1 exp −x/θ̄ . (30)
θ̄ ! (κ̄)
Likewise, the choice m = δ/2 leads to the Fokker–Planck equation with constant
coefficient of diffusion
∂f (x, t) ∂2 ∂ 2 2κ −1
= f (x, t) + x − − 1 x f (x, t) . (31)
∂t ∂x 2 ∂x θδ δ
222 G. Toscani
In this second case, the steady state of Eq. (31) is the Chi-distribution of shape κ̃ =
2κ/δ and scale θ̃ = θ δ/2
2 1 ' (
f∞ (x; θ δ , κ/δ, 2) = f∞ (x; θ̃ 2 , κ̃/2, 2) = x κ̃−1 exp −x 2/θ̃ 2 .
θ̃ κ̃/2 ! (κ̃/2)
(32)
θδ
H (f |f∞ (θ, κ, δ)) ≤ I2−δ (f |f∞ (θ, κ, δ)), (33)
δ2
where, given two probability densities f (x) and g(x), with x ∈ R+ , H (f, g)
denotes the Shannon entropy of f relative to g
f (x)
H (f |g) = f (x) log dx,
R+ g(x)
and, for a given constant β ≥ 0, Iβ (f, g) denotes the weighted Fisher information
of f relative to g
2
d f (x)
Iβ (f |g) = β
x f (x) log dx.
R+ dx g(x)
Statistical Description of Human Addiction Phenomena 223
θδ x̄Lδ x̄Lδ λ
= = .
δ2 γ μ
Note that the rate of exponential convergence towards the equilibrium density
increases with the alarm level x̄L and with the variance λ of the stochastic part of
the microscopic interaction, while it decreases with respect to the maximal amount
of variation μ of the value function (5). Also, the behaviour with respect to the
parameter δ that characterizes the value function (5) is different depending of the
value of x̄L . The rate of convergence decreases with δ if x̄L < 1, while it increases
in the opposite situation. It is remarkable that there is no dependence on δ when the
alarm level x̄L = 1.
5 Conclusions
Acknowledgments This work has been written within the activities of GNFM (Gruppo Nazionale
per la Fisica Matematica) of INdAM (Istituto Nazionale di Alta Matematica), Italy. The research
was partially supported by the Italian Ministry of Education, University and Research (MIUR)
through the “Dipartimenti di Eccellenza” Programme (2018–2022)—Department of Mathematics
“F. Casorati”, University of Pavia and through the MIUR project PRIN 2017TEXA3H “Gradient
flows, Optimal Transport and Metric Measure Structures”.
References
1. Aitchison, J., Brown, J.A.C.: The Log-Normal Distribution. Cambridge University Press,
Cambridge (1957)
2. Bellomo, N., Herrero, M.A., Tosin, A.: On the dynamics of social conflicts looking for the
Black Swan. Kinet. Relat. Models 6, 459–479 (2013)
3. Bellomo, N., Knopoff, D., Soler, J.: On the difficult interplay between life, complexity, and
mathematical sciences. Math. Models Methods Appl. Sci. 23, 1861–1913 (2013)
224 G. Toscani
4. Bellomo, N., Colasuonno, F., Knopoff, D., Soler, J.: From a systems theory of sociology to
modeling the onset and evolution of criminality, Netw. Heterog. Media 10, 421–441 (2015)
5. Ben-Naim, E.: Opinion dynamics: rise and fall of political parties, Europhys. Lett. 69, 671–677
(2005)
6. Ben-Naim, E., Krapivski, P.L., Redner, S.: Bifurcations and patterns in compromise processes.
Physica D 183, 190–204 (2003)
7. Ben-Naim, E., Krapivski, P.L., Vazquez, R., Redner, S.: Unity and discord in opinion dynamics.
Physica A 330, 99–106 (2003)
8. Bertotti, M.L., Delitala, M.: On a discrete generalized kinetic approach for modelling
persuader’s influence in opinion formation processes. Math. Comp. Model. 48, 1107–1121
(2008)
9. Bobylev, A.: The theory of the nonlinear, spatially uniform Boltzmann equation for Maxwellian
molecules. Sov. Sco. Rev. C Math. Phys. 7, 111–233 (1988)
10. Boudin, L., Salvarani, F.: The quasi-invariant limit for a kinetic model of sociological collective
behavior. Kinetic Rel. Mod. 2, 433–449 (2009)
11. Boudin, L., Salvarani, F.: A kinetic approach to the study of opinion formation. ESAIM: Math.
Mod. Num. Anal. 43, 507–522 (2009)
12. Boudin, L., Mercier, A., Salvarani, F.: Conciliatory and contradictory dynamics in opinion
formation. Physica A 391, 5672–5684 (2012)
13. Box-Steffensmeier, J.M., Jones, B.S.: Event History Modeling A Guide for Social Scientists.
Cambridge University Press, Cambridge (2004)
14. Cercignani, C.: The Boltzmann Equation and Its Applications. Springer Series in Applied
Mathematical Sciences, vol. 67. Springer, New York (1988)
15. Chakraborti, A., Chakrabarti, B.K.: Statistical mechanics of money: effects of saving propen-
sity. Eur. Phys. J. B 17, 167–170 (2000)
16. Chatterjee, A., Chakrabarti, B.K., Manna, S.S.: Pareto law in a kinetic model of market with
random saving propensity, Physica A 335, 155–163 (2004)
17. Chatterjee, A., Chakrabarti, B.K., Stinchcombe, R.B.: Master equation for a kinetic model of
trading market and its analytic solution. Phys. Rev. E 72, 026126 (2005)
18. Comincioli, V., Della Croce, L., Toscani, G.: A Boltzmann-like equation for choice formation.
Kinetic Rel. Mod. 2, 135–149 (2009)
19. Cordier, S., Pareschi, L., Toscani, G.: On a kinetic model for a simple market economy. J. Stat.
Phys. 120, 253–277 (2005)
20. Cordier, S., Pareschi, L., Piatecki, C.: Mesoscopic modelling of financial markets. J. Stat. Phys.
134(1), 161–184 (2009)
21. Dimarco, G., Toscani, G.: Kinetic modeling of alcohol consumption (2019). arXiv:1902.08198
22. Drǎgulescu, A., Yakovenko, V.M.: Statistical mechanics of money. Eur. Phys. Jour. B 17, 723–
729 (2000)
23. Düring, B., Matthes, D., Toscani, G.: Kinetic equations modelling wealth redistribution: a
comparison of approaches. Phys. Rev. E 78, 056103 (2008)
24. Düring, B., Markowich, P.A., Pietschmann, J-F., Wolfram, M-T.: Boltzmann and Fokker-
Planck equations modelling opinion formation in the presence of strong leaders. Proc. R. Soc.
Lond. Ser. A Math. Phys. Eng. Sci. 465, 3687–3708 (2009)
25. Feller, W.: Two singular diffusion problems. Ann. Math. 54(2), 173–182 (1951)
26. Furioli, G., Pulvirenti, A., Terraneo, E., Toscani, G.: Fokker–Planck equations in the modelling
of socio-economic phenomena. Math. Mod. Meth. Appl. Scie. 27(1), 115–158 (2017)
27. Furioli, G., Pulvirenti, A., Terraneo, E., Toscani, G.: Non-Maxwellian kinetic equations
modeling the evolution of wealth distribution. Math. Models Methods Appl. Sci. 30(4), 685–
725 (2020)
28. Galam, S.: Rational group decision making: a random field Ising model at T = 0. Physica A
238, 66–80 (1997)
29. Galam, S., Moscovici, S.: Towards a theory of collective phenomena: consensus and attitude
changes in groups. Euro. J. Social Psychol. 21, 49–74 (1991)
Statistical Description of Human Addiction Phenomena 225
30. Galam, S., Zucker, J.D.: From individual choice to group decision-making. Physica A 287,
644–659 (2000)
31. Galam, S., Gefen, Y., Shapir, Y.: Sociophysics: a new approach of sociological collective
behavior. I. Mean-behaviour description of a strike. J. Math. Sociol. 9, 1–13 (1982)
32. Gualandi, S., Toscani, G: Pareto tails in socio-economic phenomena: a kinetic description.
Economics 12(2018–31), 1–17 (2018)
33. Gualandi, S., Toscani, G: Call center service times are lognormal. A Fokker–Planck descrip-
tion. Math. Mod. Meth. Appl. Scie. 28(08), 1513–1527 (2018)
34. Gualandi, S., Toscani, G: Human behavior and lognormal distribution. A kinetic description.
Math. Mod. Meth. Appl. Scie. 29(4), 717–753 (2019)
35. Gualandi, S., Toscani, G: The size distribution of cities: a kinetic explanation. Physica A 524,
221–234 (2019)
36. Kahneman, D., Tversky, A.: Prospect theory: an analysis of decision under risk. Econometrica
47(2), 263–292 (1979)
37. Kahneman, D., Tversky, A.: Choices, Values, and Frames. Cambridge University Press,
Cambridge (2000)
38. Kehoe, T., Gmel, G., Shield, K.D., Gmel, G., Rehm, J.: Determining the best population-
level alcohol consumption model and its impact on estimates of alcohol-attributable harms.
Population Health Metri. 10, 6 (2012)
39. Kuss, D.J.; Griffiths, M.D.: Online social networking and addiction-A review of the psycho-
logical literature. Int. J. Environ. Res. Public Health 8, 3528–3552 (2011)
40. Ledermann, S.: Alcool, Alcoolisme, Alcoolisation, vol. I. Presses Universitaires de France,
Paris (1956)
41. Levy, M., Levy, H., Solomon, S.: A microscopic model of the stock market: CYCLES, booms
and crashes. Econ. Lett. 45, 103–111 (1994)
42. Levy, M., Levy, H., Solomon, S.: Microscopic Simulation of Financial Markets: From Investor
Behaviour to Market Phenomena. Academic, San Diego (2000)
43. Lienhard, J.H., Meyer, P.L.: A physical basis for the generalized Gamma distribution. Quarterly
Appl. Math. 25(3), 330–334 (1967)
44. Limpert, E., Stahel, W.A., Abbt, M.: Log-normal distributions across the sciences: keys and
clues. BioScience 51(5), 341–352 (2001)
45. Lux, T., Marchesi, M.: Scaling and criticality in a stocastich multi-agent model of a financial
market. Nature 397(11), 498–500 (1999)
46. Lux, T., Marchesi, M.: Volatility clustering in financial markets: a microscopic simulation of
interacting agents. Int. J. Theoret. Appl. Finance 3, 675–702 (2000)
47. Maldarella, D., Pareschi, L.: Kinetic models for socio–economic dynamics of speculative
markets. Physica A 391, 715–730 (2012)
48. Mielecka-Kubien, Z.: On the estimation of the distribution of alcohol consumption. Math.
Population Studies 25(1), 1–19 (2018)
49. Naldi, G., Pareschi, L., Toscani, G. (Eds.): Mathematical Modeling of Collective Behavior in
Socio-Economic and Life Sciences. Birkhauser, Boston (2010)
50. Otto, F., Villani, C.: Generalization of an inequality by Talagrand and links with the logarithmic
Sobolev inequality. J. Funct. Anal. 173, 361–400 (2000)
51. Pareschi, L., Toscani, G.: Interacting Multiagent Systems: Kinetic Equations and Monte Carlo
Methods. Oxford University Press, Oxford (2014)
52. Rehm, J., Kehoe, T., Gmel, G., Stinson, F., Grant, B., Gmel, G.: Statistical modeling of
volume of alcohol exposure for epidemiological studies of population health: the US example.
Population Health Metri. 8, 3 (2010)
53. Stacy, E.W.: A generalization of the Gamma distribution. Ann. Math. Statist. 33, 1187–1192
(1962)
54. Sznajd–Weron, K., Sznajd, J.: Opinion evolution in closed community. Int. J. Mod. Phys. C 11,
1157–1165 (2000)
55. Toscani, G.: Kinetic models of opinion formation. Commun. Math. Sci. 4, 481–496 (2006)
226 G. Toscani
56. Toscani, G.: Entropy-type inequalities for generalized Gamma densities. Ricerche di Matem-
atica (2020, in press). arXiv:1909.13658
57. Toscani, G., Brugna, C., Demichelis, S.: Kinetic models for the trading of goods. J. Stat. Phys
151, 549–566 (2013)
58. Toscani, G., Tosin, A., Zanella M.: Multiple-interaction kinetic modelling of a virtual-item
gambling economy. Phys. Rev. E 100, 012308 (2019)
59. Villani, C.: Contribution à l’étude mathématique des équations de Boltzmann et de Landau en
théorie cinétique des gaz et des plasmas. PhD Thesis, University Paris-Dauphine (1998)
Boltzmann-Type Description with Cutoff
of Follow-the-Leader Traffic Models
1 Introduction
A. Tosin ()
Department of Mathematical Sciences “G. L. Lagrange”, Politecnico di Torino, Torino, Italy
e-mail: andrea.tosin@polito.it
M. Zanella
Department of Mathematics “F. Casorati”, University of Pavia, Pavia, Italy
e-mail: mattia.zanella@unipv.it
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 227
G. Albi et al. (eds.), Trails in Kinetic Theory, SEMA SIMAI Springer Series 25,
https://doi.org/10.1007/978-3-030-67104-4_8
228 A. Tosin and M. Zanella
If the road is identified with the real axis and the position of the ith vehicle at
time t ≥ 0 is denoted by xi = xi (t) ∈ R, a general FTL model is expressed by the
following system of ordinary differential equations, cf. [8]:
⎧
⎨ẋi = vi
avim i = 1, 2, . . . , (1)
⎩v̇i = (vi+1 − vi ) ,
(xi+1 − xi )n
where vi = vi (t) ∈ R+ stands for the speed of the ith vehicle whereas a ∈ R+
and m, n ∈ N are parameters characterising the interaction of the ith vehicle with
the (i + 1)th vehicle ahead. In essence, (1) prescribes that the acceleration v̇i is
proportional to the relative speed of the two interacting vehicles through the non-
constant factor
avim
,
(xi+1 − xi )n
si := xi+1 − xi , (2)
Boltzmann-Type Description with Cutoff of Follow-the-Leader Traffic Models 229
i.e. the space gap between a vehicle and the vehicle ahead. The advantage is
that from the kinetic model one can then readily recover a statistical description
of the traffic distributions mentioned before, which would instead be much less
straightforward from a speed-based model.
On the other hand, we consider “collisional” models with cutoff, which is a form
of non-constant collision kernel quite rare in the kinetic literature of vehicular traffic
and also, more in general, of multi-agent systems, see [4, 7, 21, 25]. In particular,
we prove that it is still possible to obtain a precise analytical characterisation of the
asymptotic distributions in spite of the increased non-linearity of the Boltzmann-
type equation caused by the non-constant kernel. It is worth anticipating that the
introduction of a kinetic model with cutoff is not just a theoretical speculation.
As it will be clear in the sequel, it is fundamental in order to ensure the physical
consistency of the interaction schemes derived from (1).
In more detail, the paper is organised as follows. In Sect. 2 we focus on the
binary interaction schemes that may be derived from (1) for m = n and we
consider, in particular, those obtained for n = 1, 2, which will be relevant for the
subsequent development of the theory. In Sect. 3 we introduce a Boltzmann-type
kinetic model of the FTL dynamics based on the previous interaction rules and we
show explicitly that a cutoff interaction kernel is needed, in general, to guarantee
the physical consistency of the statistical description of the system. We anticipate
that the role of such a kernel will be to exclude possible interactions leading to
unphysical negative values of the post-interaction headway. In Sect. 4 we discuss the
application of the asymptotic procedure called the quasi-invariant interaction limit
to our Boltzmann-type setting with cutoff. In particular we show that, in a suitable
regime of the parameters of the binary interactions, it permits to recover a Fokker–
Planck approximation of the original “collisional” equation, whence we compute
explicitly the stationary distributions of the kinetic model. In Sect. 5 we present
some numerical tests which show that, consistently with the theoretical predictions
in the appropriate regime of the microscopic parameters, the numerical solution of
the Boltzmann-type equation approaches for large times the analytically computed
stationary solution of the Fokker–Planck equation. Finally, in Sect. 6 we summarise
the contents of the paper and we propose some concluding remarks.
We observe that, using the headway (2), we may rewrite model (1) in the form
v̇i ṡi
m = a n, i = 1, 2, . . . ,
vi si
230 A. Tosin and M. Zanella
which allows for a direct integration of the ith equation depending on the values of
the exponents m, n. Throughout the paper, we will focus in particular on the case
m = n, which for n = 1 gives
In both cases, C is an arbitrary integration constant. Since si ∈ [0, +∞) and a > 0,
we observe that in (3) vi grows unboundedly for every C > 0. Conversely, in (4) vi
1
increases from 0 to 1/C n−1 , which suggests to fix in this case C = 1 so as to obtain
a unitary maximum dimensionless speed of the vehicles.
Writing (1) with m = n = 1 for the ith and the (i + 1)th vehicle, subtracting the
corresponding equations and using (3), we determine the following equation for the
headway si :
d $ %
ṡi − C si+1
a
− sia = 0,
dt
which implies
ṡi = C si+1
a
− sia + c (5)
s = s + Ct s∗a − s a .
Since the (i + 1)th vehicle does not modify instead its headway when interacting
with the ith vehicle behind, the analogous binary rule for it reads simply s∗ = s∗ .
Boltzmann-Type Description with Cutoff of Follow-the-Leader Traffic Models 231
In order to deal more realistically with partly random binary interactions, which
model the non-deterministic aspects of driver behaviour, we further add to s a
zero-mean stochastic fluctuation, which does not modify on average the main FTL
dynamics. To this purpose, we introduce a random variable η ∈ R such that
where · denotes the expectation with respect to the law of η, and we finally write
with γ := Ct > 0 for brevity. The coefficient s δ with δ > 0 gives the intensity
of the stochastic fluctuation. We assume that it increases with s, so that when
a vehicle is close to the leading vehicle it mostly follows the deterministic FTL
model. Conversely, when it is far from the leading vehicle it is mostly prone to the
randomness of the driver behaviour.
For n = 2, which here we regard as the prototype of the cases n > 1, from (4) we
have
si
vi = . (8)
a + si
Proceeding like in Sect. 2.1, we determine now the following equation for the
headway si :
d 1 1
ṡi − a − = 0,
dt a + si a + si+1
namely
1 1
ṡi = a − +c (9)
a + si a + si+1
s = s + I (s, s∗ ) + s δ η,
(11)
s∗ = s∗ ,
where the interaction function I has the property that I (s, s∗ ) = −I (s∗ , s). In
order to be physically admissible, these rules have to be such that s , s∗ ≥ 0 for all
s, s∗ ≥ 0, which is clearly obvious for s∗ but not for s .
In general, the possibility to guarantee s ≥ 0 depends strongly on I and on
the exponent δ of the coefficient of the stochastic fluctuation η. For instance, in the
case (10) with δ = 1 it can be proved that the conditions
γ
η≥ − 1, γ < a2
a2
are sufficient to ensure a priori s ≥ 0 for all possible choices of s, s∗ ≥ 0, see [20]
for the details. They amount to saying that the support of η is bounded from the left,
however in such a way that η can take also negative values, which are essential in
order to meet the requirements (6).
The same is instead not true if, for the same interaction rule (10), we consider
e.g., δ = 12 . Indeed, assume that we bound the support of η from the left as η ≥ −η0
for some 0 < η0 < +∞. Then, no matter how small η0 is, if η takes any negative
value η = η̄ ∈ [−η0 , 0) and furthermore s = η̄2 we have
1 1
s =γ − ,
a + η̄ 2 a + s∗
thus every s∗ ∈ [0, η̄2 ) produces s < 0. A totally analogous situation occurs also
for the interaction rule (7) with δ = 12 .
These examples demonstrate that, in general, not all the interactions modelled
by (7) and (10) are physically admissible. Those which are not have to be discarded
Boltzmann-Type Description with Cutoff of Follow-the-Leader Traffic Models 233
from the statistical description of the system dynamics, in order to get the correct
aggregate trends based only on the admissible interactions. This may be achieved
by considering a Boltzmann-type description with cutoff:
d
ϕ(s)f (s, t) ds
dt R+
(12)
1
= χ(s ≥ 0)(ϕ(s ) − ϕ(s))f (s, t)f (s∗ , t) ds ds∗ ,
2λ R+ R+
plays the role of the cutoff (in particular, non-constant) collision kernel. Specifically,
it discards the interactions producing s < 0, which in this way do not contribute to
1
the evolution of f . Finally, the coefficient 2λ on the right-hand side comes from
the general form of Boltzmann-type equations with non-symmetric interactions,
cf. [18], the parameter λ > 0 representing a relaxation time (in other words, λ1
is the interaction frequency).
The presence of the non-constant collision kernel χ(s ≥ 0) makes it more
difficult to extract from (12) information on the aggregate trends of the system,
such as e.g., the evolution of the statistical moments of the distribution function f :
d
f (s, t) ds = 0,
dt R+
namely the conservation of the mass of the vehicles. This condition also implies that
it is possible to understand f as a probability density, up to possibly normalising it
with respect to the constant total mass.
Choosing instead ϕ(s) = s in (12) we discover
dM1 1
= χ(s ≥ 0)(I (s, s∗ ) + s δ η)f (s, t)f (s∗ , t) ds ds∗ .
dt 2λ R+ R+
234 A. Tosin and M. Zanella
We notice that if the binary interactions are such that the condition s ≥ 0 may be
guaranteed a priori, like in the case (10) with δ = 1, then χ(s ≥ 0) ≡ 1 and
dM1 1
= I (s, s∗ )f (s, t)f (s∗ , t) ds ds∗ = 0,
dt 2λ R+ R+
because I is antisymmetric with respect to the line s∗ = s. In this case, also the first
moment of f , namely the mean headway of the vehicles, is conserved. However,
this is in general not the case of the models that we are considering.
The difficulty to deal with the strongly non-linear Boltzmann-type equation (12)
may be bypassed in suitable asymptotic regimes, which allow one to transform (12)
in a kinetic model more amenable to analytical investigations. This does not only
include the determination of the statistical moments Mk but also the explicit
computation of the stationary distribution, say f ∞ = f ∞ (s), which in this context
plays the role of the Maxwellian distribution of the classical kinetic theory in that it
depicts the emerging trend when interactions are close to equilibrium.
4 Fokker–Planck Asymptotics
a = Var(η) = , λ= (13)
2
Remark 1
(i) Assumption 4.14.1 implies, in particular, that f has a minimum number of
moments bounded. Moreover, it implies that log s ∈ Lp (R+ ; f (·, t) ds) for
every p ∈ [0, p]. Indeed, since |log s| ≥ 1 for s ∈ (0, e−1 ) ∪ (e, +∞), we
have:
1
e
|log s|p f (s, t) ds ≤ |log s|p f (s, t) ds
R+ 0
e +∞
+ f (s, t) ds + |log s|p f (s, t) ds
1
e e
(ii) For every a ≥ 0, Assumption 4.14.1 implies that P(Y < −a) = P(Y > a),
hence in particular that P(Y < −a) = 12 P(|Y | > a).
To begin with, we observe that χ(s ≥ 0) = 1 − χ(s < 0), therefore we may
rewrite (12) as
d 1
ϕ(s)f (s, t) ds = ϕ(s ) − ϕ(s)f (s, t)f (s∗ , t) ds ds∗
dt R+ R+ R+
1
− χ(s < 0)(ϕ(s ) − ϕ(s))f (s, t)f (s∗ , t) ds ds∗
R+ R+
s∗ 1
s∗ − s = log + 2 s∗¯ log2 s∗ − s ¯ log2 s ( → 0+ ) (15)
s 2
s + γ (s∗ − s )
Y <− √ =: b (s, s∗ ), (16)
s
√ s̄ ∈ (min{s, s∗ }, max{s, s∗ }). Using (15), we see that the remainders o(1),
where
o( ) denote terms which are bounded in s, s∗ because:
(i) s is bounded away from 0 and +∞ thanks to the compactness of the support
of ϕ and all of its derivatives;
(ii) Assumption 4.14.1 and Remark 11 ensure the f -integrability of the powers of
s∗ and |log s∗ |, hence also of their products owing to Hölder’s inequality, on
R+ for p sufficiently large.
Boltzmann-Type Description with Cutoff of Follow-the-Leader Traffic Models 237
The goal is now to take → 0+ in (17). Passing formally to the limit under the
integrals, we have to handle expressions of the form |Y |k χ(Y < b (s, s∗ )) for
k = 0, . . . , 3. From Hölder’s inequality we get
1 1 1 1
|Y |k χ (Y < b (s, s∗ )) ≤ |Y |kq q χ (Y < b (s, s∗ ))r r = |Y |kq q P(Y < b (s, s∗ )) r ,
→0+
R (f, f )[ϕ] −−−→ 0.
2 We recall that Chebyshev’s inequality states that P(|X − μ| ≥ kσ ) ≤ k12 , where X is a real-
valued random variable with finite expectation μ and finite non-zero variance σ 2 and k > 0. Here
we apply it for X = Y , with μ = 0 and σ 2 = 1, and k = |b (s, s∗ )|.
238 A. Tosin and M. Zanella
1
+ ϕ (s)sf (s, t) ds. (18)
2 R+
If we denote
1 2 $ %
∂t f = ∂s (sf ) − γ ∂s (L(t) − log s)f . (20)
2
In summary, (18) and (20) represent the weak and the strong form of the asymptotic
model which approximates (12) in the quasi-invariant regime (13) of the interac-
tions (7).
Notice that, because of the compactness of suppϕ, the Fokker–Planck equa-
tion (20) comes without conditions at s = 0 and s → +∞. Boundary conditions
may be set by imposing, for instance, the fulfilment of some conservation properties.
In particular, as it will be clear in a moment, in this context it is useful to guarantee
that model (20) conserves in time the first moment of f , i.e. the mean headway of
the vehicles. To study the evolution of M1 , we multiply (20) by s and we integrate
on R+ . Recalling the definition (19), we discover:
+∞
dM1 1 2
= s ∂s f (s, t) − γ L(t)sf (s, t) + γ s log sf (s, t) ,
dt 2 0
therefore M1 is conserved if, for all t > 0, the terms sf (s, t), s 2 ∂s f (s, t) and
s log sf (s, t) vanish when s → 0+ and s → +∞. Sufficient conditions for this are
that, for all t > 0, f (s, t) and ∂s f (s, t) are bounded in s = 0 and are infinitesimal
of order greater than 2 for s → +∞.
Boltzmann-Type Description with Cutoff of Follow-the-Leader Traffic Models 239
Next, we may use (20) to obtain the stationary distribution f ∞ , which satisfies
1
∂s (sf ∞ ) − γ (L∞ − log s)f ∞ = 0,
2
where L∞ := limt →+∞ L(t) is so far unknown. This differential equation can be
easily solved by separation of variables. Its unique solution with unitary mass is the
function
√
γ ∞ 2
f (s) = √ e−γ (log s−L ) ,
∞
s π
L∞ + 4γ
1
M1∞ := sf ∞ (s) ds = e ,
R+
which, owing to the conservation in time of M1 , has to coincide with the constant
mean headway of the system, say h > 0. Therefore we can express L∞ = log h− 4γ 1
see Fig. 1.
In the transportation engineering literature, the log-normal distribution has often
been reported to fit well the empirical data of vehicle interspacings, see e.g., [12, 16].
Fig. 1 The log-normal distribution (21) predicted by model (7) in the quasi-invariant regime (13)
for: h = 1 and various γ > 0 (left); γ = 1 and various h > 0 (right)
240 A. Tosin and M. Zanella
s s 1−a
τ := = . (22)
v C
Without loss of generality, let us fix C = 1. If, consistently with the quasi-
invariant regime (13), we assume that a is small, in particular a < 1, we can use
the distribution (21) together with the transformation (22) to obtain the stationary
distribution g ∞ = g ∞ (τ ) of the time headway:
1 a
g ∞ (τ ) = τ 1−a f ∞ (τ 1/(1−a))
1−a
√ !2
γ − γ log τ −(1−a) log h− 4γ
1
= √ e (1−a)2 ,
τ (1 − a) π
approach presented here offers a more general and organic explanation grounded on
simpler and sounder first principles.
Finally, from (21) and the transformation (3) with C = 1 we derive the stationary
distribution k ∞ = k ∞ (v) of the speed v in the quasi-invariant limit (13), i.e. in
particular for a small:
1 1−a ∞ 1/a
k ∞ (v) = v a f (v )
a
√ !2
γ − γ2 log v−a log h− 4γ1
= √ e a .
va π
1
a= √ , Var(η) = , λ= , (23)
2
with 0 < 1 as usual. The scaled interaction rules take then the form
s∗ − s √
s = s + γ √ √ + sY,
(1 + s)(1 + s∗ )
s∗ = s∗ ,
After rewriting (12) in the form (14), we observe that s < 0 implies
1 s∗ − s γ − 1√
Y < −√ s + γ √ √ ≤ √ s =: b (s), (24)
s (1 + s)(1 + s∗ )
s∗ −s √
whence χ(s < 0) ≤ χ(Y < b (s)). Moreover, (1+√s)(1+ s∗ )
≤ |s∗ − s|. Thus,
∞
for ϕ ∈ Cc (R+ ) we estimate:
> &
s
|R (f, f )[ϕ](t)| ≤ χ (Y < b (s)) ϕ (s) γ |s∗ − s| + |Y |
R+ R+
1 √
+ ϕ (s) γ 2 (s∗ − s)2 + 2γ s |s∗ − s| |Y | + sY 2
2
1 √
+ ϕ (s̄) γ 3 2 |s∗ − s|3 + 3γ s(s∗ − s)2 |Y |
6
?
√ 3/2 3
+3γ s |s∗ − s| Y + s |Y |
2
f (s, t)f (s∗ , t) ds ds∗ .
where q, r ≥ 1 are chosen like in Sect. 4.1. In view of Assumption 4.14.1, it results
|Y |kq < +∞ for k = 0, . . . , 3. Furthermore, from (24) we see that we can take
so small, in particular < γ1 , that b (s) < 0 for all s > 0. Consequently, invoking
Assumption 4.14.1 and Remark 11 together with Chebyshev’s inequality, we obtain
1 1 1 1 1/r
P(Y < b (s)) r = P(|Y | > |b (s)|) r ≤ = .
2 1/r 21/r b (s)2/r (2s)1/r (γ − 1)2/r
Plugging this into the estimate of |R (f, f )[ϕ](t)|, and recalling that s ∈ suppϕ
is bounded away from 0, +∞ while the powers of s∗ are f -integrable thanks to
Assumption 4.14.1 with p sufficiently large, we conclude
→0+
R (f, f )[ϕ](t) −−−→ 0.
#
In particular, we stress that the term containing s vanishes in the limit because
√
at the denominator is compensated by the factor 1/r with r < 2.
Boltzmann-Type Description with Cutoff of Follow-the-Leader Traffic Models 243
s∗ − s
A (f, f )[ϕ](t ) = γ ϕ (s)
√ √ f (s, t )f (s∗ , t ) ds ds∗
R+ R+ (1 + s)(1 + s∗ )
1 γ 2 (s∗ − s)2
+ ϕ (s) √ 2 √ + s f (s, t )f (s∗ , t ) ds ds∗
2 R+ R+ (1 + s) (1 + s∗ )2
1 γ 3 2 (s∗ − s)3 3γ (s∗ − s)s
+ ϕ (s̄) √ √ + √ √
6 R+ R+ (1 + s)3 (1 + s∗ )3 (1 + s)(1 + s∗ )
√
+ s 3/2 Y 3 f (s, t )f (s∗ , t ) ds ds∗
→0+ 1
−−−→ γ ϕ (s)(s∗ − s) + ϕ (s)s f (s, t )f (s∗ , t ) ds ds∗ ,
R+ R+ 2
d 1
ϕ(s)f (s, t) ds = γ ϕ (s) (M1 (t) − s) f (s, t) ds + ϕ (s)sf (s, t) ds.
dt R+ R+ 2 R+
1 2
∂t f = ∂ (sf ) − γ ∂s ((M1 (t) − s)f ), (25)
2 s
which comes again without conditions at s = 0 and for s → +∞ because of the
compactness of suppϕ. Like in Sect. 4.1, it is convenient to fix these conditions in
such a way that M1 is conserved in time. To this purpose, we multiply (25) by s and
we integrate on R+ to discover:
+∞
dM1 1 2
= s ∂s f (s, t) − γ M1 (t)sf (s, t) + γ s f (s, t)
2
.
dt 2 0
From here we see that, analogously to Sect. 4.1, sufficient conditions for dM dt = 0
1
are the fact that, for all t > 0, f (s, t) and ∂s f (s, t) are bounded at s = 0 and be
infinitesimal of order greater than 2 for s → +∞.
Under such conditions we can set M1 (t) = h for all t ≥ 0, so that from (25) we
obtain in particular the following unique stationary distribution with unitary mass:
Fig. 2 The gamma distribution (26) predicted by model (10) in the quasi-invariant regime (23)
for: h = 1 and various γ > 0 (left); γ = 1 and various h > 0 (right)
namely a gamma probability density function with shape parameter 2γ h > 0 and
rate parameter 2γ > 0, see Fig. 2.
In the transportation engineering literature, also the gamma distribution is
sometimes used to fit the experimental measurements of the vehicle interspacings,
see e.g., [5]. Our derivation demonstrates that it may be justified out of Follow-
the-Leader microscopic dynamics (1) with an appropriate choice of the exponents
m, n.
Recalling (8), we see that the time headway is simply
s
τ= = a + s,
v
hence its asymptotic distribution g ∞ , which is supported in the interval [a, +∞)
because s ≥ 0 implies now τ ≥ a, is obtained by translating f ∞ rightward:
g ∞ (τ ) = f ∞ (τ − a)χ(τ ≥ a).
and is naturally supported in [0, 1], see Fig. 3. Notice that for v → 1− we have
k ∞ (v) → 0. Conversely, for v → 0+ we may have k ∞ (v) → 0 if 2γ h > 1;
k ∞ (v) → (2γ a)2γ h/ !(2γ h) if 2γ h = 1; or k ∞ (v) → +∞ if 2γ h < 1. In the
latter case, the singularity of k ∞ at v = 0 is however integrable.
Boltzmann-Type Description with Cutoff of Follow-the-Leader Traffic Models 245
Fig. 3 The speed distribution (27) with a = 10 for: h = 5 and various γ > 0 (left); γ = 1 and
various h > 0 (right)
We stress that, consistently with the quasi-invariant regime (23) motivating the
form (26) of f ∞ , in both expressions of g ∞ and k ∞ the parameter a has to be
understood as sufficiently large.
If we consider model (10) with δ = 1 then, owing to the discussion at the beginning
of Sect. 3, we can guarantee a priori the fulfilment of the condition s ≥ 0 for all
s, s∗ ≥ 0 with an appropriate choice of the parameters a, γ and of the random
variable η. This implies that χ(s ≥ 0) ≡ 1 in (12), hence, under the same
scaling (23), the quasi-invariant limit simplifies considerably (it basically requires to
deal only with the term A (f, f )[ϕ]) and yields finally the Fokker–Planck equation
1 2 2
∂t f = ∂ (s f ) − γ ∂s ((h − s)f ),
2 s
which differs from (25) only in the coefficient of f in the second order derivative.
The unique stationary solution with unitary mass is now
2γ h
(2γ h)1+2γ e− s
f ∞ (s) = · ,
!(1 + 2γ ) s 2(1+γ )
5 Numerical Tests
We present now several numerical tests, which illustrate the theoretical results
obtained in Sect. 4. In particular, they show that the large time numerical solution
to the Boltzmann-type equation with cutoff (12) is consistently approximated, for
> 0 small, by either stationary distribution (21) and (26) depending on the
assumed model of binary interactions.
For the numerical solution of the Boltzmann-type equation with cutoff (12), we
adopt a direct simulation Monte Carlo (MC) method. We refer the interested reader
to [17, 18] for an introduction. Here, we simply report an essential algorithm which
implements an MC scheme suited to our equation, see Algorithm 1. In particular,
unlike standard MC algorithms, we take into account that some binary interactions
may need to be rejected, if they produce negative post-interaction headways (see
lines 8 to 12 in Algorithm 1). It is worth remarking that, besides updating the
microscopic states of the particles with the MC scheme, we also need to reconstruct
their probability density function at every time step. For this, we recall that several
approaches are possible, such as e.g., standard histograms (which we use in this
paper), the weighted area rule or kernel density estimation-type strategies.
In the following tests, we invariably use a sample of N = 105 particles.
Moreover, for density reconstruction purposes, we take s in a bounded interval
[0, S] ⊂ R+ and we discretise the latter by means of a certain number NS of grid
points. In particular, for the model with n = 1 we use S = 20 and NS = 200, while
for the model with n = 2 we use S = 10 and NS = 100.
Boltzmann-Type Description with Cutoff of Follow-the-Leader Traffic Models 247
We consider first the binary interaction scheme (7) with δ = 12 and the quasi-
invariant scaling (13). In particular, we take for η a centred uniform law, so as to
meet Assumption 4.14.1. Moreover, we prescribe the following initial condition:
1
if 0 ≤ s ≤ 5
f (s, 0) = 5 (28)
0 otherwise,
0.4 0.6
0.3
0.4
0.2
0.2
0.1
0 0
0 5 10 15 20 0 5 10 15 20
Fig. 4 Follow-the-leader model with n = 1. Comparison of the large time numerical solution
of (12) with the Fokker–Planck equilibrium distribution (21) for a decreasing scaling parameter
and two different values of the parameter γ in (7)
248 A. Tosin and M. Zanella
106
105
104
102
100 100
0 5 10 15 20 0 5 10 15 20
We repeat the same tests as in Sect. 5.1 for the binary interaction scheme (10)
with δ = 12 under the quasi-invariant scaling (23). Hence, we compare the large
time numerical solution of the Boltzmann-type equation (12) with the gamma
equilibrium distribution (26) of the Fokker–Planck equation (25) obtained in the
quasi-invariant limit.
Figure 6 confirms that, for sufficiently small ( = O(10−3 ) in this case), the
large time Boltzmann solution approaches consistently the analytical Fokker–Planck
equilibrium. Moreover, Fig. 7 shows that, for decreasing , the cumulative number
of rejections performed by the MC algorithm 1 diminishes and remains constant in
time.
0.4 1
0.8
0.3
0.6
0.2
0.4
0.1
0.2
0 0
0 2 4 6 8 10 0 2 4 6 8 10
Fig. 6 Follow-the-leader model with n = 2. Comparison of the large time numerical solution
of (12) with the Fokker–Planck equilibrium distribution (26) for a decreasing scaling parameter
and two different values of the parameter γ in (10)
Boltzmann-Type Description with Cutoff of Follow-the-Leader Traffic Models 249
104 102
102 101
100 100
0 5 10 15 20 0 5 10 15 20
6 Conclusions
From the technical point of view, treating the cases with δ = 12 has required
to deal with “collisional” models with cutoff. This means that in the Boltzmann-
type equation we have considered a non-constant collision kernel of the form
χ(s ≥ 0), where χ denotes the characteristic function and s is the post-interaction
headway. Such a kernel discards from the statistical description of the system
possible interactions leading to unphysical negative headways and turns out to be
necessary because for δ = 12 it is impossible to rule out a priori such interactions.
On the other hand, for δ = 1 a more standard Maxwellian description may be
adopted, because a priori bounds on η and the parameters of the interactions can be
established which guarantee the non-negativity of the post-interaction headway.
The analytical determination of the stationary distributions mentioned above
has been possible in a particular regime of the microscopic parameters, called the
quasi-invariant regime. Essentially, it corresponds to the case in which each vehicle
interaction produces a very small variation of the headway but the interaction
frequency is very high. In this sense, it is reminiscent of the grazing collision regime
of the classical kinetic theory. In such a regime, the Boltzmann-type equation can
be consistently approximated by a Fokker-Planck equation, which is more amenable
to analytical investigations including the possible explicit computation of the large
time distributions. Nevertheless, the application of this theory to kinetic models with
cutoff is non-standard and has represented the main difficulty to overcome in this
paper from both the analytical and the numerical points of view.
We believe that the techniques discussed in this paper may further foster the
application of kinetic theory methods to new problems in the wide realm of multi-
agent systems, which for various reasons may require non-constant interaction
kernels, see e.g., [7, 25], and whose investigation might have been partly discour-
aged so far by the lack of proper analytical and numerical tools.
Acknowledgments This research was partially supported by the Italian Ministry for Education,
University and Research (MIUR) through the “Dipartimenti di Eccellenza” Programme (2018–
2022), Department of Mathematical Sciences “G. L. Lagrange”, Politecnico di Torino (CUP:
E11G18000350001) and Department of Mathematics “F. Casorati”, University of Pavia; and
through the PRIN 2017 project (No. 2017KKJP4X) “Innovative numerical methods for evolu-
tionary partial differential equations and applications”.
This work is also part of the activities of the Starting Grant “Attracting Excellent Professors”
funded by “Compagnia di San Paolo” (Torino) and promoted by Politecnico di Torino.
Both authors are members of GNFM (Gruppo Nazionale per la Fisica Matematica) of INdAM
(Istituto Nazionale di Alta Matematica), Italy.
References
3. Chen, X., Li, L., Zhang, Y.: A Markov model for headway/spacing distribution of road traffic.
IEEE Trans. Intell. Transp. 11(4), 773–785 (2010)
4. Cordier, S., Pareschi, L., Toscani, G.: On a kinetic model for a simple market economy. J. Stat.
Phys. 120(1), 253–277 (2005)
5. Cowan, R.J.: Useful headway models. Transp. Res. 9(6), 371–375 (1975)
6. Delitala, M., Tosin, A.: Mathematical modeling of vehicular traffic: a discrete kinetic theory
approach. Math. Models Methods Appl. Sci. 17(6), 901–932 (2007)
7. Furioli, G., Pulvirenti, A., Terraneo, E., Toscani, G.: Non-Maxwellian kinetic equations
modeling the evolution of wealth distribution. Math. Models Methods Appl. Sci. 30(4), 685–
725 (2020)
8. Gazis, D.C., Herman, R., Rothery, R.W.: Nonlinear follow-the-leader models of traffic flow.
Oper. Res. 9, 545–567 (1961)
9. Greenberg, I.: The log-normal distribution of headways. Aust. Road Res. 2(7), 14–18 (1966)
10. Gualandi, S., Toscani, G.: Human behavior and lognormal distribution. A kinetic description.
Math. Models Methods Appl. Sci. 29(4), 717–753 (2019)
11. Herty, M., Pareschi, L.: Fokker-Planck asymptotics for traffic flow models. Kinet. Relat.
Models 3(1), 165–179 (2010)
12. Jin, X., Zhang, Y., Wang, F., Li, L., Yao, D., Su, Y., Wei, Z.: Departure headways at signalized
intersections: a log-normal distribution model approach. Transp. Res. C 17(3), 318–327 (2009)
13. Kahneman, D., Tversky, A.: Prospect theory: an analysis of decision under risk. Econometrica
47(2), 263–292 (1979)
14. Klar, A., Wegener, R.: Enskog-like kinetic models for vehicular traffic. J. Stat. Phys. 87(1–2),
91–114 (1997)
15. Ni, D., Hsieh, H.K., Jiang, T.: Modeling phase diagrams as stochastic processes with
application in vehicular traffic flow. Appl. Math. Model. 53, 106–117 (2018)
16. Panichpapiboon, S., Cheng, L.: Irresponsible forwarding under real intervehicle spacing
distributions. IEEE Trans. Veh. Technol. 62(5), 2264–2272 (2013)
17. Pareschi, L., Russo, G.: An introduction to Monte Carlo method for the Boltzmann equation.
ESAIM: Proc. 10, 35–75 (2001)
18. Pareschi, L., Toscani, G.: Interacting Multiagent Systems: Kinetic Equations and Monte Carlo
Methods. Oxford University Press, Oxford (2013)
19. Paveri-Fontana, S.L.: On Boltzmann-like treatments for traffic flow: a critical review of the
basic model and an alternative proposal for dilute traffic analysis. Transp. Res. 9(4), 225–235
(1975)
20. Piccoli, B., Tosin, A., Zanella, M.: Model-based assessment of the impact of driver-assist
vehicles using kinetic theory. Z. Angew. Math. Phys. 71, 52 (2020)
21. Prigogine, I., Herman, R.: Kinetic Theory of Vehicular Traffic. Elsevier, New York (1971)
22. Puppo, G., Semplice, M., Tosin, A., Visconti, G.: Kinetic models for traffic flow resulting in a
reduced space of microscopic velocities. Kinet. Relat. Models 10(3), 823–854 (2017)
23. Toscani, G.: Kinetic models of opinion formation. Commun. Math. Sci. 4(3), 481–496 (2006)
24. Tosin, A., Zanella, M.: Kinetic-controlled hydrodynamics for traffic models with driver-assist
vehicles. Multiscale Model. Simul. 17(2), 716–749 (2019)
25. Tosin, A., Zanella, M.: Uncertainty damping in kinetic traffic models by driver-assist controls.
Math. Control Relat. Fields (2021). https://doi.org/10.3934/mcrf.2021018
26. Villani, C.: Contribution à l’étude mathématique des équations de Boltzmann et de Landau en
théorie cinétique des gaz et des plasmas. Ph.D. Thesis, Paris 9 (1998)
27. Villani, C.: On a new class of weak solutions to the spatially homogeneous Boltzmann and
Landau equations. Arch. Ration. Mech. Anal. 143(3), 273–307 (1998)
28. Wisitpongphan, N., Bai, F., Mudalige, P., Sadekar, V., Tonguz, O.: Routing in sparse vehicular
ad hoc wireless networks. IEEE J. Sel. Area Commun. 25(8), 1538–1556 (2007)
29. Yin, S., Li, Z., Zhang, Y., Yao, D., Su, Y., Li, L.: Headway distribution modeling with regard
to traffic status. In: 2009 IEEE Intelligent Vehicles Symposium, Xi’an, pp. 1057–1062. IEEE,
Piscataway (2009)