Citation for published version:

Milewski, P, Rogers, T & Worsfold, J 2023, 'Density Fluctuations in Stochastic Kinematic Flows', SIAM Journal
on Applied Mathematics, vol. 83, no. 3, pp. 1000-1024. https://doi.org/10.1137/22M1494166

DOI:
10.1137/22M1494166

Publication date:
2023

Document Version
Peer reviewed version

Link to publication

Publisher Rights
CC BY

University of Bath

Alternative formats
If you require this document in an alternative format, please contact:
openaccess@bath.ac.uk
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

Take down policy

If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately
and investigate your claim.

Download date: 16. Feb. 2024

DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS∗
JEREMY R. WORSFOLD , TIM ROGERS , AND PAUL MILEWSKI†

Abstract. At the macroscopic scale, many important models of collective motion fall into the
class of kinematic flows for which both velocity and diffusion terms depend only on particle density.
When total particle numbers are fixed and finite, simulations of corresponding microscopic dynamics
exhibit stochastic effects which can induce a variety of interesting behaviours not present in the large
system limit. In this article we undertake a systematic examination of finite-size fluctuations in a
general class of particle models whose statistics correspond to those of stochastic kinematic flows.
Doing so, we are able to characterise phenomena including: quasi-jams in models of traffic flow;
stochastic pattern formation amongst spatially-coupled oscillators; anomalous bulk sub-diffusion in
porous media; and travelling wave fluctuations in a model of bacterial swarming.

Key words. kinematic flows, interacting particle systems, traffic modelling, organism swarming,
anomalous sub-diffusion

AMS subject classifications. 60H15, 60H30, 58J65, 92C35, 76Z05

1. Introduction. Kinematic flows are systems in which the velocity field of

particles in a volume of space can be expressed entirely in terms of their density. As
well as continuous media, these flows have been employed as ad hoc descriptions of
the movement or alignment of large populations of individuals such as in the model-
ling of pedestrians, animal motion or vehicle traffic where the term kinematic waves
were first used in this context [1]. The other dominant modelling paradigm in these
fields are stochastic microscopic (sometimes individual-based or agent-based ) dynam-
ics. Making a precise connection between these is not immediate, and furthermore
there are important questions about the size and effect of fluctuations induced by the
discreteness of particles which are not present in deterministic, continuum descrip-
tions. These issues have been studied extensively since McKean’s [2] seminal paper in
1967 on stochastic particles interacting with each other and the limiting distribution
as the number of particles tends to infinity. While much of the work done focuses on
this limiting distribution [3], fluctuations for finite particle numbers have also been
examined [4, 5, 6, 7, 8].
Mean-field interacting particle models, such as the one proposed by McKean,
are applicable to a wide variety of areas of research because of the natural way in
which they model particle-like objects interacting with each other. Most notably, the
stochastic Kuramoto model [9, 10] is a specific case of this general paradigm and is
a key model in the study of synchronisation. In biology, interacting particle systems
have been used to model movement of insects such as the dispersion of grasshoppers
[11, 12]. Models for dispersal and aggregation of organisms have been proposed [13,
14, 15, 16, 17] at both the individual particle level and in the continuum limit. Dean
[18] not only found a stochastic partial integro-differential equation (SPIDE) for the
density but expressed the system in terms of its free energy, allowing for connections
to be made to real particle systems in physics. Specific partial differential equations
(PDE) from physics arise from this model such as the Porous Media Equation (PME)
[19] and the granular media equation [20, 21].
The majority of the previous work mentioned above focuses on particles with
∗ Submittedto the editors January 2, 2023.
Funding: This work was funded by EPSRC
† Department of Applied Mathematics, The University of Bath, Bath BA2 7AY, UK

(jw3286@bath.ac.uk, ma3tcr@bath.ac.uk, pam28@bath.ac.uk).

1
2 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

Fickian (independent) noise. In physical systems this is not always the case, however,
as it has been observed that organisms can change their diffusivity based on their
local density [12]. It is therefore useful to generalise the work done by Dean to allow
for systems with non-linear diffusion. By doing so, a wide range of swarming and
dispersion models can be re-interpreted in terms of their individual constituents.
Here, we examine a class of models obtained as a slight specialization of the model
proposed by McKean, in which particle interactions are based on pairwise distances.
By defining the density of particles as a sum of Dirac masses, any integrable function
can be fully expressed in terms of this empirical density. Furthermore, this system is
the same in law to the solution of a certain SPIDE containing a spatiotemporal noise
term of order O(N −1/2 ) which can have profound effects for low numbers of particles
or systems near criticality. To quantify these effects, we explore the statistics of first-
order corrections as generalized Ornstein-Uhlenbeck (OU) processes in both stationary
and non-stationary regimes.
This paper is structured as follows. First, we state our model for the behaviour of
the individual particles and recast it as an effective SPIDE for the empirical density.
Beginning with an exploration of fluctuations around uniform density states, we derive
the general expression for the power spectral density. We apply this to a simple model
for traffic corresponding in the macroscopic limit to the Lighthill Whitham Richards
(LWR) model [1]; there, fluctuations manifest as travelling density waves with a char-
acteristic velocity moving backwards relative to the flow of traffic. Demonstrating the
generalization of our technique to higher dimensions, we show how to characterize
stochastic patterning which develops in a spatially extended Kuramoto model. In the
later sections we explore fluctuations around non-stationary states including models
of porous media and biological aggregation. Our methods provide a general frame-
work for the study of noise-induced phenomena in a broad class of models relevant to
collective motion.
2. Kinematic Interacting Particle Systems. We consider N particles in a
compact one-dimensional domain A, with boundary conditions to be specified later.
Direct particle interactions are pairwise, additive, and depend only on the displace-
ment between particles, not their absolute position in the domain. Each particle also
experiences a stochastic positional noise, with the strength of the noise also given
as an additive function of pairwise particle distance. The position, Xi , of particle
i = 1, . . . , N evolves according to the stochastic differential equation (SDE)
(2.1)
 β
N N
1 X √ 1 X
dXi (t) = f (Xi (t) − Xj (t)) dt + 2D  g(Xi (t) − Xj (t)) dWi (t),
N j=1 N j=1

where D, β ∈ R+ , Wn (t) is a standard one-dimensional Wiener process independent

of other particles. Here, f and g are generalised functions which respectively describe
the contributions to the drift and diffusion of particle i from pairwise interactions
with other particles. The parameter β controls the scaling of the diffusive term, with
typical choices for most applications being β = 1 or β = 1/2.
Our main object of interest is the particle density, defined as a weighted sum of
Dirac masses:
N
1 X
(2.2) ϱ(x, t) = δ(x − Xn (t)).
N n=1
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 3

The number of particles within some region A′ ⊆ A at time t is N

R
A′
ϱ(x, t) dx. More
generally, if F is an arbitrary function, we have
N
1 X
Z
(2.3) ⟨F, ϱ⟩ = F (x)ϱ(x, t) dx = F (Xn (t)) .
A N n=1

Note that ⟨F, ϱ⟩ is a random variable; it is the empirical average of F over the (ran-
dom) empirical density ϱ. Fluctuations in particle locations will induce corresponding
fluctuations in the measurement ⟨F, ϱ⟩, and making judicious choices of F will allow
us to characterize various interesting phenomena in these models. Rather than work
directly with particle locations, we find it more convenient to consider a statistically
equivalent SPIDE formulation. Under mild smoothness assumptions one can show
(see Appendix A and [5]) that ⟨F, ϱ⟩ has the same distribution as ⟨F, ρ⟩, where ρ(x, t)
satisfies
∂ρ ∂   ∂2   r 2D ∂ √ 
2β β
(2.4) = (f ∗ ρ) ρ + D 2 (g ∗ ρ) ρ + ρ (g ∗ ρ) η .
∂t ∂x ∂x N ∂x
Here
R η is spatiotemporal white noise, and ∗ denotes the convolution (a ∗ b)(x) =
A
a(x − y)b(y) dy for some functions a, b : A 7→ R which also obey appropriate
boundary conditions. Note that, crucially, the equivalence in law ⟨F, ϱ⟩ ∼ ⟨F, ρ⟩ is
exact for any N .
In the limit of large N , it is possible to establish a law of large numbers result
ϱ → ρ∞ , where ρ∞ solves the PDE

∂ ρ∞ ∂   ∂2  2β

(2.5) = (f ∗ ρ∞ ) ρ∞ + D 2 (g ∗ ρ∞ ) ρ∞ .
∂t ∂x ∂x
A corresponding central limit theorem exists quantifying fluctuations around ρ∞ for
large but finite N [5].
In what follows we will explore the range of behaviours possible in kinematic
interacting particle systems of the type specified here. In particular, we will investigate
the role of correlations in fluctuations around the PDE limit in shaping emergent
noise-driven dynamics.
3. Fluctuations around a Uniform Density State. In many cases a uniform
distribution of particles is stable for some or all choices of the model parameters. In
this section we study the behaviour of the fluctuations about this state in the case
that β = 1. Despite the system being in this stable state, these fluctuations can
have visible structure akin to stochastic Turing patterns [22, 23]. These patterns have
been observed in cases such as the stochastic Kuramoto model [24]. If we choose
F = e−ikx /2π, k ∈ Z and impose periodic boundary conditions on A = [−π, π], then
ρ̂k := ⟨e−ikx /2π, ρ⟩ are the Fourier modes of the density. We can then write (2.4)
with β = 1 in its equivalent Fourier series form as
N
d ρ̂k 1 X
(3.1) = Ak (ρ̂) + √ Gkn (ρ̂)ηn
dt N n=1

with

fˆℓ ρ̂ℓ ρ̂k−ℓ − 4π 2 k 2 D

X X
(3.2) Ak (ρ̂) = −2πik ĝℓ ĝq ρ̂ℓ ρ̂q ρ̂k−ℓ−q .
ℓ∈Z ℓ,q∈Z
4 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

The derivation of this can be found in Appendix B. The noise term has a correlation
structure defined by
N
X X
(3.3) Bkℓ (ρ̂) = Gkn (ρ̂)(G† (ρ̂))nℓ = 4πDkℓ ĝj ĝm ρ̂j ρ̂m ρ̂k−ℓ−m−j
n=1 j,m∈Z

where G† is the Hermitian conjugate of G. (3.1) is derived by using Itô calculus on

the Fourier series expansion of the density
N
1 X −ikXn (t)
(3.4) ρ̂k (t) = e
2πN n=1

and also expressing (2.1) in terms of its Fourier expansion:

√
ρ̂ℓ fˆℓ eiℓXn dt + 2π 2D
X X
(3.5) dXn = 2π ρ̂ℓ ĝℓ eiℓXn dWn .
ℓ ℓ

We assume (2.5) admits a stable, uniform density state, ρ∗ (x, t) = 1/2π, and that the
fluctuations in each Fourier mode, k, can be expressed as
√
(3.6) ξk (t) = N (ρ̂k (t) − ρ̂∗k (t))
where ρ̂∗k (t) = δ0k /2π and we use the Kronecker delta: δkj = 1 for k = j and √ zero
otherwise for k, j ∈ Z. In real space, we can substitute ρ(x, t) = ρ∗ + ξ(x, t)/ N into
(2.4) with β = 1 and perform a Taylor expansion up to O(N −1/2 ). At O(1), we have
 
2
(3.7) ∂t ρ∗ = ∂x ((f ∗ ρ∗ ) ρ∗ ) + D∂x2 (g ∗ ρ∗ ) ρ∗

all of which are zero given that ρ∗ is stationary in time and space. Thus we are left
with the O(N −1/2 ) terms:
 
2
∂t ξ = ∂x ((f ∗ ξ) ρ∗ + (f ∗ ρ∗ ) ξ) + D∂x2 (g ∗ ρ) ξ + 2 (g ∗ ξ) ρ∗
(3.8) √ √ ∗
ρ (g ∗ ρ∗ ) η .


+ 2D∂x

By noticing that (f ∗ 1/2π) = fˆ0 , we arrive at the following equation for the fluctua-
tions at the uniform state:
    √
ˆ 1 2 2 ĝ0 ĝ0 2D
(3.9) ∂t ξ = ∂x ξ f0 + (f ∗ ξ) + D∂x ĝ0 ξ + (g ∗ ξ) + √ ∂x η.
2π π 2π
Also, a similar Taylor expansion can be performed in Fourier space on (3.1). Using
(3.6) and keeping terms up to O(N −1/2 ) gives
N
1 dξk 1 X ∂ Ak 1 X
(3.10) √ = Ak (ρ̂∗ ) + √ ξℓ + √ Gkn (ρ̂∗ )ηn .
N dt N ℓ∈Z ∂ ρ̂ℓ ρ̂∗ N n=1

We know that Ak (ρ̂∗ ) = 0 since ρ̂∗ is defined as a steady state. Therefore, defining
Jkl := ∂Ak /∂ ρ̂l as the Jacobian and using ∗ to indicate evaluation at the uniform
density state leads us to
N
dξk X
∗
X
(3.11) = Jkℓ ξl (t) + G∗kn ηn (t)
dt n=1
ℓ∈Z
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 5

where we have neglected higher order terms. More explicitly, the Jacobian and noise
correlation in this state can be found using ∂ ρ̂k /∂ ρ̂ℓ = δkℓ and substituting ρ̂∗k = δ0k .
Thus,
h i δkℓ kℓDĝ02
(3.12) ∗
Jkl = δkl −ik(fˆ0 + fˆk ) − Dk 2 ĝ0 (ĝ0 + 2ĝk ) , ∗
Bkℓ =
2π 2
which mirrors the evolution of the fluctuation in real space in (3.9). Note that the
∗
Jacobian recovers the linear stability condition that if Jkk < 0, ∀k then the uniform
state is stable in the deterministic limit N → ∞. Assuming this is the case, each
Fourier mode is damped according to the eigenvalues of the Jacobian while the noise
term continually excites each mode. When evaluated at the uniform state, J ∗ and
B ∗ are diagonal and result in spatiotemporal patterns (this can also be the case for
non-diagonal J ∗ and B ∗ ). To quantify the patterns we take the Fourier transform in
time. This produces a linear system of decoupled Fourier modes
N
∗ ˜
X X
(−iωδkℓ − Jkℓ )ξℓ (ω) = G∗kn η̃n (ω)
ℓ∈Z n=1

with ξ˜k (ω) representing the Fourier transform of ξk (t) with frequency ω and where
cov(η̃n (ω), η̃m (ω ′ )) = δnm δ(ω − ω ′ ). For convenience we write, Φk (ω) = −iω − Jkk
∗
,
and thus the fluctuations can be expressed as
N
ξ˜k (ω) = Φ−1
X
(3.13) k (ω) G∗kn η̃n (ω).
n=1

Since these fluctuations are a stationary, random process we use the Wiener-Khinchin
theorem [25] to represent this process in terms of its spectral decomposition. We
define the power spectrum of the process, ξk (t), as the Fourier transform of its auto-
correlation. Specifically,
Z ∞
1
(3.14) Sξk (ω) = Rξ (τ )e−iωτ dτ
2π −∞ k
hR i
∞
where Rξk (τ ) = E −∞ ξk (t)ξk† (t − τ ) dt . The Wiener-Khinchin theorem states that,
for a stationary random process, this formulation is equivalent to the average of the
square magnitude of the process’ frequency signals

(3.15) Sξk (ω) = E[|ξ˜k (ω)|2 ].

Using the latter formulation, the power spectrum is given by the average squared
magnitude of (3.13),
N ∗
1 X
∗ ∗† Bkk
(3.16) Sξk (ω) = Gkn (G )mk ⟨η̃ n η̃ m ⟩ = .
|Φk (ω)|2 n,m=1 |Φk (ω)|2

In general, the peak of the power spectrum at ξ˜kmax (ωmax ) will give the characteris-
tic length scale of the fluctuations, 2π/kmax , and their periodicity 2π/ωmax . These
patterns can be stationary or oscillatory, which we now illustrate with two examples.
6 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

3.1. Traffic modelling. Often, individual-based modelling of vehicle traffic is

done using discrete cellular automata or so-called follow the leader models [26, 27].
These can be effective at recreating features of traffic flow but are difficult to connect to
continuum models. These continuum models are also vital in vehicle traffic modelling
since they allow for fast simulations of large road networks through numerical PDE
solutions while still capturing many features of traffic [26]. Models based on McKean’s
mean-field interacting diffusion model [3] have not been used in the context of vehicle
traffic despite them being readily expressed as continuum flows in the limit N → ∞
and naturally incorporating stochasticity on the individual level.
It has been noted that the viscous Burgers equation arises in the case of linear
diffusion (g(x) = 1) and purely local coupling in the drift choosing the drift term to
be a Dirac delta, f (x) = δ(x) [28, 29]. While we will use the same symbol for the
Kronecker delta as the Dirac delta, these are distinct and the context will be clear
by the use of either subscripts for the former or as a function for the latter. While
the Lighthill Whitham Richards (LWR) model is able to capture the propagation of
traffic jams as kinematic waves, it cannot explain the formation of phantom jams
from uniform flow. By keeping a linear diffusion (g(x) = 1) but choosing f (x) =
v0 (1 − δ(x)/ρjam ) we obtain the LWR model with a diffusion term. Explicitly, the
density equation is

∂ρ ∂ ∂2ρ
(3.17) + (ρu) = D 2
∂t ∂x ∂x
with u(ρ) = v0 (1 − ρ/ρjam ). This diffusion term is commonly used in finite difference
schemes for the original LWR model [30] to prevent the formation of shock fronts. On
an individual level this can be understood as the particles (cars) trying to move at
the desired maximum speed, v0 , but moving at a reduced velocity which is dependent
on the local density. The linear noise term reflects the random, imperfect decisions
the particles make, irrespective of the particles around them.
The uniform density state is always stable for this model, irrespective of the
choices of ρjam ,v0 and D. Using (3.16), we expect the fluctuations about this state to
obey

k2 D
(3.18) ⟨|ξˆk (ω)|2 ⟩ = .
2π 2 [(kv0 (1 − 2ρ∗ /ρjam ) − ω)2 + k 4 D2 ]

Therefore, this model has density waves on the order of O(N −1/2 ). The speed of these
waves, vc , is determined by the peak of the fluctuation spectra, vc = ωmax /kmax =
v0 (1 − 2ρ∗ /ρjam ). This is the characteristic speed at ρ∗ for the inviscid case, D = 0,
while the average flow of particles is v0 (1 − ρ∗ /ρjam ). These waves travel backwards
relative to the flow of traffic and appear spontaneously, as is the case with phantom
jams. In Figure 1 we can see these density fluctuations moving backwards relative to
the average flow of the particles as predicted.
3.2. Spatially coupled Kuramoto model. In the case that f (x) = − sin(x)
and g(x) = 1 in (2.1), we obtain the stochastic Kuramoto model [31, 10] with no in-
trinsic frequencies where Xn ∈ [0, 2π), n = 1, . . . , N are then the phases of N coupled
oscillators. The Kuramoto model is a simple model for synchronisation but is appli-
cable to many synchronisation phenomena in the real world [32, 33]. The fluctuations
about uniform density for this globally coupled system have been quantified by [24].
Here we extend the Kuramoto model so that the oscillators have phase, ϑ, and fixed
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 7

Fig. 1. Quasi-Phantom Jams moving at the characteristic velocity, vc = v0 (1 − 2ρ∗ /ρjam ) with
ρjam = 0.2, v0 = 0.5 and D = 1. Grey lines show the trajectories in position, x, against time, t, of
N = 200 particles obeying the stochastic LWR model and forming density clusters matching those
predicted in (3.18). Also shown for comparison are the theoretical values for the characteristic wave
velocity (dotted blue); the average speed of the particles in uniform density (dashed red); and the
maximum velocity (dash-dotted green).

position, x ∈ [0, L]d . Previously, we considered particles with spatial coupling but
here the spatial position of the oscillators is fixed and only the phase is variable. The
coupling is now a combination of spatial coupling, K : [0, L]d 7→ R, and the original
phase coupling, f (ϑ). Our modified SDE is thus

N
1 X √
(3.19) dϑn = K(xn − xm )f (ϑn − ϑm ) dt + 2D dWn .
N m=1
P
In this case, we want to study the position-phase density ρ(x, ϑ) = n δ(x−xn )δ(ϑ−
ϑn )/N . While the approach follows from earlier in this section, the details can be
found in Appendix B.1. We express (3.19) in terms of the Fourier modes of the
position-phase density,

(3.20)
 
N
X 1 X
dρ̂k,ℓ = −ℓ2 Dρ̂k,ℓ − 2πiℓLd fℓ′ K̂k′ ρ̂k−k′ ,ℓ−ℓ′ ρ̂k′ ,ℓ′  dt + √ Gk,ℓ,n dWn ,
k′ ,ℓ′
N n=1

where the first index in ρ̂k,ℓ corresponds to the d-dimensional Fourier series in space;
the second index is the Fourier series of the oscillator phase; and where G obeys

Dℓℓ′
GG† (k,ℓ),(k′ ,ℓ′ ) =
 
(3.21) ρ̂k−k′ ,ℓ−ℓ′ .
πLd

−1
The phases of the oscillators are uniformly distributed, ρ̂∗k,ℓ = 2πLd δk0 δℓ0 , with
a
√ similar small perturbation about this stable state to the previous section, φk,ℓ =
N (ρ̂k,ℓ − ρ̂∗k,ℓ ). As before, we use this to linearise (3.20) and keep terms up to
O(N −1/2 ) to obtain
r
D ℓ
(3.22) dφk,ℓ = λk,ℓ φk,ℓ dt + dZk,ℓ
2 πLd
8 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

where Zk,ℓ is a standard complex Wiener process and

 
(3.23) λk,ℓ = −ℓ2 D − iℓ fˆℓ K̂k − fˆ0 K̂0 .

When evaluated at theuniform  density state, the noise becomes uncorrelated between
the Fourier modes as GG† (k,ℓ),(k′ ,ℓ′ ) = δk−k′ δℓ−ℓ′ Dℓℓ′ /2π 2 L2d . Therefore, the N
independent Wiener processes can be replaced with the single complex Wiener process,
Zk,l , for each (k, l). If we now specify that f (x) = − sin(x), g(x) = 1 as in the
Kuramoto model, then we need only consider the ℓ = 1 Fourier mode, φ1 (x, t). This
is a spatially varying version of the complex order parameter often used to describe the
Kuramoto model. At each position x, the magnitude of the order parameter gives the
level of coherence of the oscillators while its argument represents the average phase at
that location. Given the coupling functions, we have that λk,1 = −D + K̂k and thus
the system is stable when D > K̂k . When this is the case, the noise term will still
result in spatial structure of order O(N −1/2 ) for φ1 (x). The strength and lengthscale
for this structure is determined by the spatial coupling K. As an example, we take
the spatial coupling to be
|x|2 |x|2
    
κ 1 1
(3.24) K(x) = √ exp − 2 − exp − 2 .
2π σ1 2σ1 σ2 2σ2
The lengthscale, µ = 2πL/|kmax |, is determined by the closest Fourier modes to the
maxima of
2π 2 σ22 |k|2 2π 2 σ12 |k|2
    
κ
(3.25) K̂(k) = exp − − exp − ,
2πLd Ld Ld
the Fourier transform of (3.24). The maxima are located at
 
2 1 σ1
(3.26) |kmax | = d 2 2 ln .
2πL (σ1 − σ2 ) σ2
To sample a realisation of the system at a moment in time, we use the complex
Ornstein Ulhenbeck process in (3.22) with ℓ = 1. While technically a multivariate
process, we can treat each Fourier mode independently as they are decoupled. For a
complex OU process,
(3.27) dx = −γ dt + ω dZ γ, ω ∈ R+ ,
the stationary distribution is a complex Gaussian with mean zero and variance ω 2 /2γ.
Thus, in this case the variance for each mode k is
−D
(3.28) Σ2k = .
4π 2 L2d λk
(1) (2)
This can be simulated by sampling two Gaussian random variables Zk′ = Yk + iYk
(1,2)
such that Yk ∼ N (0, Σ2k /2). It can be seen from (3.28) that the modes closest to
being unstable are likely to result in large amplitude fluctuations.
In Figure 2, patterns in both the phase and the coherence can be seen showing
that certain regions of oscillators develop non-local synchronisation. Over time, these
patterns fade and reappear at different positions but always with the same separation
µ. Despite the number of particles being large, the parameters are such that the
system is close to the point of instability, and so we observe exaggerated patterns
relating to strong coherence in the oscillators. Such patterns cannot be predicted
under the assumption of deterministic behaviour.
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 9

π 3 π
2 0.20
1
0.15
0 0 0
0.10
-1
-2 0.05

−π -3 −π
−π 0 π −π 0 π

Fig. 2. Stochastic patterns for the spatially coupled Kuramoto model with d = 2. A realisation
of the phase, ψ(x), (left) and coherence, r(x), (right) for the order parameter φ1 (x) = reiψ , sampled
from the zero-mean, Ornstein-Ulhenbeck process in (3.22). The spatial coupling is given in (3.24)
with κ = 2.6602, σ1 = 0.1, σ2 = 0.05, D = 0.1 and N = 106 . Clear patterns of local coherence can
be seen on the predicted lengthscale of µ ≈ 3.06.

4. Fluctuations around Non-uniform Density States. As in Section 3, we

study the fluctuations of order O(N −1/2 ) in the density about a deterministic solution.
Previously, we considered uniform densities but here we look at fluctuations about
non-uniform, steady solutions with general β > 0. In general, we have

1
(4.1) ρ(x, t) = ρ∗ (x, t) + √ ξ(x, t)
N

where ρ∗ (x, t) is now a time-dependent solution of (2.5). Expanding (2.4) using (4.1)
and keeping terms up to O(N −1/2 ) lead us to

(4.2)
∂ξ ∂ ∂2 
[ξ(f ∗ ρ∗ ) + ρ∗ (f ∗ ξ)] + D 2 (g ∗ ρ∗ )2β ξ + 2βρ∗ (g ∗ ρ∗ )2β−1 (g ∗ ξ)

=
∂t ∂x ∂x
√ ∂ √ ∗
ρ (g ∗ ρ∗ )β η .

+ 2D
∂x
Non-uniform, time-dependent distributions are likely to arise when considering models
for aggregation of organisms such as the compact swarms in [15, 16].
4.1. Fluctuations in the Porous Media Equation. In this section we focus
on non-linear diffusion of particles by removing the deterministic interactions between
particles (f coupling) from (2.4). Also, here we will not be studying travelling wave
solutions and thus our domain is A = R meaning we require the density and coupling,
g, to decay sufficiently at infinity for the convolution to be well defined. The linear
noise case where (g ∗ ρ) = 1 is well understood as all particles behave independently
and the result is simply Fickian diffusion.
A general non-linear diffusion is proposed by Okubo and Kareiva [17] as an initial
dispersal period for a general model for aggregation of insects. Density dependent
diffusions have been observed in experiments involving grasshoppers [11] and are often
used to model the group behaviour of organisms. This effect can be replicated by
choosing g(x) = δ(x) in (2.1), meaning that the particles make more pronounced
random movements when they are in proximity to other particles in an attempt to
10 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

find space. Applying this coupling to the large N limit in (2.5) with β = 1, this
becomes the model studied by Okubo and Kareiva:

∂ρ ∂2
ρ3 .


(4.3) = 2
∂t ∂x
Note that here we set D = 1 without loss of generality as, without a drift, the diffusion
coefficient just corresponds to a rescaling of time. For general β, however, this type
of coupling gives rise to a general non-linear diffusion equation:

∂ρ ∂2 1+2β

1 ∂  (2β+1)/2 
(4.4) = ρ + √ ρ η .
∂t ∂x2 N ∂x
In order to quantify the effects of the additional stochastic term, we must first describe
the general behaviour of the thermodynamic limit. In the this limit, N → ∞, (4.4)
becomes the PME [34, 35, 36]

∂ ρ∗ ∂2
(4.5) = ((ρ∗ )m ) , m>1
∂t ∂x2
with m = 1 + 2β. Starting from a point source, this equation has a general solution
called the Barenblatt solution [35, 37] of the form
 
h
2 i1/m−1
(4.6) ρ∗ (x, t) = max t−a C − κ xt−a ,0 , t>0

where a = (m + 1)−1 and κ = a(m − 1)/2m. For m = 3, as studied in [17], this

takes the form of a semi-ellipse with a growing width, while m = 2 corresponds to
a parabola. The
R ∞constant C in (4.6) is determined by ensuring that the total mass
is normalised, −∞ ρ(x, t) dx = 1. Thus, we re-express this solution in terms of the
height and width of the distribution:
"  2 #1/m−1
∗ x
(4.7) ρ (x, t) = h(t) 1 − .
r(t)

We determine the height h(t) in terms of the width r(t) ∝ ta given that the total
mass is unitary (see Appendix C). We find that h(t) = γ/r(t) with

Γ( m + 1 )
(4.8) γ = √ m−1 m 2
πΓ( m−1 )

and the width of the distribution is given by

(4.9) r(t) = γ a(m−1) κ−a ta .

Stochastic versions of the PME have been proposed [38] and for nonlocal PMEs
[39]. Our model is distinct from these since the noise which arises from the finite
particle nature of the system is density dependent. In Figure 4(a) it can be seen that
for small numbers of particles the shape of the distribution is not well approximated
by the solution to the PME and the stochasticity of the particles results in the shape
fluctuating over time. To describe this behaviour better, we return to the stochastic
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 11

(a) (b) m
1.5
0.4 2
3
4

0.2

−2 −1 0 1 2 x
x

Fig. 3. The stochastic and deterministic PME from (4.10). (a) Comparison between the
deterministic limit and a simulation of N = 1000 particles with with β = 1 (m = 3) at time t = 1.
The distribution takes the form of a semi-ellipse. (b) The limiting distribution as N → ∞ given
in (4.5) for various values of m = 1 + 2β. As m → 1 the distribution becomes Gaussian, while for
m = 2 it is a parabola and as m increases, the shape becomes more rectangular.

partial differential equation (SPDE) of (4.4). Rewriting this equation in terms of m

we have that
∂ρ ∂2 1 ∂  m/2 
(4.10) = 2
(ρm ) + √ ρ η .
∂t ∂x N ∂x
The objective is to understand how the shape and position of distribution changes.
Since we know the deterministic solution stays centered on zero, it is most natural
to consider the cumulative moments of the density distribution: ⟨xn , ρ⟩, n ∈ N0 .
Combining this with (4.1), we define the fluctuations in the cumulative moments as
√
(4.11a) Ξn (t) := N (⟨xn , ρ⟩ − ⟨xn , ρ∗ ⟩)
Z ∞
(4.11b) = xn ξ(x, t) dx.
−∞

The moments of the distribution are then also fluctuation processes about the de-
terministic solution. It is well known that finding all the moments of a distribution
is equivalent to knowing the full distribution. However, calculating the first few mo-
ments will be sufficient in characterising the basic properties of the distribution. Here,
we see that the SPDE for the linearised fluctuations in (4.2) becomes
∂ξ ∂2 ∂  ∗ m/2 
= m 2 (ρ∗ )m−1 ξ +


(4.12) (ρ ) η .
∂t ∂x ∂x
Substituting in the deterministic solution, ρ∗ (x, t), this becomes
∂2 x2
  
∂ξ ∂  ∗ m/2 
(4.13a) = mhm−1 2 1− 2 ξ + (ρ ) η
∂t ∂x r ∂x
mhm−1  2 ∂ m

(r − x2 )ξ ′′ − 4xξ ′ − 2ξ + (ρ∗ ) 2 η

(4.13b) = 2
r ∂x
where ξ ′ = ∂ξ/∂x. It is also true that hm−1 r−2 = κt−1 and so
∂ξ mκ  2 ∂  ∗ m/2 
(r − x2 )ξ ′′ − 4xξ ′ − 2ξ +

(4.14) = (ρ ) η .
∂t t ∂x
12 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

Integrating with respect to the positions weighted by xn enables us to express the

equation above in terms of the cumulative moments
(4.15a)
Z ∞
mκ 2 ∞ n ′′
 Z Z ∞ Z ∞
∂ξ
xn dx = r x ξ dx − xn+2 ξ ′′ dx − 4 xn+1 ξ ′ dx
−∞ ∂t t −∞ −∞ −∞
Z ∞  Z ∞
∂  ∗ m/2 
(4.15b) −2 xn ξ dx − xn (ρ ) η dx.
−∞ −∞ ∂x
Using integration by parts to remove the derivates from the fluctuations gives
dΞn mκ 
n(n − 1)r2 Ξn−2 − (n + 2)(n + 1)Ξn + 4(n + 1)Ξn − 2Ξn

(4.16a) =
dt t Z ∞
xn−1 (ρ∗ )
m/2
(4.16b) −n η dx
−∞
∞
mn(n − 1)κ  2
Z
xn−1 (ρ∗ )
m/2

(4.16c) = r Ξn−2 − Ξn − n η dx.
t −∞

The fluctuations in each moment constitute a set of coupled, time dependent, OU

processes which we can express in vector form
(4.17) Ξ̇(t) = A(t)Ξ(t) + b(t)
where b is a correlated white noise vector with correlation matrix (see Appendix D)
Z r
(4.18a) Bpq (t) = 2 xp+q−2 (ρ∗ )m dx
−r

2pqhm rp+q−1 Γ( 2 )Γ(2+ m−1 )
p+q−1 1

Γ( 2 + m−1 )
p+q+3 1 p + q is even
(4.18b) =
0 otherwise.
All moments are initially zero since we start from a point source and so E [Ξ(t)] =
0, ∀t > 0. The solution to a multidimensional time dependent OU process has the
following covariance [40]
Z t  Z t 
′ ′ T ′ ′
cov (Ξ(t)) = exp A(t ) dt cov (Ξ(0)) exp A (t ) dt
0 0
(4.19) Z t Z t  Z t 
′ ′ T
+ dt exp A(s)ds B(t ) exp A (s)ds .
0 t′ t′

For now we focus on the first moment. We find the noise term, B11 (t), to be

 
1
Γ 21 Γ 2 + m−1
(4.20) B11 (t) = 2γ a(3m−1) κa(m−1) t−a(m−1)   .
Γ 52 + m−1
1

We also know that A11 (t) = 0 and so the variance for the first moment (the centre of
mass), σ12 = var (Ξ1 ) is
Z t
2
(4.21a) σ1 (t) = B11 (τ ) dτ
0
 
1
√ 2a a(3m−1) a(m−1) Γ 2 + m−1
(4.21b) = (m + 1) πt γ κ  .
Γ 52 + m−11
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 13

(a) 1.5 (b)

2 1

σ12
0.5
0
0
x

1 (c)

σ22
−2 0.5

0
0 1 2 3 4 5 0 2 4
t t

Fig. 4. Stochastic PME for m = 3 (β = 1). (a) Simulation of N = 15 particles (grey)

with the measured centre of mass in blue showing the fluctuations of the first moment, Ξ1 , on the
order N −1/2 . The support of the solution to the limiting PDE is shown in red, illustrating that
the particles can extend beyond the theoretical shape. Theoretical ensemble variances over time for
the mean displacement (b) and mean squared displacement (c) of particles (blue) compared to the
measured variances from 100 simulations of N = 500 particles (red).

Returning to m = 3, as seen in Figure 3 the density profile is elliptical with width

1/4
growing as r(t) = 2 3t/π 2 , we have γ = 2/π. This means that the variance in the
centre of mass, is
√
3t
(4.22) σ12 = .
π
For the higher moments, this becomes cumbersome to calculate for general m so
here we present only the variance in the fluctuations of the mean squared displacement
for m = 3. We have
q !
1 3
 
0 0 0
(4.23) A(t) = 1 , B(t) = 2π t
0 − 2t 0 1
2π 2

and since both matrices are diagonal, the first and second moments are independent.
Thus,
t  Z t 
1
Z
(4.24a) σ22 = var (Ξ2 ) = B22 (τ ) exp − ds dτ
0 τ s
t
(4.24b) = .
4π 2
In Figure 4 we see how the variance of the first two moments increases with time by
running an ensemble of 100 simulations compared to these results.
4.2. Swarming model. Several models for the aggregation and swarming of
organisms in terms of kinematic flows have been proposed [16, 15, 41]. The majority
of these models can be expressed as the large N limit of our model for appropriate
choices of f, g and β. For example, here we show that the model presented by Milewski
14 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

and Yang [15] is recovered if we choose our deterministic coupling to be

(
ex , x ≤ 0
(4.25) f (x) =
0, otherwise

and our noise coupling is density-dependent as in the previous section: g(x) = δ(x)
and β = 1. The deterministic coupling is intended to replicate right-sensing organisms
attempting to catch up with organisms in front of them and being unaware of those
behind them due to their field of view. The resulting PDE from (2.5) is

∂ρ ∂ ∂2
(ρ (f ∗ ρ)) − D 2 ρ3 = 0.


(4.26) +
∂t ∂x ∂x
Note that this is the same as the kinematic equation in [15] if ρ0 = 3D/2π. On
a periodic domain, this forms a swarm with consistent shape which moves with a
particular velocity. The shape is either a continuous distribution of particles with a
tail or a compact, more symmetric swarm depending on the choice of D. Figure 5
shows the case where a swarm with a long tail forms. The properties such as the speed
and formation of these shapes are studied extensively in [15]. Qualitatively, the shape
of the swarm from particle simulations matches well with the numerical solutions of
the PDE. Going beyond this, we can approximate the fluctuations about the steady
M
√ ξ∈R ,
state using the numerical solutions and a discretised form of the fluctuations,
with spacing h = 2π/M . The spatially discretised fluctuations are O( M ) as the
fluctuation for a particular bin is related to the number of particles within that bin.
Thus we write the discretised fluctuations as
r
N
(4.27) ξ(t) = (ρ(t) − ρ∗ (t)) .
M
We introduce v = (f ∗ ρ∗ ) and ψ = (g ∗ ρ∗ ) so (4.2) becomes

∂ξ ∂ ∂2
√ ∂ √
(4.28) = (vξ + ρ∗ (f ∗ ξ)) + D 2 ψ 2 ξ + 2ρ∗ ψ (g ∗ ξ) + 2D ( ρ∗ ψη).
∂t ∂x ∂x ∂x
The discretised fluctuations can be approximated as another OU process about the
steady state like in previous examples. The result is a standard matrix-vector equation
for a finite difference scheme but with an additional noise term. Because of this
difference, we show how the discretisation is done explicitly. Firstly, we take a grid of
points, i = 1, . . . , M , with equal spacing h = 2π/M . We use a central finite difference
scheme such that
1 1
(4.29) ∇h [ϕi ] = [ϕi+1 − ϕi−1 ], ∆h [ϕi ] = [ϕi+1 − 2ϕi + ϕi−1 ]
2h h2
are the discrete approximations to the gradient and Laplacian of a continuous function
ϕ interpolated at the grid points ϕi ≈ ϕ(ih). The convolutions are approximated as
Z π
(4.30a) (ϕ ∗ a) (x) = ϕ(x − y)a(y) dy
−π
M
X
(4.30b) ≈h ϕi−j aj .
j=1
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 15

Hence, for each grid point, i, we have that

   
N N
dξi X X
(4.31) = h∇h vi ξi + ρ∗i fi−j ξj  + h2 ∆h ψi2 ξi + 2ρ∗i ψi gi−j ξj 
dt j=1 j=1
√ hp i
+ 2Dh∇h ρ∗i ψi ηi

where cov(ηi , ηj ) = δij M = 2πδij /h. This dependence on M for scale of the noise
is due to the number of particles at each site i being inversely proportional to the
number of sites M .
The discretised fluctuations can then be expressed in vector form. Writing C =
circ(c1 , . . . , cM ) to indicate the circulant matrix where C1j = cj , j = 1, . . . , N and
the ith row is i − 1 circular shifts to the right of the first, the finite difference matrix
operators are thus
1
(4.32a) ∇= circ(0, 1, 0, . . . , 0, −1)
2
(4.32b) ∆ = circ(−2, 1, 0, . . . , 0, 1)

The factors of h are absent as they cancel with the factors of h from the convolution
terms in (4.31). The convolutions can also be written in terms of circulant matrices,
M
X
(4.33) ϕi−j aj = [Φa]i
j=1

where Φ = circ(ϕ1 , ϕ2 , . . . , ϕM ). We replace element-wise products of vectors with

diagonal matrix-vector products such as ϕi ai = [diag(ϕ)a]i . Using this, we introduce
the following matrices:

(4.34) Λ = diag(ρ∗ ), V = diag(v) = diag(F ρ∗ ), Ψ = diag(ψ) = diag(Gρ∗ ),

where F = circ(f1 , . . . , fM ) and similarly for G. Now, we can write the discretised
version of (4.28) as

dξ  √ √
= ∇(V + ΛF ) + D∆(Ψ2 + 2ΛΨG) ξ + 2D∇ ΛΨη.

(4.35)
dt
Because ξ is a fluctuation about the steady state, it has zero mean and thus we can
find its covariance, Σ = cov (ξ), as the solution to the continuous Lyapunov equation:

(4.36) AΣ + ΣAT + B = 0

where

(4.37) A = ∇(V + ΛF ) + D∆(Ψ2 + 2ΛΨG)

and
D  √   √ T
(4.38a) B= ∇ ΛΨ ∇ ΛΨ
πh
D
(4.38b) = ∇ΛΨ2 ∇T .
πh
16 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

ρ
ρ∗
0.4 ∗
ρ ± 2σ
ρ

0.2

0 π
−π − π2 0 − π2
x

Fig. 5. Discretised fluctuations about the deterministic limit for the swarming model proposed
by Milewski and Yang [15] with M = 64. Solid line: deterministic limit of the swarming model
presented in (4.26) for D = 0.4. Dashed lines: two standard deviations calculated from the diagonal
elements of the covariance matrix, Σ, in (4.36) from the deterministic limit with N = 1000. His-
togram: distribution of particles from a simulation at a moment in time after the swarm has formed
for N = 1000.

It is worth reiterating that this holds for general coupling functions f and g for which
a numerical approximation to a stationary or travelling wave solution can be obtained.
Figure 5 shows how this can be used for the swarming model mentioned above. The
deterministic limit was determined from a numerical simulation by finite differences
for a given discretisation. The matrices A and B were calculated using ρ∗ and used
to solve the Lyapunov equation in (4.36) to give the correlation structure of the bins
for a given number of particles in the system.
p Here, we only show how the diagonal
elements of this correlation matrix, σ = diag(Σ), illustrate the variation from the
deterministic limit.
5. Discussion. The model presented in this paper is both a specialisation of the
model introduced by McKean [2] and an extension on the model by Dean [18]. The
interpretation of the model as a SPIDE for the density allows clear links to kinematic
fluid flows to be drawn. Many kinematic flows can be understood as the limit of many
interacting particles obeying the simple rules laid out in our model, particularly flows
created to model organisms such as fish or insects. Despite focusing on just a few
specific examples of kinematic flows, we emphasise that this model encompasses a
large class of kinematic flows.
By adding a spatial correlation for phase oscillators it was also shown that this
model can demonstrate interesting pattern formation for systems close to criticality
as the finite-sized effects are amplified. We reiterate that these patterns persist even
for large particle numbers close to criticality, a phenomenon not captured by the
deterministic limit.
We studied uniform flows for linear diffusion cases and quantified the noise-driven
fluctuations. Most notably, a simple individual-based model of traffic related to the
LWR model displays quasi-phantom jams due to the finite number of particles. This
shows that careful consideration needs to be taken when modelling vehicle traffic as a
continuum flow as such effects can be missed. The existence of these quasi-phantom
jams warrants further investigation into traffic models of this type. It is possible that
extending this to be a second-order model (see [26] for examples) would allow other
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 17

features such as actual phantom jams to be explained. Beyond vehicle traffic, second-
order models of this type are used to describe more complex swarming behaviour, for
instance those based off the Cucker-Smale model [42]. Using the techniques developed
here on those models could be of use in the study of other types of swarming behaviour.
Not only fluctuations with wave patterns were studied, but also variations in the
shape of distributions due to fluctuations. In certain cases fluctuations in macroscopic
quantities, such as the cumulative moments for the stochastic PME formulated in
Section 4.1, can be calculated analytically. Crucially, a closed form of the deterministic
solution is required as well as a suitable choice of the macroscopic quantity: ⟨F, ρ⟩.
When this is not the case, we provided an example of how numerical methods can
still be used to describe the expected variation in shape of a discretised solution.
This paper examines just a few types of models which can be expressed with
specific choices of the coupling functions. In particular, both local and non-local de-
terministic couplings have been explored with localised or constant noise correlations
but non-local noise couplings have not. It is possible such non-trivial noise coupling
could result in interesting behaviour.

REFERENCES

[1] M. J. Lighthill and G. B. Whitham, On kinematic waves ii. a theory of traffic flow on long
crowded roads, Proceedings of the Royal Society of London. Series A. Mathematical and
Physical Sciences, 229 (1955), pp. 317–345, https://doi.org/10.1098/rspa.1955.0089.
[2] H. P. McKean, Propagation of chaos for a class of non-linear parabolic equations, Stochastic
Differential Equations (Lecture Series in Differential Equations, Session 7, Catholic Univ.,
1967), (1967), pp. 41–57.
[3] L.-P. Chaintron and A. Diez, Propagation of chaos: a review of models, methods and appli-
cations, 2021, https://arxiv.org/abs/2106.14812.
[4] M. Hitsuda and I. Mitoma, Tightness problem and stochastic evolution equation arising
from fluctuation phenomena for interacting diffusions, Journal of Multivariate Analysis,
19 (1986), pp. 311–328, https://doi.org/10.1016/0047-259X(86)90035-7.
[5] A.-S. Sznitman, Nonlinear reflecting diffusion process, and the propagation of chaos and
fluctuations associated, Journal of Functional Analysis, 56 (1984), pp. 311–336, https:
//doi.org/10.1016/0022-1236(84)90080-6.
[6] A.-S. Sznitman, A propagation of chaos result for Burgers’ equation, Probability theory and
related fields, 71 (1986), pp. 581–613, https://doi.org/10.1007/BF00699042.
[7] P.-H. Chavanis, The generalized stochastic Smoluchowski equation, Entropy, 21 (2019), https:
//doi.org/10.3390/e21101006.
[8] P. H. Chavanis and L. Delfini, Random transitions described by the stochastic Smoluchowski-
Poisson system and by the stochastic Keller-Segel model, Phys. Rev. E, 89 (2014),
p. 032139, https://doi.org/10.1103/PhysRevE.89.032139.
[9] S. H. Strogatz and R. E. Mirollo, Stability of incoherence in a population of coupled os-
cillators, Journal of Statistical Physics, 63 (1991), pp. 613–635, https://doi.org/10.1007/
BF01029202.
[10] J. A. Acebrón, L. L. Bonilla, C. J. Pérez Vicente, F. Ritort, and R. Spigler, The
Kuramoto model: A simple paradigm for synchronization phenomena, Rev. Mod. Phys.,
77 (2005), pp. 137–185, https://doi.org/10.1103/RevModPhys.77.137.
[11] D. Aikman and G. Hewitt, An experimental investigation of the rate and form of dispersal in
grasshoppers, Journal of Applied Ecology, 9 (1972), pp. 807–817, https://doi.org/10.2307/
2401906.
[12] E. Luçon, Large population asymptotics for interacting diffusions in a quenched random
environment, in Particle Systems and PDEs II, vol. 129 of Springer Proceedings in
Mathematics and Statistics, Braga, Portugal, Dec 2013, Springer, pp. 231–251, https:
//doi.org/10.1007/978-3-319-16637-7.
[13] D. Grünbaum and A. Okubo, Modelling social animal aggregations, in Frontiers in Math-
ematical Biology, S. A. Levin, ed., Berlin, Heidelberg, 1994, Springer Berlin Heidelberg,
pp. 296–325, https://doi.org/10.1007/978-3-642-50124-1 18.
[14] D. Grünbaum, Translating stochastic density-dependent individual behavior with sensory con-
18 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

straints to an Eulerian model of animal swarming, Journal of Mathematical Biology, 33

(1994), pp. 139–161, https://doi.org/10.1007/BF00160177.
[15] P. A. Milewski and X. Yang, A simple model for biological aggregation with asymmetric
sensing, Communications in Mathematical Sciences, 6 (2008), pp. 397 – 416, https://doi.
org/cms/1214949929.
[16] A. Mogilner and L. Edelstein-keshet, A non-local model for a swarm, J. Math. Biol, 38
(1999), pp. 534–570, https://doi.org/10.1007/s002850050158.
[17] A. Okubo and P. Kareiva, Some Examples of Animal Diffusion, vol. 14, Springer, 2001,
https://doi.org/10.1007/978-1-4757-4978-6 6.
[18] D. S. Dean, Langevin equation for the density of a system of interacting Langevin processes,
Journal of Physics A: Mathematical and General, 29 (1996), pp. L613–L617, https://doi.
org/10.1088/0305-4470/29/24/001.
[19] P. Blanchard, M. Röckner, and F. Russo, Probabilistic representation for solutions of an
irregular porous media type equation, The Annals of Probability, 38 (2010), pp. 1870 –
1900, https://doi.org/10.1214/10-AOP526.
[20] D. Godinho and C. Quiñinao, Propagation of chaos for a subcritical Keller–Segel model,
Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, 51 (2015), pp. 965 – 992,
https://doi.org/10.1214/14-AIHP606.
[21] F. Bolley, I. Gentil, and A. Guillin, Uniform convergence to equilibrium for granular
media, Archive for Rational Mechanics and Analysis, 208 (2013), pp. 429–445, https://doi.
org/10.1007/s00205-012-0599-z.
[22] A. J. McKane, T. Biancalani, and T. Rogers, Stochastic pattern formation and sponta-
neous polarisation: The linear noise approximation and beyond, Bulletin of Mathematical
Biology, 76 (2014), pp. 895–921, https://doi.org/10.1007/s11538-013-9827-4.
[23] T. Biancalani, D. Fanelli, and F. Di Patti, Stochastic Turing patterns in the Brusselator
model, Phys. Rev. E, 81 (2010), p. 046215, https://doi.org/10.1103/PhysRevE.81.046215.
[24] E. Luçon, Quenched limits and fluctuations of the empirical measure for plane rotators in
random media., Electronic Journal of Probability, 16 (2011), https://doi.org/10.1214/ejp.
v16-874.
[25] N. Wiener, Generalized harmonic analysis, Acta mathematica, 55 (1930), pp. 117–258.
[26] D. Helbing, Traffic and related self-driven many-particle systems, Rev. Mod. Phys., 73 (2001),
pp. 1067–1141, https://doi.org/10.1103/RevModPhys.73.1067.
[27] M. Bando, K. Hasebe, A. Nakayama, A. Shibata, and Y. Sugiyama, Dynamical model
of traffic congestion and numerical simulation, Phys. Rev. E, 51 (1995), pp. 1035–1042,
https://doi.org/10.1103/PhysRevE.51.1035.
[28] L. L. Bonilla, C. J. Pérez Vicente, F. Ritort, and J. Soler, Exactly solvable phase
oscillator models with synchronization dynamics, Phys. Rev. Lett., 81 (1998), pp. 3643–
3646, https://doi.org/10.1103/PhysRevLett.81.3643.
[29] M. Bossy and D. Talay, A stochastic particle method for the Mckean-Vlasov and the Burgers
equation, Mathematics of Computation, 66 (1997), pp. 157–192, https://doi.org/10.2307/
2153648.
[30] C. F. Daganzo, A finite difference approximation of the kinematic wave model of traffic flow,
Transportation Research Part B: Methodological, 29 (1995), pp. 261–276, https://doi.org/
10.1016/0191-2615(95)00004-W.
[31] Y. Kuramoto, Self-entrainment of a population of coupled non-linear oscillators, in Inter-
national Symposium on Mathematical Problems in Theoretical Physics, H. Araki, ed.,
Berlin, Heidelberg, 1975, Springer Berlin Heidelberg, pp. 420–422, https://doi.org/10.
1007/BFb0013365.
[32] B. Eckhardt, E. Ott, S. H. Strogatz, D. M. Abrams, and A. McRobie, Modeling walker
synchronization on the millennium bridge, Phys. Rev. E, 75 (2007), p. 021110, https:
//doi.org/10.1103/PhysRevE.75.021110.
[33] B. Ermentrout, An adaptive model for synchrony in the firefly pteroptyx malaccae, Journal
of Mathematical Biology, 29 (1991), pp. 571–585, https://doi.org/10.1007/BF00164052.
[34] J. L. Vázquez, The porous medium equation: mathematical theory, Oxford University Press
on Demand, 2007.
[35] D. G. Aronson, The porous medium equation, Springer Berlin Heidelberg, Berlin, Heidelberg,
1986, pp. 1–46, https://doi.org/10.1007/BFb0072687.
[36] O. A. Arqub and N. Shawagfeh, Application of reproducing kernel algorithm for solving
dirichlet time-fractional diffusion-gordon types equations in porous media, Journal of Por-
ous Media, 22 (2019), pp. 411–434, https://doi.org/10.1615/JPorMedia.2019028970.
[37] B. Gilding and L. Peletier, On a class of similarity solutions of the porous media equation,
Journal of Mathematical Analysis and Applications, 55 (1976), pp. 351–364, https://doi.
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 19

org/10.1016/0022-247X(76)90166-9.
[38] J. U. Kim, On the stochastic porous medium equation, Journal of Differential Equations, 220
(2006), pp. 163–194, https://doi.org/10.1016/j.jde.2005.02.006.
[39] A. D. Gregorio, Stochastic models associated to a nonlocal porous medium equation, Modern
Stochastics: Theory and Applications, 5 (2018), pp. 457–470, https://doi.org/10.15559/
18-VMSTA112.
[40] C. Gardiner, Stochastic methods, vol. 4, Springer Berlin, 2009.
[41] C. M. Topaz, A. L. Bertozzi, and M. A. Lewis, A nonlocal continuum model for biological
aggregation, Bulletin of mathematical biology, 68 (2006), p. 1601, https://doi.org/10.1007/
s11538-006-9088-6.
[42] F. Cucker and S. Smale, Emergent behavior in flocks, IEEE Transactions on Automatic
Control, 52 (2007), pp. 852–862, https://doi.org/10.1109/TAC.2007.895842.

Appendix A. Stochastic PDE.

For any sufficiently smooth function F supported on A ⊆ R,
N
" #
1 X
Z
(A.1) E F (Xi ) − ϱ(x, t)F (x)dx = 0
N i=1 A

where ϱ is the solution to the SPIDE in (2.4) if E N1 i F (Xi (0)) − ⟨ϱ0 , F ⟩ = 0.

 P 

Firstly, using Itô calculus we know that

h i
2β
(A.2a) dF (Xi ) = F ′ (Xi ) (f ∗ ρ) (Xi ) + DF ′′ (Xi ) (g ∗ ρ) (Xi ) dt
√ β
(A.2b) + 2DF ′ (Xi ) (g ∗ ρ) (Xi )dWi (t)

and we have used the fact that N1 i g(x − Xi ) = A ρ(y, t)g(x − y)dy =: (g ∗ ρ)(x, t)
P R

and similarly for f . If we then add the contributions of all particles and write in its
Langevin form this becomes
1 X d 1 Xh ′ 2β
i
F (Xi ) = F (Xi ) (f ∗ ρ) (Xi ) + DF ′′ (Xi ) (g ∗ ρ) (Xi )
N i dt N i
(A.3) √
2D X ′ β
+ F (Xi ) (g ∗ ρ) (Xi )ηi .
N i

Furthermore, we can express the first two terms as integrals of the empirical density
d
Z Z n o
2β
dxρ(x, t)F (x) = ρ(x, t) (f ∗ ρ)F ′ (x) + D (g ∗ ρ) F ′′ (x) dx
dt A A
(A.4) √
2D X ′ β
+ F (Xi ) (g ∗ ρ) (Xi )ηi .
N i

We can then put the spatial derivatives onto the density terms by integration by parts
so that
 
d
Z Z 
2β
dxρ(x, t)F (x) = dxF (x) ∂x2 D (g ∗ ρ) ρ − ∂x ((f ∗ ρ)ρ)
dt A A
(A.5) √ 
2D X  β
− ∂x δ(Xi − x) (g ∗ ρ) (x)ηi .
N i

It remains to be shown that the stochastic term above can be replaced with a Gaussian
random field which is independent of the individual particle locations. We follow the
20 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

approach of Dean [18] by showing that the two formulations have the same correlation
function. Firstly we write the noise term in the density evolution as,
N
X
∂x δ(Xi − x)ηi (g ∗ ρ)β (x, t)


(A.6) ξ(x, t) = −
i=1

which has the following correlation function

cov(ξ(x, t), ξ(y, s)) = 2Dδ(t − s)∂x ∂y ρ(x, t)(g ∗ ρ)2β (x, t)δ(y − x) .


(A.7)
√
If we choose ξ ′ (x, t) = ∂x η(x, t) ϱ(x, t)(g ∗ ϱ)β (x, t) , where

(A.8) cov(η(x, t)η (y, s)) = δ (t − s) δ(x − y),

we regain the our original noise correlation since
√
cov(ξ ′ (x, t), ξ ′ (y, s)) = cov ∂x η(x, t) ρ(x, t)(g ∗ ϱ)β (x, t) ,
 
(A.9a)
√
∂y η(y, s) ρ(y, s)(g ∗ ϱ)β (y, s)
 

(A.9b)
= δ(t − s)∂x ∂y δ(x − y)ϱ(x, t)(g ∗ ϱ)2β (x, t) .


(A.9c)
Now ξ ′ (x, t) has no explicit dependency on individual particles, Xn , and thus we
replace the noise term with this new noise field:
d
Z Z n  
2β
dxρ(x, t)F (x) = dxF (x) ∂x2 D (g ∗ ρ) ρ − ∂x ((f ∗ ρ)ρ)
dt A A
(A.10) r )
2D √ β


− ∂x ρ(g ∗ ρ) η .
N

Therefore, if ϱ is the solution to (2.4) then (A.1) holds.

This result is a generalisation of the case of constant noise on each agent from
Dean [18] and is recovered when (g ∗ ρ) = 1. Importantly, the stochastic term is of
O(N −1/2 ) and so will be significant when the number of agents is not large.
Appendix B. Fourier representation of Density. When the particles are
compactly supported in a region A, we can express the density in terms of its Fourier
series coefficients. Here we use A = [−π, π) for simplicity. To find the evolution of
the Fourier modes of the density,
N
1 X −ikXn (t)
(B.1) ρ̂k (t) = e ,
2πN n=1

we first express the coupling between particles in terms of the Fourier modes of the
coupling functions:
N Z π
1 X
(B.2a) f (Xn − Xm ) = ρ(x, t)f (Xn − x) dx
N m=1 −π
XZ π
(B.2b) = ρ̂k eikx fˆℓ eiℓ(Xn −x) dx
k,ℓ −π

δk,ℓ ρ̂k fˆℓ eiℓXn

X
(B.2c) = 2π
k,ℓ

ρ̂k fˆk eikXn

X
(B.2d) = 2π
k
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 21

and similarly for g. We can then rewrite (2.1) as

√
ρ̂ℓ fˆℓ eiℓXn dt + 2π 2D
X X
(B.3) dXn = 2π ρ̂ℓ ĝℓ eiℓXn dWn .
ℓ ℓ

Next we use Itô’s lemma,

N N
X ∂ ρ̂k 1 X ∂ 2 ρ̂k
(B.4) dρ̂k = dXn + dXn · dXm
n=1
∂Xn 2 n,m=1 ∂Xn ∂Xm
(B.5)
to write the Fourier modes of the density independently of the individual particle
positions
N N
−ik X −ikXn 2πk 2 D X X
dρ̂k = e dXn − ĝℓ ĝm ρ̂ℓ ρ̂m e−i(k−m−ℓ)Xn dt
2πN n=1 N n=1
ℓ,m∈Z
 

fˆℓ ρ̂ℓ ρ̂k−ℓ − 4π 2 k 2 D

X X
(B.6) = −2πik ĝℓ ĝm ρ̂ℓ ρ̂m ρ̂k−ℓ−m  dt
ℓ∈Z ℓ,m∈Z
√ N X
ik 2D X
− ĝℓ ρ̂ℓ e−i(k−ℓ)Xn dWn .
N n=1 ℓ∈Z

Now we define
fˆℓ ρ̂ℓ ρ̂k−ℓ − 4π 2 k 2 D
X X
Ak (ρ̂) = −2πik ĝℓ ĝm ρ̂ℓ ρ̂m ρ̂k−ℓ−m ,
ℓ∈Z ℓ,m∈Z
(B.7) √
ik 2D X
Gkn (ρ̂) = − √ ĝℓ ρ̂ℓ e−i(k−ℓ)Xn
N ℓ∈Z

so that
N
d ρ̂k 1 X
(B.8) = Ak (ρ̂) + √ Gkn (ρ̂)ηn (t)
dt N n=1

where ηn (t) are zero mean Gaussian random variables with correlator ⟨ηi (t)ηj (t′ )⟩ =
δij δ(t − t′ ). We find the noise correlation matrix in Fourier space to be,
N
Gkn G†nℓ
X
Bkℓ (ρ̂) =
n=1
N
2Dkℓ X X
= ĝm ρ̂m e−i(k−m)Xn ĝj† ρ̂†j ei(ℓ−j)Xn
N n=1
(B.9) j,m∈Z

ĝm ρ̂m ĝj† ρ̂†j ρ̂k+j−m−ℓ

X
= 4πDkℓ
j,m∈Z
X
= 4πDkℓ ĝj ĝm ρ̂j ρ̂m ρ̂k−j−m−ℓ
j,m∈Z

†
where in the final step we have relabelled j → −j and used that gk = g−k and
similarly for ρ since both are real functions.
22 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

B.1. Fourier representation with spatial coupling. We now extend the

formulation from the previous section to allow for additional coupling in extra spatial
dimensions K : A′ 7→ R with A′ = [0, L]d as well as the original coupling in f :
[−π, π) 7→ R. The phase of the oscillators, ϑ evolves according to
N
1 X √
(B.10) dϑn = K(xn − xm )f (ϑn − ϑm ) dt + 2D dWn .
N m=1
PN
We define a new form of the empirical density: ρ(x, ϑ) n=1 δ(x − xn )δ(ϑ − ϑn ). To
find the evolution of the Fourier modes of this density,
N
1 X −i(k.xn +ℓϑn )
(B.11) ρ̂k,ℓ (t) = e
2πN n=1

where the dot refers to the inner product. We again express the coupling between
particles in terms of the Fourier modes of the coupling functions:
(B.12a)
N π
1 X
Z Z
K(xn − xm )f (ϑn − ϑm ) = ρ(x, ϑ, t)K(xn − x)f (ϑn − ϑ) dx dϑ
N m=1 A′ −π
X Z Z π
(B.12b) = ρ̂k,ℓ ei(k.x+ℓϑ)
k,k′ ,ℓ,ℓ′ A′ −π
′
(B.12c) × fˆℓ′ eik.(xn −x)+iℓ (ϑn −ϑ) dx dϑ
′ ′
δkk′ δℓℓ′ ρ̂k,ℓ K̂k′ fˆℓ′ ei(k .xn +ℓ ϑn )
X
(B.12d) = 2πLd
k,k′ ,ℓ,ℓ′

ρ̂k,ℓ K̂k fˆℓ ei(k.xn +ℓϑn ) .

X
(B.12e) = 2πLd
k,ℓ

We can then rewrite (B.10) as

√
ρ̂k,ℓ K̂k fˆℓ ei(k.xn +ℓϑn ) dt + 2D dWn .
X
(B.13) dϑn = 2πLd
k,ℓ

Next, we use Itô’s lemma where only the phases ϑn change but their positions do not,
N N
X ∂ ρ̂k,ℓ 1 X ∂ 2 ρ̂k,ℓ
(B.14) dρ̂k,ℓ = dϑn + dϑn · dϑm ,
n=1
∂ϑn 2 n,m=1 ∂ϑn ∂ϑm

to write the Fourier modes of the density independently of the individual particle
positions
N N
−iℓ X −i(k.xn +ℓϑn ) ℓ2 D X −i(k.xn +ℓϑn )
dρ̂k,ℓ = e dϑ n − e dt
2πLd N n=1 N n=1
 
X
(B.15) = −ℓ2 Dρ̂k,ℓ − 2πiℓLd fℓ′ K̂k′ ρ̂k−k′ ,ℓ−ℓ′ ρ̂k′ ,ℓ′  dt
k′ ,ℓ′
√ N
−iℓ 2D X −i(k.xn +ℓϑn )
+ e dWn .
2πLd N n=1
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 23

Thus, writing
√
−iℓ 2D −i(k.xn +ℓϑn )
(B.16) Gk,ℓ,n = √ e ,
2πLd N
we obtain the result in (3.20) where

Dℓℓ′ X −i(k.xn +ℓϑn ) i(k′ .xn +ℓ′ ϑn )

GG† (k,ℓ),(k′ ,ℓ′ ) =
 
(B.17) e e
2π 2 L2d N n
Dℓℓ′ X −i[(k−k′ ).xn +(ℓ−ℓ′ )ϑn ]
(B.18) = e
2π 2 L2d N n
Dℓℓ′
(B.19) = ρ̂k−k′ ,ℓ−ℓ′ .
πLd
Appendix C. Barenblatt Solution to the Porous Media Equation.
Firstly, we wish to find the relationship between the height and the width of the
deterministic distribution. Since the density, ρ∗ (x), is normalised and non-zero for
|x| < r(t), we know that
r 1/m−1
x2
Z 
(C.1) 1= h(t) 1 − 2 dx,
−r r

thus the height is given by

"Z 1/m−1 #−1
r
x2

(C.2a) h(t) = 1− 2 dx
−r r
Z 1 −1
1 2

1/m−1 γ
(C.2b) = 1−y dy =: .
r −1 r

Now, carrying out this integration, we find that

Z 1 √ m
2 1/m−1

πΓ( m−1 )
(C.3) 1−y dy = m 1
−1 Γ( m−1 + 2 )

and so the explicit form of the normalising constant is

Γ( m + 1 )
(C.4) γ = √ m−1 m 2 .
πΓ( m−1 )

In order to relate this to the Barenblatt solution, we return to (4.6),

h i1/m−1
(C.5a) ρ(x, t) = Ct−a(m−1) − κx2 t−a(m+1)
h i1/m−1
(C.5b) = Ct−a(m−1) − κx2 t−1 .

Similarly for (4.7), we see that

1/m−1
hm−1

(C.6) ρ(x, t) = hm−1 − x2 2 .
r
24 J. WORSFOLD, T. ROGERS, AND P. MILEWSKI

From these two expressions for the distribution it is clear that

(C.7) hm−1 r−2 = κt−1 , hm−1 = Ct−a(m−1)

from which we find

1
(C.8) h(t) = C (m−1) t−a

and thus the width of the distribution is given by

(C.9a) r2 = κ−1 Ct−a(m−1)+1

(C.9b) r(t) = κ−1/2 C 1/2 ta .

With these results, we can express the constant C in terms of our other constants,
κ, γ:
1 1 1
(C.10a) γ = h(t)r(t) = C 2 + m−1 κ− 2
1 m+1
(C.10b) = κ− 2 C 2(m−1)

a(m−1)
(C.10c) C = γ2κ .

Consequently, the width, r(t), can also now be expressed solely in terms of time and
the parameters determined by m,
 1 a(m−1) 1
(C.11a) r(t) = γκ 2 κ− 2 ta

= γ a(m−1) κ 2 [ m+1 −1] ta

1 m−1
(C.11b)
(C.11c) = γ a(m−1) κ−a ta .

Appendix D. Porous Media Fluctuations. To find the variance in the

centre of mass of the distribution, we first calculate the correlation between any two
given moments of the distribution
Z r
(D.1a) Bpq (t) = 2 xp+q−2 (ρ∗ )m dx
−r
m
Z r   x 2  m−1
m p+q−2
(D.1b) = 2pqh x 1− dx
−r r
Z 1
 m
= 2pqhm rp+q−1 y p+q−2 1 − y 2 m−1 dy

(D.1c)
−1

2pqhm rp+q−1 Γ( 2 )Γ(2+ m−1 )
p+q−1 1

Γ( 2 + m−1 )
p+q+3 1 p + q is even
(D.1d) =
0 otherwise.

We can also simplify this further by noticing that

2
(D.2a) hm r = γ m r−(m−1) = γ m−a(m−1) κa(m−1) t−a(m−1)
(D.2b) = γ a(3m−1) κa(m−1) t−a(m−1) .
 DENSITY FLUCTUATIONS IN STOCHASTIC KINEMATIC FLOWS 25

This means we can write the variance of the fluctuations in the centre of mass as
Z t
(D.3a) 2
σ1 = dt′ B11 (t′ )
0

 
Z t Γ 21 Γ 2 + m−1 1

(D.3b) =2 dτ τ a(1−m) γ a(3m−1) κa(m−1)  

0 Γ 52 + m−1
1

 
√ Γ 2 + m−1 1
π
(D.3c) = t2a γ a(3m−1) κa(m−1)  
a Γ 52 + m−11

 
1
√ Γ 2 + m−1
(D.3d) σ12 = (m + 1) πt2a γ a(3m−1) κa(m−1)  .
5 1
Γ 2 + m−1