N - Body Simulations of Gravitational Dynamics

a
r
X
i
v
:
1
1
0
5
.
1
0
8
2
v
1

[
a
s
t
r
o
-
p
h
.
I
M
]

5

M
a
y

2
0
1
1
EPJ manuscript No.
(will be inserted by the editor)
N-body simulations of gravitational dynamics
Walter Dehnen
1
and Justin I. Read
1,2
1
Department of Physics & Astronomy, University of Leicester, Leicester LE17RH, United Kingdom
2
Institute for Astronomy, Department of Physics, ETH Z urich, Wolfgang-Pauli-Strasse 16, CH-8093 Z urich, Switzerland.
Received: / Revised version:
Abstract. We describe the astrophysical and numerical basis of N-body simulations, both of collisional
stellar systems (dense star clusters and galactic centres) and collisionless stellar dynamics (galaxies and
large-scale structure). We explain and discuss the state-of-the-art algorithms used for these quite dierent
regimes, attempt to give a fair critique, and point out possible directions of future improvement and
development. We briey touch upon the history of N-body simulations and their most important results.
PACS. 98.10.+z Stellar dynamics and kinematics 95.75.Pq Mathematical procedures and computer
techniques 02.60.Cb Numerical simulation; solution of equations
1 Introduction
Of the four fundamental forces, gravity is by far the weakest. Yet on large distances it dominates all other interactions
owing to the fact that it is always attractive. Most gravitational systems are well approximated by an ensemble of
point masses moving under their mutual gravitational attraction and range from planetary systems (such as our own)
to star clusters, galaxies, galaxy clusters and the universe as a whole.
Systems dominated by long-range forces such as gravity are not well treatable by statistical mechanical methods:
energy is not extensive, the canonical and micro-canonical ensembles do not exist, and heat capacity is negative [1].
Moreover, gravitational encounters are inecient for re-distributing kinetic energy, such that many such encounters
are required for relaxation, i.e. equipartition of kinetic energy. Gravitational systems, where this process is potentially
important over their lifetime are called collisional as opposed to collisionless stellar systems
1
. The timescale over
which this so-called two-body relaxation is important is roughly [1]
t
relax

N
8 ln
t
dyn
(1)
where N is the number of particles; t
dyn
is the dynamical time
2
; ln = ln(b
max
/b
min
) is called the Coulomb Logarithm;
and b
max
and b
min
are the maximum and minimum impact parameters for the system, respectively
3
. Collisional systems
usually have a high dynamic age (t
dyn
short compared to their lifetime) and high density, and include globular star
clusters and galactic centres. The majority of stellar systems, however, are collisionless.
The numerical simulation of N-body systems has a long history, dating back to the very rst light-bulb experiments
of Holmberg [3] in 1941. The rst computer calculations were performed by von Hoerner [4] in 1960 who reached
N = 16, paving the way for the pioneering work of Aarseth [5] in 1963 with N = 100. Since these early works, N has
nearly doubled every two years in accordance with Moores law (see Fig. 1). The latest collisional N-body calculations
1
We shall use the term stellar system for a system idealised to consist of gravitating point masses, whether these are stars,
planets, dark-matter particles, or whatever.
2
The dynamical or crossing time is a characteristic orbital time scale: the time required for a signicant fraction of an orbit.
A useful approximation, valid for orbits in inhomogeneous stellar systems, is t
dyn
(G )
1/2
with the mean density interior
to the particles current radius [1]. Note that the complete orbital period is longer by a factor 3.
3
The minimum impact parameter is well dened for a system containing equal mass point particles: bmin = 2Gm/
2
, where
is the velocity dispersion of the background particles [2]. The maximum impact parameter is related in some way to the extent
of the system, but is less well dened. Fortunately, even an order of magnitude uncertainty in bmax has only a small eect since
it appears inside a logarithm. Notice that the form of the Coulomb Logarithm has a profound implication: each octave of b
contributes equally to ln such that relaxation is driven primarily by weak long-range interactions with b bmin.
2 Walter Dehnen, Justin I. Read: N-body simulations of gravitational dynamics
have reached over 10
6
particles [6], while collisionless calculations can now reach more than 10
9
particles [710]. This
disparity reects the dierence in complexity of these rather dissimilar N-body problems. The signicant increase in
N in the last decade was driven by the usage of parallel computers.
Fig. 1. The increase in particle number over the past
50 years for selected collisional (red, [46, 1116] taken
from [17]) and collisionless (blue, [3, 79, 1824]) N-body
simulations. The line shows the scaling N =N02
(yeary
0
)/2
expected from Moores law if the costs scale N.
In this review, we discuss the state-of-the art software algo-
rithms and hardware improvements that have driven this dra-
matic increase in N. We consider the very dierent challenges
posed by collisional (2) versus collisionless (3) systems, and
we attempt to give a fair critique of the methods employed,
pointing out where there is room for improvement, and dis-
cussing some interesting future research directions. Our focus
is primarily on gravity; we do not consider the important role
that the other fundamental forces play
4
. Since our goal is to
elucidate the numerics, we will only touch upon the many in-
teresting and important results that have come out of N-body
modelling over the past 50 years. We must therefore apologise
in advance for all that is missed out in this brief review. We do
not have space to discuss modelling gas physics and its many
complications. Nor will we discuss the art of setting up initial
conditions for N-body simulations.
There are already several N-body reviews and books in the
literature. Aarseth [25] and Heggie & Hut [26] give excellent
reviews of the N-body problem, focusing mainly on collisional
N-body simulations. Hockney & Eastwood [27] cover many
aspects of particle-based simulations, focusing on collisionless
applications. Trenti & Hut [28] give a general overview of the
N-body problem, covering both the collisional and collision-
less regimes. Our review takes a somewhat dierent approach
to these previous works. We attempt to review numerical tech-
niques for both collisional and collisionless N-body simula-
tions, focusing on the very latest techniques, and with a view to assessing interesting future research directions. As
much as possible, we present the key equations and describe the methodology at a level where we hope the reader will
be able to obtain a relatively deep understanding of the algorithms and their hidden gremlins.
This paper is organised as follows. In 2, we review numerical methods for collisional N-body simulations: the
astrophysical (2.1) and numerical foundations (2.2) with special treatment of the time integration (2.3), recent
hardware-driven developments (2.4); and give a critique of the current state of the art (2.5), summarise alternatives
to N-body methods (2.6), as well as present a (brief and biased) overview over past and recent astrophysical results
with an outlook for the future (2.7). In 3, we review numerical methods for collisionless simulations: the astrophysical
(3.1) and numerical foundations (3.2), the basis of cosmological N-body simulations (3.3), force softening (3.4) and
the various force solvers (3.5), recent developments and challenges (3.6), and a very brief overview of astrophysical
results (3.7). 4 describes methods for validation of N-body simulations, and in 5 we present our conclusions and
outlook for the future of the eld.
2 N-body methods for collisional systems
Collisional systems are dynamically old such that t
dyn
is short compared to their age. This applies mainly to massive
star clusters, and for this reason we will mostly refer to the N-body particles in this section as stars within a star
cluster. Such clusters typically orbit deep within the potential of a host galaxy like our own Milky Way such that
their dynamics is aected by the tidal eld of their host.
Over many dynamical times, the accumulated eect of many small encounters between stars signicantly aects
the evolution of collisional systems. Relaxation-driven equipartition of energy causes heavier stars to sink towards the
centre, while low-mass stars are pushed to the outskirts, where they are susceptible to being skimmed o by Galactic
4
Many stellar systems contain signicant amounts of gas, which in addition to gravity also interact electromagnetically. This
gives rise to a large number of complicated eects from radiative cooling and the formation of stars (inside of which the strong
and weak interactions become important, too), to active galactic nuclei and outows driven by radiative heating. While for
most gravitational systems these non-gravitational eects play an important role only for brief periods of their lifetime, their
understanding and appropriate modelling is at the forefront of many contemporary challenges in astrophysics. This is, however,
beyond the scope of this paper.
Walter Dehnen, Justin I. Read: N-body simulations of gravitational dynamics 3
tides. Such relaxation processes provide a mechanism for transporting specic energy outwards, resulting in what is
called the gravothermal catastrophe: a dramatic increase in the central density resulting in core collapse [2932].
In addition to relaxation which is largely driven by distant encounters short-range interactions can form bound
particle pairs. Such binaries are called hard if their binding energy exceeds the typical kinetic energy of surrounding
stars in the cluster. Close encounters with such eld stars result in a further hardening of hard binaries, while soft
binaries loose binding energy and become even softer (known as Heggies law [33, 34]). Hard binary interactions are
of great importance, as they act as a source of kinetic energy [35], heating the core of the cluster (where most binaries
have sunk due to their larger mass), counter-acting the relaxation-driven energy ux from the core to the outskirts, and
thus prolonging the time until core collapse [36, 37]. Core collapse could perhaps even be reversed by binaries freshly
formed in three-body encounters (or two-body encounters involving strong stellar tides), a process most likely to occur
in the very high densities reached during core collapse. The formation and evolution of hard binaries (and higher order
multiples of stars) presents a unique numerical challenge, because their short dynamical times and extreme forces lead
to very small integration time steps [25]; they require special treatment that we discuss in 2.3.
Another reason binary interactions are important astrophysically is that they provide a route to forming close
binary stars of a constitution and type unlikely to form under ordinary star-formation conditions. Blue Straggler
stars, for example, may form in this way [3840], as well as ultra compact X-ray binaries, which are over-abundant in
globular clusters [41, and references therein].
2.1 Equations governing collisional stellar systems
The physics of collisional stellar systems is in principle quite simple: the motion of N point masses under their mutual
gravitational attraction. This is in fact a Hamiltonian system, with a Hamiltonian H and equations of motion
H =
i
p
2
i
2m
i
G
j>i
m
i
m
j
|x
i
x
j
|
, a
i
=
p
i
m
i
=
1
m
i
H
x
i
= G
j=i
m
j
x
i
x
j
|x
i
x
j
|
3
. (2)
where p
i
= m
i
x
i
is the momentum; x
i
is the position; and a
i
= x
i
is the acceleration of particle i. N-body simulations
of collisional systems simply try to solve these equations directly by brute force.
For problems involving massive central black holes, general relativistic (GR) corrections to the force can become
important. This is because, although we are almost always in the weak eld regime, if the black hole dominates the
central potential, then the potential will be close to Keplerian and super-resonant. GR eects act to break resonance
by inducing orbital precessions. Over many dynamical times, or if stars orbit very close to the central black hole, such
eects can become important [42, 43]. We will not discuss such post-Newtonian corrections to the force further in
this review, but refer the interested reader to Aarseth [42] and references therein.
2.2 Numerics of collisional N-body simulations
While equations (2) are conceptually straightforward, they are anything but straightforward to solve numerically. One
obvious problem is that the computational cost for the force calculation for all particles grows as O(N
2
), implying
that realistic simulations with N 10
6
(the number of stars in massive star clusters) are challenging. Such high N
has only recently been achieved for a small number of simulations using special hardware chips operated in parallel
[6] (see also 2.4). A second serious problem is the enormous range of time scales, ranging from days for the periods of
tight binaries
5
to millions of years for the orbits of most stars, and 10
10
years for the age of the whole stellar system.
To make things worse, the time step required to accurately integrate the trajectory of any individual star can change
considerably along its orbit and abruptly during close encounters. Given this range of formidable problems, it is not
surprising that there are very few N-body codes which can cope with them.
As already alluded to, the computation of the forces dominates the computational cost of existing collisional N-
body codes, all of which use the brute-force direct-summation approach (i.e. a straightforward implementation of
equations 2). This is motivated by the need for an accurate force computation to ensure correct modelling of both
close and distant encounters. Currently, several approaches are applied to reduce the frequency of force computations
for any individual star. First, the force is split into contributions from near neighbours and the far eld. While the
former varies on short time scales, it is quite cheap to compute, whereas the latter is expensive to evaluate but much
smoother in time. Thus, splitting the force into two components like this allows us to reduce the need for the expensive
computation of the far-eld force. This Ahmad-Cohen [45] scheme necessarily requires individual timesteps for each
particle which we discuss in 2.3.
5
The shortest observed binary period is 11 minutes for the X-ray binary 4U1820-30 in the globular cluster NGC6624 [44],
but such very short periods are not simulated in contemporary N-body simulations.
Second, the frequency for computing the acceleration a
i
can be further reduced by computing its time derivative,
the jerk:
a
i
= G
j=i
m
j
x
2
ij
x
ij
3x
ij
(x
ij
x
ij
)
|x
ij
|
5
with x
ij
x
i
x
j
, (3)
which is used to predict a
i
into the future or, equivalently, to employ a higher-order time integrator with a larger time
step (see 2.3). Ever higher order schemes require ever higher derivatives of a
i
to obtain forward interpolations of the
force [46].
2.3 Time integration
The accurate time integration of close encounters is the most dicult part of collisional N-body methods, while
for collisionless N-body methods force softening (see 3.4) alleviates this problem substantially. Here, we review the
various time integration methods employed in both types of N-body methods. Let us begin our considerations by the
simple Euler method, which updates the position and velocity for a given particle by timestep t via
x(t + t) = x(t) + xt (4a)
x(t + t) = x(t) +a(t) t. (4b)
While conceptually straightforward, this scheme performs very poorly in practice. The Euler method is just a Taylor
expansion to rst order in t and the errors are proportional to t
2
. We can signicantly improve on this at little
additional computational cost either by increasing the expansion order and thus the accuracy, or by integrating a
near-by Hamiltonian exactly using a low-oder scheme. We now compare and contrast a popular example of each type
of approach: the second-order leapfrog integrator, which is heavily used in collisionless N-body applications, and the
fourth-order Hermite scheme, which has become the integrator of choice for collisional applications.
2.3.1 The Leapfrog integrator
The leapfrog integrator is an example of a symplectic integrator. Symplectic integrators exactly solve an approximate
Hamiltonian. As a consequence, the numerical time evolution is a canonical map and preserves certain conserved
quantities exactly, such as the total angular momentum, the phase-space volume, and the Jacobi constants. The idea
is to approximate the Hamiltonian H in equation (2) with
H = H +H
err
(5)
where H
err
is the error Hamiltonian. Provided that

H and H are time-invariant, the energy error is bounded at all
times [47]
6
. The goal now is to nd

H that can be solved exactly by simple numerical means and minimises H
err
.
Dening the combined phase-space coordinates w = (x, p) we can re-write Hamiltons equations as:
w = Hw, (6)
where H {, H} (with the Poisson bracket {A, B}
x
A
p
B
x
B
p
A) is an operator acting on w. Equation (6)
has the formal solution
w(t +t) = e
t H
w(t), (7)
where we can think of the operator e
t H
as a symplectic map from t to t + t. This operator can be split, in an
approximate sense, into a succession of discrete but symplectic steps, each of which can be exactly integrated. The
most common choice is to separate out the kinetic and potential energies, H = T(p) +V (x), such that we can split
e
t H
= e
t (T +V)
e
t V
e
t T
= e
t

H
. (8)
Because the dierential operators T {, T} and V {, V } are non-commutative, the central relation in equation (8)
is only approximately true. This operator splitting is extremely useful, because, while equation (6) has in general no
simple solution, the equivalent equations for each of our new operators do:
e
t T
_
x
p
_
=
_
x +t p
p
_
and e
t V
_
x
p
_
=
_
x
p t V (x)
_
. (9)
6
A symplectic integrator which obtains zero energy error is exact.
These operations are also known as drift and kick operations, because they only change either the positions (drift) or
velocities (kick). Note that the drift step in (8) is identical to the simple Euler method (4a), while its kick step is not
identical, because the acceleration is calculated using the drifted rather than the initial positions. The integrator that
applies a drift followed by a kick (equation 8) is called modied Euler scheme and is symplectic.
It is clear from the similarity between the simple and modied Euler schemes that both are only rst order accurate.
We can do better by concatenating many appropriately weighted kick and drift steps:
e
t

H
=
N
i
e
aitV
e
bitT
= e
t H+O(t
n+1
)
(10)
with coecients a
i
and b
i
chosen to obtain the required order of accuracy n. From equation (10) we see that: (i) the
approximate Hamiltonian

H is solved exactly by the successive application of the kick and drift operations, and (ii)
H approaches H in the limits t 0 or n . At second order (n = 2), and choosing coecients that minimise
the error, we derive the leapfrog integrator:
e
t H+O(t
3
)
= e
1
2
t V
e
t T
e
1
2
t V
. (11)
Applying equations (9), this becomes (subscripts 0 and 1 refer to times t and t +t, respectively):
x
= x
0
+
1
2
a
0
t (12a)
x
1
= x
0
+ x
t (12b)
x
1
= x
+
1
2
a
1
t (12c)
where a
0
= V (x
0
) and a
1
= V (x
1
), while the intermediate velocity x
serves only as an auxiliary quantity.

Combining equations (12) we nd the familiar Taylor expansions of the positions and velocities to second order
7
:
x
1
= x
0
+ x
0
t +
1
2
a
0
t
2
, (13a)
x
1
= x
0
+
1
2
(a
0
+a
1
)t. (13b)
In principle, one may combine as many kick and drift operations as desired to raise the order of the scheme. However,
it is impossible to go beyond second order without having at least one a
i
and one b
i
coecient in equation (10)
be negative [48, 49, see also 50]. This involves some backwards integration and is problematic when using variable
timestepsespecially if time symmetry is required
8
.
As mentioned previously, variable individual timesteps are a requirement for most N-body applications. Unfortu-
nately, once we allow t to vary in space and time, the symplectic nature of the leapfrog is broken. But, we can do
almost as well as symplectic by ensuring that the scheme remains time symmetric [52]. One route to time symmetry
is to solve the implicit leapfrog equations (13) using a symmetrised timestep:
t =
1
2
_
T(x
0
, x
0
, a
0
...) +T(x
1
, x
1
, a
1
...)
, (14)
where the function T generates the step size depending on the position, velocity, acceleration etc. (we will discuss
possible functional forms in 2.3.3). Since t is a function of the force evaluated at t +t, which in turn is a function
of t, such a scheme is implicit and in general requires an iterative solution [53]. Since each iteration involves another
expensive force evaluation, implicit schemes are not used in practice. Fortunately, there are explicit and time-symmetric
methods for adapting the time step [54], for example
t
old
t
new
= T(x, x, a...)
2
, (15)
where t
old
and t
new
are the time steps used to evolve to and from the arguments of T. Clearly, equation (15) is
time symmetric by construction and requires no iteration to solve.
7
This is the kick-drift-kick (KDK) leapfrog. An alternative is the drift-kick-drift (DKD) version. In practice, the KDK is
preferable, because acceleration and potential are known at the second-order accurate positions (not at an auxiliary intermediate
position as for the DKD), facilitating their usage in time-step control. Moreover, with the block-step scheme (see 2.3.3 and
Fig. 3) the KDK results in synchronised force computations for all active particles.
8
Recently, Chin & Chen [51] have constructed fourth-order symplectic integrators which require only forwards integration.
To achieve this, rather than eliminate all the errors by appropriate choice of the coecients ai and bi, they integrate one of the
error terms thus avoiding any backward step. Their method requires just two force and one force gradient evaluation per time
step. It has not yet found application in N-body dynamics, but could be a very promising avenue for future research.
Fig. 2. Left Comparison of the leapfrog integrator (black); a 4th order Hermite scheme (red); and time symmetric leapfrog
using variable timesteps (blue) for the integration of an elliptic (e = 0.9) Kepler orbit over 100 periods. In the rst two cases a
xed timestep of 0.001 of the period was used; in the latter case we use the timestep criteria in equation (20). Middle Fractional
change in energy for the Kepler problem for the leapfrog integrator with xed timesteps (black), variable timesteps (equation
20; red), and symmetric variable timesteps (blue). Right Fractional change in energy for the Kepler problem for the 4th order
Hermite integrator with xed timesteps (black) variable timesteps (equation 25; red), and variable timesteps (equation 26; red
dotted). The blue curve shows the energy error for the same Kepler orbit calculated using the K-S regularised equations of
motion (see text for details). All calculations with variable timesteps were run at the same computational cost ( 250 force and
jerk evaluations per orbit, which is about a quarter of the cost of the xed-timestep calculations).
In the middle panel of Fig. 2, we compare energy conservation for the leapfrog using xed timesteps, variable
timesteps, and time symmetric variable timesteps of equation (15) for a Kepler orbit with eccentricity e = 0.9. With
xed timesteps (black) the energy uctuates on an orbital time scale, but is perfectly conserved in the long term;
with variable timesteps, manifest energy conservation is lost (red); while with the time symmetric variable time step
scheme, we recover excellent energy conservation (blue). The time symmetric variable time step leapfrog used about a
quarter of the force calculations required for the xed-step integration while giving over an order of magnitude better
energy conservation. This is why variable timesteps are an essential ingredient in modern N-body calculations.
2.3.2 Hermite integrators
The leapfrog integrator is a popular with collisionless N-body applications because of its simplicity, manifest energy
conservation, and stability. However, its integration errors on short time scales make it less useful for studying collisional
systems, where one must correctly track chaotic close encounters, and such errors would rapidly ruin the integration.
Shrinking the timestep helps, but as the leapfrog is only a second-order scheme, the step sizes required become
prohibitively small. This motivates considering higher-order non-symplectic integrators for collisional applications.
The current state of the art are fourth-order non-symplectic integrators, so called Hermite schemes (for higher
orders see [46]). From the Taylor expansion of the position and its time derivatives at time t +t
x
1
= x
0
+ x
0
t +
1
2
a
0
t
2
+
1
6
a
0
t
3
+
1
24
a
0
t
4
, (16a)
x
1
= x
0
+a
0
t +
1
2
a
0
t
2
+
1
6
a
0
t
3
+
1
24
...
a
0
t
4
, (16b)
a
1
= a
0
+ a
0
t +
1
2
a
0
t
2
+
1
6
...
a
0
t
3
, (16c)
a
1
= a
0
+ a
0
t +
1
2
...
a
0
t
2
, (16d)
we can eliminate a
0
and
...
a
0
to obtain
x
1
= x
0
+
1
2
( x
1
+ x
0
)t +
1
12
(a
0
a
1
)t
2
+O(t
5
), (17a)
x
1
= x
0
+
1
2
(a
1
+a
0
)t +
1
12
( a
0
a
1
)t
2
+O(t
5
). (17b)
These equations are not only fourth-order accurate but also time symmetric (though not symplectic) and will therefore
give excellent energy conservation. The only snag is their circularity: in order to obtain x
1
and x
1
we need to know
the acceleration a and jerk a not only at time t (where we can readily compute them from equations 2 and 3) but also
at time t +t, when they in turn depend on x
1
and x
1
. In order words, equations (17) dene an implicit scheme. In
practice, this diculty is side-stepped by rst predicting positions and velocities
x
p
= x
0
+ x
0
t +
1
2
a
0
t
2
+
1
6
a
0
t
3
, (18a)
x
p
= x
0
+a
0
t +
1
2
a
0
t
2
; (18b)
then estimating acceleration a
1
and jerk a
1
using equations (2) and (3) with the predicted positions and velocities;
and nally obtaining the corrected x
1
and x
1
from equations (17). A single iteration of this method is is called a
Predict-Evaluate-Correct (PEC) scheme; further iterations are denoted P(EC)
n
, with n the number of iterations [55].
In the limit n , we converge on the implicit Hermite solution (17). In practice, implicit integration schemes are
not employed because of the numerical cost of calculating the acceleration and jerk over several iterations. However,
unlike the implicit Hermite scheme, the explicit PEC scheme is not time symmetric.
A comparison of various avours of the 4th order Hermite integrator are given in Figure 2, for the integration of an
elliptical Kepler orbit with e = 0.9 over 100 orbits. Notice that the leapfrog integrator with xed timestep conserves
energy exactly (in the long-term), but that the peak of the oscillations is initially two orders of magnitude worse
than for the 4th order Hermite scheme with xed steps (compare black lines in middle and right panels). Over time,
the energy losses accumulate for the Hermite integrator, causing the apocentre of the orbit to decay, but the orbital
precession (both are numerical errors) is signicantly less than for the leapfrog (compare black and red orbits in the
left panel). It is the orbital stability and excellent energy accuracy that have made Hermite integrators popular for
use in collisional N-body problems.
2.3.3 The choice of time-step
Given the enormous dynamic range in time involved in collisional N-body problems (ranging from days to giga-years),
it has become essential to use variable timestep schemes [25]. Early schemes used an individual time step for each
particle. However, it is better to arrange the particles in a hierarchy of timesteps organised in powers of two, with
reference to a base step t
0
[56]:
t
n
= t
0
/2
n
(19)
Fig. 3. Schematic illustration of a block
time-stepping scheme. Particles are organised on
timesteps in a hierarchy of powers of two rela-
tive to a base time t0. The time step level, de-
noted n0,1,2... is called the timestep rung. Particles
can move up and down rungs at synchronisation
points marked by the red arrows.
for a given timestep rung n. Particles can then move between rungs at
synchronisation points as shown in Fig. 3. This block-step scheme leads
to signicant eciency savings because particles on the same rung are
evolved simultaneously. However, time symmetry with block stepping
presents some challenges [57]. A key problem is that, in principle,
particles can move to lower timestep rungs whenever they like, but
they may only move to higher rungs at synchronisation points where
the end of the smaller step overlaps with the end step of a higher
rung (see Fig. 3). This leads to an asymmetry in the timesteps, even
if some discrete form of equation (15) is used. Makino et al. [57] show
that it is possible to construct a near-time symmetric block time step
scheme, provided some iteration is allowed in determining the time
step. Whether a non-iterative scheme is possible remains to be seen.
We now need some criteria to decide which rung a particle should
be placed on. For low-order integrators like the leapfrog, we have only
the acceleration to play with. In this case, a possible timestep criterion
can be found by analogy with the Kepler problem:
t
i
=
_
|
i
|/|a
i
|, (20)
where
i
is the gravitational potential of particle i, and is a
dimensionless accuracy parameter. Substituting
i
= GM/r
i
and
|a
i
| = GM/r
2
i
, valid for a particle at radius r
i
orbiting a point mass
M, we see that this gives t
i
=
_
r
3
i
/GM, i.e. exactly proportional
to the dynamical time. However, a timestep criteria that depends on
the potential is worrisome since the transformation + const. has no dynamical eect, but would alter the
timesteps. In applications like cosmological N-body simulations, where the local potential has signicant external
contributions, simulators have typically employed:
t
i
=
_
/|a
i
| (21)
and similar, where is the force softening length (see 3). Equation (21) is really only dened on dimensional grounds:
it creates a quantity with dimensions of time from a local length scalethe force softeningand the local acceleration.
It is clear that this time step criteria would be of no use for, say, the Kepler problem, where it will lead to too small
steps at large radii, and too large steps at small radii.
In a recent paper, Zemp et al. [58] have attempted to solve the above conundrum by trying to determine what a
particle is orbiting about. If this is known, then the dynamical time itself makes for a natural timestep criteria:
t
i
=
_
r
3
i
/GM(r
i
), (22)
where M(r
i
) is the mass enclosed within the particles orbit from some attractor at distance r
i
. Zemp et al. attempted
to dene a M(r) based on information taken from a gravitational tree structure (see 3). Such ideas lend themselves
naturally to collisionless simulations, where a tree is often readily available as a by-product of the force calculation (see
3.5.1). But it remains to be seen if such a timestep criteria can be competitive for collisional N-body applications.
Unlike many collisionless applications, in collisional N-body applications, it is often well-dened from the outset what
particles are orbiting aboutat least until close interactions occur when case special treatment is required anyway.
In addition, the higher-order integrators typically employed provide a wealth of additional free information that can
be used to determine the timestep. The fourth order Hermite integrator, for example, gives us a, a and
...
a (the latter
two from nite dierences). Such considerations have motivated higher-order time-stepping criteria.
For example, an immediately obvious choice for a higher-order criterion is to set the time step based on the
truncation error in the Hermite expansion:
t
i
=
_
|a
i
|/| a
i
|
_
1/2
or t
i
=
_
|a
i
|/|
...
a
i
|
_
1/3
(23)
or to use the error in the predictor step, as suggested by Nitadori & Makino [46]:
t
i
= t
old,i
_
|a
i
|/|a
i
a
p,i
|
_
1/p
(24)
where p is the order of the expansion, a
p,i
the predicted acceleration, and t
old,i
the previous timestep. However,
while such criteria seem sensible, they are all out-performed by the seemingly mystic Aarseth [25] criterion
t
i
=
_
|a
i
|| a
i
| +| a
i
|
2
| a
i
||
...
a
i
| +| a
i
|
2
_
1/2
(25)
with 0.02 (which can be generalised for higher-order schemes, see [46, 59]).
The success of equation (25) probably lies in the fact that it conservatively shrinks the time step if either a or
...
a
are large compared to the smaller derivatives, and it requires no knowledge of the previous timestep. Equations (23)
give poorer performance because they do not use information about all known derivatives of a. Equation (24) gives
poorer performance for large timesteps. It too uses information about all calculated derivatives of a (since it is based
on the error between predicted and true accelerations). But the problem is that it relies on the previous timestep for
its normalisation. If t
old,i
is too large, the criteria will not respond fast enough, leading to overly large timesteps and
large energy losses [46].
The above suggests that a conservative truncation error-like criteria that encompasses all derivatives of a might
perform at least as well as the Aarseth criteria while being (perhaps) more theoretically satisfying:
t
i
= min
_
_
_
_
|a
i
|
| a
i
|
_
,
_
2
|a
i
|
| a
i
|
_
1/2
,
_
3
|a
i
|
|
...
a
i
|
_
1/3
, ...,
_
p
|a
i
|
|a
(p)
i
|
_
1/p
_
_
_
(26)
where p is the highest order of a calculated by the integrator; and a
(p)
is the p
th
derivative of a.
In the right panel of Fig. 2, we compare 4th order Hermite integrators with dierent variable timestep criteria for
a Kepler orbit problem with e = 0.9. The black curve shows results for a xed timestep with t = 0.001 periods; the
red curve shows results using a variable timestep and the Aarseth criteria (equation 25); and the red dotted curve
shows results using equation (26). For the Aarseth criteria we use = 0.02; for our new criteria in equation (26) we
set in all cases such that exactly the same number of steps are taken over ten orbits as for the Aarseth criteria.
The truncation error-like criteria (equation 26) appears to give very slightly improved performance for the same cost.
However, whether this remains true for full N-body applications remains to be tested.
The above timestep criteria have been well tested for a wide range of problems and so appear to work wellat
least for the types of problem for which they were proposed. However, there remains something unsatisfying about all
of them. For some, changing the velocity or potential can alter the timestep and, with the possible exception of the
criterion by Zemp et al. [58], all are aected by adding a constant to the acceleration. This is unsatisfactory, since
the internal dynamics of the system is not altered by any of these changes. Applying a constant uniform acceleration,
generated for example by an external agent, to a star cluster is allowed by the Poisson equation and does not alter the
internal dynamics, and thus should not drastically alter the timesteps. Only if the externally generated acceleration
varies across the cluster does it aect its internal dynamics, an eect known as tides. This suggests
t
i
= (/||(a)
i
||)
1/2
(27)
where a is the gradient of the acceleration and |||| denotes the matrix norm. Remarkably, for the Kepler problem this
agrees with equation (20), while for isolated systems with power-law mass proles it is very similar to equation (22).
However, computing the gradient of a merely for the sake of the timestep seems extravagant.
2.3.4 Close encounters and regularisation
A key problem when modelling collisional dynamics is dealing with the divergence in the force for x
i
x
j
in
equation (2), requiring prohibitively small timesteps (or large errors) with any of the above schemes. Consider our
simple Kepler orbit problem. For a timestep criteria as in equation (20), this gives a timestep at pericentre r
p
of
t
2
r
3
p
/GM. Thus, for increasingly eccentric orbits, the timesteps will rapidly shrink, leading to a few highly eccentric
particles dominating the whole calculation. To avoid this problem, collisional N-body codes introduce regularisation
for particles that move on tightly bound orbits. The key idea is to use a coordinate transformation to remove the force
singularity, solve the transformed equations, and then transform back to physical coordinates. Consider the equations
of motion for a perturbed two-body system with separation vector R = x
1
x
2
(using R |R|):
R = G(m
1
+m
2
)
R
R
3
+F
12
, (28)
where F
12
= F
1
F
2
is the external perturbation. This, of course, still has the singularity at R = 0. Now, consider
the time transformation dt = Rd:
R
=
1
R
R
G(m
1
+m
2
)
R
R
+R
2
F
12
(29)
where

denotes dierentiation w.r.t. . Note that we have removed the R
2
singularity in the force, but gained another
in the term involving R
. To eliminate that, we must also transform the coordinates. The current transformation of
choice is the KustaanheimoStiefel (K-S) transformation [25, 60, 61], which requires a move to four spatial dimensions.
We introduce a dummy extra dimension in R = (R
1
, R
2
, R
3
, R
4
), with R
4
= 0, and transform this to a new four vector
u = (u
1
, u
2
, u
3
, u
4
) such that R = L(u)u, with:
L =
_
_
u
1
u
2
u
3
u
4
u
2
u
1
u
4
u
3
u
3
u
4
u
1
u
2
u
4
u
3
u
2
u
1
_
_ (30)
The inverse transformation is non-unique, since one of the components of u is arbitrary. In general, we may write:
u
2
1
=
1
2
(R
1
+R) cos
2
; u
2
=
R
2
u
1
+R
3
u
4
R
1
+R
(31a)
u
2
4
=
1
2
(R
1
+R) sin
2
; u
3
=
R
3
u
1
R
2
u
4
R
1
+R
(31b)
where is a free parameter. It is a straightforward exercise to verify that equations (31) satisfy the transformation
equation R = L(u)u. We also require a transformation between the velocities

R and u
. Writing R
= L(u
)u +
L(u)u
= 2L(u)u
, and using the relation L

T
L = RI gives:
u
=
1
2
L
T
R
R
=
1
2
L
T

R (32)
where the last relation follows from the time transformation dt = Rd. Substituting the K-S coordinate transform
into equation (29) gives [25]:
u
1
2
Eu =
1
2
RL
T
F
12
(33a)
where E is the specic binding energy of the binary, which is evolved as:
E
= 2u
L
T
F
12
. (33b)
(Note that the transformed time is given by t
= |u|
2
, which follows from equation 29.) We can now see two important
things. Firstly, there are no longer any coordinate singularities in equations (33). Secondly, in the absence of an external
eld (F
12
= 0), E = const. and our transformed equations correspond to a simple harmonic oscillator.
We can evolve the above regularised equations of motion using the Hermite scheme (2.3.2), so long as we can
calculate u
and E
. These follow straightforwardly from the transformed time derivatives of equations (33):
u
=
1
2
(E
u +Eu
+R
Q+RQ
) , E
= 2u
Q+ 2u
(34)
where Q = L
T
F
12
describes the external interaction term.
In Figure 2(c), we show results for a Kepler orbit with eccentricity e = 0.9 integrated over 100 orbits using the K-S
regularisation technique (blue). We use a Hermite integrator with variable timesteps, and timestep criterion (26). For
as many force calculations as the variable timestep Hermite integration scheme, the results are over 100 times more
accurate. This is why K-S regularisation has become a key element in modern collisional N-body codes.
K-S regularisation as presented above works only for a perturbed binary interaction. However, it is readily gener-
alised to higher order interactions where for each additional star, we must transform away another potential coordinate
singularity [25, 62, 63]. In practice, this means introducing N coupled K-S transformations, which requires 4N(N1)+1
equations, making extension to large N inecient. For this reason, chain regularisation has become the state-of-the
art [64, 65]. The idea is to regularise only the close interactions between the N particles, rather than all inter-particle
distances, which reduces the number of equations to just 8(N 1) + 1, paving the route to high N. For interactions
involving large mass ratio, other regularisation techniques can become competitive with the K-S chain regularisation
[66]. This is particularly important for interactions between stars and supermassive black holes.
2.4 Recent numerical developments
Many of the key algorithmic developments for collisional N-body simulations were advanced very early on in the
1960s and 1970s [5, 45, 62, 63]. As a result, the eld has been largely driven by the extraordinary improvement in
hardware. From the early 1990s onwards, the slowest part of the calculation the direct N-body summation that
scales as N
2
was moved to special hardware chips called GRAPE processors (GRAvity PipE; Ito et al. [67]). The latest
GRAPE-6 processor manages an impressive 1 Teraop [68], allowing realistic simulations of star clusters with up to
10
5
particles [16]. However, to move toward the million star mark (relevant for massive star clusters), several GRAPE
processors must be combined in parallel. This became possible only very recently with the advent of the GRAPE-6A
chip [69]. The GRAPE-6A is lower performance (and cheaper) than the GRAPE-6, but specially designed to be used in a
parallel cluster. Such a cluster was recently used by Harfst et al. [6] to model a star cluster with N = 4 10
6
.
The GRAPE processors have been invaluable to the direct N-body community, and with the recently developed
GRAPE-DR, they will continue to drive the eld for some time to come [70]. However, concurrent with the further devel-
opment of the GRAPE chips, signicant interest is now shifting towards Graphical Processor Units (GPUs) for hardware
acceleration. This is driven primarily by cost. Even the smaller and cheaper GRAPE-6A costs several thousand dollars
at the time of writing and delivers 150 GigaFlops of processing power. By contrast GPUs deliver 130 GigaFlops
for just a couple of hundred dollars. The advent of a dedicated N-body library for GPUs makes the switch to GPUs
even easier [71]. Whether the future of direct N-body calculations lies in dedicated hardware, or GPUs remains to be
seen. A third way is entirely possible if new algorithms can make the force calculations more ecient. We discuss the
prospects for this, next.
2.5 Critique and numerical alternatives
One of the main problems with contemporary N-body codes for collisional stellar systems is their great similarity. An
immediate consequence is that the usual method for the validation of simulation results (see 4) by comparing inde-
pendent approaches is hardly possible. In addition, splitting the force computation into near and far eld components
is intimately connected with the time integration such that one cannot simply change one without the other.
The requirement for an accurate force computation by no means implies the need for the costly brute-force approach
currently employed. Alternatively, an approximative method with high accuracy, such as the fast-multipole method
(see 3.5.1), requires only O(N) instead of O(N
2
) operations to compute the forces for all N particles.
Time integration could perhaps also be improved. The Hermite scheme typically employed is neither symplectic
nor time-reversible. While this does not necessarily imply that the time integration method causes uncontrolled errors,
it would certainly be desirable to compare to an alternative method. One interesting option is to use fourth-order
forward symplectic integrators ([51], see also footnote 8) in conjunction with a time-symmetric method for adapting
the time steps, resulting in an overall time-reversible scheme.
2.6 Alternatives to N-body simulations
Collisional stellar systems are typically in dynamical, or virial, equilibrium where their overall properties, such as the
mean density and velocity distributions of stars, remain unchanged over dynamical time scales. However, owing to
stellar encounters, the systems evolves on much longer time scales. N-body methods follow the dynamics on all time
scales and thus do not exploit the fact that the system is almost in dynamical equilibrium.
An alternative is to use a mean-eld approach where a dynamical equilibrium is evolved to another dynamical
equilibrium, using some prescription for the processes, such as stellar encounters, driving this evolution. This is the
gist of Fokker-Planck codes, which approximately solve the collisional Boltzmann equation
df
dt
=
f
t
+ x
f
x

tot
x
f
x
= [f]. (35)
Here f(x, x, t) is the density of stars in six-dimensional phase-space {x, x}, also known as the distribution function
(see 3.1 for more on this). The encounter operator [f] describes the interaction between stars. In the limit [f] 0,
equation (35) recovers the collisionless Boltzmann equation (39). In general [f] is a complicated non-trivial functional
of f(x, x, t). When replacing [f] with an approximation obtained using certain simplifying assumptions, we obtain
the Fokker-Planck equation (FPE, not given here, see [1] for more details). The combined solution of the FPE and
the Poisson equation (38) is still a formidable problem: solving it in six phase-space coordinates plus time is simply
unfeasible. Several methods have been employed to tackle this problem, very briey summarised below.
Orbit averaging averages the net eect of the FPE over each orbit (assuming they are regular), thus reducing from
six to three dimensions. The resultant equations may then be solved on a mesh. Following the pioneering work of Cohn
[72], typically this has been done in two dimensions, integrating the specic energy and z-component of the angular
momentum (e.g. [73], we are unaware of any fully 3D calculation).
Monte-Carlo methods, pioneered by Henon [74], sample the stellar system using tracer particles, whose trajectories
are followed including not only the background potential but also the diusion in velocity according to the FPE.
Current implementations assume spherical symmetry to boost the resolution (for a recent example see [75]).
Fluid Models take velocity moments of the FPE, resulting in uid-like equations [76]. This avoids the need to orbit-
average the FPE, but (1) implicitly assumes that scattering events are local and (2) requires some assumption on the
distribution function to close the hierarchy of moment equations (i.e. an eective equation of state).
Despite the large number of assumptions that go into workable Fokker-Planck codes (independent local 2-body
encounters only, assumed Coulomb logarithm, etc.), the agreement with full N-body models is remarkable [7779].
Currently, with such a large number of assumptions Fokker-Planck codes mainly increase our understanding of the
N-body simulations, rather than act as an independent cross-check of the results. However, given the rapid advance
in computational power, it is perhaps time to revisit this approach. One may also consider alternatives to the FPE
(but still based on equation 35), for example similar to the Balescu [80]-Lenard [81] equation for plasma physics [82].
There remain signicant numerical challenges to such alternative approaches, but they hold the promise of a robust
method for solving collisional N-body dynamics that relies on very dierent assumptions to the N-body method.
2.7 Past, recent, and future astrophysical modelling
While the focus of this review rests rmly on numerical methods, we briey discuss our personal highlights of previous
astrophysical results, as well as recent and possible future developments.
Perhaps the earliest result was the numerical demonstration of core-collapse [5], which inspired the theory of
gravothermal-catastrophe [30]. Since then the increase in N to nearly 10
5
[14, 15, 32, 83] has conrmed the onset of
gravo-thermal oscillations after core-collapse, rst discovered using a uid code [84]. Exploring the eect of binaries,
several studies found that binaries may delay or even reverse core collapse [37, 85, 86].
Apart from being fascinating from a theorists point of view, core collapse may explain the rich variety of Globular
Clusters observed in our Galaxy [87], as well as potentially raising the central density enough to promote stellar
collisions. This latter process can seed the onset of runaway growth, leading to the formation of intermediate mass
black holes (IMBHs)
9
[93, 94]. The rm detection IMBHs in star clusters in the basis of dynamical evidence is quite
challenging [95, 96] and requires understanding of the collisional cluster dynamics [97, 98].
The most important force acting in collisional N-body systems is gravity, and solving the gravitational force
equation has been the focus of this review so far. However, many other interesting physical processes are at play
within the stars. Stars evolve over time, moving o the main sequence onto the giant branch and then eventually
ending their lives as stellar remnants [99]. Binary stars have even more complex lives, and all sorts of interesting
physics can occur due to mass transfer between them [100, 101]. In the very dense centres of star clusters, physical
stellar collisions can drive stellar evolution [102104]. And nally, the dearth of observed gas in star clusters may
provide interesting constraints on stellar evolution models, in which case gas must also be included in cluster models
[105, 106].
The rst attempt to mesh astrophysical processes with N-body models was presented in 1987 by Terlevich [107].
Since then, the number of stars modelled and the complexity of the stellar and binary evolution models has continued
to grow [108111]. Recently, there has been a dedicated drive to combine dierent physical simulation codes together
within one framework. This is the goal of the MODEST collaboration and its o-shoots
10
[Modelling and Observing
DEnse STellar systems; 17, 112115].
With ever-improving hardware and software, the challenges in modelling N-body systems will shift from solving
gravity, to building ever more realistic models for the physics beyond the particle resolution (stellar evolution, binary
evolution, collisions, gas physics, and feedback processes from stars and stellar remnants). Understanding these pro-
cesses, and building believable models, which can explain objects such as the 11 minute X-ray binary 4U1820-30 (see
footnote 5) as well as their distribution and evolution, will become a new frontier of computational astrophysics in the
coming decades.
3 N-body methods for collisionless systems
In collisionless stellar systems the long-term eects of two-body encounters are negligible. In other words, the gravi-
tational potential governing the stellar motions, which is actually the sum of many individual point-mass potentials,
is well approximated by a smooth mean potential (x, t). If unperturbed, a collisionless stellar system quickly (within
a few dynamical times) settles into dynamic equilibrium, when changes in the mean density and potential become
negligible and no further evolution occurs. In this situation, the virial theorem
2T +W = 0, with T =
i
1
2
m
i
x
2
i
and W =
i
m
i
x
i
x
i
(36)
holds and one speaks of virial equilibrium. Because relaxation in the thermodynamic sense does not occur, such
virial equilibria have generically non-Maxwellian and anisotropic velocity distributions, corresponding to a tensor-like
pressure in the uid picture.
The main exponent of collisionless stellar systems are galaxies and systems of galaxies (clusters and the universe
as a whole), which have number densities too small and dynamical times scales too long for stellar encounters to
be important. However, whereas for collisional systems external perturbations are usually weak, such perturbations,
including galaxy encounters and mergers, are frequent and often signicant for collisionless system. As a consequence,
collisionless systems are frequently perturbed away from equilibrium, resulting in evolution to a new equilibrium.
Another related process is secular evolution, when the perturbation originates from an instability of the system itself.
3.1 Equations governing collisionless stellar systems
Because the graininess of the stellar dynamics has negligible eect, collisionless stellar systems are commonly described
using continuum methods. The fundamental quantity describing the state of the system at any time is its distribution
function f(x, x, t), which is the mass density of stars (in the continuum limit) in six-dimensional phase-space {x, x} at
time t. This is a signicant simplication compared to collisional systems (whose phase-space is 6N dimensional) and
as a consequence any correlations between particles, such as binaries and encounters, are ignored
11
. The continuous
9
There are a number of stellar-mass black-hole candidates [88], as well as super-massive black-hole candidates (10
69
M,
found at the centres of galaxies like our own Milky Way, e.g. [89, 90]). However, there is no conrmed discovery of an IMBH
with 10
34
M, which are dicult to detect unambiguously, but would provide a missing link [91, 92].
10
See also http://www.manybody.org/modest/.
11
In other words, the BBGKY hierarchy (after their discoverers Bogoliubov, Born, Green, Kirkwood, and Yvon, see also [1])
of the 1-body, 2-body, and higher order distribution functions is truncated at the lowest order.
Fig. 4. The eect of mixing demonstrated with simple phase-mixing of 10
4
points in the Hamiltonian H =
1
2
p
2
+ |q|. This
corresponds to massless tracer particles orbiting a central point mass in one-dimensional gravity (where phase-space is two-
dimensional). The ne-grained distribution function is either 1 or 0, but at late times a smooth coarse grained distribution
appears. Note that dynamical mixing in 3D self-gravitating systems is much faster and stronger than with this toy model.
spatial density and gravitational potential generated by the system are obtained as
(x, t) =
_
d xf(x, x, t), (37)
(x, t) = G
_
dx
(x
, t)
|x x
|
= G
__
d xdx
f(x
, x, t)
|x x
|
, (38)
respectively. Because, according to Liouvilles theorem, phase-space volume remains constant along the ow and
because of mass conservation, the ratio f = dM/d{x, x} is constant too, thus satisfying
0 =
df
dt
=
f
t
+ x
f
x

tot
x
f
x
, (39)
known as the collisionless Boltzmann equation (CBE). Here, we have used the equation of motion x =
tot
with
the total potential
tot
= +
ext
, which includes contributions from external agents not modelled by the distribution
function f. The evolution of a collisionless system is thus governed by the CBE in conjunction with the Poisson
equation (38) and, possibly, an external potential.
For equilibrium systems f/t = 0 and thus /t = 0. In this case, any distribution function which depends
on the phase-space co-ordinates only through isolating integrals of motion
12
solves the CBE. This is known as the
Jeans theorem and allows the construction of simple equilibrium models. Unfortunately, the most general collisionless
equilibria are triaxial and even though most of their orbits are regular and respect three isolating integrals, only the
orbital energy allows simple treatment. Therefore, the Jeans theorem is of practical usage only for systems of higher
symmetry, where the angular momentum (or one of its components) is also an isolating integral.
For near-equilibrium systems, approximate solutions can be obtained by writing f = f
0
+ f
1
(with f
1
f
0
and
f
0
describing a given equilibrium model), and ignoring the term of the CBE quadratic in f
1
. The resulting linear
perturbation analysis gives insight into the stability properties of simple equilibrium models. However, most galaxies
frequently undergo strong perturbations and considerable deviations from equilibrium, when these methods fail and
instead a full numerical treatment is required.
The direct numerical solution of the CBE, a non-linear PDE in seven dimensions, is not feasible. This is not just
because of the vastness of six-dimensional phase-space and the inhomogeneity of f (typically most of the mass resides
12
Any function I(x, x), such that

I = 0 and the constraint I(x, x) = I0 reduces the dimensionality of phase-space by
one (isolates) is an isolating integral of motion. Examples are the orbital energy for static potentials, or the orbital angular
momentum in case of a spherical potential.
in a tiny fraction of the bound phase space). A much more severe problem is that under the CBE f develops ever
stronger gradients, even when evolving towards equilibrium. This is because f is conserved along the ow, so that
initial uctuations are not averaged away, but mixed leading to ever thinner layers of dierent phase-space density.
These require ever higher numerical resolution even though the system hardly evolves observably, as demonstrated
in Fig. 4. This gure also demonstrates that the mixing invalidates the continuum limit: at late times the concept
of the distribution function becomes useless. An alternative is to average over the uctuations by considering the
coarse-grained phase-space density

f, a local mean of f. However, for

f no evolution equation exists
13
.
3.2 Numerics of collisionless N-body simulations
Fortunately enough, the above problems are easily overcome by the usage of N-body techniques. The basic idea is to
model the distribution function by an ensemble of N phase-space points {x
i
, x
i
}, i = 1 . . . N with weights
i
, which are
randomly chosen to represent f(x, x, t = 0). The conservation of f along the ow implied by the CBE then means that
the weights
i
remain unchanged along each trajectory, such that the task is reduced to integrating all N trajectories.
Thus, in one sense an N-body code solves the CBE by the method of characteristics for solving PDEsthe particle
trajectories are the characteristics of the CBE. In another sense, an N-body code is a Monte-Carlo technique: any
initial sample of N phase-space points drawn from the same distribution function at t = 0 results in another, equally
valid, N-body model for the time evolution of f(x, x, t). An important consequence is that the N simulated particles
are not modelling individual stars, nor is it helpful to consider them as super stars of enormous mass. The only
correct interpretation is that the ensemble of all N particles together represents the continuous distribution function
f (and in fact provides an implied coarse-graining dependant on the numerical resolution). Thus unlike the situation
for collisional N-body methods, the number N of particles is a numerical parameter, controlling the resolution and by
implication the accuracy for certain predictions of the model.
With the N-body method one has no knowledge of the distribution function f at t > 0, except at the phase-space
positions of the N particles (where f remains at its original value by virtue of the CBE). However, this is not a problem
at all, not least because the structure of f is not necessarily very useful (see above), but also because any moments of
the distribution function can be estimated from the N particles in the usual Monte-Carlo way
g(t)
__
dxd xg(x, x) f(x, x, t)
i
g
_
x
i
(t), x
i
(t)
_
. (40)
Here, the function g species the moment, for example g = 1 obtains the total mass, while g =
1
2
x
2
results in the total
kinetic energy. Such moments are in fact all one ever wants to know: all observable properties, including the coarse-
grained distribution function, are just moments of f. However, one must not forget that this estimation procedure is
always subject to shot noise, the amplitude of which depends on the number of particles eectively contributing, and
hence on the numerical resolution and the width of the moment function g.
3.3 Cosmological N-body simulations
For cosmological simulations, in principle we should switch to general relativistic (GR) equations of motion. However,
we can take advantage of a very useful approximation. On large scales, the Universe is very nearly isotropic and
described well by a smooth Friedmann-Lamatre-Robertson-Walker (FLRW) spacetime [116]. However, on small scales
the Universe must tend towards a locally inertial frame in which Newtons laws are valid. This fact can be used to
show that the cosmological expansion has essentially no eect on the local dynamics, even on galaxy cluster scales
[117]. Thus, we may think of the Universe as being lled with a self-gravitating uid that locally obeys Newtons law
of gravitation, while expanding as an FLRW metric on large scales
14
. This leads to the following Hamiltonian for any
orbiting test-particle [122, 123]:
H =
p
2
i
2m
i
a
2
+
m
i
2a
(x
i
), (41)
13
One option is to use the CBE also for

f and combine it with some averaging or coarse-graining operation, depending on the
level of uctuations developing under the CBE. However, the choice of an appropriate coarse-graining operation is a fundamental
problem with such an approach, partly because phase-space has no natural metric. One would have to use

f itself to dene
the coarse-graining and it is not clear how this should be done, in particular for cold models, when f is non-zero only on a
hypersurface. This is the common situation with initial conditions for cosmological simulations.
14
Recently, doubts have been raised about the validity of this approximation. Various authors suggested that local inhomo-
geneities aect the large-scale dynamics, for example mimicking cosmic acceleration which traditionally has been attributed to
dark energy [118, 119]. This seems to be unlikely [120], though the necessary corrections due to inhomogeneities could aect
attempts to use N-body simulations to precisely determine cosmological parameters [121].
where x
i
and p
i
= a
2
m
i
x
i
are now co-moving canonical coordinates (with respect to the isotropic, smooth, expanding
background spacetime); a a(t) is the scalefactor that follows from the FLRW model; and the equations of motion
follow in the usual way from Hamiltons equations. Since we are now working in co-moving space, the potential is
really the peculiar potential with respect to the smooth background; while the forces are peculiar forces with respect
to the expansion. Note that the above Hamiltonian now explicitly depends on time and thus does not conserve energy.
This is the standard problem of ambiguous energy conservation in GR [for a discussion see e.g. 122].
The usual method for approximating isotropy and homogeneity of the universe on large scales is to simulate a small
cubic patch of size L of the universe (to satisfy the local-Newtonian approximation), and apply periodic boundary
conditions in co-moving co-ordinates
(x, t) = G
n
_
dx
(x
+ nL, t)
|x x
nL|
= G
n
__
d xdx
f(x
+ nL, x, t)
|x x
nL|
, (42)
where the sum over n = (n
x
, n
y
, n
z
) accounts for all periodic replica. In practice, the periodic sum is approximated
using Ewald [124]s method, which was originally invented for solid-state physics and imported to this eld by Hernquist
et al. [125] (but note an error in their eq. 2.14b as pointed out by [126]). Alternatively, Fourier methods, which naturally
provide periodic boundary conditions, can be used, see 3.5.2.
3.4 Force softening
In order to obtain the particle trajectories, one just has to integrate the equations of motion x
i
=
tot
(x
i
) for all N
particles. This is usually done using the leapfrog integrator possibly with individual timesteps arranged in the block-step
scheme (2.3.1). The self-potential must be estimated from the positions and masses of the particles themselves. By
virtue of equations (38) or (42), is a moment of f and so we can estimate it. However, the straightforward application
of equation (40) results simply in the equations of motion (2) for a collisional system with N stars of masses
i
(and
possibly immersed in an external gravitational eld). This is not what we want to model and, moreover, integrating
these equations numerically is rather dicult, as we have discussed in 2.3 above.
The problem is that the function g = G/|x x
| for estimating the potential via equation (40) is quite localised

and even diverges for x x
, such that shot noise becomes a serious issue with the simple Monte-Carlo integration
of equation (40). This is closely related to estimating the spatial density . The application of equation (40) with
g = (xx
) obtains a sum of -functions at the instantaneous particle positions, consistent with the accelerations (2)
from the Poisson equation 4G = a. The problem of estimating a smooth density from scattered data points
(in several dimensions) is generic to many applications in science. A number of solutions are known. One of them is
to widen the -spikes in the density estimate to a nite size, resulting in the estimator [127]
(x) =
3

_
|x x
i
|
_
, (43)
where () is the dimensionless kernel function (normalised to unit integral) and the softening length. The corre-
sponding estimator for the potential then follows by application of the Poisson equation, or equivalently, by replacing
the Greens function G/|x x
| with the potential generated by a mass distribution of the density kernel
(x) =
i
G
i

_
|x x
i
|
_
(44)
where 4 =
2
(
2
(a prime denoting dierentiation). For example, the commonly used Plummer softening
(x) = G
i
_
|x x
i
|
2
+
2
(45)
corresponds to replacing each particle by a Plummer [128] sphere of scale radius and mass
i
, i.e.
() = (3/4) (
2
+ 1)
5/2
and () = (
2
+ 1)
1/2
. (46)
3.4.1 Softening as a method to suppress close encounters
The softening of gravity has two important aspects. First, at close distances r = |x
i
x| the force no longer diverges
as r
2
, but for r actually decays to zero. As a consequence, the force estimated from all particles is a smooth
and continuous function, such that integrating the equations of motion is much simpler than for the un-softened case
(required with collisional N-body methods). This also means that the eect and importance of close particle encounters
is suppressed. This is exactly what we want, because close encounters are not described by the one-body distribution
function f, but by the 2-body distribution function, the next in the BBGKY hierarchy (see also footnote 11). The
minimum softening length required to prevent large-angle deections during close encounters is
2body
G/
2
(47)
with the particle mass and the typical velocity dispersion [129].
In other words, close encounters between the integrated trajectories are numerical artifacts and by softening the
forces we eliminate their (articial) eects and, at the same time, considerably simplify the task of time integration
compared to collisional N-body codes. Unfortunately, two-body relaxation is driven not only by close encounters, but all
octaves of impact parameter contribute equally. Therefore, force softening does not, contrary to some common opinion,
much reduce the articial two-body relaxation (only by 2, [130]). This implies that articial relaxation eects may
aect the high-density regions of N-body simulations (where t
dyn
and hence t
relax
is shortest, see equation 1). Thus
while we aim to model collisionless systems, the simulations themselves may still suer from small-angle deections.
3.4.2 Softening as optimal force estimation: the choice of the softening kernel
The second aspect of force softening is a systematic reduction of gravity at close distances: in the limit of N but
xed , the gravity estimated by (44) disagrees with that of the system modelled. This bias of the average N-body
force is the price for the reduction in shot noise (i.e. suppression of close encounters). The presence of this force bias
implies that simulation results on scales smaller than a few are unreliable.
The overall force error is a combination of this bias and the noise, measured by the variance of the force estimate,
and depends on the system modelled, the number N of particles, the softening length and the softening kernel .
For small , the bias and variance for the estimated force

F can be approximated analytically using a local Taylor
expansion [131]
bias{
F(x)} = a
0
2
G(x) +a
2
4
G
2
(x) +O(
6
) with a
k
=
(4)
2
(k + 3)!
_

0
d
k+4
(), (48)
Nvar{
F(x)} = b G
2
M
1
(x) +O(
0
) with b = (4)
2
_

0
d
2
() (). (49)
with (x) and M the (smooth) density of the system modelled and its total mass, respectively. Thus, the total force
error, bias{
F(x)}
2
+var{
F(x)}, becomes minimal for

opt
N
1/5
, depending on the system and the softening kernel.
In particular, it is required that
5
() 0 as for the constant a
0
to be nite. For kernels, such as Plummer
softening (for which
5
at 1), which do not satisfy this condition, the expansion above does not work and
the force bias grows faster than quadratic.
While these static considerations may not be directly relevant for the dynamical evolution of the N-body system,
it appears best to reduce the force bias at least to O(
2
) by using softening kernels with either nite density support
(corresponding to exact 1/r gravity and = 0 for r > ) or a decay steeper than
5
at .
In view of equation (48), one may even reduce the bias further by designing softening kernels with a
0
= 0 [131].
Such kernels have < 0 over some radial range (usually at large ), and compensate the rare under-estimation of
gravity at small radii with a slight over-estimation at larger radii. The resulting reduction in bias allows a larger
softening length to suppress the noise. However, equation (48) is applicable only when the density (x) is smooth on
scales smaller than . Since many collisionless systems have quite steep central density gradients, where may even
formally diverge like a power law, techniques (like this) which require larger softening length appear less advisable.
3.4.3 The choice of the softening length
While
2body
(equation 47) represents a minimum softening length required to suppress articial large-angle deections,
larger values simplify the time integration further and hence reduce the computational costs. Another useful criterion
is that the maximum inter-particle force shall not exceed the typical mean-eld strength. For the situation of a single
stellar system of mass M and radius R, this translates to G/
2
GM/R
2
. Replacing M = N, we nd [132]
min
R/
N. (50)
This is signicantly larger than
2body
2R/N (which follows from equation 47 with
2
GM/2R). Most practitioners
are guided by these simple considerations, and practical convergence studies regarding the best choice for are lacking.
A notable exception is the work of Power et al. [132], who suggest for cosmological simulations 4
min
.
3.4.4 Adaptive individual softening
The local resolution of an N-body system is determined by the particle number density and an obvious idea is to use
smaller softening lengths in high-density regions. This leads to the concept of individual
i
with
2
ij
= (
2
i
+
2
j
)/2 the
softening used in the interaction between particles i and j (other symmetrisations are possible, but this particular one
allows an ecient approximation when using the tree code [133]). The individual softening lengths are adapted, for
example such that
3
i

i
remains constant, equivalent to the adaption of individual smoothing lengths in smoothed-
particle hydrodynamics (SPH, [134] see also Lodato & Cossins, this issue).
When implementing such a scheme, two things must be ensured in order to guarantee the validity of the N-
body method. First, the adaption of the softening lengths must be time-reversible. This can be achieved to sucient
accuracy by a technique equivalent to that used in SPH [135, appendix A1]. Second, since
i
depends (implicitly) on
the positions of particle i and its neighbours, the N-body force
/x
i
contains additional terms, which must be
included, as outlined by Price & Monaghan [136], to preserve the Hamiltonian character of the method and hence
energy conservation.
3.5 Force computation
In collisionless N-body methods the force is only ever an estimate, which unavoidably carries with it an estimation
error (which is reduced by force softening, see above). Therefore, we may as well use less accurate methods for the ac-
tual calculation of the estimated forces than the computationally expensive direct summation, i.e. the straightforward
implementation of equation (44) or (45). Because the computational eort of any N-body method is always dominated
by the calculation of the gravitational forces, considerable eort has been invested into the design of fast force calcu-
lation algorithms, resulting in many dierent methods, which we describe in some detail below. All of these methods
are substantially faster than direct summation and together with the simplications for the time-integration due to
force softening, allow N to be 4 orders of magnitude higher in collisionless than collisional N-body simulations.
3.5.1 Approximating direct summation
A number of methods are based on approximating the direct summation
(x
b
) =
a
(x
b
x
a
) (51)
(corresponding to equation 44 with G and absorbed into ) by replacing the contributions from all particles within
a local group by a single expression.
The tree code was pioneered by Barnes & Hut [137] in 1986 and uses a hierarchical spatial tree to dene localised
groups of particles. Because stellar systems are often highly inhomogeneous, this is much better than dening groups
by cells of an equidistant mesh (when few cells would contain most of the particles). With the usual oct-tree each cubic
cell containing fewer than n
max
particles is split into up to eight child cells of half their parents size. This results in a
tree-like hierarchy of cubic nodes with the root box, containing all particles, at its bottom. The particles within each
of the tree nodes constitute a well-dened and localised group. The approximation used in the tree code is formally
obtained by Taylor expanding the kernel function in x
a
around some expansion centre z
A
of the group
15
(x
b
x
a
)
|n|p
1
n!
(x
a
z
A
)
n
n
(x
b
z
A
), (52)
where p is the expansion order. By inserting this expansion into (51) we obtain for the potential from the group A the
multipole expansion
A
(x
b
) =
aA
a
(x
b
x
a
)
|n|p
M
n
(z
A
) D
n
(x
b
z
A
) (53)
with the derivatives D
n
(r)
n
(r) and the multipoles of group A w.r.t. its expansion centre z
A
M
n
(z
A
) =
aA
a
(1)
n
n!
(x
a
z
A
)
n
. (54)
15
Using the multi-index notation n (nx, ny, nz) with ni 0, n |n| nx + ny + nz, r
n
r
nx
x
r
ny
y
r
nz
z
, and n! nx! ny! nz!.
Fig. 5. Left: computation of the force for one of 100 particles (asterisks) in two dimensions (for graphical simplicity) using
direct summation: every line corresponds to a single particle-particle force calculation. Middle: approximate calculation of the
force for the same particle using the tree code. Cells opened are shown as black squares with their centres z indicated by solid
squares and their sizes w by dotted circles. Every green line corresponds to a cell-particle interaction. Right: approximate
calculation of the force for all 100 particles using the tree code, requiring 902 cell-particle and 306 particle-particle interactions
( = 1 and nmax = 1), instead of 4950 particle-particle interactions with direct summation.
For the un-softened case ( = 1/|r|), this series converges in the limit p if |x
b
z
A
| > max
a
{|x
a
z
A
|}, i.e. if x
b
is outside a sphere centred on z
A
and containing all x
a
. Furthermore, if the centre of mass of the group A is chosen as
its expansion centre z
A
, the dipole vanishes such that the zero-order expansion (p = 0) is actually rst-order accurate.
Because of its great simplicity, this approach is in fact a common choice for practical implementations of the tree code.
In a preparatory step, the multipole moments (54) and sizes w
A
satisfying w
A
max
a
{|x
a
z
A
|} are computed
for each tree cell. (Both is best done recursively, exploiting the results from the daughter cells via the shifting formula
M
n
(z +x) =
|k||n|
x
k
k!
M
nk
(z), (55)
sometimes called the upward pass.) The gravity at any position x from all the particles within tree cell A is then
simply approximated by applying equation (53) if
16
w
A
< r = |x z
A
| with some opening angle 1. Otherwise,
the sum of the potentials obtained by applying the same algorithm to the daughter cells is used (or, if the cell is a tree
leaf, from direct summation over all particles within the cell). Of course, the force is computed as the derivative of

.
Fig. 5 demonstrates the working of the tree code and compares it graphically to the direct summation approach,
by applying both to the task of computing the force for one of 100 particles (left and middle panels). Obviously, the
tree code requires much fewer calculationsit is straightforward to show that the number of force computations per
particle scales like the depth of the tree, i.e. ln N, such that the cost of computing all N forces is O(N ln N).
The fast multipole method (FMM) also works with localised particle groups and for stellar systems is best imple-
mented using a tree structure [139, 140], although the original proposal used a xed grid [141]. In addition to expanding
the Greens function (x
b
x
a
) at the source positions x
a
(as for the tree code), it also expands it at the sink positions
x
b
(with z
B
the centre of the group B of sink particles, see also the left panel of Fig. 6):
(x
b
x
a
)
|n|p
|m|pn
(1)
n
n!m!
(x
b
z
B
)
n
(x
a
z
A
)
m
n+m
(z
B
z
A
). (56)
Inserting this into (51), yields for the potential generated by all particles in A and at any position x
b
within B
AB
(x
b
)
|n|p
1
n!
(x
b
z
B
)
n
F
n
(z
B
), (57)
F
n
(z
B
) =
|m|pn
M
m
(z
A
) D
n+m
(z
B
z
A
) (58)
16
Ensuring convergence of the series. Curiously, some early implementations used for wA simply the linear size of the cubic
cell, when wA maxa{|xa zA|} is not guaranteed and the approximated forces can be catastrophically wrong, resulting in
the infamous exploding galaxies bug [138].
Fig. 6. Left: Illustration of the geometry for the Taylor expansion used with the fast multipole method. Right: approximate
calculation of the force for the same 100 particles as in Fig. 5 using the FMM, requiring 132 cell-cell (blue), 29 cell-particle
(green), and 182 particle-particle (red) interactions ( = 1 and nmax = 1).
with the multipole moments M
m
(z
A
) dened in equation (54). The eld tensors F
n
(z
B
) are the coecients of the
Taylor series (57) for the potential around z
B
. This dual expansion at both ends of all interactions considerably
speeds up the simultaneous computation of gravity for all particles, but brings no advantage over the tree code when
computing the force at a single position. The FMM algorithm works these equations backwards, starting with an
upward pass (as for the tree code) to compute the multipole moments M
m
(z
A
) and widths w
A
for all cells.
The second part is the interaction phase, when the eld tensors are evaluated for each cell. This is achieved by the
following algorithm starting with the root-root interaction
17
. If for a mutual cell-cell interaction (w
A
+ w
B
) < r =
|z
A
z
B
|, then the eld tensors for the interactions A B and B A are computed according to equation (58) and
added to F
n
(z
B
) and F
n
(z
A
), respectively. Otherwise, the interaction is split into up to 8 new interactions by opening
the bigger (in terms of w) of the two cells. A cell self-interaction (like the initial root-root interaction) is performed by
simple direct summation if the cell contains only a few particles, otherwise it is split into up to 36 new interactions.
Finally, in a down-ward pass the contributions from the parent cells are added to the eld-tensors of their daughters
after applying the shifting formula
F
n
(z +x) =
|k|pn
x
k
k!
F
n+k
(z), (59)
followed by the evaluation of gravity via equation (57) at the sink positions within leaf cells. The total computational
costs of this algorithm are dominated by the interaction phase, which only requires O(N) interactions for the compu-
tation of all N particle forces. This represents a substantial reduction from the O(N
2
) for direct summation. It is also
a factor 10 faster than the tree code for typical N 10
6
.
Some practical considerations
1. The simple convergence criterion |r| < w commonly used with the tree code and the FMM is merely geometric
and therefore controls the relative error for each interaction [140]
|
AB
|
|
AB
|

(p + 1)
p
(1 )
2
. (60)
However, usually the forces from small (and hence nearby) cells are smaller than from bigger distant cells such
that the absolute force error is dominated by the few interactions with big cells. A better approach is therefore to
try to balance the force errors by a non-geometric opening criterion, for example using a mass-dependent opening
angle. With this, one can obtain a total cost < O(N) at xed approximation error [140].
2. For the un-softened Greens function = 1/r, the multipole expansion is reduced from three to two-dimensional
indices when using spherical harmonics. Essentially, this exploits the fact that D
n+2x
+D
n+2y
+D
n+2z
= 0 for any
n (as a consequence of 0 =
2
), such that the number of independent terms is reduced from
_
p+3
p
_
to (p + 1)
2
.
This reduces the costs for the FMM interaction operation (58) from O(p
6
) to O(p
4
), which may be further reduced
to O(p
3
) (by fast rotation methods, e.g. [142]), O(p
2
ln p) (by fast Fourier methods [143]), or even O(p
2
) (using
plane wave approximations [144]).
17
Here, we describe the falcON algorithm [140], but essentially any multipole-based method which expands the Greens functions
both at the source and sink positions and therefore performs cell-cell interactions qualies as FMM.
Fig. 7. Logical interaction plots for the computation of all mutual forces between the same N = 100 particles as in Fig. 5.
Left: direct summation: every red dot corresponds to the force computation for one of N(N 1)/2 particle pairs. Middle: tree
code: the approximated force for a particle i from a cell containing many j is represented by a green line (particles are ordered in
their tree order: all particles within a tree cell are contiguous). Forces between close neighbours often cannot be approximated,
resulting in the clustering of red dots along the diagonal. Right: using the FMM code falcON: each blue box corresponds to
the approximation for the force between two cells. Unlike with the tree code, green lines represent mutual interactions.
3. Both the distribution of force errors and the total computational costs depend on the opening criterion, the expan-
sion order, and other numerical details. For an optimal choice of these numerical parameters, the computational cost
depends non-trivially on the required approximation error. With collisionless N-body codes one usually requires
only a modest relative force accuracy of 10
3
, in which case low-order techniques are sucient.
4. Fig. 7 gives an alternative graphical comparison of direct summation, the tree code, and FMM for the computation
of all N forces. With direct summation, each pair-wise interaction is considered only once because the mutual
forces satisfy F
ij
= F
ji
, according to Newtons third law.
With the tree code, this natural symmetry between sinks and sources is broken: each interaction is one-sided.
As a consequence, the full interaction matrix (except for the diagonal) has to be approximated and Newtons third
law is violated, implying the loss of total-momentum conservation.
With the FMM on the other hand, each interaction is mutual, exploiting the natural symmetry of the problem
and satisfying Newtons third law for each particle pair (though the approximated F
ij
is not exactly aligned with the
separation vector x
i
x
j
). Only the upper half of the interaction matrix needs to be approximated. Further away
from the diagonal, a single cell-cell interaction (blue box) approximates ever more particle-particle interactions.
5. From Fig. 7 one can also see that for the tree code the workload can be easily split between many processors.
This is more dicult with direct summation or the FMM. For example, when distributing the particles amongst
processors, it is not obvious which computes the forces between particles (or cells) residing on dierent processors
(but see [9]).
Critique Apart from the violation of Newtons third law with the tree code (but not the FMM), the only potential
short-coming is a discontinuity of the forces: there are some magic boundaries in space on either side of which a
dierent approximation will be used, depending on the opening criterion. As a consequence, the approximated forces
are not conservative across these boundaries and the total energy of the N-body system is not exactly conserved but
exhibits some uctuations, even in the limit of zero time step. Of course, the amplitude of this eect can be reduced
by decreasing the opening angle .
It appears that this issue has little if any eect on the validity of the N-body model (there are no obvious
dierences between results obtained using the tree code and methods which do not suer from this problem). In
principle it should be possible to design a tree code or FMM without this problem by smoothly interpolating between
the force approximations on either side of the magic boundaries. The natural adaptivity and eciency of the tree code
and FMM make them versatile and powerful force solvers for general-purpose N-body codes.
3.5.2 Grid-based methods
Instead of solving the integral form (38) or (42) of the Poisson equation, as with direct summation and its approxi-
mations, grid-based methods solve its dierential form
2
(x) = 4G(x) (61)
after discretisation on a grid.
Fast-Fourier-transform-based methods achieve this in the Fourier domain where the Poisson equation (61) becomes
k
2
(k) = 4G(k). The method rst estimates the density on the vertices of an equidistant grid by a technique which
spreads the mass of each particle across neighbouring cells (see [27] for details). Next, the Poisson equation is solved,
using the fast Fourier transform (FFT), for the value of the potential on each grid vertex; and nally the potential
and force for each particle is found by interpolation between vertices.
This method is fast (the cost for the FFT and the interpolation are O(n
grid
ln n
grid
) and O(N), respectively), but
not eective for inhomogeneous particle distributions typical for most N-body simulations, when in order to resolve
the high-density regions n
3
grid
N is required.
The Fourier approach implicitly assumes a periodic tiling of all space with the computational domain, such that it
actually (approximately) solves
2
(x) = 4G
n
(x + nL). (62)
This corresponds not to equation (38) but (42), exactly as desired for cosmological N-body simulations of large-scale
structure formation. Therefore, FFT-based methods (or hybrid methods using FFT, see below) are predominantly
used with such simulations.
The periodic boundary conditions can be avoided either by simply doubling the computational domain in each
dimension or better by James [145] method, which subtracts the contributions from the periodic replicas of the
computational box via a Fourier technique involving surface charges. Combining this with a set of nested grids of
increasing resolution enables an ecient FFT-based force solver for inhomogeneous single stellar systems, such as
galaxies [146].
Multi-grid techniques were pioneered in the west by Brandt [147] in 1977; they also interpolate the density and
potential between grid and particles, but solve the discretised form of the Poisson equation, i.e. a large but sparse
matrix equation, using relaxation methods, such as Gauss-Seidel iteration. The basic idea exploits the fact that
on a coarser grid relaxation occurs faster because information travels faster. The distribution of errors (the dierence
between the actual density and that obtained via the discretisation of
2
from the current estimate for the potential)
is rst smoothed on the nest grid by a few Gauss-Seidel iterations. After transferring the problem to a coarser grid,
the process is repeated on coarser and coarser grids until, on the coarsest grid convergence is achieved. Then the
problem is transferred back to ner and ner grids, each time iterating until convergence.
Since the costs of an iteration shrinks by a factor eight when going to the next coarser grid, the total cost is
essentially determined by the costs of a xed number of iterations on the nest grid, and hence O(n
grid
) in theory
[148]. However, as far as we are aware, in practical applications for N-body simulations, the method is not signicantly
more ecient than others.
The advantage of this technique over the FFT approach is that the grid does not need to be equidistant, but can be
locally adapted according to the particle density. In fact, the structure of such an adaptively rened mesh is identical
to that of a shallow oct-tree, as used with the tree-code and FMM.
Critique Gravitational softening is implicit with grid methods, in contrast to direct summation and its approximations,
and depends on the grid size. This means that for adaptive meshes the softening may vary along a particle orbit,
depending on the local mesh resolution. Since softening unavoidably leads to a reduction in the estimated binding
energy of a particle (3.4.2), these variations introduce an unphysical uctuation of particle binding energies. In
non-equilibrium simulations, for example of gravitational collapse, this leads to violation of energy conservation and
articial secular evolution. This problem also occurs with individually adapted softening lengths with explicit softening
(as with the tree code or FMM), but there a solution is known, see 3.4.4.
With adaptive multi-grid methods, it is natural to use dierent time steps (within the block-step scheme) for par-
ticles living on dierent renement levels. Such a scheme can be substantially accelerated by advancing the particles
asynchronously: the particles on the coarser grid remain xed while those on the ner grid move and their gravity is ap-
proximated by using the coarser-grid potential as boundary condition. However, such a method is neither Hamiltonian
nor time symmetric, potentially resulting in articial secular evolution (and violation of energy conservation).
3.5.3 Basis function methods
The idea of basis-function methods, pioneered by Clutton-Brock [149] in 1972 and later dubbed self-consistent eld
code [150], is to expand the mass density into basis functions
n
(x) with coecients C
n
(x) =
n
C
n
n
(x), (63)
where n = (n, l, m) is a three-dimensional set of indices, and estimate the potential as
(x) = G
n
C
n
n
(x), (64)
where 4
n
(x) =
2
n
(x). Usually, the sets of basis functions used are complete and bi-orthogonal, i.e. satisfy
nn
=
_
dx
n
(x)
n
(x), (65)
(x x
) =
n
(x)
n
(x
), (66)
1
|x x
|
=
n
(x)
n
(x
), (67)
where the integral is over all space and the sums include all terms |n| . Applying the bi-orthogonality relation
(65) to the density estimate (63) gives C
n
=
_
dx (x)
n
(x), which upon inserting of the Monte-Carlo estimator for
the mass density (equation 40 with g a delta-spike) yields
C
n
=
n
(x
i
). (68)
Thus, the basis-function force solver rst calculates the coecients C
n
via equation (68) and subsequently computes
potential and force at any position via equation (64) and its derivative. Since both of these operations are trivially split
between multiple processes and require minimal communication, this method is ideal for computational parallelism.
Of course, in practice one has to truncate the expansion at some nite order n
max
. Ideally, the error made by this
truncation is small, such that the signal in the neglected coecients C
n>nmax
is negligible. This is usually achieved by
choosing the parameters of a given basis-function set such that
0
already closely matches the system modelled and
higher-order terms merely describe deviations. A more systematic approach is to design the basis functions to match
the system at hand [151] but also to truncate the expansion smoothly according to the estimated signal-to-noise in
the neglected coecients [152].
Critique Of course, the basis-function approach is not suitable for modelling wildly dynamic situations, such as
galaxy mergers, but only for near-equilibrium dynamics, when the stellar system deviates only slightly from the
smooth zeroth-order basis function. An advantage of the method is the eective softening, which usually is small in
high-density regions where the the basis functions vary mostly. Unlike Greens-function softening, the method does not
bias the force of a density cusp ( r
) if the basis functions contain the same power-law cusp and the N-body
system is centred on the origin of the expansion. In fact, o-centring is a signicant problem with this method.
It has been argued that the basis-function method is ideally suited for such near-equilibrium systems, because the
computational cost scales linear with N. However, this view is too simplistic, since it is futile to increase N but not
also the force resolution, i.e. n
max
. Simple arguments based on equivalence to Greens-function softening suggest that
the optimum resolution for given N requires an increase of n
max
such that the overall computational cost scale like
O(N
2
), as for straightforward direct summation [131].
One may hope to alleviate this problem by a continually adapting the lowest-order basis function [151] such that
n
max
can be kept low. However, this approach changes the approximated Greens function
(x x
) =
nmax
n
(x)
n
(x
) (69)
during the simulation and therefore introduces an articial time dependence into the approximate Hamiltonian. At
best, this only destroys energy conservation, but more likely has other adverse eects which are less obvious to identify.
The basis-function method may be most useful for simulations with constrained symmetry. For example, enforcing
spherical symmetry simply amounts to setting C
nlm

l0
m0
when using a basis based on spherical-harmonics.
This reduces the computational costs substantially, even when including many radial basis functions, thus enabling
extremely fast spherically symmetric simulations with large N. The basis-function method is certainly very useful as
force solver for other purposes, for example to approximate the density and/or potential of an external system or to
model the potential in perturbation analyses.
3.5.4 Hybrid methods
The advantages and disadvantages of the various force solvers naturally lead to the concept of hybrid methods, to
avoid the respective disadvantages. Most relevant in this context is presumably the usage of FFT-based methods to
obtain periodic boundary conditions, which are desired in cosmological simulations. The disadvantage of the FFT-
based method (also called PM particle-mesh) is the lack of resolution and adaptivity on small scales. In the 1980s, the
early days of cosmological simulations, the combination of PM with direct summation (PP: particle-particle) for the
computation of the near-neighbour force decit (the dierence between the average FFT force and that for the desired
softening length) enjoyed some popularity as P
3
M codes [153]. Later, this approach has been improved by replacing
direct summation with a tree code (TreePM [154]), but we are not aware of the obvious FMM-PM combination.
Combinations of the multi-grid method with the FFT are straightforward (the FFT is used as force solver on the
coarsest grid [155]).
Other hybrid methods combine the basis-function approach in a subset of the spatial dimensions with a grid in
the remaining, for example using spherical harmonics with a radial grid [156, 157].
3.6 Recent numerical developments and challenges
Collisionless N-body simulations have benetted enormously from the incredible increase in computer power, in
particular as the computational costs only increase like N ln N (for the tree code and grid methods). This combined
with massive parallelisation on many thousand cores has allowed very large N simulations. A further increase by a
factor 10 should be possible just by using the FMM as force solver (though implementing the FMM eciently
in parallel is challenging). Another welcome numerical advancement would be better time-stepping methods (see the
discussion at the end of 2.3.3).
However, the main challenge of contemporary applications of collisionless N-body methods lie not in the method,
but in the astrophysics it misses out: any non-gravitational interaction. Baryonic matter contributes only 16% of all
matter on large scales [116] and inter-stellar gas contributes only little to the mass of individual galaxies (in the Milky
Way the mass ratio between gas and stars is about 1:9). However, the gas not only interacts gravitationally (with all
other matter), but can directly dissipate (and absorb) energy in form of radiation, and therefore behaves fundamentally
dierently from point-mass particles.
One big challenge of contemporary astrophysics is to model the formation of galaxies ab initio. To this end, many
complex astrophysical processes and phenomena must be modelled, such as the multi-phase nature of the gas, its
condensation to stars and active galactic nuclei (AGN), and their feedback (via winds and radiation) onto the gas
(all these process may be summarised as baryon physics). While hydrodynamics is comparatively straightforward
to add to the method, e.g. via smoothed particle hydrodynamics [134], most of the baryonic physics is not. This
is because these processes are themselves not properly understood and can only be modelled via parametric sub-
resolution models. For example, a gas particle is turned into a star particle according to some parametric model of
our current understanding of the star formation process. While substantial progress has been made, the challenges
are still formidable, not only because of our limited astrophysical understanding of the baryon physics, but also since
much higher numerical resolution may be required for convergence than with simple gravity-only N-body experiments.
3.7 Past, recent, and future astrophysical modelling
As in 2.7, we provide a very brief summary of selected (by personal opinion) highlights of astrophysical results based
on collisionless N-body simulations (apology to anybody who feels missed-out).
One of the earliest results was the modelling by (author?) in 1972 [158, see also 3, 159]
18
of tidal interactions
of rotationally supported disc galaxies, demonstrating a large variety of observed phenomena, such as tidal arms and
bridges. Simulations of galaxy interactions could explain many other observed phenomena such as shells, ripples, and
other ne structure [160163], as well as ring galaxies [164]. Also, the merger origin of elliptical galaxies received
strong support from N-body simulations [165, 166].
The bar instability of isolated disc galaxies was discovered in the rst simulations of such systems [167, 168],
providing a natural explanation of the high frequency of barred galaxies. More recent simulations suggest a close
connection between the dynamics of bars and outer rings [169] and demonstrate the importance of a dark-matter halo
as an angular-momentum absorber [170, 171]. The buckling instability of bars was rst seen in N-body simulations
[172], though the identication of boxy or peanut-shaped bulges as side-on bars was only made later. Simulations of
18
Sadly, this excellent work is only little-known (63 citations), largely because publishing in the west was extremely dicult
for Soviet scientists during the cold-war era. The later study by Toomre & Toomre (ApJ 178, 623) with essentially the same
simulation set-up but only N = 120 (compared to N = 2000) is much more widely known (over 1700 citations).
disc galaxies also showed spontaneous formation of transient and/or long-lived spiral features [173], which prompted
the theoretical model of swing-amplication by Toomre & Zang [174].
Cosmological simulations based on cold dark-matter match the observed large-scale structure of the universe
[175, 176], while hot (e.g. neutrino) dark-matter is ruled out [177]. The dark-matter haloes formed in such simulations
possess a universal density prole [21, 178], which varies between r
1
at r 0 and r
3
to r
4
at large radii,
and triaxial shapes [21, 179]. This is now observationally conrmed on galaxy cluster scales [180182]. Finally, such
simulations predict a wealth of substructure that should be present orbiting the Milky Way, and in the Local Group
that is not (yet) observed [183, 184].
Confronting such predictions, as well as those of almost all applications of collisionless N-body simulations, with
observations requires an extension of the method to include baryonic physics (see 3.6). The goal for the coming
decades will be realistic simulations of galaxy formation and evolution, which will allow us to test both our current
cosmological model, and galaxy formation theories.
4 Validation
A very important issue with N-body techniques is the validation of the numerical method in general and for the specic
purpose in particular. While a 100% validation is never possible, there are several powerful validation techniques
available. In particular for collisionless N-body techniques, every project requires a validation to ensure that the
numerical method (which includes the values of numerical parameters such as N) guarantees the desired numerical
accuracy, thus avoiding systematic errors driven by numerical artifacts.
Validation of the method Before even considering the application of a certain N-body program (collisional or colli-
sionless) for a particular modelling purpose, the values of numerical parameters (controlling time step, force approx-
imation, etc.) required for correctness must be established. This can be done in two ways: rst by simulating simple
but non-trivial situations for which the correct outcome is known from other means (such a perturbation analysis or,
in the case of collisional N-body code by using any of the alternative methods of 2.6); and second (for collisionless
N-body methods) by re-simulating a certain situation with ever better numerical accuracy, which should converge to
a (hopefully correct) answer.
Validation by Monitoring One of the most straightforward validation requirements is the actual numerical conserva-
tion of total energy, momentum, and angular momentum. While with a nite time step energy conservation can never
be fully achieved, a relative energy conservation to 3-4 digits is usually considered sucient in practice for collisionless
N-body techniques, while for collisional methods 5-6 digits appear to be the norm. A systematic trend of energy hints
at an articial secular evolution and is much more problematic than mere uctuations. The same applies to momentum
and angular momentum.
In order to monitor whether gravitational softening introduces signicant bias (if for instance many particles overlap
within the softening length), one can compare the virial W and the potential energy V , dened as
W =
i
x
i

_
self
(x
i
) +
ext
(x
i
)
_
, V =
i
_
1
2
self
(x
i
) +
ext
(x
i
)
_
(70)
with

self
and
ext
denoting, respectively, the (approximated and softened) N-body potential and any external po-
tential. For pure Newtonian gravity (unsoftened), one has W = V and deviations from this equality are indicative of
the global importance of the gravity reduction due to softening.
Validation by simulation When the eect of some particular perturbation is to be modelled, for example of an infalling
satellite, it is important to ensure that, at the resolution at which the simulation results will be interpreted, a control
simulation without the perturbation behaves as expected. Sometimes, this may not be sucient, for example when
the perturber gives rise to additional astrophysical eects not present in and hence not scrutinised by the control
simulation. In such a situation one must employ convergence tests: an equivalent simulation at signicantly higher
resolution should give consistent results. Even this may fail when the simulations converge to the wrong answer.
Validation by comparison The only way to to shield against this problem is to re-simulate with a completely dif-
ferent N-body method. For example, various N-body tests have been done within the ESF-funded astro-sim project
(http://www.astrosim.net/code). With collisionless N-body methods this is often the only reliable validation tech-
nique, since there are hardly any non-linear problems with solutions known from other methods (against which one
could compare). Such projects are important and have demonstrated that cosmological codes now agree at the 10%
level for important metrics [185]. This is excellent progress, but not good enough for next generation cosmological
probes like Euclid (http://sci.esa.int/euclid).
Validation and interpretation Simulation results, in particular when unexpected, should never be trusted at face
value, but an eort must be made to understand them in astrophysical terms. This is very often non-trivial, because
the highly non-linear and complex dynamics is not compatible with any simplifying assumptions that would allow
insight by means of approximate analytical models. However, unlike the situation with real stellar systems and galaxies,
the N-body model can be analysed (observed) in full six-dimensional phase-space allowing a much better handle
on the dynamical processes than nature itself. Contemporary analysis techniques for N-body data are still somewhat
immature, but their discussion beyond the scope of this paper.
5 Conclusion
We have reviewed some of the latest numerical techniques for modelling collisional and collisionless N-body systems.
Our focus throughout has been on purely gravitational N-body simulations, with a view to presenting the key numerical
algorithms. Over the past 50 years of N-body calculations, the eld has undergone dramatic changes. Improved
software algorithms, specialised hardware, and ecient parallel programming have meant that N has kept pace with
Moores Law, nearly doubling every two years. This has allowed us to simulate the most massive star clusters, galaxies,
and the Universe as a whole with increasing precision.
It is clear that our modelling of gravity is in good shape. Dierent codes and techniques give converged results,
while both collisional and collisionless simulations continue to push to larger N. However, the coming challenge will
be building believable models of baryonic processes in the Universe. This is beyond the scope of this short review, but
is the key challenge for future N-body simulations on all scales.
References
1. J.J. Binney, S. Tremaine, Galactic dynamics. 2nd ed (Princeton University Press, Princeton 2008)
2. S.D.M. White, MNRAS 174, 19 (1976)
3. E. Holmberg, ApJ 94, 385 (1941)
4. S. von Hoerner, Zeitschrift f ur Astrophysik 50, 184 (1960)
5. S.J. Aarseth, MNRAS 126, 223 (1963)
6. S. Harfst, A. Gualandris, D. Merritt, R. Spurzem, S.F. Portegies Zwart, P. Berczik, New Astronomy 12, 357 (2007)
7. V. Springel, S.D.M. White, A. Jenkins, C.S. Frenk, N. Yoshida, L. Gao, J. Navarro, R. Thacker, D. Croton, J. Helly, et
al., Nature 435, 629 (2005)
8. R. Teyssier, S. Pires, S. Prunet, D. Aubert, C. Pichon, A. Amara, K. Benabed, S. Colombi, A. Refregier, J. Starck, A&A
497, 335 (2009)
9. J. Stadel, D. Potter, B. Moore, J. Diemand, P. Madau, M. Zemp, M. Kuhlen, V. Quilis, MNRAS 398, L21 (2009)
10. I.T. Iliev, B. Moore, S. Gottlober, G. Yepes, Y. Homan, G. Mellema, MNRAS tmp, 296, arXiv:1005.3139 (2011).
11. E. Terlevich, in Star Clusters, edited by J.E. Hesser (IAU Symposium 85, 1980)
12. S. Inagaki, Pub. Astron. Soc. Japan 38, 853 (1986)
13. S.J. Aarseth, D.C. Heggie, in The Globular Cluster-Galaxy Connection, edited by G.H. Smith & J.P. Brodie (Astronomical
Society of the Pacic Conference Series 48, 1993)
14. R. Spurzem, S.J. Aarseth, MNRAS 282, 19 (1996)
15. J. Makino, ApJ 471, 796 (1996)
16. H. Baumgardt, J. Makino, MNRAS 340, 227 (2003)
17. P. Hut, New Astronomy Review 54, 163 (2010)
18. P.J.E. Peebles, AJ 75, 13 (1970)
19. G. Efstathiou, J.W. Eastwood, MNRAS 194, 503 (1981)
20. C.S. Frenk, S.D.M. White, G. Efstathiou, M. Davis, Nature 317, 595 (1985)
21. J. Dubinski, R.G. Carlberg, ApJ 378, 496 (1991)
22. B. Moore, F. Governato, T. Quinn, J. Stadel, G. Lake, ApJ 499, L5 (1998)
23. S. Ghigna, B. Moore, F. Governato, G. Lake, T. Quinn, J. Stadel, ApJ 544, 616 (2000)
24. V. Springel, J. Wang, M. Vogelsberger M., A. Ludlow, A. Jenkins, A. Helmi, J.F. Navarro, C.S. Frenk, S.D.M. White,
MNRAS 391, 1685 (2008)
25. S.J. Aarseth, Gravitational N-Body Simulations (Cambridge University Press, Cambridge 2003)
26. D.C. Heggie, P. Hut, The Gravitational Million-Body Problem: A Multidisciplinary Approach to Star Cluster Dynamics
(Cambridge University Press, Cambridge 2003)
27. R.W. Hockney, J.W. Eastwood, Computer simulation using particles, (Hilger, Bristol 1988)
28. M. Trenti. P. Hut, scholarpedia http://www.scholarpedia.org/article/N-body simulations (arXiv:0806.3950 2008)
29. V.A. Antonov, Solution of the problem of stability of stellar system Emdens density law and the spherical distribution of
velocities (Vestnik Leningradskogo Universiteta, Leningrad 1962)
30. D. Lynden-Bell, R. Wood, MNRAS 138, 495 (1968)
31. L. Spitzer, Dynamical evolution of globular clusters (Princeton University Press, Princeton 1987)
32. H. Baumgardt, D.C. Heggie, P. Hut, J. Makino, MNRAS 341, 247 (2003)
33. D.C. Heggie, MNRAS 173, 729 (1975)
34. J.G. Hills, AJ 80, 809 (1975)
35. S.J. Aarseth, J.G. Hills, A&A 21, 255 (1972)
36. J.M. Fregeau, M.A. G urkan, K.J. Joshi, F.A. Rasio, ApJ 593, 772 (2003)
37. D.C. Heggie, M. Trenti, P. Hut, MNRAS 368, 677 (2006)
38. C.D. Bailyn, ARA&A 33, 133 (1995)
39. H.B. Perets, D.C. Fabrycky, ApJ 697, 1048 (2009)
40. B. Lanzoni, N. Sanna, F.R. Ferraro, E. Valenti, G. Beccari, R.P. Schiavon, R.T. Rood, M. Mapelli, S. Sigurdsson, ApJ
663, 1040 (2007)
41. A. King, ApJ 732, L28 (2011)
42. S.J. Aarseth, MNRAS 378, 285 (2007)
43. S. Harfst, A. Gualandris, D. Merritt, S. Mikkola, MNRAS 389, 2 (2008)
44. E.H. Morgan, R.A. Remillard, M.R. Garcia, ApJ 324, 851 (1988)
45. A. Ahmad, L. Cohen, J. Comp. Phys. 12, 389 (1973)
46. K. Nitadori, J. Makino, New Astronomy 13, 498 (2008)
47. H. Yoshida, Celestial Mechanichs and Dynamical Astronomy 56, 27 (1993)
48. Q. Sheng, IMA J. Num. Analysis 9, 199 (1989)
49. M. Suzuki, J. Math. Phys. 32, 400 (1991)
50. B. Leimkuhler, S. Reich, Simulating Hamiltonian Dynamics, (Cambridge University Press, Cambridge 2005)
51. S. Chin, C. Chen, Celestial Mechanichs and Dynamical Astronomy 91, 301 (2005)
52. G.D. Quinlan, S. Tremaine, AJ 100, 1694 (1990)
53. P. Hut, J. Makino, S.L.W. McMillan, ApJ 443, L93 (1995)
54. T. Holder, B. Leimkuhler, S. Reich, Applied Numerical Mathematics 39, 367 (1999)
55. E. Kokubo, K. Yoshinaga, J. Makino, MNRAS 297, 1067 (1998)
56. J. Makino, Pub. Astron. Soc. Japan 43, 859 (1991)
57. J. Makino, P. Hut, M. Kaplan, H. Saygn, New Astronomy 12, 124 (2006)
58. M. Zemp, J. Stadel, B. Moore, C.M. Carollo, MNRAS 376, 273 (2007)
59. J. Makino, ApJ 369, 200 (1991)
60. P. Kustaanheimo, A. Schinzel, H. Davenport, E. Stiefel, Journal f ur die reine und angewandte Mathematik 218, 204 (1965)
61. H. Yoshida, Celestial Mechanics 28, 239 (1982)
62. S.J. Aarseth, K. Zare, Celestial Mechanics 10, 185 (1974)
63. D.C. Heggie, Celestial Mechanics 10, 217 (1974)
64. S. Mikkola, S.J. Aarseth, Celestial Mechanichs and Dynamical Astronomy 47, 375 (1990)
67. T. Ito, J. Makino, T. Ebisuzaki, D. Sugimoto, Computer Physics Communications 60, 187 (1990)
68. J. Makino, T. Fukushige, M. Koga, K. Namura, Pub. Astron. Soc. Japan 55, 1163 (2003)
69. T. Fukushige, J. Makino, A. Kawai, Pub. Astron. Soc. Japan 57, 1009 (2005)
70. J. Makino, in Dynamical Evolution of Dense Stellar Systems, edited by E. Vesperini, M. Giersz, & A. Sills (IAU Symposium
246, 2008)
71. E. Gaburov, S. Harfst, S.F. Portegies Zwart, New Astronomy 14, 630 (2009)
72. H. Cohn, ApJ 234, 1036 (1979)
73. K. Takahashi, Pub. Astron. Soc. Japan 47, 561 (1995)
74. M. Henon, Ap&SS 13, 284 (1971)
75. M. Freitag, W. Benz, A&A 375, 711 (2001)
76. R.B. Larson, MNRAS 147, 323 (1970)
77. R. Spurzem, K. Takahashi, MNRAS 272, 772 (1995)
78. D.C. Heggie, M. Giersz, R. Spurzem, K. Takahashi, Highlights of Astronomy 11, 591 (1998)
79. E. Kim, I. Yoon, H.M. Lee, R. Spurzem, MNRAS 383, 2 (2008)
80. R. Balescu, Physics of Fluids 3, 52 (1960)
81. A. Lenard, Annals of Physics 10, 390 (1960)
82. J. Heyvaerts, MNRAS 407, 355 (2010)
83. M. Giersz, D.C. Heggie, MNRAS 268, 257 (1994)
84. D. Sugimoto, E. Bettwieser, MNRAS 204, 19P (1983)
85. S.L.W. McMillan, P. Hut, J. Makino, ApJ 362, 522 (1990)
86. A. Tanikawa, T. Fukushige, Pub. Astron. Soc. Japan 61, 721 (2009)
87. W.E. Harris, AJ 112, 1487 (1996)
88. C.D. Bailyn, R.K. Jain, P. Coppi, J.A. Orosz, ApJ 499, 367 (1998)
89. J. Magorrian, S. Tremaine, D. Richstone, R. Bender, G. Bower, A. Dressler, S.M. Faber, K. Gebhardt, R. Green, C. Grill-
mair, J. Kormendy, T. Lauer, AJ 115, 2285 (1998)
90. J.E. Greene, C.Y. Peng, M. Kim, C. Kuo, J.A. Braatz, C.M. Violette Impellizzeri, J.J. Condon, K.Y. Lo, C. Henkel,
M.J. Reid, ApJ 721, 26 (2010)
91. P. Madau, M.J. Rees, ApJ 551, L27 (2001)
92. M. Volonteri, F. Haardt, P. Madau, ApJ 582, 559 (2003)
93. S.F. Portegies Zwart, S.L.W. McMillan, ApJ 576, 899 (2002)
94. S.F. Portegies Zwart, H. Baumgardt, P. Hut, J. Makino, S.L.W. McMillan, Nature 428, 724 (2004)
95. K. Gebhardt, R.M. Rich, L.C. Ho, ApJ 578, L41 (2002)
96. K. Gebhardt, R.M. Rich, L.C. Ho, ApJ 634, 1093 (2005)
97. H. Baumgardt, J. Makino, P. Hut, S.L.W. McMillan, S.F. Portegies Zwart, ApJ 589, L25 (2003)
98. H. Baumgardt, J. Makino, P. Hut, ApJ 620, 238 (2005)
99. A.C. Phillips, The Physics of Stars, 2nd Edition, (Wiley, 1999)
100. R.P. Church, J. Dischler, M.B. Davies, C.A. Tout, T. Adams, M.E. Beer, MNRAS 395, 1127 (2009)
101. C. Lajoie, A. Sills, ApJ 726, 66 (2011)
102. A. Sills, J.C. Lombardi, Jr., C.D. Bailyn, P. Demarque, F.A. Rasio, S.L. Shapiro, ApJ 487, 290 (1997)
103. I.A. Bonnell, M.R. Bate, H. Zinnecker, MNRAS 298, 93 (1998)
104. H. Baumgardt, R.S. Klessen, MNRAS tmp, 321, arXiv:1009.1189 (2011)
105. D.J. Faulkner, K.C. Freeman, ApJ 211, 77 (1977)
106. W. Priestley, M. Ruert, M. Salaris, MNRAS 411, 1935 (2011)
107. E. Terlevich, MNRAS 224, 193 (1987)
108. S.F. Portegies Zwart, S.L.W. McMillan, P. Hut, J. Makino, MNRAS 321, 199 (2001)
109. H. Baumgardt, P. Hut, J. Makino, S.L.W. McMillan, S.F. Portegies Zwart, ApJ 582, L21 (2003)
110. J.R. Hurley, O.R. Pols, S.J. Aarseth, C.A. Tout, MNRAS 363, 293 (2005)
111. R.P. Church, C.A. Tout, J.R. Hurley, Pub. Astron. Soc. Australia 26, 92 (2009)
112. P. Hut, M.M. Shara, S.J. Aarseth, R.S. Klessen, J.C. Lombardi, Jr., J. Makino, S. McMillan, O.R. Pols, P.J. Teuben,
R.F. Webbink, New Astronomy 8, 337 (2003)
113. A. Sills, S. Deiters, P. Eggleton, M. Freitag, M. Giersz, D.C. Heggie, J. Hurley, P. Hut, N. Ivanova, R.S. Klessen, et al.,
New Astronomy 8, 605 (2003)
114. M.B. Davies, P. Amaro-Seoane, C. Bassa, J. Dale, F. de Angeli, M. Freitag, P. Kroupa, D. Mackey, M.C. Miller, S.F. Porte-
gies Zwart, New Astronomy 12, 201 (2006)
115. R. Spurzem, in JENAM-2007, Our Non-Stable Universe (2007)
116. E. Komatsu, K.M. Smith, J. Dunkley, C. L. Bennett, B. Gold, G. Hinshaw, N. Jarosik, D. Larson, M.R. Nolta, L. Page,
et al., ApJS 192, 18 (2011)
117. F.I. Cooperstock, V. Faraoni, D.N. Vollick, ApJ 503, 61 (1998)
118. V.F. Mukhanov, L.R.W. Abramo, R.H. Brandenberger, Physical Review Letters 78, 1624 (1997)
119. T. Buchert, M. Kerscher, C. Sicka, Physical Review D 62, 043525 (2000)
120. A. Ishibashi, R.M. Wald, Classial and Quantum Gravity 23, 235 (2006)
121. A. Enea Romano, P. Chen, arXiv:1104.0730 (2011)
122. P.J.E. Peebles, The large-scale structure of the universe, (Princeton University Press, 1980)
123. V. Springel, MNRAS 364, 1105 (2005)
124. P.P. Ewald, Annalen der Physik 369, 253 (1921)
125. L. Hernquist, F.R. Bouchet, Y. Suto, ApJS 75, 231 (1991)
126. R. Klessen, MNRAS 292, 11 (1997)
127. B.W. Silverman, Density estimation for statistics and data analysis, (Chapman & Hall, London, 1986)
128. H.C. Plummer, MNRAS 71, 460 (1911)
129. S.D.M. White, MNRAS 189, 831 (1979)
130. C. Theis, A&A 330, 1180 (1998)
131. W. Dehnen, MNRAS 324, 273 (2001)
132. C. Power, J.F. Navarro, A. Jenkins, C.S. Frenk, S.D.M. White, V. Springel, J. Stadel, T. Quinn, MNRAS 338, 14 (2003)
133. T.R. Saitoh, J. Makino, arXiv:1005.1752 (2010)
134. D.J. Price, J. Comp. Phys., in press (arXiv:1012.1885, 2011)
135. L. Cullen, W. Dehnen, MNRAS 408, 669 (2010)
136. D.J. Price, J.J. Monaghan, MNRAS 374, 1347 (2007)
137. J.E. Barnes, P. Hut, Nature 324, 446 (1986)
138. J.K. Salmon, M.S. Warren, J. Comp. Phys. 111, 136 (1994)
139. W. Dehnen, ApJ 536, L39 (2000)
140. W. Dehnen, J. Comp. Phys. 179, 27 (2002)
141. L. Greengard, V. Rokhlin, J. Comp. Phys. 73, 325 (1987)
142. D. Pinchon, P.E. Hoggan, J. Phys. A: Math. Theory 40, 1597 (2007)
143. W.D. Elliott, J.A. Board, Jr., SIAM J. Sci. Comp. 17, 398 (1996)
144. H. Cheng, L. Greengard, V. Rokhlin, J. Comp. Phys. 155, 468 (1999)
145. R.A. James, J. Comp. Phys. 25, 71 (1977)
146. J. Magorrian, MNRAS 381, 1663 (2007)
147. A. Brandt, Mathematics of Computation 31, pp. 333 (1977)
148. U. Trottenberg, C.W. Oosterlee, A. Sch uller, Multigrid (Academic Press, 2001)
149. M. Clutton-Brock, Ap&SS 16, 101 (1972)
150. L. Hernquist, J.P. Ostriker, ApJ 386, 375 (1992)
151. M.D. Weinberg, AJ 117, 629 (1999)
152. M.D. Weinberg, ApJ 470, 715 (1996)
153. G. Efstathiou, M. Davis, S.D.M. White, C.S. Frenk, ApJS 57, 241 (1985)
154. P. Bode, J.P. Ostriker, ApJS 145, 1 (2003)
155. A.V. Kravtsov, A.A. Klypin, A.M. Khokhlov, ApJS 111, 73 (1997)
156. T.A. McGlynn, ApJ 281, 13 (1984)
157. M. Trenti, G. Bertin, T.S. van Albada, A&A 433, 57 (2005)
158. N.N. Kozlov, R.A. Syunyaev, T.M.

Eneev, Soviet Physics Doklady 17, 413 (1972)
159. T.M.

Eneev, N.N. Kozlov, R.A. Sunyaev, A&A 22, 41 (1973)
160. P.J. Quinn, ApJ 279, 596 (1984)
161. C. Dupraz, F. Combes, A&A 166, 53 (1986)
162. L. Hernquist, P.J. Quinn, ApJ 331, 682 (1988)
163. L. Hernquist, P.J. Quinn, ApJ 342, 1 (1989)
164. R. Lynds, A. Toomre, ApJ 209, 382 (1976)
165. J.E. Barnes, ApJ 331, 699 (1988)
166. J.E. Barnes, Nature 344, 379 (1990)
167. R.H. Miller, K.H. Prendergast, W.J. Quirk, ApJ 161, 903 (1970)
168. F. Hohl, ApJ 168, 343 (1971)
169. E. Athanassoula, M. Romero-G omez, J.J. Masdemont, MNRAS 394, 67 (2009)
170. E. Athanassoula, MNRAS 341, 1179 (2003)
171. J. Dubinski, I. Berentzen, I. Shlosman, ApJ 697, 293 (2009)
172. F. Combes, R. H. Sanders, A&A 96, 164 (1981)
173. T. Zang, PhD thesis (Massachusetts Institute of Technology, 1976)
174. A. Toomre, in Structure and Evolution of Normal Galaxies, edited by S.M. Fall & D. Lynden-Bell (1981)
175. M. Davis, G. Efstathiou, C.S. Frenk, S.D.M. White, ApJ 292, 371 (1985)
176. V. Springel, C.S. Frenk, S.D.M. White, Nature 440, 1137 (2006)
177. S.D.M. White, C.S. Frenk, M. Davis, ApJ 274, L1 (1983)
178. J.F. Navarro, C.S. Frenk, S.D.M. White, ApJ 462, 563 (1996)
179. M.S. Warren, P.J. Quinn, J.K. Salmon, W.H. Zurek, ApJ 399, 405 (1992)
180. P. Saha, J.I. Read, L.L.R. Williams, ApJ 652, L5 (2006)
181. P. Saha, J.I. Read, ApJ 690, 154 (2009)
182. K. Umetsu, T. Broadhurst, A. Zitrin, E. Medezinski, D. Coe, M. Postman, arXiv:1105.0444 (2011)
183. A. Klypin, A.V. Kravtsov, O. Valenzuela, F. Prada, ApJ 522, 82 (1999)
184. B. Moore, S. Ghigna, F. Governato, G. Lake, T. Quinn, J. Stadel, P. Tozzi, ApJ 524, L19 (1999)
185. K. Heitmann, Z. Lukic, P. Fasel, S. Habib, M.S. Warren, M. White, J. Ahrens, L. Ankeny, R. Armstrong, B. OShea,
P.M. Ricker, V. Springel, J. Stadel, H. Trac, Computational Science and Discovery 1, 015003 (2008)

N - Body Simulations of Gravitational Dynamics

Uploaded by

Copyright:

Available Formats

N - Body Simulations of Gravitational Dynamics

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

N - Body Simulations of Gravitational Dynamics

Uploaded by

Copyright:

Available Formats

a

serves only as an auxiliary quantity.

, and using the relation L

| for estimating the potential via equation (40) is quite localised

| with the potential generated by a mass distribution of the density kernel

F(x)}, becomes minimal for

You might also like