2 From Boltzmann to Navier-Stokes to Euler

Reading: Ryden ch. 1, Shu chs. 2 and 3

2.1 The distribution function and the Boltzmann equation

Define the distribution function f (~x, ~v , t) such that f (~x, ~v , t)d3 xd3 v = probability of finding a
particle in phase space volume d3 xd3 v centered on ~x, ~v at time t.
The normalization is
Z Z
f (~x, ~v , t)d3 xd3 v = N (# particles in system)

and
[f ] = cm−3 ( cm s−1 )−3 .

Particles are not created or destroyed, so continuity implies

∂f ∂f ∂f df
X  
+ ẋi + v̇i = |c , (7)
∂t i=1,3 ∂xi ∂vi dt

where dfdt |c represents discontinuous motion of particles through phase space because of collisions.
(Collisions cannot instantaneously change particle positions, but they can instantaneously change
particle velocities.)
Substituting ẋi = vi and v̇i = gi leads to the Boltzmann equation
∂f X ∂f X ∂f df
+ vi + gi = |c . (8)
∂t i
∂xi i
∂vi dt

Stellar dynamics is based on the collisionless Boltzmann equation, with the RHS=0.
In the fluid (λ ≪ L) limit, on the other hand, the collision term makes f (~v ) approximately
Maxwellian while locally conserving mass, momentum, and energy.
The Boltzmann equation is hard to manage because it is 6-d, but it tells us more than we really
want to know in most cases.
We are usually happy with the density, mean velocity, and velocity dispersion as a function of ~x,
since the velocity distribution function is close to Maxwellian.
We can therefore get more useful equations by taking moments of the Boltzmann equation.
A similar procedure is used in stellar dynamics to obtain the Jeans equations. However, these are
less powerful than the hydrodynamic equations because in the absence of collisions the velocity
distribution function may be far from Maxwellian (in particular, it may be anisotropic).

The mass density is Z

ρ(~x, t) = mf (~x, ~v , t)d3 v, (9)

5
where m is the particle mass. If necessary, one can sum over different f ’s for different particle
types, but we will assume a single particle species here.
The mass-weighted average of a quantity Q at position ~x is
1
Z
hQi = Qmf (~x, ~v , t)d3 v. (10)
ρ

2.2 The continuity equation

Multiply both sides of the Boltzmann equation by m and integrate over d3 v:
∂ X ∂ Z ∂ df 3
Z Z X Z
mf d3 v + mf vi d3 v + m (gi f )d3 v = m |c d v.
∂t i
∂xi i
∂vi dt

First term is
∂ρ
.
∂t

Second term is
X ∂
~ · (ρ~u) ,
(ρhvi i) ≡ ∇
i
∂xi
where

~u ≡ h~v i = mean fluid velocity at (~x, t) [ cm s−1 ]

w
~ ≡ ~v − ~u = particle random velocity [ cm s−1 ].

Third term is Z Z
m ~ v · (~g f )d3 v = m
∇ n̂ · (~g f )dA = 0.
V S
The first equality follows from the divergence theorem and the second from the assumption that f
vanishes as v −→ ∞.

RHS vanishes because of local mass conservation: collisions do not create or destroy particles at a
fixed position, only shift them in velocity space.

Result:
∂ρ ~
+ ∇ · (ρ~u) = 0.
∂t
More transparently
∂ρ ~ + ρ∇
~ · ~u = 0,
+ ~u · ∇ρ
∂t
implying a continuous change of mass density.

6
2.3 The momentum equation
Multiply Boltzmann equation by m~v and integrate.
∂ X ∂ Z ∂f 3 df 3
Z Z X Z
3 3
mvj f d v + mf vj vi d v + m gi vj d v= mvj |c d v.
∂t i
∂xi i
∂vi dt

First term is
∂
(ρuj ).
∂t

Second term is
X ∂ X ∂
(ρhvj vi i) = (ρui uj + ρhwi wj i) .
i
∂xi i
∂xi

For the third term, note that

∂ ∂f
(vj f ) = vj + δij f
∂vi ∂vi
to obtain
∂
X Z   X Z
mgi (vj f ) − δij f d3 v = − gi δij mf d3 v = −ρgj ,
i
∂vi i
where δij ≡ 1 for i = j, 0 for i 6= j, and we have used the divergence theorem to get rid of the first
term in the brackets.

RHS vanishes because collisions conserve momentum. (Note: Ryden sets the collision term to zero
before integrating on the grounds that it will be zero in equilibrium, but I think this is not justified
in all circumstances.)

Result:
∂ X ∂
(ρuj ) + (ρui uj + ρhwi wj i) = ρgj .
∂t i
∂xi

The diagonal terms of hwi wj i are generally much larger than the off-diagonal terms, since random
velocities in different directions are usually almost uncorrelated.
It therefore makes sense to divide the ρhwi wj i term into a contribution from pressure and a con-
tribution from viscosity:
1
P ≡ pressure
= ρh|w| ~ 2 i [ dyne cm−2 or erg cm−3 ] (11)
3
πij ≡ viscous stress tensor = P δij − ρhwi wj i [ dyne cm−2 or erg cm−3 ] (12)

to obtain
∂ X ∂
(ρuj ) + (ρui uj + P δij − πij ) = ρgj ,
∂t i
∂xi
or, in tensor form
∂
 ↔

~ · ρ~u~u + P I − ↔
(ρ~u) + ∇ π = ρ~g.
∂t

7
 ↔
Here (~u~u)ij = ui uj , I ij = δij .
↔
All three tensors are symmetric, and π is traceless.

One can combine this form of the momentum equation with the continuity equation to get the form
∂uj X ∂uj 1X ∂
+ ui = gj − (P δij − πij )
∂t i
∂xi ρ i ∂xi

or
∂~u  ~  1~ 1~ ↔
+ ~u · ∇ ~u = ~g − ∇P + ∇· π.
∂t ρ ρ
Viscosity acts to oppose shearing motion and interpenetration.

2.4 The energy equation

One can multiply by mv 2 and go through a similar procedure (see Ryden and/or Shu) to derive
the internal energy equation
∂ǫ ~ = −P ∇
~ · ~u − 1 ∇ ~ + 1Ψ
~ ·F
+ ~u · ∇ǫ
∂t ρ ρ ρ

1
ǫ ≡ specific internal energy = h|w|~ 2 i [ erg g−1 ] (13)
2
~ ≡ 1
F conduction heat flux = ρhw| ~ w|~ 2 i [ erg cm−3 s−1 cm] (14)
2
X ∂ui
Ψ ≡ viscous dissipation rate = πij [ erg cm−3 s−1 ]. (15)
i,j
∂x j

One again makes use of the divergence theorem and of the fact that collisions conserve energy (as
well as mass and momentum).

Note that if the distribution of w ~ is symmetric about zero, then F ~ vanishes. If the distribution
is skewed, then hot particles have a drift relative to cold particles, producing a heat flux in the
direction of the drift.
In most cases, a temperature gradient produces a conductive flux F~ ∝ ∇T ~ .
However, if F~ is uniform, heat flowing out is replaced by heat flowing in. The local thermal energy
changes only if ∇~ · F~ 6= 0.

Ψ represents conversion of bulk motion of the fluid into internal energy via viscous dissipation. It
is the viscous analog of heating by P dV work.

2.5 Eulerian and Lagrangian derivatives

An Eulerian reference frame is one that stays fixed in space. (This has nothing to do with “Euler
equations,” except that it is the same Euler.)
A Lagrangian reference frame is one that moves with the fluid.

8
∂Q
∂t is the time derivative of Q at a fixed Eulerian position.
The time derivative along an element moving with the mean flow ~u,
DQ ∂Q  ~ 
= + ~u · ∇ Q, (16)
Dt ∂t
is known as the Lagrangian derivative.

The physical interpretation of hydrodynamics equations may be more transparent in Lagrangian

formulation.
Numerical calculations can be performed in either Eulerian or Lagrangian frames. Often one is
more powerful than the other for a particular application.
Lagrangian codes tend to be powerful when there is a large dynamic range in density and one needs
higher resolution in high density regions than in low density regions.
Smoothed particle hydrodynamics (SPH) is an example of a Lagrangian hydrodynamics algorithm,
where one follows fluid properties at the locations of particles, which move with the flow.

2.6 The Navier-Stokes equations

Collecting our results, we have the Navier-Stokes equations:
Dρ ∂ρ ~ ~ · ~u,
= ∂t + ~u · ∇ρ = −ρ∇ (17)
Dt
D~u ∂~
u

~ ~u = ~g − 1 ∇P

~ + 1 ∇· ~ ↔
= ∂t + ~u · ∇ π, (18)
Dt ρ ρ
Dǫ ~ P~ 1~ ~ 1 Γ−Λ
= ∂ǫ
∂t + ~
u · ∇ǫ =− ∇ · ~u − ∇ ·F + Ψ+ (19)
Dt ρ ρ ρ ρ

Γ ≡ volumetric radiative heating rate [ erg cm−3 s−1 ]

Λ ≡ volumetric radiative cooling rate [ erg cm−3 s−1 ].

(17) represents local mass conservation. The change in density of a Lagrangian fluid element is
produced by a change in the specific volume.

(18) represents local momentum conservation. Accelerations are produced by gravity, pressure
gradients, and viscous forces.

(19) represents local energy conservation. Changes in internal energy are produced by P dV work,
by conduction, by viscous heating (conversion of bulk kinetic energy), and by radiative heating or
cooling, which we have added to our previous results.

~g may be specified externally or, if self-gravity is important, computed from Poisson’s equation
~ · ~g = −4πGρ.
∇
↔
Ψ is determined once ~u and π are specified.
Γ − Λ can be computed from properties of the gas and the radiation field.

9
This leaves us with 5 equations and 14 unknowns: ρ, ~u, P , π (five independent elements), ǫ, F~ .
↔

To get a closed set of equations, we must find relations among ρ, P , ǫ, π , and F~ using constitutive
↔

relations for the gas.

We have already discussed the simplest example, the equation of state for a monatomic gas, ǫ = 32 Pρ .
We will return to a more general discussion of the equation of state and to viscosity and heat
conduction soon.

2.7 The Euler equations

Formally, the Navier-Stokes equations can be derived from the first-order expansion of the Boltz-
mann equation in the parameter λ/L.
In many cases, it is adequate to use the zero’th-order approximation λ/L = 0, in which case one
gets the eight “constraint” equations π = 0, F~ = 0.
↔

Viscosity and heat conduction are diffusive effects (arising because particles diffuse out of their fluid
elements), and these are often small compared to dynamical effects, when flows are a significant
fraction of (or greater than) the sound speed, or when diffusion times are long compared to system
lifetimes.
Some important exceptions in astrophysical situations are shocks, viscosity in accretion disks, con-
ductive evaporation of cold clouds in a hot medium, and conduction in some kinds of stars and
planets.

If we neglect diffusive terms in the Navier-Stokes equations, we get the Euler equations
Dρ ∂ρ ~ ~ · ~u,
= ∂t + ~u · ∇ρ = −ρ∇ (20)
Dt
D~u ∂~
u

~ ~u = ~g − 1 ∇P

~
= ∂t + ~u · ∇ (21)
Dt ρ
Dǫ ∂ǫ ~ P~ Γ−Λ
= ∂t + ~
u · ∇ǫ =− ∇ · ~u + (22)
Dt ρ ρ

Together with the equation of state ǫ = 32 Pρ , the Euler equations describe the dynamics of a perfect
monatomic gas.
In a perfect gas, collisions ensure that the local distribution of random velocities acquires the
maximum entropy form, in which the velocity distribution in each dimension is independently
Gaussian with variance hwi2 i = kT
m.
This is a Maxwellian velocity distribution,
3/2 !
m mw2

3
fMax (w)d
~ w= exp − d3 w. (23)
2πkT 2kT

The Maxwellian distribution has hwi wj i = 0 for i 6= j and hw2 wi i = 0, implying π = 0 and F~ = 0.
↔

10
2.8 The Jeans Equations

It is worth briefly noting the analogy between the Euler equations of fluid dynamics and the Jeans
equations of stellar dynamics. These are also derived by taking moments of the Boltzmann equation,
in this case the collisionless Boltzmann equation (Binney & Tremaine, section 4.2.)

The first Jeans equation is just the continuity equation for stellar number density, exactly equivalent
to the fluid continuity equation for mass density.

The second Jeans equation, in Binney & Tremaine’s notation, is

2)
∂v j X ∂v j ∂Φ X ∂(νσij
ν + νv i = −ν − , (24)
∂t i
∂xi ∂xj i
∂xi

where v is the mean fluid velocity, ν is the stellar number density, Φ is the gravitational potential,
2 = hw w i where w
and σij ~ is the random velocity.
i j
2 in place
This is very similar to the second Euler equation, but with an anisotropic stress tensor νσij
~ .
of ∇P

The limitation of the Jeans equations is that there is no equation of state relation between den-
sity and σij2 , no equivalent of the thermal energy equation, and no strong reason for assuming a

Maxwellian random velocity distribution.

Thus, stellar dynamical models that use the Jeans equations must usually assume some form of
2 , and the resulting models are only as accurate as the assumptions.
σij

11