Arch. Rational Mech. Anal.

192 (2009) 217–274

Digital Object Identifier (DOI) 10.1007/s00205-008-0139-z

Stochastic Formulation
for the Initial-Boundary Value Problems
of the Boltzmann Equation
Shih-Hsien Yu

Communicated by T.-P. Liu

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
2. Preliminaries on collision operators and stochastic processes
generated by boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
2.1. Properties of collision operators . . . . . . . . . . . . . . . . . . . . . . . . 226
2.2. Diffuse boundary condition and conservation law . . . . . . . . . . . . . . . 229
2.3. A probability interpretation of free transport equation
with diffuse boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . 231
2.4. Representation of a free transport equation from PDE (an alternative approach) 236
3. The free transport equation: maximal principles . . . . . . . . . . . . . . . . . . . 238
3.1. Maximal principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
3.2. L 1 in space to L ∞ at boundary . . . . . . . . . . . . . . . . . . . . . . . . . 240
4. The free transport equation: probability fluctuation
and rate of convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
4.1. Probability estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
4.2. Fluctuation in the boundary flux function ρ± (t) . . . . . . . . . . . . . . . . 247
4.3. From the fluctuation in the boundary flux function to fluctuation in solution . 248
4.4. Convergence of the full free transport equation . . . . . . . . . . . . . . . . . 249
5. The linearized Boltzmann equation, (Maxwell molecule collision model) . . . . . 250
5.1. Rate of convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
5.2. Convergence to a stationary solution . . . . . . . . . . . . . . . . . . . . . . 254
5.3. Existence of non-trivial nonlinear stationary solutions and its stability . . . . 256
6. Hard potential with angular cutoff collision models . . . . . . . . . . . . . . . . . 260
6.1. Transport equation with damping . . . . . . . . . . . . . . . . . . . . . . . . 260
6.2. On the exponential converging property of the solution of a linearized Boltzmann
equation for angular cutoff hard potential collision models . . . . . . . . . . 265
6.3. On linear and nonlinear stability . . . . . . . . . . . . . . . . . . . . . . . . 268

1. Introduction

The Boltzmann equation is a basic equation for both rarefied gas theory and
the study of thermal non-equilibrium flow. It allows for the physical formulation
218 Shih-Hsien Yu

of boundary conditions that model the effects of the boundary thermal information
on the gas flows. For instance, the basic diffuse reflection boundary condition was
introduced by Maxwell in his study of the Crookes radiometer on the flow induced
by a boundary thermal gradient, [29].
Various basic boundary effects have been investigated extensively using asymp-
totic expansions, numerical computations, and through physical experiments by
Sone, Aoki, etc., [34]. They discovered many interesting phenomena, such as the
ghost effects, that can be understood only through kinetic theory. These works
motivate the present paper.
Understanding the boundary phenomena requires quantitative description of
the Boltzmann solutions. For this, we introduce a new stochastic process approach
for treating the diffuse boundary condition.
Consider the Boltzmann equation in a domain Ω ⊂ R3 . The diffuse boundary
condition is a prescription for the particle velocity distribution function F(x, t, ξ )
given as follows:
⎧ +
⎪
⎪ ∂t F + ξ · ∇x F = Q(F), x ∈  Ω, ξ√ ∈ R , t ∈ R , 
3
⎪
⎨ 
F(x, t, ξ )|ξ ·n(x)<0,x∈∂Ω = ρ(x,t)
√
θ(x)
2π
M[1,0,θ(x)] (ξ )  ,
⎪  x∈∂Ω,ξ ·n(x)<0
⎪
⎪
⎩ρ(x, t) = F(x, t, ξ )|ξ · n(x)|dξ,
ξ ·n(x)>0,ξ ∈R3

where M[1,0,θ] is a local Maxwellian state defined in terms of the given boundary
temperature θ = θ (x):
|ξ −u|2
ρe− 2θ
M[ρ,u,θ] ≡ .
(2π θ )3/2
In this paper we assume that the collision operator is either (i) a hard sphere collision
model, (ii) an angular cutoff hard potential collision model or (iii) an angular
cutoff Maxwell molecule model, [21]. Here, n is the outward unit normal vector
of the domain Ω, and θ (x) = θw (x), x ∈ ∂Ω, is the temperature imposed on the
boundary ∂Ω as illustrated in Fig. 1.
Mathematical studies on the Boltzmann equation with diffuse boundary con-
ditions have been investigated based on the Darrozes–Guiraud inequality [13] and
compensated-compactness approaches, see [3–7,13,24] and references therein.
Sone [36], and [15,18,30,32] investigate the physical mechanisms underlying
thermal transpiration flow. Furthermore, there is a summary article about the Sone’s
work in [2].
The diffuse boundary condition allows for explicit stochastic formulation and
thereby the explicit pointwise estimates of the solutions. There are works on the
boundary effects for other types of boundary conditions. In [17], specular reflection
and periodic boundary problems have been studied and global existence results have
been obtained.
Our starting point is the introduction of a stochastic formulation and its analysis
for the free transport equation
∂t F + ξ · ∇x F = 0.
 Stochastic Formulation for the Initial-Boundary Value Problems 219

Ω
ρ ( x) 2π
F ( x, t , ξ ) |ξ ⋅n ( x )<0, x∈∂Ω = M [1, 0,θ ( x )] (ξ ) |ξ ⋅n ( x )<0, x∈∂Ω
θ (x) x∈∂Ω

ρ ( x) = ∫ F ( x , t , ξ ) ξ ⋅ n ( x ) dξ
ξ ⋅n ( x ) > 0
ξ ∈R 3

θ ( x ) : Temperature at boundary
ξ ⋅n ( x ) < 0
x∈∂Ω
n : Outwards unit normal vector

Fig. 1

The free transport equation has been studied for other kinetic models, there are many
interesting analyses and applications of the free transport equation with specular
reflection boundary conditions to various problems in mathematical physics. In a
sequel of works [8,19,20] the Lorentz gas theory can be derived as Boltzmann–
Grad hydrodynamic limits. In [11,12] microscopic friction, which is related to the
microscopic Ohm’s law, has been investigated. An algebraic convergence rate due
to particles re-collisions is discovered. The slow rate is explained as no loss of
memory after specular reflection collision.
The purpose of this paper is to study the interaction between the diffuse bound-
ary and the collision operator which work together to enhance the rate at which
rarefied gas equilibrates. In a very rarefied gas environment, binary particle colli-
sions become less important for particles with non-small velocity. Such particles
likely collide with the boundary before scattering with other particles. The effect
from the diffuse boundary is more important for such particles. In contrast, for slow
particles, the binary collisions between particles are more important than the colli-
sions with the physical boundary. This is because such slow particles will stay in the
domain longer before colliding with the physical boundary. The chance of being hit
by the other particles becomes bigger in this case. Heuristically, one can see that the
equilibrating mechanisms due to particle-particle collisions and particle-boundary
collisions complement each other.
The main goal of the present paper is to introduce a new probabilistic approach
for the study of this enhanced equilibrating mechanism. To explain our main ideas,
we formulate the problem and carry out our analysis for the simplest situation of
the planar wave solution of the Boltzmann equation in a slab,

∂t F + ξ 1 ∂x F − Q(F) = 0, x ∈ [−, ], ξ ∈ R3 , (1)

with the diffuse boundary conditions at x = ±

220 Shih-Hsien Yu

⎧
⎪
⎪ 2π
⎪
⎪ F(±, t, ξ )|±ξ 1 <0 = ρ± (t) M± (ξ )|±ξ 1 <0
⎪
⎪ θ±
⎪
⎪ 
⎨
ρ± (t) ≡ |ξ 1 |F(±, t, ξ )dξ, (BC)
⎪
⎪ ±ξ 1 >0,ξ ∈R3
⎪
⎪ |ξ |2
⎪
⎪ − 2θ
⎪
⎪ e ±
⎩M± (ξ ) ≡ ,
(2π θ± )3/2

and with initial data

F(x, 0, ξ ) = F0 (x, ξ ).

The functions ρ± (t) are called the boundary flux functions, which play a key role
in our formulation of the problem.
The diffuse boundary conditions are posed so that the conservation law holds:
  
d
F(x, t, v) dv dx = 0.
dt − R

The physical condition F(x, t, v)
0 holds if it holds for the initial data. Our starting
point is to view the diffuse boundary conditions as modeling the boundary random-
ness according to the thermal conditions. This physical interpretation allows us to
make use of the conservation law and the non-negativity to offer a probabilistic
point view for the free transport equation

∂t F + ξ 1 ∂x F = 0

with the diffuse boundary conditions. Without loss of generality one can assume

− R F(x, 0, ξ ) dξ dx = 1, then the measure F(x, t, ξ ) dx dξ can be inter-
1 1 1 1

preted as the probability measure that a random particle x = ω(t), ẋ = ξ 1 presents

at (x, x + dx)×(ξ 1 , ξ 1 + dξ 1 ) at time t. The random particle x = ω(t) is determin-
istic in the interior domain and disassociates randomly after every collision with
the boundary. This gives a sequence of independent random variables {σi }i∈N∪{0} .
The random variable σi stands for the time lapse of the random particle between
the ith collision and i + 1th collision. One can relate boundary flux functions ρ± (t)
in terms of the following stochastic processes:
⎧ ∞
⎪
⎪
⎪
⎪ ρ (t) dt = P (Wn ∈ (t, t + dt)|ω(Wn ) = ±) ,
⎨ ±
n=1
n
⎪
⎪
⎪
⎩ Wn ≡
⎪ σj.
j=0

Thus Wn denotes the random variable for the time lapse of an n-consecutive col-
lisions. With this probabilistic interpretation of ρ± (t), one can develop integral
 Stochastic Formulation for the Initial-Boundary Value Problems 221

equations for ρ± (t) as follows, Section 2,

⎧
⎪
⎪ ρ+ (t) = H− (t) ∗ H+ (t) ∗ ρ+ (t)
⎪
⎪  0 
⎪
⎪
2/t
⎪
⎨ + H− (t) ∗ −vF(− − vt, 0, v) dv + vF( − vt, 0, v) dv ,
−2/t 0
(2)
⎪
⎪ ρ− (t) = H+ (t) ∗ H− (t) ∗ ρ− (t)
⎪
⎪  2/t 
⎪
⎪ 0
⎪
⎩ + H+ (t) ∗ vF( − vt, 0, v) dv + −vF(− − vt, 0, v) dv ,
0 −2/t

where H− (t) and H+ (t) are the conditional probability density functions of the
random σn with probability laws

P (σn + ω(Wn−1 ) ∈ (t, t + dt)|ω(Wn−1 ) = −) = H− (t) dt,
P (σn + ω(Wn−1 ) ∈ (t, t + dt)|ω(Wn−1 ) = ) = H+ (t) dt, n = 1, 2, . . . ,
where
⎧
⎪ 22
⎨ 42 − t 2 θ±2
e for t
0,
H± (t) ≡ t 3 θ±
⎪
⎩0 else,
 t
U (t) ∗ V (t) ≡ U (t − r )V (r ) dr.
0
The slab length is assumed to be small
 1,
and the temperature variation θ− − θ+ is assumed small as well,
|θ− − θ+ | 1.
without loss of generality, we fix
θ− = 1.
When the slab length  is sufficiently small, particles with non-small velocity will
likely travel through the slab without colliding with any particle. In such a situation,
the free transport equation with (BC)
⎧
⎪
⎪ ∂t t(x, t, ξ ) + ξ 1 ∂x t(x, t, ξ ) = 0, (x, t) ∈ (−, ) × R+ ,
⎪
⎪
⎪
⎪ ξ = (ξ 1 , ξ 2 , ξ 3 ) ∈ R3 ,
⎪
⎨
⎪  0, ξ ) = t0 (x, ξ ),
t(x, (3)
⎪
⎪  |t (x, ξ )| 
⎪
⎪  0 
⎪
⎪    1,
⎩ M− (ξ )  L ∞ (L ∞ )
x ξ

is an accurate approximation for the study of the collisions between particles and
the diffuse boundary. Here, the norm · L ∞ ∞ is given by
x (L ξ )

g L∞ ∞
x (L ξ )
≡ sup |g(x, ξ )|.
x∈(−,),ξ ∈R3
222 Shih-Hsien Yu

In the free transport equation the velocity distribution described on the boundary
is three dimensional. One can integrate out the two velocity variables tangential to
the boundary to obtain a free transport equation defined in one-dimensional velocity
v, T(x, t, v) = R2 t(x, t, v, ξ 2 , ξ 3 ) dξ 2 dξ 3 ,
⎧
⎪
⎪ ∂t T(x, t, v) + v∂x T(x, t, v) = 0, (x, t, v) ∈ (−, ) × R+ × R,
⎪
⎪
⎨
ρ± (t) ≡ |v|T(±, t, v) dv,
±v>0 √ (4)
⎪
⎪
⎪
⎪ ρ ± (t) 2π
⎩T(±, t, v) ≡ √ G± (v),
θ±
where
v2
−
e 2θ±
G± (v) ≡ √ .
2π θ±
The probabilistic interpretation of (2) for ρ± (t) is physically natural for the
nonnegative function F(x, v, t). It can also be applied to obtain the ρ± (t) for the
perturbation T(x, t, v), for which T(x, 0, v)
0 does not hold in general. This pos-
itivity condition can be removed, and still keeping the probabilistic interpretation,
since any function T(x, 0, v) can be expressed as a difference of two positive-valued
functions. Of course, once (2) is derived, either by probabilistic or PDE methods,
it holds for any function F(x, v, t) by linearity.
With the integral equations (2) one can make use of maximal principles to
study the global upper bounds of the boundary flux functions ρ± (t). We will need
a combinatorics approach and some translation properties of the free transport
equation to analyze the contribution from slow velocity particles to the boundary
flux functions globally in time. This is the main interest in Section 3.
Furthermore, by the integral equations, maximal principles, the law of large
number theory on W2n and the central limit theorem on W2n , one can estimate the
fluctuation in the boundary flux functions ρ± (t):
 
1/10
 t 1/10
|ρ± (t + ∆t) − ρ± (t)|  O(1) for ∆t ∈ 0, .
+t 

We will see that, due to E[|σi |2 ] = ∞, the law of large number theory and the
central limit theorem are somewhat non-standard. We derive them here in order to
obtain refined estimates for the study of fluctuation.
By combining this fluctuation estimate for the boundary flux functions ρ± (t)
and the conservation law for T(x, t, v), we obtain the following theorem in Sec-
tion 4.

Theorem 1. (Main theorem for free transport equation, I) Suppose that the initial
data T(x, 0, v) of (4) satisfies
 
 T(x, 0, v) − S(x, v) 
 
    1.
 G± (v)  ∞ ∞
L x (L v )
 Stochastic Formulation for the Initial-Boundary Value Problems 223

Then,
⎧  1/10
⎪
⎪ 
G− (v) for v > x+
⎪
⎪ +t t
⎨
|T(x, t, v) − ρ0 S(x, v)|  O(1)  1/10

⎪
⎪ G+ (v) for v < t ,
x−
⎪
⎪  + t  + t
⎩
1 else,

where S(x, v) is a stationary solution of (4),

⎧
⎪
⎨ 2π θ− G− (v) for v > 0,
S(x, v) = 
⎪
⎩ 2π G+ (v)
θ+ for v < 0,

and ρ0 is set so that the conservation of mass is satisfied:

 
T(x, 0, v) − ρ0 S(x, v) dv dx = 0.
− R

After obtaining the boundary flux functions ρ± (t) from the above theorem, the
theorem below follows immediately. First denote the solution t(x, t, ξ ) of (3) in a
form to relate its dependence on the initial data t0 (x, ξ ):

Tt0 [t0 ](x, ξ ) = t(x, t, ξ ).

Theorem 2. (Main theorem for free transport equation, II) Suppose that the initial
data t0 (x, v) of (3) satisfies
 
 t (x, ξ ) − s(x, ξ ) 
0 
    1,
 M− (ξ )  ∞ ∞
L (L )
x ξ

Then,
⎧ 
⎪
⎪ 
1/10
⎪ +t
⎨ M− (ξ ) for ξ 1 > t ,
x+
 1/10  
|T0 [t0 ](x, ξ ) − ρ0 s(x, ξ )|  O(1)
t

(ξ ) 
for ξ 1 < t ,
x−
⎪
⎪ +t M+ +t
⎪ 
⎩ M (ξ )
− else,
(5)
|ρ± (t) − ρ0 |  O(1)1/(( + t)/) 1/10
, (6)

|Tt0 [t0 ](x, ξ ) − ρ0 s(x, ξ )|  O(1) for |ξ 1 |  /(( + t)/)1/10 , (7)

where s(x, ξ ) is the stationary solution of (3),

⎧
⎪
⎨ 2π θ− M− (ξ ) for ξ > 0,
1
s(x, ξ ) = 
⎪
⎩ 2π M+ (ξ ) for ξ 1 < 0,
θ+
224 Shih-Hsien Yu

and the constant ρ0 is determined by the conservation law:

 
t0 (x, ξ ) dξ dx
− R3
ρ0 =    .
s(x, ξ ) dξ dx
− R
Theorem 2 asserts how fast the diffuse boundary equilibrates the particle √ dis-
tribution towards its stationary distribution. In a small time interval (0, ), the
solution T(x, t, v) is within a close distance of the order √ 
1/20 to the stationary

solution for a particle with non-small velocity |ξ |
. In contrast to this fast

1

equilibrating mechanism, the global decay toward the stationary distribution is of

algebraic rate. This is due to the fact that the diffuse boundary has less effect on
particles with slow velocity. For particles with slow velocity, the collision operator
Q is an important equilibrating mechanism on a time scale of O(1). One of the
main tasks of our analysis is to find an effective way of combining these two mech-
anisms together to obtain a strong overall equilibration. With this total equilibrating
mechanism, the methodology of Grad [23] in constructing the local time solution of
the Boltzmann equation can be refined for the construction of the global solution.
This approach to extend the local solution to the global solution follows the same
strategy as in [37] for the initial value problem.
The strategy to understand how the two mechanisms complement each other is
to replace the nonlinear collision operator Q by a linearized collision operator, L,
around a global Maxwellian state M− first. The linearized collision operator L can
be decomposed into the form L = −ν + K, [13,14,21,25], where
⎧
⎪
⎨ν(ξ )h(ξ ) = O(1)(1 + |ξ |)γ h(ξ ),
 h(η)
⎪
⎩ Kh(ξ ) = M− (ξ ) K (ξ, η)  dξ,
R 3 M0 (η)

and γ = s−4 −s of the collision

s depend on the inverse power law potential r
operator Q, see Section 2. For the Maxwell molecule collision model s = 4,
for the hard potential model s ∈ (4, ∞), and for the hard sphere model s = ∞.
Let Ttν denote the solution operator of the linear transport equation with damping
term ν(ξ ):

∂t tν + ξ 1 ∂x tν + νtν = 0,
tν (x, 0, ξ ) ≡ t0 (x, ξ )
and with the diffuse boundary condition (BC)
Ttν [t0 ](x, ξ ) = tν (x, t, ξ ).
When both the time variable t and width of the slab  are chosen small enough,
one can effectively compare the operator Ttν with Tt0 ; use Theorem 2 and Grad’s
lemmas in the next section to get the local existence theory of

∂t f + ξ 1 ∂x f − Lf = 0,
(8)
f(x, 0, ξ ) = f0 (x, ξ ).
 Stochastic Formulation for the Initial-Boundary Value Problems 225


For this local theory, we need only the conditions (BC), f0 / M− L ∞ ∞  1,
x (L ξ )
and the choice of the Maxwellian M− to have the same mass as the initial data:
 
f0 (x, ξ ) dξ dx = 0.
− R3

The key to extending the local existence theory to the global exponential converg-
ing theory is based on the uneven velocity distribution in (5) when ρ0 = 0, (this
condition is a consequence of (1, Q) = 0, see Theorem 5). The uneven velocity
distribution asserts that either the value of the velocity distribution function is small
or |ξ 1 | is small. So, when K acts on such an uneven velocity distribution function,
the resulting function is small. A bootstrap argument can be used to conclude the
global exponential converging property of f(x, t, ξ ). We also denote the solution
operator of (8) by St ,

St [f0 ](x, ξ ) = f(x, t, ξ ).

We have carried out this strategy to yield the following theorem for the case of the
angular cutoff hard potential model in Section. 6.

Theorem 3. (Main theorem on linear Boltzmann equation) There exists SL (x, ξ )

satisfying (BC) and
⎧ 1
⎨ξ ∂x SL − LSL = 0,
 
⎩ SL (x, ξ ) − s(x, ξ ) dξ dx = 0.
− R3

For any f0 satisfying f0 / M− L∞ ∞
x (L ξ )
 1 it follows

|St [f0 − ρ0 SL ](x, ξ )|

⎧ 1/60
 ⎨ for |ξ 1 | > 1/2 , t > 3/7 /| log |
 O(1)e−ν0 t/2 M− (ξ ) 1 for |ξ 1 | < 1/2
⎩
1 for t  3/7 /| log |,

where ν0 = ν(0) and where ρ0 is given by

 
f0 (x, ξ ) − ρ0 s(x, ξ ) dξ dx = 0. (9)
− R3

After obtaining the linear exponential converging property of St in the above

theorem, the nonlinear theory is a direct application of Grad’s methodology. We
obtain the following theorem for the case of the angular cutoff hard potential model
in Section 6.3.

Theorem 4. (Main theorem on Boltzmann equation) Suppose that both the tem-
perature difference |θ+ − θ− | and the slab length  in (1) are sufficiently small.
Then, there exist non trivial stationary solutions φ of (1). The stationary solution
is nonlinearly stable in the sense that
226 Shih-Hsien Yu

if ⎧ 
⎪ 
⎪
⎪
⎨− R3 (φ(x, ξ )) − F(x,
⎪

0, ξ )) dξ dx = 0,
 F(x, 0, ξ ) − φ(x, ξ )  (10)
⎪
⎪ 

 ε
⎪
⎪  ∞ ∞
1,
⎩ M− (ξ ) L (L )
x ξ

then
|F(x, t, ξ ) − φ(x, ξ )|
⎧ 1/60
 ⎨ for |ξ 1 | > 1/2 , t > 3/7 /| log |
 O(ε)e−ν0 t/8 M− (ξ ) 1 for |ξ 1 | < 1/2
⎩
1 for t  3/7 /| log |.
The condition (10) is a generic condition. For any constant c ∈ R the profile cφ
is still a stationary solution of (1). Thus, for any given initial data F(x, 0, ξ ) there
exists a unique c0 such that
 
(c0 φ(x, ξ )) − F(x, 0, ξ )) dξ dx = 0.
− R3

Remark 1. Theorem 4 in particular establishes the existence of the stationary solu-

tion φ(x, ξ ) for the classical rarefied gas dynamics problem of heat transfer between
two plates. Previous works, Aoki (83) and Cercignani (60’s) [1] are for the linear
BGK models.

For the case of the Maxwell molecule model, the proofs of both Theorems 3
and 4 are given in Section 5 right after the completion of Theorem 2. For this
collision model the collision operator is a bounded operator. It can be treated as
perturbation of the free transport equation without too much complication. And it
also serves as a model problem for the development in Section 6.
One of the basic features of the present paper is the pointwise understanding of
the Boltzmann solutions, with emphasis on the combined boundary and collision
effects for the rarefied gases. The emphasis differs markedly from the pointwise
estimates on Boltzmann solutions near global Maxwellian state, see [26–28].

2. Preliminaries on collision operators and stochastic processes

generated by boundaries

2.1. Properties of collision operators

The collision operator Q with Grad’s angular cutoff potential with an inverse
power law r −s is given in the form:
⎧
⎪
⎪ Q( f ) ≡ B( f, f ),

⎨ 1
B(g, h)(ξ ) ≡ q(V, θ )(f(ξ )g(ξ∗ ) + g(ξ )h(ξ∗ )
⎪
⎪ 2 R3 ×S 2
⎩
−f(ξ )g(ξ∗ ) − g(ξ )h(ξ∗ )) dξ∗ dω,
 Stochastic Formulation for the Initial-Boundary Value Problems 227

where

ξ = ξ − ((ξ − ξ∗ ) · ω)ω,
ξ∗ = ξ∗ + ((ξ − ξ∗ ) · ω)ω,
(ξ − ξ∗ )
V = |ξ − ξ∗ |, θ = cos−1 ·ω .
|ξ − ξ∗ |

Here, the condition for Grad’s angular cutoff potential for an inverse power law
imposes the following on q(V, θ ):

q0 V γ | cos θ | < q(V, θ ) < q1 V γ | cos θ | for some 0 < q0 < q1 ,

(s − 4)
γ ≡ .
s
Those collision operators are classified as follows:

s<4 s=4 s>4 s=∞

Collision model: Q soft potential Maxwell molecule hard potential hard sphere

Maxwellians and collision invariants

Theorem 5. (Boltzmann 1872) If Q(M) ≡ 0, then the state M is called a Maxwellian

and it is in the form
|ξ −u|2
e− 2θ
M(ξ ) = M[ρ,u,θ] ≡ ρ .
(2π θ )3

The collision invariants χi , i = 0, 1, 2, 3, 4, are the five polynomials

⎧
⎪
⎨χ0 (ξ ) = 1,
χi (ξ ) = ξ i for i = 1, 2, 3,
⎪
⎩
χ4 (ξ ) = |ξ |2 .

They are orthogonal to the collision operator Q:


χi (ξ )Q(h) dξ = 0 for i = 0, . . . , 4, for any h under consideration.
R3

Let M− denote the absolute Maxwellian M[1,0,1] ,

|ξ |2
e− 2
M− (ξ ) ≡ ,
(2π )3/2

and L means the linearized collision operator of Q at the absolute Maxwellian state
M− ,

Lh = 2B(M− , h).
228 Shih-Hsien Yu

From the definition of Q , the operator Lh(ξ ) can be rewritten as follows


1
Lh(ξ ) ≡ q(v, θ )(M− (ξ )h(ξ∗ ) + h(ξ )M− (ξ∗ ) − h0 (ξ )M− (ξ∗ )
2 R3 ×S 2
−M− (ξ )h(ξ∗ )) dξ∗ dω
 
1
= −ν(ξ )h(ξ ) M− (ξ ) + q(v, θ )(M− (ξ )h(ξ∗ ) + h(ξ )M− (ξ∗ )
2 R3 ×S 2
−M− (ξ )h(ξ∗ )) dξ∗ dω, (11)
where
⎧
⎪
⎪ (ξ − ξ∗ ) · ω
⎨v = |ξ − ξ∗ |, θ = cos−1 ,
 v (12)
⎪
⎪
1
⎩ν(ξ ) ≡ q(v, θ )M− (ξ∗ ) dξ∗ dω.
2 R3 ×S 2
From this, the linearized collision operator L can be abbreviated in an operator
form:
Lh(ξ ) = ν(ξ )h(ξ ) + Kh(ξ ) (13)
where
 
h(η)
Kh(ξ ) = M− (ξ ) K (ξ, η)  dη.
R3 M− (η)
For a hard sphere collision model, the functions ν(ξ ) and K (ξ, η) were first com-
puted explicitly by Hilbert compare[13,25]:
2
 |ξ |
1 − |ξ2| 1 u2
ν(ξ ) = √ 2e + 2 |ξ | + e− 2 du ,
2π |ξ | 0
2 (|ξ | − |η|2 )2
2 |ξ − η|2
K (ξ, η) = √ exp − −
2π |ξ − η| 8|ξ − η|2 8
|ξ − η| (|ξ |2 + |η|2 )
− exp − .
2 4
For an angular cutoff potential model, the function K (ξ, η) was given in the exact
form by Grad [21]:
K (ξ, η) = −K 1 (ξ, η) + K 2 (ξ,
 η),

K 1 (ξ, η) ≡ M− (ξ )M− (η) q(V, θ ) dω,
S2

|ξ −η|2 |ξ −η|2 (|ξ |2 −|η|2 )2 

2 e− 8 − 8 |w+ζ |2
K 2 (ξ, η) ≡ e− 2 n(η−ξ, w) dw, (14)
(2π )3/2 |ξ −η|2
w∈R3
 w·(η−ξ )=0
ζ = 21 (η + ξ ) − 21 (|η|2 − |ξ |2 )(η − ξ ),
|n(ξ − η, w)|  C|η − ξ |(|η − ξ |2 + |w|2 )−(1−γ )/2 .
 Stochastic Formulation for the Initial-Boundary Value Problems 229

Remark 2. Many refined estimates on K (ξ, η) are also available in [9,10].

Lemma 1. (Grad) [22] For a collision operator Q with an angular cutoff potential
with an inverse power law r −s , there exists a constant C > 0 such that

0 < ν(ξ ) = O(1 + |ξ |)γ , (15)

     
 ν(ξ )−1 B(f, g)   f   g 
     
  C 1   1 
, (16)
 M−
1
  M− L ∞ M−  L ∞
 
L∞
2 2 2
ξ ξ ξ

where
s−4
γ ≡ .
s
Lemma 2. (Grad) [22] The kernel function K (ξ, η) of K given in (14) is symmetric
in (ξ, η) and satisfies

O(1)
|K (ξ, η)|dη  , (17)
(1 + |ξ |)
 R
3

|K (ξ, η)|2 dη  O(1). (18)

R3

Lemma 3. The same integral kernel function K (ξ, η) in Lemma 2 satisfies that when
ε 1

sup |K (ξ, η)| dη < O(1)ε.
ξ ∈R |η1 |<ε

This lemma is a direct consequence of (14). The proof is omitted.

2.2. Diffuse boundary condition and conservation law

We consider a planar wave solution of a free transport equation with diffuse

boundary conditions in a slab:

∂t t + ξ 1 ∂x t = 0, x ∈ [−, ], t > 0, ξ ∈ R3 ,

t(x, 0, ξ ) = t0 (x, ξ ),
⎧
⎪
⎪ 2π
⎪
⎪ t(±, t, ξ )|±ξ 1 <0 = ρ± (t) M± (ξ )|±ξ 1 <0
⎪
⎪ θ±
⎪
⎪ 
⎨ (19)
ρ± (t) ≡ |ξ 1 |t(±, t, ξ ) dξ,
⎪
⎪ ±ξ 1 >0,ξ ∈R3
⎪
⎪ 2
⎪
⎪ − |ξ |
⎪
⎪ e 2θ±
⎩M± (ξ ) ≡ ,
(2π θ± )3/2
where θ± are the accommodated temperatures at x = ± and t0 (x, ξ ) is the imposed
initial data. The major objective in this free transport equation is to analyze the
boundary flux functions ρ± (t) which are important in determining the solution of
t(x, t, ξ ) by the characteristic curve method.
230 Shih-Hsien Yu

We can reduce the independent variable ξ ∈ R3 in t(x, t, ξ ) into a one-

dimensional variable and still retain the original boundary flux functions ρ± (t)
as follows. Integrate t(x, t, v, ξ 2 , ξ 3 ) over (ξ 2 , ξ 3 ) ∈ R2 and denote the resulting
function as T(x, t, v):

T(x, t, v) ≡ t(x, t, v, ξ 2 , ξ 3 ) dξ 2 dξ 3 .
(ξ 2 ,ξ 3 )∈R2

The function T(x, t, v) satisfies

∂t T(x, t, v) + v∂x T(x, t, v) = 0 for x ∈ [−, ], t > 0, v ∈ R, (20)
with the diffuse boundary conditions at |x| = ,
⎧
⎪
⎪ 2π
⎪
⎪T(±, t, v)|±v<0 = ρ± (t) G± (v)|±v<0 ,
⎪
⎪ θ±
⎪
⎪ 
⎨
ρ± (t) ≡ |v|T(±, v, t) dv, (21)
⎪
⎪ ±v>0
⎪
⎪
⎪
⎪ 1 − 2θv2
⎪
⎪ G (v) ≡ e ±,
⎩ ±
2π θ±

and the initial data T(x, 0, v) = T0 (x, v) is


T0 (x, v) = t0 (x, v, ξ 2 , ξ 3 ) dξ 2 dξ 3 .
R2
Note that the boundary flux functions ρ± (t) for (20) and (19) are identical:
 
|v|t(±, t, v, ξ 2 , ξ 3 ) dξ 2 dξ 3 dv = ρ± (t) = |v|T(±, t, v) dv.
±v>0,(ξ 2 ,ξ 3 )∈R2 ±v>0
(22)
Lemma 4. (Conservation law) The solution of (20), T(x, v, t), satisfies the conser-
vation law
 
d 
T(x, t, v) dv dx = 0. (23)
dt − R
Proof. The boundary condition (21) implies
 

vT(x, t, v) dv  = 0.
R x=±

This yields that

 
d 
T(x, t, v) dv dx
dt − R
   x=

= −v∂x T(x, t, v) dv dx = vT(x, t, v) dv  = 0.
− R R x=−
 Stochastic Formulation for the Initial-Boundary Value Problems 231

2.3. A probability interpretation of free transport equation

with diffuse boundary conditions

The problem (20) can be regarded as the evolution of a probability density

function at (x, t, v) by normalizing the initial data T0 to satisfy,
  
T0 (x, v) dv dx = 1,
− R

and is non-negative. The evolution is due to following two mechanisms:

1. Particle transport in the free space. (Deterministic.)
2. Collision with the diffuse boundary. (Random.)
From these two mechanisms we can devise a stochastic process by considering
a random particle with trajectory x = x(t) with initial configuration (x(0), ẋ(0)) =
(x0 , v0 ). This particle travels freely in the interval (−, ). This is the deterministic
part. When it collides with the boundary |x| = , the velocity renews randomly
after collision according to the temperature accommodated. One can record the
history of velocities {vn }n∈N∪{0} , which is a sequence of velocities of this random
particle bouncing backwards and forwards in the slab. Since each collision will not
correlate to the previous history, the sequence of velocities can be regarded as a
sequence of independent random variables. From the sequence of velocities, one
also can recover {σn }n∈N∪{0} a sequence of time lapses between two collisions
We denote σ0 (x0 , v0 ) the boundary hitting time for each (x0 , v0 ) ∈ (−, ) ×
(R\{0}), and {vn (x0 , v0 )}n∈N is the sequence of random velocities. From Fig. 2,
⎧
⎨ −x
v for v0 > 0,
0
σ0 (x0 , v0 ) =
⎩ −−x for v0 < 0.
v0

t
σn x ' = vn

σ2 x' = v2
σ1 x' =
v1

σ0 v0
x' = x
− x0
Fig. 2
232 Shih-Hsien Yu

The probability density function of each random variable vn (x0 , v0 ) is given by

{vn (x0 , v0 )}n∈N :
⎧
⎨ P{vn (x0 , v0 ) ∈ [v, v + dv]} = H 1+(−1)n+1 (v) dv for v0 > 0,
2 (24)
⎩ P {(vn (x0 , v0 ) ∈ [v, v + dv]} = H 1+(−1)n (v) dv for v0 < 0.
2

Here,
⎧  v

⎪
⎪ v − v2 1 − |v|
⎪
⎨H1 (v) = − θ e
⎪ ,
2θ+
+ 2
 v
 (25)
⎪
⎪
⎪ v − 2θv2 1 + |v|
⎩H0 (v) = θ e
⎪ .
−
− 2

The time lapse of the random particle is given by LENGTH/VELOCITY. Thus, the
sequence of the time lapse random variable is

2l
{σn (x0 , v0 )}n∈N∪{0} : σn ≡ . (26)
vn

Remark 3. The diffuse boundary condition disassociates the boundary flux function
into a certain law of probability. In terms of the density function of particles the prob-
ability law shall be in a Gaussian form. However, the random process is in regards to
the flux function, whose physical quantity is momentum. Thus, the law shall pertain
to the momentum density, and it shall be in the form GAUSSIAN*VELOCITY.
This gives rise to the distribution laws in (24).

Remark 4. The variance of the random variable σn , n
1, is unbounded

⎧  
∞ 2
v − 2θv2 ∞
⎪
⎪ 2 42 − 2θv2
⎪
⎨Var(σ2n ) = e − dv = e − dv = ∞,
0 v θ− 0 vθ−
⎪  2  (27)
⎪
⎪
0 2 −v − 2θv2 0 42 − 2θv2
⎩Var(σ2n+1 ) = e + dv = e + dv = ∞.
−∞ v θ+ −∞ −vθ+

with the sequence of random velocity {vn (x0 , v0 )}n∈N∪{0} we can reconstruct the
trajectory and calculate the probability density function of the random particle.

Let {Wn (x0 , v0 )}n∈N∪{0} be the sequence of random variables for the time lapse
of n -consecutive collisions.

n
Wn (x0 , v0 ) ≡ σ j (x0 , v0 ).
j=0
 Stochastic Formulation for the Initial-Boundary Value Problems 233

The trajectory of the random particle ω(t) is defined by

⎧
⎪
⎪x0 + v0 t for t ∈ (0, σ0 ),
⎪
⎪
⎪
⎪  + v1 t − σ0 for t ∈ (σ0 , W0 ), v0 > 0,
⎪
⎪
⎪
⎪
⎪
⎨− + v2k (t − W2k ) for t ∈ (W2k , W2k+1 ), v0 > 0, k
0,
ω(t) =  + v2k+1 (t − W2k−1 ) for t ∈ (W2k−1 , W2k ), v0 > 0, k
1,
⎪
⎪
⎪
⎪ − + v1 t − σ0 for t ∈ (σ0 , W0 ), v0 < 0,
⎪
⎪
⎪
⎪
⎪
⎪  + v2k (t − W2k ) for t ∈ (W2k , W2k+1 ), v0 < 0, k
0,
⎪
⎩− + v
2k+1 (t − W2k−1 ) for t ∈ (W2k−1 , W2k ), v0 < 0, k
1.

Let T̃(x, t, v) denote the probability density function of the random variable
(ω(t), ω̇(t)):

P(ω(t) ∈ (x, x + dx), ω̇(t) ∈ (v, v + dv)) = T̃(x, t, v) dx dv.

Since the trajectories consist of straight line segments connecting x = − 

and x = , the probability density function T̃(x − v∆t, t + ∆t, v) dx dv =
T̃(x, t, v) dx dv, for almost all t , and the function T̃(x, t, v) satisfies the free trans-
port equation

∂t T̃(x, t, v) + v∂x T̃(x, t, v) = 0.

Next, we need to show that T̃ will satisfy the diffuse boundary conditions. First
we need to identify the probability meaning of the boundary flux functions ρ± (t)
which is the flux of the total number of particles leaving the boundary at x = ±.
Thus, the measure ρ± (t) dt is identical to the probability of the particle colliding
with the boundary in the interval (t, t + dt). We can use the random variables Wn
to compute ρ± (t) dt as follows
⎧∞
⎪
⎪
⎪
⎪ P {W2n+1 ∈ (t, t + dt)} for v0 > 0,
⎨
ρ− (t) dt = ∞ n=0
⎪
⎪
⎪
⎪ P {W2n ∈ (t, t + dt)} for v0 < 0,
⎩
⎧∞n=0 (28)
⎪
⎪
⎪
⎪ P {W2n ∈ (t, t + dt)} for v0 > 0,
⎨
ρ+ (t) dt = ∞ n=0
⎪
⎪
⎪
⎪ P {W2n+1 ∈ (t, t + dt)} for v0 < 0.
⎩
n=0

The measures ρ± (t) dt will be re-distributed into the measures in the state space
according to probability laws in (24), (25):

ρ− (t)H0 (v) dt dv for v > 0,
(29)
ρ+ (t)H1 (v) dt dv forv < 0.
The probability that a particle emits from x =  in the time interval [t, t + dt]
with velocity in [v − dv, v] is the same as the probability that a particle in the
234 Shih-Hsien Yu

dx
= v + dv
dt

dx ~
=v T ( , t , v)dxdv
dt
ρ + (t ) H1 (v)dtdv
dx
dt

Fig. 3

interval [ − vdt, ] at time t with velocity is in [v − dv, v], since these two
sets are equivalent moduli by the characteristic curves dx dt = v . This gives rise to
T̃(, t, v) dx dv = H1 (t)ρ+ (t) dt dv . Next, replace v = dx/dt into the above. See
Fig. 3. It follows

T̃(, t, v) dx dv
√
dx ρ+ (t) 2π
= ρ+ (t)H1 (v) dv dt = ρ+ (t)H1 (v) dv = √ G+ (v) dx dv.
|v| θ+

From this we conclude that the function T̃ satisfies the diffuse boundary conditions:
⎧ √
⎪
⎪ 2π
⎪
⎨T̃(−, t, v) = ρ− (t) √ G− (v) for v > 0,
θ−
√ (30)
⎪
⎪ 2π
⎪
⎩T̃(, t, v) = ρ+ (t) √ G+ (v) for v < 0.
θ+

Thus, the probability density function of this stochastic process ω(t) solves the free
transport equation with the diffuse boundary condition, that is

T̃(x, t, v) = T(x, t, v)

where T(x, t, v) is the solution of (20) For the convenience of studying the behavior
of the stochastic processes Wn , we will use the time lapse sequence {σn }n∈N∪{0} to
continue the study. The time lapse of i + 1-th collision is σi = 2|/vi |. The time
lapse random variable is
⎧
⎪
⎪ − 22
⎪
⎪ 2 2 42 e σ 2 θ−
⎪
⎨H0 (v) dv = H0 d = dσ ≡ H− (σ ) dσ,
σ σ θ− σ3 (31)
⎪
⎪ − 22
⎪
⎪ 2 2 42 e σ 2 θ+
⎪
⎩H1 (v) dv = H1 d = dσ ≡ H+ (σ ) dσ,
σ σ θ+ σ3
 Stochastic Formulation for the Initial-Boundary Value Problems 235

where
⎧
⎪
⎨ 4
2 2 2π
G± for t > 0,
H± (t) ≡ t 3 t θ±
⎪
⎩
0 else,
+
Qβ = H+ ∗ H− ∗ H+ ∗ H− ∗ · · · ∗ H− , (32)
  
2β+2
Q−
β = H− ∗ H+ ∗ H− ∗ H+ ∗ · · · ∗ H+ .
  
2β+2

Substitute (31) into the random variable Wn (x0 , v0 ). Then plug it into (28) to get
that for v0 > 0
⎛ ⎞
∞
 − x0
ρ+ (t) = ⎝1 + Q− ⎠
β (t) ∗ δ t − ,
v0
β=0⎛ ⎞ (33)
∞
 − x
ρ− (t) = H+ (t) ∗ ⎝1 + Q β (t)⎠ ∗ δ t −
+ 0
.
v0
β=0

For general initial data T0 with a restrictive condition T0 (x, v) ≡ 0forv < 0, one
has
 ∞ 
T0 (x, v) = T0 (x0 , v0 )δ(x0 − x)δ(v0 − v) dx0 dv0 .
0 −

From this decomposition and (33),

⎛ ⎞
∞  ∞ 
 − x0
ρ+ (t) = ⎝1 + Q− ⎠
β (t) ∗ δ t− ∗ δ
v0 0 −
β=0
 − x0
× t− T0 (x0 , v0 ) dx0 dv0
v0
⎛ ⎞
∞  2/t
= ⎝1 + Q−
β (t) ⎠∗ v0 T0 ( − v0 t, v0 ) dv0 . (34)
β=0 0

By a symmetric reasoning for a general initial data T0 (x, v),

⎛ ⎞
∞
 + 
ρ+ (t) = ⎝1 + Q− ⎠ −
β (t) ∗ T0 (t) + H− (t) ∗ T0 (t) , (35)
β=0
⎛ ⎞
∞
 − 
ρ− (t) = ⎝1 + Q+ ⎠ +
β (t) ∗ T0 (t) + H+ (t) ∗ T0 (t) , (36)
β=0
236 Shih-Hsien Yu

where
⎧  2/t
⎪
⎪ +
⎨ 0
T (t) ≡ vT0 ( − vt, 0, v) dv,
0
0
⎪
⎪
⎩T0− (t) ≡ − vT0 (− − vt, 0, v) dv.
−2/t

2.4. Representation of a free transport equation from PDE (an alternative

approach)

The function T(x, t, v) satisfies a scalar hyperbolic equation so that the solution
T(x, t, v) can be represented in terms of the boundary data ρ± (t) as follows
⎧
⎪
⎪ 2π
⎪
⎪ ρ− (t − (x + )/v)G− (v) if v > 0, t > (x + )/v,
⎪
⎪
⎨ θ−
T(x, t, v) = 2π
⎪
⎪ ρ+ (t + (−x + )/v)G+ (v) if v < 0, t > −(x − )/v,
⎪
⎪ θ+
⎪
⎪
⎩
T0 (x − vt, v) else.
(37)

This gives rise the representation ρ± (t):

  2/t 
ρ+ (t) = vT(, t, v) dv = vT(, t, v) dv + vT(, t, v) dv
v>0 0 v>2/t
 2/t
= vT0 ( − vt, 0, v) dv
0
 t 2 2π2 2
+ G− ρ− (τ ) d
0 t −τ t −τ
θ− t −τ
 2/t  t
42 2 2π
= vT0 ( − vt, 0, v) dv + G− ρ− (τ ) dτ.
0 (t − τ ) t −τ θ−
3
0
(38)
 0
ρ− (t) = − vT0 (− − vt, 0, v) dv
−2/t
 t 42 2 2π
+ G+ ρ+ (τ ) dτ. (39)
0 (t − τ )3 t −τ θ+

Here, the function H− (t) ∗ H+ (t) will satisfy that

2
−
2 e 2θ± t 2
|H− (t) ∗ H+ (t)|, |H+ (t) ∗ H− (t)|  O(1) . (40)
t3
 Stochastic Formulation for the Initial-Boundary Value Problems 237

Then, (38) and (39) can be rewritten as

ρ+ (t) = T0+ (t) + H− (t) ∗ ρ− (t), (41)
ρ− (t) = T0− (t) + H+ (t) ∗ ρ+ (t), (42)
and with the above two one can expand ρ± (t) into
⎛ ⎞
∞
 + 
ρ+ (t) = ⎝1 + Q− ⎠ −
β (t) ∗ T0 (t) + H− (t) ∗ T0 (t) , (43)
β=0
⎛ ⎞
∞
 − 
ρ− (t) = ⎝1 + Q+ ⎠
β (t) ∗ T0 (t) + H+ (t) ∗ T0+ (t) . (44)
β=0

It is also clear that from the definitions in (32) that the functions in (43) and
(44) satisfy
⎧
⎪
⎪ H± (t), Q ±β (t) > 0 for t > 0,
⎨
H (t), Q ± β (t) = 0 for t  0,
 ±∞ (45)
⎪
⎪
∞
⎩ H± (t) dt = ±
Qβ (t) dt = 1,
0 0
and
2

|T0± (t)| = O(1) . (46)
+t
Lemma 5. (Global existence of ρ± ) Suppose that T0 (x, v)
0 and satisfies that
 
 T 
 0 
   1. (47)
 G−  ∞ ∞
L (L ) x v

Then, there exists C0 such that

ρ± (t)  O(1)et . (48)
Furthermore, if T0 (x, v)
0 for all x and v, then ρ± (t)
0 for all t
0.
Proof. The condition (47) yields
 ∞
|T0+ (t)| + |T0− (t)| dt < ∞. (49)
0
∞
Since 0 |H± (t)|dt = 1, there exists C0 > 0
C0 t β
|Q β (t)|  .
β!
Thus, (43), (44), and (49) conclude that
∞
tβ
ρ± (t)  O(1) = O(1)et .
β!
β=0

The positive-valued property is a consequence of H± and T0 being positive-valued

functions and the expansions (43) and (44).
238 Shih-Hsien Yu

3. The free transport equation: maximal principles

3.1. Maximal principle

The function S(x, v) given by

⎧
⎪
⎪ 2π
⎪
⎪
⎨ θ G+ (v) for v < 0,
+
S(x, v) ≡ (50)
⎪
⎪ 2π
⎪
⎪
⎩ θ G− (v) for v > 0,
−

is a stationary solution of (20).

Lemma 6. (First maximal principle) Suppose that the initial data T0 (x, v) of (20)
satisfies

0  T0 (x, v)  S(x, v) for all x and v.

Then, the solution T(x, t, v) of (20) satisfies

T(x, t, v)  S(x, v) for all t > 0, x, and v under consideration. (51)

Proof. The function S(x, v) − T(x, t, v) is a solution of (20) with positive-valued

initial data S(x, v)−T0 (x, v) > 0. Lemma 5 can be applied. The function S(x, v)−
T(x, t, v) is a positive-valued function.

The condition stated in Lemma 6 requires the initial data T0 bounded by S from
the above. This condition is too strong to implement this maximal principle to a
rather general setting required by the Boltzmann equation. However, such a strong
condition is generically true after some time lapse. It is assured by the following
lemma.

Lemma 7. (Second maximal principle) Suppose that |θ− − θ+ | is sufficiently small

and that the initial data T0 (x, v) of (20) satisfies
 
 T 
 0 
   1. (52)
 G−  ∞
L∞
x (L ξ )

Then, there exist t0 and C such that

T(x, t, v)
sup  C. (53)
t
t0 S(x, v)
x∈[−,]
v∈R
 Stochastic Formulation for the Initial-Boundary Value Problems 239

Proof. Any particle with velocity |v|
1 will collide with the boundaries |x| = 
within a time interval less than 2. Thus, when t = t0 ≡ 3, all particles with velocity
|v| > 1 can be connected to the boundaries |x| =  instead of the initial data. It
follows
T(x, 3, v)
⎧ 
⎪
⎪ 2π
⎪
⎪ max ρ− (t) G− (v) for v > 1,
⎪
⎪ θ−
⎪
⎪
t∈[0,3]
⎪
⎪
⎪
⎪
⎪
⎪ 
⎪
⎪
⎪
⎪ 2π
⎪
⎪ max ρ (t) G+ (v) for v < −1,
⎪ t∈[0,3] +
⎪ θ+
⎪
⎪
⎨
   (54)
⎪
⎪ 2π
⎪
⎪ max ρ− (t) G− (v) + max T0 (x, v) for 0 < v < 1,
⎪
⎪
⎪
⎪ t∈[0,3] θ− |v|1
⎪
⎪
⎪
⎪
x∈[−,]
⎪
⎪
⎪
⎪  
⎪
⎪
⎪
⎪ 2π
⎪ max ρ+ (t)
⎪ G+ (v) + max T0 (x, v) for 0 > v > −1.
⎪
⎪ θ+
⎩ t∈[0,3] |v|1
x∈[−,]

Due to (48),
sup |ρ± (τ )| = O(1). (55)
τ ∈[0,3]

From (55) and (54) there exists C0 > 0 such that

ρ− (t)2π ρ+ (t)2π
sup + < C0 . (56)
0<t<3 θ− θ+
From (54) and (56), there exists C > 0 such that
T(x, 3, v)
sup < C. (57)
x∈[−,] S(x, v)
v∈R

Then, by Lemma 6 and (57) this lemma follows for t
3.

We introduce another maximal principle for initial data with non-small velocity
distribution.
Lemma 8. (Third maximal principle) Let δ0 > 0. Suppose that the initial data T0 of
(20) satisfying that
T0 = 0 for v  δ0 . (58)
Then,
ρ+ (t)  sup ρ+ (τ ) for t > 0. (59)
τ ∈[0,2/δ0 ]
240 Shih-Hsien Yu

Proof. From (58) one has that T0 (x, v) = 0 for v  −δ0 . This yields that
 0
T0− (t) = T0 (− − vt, v) dv = 0 for t > 0, (60)
−2/t
 2/t
T0+ (t) = T0 ( − vt, v) dv = 0 for t > 2/δ0 . (61)
0

Thus T0− ≡ 0 and T0+ (t) = 0 for t > 2/δ0 . This, (41) and (42) result in that

ρ+ (t) = H− (t) ∗ ρ− (t) + T0+ (t) = H− (t) ∗ H+ (t) ∗ ρ+ (t) + T0+ (t)
= H− (t) ∗ H+ (t) ∗ ρ+ (t) for t > 2/δ0 . (62)

Due to (45) one has

H− (t) ∗ H+ (t) L1 = 1.

From this and (62), one can conclude that the maximum of ρ+ (t) can not be attained
in the domain (2/δ0 , ∞). This lemma follows.

3.2. L 1 in space to L ∞ at boundary

Lemma 9. Let ε0 > 0 be a given positive number. Suppose that the initial data T0
of (20) satisfying

T0
0, 
0 for |v| > ε0 , (63)
T0 (x, v) =
O(1) else.

Then, the boundary flux functions ρ± (t) satisfy

ρ± ∞  O(1)ε0 . (64)

Proof. Repartition the initial data T0 :


T0 (x, v) = −n  j n χ j,n (x, v),
⎧ 1/ε0 n<∞
⎪
⎨1 if x ∈ (− + ( j − 1)2/n, − + j2/n] (65)
χ j,n (x, v) ≡ and v ∈ (1/(n + 1), 1/n],
⎪
⎩
0 else,

and
 2/t
χ+
j,n (t) ≡ vχ j,n ( − vt, v) dv,
0
 0
χ−
j,n (t) ≡ −vχ j,n (− − vt, v) dv.
−2/t
 Stochastic Formulation for the Initial-Boundary Value Problems 241

T j,n denotes the solution of (20) with initial data χ j,n . The solution T(x, t, v) can
be decomposed into
∞ n
T(x, t, v) = T j,n (x, t, v). (66)
n=1 j=1

For each given ( j, n) with 1  j  n, one has the translation property for T j,n :
⎧
⎪
⎨T (x, v, t) = 0 for t < (n − j)2
j,n

T j,n (x, v, t)  2(T n−1,n + T n,n )(x, v, t − (n − j)2) (67)

⎪
⎩
for t
(n − j)2 (time-shift property).

This translation property for T j,n can be illustrated by the following diagram:

t 4 (n + 1) /n

1
v=
T j ,n = 0 n +1
(n − j)2
1
v=
n
T j ,n = 0
j2
x =− +
n
x= x
2
n

With this translation property, this problem is reduced to the studying of the
time-asymptotic property of Tn,n and Tn−1,n . Since Tn−1,n is next to Tn,n . It is
sufficient to study Tn,n .
The initial datum T n,n satisfies

T n,n (x, 0, v) = χn,n (x, v)


O(1) for all (x, v) ∈ [ − 2/n, ] × [1/(n + 1), 1/n],
= (68)
0 else.

Denote the boundary flux of Tn,n by ρn,± (t)


ρn,± (t) ≡ |v|Tn,n (±, t, v) dv.
±v>0
242 Shih-Hsien Yu

Then, by (48)
⎛ ⎞ 
∞  
2/t 
ρn,+ (8) = ⎝1 + H− (t) ∗ H+ (t) + Q− ⎠
β (t) ∗ vχn,n ( − vt, v) dv 
β=1 0 
t=8
 1/n 1
= O(1)e 8
vdv  C0 3 . (69)
1/(n+1) n
The velocity distribution of the initial data χn,n is confined in the region
[1/(n + 1), 1/n]. Thus,
−
χn,n ≡ 0,

+ 0 for t > 8,
χn,n (t) = O(1)
n3
for t  8.
Then from the above and by the identity from (41) and (42) one has that
+
ρn,+ (t) = H− (t) ∗ ρ− (t) + χn,n (t) = H− (t) ∗ H+ (t) ∗ ρn,+ (t)
− +
+H− (t) ∗ χn,n (t) + χn,n (t)
= H− (t) ∗ H+ (t) ∗ ρn,+ (t) for t
8. (70)
Combining this with maximal principle and (69), we have
C0
ρn,+ (t)  for t > 8. (71)
n3
Similarly,

1
vTn−1,n (, t, v) dv = O(1) for t
0. (72)
v>0 n3
Finally, we combine the decomposition (66), time translation property (67), and the
estimates (71), (72) together to obtain
 
ρ+ (t) = vT(, t, v) dv = v T j,n (, t, v) dv
v>0 v>0
−n  j n
1/ε0 n<∞
⎛ ⎞
∞
t
n− 2 n 
⎜ ⎟
= ⎝ + ⎠ vT j,n (, t, v) dv
n=1/ε0 j=0 t v>0
j=n− 2
∞ n 
 vTn,n (, t − (n − j)2, v)dv
n=1/ε0 j=n− t v>0
2
1 1
 O(1)  O(1) = O(1)ε0 . (73)
n3 n3
−n  j n −n  j n
1/ε0 n<∞ 1/ε0 n<∞
(n− j+4)<t

This proves (64).

 Stochastic Formulation for the Initial-Boundary Value Problems 243

4. The free transport equation: probability fluctuation

and rate of convergence

In the previous section we have used PDE-type and combinatorics analysis to

obtain upper bounds of the free transport equation. In order to obtain the conver-
gence and its rate of convergence, we will return to probability theories.

4.1. Probability estimates

4.1.1. Law of large numbers Denote

{X i = σ2i (x0 , v0 ) + σ2i−1 (x0 , v0 )}i∈N ,

where {σi }i∈N is the random sequence of the time lapse of a random particle bounc-
ing between the slabs given in (26) and with the condition

v0 > 0.

It is a sequence of independent identical random variables with the probability

density function H− (t) ∗ H+ (t):

P{X i ∈ [t, t + dt]} = (H+ (t) ∗ H− (t)) dt;

and we also introduce the following random variables,

n
Sn ≡ X j,
j=1

γ Xj if |X j |  γ ,
Yj ≡
0 if |X j |
γ ,
n
γ γ
Tn ≡ Yj ,
j=1

where γ is a positive parameter.

Lemma 10. For any γ
n 1/2 one has

P {|Sn − n E[X 1 ]| > γ }  O(1)nγ −2 log γ . (74)

In particular, for any α ∈ (1/2, 3/2) one has that

  "
 Sn 
P  − E[X 1 ] > n −1+α  O(1)(n 1−2α log n + n −3 ), (75)
n

where the constant O(1) is independent of n but α. The above is equivalent to

# $  
P |Sn − n E[X 1 ]| > n α  O(1) n 1−2α log n + n −3 . (76)
244 Shih-Hsien Yu

Proof. Since Var(σi ) are unbounded due to (27), we need to introduce the equivalent
γ
class Yn due to Khintchine, compare[16]. For a fixed n we consider the equivalent
γ
class {Y j } j∈N . From (40) one has that
2
−
2 e (θ− +θ+ )t 2
|H− (t) ∗ H+ (t)|  O(1) for t > 0. (77)
t3
γ
The random variable Y j satisfies
⎧ % &  ∞
⎪
⎪ γ
⎪
⎨ P Y j  = X j = (H− (t) ∗ H+ (t)) dt = O(1)γ −2 ,
γ
γ (78)
⎪
⎪ E[|Y1 − X 1 |] = O(1) γ −1 ,
⎪
⎩ E[|Y γ − X |2 ] = O(1)2 | log(γ )|.
1 1
%  &
 
We decompose the event  Snn − E[X 1 ] > n −1 γ :
⎧ ⎫
  " ⎨  "' ⎬+ +
 Sn   S  n
γ
n
 − E[X 1 ] > n −1 γ =  n − E[X 1 ] > n −1 γ {Y = X }
n  ⎩ n  j j
⎭
j=1 j=1
  "' "
 Sn  γ
×  − E[X 1 ] > n −1 γ {Y j  = X j }
n
⎧ ⎫
⎨ T γ  "' ⎬+ +
 n
γ
n
=  n
− E[X ]  > n −1 γ {Y = X }
⎩  n 
1 j
j ⎭
j=1 j=1
  "' "
 Sn  γ
×  − E[X 1 ] > n −1 γ {Y j  = X j } .
n
(79)
From the above decomposition and (78),
  "
 Sn 

P  − E[X 1 ] > n γ −1
n
 γ  " % &
 Tn  n
γ
 P  − E[X 1 ] > n −1 γ + P Y j = X j
n
j=1
 γ  "
 Tn 
 P  − E[X 1 ] > n −1 γ
n
 γ  "
n
1  Tn  n
+ = P  − E[X ]  > n −1
γ + 2
γ2  n 1 
γ
j=1
 γ  "
 Tn   γ 
P   − E[Y1 ] > n −1 γ −  E[Y1 ] − E[X 1 ]
n
 γ  "
n  Tn  O(1) n
+ 2  P  − E[Y1 ] > n −1 γ − + 2
γ n γ γ
 γ  "
 Tn  n
 P  − E[Y1 ] > n −1 γ /2 + 2 . (80)
n γ
 Stochastic Formulation for the Initial-Boundary Value Problems 245

We have the following from the Chebychev’s inequality and independence of

γ
{Y j } j :
% &
γ γ 2
# γ  $ E Tn − n E[Y1 ]
γ
P Tn − n E[Y1 ] > A 
% A2
γ γ 2 &
n E Yn − E[Y1 ] n log γ
= = O(1)2 . (81)
A2 A2
Substitute A = γ into the above, then (74) follows. Substitute A = n α and γ = n α
with α ∈ (1/2, 3/2) into (81), then (75) follows.

4.1.2. Central limit theorem

Lemma 11. There exists C > 0 such that
 
 k  (1+k)
∂x Q n−1 (x)  o(1)n − 2 for k = 0, 1. (82)
∞

Proof. Since Q n−1 is defined as the nth iteration of the convolution of H+ ∗ H− ,

we start to consider the Fourier transformation of H

+ ∗ H− (η):

M(η) ≡ H
+ ∗ H− (η) = e−iηx H+ ∗ H− (x) dx,
R

and estimate ∂xk Q n−1 (x) in terms of the Fourier transformation


eiηx ηk (M(η))n dη for k = 0, 1.
R

Denote
22
 ∞ −
42 e x 2 θ±
M± (η) ≡ (cos(ηx) + i sin(ηx)) dx,
0 x 3 θ±
and it follows

M(η) = M− (η)M+ (η).

We have
42 42
 ∞ −  ∞ 1 −
42 e x 2 θ± d 42 e x 2 θ±
cos(ηx) dx = 1 + cos(ηxt) dt dx
0 x 3 θ± 0 0 dt x 3 θ±
42
 1 ∞ −
42 e x 2 θ±
= 1−η sin(ηxt) dx dt
0 0 x 2 θ±
2 η2
 1 ∞ − 4
42 e x̄ 2 θ±
= 1 − η2 sin(x̄t) d x̄ dt,
0 0 x̄ 2 θ±
(83)
246 Shih-Hsien Yu

and
42 2 η2
 ∞ −  ∞ − 4
42 e x 2 θ± 42 e x̄ 2 θ±
sin(ηx) dx = η2 sin(x̄) d x̄. (84)
0 x 3 θ± 0 x̄ 3 θ±
By a direct computation, one has that
2 η2
 1 ∞ − 4
42 e x̄ 2 θ±
M±,a ≡ lim | log η|−1 sin(x̄t) d x̄ dt ∈ (0, ∞) (85)
η→0 0 0 x̄ 2 θ±
2 η2
 ∞ − 4
42 e x̄ 2 θ±
M±,b ≡ lim η sin(x̄) d x̄ ∈ (0, ∞), (86)
η→0 0 x̄ 3 θ±
M±,a = O(1), M±,b = O(1). (87)
From the fact that
  
 ∞  ∞
|M± (η)| =  e iηx
H± (x) dx   H± (x) dx = 1,
0 0
the equality holds only when η = 0, M± (η) are continuous functions in η, (85) and
(86), there exist η0 and α0 ∈ (0, 1) such that
|M± (η)| < α0 < 1 for |η| > η0 , (88)
M± (η) = 1−(M±,a + o(1))η2 | log η|+i(M±,b +o(1))η for |η|  η0 . (89)
It follows for |η|  η0
log M(η) = log M− (η) + log M+ (η)
= −η2 | log η|(M+,a + M−,a + o(1))+iη(M+,b + M−,b + o(1)). (90)
The above three estimates lead to the estimate
 
 
 iηk ei xη M(η)n dη
 
R
   
 
 
= + η M(η) dη
k n
 |η|η0 |η|
η0 
 
 
 
 |η|k |M(η)n | dη + O(1)α0n−2
 |η|η0 
 
 
 k n(−(Ma,+ Ma,− +o(1))| log η|η2 +i(Mb,+ +Mb,− +o(1))η) 
 |η| |e | dη + O(1)α0n−2
 |η|η0 

|η|k en(−(Ma,+ Ma,− +o(1))| log η|η ) dη + O(1)α0n−2
2

|η|η0
 
 
 
= n −(1+k)/2  k (−Ma,+ −Ma,− +o(1))| log η0 |η̄ 2
|η̄| e d η̄  + O(1)α0n−2
 |η̄|√nη0 
= o(1)n −(1+k)/2 + α0n−1 . (91)
 Stochastic Formulation for the Initial-Boundary Value Problems 247

4.2. Fluctuation in the boundary flux function ρ± (t)

We continue to estimate ρ+ (t) − ρ+ (t ) with t ∈ [t − (t/)1/10 , t].

For any n ∈ N one reformulates (43) as follows
 n−1

 
ρ+ (t) = 1 + H− (t) ∗ H+ (t) + Q l (t) ∗ T0+ (t) + H− (t) ∗ T0− (t)
−

l=1
+Q −
n (t) ∗ ρ+ (t). (92)

From (74) one has that when t/  1 and l t/

 ∞ 
 
  O(1)l log(/t)  .
2
 Q ±
(τ ) dτ (93)
 l  t2
t/2

It follows that
√
t/  t
 
Q l± (t − τ ) |T0+ (τ )| + |H− (τ ) ∗ T0− (τ )| dτ
l=1 0
√
t/  
t t/2  
= + Q l± (t − τ ) |T0+ (τ )| + |H− (τ ) ∗ T0− (τ )| dτ
l=1 t/2 0
√
t/
 
= O(1) max |T0+ (τ )| + |H− (τ ) ∗ T0− (τ )|
τ ∈(t/2,t)
l=1
 t/2
+ max |T0+ (τ )| + |H− (τ ) ∗ T0− (τ )| Q l± (t − τ ) dτ
τ ∈(0,t/2) 0
√
t/
  2
 2

= O(1) t/ + O(1) l log(t/) = O(1) log(t/) .
(+t) +t t
l=1
(94)
√
Substitute n = [ t/]
1 into (92). Then, it follows from (76), (82), and (94)
that
 t
|ρ+ (t) − ρ+ (t )|  O(1) Q n (t − τ )|ρ+ (τ )| dτ
t
 t
+ |Q n (t − τ ) − Q n (t − τ )|ρ+ (τ ) dτ
t −(t/)3/8
 t −(t/)3/10
+ |Q n (t − τ ) − Q n (t − τ )|ρ+ (τ ) dτ + O(1)t −1 log(1 + t)
0
(t − t ) |t − t |(t/)3/10 n log(t/)
 O(1) √ + + ρ+ ∞
n n (t/)3/5
+O(1) log(t/)(t/)−1
248 Shih-Hsien Yu

√
(t/)1/10 (t/)1/10 (t/)3/10 log(t/) t/
 O(1) √ + √ + ρ+ ∞
( t/)1/2 t/ (t/)3/5
+O(1) log(t/)(t/)−1
 O(1)(t/)−1/10 ρ+ ∞ + O(1)t −1 log(t/). (95)

4.3. From the fluctuation in the boundary flux function to fluctuation in solution

Proof of Theorem 1. For the purpose of relating the fluctuation in the boundary flux
functions to solutions in the interior domain, we introduce the function S∗ (x, v; τ )
to measure the fluctuation in the interior:
S∗ (x, v; τ ) = ρ+ (τ )S(x, v).
From (37), the function T(x, t, v) − S∗ (x, v; t) can be represented through the
boundary flux functions ρ± (t):
T(x, t, v) − S∗ (x, v; t)
⎧
⎪
⎨  θ− (ρ− (t − (x + )/v) − ρ+ (t))G− (v) for v > (x + )/t > 0,
2π
⎪
=
θ (ρ+ (t + (−x + )/v) − ρ+ (t))G+ (v) for v < (x − )/t < 0,
2π
⎪
⎪
⎩ +
T0 (x, v) − S∗ (x, v; t) else.
(96)
From (41) and (95), one has the variation of the boundary flux functions across the
interior domain, namely |ρ− (t) − ρ+ (t¯)|,
1/10

|ρ+ (t) − ρ− (t )|  O(1) for t ∈ [t − (t/)1/10 , t]. (97)
t +
Equations (95)–(97) yield
|T(x, v, t) − S∗ (x, v; t)|
 O(1)(t/)−1/10 ρ+ ∞ max(G+ (v), G− (v)) for |v| > /(t/)1/10 . (98)
From Lemma 7 one can conclude that
|T(x, t, v) − S∗ (x, v; t)|  O(1) for |v|  /(t/)1/10 . (99)
From conservation laws (23) there exists a unique ρ0 such that
 − 
T(x, t, v) − ρ0 S(x, v) dv dx = 0 for all t
0. (100)
 R
This yields that
     
0 = (ρ+ (t) − ρ0 ) S(x, v) dv dx + T(x, t, v) − S∗ (x, v; t) dv dx.
− R − R
(101)
 Stochastic Formulation for the Initial-Boundary Value Problems 249

By combining (98), (99), and (101) together one can conclude

|ρ+ (t) − ρ0 |  O(1)1/(t/)1/10 . (102)
(96) and (102) conclude the theorem.

4.4. Convergence of the full free transport equation

With the converging rate estimate of ρ± (t) in (102), we can return to the problem
(19). We denote the stationary solution of (19) by s(x, ξ ) for (x, ξ ) ∈ [−, ] × R3 :
⎧
⎪
⎪ 2π
⎪
⎪
⎨ θ M− (ξ ) for ξ > 0,
1
−
s(x, ξ ) = (103)
⎪
⎪ 2π
⎪
⎪
⎩ θ M+ (ξ ) for ξ < 0.
1
+

Proof of Theorem 2. In the reduction from (19) to (20), the functions ρ± (t) remain
unchanged. The estimate in (102) concludes that ρ± (t) converges to its limits with
a rate O(1)/((t + )/)1/10 in (6). The formula (37) can be generalized to
t(x, t, ξ )
⎧
⎪
⎨ θ− ρ− (t − (x + )/ξ )M− (ξ , ξ , ξ ) if ξ > 0, t > (x + )/ξ ,
2π 1 1 2 3 1 1
⎪
=
θ+ ρ+ (t +(−x + )/ξ )M+ (ξ , ξ , ξ ) if ξ < 0, t > −(x −)/ξ ,
2π 1 1 2 3 1 1
⎪
⎪
⎩
t0 (x − ξ 1 t, ξ ) else.
(104)
Combining (104) and (97), one has (5) and (7).

A control problem, I
The function I αt,Λ (x, t) with α
0 and Λ
0 denotes an external input function
with the properties
⎧ 
⎪ 
⎪
⎪ α
⎨− R3 I t,Λ (x, ξ ) dξdx = 0 for t
0,
⎪
 |I αt,Λ (x, ξ )|  (105)
⎪
⎪   −Λt
⎪
⎪     e .
⎩ (1 + |ξ |)α M− (ξ )  ∞ ∞
L x (L ξ )

We consider the following control problem with the input function I 0t,0 (x, ξ ):

∂t f + ξ 1 ∂x f = I 0t,0 ,
f(x, 0, ξ ) ≡ 0,
⎧
⎪
⎪ 2π
⎪
⎨f(±, t, ξ )|±ξ 1 <0 = ρ± (t) M± (ξ )|±ξ 1 <0 (106)
 θ±
⎪
⎪
⎪
⎩ρ± (t) ≡ |ξ 1 |f(±, t, ξ ) dξ.
±ξ 1 >0,ξ ∈R3
250 Shih-Hsien Yu

Proposition 1. The boundary flux functions ρ± (t) of (106) satisfy

 t 1/10

ρ± (t)  O(1) sup
0 +t −s x∈[−,]
  
× sup |ξ | 1
|I s,0 (x, s, ξ )| dξ dξ ds.
0 2 3
(107)
ξ 1 ∈R (ξ 2 ,ξ 3 )∈R2

Proof. By the Duhamel’s principle the solution f(x, t, ξ ) of (106) and the boundary
flux function can be represented as
 t
f(x, t, ξ ) = Tt−s
0 I s,0 (x, ξ ) ds,
0
0
and
 t
ρ± (t) = |ξ 1 |Tt−s
0 I s,0 (x, ξ ) dξ ds.
0
(108)
0 ±ξ 1 >0

Due to the zero total mass condition − R3 I 0s,0 (x, ξ ) dξ dx = 0 for s > 0, from
Theorem 2 one has that
 
   1/10
 |ξ 1 t−s 0
|T I (±, ξ ) dξ   O(1) sup
 0 s,0  ( + t − s)
±ξ 1 >0 x∈[−,]
  
× sup |ξ | 1
|I s,0 (x, ξ )| dξ dξ .
0 2 3
ξ 1 ∈R (ξ 2 ,ξ 3 )∈R2

This and (108) conclude this proposition.

5. The linearized Boltzmann equation, (Maxwell molecule collision model)

After estimating the equilibrating rate due to the diffuse boundary in the free
transport equation, we start to consider the mechanism due to particle-particle
collision. We consider as a simple case the linearized Boltzmann equation for the
Maxwell molecular model first:
∂t f + ξ 1 ∂x f = Lf,
⎧ 0, ξ ) = I0 (x, ξ ),
f(x,
⎪
⎪ 2π
⎪
⎪ f(±, t, ξ )|±ξ 1 <0 = ρ± (t) M±
⎪
⎪ θ±
⎪
⎪ 
⎨ (109)
ρ± (t) ≡ |ξ 1 |f(±, t, ξ ) dξ,
⎪
⎪ ±ξ >0,ξ ∈R
1 3
⎪
⎪ |ξ |2
⎪
⎪ − 2θ
⎪
⎪M (ξ ) ≡ e ±
⎩ ± ,
(2π θ± )3/2
where L = −ν + K is the linear collision operator given in (13). For the Maxwell
molecule gas model, the function ν(ξ ) is a constant
ν(ξ ) = ν0 > 0.
 Stochastic Formulation for the Initial-Boundary Value Problems 251

5.1. Rate of convergence

In this subsection we will establish the global estimate of the solution operator
St of (109) on the initial data I0 which satisfies the zero total mass condition.

Lemma 12. Suppose that the initial data I0 satisfies

⎧ 
⎪ 
⎪
⎪
⎨− R3 I0 (x,
⎪

ξ ) dξ = 0,
 I (x, ξ )  (110)
⎪
⎪  0 
⎪
⎪    O(1).
⎩ M− (ξ )  ∞ ∞
L (L ) x ξ

Then, the solution operator St of the linearized√Boltzmann equation for the

Maxwell molecule model (109) satisfies that for t

 √
−ν0 t/2 M− for |ξ 1 |  √,
|S I0 (x, ξ )|  O(1)e
t
(111)
1/40 M− for |ξ 1 |
.

Proof. Use the following iteration scheme to construct a solution local in time:

⎧ f0 (x, t, ξ ) ≡ T [I ](x, ξ ),
t
 0t 0
⎪
⎪
⎨d0 (x, t, ξ ) ≡ Tt−s
0 [Lf0 (·, s, ·)](x, ξ ) ds,
 t
0 (112)
⎪
⎪
⎩dn (x, t, ξ ) ≡ T0t−s [Ldn−1 (·, s, ·)](x, ξ ) ds for n
1,
0
ρn,± (t) ≡ ±ξ 1 >0,ξ ∈R3 |ξ 1 |dn (±, t, ξ ) dξ for n
0.

From the zero total mass condition in the initial data and from Theorem 2, there
exists τ0 > 0 independent of  such that
  √
1 for t  √,
ρ̄± (t) ≡ |ξ f0 (±, t, ξ )| dξ  O(1) 1/20
1
(113)
±ξ 1 >0,ξ ∈R3  for t ∈ [ , τ0 ],
⎧ √
 ⎨1 for t  ,√ √
f0 (x, t, ξ )  O(1) M− (ξ ) 1 for |ξ 1 |  √, t ∈ [√, τ0 ], (114)
⎩ 1/20
 1 for |ξ 1 |
, t ∈ [ , τ0 ].

The linear collision operator is orthogonal to the collision invariants. It results in the
zero total mass property
 
Lf0 (x, t, ξ ) dξ dx = 0 for t
0. (115)
− R3

Here, L is a bounded operator in · L ∞

ξ
. Thus, Proposition 1 can be applied together
with (114) to obtain that
√ √
 for t  [0, ],
|ρ0,± (t)|  O(1) √ √ (116)
1/20 (  + t)9/10 + 1/2 for t ∈ [ , τ0 ].
252 Shih-Hsien Yu

The function d0 (x, t, ξ ) can be represented as follows

 t
d0 (x, t, ξ ) = (Lf0 )(x − ξ 1 (t − τ ), τ, ξ 1 ) dτ
τ (x,t,ξ )
⎧
⎨0 if τ (x, t, ξ ) = 0,
+ ρ0,− (τ (x, t, ξ ))M− (ξ ) if τ (x, t, ξ ) > 0, ξ 1 > 0, (117)
⎩
ρ0,+ (τ (x, t, ξ ))M+ (ξ ) if τ (x, t, ξ ) > 0, ξ 1 < 0,
where τ (x, t, ξ ) is the backward intersecting time of the characteristic curve
Γ (s; x, t, ξ ) = x − ξ 1 (t − s) with the slab boundary, that is
Γ (τ (x, t, ξ ); x, t, ξ ) ∈ {−, } if τ (x, t, ξ ) > 0 else τ (x, t, ξ ) = 0.
It follows that
√ √
0 < t − τ (x, t, ξ )  2  when |ξ 1 | > . (118)
√
This gives rise to that when |ξ 1 | > ,
√ √
|t − τ (x, t, ξ )|  O(1)/  = O(1) .
√
This and (17) results in that for |ξ 1 |

 t
Lf0 (x − ξ 1 (t − τ ), τ, ξ ) dτ
τ (x,t,ξ )
 t
= − ν0 f0 (x − ξ 1 (t − τ ), τ, ξ )
τ (x,t,ξ )

 
M− (ξ )K (ξ, ξ∗ )f0 (x − ξ 1 (t − τ ), τ, ξ∗ )
+  dξ∗ dτ
R3 M− (ξ∗ )
 
√  f (x, τ, ξ )   √ 
0 
 O(1)ν0  sup    M− = O(1)  M− . (119)
τ ∈[0,t]  M−  ∞ ∞ L x (L ξ )
√
It follows directly with (17) when |ξ 1 | < 
 t
Lf0 (x − ξ 1 (t − τ ), τ, ξ ) dτ
τ (x,t,ξ )
 t

= −ν0 f0 (x − ξ 1 (t − τ ), τ, ξ )
τ (x,t,ξ )

K (ξ, ξ∗ )f0 (x − ξ 1 (t − τ ), τ, ξ∗ ) dξ∗ dτ
+
R3
 
 f (x, τ, ξ )  
0 
 O(1)ν0 t sup    = O(1)t M− . (120)
τ ∈[0,t]  M−  ∞ ∞ L x (L ξ )

From (116), (117), (119), and (120), one has that for t  τ0 1
  √
τ M (ξ ) if |ξ 1 |  √
,
|d0 (x, t, ξ )|  O(1) 01/10 − (121)
 M− (ξ ) for |ξ 1 |
.
 Stochastic Formulation for the Initial-Boundary Value Problems 253

with the same computation, one can show that for t ∈ [0, τ0 ]
  √
τ M (ξ ) if |ξ 1 |  √,
|dn (x, t, ξ )|  O(1)(τ0 )n 01/10 − (122)
 M− (ξ ) for |ξ 1 |
.
This geometric sequence shows  that the iteration scheme in (112) converges for
0 < t < τ0 1. The series f0 + ∞ n=0 dn defines the solution of (109) in the local
time interval [0, τ0 ] with τ0 1. With the estimates in (114), (121), and (122), the
solution f(x, t, ξ ) satisfies
 
 ∞ 
 
|f(x, t, ξ )| = f0 (x, t, ξ ) + dn (x, t, ξ )  O(1)
 
n=0
⎧ √
⎨  M− for t  √, √
× M− for t ∈ [√, τ0 ], |ξ 1 |  √, (123)
⎩ 1/10
 M− for t ∈ [ , τ0 ], |ξ 1 |
.
With this local time estimate (123), we continue to use a bootstrap procedure
to obtain a global exponential decaying estimate for the solution of (112) for
t ∈ [(l − 1/2)τ0 , (l + 1/2)τ0 ] with l ∈ N in the following.
From (123) there exists C0 such that
⎧  √ √
⎨ |f(x, t, ξ )|  C0 M− (ξ ) for |ξ 1 |  √, t ∈ [√, τ0 ],
|f(x, t, ξ )|  C0 1/40 M− (ξ ) for |ξ 1 |
√ , t ∈ [ , τ0 ], (124)
⎩
|ρ± (t)|  C0 1/40 for t ∈ [ , τ0 ].
Rewrite (109) as
∂t f + ξ 1 ∂x f + ν0 f = Kf.
By using the characteristic curve method one has the following representation of
f(x, t + τ0 , ξ ):
 τ0 
f x, t + , ξ
2 
⎧
−ν0 t τ0 
⎪
⎪ e f x − ξ 1 t, , ξ
⎪
⎪  t 2   
⎪
⎪  K (ξ, ξ∗ ) f x − ξ 1 (t − τ ) , τ + τ20 , ξ∗
⎪
⎪ −ν0 (t−τ )
⎪
⎪ + e M− (ξ )  dξ∗ dτ
⎪
⎪ R M− (ξ∗ )
⎪
⎪
0
⎪
⎪ if τ (x, t, ξ ) = 0,
⎪ 
⎪
⎪
⎪ −ν0 (t−τ (x,t,ξ )) ρ
τ0 
⎪
⎪ e − τ (x, t, ξ ) + M− (ξ )
⎪
⎨  t  2  
= −ν0 (t−τ )
 K (ξ, ξ∗ ) f x − ξ 1 (t − τ ) , τ + τ20 , ξ∗
⎪ + e M− (ξ ) 
⎪
⎪ τ (x,t,ξ ) M− (ξ∗ )
⎪
⎪
R
⎪
⎪ × τ (x, ξ ) > ξ 1 > 0,
⎪
⎪ dξ ∗ dτ if t, 0 and
⎪
⎪
⎪
⎪ e−ν0 (t−τ (x,t,ξ )) ρ+ (τ (x, t, ξ )) M+ (ξ )  
⎪
⎪  t 
⎪
⎪
⎪ + −ν0 (t−τ )
 K (ξ, ξ∗ ) f x −ξ 1 (t −τ ) , τ + τ20 , ξ∗
⎪
⎪ e M− (ξ )  dξ∗
⎪
⎪ τ (x,t,ξ ) R M− (ξ∗ )
⎪
⎩
× dτ if τ (x, t, ξ ) > 0 and ξ 1 < 0,
(125)
254 Shih-Hsien Yu

with this representation, (124), and Lemma 3, one can improve (124) by gaining a
small factor 1/40 from the integration operator Kf(x, t, ξ ). It follows
  −ν τ /2  √
e 0 0 + O(1)1/40 M− for |ξ 1 |  ,
|f(x, τ0 , ξ )|  C0  √ (126)
O(1)1/40 M− for |ξ 1 |
.

By assuming that  τ0 , one has that

 −ν τ /4  √
e 00 M
− for |ξ 1 |  √,
|f(x, τ0 , ξ )|  C0 (127)
O(1)1/40 M− for |ξ 1 |
.

Similarly the scheme will yield that from the data of f(x, τ0 /2, ξ ) given in (124),
√
|f(x, τ0 , ξ )|  O(1)1/20  C0 e−ν0 /4 1/40 for |ξ 1 |
. (128)

Finally with the assumption  1 we have that

 √
−ν0 τ0 /4 M− for |ξ 1 |  √,
|f(x, τ0 , ξ )|  C0 e (129)
 1/40 M− for |ξ 1 |
.

From the estimate (129) one can conclude that the solution f(x, t, ξ ) decays
exponentially in time by repeating the procedure to obtain the estimates (128) from
(124) in each time interval [kτ0 /2, (k + 1)τ0 /2] to obtain that when  1
 √
M for |ξ 1 |  √,
|f(x, kτ0 /2, ξ )|  e−(k−1)ν0 τ0 /4 C0 1/40− (130)
 M− for |ξ 1 |
.

5.2. Convergence to a stationary solution

We continue to establish the stationary solution SL (x, ξ ) of the problem (109):

ξ 1 ∂x SL − LSL = 0. (131)

Rewrite SL as

SL = s + D,

where s(x, ξ ) is the stationary solution of the free transport equation given in (103).
Thus,

0 = ξ 1 ∂x (s + D) − L(s + D).

This gives rise to

ξ 1 ∂x D = LD + Ls. (132)
 Stochastic Formulation for the Initial-Boundary Value Problems 255

Lemma 13. Let St denote the solution operator of (109). The function D(x, ξ ) =
∞ t
0 S Ls(x, ξ ) dt solves (132). Furthermore,
 √
 1 for |ξ 1 |  √,
|(SL − s)(x, ξ )|  O(1)|θ− − θ+ | M− (ξ ) · 1/40 (133)
 for |ξ 1 |
,

and
  
(SL − s)(x, ξ ) dξ dx = 0.
− R3

Proof. Consider the function D (x, t, ξ ):

, t -
D (x, t, ξ ) ≡ S Ls dτ (x, ξ ).
t−τ
(134)
0

This function D (x, t, ξ ) satisfies

∂ D + ξ 1 ∂x D = LD + Ls,
⎧t
⎪
⎪ D (x, 0, ξ ) = 0,
⎪
⎪
⎪
⎨ 2π
D (±, t, ξ )|±ξ 1 <0 = ρ± (t) M± (ξ )|±ξ 1 <0 , (135)
⎪  θ±
⎪
⎪
⎪
⎪
⎩ρ± (t) = |ξ 1 |D (±, t, ξ )dξ.
±ξ 1 >0

If the limit limt→∞ D (x, t, ξ ) exists, then (135) concludes that

D(x, ξ ) ≡ lim D (x, t, ξ ) (136)

t→∞

defines a solution of (132).

Since the linear operator L is orthogonal to the collision invariants,
 
Ls(x, ξ ) dξ dx = 0.
− R3

Lemma 12 can be applied to show that the function St Ls decays exponentially in

time pointwise. It yields that the limit limt→∞ D (x, t, ξ ) exists, and
 
 
|D(x, ξ )| =  lim D (x, t, ξ )
t→∞
 ∞    √
  Ls 
−ν0 τ/2   1 for |ξ 1 |  √,
 O(1) M− (ξ ) e dτ    · 1/40 (137)
0  M−  ∞  for |ξ 1 |
.
L ξ

Next, we need to estimate the structure of D(x, ξ ). By assuming that |θ− − θ+ | 1,

one has from Theorem 5 and Lemma 1 that

|Ls(ξ )|  O(1)|θ− − θ+ | M− (ξ ), (138)

|Q(s)(ξ )|  O(1)|θ− − θ+ | M− (ξ ). (139)
256 Shih-Hsien Yu

From (134), (137) and (138), one has that

 √
 1 for |ξ 1 |  √,
|D(x, ξ )|  O(1)|θ− − θ+ | M− · 1/40
 for |ξ 1 |
.
From (135), one also has that
    
D (x, t, ξ ) dξ dx = D (x, 0, ξ ) dξ dx = 0. (140)
− R3 − R3

This concludes
  
D(x, ξ ) dξ dx = 0.
− R3

Theorem 6. f(x, t, ξ ) the solution of the initial-boundary problem of (109) con-

verges to ρ0 SL exponentially:
 √
−ν0 t/2
 1 for |ξ 1 |  √,
|f(x, t, ξ ) − ρ0 SL (x, ξ )|  O(1)e M− (ξ ) 1/40
 for |ξ 1 |
,
where
   .  
ρ0 = I0 (x, ξ ) dξ dx s(x, ξ ) dξ dx.
− R3 − R3

Proof. V(x, t, ξ ) ≡ f(x, t, ξ ) − ρ0 SL (x, ξ ) denotes the perturbation. This function

V(x, t, ξ ) can be represented as St (I0 − ρ0 SL )(x, ξ ). Due to the choice of ρ0 , the
zero total mass condition is valid for V(x, 0, ξ ). Thus, Lemma 12 can be applied.

5.3. Existence of non-trivial nonlinear stationary solutions and its stability

We return to the problem (1) for a Maxwell molecule collision model. Denote
the nonlinear stationary solution of (1) in the form
φ = Dd + SL ,
where SL is defined in (131). It stands for a stationary solution of a linearized
problem. Then, the equation of the stationary solution ξ 1 ∂x F = Q(F) can be
rewritten as
ξ 1 ∂x Dd = LDd + Q(Dd + (SL − M− )). (141)
Theorem 7. When |θ− − θ+ | and  are sufficiently small, there is a nontrivial non-
linear stationary solution φ. Furthermore,
⎧ √
 ⎨1 for |ξ 1 |  ,
|φ(x, ξ ) − s(x, ξ )|  O(1)|θ+ − θ− | M− (ξ ) √ (142)
⎩ 1/40
 for |ξ |
.
1
 Stochastic Formulation for the Initial-Boundary Value Problems 257

Proof. One can consider the iteration scheme

⎧  ∞
⎪
⎪ 0
= St Q(SL − M− ) dt,
⎪
⎪ D
⎨ d 0
(143)
⎪
⎪  ∞  
⎪
⎪
⎩Dnd = St Q SL − M− + Ddn−1 dt for n
1.
0

Due to (133), (138), (139), and Lemma 1 one has that

  
 
Q(SL − M− )/ M−  ∞ ∞  O(1)|θ− − θ+ |2 1. (144)
L x (L ξ )

Since the collision operator Q is orthogonal to the collision invariant, St Q(Dd +

(s + D − M− )) decays exponentially in time. This, (144), and Lemma 12 yield that
⎧ √
 ⎨1 for |ξ 1 |  ,
D0d (x, ξ ) = O(1)|θ− − θ+ |2 M− (ξ ) · √ (145)
⎩ 1/40
 for |ξ |
.
1

Similarly, one has

⎧ √
 ⎨1 for |ξ 1 |  ,
D1d (x, ξ ) = O(1)|θ− − θ+ |2 M− (ξ ) · √ (146)
⎩ 1/40
 for |ξ 1 |
,
for n
2
 ∞ / 0
Dnd − Ddn−1 = St B 2(SL − M− ) + Ddn−1 + Ddn−2 , Dn−1
d − Ddn−2 dt.
0
(147)
We make the a priori assumption that for n
0
 
 n  
Dd / M−  ∞ ∞  |θ− − θ+ |. (148)
L x (L ξ )

This assumption holds when n = 0 due to (145). Under this assumption (133) and
(148), from (16) one has that
  
 
B(SL − M− + Dnd + Ddn−1 , Dnd − Ddn−1 )/ M−  ∞ ∞  O(1)|θ− − θ+ |
L x (L ξ )
 

×(Dnd − Ddn−1 )/ M−  ∞ ∞ . (149)
L x (L ξ )

This and (147) yields that the sequence {Dnd − Ddn−1 }n
1 is a geometric sequence

|Dnd (x, ξ ) − Dn−1

d (x, ξ )|
   
 
 O(1)|θ− − θ+ | M− (ξ ) (Ddn−1 − Ddn−2 )/ M−  ∞
L∞
x (L ξ )
 √
1 for |ξ 1 |  √,
· (150)
1/40 for |ξ 1 |

258 Shih-Hsien Yu

under the condition |θ− − θ+ | 1 and (148). This yields that for all n
1
     j 
 Dn   D0  n  D − D j−1 
 d   d   d 
     +   d

 M−  ∞
 M−  ∞
 M−  ∞
L∞
x (L ξ ) L∞
x (L ξ ) j=1 L∞
x (L ξ )
 
 D0 
 d 
 (1 + O(1)|θ− − θ+ |)    = O(1)|θ− − θ+ |2 . (151)
 M−  ∞
L∞
x (L ξ )

Thus, the assumption (149) is true and

lim Dnd exists.

n→∞

Thus,

φ(x, ξ ) ≡ SL + lim Dnd (x, ξ )

n→∞

is a non-trivial stationary solution of (1):

ξ 1 ∂x φ = Q(φ).

The property in (142) is a consequence of the property of the solution operator St

described in (111). The details are omitted.

Theorem 8. The stationary profile φ constructed in Theorem 7 is nonlinearly time-

asymptotically stable.

Proof. Consider the problem (1) with initial data F0 (x) satisfying
⎧  
⎪  
⎪
⎨ (F 0 (x, ξ ) − φ(x, ξ ))/ M − (ξ ) ∞ ∞ ε 1,
L x (L ξ )
 
⎪
⎪
⎩ (F0 (x, ξ ) − φ(x, ξ )) dξ dx = 0.
− R3

Let f(x, t, ξ ) denote F(x, t, ξ ) − φ(x, ξ ). The equation for f is

∂t f + ξ 1 ∂x f = Lf + 2B(φ − M− , f) + Q(f),
⎧
⎪
⎪ 2π
⎪
⎪ f(±, t, ξ )|±ξ 1 <0 = ρ± (t) M±
⎪
⎨ θ±
 (152)
⎪
⎪ ρ± (t) ≡ |ξ 1 |f(±, t, ξ )dξ,
⎪
⎪ ±ξ 1 >0,ξ ∈R3
⎪
⎩
f(x, 0, ξ ) = I0 (x, ξ ) ≡ F(x, 0, ξ ) − φ(x, ξ ).
 Stochastic Formulation for the Initial-Boundary Value Problems 259

We consider the following iteration scheme similar to (216) in constructing the

stationary solution φ(x, ξ ):
⎧ 0
⎪
⎪ Dd (x, t, ξ ) = St I0 (x, ξ ),
⎪
⎪  t
⎨ 
Dnd (x, t, ξ ) = St I0 (x, ξ ) + St−s B(φ − M− , Ddn−1 (·, s, ·)) (153)
⎪
⎪
⎪
⎪   0
⎩ +Q Dn−1 (·, s, ·) (x, ξ ) ds for n
1.
d
 t
Ddn+1 (x, t, ξ ) − Ddn (x, t, ξ ) = St−τ B(φ − M−
0

+ Ddn (·, τ, ·) + Ddn−1 (·, τ, ·), Ddn (·, τ, ·) − Ddn−1 (·, τ, ·))(x, ξ ) dτ. (154)
We introduce the norm |||| · ||||:
 
 eν0 t/8 |h(x, t, ξ )| 
 
||||h|||| = sup    .
t>0  M− (ξ )  ∞
L∞
x (L ξ )

From Lemma 12,

||||D0d ||||  O(1)ε. (155)
We make the a priori assumption that
√
||||Dnd ||||  O(1) ε for n
0. (156)
From this assumption, Theorem 7, and Lemma 1 we obtain
||||2B(φ − M− + Dnd + Dnd + Ddn−1 , Dnd − Dn−1 )||||
 √ 
 O(1) |θ− − θ+ | + ε ||||Dnd − Ddn−1 ||||. (157)
n−1
√ sequence {Dd − Dd }n∈N is a geometric
Under the a priori assumption of(156), the n

sequence with a contraction rate O(1)( ε + |θ− − θ+ |) 1. This and (155) allow
us to conclude that
n
√
||||Dnd ||||  O(1)( ε + |θ+ − θ− |) j−1 ||||D0d |||| = O(1)ε.
j=1

This assures that (156) is true for all n ∈ N. It shows that the sequence {Dnd }n∈N
converges in the norm |||| · ||||. This shows that the function f(x, t, ξ ) converges to
φ(x, ξ ) in time exponentially with a rate not slower than e−ν0 t/8 .

Corollary 1. If the initial data F(x, 0, ξ ) satisfies

⎧  
⎪
⎪
⎪
⎨ (φ(x, ξ ) − F(x, 0, ξ )) dξ dx = 0,
− R
3
 (158)
⎪  
⎪  F(x,0,ξ
⎪ √ )−φ(x,ξ ) 
 ε,
⎩ M− (ξ ) ∞
L∞
x (L ξ )
260 Shih-Hsien Yu

then

|F(x, t, ξ ) − φ(x, ξ )|  O(ε)e−ν0 t/8 M− (ξ )
⎧ 1/60
⎪
⎪  for |ξ 1 | > 1/2 , t > 3/7 /| log |
⎨
1 for |ξ 1 | < 1/2
⎪
⎪
⎩
1 for t  3/7 /| log |.

Proof. By applying Lemma 12 to the iteration scheme in (153), this corollary

follows.

6. Hard potential with angular cutoff collision models

In contrast to the Maxwell molecule collision model, the collision frequencies

ν(ξ ) grow to infinity as |ξ | → ∞ for angular cutoff hard potential collision models.
One shall not treat L = −ν + K as a perturbation to the free transport equation
in constructing the global solutions for the angular cutoff hard potential collision
models. We will need to analyze the transport equation with damping term.

6.1. Transport equation with damping

We consider the following the transport equation with damping:

∂t tν + ξ 1 ∂x tν = −ν(ξ )tν , x ∈ [−, ]

⎧
⎪
⎪ 2π
⎪
⎪ tν (±, t, ξ )|±ξ 1 <0 = ρ ν (t)± M± (ξ )|±ξ 1 <0
⎪
⎪ θ±
⎪
⎪ 
⎨ (159)
ν
ρ± (t) ≡ |ξ 1 |tν (±, t, ξ ) dξ,
⎪
⎪ ±ξ 1 >0,ξ ∈R3
⎪
⎪ |ξ |2
⎪
⎪ − 2θ
⎪
⎪ e ±
⎩M± (ξ ) ≡ ,
(2π θ± )3/2
tν (x, 0, ξ ) = t0 (x, ξ ),
 
 t 
 0 
    1, (160)
 ν M−  ∞ ∞
L x (L ξ )

where the function ν(ξ ) satisfies

⎧
⎨ inf ν(ξ ) = ν(0) > 0,
ξ ∈R 3
⎩ν(ξ ) = O(1)(1 + |ξ |)γ ,
 Stochastic Formulation for the Initial-Boundary Value Problems 261

where γ = s−4 ν
s ∈ (0, 1]. The representation of ρ± (t) in (43) and (44) can be modified
into
 
∞  
ν ν ν
ρ+ (t) = 1 + H− (t) ∗ H+ (t) + Q n (t) ∗ T0ν,− (t) + H−ν (t) ∗ T0ν,+ (t) ,
ν−

n=1
(161)
 
∞  
ν ν,+ ν,−
ρ− (t) = 1 + H+ν (t) ∗ H−ν (t) + Q ν+ ν
n (t) ∗ T0 (t) + H+ (t) ∗ T0 (t) .
n=1
(162)

ν,±
Here, the functions H±ν , Qν±
n , and T0 are defined as
⎧  
⎪  ∞  ∞ 2 2 +|ξ 2 |2 +|ξ 3 |2  
⎨ (2)2 − t ∗ ∗
−ν 2 2 3
t ,ξ∗ ,ξ∗ t
H±ν (t) ≡ e 2θ± e dξ∗2 dξ∗2 for t > 0,
⎪ −∞ −∞ t 3 2πθ±2
⎩
0 else,
Q ν+ ν ν ν ν ν
n = H+ ∗ H− ∗ H+ ∗ H− ∗ · · · ∗ H− ,
  
2n+2
Q ν− ν ν ν ν
n = H− ∗ H+ ∗ H− ∗ H+ ∗ · · · ∗ H+ ,
ν (163)
  
 2/t  ∞ 2n+2
 ∞
T0ν,+ (t) ≡ ξ∗1 e−ν(ξ∗ )t t0 ( − ξ∗1 t, ξ∗ ) dξ∗2 dξ∗3 dξ∗1 ,
0 0 −∞ ∞  ∞
−∞

T0ν,− (t) ≡ −ξ∗1 e−ν(ξ∗ )t t0 (− − ξ∗1 t, ξ∗ ) dξ∗2 dξ∗3 dξ∗1 .

−2/t −∞ −∞

Note that the factor 1

θ±
2 in H±ν is due to the fact that the dimension of the integration
domain of ξ∗ is three.

Lemma 14. Suppose that the initial data of (159) satisfies the zero total mass
condition
  
t0 (x, ξ ) dξ dx = 0. (164)
− R3

Then, for t < 3/7 /| log |

1/15
ν 
|ρ± (t)|  O(1) , (165)
+t

and
 1/15  √

ν M− for |ξ 1 |

,
|t (x, t, ξ )|  O(1) +t
√ (166)
ν(ξ ) M− for |ξ |  .
1
262 Shih-Hsien Yu

Proof. First we rescale the space-time variable:

⎧
⎪
⎪ x = x/, t = t/, ξ = ξ,
⎪
⎪
⎪
⎪t̄ν (x , t , ξ ) ≡ tν (x, t, ξ ),
⎪
⎪
⎪
⎪ H̄± (t ) dt ≡ H± (t) dt,
⎪
⎪
⎪
⎪ ν ν
⎪
⎪ H̄± (t ) dt ≡ H± (t) dt,
⎪
⎪
⎪
⎪ Q̄ ν+ ν ν ν ν ν
⎪
⎪ n (t ) = H̄+ (t ) ∗ H̄− (t ) ∗ H̄+ (t ) ∗ H̄− (t ) ∗ · · · ∗ H̄− (t ),
⎪
⎪   
⎪
⎪
⎨ ν− 2n+2
Q̄ n (t ) = H̄−ν (t ) ∗ H̄+ν (t ) ∗ H̄−ν (t ) ∗ H̄+ν (t ) ∗ · · · ∗ H̄+ν (t ), (167)
⎪   
⎪
⎪
⎪
⎪ 2n+2
⎪
⎪ Q̄ +
⎪
⎪ n (t ) = H̄+ (t ) ∗ H̄− (t ) ∗ H̄+ (t ) ∗ H̄− (t ) ∗ · · · ∗ H̄− (t ),

⎪
⎪
⎪
⎪
⎪ 2n+2
⎪ Q̄ − (t ) = H̄ (t ) ∗ H̄ (t ) ∗ H̄ (t ) ∗ H̄ (t ) ∗ · · · ∗ H̄ (t ),
⎪
⎪
⎪ n
⎪
⎪ − + −
 + +

⎪
⎪ 
⎪
⎪
2n+2
⎪
⎪ ν
⎩ρ̄± (t ) ≡ |ξ 1 |t̄ν (±1, t , ξ ) dξ.
±ξ 1 >0

Similar to (41), (42), and (92) one has that

 n−1

ν
ρ̄+ (t ) = 1+ H̄−ν (t )∗ H̄+ν (t )+ Q̄ lν− (t ) ∗
l=1
 
× T̄0ν,+ (t ) + H̄−ν (t ) ∗ T̄0ν− (t ) + Q̄ ν− ν
n (t ) ∗ ρ̄+ (t ), (168)
 n−1

ν
ρ̄− (t ) = 1 + H̄+ν (t ) ∗ H̄−ν (t ) + Q̄ lν+ (t ) ∗
l=1
 
× T̄0ν,− (t ) + H̄+ν (t ) ∗ T̄0ν+ (t ) + Q̄ ν+ ν
n (t ) ∗ ρ̄− (t ), (169)
ν
ρ̄+ (t )= T̄0ν,− (t )+ H̄−ν (t ν
) ∗ ρ̄− (t ), (170)
ν
ρ̄− (t )= T̄0ν,+ (t )+ H̄+ν (t ν
) ∗ ρ̄+ (t ). (171)
From (163) one can have the comparison for all t ∈ R
⎧ ν
⎪
⎨ H̄± (t )  H̄± (t ),
Q̄ ν± ±
n (t )  Q̄ n (t ),
⎪  (172)
⎩ T̄ν,± (t )  O(1) 1 2 .
0 1+t

With this comparison one can apply the probability estimates (law of large numbers
theory) in (93) to obtain that for t
l
 ∞  ∞
l| log t |
Q̄ lν± (τ ) dτ  Q̄ l± (τ ) dτ  O(1) , (173)
t t t2
ν (t ) that
and the global existence of ρ̄±
ν
ρ̄± L∞  O(1).
 Stochastic Formulation for the Initial-Boundary Value Problems 263

ν (t¯).
Next, we consider the fluctuations in the boundary flux functions ρ̄±
σ
When t <  0 for some σ0 ∈ (0, 1/2), one has that
sup (1 − e−ν(ξ∗ )t )(M− (ξ∗ ))1/4  O(1)t| log()|. (174)
ξ ∗ ∈R 3

This yields that for t < σ0 for some σ0 ∈ (0, 1)

 0  ∞ ∞
|ξ∗1 (e−ν(ξ∗ )t − 1)t0 (− − ξ∗1 t, ξ )| dξ∗2 dξ∗3 dξ∗1
−2/t −∞ −∞
 0  ∞  ∞
= O(1)t| log | |ξ∗1 |t0 (− − ξ∗1 t, ξ )|/(M− (ξ∗ ))1/4 dξ∗2 dξ∗3 dξ∗1 ,
−2/t −∞ −∞
(175)
and
|H±ν (t) − H± (t)|  O(1)t| log |H± (t). (176)
It leads to for t  −1+σ0 and σ0 ∈ (0, 1)
2
−
e θ± |t |2
H̄± (t ) = O(1) , (177)
|t |3
1
−
  (C0 t )2
 H̄± (t ) − H̄ ν (t )  O(1)| log | e = O(1)(| log |t ) H̄± (t ). (178)
± 2
t
This yields that when t = −β /| log | with β ∈ (0, 2/3)
 
H̄±ν (t ) = 1 + O(1)1−β H̄± (t ) for t = −β . (179)
Let
1 2
n = −β/2 .
This and (179) result in that
 
1−β n ±
 
Q̄ ν±
n (t ) = 1 + O(1) Q̄n (t ) = 1 + O(1)n1−β Q̄±
n (t )
  ± −β
= 1 + O(1) 1−3β/2
Q̄ n (t ) for t   . (180)

This also results in that for t¯, s̄ ∈ [1, −β ] and for l  n,
 
| Q̄ lν± (t¯)− Q̄ lν± (s̄)| = | Q̄ l± (t¯)− Q̄ l± (s̄)|+ O(1)1−3β/2 Q̄ l± (t¯)+ Q̄ l± (s̄) . (181)

From (172), (173), (179), and (181) one can reduce the estimates in H̄±ν and
Q̄ lν±to H̄± and Q̄ l± . Thus one applies the procedure to obtain the fluctuation of
ν (t ): for s ∈ [t − t 1/10 , t ]
ρ± (t) in (95) from (168) and (169) to obtain that for ρ̄±
ν ν
|ρ̄± (t ) − ρ̄± (s )|  O(1)
 
−3/20 −1/10 33/20 −1/10 ν −1
× t + [t +t ] + | log t |t ρ̄± ∞ +t | log t |.
(182)
264 Shih-Hsien Yu

Thus, when β < 4/7, (t = −β /| log |),

ν ν
|ρ̄± (t ) − ρ̄± (s )|  O(1)| log(1 + t )|(1 + t )−1/10
 (1 + t )−1/15 for t = −β , s ∈ [t − (t )1/10 , t ]. (183)

With this fluctuation estimate, (170), and (171), one has that
ν ν
|ρ̄− (s ) − ρ̄+ (s )|  O(1)(1 + t )−1/15 ρ̄±
ν
∞ for t  −4/7 /| log |. (184)

Under the scaling in (167), the equation (159) becomes

∂t t̄ν + ξ 1 ∂x t̄ν = −ν(ξ )t̄ν , x ∈ [−1, 1]

ν
t̄⎧ (x , 0, ξ ) = t0 (x, ξ ),
⎪
⎪ 2π
⎪
⎨t̄ν (±1, t , ξ )|±ξ 1 <0 = ρ̄±ν
(t ) M± (185)
 θ±
⎪
⎪ ν
⎪
⎩ρ̄± (t ) ≡ |ξ 1 |t̄ν (±1, t , ξ )dξ.
±ξ 1 >0,ξ ∈R3

The solution t̄ν (x , t , ξ ) can be represented as

t̄ν (x , t , ξ )
⎧
⎪ −ν(ξ )t  ν
⎨e t̄ (x − ξ 1 t , 0, ξ ) if τ (x , t , ξ ) = 0,
= e−ν(ξ )(t −τ (x ,t ,ξ )) ρ− (τ (x , t , ξ ))M− (ξ ) if τ (x , t , ξ ) > 0 and ξ 1 > 0,
⎪
⎩ −ν(ξ )(t −τ (x ,t ,ξ ))
e ρ+ (τ (x , t , ξ ))M+ (ξ ) if τ (x , t , ξ ) > 0 and ξ 1 < 0.
(186)

This system does not satisfy the conservation law, but it satisfies
⎧  1  1
⎪
⎪ d
⎨ t̄ν (x , ξ, t ) dx = − ν(ξ )t̄ν (x , ξ, t ) dx = O(1),
dt
 1 −1 R −1 R (187)
⎪
⎪ ν
⎩ t̄ (x , ξ, 0) dx = 0.
−1 R

This yields that when t = −4/7 /| log |

 1
t̄ν (x , ξ, t ) dx = O(1)3/7 . (188)
−1 R

From this, from fluctuation estimates (182), (184) and from the representation (186),
one has that for t  −4/7 /| log |
⎧ 1/15 
⎨ 1 √
ν M− for |ξ 1 |
,
|t̄ (x , t , ξ )|  O(1)
⎩ 1 +t
(189)
√
ν(ξ ) M− for |ξ 1 |  .

By rescaling the parameters (x , t ) to the original one (x, t), the estimates (189)
concludes this lemma.
 Stochastic Formulation for the Initial-Boundary Value Problems 265

6.2. On the exponential converging property of the solution of a linearized

Boltzmann equation for angular cutoff hard potential collision models

Lemma 14 gives the structure of the solution tν for t  3/7 /| log |. It is con-
venient to denote Ttν [t0 ](x, t, ξ ) in order to relate tν (x, t, ξ ) and the initial data
t0 (x, ξ ):

Ttν [t0 ](x, ξ ) ≡ tν (x, t, ξ ), (190)

where tν is the solution of the transport equation with damping (159).

In this subsection we are interested in the global exponential decay structure in
the solution of a linear Boltzmann equation with total zero mass initial data:

∂⎧t f + ξ 1 ∂x f = Lf, x ∈ [−, ], t > 0,

⎪
⎪ 2π
⎪
⎨f(±, t, ξ )|±ξ 1 <0 = ρ± (t) M± (ξ )|±ξ 1 <0 ,
 θ± (191)
⎪
⎪
⎪
⎩ρ± (t) ≡ |ξ 1 |f(±, t, ξ ) dξ, f(x, 0, ξ ) = I0 (x, ξ ),
±ξ 1 >0,ξ ∈R3
   
 I 
 0 
I0 (x, ξ ) dξ dx = 0,     1.
− R 3  ν M−  L ∞ (L ∞ )
x ξ

Lemma 15. (Local existence) There exists C0 > 0 such that the solution f(x, t, ξ )
of (191) satisfying that for t  3/7 /| log |
⎧  1/15  √
⎨ ν(ξ )  M− (ξ ) for |ξ 1 |
,
+t
|f(x, t, ξ )|  O(1)  (192)
⎩ √
ν(ξ ) M− (ξ ) for |ξ 1 |  .

Remark 5. The construction for this problem is different from the Maxwell mole-
cule collision model. In the Maxwell molecule collision model, the linearized colli-
sion operator is treated as a perturbation of the free transport equation so that the
Picard’s iteration described in (112) works. In contrast to this only K, (L = −ν + K)
can be treated as perturbation of the transport equation with damping −ν(ξ ) ⊗ I .

Proof. Consider the following Picard’s iteration scheme:

for t  3/7 /| log |
⎧
⎨i0 (x, t, ξ ) ≡ Ttν [I0 ](x, ξ ),
 t
(193)
⎩in (x, t, ξ ) ≡ i0 (x, t, ξ ) + Tt−τ
ν [Kin−1 (·, τ, ·)](x, ξ ) dτ for n
1.
0

From Lemma 14, and the zero total mass condition (191) one has that for
t  3/7 /| log |
 1/15  √

M− for |ξ 1 |

,
|i0 (x, t, ξ )|  O(1) +t
√ (194)
ν(ξ ) M− for |ξ |  .
1
266 Shih-Hsien Yu

This and Lemma 3 yield that

 1/15 
|K[i0 ](x, t, ξ )|  O(1) ν(ξ ) M− for t < 3/7 /| log |. (195)
+t

Since the function K[i0 (x, t, ·)](ξ ) does not necessary satisfy the total zero mass
condition, one only can conclude that
 t−τ 
T [Ki0 (·, τ, ·)](x, ξ )
ν
 1/15 
= O(1) ν(ξ ) M− (ξ ) for τ < t < 3/7 /| log |. (196)
+τ
(Note that the right-hand side does depend on t.) This results in
 1/15 
t 
|i0 (x, t, ξ ) − i1 (x, t, ξ )|  O(1) dτ ν(ξ ) M− (ξ )
0 +τ
 1/15 
= O(1)3/7 ν(ξ ) M− (ξ ) for t  3/7 /| log |. (197)
+t

Finally, one can prove that there exists C > 0 such that for t  3/7 /| log |

 1/15 
|in (x, t, ξ ) − in−1 (x, t, ξ )|  C3n/7 ν(ξ ) M− (ξ ) for n
1.
+t
(198)

This concludes that the iteration scheme in (193) converges when t  3/7 /| log |
and the solution f(x, t, ξ ) of (191) satisfies that for t  3/7 /| log |
 1/15  √

ν(ξ ) M− (ξ ) for |ξ 1 |

,
|f(x, t, ξ )|  O(1) +t
√ (199)
ν(ξ ) M− (ξ ) for |ξ |  .
1

Theorem 9. (Exponential decay) The solution f(x, t, ξ ) of (191) satisfies the fol-
lowing global estimates for t > 3/7 /| log |
  √
−ν0 t/2 ν(ξ ) M− (ξ ) for |ξ 1 |  √,
|f(x, t, ξ )|  O(1)e (200)
1/60 ν(ξ ) M− (ξ ) for |ξ 1 |
,
|ρ± (t)|  O(1)1/60 e−ν0 t/2 , (201)

where ν0 ≡ ν(0).

Proof. Similar to the proof of Lemma 111, we need a local existence structure to
construct the global exponential decay structure.
Denote

τ0 ≡ 3/7 /| log |.

 Stochastic Formulation for the Initial-Boundary Value Problems 267

Lemma 15 asserts the local time estimates of f(x, t, ξ ) for t  3/7 /| log | :
  √
ν(ξ ) M− (ξ ) for |ξ 1 |  ,
|f(x, t, ξ )|  O(1)   1/15  √ (202)
+t ν(ξ ) M− (ξ ) for |ξ 1 |
.

Similar to (124), from (202) there exists C0 such that for t ∈ [τ0 /2, τ0 ]
  √
ν(ξ ) M− (ξ ) for |ξ 1 |  √,
|f(x, t, ξ )|  C0 1/60
 ν(ξ ) M− (ξ ) for |ξ 1 |
, (203)
|ρ± (t)|  C0 1/60 .
Similar to (125), one has
 τ0 
f x, t + , ξ
2 
⎧
−ν(ξ )t τ0 
⎪
⎪ e f x − ξ 1 t, , ξ
⎪
⎪  t 2   
⎪
⎪  K (ξ, ξ∗ ) f x − ξ 1 (t − τ ) , τ + τ20 , ξ∗
⎪
⎪ −ν(ξ )(t−τ )
⎪
⎪ + e M− (ξ ) 
⎪
⎪ R M− (ξ∗ )
⎪
⎪
0
⎪
⎪ × dξ dτ if τ (x, t, ξ ) = 0,
⎪ ∗ 
⎪
⎪
⎪ −ν(ξ )(t−τ (x,t,ξ )) τ0 
⎪
⎪ e ρ − τ (x, t, ξ ) + M− (ξ )
⎪
⎨ 2  
 t 
= −ν(ξ )(t−τ )
 K (ξ, ξ∗ ) f x − ξ 1 (t − τ ) , τ + τ20 , ξ∗
⎪ + e M− (ξ ) 
⎪
⎪ τ (x,t,ξ ) R M− (ξ∗ )
⎪
⎪
⎪
⎪ × τ (x, ξ ) > ξ 1 > 0,
⎪
⎪ dξ ∗ dτ if t, 0 and
⎪
⎪ −ν(ξ )(t−τ (x,t,ξ ))
⎪
⎪ e ρ+ (τ (x, t, ξ )) M+ (ξ )
⎪
⎪  t   
⎪
⎪
⎪ −ν(t−τ )
 K (ξ, ξ∗ ) f x − ξ 1 (t − τ ) , τ + τ20 , ξ∗
⎪
⎪ + e M− (ξ ) 
⎪
⎪ τ (x,t,ξ ) R M− (ξ∗ )
⎪
⎩
× dξ∗ dτ if τ (x, t, ξ ) > 0 and ξ 1 < 0.
(204)
√
This, (203), and Lemma 3 together lead to the improved factor 1/60 for |ξ 1 |  
⎧  −ντ /2   √
⎨ e 0 + O(1)τ0 1/60 ν(ξ ) M− (ξ ) for |ξ 1 |  ,
|f(x, τ0 , ξ )|  C0  √
⎩
O(1)1/60 ν(ξ ) M− (ξ ) for |ξ 1 |
.
(205)
Lemma 15 is sufficient to obtain that from the data f(x, τ0 , ξ ) stated in (203) to yield
that
 √
|f(x, τ0 /2, ξ )|  O(1)4/105 ν(ξ ) M− (ξ )  C0 e−ν0 τ0 /4 1/60 for |ξ 1 |
.
(206)
The condition  1, (205), and (206) yield that
  √
ν(ξ ) M− (ξ ) for |ξ 1 |  √,
|f(x, t, τ0 )|  C0 e−ντ0 /4 1/60 (207)
 ν(ξ ) M− (ξ ) for |ξ 1 |
.
268 Shih-Hsien Yu

By repeating the local existence lemma and bootstrap method, one can prove that
for k ∈ N
⎧  √
⎨ ν(ξ ) M− (ξ ) for |ξ 1 |  ,
|f(x, t, (1 + k)τ0 /2)|  C0 e−νkτ0 /4  √ (208)
⎩ 1/60
 ν(ξ ) M− (ξ ) for |ξ 1 |
.

This and the local existence lemma give rise to the theorem.

6.3. On linear and nonlinear stability

Similar to the Maxwell molecule model, SL denotes the linear stationary solu-
tion of ∂t SL + ξ 1 ∂x SL − LSL = 0 for an angular cutoff hard potential model.
t 1 2
Lemma 16. (Linear stationary solution) The limit D(x, ξ ) ≡ limt→∞ 0 St Ls
(x, ξ ) dt exists for all (x, ξ ), and it satisfies
 √
M− (ξ ) for |ξ 1 |  ,
|SL (x, ξ )−s(x, ξ )|  O(1)|θ− −θ+ |  √ (209)
1/60 M− (ξ ) for |ξ 1 |
,
 
SL (x, ξ ) − s(x, ξ ) dξ dx = 0. (210)
− R3

SL = s + D defines a nontrivial stationary solution, where s is given in (103).

The proof of this lemma is identical to that of Lemma 13. The details are omitted.

Lemma 17. (Linear stability) Suppose that

  
J0 (x, ξ ) − s(x, ξ ) dξ dx = 0.
− R3

Then, for t
3/7 /| log |

 t 
S [J0 ](x, ξ ) − D(x, ξ ) − s(x, ξ )  O(1)e−ν0 t/2
  √
ν(ξ ) M− (ξ ) for |ξ 1 |  √,
· 1/60 (211)
 ν(ξ ) M− (ξ ) for |ξ 1 |
.

Proof. This lemma is a consequence of

  
D(x, ξ ) dξ dx = 0
− R3

and Theorem 9.

A control problem, II
 Stochastic Formulation for the Initial-Boundary Value Problems 269

Lemma 18. Let Λ0 ∈ [0, ν0 /4). For the solution operator St of (191), the following
holds
 t
γ
O t (x, ξ ) ≡ St−s I s,Λ0 (x, ξ ) ds
0
 √
−Λ0 t
 1 for |ξ 1 |  √,
 O(1)e M− (ξ ) · 1/60 (212)
 for |ξ 1 |
,

where γ = s−4
s , and where I αs,Λ0 (x, ξ ) is given in (105).

Proof. Denote

ρ±
o
(t) ≡ |ξ 1 |O t (±, ξ ) dξ.
±ξ 1 <0

By Theorem 9 one has that for t
2τ0 , τ0 ≡ 3/7 /| log |,

   
 t t−τ0
γ 
|ρ±
o
(t)| =  + |ξ 1 |St−s I s,Λ0 (±, ξ ) ds dξ 
±ξ 1 >0 t−τ0 0
 t−τ0
= O(1)τ0 e−Λ0 t + 1/60 e−ν0 (t−s)/2 e−Λ0 s ds
0
1/60 e−Λ0 t
= O(1) , (213)
ν0 /2 − Λ0

and
 t  t−τ0
γ
O t (x, ξ ) = O(1) + St−s I s,Λ0 (x, ξ ) ds dξ
t−τ0 0
 √
−Λ0
 1 for |ξ 1 |  √,
= O(1)e t
M− (ξ )·ν(ξ ) 1/60 (214)
 (1 + |ξ |)γ for |ξ 1 |  .
γ
The function O t (x, ξ ) can be represented by ∂t O t + ξ 1 ∂x O t − L O t = I t,Λ0 with
initial data O 0 ≡ 0. This gives rise to the characteristic curve representation
 t  
γ
O t (x, ξ ) = e−ν(ξ )(t−s) K O s + I s,Λ0 (x − ξ 1 (t − τ ), ξ ) ds
τ (x,t,ξ )
⎧
⎪
⎪ 0 if τ (x, t, ξ ) = 0,
⎨
−ν(ξ )(t−τ (x,t,ξ )) ρ− (τ (x,t,ξ ))M− (ξ )
o
+ e √
2π θ−
if ξ 1 > 0, τ (x, t, ξ ) > 0,
⎪
⎪
⎩ e−ν(ξ )(t−τ (x,t,ξ )) ρ+ (τ (x,t,ξ ))M+ (ξ )
o
√
2π θ
if ξ 1 < 0, τ (x, t, ξ ) > 0.
+
(215)

Substitute (214) and (213) together with ν(ξ ) = O(1)(1+|ξ |)γ into (215) to gain the
extra (1 + |ξ |)−γ when integrating out the time variable. Then this lemma follows.
270 Shih-Hsien Yu

We will follow almost the same procedure as in Section 5.3 for constructing the
stationary solution for the Maxwell molecule model to build the nonlinear stationary
solution φ for this angular cutoff hard potential model. Recall the iteration scheme
(143)
⎧  ∞
⎪
⎨ Dd (x, ξ ) =
0 St Q(SL − M− )(x, ξ ) dt,
0 ∞   (216)
⎪
⎩ Dn (x, ξ ) =
d St Q SL − M− + Ddn−1 (x, ξ ) dt for n
1,
0

and St and Q stand for the solution operator of (191) and the collision operator of
the angular cutoff hard potential model.

Theorem 10. (Nonlinear stationary solution) When |θ− − θ+ | and  are sufficiently
small, there is a nontrivial nonlinear stationary solution φ for the angular cutoff
hard potential model. Furthermore,
 √
 1 for |ξ 1 |  √,
|φ(x, ξ ) − s(x, ξ )|  O(1)|θ+ − θ− | M− (ξ ) 1/40
 for |ξ 1 |
,

and the solution φ(x, ξ ) is nonlinearly exponentially time-asymptotically stable.

Proof. By (209) and by Lemma 1, one has that

 
 Q(S − M ) 
 L − 
    O(1)|θ− − θ+ |2 . (217)
 ν(ξ ) M− (ξ )  ∞ ∞
L (L )
x ξ

(217) and Lemma 18 with Λ0 = 0 lead to

 ∞
St Q(SL − M− )(x, ξ ) dt
0
 √
 1 for |ξ 1 |  √,
= O(1)|θ− − θ+ |2 M− (ξ ) · 1/60 (218)
 for |ξ 1 |
.

This leads to D0d in (216) satisfying that

 
 D0 
 d 
  = O(1)|θ− − θ+ |2 . (219)
 M−  ∞
L∞
x (L ξ )

We may assume that

 
 Dn 
 d 
   O(1)|θ− − θ+ | for n
0. (220)
 M−  ∞
L∞
x (L ξ )
 Stochastic Formulation for the Initial-Boundary Value Problems 271

Under the assumption (220), from Lemma 1

 
 Q(S − M + Dn ) − Q(S − M + Dn−1 ) 
 L − L − 
 d
 d

 ν M−  ∞ ∞
L x (L ξ )
 
 B(S − M + Dn + Dn−1 , Dn − Dn−1 ) 
 L − 
= d
 d d d

 ν M−  ∞ ∞
L x (L ξ )
 
 Dn − Dn−1 
 
 O(1)|θ+ − θ− |  d  d  . (221)
 M−  L ∞ (L ∞ )
x ξ

Again (221) and Lemma 18 with Λ0 = 0 lead to

 ∞ 
 
 St
(Q(S − M − + Dn
) − Q(S − M − + Dn−1
)) dt 
 0 L d L d 
  
 
 M− 
  ∞ ∞
L x (L ξ )
 
 Dn − Dn−1 
 
= O(1)|θ− − θ+ |  d  d  , (222)
 M−  L ∞ (L ∞ )
x ξ

and
  
∞ Dn + Ddn−1 d
(Ddn+1 − Dnd )(x, ξ )
= S 2B s + d − M− + d
t
, Dn − Ddn−1 dt
0 2
   √
   1 for |ξ 1 |  √,
= O(1)|θ− −θ+ | Dnd −Ddn−1  ∞ ∞ M− (ξ ) · 1/60
L x (L ξ )  for |ξ 1 |  .
(223)

Thus the sequence {(Dnd −Ddn−1 )/ M− }n∈N is a geometric sequence with a contract
rate O(1)|θ− − θ+ | in the norm · L ∞ ∞ under the assumption (220). This
x (L ξ )
geometric sequence, (219), and the assumption (220) lead to
   
 Dn   D0 
 d   d 
   (1 + O(1)|θ− − θ+ |)    = O(1)|θ− − θ+ |2 .
 M−  ∞ ∞  M−  ∞ ∞
L (L )
x ξ L (L ) x ξ
(224)

Thus, the assumption (220) is valid for all n ∈ N as well and it follows
∞ 
(Dnd − Ddn−1 ) + D0d (x, ξ )
n=1
 √
 1 for |ξ 1 |  ,
= O(1)|θ− − θ+ | M− (ξ ) ·
2
√ (225)
1/60 for |ξ 1 |
.
272 Shih-Hsien Yu

This concludes the existence of a nontrivial nonlinear stationary solution

φ = (SL + M− + limn→∞ Dnd and it satisfies
 √
M− (ξ ) for |ξ 1 |  ,
|φ(x, ξ ) − SL (x, ξ )|  O(1)|θ− −θ+ | 2
 √ (226)
1/60 M− (ξ ) for |ξ 1 |
.

Theorem 11. (Nonlinear stability) The nonlinear stationary solution φ(x, ξ )

constructed in (10) is nonlinearly stable in the sense that the initial data F(x, 0, ξ )
of (1) satisfying that
 
 F(x, 0, ξ ) − φ(x, ξ ) 
 
   1.
 M−  ∞ ∞
L (L )
x ξ

If the initial data F(x, 0, ξ ) satisfies that

⎧ 
⎪ 
⎪
⎪
⎨− R3 (φ(x, ξ ) − F(x,
⎪

0, ξ )) dξ dx = 0,
 F(x, 0, ξ ) − φ(x, ξ )  (227)
⎪
⎪ 

  ε,
⎪
⎪  ∞ ∞
⎩ M− (ξ ) L (L )
x ξ

then

|F(x, t, ξ ) − φ(x, ξ )|  O(ε)e−ν0 t/8 M− (ξ )
⎧ 1/60
⎨ for |ξ 1 | > 1/2 , t > 3/7 /| log |
× 1 for |ξ 1 | < 1/2
⎩
1 for t  3/7 /| log |.

This theorem is the angular cutoff hard potential model version for Theorem 8
including Corollary 1. The proof of this theorem is a combination of that in Theo-
rem 8 with a slight modification from Theorem 10. The details are omitted.

Proof of Theorem 3. Lemma 6 concludes this theorem for Maxwell molecule

model; and Lemma 17 concludes this theorem for the hard potential model.

Proof of Theorem 4. Theorem 8 and Corollary 1 conclude this theorem for Maxwell
molecule model, and Theorem 11 concludes the theorem for the hard potential
model.

Acknowledgments. The author would like to thank Professor Yoshio Sone for his interest
in this work. This research is supported by the National University of Singapore start-up
research grant R-146-000-108-133.
 Stochastic Formulation for the Initial-Boundary Value Problems 273

References

1. Aoki, K., Cercignani, C.: Evaporation and condensation on two parallel plates at finite
Reynolds numbers. Phys. Fluids 26, 1163–1164 (1983)
2. Aoki, K.: Introduction (Yashio Sone). Bull. Inst. Math. Acad. Sin. (NS) 2(4), 805–822
(2007)
3. Arkeryd, L., Cercignani, C.: A global existence theorem for the initial-boundary
value problem for the Boltzmann equation when the boundaries are not isothermal.
Arch. Rat. Mech. Anal. 125, 271–287 (1993)
4. Arkeryd, L., Cercignani, C., Illner, R.: Measure solutions of the steady Boltzmann
equation in a slab. Comm. Math. Phys. 142, 285–296 (1991)
5. Arkeryd, L., Maslova, N.: On diffuse reflection at the boundary for the Boltzmann
equation and related equations. J. Stat. Phys. 77 (1994)
6. Arkeryd, L.: Nouri, A compactness result related to the stationary Boltzmann equation
in a slab, with applications to the existence theory. Ind. Univ. Math. J. 44(Nb3), 815–839
(1995)
7. Arkeryd, L.: Nouri, Boltzmann asymptotics with diffuse reflection boundary condi-
tions. Monatsh. Math. 123(4), 285–298 (1997)
8. Bourgain, J., Golse, F., Wennberg, B.: On the distribution of free paths lengths in
the periodic Lorentz gas. Commun. Math. Phys. 190, 491–508 (1998)
9. Caflisch, R.E.: The Boltzmann equation with a soft potential, I and II. Commun. Math.
Phys. 74, 71–95, 97–109 (1980)
10. Caflisch, R.E., Nicolaenko, B.: Shock profile solutions of the Boltzmann equation.
Commun. Math. Phys. 89, 161–194 (1982)
11. Caprino, S., Cavallaro, G., Marchioro, C.: On a microscopic model of viscous
friction. Math. Models Methods Appl. Sci. 17(9), 1369–1403 (2007)
12. Caprino, S., Marchioro, C., Pulvirenti, M.: Approach to equilibrium in a micro-
scopic model of friction. Commun. Math. Phys. 264(1), 167–189 (2006)
13. Cercignani, C.: The Boltzmann Equation and its Applications. Springer, Berlin, 1988
14. Cercignani, C., Illner, R., Pulvirenti, M.: The mathematical theory of dilute gases.
Applied Mathematical Sciences, Vol. 106. Springer, New York, 1994
15. Chen, C.-C., Chen, I.-K., Liu, T.-P., Sone, Y.: Thermal transpiration for the linearized
Boltzmann equation. Commun. Pure Appl. Math. 60(2), 147–163 (2007)
16. Chung, K.-L.: A Course in Probability Theory. Academic Press, New York, 1974
17. Desvillettes, L., Villani, C.: On the trend to global equilibrium for spatially inhomo-
geneous kinetic systems: the Boltzmann equation. Invent. math. 159, 245–316 (2005)
18. Golse, F.: Golse’s theorem on a one-way flow. Appendix A.4 in Kinetic Theory and
Fluid Dynamics, 1st edn, August 12. (Ed. Sone Y.) Birkhauser, Boston, 2002
19. Golse, F.: The periodic Lorentz gas in the Boltzmann-Grad limit. Proceedings Inter-
national Congress of Mathematicians, Madrid, 2006, vol. 3. (Eds. M. Sanz-Solé,
J. Soria, J.-L. Varona and J. Verdera) European Math. Soc., Zürich, 183–201, 2006
20. Golse, F., Caglioti, E.: On the distribution of free paths lengths in the periodic Lorentz
gas III. Commun. Math. Phys. 236, 199–221 (2003)
21. Grad, H.: Asymptotic theory of the Boltzmann equation, II. Int. Symp. on Rarefied Gas
Dynamics, Third Symp., 25–59 (1962)
22. Grad, H.: Asymptotic equivalence of the Navier-Stokes and nonlinear Boltzmann equa-
tions. Proc. Symp. Appl. Math., Vol. 17 (Ed. R. Finn), AMS, Providence, 154–183, 1965
23. Grad, H.: Proc. Symp. in Appl. Math. 17, 154–183 (1965), Rarified Gas Dynamics, I,
26–59 (1963)
24. Hamdache, K.: Weak solutions of the Boltzmann equation. Arch. Rat. Mech. Anal. 119,
309–353 (1992)
25. Hilbert, D.: Grundzüge einer Allgemeinen Theorie der Linearen Integralgleichungen,
Chap. 22, Teubner, Leipzig
274 Shih-Hsien Yu

26. Liu, T.-P., Yu, S.-H.: The Green’s function and large-time behavior of solutions for
one-dimensional Boltzmann equation. Commun. Pure. Appl. Math. 57l(12), 1543–1608
(2004)
27. Liu, T.-P., Yu, S.-H.: Green’s function of Boltzmann equation, 3-D waves. Bull. Inst.
Math. Acad. Sin. (NS) 1(1), 1–78 (2006)
28. Liu, T.-P., Yu, S.-H.: Initial-boundary value problem for one-dimensional wave solu-
tions of the Boltzmann equation. Commun. Pure Appl. Math. 60(3), 295–356 (2007)
29. Maxwell, J.C.: On stresses in rarefied gases arising from inequalities of temperature.
R. Soc. Phil. Trans. 27, 304–308 (1878)
30. Niimi, H.: Thermal creep flow of rarefied gas between two parallel plates. J. Phys. Soc.
Jpn 30, 572–574 (1971)
31. Pao, Y.P.: Application of kinetic theory to the problem of evaporation and condensation.
Phys. Fluids 14, 306–312 (1971)
32. Sone, Y.: Thermal creep in rarefied gas. J. Phys. Soc. Jpn 21, 1836–1837 (1966)
33. Sone, Y.: Flow induced by thermal stress in rarefied gas. Phys. Fluids 15, 1418–1423
(1972)
34. Sone, Y.: Kinetic Theory and Fluid Dynamics, 1st edn. August 12, Birkhauser, Boston,
2002
35. Sone, Y.: Molecular Gas Dynamics: Theory, Techniques, and Applications. Birkhauser,
Boston, 2007
36. Sone, Y., Yamamoto, K.: Flow of rarefied gas through a circular pipe. Phys Fluids 11,
1672–1678 (1968) Erratum: Phys Fluids 13, 1651 (1970)
37. Ukai, S.: On the existence of global solutions of mixed problem for non-linear Boltz-
mann equation. Proc. Jpn Acad. 50, 179–184 (1974)

Department of Mathematics,
National University of Singapore,
Singapore, Singapore.
e-mail: [email protected]

(Received December 5, 2007 / Accepted March 4, 2008)

Published online July 5, 2008 – © Springer-Verlag (2008)