Lectures Notes

on
Nonlinear Diffusion Equations
by
Yin Jingxue

September, 2002
Contents

1 Degenerate Parabolic Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Classical Diffusion Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Nonlinear Diffusion Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Other Special Solutions of Porous Medium Equation . . . . . . . . . . . . . 3
1.4 Degenerate Parabolic Equations . . . . . . . . . . . . . . . . . . . . . . . . 3
1.5 Definitions of Generalized Solutions . . . . . . . . . . . . . . . . . . . . . . 4
2 Existence of Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1 Weakly Degenerate Equations without Lower Order Terms . . . . . . . . . 8
2.2 Weakly Degenerate Equations with Lower Order Terms . . . . . . . . . . . 11
2.3 Strongly Degenerate Equations without Terms of Lower Order . . . . . . . 17
2.4 General Strongly Degenerate Equations . . . . . . . . . . . . . . . . . . . . 19
3 Uniqueness of Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1 Comparison Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 Testing Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3 Holmgren’s Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4 Approximate Holmgren’s Approach . . . . . . . . . . . . . . . . . . . . . . . 32
3.5 Approach by Characteristic Operators . . . . . . . . . . . . . . . . . . . . . 38
3.6 Approach by the Theory of BV Spaces . . . . . . . . . . . . . . . . . . . . . 42
4 Finite Propagation of Disturbances . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1 Comparison Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 Approach by Moser Iteration . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.3 Approach by Energy Estimates . . . . . . . . . . . . . . . . . . . . . . . . . 56
5 Free Boundary of Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.1 Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.2 Entropy Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3 A Free Boundary Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6 Properties of Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.1 Regularity of Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2 Localization and Extinction of Disturbances . . . . . . . . . . . . . . . . . . 92
6.3 Asymptotic Behaviors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
7 Appendix Classes BV and BVx . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

i
Nonlinear Diffusion Equations 1

1 Degenerate Parabolic Equations

1.1 Classical Diffusion Equations
Many diffusion processes can be described by evolutionary partial differential equations. In this
section, we consider the mathematical formulation of the heat conduction of a body. Suppose that
there is a body with sufficiently large volume, which we may say that it occupies the whole space
RN , u(x, t) denotes the temperature distribution of the body at the location x and at the time t.
Conservation law gives
∂u →
+ div J = 0,
∂t
→
where J denotes the heat flux. In general, the heat flux depends on the temperature distribution
and its derivatives. Since the temperature always propagates along the fast descending direction,
that is the direction along −∇u, a simple model for the homogeneous diffusion can be formulated
by
→
J = −a∇u,

where a is a positive constant, called the diffusion coefficient. If we assume that a = 1, then the
temperature distribution satisfies

∂u
= ∆u, (x, t) ∈ RN × (0, T ), (1.1)
∂t
u(x, 0) = u0 (x), x ∈ RN . (1.2)

Here, the equation (1.1) is the standard classical heat equation.

For a special initial datum u0 ∈ C0 (RN ) with u0 (x) ≥ 0, that is, heating the body locally at
the initial time, then by the Gauss kernel
 
1 |x|2
E(x, t) = √ exp − ,
(2 πt)N 4t

we see that the distribution can be expressed explicitly

u(·, t) = E(·, t) ∗ u0 (·),

that is Z
u(x, t) = E(x − y, t)u0 (y)dy.
RN

Then we have

Proposition 1.1 Let u0 ∈ C0 (RN ) with u0 (x) ≥ 0 and u0 (x) 6≡ 0, and u be the solution of the
problem (1.1)–(1.2). Then
u(x, t) > 0, ∀x ∈ RN , t > 0.

From the point of view of physics, this shows that when heating a local part of a sufficiently
large body initially, the heat flux would propagate to the whole body instantaneously, which is
impossible for the physical reality. So, the classical model has some disadvantages and inaccuracy
for describing the heat conduction.
2 Degenerate Parabolic Equations

1.2 Nonlinear Diffusion Equations

For the above mentioned classical model for describing the heat conduction, the inaccuracy lies
in that we do not consider the complicated dependence of the heat flux with the temperature
distribution and its derivatives. If we assume that the heat conduction coefficient depends on the
distribution in a manner as a = a(u), then the equation becomes
∂u
= div(a(u)∇u), (1.3)
∂t
which has been always adopted for relatively slow diffusion processes. In particular, if we consider
the coefficient a(u) is a power-like function, that is, a(u) = mum−1 for some m > 1, then the
equation becomes
∂u
= ∆um . (1.4)
∂t
This is the famous porous medium equation, which is proposed when discussing the diffusion
processes of flows through the porous medium. If u = 0, then the equation is not of parabolic type.
However, whether or not the above equation can accurately describe the temperature distribu-
tion than the heat equation? To see this, let us first consider a class of special solution for this
kind of equation, that is the Barenblatt solution:
  !1/(m−1)
−k k(m − 1) |x|2
Bm (x, t) = t 1− ,
2mN t2k/N +
 −1
2
where k = m−1+ , u+ = max(u, 0). If we take t = 1 as the initial time, then the value
N
of Bm (x, t) is
  !1/(m−1)
k(m − 1) 2
Bm (x, 1) = 1− |x| ,
2mN +

which has the following properties:

Bm (x, 1) > 0 for |x| < αm ,
Bm (x, 1) = 0 for |x| ≥ αm ,
r
2mN
where αm = . Here the support of Bm is suppBm = Bαm (0), while the support of
k(m − 1)
solution is Bαm (t) (0), with the radius αm (t) = αm tk/N , which is increasing as the time increases.
Still from the point of view of physics, this shows that when heating a body locally, the heat
flux will propagates in a finite speed. So, to some extend, this model reflex a special manner of
physics reality. The above mentioned property will be called the Finite Speed of Propagation of
Perturbations.
It is worth mentioned that Bm is not the usual solution of the equation (1.1), which should
be explained in detail later. Before we give an exact explanation, we first take a look at some
properties of the function Bm :
1. In the region {(x, t); Bm (x, t) > 0}, Bm (x, t) satisfies (1.1) classically;
∂ m
2. On the lateral boundary |x| = αm tk/N , Bm (x, t) = 0, and B (x, t) = 0(i = 1, · · · , N );
∂xi m
∂
3. The derivative Bm (x, t)(i = 1, · · · , N ) may not be existent on the lateral boundary
∂xi
k/N
|x| = αm t . Even if it is existent, it may not be continuous.
Nonlinear Diffusion Equations 3

If we consider the relatively fast flows, then there is mode taking the form
→
J = −|∇u|p−2 ∇u, (p > 2)

and the equation becomes the well-known p-Laplacian equation

∂u
= div(|∇u|p−2 ∇u).
∂t

1.3 Other Special Solutions of Porous Medium Equation

1. Pressure solution of degree two
 1/(m−1)
T0 |x|2
u(x, t) = ,
T0 − t
m−1
where T0 = . Its initial value is
2m(2 + N (m − 1))

u(x, 0) = |x|2/(m−1) .

In fact, this function is a classical solution of the equation (1.1) in QT0 ≡ RN × T0 , by a

direct calculation. By this special solution, we see that when heating a body with large initial
temperature, then there may appear the Blow-up phenomena.
2. Pressure solution of degree one (N = 1)
 1/(m−1)
m−1
u(x, t) = c (ct ± x)+ ,
m
where c > 0 is an arbitrary constant. Its initial value is
 1/(m−1)
m−1
u(x, 0) = (±cx)+ .
m
Such solution is quite similar to the travelling wave solutions for the wave equation.

1.4 Degenerate Parabolic Equations

Consider the equation
∂u →
= ∆A(u) + divB(u), (1.5)
∂t
→
where A(u) ∈ C 1 (R), B(u) ∈ C 1 (R, RN ), satisfying A0 (u) ≥ 0.
If A0 (u) > 0, then the equation is of parabolic type, while if A0 (u) ≥ 0, it becomes hyperbolic
type at the points where A0 (u(x, t)) = 0. We call this equation degenerates in such points, and
call the equation is a degenerate parabolic equation. Consider the set

E = {s; A0 (s) = 0}.

Then if u(x, t) ∈ E, then the equation degenerates at the point (x, t). So, the set E is called to
be the degenerate set. As for the porous medium equation (1.1), the degenerate set has only one
point E = {0}.
4 Degenerate Parabolic Equations

The degenerate set affects the properties of solutions, especially the regularity. We will distin-
guish two kind of degeneracies. If the set has no interior point, then we call the equation is of
weakly degenerate type. The degenerate set of the porous medium equation contains one point,
and hence is weakly degenerate. However, for the weakly degenerate equation, the set may have
infinite points, even with the positive measure. If the degenerate set has interior point, then the
equation is called to be strongly degenerate. Suppose that E ⊃ [α, β], and u ∈ [α, β], then the
equation becomes
∂u →
= divB(u). (1.6)
∂t
This is a first order equation, a hyperbolic equation. Since the hyperbolic equation may admit
discontinuous solutions, the strongly parabolic equations have discontinuous solutions too. For the
weakly degenerate equations the solutions, may have good regularity than the stronger one.

1.5 Definitions of Generalized Solutions

Since quasilinear parabolic equations with degeneracy do not have classical solutions in general, it
is necessary to generalize the notion of solutions.
Consider the Cauchy problem
∂u
= ∆A(u), (x, t) ∈ QT ≡ RN × (0, T ), (1.7)
∂t
u(x, 0) = u0 (x), x ∈ RN , (1.8)
where A(u) ∈ C 1 (R), satisfying A0 (u) ≥ 0.
Let G be a subdomain of QT . If u is a classical solution of the equation (1.7) on G, then all
the derivatives appeared in the equation are the usual derivatives, and the equation holds in the
usual sense too. So, for any ϕ ∈ C0∞ (G), we have
ZZ ZZ
∂u
· ϕdxdt = ∆A(u) · ϕdxdt.
G ∂t G

Integrating by parts, we see that

ZZ ZZ
∂ϕ
u dxdt = − A(u)∆ϕdxdt.
G ∂t G

However, the function u satisfying the above integral equality needs not to be having the above
mentioned smoothness, and even having no derivatives. It is enough for u to be locally integrable.
So, we introduce the following
Definition 1.1 A function u is called a generalized solution of the equation (1.7) on G, if u,
A(u) ∈ L1loc (G) and u satisfies
ZZ  
∂ϕ
u + A(u)∆ϕ dxdt = 0
G ∂t
for any ϕ ∈ C0∞ (G).
∂A(u)
Obviously, if u is a generalized solution of (1.7) on G, and the weak derivatives ∈ L1loc (G)
∂xi
(i = 1, 2, · · · , N ), then the above integral equality can be transformed to the form
ZZ  
∂ϕ
u − ∇A(u) · ∇ϕ dxdt = 0.
G ∂t
Nonlinear Diffusion Equations 5

Definition 1.2 A function u is said to be a generalized solution of the Cauchy problem (1.7),
(1.8) on QT , if u is a generalized solution of the equation (1.7) on QT and u satisfies
Z Z
lim u(x, t)h(x)dx = u0 (x)h(x)dx
t→0+ RN RN

for any h ∈ C0∞ (RN ).

If u, A(u) ∈ L1loc (QT ) (we mean u, A(u) ∈ L1 (G) for any bounded subdomain G of QT .), then
the above definition is equivalent to the following
Definition 1.3 A function u is called a generalized solution of the Cauchy problem (1.7), (1.8)
on QT , if u, A(u) ∈ L1loc (QT ) and u satisfies
ZZ   Z
∂ϕ
u + A(u)∆ϕ dxdt + u0 (x)ϕ(x, 0)dx = 0.
QT ∂t RN

for any ϕ ∈ C ∞ (QT ),which vanishes for large |x| and t = T .

∂A(u)
If the weak derivatives ∈ L1loc (QT ), then the above integral equality can be transformed
∂xi
to the form ZZ   Z
∂ϕ
u − ∇A(u) · ∇ϕ dxdt + u0 (x)ϕ(x, 0)dx = 0.
QT ∂t RN
Now we define generalized solutions of the boundary value problem.
Let Ω ⊂ RN be a domain with appropriately smooth boundary ∂Ω. Denote ΩT = Ω × (0, T ),
Γ = ∂Ω × (0, T ). Consider the first boundary value problem for the equation (1.7) on Ω T , with
boundary value condition
u(x, t)|Γ = g(x, t) (1.9)
and initial value condition
u(x, 0) = u0 (x), x ∈ Ω. (1.10)
Definition 1.4 A function u is called a generalized solution of the boundary value problem
(1.7), (1.9), (1.10) on ΩT , if u, A(u) ∈ L1 (ΩT ) and u satisfies
ZZ   Z Z
∂ϕ ∂ϕ
u + A(u)∆ϕ dxdt + u0 (x)ϕ(x, 0)dx − A(g) dS = 0, (1.11)
Ωτ ∂t Ω Γ ∂n

for any ϕ ∈ C ∞ (ΩT ) which vanishes for (x, t) ∈ Γ, t = T and large |x|, where n denotes the
outward normal to Γ.
Definition 1.5 A function u is called a generalized solution of the boundary value problem
∂A(u)
(1.7), (1.9), (1.10) on ΩT , if u, A(u) ∈ L1 (ΩT ), ∈ L1 (ΩT ), and u satisfies
∂xi
ZZ   Z
∂ϕ
u + A(u)∆ϕ dxdt + u0 (x)ϕ(x, 0)dx = 0, (1.12)
Ωτ ∂t Ω

for any ϕ ∈ C ∞ (ΩT ) which vanishes when |x| is large enough, t = T and in a neighborhood of Γ,
lim u(y, t) exists for almost all (x, t) ∈ Γ, and
and u has boundary value u|Γ , namely, the limit y→x
y∈Ω

lim u(y, t) = g(x, t).

y→x
(1.13)
y∈Ω
6 Degenerate Parabolic Equations

∂A(u)
It is not difficult to check that if ∈ L1 (ΩT ) (i = 1, 2, · · · , N ), then Definition 1.4 is
∂xi
equivalent to Definition 1.5.
∂A(u)
To prove, we first notice that the integrability of implies that A(u) and u have boundary
∂xi
values on Γ, in the sense of trace.
Let u be the generalized solution in Definition 1.5, then from (1.12), we first have
ZZ   Z
∂ϕ
uξ + A(u)∆(ϕξ) dxdt + u0 (x)ϕ(x, 0)ξ(x)dx = 0,
ΩT ∂t Ω

where ϕ is the test function in Definition 1.4, ξ ∈ C0∞ (Ω) satisfies

C
0 ≤ ξ ≤ 1, |∇ξ(x)| ≤ , and ξ(x) = 1 if x ∈ Ωε .
ε
Here, C is a constant independent of ε > 0, Ωε = {x ∈ Ω; dist(x, ∂Ω) ≥ ε}. Letting ε → 0, and
noticing that
ZZ
A(u)∆(ϕξ)dxdt
ΩT
ZZ ZZ
=− ξ∇A(u) · ∇ϕdxdt − ϕ∇A(u) · ∇ξdxdt
ΩT ΩT
ZZ
→− ∇A(u) · ∇ϕdxdt, (ε → 0),
ΩT

we see that
ZZ   Z
∂ϕ
u − ∇A(u) · ∇ϕ dxdt + u0 (x)ϕ(x, 0)dx = 0.
QT ∂t RN

which implies (1.11) by integrating by parts.

Conversely, let u be the generalized solution in Definition 1.4, then (1.12) is valid obviously.
Let ϕ be the test function in Definition 1.4, and ϕ|t=0 = 0. In (1.11), decomposing ϕ into ϕ1 + ϕ2 ,
where ϕ1 = ϕξ, ϕ2 = ϕ(1 − ξ), and using (1.11), we have
ZZ   Z
∂ϕ2 ∂ϕ
u + A(u)∆ϕ2 dxdt − A(g) dS = 0.
ΩT ∂t Γ ∂n
Integrating by parts, and letting ε → 0, we then obtain
Z
∂ϕ
(A(u)|Γ − A(g)) dS = 0.
Γ ∂n
Due to the arbitrariness of ϕ, we see that (1.13) holds.
The following assertion is valid obviously.

Proposition 1.2 Let u be a generalized solution of the Cauchy problem (1.7), (1.8), and
∂A(u)
∈ L1loc (QT ), (i = 1, 2, · · · , N ). Then for any smooth domain Ω ⊂ RN , u is a generalized
∂xi
solution of the boundary value problem for the equation (1.7) on Ω T = Ω × (0, T ) with boundary
value
g(x, t) = u(x, t)|Γ
and initial value u0 (x).
Nonlinear Diffusion Equations 7

Remark 1.1 Sometimes we need to use the notion of generalized super (sub)-solutions. To de-
fine the generalized super-solutions(sub-solutions), it suffices to replace ”=” in the integral equality
by ”≤” (”≥”) and require the test function ϕ to be nonnegative. Similarly, we may define the
generalized super(sub)-solutions for the Cauchy problem and the boundary value problem.

Remark 1.2 If u is a generalized solution of the Cauchy problem (1.7), (1.8), then for any
τ ∈ (0, T ) there holds
ZZ   Z
∂ϕ
u + A(u)∆ϕ dxdt − u(x, τ )ϕ(x, τ )dx
Qτ ∂t RN
Z
+ u0 (x)ϕ(x, 0)dx = 0, (1.14)
RN

for any ϕ ∈ C ∞ (QT ) which vanishes when |x| is large enough, where Qτ = RN × (0, τ ).

To prove, taking an arbitrary function η ∈ C ∞ [0, T ] with η(t) = 1 for t ∈ [0, τ − ε], and
η(t) = 0 for t ∈ [τ, T ], and choosing the test function as ϕη in the integral equality of the definition
of generalized solutions, we obtain
ZZ   ZZ
∂ϕ
η u + A(u)∆ϕ dxdt + uϕη 0 dxdt (1.15)
Qτ ∂t Qτ
Z
+ u0 (x)ϕ(x, 0)dx = 0. (1.16)
RN

Letting ε → 0, we obtain (1.14).

Remark 1.3 Similar conclusion (1.14) can be obtained for the boundary value problem.
8 Existence of Solutions

2 Existence of Solutions
2.1 Weakly Degenerate Equations without Lower Order Terms
In this section, we study the Cauchy problem for the filtration equation in one dimension without
lower order terms:
∂u ∂ 2 A(u)
= , (2.1)
∂t ∂x2
u(x, 0) = u0 (x), (2.2)

where u0 (x) ≥ 0 is a locally integrable function on R and A(s) ∈ C 1 [0, ∞) satisfies the conditions

A(s) > 0, A0 (s) > 0 for s > 0, A(0) = A0 (0) = 0. (2.3)

Under these conditions, the equation is weakly degenerate, and the degenerate set contains only
one point.
Denote QT = R × (0, T ).

Theorem 2.1 Assume that u0 is a nonnegative, continuous and bounded function on R, A(u0 )
satisfies the Lipschitz condition, A(s) is appropriately smooth and A(s) → +∞ as s → +∞. Then
for any T > 0 the Cauchy problem (2.1), (2.2) admits a continuous and bounded generalized
∂A(u)
solution u on QT such that is bounded. Moreover, the solution u is classical in the domain
∂x
{(x, t) ∈ QT , u(x, t) > 0}.

Proof. Denote v0 = A(u0 ) and choose a sequence of smooth functions {v0n (x)} such that
{v0n (x)} uniformly converges to v0 (x) as n → ∞ and

d n
v (x) ≤ K0 , 0 < v0n (x) ≤ M (n = 1, 2, · · · )
dx 0

with some constants K0 and M . Construct a sequence of smooth functions {wn (x)} such that


 wn (x) = v0n (x), |x| ≤ n − 2,




 wn (x) = M,
 |x| ≥ n − 1,

0 < wn (x) ≤ M, (2.4)







 d
 wn (x) ≤ N = max{K0 , M }, (n = 1, 2, · · · ).
dx

Now denote v = A(u), Φ(v) = A−1 (v). Then (2.1) becomes

∂v ∂2v
Φ0 (v) = . (2.5)
∂t ∂x2
Consider the initial-boundary value problem for (2.5) with conditions

v(x, 0) = wn (x), v(±n, t) = M, (2.6)

which admits a classical solution vn (x, t) on Gn = (−n, n) × (0, T ) by virtue of the standard theory
for parabolic equations.
Nonlinear Diffusion Equations 9

From the maximum principle for classical solutions, we have

0 < inf wn (x) ≤ vn (x, t) ≤ M. (2.7)

We will further prove that

∂vn
≤N on Gn . (2.8)
∂x
∂vn
For this purpose, let Pn = which satisfies
∂x
∂Pn ∂ 2 Pn 1 ∂ 0 ∂Pn
Φ0 (vn ) = − 0 Φ (vn ) .
∂t ∂x2 Φ (vn ) ∂x ∂x

The maximum principle shows that

∂vn ∂vn
max ≤ max ,
Gn ∂x Γn ∂x

∂vn
where Γn = ∂p Gn denotes the parabolic boundary of Gn . Since = |wn0 (x)| ≤ N , to prove
∂x t=0
∂vn
(2.8) it suffices to prove ≤ N.
∂x x=±n
Notice that vn achieves its maximum M on the lateral boundary x = n, 0 ≤ t ≤ T . Hence

∂vn
≥ 0. (2.9)
∂x x=n

Consider the auxiliary function

zn = vn − M (x − n + 1),
which satisfies
∂zn ∂ 2 zn
Φ0 (vn ) = , on Dn = {n − 1 < x < n, 0 < t < T }
∂t ∂x2
zn (x, 0) = wn (x) − M (x − n + 1) = M (n − x) ≥ 0 on [n − 1, n]

zn (n, t) = 0, zn (n − 1, t) = vn (n − 1, t) > 0 on (0, T ].

zn achieves its minimum min zn on x = n, 0 ≤ t ≤ T . Hence

Dn

∂zn ∂vn
= − M ≤ 0,
∂x x=n ∂x x=n

∂vn
which combining with (2.9) yields 0 ≤ ≤ M ≤ N.
∂x x=n
∂vn
Similarly we can prove ≤ M.
∂x x=−n
The estimate (2.8) implies the uniform Lipschitz continuity of vn in x: for any (x, t), (y, t) ∈ QT )
and n large enough such that (x, t), (y, t) ∈ Gn ),

|vn (x, t) − vn (y, t)| ≤ N |x − y|. (2.10)

10 Existence of Solutions

Denote un = A−1 (vn ). Note that the assumption A(s) → ∞ as s → ∞ and A0 (s) > 0 imply
the existence of the inverse function A−1 (vn ). From (2.5) we have

∂un ∂ 2 vn
= . (2.11)
∂t ∂x2

For any (x, t), (y, t) ∈ QT ), choose n large enough such that (x, t), (y, t) ∈ Gn ), x+|∆t|1/2 ∈ [−n, n]
with ∆t = t − s. Integrating (2.11) yields
Z x+|∆t|1/2
(un (z, t) − un (z, s))dz
x
Z t Z x+|∆t|1/2
∂un
= dydτ
s x ∂t
Z tZ x+|∆t|1/2
∂ 2 vn
= dydτ
∂x2
Zs t x 
∂v ∂vn
= (x + |∆t|1/2 , τ ) − (x, τ ) dτ.
s ∂x ∂x

Hence from (2.8) we obtain

Z x+|∆t|1/2
(un (z, t) − un (z, s))dz ≤ 2N |∆t|.
x

Using the mean value theorem for integrals, we see that there exists x∗ ∈ [x, x + |∆t|1/2 ] such that
Z x+|∆t|1/2
(un (z, t) − un (z, s))dz = (un (x∗ , t) − un (x∗ , s))|∆t|1/2 .
x

Thus
|un (x∗ , t) − un (x∗ , s)| ≤ 2N |∆t|1/2
and hence for some constant C,

|vn (x∗ , t) − vn (x∗ , s)| = |A(un (x∗ , t)) − A(un (x∗ , s))|
=|A0 (ξn )||un (x∗ , t) − un (x∗ , s)| ≤ C|∆t|1/2 . (2.12)

Combining (2.10) with (2.12) deduces

|vn (x, t) − vn (y, s)|

≤ |vn (x, t) − vn (x∗ , t)| + |vn (x∗ , t) − vn (x∗ , s)|
+|vn (x∗ , s) − vn (y, s)|
 
≤ N |x − x∗ | + |∆t|1/2 + |x∗ − y|
e − y| + |∆t|1/2 )
≤ C(|x

e and proves that {vn } is uniformly Hölder continuous on Gn with Hölder

for some constant C
1
exponent {1, }. This together with (2.7), (2.8) implies that there exists a subsequence of {v n },
2
supposed to be {vn } itself, such that {vn } converges uniformly to a certain function v on any
Nonlinear Diffusion Equations 11

 
∂vn ∂v
compact subset of QT and weak star converges to on any bounded domain of QT .
∂x ∂x
Furthermore it is easy to see that {un } (un = A−1 (vn )) converges uniformly to u = A−1 (v) on any
compact subset of QT .
∂A(u)
Obviously u and are bounded. Given any test function ϕ, i.e. ϕ ∈ C ∞ (QT ) such
∂x
that ϕ = 0 when |x| is large enough and t = T . Let n be large enough such that supp ϕ ⊂ Gn .
Multiplying (2.11) by ϕ, integrating over QT and integrating by parts, we obtain
ZZ   Z ∞
∂ϕ ∂A(un ) ∂ϕ
un − dxdt + ϕ(x, 0)A−1 (wn (x))dx = 0
QT ∂t ∂x ∂x −∞

from which it follows by letting n → ∞ and noticing that for large n, wn (x) = v0n (x) and wn (x)
converges uniformly to v0 = A(u0 ) on any finite interval, that u satisfies the integral equality of
generalized solutions, i.e u is a generalized solution of (2.1), (2.2).
Finally, we prove that the solution u is classical in {(x, t) ∈ QT , u(x, t) > 0}. Let (x0 , t0 ) ∈ QT ,
u(x0 , t0 ) > 0. Then there is a neighborhood U ⊂ QT of (x0 , t0 ) and constant α0 > 0 such that

u(x, t) ≥ α0 > 0, for (x, t) ∈ U .

Hence
α0
un (x, t) ≥ > 0 for (x, t) ∈ U and large n.
2
This means that for large n the solution un satisfies
 
∂un ∂ ∂un
= a(x, t)
∂t ∂x ∂x
with a(x, t) = A0 (un ) being uniformly parabolic on U . From the standard theory for parabolic
equations, it follows that for large n, un is uniformly bounded and equi-continuous in C 2 (U ). Thus
u ∈ C 2 (U ) and satisfies (2.1) in the classical sense. The proof is complete. 

2.2 Weakly Degenerate Equations with Lower Order Terms

This section is devoted to weakly degenerate equations in one dimension with lower order term
∂u ∂ 2 A(u) ∂B(u)
= + , (2.13)
∂t ∂x2 ∂x
where A(s), B(s) ∈ C 1 (R), A(0) = 0, A0 (s) ≥ 0, however the set E = {s; A0 (s) = 0} does not
contain any interior point. In contrast to the previous sections, the equations we consider is
much general. Only the first boundary value problem for (2.13) will be discussed in detail, the
corresponding initial and boundary conditions are

u(0, t) = u(1, t) = 0, (2.14)

u(x, 0) = u0 (x). (2.15)

Denote QT = (0, 1) × (0, T ). We introduce the following definition.

Definition 2.1 A function u ∈ L1 (QT ) is called a generalized solution of the boundary value
problem (2.13)–(2.15), if A(u), B(u) ∈ L1 (QT ) and the integral equality
ZZ   Z 1
∂ϕ ∂2ϕ ∂ϕ
u + A(u) 2 − B(u) dxdt + u0 (x)ϕ(x, 0)dx = 0
QT ∂t ∂x ∂x 0
12 Existence of Solutions

is fulfilled for any function ϕ ∈ C ∞ (QT ) withTϕ(0, t) = ϕ(1, t) = ϕ(x, T ) = 0.

It is easy to check that if u ∈ C 1 (QT ) C 2 (QT ) is a generalized solution of the boundary
value problem (2.13)–(2.15), then u is a classical solution of the problem. Here we notice that
A(u(x, t))|x=0,1 = 0 implies u(x, t)|x=0,1 = 0 because of the strict monotonicity of A(s).

Theorem 2.2 Assume that u0 ∈ Lip[0, 1] with u0 (0) = u0 (1) = 0, A(s), B(s) are appropriately
smooth, A0 (s) ≥ 0, lim A(s) = ±∞ and the set E = {s; A0 (s) = 0} has no interior point. Then
s→±∞
the first boundary value problem (2.13)–(2.15) admits a continuous solution.

To prove the theorem, we consider the following regularized problem

∂uε ∂ 2 Aε (uε ) ∂B(uε )

= + , (2.16)
∂t ∂x2 ∂x
uε (0, t) = uε (1, t) = 0, (2.17)
uε (x, 0) = u0ε (x), (2.18)

where Aε (s) = εs + A(s)(ε > 0) and u0ε is a smooth function approximating u0 uniformly with
(k) (k)
u0ε (0) = u0ε (1) = 0(k = 0, 1, 2) and |u00ε | uniformly bounded.
Let uε be a smooth solution of the problem, whose existence follows from the classical theory.
We need some estimates for uε to ensure the compactness of {uε }.
First, the maximum principle implies that

sup |uε (x, t)| ≤ M (2.19)

QT

with constant M independent of ε.

Next, we have

Lemma 2.1 Let uε be a solution of the problem (2.16)–(2.18). Then

∂Aε (uε )
≤C (2.20)
∂x x=0,1

with constant C independent of ε.

Proof. Let Z uε
A0ε (s)
λε (s) = , wε = λε (s)ds,
θ(s) 0

where θ(s) is an auxiliary function of the form θ(s) = α + s with an arbitrary constant α greater
than M , the constant in (2.19). For example, we may choose α = M + 1. Then

∂wε ∂Aε (uε ) ∂uε

= /θ(uε ) = A0ε (uε ) /θ(uε ),
∂x ∂x ∂x

∂wε ∂Aε (uε ) ∂uε

= /θ(uε ) = A0ε (uε ) /θ(uε ).
∂t ∂t ∂t
Using (2.16), one can easily check that wε satisfies

∂wε ∂ 2 wε ∂wε ∂wε

− A0ε (uε ) 2
−( + B 0 (uε )) = 0,
∂t ∂x ∂x ∂x
Nonlinear Diffusion Equations 13

∂wε 2
in which we have a term ( ) ; as will be seen below, this term plays an important role in our
∂x
∂Aε (uε )
proof. If we set wε = , then in the equation which wε satisfies, this term disappears. This
∂x
is just why we introduce the auxiliary function θ(s).
Define an operator H as follows:
∂w ∂2w ∂w ∂w
H[w] ≡ − A0ε (uε ) 2 − ( + B 0 (uε ) .
∂t ∂x ∂x ∂x
Then H[wε ] = 0. Let
vε = K(x − 1) − wε
where the constant K is to be determined. By calculation we see that, for sufficient large K > 0,
∂wε K K2
H[vε ] = −2( − )2 − − KB 0 (uε ) − H[wε ]
∂x 2 2
K2
≤− − KB 0 (uε ) < 0 in QT .
2
From this it follows that vε can not achieve its maximum at any point inside QT . In addition,
since from (2.17), (2.18) and the uniform boundedness of u00ε we have
vε (0, t) = −K < 0, vε (1, t) = 0,
∂vε
= K − wε0 (x, 0) > 0
∂x t=0
for large K, we can assert that the maximum of vε must be zero and must be achieved at x = 1.Then
∂vε
≥0
∂x x=1

and hence
∂Aε (uε ) ∂wε
= (α + uε ) ≤ (α + uε ) K ≤ C.
∂x x=1 ∂x x=1 x=1
Similarly, we can prove that
∂Aε (uε )
≥ −C.
∂x x=1
Therefore the conclusion (2.20) for x = 1 is proved. Similarly, we can prove another part of the
conclusion (2.20). 
Lemma 2.2 Let uε be a solution of the problem (2.16)–(2.18). Then
∂Aε (uε )
≤C (2.21)
∂x
with constant C independent of ε.
Proof. Let Z uε
A0ε (s)
λε (s) = , wε = λε (s)ds,
θ(s) 0
where θ(s) is an auxiliary function to be determined, the first requirement is that it has positive
upper bound and lower bound on |s| ≤ M (M is the constant in (2.19)). Then from (2.16) we see
that wε satisfies  
∂wε ∂ 2 wε ∂wε ∂wε
− A0ε (uε ) − θ 0
(u ε ) + B 0
(u ε ) =0
∂t ∂x2 ∂x ∂x
14 Existence of Solutions

∂wε
and vε = satisfies
∂x
 
∂vε ∂ 2 vε ∂uε ∂vε
− A0ε (uε ) 2 − 2θ0 (uε )vε + B 0 (uε ) + A00ε (uε )
∂t ∂x ∂x ∂x
θ00 (uε ) 3 B 00 (uε ) 2
− v − v = 0.
λε (uε ) ε λε (uε ) ε

Multiplying this equality by vε gives

1 ∂vε2 ∂ 2 vε
− A0ε (uε )vε
2 ∂t ∂x2
 
1 ∂uε ∂vε2
− 2θ0 (uε )vε + B 0 (uε ) + A00ε (uε ) (2.22)
2 ∂x ∂x
θ00 (uε ) 4 B 00 (uε ) 3
− v − v = 0.
λε (uε ) ε λε (uε ) ε

If vε2 achieves its maximum at some point of the parabolic boundary, then, by Lemma 2.1,
(2.21) holds clearly. Suppose that vε2 achieves the maximum at some point (x0 , t0 ) not on the
parabolic boundary. Then at (x0 , t0 ) the sum of the first three terms on the left side of (2.22) is
nonnegative and hence
θ00 (uε ) 4 B 00 (uε ) 3
− v − v ≤ 0,
λε (uε ) ε λε (uε ) ε
namely
−θ00 (uε )vε2 − B 00 (uε )vε ≤ 0
from which it follows by using Young’s inequality that for δ > 0,

(B 00 (uε ))2
−θ00 (uε )vε2 ≤ δvε2 + ,
4δ
namely
(B 00 (uε ))2
(−θ00 (uε ) − δ)vε2 ≤ .
4δ
If θ(s) is chosen such that θ 00 (s) has negative upper bound on |s| ≤ M , then we can choose δ > 0
so small that
vε2 ≤ C
with constant C independent of ε. The last inequality implies (2.21), if θ(s) is required to have
positive upper bound and lower bound on |s| ≤ M . The choice of such functions θ(s) is quite free,
for example, we can choose θ(s) = 1 + (M − s)(M + s). This complete the proof of our lemma. 

Lemma 2.3 Let uε be a solution of the problem (2.16)–(2.18). Then for any (x1 , t1 ), (x2 , t2 )
∈ QT ,  
1/2
|Aε (uε (x1 , t1 )) − Aε (uε (x2 , t2 ))| ≤ C |x1 − x2 | + |t1 − t2 | , (2.23)

where the constant C is independent of ε.

Proof. Since Lemma 2.2 implies that

|Aε (uε (x1 , t)) − Aε (uε (x2 , t))| ≤ C |x1 − x2 | , ∀(x1 , t), (x2 , t) ∈ QT , (2.24)
Nonlinear Diffusion Equations 15

it remains to further prove

1/2
|Aε (uε (x, t1 )) − Aε (uε (x, t2 ))| ≤ C |t1 − t2 | , ∀(x, t1 ), (x, t2 ) ∈ QT . (2.25)

Suppose, for example, ∆t = t2 − t1 > 0. Given α ∈ (0, 1) arbitrarily and denote d = ∆tα . We
may suppose that d ≤ 1/2; otherwise, (2.25) follows immediately from the uniform boundedness
of {uε }.
In case x + d ≤ 1, we integrate (2.16) over (x, x + d) × (t1 , t2 ). Integrating by parts gives
Z x+d Z t2 Z t2
∂Aε (uε ) x+d x+d
(uε (ξ, t2 ) − uε (ξ, t1 )) dξ = dt + B(uε ) dt.
x t1 ∂x x t1 x

Using the mean value theorem for integrals, we see that

Z x+d
(uε (ξ, t2 ) − uε (ξ, t1 )) dξ = d (uε (x∗ , t2 ) − uε (x∗ , t1 ))
x

for some x∗ ∈ [x, x + d]. Combining this with the above equality and using (2.19) and Lemma 2.2,
we obtain
|uε (x∗ , t2 ) − uε (x∗ , t1 )| ≤ C∆t1−α .
This, together with (2.24) gives

|Aε (uε (x, t2 )) − Aε (uε (x, t1 ))|

≤ |Aε (uε (x, t2 )) − Aε (uε (x∗ , t2 ))|

+ |Aε (uε (x∗ , t2 )) − Aε (uε (x∗ , t1 ))|

+ |Aε (uε (x∗ , t1 )) − Aε (uε (x, t1 ))|

≤ C∆tα + C∆t1−α + C∆tα = C(2∆tα + ∆t1−α )

1
which implies (2.25), if we take α = .
2
1 1
In case x + d > 1, since d ≤ , we have x > 1 − d ≥ and can obtain the same conclusion by
2 2
integrating (2.16) over (x − d, x) × (t1 , t2 ). The proof is complete. 
Proof of Theorem 2.2. Denote wε = Aε (uε ). Lemma 2.3 and (2.19) imply the uniform
boundedness and equicontinuity of {wε } on QT . Hence there exists a subsequence, still denoted
by {wε }, and a function w ∈ C 1,1/2 (QT ), such that

lim wε (x, t) = w(x, t), uniformly in QT .

ε→0

Let ψ(s) be the inverse function of A(s), whose existence for s ∈ R follows from the assumption
lim A(s) = ±∞ and the strict monotonicity of A(s). Then
s→±∞

u(x, t) = lim uε = lim ψ(wε − εuε )

ε→0 ε→0

exists and u ∈ C(QT ). To prove that u is a generalized solution of the problem (2.13)–(2.15),
notice that from (2.16)–(2.19), for any ϕ ∈ C ∞ (QT ) with ϕ(0, t) = ϕ(1, t) = ϕ(x, T ) = 0, one has
ZZ   Z 1
∂ϕ ∂2ϕ ∂ϕ
uε + wε 2 − B(uε ) dxdt + u0ε (x)ϕ(x, 0)dx = 0
QT ∂t ∂x ∂x 0
16 Existence of Solutions

and hence ZZ   Z 1
∂ϕ ∂2ϕ ∂ϕ
u + w 2 − B(u) dxdt + u0 (x)ϕ(x, 0)dx = 0
QT ∂t ∂x ∂x 0
by letting ε → 0. Since w = A(u), by definition, u is a generalized solution of (2.13)–(2.15). The
proof of Theorem 2.2 is complete. 
Theorem 2.3 If in addition to the conditions of Theorem 2.2, assume that
|A(s1 ) − A(s2 )| ≥ λ|s1 − s2 |m (2.26)
for some constants m > 1, λ > 0, then the generalized solution u of the problem (2.13)–(2.15)
given in Theorem 2.2, is Hölder continuous, precisely, u ∈ C 1/m,1/(m+1) (QT ).
Proof. In the proof of Theorem 2.2, in fact, we have reached A(u(x, t)) ∈ C 1,1/2 (QT ) which
follows from (2.23) by letting ε → 0. Thus, using the assumption (2.26), we obtain
|u(x1 , t1 ) − u(x2 , t2 )|
1/m
≤ λ−1/m |A(u(x1 , t1 )) − A(u(x2 , t2 ))|

1/m
≤ C |x1 − x2 | + |t1 − t2 |1/2


≤ C |x1 − x2 |1/m + |t1 − t2 |1/2m ∀(x1 , t1 ), (x2 , t2 ) ∈ QT ,
namely u ∈ C 1/m,1/2m (QT ). We further prove that u ∈ C 1/m,1/m+1 (QT ).
First, using (2.26) and Lemma 2.3 gives
|uε (x1 , t) − uε (x2 , t)|
≤λ−1/m |A(uε (x1 , t)) − A(uε (x2 , t))|1/m
1/m
≤C |Aε (uε (x1 , t)) − Aε (uε (x2 , t))|
+ Cε1/m |uε (x1 , t) − uε (x2 , t)|1/m
1/m
≤C |x1 − x2 | + Cε1/m . (2.27)
Next, for any given α ∈ (0, 1), by an argument similar to the proof of Lemma 2.3, we can assert
that for any x ∈ (0, 1), there exists x∗ ∈ (0, 1) with |x − x∗ | ≤ d = ∆tα , such that
|uε (x∗ , t1 ) − uε (x∗ , t2 )| ≤ C∆t1−α .
Combining this with (2.27) gives
|uε (x, t1 ) − uε (x, t2 )|

≤ |uε (x, t1 ) − uε (x∗ , t1 )| + |uε (x∗ , t1 ) − uε (x∗ , t2 )|

(2.28)
+ |uε (x∗ , t2 ) − uε (x, t2 )|


≤ C ∆tα/m + εα/m + ∆t1−α .
m
Let ε → 0 in (2.27), (2.28) and choosing α = yield
m+1
1/m
|u(x1 , t) − u(x2 , t)| ≤ C |x1 − x2 | , ∀(x1 , t), (x2 , t) ∈ QT ,
1/(m+1)
|u(x, t1 ) − u(x, t2 )| ≤ C |t1 − t2 | , ∀(x, t1 ), (x, t2 ) ∈ QT
1/m,1/m+1
which imply u ∈ C (QT ). 
Nonlinear Diffusion Equations 17

2.3 Strongly Degenerate Equations without Terms of Lower Order

Now, we turn to the discussion of strongly degenerate parabolic equations without lower order
terms in multi-dimensional case
∂u
= ∆A(u) (2.29)
∂t
with A0 (s) ≥ 0 and E = {s; A0 (s) = 0} being allowed to have interior points. Assume the initial
value condition
u(x, 0) = u0 (x). (2.30)
Due to the strong degeneracy, we need a lemma first, in order to treat some compactness
arguments during the existence proof.

Lemma 2.4 Let Ω ⊂ RN be a domain, {uk } be a uniformly bounded sequence in L∞  (Ω),

namely, |uk (x)| ≤ M a.e. in Ω, with M independent of k. Then there exist a subsequence ukj
of {uk }, and a family of probability measures {µx }x∈Ω with suppµx ⊂ [−M, M ] such that for any
F ∈ C(R), Z
∗
F (ukj (x)) * F (λ)µx (dλ), in L∞ (Ω).
R

Using this lemma, we can prove the following result on the weak convergence.

Lemma 2.5 Let Ω ⊂ RN be a bounded domain, |uk (x)| ≤ M a.e in Ω and

∗
uk (x) * u(x), in L∞ (Ω).

Assume that A(s), B(s) ∈ C(R) and A(s) is nondecreasing. If for any α ∈ A(R), B(A −1 (α))
contains only one single point, and

A(uk (x)) → w(x), B(uk (x)) → v(x), a.e. in Ω,

then
A(u(x)) = w(x), B(u(x)) = v(x), a.e. in Ω.

Theorem 2.4 Assume that A(s) ∈ C 1 (R) and u0 (x) ∈ L∞ (RN ) with A0 (s) ≥ 0 and A(u0 ) ∈
1,2
Wloc (RN ).
Then the Cauchy problem (2.29), (2.30) admits a bounded generalized solution.

Proof. Consider the regularized problem

∂uε
= ∆Aε (uε ), (2.31)
∂t
uε (x, 0) = u0ε (x), (2.32)

where Aε (s) and u0ε are the locally uniform smooth approximation of A(s) and u0 respectively,
with A0ε (s) > 0, Aε (0) = 0 and sup |u0ε (x)| ≤ M .
From the maximum principle, we first have

sup |uε (x, t)| ≤ M. (2.33)

QT

∂
Next, multiply (2.31) by Aε (uε )ωλ (x) and integrate over Qt , where
∂t
 p 
ωλ (x) = exp −λ 1 + |x|2 .
18 Existence of Solutions

Then we obtain
ZZ  2 ZZ
∂uε ∂Aε (uε )
A0ε (uε ) ωλ (x)dxds = 4 Aε (uε )ωλ (x)dxds.
Qt ∂t Qt ∂t

Integrating by parts, noticing that |∇ωλ (x)| ≤ Cωλ (x) and using Young’s inequality yield
ZZ  2
∂uε
A0ε (uε ) ωλ (x)dxds
Qt ∂t
ZZ
1 ∂

= − | 5 Aε (uε )|2 ωλ (x) dxds
2
Qt ∂t
ZZ
∂uε
− A0ε (uε ) 5 Aε (uε ) 5 ωλ (x)dxds
Q ∂t
Z t Z
1 2 1
≤ − | 5 Aε (uε )| ωλ (x)dx + | 5 Aε (u0ε )|2 ωλ (x)dx
2 RN 2 RN
ZZ  2 ZZ
1 ∂uε
+ A0ε (uε ) ωλ (x)dxds + C | 5 Aε (uε )|2 ωλ (x)dxds,
2 Qt ∂t Qt

namely,
ZZ  Z 2
∂uε
A0ε (uε )
ωλ (x)dxds + |∇Aε (uε )|2 ωλ (x)dx
∂t N
Z Qt ZZ R
2 2
≤ |∇Aε (u0ε )| ωλ (x)dx + C |∇Aε (uε )| ωλ (x)dxds,
RN Qt

from which, using Gronwall’s inequality, it follows that

ZZ  2
∂uε
A0ε (uε ) ωλ (x)dxdt ≤ C, (2.34)
QT ∂t
ZZ
|∇Aε (uε )|2 ωλ (x)dxdt ≤ C. (2.35)
QT

Denote vε = Aε (uε ). From (2.33), (2.34), (2.35) we see that there exists a subsequence of {u ε },
denoted still by {uε }, such that
∗
uε (x, t) * u(x, t), in L∞ (QT ), (2.36)
vε (x, t) → w(x, t), a.e. in QT , (2.37)
1,2
where u ∈ L∞ (QT ), w ∈ Wloc (QT ) with
ZZ
2
|∇w| ωλ (x)dxdt < +∞.
QT

Since Aε (s) approximates A(s) locally uniformly, from the uniform boundedness of {uε } and (2.37),
we have
wε ≡ A(uε ) → w(x, t), a.e in QT . (2.38)
Using Lemma 2.5, we then conclude from (2.37) and (2.38) that w(x, t) = A(u(x, t)) a.e. in Q T .
Finally, it is easy to check that u satisfies the integral equality and the initial value condition
in the sense of definition of generalized solutions. The proof is complete. 
Nonlinear Diffusion Equations 19

2.4 General Strongly Degenerate Equations

Now we turn to equations with arbitrary degeneracy in general form
∂u ~
= ∆A(u) + divB(u), (2.39)
∂t
where Z u Z u
A(u) = a(s)ds, ~
B(u) = ~b(s)ds
0 0

with a(s) ≥ 0 and a(s) and ~b(s) being continuous functions, but E = {s; a(s) = 0} being arbitrary.
As we have already seen in the previous sections, strongly degenerate parabolic equations are
in fact parabolic-hyperbolic mixed type equations. As a direct consequence, such equations may
have discontinuous solutions, which should be defined in a suitable manner. However, it is helpful
to recall something related to the first order equations first. So, we consider a special case of the
equation (2.39), namely, A(s) ≡ 0,
∂u ~
= divB(u). (2.40)
∂t
Definition 2.2 Let G be a subdomain in RN × (0, +∞). A function u ∈ L∞ (G) ∩ BV (G) is
called a BV solution of the equation (2.40) in G, if for any ϕ ∈ C0∞ (G), the following integral
equality is fulfilled: ZZ  
∂ϕ ~
u − B(u)∇ϕ dxdt = 0.
G ∂t
However, the solutions thus defined may not be unique, both for the Cauchy problem and for
the boundary value problem. Usually, the so-called entropy condition should be attached to the
solution, namely,

[sgn(u+ − k) − sgn(u− − k)][(u − k)γt − (Bi (u) − Bi (k))γxi ] ≤ 0 (2.41)

holds on the jump set Γu , for any k ∈ R, where ~γ ≡ (γt , γx1 , · · · , γxN ) denotes the normal to Γu ,
u+ and u− denote the approximate limits of u along the direction of ~γ and −~γ , u denotes the
symmetric mean value of u. Solutions satisfying (2.41) are called BV entropy solutions, which can
also be defined by using one integral inequality:
ZZ  
∂ϕ ~ ~
sgn(u − k) (u − k) − (B(u) − B(k))∇ϕ dxdt ≥ 0
G ∂t
holds for any 0 ≤ ϕ ∈ C0∞ (G).
Motivated by this idea, Vol’pert and Hujaev gave an extension of the definition to the solutions
of the equation (2.39).

Definition 2.3 Let G be a subdomain in RN × (0, +∞). A function u ∈ L∞ (G) ∩ BV (G)

is called a BV entropy solution of the equation (2.39) in G, if ∇A(u) ∈ L 2loc (G), and for any
0 ≤ ϕ ∈ C0∞ (G), the following integral inequality is fulfilled:
ZZ  
∂ϕ ~ ~
sgn(u − k) (u − k) − (B(u) − B(k) − ∇A(u))∇ϕ dxdt ≥ 0.
G ∂t

Now, we consider the first boundary value problem. Let Ω be a bounded domain of RN with
smooth boundary ∂Ω. The simplest boundary value condition is

u(x, t) = 0 (x, t) ∈ ∂Ω × (0, T ).

20 Existence of Solutions

However, because of the possible discontinuity, the condition thus proposed might not be well-
posed. Instead of this, we consider the boundary value condition as

A(u(x, t)) = 0 (x, t) ∈ ∂Ω × (0, T ). (2.42)

Since the discontinuity may appear on the boundary, the condition (2.42) alone may not ensure
the uniqueness of solutions. So, we should attach the so-called boundary entropy condition to
solutions. This can also be done by considering one integral inequality, see the definition below.
The initial value condition can be given usually

u(x, 0) = u0 (x) x ∈ Ω. (2.43)

Denote QT = Ω × (0, T ).

Definition 2.4 A function u ∈ L∞ (QT ) ∩ BV (QT ) is called a BV entropy solution of the

boundary value problem (2.39), (2.42), (2.43), if the following conditions are fulfilled:
(1) ∇A(u) ∈ L2 (QT );
(2) ess lim u(x, t) = u0 (x) for almost all x ∈ Ω and A(γu(x, t)) = 0 for almost all (x, t) ∈
t→0+
∂Ω × (0, T ), where γu(x, t) is the trace of u at (x, t);
(3) For any 0 ≤ ϕ1 , ϕ2 ∈ C ∞ (QT ) with suppϕ1 , ϕ2 ⊂ Ω × (0, T ), ϕ1 = ϕ2 and k ∈ R,
∂Ω ∂Ω
there holds ZZ  
∂ϕ1 ~ ~
sgn(u−k) (u−k) −(B(u)− B(k)−∇A(u))∇ϕ 1 dxdt
QT ∂t
ZZ  
∂ϕ2 ~ ~
+ sgnk (u−k) −(B(u)− B(k)−∇A(u))∇ϕ 2 dxdt ≥ 0.
QT ∂t

Lemma 2.6 Assume that A(s) and f (x) are appropriately smooth with A 0 (s) > 0. Let u be a
solution of the problem
∆A(u) = f, x ∈ Ω,

u = 0.
∂Ω

Then Z Z
∂u
A0 (0) dσ ≤ |f (x)| dx,
∂Ω ∂n Ω

∂u
where is the derivative of u along the outward normal vector.
∂n
Proof. Consider the problem

div(A0 (u(x))∇u1 ) = f + (x), x ∈ Ω,

u1 = 0.
∂Ω

Here and below, as before, f + = max{f, 0}, f − = max{−f, 0}. The maximum principle shows
∂u1
that u1 (x) ≤ 0 in Ω and hence ≥ 0 on ∂Ω. Integrating the equation that u1 satisfies, we
∂n
obtain Z Z Z
∂u1 ∂u1
f + (x)dx = A0 (0) dσ = A0 (0) dσ.
Ω ∂Ω ∂n ∂Ω ∂n
Nonlinear Diffusion Equations 21

Similarly, if we consider the problem

div(A0 (u(x))∇u2 ) = f − (x), x ∈ Ω,

u2 = 0,
∂Ω

then we obtain Z Z Z
− 0 ∂u2 ∂u2
f (x)dx = A (0) dσ = A0 (0) dσ.
Ω ∂Ω ∂n ∂Ω ∂n
By the uniqueness of solutions, we have u = u1 − u2 . Thus
Z
∂u
A0 (0) dσ
∂Ω ∂n
Z Z
∂u1 ∂u2
≤ A0 (0) dσ + A0 (0) dσ
Z ∂Ω ∂n Z ∂Ω
Z
∂n
= f + (x)dx + f − (x)dx = |f (x)|dx.
Ω Ω Ω

The proof of the lemma is complete. 

1/2
Lemma 2.7 Denote ξ = (ξ1 , · · · , ξN ), |ξ| = ξ12 + · · · + ξN
2
and
p
Iη (ξ) = η + |ξ|2 , (η > 0).
 
∂ 2 Iη
Then Iη (ξ) is a strictly concave function, namely, the matrix is positively definite, and
∂ξi ∂ξj
for any j = 1, · · · , N , ξ ∈ RN ,
∂ 2 Iη
lim ξi = 0. (2.44)
η→0 ∂ξi ∂ξj

Proof. By an immediate calculation, we first obtain

∂Iη ξi
=p ,
∂ξi η + |ξ|2
∂ 2 Iη δij ξi ξj
=p − .
∂ξi ∂ξj η + |ξ|2 (η + |ξ|2 )3/2

Hence, for any α = (α1 , · · · , αN ) 6= 0,

XN  
∂ 2 Iη 1 2 |α · ξ|2
αi αj = p |α| −
i,j=1
∂ξi ∂ξj η + |ξ|2 η + |ξ|2
 2 2

1 |α| |ξ|
≥ p |α|2 −
η + |ξ|  2 η + |ξ|2
2

|α| |ξ|2
= p 1− > 0,
η + |ξ|2 η + |ξ|2

which shows the strict convexity of Iη (ξ).

22 Existence of Solutions

Next we have
∂ 2 Iη
lim ξi
η→0 ∂ξi ∂ξj
ξj |ξ|2 ξj
= lim p − lim 2 )3/2
η→0 η + |ξ|2  η→0 (η + |ξ|
2
ξj |ξ|
= lim p 1−
η→0 η + |ξ|2 η + |ξ|2
= 0,
which shows (2.44). 

Theorem 2.5 Assume that A(s), B(s) ~ and u0 are appropriately smooth and u0 (x) satisfies
suitable compatibility conditions on ∂Ω × {t = 0}. Then the boundary value problem (2.39), (2.42),
(2.43) admits a BV entropy solution.

Proof. Consider the regularized problem

∂uε ~ ε ),
= ∆Aε (uε ) + divB(u (2.45)
∂t
uε (x, t) = 0, (x, t) ∈ ∂Ω × (0, T ), (2.46)
uε (x, 0) = u0 (x), x ∈ Ω (2.47)

where Aε (s) = εs + A(s) (ε > 0). The existence of a appropriately smooth solution u ε follows
from the standard theory.
First, by the maximum principle, we have

sup |uε (x, t)| ≤ M (2.48)

QT

provided sup |u0 (x)| ≤ M .

RN
Next, we estimate the L2 norm of ∇Aε (uε ). To this purpose, multiply (2.45) by Aε (uε ) and
integrate over Qt ,
ZZ
∂uε
Aε (uε ) dxds
Z ZQ t ∂t ZZ
= ∆Aε (uε )Aε (uε )dxds + ~ ε )Aε (uε )dxds.
divB(u
Qt Qt

Integrating by parts yields

Z ZZ
2
Ψε (uε )dx + |∇Aε (uε )| dxds
Ω Qt
Z ZZ
= Ψε (u0 )dx − ~ ε )∇Aε (uε )dxds
B(u
Ω Qt
ZZ ZZ 2
1 2 1 ~ ε ) dxds + C
≤ |∇Aε (uε )| dxds + B(u
2 Q 2 Qt
ZZ t
1 2
≤ |∇Aε (uε )| dxds + C,
2 Qt

where Z s
Ψε (s) = Aε (s)ds.
0
Nonlinear Diffusion Equations 23

Hence ZZ
2
|∇Aε (uε )| dxdt ≤ C. (2.49)
QT

∂uε
Finally, we estimate the L1 norm of the derivatives of uε . Doing this for is easy. To this
∂t
∂uε
purpose, differentiate (2.45) with respect to t and denote vε = . Then
∂t
∂vε  
= ∆ (aε (uε )vε ) + div ~b(uε )vε .
∂t
Let Hη (s) be defined as follows:
Z s  
2 |s|
Hη (s) = hη (σ)dσ with hη (s) = 1− .
η η +

Multiplying the above equality by Hη (vε ) and then integrating over Ω with respect to x, yield
Z
∂
θη (vε (x, t))dx
∂t
Z Ω   
= Hη (vε ) ∆ (aε (uε )vε ) + div ~b(uε )vε dx
ΩZ

= − aε (uε )Hη0 (vε ) (∇vε )2 dx (2.50)

ZΩ

− a0ε (uε )vε Hη0 (vε )∇vε ∇uε dx

ZΩ

− ~b(uε )vε Hη0 (vε )∇vε dx,

Ω

where Z s
θη (s) = Hη (σ)dσ.
0

Remove the first term on the right hand side of (2.50), which is nonpositive and then integrate
with respect to t and let η → 0. We obtain
Z
∂uε (x, t)
sup dx ≤ C. (2.51)
0<t<T Ω ∂t

The L1 norm of |∇uε | is a little difficult to estimate. Let Iη (ξ) be the function defined in
∂Iη (∇uε )
Lemma 2.7. Differentiate (2.45) with respect to xi multiply the resulting equality by
∂ξi
and sum up for i from 1 up to N , and then integrate with respect to x on Ω,
Z Z
∂ ∂Iη (∇uε ) ∂uεxi
Iη (∇uε )dx = dx
∂t Ω Ω ∂ξi ∂t
Z Z
∂Iη (∇uε ) ∂ ∂Iη (∇uε ) ∂ ~ ε )dx.
= 4 Aε (uε )dx + divB(u
Ω ∂ξ i ∂x i Ω ∂ξi ∂xi
Integrating by parts gives
Z Z
∂ ∂ 2 Iη (∇uε ) ∂ 2 uε ∂ 2 Aε (uε )
Iη (∇uε )dx = − dx
∂t Ω Ω ∂ξi ∂ξj ∂xj ∂xp ∂xi ∂xp
24 Existence of Solutions

Z
∂ 2 Iη (∇uε ) ∂ 2 uε ∂Bp (uε )
− dx
Ω ∂ξi ∂ξj ∂xj ∂xp ∂xi
Z
∂Iη (∇uε ) ∂
+ ~n · ∇ Aε (uε )dσ
∂Ω ∂ξi ∂xi
Z
∂Iη (∇uε ) ∂ ~
+ ~n · B(uε )dσ
∂Ω ∂ξi ∂xi
=J1η + J2η + J3η + J4η . (2.52)
By Lemma 2.7, we have
lim J1η
η→0
Z
∂ 2 Iη (∇uε ) ∂ 2 uε ∂ 2 uε
= − lim aε (uε ) dx
η→0 Ω ∂ξi ∂ξj ∂xj ∂xp ∂xi ∂xp
Z
∂ 2 Iη (∇uε ) ∂ 2 uε ∂uε ∂uε
− lim a0ε (uε ) dx
η→0 Ω ∂ξi ∂ξj ∂xj ∂xp ∂xi ∂xp
Z
∂ 2 Iη (∇uε ) ∂ 2 uε ∂uε ∂uε
≤ − lim a0ε (uε ) dx = 0. (2.53)
η→0 Ω ∂ξi ∂ξj ∂xj ∂xp ∂xi ∂xp
Similarly Z
∂ 2 Iη (∇uε ) ∂ 2 uε ∂uε
lim J2η = − lim bp (uε ) dx = 0. (2.54)
η→0 η→0 Ω ∂ξi ∂ξj ∂xj ∂xp ∂xi
∂uε ∂uε
To estimate J3η and J4η , notice that uε |∂Ω = 0 implies = ni on ∂Ω, where ni is the
∂xi ∂~n
i-component of ~n. Hence
∂ ~ ∂
~n · B(uε ) = np Bp (uε )
∂xi ∂xi
∂uε ∂uε
= np bp (uε ) = bp (uε ) ni np ,
∂xi ∂~n
∂ ∂ 2 Aε (uε )
~n · ∇ Aε (uε ) = np
∂xi ∂xp ∂xi
 
∂ ∂Aε (uε )
= np ni
∂xp ∂~n
 
∂ ∂Aε (uε ) ∂Aε (uε ) ∂ni
= ni np + np .
∂xp ∂~n ∂~n ∂xp
On the other hand, since

~ ε) = ∂uε
∆Aε (uε ) + ∇B(u = 0, on ∂Ω,
∂t
we have  
∂uε ∂ ∂Aε (uε )
bp (uε ) np = −∆Aε (uε ) = − np
∂~n ∂xp ∂~n
 
∂ ∂Aε (uε ) ∂Aε (uε ) ∂np
= − np − .
∂xp ∂~n ∂~n ∂xp
Hence for (x, t) ∈ Ω,
 
∂ ~ ∂ ∂Aε (uε ) ∂ni ∂np
~n · B(uε ) + ~n · ∇ Aε (uε ) = np − ni .
∂xi ∂xi ∂~n ∂xp ∂xp
Nonlinear Diffusion Equations 25

Therefore

|J3η + J4η |
Z
∂Iη (∇uε ) ∂ ∂ ~
≤ ~n · ∇ Aε (uε ) + ~n · Bε (uε ) dσ
∂Ω ∂ξi ∂xi ∂xi
Z
∂Aε (uε ) ∂ni ∂np
≤ np − ni dσ
∂Ω ∂~n ∂x p ∂xp
Z
∂uε
≤ CA0ε (0) dσ.
∂Ω ∂~ n

Using Lemma 2.6, (2.45) and (2.48), (2.51), we further obtain

Z
|J3η + J4η | ≤ |∆Aε (uε )|dx
Z ZΩ
∂uε ~ ε ) dx
≤ dx + ∇B(u
Ω ∂t Ω
Z
≤C + C |∇uε | dx. (2.55)
Ω

Combining (2.53), (2.54), (2.55) with (2.52) we are led to

Z ZZ
|∇uε (x, t)| dx ≤ C + C |∇uε (x, s)|dxds,
Ω Qt

and then using Gronwall’s inequality, finally obtain

Z
sup |∇uε (x, t)| dx ≤ C. (2.56)
0<t<T Ω

Just as in one dimensional case, we may apply the estimates (2.48), (2.50), (2.51), (2.56) to
conclude the existence of a subsequence of {uε }, whose limit function u is a BV solution of the
problem (2.39),(2.42),(2.43), and thus complete the proof of our theorem. 
26 Uniqueness of Solutions

3 Uniqueness of Solutions
3.1 Comparison Approach
Consider the Cauchy problem for Porous Medium equation
∂u
= ∆um , (3.1)
∂t
u(x, 0) = u0 (x). (3.2)
Denote QT = RN × (0, T ).

Theorem 3.1 Let ui ∈ L1 (QT ) ∩ L∞ (QT ) (i = 1, 2) be generalized solutions of the Cauchy

problem for (3.1) with initial data u0i (x), (i = 1, 2). If 0 ≤ u01 (x) ≤ u02 (x) a.e on RN , then
u1 (x, t) ≤ u2 (x, t) a.e on QT .

Proof. From the definition of generalized solutions, we have

Z Z
ui (x, τ )ϕ(x, τ )dx − ui0 (x)ϕ(x, 0)dx
RN   R N
ZZ
∂ϕ
= ui + um i ∆ϕ dxdt (i = 1, 2)
Qτ ∂t

for any τ ∈ (0, T ) and ϕ ∈ C ∞ (QT ) which vanishes when |x| is large enough.
Let z = u1 − u2 , z0 = u01 − u02 . Then
Z Z
z(x, τ )ϕ(x, τ )dx − z0 (x)ϕ(x, 0)dx.
RN RN
ZZ  
∂ϕ
= z + (um 1 − u m
2 )∆ϕ dxdt.
Qτ ∂t
ZZ  
∂ϕ
= z + a∆ϕ dxdt, (3.3)
Qτ ∂t

where  m m
 u1 (x, t) − u2 (x, t) , if u1 (x, t) 6= u2 (x, t),
a(x, t) = u1 (x, t) − u2 (x, t)

mu1m−1 (x, t), if u1 (x, t) = u2 (x, t).
If for any nonnegative function g ∈ C0∞ (RN ), the problem

 ∂ϕ + a∆ϕ = 0,
∂t (3.4)

ϕ(x, τ ) = g(x),

had a solution ϕ ∈ C ∞ (Qτ ) which vanishes when |x| is large enough, then from (3.3) we would
obtain Z
z(x, t)g(x)dx ≤ 0. (3.5)
RN

Here we have used the assumption z0 (x) ≤ 0 and the fact ϕ(x, t) ≥ 0 which follows from the
maximum principle. Therefore by the arbitrariness of g(x) we get z(x, τ ) ≤ 0 or u1 (x, τ ) ≤ u2 (x, τ )
a.e for x ∈ RN and this is what we want to prove.
Nonlinear Diffusion Equations 27

However since the coefficient a in (3.4) is merely a nonnegative and locally integrable function,
(3.4) does not admit any smooth solution in general and even if (3.4) does admit, the solution can
not have compact support in x in general. In view of this point, we replace a by
1
an = ρ n ∗ a + ,
n
where ρn is a mollifier on QT and consider the boundary value problem
 ∂ϕ
 for |x| < R, 0 < t < τ,
 ∂t + an ∆ϕ = 0


ϕ=0 for |x| = R, 0 < t < τ, (3.6)




ϕ(x, τ ) = g(x) for |x| ≤ R,

where R > R0 + 1 such that supp g(x) ⊂ BR0 = {x ∈ RN ; |x| < R0 }. We choose ρn such that
Z τZ
2 1
(a − ρn ∗ a) dxdt ≤ 2 . (3.7)
0 BR n
Let ϕn be a solution of (3.6). Extend the definition of ϕn to the whole Qτ by setting ϕn = 0
outside the domain |x| ≤ R, 0 ≤ t ≤ τ . Since the extended function ϕn may not necessarily be a
sufficiently smooth function on Qτ , we use a function ξR ∈ C0∞ (RN ) with the following properties
to ”cut-off” ϕn : 

 0 ≤ ξR (x) ≤ 1,





 ξR (x) = 1, for |x| < R − 1,
(3.8)

 1
 ξR (x) = 0, for |x| > R − ,


 2


|∇ξR (x)|, |∆ξR (x)| ≤ C.
Here and below, we always use C to denote a universal constant independent of R, n, which may
take different values on different occasions. Choosing ϕ = ξR ϕn in (3.3) yields
Z Z
z(x, τ )g(x)ξR (x)dx − z0 (x)ξR (x)ϕn (x, 0)dx
N RN
ZRZ
= (um m
1 − u2 ) (2∇ξR ∇ϕn + ϕn ∆ξR ) dxdt.
Qτ
ZZ
+ zξR (a − an )∆ϕn dxdt ≡ In + Jn . (3.9)
Qτ

Now we are ready to estimate In and Jn . Multiplying the equation in (3.6) by ∆ϕn , integrating
over BR × (t, τ ) and applying Green’s formula, we obtain
Z Z τZ Z
1 1
|∇ϕn (x, t)|2 dx + an (∆ϕn )2 dxds = |∇g|2 dx,
2 BR t BR 2 BR
from which it follows in particular
Z τ Z
|∇ϕn |2 dxds ≤ C, (3.10)
t BR
Z τ Z
an (∆ϕn )2 dxds ≤ C. (3.11)
t BR
28 Uniqueness of Solutions

Using (3.8), (3.10) and noting that ui ∈ L∞ (QT ) (i = 1, 2) and ϕn is uniformly bounded yield
Z τZ
|In | ≤ C (um m
1 + u2 )(|∇ϕn | + 1)dxdt
0
Z τZ B R \BR−1
(3.12)
≤C (u1 + u2 )dxdt.
0 BR \BR−1

Using (3.11) and noting that ui ∈ L∞ (QT ) (i = 1, 2) yield

Z τ Z 1/2 Z τ Z 1/2
(a − an )2 2
|Jn | ≤ C dxdt an (∆ϕn ) dxdt
0 BR an 0 BR
Z τ Z 1/2
(a − an )2
≤C dxdt .
0 BR an

By virtue of (3.7) we further obtain

Z Z  2 !1/2
√ τ
1 C
|Jn | ≤ C n a − ρn ∗ a − dxdt ≤√ . (3.13)
0 BR n n

Combining (3.12), (3.13) with (3.9) and noting that z0 (x) ≤ 0, ϕn ≥ 0, ξR ≥ 0, we finally arrive at
Z
z(x, τ )g(x)ξR (x)dx
N
ZR
≤ z0 (x)ϕn (x, 0)ξR (x)dx + |In | + |Jn |
RN
Z τZ
C
≤|In | + |Jn | ≤ C (u1 + u2 )dxdt + √ .
0 BR \BR−1 n

Letting n → ∞ and then R → ∞ and noting that ui ∈ L1 (QT ) (i = 1, 2), we see that the right
hand side tends to zero and hence (3.5) holds. The proof is complete. 

Remark 3.1 Checking the proof of Theorem 3.1, we see that the method applied is adapted to
more general equations of the form
∂u
= ∆A(u) (3.14)
∂t
with A(u) ∈ C 1 [0, ∞) and

A(s) > 0, A0 (s) > 0 for s > 0, A(0) = A0 (0) = 0.

. In other words, using the same method we can prove that for generalized solutions of (3.14) in
L1 (QT ) ∩ L∞ (QT ), the comparison theorem is valid.

Remark 3.2 It should be pointed out that requiring a generalized solution u to belong to
L1 (QT ) ∩ L∞ (QT ) is too restrictive, which means that u must be ”small” at infinity and ex-
cludes even the nonzero constant solution. Fortunately those generalized solutions determined by
the initial data with compact support satisfy such condition.

As an immediate corollary of Theorem 3.1, we have

Theorem 3.2 Suppose 0 ≤ u0 ∈ L1 (RN ) ∩ L∞ (RN ). Then the Cauchy problem (3.1), (3.2)
admits at most one generalized solution in L1 (QT ) ∩ L∞ (QT ).
Nonlinear Diffusion Equations 29

Similar to the proof of Theorem 3.1, we may obtain the following comparison theorem for the
boundary value problem.

Theorem 3.3 Let ui ∈ L1 (ΩT ) ∩ L∞ (ΩT ) (i = 1, 2) be generalized solutions of (3.1) satisfying

the initial value condition and boundary value condition

ui (x, 0) = ui0 (x), for x ∈ Ω (i = 1, 2)

and
ui (x, t)|Γ = gi (x, t) for (x, t) ∈ Γ (i = 1, 2),
where Ω ⊂ RN is a smooth domain, Γ = ∂Ω × (0, T ), ΩT = Ω × (0, T ). If u1 (x) ≤ u2 (x) a.e on Ω
and g1 (x, t) ≤ g2 (x, t) a.e on Γ, then u1 (x, t) ≤ u2 (x, t) a.e on ΩT .

Remark 3.3 Since the generalized solution u of the Cauchy problem for (3.1) on Q T with
∂A(u)
locally integrable derivatives (i = 1, 2, · · · , N ) is also a generalized solution of the boundary
∂xi
value problem for (3.1) on any domain of the form ΩT = Ω × (0, T ), we can use the comparison
theorem on any ΩT for generalized solutions of the Cauchy problem.

3.2 Testing Techniques

In this section, we study the Cauchy problem for the filtration equation in one dimension:

∂u ∂ 2 A(u)
= , (3.15)
∂t ∂x2
u(x, 0) = u0 (x), (3.16)

where u0 (x) ≥ 0 is a locally integrable function on R and A(s) ∈ C 1 [0, ∞) satisfies the conditions

A(s) > 0, A0 (s) > 0, for s > 0, A(0) = A0 (0) = 0. (3.17)

Denote QT = R × (0, T ).

Theorem 3.4 The Cauchy problem (3.15), (3.16) has at most one generalized solution u
∂A(u)
bounded together with the weak derivative .
∂x
∂A(u)
Proof. First notice that if both u and are bounded, then by approximation, the
∂x
integral identity in the definition of generalized solutions holds for any ϕ ∈ W 1,∞ (QT ) vanishing
when |x| is large enough and t = T . Such functions will be also called test functions.
Now let u1 , u2 be generalized solutions of (3.15), (3.16). Then u1 and u2 satisfy the integral
equality of the definition of generalized solutions. Hence we have
ZZ ZZ  
∂ϕ ∂ϕ ∂A(u1 ) ∂A(u2 )
(u1 − u2 )dxdt = − dxdt (3.18)
QT ∂t QT ∂x ∂x ∂x

for any test function ϕ.

If Z t
ϕ(x, t) = (A(u1 (x, τ )) − A(u2 (x, τ )))dτ (3.19)
T
30 Uniqueness of Solutions

could be chosen as a test function, then from (2.4) we would have

ZZ
(A(u1 ) − A(u2 ))(u1 − u2 )dxdt
QT
ZZ Z t  
∂A(u1 (x, τ )) ∂A(u2 (x, τ ))
= − dτ
QT  T ∂x ∂x

∂A(u1 (x, t)) ∂A(u2 (x, t))
· − dxdt
∂x ∂x
ZZ Z t   2
1 ∂ ∂A(u1 ) ∂A(u2 )
= − dτ dxdt
2 QT ∂t T ∂x ∂x
Z Z 0   2
1 ∞ ∂A(u1 ) ∂A(u2 )
= − − dτ dx ≤ 0
2 −∞ T ∂x ∂x
and hence u1 = u2 a.e. on R due to the condition (3.17).
However the function ϕ defined by (3.19) can not play the role of a test function since in general
it does not vanish for large |x|, although we have ϕ ∈ W 1,∞ (QT ) and ϕ(x, T ) = 0.
A natural idea is to consider the following cut-off function instead of ϕ:
Z t
ϕn (x, t) = αn (x) (A(u1 ) − A(u2 ))dτ
T

where αn (x) is a smooth function such that

αn (x) = 1, when |x| ≤ n − 1;

αn (x) = 0, when |x| ≥ n;

0 ≤ αn (x) ≤ 1, when n − 1 < |x| < n;

|α0n (x)| is bounded uniformly.

Substituting ϕ = ϕn into (3.18) yields

ZZ
αn (x)(A(u1 ) − A(u2 ))(u1 − u2 )dxdt
QT
Z ∞ Z 0   2
1 ∂A(u1 ) ∂A(u2 )
=− αn (x) − dτ dx
2 −∞ T ∂x ∂x
ZZ Z t 
+ α0n (x)
(A(u1 ) − A(u2 )) dτ
Qn T
 
∂A(u1 ) ∂A(u2 )
· − dxdt
∂x ∂x
ZZ Z t 
≤ α0n (x) (A(u1 ) − A(u2 )) dτ
Qn T
 
∂A(u1 ) ∂A(u2 )
· − dxdt, (3.20)
∂x ∂x
where Qn = {(x, t); n − 1 < |x| < n, 0 < t < T }.
∂A(ui )
Since ui and (i = 1, 2) are bounded and αn (x) is bounded uniformly in n, the right
∂x
hand side of (3.20), denoted by I2n , is bounded in n, so is the left hand side, denoted by I1n . Since
Nonlinear Diffusion Equations 31

αn (x) increases with n, so does I1n . Thus lim I1n exists and we have
n→∞
ZZ
lim I1n = (A(u1 ) − A(u2 ))(u1 − u2 )dxdt.
n→∞ QT

Furthermore we can prove

lim I2n = 0. (3.21)
n→∞

∂A(ui )
In fact, from the boundedness of ui and and the uniform boundedness of α0n (x), we
∂x
have ZZ
|I2n | ≤ C |A(u1 ) − A(u2 )|dxdt
Qn
Z Z 1/2
2
≤C (A(u1 ) − A(u2 )) dxdt
Qn
Z Z 1/2
≤C (A(u1 ) − A(u2 ))(u1 − u2 )dxdt ,
Qn
ZZ
where C is a constant independent of n. The finiteness of the integral (A(u1 ) − A(u2 ))(u1 −
QT
u2 )dxdt implies that the right hand side of (3.21) tends to zero as n → ∞.
Letting n → ∞ in (3.20) yields
ZZ
(A(u1 ) − A(u2 ))(u1 − u2 )dxdt = 0,
QT

and hence u1 = u2 a.e. on QT . This completes the proof of our theorem. 

3.3 Holmgren’s Approach

The purpose of this section is to show the basic idea of a method based on adjoint equations for
proving the uniqueness of solutions, whose details will be given in next subsection. So, we will
discuss only the Hölder continuous solutions of a uniformly parabolic equation, namely, consider
the boundary value problem for the equation
∂u ~
= ∆A(u) + divB(u), in QT ≡ Ω × (0, T ), (3.22)
∂t
where Z u Z u
A(u) = a(s)ds, ~
B(u) = ~b(s)ds
0 0

with a(s) ≥ 0, a(s) and ~b(s) being Hölder continuous. The boundary value condition and initial
value condition are
u(x, t) = 0, on x ∈ ∂Ω,
u(x, 0) = u0 (x). (3.23)

Definition 3.1 A function u ∈ C α,α/2 (QT ) is called a Hölder continuous solution of the bound-
ary value problem (3.22)–(3.23), if for any ϕ ∈ C 2+α,1+α/2 (QT ) with ϕ(x, t) = 0 for x ∈ ∂Ω and
t = T , the following integral equality holds:
Z ZZ ZZ ZZ
∂ϕ ~
u0 (x)ϕ(x, 0)dx + u dxdt + A(u)∆ϕdxdt − B(u)∇ϕdxdt = 0.
Ω QT ∂t QT QT
32 Uniqueness of Solutions

Theorem 3.5 Assume that u0 ∈ C α (Ω), A0 (s) ≥ a0 > 0 for some constant a0 . Then the first
boundary value problem (3.22)–(3.23) has at most one Hölder continuous solution.

Proof. Let u1 , u2 ∈ C α,α/2 (QT ) be solutions of the boundary value problem (3.22)–(3.23).
By the definition of Hölder continuous solutions, we have
ZZ  
∂ϕ ~
(u1 − u2 ) +ea∆ϕ − β · ∇ϕ dxdt = 0 (3.24)
QT ∂t

for any ϕ ∈ C 2+α,1+α/2 (QT ) with ϕ(x, t) = 0 for x ∈ ∂Ω and t = T , where

Z 1
a=e
e a(u1 , u2 ) = A0 (θu1 + (1 − θ)u2 )dθ,
0
Z 1
β~ = β(u
~ 1 , u2 ) = ~ 0 (θu1 + (1 − θ)u2 )dθ.
B
0

Consider the adjoint problem

∂ϕ
+e a∆ϕ − β~ · ∇ϕ = u1 − u2 ,
∂t
ϕ(x, t) = 0, on x ∈ ∂Ω

ϕ(x, T ) = 0.

Since u1 − u2 ∈ C α,α/2 (QT ), from the classical theory for parabolic equations, the above problem
admits a unique solution ϕ ∈ C 2+α,1+α/2 (QT ). Substituting such ϕ into (3.24), we see that
ZZ
(u1 − u2 )2 dxdt = 0,
QT

from which it follows that u1 = u2 . The proof is complete. 

3.4 Approximate Holmgren’s Approach

Consider weakly degenerate equations in one dimension of the form

∂u ∂ 2 A(u) ∂B(u)
= + , (3.25)
∂t ∂x2 ∂x
where A(s), B(s) ∈ C 1 (R), A(0) = 0, A0 (s) ≥ 0, however the set E = {s; A0 (s) = 0} does not
contain any interior point. Only the first boundary value problem for (3.25) will be discussed in
detail, the corresponding initial and boundary conditions are

u(0, t) = u(1, t) = 0, (3.26)

u(x, 0) = u0 (x). (3.27)

Denote QT = (0, 1) × (0, T ).

Definition 3.2 A function u ∈ L1 (QT ) is called a generalized solution of the boundary value
problem (3.25)–(3.27), if A(u), B(u) ∈ L1 (QT ) and the integral equality
ZZ   Z 1
∂ϕ ∂2ϕ ∂ϕ
u + A(u) 2 − B(u) dxdt + u0 (x)ϕ(x, 0)dx = 0
QT ∂t ∂x ∂x 0
Nonlinear Diffusion Equations 33

is fulfilled for any function ϕ ∈ C ∞ (QT ) withTϕ(0, t) = ϕ(1, t) = ϕ(x, T ) = 0.

It is easy to check that if u ∈ C 1 (QT ) C 2 (QT ) is a generalized solution of the boundary
value problem (3.25)–(3.27), then u is a classical solution of the problem. Here we notice that
A(u(x, t))|x=0,1 = 0 implies u(x, t)|x=0,1 = 0 because of the strict monotonicity of A(s).
Theorem 3.6 Assume that u0 ∈ L∞ (0, 1), A(s), B(s) ∈ C 1 (R) with A0 (s) ≥ 0 and the set
E = {s; A0 (s) = 0} has no interior point. Then the first boundary value problem (3.25)–(3.27) has
at most one bounded and measurable solution.
We will prove the theorem by means of Holmgren’s approach. The crucial step is to establish
the L1 estimate for the derivatives of the solutions of the adjoint equation. Our proof will be
completed by using this estimate, together with some L2 -type estimates for the solutions.
Let u1 , u2 ∈ L∞ (QT ) be solutions of the boundary value problem (3.25)–(3.27). By the
definition of generalized solutions, we have
ZZ  
∂ϕ e ∂2ϕ e ∂ϕ
(u1 − u2 ) +A 2 −B dxdt = 0
QT ∂t ∂x ∂x

for any ϕ ∈ C ∞ (QT ) with ϕ(0, t) = ϕ(1, t) = ϕ(x, T ) = 0, where

Z 1
e = A(u
A e 1 , u2 ) = A0 (θu1 + (1 − θ)u2 )dθ,
0
Z 1
e = B(u
B e 1 , u2 ) = B 0 (θu1 + (1 − θ)u2 )dθ.
0

If for any f ∈ C0∞ (QT ), the adjoint problem

2
∂ϕ e∂ ϕ − B e ∂ϕ = f,
+A 2
∂t ∂x ∂x
ϕ(0, t) = ϕ(1, t) = 0,

ϕ(x, T ) = 0

had a solution in C ∞ (QT ), then we would have

ZZ
(u1 − u2 )f dxdt = 0
QT

and the uniqueness would follow from the arbitrariness of f . However, since the coefficients A e
e
and B are merely bounded and measurable, it is difficult to discuss the solvability of the adjoint
problem directly. Even if we have established the existence of solutions, the solutions is not smooth
in general. This situation competes us to consider some approximation of the adjoint equation.
For sufficiently small η > 0 and δ > 0, let

e −1/2 B,
 (η + A) e if |u1 − u2 | ≥ δ,
λδη =
 0, if |u1 − u2 | < δ.

Since A(s) is strictly increasing and u1 , u2 ∈ L∞ (QT ), there must be constants L(δ) > 0, K(δ) > 0
depending on δ, but independent of η, such that

Ae = A(u1 ) − A(u2 ) ≥ L(δ), whenever |u1 − u2 | ≥ δ,

u1 − u 2
δ
|λη | ≤ K(δ).
34 Uniqueness of Solutions

eε and λδη,ε be a C ∞ approximation of A

Let A e and λδη respectively, such that

eε = A,
lim A e a.e. in QT ,
ε→0

lim λδη,ε = λδη , a.e. in QT ,

ε→0
eε ≤ C,
A

|λδη,ε | ≤ K(δ).

Denote
eη,ε
B δ eε )1/2 .
= λδη,ε (η + A
For given f ∈ C0∞ (QT ), consider the approximate adjoint problem
2
∂ϕ eε ) ∂ ϕ − B δ ∂ϕ
eη,ε
+ (η + A 2
= f, (3.28)
∂t ∂x ∂x
ϕ(0, t) = ϕ(1, t) = 0, (3.29)
ϕ(x, T ) = 0. (3.30)

Lemma 3.1 The solution ϕ of (3.28)–(3.30) satisfies

sup |ϕ(x, t)| ≤ C, (3.31)

QT
ZZ  2
eε ) ∂2ϕ
(η + A dxdt ≤ K(δ)η −1 , (3.32)
QT ∂x2
ZZ  2
∂ϕ
dxdt ≤ K(δ)η −1 . (3.33)
QT ∂x

Here and in the sequel, we use C to denote a universal constant, independent of δ, η and ε, and
K(δ) a constant, depending only on δ, which may take different values on different occasions.

Proof. (3.31) follows from the maximum principle. To prove (3.32) and (3.33), we multiply
∂2ϕ
(3.28 by and integrate over QT . Integrating by parts and using (3.29), (3.30) yield
∂x2
Z  2 ZZ    ∂ 2 ϕ 2
1 1 ∂ϕ(x, 0) eε
dx + η+A dxdt
2 0 ∂x QT ∂x2
ZZ 2 ZZ
e δ ∂ϕ ∂ ϕ dxdt = ∂2ϕ
− B η,ε 2
f 2 dxdt.
QT ∂x ∂x QT ∂x

Using Young’s inequality and noticing that |λδη,ε | ≤ K(δ), we obtain

ZZ  2
eε ) ∂2ϕ
(η + A dxdt
QT ∂x2
ZZ 2 ZZ
eε )1/2 ∂ϕ ∂ ϕ dxdt + ∂2ϕ
≤ λδη,ε (η + A f dxdt
QT ∂x ∂x2 QT ∂x2
ZZ  2 2 ZZ
1 e ∂ ϕ ∂ϕ
≤ (η + Aε ) 2
dxdt + K(δ) ( )2 dxdt + Cη −1 . (3.34)
4 QT ∂x QT ∂x
Nonlinear Diffusion Equations 35

Using (3.29) and Young’s inequality again gives

ZZ ZZ
∂ϕ 2 ∂2ϕ
( ) dxdt = − ϕ 2 dxdt
QT ∂x QT ∂x
ZZ  2 2
≤α eε ) ∂ ϕ dxdt + Cα−1 η −1
(η + A (3.35)
QT ∂x2

for any α > 0. Substituting this into (3.34) and choosing α > 0 small enough, we derive (3.32).
(3.33) follows from (3.32) and (3.35). The proof is complete. 

Lemma 3.2 The solution ϕ of (3.28)–(3.30) satisfies

Z 1
∂ϕ(x, t)
sup dx ≤ C, (3.36)
0<t<T 0 ∂x
where the constant C is independent of δ, η and ε.

Proof. For small β > 0, let



 1, if s ≥ β,

 Z s
s
sgnβ s = , if |s| < β, Iβ (s) = sgnβ θdθ.

 β 0


−1, if s ≤ −β,

∂ϕ
Differentiate (3.28) with respect to x, multiply the resulting equality by sgnβ and integrate over
∂x
St ≡ (0, 1) × (t, T ). Then we obtain
ZZ
∂ ∂ϕ
Iβ ( )dxdτ
St ∂t ∂x
ZZ 2
∂ eε ) ∂ ϕ ]sgnβ ( ∂ϕ )dxdτ
+ [(η + A
∂x ∂x2 ∂x
Z ZS t
∂ e δ ∂ϕ ∂ϕ
− [B ]sgnβ ( )dxdτ
∂x η,ε ∂x ∂x
Z ZSt
∂ϕ ∂f
= sgnβ ( ) dxdτ.
St ∂x ∂x

Hence, integrating by parts and using (3.30) yield

Z 1
∂ϕ(x, t)
Iβ ( )dx
0 ∂x
ZZ 2
=− (η + A eε )( ∂ ϕ )2 sgn0 ( ∂ϕ )dxdτ
β
St ∂x2 ∂x
ZZ 2
+ Be δ ∂ϕ ∂ ϕ sgn0 ( ∂ϕ )dxdτ
η,ε β
St ∂x ∂x2 ∂x
Z T  2

e ∂ ϕ e δ ∂ϕ ∂ϕ x=1
+ (η + Aε ) 2 − Bη,ε sgnβ ( ) dτ
∂x ∂x ∂x x=0
ZtZ
∂ϕ ∂f
− sgnβ ( ) dxdτ. (3.37)
St ∂x ∂x
36 Uniqueness of Solutions

The first term on the right side is nonpositive. The last term is bounded. Using (3.28), (3.29) and
the fact f ∈ C0∞ (QT ) we see that
 2

x=1
(η + Aeε ) ∂ ϕ − B e δ ∂ϕ sgn ( ∂ϕ )
2 η,ε β
∂x ∂x ∂x x=0
 
∂ϕ ∂ϕ x=1
= f− sgnβ ( ) =0
∂t ∂x x=0

which shows that the third term on the right side of (3.37) vanishes. Therefore we have
Z 1   Z T Z
∂ϕ(x, t) ∂2ϕ
Iβ dx ≤ C(η, ε, δ) dτ | |dx + C
0 ∂x t [0,1]
T
{x;|∂ϕ/∂x|≤β} ∂x2

from which (3.36) follows by letting β → 0 and using a known result (see S. Sakes, Theory of
Integration, 131–133) to conclude that
Z
∂2ϕ
| |dx → 0 as β → 0.
[0,1]
T
{x;|∂ϕ/∂x|≤β} ∂x2

Proof of Theorem 3.6. Given f ∈ C0∞ (QT ). Let ϕ be a solution of (3.28)–(3.30). Then
ZZ
(u1 − u2 )f dxdt
QT
ZZ  2

∂ϕ eε ) ∂ ϕ − B
e δ ∂ϕ dxdt.
= (u1 − u2 ) + (η + A η,ε
QT ∂t ∂x2 ∂x

As indicated above, from the definition of generalized solutions, we have

ZZ  2

∂ϕ e∂ ϕ − B e ∂ϕ dxdt = 0.
(u1 − u2 ) +A
QT ∂t ∂x2 ∂x

Thus
ZZ
(u1 − u2 )f dxdt
QT
ZZ ZZ
∂2ϕ 2
e ∂ ϕ dxdt
eε − A)
= (u1 − u2 )η 2 dxdt + (u1 − u2 )(A
QT ∂x QT ∂x2
ZZ
− e δ − B)
(u1 − u2 )(B e ∂ϕ dxdt. (3.38)
η,ε
QT ∂x

Now we are ready to estimate all terms on the right side of (3.38).
First, from Lemma 3.1,
ZZ 2
(u1 − u2 )(A e ∂ ϕ dxdt
eε − A)
QT ∂x2
Z Z 1/2 Z Z 1/2
e e 2 ∂2ϕ 2
≤ C (Aε − A) dxdt ( 2 ) dxdt
QT QT ∂x
Z Z 1/2
≤ K(δ)η −1 (Aeε − A)
e 2 dxdt .
QT
Nonlinear Diffusion Equations 37

Hence ZZ 2
lim (u1 − u2 )(A e ∂ ϕ dxdt = 0.
eε − A) (3.39)
ε→0 QT ∂x2
Denote
Gδ = {(x, t) ∈ QT ; |u1 − u2 | < δ} ,

Fδ = {(x, t) ∈ QT ; |u1 − u2 | ≥ δ} .
Using Lemma 3.1 and 3.2, we have
ZZ
δ ∂ϕ
(u1 − u2 )(B̃ηε − B̃) dxdt
QT ∂x
ZZ ZZ
δ ∂ϕ ∂ϕ
≤ (u1 − u2 )B̃ηε dxdt + (u1 − u2 )B̃ dxdt
Gδ ∂x Gδ ∂x
ZZ
δ ∂ϕ
+ (u1 − u2 )(B̃ηε − B̃) dxdt
Fδ ∂x
Z Z 1/2 Z Z 1/2
δ 2 ∂ϕ 2
≤C (ληε ) dxdt ( ) dxdt +
Gδ Gδ ∂x
Z Z 1/2 Z Z 1/2
δ 2 ∂ϕ 2
+C (B̃ηε − B̃) dxdt ( ) dxdt + Cδ
Fδ Fδ ∂x
Z Z 1/2
−1 δ 2
≤CK(δ)η (ληε ) dxdt + Cδ+
Gδ
Z Z 1/2
+ CK(δ)η −1 δ
(B̃ηε − B̃)2 dxdt .
Fδ

eδ = B
Since lim λδη,ε = 0 a.e. on Gδ , and lim B e a.e. on Fδ , it follows that
η,ε
ε→0 ε→0
ZZ
e δ − B)
e ∂ϕ
lim (u1 − u2 )(B η,ε dxdt ≤ Cδ. (3.40)
ε→0 QT ∂x
On the other hand, for any fixed γ > 0, using the definition of Fδ , Gδ and Cauchy’s inequality,
we have
ZZ
∂2ϕ
η(u1 − u2 ) 2 dxdt
QT ∂x
ZZ ZZ
∂2ϕ ∂2ϕ
≤ η(u1 − u2 ) 2 dxdt + η(u1 − u2 ) 2 dxdt
Fγ ∂x Gγ ∂x
ZZ ZZ 2
!1/2
2 2
e−1/2 e1/2 ∂ ϕ ∂ ϕ
≤Cη sup |A | A dxdt + Cηγ dxdt
Fγ Fγ ∂x2 QT ∂x
2
ZZ 2
e−1/2 | e1/2 e 1/2 ∂ ϕ
≤Cη sup |A (A ε − A ) dxdt+
Fγ Fγ ∂x2
ZZ 2
+ Cη sup |A e−1/2
| e1/2 ∂ ϕ dxdt + CγK(δ)1/2
A ε
Fγ Fγ ∂x2
ZZ !1/2 Z Z 1/2
e−1/2 e −A
1/2 e ) dxdt
1/2 2 ∂2ϕ 2
≤Cη sup |A | (A ε ( 2
) dxdt +
Fγ Fγ QT ∂x
38 Uniqueness of Solutions

ZZ  2 !1/2
Cη eε ∂2ϕ
+ A dxdt + CγK(δ)1/2
L(γ)1/2 QT ∂x2
ZZ !1/2
K(δ)1/2 e1/2 − A
e1/2 )2 dxdt K(δ)1/2
≤C (A ε + Cη 1/2 + CγK(δ)1/2
L(γ)1/2 Fγ L(γ)1/2

eε = A
Since lim A e a.e on QT , letting ε → 0 in the above inequality yields
ε→0
ZZ
∂2ϕ 1/2 K(δ)
1/2
lim η(u1 − u2 ) dxdt ≤ Cη + CγK(δ)1/2 .
ε→0 QT ∂x2 L(γ)1/2

δ
It follows by setting γ = that and hence
K(δ)1/2
ZZ
∂2ϕ K(δ)1/2
lim η(u1 − u2 ) 2
dxdt ≤ Cη 1/2 + Cδ. (3.41)
ε→0 QT ∂x L(δ/K(δ)1/2 )1/2

Combining (3.38)–(3.41) we finally obtain

ZZ
K(δ)1/2
(u1 − u2 )f dxdt ≤ Cη 1/2 + Cδ,
QT L(δ/K(δ)1/2 )1/2

which implies that ZZ

(u1 − u2 )f dxdt = 0
QT

by letting η → 0 and then δ → 0. The proof of our theorem is complete. 

3.5 Approach by Characteristic Operators

Consider equations of the form
∂u
= ∆A(u) (3.42)
∂t
with A0 (s) ≥ 0 and E = {s; A0 (s) = 0} being allowed to have interior points. Assume the initial
value condition
u(x, 0) = u0 (x). (3.43)
The solutions of the Cauchy problem (3.42), (3.43) has already been defined. We will see below
that, just as in the case of weakly degenerate equations, the formulation of the Cauchy problem
for strongly degenerate equations of the form (3.42) with solutions defined by integral equality is
correct, namely, solutions of the problem thus defined are uniquely determined.

Theorem 3.7 Assume that A(s) ∈ C 1 (R) and A0 (s) ≥ 0. Then the Cauchy problem (3.42),
(3.43) admits at most one bounded and integrable generalized solution.

Proof. Let u1 and u2 be bounded and integrable generalized solutions of (3.42), (3.43).
Denote z = u1 − u2 , v = A(u1 ) − A(u2 ). Then from the definition of generalized solutions, we have
ZZ  
∂ϕ
z + v∆ϕ dxdt = 0 (3.44)
QT ∂t

for any ϕ ∈ C0∞ (QT ).

Nonlinear Diffusion Equations 39

We wish to assert that z ≡ 0 from this equality by choosing special test functions ϕ.
From the Lp theory for elliptic equations, for any f ∈ Lp (RN ) (1 ≤ p ≤ ∞), the problem

−∆u + λu = f, in D0 (RN ) (λ > 0) (3.45)

has a unique solution uλ ∈ Lp (RN ). Defining an operator Tλ by uλ = Tλ f , one has the estimate

λkTλ f kp ≤ kf kp , (3.46)

where k · kp will denote either the norm of Lp (RN ) or the norm of Lp (QT ) depending on the
context. Because of (3.46), Tλ also defines an operator Tλ : Lp (QT ) → Lp (QT ) (1 ≤ p ≤ ∞)
and (3.46) holds equally for f ∈ Lp (RN ) and f ∈ Lp (QT ). For any γ ∈ C0∞ (QT ), we wish to set
ϕ = Tλ γ in (3.44). Since Tλ commutes with differentiations, we have Tλ γ ∈ C ∞ (QT ). Clearly
(Tλ γ)(x, t) = 0 for t near 0 and T . Moreover, since z, v ∈ L∞ (QT ), (3.44) clearly continues to
T ∂ϕ
hold for ϕ ∈ C ∞ (QT ) L1 (QT ) with ϕ(x, t) = 0 for t near 0 and T provided that , ∆ϕ,
∂t
∂Tλ γ ∂γ
|∇ϕ| ∈ L1 (QT ). ϕ = Tλ γ has these properties. Moreover, ∆Tλ γ = λTλ γ − γ, = Tλ .
∂t ∂t
Thus, from (3.44), it follows that for γ ∈ C0∞ (QT ),
ZZ  
∂γ
zTλ + v (λTλ γ − γ) dxdt
∂t
Z ZQ T  
∂γ
= Tλ z + (λTλ v − v) γ dxdt = 0,
QT ∂t

where the first equality is due to the obvious symmetry of Tλ and the absolute convergence of all
integrals involved. This means that in the sense of distributions,

∂Tλ z
= λTλ v − v. (3.47)
∂t
Denote Z
gλ (t) = z(x, t)Tλ z(x, t)dx.
RN

Since z, Tλ z ∈ L∞ (QT ) ∩ L1 (QT ), gλ (t) is well defined for almost all t ∈ (0, T ). If we can prove
that
lim gλ (t) = 0, a.e. t ∈ (0, T ), (3.48)
λ→0

then, since Z
gλ (t) = (λTλ z − ∆Tλ z) Tλ zdx = λkTλ z(·, t)k22 + k∇Tλ z(·, t)k22 ,
RN

we have, for almost all t ∈ (0, T ),

lim λTλ z(·, t) = 0, lim ∇Tλ z(·, t) = 0 in L2 (RN ).

λ→0 λ→0

It follows that, in particular, in the sense of distributions,

lim λTλ z(·, t) = 0, lim ∆Tλ z(·, t) = 0

λ→0 λ→0

and hence
z(·, t) = lim (λTλ z(·, t) − ∆Tλ z(·, t)) = 0.
λ→0
40 Uniqueness of Solutions

This shows that z = u1 − u2 = 0 a.e in QT .

Now we turn to the proof of (3.48). Let Jε be the standard mollifier in t. Then from (3.47)
and the symmetry of Tλ , we have
Z Z
d ∂
Tλ (Jε z)Jε zdx = 2 Tλ (Jε z)Jε zdx
dt RN RN ∂t
Z   Z
∂
= 2 Jε Tλ z Jε zdx = 2 Jε (λTλ v − v) Jε zdx.
RN ∂t RN

Integrating with respect to t gives

Z
Tλ (Jε z(x, t))Jε z(x, t)dx
N
ZR ZZ
= Tλ (Jε z(x, 0))Jε z(x, 0)dx + 2 Jε (λTλ v − v) Jε zdxds.
RN Qt

Letting ε → 0 and noticing that the nondecreasingness of A(s) implies z(x, t)w(x, t) ≥ 0, we derive
ZZ
gλ (t) = 2 (λTλ v − v) zdxds
QZt Z ZZ (3.49)
≤ 2λ zTλ v = 2λ vTλ zdxds.
Qt Qt

Since z ∈ L∞ (QT ) and A(s) ∈ C 1 (R) imply |v| ≤ C|z|, we further obtain
ZZ
gλ (t) ≤ Cλ |zTλ z|dxds. (3.50)
Qt

From (3.45) it is easily seen that

Z Z Z
2 2
|∇Tλ f | dx + λ |Tλ f | dx ≤ C |f |2 dx.
RN RN RN

Using Sobolev’s inequality gives

Z Z
2
|Tλ f | dx ≤ C |f |2 dx.
RN RN

Thus from (3.50) it follows that

Z
gλ (t) ≤ Cλ |z|2 dx
RN

which and z ∈ L2 (QT ) imply (3.48). The proof of Theorem 3.7 is complete. 

In order to prove the uniqueness of bounded and integrable solutions of (3.42), (3.43) under
the assumption weaker than A(s) ∈ C 1 (R), we will apply the following lemma.

Lemma 3.3 Let Tλ be the operator introduced in the proof of Theorem 3.7 and f ∈ L ∞ (RN ) ∩
L (RN ). Then
1

lim λkTλ f k∞ = 0.
λ→0
Nonlinear Diffusion Equations 41

Proof. Let K(x) be a solution of the equation

−∆u + u = δ(x)

where δ(x) is the Dirac function. Then

C(r) ≡ sup K(x) < ∞, K ∈ L1 (B),

|x|≥r

where B = {x; |x| ≤ 1}. Using K(x), it is not difficult to verify that the solution Tλ f of (3.45) can
be expressed by Z
n/2
√
λ(Tλ f )(x) = λ K( λ(x − y))f (y)dy.
RN

Thus, for any fixed r > 0,

Z √
λ|(Tλ f )(x)| ≤ λn/2 √ |K( λ(x − y))f (y)|dy
{ λ|x−y|≥r}
Z √
+λn/2 √ |K( λ(x − y))f (y)|dy
{ λ|x−y|≤r}
Z
≤ λn/2 C(r)kf k1 + kf k∞ K(y)dy.
{|y|≤r}

Letting λ → 0 gives Z
lim λ|(Tλ f )(x)| ≤ kf k∞ K(y)dy,
λ→0 {|y|≤r}

and the conclusion of Lemma 3.3 by setting r → 0. 

Theorem 3.8 Assume that A(s) ∈ C(R) and A(s) is nondecreasing. Then the Cauchy problem
(3.42), (3.43) admits at most one bounded and integrable generalized solution.

Proof. Checking the proof of Theorem 3.7, we see that, it suffices to verify (3.48) under the
assumption of the present theorem.
From the boundedness of u1 , u2 and the continuity of A(s), it is easy to see that for any ξ > 0,
there exists an η > 0 such that

{(x, t); |w(x, t)| > ξ} ⊂ {(x, t); |z(x, t)| > η}

up to a set of measure zero. This and z ∈ L1 (QT ) imply

C(ξ, t) ≡ mes {(x, t); |w(x, t)| > ξ} < ∞

for almost all t ∈ (0, T ).

Now we estimate the right hand side of (3.49) which still holds under the assumption A(s) ∈
C(R). For fixed ξ > 0, we have
Z
λ wTλ zdx
RN
ZZ ZZ
≤ λ wTλ z λ wTλ z
{x;|w(x,t)|>ξ} {x;|w(x,t)|≤ξ}
≤ C(ξ, t)kw(·, t)k∞ kλTλ z(·, t)k∞ + ξλkTλ z(·, t)k1
42 Uniqueness of Solutions

and hence, by Lemma 3.3,

Z
lim λ wTλ zdx ≤ ξλkTλ z(·, t)k1 .
λ→0 RN

Thus Z
lim λ wTλ zdx = 0 (3.51)
λ→0 RN

for almost all t ∈ (0, T ).

On the other hand,
Z Z
λ w(x, t)Tλ z(x, t)dx = λ z(x, t)Tλ w(x, t)dx
RN RN (3.52)
≤ kλTλ w(x, t)k∞ kz(x, t)k1 ≤ kw(x, t)k∞ kz(x, t)k1 .

Letting λ → 0 in (3.49) and using (3.51), (3.52) and the dominated convergence theorem yield
(3.48) and complete the proof of our theorem. 

3.6 Approach by the Theory of BV Spaces

In this section, we consider equations of the form
∂u ~
= ∆A(u) + divB(u), (3.53)
∂t
where Z u Z u
A(u) = a(s)ds, ~
B(u) = ~b(s)ds
0 0

with a(s) ≥ 0, a(s) and ~b(s) being continuous and E = {s; a(s) = 0} containing no interior point.
Without loss of generality, we may assume that A(0) = 0. We will discuss the Cauchy problem for
(3.53) with initial value condition
u(x, 0) = u0 (x). (3.54)
Denote QT = RN × (0, T ).

Definition 3.3 A function u ∈ L1loc (QT ) is called a generalized solution of the Cauchy problem
~
(3.53)–3.54), if A(u) ∈ L1loc (QT ), B(u) ∈ L1loc (QT ), and for any ϕ ∈ C0∞ (QT ),
ZZ ZZ ZZ
∂ϕ ~
u dxdt + A(u)∆ϕdxdt − B(u)∇ϕdxdt = 0;
QT ∂t QT QT

furthermore, for any h ∈ C0∞ (RN ),

Z Z
lim u(x, t)h(x)dx = u0 (x)h(x)dx.
t→0+ RN RN

If in addition, u ∈ L∞ (QT ) ∩ BV (QT ), with ∇A(u) ∈ L2loc (QT ), then we will simply say that
u is a BV solution.

Here and below, we denote by BV (QT ) the set of all functions of locally bounded variation on
QT . In the appendix of this chapter, we will list all results on the class BV (QT ) and more general
class BVx (QT ) needed in this chapter without proofs.
Nonlinear Diffusion Equations 43

Theorem 3.9 Assume that E = {s; a(s) = 0} has no interior point. Let u1 , and u2 be BV
solutions of (3.53) with initial value u01 , and u02 respectively. Then for almost all t ∈ (0, T ),
Z Z
Kλ t
|u1 (x, t) − u2 (x, t)| ωλ (x)dx ≤ e u01 (x) − u02 (x) ωλ (x)dx,
RN RN

where λ > 0, Kλ is a constant depending only on λ and the bound of u1 , and u2 , and
 p 
ωλ (x) = exp −λ 1 + |x|2 .

We first prove the following lemma which plays an important role in the proof of Theorem 3.9.

Lemma 3.4 Assume that E = {s; a(s) = 0} has no interior point. Let u be a BV solution of
(3.53)–3.54). Then H(Γu ) = 0, where H(S) denotes the Hausdorff measure of S.

Proof. Denote

Γ1 = {(x, t) ∈ Γu ; γ1 (x, t) = · · · = γN (x, t) = 0} ,


Γ2 = (x, t) ∈ Γu ; γ12 (x, t) + · · · + γN
2
(x, t) > 0 .

Clearly, Γu = Γ1 ∪ Γ2 .
First prove H(Γ1 ) = 0. Since any measurable subset of Γ1 can be expressed as the union of a
Borel set and a set of measure zero, it suffices to prove H(S) = 0 for any Borel subset S ⊂ Γ 1 .
We may suppose that S is compact. By Lemma 7.8, for any bounded function f (x, t) which is
∂u
measurable with respect to the measure , we have
∂xi
ZZ Z T Z
∂u ∂u(·, t)
f (x, t) = dt f (x, t) , (3.55)
S ∂xi 0 St ∂xi

where S t = {x; (x, t) ∈ S}. Since by Lemma 7.4, for any Borel subset S1 ⊂ S,
Z
∂u
(S1 ) = (u+ (x, t) − u− (x, t))γi dH,
∂xi S1
Z
∂u(·, t) t
(S1 ) = (ut+ (x, t) − ut− (x, t))γit dH t ,
∂xi t
S1

3.55) is equivalent to
ZZ
f (x, t)(u+ (x, t) − u− (x, t))γi dH
S
Z T Z
= dt f (x, t)(ut+ (x, t) − ut− (x, t))γit dH t .
0 St

The definition of Γ1 implies that the left hand side vanishes, so we have
Z T Z
dt f (x, t)(ut+ (x, t) − ut− (x, t))γit dH t = 0.
0 St

Choose
f (x, t) = χS (x, t)sgn(ut+ (x, t) − ut− (x, t))sgnγit
44 Uniqueness of Solutions

in the above equality, where χS (x, t) denotes the characteristic function of S, and sum up for i
from 1 up to N . Then we obtain
Z Z


dt ut+ (x, t) − ut− (x, t) |γ1t | + · · · + |γN
t
| dH t = 0, (3.56)
G St

where G is the projection of S on the t-axis. (3.56) implies that for almost all t ∈ G,
Z


ut+ (x, t) − ut− (x, t) |γ1t | + · · · + |γN
t
| dH t = 0,
St

and hence for almost all t ∈ G,

γ1t = · · · = γN
t
=0
H t -almost everywhere on S t , which is impossible unless mesG = 0.
For any α, β with 0 < α < β < T , we can choose ψj (t) ∈ C0∞ (0, T ) such that

0 ≤ ψj (t) ≤ 1, lim ψj (t) = χ(α,β] (t) ∀ t ∈ (0, T ).

j→∞

By Lemma 7.6, we can choose ϕk (x, t) ∈ C0∞ (QT ) such that

 
1 ∂u
|ϕk (x, t)| ≤ 1, lim ϕk = χS in L QT , .
k→∞ ∂t

Now from the definition of BV solutions, we have

ZZ
∂u
ϕk (x, t)ψj (t)
∂t
Z ZQ T ZZ
= A(u)ψj (t)∆ϕk dxdt − ~
B(u)ψ j (t)∇ϕk dxdt.
QT QT

Letting j → ∞ leads to
ZZ
∂u
ϕk (x, t)χ(α,β] (t)
∂t
Z ZQ T ZZ
= A(u)χ(α,β] (t)∆ϕk dxdt − ~
B(u)χ (α,β] (t)∇ϕk dxdt.
QT QT

Clearly, this equality also holds if (α, β] is replaced by (α, β) and hence it holds even if (α, β] is
replaced by any open set I with I ⊂ (0, T ). Since G is a Borel set, by approximation, we may
conclude that ZZ
∂u
ϕk (x, t)χG (t)
∂t
Z ZQ T ZZ
= A(u)χG (t)∆ϕk dxdt − ~
B(u)χ G (t)∇ϕk dxdt.
QT QT

Since mesG = 0, the two terms on the right hand side vanish and
ZZ
∂u
ϕk (x, t)χG (t) = 0.
QT ∂t

Letting k → ∞ gives ZZ ZZ
∂u ∂u
= χS (x, t)χG (t) = 0.
S ∂t QT ∂t
Nonlinear Diffusion Equations 45

Hence Z
(u+ (x, t) − u− (x, t))γt dH = 0 (3.57)
S
which implies H(S) = 0 and H(Γ1 ) = 0 by the arbitrariness of S.
Next we prove H(Γ2 ) = 0. Let S be any Borel subset of Γ2 , which is compact in QT . Since S
is a set of N + 1-dimensional measure zero and ∇A(u) ∈ L2loc (QT ), we have
ZZ
∂
A(u)dxdt = 0, i = 1, · · · , N,
S ∂xi

and hence Z
(A(u+ (x, t)) − A(u− (x, t)))γi dH = 0, i = 1, · · · , N.
S
From this it follows by the definition of Γ2 and the strict monotonicity of A(S) that H(S) = 0 and
hence H(Γ2 ) = 0 by the arbitrariness of S. Thus the lemma is proved. 
Proof of Theorem 3.9. By the definition of generalized solutions, we have
ZZ
∂ϕ ~ 1 ) − B(u
~ 2 ))∇ϕ}dxdt = 0
{(u1 − u2 ) + (A(u1 ) − A(u2 )) 4 ϕ − (B(u
QT ∂t
or ZZ
∂ϕ ~
(z + αz∆ϕ − βz∇ϕ)dxdt =0 (3.58)
QT ∂t
for any ϕ ∈ C0∞ (QT ), where
Z 1 Z 1
z = u 1 − u2 , α= a(λu1 + (1 − λ)u2 )dλ, β~ = ~b(λu1 + (1 − λ)u2 )dλ.
0 0

Here for convenience of the following discussion we have replaced ui by ui , the symmetric mean
value of ui . Note that doing so does not change the value of the related integrals. Since ui (i = 1, 2)
are BV solutions, from the properties of BV functions we see that

αz ∈ BV (QT ), ~ ∈ BV (QT )
βz

and (3.58) can be written as

ZZ
∂z ~
(−ϕ − ∇(αz)∇ϕ + ϕdiv(βz))dxdt =0 (3.59)
QT ∂t

for any ϕ ∈ C0∞ (QT ).

The crucial step in proving Theorem 3.9 is to establish the following inequality
ZZ  
∂ϕ ~
J(u1 , u2 , ϕ) ≡ sgnz z − ∇(αz)∇ϕ − βz∇ϕ dxdt ≥ 0 (3.60)
QT ∂t

for any 0 ≤ ϕ ∈ C0∞ (QT ).

To this purpose, we define Z s
Hε (s) = hε (ρ)dρ
0
for small ε > 0, where  
2 |s|
hε (s) = 1− .
ε ε +
46 Uniqueness of Solutions

Obviously hε ∈ C(R) and for all s ∈ R,

hε (s) ≥ 0, |shε (s)| ≤ 1, |Hε (s)| ≤ 1,

lim Hε (s) = sgns, lim shε (s) = 0.

ε→0 ε→0

From the properties of BV solutions, one has

Hε (αz) ∈ BV (QT ), ∇Hε (αz) ∈ L2loc (QT ).

Instead of J(u1 , u2 , ϕ), we first consider

ZZ  
∂ϕ ~
Jε (u1 , u2 , ϕ) ≡ Hε (αz) z − ∇(αz)∇ϕ − βz∇ϕ dxdt.
QT ∂t

By the strict monotonicity of A(s), αz = A(u1 ) − A(u2 ) and z = u1 − u2 have the same sign, so
sgnz = sgn(αz) and hence
lim Jε (u1 , u2 , ϕ) = J(u1 , u2 , ϕ).
ε→0

Replace ϕ by Hε (αz)ϕ in 3.59), (This is possible by approximation)

ZZ
∂z ~
(−Hε (αz)ϕ − ∇(αz)∇(Hε (αz)ϕ) + Hε (αz)ϕdiv(βz))dxdt = 0.
QT ∂t
Using this we obtain
Jε (u1 , u2 , ϕ)
ZZ ZZ
∂ ~
= Hε (αz) (ϕz) − Hε (αz)div(ϕβz)
QT ZZ ∂t QT
2
+ ĥε (αz) |∇(αz)| ϕdxdt
ZZ Q T ZZ
∂ ~
≥ Hε (αz) (ϕz) − Hε (αz)div(ϕβz)
QT ∂t QT

for 0 ≤ ϕ ∈ C0∞ (QT ), where ĥε (w) is the functional superposition of hε (s) and w(x, t), namely,
Z 1
ĥε (w) = hε (λw+ + (1 − λ)w− )dλ.
0

Hence

J(u1 , u2 , ϕ) = lim Jε (u1 , u2 , ϕ)

ε→0
ZZ ZZ
∂ ~
≥ sgn(αz) (ϕz) − sgn(αz)div(ϕβz)
QT ∂t QT
ZZ ZZ
∂ ~
= sgn(z) (ϕz) − sgn(z)div(ϕβz). (3.61)
QT ∂t QT

By Lemma 3.4, we have

H(Γu1 ) = H(Γu2 ) = 0
which implies, in particular, up to a set of N -dimensional measure zero ( Note that QT is an
N + 1-dimensional set),
Hε (z) = Hε (z), hε (z) = hε (z) = ĥε (z).
Nonlinear Diffusion Equations 47

Also notice that since z ∈ BV (QT ), we must have

∂ ~ (G) = 0
(ϕz) (G) = 0, div(ϕβz)
∂t
for any set G of N -dimensional measure zero. Thus from (3.61) and the fact that lim shε (s) = 0,
ε→0
we obtain
J(u1 , u2 , ϕ)
ZZ ZZ
∂ ~
≥ lim Hε (z) (ϕz) − lim Hε (z)div(ϕβz)
ε→0 QT ∂t ε→0 QT
ZZ ZZ
∂ ~
= lim Hε (z) (ϕz) − lim Hε (z)div(ϕβz)
ε→0 QT ∂t ε→0 QT
ZZ ZZ
∂ ~
= − lim ϕz Hε (z) + lim ϕβz∇H ε (z)
ε→0 QT ∂t ε→0 QT
ZZ ZZ
∂z ~ ĥε (z)∇z
= − lim ϕz ĥε (z) + lim ϕβz
ε→0 QT ∂t ε→0 QT
= 0
and (3.60) is proved.
Now we use (3.60) to complete the proof of our theorem. Note that
ZZ
sgnz∇(αz)∇ϕdxdt
QT
ZZ
= lim Hε (αz)∇(αz)∇ϕdxdt
ε→0 QT
ZZ
= lim ∇ [αzHε (αz)] ∇ϕdxdt
ε→0 QT
ZZ
− lim αzhε (αz)∇(αz)∇ϕdxdt
ε→0 QT
ZZ
= − lim αzHε (αz)∆ϕdxdt
ε→0 QT
ZZ
= − αzsgn(αz)∆ϕdxdt
QT
ZZ
= − α|z|∆ϕdxdt.
QT

From this and (3.60) we see that for 0 ≤ ϕ ∈ C0∞ (QT ),

ZZ ZZ ZZ
∂ϕ ~
|z| dxdt ≥ β|z|∇ϕdxdt − α|z|∆ϕdxdt. (3.62)
QT ∂t QT QT

Given τ , s ∈ (0, T ), τ < s. Let

Z s−t
ψε (t) = αε (σ)dσ ε < min{τ, T − s},
τ −t

where αε (t) is the kernel of a mollifier with αε (t) = 0 for t ∈/ (−ε, ε). In particular, choose
ϕR ∈ C0∞ (RN ) such that ϕR (x) = 1 for |x| < R, ϕR (x) = 0 for |x| > R + 1 and
0 ≤ ϕR (x) ≤ 1, |∇ϕR (x)| ≤ C, |∆ϕR (x)| ≤ C, f or x ∈ RN .
48 Uniqueness of Solutions

Take ϕ = ψε (t)ϕR (x)ωλ (x) in 3.62) and notice that

|∇ωλ (x)| ≤ Cλ ωλ (x), |∆ωλ (x)| ≤ Cλ ωλ (x).

Then we obtain
Z Z T
ϕR (x)ωλ (x)dx |z(x, t)| [αε (s − t) − αε (τ − t)] dt
RN 0
Z s+ε Z
≤ Cλ dt |z(x, t)|ωλ (x)dx
0 RN

which implies, by letting ε → 0 and then R → ∞, that for almost all τ , s ∈ (0, T ),
Z
|z(x, s)|ωλ (x)dx
RNZ Z Z
s
≤ Cλ dt |z(x, t)|ωλ (x)dx + |z(x, τ )|ωλ (x)dx.
0 RN RN

Since BV functions have trace on the superplane t = 0 and the trace of ui must be u0i (x), we
further obtain Z
|z(x, s)|ωλ (x)dx
RN
Z s Z Z
≤ Cλ dt |z(x, t)|ωλ (x)dx + |z 0 (x, 0)|ωλ (x)dx
0 RN RN
0
by letting τ → 0, where z (x) = u01 (x) − u02 (x). Finally we may use Gronwall’s inequality to
complete the proof of the theorem. 
Nonlinear Diffusion Equations 49

4 Finite Propagation of Disturbances

4.1 Comparison Approach
Consider the equation
∂u ∂ 2 A(u)
= (4.1)
∂t ∂x2
with initial data
u(x, 0) = u0 (x) x ∈ R. (4.2)
2
We always assume that A(u) ∈ C 1 [0, ∞) ∩ C (0, ∞),

A(u) > 0, A0 (u) > 0, A00 (u) > 0 for u > 0,

A(0) = A0 (0) = 0,

and u0 is nonnegative, bounded and continuous on R with A(u0 ) satisfying the Lipschitz condition.
By virtue of existence and uniqueness theorems, the Cauchy problem (4.1), (4.2) admits exactly
one nonnegative, bounded and continuous generalized solution u on Q = R × (0, ∞) with bounded
∂A(u)
weak derivative . As we have seen from the Barenblatt solutions, for the porous medium
∂x
equation
∂u ∂ 2 um
= , (m > 1) (4.3)
∂t ∂x2
which corresponds to the slow diffusion, the speed of propagation of disturbances is finite. The
mathematical description of this fact is that if supp u0 is bounded, then for any t > 0, supp u(x, t)
is also bounded.
We have the following general result.

Theorem 4.1 Assume that for any u > 0,

Z u 0
A (s)
Ψ(u) = ds < +∞. (4.4)
0 s

Let u be a generalized solution of the Cauchy problem (4.1), (4.2) on Q. If supp u 0 (x) is bounded,
then for any t > 0, supp u(·, t) is also bounded.

Proof. Consider the function of the form

u(x, t) = Ψ−1 (c(ct + t1 ) − x)+ ).

It is easy to verify that for any t1 > 0, c > 0, u(x, t) is a generalized solution of (4.1) with initial
data
u(x, 0) = Ψ−1 (c(ct1 − x)+ ).
Let x0 = sup{supp u0 (x)}. Then

u(x, 0) = u0 (x) = 0 ≤ u(x, 0) for x ≥ x0 .

Besides, since Ψ and Ψ−1 are increasing, we have

u(x, t) = Ψ−1 (c(c(t + t1 ) − x0 )+ )

≥ Ψ−1 (c(ct1 − x0 )+ ) ≥ M = sup u,

Q
50 Finite Propagation of Disturbances

provided ct1 − x0 > 0 and c(ct1 − x0 ) ≥ Ψ(M ). Hence

u(x0 , t) ≤ M ≤ u(x0 , t) for t ≥ 0.

Now we apply the comparison theorem on GT = (x0 , ∞) × (0, T ) and then obtain

u(x, t) ≤ u(x, t), on GT ,

from which it follows that u(x, t) = 0 when x ≥ Xt = c(t + t1 ), since u(x, t) = 0 for x ≥ Xt =
c(t + t1 ).
Notice that in applying the comparison theorem on GT , we need to check u ∈ L1 (GT ). From
the definition of solutions, for any τ ∈ (0, T ) and ϕ ∈ C ∞ (QT ) which vanishes when |x| is large
enough, we have Z Z
u(x, τ )ϕ(x, τ )dx − u0 (x)ϕ(x, 0)dx
ZRZ  R

∂ϕ ∂2ϕ
= u + A(u) 2 dxdt.
Qτ ∂t ∂x
In particular, we take ϕ = ϕX ∈ C ∞ (R) such that

0 ≤ ϕX (x) ≤ 1, ϕX (x) = 1, for |x| ≤ X,

ϕX (x) = 0, for |x| ≥ X + 1, |ϕ00X (x)| ≤ C,

where the constant C is independent of X. Then we obtain

Z ZZ Z
u(x, τ )dx + A(u)ϕ00X (x)dxdt ≤ u0 (x)dx.
|x|≤X X≤|x|≤X+1 R
0≤t≤T

Z
Since the right hand side is bounded uniformly in X, letting X → ∞, we see that u(x, τ )dx is
R
bounded and hence u ∈ L1 (QT ).
Similarly we can prove that there exists Xt0 such that u(x, t) = 0 when x ≤ Xt0 . 

Remark 4.1 Theorem 4.1 shows that (4.4) is a sufficient condition for (4.1) to possess the
property of finite speed of disturbances. One can prove that (4.4) is also necessary for (4.1) to
possess such property. Oleinik guessed the necessity of this condition at a symposium held in
Moscow university. Since then people verified this supposition for some special cases or under
some additional condition. Finally Peletier gave a satisfactory result.

When A(u) = um , the condition (4.4) is equivalent to m > 1, the slow diffusion case.
Assume that supp u0 = [x1 , x2 ] (−∞ < x1 < x2 < +∞) and u is a generalized solution of (4.3),
(4.2) on Q = R × (0, ∞). Denote

Ω = {(x, t); u(x, t) > 0, t > 0}

Ω(t) = {x; u(x, t) > 0}

ζ1 (t) = inf Ω(t), ζ2 (t) = sup Ω(t).

By the continuity of u, Ω is an open set. Theorem 4.1 implies that for any t > 0, ζi (t) (i = 1, 2) is
finite. We call x = ζi (t) (i = 1, 2) the free boundary or interface of u.
Nonlinear Diffusion Equations 51

Theorem 4.2 Assume that m > 1. Then (−1)i ζi (t) (i = 1, 2) is increasing and

lim (−1)i ζi (t) = +∞, (i = 1, 2). (4.5)

t→∞

Theorem 4.2 means that in case of slow diffusion, disturbances will be propagated to infinite
scope, although the speed of propagation is finite.
Before proving Theorem 4.2, we first prove the following proposition which is also very useful
in the sequel.

Proposition 4.1 Assume that m > 1 and u is the generalized solution of the Cauchy problem
(4.3), (4.2)on Q = R × (0, ∞). Then

∂u ku
≥− , (4.6)
∂t t
∂v (m − 1)kv
≥− , (4.7)
∂t t
∂ 2v k
2
≥− (4.8)
∂x t
in the sense of distributions.

Remark 4.2 As will be seen from the proof stated below, this proposition is valid for any initial
data u0 ∈ L∞ (RN ).

Proof of Proposition 4.1. The given generalized solution u can be obtained as the limit
of a sequence of classical solutions which are positive, and uniformly bounded on Q and whose
derivatives up to second order are bounded. To prove the proposition, we may simply suppose
that u is just its approximate smooth solution.
We first verify (4.8). Notice that v satisfies
 2
∂v ∂2v ∂v
= (m − 1)v 2 + (4.9)
∂t ∂x ∂x

∂2v
and w = satisfies
∂x2
∂w ∂2w ∂v ∂w
Lw = − (m − 1)v 2 − 2m · − (m + 1)w2 = 0. (4.10)
∂t ∂x ∂x ∂x
k ∂2v
e = − also satisfies the equation (4.10) on Q. Since w =
Clearly the function w is bounded,
t ∂x2
we have
k
w(x, ε) > −
ε
provided ε > 0 is small enough. Thus we may use the comparison theorem on R×(ε, ∞) to assert

k
w(x, t) ≥ w(x,
e t) = − for (x, t) ∈ R × (ε, ∞),
t
from which (4.8) follows by virtue of the arbitrariness of ε.
(4.7) follows from (4.8) and (4.9).
52 Finite Propagation of Disturbances

To verify (4.6), it suffices to note that

 1/(m−1)  2 !
∂u m−1 ∂2v 1 ∂v
= 2
+ v (2−m)/(m−1)
∂t m ∂x m−1 ∂x
∂2v ku
≥u 2 ≥− .
∂x t 
Proof of Theorem 4.2. From (4.6) we have
∂ k
(t u) ≥ 0.
∂t
From this and the continuity of u, it is easy to see that for any x ∈ R, tk u is increasing in t.
Therefore
Ω(t1 ) ⊂ Ω(t2 ), for t1 < t2 ,
in other words, (−1)i ζi (t) (i = 1, 2) is increasing.
Next we prove (4.5). For simplicity we suppose that x1 < 0 < x2 , u0 (0) > 0. Then there exist
δ > 0, ε0 > 0 such that
u0 (x) ≥ ε0 for |x| ≤ δ.
Consider the function
Bτ,L (x, t) = L1/m−1 Bm (x, L(t + τ )) (L, τ > 0)
where Bm (x, t) is the Barenblatt solution of (4.3). It is easy to check by direct calculation that for
any L, τ > 0, Bτ,L (x, t) is a generalized solution of (4.3) with initial data

Bτ,L (x, 0) = L1/(m−1) B(x, Lτ )

  !1/(m−1)
L1/(m−1) m−1 x2
= 1− .
(Lτ )1/(m+1) 2m(m + 1) (Lτ )2/(m+1) +

Obviously
2
L2/(m −1)
0 ≤ Bτ,L (x, 0) ≤ 1/(m+1) ,
 τ 
2 2m(m + 1) 2/(m+1)
supp Bτ,L (x, 0) = x : |x| ≤ (Lτ ) .
m−1
Choose L, τ > 0 such that
2 2m(m + 1)
L2/(m −1)
= ε0 τ 1/(m+1) , (Lτ )2/(m+1) = δ 2 .
m−1
It is easy to see that this is possible. For such L, τ > 0, we have
Bτ,L (x, 0) ≤ u(x, 0),
and thus by the comparison theorem
Bτ,L (x, t) ≤ u(x, t).
In particular  
2 2m(m + 1) 2/(m+1)
Ω(t) ⊃ x : |x| ≤ (L(t + τ )) .
m−1
This completes the proof of (4.5). 
Nonlinear Diffusion Equations 53

4.2 Approach by Moser Iteration

Consider equations of the form
∂u
= ∆A(u) (4.11)
∂t
with A0 (s) ≥ 0 and E = {s; A0 (s) = 0} being allowed to have interior points. Assume the initial
value condition
u(x, 0) = u0 (x). (4.12)
The solutions of the Cauchy problem (4.11), (4.12) has already been defined.
There are a lot of papers discussing the properties of solutions of weakly degenerate equations.
However the study in this aspect for strongly degenerate equations is very few. In what follows
we present a result on the propagation of disturbances for such equations. To prove the result, we
need the following embedding theorem.

Lemma 4.1 Let Ω be a bounded domain of RN with smooth boundary and N > 2. Then for
any u ∈ H01 (Ω), there holds
Z 1/m Z 1/2
m 1/m−1/2∗ 2
|u| dx ≤ C|Ω| |∇u| dx ,
Ω Ω

where |Ω| denotes the measure of Ω, C is a constant independent of Ω and u, and

2N
2∗ = , m ∈ [1, 2∗ ].
N −2
We need also the following iteration lemma.

Lemma 4.2 Assume that F (s) is a nonnegative and bounded function on [0, +∞) satisfying

C
F (R) ≤ F (2R)β , ∀0 < R ≤ R0 ,
Rα
where α > 0, β > 1 and C is a constant independent of R. Then for large R 0 ,
 
R0
lim F = 0.
m→∞ 2m

Proof. From the assumption, we have

   β
R0 2mα R0
F ≤C F .
2m R0α 2m−1

By iteration we obtain    αm

R0 C βm m
F ≤ (2α ) F (R0 )β ,
2m R0α
where
m−1
X βm − 1
αm = βi = ≥ β m−1 ,
i=0
β−1
m−1
X β(β m − 1) − m(β − 1) β
βm = (m − i)β i = ≤ βm .
i=0
(β − 1)2 (β − 1)2
54 Finite Propagation of Disturbances

If we choose R0 such that

 1/β
C β/(β−1)2
λ≡ (2α ) F (R0 ) < 1,
R0α

then we have  
R0 m
F ≤ λβ → 0, as m → ∞
2m
and thus the lemma is proved. 

Theorem 4.3 Assume that A(s) and u0 (x) are appropriately smooth with A0 (s) ≥ 0 and
Z s
Ψ(s) ≡ A(σ)dσ ≥ γ|A(s)|p (4.13)
0

for some constants γ > 0 and 1 ≤ p < 2. Let u be a BV solution of the Cauchy problem (4.11),
(4.12). If
suppu0 ⊂ BL
for some constant L, then there exists a constant L1 , such that

suppu(0, t) ⊂ BL1

for almost all t ∈ (0, T ).

Proof. From the definition of generalized solutions, we have

ZZ ZZ
∂u
ϕ + ∇w∇ϕdxdt = 0 (4.14)
QT ∂t QT

for any ϕ ∈ C0∞ (QT ), where w = A(u). Denote by BR = BR (x0 ) the ball of radius R centered at
x0 . Let 0 ≤ ψ(t) ∈ C0∞ (0, T ) and 0 ≤ h(x) ∈ C0∞ (B2R ), h(x) = 1 in BR and

∂h C ∂2h C
≤ , ≤ 2, (i, j = 1, · · · , N ).
∂xi R ∂xi ∂xj R

Choosing ϕ = h2 ψ ŵ in (4.14), we derive

ZZ ZZ ZZ
2 ∂Ψ(u) 2 2
h ψ + ψ |∇w| h dxdt − ψ|w|2 div(h∇h)dxdt = 0,
QT ∂t QT QT

from which it follows by the arbitrariness of ψ(x) that for any t ∈ (0, T ),
ZZ ZZ ZZ
∂Ψ(u) 2
h2 + |∇w| h2 dxds − |w|2 div(h∇h)dxdt = 0,
Qt ∂t Qt Qt

or Z Z Z tZ
h2 Ψ(u)dx − h2 Ψ(u0 )dx + | 5 w|2 h2 dxds
B2R B2R 0 B2R
Z tZ
− |w|2 div(h 5 h)dxds = 0.
0 B2R
Nonlinear Diffusion Equations 55
Z
Thus, if B2R ∩ suppu0 = ∅, then h2 Ψ(u0 )dx = 0 and using (4.13) yields
B2R

Z Z TZ
2
sup |w|p h2 dx + |∇w| h2 dxdt
0<t<T B2R 0 B
Z TZ 2R Z TZ
2 C
≤ C |w| |div(h∇h)| dxdt ≤ 2 |w|2 dxdt,
0 B2R R 0 B2R

which implies, in particular, that

Z Z T Z
Cp
sup |w| dx ≤ 2 |w|2 dxdt, (4.15)
0<t<T BR R 0 B2R
Z TZ Z TZ
C
|∇w|2 h2 dxdt ≤ |w|2 dxdt. (4.16)
0 B2R R2 0 B2R

By Lemma 4.1, we have

Z 1/2∗ Z 1/2∗
∗ ∗
|w|2 dx ≤ |wh|2 dx
BR B2R
Z 1/2
≤ C |∇(wh)|2 dx
B2R
Z 1/2  Z 1/2
C
≤ C |∇w|2 h2 dx + |w|2 dx .
B2R R2 B2R

From this, using Hölder’s inequality and (4.15), (4.16) we derive

Z T Z
|w|2+2p/N dxdt
0 BR
Z T Z 2/N Z 2/2∗
p 2∗
≤ |w| dx |w| dx dt
0 BR BR
 Z 2/N Z T Z 2/2∗
p 2∗
≤ sup |w| dx |w| dx dt
0<t<T BR 0 BR
Z Z !1+2/N
T
C
≤ |w|2 dxdt .
R2 0 B2R

Using Hölder’s inequality again gives

Z TZ
|w|2 dxdt
0 BR
Z Z !N/(N +p) Z Z !p/(N +p)
T T
2+2p/N
≤ |w| dxdt dxdt
0 BR 0 BR
Z Z !(N +2)/(N +p)
T
≤ CR(N p−2N −4)/(N +p) |w|2 dxdt ,
0 B2R

or
F (R) ≤ CR−4/(N +p) F (2R)(N +2)/(N +p) ,
56 Finite Propagation of Disturbances

where
Z T Z
1
F (R) = N |w|2 dxdt.
R 0 BR

N +2
Since 1 ≤ p < 2 implies that > 1, from Lemma 4.2 we see that for large R0 ,
N +p
 
R0
lim F = 0. (4.17)
m→∞ 2m

From (4.17) and the definition of F (R), using Lebesgue’s Theorem, we conclude that for large
R0 , w(x, t) = 0 almost everywhere in the set QT \(BL+2R0 (0) × (0, T )). This and the equality
ZZ ZZ
∂ϕ
u dxdt + w 4 ϕdxdt = 0
QT ∂t QT

holding for any ϕ ∈ C0∞ (BL+2R0 (0)×(0, T )) imply that for large R0 , u(x, t) = 0 almost everywhere
in the set QT \(BL+2R0 (0) × (0, T )). The proof of the theorem is thus completed. 

4.3 Approach by Energy Estimates

In this section, we consider a fourth order equation with double degeneracy, namely,
 
∂B(u) ∂2 ∂2u
+ 2A = 0, (x.t) ∈ QT , (4.18)
∂t ∂x ∂x2

where A(s) = |s|p−2 s, B(s) = |s|q−2 s(p, q > 1), QT = (0, 1) × (0, T ). We will consider the first
boundary value problem as an example. The corresponding initial and boundary value conditions
are

u(x, 0) = u0 (x), (4.19)

u(0, t) = u(1, t) = ux (0, t) = ux (1, t) = 0. (4.20)

Definition 4.1 A function u is said to be a generalized solution of the problem (4.18)–(4.20),

if the following conditions are fulfilled:
0
(1) u ∈ L∞ (0, T ; W02,p (I)), B(u) ∈ C(0, T ; Lq (I)),

∂B(u) 0
∈ L∞ (0, T ; W −2,p (I)), B(u(x, 0)) = B(u0 ),
∂t
where I = (0, 1) and p0 and q 0 are the conjugate exponents of p, q respectively;
(2) For any ϕ ∈ C0∞ (QT ), the following integral equality holds
ZZ ZZ  
∂ϕ ∂2u ∂2ϕ
− B(u) dxdt + A dxdt = 0.
QT ∂t QT ∂x2 ∂x2

Theorem 4.4 Let u0 ∈ W02,p (I), q ≤ p. Then the boundary value problem (4.18)–(4.20)
admits at least one generalized solution.

We omit the details of the proof.

Nonlinear Diffusion Equations 57

Theorem 4.5 Let u be the solution of the problem (4.18)–(4.20) obtained in Theorem 4.4.
Then for any 0 ≤ ρ ∈ C 2 (I),
Z 1 Z
1 1 1 q
ρ(x)|u(x, t)| dx − 0 ρ(x)|u0 (x)|q dx
0q0 q 0
ZZ  2  2
∂ u ∂
=− A 2
(ρ(x)u(x, τ )) dxdτ, (4.21)
Qt ∂x ∂x2

where Qt = (0, 1) × (0, t).

We omit the details of the proof.

Lemma 4.3 (Nirenberg’s inequality) Let Ω ⊂ RN be a bounded domain with smooth boundary,
u ∈ W m,r (Ω), Then
1−a
kDj ukLp ≤ C1 kDm ukaLr kukL q + C2 kukLq ,
where  
j 1 j 1 m 1
≤ a < 1, = +a − + (1 − a) .
m p n r n q

Lemma 4.4 (Weighted Nirenberg’s inequality)

Z 1/p
(x)k+ |Du|p dx
R
Z a/p Z (1−a)/q
≤ C (x)k+ |D2 u|p dx (x)k+ |u|q dx ,
R R

where k is a nonnegative integer, (x)+ = max{x, 0}, provided that the integral on the right hand
side exists and  
1 1 1 1 2 1
≤ a < 1, = +a − + (1 − a) .
2 p 1+k p 1+k q

Lemma 4.5 (Hardy’s inequality)

Z Z
k p k+p
x+ |u| dx ≤ C x+ |Du|p dx,
R R

where k ≥ 0, p > 1, provided that the integrals on both sides exist.

Lemma 4.6 (Kjellberg’s inequality)

Z 1/p Z 1/p(2p+3) Z 2/(2p+3)
|u|p dx ≤C x2 |u|p dx |u|p+1 dx ,
R R R

where p > 0, provided that the integral on the right hand side exists.

Remark 4.3 In Lemma 4.4–4.6, R can be replaced by any finite interval, provided that u can
be extended via some manner to be defined on R, for example, the functions in H 02 (I) can be zero
extended.
58 Finite Propagation of Disturbances

Theorem 4.6 Let u be the solution obtained in Theorem 4.4 with initial data u 0 . If 1 < q < p
and supp u0 ⊂ [x1 , x2 ], 0 < x1 < x2 < 1, then
supp u(x, ·) ⊂ [x1 (t), x2 (t)] ∩ [0, 1], a.e. t ∈ (0, T ),
where x1 (t), x2 (t) can be expressed by
x1 (t) = x1 − C1 tµ , x2 (t) = x2 + C2 tµ
with positive constants C1 , C2 , µ depending only on p, q and u0 .
Proof. Setting ρ(x) = (x − y)k+ in Theorem 4.5, where y ∈ [x2 , 1) is any fixed constant, and
using the assumption suppu0 ⊂ [x1 , x2 ], we have
Z
1 1
(x − y)k+ |u(x, t)|q dx
q0 0
ZZ  2  2
∂ u ∂

= − A 2 2
(x − y)k+ u(x, τ ) dxdτ.
Qt ∂x ∂x
We use Young’s inequality to the right hand side to obtain
ZZ  2  2
∂ u ∂

− A 2 2
(x − y)k+ u(x, τ ) dxdτ
Qt ∂x ∂x
ZZ p
∂2u
= − (x − y)k+ dxdτ +
Qt ∂x2
ZZ  2 
k−1 ∂ u ∂u
−2k (x − y)+ A dxdτ
Qt ∂x2 ∂x
ZZ  2 
k−2 ∂ u
−k(k − 1) (x − y)+ A udxdτ
Qt ∂x2
ZZ p
1 ∂2u
≤ − (x − y)k+ dxdτ −
2 Qt ∂x2
ZZ p
k−p ∂u
+C1 (x − y)+ dxdτ
Qt ∂x
ZZ
+C2 (x − y)k−2p
+ |u|p dxdτ.
Qt

From this and ZZ ZZ p

∂u
(x − y)k−2p
+ |u|p dxdτ ≤ C k−p
(x − y)+ dxdτ,
Qt Qt ∂x
which is a consequence of Lemma 4.5, we obtain
Z 1 ZZ p
∂2u
sup (x − y)k+ |u(x, τ )|q dx ≤ C (x − y)k+ dxdτ, (4.22)
0<τ ≤t 0 Qt ∂x2
ZZ p ZZ p
∂2u k−p ∂u
(x − y)k+ dxdτ ≤ C (x − y)+ dxdτ. (4.23)
Qt ∂x2 Qt ∂x
Set ZZ p
∂2u
fm (y) = (x − y)m
+ dxdτ, m = 1, 2, · · ·
Q ∂x2
Z t Zt1 2 p
∂ u
f0 (y) = dxdτ.
0 y ∂x2
Nonlinear Diffusion Equations 59

From (4.22),(4.23), Lemma 4.4 and Hölder’s inequality, we have

ZZ p
∂u
f2p+1 (y) ≤C (x − y)p+1
+ dxdτ
Qt ∂x
Z t Z 1 p a
∂2u
≤C (x − y)p+1
+ dx ·
0 0 ∂x2
Z 1 (1−a)p/q
· (x − y)p+1
+ |u| q
dx dτ
0
Z t Z 1 p a
∂2u
≤ C [fp+1 (y)](1−a)p/q (x − y)p+1
+ dx dτ
0 0 ∂x2
a+(1−a)p/q
≤ Ct1−a (fp+1 (y)) ,

where
1 1 1
+ −
q p+2 p
a= .
1 2 1
+ −
q p+2 p
p
Set ν = a + (1 − a) . Applying Hölder’s inequality to the right hand side of the above inequality,
q
we further obtain
Z Z 2 p ν
1−a p+1 ∂ u
f2p+1 (y) ≤ Ct (x − y)+ dxdτ
Qt ∂x2
Z Z p (p+1)ν/(2p+1)
∂2u
≤ Ct1−a (x − y)2p+1
+ dxdτ ·
Qt ∂x2
Z t Z 1 2 p pν/(2p+1)
∂ u
· dxdτ
0 y ∂x2
(p+1)ν/(2p+1)
≤ Ct1−a (f2p+1 (y)) f0 (y)pν/(2p+1) .

Therefore
f2p+1 (y) ≤ Ct(1−a)/σ [f0 (y)]pν/(2p+1)σ ,

where
p+1
σ =1− ν > 0.
2p + 1
Using Hölder’s inequality again gives
1/(2p+1)
f1 (y) ≤ (f2p+1 (y)) [f0 (y)](2p)/(2p+1) ≤ Ctλ [f0 (y)]θ+1 ,

where
1−a pν 1
λ= , θ= 2
− > 0.
σ(2p + 1) σ(2p + 1) 2p + 1
Since f10 (y) = −f0 (y), we have

f10 (y) ≤ −Ct−λ/(θ+1) [f1 (y)]1(θ+1) .

60 Finite Propagation of Disturbances

∂ 2 u(x, t)
If f1 (x2 ) = 0, then = 0 for x ∈ [x2 , 1], and hence from the boundary value condition,
∂x2
we see that u(x, t) = 0 for x ∈ [x2 , 1], i.e., suppu(·, t) ⊂ [0, x2 ]. If f1 (x2 ) 6= 0, then there exists an
interval (x2 , x∗2 ), such that f1 (y) > 0 in (x2 , x∗2 ), but f1 (x∗2 ) = 0. So, for y ∈ (x2 , x∗2 ),
 0 θ f10 (y)
f1 (y)θ/(θ+1) = ≤ −Ct−λ/(θ+1) .
θ + 1 f1 (y)1/θ

Integrating the above inequality over (x2 , x∗2 ), we obtain

f1 (x∗2 )θ/(θ+1) − f1 (x2 )θ/(θ+1) ≤ −Ct−λ/(θ+1) (x∗2 − x2 ).

Therefore
x∗2 ≤ x2 + Ctλ/(θ+1) (f1 (x2 ))θ/(θ+1) ≡ x2 + C2 tµ ,
which implies
suppu(·, t) ⊂ [0, x2 + C2 tµ ].
Similarly
suppu(·, t) ⊂ [x1 − C1 tµ , 1].
The proof is complete. 
Nonlinear Diffusion Equations 61

5 Free Boundary of Solutions

5.1 Interfaces
Consider the one-dimensional porous medium equation

∂u ∂ 2 um
= (5.1)
∂t ∂x2
with m > 1 which corresponds to the slow diffusion, subject to the initial value condition

u(x, 0) = u0 (x) x ∈ R. (5.2)

Assume that supp u0 = [x1 , x2 ](−∞ < x1 < x2 < +∞) and u is a generalized solution of (5.1),
(5.2) on Q = R × (0, ∞). Denote

Ω = {(x, t); u(x, t) > 0, t > 0}

Ω(t) = {x; u(x, t) > 0}

ζ1 (t) = inf Ω(t), ζ2 (t) = sup Ω(t).

By the continuity of u, Ω is an open set. As already shown in the previous sections, for any t > 0,
ζi (t) (i = 1, 2) is finite. We call x = ζi (t) (i = 1, 2) the free boundary or interface of u.
In what follows we will further investigate the properties of free boundaries x = ζi (t) (i = 1, 2)
of the generalized solution u of the equation (5.1).
Naturally we expect the free boundaries to move with the local velocity. Set
m
v= um−1 .
m−1
Then, in view of the equation of state, which we have used to derive the equation (5.1), v is
essentially the pressure, and, by Darcy’s law, we expect

∂v(x, t)
ζi0 (t) = − lim ,
(x,t)∈Ω ∂x
x→ζi (t)

where Ω = {(x, t) ∈ Q; u(x, t) > 0}. The next result shows that this is almost true.

Theorem 5.1 The limits

∂v(x, t)
vx (ζi (t), t) ≡ lim (i = 1, 2),
(x,t)∈Ω ∂x
x→ζi (t)
ζi (t + ∆t) − ζi (t)
ζi0 (t + 0) ≡ lim (i = 1, 2)
∆t→0+ ∆t

exist for all t > 0 and

ζi0 (t + 0) = −vx (ζi (t), t) (i = 1, 2).

Proof. From Theorem 2.1, u ∈ C ∞ (Ω). By Proposition 4.1, for any τ > 0, there exists a
constant β depending only on τ such that

∂2v
≥ −β for (x, t) ∈ Ω, t ≥ τ
∂x2
62 Free Boundary of Solutions

∂2f
in the sense of distributions. Hence the function f (x, t) = v(x, t) + βx2 satisfies ≥ 0. This
∂x2
∂f
means that is an increasing function of x for each fixed t ≥ τ . Since for any t ≥ τ , u has
∂x
∂v ∂f
compact support, it follows from Lemma 6.1 that for any finite T > τ , and are bounded
∂x ∂x
∂f (x, t) ∂v(x, t)
on Ω ∩ {(x, t); t ∈ [τ, T ]}. Thus the limits lim (i = 1, 2) and lim (i = 1, 2)
x→ζi (t) ∂x x→ζi (t) ∂x
(x,t)∈Ω (x,t)∈Ω
exist for any t ∈ [τ, T ]. Since T , τ are arbitrary, these limits exist for any t > 0.
We are ready to prove the rest part of the theorem for any t. For simplicity we take t = 0 and
treat ζ2 (t) only. Set a = ζ2 (0), vx (a, 0) = α. Then either α = 0 or α < 0.
Case 1. α < 0.
We will show that for any sufficiently small ε > 0, there exists a δ > 0 depending only on ε
such that
ζ2 (∆t) − a
+α <ε (5.3)
∆t
∂v(x, 0)
whenever 0 < ∆t < δ. Since lim− = α, it follows that for any 0 < ε < −α, there exists a
x→a ∂x
δ0 > 0 such that
∂v(x, 0)
α−ε< < α + ε, (a − δ0 < x < a).
∂x
Using the mean value theorem we get

(α + ε)(x − a) < v(x, 0) < (α − ε)(x − a)

and hence   1/(m−1)


 m−1

 u(x, 0) > (α + ε)(x − a) ,
m
 1/(m−1) (5.4)

 m−1

 u(x, 0) < (α − ε)(x − a)
m
wherever a − δ0 < x < a.
We will use the following pressure solutions of (5.1)
 1/(m−1)
m−1
u1 (x, t) = (α + ε)((α + ε)t + (x − a))− ,
m
 1/(m−1)
m−1
u2 (x, t) = (α − ε)((α − ε)t + (x − a))− ,
m

which are some modification of (4.37). From (5.4) it follows that

u1 (x, 0) < u(x, 0) < u2 (x, 0), for x ≥ a − δ0 .

By continuity there exists τ > 0 such that,

u1 (a − δ0 , t) < u(a − δ0 , t) < u2 (a − δ0 , t), for 0 < t < τ.

Therefore the comparison theorem gives

u1 (x, t) ≤ u(x, t) ≤ u2 (x, t), for x ≥ a − δ0 , 0 < t < τ

Nonlinear Diffusion Equations 63

and from this it is easy to see that

a − (α + ε)t ≤ ζ2 (t) ≤ a − (a − ε)t, for 0 < t < τ

or
ζ2 (t) − a
+ α ≤ ε, for 0 < t < τ,
t
which completes the proof in Case 1.
20 . α = 0.
From the proof in Case 1, we see that we still have

u1 (x, t) ≤ u(x, t), for x ≥ a − δ0 , 0 < t < τ

and hence
ζ2 (t) − a
≤ ε, for 0 < t < τ.
t
On the other hand, by the monotonicity of ζ2 (t),

ζ2 (t) − a
≥ 0 for t > 0.
t
Combining these two inequalities completes the proof in Case 2. 
 
1
Proposition 5.1 For any 0 < δ < 1, ζi (t) (i = 1, 2) is Lipschitz continuous on δ, .
δ
 
1 ∂v(x, t)
Proof. Let t1 ∈ δ, . Denote α0 = ζ2 (t1 ) and C = max which is finite by
δ δ≤t≤1/δ ∂x
x∈R
Lemma 6.1 (see Section 6).
Consider the pressure solution of (5.1),
 1/(m−1)
m−1
U (x, t) = c(c(t − t1 ) + α0 − x)+ .
m

Since v(x, t) ≤ C(α0 − x) for x ≤ α0 , we have

u(x, t1 ) ≤ U (x, t1 ) for x ∈ R.

Therefore by the comparison theorem,

1
u(x, t) ≤ U (x, t) for x ∈ R, t1 ≤ t ≤ .
δ
From this it follows that
1
ζ2 (t) ≤ α0 + c(t − t1 ) for t1 ≤ t ≤ ,
δ
and hence, since ζ2 (t) is increasing, we obtain

1
0 ≤ ζ2 (t) − ζ2 (t1 ) ≤ C(t − t1 ), for t1 ≤ t ≤ .
δ
The Lipschitz continuity of ζ1 can be proved similarly.
64 Free Boundary of Solutions

In the sequel we will use the Barenblatt solutions (see (1.37))

  !1/(m−1)
1 x2
w(x, t) = Bm (x, t) = 1− 2 ,
λ(t) λ (t) +

where  1/(m+1)
1/(m+1) 2m(m + 1)
λ(t) = Cm (t + 1) , Cm = .
m−1
An immediate calculation shows that for any c, L > 0,

wc,L (x, t) = cw(Lx, cm−1 L2 t)

is a generalized solution of (5.1). Denote

m
vc,L (x, t) = wm−1 (x, t),
m − 1 c,L
or !
m cm−1 L 2 x2
vc,L (x, t) = m−1 1− 2 , (5.5)
m − 1 λc,L (t) λc,L (t)
+
where
1/(m+1)
λc,L (t) = Cm cm−1 L2 t + 1 .
(i)
Let x = ζc,L (t) (i = 1, 2) be the free boundaries of vc,L . Then

(1) Cm m−1 2
1/(m+1)

ζc,L (t) = c L t+1 ,
L
(2) Cm m−1 2
1/(m+1)
ζc,L (t) =− c L t+1 .
L 

Proposition 5.2 For any δ > 0 there exists a convex function ξi (t) (t > 0) and a C 1,1 function
ηi (t) such that
(−1)i ζi (t) = ξi (t) + ηi (t), (t > δ, i = 1, 2). (5.6)
Here C 1,1 denotes the class of all functions whose derivatives are Lipschitz continuous.

Proof. We only prove (5.6) for i = 2. We set ζ(t) = ζ2 (t) and want to study the free
boundary x = ζ(t) near a point (x0 , t0 ) with t0 > δ. For the sake of simplicity, we perform a
translation of x, t variables so that, in new x, t variables,
cm
t0 = 0, x0 = ζ(t0 ) = .
L
c 
m
Then (x0 , t0 ) = , 0 lies on the free boundary of both v(x, t) and vc,L (x, t) defined by (5.5).
L
We wish to choose c, L so that

ζ 0 (0 + 0) = ζe0 (0), (5.7)

vc,L (x, 0) ≤ v(x, 0), (5.8)

e = ζ (2) (t). In view of Theorem 5.1, (5.7) is equivalent to

where ζ(t) c,L

∂  cm  ∂ c
m

v ,0 = vc,L ,0 , (5.9)
∂x L ∂x L
Nonlinear Diffusion Equations 65

∂ 2 v(x, t)
≥ −2P, for x ∈ R, t ≥ δ. (5.10)
∂x2
If c, L are chosen so that

∂ 2 vc,L (x, 0)
= −2P, whenever vc,L (x, 0) > 0, (5.11)
∂x2
then since also    
Cm Cm
vc,L ,0 =0=v ,0 ,
L L
we then deduce the inequality (5.8).
By an immediate calculation, (5.7), (5.11) turn out to be

1
ζ 0 (0 + 0) = Cm cm−1 L, (5.12)
m+1

m 2L2 cm−1
m+1 = 2p, (5.13)
m − 1 Cm
which clearly have a unique positive solution (c, L). For such (c, L), (5.8) holds and hence by the
comparison theorem,
vc,L (x, t) ≤ v(x, t),
from which it follows that
e ≤ ζ(t).
ζ(t)
Using this inequality and recalling (5.7) we obtain

e − ζ(0)
ζ(h) − ζ(0) − hζ 0 (0 + 0) ≥ ζ(h) e − hζe 0 (0). (5.14)

By a direct calculation, using (5.7), (5.13) yields

m
ζe 00 (0) = −Cm c2(m−1) L3
(m + 1)2
m+2
(m − 1)Cm
= − P cm−1 L = −γP ζ 0 (0 + 0),
(m + 1)2
m+1
(m − 1)Cm
where γ = . It is clear that
m+1

ζe00 (t) = ζe00 (0) + O(t) (t → 0+ ).

Therefore the right hand side of (5.14) is equal to

1 2 e00 1
h ζ (0) + O(h3 ) = − γh2 P ζ 0 (0 + 0) + O(h3 ),
2 2
and (5.14) gives
1
ζ(h) − ζ(0) − hζ 0 (0 + 0) ≥ − h2 γζ 0 (0 + 0) + O(h3 ).
2
Going back to the original (x, t) coordinates, we obtain

Φh (t) ≥ −γP ζ 0 (t + 0) + O(h) (5.15)

66 Free Boundary of Solutions

at any point t = t0 > 0, where

ζ(t + h) − ζ(t) − hζ 0 (t + 0)
Φh (t) = (h > 0).
h2 /2
     
1 δ 1
Now, in any interval δ, and for any h ∈ 0, , Φh (t) ∈ L∞ δ, , since by Proposition
δ   2 δ
1
5.1, ζ(t) is Lipschitz continuous on δ, . Also, for any 0 < s < T < ∞,
δ
Z T Z T +h Z s+h !
2
Φh (t)dt = ζ(t)dt − ζ(t)dt − h(ζ(T ) − ζ(s))
s h2 T s
Z T +h Z s+h
2 2
= (ζ(t) − ζ(T ))dt − 2 (ζ(t) − ζ(s))dt ≤ C.
h2 T h s
Here and below C > 0 is a universal constant independent of h. Since, by (5.15), ζ(t) is Lipschitz
continuous,
Φh (t) ≥ −C, for t ∈ (s, T ),
we conclude that Z T
|Φh (t)|dt ≤ C.
s
We can therefore choose a sequence h = hn → 0 such that Φhn weakly converges to a measure
µ0 . Using (5.15) we then have

Φhn (t) + γP ζ 0 (t + 0) → µ, weakly asn → ∞,

where µ is a nonnegative measure. Since the weak convergence implies the convergence in the sense
of distributions, we conclude from the definition of Φh , that

ζ 00 + γP ζ 0 = µ (5.16)
in the sense of distributions. Set
Z tZ τ
ξ(t) = dµ(s)dτ,
δ δ
Z t
η(t) = ζ(δ) + ζ 0 (δ + 0)(t − δ) − γP (ζ(s) − ζ(δ))ds.
δ

Then ξ(t) is convex, η ∈ C 1,1 , and ζ̂(t) = ξ(t) + η(t) is a solution of (5.16) with ζ̂(δ) = ξ(δ),
ζ̂ 0 (δ) = ζ(δ + 0). By uniqueness we must have ζ̂ ≡ ζ. The proof is complete. 
Corollary 5.1 For any t > 0, ζi0 (t ± 0) exists and

(−1)i ζi0 (t − 0) ≤ (−1)i ζi0 (t + 0). (5.17)

Furthermore, for any δ > 0, there exists P > 0 such that for t2 > t1 > δ,

(−1)i ζi0 (t2 − 0)eγP t2 ≥ (−1)i ζi0 (t1 + 0)eγP t1 , (5.18)

which implies that there exists a constant t∗i ≥ 0 such that ζ 0 (t) is strictly increasing for t > t∗ and
ζi (t) ≡ xi for 0 ≤ t ≤ t∗ .
If t∗ > 0, then the free boundary x = ζ(t) remains stationary (vertical) for t ∗ units of time
after which it begins to move without further stops. We will call t∗i the waiting time.
Nonlinear Diffusion Equations 67

Proof. (5.17) follows from the convexity of ξ(t) and η ∈ C 1,1 . The existence of ζi (t − 0) can
be proved similarly.
From (5.16), we get
(ζ 0 eγP t )0 = eγP t µ ≥ 0,
which implies (5.18).
If ζ(t) ≡const. on the interval 0 < s1 < t < s2 , then ζ 0 (t) ≡ 0 on this interval. Using(5.18) we
deduce ζ 0 (t) ≡ 0 on 0 < t < s1 . This completes the proof.
Now we are ready to discuss the continuous differentiability of ζi (t) for t > t∗i . We will treat
ζ(t) = ζ2 (t) only.
Let (x0 , t0 ) = (ζ(t0 ), t0 ) with t0 > t∗2 and Nδ be the intersection of a δ-neighborhood of (x0 , t0 )
and Ω = {(x, t) ∈ Q; u(x, t) > 0}. Take δ > 0 so small that N 2δ stays away from t = t∗2 and from
the free boundary x = ζ1 (t). Denote by d(x, t) the distance from (x, t) to the free boundary. 

Lemma 5.1 There exist positive constants C1 , C2 depending only on m, δ such that
v(x, t)
C1 ≤ ≤ C2 in Nδ .
d(x, t)
Proof. By Corollary 5.1 and our assumption on Nδ , there exist constants c > 0,C > 0,
depending only on δ, such that for all t with (ζ(t), t) ∈ ∂Nδ ,

c ≤ ζ 0 (t ± 0) ≤ C. (5.19)

Hence there exist constants C 1 > 0, C 2 > 0 depending only on δ , such that
d(x, t)
C1 ≤ ≤ C 2, for (x, t) ∈ Nδ .
|x − ζ(t)|

Since v(x, t) is Lipschitz continuous,

v(x, t) = v(x, t) − v(ζ(t), t) ≤ C 0 |x − ζ(t)| ≤ C2 d(x, t).

∂2v
Next, since ≥ −2P (see (5.10)), using the Taylor formula, Theorem 5.1 and (5.19), we see
∂x2
that for δ > 0 small enough,

v(x, t) ≥ |vx (ζ(t), t)||x − ζ(t)| − P |x − ζ(t)|2

= |ζ 0 (t + 0)||x − ζ(t)| − P |x − ζ(t)|2 ≥ c0 |x − ζ(t)|

≥ C1 d(x, t)

with a constant C1 > 0 depending only on δ . The rest part of the lemma can be proved similarly.


Lemma 5.2 There exists a constant C depending only on δ, such that

∂v ∂2v C
≤ C, ≤ , for (x, t) ∈ Nδ .
∂t ∂t2 d(x, t)
1
Proof. Let (x, t) ∈ Nδ and S be a square with center (x, t) and side γ = d(x, t). For
2
simplicity, we suppose x = 0, t = 0. Consider the function

w(x, t) = γ −1 v(γx, γt),

68 Free Boundary of Solutions

where (x, t) varies in a unit square S0 (so that (γx, γt) ∈ S). Clearly, in S0
 2
∂w ∂2w ∂w
= (m − 1)w 2 + , (5.20)
∂t ∂x ∂x
and in view of Lemma 5.1, for some constants C1 > 0, C2 > 0,

C1 ≤ (m − 1)w ≤ C2 .
 2
∂v
Thus (5.20) is uniformly parabolic. Noting that is a bounded function, we can apply the
∂x
Nash-Moser estimate and conclude that for some α ∈ (0, 1),

|w|C α (S 0 ) ≤ C,
3
where S 0 is a square concentric with S0 and with side , and C > 0 is a constant depending only
0
4
on C1 , C2 , dist (S , ∂S0 ). We can now construct a fundamental solution for the equation

∂W ∂2W
= (m − 1)w
∂t ∂x2
∂w
and use it to express w, from which we see that is Hölder continuous. Applying Schauder’s
∂x
∂w ∂ 2 w
estimates we conclude that , (in fact, any derivative of w) are bounded by a constant
∂t ∂t2
1
C ∗ , depending only on δ, in a square S 00 concentric with S0 and with side . Going back to the
2
∂v(x, t) ∂ 2 v(x, t) ∗
function v we deduce that , are bounded by C ; this completes the proof. 
∂t ∂t2
∂v(x0 , t0 )
Lemma 5.3 Let x0 = ζ(t0 ),t0 > 0, = −ζ 0 (t0 + 0) = −b < 0. Then there exists a
∂x
neighborhood Nδ of (x0 , t0 ), such that for any constant a,

v(x, t) = Lb (x − x0 , t − t0 ) + o(|x − x0 | + |t − t0 |)
x − x0
for (x, t) ∈ Nδ with = a, where Lb (x, t) = b(bt − x)+ .
t − t0
Proof. For η > 0, define

vδ (x, t) = η −1 v(ηx − x0 , ηt − t0 ).
α
Fix α ∈ (0, t0 ). Then v is defined in the -neighborhood of the origin and v is also a generalized
η
 1/(m−1)
m−1
solution of the equation (5.20). (This means that v0 is a generalized solution of
m
α
(5.1)). By Proposition 5.1, for t ≥ − ,
η
∂ 2 vη ∂ 2 vη ηk ηk
2
=η (ηx + x0 , ηt + t0 ) ≥ − ≥− . (5.21)
∂x ∂x2 ηt + t0 t0 − η

From the proof of Theorem 6.1, we see that vη ∈ C 1,1/2 with the Hölder coefficient independent
of η. Therefore there exists a subsequence {vηn } with ηn → 0 such that vηn → v in R2 uniformly
Nonlinear Diffusion Equations 69

on compact sets and v is a generalized solution of (5.20). We may simply suppose that v η → v in
R2 as η → 0 uniformly on compact sets. It is easily seen that all of the derivatives of v η converge
to the corresponding derivatives of v.
In view of our hypothesis, for each fixed x < 0,

∂v ∂vη ∂v
(x, 0) = lim (x, 0) = (x0 , t0 ) = −b;
∂x η→0 ∂x ∂x
and for each fixed x ≥ 0,
1
v(x, 0) = lim vη (x0 , 0) = lim v(ηx + x0 , t0 ) = 0.
η→0 η→0 η

Thus (
−bx for x < 0
v(x, 0) =
0 for x ≥ 0.
Moreover,
C
vη (x, t) ≤ |ηx + x0 − ζ(ηt + t0 )|
η
C
≤ (η|x| + |x0 − ζ(ηt + t0 )|) ≤ C(|x| + 2bt)
η
provided x ∈ R and 0 < t < ε with ε > 0 small enough. Therefore we can use the corresponding
uniqueness result to conclude that for x ∈ R, 0 < t < ε,

v(x, t) = Lb (x, t). (5.22)

Repeating the argument we can further prove that (5.22) holds on the whole Q.
By Lemma 5.2, it is easy to see that there exists a constant C depending only on δ, such that

∂ 2 vη (x, t) ∂2v C1
2
=η (ηx + x0 , ηt + t0 ) ≤
∂t ∂t2 −x

δ2
for x < 2bt < 0 with t2 + x2 < . Therefore
η2

∂ 2 v η (x, t) C1
≤ for x < 2bt < 0. (5.23)
∂t2 −x

For x < 2bt < 0,

∂v(x, 0) t2 ∂ 2 v(x, et)
v(x, t) = v(x, 0) + t +
∂t 2 ∂t2
∂Lb (x, 0) t2 ∂ 2 v(x, e
t)
= Lb (x, 0) + t +
∂t 2 ∂t2
where t < e
t < 0. Thus
t2 ∂ 2 v(x, e
t)
v(x, t) − Lb (x, t) =
2 ∂t
and it follows by using (5.23) that

lim (v(x, t) − Lb (x, t)) = 0 for t < 0. (5.24)

x→−∞
70 Free Boundary of Solutions

Next we show that

v(x, t) ≥ Lb (x, t) for t < 0. (5.25)
Since Lb (x, t) = 0 for x ≥ bt, it suffices to prove (5.25) for x < bt, t < 0. In view of (5.21), we
∂2v
have ≥ 0. Thus for x < bt, t < 0,
∂x2
∂v(x, t) ∂v(bt, t)
≤ = −b,
∂x Z x ∂x
∂v
v(x, t) = v(bt, t) + (y, t)dy
bt ∂x
≥ v(x, t) − b(x − bt) = v(bt, t) + Lb (x, t)

from which (5.25) follows.

Now we prove that (5.22) holds for t < 0. Consider a fixed point (x, t) with x < bt. Let S be
a rectangle with center (x, t) and boundaries parallel to the coordinate axis, such that (x, t) ∈ S
implies that x < bt and S contains points with t > 0. In S, v − Lb ≥ 0 and achieves its minimum
value, 0, on S ∩ {t > 0}. By the strong minimum principle, we must have v ≡ Lb in S. Therefore
(5.22) holds for x < bt, t < 0.
Suppose that v(x, t) > 0 for some (x, t) with x > bt, t < 0. Then from Theorem 4.2, we have
v(x, t) > 0 for all t > t. However this contradicts the fact that the line x = x must intersect the
line x = bt for some t > t and v = Lb = 0 at that intersection. Therefore (5.22) also holds for
x > bt, t < 0.
Given a > 0. Let η = t − t0 6= 0. Since v|η| → v ≡ Lb as |η| → 0, given ε > 0 there exists a
η0 = η0 (ε) > 0 such that

1
v(ηa + x0 , η + t0 ) − Lb (asgn η, sgn η) < ε
|η|
x − x0
provided |η| = |t − t0 | ∈ (0, η0 ), = a. Namely
t − t0
|v(x, t) − Lb (x − x0 , t − t0 )| < ε|t − t0 |.

The proof of Lemma 5.3 is complete. 

Theorem 5.2 ζi (t) (i = 1, 2) is continuously differentiable at any t 6= t∗i (i = 1, 2), where t∗i
(i = 1, 2) is defined in Corollary 5.1.

Proof. We will prove the assertion only for ζ(t) = ζ2 (t). Since a convex function which is dif-
ferentiable everywhere on an interval is necessarily continuously differentiable there, by Proposition
5.2, it suffices to prove that ζ(t) is differentiable everywhere at any t = t0 > t∗2 , i.e

ζ 0 (t0 − 0) = ζ 0 (t0 + 0) = b. (5.26)

If b = 0, then by Corollary 5.1, (5.26) is trivial. We suppose b > 0. Denote x0 = ζ(t0 ). From
the definition of t∗2 , there must be a τ ∈ (t∗2 , t0 ) such that ζ 0 (τ + 0) > 0.
First we prove that for any a > b, there exists ε > 0 such that

ζ(t) > x0 + a(t − t0 ), t0 − ε < t < t 0 . (5.27)

If (5.27) is false, then there exists a sequence εn with εn ↓ 0 as n → ∞, such that

ζ(t0 − εn ) ≤ x0 − εn a
Nonlinear Diffusion Equations 71

and hence, by the definition of ζ(t)

v(x0 − εn a, t0 − εn ) = 0.

However, from Lemma 5.3 we have

v(x0 − εn a, t0 − εn ) = Lb (−εn a, −εn ) + o(εn )

= −b2 εn + baεn + o(εn ).

Hence
−b2 + ba + o(1) = −b(b − a) + o(1) = 0 (n → ∞)
which contradicts the fact that a > b > 0. Therefore (5.27) holds and we conclude that

x0 − ζ(t)
lim ≤ a.
t→t0 −0 t0 − t
Since a > b is arbitrary, it follows that

x0 − ζ(t)
lim ≤ b. (5.28)
t→t0 −0 t0 − t
Next we prove that for any c < b, there exists η > 0 such that

ζ(t) < x0 + c(t − t0 ), for t0 − η < t < t0 . (5.29)

If (5.29) is false, then there exists a sequence {ηn } with ηn ↓ 0 as n → ∞, such that

ζ(t0 − ηn ) ≥ x0 − cηn > x0 − bηn .

e ∈ (x0 − bηn , x0 − cηn ), such that

By the mean value theorem there exists x

∂v v(x0 − cηn , t0 − ηn ) − v(x0 − bηn , t0 − ηn )

x , t0 − η n ) =
(e .
∂x (b − c)ηn

Since c < b, we have, by Lemma 5.3,

∂v o(ηn )
x, t0 − ηn ) =
(e = o(1) (n → ∞).
∂x (b − c)ηn

According to Proposition 4.1,

k
v(x, t) + (ζ(t) − x)2
2t
is a convex function of x. Thus
∂v
(ζ(t0 − ηn ), t0 − ηn )
∂x
∂v k(ζ(t0 − ηn ) − x
e)
≥ x , t0 − η n ) +
(e = o(1) (n → ∞),
∂x t0 − η n

so that
∂v
lim sup (ζ(t), t) ≥ 0.
t→t0 −0 ∂x
72 Free Boundary of Solutions

On the other hand, by Corollary 5.1,

∂v
(ζ(t), t) = −ζ 0 (t0 + 0)
∂x
≤ −ζ 0 (τ + 0)eγP (τ −t)

≤ −ζ 0 (τ + 0)eγP (τ −t0 ) < 0 for t ∈ (τ, t0 ).

This contradiction shows that (5.29) holds. Since c < b is arbitrary, we conclude that

x0 − ζ(t)
lim inf ≥ b.
t→t0 −0 t0 − t

This, combining with (5.28) gives (5.26). 

5.2 Entropy Solutions

In this section we study equations with strong degeneracy, which may admit discontinuous solutions
in general, and the so called entropy solutions should be considered. For simplicity, we consider
equations in one dimensional case of the form

∂u ∂ 2 A(u) ∂B(u)
= + (5.30)
∂t ∂x2 ∂x
where A(s) and B(s) are appropriately smooth with A(0) = B(0) = 0 and A 0 (s) ≥ 0. Strong
degeneracy means that E = {s; A0 (s) = 0} may have interior points.
Denote QT = I × (0, T ) with I = R = (−∞, +∞) for the Cauchy problem and I = (0, 1) for
the boundary value problem.
The remarkable situation in treating strongly degenerate equations is that the solutions of such
equations might be discontinuous. This can be exposed in the following consideration. Suppose
E ⊃ [a, b](a < b). Then for u ∈ [a, b], (5.30) becomes the first order quasilinear equation

∂u ∂B(u)
= (5.31)
∂t ∂x
whose solutions, as is well-known, might have discontinuity, even if the initial value is smooth
enough.
The first problem is how to define solutions with discontinuity for (5.30). Motivated by the
theory of shock waves, a meaningful discontinuous solution u of (5.31) should satisfy the so-called
entropy condition
(u − k)γt ≤ (B(u) − B(k))γx , ∀ k∈R (5.32)
at the points of discontinuity of u in addition to the integral equality
ZZ  
∂ϕ ∂ϕ
u − B(u) dxdt = 0, ∀ ϕ ∈ C0∞ (QT ). (5.33)
QT ∂t ∂x

1 +
Here and below, u = (u + u− ) denotes the symmetric mean value and u± the approximate
2
limits of u at the points of discontinuity. It is not difficult to see that (5.32) and (5.33) imply
ZZ  
∂ϕ ∂ϕ
sgn(u − k) (u − k) − (B(u) − B(k)) dxdt ≥ 0
QT ∂t ∂x (5.34)
∀ 0 ≤ ϕ ∈ C0 (QT ), k ∈ R.
∞
Nonlinear Diffusion Equations 73

In fact, at least for piecewise continuous functions (5.32),(5.33) are equivalent to (5.34). However
the integral in (5.34) makes sense for any u ∈ L1loc (QT ), so we can use (5.34) to define more general
solutions. It was Kruzhkov who first defined generalized solutions of (5.31) in this way and proved
the solvability of the Cauchy problem in L∞ (QT ).
Inspired by Kruzhkov’s idea, Vol’pert and Hudjaev (1969) defined generalized solutions for
(5.30) as follows.

Definition 5.1 A function u ∈ BV (QT ) ∩ L∞ (QT ) is called a generalized solution of the

∂A(u)
equation (5.30), if ∈ L1loc (QT ) and
∂x
ZZ  
∂ϕ ∂A(u) ∂ϕ
sgn(u − k) (u − k) − dxdt
QT ∂t ∂x ∂x
ZZ  
∂ϕ (5.35)
− sgn(u − k) (B(u) − B(k)) dxdt ≥ 0,
QT ∂x
∀ 0 ≤ ϕ ∈ C0∞ (QT ), k ∈ R.

The existence and uniqueness of generalized solutions thus defined is discussed by Vol’pert and
Hujaev. However, as pointed out by Wu Zhuoqun, the proof of uniqueness given there is incorrect
due to the adoption of the wrong form of a discontinuity condition for the solutions considered,
which plays an essential role in the proof. Wu and Yin revised this condition. On the basis of the
correct form of the discontinuity condition and a deep study of the properties of BV functions and
BVx functions, they finally completed the proof of uniqueness.

Theorem 5.3 Let u be a BV solution of (5.30). Then H-almost everywhere on Γ ∗u ,

(u+ −u− )γt −(B(u)−B(k))γx −(wr −wl )|γx | = 0, (5.36)

A0 (s) = 0, ∀s ∈ [u∗ , u∗ ], (5.37)

where
∂A(u)
w= , u∗ = min{u+ , u− }, u∗ = max{u+ , u− }.
∂x
In this section, by a BV solution we always mean a generalized solution in the sense of Definition
3.4.1.
Proof. Taking k > |u|L∞ and k < −|u|L∞ respectively, we are led from the inequality (5.35)
to the measure equality
∂u ∂w ∂B(u)
= + (5.38)
∂t ∂x ∂x
∂w
which implies, in particular, that is a Randon measure on QT . The discontinuity conditions
∂x
(5.36), (5.37) will be proved based on the measure equality (5.38).
Without loss of generality, we may assume that I = (0, 1). Denote
 Z Z 
∂w(·, t)
EN = t ∈ (0, T ); |w(x, t)| dx + ≤N .
I I ∂x

From the properties of functions of bounded variation of one variable it follows that
Z Z
∂w(·, t)
sup |w(x, t)| ≤ |w(x, t)| dx +
x∈I I I ∂x
74 Free Boundary of Solutions

and there exists a set F ⊂ (0, T ) with mesF = 0, such that for t ∈ (0, T ) \ F ,

lim χEN (t) = 1.

N →∞

By Lemma 7.7, we have H[(I × F ) ∩ Γ∗u ] = 0. Set

DN = I × E N .

We are ready to show that (5.36) holds H-almost everywhere on DN ∩ Γ∗u for any N . Once this is
done, since the set DN ∩ Γ∗u is increasing with N and tends to the set (I × ((0, T ) \ F )) ∩ Γ∗u as
N → ∞, using the fact H((I × F ) ∩ Γ∗u ) = 0, we conclude that (5.36) holds H-almost everywhere
on Γ∗u .
Let S be an arbitrary bounded and measurable subset of DN ∩ Γ∗u . Integrating (5.38) on S and
using Lemma 7.4 yield
Z ZZ Z
∂w
(u+ − u− )γt dH = + (B(u+ ) − B(u− ))γx dH.
S S ∂x S

Since by Lemma 7.8 and Corollary 7.3, we have

ZZ Z T Z
∂w ∂w(·, t)
= dt
S ∂x 0 St ∂x
Z T X Z



= dt wr (x, t) − wl (x, t) = wr (x, t) − wl (x, t) |γx |dH,
0 x∈S t S

from the arbitrariness of S, we obtain the desired conclusion.

∂A(u)
Now we prove (5.37). Since w = ∈ L1loc (QT ), for any bounded and measurable subset
∂x
S of Γ∗u , we have
ZZ ZZ
∂A(u)
w(x, t)dxdt = = 0.
S S ∂x
On the other hand, using Lemma 7.4 gives
ZZ Z Z u+
∂A(u)
= A0 (s)dsγx dH.
S ∂x S u−

Thus, from the arbitrariness of S, there holds H-almost everywhere on Γ∗u ,

Z u+
γx A0 (s)ds = 0
u−

which proves (5.37), since (5.36) implies, in particular, that γx 6= 0 H-almost everywhere on Γ∗u .
The theorem is proved. 

Theorem 5.4 Let u be a BV solution of (5.30). Then there exists a subset G ⊂ Γ u with
H(G) = 0 such that for any (x, t) ∈ Γu \ G and k ∈ R,

 
sgn(u+ − k) − sgn(u− − k) (u − k)γt − (B(u) − B(k))γx − wγ
e x ≤ 0. (5.39)
Nonlinear Diffusion Equations 75

Proof. Denote z = u − k. Notice that we may replace sgnz by

1

σ= sgnz + + sgnz − .
2

Using the measure equality (5.38) we can derive from the inequality (5.35),
ZZ   ZZ
∂ϕz ∂ϕβz ∂
J(u, k, ϕ) ≡ σ − − σ (ϕw) ≥ 0,
QT ∂t ∂x QT ∂x

where
Z 1
β= B 0 (λu + (1 − λ)k)dλ.
0

By Lemma 7.5 we have

ZZ  
∂ϕz ∂ϕβz
σ −
∂t ∂x
ZQT

 
= − ϕ sgnz + − sgnz − zγt − (B(u) − B(k))γx dH
ZΓ u
≡ − ϕh(k, x, t)dH.
Γu

Thus Z ZZ
∂
ϕh(k, x, t)dH + σ (ϕw) ≤ 0. (5.40)
Γu QT ∂x

For fixed N , choose a sequence {gj (t)} ⊂ C0∞ (0, T ) such that

0 ≤ gj (t) ≤ 1, lim gj (t) = ψN (t) = χEN (t)

j→∞

for almost all t ∈ (0, T ). Replacing ϕ by ϕgj in (5.40) and letting j → ∞, we obtain, by Lemma
7.7 and the dominated convergence theorem,
Z ZZ
∂
ψN ϕh(k, x, t)dH + ψN σ (ϕw) ≤ 0.
Γu QT ∂x

Multiplying this inequality by f (k) with 0 ≤ f (λ) ∈ C0∞ (R) and integrating the resulting inequality
with respect to k over R yield
Z Z 
ψN ϕ f (k)h(k, x, t)dk dH
Γu R Z 
ZZ (5.41)
∂
+ ψN f (k)σdk (ϕw) ≤ 0.
QT R ∂x

Noticing that
Z Z s
f (k)σdk = 2F (u) − F (∞), F (s) = f (λ)dλ)
ZR −∞
∂
(ϕw(·, t)) = 0
I ∂x
76 Free Boundary of Solutions

and using Lemma 7.3 and Lemma 7.8, we obtain

ZZ Z 
∂
ψN f (k)σdk (ϕw)
∂x
ZQZT R
∂
= 2 ψN F (u) (ϕw)
QT ∂x
Z T Z
∂
= 2 ψN (t)dt F (u(x, t)) (ϕw(·, t))
0 I ∂x
Z T Z
= 2 ^t)) ∂ (ϕw(·, t)) .
ψN (t)dt F (u(x,
0 I ∂x
Integrating by parts and using Lemma 7.8 yield
Z T Z
2 ψN (t)dt F (u(x,^t)) ∂ (ϕw(·, t))
0 I ∂x
Z T Z
^ ∂F (u(·, t))
= −2 ψN (t)dt ϕ(x, t)w(x, t)
0 I ∂x
ZZ
∂F (u)
= −2 ψN ϕw
e .
QT ∂x

Thus (5.41) can be written as

Z Z  ZZ
∂F (u)
ψN ϕ f (k)h(k, x, t)dk dH − 2 ψN ϕw
e ≤ 0. (5.42)
Γu R QT ∂x

For any measurable subset S of Γu , similar to Lemma 7.6, we can select a sequence {ϕj } ⊂
C0∞ (QT ) such that
 
∂u ∂u
|ϕj (x, t)| ≤ 1, lim ϕj = χS in L1 QT , + .
j→∞ ∂x ∂t
∂u ∂u
From this and Corollary 7.1 ( for both | | and | | ) it follows that (precisely along a subsequence)
∂x ∂t
lim ϕj (x, t) = χS (x, t)
j→∞

H-almost everywhere on Γu . Then replacing ϕ by ϕj in (5.42), letting j → ∞ and using the

dominated convergence theorem and Lemma 7.4, we obtain
Z Z  Z Z + !
u
ψN f (k)h(k, x, t)dk dH − 2 ψN w
e f (k)dk γx dH ≤ 0.
S R S u−

By the arbitrariness of N and S, there exists a set Gf ⊂ Γu with H(Γu \ Gf ) = 0, such that for
each (x, t) ∈ Gf ,
Z Z u+
− f (k)h(k, x, t)dk + 2w e f (k)dkγx ≥ 0.
R u−

Choose f (λ) ≥ 0 to be a smooth function with f (0) = 1 and suppf = [−1, 1]. Set
 
s−λ
fs,ρ (λ) = f ,
ρ
Nonlinear Diffusion Equations 77

where s, ρ are rational numbers with s ∈ R, 0 < ρ < 1. Denote

\
G0 = Gfs,ρ .
s,ρ

Clearly, H(Γu \ G0 ) = 0, and for any (x, t) ∈ G0 ,

Z Z u+
− fs,ρ (k)h(k, x, t)dk + 2w
e fs,ρ (k)dkγx ≥ 0.
R u−

Noticing that
h(k, x, t) = 0, ∀k∈[u∗ , u∗ ],

we derive at once the following relations: for any (x, t) ∈ G0 ,

−h(k, x, t) + 2w(x,
e t)γx ≥ 0, if k ∈ (u− , u+ ), u− < u+ ,

−h(k, x, t) − 2w(x,
e t)γx ≥ 0, if k ∈ (u+ , u− ), u+ < u− .

From this it follows that (5.39) holds for any k 6= u+ , u− with G = Γu \ G0 . To prove (5.39) for
k = u+ and k = u− , it suffices to let k tend to u− and u+ from the interval with u− and u+ as
endpoints. The proof of Theorem 5.4 is complete. 

As an immediate consequence of (5.36) and (5.39), we have

Corollary 5.2 There exists a set G ⊂ Γ∗u with H(G) = 0, such that for any (x, t) ∈ Γ∗u \ G
and any k ∈ R, there holds

sgn(u+ − k) {(u+ − k)γt − (B(u+ ) − B(k))γx

− (wr sgn+ γx − wl sgn− γx )γx
(5.43)
≤ sgn(u− − k) {(u− − k)γt − (B(u− ) − B(k))γx
− (wl sgn+ γx − wr sgn− γx )γx

Corollary 5.3 H-almost everywhere on Γu ,

γx 6= 0. (5.44)

Proof. Let
Γ0u = {(x, t) ∈ Γu ; γx = 0}.

If (5.44) were not true, then from (5.39), we would have H(Γ0u ) > 0 and for any k ∈ R,

(sgn(u+ − k) − sgn(u− − k))(u − k)γt ≤ 0

holds H-almost everywhere on Γ0u , which is impossible due to the arbitrariness of k and the fact
that γt 6= 0 H-almost everywhere on Γ0u . 

Remark 5.1 From Corollary 5.3 it follows that Γu = Γ∗u except for a set of Hausdorff measure
zero and hence (5.36), (5.37) and (5.43) hold H-almost everywhere on Γ u .
78 Free Boundary of Solutions

5.3 A Free Boundary Problem

As indicated in the previous section, the strongly degenerate parabolic equations may admit dis-
continuous solutions in general. In this section, we consider a special problem of the form

∂u ∂ 2 A(u)
= ,
∂t ∂x2
u(x, 0) = u0 (x),

where

A(u) = 0, for u < 0, and A(u) = um , for u > 0,

u0 (x) = α < 0 for x < 0, and u0 (x) = β > 0, for x > 0,

and m > 1 and α, β are constants. For this equation, the degenerate set becomes E = (−∞, 0],
and hence is of strongly degenerate type. We have prescribed an initial value with a discontinuous
point x = 0, and hence there should be a line of discontinuity x = λ(t) with λ0 (t) < 0 starting
from (0, 0). The purpose lies in getting a result how does the discontuity emerge and develop?
First of all, let us recall that at the jump point, there is a discontinuity condition

A0 (s) = 0, for u ∈ [u∗ , u∗ ],

which implies that

u+ = 0.
x=λ(t)

Next, taking into account the discontinuity equality

" r  l #
∂A(u) ∂A(u)
(u+ − u− )νt − − νx = 0,
∂x ∂x

we get
 r
∂A(u)
= αλ0 (t).
∂x x=λ(t)

So, in the region {(x, t); x < λ(t)}, it follows from E = (−∞, 0] that u(x, t) ≡ β, while in the
region {(x, t); x > λ(t)}, u satisfies the following free boundary problem

∂u ∂ 2 um
= , x > λ(t), 0 < t < ∞, (5.45)
∂t ∂x2
u = 0, 0 < t < ∞, (5.46)
x=λ(t)
m
∂u
= αλ0 (t), 0 < t < ∞, (5.47)
∂x x=λ(t)
u(x, 0) = β, x > 0. (5.48)

We seek solutions of the form

x
u = u(ξ), ξ= . (5.49)
t1/2
Nonlinear Diffusion Equations 79

Then we arrive at a free boundary problem for an ordinary differential equation, namely
1
(um )00 = − ξu0 ,
2
u = 0,
ξ=ξ0
α
(um )0 = ξ0 ,
ξ=ξ0 2
u = β.
ξ=∞

Let v = um . Then the problem transformed into

ξ 1/m−1 0
v 00 = − v v, (5.50)
2m
v = 0, (5.51)
ξ=ξ0
α
v0 = ξ0 , (5.52)
ξ=ξ0 2
v = βm. (5.53)
ξ=∞

To prove the existence of solutions of the free boundary problem, we need a series of estimates
on the solutions. First, we have
Proposition 5.3 For ξ0 ≤ ξ ≤ 0, there hold the following estimates
αξ0 αRξ0
(ξ − ξ0 ) ≤ v(ξ) ≤ (ξ − ξ0 ), (5.54)
2 2
αξ0 αRξ0
≤ v 0 (ξ) ≤ , (5.55)
2 2
where "   #
1/m−1
1 |α| 2/m
R = exp |ξ0 | .
2 2

Proof. Integrating (5.50) and using (5.51), (5.52), we get

Z ξ !
0 αξ0 1 1/m−1
v (ξ) = exp − τv (τ )dτ ,
2 2m ξ0
Z  Z τ 
αξ0 ξ 1 1/m−1
v(ξ) = exp − sv (s)ds dτ.
2 ξ0 2m ξ0
The first equality clearly implies the first part of (5.55), and consequently the first part of (5.54).
Further, by noticing that ξ0 ≤ ξ ≤ 0, it follows that
" Z ξ  1/m−1 #
0 αξ0 1 αξ0 1/m−1
v (ξ) ≤ exp − τ (τ − ξ0 ) dτ
2 2m ξ0 2
"  1/m−1 Z ξ #
αξ0 1 |α|
≤ exp |ξ0 |1/m (τ − ξ0 )1/m−1 dτ
2 2m 2 ξ0
"   #
1/m−1
αξ0 1 |α| 2/m
≤ exp |ξ0 | ,
2 2 2
80 Free Boundary of Solutions

which implies the second part of (5.55), and consequently the second part of (5.54). The proof is
complete. 

Proposition 5.4 Let [ξ0 , ξ1 ) be the maximal existence interval of v(ξ). Then for any ξ ∈
(ξ0 , ξ1 ),

0 < v(ξ) < αM ξ0 , (5.56)

 
αξ0 1
0 < v 0 (ξ) ≤ R exp − (αM ξ0 )1/m−1 ξ 2 , (5.57)
2 4m

where M is a constant, continuously in ξ0 ∈ (−∞, 0] and determined by

 Z ∞ 
R|ξ0 | 1/2 (m−1)/2m (m−1)/2m M
+ Rm |α| |ξ0 | exp(−η )dη M (m−1)/2m ≤
2
.
2 0 2

Proof. It is obvious that v(ξ) > 0 and v 0 (ξ) > 0 for any ξ ∈ (ξ0 , ξ1 ). We first show the
second part of (5.56). If ξ1 ≤ 0, then the conclusion follows at once from (5.54). Now, we assume
that ξ1 > 0. Since v(0) < αM ξ0 , v(ξ) < αM ξ0 holds at least in some right neighborhood of ξ = 0.
If the conclusion were not true, then there exist some ξ2 ∈ (0, ξ1 ), such that

v(ξ) < αM ξ0 for 0 ≤ ξ < ξ2 and v(ξ2 = αM ξ0 ).

Express v(ξ) as the following

Z  Z τ 
αξ0 0 1 1/m−1
v(ξ) = exp − sv (s)ds dτ
2 ξ0 2m ξ0
Z  Z τ 
αξ0 ξ 1
+ exp − sv 1/m−1 (s)ds dτ ≡ I1 + I2 .
2 0 2m ξ0

Using the first part of (5.54), we get

R|ξ0 |
I1 ≤ αξ0 ,
2
and by virtue of the definition of ξ2 , we get for 0 ≤ ξ < ξ2 ,
Z  Z τ 
αξ0 ξ 1 1/m−1
I2 = exp − sv (s)ds dτ
2 0 2m ξ0
Z  Z 0   Z τ 
αξ0 ξ 1 1/m−1 1 1/m−1
= exp − sv (s)ds exp − sv (s)ds dτ
2 0 2m ξ 2m 0
"   0 #Z  Z τ 
1/m−1 ξ
αξ0 1 |α| 2/m 1 1/m−1
≤ exp |ξ0 | exp − sv (s)ds dτ
2 2 2 0 2m 0
Z ξ  Z τ 
αξ0 1 1/m−1
≤ R exp − s(αM ξ0 ) ds dτ
2 0 2m 0
Z ξ  
αξ0 1
= R exp − (αM ξ0 )1/m−1 τ 2 ds dτ
2 0 4m
Z γξ
αξ0
= Rγ −1 exp(−s2 )ds
2 0
 Z ∞ 
<αξ0 Rm |α|(m−1)/2m |ξ0 |(m−1)/2m
1/2
exp(−s2 )ds · M (m−1)/2m ,
0
Nonlinear Diffusion Equations 81

where  1/2
1
γ= (αM ξ0 )1/m−1 .
4m
Hence, by the choice of M ,
 Z ∞ 
R|ξ0 | M
v(ξ2 ) ≤ αξ0 + Rm1/2 |α|(m−1)/2m |ξ0 |(m−1)/2m exp(−s2 )ds · M (m−1)/2m ≤ αξ0 ,
2 0 2
which contradicts the definition of ξ2 , and the estimate (5.56) is proved. Moreover, the estimate
(5.57) follows from (5.56) immediately. The proof is complete. 
Using the above established estimates, we can now discuss the solvability of solutions of the
free boundary problem. First, we have
Theorem 5.5 Let ξ0 < 0 be fixed. Then the problem (5.45)–(5.47) admits a unique solution
v ∈ C 1 [ξ0 , ∞) ∩ C 2 (ξ0 , ∞) with v(ξ) > 0 for ξ > ξ0 , which is continuously depends on ξ0 and α.
Proof. First we prove the existence. By virtue of the estimates in Proposition 5.3 and
Proposition 5.4, it suffices to prove (5.45)–(5.47) has a solution on ξ0 ≤ ξ ≤ 0, which is positive
except at the point ξ = ξ0 .
We adopt Schauder’s fixed point theorem to show the existence. For this purpose, we denote
by M the class of functions v ∈ C[ξ0 , 0] such that
0 < v(ξ) ≤ N for ξ > ξ0 ,
where N is a constant satisfying
1
(|α| + 3N 1/m )|ξ0 |2 ≤ N.
2
Clearly, M is a closed convex subset of C[ξ0 , 0].
Consider the operator
Z ξ
αξ0 1
w = Tv = (ξ − ξ0 ) + (ξ − 2τ )v 1/m (τ )dτ.
2 2 ξ0

For any v ∈ M, w = T v ∈ C[ξ0 , 0] and w = T v > 0 except at ξ = ξ0 . Moreover

|α| 3
w≤ |ξ0 |2 + |ξ0 |2 N 1/m ≤ N.
2 2
So, w ∈ M. It is easy to see that T M is compact and T is continuous on M. Thus, Schauder’s
fixed point theorem gives the existence of solutions of the fllowing integral equation
Z
αξ0 1 ξ
v(ξ) = (ξ − ξ0 ) + (ξ − 2τ )v 1/m (τ )dτ (5.58)
2 2 ξ0
and hence to the orginal problem (5.45)–(5.47).
To prove the uniqueness of solutions, let v1 , v2 be two solutions of the problem (5.45)–(5.47).
It suffices to prove that v = v1 − v2 = 0 in a neighborhood of ξ0 , say ξ0 ≤ ξ + δ with small δ.
From (5.58), we have
Z
1 ξ 1/m 1/m
v(ξ) = (ξ − 2τ )(v1 (τ ) − v2 (τ ))dτ
2 ξ0
Z
1 ξ
= (ξ − 2τ )ṽ 1/m−1 (τ )v(τ ))dτ,
2 ξ0
82 Free Boundary of Solutions

where ṽ(τ ) is a certain point between v1 (τ ) and v2 (τ ) and hence from Proposition 5.3,
αξ0
ṽ(τ ) ≥ (τ − ξ0 ).
2
Thus for ξ0 ≤ ξ ≤ ξ0 + δ,
Z ξ  1/m−1
3 1 αξ0
|v(ξ)| ≤ |ξ0 | (τ − ξ0 )1/m−1 |v(τ )|dτ
2 ξ0 m 2
 1/m−1
3 |α|
≤ (δ|ξ0 |)1/m max |v(ξ)|,
2 2 ξ0 ≤ξ≤ξ0 +δ

which is impossible for δ small enough, unless v ≡ 0 on ξ0 ≤ ξ ≤ ξ0 + δ.

Finally, we prove the continuous dependence of solutions upon ξ0 < 0 and α < 0. Denote by
v(ξ; ξ0 , α) the solution of (5.45)–(5.47). Let ξ0 < 0, α < 0 be fixed. What we want to prove is that
for any ξ1 > ξ0 and ε > 0, there exists a constant δ > 0 such that
|v(ξ; ξ00 , α0 ) − v(ξ; ξ0 , α)| < ε, |v 0 (ξ; ξ00 , α0 ) − v(ξ; ξ0 , α)| < ε
whenever |ξ00 − ξ0 | < δ, |α0 − α| < δ and max(ξ00 , ξ0 ) ≤ ξ ≤ ξ1 .
(n) (n)
If it were not the case, then there would exist ε0 > 0 and ξ0 , α0 , ξ (n) such that
(n)
ξ0 → ξ0 , α(n) → α (n → ∞),
(n)
max(ξ0 , ξ0 ) ≤ ξ (n) ≤ ξ1
and
(n)
|v(ξ (n) ; ξ0 , α(n) ) − v(ξ (n) ; ξ0 , α)| ≥ ε0 (5.59)
or
(n)
|v 0 (ξ (n) ; ξ0 , α(n) ) − v 0 (ξ (n) ; ξ0 , α)| ≥ ε0 .
Suppose, for example, the case is the former, namely, (5.59). We have
(n) Z
(n) α(n) ξ0 1 (n) 1 ξ 1/m (n)
v 0 (ξ; ξ0 , α(n) ) = − ξv 1/m (ξ; ξ0 , α(n) ) + v (τ ; ξ0 , α(n) )dτ
2 2 2 ξ0(n)
(n) Z
(n) α(n) ξ0 (n) 1 ξ (n)
v(ξ; ξ0 , α(n) ) = (ξ − ξ0 ) + (ξ − 2τ )v 1/m (τ ; ξ0 , α(n) )dτ. (5.60)
2 2 ξ0(n)

From Proposition 5.3 and Proposition 5.4, we see that v(ξ; ξ00 , α0 ) and v 0 (ξ; ξ00 , α0 ) are bounded
uniformly in (ξ 0 , α0 ) in a small neighborhood of (ξ0 , α) and hence from (5.59), it is easy to
(n) (n) (n)
see that {v(ξ; ξ0 , α(n) )} and {v 0 (ξ; ξ0 , α(n) )} are equicontinuous. Thus {v(ξ; ξ0 , α(n) )} and
(n) (n)
{v 0 (ξ; ξ0 , α(n) )} have subsequences converging unformly. Suppose they are {v(ξ; ξ0 , α(n) )} and
(n)
{v 0 (ξ; ξ0 , α(n) )} themselves and
(n) (n)
lim {v(ξ; ξ0 , α(n) )} = w(ξ), lim {v 0 (ξ; ξ0 , α(n) )} = w̃(ξ). (5.61)
n→∞ n→∞

From (5.59), (5.59) we get

Z
αξ0 1 1 ξ 1/m
w̃(ξ) = − ξw1/m (ξ) + w (τ )dτ
2 2 2 ξ0
Z
αξ0 1 ξ
w(ξ) = (ξ − ξ0 ) + (ξ − 2τ )w1/m (τ )dτ. (5.62)
2 2 ξ0
Nonlinear Diffusion Equations 83

whence w(ξ) is a solution of (5.45)–(5.48) and w̃(ξ) = w 0 (ξ). By uniqueness, w(ξ) ≡ v(ξ; ξ0 , α).
On the other hand, we may suppose that {ξ (n) } converges:
˜
lim ξ (n) = ξ.
n→∞

Letting n → ∞ in (5.58), from the uniform convergence we get

˜ − v(ξ; ξ0 , α)| ≥ ε0 .
|w(ξ)

The contradiction means that what we want to prove is true. The proof of the theorem is complete.


Theorem 5.6 The free boundary problem (5.45)–(5.48) admits at least one solution (v(ξ), ξ 0 )
with ξ0 < 0 and v ∈ C 1 [ξ0 , ∞) ∩ C 2 (ξ0 , ∞), with v(ξ) and −ξ0 strictly decreasing in α and strictly
increasing in β.

Proof. For any fixed ξ00 < 0, let v(ξ; ξ00 ) be the solution of the problem (5.45)–(5.47) corre-
sponding to ξ = ξ00 . By Proposition 5.4, we see that v(ξ) is strictly increasing and bounded above.
So, the limit
lim v(ξ; ξ00 ) = γ(ξ00 )
ξ→∞

exists. Let
E = {ξ00 < 0; γ(ξ00 ) < β m }.
Still from Proposition 5.4, we see that

lim γ(ξ00 ) = 0,
ξ00 →0−

which implies that the set E is nonempty. On the other hand, by virtue of the strict monotonicity
of v(ξ) and Proposition 5.3,
α(ξ00 )2
γ(ξ00 ) ≥ v(0; ξ00 ) ≥ − ,
2
which implies that
0
lim γ(ξ00 ) = +∞.
ξ0 →−∞

Therefore, the set E is bounded below, and so, we may define

ξ0 = inf E.

We conclude that
lim v(ξ; ξ0 ) = γ(ξ0 ) = β m .
ξ→+∞

Suppose the contrary, then there would be γ(ξ0 ) 6= β m . Denote by δ = |γ(ξ0 ) − β m |. From
Proposition 5.3, it is clear that we can choose positive constants c1 and c2 such that

0 < v 0 (ξ; ξ00 ) ≤ c1 exp(−c2 ξ 2 ), ξ≥0 (5.63)

for any ξ00 ∈ [3ξ0 /2, ξ0 /2]. Then we choose ξ1 such that
Z ∞
δ
c1 exp(−c2 ξ 2 )dξ < .
ξ1 4

By the continuous dependence of solutions on ξ00 , there exists η ∈ (0, |ξ0 |/2, such that
δ
|v(ξ1 ; ξ00 ) − v(ξ1 ; ξ0 )| < (5.64)
4
84 Free Boundary of Solutions

whenever ξ00 ∈ [ξ0 − η, ξ0 + η].

Now, for ξ00 ∈ [ξ0 − η, ξ0 + η], integrating (5.63) over (ξ1 , ∞), we get
δ
0 < γ(ξ00 ) − v(ξ1 ; ξ00 ) <
4
and, in particular,
δ
0 < γ(ξ0 ) − v(ξ1 ; ξ0 ) < ,
4
which, together with (5.64), implies
δ
|γ(ξ00 ) − γ(ξ0 )| < .
2
If γ(ξ0 ) = β m − δ, then
δ
γ(ξ00 ) = γ(ξ0 ) + γ(ξ00 ) − γ(ξ0 ) < β m −
2
for any ξ00 ∈ [ξ0 − η, ξ0 + η]. This means that [ξ0 − η, ξ0 + η] ⊂ E, which contradicts the definition
of ξ0 .
If γ(ξ0 ) = β m + δ, then
δ
γ(ξ00 ) = γ(ξ0 ) + γ(ξ00 ) − γ(ξ0 ) > β m +
2
for any ξ00 ∈ [ξ0 − η, ξ0 + η]. This means that no point on [ξ0 − η, ξ0 + η] is in E, which also
contradicts the definition of ξ0 . The proof is complete. 

Theorem 5.7 The free boundary problem (5.45)–(5.48) admits at most one solution (v(ξ), ξ 0 ).

To prove this theorem, we first need a proposition, which will be used in the proof of the
theorem.

Proposition 5.5 Let (v1 (ξ), ξ0,1 ) and (v2 (ξ), ξ0,2 ) be two different solutions of the free bound-
ary problem (5.45)–(5.48).
(1) v1 (0) 6= v2 (0);
(2) If v1 (0) > v2 (0), then v10 (0) < v20 (0);
(3) If v1 (0) < v2 (0), then v10 (0) > v20 (0).

Proof. Since the arguments are quite similar for the three conclusions, as an example, we
show the first one. If the conclusion were not true, then the standard uniqueness theorem implies
that v10 (0) 6= v20 (0). Without loss of generality, we may assume that v10 (0) > v20 (0), which implies
that in a right neighborhood of ξ = 0, v1 (ξ) > v2 (ξ). Hence
Z ξ !
0 0 1 1/m−1
v1 (ξ) =v1 (0) exp − τ v1 (τ )dτ
2m 0
Z ξ !
1 1/m−1
>v20 (0) exp − τ v1 (τ )dτ
2m 0
Z ξ !
0 1 1/m−1
>v2 (0) exp − τ v2 (τ )dτ
2m 0
=v20 (ξ)
Nonlinear Diffusion Equations 85

holds in the same neighborhood, and so for all ξ > 0. In addition,

Z !
ξ
1 1/m−1
v10 (ξ) − v20 (ξ) ≥ (v10 (0) − v20 (0)) exp − τ v1 (τ )dτ ,
2m 0

which implies
β m = lim v1 (ξ) > lim v2 (ξ) = β m ,
ξ→∞ ξ→∞

which is impossible. The proof is complete. 

Proof of Theorem 5.7. We argue by contradiction. Suppose that there were two solutions
(v1 (ξ), ξ0,1 ) and (v2 (ξ), ξ0,2 ) of the problem (5.45)–(5.48). Then by the uniqueness result in Theo-
rem 5.5, we must have ξ0,1 6= ξ0,2 . Now, we apply Proposition 5.5. Since v1 (0) 6= v2 (0), we discuss
in the following two cases.
(1) If v1 (0) > v2 (0), then v10 (0) < v20 (0). We first conclude that

ξ0,1 < ξ0,2 . (5.65)

In fact, if this were not true, then there would be ξ0,1 > ξ0,2 . Let ξ1 (v) and ξ2 (v) be the inverse
functions of v1 (ξ) and v2 (ξ) respectively, and set ξ(v) = ξ1 (v) − ξ2 (v). We consider the function
ξ(v) in the interval [0, v2 (0)]. Using the strict monotonicity of ξ1 (v), we see that

ξ1 (0) = ξ0,1 > ξ0,2 = ξ2 (0),

ξ1 (v2 (0)) < ξ1 (v1 (0)) = 0 = ξ2 (v2 (0)),

which implies that there exists a point v0 ∈ (0, v2 (0)) such that

ξ(v) > 0 for 0 ≤ v < v0 and ξ(v0 ) = 0.

Noticing that
2 2
ξ 0 (0) = − > 0,
αξ0,1 αξ0,2
we may conclude that for some point v ∈ (0, v0 ), ξ(v) achieves its maximum, and hence

ξ(v) > 0, ξ 0 (v) = 0, ξ 00 (v) ≤ 0.

On the other hand, using the equation satisfied by ξ1 , ξ2 , we have

1 1/m−1
ξ 00 =ξ100 − ξ200 = v (ξ1 ξ10 − ξ2 ξ20 )
2m
1 1/m−1 1 1/m−1
= v (ξ1 − ξ2 )ξ10 + v ξ2 (ξ10 − ξ20 )
2m 2m
1 1/m−1 0 1 1/m−1 0
= v ξξ1 + v ξ2 ξ . (5.66)
2m 2m
Therefore, ξ 00 (v) > 0. The contradiction proves (5.65).
However, using (5.65), we get
2 2
ξ(0) < 0, ξ 0 (0) = − < 0.
αξ0,1 αξ0,2
Recalling that v10 (0) > v20 (0), and noticing the strict monotonicity of ξ10 (v), we have
1 1
ξ10 (v2 (0)) > ξ10 (v1 (0)) = > 0 = ξ20 (v2 (0)),
v10 (0) v2 (0)
86 Free Boundary of Solutions

which implies ξ 0 (v2 (0)) > 0. So, just as dis above, the function ξ(v) achieves its minimum at some
=
point v, at which
= = =
ξ( v) < 0, ξ 0 ( v) = 0, ξ 00 ( v) ≥ 0.
=
While from (5.66), ξ 00 ( v) < 0, which is a contraction.
(2) If v1 (0) < v2 (0), then v10 (0) > v20 (0), and the similar arguments as above may result in
another contradiction.
Summing up, we have proved the uniqueness. 
Nonlinear Diffusion Equations 87

6 Properties of Solutions
6.1 Regularity of Solutions
Consider the porous medium equation
∂u ∂ 2 um
= (6.1)
∂t ∂x2
with m > 1, subject to the initial value condition

u(x, 0) = u0 (x) (x ∈ R). (6.2)

It is shown that generalized solutions of the above problem are all continuous, which satisfy
the equation in classical sense in a neighborhood of any “point of non-degeneracy”. However, as
∂u
pointed by Kalashnikov, solutions of the equation may have discontinuous derivative , even if
∂x
its initial data are sufficiently smooth.
In this section a thorough investigation on the regularity of generalized solutions of the equation
(6.1) will be made.
To this purpose we first prove

Lemma 6.1 Let m > 1 and u be a positive and smooth solution of (6.1) on G = (a, b) × (0, T ].
∂3u
Here by ”smooth” we mean u ∈ C 2 (G) ∩ C(G), such that ∈ C(G). Then for any τ ∈ (0, T ),
∂x3
a1 , b1 ∈ (a, b) (a1 < b1 ), we have on G1 = (a1 , b1 ) × [τ, T ],

∂um−1 (x, t)
≤ C, (6.3)
∂x
where C = C(τ, m, M, a1 − a, b1 − b), M = sup u. If
G

du0m−1 (x)
M1 = sup < ∞,
(a,b) dx

then (6.3) holds on (a1 , b1 ) × [0, T ] where C depends on M1 instead of τ .

Proof. Let v = um−1 . Then v satisfies

 2
∂v ∂2v m ∂v
= mv 2 + .
∂t ∂x m−1 ∂x

∂v ∂2v
Here we have used the assumption u > 0, otherwise and may not exist everywhere.
∂x ∂x2
∂v ∂v
Since our goal is to estimate , a natural approach is to derive and use the equation for .
∂x ∂x
However it seems difficult to achieve our goal in this way. So we consider a transformation of v:
v = ϕ(w) with ϕ to be determined later. We first require that

ϕ : [0, 1] → [0, M ∗ ], ϕ0 (r) > 0, (6.4)

where M ∗ = M m−1 . Then w satisfies

 2  2
∂w ∂2w ϕ00 ∂w m ∂w
= mϕ 2 + mϕ 0 + ϕ0 .
∂t ∂x ϕ ∂x m−1 ∂x
88 Properties of Solutions

∂w
Let p = . Then p satisfies
∂x
 00 0 !
∂p ∂2p m2 00 ϕ
= mϕ 2 + ϕ + mϕ p3
∂t ∂x m−1 ϕ0
 
m(m + 1) 0 ϕ00 ∂p
+ ϕ + wmϕ 0 p
m−1 ϕ ∂x

or after multiplying both sides by p,

 00 0 !
1 ∂p2 ∂2p m2 00 ϕ
=mϕp + ϕ + mϕ p4
2 ∂t ∂x m−1 ϕ0
 
m(m + 1) 0 ϕ00 ∂p
+ ϕ + 2mϕ 0 p2 . (6.5)
m−1 ϕ ∂x

Now we choose ζ = ζ(x, t) ∈ C 2 (G) satisfying

ζ = 1 for (x, t) ∈ G1 ,

ζ = 0 for x = a, b and t = 0,

0 ≤ ζ ≤ 1 for (x, t) ∈ G,
 
∂ζ C ∂ζ 1 1
≤ , ≤ C max , ,
∂t τ ∂x a 1 − a b1 − b
 
∂2ζ 1 1
≤ C max ,
∂x2 (a1 − a)2 (b1 − b)2

and use ζ(x, t) to cut off the function p2 , namely to consider the function z = ζ 2 p2 instead of p2 .
Multiplying (6.5) by ζ 2 , we can check that z satisfies
 
1 ∂z ∂2z
− mϕ 2
2 ∂t ∂x
 
p m(m + 1) 0 ϕ00 ∂z
= ϕ + 2mϕ 0
2 m−1 ϕ ∂x
 2 !
∂ζ ∂ζ ∂2ζ
+ ζ − mϕ − mϕζ 2 p2
∂t ∂x ∂x
 00

m(m + 1) 0 ϕ ∂ζ
− ϕ + 2mϕ 0 ζ p3
m−1 ϕ ∂x
 00 0 !
m2 00 ϕ
+ ϕ + mϕ ζ 2 p4
m−1 ϕ0
 2
2 ∂p ∂ζ ∂p
− mϕζ − 4mϕζ p . (6.6)
∂x ∂x ∂x

Let (x0 , t0 ) ∈ G be a point where z achieves its maximum. Then at this point

∂z ∂z ∂2z
≥ 0, = 0, 2 ≤ 0. (6.7)
∂t ∂x ∂x
Nonlinear Diffusion Equations 89

Since
∂z ∂p ∂ζ
= 2ζ 2 p + 2ζ p2 ,
∂x ∂x ∂x
at (x0 , t0 ) we have
∂p ∂ζ
ζ = −p . (6.8)
∂x ∂x
Using (6.7) and (6.8), from (6.6) we obtain
 2 !  2
∂ζ ∂ζ ∂2ζ ∂ζ
ζ − mϕ − mϕζ 2 p2 + 3mϕ p2
∂t ∂x ∂x ∂x
 00 0 !
m2 00 ϕ
+ ϕ + mϕ ζ 2 p4
m−1 ϕ0
 
m(m + 1) 0 ϕ00 ∂ζ
− ϕ + 2mϕ 0 ζ p3 ≥ 0,
m−1 ϕ ∂x
namely
 00 0 !
m2 00 ϕ
− ϕ + mϕ ζ 2 p4
m−1 ϕ0
 2 !
∂ζ ∂ζ ∂2ζ
≤ ζ + 2mϕ − mϕζ 2 p2
∂t ∂x ∂x
 00

m(m + 1) 0 ϕ ∂ζ
− ϕ + 2mϕ 0 ζ p3 . (6.9)
m−1 ϕ ∂x
We hope that the coefficient of the left side is positive. It suffices to require
 00 0
00 ϕ ϕ000 ϕ0 − (ϕ00 )2
ϕ < 0, = < 0,
ϕ0 (ϕ0 )2
or ϕ000 ϕ0 − (ϕ00 )2 < 0. (6.10)
Clearly ϕ(r) = αr 2 + (M ∗ − α)r with −M ∗ < α < 0 satisfies both (6.4) and (6.10). In particular,
we choose
M ∗r
ϕ(r) = (4 − r).
3
Substituting this function into (6.9) and using Young’s inequality, we obtain
c22
ζ 2 p2 ≤ C1 + C2 ζ|p| ≤ C1 + εζ 2 p2 + (ε > 0)
4ε
or
c22
(1 − ε)ζ 2 p2 ≤ C1 +
4ε
1
with Ci = Ci (τ, m, M, a1 − a, b1 − b) (i = 1, 2). Taking ε = yields
2
ζ 2 p2 ≤ C.
Hence
∂v ∂w
= ϕ0 (w) = ϕ0 (w)|p| ≤ C on G1 .
∂x ∂x
This complete the proof of the first conclusion. 
90 Properties of Solutions

Similarly we can prove the second conclusion. To do this we choose ζ(x) ∈ C02 (a, b) such that
0 ≤ ζ(x) ≤ 1, ζ = 1 for x ∈ (a1 , b1 ).

Theorem 6.1 Suppose that u0 (x) is nonnegative and bounded, um 0 (x) is Lipschitz continuous
and u(x, t) is a generalized solution of (6.1), (6.2). Then
 
α,α/2 1
1. for any τ ∈ (0, T ), u ∈ C (R × [τ, T ]) with α = min 1, and u ∈ C α,α/2 (R ×
m−1
[0, T ]) provided u0m−1 (x) is Lipschitz continuous;
∂um ∂um
2. exists and is continuous in x, in particular, = 0 whenever u(x, t) = 0;
∂x ∂x
∂u ∂u
3. exists and is continuous in x if 1 < m < 2, in particular, = 0 whenever u(x, t) = 0.
∂x ∂x

α,α/2
Remark 6.1 The Barenblatt solution (1.37) shows  that u ∈ C (R × [τ, T ]) is the best
1
possible global result and the Hölder exponent α = min 1, cannot be increased.
m−1

Proof of Theorem 6.1. The case u0 ≡ 0 is trivial. We suppose that sup u0 > 0.
1. From the proof of existence theorem and also using the uniqueness theorem we assert that
u(x, t) can be obtained as the limit of the classical solution un (x, t) of the problem


 ∂u ∂ 2 um

 = |x| < n, t > 0,
 ∂t ∂x2
 u(x, 0) = u0n (x) |x| ≤ n,




u(±n, t) = M = sup u0 0 < t < T,

and un (x, t) uniformly converges to u(x, t) on any compact subset of QT . Here u0n (x) > 0 and
{u0n (x)} uniformly converges to u0 (x) on any finite interval.
Given any τ ∈ (0, T ) and a < b. By Lemma 6.1, for any large n and (x, t) ∈ Gτ = [a, b] × [τ, T ],
we have
∂unm−1
≤C (6.11)
∂x

with some constant C = C(τ, m, M ) independent of a, b.

It follows by virtue of the uniform boundedness of un

∂um
n
≤C on Gτ . (6.12)
∂x

Checking the proof of existence theorem we see that in fact this estimate has been established
there.
Let (x, t), (y, s) ∈ Gτ . Based on (6.12) we see from the proof of Theorem 2.1 that for some
x ∈ [x, x + |t − s|1/2 ],
∗

|un (x∗ , t) − un (x∗ , s)| ≤ C|t − s|1/2 . (6.13)

Therefore if m > 2, then

|unm−1 (x∗ , t) − unm−1 (x∗ , s)| ≤ C|t − s|1/2 .

Nonlinear Diffusion Equations 91

This and (6.11) imply

|unm−1 (x, t) − unm−1 (y, s)|

≤ |unm−1 (x, t) − unm−1 (x∗ , t)| + |unm−1 (x∗ , t) − unm−1 (x∗ , s)|

+|unm−1 (x∗ , s) − unm−1 (y, s)|

≤ C(|x − x∗ | + |t − s|1/2 + |x∗ − y|) ≤ C(|x − y| + |t − s|1/2 ).

Using the inequality |a − b|α ≤ |aα − bα | (a > 0, b > 0, α ≥ 1) we further obtain

|un (x, t) − un (y, s)|m−1 ≤ |um−1 (x, t) − um−1 (y, s)| ≤ C(|x − y| + |t − s|1/2 )

or
|un (x, t) − un (y, s)| ≤ C(|x − y|α + |t − s|α/2 )
1
with α = .
m−1
If 1 < m < 2, then using (6.11) again yields

∂un 1 ∂um−1 1 ∂unm−1

≤ un2−m n ≤ un2−m ≤C
∂x m−1 ∂x m−1 ∂x

and hence
|un (x, t) − un (y, t)| ≤ C|x − y|.
This and (6.13) give

|un (x, t) − un (y, s)|

≤ |un (x, t) − un (x∗ , t)| + |un (x∗ , t) − un (x∗ , s)|

+|un (x∗ , s) − un (x, s)|

≤ C(|x − x∗ | + |t − s|1/2 + |x∗ − y|) ≤ C(|x − y| + |t − s|1/2 ).

Summing up, in either case we have

|un (x, t) − un (y, s)| ≤ C(|x − y|α + |t − s|α/2 )

 
1
with α = min 1, , which implies that u ∈ C α,α/2 (R × [τ, T ]) by letting n → ∞.
m−1
If u0m−1 is Lipschitz continuous, then by Lemma 6.1, (6.11) holds for (x, t) ∈ [a, b] × [0, T ].
Based on this fact we can prove u ∈ C α,α/2 (R × [0, T ]).
2. If u(x0 , t0 ) > 0, then from the existence theorem, we see that u is a classical solution in
∂um
some neighborhood of (x0 , t0 ), in particular, exists and is continuous near (x0 , t0 ).
∂x
If u(x0 , t0 ) = 0 (t0 > 0), then for any given δ > 0, from the first conclusion of our theorem

0 ≤ u(x, t0 ) = u(x, t0 ) − u(x0 , t0 ) ≤ Cδ α

for x ∈ Iδ = [x0 − δ, x0 + δ]. Hence for large n, we have

un (x, t0 ) ≤ C1 δ α for x ∈ Iδ ,
92 Properties of Solutions

and Z y
∂um
n (ξ, t0 )
|um
n (x, t0 ) − um
n (y, t0 )| = dξ
x x
Z y
m ∂um−1
= un (ξ, t0 ) n (ξ, t0 )dξ
m−1 x ∂x
m α
≤ C1 δ |x − y|
m−1
for x, y ∈ Iδ . Here we have used (6.11) again. Letting n → ∞ gives
mC1 α
|um (x, t0 ) − um (y, t0 )| ≤ δ |x − y| for x, y ∈ Iδ
m−1
∂um (x, t0 )
with some constant C1 , which implies that is continuous at x = x0 .
∂x
3. Suppose 1 < m < 2. Since u is a classical solution of (6.1) whenever u(x, t) > 0, we need to
treat only the case that u(x0 , t0 ) = 0 for some (x0 , t0 ). We have

∂unm−1 (x, t0 )
≤ C1 un2−m (x, t0 ) ≤ C1 δ α(2−m) for x ∈ Iδ .
∂x

∂u(x, t0 )
From this it is easily seen that is continuous at x = x0 . 
∂x

6.2 Localization and Extinction of Disturbances

Consider equations of the form

∂u ∂ 2 A(u)
= − c(u) (x, t) ∈ Q, (6.14)
∂t ∂x2
subject to the initial value condition

u(x, 0) = u0 (x) x ∈ R. (6.15)

where A(u), c(u) are appropriately smooth. In addition to the assumptions on A(u) and u 0 given
at the beginning of this section, we assume that c0 (u) > 0 for u > 0 and c(0) = 0.
We can prove the existence and uniqueness of generalized solutions (which can be defined in
an obvious way) of the Cauchy problem for (6.14). Also the comparison theorem is valid. In the
sequel, we will use these results without proof.

Theorem 6.2 Let u be a generalized solution of the Cauchy problem (6.14), (6.15) on Q. If
u0 (x) = 0 for |x| ≥ X > 0 and
Z 1 Z v −1/2
Ψ(ξ)dξ dv < +∞, (6.16)
0 0

where Ψ(v) = c(Φ(v)), Φ(v) = A−1 (v), then there exists X1 > X such that u(x, t) = 0 for |x| ≥ X,
t ≥ 0.

Proof. To prove that for some X1 ≥ X, u(x, t) = 0 for x ≥ X, t ≥ 0, it suffices to construct

a generalized solution w(x, t) on the domain GX = {(x, t); x > X, t > 0} such that

u(x, t) ≤ w(x, t) for (x, t) ∈ GX (6.17)

Nonlinear Diffusion Equations 93

and w(x, t) = 0 for x ≥ X1 , t ≥ 0.

We try to seek such generalized solution w(x, t) among functions which depend only on x. First
we require w(x) to satisfy (6.14) on GX \GX1 , namely, for X < x < X1 ,
d2
A(w(x)) = c(w(x))
dx2
or
d2 j(x)
= c(Φ(j(x))) = Ψ(j(x)), (6.18)
dx2
where j(x) = A(w(x)) and X1 > X is a constant to be determined later. Denote the inverse
function of j(s) by J(v). Then
dj(x) 1
= 0 ,
dx J (v)
 0
d2 j(x) J 00 (v) dj(x) J 00 (v) 1
=− 0 · =− 0 = ,
dx2 (J (v))2 dx (J (v))3 2(J 0 (v))2
and (6.18) turns out to be
 0
1
= Ψ(v).
2(J 0 (v))2
Integrating this equality yields
Z v  Z η −1/2
J(v) = 2 Ψ(ξ)dξ dη. (6.19)
0 0

From the above analysis, it is natural to define J(v) by (6.19) and then to consider its inverse
function j(x). The condition (6.16) ensures the definition of J(v) for all v ≥ 0. Since J(v) is
increasing and J(+∞) = +∞, j(x) is well-defined for all x ≥ 0. j(x) is a solution of (6.18) for
x > 0; so is j(X1 − x) for any X1 and x < X1 . Hence for any X1 ≥ X, Φ(j(X1 − x)) is a (classical)
solution for x < X1 . We choose X1 such that
u(x, t) ≤ w(X). (6.20)
This is possible; for example, we may take X1 = X + J(A(M )), where M = sup u.
Now we define
(
Φ(j(X1 − x)) for X ≤ x ≤ X1 ,
w(x, t) = w(x) =
0 for x ≥ X1 .
dA(w)
Since w and equal zero at x = X1 , it is easy to check that w is a generalized solution of
dx
(6.14) on GX . Using the comparison theorem for u and w on GX and noticing (6.20) and that
u0 (x) = 0 for x ≥ X, we arrive at (6.19) and that u(x, t) = 0 for x ≥ X, t > 0.
Similarly we can prove that u(x, t) = 0 for x ≤ −X, t > 0. 
Remark 6.2 For equations of the form
∂u ∂ 2 um
= − cun (m ≥ 1, n > 0, c > 0),
∂t ∂x2
the condition (6.16) becomes
 1/2 Z 1
m+n
v −(m+n)/(2m) dv < +∞
mc 0

which is equivalent to n < m.

94 Properties of Solutions

Theorem 6.3 Let u(x, t) be a generalized solution of the Cauchy problem (6.14), (6.15) on
Q = R × (0, ∞). If
Z 1
dy
< ∞, (6.21)
0 c(y)
then there exists T ∈ (0, ∞) such that u(x, t) ≡ 0, for x ∈ R, t ≥ T .
In this case we will say that extinction occurs for the solution u(x, t) at the time t = T .

Proof. To prove our theorem, we first choose a function w(t) such that

dw(t)
= −c(w(t)) for 0 < t < T, (6.22)
dt
w(T ) = 0 (6.23)

with T > 0 to be determined later. Integrating (6.22) and using (6.23) we obtain
Z w(t)
dy
= T − t, for 0 < t < T. (6.24)
0 c(y)
Z v
dy
Here we notice that the condition (6.21) ensures that the integral is defined for any v ≥ 0
0 c(y)
and is increasing in v.
Extend the function w(t) defined by (6.24) to (0, ∞) with w(t) = 0 for t ∈ [T, ∞). Then it is
easy to check that w(t) is a generalized solution of (6.14) on (0, ∞).
If we have
u(x, 0) = u0 (x) ≤ w(0), (6.25)
then the comparison theorem gives

u(x, t) ≤ w(t) for (x, t) ∈ Q

and hence u(x, t) = 0 for x ∈ R, t ≥ T . Since

Z w(0)
dy
= T,
0 c(y)
Z M
dy
for (6.25) to be held, it suffices to take T = where M ≥ sup u0 (x). The proof is thus
0 c(y)
completed. 

Remark 6.3 Kalashnikov proved that for the generalized solution of (6.14)to have the property
of extinction, the condition (6.21) is also necessary.

6.3 Asymptotic Behaviors

In a variety of diffusion phenomena, people are required to study equations of the form
∂u
= ∆um + λup , (6.26)
∂t
where m > 0, p > 0, λ are constants. The nonlinear term λup in the equation describes the
nonlinear source in the diffusion process; it is called ”heat source” whenever λ > 0 and ”cold
source” or absorption term whenever λ < 0. Just as shown in the diffusion process, the appearance
Nonlinear Diffusion Equations 95

of nonlinear sources will exert a great influence to the properties of solutions. For instance, when
the ”heat source” occurs, the solutions of the equation (6.26) might be blowing-up, namely, the
solutions might be unbounded at finite time. In this case, in order that the equation (6.26) has
a generalized solution (its definition is similar to the case λ = 0), the condition on the growth of
initial data
u(x, 0) = u0 (x) ≥ 0 for x ∈ RN (6.27)
should be more restrictive, compared with the case λ = 0. Contrary, if the ”cold source” occurs,
then in order that the Cauchy problem (6.26), (6.27) has a generalized solution, the condition on
the growth of u0 (x) might be less restrictive, compared with the case λ = 0; in some circumstances
the solutions even exist for any u0 ∈ L1loc (RN ).
In the study of asymptotic properties of solutions, one of the basic problems is to analyze the
relationship between the behavior of solutions for large time and the asymptotic behavior of the
initial value u0 as |x| → ∞.
Assume that
lim |x|α u0 (x) = A,
|x|→∞

where α > 0 and A > 0 are constants. Denote

2 2(p − 1)
γ = α(m − 1) + 2, µ =m−1+ , β= .
N p−m
 
2
Again assume that m > 1− .
N +
In the case λ = 0, we have the following results:
(1) If 0 ≤ α < N , then the solution u of (6.26), (6.27) satisfies

tα/γ |u(x, t) − wA (x, t)| → 0 as t → ∞

uniformly on {x ∈ RN ; |x| ≤ bt1/γ } with some constant b > 0, where wA (x, t) is the solution of
(6.26) with initial data
u(x, 0) = A|x|−α . (6.28)
By the uniqueness of solutions of (6.26),(6.28),

WA = t−α/γ f (η, A),

where η = |x|t−1/γ and f is the solution of the problem


 N −1 m 0 1 0 α
 (f m )00 + (f ) + f + f = 0, η > 0,
η η γ

 f 0 (0) = 0, f ≥ 0, lim η α f (η) = A.
η→∞

(2) If α > N , then the solution of (6.26),(6.27) satisfies

t1/µ |u(x, t − Ec (x, t))| → 0, as t → ∞

uniformly on {x ∈ RN ; |x| ≤ bt1/N µ } with some constant b > 0, where c = ku0 kL1 (RN ) , Ec is the
Barenblatt solution of (6.26),
 1/(m−1)
−1/µ (m − 1)|x|2
Ec = t c0 − (6.29)
2mN µt2/µN +
96 Properties of Solutions
Z
with a constant c0 > 0 such that Ec (x, t)dx = c.
RN
In the case λ < 0, in view of the occurrence of the absorption, the behavior of the solution of
(6.26), (6.27) as t → ∞ depends not only on the asymptotic behavior of u0 as |x| → ∞, but also
on the ”competition” of diffusion and absorption. We have the following results:
2
(1)If p > max{m, 1}, 0 < α < , then the solution u of (6.26), (6.27) satisfies
p−m
 1/(p−1)
1
t1/(p−1) u(x, t) → as t → ∞
p−1

uniformly on {x ∈ RN ; |x| < bt1/β } with some constant b > 0.

2 2
(2) If p > m + , < α < N , then the solution u of (6.26),(6.27) satisfies
N (p − m)

tα/γ |u(x, t) − wA (x, t)| → 0 as t → ∞

uniformly on {x ∈ RN ; |x| ≤ bt1/γ } with some constant b > 0, where wA is the solution of (6.26),
(6.28) with λ = 0.
2
(3) If p > m + , α > N , then the solution u of (6.26), (6.27) satisfies
N
t1/ν |u(x, t) − Ec (x, t)| → 0 as t → ∞
Z
uniformly on {x ∈ RN ; |x| ≤ bt1/N ν } with some constant b > 0, where c = u0 (x, t)dx +
Z Z RN

λ up (x, t)dxdt, Ec is given by (6.29) with c0 such that c0 Ec (x, t)dx = C.

RN RN
2 2
(4) If max{m, 1} < p < m + , α > , then the solution u of (6.26), (6.27) satisfies
N p−m

t1/(p−m) |u(x, t) − U (x, t)| → 0 as t → ∞

uniformly on {x ∈ RN ; |x| ≤ bt1/β } with some constant b > 0, where U is a very singular solution
of (6.26).
Another basic problem in the study of asymptotic properties is to discuss whether the solutions
of the evolution equation considered ”tend” and in what sense ”tend” to the solutions of the
corresponding stationary equation.
As an example, we introduce a result. Consider the boundary value problem


 ∂u ∂ 2 um

 = + ψ(u), (m > 1)
 ∂t ∂x2
u|x=0,1 = 0, (6.30)





u|t=0 = u0 (x, t),

where 0 ≤ u0 ≤ 1. Let u = u(t; u0 ) be its solution. Define the distance

∂um ∂v m
d(u, v) = ku − vkL1 (0,1) + −
∂x ∂x L2 (0,1)

in the space
∂um
X = {u ∈ L∞ (0, 1); 0 ≤ u ≤ 1, ∈ L2 (0, 1)}.
∂x
Nonlinear Diffusion Equations 97

Let ω(u0 ) be the ω-limit set of u(t, u0 ):

ω(u0 ) = {ω ∈ X; there exists tn → ∞ such that u(tn , u0 ) → ω in X}.

Denote
γτ (u0 ) = {u(t, u0 ); t ≥ τ } (τ ≥ 0).
It is proved that for each τ > 0, γτ (u0 ) is precompact in X and ω(u0 ) ⊂ E, where E is the set of
solutions of the stationary problem
 2 m
 d u + ψ(u) = 0,
dx2

u|x=0,1 = 0.

The proof is based on a usage of the Liapunov functional

Z 1 Z 1
1 2
V (w) = (wm )0 dx − F ∗ (w)dx,
2 0 0
Z
1 w m−1
where F ∗ (w) = ρ ψ(ρ)dρ. Under some conditions on u0 , the ω-limit set of u(t, u0 )
m 0
contains only one point.
98 Appendix Classes BV and BVx

7 Appendix Classes BV and BVx

In this section we list a series of properties of BV functions and BVx functions without proofs,
which are needed in the previous sections.
Let Ω be a domain of RN , QT = Ω × (0, T ). Denote by BV (QT ) the set of all functions of
locally bounded variation, namely, a subset of L1loc (QT ), in which the weak derivatives of each
function are Radon measures on QT . A little general class, denoted by BVx (QT ), is another subset
of L1loc (QT ), in which only the derivatives in x of each function are Radon measures on QT . Clearly

BV (QT ) ⊂ BVx (QT ).

The following lemma is a basic result in measure theory.

Lemma 7.1 Let µ be a Radon measure in the measure space X. For any Borel set G, define
Z
ν(G) = f dµ,
G

where f ∈ L1 (X, |µ|). Then ν is also a finite Radon measure on X and for any function g which
is essentially bounded and measurable with respect to ν, there holds
Z Z
gdν = gf dµ.
X X

Lemma 7.2 Assume that u ∈ BVx (QT ). Then for almost all t ∈ (0, T ), u(·, t) ∈ BV (Ω). If
we define
∂u(·, t)
µi (t) ≡
∂xi
and use |µi (t)|(G) to denote the total variation of the measure |µi (t)| on a rectangle G in Ω, then
Z
∂u(·, t)
|µi (t)|(G) ≡ , (i = 1, · · · , n)
G ∂xi

is a locally integrable function on (0, T ).

Now we introduce some related notations. For u ∈ BV (QT ), denote by Γu the set of all point
of discontinuity of u, γ = (γ1 , · · · , γn , γt ) the unit normal vector to Γu , u+ (x0 , t0 ) and u− (x0 , t0 )
the approximate limits of u at (x0 , t0 ) ∈ Γu taking from the half space

(x1 − x01 )γ1 + · · · + (xn − x0n )γn + (t − t0 )γt > 0

and
(x1 − x01 )γ1 + · · · + (xn − x0n )γn + (t − t0 )γt < 0
1 +
respectively, u(x, t) ≡ (u (x, t) + u− (x, t)) the symmetric mean value of u at the regular point
2
(x,t). By a regular point, we mean either a point of approximate continuity or a point of disconti-
nuity.
For BV functions we have the following formula for derivatives of composites: If f ∈ C 1 (R)
and u ∈ L∞ (QT ) ∩ BV (QT ), then f (u) ∈ BV (QT ) and

∂f (u(x, t)) ∂u(x, t)

= fˆ0 (u(x, t))
∂ξ ∂ξ
Nonlinear Diffusion Equations 99

where ξ denotes t or xi (i = 1, · · · , N ), ĝ(u(x, t)) denote the functional superposition of g(u) and
u(x, t) which is defined by
Z 1
ĝ(u(x, t)) = g(τ u+ (x, t) + (1 − τ )u− (x, t))dτ.
0

Also we have the product rule for BV functions: If u, v ∈ L∞ (QT ) ∩ BV (QT ), then u, v ∈
BV (QT ) and
∂uv ∂v ∂u
=u +v
∂ξ ∂ξ ∂ξ
where ξ denotes t or xi (i = 1, · · · , N ).
The Hausdorff measure of a subset S of Γu is denoted by H(S).
For fixed t, we use Γtu , H t , (γ1t , · · · , γnt ) and ut± to denote the set of points of discontinuity of
u(·, t), the Hausdorff measure on Γtu , the unit normal vector to Γtu and the approximate limits of
u(·, t).
For simplicity, all results are stated merely in one space variable in the sequel and Ω is assumed
to be an open interval which may be infinite. For fixed t ∈ (0, T ), we use ũ(·, t), ur (·, t) and ul (·, t)
to denote the symmetric mean value, the right approximate limit and left approximate limit of
u(·, t). From Lemma 7.2, we see that if u ∈ BVx (QT ), then for almost all t ∈ (0, T ) and any x ∈ I,
ũ(x, t), ur (x, t) and ul (x, t) exist. The following result explains the relation between u(x, t) and
ũ(x, t) and the relation among u+ (x, t), ur (x, t), u− (x, t), and ul (x, t).
Lemma 7.3 Assume u ∈ BV (QT ). Then there exists a subset Et ⊂ (0, T ) with mesEt = 0,
such that for any (x, t) ∈ I × [(0, T ) \ Et ],
u(x, t) = u
e(x, t), u+ (x, t) − u− (x, t) = ur (x, t) − ul (x, t) .
Lemma 7.4 Assume that u ∈ BV (QT ) and S is a bounded and H-measurable subset of Γu
∂u
with S ⊂ Γu . Then S is measurable with respect to the measure and
∂x
ZZ Z
∂u
= (u+ − u− )γx dH.
S ∂x S

Corollary 7.1 Assume that u ∈ BV (QT ) and

S ⊂ Γ∗u ≡ {(x, t) ∈ Γu ; γx 6= 0} .
Then
∂u
(S) = 0 iff H(S) = 0.
∂x
∂u
Corollary 7.2 Assume that u ∈ BV (QT ). If f (x, t) is measurable with respect to on QT .
∂x
Then f (x, t) is H-measurable on Γ∗u .
Lemma 7.5 Assume that u, v ∈ BV (QT ) and |v(x, t)| ≤ K|u(x, t)| for almost all (x, t) ∈ QT ,
where K is a constant. Then for any ϕ ∈ C0∞ (QT ),
ZZ Z
∂
σ (ϕv) = − ϕ(sgnu+ − sgnu− )vγt dH,
∂t
Z ZQ T ZΓu
∂
σ (ϕv) = − ϕ(sgnu+ − sgnu− )vγx dH,
QT ∂x Γu

where
1
σ= (sgnu+ + sgnu− ).
2
100 Appendix Classes BV and BVx

Lemma 7.6 Assume that u ∈ BVx (QT ), G is a bounded rectangle with G ⊂ QT and f (x, t) is
∂u
bounded and measurable with respect to | | on G. Then there exist a sequence {fn } ⊂ C0∞ (G)
∂x
and a constant M , such that

|fn (x, t)| ≤ M, lim kfn − f kX = 0,

n→∞
 
1 ∂u
where X = L G, .
∂x
Lemma 7.7 Assume that u ∈ BVx (QT ), Et ⊂ (0, T ) is a set of Lebesgue measure zero and
∂u
G = I × Et . Then (G) = 0. If u ∈ BV (QT ), then H(G ∩ Γ∗u ) = 0.
∂x
Lemma 7.8 Assume that u ∈ BVx (QT ) and f (x, t) is measurable with respect to the measure
∂u
, with compact support. Then for almost all t ∈ (0, T ), f (x, t) is measurable with respect to the
∂x
∂u(·, t)
measure . Moreover,
∂x
Z Z
∂u(·, t) ∂u(·, t)
f (x, t) and f (x, t)
I ∂x I ∂x
are Lebesgue integrable on (0, T ) and there hold
ZZ Z T Z
∂u ∂u(·, t)
f (x, t) = dt f (x, t) ,
Q ∂x 0 I ∂x
ZZ T Z T Z (7.1)
∂u ∂u(·, t)
f (x, t) = dt f (x, t) .
QT ∂x 0 I ∂x

As formulas transforming the double integral into an iterated integral, (7.1) are very useful
in the study of discontinuity conditions. For convenience of applications we need to discuss the
integrability of the integrand f (x, t). First let us recall the definition of the Caratheodory functions.
f (x, t) is called a Caratheodory function, if
(1) for almost all t ∈ (0, T ), f (x, t) is continuous in x,
(2) for any x ∈ I, f (x, t) is Lebesgue measurable in t.
Now we introduce a new class of functions, denote by C a (QT ): f ∈ C a (QT ) if and only if there
exist a sequence {fn (x, t)} of Caratheodory functions and a constant M > 0, such that

|fn (x, t)| ≤ M, lim fn (x, t) = f (x, t),

n→∞

for almost all t ∈ (0, T ) and any x ∈ I.

Remark 7.1 If u ∈ L∞ (QT ) ∩ BVx (QT ), then

e(x, t), ur (x, t), ul (x, t) ∈ C a (QT ).

u

u, sgnur and sgnul are measurable with respect to the

If, in addition, w ∈ BVx (QT ), then sgne
∂w
measure .
∂x
Lemma 7.9 Assume that u ∈ BVx (QT ) and f ∈ C a (QT ) with compact support. Then f (x, t)
∂u
is integrable on QT with respect to the measure .
∂x
Nonlinear Diffusion Equations 101

In applications, the following formula which transforms the double integral into a curve integral,
is also very useful.

Lemma 7.10 Assume that u ∈ L∞ (QT ) ∩ BV (QT ) and S is a subset of Γ∗u , bounded and
∂u
H-measurable. Then for any bounded and measurable (with respect to ) function f (x, t) on QT
∂x
with compact support, there holds
ZZ Z T X
∂u 
f (x, t) = dt ur (x, t) − ul (x, t) f (x, t)
∂x
ZS 0 x∈S t
 r 
= f (x, t) u (x, t) − ul (x, t) |γx |dH.
S

Corollary 7.3 Under the assumptions of Lemma 7.10, there holds

Z T X Z
dt f (x, t) = f (x, t)|γx |dH.
0 x∈S t S

Corollary 7.4 If u ∈ L∞ (QT ) ∩ BV (QT ), then

 
u+ (x, t) − u− (x, t) = ur (x, t) − ul (x, t) sgnγx

holds H-almost everywhere on Γ∗u

Corollary 7.5 If u ∈ L∞ (QT ) ∩ BV (QT ), then

(
u+ (x, t) = ur (x, t)sgn+ γx − ul (x, t)sgn− γx ,

u− (x, t) = ul (x, t)sgn+ γx − ur (x, t)sgn− γx

holds H-almost everywhere on Γ∗u , where

( (
1, for s > 0, 0, for s ≥ 0,
+ −
sgn s = sgn s =
0, for s ≤ 0, −1, for s < 0.