manuscripta math.

147, 101–138 (2015) © Springer-Verlag Berlin Heidelberg 2014

Naian Liao

Existence and nonexistence of solutions to a logarithmic

diffusion equation in bounded domains
Received: 10 July 2014 / Accepted: 23 October 2014
Published online: 6 November 2014

Abstract. We construct solutions to the Dirichlet problem of a logarithmic diffusion equa-

tion with boundary value that could vanish somewhere. We also provide nonexistence results
that show our existence theorems are, in some sense, optimal. Based on these results, we
are able to construct a number of examples that complement our previous study of local
behaviors of solutions to such an equation. Some preliminary results on the geometry of the
vanishing set of local solutions are also reported.

1. Introduction and main results

Consider the following Dirichlet problem for the logarithmic diffusion equation
u t −  ln u = 0 in E T ;
ln u = ln g on ∂ E × (0, T ); (1.1)
u(·, 0) = u o .
Here E is a smooth domain in R N with N ≥ 2, and E T = E × (0, T ) with T > 0.
Let ∂ p E T = [E × {0}] ∪ [∂ E × [0, T )] be the parabolic boundary of E T and
let ST = ∂ E × (0, T ) be its lateral boundary. We will use Bρ (x) (or K ρ (x)) to
denote a ball (or a cube) centered at x with radius (or edge) ρ. We are interested in
solving (1.1) when g is permitted to vanish on a subset of ST . Assume momentarily
g ≥ 0 and u o ≥ 0 are bounded and measurable so that ln g is well-defined as a
measurable function on ST .
The existence or nonexistence of solutions to (1.1) hinges on the notion of
solutions. A bounded measurable function u is called a weak sub(super)-solution
to (1.1) if ln u ∈ L 2 (0, T ; W 1,2 (E)), and for almost all 0 < t < T
  t
 
uη(x, t) d x + − uητ + D ln u Dη d xdτ
E
 0 E

≤ (≥) u o η(x, 0) d x
E

N. Liao (B): Department of Mathematics, Vanderbilt University, Nashville TN 37240,

USA. e-mail: liaonaian@gmail.com
Present Address:
N. Liao: College of Mathematics and Statistics, Chongqing University, Chongqing, 401331,
China. e-mail: liaon@cqu.edu.cn
Mathematics Subject Classification: Primary 35K20, 35K65; Secondary 35B45

DOI: 10.1007/s00229-014-0717-3
102 N. Liao

for all nonnegative testing functions

η ∈ W 1,1 (0, T ; L 1 (E)) ∩ L 2 (0, T ; Wo1,2 (E)).
In addition, ln u(·, t) ≤ (≥) ln g(·, t) in the sense of traces on ∂ E for a.e. 0 < t < T .
Since we are only interested in the existence and nonexistence of solutions when g
vanishes somewhere, we always assume the boundedness of g.
One way of constructing a solution to (1.1) is by a semigroup technique of
time-discretization. Let λ > 0 and assume momentarily f ≥ 0 and  ≥ 0 be
measurable so that ln  is well-defined as a measurable function on ∂ E. This leads
to studying the Dirichlet problem
λu −  ln u = f in E;
(1.2)
ln u = ln  on ∂ E.
A measurable function u is called a sub(super)-solution to (1.2) if ln u ∈
W 1,2 (E) and
  
λ uζ d x + D ln u Dζ d x ≤ (≥) f ζ dx
E E E
for all nonnegative
ζ ∈ Wo1,2 (E).
In addition, ln u ≤ (≥) ln  in the sense of traces on ∂ E. We state our main result
here concerning existence of solutions to (1.1).
Theorem 1.1. Let 0 ≤ u o ∈ L ∞ (E) and 0 ≤ g ∈ L ∞ (∂ E × (0, T )) and ln g
admits an extension to E T which we still denote as ln g such that
ln g ∈ W 1,1 (0, T ; L 1 (E)) ∩ L 2 (0, T ; W 1,2 (E)).
Then there is a unique bounded solution to (1.1). If, in addition,
g(·, t)
is decreasing on ∂ E × (0, T ), (1.3)
t
then the solution obtained is smooth and positive in E T and
u
ut ≤ in E T .
t
The proof will be given in Sect. 5. We will also give examples of locally continuous
solutions that do not satisfy the semiconvexity in time. Note the assumption that
ln g admits an extension to E T with some differentiability in time is natural when
seeking for a solution with D ln u ∈ L 2 (E T ). However, we can dispense with such
a requirement if we consider a weaker notion of solutions.
A bounded function u is called a very weak solution to (1.1) if ln u ∈ L 1 (E T )
and
  
∂ϕ
uϕt + ln uϕ d xdt = ln g dσ dt − u o ϕ(·, 0) d x (1.4)
ET ∂E ∂ν E

for all ϕ ∈ C ∞ ( Ē × [0, T ]) and ζ = 0 on [∂ E × (0, T )] ∪ [E × {T }]. Then we

have the following
Logarithmic diffusion equation in bounded domains 103

Theorem 1.2. Let 0 ≤ u o ∈ L ∞ (E) and 0 ≤ g ∈ L ∞ (∂ E × (0, T )) and ln g ∈

L 1 (∂ E × (0, T )). Then there exists a very weak solution to (1.1).

Note we do not know if such a solution is unique.

Since we will use a technique of time discretization to construct solutions, we
develop a relevant statement concerning the existence of solutions to (1.2). Here
and in the sequel, we require τ > max1, N /2 whenever appearing in L τ .

Lemma 1.1. Let 0 ≤ f ∈ L r (E) and 0 ≤  ∈ L r (∂ E) such that ln  ∈

1 β
W 2 ,2 (∂ E). Then there exists a unique solution u to (1.2). Moreover, u ∈ Cloc (E)
for some 0 < β < 1 depending on

{N , λ, diam(E),  f r , r } .

1
The condition ln  ∈ W 2 ,2 (∂ E) yields an extension of ln  to E and it is a
natural condition to impose when seeking for a solution with D ln u ∈ L 2 (E). We
can consider a weaker notion of solutions. A function u ∈ L 1 (E) is called a very
weak solution to (1.2) if ln u ∈ L 1 (E) and
   
∂ζ
λ uζ d x − ln uζ d x = f ζ dx − ln  dσ (1.5)
E E E ∂E ∂ν

for all ζ ∈ C ∞ ( Ē) and ζ = 0 on ∂ E. Then we have

Lemma 1.2. Let 0 ≤ f ∈ L r (E) and 0 ≤  ∈ L r (∂ E) such that ln  ∈ L 2 (∂ E).

Then there exists a very weak solution u ∈ L r (E) to (1.2).

Note that we do not know whether such a solution is unique.

1
The requirement that ln g ∈ W 2 ,2 (∂ E) allows g to vanish somewhere on the
boundary. In fact, for a fixed t, the set [g = 0] could be of positive H N −2 -measure.
See Sect. 3 for examples. A larger size of the set of zeros on the boundary is in fact
forbidden in the sense of the following theorem.
We call u a local weak solution of the logarithmic diffusion equation if
   1,2 
u ∈ Cloc 0, T ; L 2loc (E) , ln u ∈ L 2loc 0, T ; Wloc (E) ;
(1.6)
u t −  ln u = 0 weakly in E T .

Theorem 1.3. If g vanishes on an open subset of ST , then there is no solution

to (1.1) in the sense that (1.6) is satisfied in the interior and the zero boundary
datum is taken in the sense of traces.

A similar statement for the elliptic case is considered in Sect. 4. Theorem 1.3
indicates that a patch of zeros with positive H N -measure on the boundary pro-
pogates into the interior even if we only use u to take the boundary trace and the
equation is satisfied in the interior. In particular, a solution cannot be sought in
L 2 (0, T ; Wo1,2 (E)).
104 N. Liao

1.1. Novelty and significance

The first equation of (1.1) arises from different physical models and a particularly
interesting one is thin film dynamics. Suppose a viscous liquid film lies on a rigid
plate and the thickness of the film is between 100Å and 1,000Å. Given an initial
disturbance to the film and the van der Waals force eventually leads to the rupture
of the film within a finite time. It is derived in [16] that

u t −  ln u + div(u 3 ∇(u)) = 0 in R2 × R+ .

The fourth order term reflects the stabilizing effects of surface tension on the liquid-
gas interface. Numerical studies in [16] suggest it is negligible.
This logarithmic diffusion equation also models the evolution of Ricci flow for
complete R2 . See [4].
The investigation of local behaviors of local solutions to (1.6) is initiated in [5].
A Harnack-type inequality is established assuming a proper higher integrability of
ln u. The modulus of ellipticity 1/u tends to infinity when u vanishes and this leads
to the major difficulty.
This difficulty also appears in the Dirichlet problem (1.1) when g is allowed to
vanish. When assigning positive boundary data, such a difficulty disappears because
of the maximum principle and it is considered in [10]. Nevertheless, it is far from
clear whether one could still obtain a solution if the datum g vanishes on part of
the boundary. The main difficulty lies in obtaining an apriori L 1 estimate of ln u in
terms of the given data.
Theorem 1.1 is the first contribution in this direction. We give conditions on g
that allow g to vanish on part of ST to generate a solution. The main tool we use
is Poisson’s representation formula. As a matter of fact, most novel discoveries of
this note hinge upon this tool.
Theorem 1.3 is also new. An essentially global version of nonexistence results
concerning continuous solutions taking zero boundary datum is claimed in [14]. It
should be pointed out that our approach is entirely local and also independent of
the interior continuity of solutions.
Moreover, we will examine the interior behavior of the local solutions to (1.6)
by examples in Sect. 4, and this complements our previous studies on this issue.
Indeed, we have established a number of local results in [5–7]. These, more or
less, parallel to the results for the porous medium equation bearing in mind that the
logarithmic diffusion equation is a formal limit of the porous medium equation as
m → 0+ . However, this formal limit does not grant to solutions to (1.6) any local
property similar to that of solutions to the porous medium equation. In particular,
an estimate of the modulus of continuity of u over a compact subset K of E T
depending only on the bound of u and the distance from K to the boundary, in
general, does not hold. See the discussion after Proposition 5.1. On the other hand,
it is still unclear whether an explicit discontinuous solution can be constructed.
Although this is claimed in [15], the notion of solution and the topology by which
the constructive approximation takes place are unclear to me.
The set of interior zeros is where the singular behaviors of the solutions to (1.6)
occur. A full characterization of such a set is not known. However, we will show
Logarithmic diffusion equation in bounded domains 105

by examples that it can be a hyperplane normal to the t-axis. On the other hand, we
show it cannot occupy a line parallel to the t-axis.

2. Uniqueness and comparison principles

In this section we list some basic facts about uniqueness and comparison principles
these equations satisfy under different assumptions.
Proposition 2.1. Let u and v be two bounded weak solutions to (1.1) with the same
boundary and initial datum. Then u = v a.e. in E × (0, T ).
Proof. Take the difference of the weak formulations for u and v; we obtain
 t
 
− (u − v)ητ (x, τ ) + D(ln u − ln v)Dη d xdτ = 0
0 E
for all
η ∈ W 1,1 (0, T ; L 1 (E)) ∩ L 2 (0, T ; Wo1,2 (E)).
Now take
⎧ t
⎨
(ln u(x, s) − ln v(x, s)) ds, 0 ≤ τ < t;
η(x, τ ) = τ
⎩
0, τ ≥ t.
It is straightforward to verify this function is an admissible test function. Then
 t   t 2
1
(u − v)(ln u − ln v) d xdτ + D(ln u − ln v) ds d x = 0
0 E 2 E 0
Thus
 t
(u − v)(ln u − ln v) d xdτ = 0
0 E
and u = v a.e. in E T .
Proposition 2.2. Let u be a super-solution and v be sub-solution to (1.1). If u ≥ v
on ∂ p E T and vt , u t ∈ L 1 (E T ), then u ≥ v a.e. in E T .
Proof. Let Sn (·) be an approximation to the Heaviside function. Namely, Sn (c)
equals 0 when c ≤ 0 and 1 when c ≥ n1 and it is linear when 0 ≤ c ≤ n1 . Take
Sn ((ln v − ln u)+ ) as a test function then

(v − u)t Sn ((ln v − ln u)+ ) d xdt
Et

=− Sn ((ln v − ln u)+ )|D(ln v − ln u)+ |2 d xdt ≤ 0.
Et
Then letting n → ∞ we have for any 0 < t < T

(v − u)+ (·, t) d x ≤ 0.
E
106 N. Liao

Similarly, we have an analogous result for the elliptic case.

Proposition 2.3. Let u be a super-solution and v be sub-solution to (1.2). If u ≥ v
on ∂ E, then u ≥ v a.e. in E.

3. Existence of solutions to (1.2)

When  is bounded and positive, the obstacle to existence generated by the logarith-
mic term disappears. We give a proof of this case for the sake of completeness. For
our convenience of applying the Fixed Point Theorem A.1, we use a transformation
v = ln u to recast it into the following problem.
λev − v = f in E;
v = ln  in ∂ E.
Then we have

Lemma 3.1. Let f ∈ L r (E) be nonnegative, and  ∈ C(∂ E) such that ln  ∈

1
W 2 ,2 (∂ E) and μ1 ≤  ≤ μ2 on ∂ E for some μ1 , μ2 > 0. Then there is a unique
solution u to (1.2) and u ∈ C( Ē).

The proof is given in the Appendix A.

3.1. When  contains zeros

Let F(x, y) be the fundamental solution of the Laplacian in R N with pole at y.

Green’s function in E is defined as

G(x, y) = F(x, y) − (x, y), N ≥2

where (x, ·) ∈ C ∞ ( Ē) is the unique solution to

 y (x, y) = 0 in E;
(x, y) = F(x, y) for y ∈ ∂ E.
Poisson’s kernel on ∂ E is defined as
∂
P(x, y) = − G(x, y) for x ∈ E, y ∈ ∂ E.
∂n(y)
We will use the well-known Poisson representation formula
 
w(x) = w(y)P(x, y) dσ − w(y)G(x, y) dy when x ∈ E
∂E E

for all w ∈ C 2 ( Ē).

To proceed, we need to discuss some integrability properties of G(x, y) and
P(x, y). First of all, we note

0 ≤ G(x, y) ≤ F(x, y) for all x, y ∈ E.

Logarithmic diffusion equation in bounded domains 107

Moreover, there is a constant C(r, N , diam(E)) such that

N
r ≤ C for all r >
G(x, ·) r −1 .
2
Next, Poisson’s kernel satisfies the following asymptotic behavior
diam(x, ∂ E)
P(x, y) ≈ for all x ∈ E, y ∈ ∂ E.
|x − y| N
See [11] for an elementary proof of this fact. Therefore, there is a constant C(N ,
diam(E)) such that for all y ∈ ∂ E

P(x, y) d x ≤ C.
E

Now we are ready to present

Proof of Lemma 1.1. Let k > 0 and

⎧ −1
⎨k ,  < k −1 ;
k = , k −1 ≤  ≤ k;
⎩
k,  > k.
1 1
Since ln  ∈ W 2 ,2 (∂ E), we also have ln k ∈ W 2 ,2 (∂ E). See [2]. As a result, it
admits an extension to a function in W 1,2 (E) which we still denote as ln k .
Moreover we can find a sequence of functions (ln k ) ∈ C ∞ ( Ē) such that as
→0

D(ln k ) → D ln k in L 2 (E);
(ln k ) → ln k in L q (E) ∀1 ≤ q < ∞;
(ln k ) → ln k in L r (∂ E);
− ln(2k) ≤ (ln k ) ≤ ln(2k) on ∂ E;
e(ln k ) → k in L r (∂ E).

The first two convergences follow from the proof of the well-known approximation
theorem for Sobolev functions when the boundary satisfies the segment property.
The third one follows from the previous two and the trace inequality
1 1− r1
wr,∂ E ≤ γ (N )(Dw2 + w2 ) r wq (3.1)

for all w ∈ W 1,2 (E) ∩ L q (E) where q = 2(r − 1). The forth one comes from the
previous one and the upper bound of k . The last one follows from the third one
and the forth one in view of
 
|e(ln k ) − k |r dσ ≤ (2k)r |(ln k ) − ln k |r dσ.
∂E ∂E

See all basic theories of Sobolev functions in [2].

108 N. Liao

Assume first that f ∈ C ∞ ( Ē). According to the previous lemma, there is a

unique solution v k, ∈ C ∞ ( Ē) to
λev − v = f in E;
(3.2)
v = (ln k ) on ∂ E.
We are going to use repeatedly Poisson’s representation formula for any x ∈ E
 
G(x, y)( f (y) − λev (y) ) dy. (3.3)
k,
v k, (x) = (ln k ) (y)P(x, y) dσ +
∂E E
First of all, note that P(·, y) dσ (y) is a probability measure on ∂ E; then, by
Jensen’s inequality
 
v k, (x) ≤ (ln k ) (y)P(x, y) dσ + G(x, y) f (y) dy
∂E E

≤ ln e(ln k ) (y) P(x, y) dσ + G(x, ·) r −1r  f r
∂E

≤ ln e(ln k ) (y) P(x, y) dσ + γ (N , r, diam(E)) f r .
∂E

This gives
   r
r v k, (x) r  f r
e d x ≤ γ (N , r, diam(E))e e(ln k ) (y) P(x, y) dσ d x.
E E ∂E

The integral on the right hand side is estimated by Hölder’s inequality

  r
e(ln k ) (y) P(x, y) dσ d x
E ∂E
  r −1 
≤ P(x, y) dσ P(x, y)er (ln k ) (y) dσ d x
∂E ∂E
E 
r (ln k ) (y)
= P(x, y) d x e dσ
∂E
E

≤ γ (N , diam(E))  r dσ.
∂E
Combining all these estimates we arrive at
 
r v k, (x) r  f r
e d x ≤ γ (N , r, diam(E))e  r (y) dσ (3.4)
E ∂E
Secondly, by taking power p = 2 at both sides of (3.3) and integrating in d x
over E we obtain
   2
|v k, (x)|2 d x ≤ γ (ln k ) (y)P(x, y) dσ d x
E E ∂E
  2
G(x, y)( f (y) − λev
k, (y)
+γ ) dy dx
E E
= I1 + I2 .
Logarithmic diffusion equation in bounded domains 109

By using the estimate of ev r , I2 is easily seen to be bounded by

k,

2
 f r + λev r
k,
I2 ≤ γ |E|G(x, ·)2 r
r −1
 
≤ γ (N , r, diam(E)) λ2 e2 f r r,∂
2
E +  f r .
2

On the other hand, I1 is estimated by Hölder’s inequality as

  
I1 ≤ γ P(x, y) dσ |(ln k ) (y)|2 P(x, y) dσ d x
E ∂E ∂E
 
=γ |(ln k ) (y)| dσ
2
P(x, y) d x
∂E
 E

≤ γ (N , diam(E)) |(ln k ) (y)|2 dσ.

∂E

Thus there is a constant γ depending on {N , r, diam(E)} such that


 
|v k, (x)|2 d x ≤ γ λ2 e2 f r r,∂
2
E +  f r +  ln 2,∂ E .
2 2
E

Now we show the L 2 norm of Dv k, is also bounded. Indeed, if we take (v k, −
(ln k ) ) as a test function in (3.2), a standard calculation yields
   
1 1
f v k, d x − λ ev v k, d x
k,
|Dv k, |2 d x ≤ |D(ln k ) |2 d x +
2 E 2 E
  E E

f (ln k ) d x + λ ev (ln k ) d x
k,
−
E E
≤ C( ln 1,2 , r,∂ E , λ,  f r , diam(E), N , r ).

Here, we have estimated the second integral on the right by


f v k, d x ≤  f r v k,  r −1
r
E
≤ C(N , r, |E|) f r v k,  2N
N −2
 
≤ C(N , r, |E|) f r v k, 2 + Dv k, 2 ,

and the term with the gradient is easily absorbed to the left-hand side by the Cauchy-
Schwarz inequality.
Thus, by the Compact Imbedding Theorem we can conclude that there is some
v ∈ W 1,2 (E) and a subsequence of ev , which we still use the same symbol to
k,

denote, such that as  → 0 and k → ∞

v k, → v a.e. in E;
v k, → v in L 2 (E);
Dv k, → Dv weakly in L 2 (E).
110 N. Liao

Note the first convergence also implies that for any ev → ev a.e in E. The uniform
k,

boundedness of ev r implies that there is some u ∈ L r (E) and a subsequence

k,

of ev , which we still use the same symbol to denote, such that

k,

ev
k,
→u weakly in L r (E).

Then u = ev a.e. in E. As a result, we are able to conclude that

v ∈ W 1,2 (E) and ev ∈ L r (E),

and for all ζ ∈ Wo1,2 (E)

  
v
λ e ζ dx + Dv Dζ d x = f ζ d x.
E E E

Finally the boundary datum ln  is taken by v ∈ W 1,2 (E). In fact, we see from
the trace inequality (3.1) that, when k → ∞ and  → 0,

(ln k ) − v2,∂ E ≤ Cv k, − v2 → 0

where C depends on the uniform bound of the v k, 1,2 . On the other hand,

(ln k ) → ln k in L r (∂ E) as  → 0;
ln k → ln  in L r (∂ E) as k → ∞.

Thus v = ln u = ln  on ∂ E.
For the local Hölder continuity of v in E we only need to observe that the first
equation in (3.2) can be written as
def
−v = F = f − λev with F ∈ L r (E).
β
Thus the classical theory of elliptic equations gives v ∈ Cloc (E) for some β de-
pending on

{N , λ, diam(E), ev r ,  f r }.

In view of the bound on ev r , β depends on

{N , λ, diam(E), r,∂ E ,  f r }.

In view of all estimates depending only on  f r , the assumption f ∈ C ∞ ( Ē)

can be removed by a proper approximation.

Global Boundedness of u
It should be remarked that the mere requirement  ∈ L r (∂ E) is not enough to
insure global boundedness of u in E. However, if f ∈ L ∞ (E) and 0 ≤  ≤ M
for some M then v k, is uniformly bounded above. Indeed, let us take

 f ∞
l ≥ max ln M, ln .
λ
Logarithmic diffusion equation in bounded domains 111

Multiply the equation by (v k, − l)+ ∈ Wo1,2 (E) and integrate in d x over E; we
obtain
    
2
λev − f v k, − l d x +
k,
D(v k, − l)+ d x = 0.
E + E
This implies by our assumptions on l that
       
v k,
ev − el v k, − l
k,
0≥ λe − f v −l
k,
dx ≥ λ d x ≥ 0.
E + E +

Hence 
def  f ∞
v k, ≤  = max ln M, ln . (3.5)
λ
Similarly, if f is strictly positive and  ≥ δ for some δ > 0 then

inf f
v ≥ min ln δ, ln
k,
. (3.6)
λ
Interior Positivity of u in Terms of  and f
The representation (3.3) implies that when K is a compact subset of E and
x ∈ K there exists a constant C1 (N , dist(K , ∂ E)) such that
   k, 
v k, (x) ≥ −C1 |ln | dσ − G(x, y) λev (x) − f (x) d x. (3.7)
∂E E
Then, since f ≥ 0, (3.4) implies that there is C2 (N , diam(E), r ) such that
 
 v k, (x)
v (x) ≥ −C1
k,
|ln | dσ − λG(x, ·) r −1
r e r
∂ E
≥ −C1 |ln | dσ − C2 λe f r r,∂ E .
∂E
We see that the positivity of solutions constructed in this way hinges upon  ln 1,∂ E .
In fact, we can formulate the following
Proposition 3.1. Let u  ∈ C 2 ( Ē) be a solution to (1.2) corresponding to f  ∈
L r (E) uniformly and M ≥  ≥ 0. If

lim |ln  (y)| dσ = ∞,
→0 ∂ E

then u  → 0 uniformly in any compact subset K of E.

Proof. By Poisson’s representation formula, for any x ∈ K
 

 
ln u (x) = ln  (y)P(x, y) dσ + G(x, y) f  − λu 
∂ E E

≤ ln  (y)P(x, y)χ[ <1] dσ + ln  (y)P(x, y)χ[ ≥1] dσ
∂E ∂E
r  f  r
+ G(x, ·) r −1

≤ C(N , dist(K , ∂ E)) ln  (y) dσ + ln M + G(x, ·) r −1
r  f  r
∂E
The right-hand side tends to −∞ as  → 0.
112 N. Liao

Zeros are Allowed on the Boundary

1
The requirement ln  ∈ W 2 ,2 (∂ E) allows  to be not essentially bounded
away from zeros on the boundary. For example, let ∂ E have a local representation
x N = φ(x̄) ∈ C 1 (|x̄| < 4R); x̄ = (x1 , . . . , x N −1 ).
Then the functions
(x) = |ln |x||−1 , N ≥ 2;
(x) = |x| p , p > 0, N ≥ 3;
− 1 N −3
(x) = e |x|2α , 0 < α < , N ≥4
4
1
will satisfy ln  ∈ W 2 ,2 (∂ E) and (x) → 0 as x → 0.
Remark 3.1. Let E = (a, b), if we assign (a) = 0 and (b) = 1 then there
is no bounded solution u such that ln u ∈ W 1,2 (a, b). Indeed, this implies also
u ∈ W 1,2 (a, b). As a result u and ln u are absolutely continuous in (a, b) and
lim u(x) = 0;
x→a +
lim ln u(x) is finite.
x→a +

A contradiction. Note this is completely independent of the equation. An analogous

argument for the case of multiple dimensions follows similarly. That is, if u, ln u ∈
W 1,2 (E) then u cannot vanish on a set of positive H N −1 -measure of ∂ E. Indeed,
one only has to realize that when we fix a point a ∈ ∂ E and the outer normal vector
ν at a, the function f (t) = u(a + t · ν) is absolutely continuous in [−δ, 0] for a
small number δ. So is g(t) = ln u(a + t · ν). See [17], Theorem 2.1.4.
However, Lemma 1.1 indicates [ = 0] is allowed to be of dimension N − 2.
In fact, when N ≥ 2 and
−1
(x) = ln |x̂| , x̂ = (x1 , x2 ),
ln  will be in W 1,2 (B R (0)). Note the set of zeros of this  is an N −2 dimensional
smooth manifold in R N . Let S be an N − 1 dimensional smooth manifold that
1
contains [ = 0]. Then ln  ∈ W 2 ,2 (S) and  contains a set of N −2 dimensional
zeros. This mainly exhibits, up to introducing a local coordinate x N = φ(x̄) as
above, that if  is defined on the sphere S 2 in R3 , then it is allowed to have a set
of zeros that occupies a one dimensional curve on S 2 .
Notion of Distributional Solutions and Positivity
Now suppose we are given a distributional solution u to
u, ln u ∈ L 1loc (E); λ > 0
(3.8)
λu −  ln u = f ∈ L 1loc (E) distributionally in E.
Denote (F) as the usual mollification of a function F. Let f ∈ L rloc (E); then,
discarding the term containing λ, we have
−(ln u) ≤ f  .
Logarithmic diffusion equation in bounded domains 113

The classical theory of elliptic equations yields that

γ N
(ln u),+ ∞,Bσρ ≤ 2Nr
(ln u),+ 2,Bρ + γρ 1− 2r  f  r,Bρ
(1 − σ ) 2r −N
An interpolation yields
N
(ln u),+ ∞,Bρ ≤ γ (N , r )(ln u),+ 1,B2ρ +γρ 1− 2r  f  r,B2ρ
Letting  → 0 yields
N
(ln u)+ ∞,Bρ ≤ γ (N , r )(ln u)+ 1,B2ρ + γρ 1− 2r  f r,B2ρ
On the other hand, this implies
−(ln u) = f  − λu  ∈ L rloc (E)
Another application of the classical elliptic theory implies that
γ N  
(ln u) ∞,Bσρ ≤ 2Nr
(ln u) 2,Bρ + γρ 1− 2r  f  r,Bρ + λu  r,Bρ
(1 − σ ) 2r −N
An interpolation yields
N  
(ln u) ∞,Bρ ≤ γ (N , r )(ln u) 1,B2ρ + γ (N , r )ρ 1− 2r  f  r,B2ρ +λu  r,B2ρ
Thus when letting  → 0
N  
 ln u∞,Bρ ≤ γ (N , r ) ln u1,B2ρ + γ (N , r )ρ 1− 2r  f r,B2ρ + λur,B2ρ .
Since ur can be estimated using  f r as above, we conclude that u is bounded
above and below in Bρ by constants depending on
{N , λ,  f r,B2ρ ,  ln u1,B2ρ }

4. Nonexistence of solutions for the elliptic equation (1.2)

Lemma 1.1 implies that solutions could be generated even if the datum  vanishes
1
on a set of positive H N −2 -measure of ∂ E provided ln  ∈ W 2 ,2 (∂ E). The trace is
taken by ln u ∈ W 1,2 (E). Nevertheless, if  vanishes on a set of positive H N −1 -
measure of ∂ E then it is impossible to generate a solution u such that both u and
ln u take the boundary traces  and ln  in the sense that u, ln u ∈ W 1,2 (E).This
was explained earlier independent of the equation. Now it is natural to ask if there
exists a function u that solves the Dirichlet problem in a weaker sense. This is what
we will explore next.
Define the notion of local weak solutions to the first of (1.2) irrespective of
boundary data as
1,2
u ∈ L loc
2
(E), ln u ∈ Wloc (E), λ > 0;
(4.1)
λu −  ln u = f ∈ L (E)
r
weakly in E.
We have the following
114 N. Liao

Theorem 4.1. If  vanishes on an open subset of ∂ E, then there is no solution

to (1.2) in the sense that (4.1) is satisfied in the interior and the zero boundary
datum is taken in the sense of traces.
This statement is entirely local to the boundary and independent of the interior
continuity or positivity of the solution. It indicates that zeros on the boundary
propogate into the interior even if we only use u to take the trace in the Sobolev
sense.
The proof of Theorem 4.1 hinges on the uniform continuity of the solution at
O and this is the content of the following proposition.
Proposition 4.1. Let u be a nonnegative, locally bounded, local, weak solution
to (1.2) in E. Assume O is an open subset of ∂ E. If u vanishes in the sense of trace
on O, then there exist constants γ and α so that u satisfies
|u(x)| ≤ γ |x − y|α
for any x ∈ E and y ∈ O.
This proposition has a parabolic counterpart in Sect. 6. We omitt the proof for the
elliptic equation while giving the proof of the parabolic case in the Appendix.
Proof of Theorem 4.1. In the weak formulation of the local weak solutions to (4.1)
we take ϕ as a test function which is a usual smooth mollification of ϕ ∈ Co∞ (E).
Then we have
λu  − (ln u) = f  in E.
For δ > 0, we define an interior region
E δ = {x ∈ E : dist(x, ∂ E) > δ}.
Moreover, define a subset of ∂ E δ that corresponds to O as
∂ E δ,O = {a − δν : a ∈ O, ν is the outer normal at a}.
Note that, by Proposition 4.1, for any a ∈ O and the outer normal ν at a
u(a − δν) ≤ γ δ α .
Now Poisson’s representation formula yields that for any x ∈ E δ
 
(ln u) (x) = P(x, y)(ln u) (y) dy + G(x, y)( f  (y) − λu  (y)) d x
∂ Eδ Eδ
 
≤ P(x, y)(ln u) (y) dy + P(x, y)(ln u) (y) dy
∂ E δ,O ∂ E δ \∂ E δ,O
+ G(x, ·) r −1
r  f r

≤ ln(γ δ α ) + C(N , r, diam(E), u∞ ,  f r ).

Since the right-hand side is independent of , we have for a.e. x ∈ E δ
ln u(x) ≤ ln(γ δ α ) + C(N , r, diam(E), u∞ ,  f r )
→ −∞ as δ → 0.
Hence u = 0 a.e. in E.
Logarithmic diffusion equation in bounded domains 115

The main ingredients of the proof are the continuity at the portion of the bound-
ary where u = 0 and a mollification of the PDE, which allows us to apply Poisson’s
representation in E δ . A similar nonexistence result actually holds for distributional
solutions if we know apriori that u takes zero boundary value on O uniformly
continuously.
Proposition 4.2. If u is a nonnegative, bounded, distributional solution to (3.8)
such that it takes zero boundary datum at O uniformly continuously. Then u = 0
in E.

5. Existence of solutions to (1.1)

5.1. Proof of Theorem 1.1

5.1.1. The δ-problem Uniqueness follows from Proposition 2.1. By assumptions,

ln g admits an extension to E × (0, T ), which we still denote as ln g, and it satisfies
ln g ∈ L 2 (0, T ; W 1,2 (E)) ∩ W 1,1 (0, T ; L 1 (E)). Assume momentarily that ln g ≥
ln δ u o ≥ δ for a small δ > 0.
For a positive integer n, slice the time interval (0, T ) into n equal subintervals
1
with length h = T /n. Since ln g(·, t) ∈ W 2 ,2 (∂ E) for a.e. 0 < t < T we may
assume this is the case for any t of the form

iT
: 1 ≤ i ≤ n, n = 1, 2, 3, . . . .
n
Construct a sequence of approximating solutions by setting u(·, 0) = u o and for
k = 0, 1, 2, . . . , (n − 1), the function u(x, (k + 1)h) is the solution to
 u(x, (k + 1)h) u(x, kh)
−  ln u(x, (k + 1)h) = ;
h h (5.1)
u(x, (k + 1)h) = g(x, (k + 1)h) on ∂ E.
By the results from the elliptic problem, especially (3.5) and (3.6), there exist
u(x, kh) such that
def
δ ≤ u(x, kh) ≤  = max{u o ∞ , g∞ }
and ln u(x, kh) ∈ W 1,2 (E) for all 1 ≤ k ≤ n; moreover for any ϕ ∈ Wo1,2 (E)
 
u t¯(x, kh)ϕ d x + D ln u(x, kh)Dϕ d x = 0. (5.2)
E E
Here
u(x, kh) − u(x, (k − 1)h)
u t¯(x, kh) = .
h
The following identity is a discrete version of integration by parts in (0, T ).

n 
n−1
h et¯(k) f (k) = e(n) f (n) − e(0) f (0) − h f t¯(k + 1)e(k). (5.3)
k=1 k=0
116 N. Liao

Here f and e are mappings from {0, 1, . . . , n} to R. Now choose ϕ to be hη(x, t) ∈

Co∞ (E × [0, T )). Sum over k from 1 to n and use the identity (5.3) to obtain

n−1 
 
−h u(x, kh)ηt¯(x, (k + 1)h) d x − u o η(x, 0) d x
k=0 E E

n 
+h D ln u(x, kh)Dη(x, kh)d x = 0
k=1 E

If we denote by (F)n (x, t) a function that equals F(x, kh) in the interval [kh, (k +
1)h). Note (ln u)n = ln(u)n and (D ln u)n = D(ln u)n = D ln(u)n then, the above
equality can be rewritten as
 T    T 
− (u)n (ηt¯)n (·, t + h) d xdt − u o η(x, h) d x + D ln(u)n D(η)n d xdt = 0.
0 E E h E
(5.4)

As n → ∞, D(η)n and (ηt¯)n will converge to Dη and ηt uniformly in E T .

In order to pass to the limit, we need to identify the weak convergence of (u)n
and D ln(u)n .
First of all, we find a uniform bound for D ln(u)n 2 . Use ϕ = h(ln u(·, kh) −
g(·, kh)) in (5.2) and sum over k from 1 to n to obtain
n 
 n 

h |D ln u(x, kh)|2 d x = h D ln u(x, kh)Dg(x, kh) d x
k=1 E k=1 E
n 
+h u t¯(x, kh) ln g(x, kh) d x
k=1 E
n 
−h u t¯(x, kh) ln u(x, kh) d x
k=1 E

= I1 + I2 + I3 .

The first integral is estimated by Young’s inequality as

n  n 
h h
I1 ≤ |D ln u(x, kh)|2 d x + |D ln g(x, kh)|2 d x.
2 E 2 E
k=1 k=1

The second integral is estimated by using the identity (5.3). Then

 
I2 = u(x, nh) ln g(x, nh) d x − u o (x) ln g(x, 0) d x
E E
n−1 

−h (ln g)t¯(x, (k + 1)h)u(x, kh) d x.
k=0 E
Logarithmic diffusion equation in bounded domains 117

Finally
 
I3 = u o (x) ln u o (x) d x − u(x, nh) ln u(x, nh) d x
E E
n−1 

+h (ln u)t¯(x, (k + 1)h)u(x, kh) d x.
k=0 E

We estimate the last term using the elementary inequality

ln(x + 1) ≤ x, ∀x > −1.

In fact,
n−1 

h (ln u)t¯(x, (k + 1)h)u(x, kh) d x
k=0 E
n−1 

= u(x, kh)(ln u(x, (k + 1)h) − ln u(x, kh)) d x
k=0 E
n−1 
 u(x, (k + 1)h)
= u(x, kh) ln − 1 + 1 dx
u(x, kh)
k=0 E
n−1 
 u(x, (k + 1)h)
≤ u(x, kh) − 1 dx
u(x, kh)
k=0 E
n−1 

= [u(x, (k + 1)h) − u(x, kh)] d x
k=0 E
 
= u(x, nh) d x − u o (x) d x.
E E

Thus, collecting all of these we have

n 
h
|D ln u(x, kh)|2 d x
2 E
k=1
n 
h
≤ |D ln g(x, kh)|2 d x
2 E
k=1
 
+ u(x, nh) ln g(x, nh) d x − u o (x) ln g(x, 0) d x
E E
n 

−h (ln g)t¯(x, kh)u(x, kh) d x
k=1 E
 
+ u o (x) ln u o (x) d x − u(x, nh) ln u(x, nh) d x
E  E

+ u(x, nh) d x − u o (x) d x.

E E
118 N. Liao

By our assumption that ln g ∈ W 1,1 (0, T ; L 1 (E)), ln g(·, t) ∈ L 1 (E) for all 0 ≤
t ≤ T . Hence the second and the third integrals on the right-hand side are bounded
by ( ln g(·, 0+ )1 +  ln g(·, T − )1 ). The first and forth integrals on the right-
hand side are bounded in view of the assumption on the extension of ln g. The last
four integrals on the right-hand side are easily seen to be bounded by a constant
depending on {, |E|}.
Hence
D ln(u)n 2 ≤ C (5.5)
for some constant independent of δ and dependent on

{D ln g2 , (ln g)t 1 , , |E|,  ln g(·, 0+ )1 ,  ln g(·, T − )1 }

The estimate (5.5) implies

D ln u δ 2 ≤ C.

Here C is independent of δ and depending on

{D ln g2 , (ln g)t 1 , , |E|,  ln g(·, 0+ )1 ,  ln g(·, T − )1 }.

Since we know D ln u δ − D(ln g)δ 2 is uniformly bounded in L 2 (0, T ; Wo1,2 (E)),

the imbedding theorem implies ln u δ − (ln g)δ is uniformly bounded in L 2 (E T ).
Thus ln u δ is uniformly bounded in L 2 (E T ) by a constant depending only on

{D ln g2 , (ln g)t 1 ,  ln g2 , , |E|,  ln g(·, 0+ )1 ,  ln g(·, T − )1 }.

Thus we have shown that the following quantities are bounded by quantities
independent of δ.

(u)n ∞ ;  ln(u)n 2 ; D ln(u)n 2 .

By using the boundedness of the last quantity we show in Appendix B, for some
h o > 0,

(u)n (x, t + s) − (u)n (x, t) L 2 (0,T −h o ;W −1,2 (E)) → 0 as s → 0

uniformly in n. Here W −1,2 (E) is the dual space of Wo1,2 (E). This implies, by an
argument in [13], there is a subsequence of (u)n converging to some u ∈ L ∞ (E T )
in L 2 (E T −h o ) and a.e. in E T . We give a proof of this fact in Appendix B.
By the boundedness of the last quantity, there is a vector a with each component
in L 2 (E T ) such that

D ln(u)n → a weakly in L 2 (E T ).

Noting ln(u)n is uniformly bounded by a constant depending on δ and  and

ln(u)n → ln u a.e. in E T , we have by the Dominated Convergence Theorem

ln(u)n → ln u in L 2 (E T ).

Thus a = D ln u.
Logarithmic diffusion equation in bounded domains 119

Using these, we can pass to the limit in (5.4) to obtain

  
− uηt − u o η(x, 0) d x + D ln u Dη d xdt = 0.
ET E ET

Next, we show ln u takes ln g as its trace on ∂ E. To this end, we consider the

trace inequality (3.1). A time integration over (0, T ) yields that

 ln(g)n − ln u2,ST ≤ C ln(u)n − ln u2,E T → 0 as n → ∞.

Here C depends on the uniform bound of ln(u)n in L 2 (0, T ; W 1,2 (E)). On the
other hand,

ln(g)n → ln g a.e. in ∂ E

and in view of the assumption δ ≤ g ≤  and the dominated convergence theorem,

we have for any 1 ≤ p < ∞

ln(g)n → ln g in L p (ST ).

Then an application of the triangle inequality yields ln u = ln g on ST .

5.1.2. When δ → 0 Now let us consider the case when g vanishes somewhere on
the boundary. By our assumptions, ln g admits an extension to E T , which we still
denote as ln g, such that

ln g ∈ L 2 (0, T ; W 1,2 (E)) ∩ W 1,1 (0, T ; L 1 (E)).

We take the truncations (u o )δ = max{u o , δ} and (ln g)δ = ln max{g, δ}, then
(ln g)δ is in the same functional spaces as above. By our previous argument, (u o )δ
and (ln g)δ generate a solution u δ and it satisfies
  
− u δ ηt − (u o )δ η(x, 0) d x + D ln u δ Dη d xdt = 0
ET E ET

for all η ∈ Co∞ (E × [0, T )).

Note also u δ1 ≤ u δ2 if δ1
≤ δ2 . Indeed, let (u)δ1 ,n and (u)δ2 ,n be the approxi-
mating solutions corresponding to the initial-boundary data truncated by δ1 and δ2
respectively. Since a comparison principle holds for the elliptic equation (1.2), we
have for any fixed n, (u)δ1 ,n ≤ (u)δ2 ,n in E T . Then letting n → ∞ yields u δ1 ≤ u δ2 .
Suppose the limit of u δ is u and then ln u δ → ln u a.e. in E T . From the uniform
boundedness of ln u δ in L 2 (0, T ; W 1,2 (E)), we can extract a subsequence and a
vector a with all components in L 2 (E T ) such that

D ln u δ → a weakly in L 2 (E T ).

On the other hand, by the Dominated Convergence Theorem ln u δ → ln u in

L 2 (E T ). Hence a = D ln u.
Finally, ln u(·, t) takes trace ln g(·, t) and this is proved in a similar fashion as
in the case of u δ .
120 N. Liao

Now we consider the trace inequality (3.1) applied to ln u(·, t) − ln u δ (·, t). A
time integration over (0, T ) yields that

(ln g)δ − ln u2,ST ≤ C ln u δ − ln u2,E T → 0 as n → ∞.

Here C depends on the uniform bound of ln u δ in L 2 (0, T ; W 1,2 (E)). On the other
hand, the dominated convergence theorem yields

(ln g)δ → ln g in L 2 (ST ).

Then an application of the triangle inequality yields ln u = ln g on ST .

We are unable to claim the constructed solution is locally continuous. Although
we have local positivity and local continuity in the space variable for each element
u(x, kh), we lose both properties when passing to the limit. In fact, as we will see
later in the discussion following Proposition 5.1, one cannot expect a general local
estimate of the modulus of continuity that depends only on the boundedness of the
solution. Additional assumptions seem necessary, and that is what we discuss next.

5.1.3. Locally smooth solutions with semi-convexity in t The construction runs ex-
actly the same as for Theorem 1.3. The main difference is brought by the monotonic-
ity of g/t. Indeed, it implies a discrete version of the semiconvexity inequality (1.3);
for any 0 ≤ k ≤ n − 1

u(x, (k + 1)h) u(x, kh)

≤ ∀x ∈ E.
k+1 k

This can be proved using induction. See [4]. Now, we use (3.7) to conclude that for
any compact subset K of E there is a constant C1 (N , dist(K , ∂ E)) such that
 
1
ln u(x, kh) ≥ −C1 | ln g(x, kh)| dσ − G(x, y)[u(y, kh) − u(y, (k − 1)h)] dy
∂E h E

for all x ∈ K . Multiplying both sides by h and summing over k from jo > 0 to
j1 ≤ n to obtain


j1 j1 

h ln u(x, kh) ≥ −C1 h | ln g(x, kh)| dσ
k= jo k= jo ∂ E

− G(x, y)[u(y, j1 h) − u(y, ( jo − 1)h)] dy
E
 T 
≥ −C1 | ln g| dσ dt − C2
0 ∂E

where C2 depends on {N , diam(E), }.

Logarithmic diffusion equation in bounded domains 121

On the other hand, the discrete semiconvexity inequality implies that for any
0 < jo < j1 ≤ n


j1 
j1
u(x, kh)
h ln u(x, kh) = h ln + ln k
k
k= jo k= jo
u(x, jo h)
≤ h( j1 − jo + 1) ln + h( j1 − jo + 1) ln j1
jo
u(x, jo h)
≤ T ln + T ln j1
jo

Combining these estimates yields for any 0 < jo < j1 ≤ n

 T 
C1 C2 jo
ln u(x, jo h) ≥ − | ln g| dσ dt − + ln .
T 0 ∂E T j1

Since this estimate is independent of n, we conclude, for any compact subset K ×

[t1 , t2 ], there is a constant C depending on
 
ln g1,∂ E×(0,T ) , T, C1 , C2 , t1 , t2 , 

such that

(u)n (x, t) ≥ e−C ∀(x, t) ∈ K × [t1 , t2 ].

The rest of the proof follows easily.

The bound from below in the above proof hinges upon the L 1 (∂ E × (0, T ))
norm of ln g. It is natural to ask what happens if ln g is not in L 1 (∂ E × (0, to )) for
some to < T . Let g be a smooth function on ST and u o, is a nonnegative bounded
function in E. Let us consider a family of solutions to

u ,t −  ln u  = 0 in E T ;
ln u  = ln g on ∂ E × (0, T ); (5.6)
u  (·, 0) = u o, .

We will see in the following proposition that this sequence of approximating prob-
lems does not give any meaningful solution when t ≥ to as the limit function will
always be zero in the interior for t ≥ to .

Proposition 5.1. Suppose u  is a decreasing sequence of C 2 ( Ē × (0, T )) solutions

to (5.6). Assume u  is uniformly bounded by M and satisfies the semiconvexity
inequality (1.3). If there is a 0 < to < T such that
 to 
lim | ln g (y, t)| dσ dt = ∞,
→0 0 ∂E

then the limit function u(x, t) = 0 for all (x, t) ∈ E × [to , T ).

122 N. Liao

Proof. Fix xo ∈ E; it suffices to show u(xo , τ ) = 0 for all τ ≥ to . By the Poisson’s

representation formula
 
ln u  (xo , t) = ln g (y, t)P(xo , y) dσ − G(xo , y)u ,t (y, t) dy.
∂E E

Integrate this formula in dt over (0, τ ) for τ ≥ to to obtain

 τ  τ   τ 
ln u  (xo , t) dt = ln g (y, t)P(xo , y) dσ dt − G(xo , y)u ,t (y, t) dy
0
0 τ ∂ E 0 E

≤ ln g (y, t)P(xo , y) dσ dt + G(xo , y)u o, (y) dy

0 ∂E E

The left-hand side is estimated by using the semiconvexity inequality (1.3), namely,
 τ  τ  τ
u  (xo , t)
ln u  (xo , t) dt = ln dt + ln t dt
0 0 t 0
 τ
≥ τ ln u  (xo , τ ) − τ ln τ + ln t dt.
0

Thus
 τ 
ln u  (xo , τ ) ≤ ln g (y, t)P(xo , y) dσ dt + C(N , τ, M, diam(E)).
0 ∂E

From this we conclude that u(xo , τ ) = 0 for all τ ≥ to .

Remark 5.1. When g =  and N = 1 Proposition 5.1 has been reported in [12].
However, we have given a more general criterion for all dimensions.

Remarks on the Local Behaviors of Solutions to (1.6)

It is well-known that a nonnegative, locally bounded, local weak solution to the
porous medium equation or fast diffusion equation
   1,2 
u ∈ Cloc 0, T ; L 2loc (E) , u m ∈ L 2loc 0, T ; Wloc (E) ;
(5.7)
u t − div u m−1 Du = 0 weakly in E T = E × (0, T ]
is locally Hölder continuous for all 0 < m < 1 and the modulus of continuity ω(·)
over a compact subset K of E T is estimated by the bound of the local solution on
K and the distance of K to ∂ p E T . See [8]. Nevertheless, this need not be true for
the logarithmic diffusion equation, even though it is a formal limit of the porous
medium equation as m → 0. The main technical reason lies in that it is not know
how to establish a DeGiorgi-type lemma for (u − k)− , without assuming sufficient
integrability of ln u; see Lemma C.1 in Appendix C and the remark following it.
To see the failure of such a local continuity estimate, let us consider (1.1) with
the initial datum u o = 1 and the boundary datum
⎧
⎪ 1, 0 ≤ t < 1;
⎨
1
g (x, t) = − (t − 1) + 1, 1 < t < 1 +  −  2 ; (5.8)
⎪
⎩ 
, t ≥1+− . 2
Logarithmic diffusion equation in bounded domains 123

We can further use a mollification to generate a smooth decreasing g that equals

one when t ≤ 1 and  when t ≥ 1 +  −  2 . By Theorem 1.3, this generates a
family of smooth solutions u  that satisfy the semiconvexity inequality (1.3). By
the maximum principle, the sequence of solutions u  decreases to a function u. If
the modulus of continuity of u  over a compact subset K of E T could be estimated
only in term of the uniform bound and the distance of K to ∂ p E T , then u should
be a continuous function in the interior by Arzela-Ascoli’s theorem. However, we
clearly have u = 0 when t > 1 by Proposition 5.1 and, by the maximum principle,
u ≥ 1 when 0 ≤ t ≤ 1. A contradiction. Thus, such a local continuity estimate
can never be attained. It would be interesting to see an explicit solution that shows
the local continuity fails irrespective of possible boundary data. A construction is
claimed in [15]. However, the intended notion of solutions seems different from
ours.
Although we have proved a Harnack-type inequality for locally bounded, non-
negative, local weak solutions to the logarithmic diffusion equation in [5] assuming
∞ p
ln u ∈ L loc (0, T ; L loc (E)) (5.9)
for some p > N + 2, we will show by an explicit example in the next section that
there exist continuous solutions that violate such a condition. Thus, it is interesting
to ask what are the minimal conditions for proving local continuity.
We have established the following Harnack inequality in the topology of L loc1 (E)

in [7] for local solutions u defined as (1.6).

   
t −s
sup u(x, τ )d x ≤ γ inf u(x, τ )d x + 2−N . (5.10)
s<τ <t K ρ (y) s<τ <t K (y)
2ρ
ρ
It is assumed (5.9) holds for some p ≥ 2. The constant γ depends on N and the
∞ (0, T ; L p (E)) norm of ln u.
L loc loc
1 Harnack inequality, if such a requirement on ln u is not
The failure of the L loc
assumed, also follows from this example. Indeed, suppose the conclusion holds
with a constant γ depending only on N . Then (5.10) is satisfied by u  which is
constructed using the boundary datum (5.8). Fix ρ and take s = 1 and t = 1 + σ
then
 
sup u  (x, τ )d x ≥ u  (x, 1)d x ≥ |K ρ |.
s<τ <t K ρ (y) K ρ (y)
On the other hand, for any given δ > 0, σ can be chosen so small that
t −s
< δ.
ρ 2−N
Now fix such a σ ; we can choose  so small that
 
inf u  (x, τ )d x ≤ u  (x, 1 + σ )d x < δ.
s<τ <t K 2ρ (y) K 2ρ (y)
Thus, when ρ is fixed we have
|K ρ | < 2γ δ ∀δ > 0.
A contradiction.
124 N. Liao

5.2. Locally continuous solutions with quenching times

In this section, we present a construction of continuous solutions that is not covered

by the previous section. Continue to assume ∂ E is smooth. Consider the function
η(t)
S(x, t) =
C1 + C2 φ(x)
where φ is the positive solution to the first eigenvalue problem of − in E with
eigenvalue λ1 and η is a nonnegative absolutely continuous function satisfying
 T
|η (t)| ≤ M and | ln η(t)|2 dt < ∞.
0

For example, η(t) = (1 − t)2 will satisfy the above conditions.

Note that −φ = λ1 φ ≥ 0 and by Hopf’s lemma |Dφ(x)| ≤ −Dφ(x)·ν(x) >
0 in ∂ E for any x ∈ ∂ E. See Proposition 5.1 on p.53 of [1]. Thus there is δ > 0
such that λ1 φ 2 + |Dφ|2 ≥ δ in Ē. A direct calculation yields
 
1  
St −  ln S = (C1 + C2 φ) η (t) − λ1 C2 φ − C2 |Dφ| 2 2
(C1 + C2 φ)2
 
1
≤ (C 1 + C 2 φ) (M − λ 1 C 2 φ) − C 2
2 |Dφ| 2
(C1 + C2 φ)2
  
1
≤ M (C 1 +C 2 φ) − λ 1 C 1 C 2 φ − C 2
2
λ 1 φ 2
+ |Dφ| 2
(C1 +C2 φ)2
 
1
≤ M (C 1 + C 2 φ) − λ C
1 1 2C φ − C 2 .
2
δ
(C1 + C2 φ)2
Now the right-hand side is easily seen to be nonpositive if C2 is large enough. More
precisely,

Mφ∞ + M 2 φ2∞ + 4δ MC1
C2 ≥ .
2δ
Let Co > 0 and functions η and φ as above. Consider another function

U (x, t) = Co η(t)(1 + φ(x)) in E T .

A similar calculation yields

 
Ut −  ln U = Co η (t)(1 + φ) + (1 + φ)−2 λ1 (1 + φ)φ + |Dφ|2
 
≥ (1 + φ)−2 −Co M(1 + φ)3 + λ1 φ(1 + φ) + |Dφ|2 .

Thus, in order to guarantee U to be a supersolution we only need to choose Co such

that
λ1 φ(1 + φ) + |Dφ|2
Co ≤ inf (5.11)
E M(1 + φ)3
Logarithmic diffusion equation in bounded domains 125

or
δ
Co ≤
M(1 + φ∞ )3
Typically η is a nonnegative function, which might be zero at discrete points such
that ln η ∈ L 2 (0, T ). In such a case, it is worth noting that U, S, ln U, ln S ∈
L 2 (0, T ; W 1,2 (E)) and the boundary trace of U (·, t) and S(·, t) can be taken for
every t, while the boundary trace of ln S(·, t) and ln U (·, t) can be taken almost
everywhere except on the set [η(t) = 0].
Recall that C1 is still left to be chosen. We choose Co first according to (5.11)
and then choose C1 so that C1 Co ≥ 1. With the aid of the subsolution S and the
supersolution U we are able to establish the following theorem.
1
Theorem 5.1. Let u o (x) = 0 in E and ln g(x, t) ∈ L 2 (0, T ; W 2 ,2 (∂ E)) satisfy
η(t)
Co η(t) ≥ g(x, t) ≥ in ∂ E × (0, T ).
C1
Then there is a unique nonnegative, bounded, locally continuous solution u to (1.1)
such that

S≤u≤U in E T .

Proof. The construction of the unique solution is in Theorem 1.1. In addition, since
S(x, t) ≤ u(x, t) ≤ U (x, t) for (x, t) ∈ ∂ E T , this is true in E T by the comparison
principle. If η(to ) = 0 for some 0 ≤ to ≤ T , then u will be forced to tend to zero
continuously with the same rate.

An Explicit Example
Let λ > 0. We can solve the following equation using separation of variables.
⎧
⎨ u t −  ln u = 0;
u(x, t) = λt on ∂ E × (0, ∞);
⎩
u(x, 0) = 0.
Indeed, u = λtφ(x) is a solution where φ is the unique solution to
λφ −  ln φ = 0;
φ = 1 on ∂ E.

The unique solution u satisfies u ∈ C ∞ (E ∞ ) and ln u ∈ L 2loc (R+ ; W 1,2 (E)).

Consider next the explicit solution
N
2(N − 2)(−t)+N −2
v= 2
, λo > 0, t ≤ 0, N ≥ 3.
λo + (−t)+N −2 |x|2

Denote by w the patched function over E × R. One verifies that ln w ∈ L 2loc

(R; W 1,2 (E)) and w ∈ C(R; L 2 (E)). Then w will be a continuous solution to (1.1)
across t = 0, though wt is not continuous. Thus we have an explicit continuous
126 N. Liao

solution which violates our assumption for the Harnack-type inequality from [5],
namely, ln u ∈ L loc ∞ (R; L p (E)) for p > N + 2. Accordingly, the conclusion
loc
of the Harnack-type inequality fails since u(·, 0) = 0. On the other hand, further
regularity, C 1 in time, may not be obtained if such an assumption is missing. See
[7].

6. Nonexistence of solutions for the parabolic equation

Now we turn to the proof of Theorem 1.3. It hinges on the continuity of the solution
at O and this is the content of the following proposition.
Proposition 6.1. Let u be a nonnegative, bounded, local, weak solution to (1.6) in
E T . Assume there is an open subset O of the smooth boundary ST . If u vanishes in
the sense of trace on O, then there exist constants γ and α depending only on N
so that u satisfies
−1 1
|u(x, t)| ≤ γ (|x − y| + u∞,E
2
T
|t − s| 2 )α

for any (x, t) ∈ E T and (y, s) ∈ O.

Local boundedness is enough to reach the same conclusion. The proof is a bound-
ary version adaption of the interior arguments in [8], where the interior Hölder
continuity is proved for the porous medium type equations. The interior Hölder
continuity of local weak solutions to the logarithmic diffusion equation cannot be
shown as in [8], and the main difficulty is generated by working with the truncated
function (u − k)− . In order to derive a DeGiorgi-type lemma for (u − k)− we have
to assume a proper extra integrability of ln u; see [5] for such a lemma and see how
it fails in the remark following Lemma C.1 in Appendix C. However, the situation
for our current case is much simpler. First of all, the zero boundary trace allows
us to circumvent working with (u − k)− . Also, the local logarithmic estimates are
avoided since we assume the boundary satisfies the property of positive geometric
density. The complete proof will be reported in the Appendix C.
Proof of Theorem 1.3 Assuming Proposition 6.1. Assume without loss of gener-
ality that O =  × (t1 , t2 ) where  is a open subset of E. Let ϕ ∈ Co∞ (E T ) and
K  (x, t) be a mollifying kernel. Let ϕ be the space-time convolution of ϕ with
K  . In the weak formulation of the local weak solutions to (1.6) we take ϕ as a
test function. Then we have

u ,t − (ln u) = 0 in E T .

Define an interior region in E as

E δ = {x ∈ E : dist(x, ∂ E) > δ} for some δ > 0.

Moreover, define

∂ E δ,O = {a − δν : a ∈ , ν is the outer normal at a}.

Logarithmic diffusion equation in bounded domains 127

t1 ,t2
Note that, by Proposition 6.1, for any (y, t) ∈ ∂ E δ,O

u(y, t) ≤ γ δ α .

Continue to denote by P the Poisson kernel and G the Green function for E. By
Poisson’s representation formula and a time integration over (τ1 , τ2 ) ⊂ (t1 , t2 )
yield that for any x ∈ E δ1 with δ1 > δ
 τ2  τ2 
(ln u) (x, t) dt = P(x, y)(ln u) (y, t) dσ dt
τ1 τ1 ∂ Eδ
 τ2 
− G(x, y)u ,t (y, t) dydt
τ Eδ
 τ2 1
≤ P(x, y)(ln u) (y, t) dσ dt + C (N , diam(E), u∞ )
τ1∂ E δ,O
 
≤ ln γ δ α |τ2 − τ1 | + C (N , diam(E), u∞ ) .

Since the right-hand side is independent of , we have for a.e. x ∈ E δ1

 τ2
ln u(x, t) dt ≤ ln(γ δ α )|τ2 − τ1 | + C(N , diam(E), u(·, τ )∞ )
τ1
→ −∞ as δ → 0.

As a result, for any x ∈ E δ1 and arbitrary (τ1 , τ2 ) ⊂ (t1 , t2 )

 τ2
ln u(x, t) dt = −∞.
τ1

Hence u = 0 a.e. in E × (t1 , t2 ).

As in the elliptic case, the main ingredients of the proof are the continuity at the
portion of the boundary where u = 0 and a mollification of the PDE, which allows
us to apply Poisson’s representation in E δ . A similar nonexistence result actually
holds for distributional solutions if we know apriori that u takes zero boundary
value uniformly continuously. A distributional solution is defined as

u, ln u ∈ L 1loc (E T );
(6.1)
u t −  ln u = 0 distributionally in E T .

Proposition 6.2. If u is a nonnegative, locally bounded, distributional solution

to (1.6) such that it takes zero boundary datum at O uniformly continuously. Then
u = 0 a.e. in E × (t1 , t2 ).

7. Zeros in the interior

We have seen in Sect. 5.2 that locally continuous solutions may have a set of zeros in
the interior of the form of a horizontal hyperplane K ρ (xo ) × {to }. This is permitted,
roughly speaking, because we have less restriction in the time direction. As we will
see next, the geometry of an admissible set where u = 0 is much more restricted
along the t-axis.
128 N. Liao

Since we do not know a local weak solution to (1.6) is locally continuous, the
meaning of the set [u = 0] is ambiguous. Here we define
   
t+ρ
def
[u = 0] = (x, t) ∈ E T : lim u(y, s) dyds = 0 .
ρ→0 t−ρ Bρ (x)

A point (x, t) is called a vanishing point, denoted as u(x, t) = 0, if the above limit
holds.

Proposition 7.1. Let u be a nonnegative, locally bounded, distributional solution

to (6.1). Then, for any x ∈ E, the set V (x) = {t ∈ (0, T ) : u(x, t) = 0} cannot
occupy a set of positive H 1 -measure.

Proof. We may assume that u < 1. Fix x ∈ E and suppose V (x) contains a set
of positive H 1 -measure. In the weak formulation of the distributional solutions
to (6.1) we take ϕ as a test function where ϕ ∈ Co∞ (E T ) and ϕ is a smooth
mollification of ϕ in space and time. Then we have

u ,t − (ln u) = 0 in E T .

Now, if Bρ (x) is a ball centered at x with radius ρ, Poisson’s representation and a

time integration over (τ1 , τ2 ) yield
 τ2  τ2 
(ln u) (x, t) dt = P(x, y)(ln u) (y, t) dσ dt
τ1 τ1 ∂ Bρ (x)
 τ2 
− G(x, y)u ,t (y, t) dydt
τ1 Bρ (x)
 τ2 
≥ (ln u) (y, t) dσ dt − G(x, ·)1 u(·, τ2 )∞ .
τ1 ∂ Bρ (x)

Integrate both sides in dρ over (0, R); we obtain that

 τ2  τ2 
(ln u) (x, t) dt ≥ (ln u) (y, t) dydt
τ1 τ1 B R (x)
−G(x, ·)1 u(·, τ2 )∞ .

This implies that

 τ2  τ2 
lim sup (− ln u) (x, t) dt ≤ lim inf (− ln u) (y, t) dydt
→0 τ1 →0 τ1 B R (x)
+G(x, ·)1 u(·, τ2 )∞ .

Fatou’s lemma yields

 τ2  τ2
lim inf (− ln u) (x, t) dt ≤ lim inf (− ln u) (x, t) dt
τ1 →0 →0 τ1
Logarithmic diffusion equation in bounded domains 129

If (x, t) is a vanishing point, then (− ln u) (x, t) → ∞. If there is a set of positive

H 1 -measure of t in (τ1 , τ2 ) such that u(x, t) = 0, then we conclude, by combining
the previous two estimates, that
 τ2 
lim inf (− ln u) (y, t) dydt = ∞.
→0 τ1 B R (x)

Since ln u ∈ L 1loc (E T ), we reach a contradiction.

Acknowledgement I wish to thank Professor Emmanuele DiBenedetto for his continuous

advice and support. This paper grew out of numerous conversations with him. He also read
the early version of this paper and came up with many helpful comments. I also need to
thank Professor Ugo Gianazza. He suggested me working on these problems.

Appendices

A. Proof of Lemma 3.1

We are going to use

Theorem A.1. (Fixed Point Theorem [9]) Let (X,  · ) be a Banach space and H
be a closed and convex subset of X , x o ∈ H , and T : H ×[0, 1] → H is continuous
and compact with T (·, 0) = xo . If there is a constant M such that x < M apriori
for all x ∈ H and 0 ≤ σ ≤ 1 satisfying T (x, σ ) = x. Then there is a fixed point
for T (x, 1).
Let H = {v ∈ L ∞ (E) : |v| ≤ M} with M to be chosen. Then H is a closed and
convex subset of L ∞ (E). For a measurable function v on E, define vn as
⎧
⎨ n if v > n;
vn = v if − n ≤ v ≤ n;
⎩
−n if v < −n.

Fix n and define an operator w = T (v, σ ) by the unique solution to

⎧
⎨ σ λevn − w = σ f ;
⎩
w = σ ln  on ∂ E.

Here  is defined in Lemma 3.1 and w ∈ W 1,2 (E) ∩ C( Ē) is a solution in the
weak sense.
We apply the stated Fixed Point Theorem to show Lemma 3.1. First of all, an
application of DeGiorgi’s method (see [1]) gives for some C > 0
def
|w| ≤ sup | ln | + C( f τ + λen ) = M(n).
∂E

With such a choice of M(n), T (·, ·) maps H × [0, 1] to H .

130 N. Liao

Secondly, we show T (·, σ ) : H → H is continuous. Let w = T (v ; σ ) and

w = T (v ; σ ) for v , v ∈ H , then

−(w − w ) + σ λ(evn − evn ) = 0, and w − w = 0 on ∂ E.

An application of the DeGiorgi’s method yields that

w − w ∞ ≤ Cen vn − vn ∞ .

Then the left hand side is small if v − v ∞ is made small. The continuity with
respect to σ is proved similarly.
Next, we show T (·, σ ) : H → H is compact uniformly in σ . In fact, since
elements v ∈ H are uniformly bounded by M(n), then functions from T (H, σ ) are
equi-countinuous. Thus Arzela-Ascoli’s theorem yields compactness.
As a result of the fixed point theorem, there exists a solution v (n) ∈ H ∩ W 1,2 (E) ∩
C( Ē) satifying in the weak sense

⎧ (n)
⎨ λevn − v (n) = f ;
(A.1)
⎩
v (n) = ln  on ∂ E.

In order to send n → ∞ we show |v (n) | is uniformly bounded. We discard the

first term in equation (A.1) by its nonnegativity and an application of DeGiorgi’s
method gives the upper bound A = sup∂ E ln  + C f r for some C. Then another
application of DeGiorgi’s method shows we actually have

def
|v| ≤ sup | ln | + C(e A +  f r ) = A1 .
∂E

As a result, v (n) converges to some v ∈ C( Ē) uniformly.

1
Let ln  ∈ W 1,2 (E) denote an extension of ln  ∈ W 2 ,2 (∂ E). Using (v (n) −ln )
as a test function in the weak formulation of v (n) we have

|∇v (n) |2 d x
 E
 
(n)
= ∇v (n) ∇ ln d x + (v (n) − ln ) f d x − evn (v (n) − ln )d x
E
 E E
1 (n) 2
≤ |∇v | d x + C(N , A1 ,  f r )
2 E

We conclude ∇v (n) converges weakly in L 2 to some d. By the uniform convergence

of v (n) we then have d = ∇v. Hence v ∈ C( Ē) will be the desired solution.
Logarithmic diffusion equation in bounded domains 131

B. Convergence of (u)n

Let W −1,2 (E) be the dual space of Wo1,2 (E). Then any f ∈ L 2 (E) can be seen as
an element in W −1,2 (E) in the sense that

[ f, ϕ] = ( f, ϕ).

Here [·, ·] means the pairing of W −1,2 (E) and Wo1,2 (E) and (·, ·) means the inner
product in L 2 (E). The norm in W −1,2 (E) is defined as

 f W −1,2 (E) = sup [ f, ϕ].

ϕW 1,2 (E) ≤1

Hence  f W −1,2 (E) ≤  f 2 .

Lemma B.1. The discrete function (u)n constructed in the proof of Theorem 1.1
converges to some u in L 2 (E T −h 1 ) for any 0 < h 1 < T .

Proof. Let (u)n,h be the Steklov average of (u)n . For h fixed, {(u)n,h } is precompact
in C(0, T − h; L 2 (E)). Indeed, by the general Ascoli-Arzela’s theorem (p291,
[2]), we only need to verify (u)n,h (·, t) is equibounded in L 2 (E) for any fixed
t ∈ (0, T − h) and (u)n,h (·, t) is equicontinuous at t in the topology of L 2 (E). All
of them are clear since (u)n is uniformly bounded by .
Then {(u)n,h } is precompact in L 2 (0, T − h; W −1,2 (E)) automatically. Next we
show (u)n,h → (u)n in L 2 (0, T − h; W −1,2 (E)) uniformly in n. Indeed, from the
equation (5.2), we have

u(·, kh) − u(·, (k − 1)h)W −1,2 (E) ≤ hD ln u(·, k)2 .

For an integer l, let s = (l + 1)h. Summing over k from jo + 1 to jo + l to obtain

jo +l

u(·, ( jo + l)h) − u(·, jo h)W −1,2 (E) ≤ hD ln u(·, kh)2
k= jo +1
⎛ ⎞1
jo +l

2

≤ s2 ⎝ hD ln u(·, kh)22 ⎠ .

1

k= jo +1

Then taking power 2 and multiplying by h at both sides and summing over jo from
0 to n − l to obtain


n−l
hu(·, jo h + s) − u(·, jo h + h)2W −1,2 (E)
jo =0


n−l  jo h+s 
≤s h |D ln(u)n |2 d xdt
jo =0 jo h E

≤ sT |D ln(u)n |2 d xdt.
ET
132 N. Liao

This gives

(u)n (x, t + s) − (u)n (x, t)2L 2 (0,T −s;W −1,2 (E)) = O(s) → 0 as s → 0

uniformly in n. Thus an application of the triangle inequality yields that {(u)n } is

precompact in L 2 (0, T − h 1 ; W −1,2 (E)) for any 0 < h 1 < T .
In order to show {(u)n } is precompact in L 2 (E T −h 1 ), we show that for any  > 0
there is a constant C depending only on  such that

v L 2 (E) ≤ vW 1,2 (E) + C vW −1,2 (E) for all v ∈ W 1,2 (E).

Suppose this is false, then there exist o > 0, Ci → ∞ and vi ∈ W 1,2 (E) such that

vi  L 2 (E) ≥ o vi W 1,2 (E) + Ci vi W −1,2 (E) .

Let wi = vi /vi W 1,2 (E) then we have

wi  L 2 (E) ≥ o + Ci wi W −1,2 (E) . (B.1)

Since the left-hand side is bounded independent of i, we get wi W −1,2 (E) → 0.
However the boundedness of wi in W 1,2 (E) and the Compact Imbedding Theorem
imply that wi converges to some w in L 2 (E), and hence in W −1,2 (E). This forces
w = 0. Thus we reach a contradiction in (B.1).
Since {(u)n } is precompact in L 2 (0, T − h 1 ; W −1,2 (E)), it is totally bounded.
See Proposition 17.6 on p48 of [2]. That means for any δ > 0 we have a finite set
{(u)n i } ⊂ {(u)n } such that for any (u)n there is (u)n i satisfying

(u)n − (u)n i  L 2 (0,T −h 1 ;W −1,2 (E)) < δ.

This joint with the previous interpolation inequality yield that

(u)n − (u)n i  L 2 (E T −h < (u)n − (u)n i  L 2 (0,T −h 1 ;W 1,2 (E))

1)
+C (u)n − (u)n i  L 2 (0,T −h 1 ;W −1,2 (E))
≤  M + C δ.

Here M is the uniform bound of {(u)n } in L 2 (0, T ; W 1,2 (E)). Choosing  and δ
appropriately we find {(u)n } is totally bounded in L 2 (0, T − h 1 ; W 1,2 (E)). Hence
the proof is concluded by Proposition 17.6 on p48 of [2].

C. Proof of Proposition 6.1

C.1. An energy estimate

Let u be a local solution to (1.6). The logarithmic diffusion equation satisfies the
notion of parabolicity defined in [3] and [8]. Thus, (u − k)+ is a local subsolution
to the logarithmic diffusion equation in the sense that
  
∂ D(u − k)+
(u − k)+,h ϕ + Dϕ d x ≤ 0
K ∂t u h
Logarithmic diffusion equation in bounded domains 133

for any ϕ ∈ Wo1,2 (K ) and any compact set K ⊂ E. Here, we have used Steklov
averages. It is referred to [3] for such a notion and its basic properties.
1,2
If we know u(·, t) ∈ Wloc (E ∪ Ot ), then since [. . . ]h is always in L 2loc (E ∪ Ot ),
the above integral inequality holds for any ϕ ∈ Wo1,2 () and any compact set
 × {t} ⊂ E ∪ Ot .
def
Let us assume (y, s) ∈ O and consider the cylinder (y, s) + Q ρ (θ ) = K ρ (y) ×
(s − θρ 2 , s] with θ , ρ > 0 so small that {[y + K 2ρ ] ∩ ∂ E} × (s − θ (2ρ)2 , s] ⊂ O.
By a translation we may assume (y, s) coincides with (0, 0). We may obtain an
energy estimate by taking the test function

ϕh = (u − k)+,h ζ 2 ,

in the weak formulation of (1.1) and integrating over Q ρ (θ ) and then letting h → 0.
Such a choice of test function is admissible since for a.e. t ∈ (−θρ 2 , 0] we know
x → ζ (x, t) vanishes on the boundary of K ρ but not on the boundary of K ρ ∩ E,
and for any k ≥ 0

(u − k)+ (·, t) = 0 in the sense of trace on ∂ K ρ ∩ E.

Thus
(u(·, t) − k)+ ζ 2 (·, t) ∈ Wo1,2 (K ρ ∩ E). (C.1)
With such a choice of k we can establish the following energy estimate near ST .

Proposition C.1. Let u be a nonnegative, local, weak solution to (1.1) in Q ρ (θ )

and ζ is a cutoff function vanishing on the parabolic boundary of Q ρ (θ ). There
exists a constant γ depending only on N such that for every (y, s) ∈ O, for every
cylinder (y, s)+ Q(θ, ρ) such that s −θρ 2 > 0 and every level k ≥ 0, the following
inequality holds:

ess sup (u − k)2+ ζ 2 (x, t) d x
s−θρ 2 <t<s [y+K ρ ]∩E

+ u −1 |D(u − k)+ |2 ζ 2 d xdt
[(y,s)+Q ρ (θ)]∩E T
 (C.2)
≤γ (u − k)2+ ζ |ζt | d xdt
[(y,s)+Q ρ (θ)]∩E T

+ u −1 (u − k)2+ |Dζ |2 d xdt.
[(y,s)+Q ρ (θ)]∩E T

C.2. Proof of the Proposition

def
For a cylinder Q 2ρ (θ ) = K 2ρ × (−θ (2ρ)2 , 0] and a point (y, s) ∈ O we define

μ+ = ess sup u.
[(y,s)+Q 2ρ (θ)]∩E T
134 N. Liao

Since we always have

ess inf u=0

[(y,s)+Q 2ρ (θ)]∩E T

the essential oscillation ω over the cylinder Q 2ρ (θ ) satisfies ω = μ+ . Let ξ and a

be constants in (0, 1).
Lemma C.1. Let u be a nonnegative, locally bounded, local, weak solution to (1.1)
in E T . There exists a positive number ν, depending on ω, θ, ξ, a and N such that
if

|[u ≥ μ+ − ξ ω] ∩ [(y, s) + Q 2ρ (θ )] ∩ E T | ≤ ν|Q 2ρ (θ ) ∩ E T |,

then

u ≤ μ+ − aξ ω a.e. in [(y, s) + Q ρ (θ )] ∩ E T .

Proof. Assume without loss of generality that (y, s) = (0, 0) and for n = 0, 1, . . . .
Set
ρ
ρn = ρ + n , K n = K ρn , Q n = K n × (−θρn2 , 0].
2
Apply (C.2) over K n and Q n to (u − kn )+ , for the levels
1−a
kn = μ+ − ξn ω where ξn = aξ + ξ.
2n
The cutoff function ζ (x, t) takes value 1 on Q n+1 and vanishes on the parabolic
boundary of Q n such that

2n+1 22(n+1)
|Dζ | ≤ and |ζt | ≤ .
ρ θρ 2
Then (C.2) gives
 
ess sup (u − kn )+ ζ (x, t) d x +
2 2
u −1 |D(u − kn )+ ζ ]|2 d xdt
−θn ρn2 <t<0 K n ∩E Q n ∩E T

22n 1 1
≤ γ 2 (ξ ω)2 + χ[u>kn ] d xdt
ρ Q n ∩E T (1 − ξ )ω θ
22n 1 ω
≤γ (ξ ω)2 1+ |[u > kn ] ∩ Q n ∩ E T |.
ρ2 ω(1 − ξ ) θ
The second integral on the left hand side is estimated by
 
−1 −1
u |D(u − kn )+ ζ ]| d xdt ≥ ω
2
|D[(u − kn )+ ζ ]|2 d xdt.
Q n ∩E T Q n ∩E T

Setting
|An |
An = [u > kn ] ∩ Q n ∩ E T and Yn = ,
|Q n ∩ E T |
Logarithmic diffusion equation in bounded domains 135

then
 
−1
ess sup (u − kn )2+ ζ 2 (x, t) d x +ω |D(u − kn )+ ζ ]|2 d xdt
−θn ρn2 <t<0 K n ∩E Q n ∩E T

22n 1 ω
≤γ (ξ ω)2 1+ |An |.
ρ 2 ω(1 − ξ ) θ
An application of the parabolic embedding yields

1−a 2
(ξ ω) 2
|A n+1 | ≤ (u − kn )2+ d xdt
2n+1 Q n+1 ∩E T
 N
N +2
N +2 2
≤ [(u − kn )+ ζ ]2 N d xdt |An | N +2
Q n ∩E T
 N
N +2
≤γ |D[(u − kn )+ ζ ]| d xdt2
Q n ∩E T
 2
N +2 2
× ess sup [(u − kn )+ ζ ] (x, t) d x
2
|An | N +2
−θρn2 <t<0 K n

for a constant γ depending only on N . This joint with the previous estimate gives
−2
γ 24n ω N +2 ω 2
|An+1 | ≤ 1+ |An |1+ N +2 .
(1 − a) ρ 1 − ξ
2 2 θ
In terms of Yn this can be rewritten as
γ 24n 1 + θ ω−1 1+ N 2+2
Yn+1 ≤ Yn .
(1 − a)2 (1 − ξ ) (θ ω−1 ) NN+2

Thus using Lemma 4.1 on p12 of [3] we conclude that Yn → 0 provided

  N +2 N
|Ao | (1 − a)2 (1 − ξ ) 2 (θ ω−1 ) 2 def
Yo = ≤ = ν.
|Q o ∩ E T | γ 4 N +2 −1
(1 + θ ω ) 2
N +2

Remark C.1. In general, we do not have an analogous lemma that states there is a
constant ν− depending on the same set of parameters, such that, if
|[u < ξ ω] ∩ [(y, s) + Q 2ρ (θ )]| ≤ ν− |Q 2ρ (θ )|, (C.3)
then
u ≥ aξ ω a.e. in [(y, s) + Q ρ (θ )].
Indeed, let us consider the sequence of solutions u  constructed after Proposi-
tion 5.1. Fix a = ξ ω = θ = 21 and choose (y, s) ∈ E × (1, T ) such that
(y, s) + Q 2ρ (θ ) ⊂ E T
136 N. Liao

and
|[(y, s) + Q 2ρ (θ )] ∩ [E × (1, T )]| ≤ ν− |Q 2ρ (θ )|.
By Proposition 5.1 and the discussion that followed, u  will satisfy (C.3), while at
the same time the intended conclusion
1
u ≥ in (y, s) + Q ρ (θ )
4
does not hold as
u → 0 uniformly in [(y, s) + Q ρ (θ )] ∩ E × (1, T ).
Since O is smooth, there is a constant β such that
|{x ∈ [y + K ρ ] ∩ E : u(x, t) > 0}| ≤ |K ρ ∩ E| ≤ (1 − β)|K ρ | (C.4)
for all
s − θρ 2 < t ≤ s.
Lemma C.2. Let θ = ω. For every ν ∈ (0, 1) there exists a positive integer q
depending on ν and N such that
ω
|[u > μ+ − q ] ∩ [(y, s) + Q 2ρ (θ )] ∩ E T | < ν|Q 2ρ (θ ) ∩ E T |.
2
Proof. Assume (y, s) = (0, 0) and set Q = Q ρ (θ ) and Q = Q 2ρ (θ ), and use the
energy estimate for the functions
ω
(u − k j )+ where k j = μ+ − j for j = 1, . . . , q,
2
over the pair of cylinders Q and Q . The cutoff function ζ is taken to be one on Q,
vanishing on the parabolic boundary of Q such that
1 2
|Dζ | ≤ and 0 ≤ ζt ≤ , θ = ω.
ρ θρ 2
Then the energy estimate (C.2) gives

−1 ω−1 ω 2
ω |D(u − k j )+ |2 d xdt ≤ γ 2 |Q|. (C.5)
Q∩E T ρ 2j
Now apply the discrete isoperimetric inequality (see p5, [3]) to the function x →
u(x, t), for −θρ 2 < t < 0, over the cube K ρ , for the levels k j and k j+1 . Taking
into account (C.4) this gives
ω
j+1
|[u(·, t) > k j+1 ] ∩ K ρ ∩ E|
2

γρ N +1
≤ |Du| d x
|[u(·, t) < k j ] ∩ K ρ ∩ E T | [k j <u(·,t)<k j+1 ]∩K ρ ∩E
 1
γ 2
≤ ρ |Du(·, t)| d x
2
β [k j <u(·,t)<k j+1 ]∩K ρ
1
×|([u(·, t) > k j ] − [u(·, t) > k j+1 ]) ∩ K ρ ∩ E| 2 .
Logarithmic diffusion equation in bounded domains 137

Set
 0
|A j | = |[u > k j ] ∩ Q ∩ E T | = |[u(·, t) > k j ] ∩ K ρ ∩ E| dt
−θρ 2

and integrate the above estimate in t over the interval (−θρ 2 , 0). We have
 1
ω γ 2 1
|A j+1 | ≤ ρ |D(u − k j )+ |2 d xdt (|A j | − |A j+1 |) 2
2j β Q∩E T

This together with (C.5) gives

γ
|A j+1 |2 ≤ |Q|(|A j | − |A j+1 |).
β2
Add these from j = 1 to j = q − 1; we have


q
γ
q|Aq |2 ≤ |A j+1 |2 ≤ |Q|2 .
β2
j=0

From this
γ
|Aq | ≤ √ |Q|.
qβ

Now, we can choose q from

γ
√ =ν
qβ

Next we choose ν as in the Lemma C.1 and note ν depends only on N since we
choose θ = ω. Then fix q as in the Lemma C.2 and
1 1
ξ= and a=
2q 2
then we have
ω 1
u ≤ μ+ − = 1 − q+1 ω in (y, s) + Q ρ (θ ) ∩ E T .
2q+1 2

This quantitative bound yields

1
ess osc u= ess sup u ≤ 1− ω.
(y,s)+Q ρ (θ)∩E T (y,s)+Q ρ (θ)∩E T 2q+1

We conclude by using a standard induction argument based on the above recursive

inequality. See p45, [3].
138 N. Liao

References

[1] DiBenedetto, E.: Partial Differential Equations. 2nd edn. Birkhäuser, Boston (2009)
[2] DiBenedetto, E.: Real Analysis. Birkhäuser, Boston (2002)
[3] DiBenedetto, E.: Degenerate Parabolic Equations, Universitex. Springer, New
York (1993)
[4] DiBenedetto, E., Diller, D.J.: About a singular parabolic equation arising in thin film
dynamics and in the Ricci flow for complete R2 , Partial differential equations and
applications, Lecture Notes in Pure and Appl. Math., vol. 177, Dekker, New York, pp.
103–119 (1996)
[5] DiBenedetto, E., Gianazza, U., Liao, N.: On the local behavior of a logarithmically
singular equation. Discrete Continuous Dyn. Syst. 17(6), 1841–1858 (2012)
[6] DiBenedetto, E., Gianazza, U., Liao, N.: Logarithmically singular parabolic equa-
tions as limits of the porous medium equation. Nonlinear Anal. TMA 75(12), 4513–
4533 (2012)
[7] DiBenedetto, E., Gianazza, U., Liao, N.: Two remarks on the local behavior of so-
lutions to logarithmically singular diffusion equations and its porous-medium type
approximations, to appear at Rivista di Matematica Univ. di Parma
[8] DiBenedetto, E., Gianazza, U., Vespri, V.: Harnack’s Inequality for Degenerate and
Singular Parabolic Equations, Springer Monographs in Mathematics. Springer, New
York (2012)
[9] Granas, A., Dugundji, J.: Fixed Point Theory, Springer Monographs in Mathemat-
ics. Springer, Berlin (2003)
[10] Hui, K.M.: On some Dirichlet and Cauchy problems for a singular diffusion equa-
tion. Differ. Integral Equ. 15(6), 769–804 (2002)
[11] Krantz, S.G.: Calculation and estimation of the Poisson kernel. J. Math. Anal.
Appl. 302(1), 143–148 (2005)
[12] Rodriguez, A., Vazquez, J.L.: Obstructions to existence in fast-diffusion equations. J.
Differ. Equ. 184(2), 348–385 (2002)
[13] Simon, J.: Compact sets in the space L p (0, T ; B). Ann. Mat. Pura Appl. 146(4), 65–
96 (1987)
[14] Vazquez, J.L.: Nonexistence of solutions for nonlinear heat equations of fast diffusion
type. J. Math. Pures Appl. 71, 503–526 (1992)
[15] Vazquez, J.L.: Failure of the strong maximum principle in nonlinear diffusion. Com-
mun. Partial Differ. Equ. 30, 1263–1303 (2005)
[16] Williams, M.B., Davis, S.H.: Nonlinear theory of film rupture. J. Colloidal Interface
Sci. 90, 220–228 (1982)
[17] Ziemer, W.: Weakly Differentiable Functions, GTM. Springer, Berlin (1989)