Vous êtes sur la page 1sur 108

Journal of Mathematical Sciences, Vol. 108, No.

4, 2002
LINEAR SECOND-ORDER PARTIAL DIFFERENTIAL EQUATIONS
OF THE PARABOLIC TYPE
A. M. Ilyin, A. S. Kalashnikov, and O. A. Oleynik UDC 517.9
The article is a review of the theory of linear partial dierential equations of the parabolic type, where
the classical facts of this theory are described in detail.
Bibliography: 128 titles.
CONTENTS
Introduction 436
1 Maximum Principle. Uniqueness of Solutions of the Main Boundary-Value Problems 437
2 A priori Estimates 448
3 Solving of Boundary-Value Problems by the Rothe Method. Cauchy Problem 463
4 The Fundamental Solution of a Linear Parabolic Equation. Greens Function.
Method of Integral Equations for Solving Boundary-Value Problems 480
5 Generalized Solutions of Boundary-Value Problems. Uniqueness Theorems.
Some Auxiliary Propositions 501
6 Method of Finite Dierences 507
7 Some Functional Methods of Solving Boundary-Value Problems 513
8 Solution of the First Boundary-Value Problem
by the Method of Continuation with Respect to a Parameter 517
9 Application of Galerkins Method
to the Construction of a Solution of the First Boundary-Value Method 518
10 Generalized Solution for the Cauchy Problem 520
11 On the Smoothness of Generalized Solutions 527
12 Behavior of Solutions with Innite Increasing of Time 533
References 537
Translated from Trudy Seminara imeni I. G. Petrovskogo, No. 21, pp. 9193, 2001. The Russian version of this
article is a reprint from the journal Uspekhi Matematicheskikh Nauk (1962, Vol. 17, issue 3 (105), pp. 3146) with
the consent of the editorial board.
10723374/02/10840435 $ 27.00 c _ 2002 Plenum Publishing Corporation 435
INTRODUCTION
In the general theory of partial dierential equations, that of second-order equations occupies a special place.
At present, the theory of second-order equations is the best developed. Many methods invented to study second-
order equations turned out to be useful also for the study of higher-order equations and systems of partial dierential
equations. The strong interest in the study of second-order equations, and parabolic equations in particular, is also
motivated by their importance in mechanics, heat conduction, probability theory, and other elds of mathematics
and physics. These applications constantly produce new problems for partial dierential equations. In particular,
one may read in E. B. Dynkins article [1] about new problems for second-order partial dierential equations arising
in connection with probability theory.
While there are monographs and large reviews on the theory of elliptic and hyperbolic equations (among
them, such a detailed review as the work of C. Miranda [2]), up to now there has been no survey of the most
important facts concerning parabolic second-order equations.
The present article was written mainly owing to the inuence of numerous questions that people have asked
the authors and the fact that we had diculty in indicating the corresponding literature.
This article is a review of the theory of second-order parabolic partial dierential equations with detailed
proofs of many basic facts of this theory. We pay special attention to the methods of construction of classical and
generalized solutions for boundary-value problems and the Cauchy problem. The majority of these methods have
a larger range of applicability and can be used in the study of higher-order equations and systems of equations
of dierent types. In each section, we indicate papers where the method we discuss nds application in dierent
questions connected with the theory of partial dierential equations.
We do not consider methods applicable only to equations of some special form. Thus, we do not consider
the application of the Fourier method or the Laplace transform to solving boundary-value problems; neither do we
consider the application of the Fourier transform to solving the Cauchy problem. These questions are considered in
[35]. Neither do we describe methods of solving boundary-value problems connected with the theory of semigroups:
to do this it is necessary to use some special knowledge in the elds of semigroup theory and functional analysis.
Recently, semigroup theory has been actively developed, and there are a large number of papers in this eld,
e.g., [68].
We have tried to give our article an elementary character and to use only well-known facts of functional
analysis and function theory. As we did not want to enlarge the article, we have omitted many special ques-
tions concerning the parabolic second-order equations, as, for example, degenerated parabolic equations, equations
with a small parameter in highest terms, the qualitative behavior of solutions in unbounded domains, equations
with piecewise-continuous coecients and the corresponding conjunction conditions on the discontinuity surfaces,
the uniqueness of the Cauchy problem with initial conditions given on dierent manifolds, and many other ques-
tions. Some of these questions are treated in [917]. We also do not consider nonlinear dierential equations and
their numerous applications. These questions are considered in many papers (e.g., in review [18]).
Let us briey indicate the contents of every section of this article. In Sect. 1, we give various forms of
the maximum principle for homogeneous and nonhomogeneous equations in bounded and unbounded domains; we
pose the main boundary-value problems and the Cauchy problem and prove the uniqueness theorems for these
problems. In Sect. 2, we formulate the main a priori estimates for solutions of boundary-value problems in terms of
the data corresponding to each problem (initial or boundary conditions, coecients of the equation, the boundary
of the domain); then we give the proof of S. N. Bernsteins estimates for closed or nonclosed domains and the proof
of the main a priori estimates for the solutions, the latter also in integral norms. In Sect. 3, the solutions of the rst
and second boundary-value problems are constructed by means of the Rothe method; then we investigate how
the smoothness of the solution of the rst boundary-value problem depends on the smoothness of the problem data.
After that, the rst boundary-value problem is investigated in domains of various types with weak restrictions on
the boundary (conditions for the existence of barrier functions). In Sect. 4, we construct the fundamental solution for
the second-order parabolic equation of general form under the condition that the coecients are H older continuous.
By means of the fundamental solution we construct the solution of the Cauchy problem and the Green function for
the solution of the rst boundary-value problem. The application of integral equations to solving boundary-value
problems is given in outline. In Sect. 5, we give the denition of the generalized solution of the rst boundary-
value problem; then we prove the corresponding uniqueness theorems and give the proofs of numerous auxiliary
propositions that are used in the following sections. In Sects. 69, the generalized solution of the rst boundary-
value problem is constructed by dierent methods (in Sect. 6 by the method of nite dierences, in Sect. 7 by
436
some functional methods, in Sect. 8 by the method of continuation with respect to a parameter, and in Sect. 9 by
Galerkin method). In Sect. 10, we consider the problem of existence and uniqueness for the generalized solutions
of the Cauchy problem in classes of rapidly growing functions. In Sect. 11, we give a method for the investigation
of the smoothness properties of the solution in terms of the smoothness of the data. In Sect. 12, we prove some
simple theorems concerning the behavior of the solutions of parabolic equations as t .
This article is an extended version of the special course of lectures that O. A. Oleynik gave at the Department
of Mechanics and Mathematics of the Moscow State University in 1960.
Note also that all of the main notations and denitions are given in Sects. 1 and 5.
We discussed some sections with V. P. Mikhaylov; he also wrote Sect. 7. We are grateful for his help and
useful suggestions. We also thank E. B. Dynkin for his advice and useful comments.
1 MAXIMUM PRINCIPLE.
UNIQUENESS OF SOLUTIONS OF THE MAIN BOUNDARY-VALUE PROBLEMS
1. Some Notation. By E
n
we denote the Euclidean space of dimension n, the point coordinates being
(x
1
, . . . , x
n
); (x, t) will denote an arbitrary point of an (n + 1)-dimensional space R
n+1
= E
n
(, +); D will
denote a bounded domain in R
n+1
lying between the planes t = 0 and t = T, and

D is the closure of D. We suppose
that the intersection D
0
of

D and the plane t = 0 is not empty. By

S we denote the closure of the set of boundary
points of D having time coordinates t ,= 0 and t ,= T. The points of

S not lying on the plane t = 0 we shall denote
by S. denotes the union of S and D
0
.
In the special case where D is a cylinder (0, T) with elements parallel to the t-axis and a domain
of the plane t = 0 as a base, we shall use the notation Q instead of D. If denotes the boundary of , we have
D
0
= and = S .
The domain consisting of points lying in Q and such that their distance from S is greater than will be
called Q

will denote the domain of points in such that their distance from is greater than . Let be
the direction of the outer normal to the surface S at the points where this normal exists. We also shall consider
the layer H of points in the space R
n+1
such that 0 < t T.
In this and all the following sections, we shall study the parabolic equation
L(u)
n

i,j=1
a
ij
(x, t)

2
u
x
i
x
j
+
n

i=1
b
i
(x, t)
u
x
i
+c(x, t)u
u
t
= f(x, t), (1.1)
where the functions a
ij
, b
i
, c, and f are real, with nite values, a
ij
= a
ji
and
n

i,j=1
a
ij
(x, t)
i

j
> 0 for
n

i=1

2
i
> 0.
We shall say that the function u(x, t) satises Eq. (1.1) at a point (x, t) if u(x, t) is continuous at this point together
with its derivatives
u
xi
,
u
t
,

2
u
xi xj
(i, j = 1, . . . , n) and (1.1) is satised.
Now let us dene some classes of functions. We shall say that a function v(x, t) is of class C
m
(c) (m = 0, 1, . . .)
on a set c if at the points of c this function has continuous derivatives up to the order m. A function v(x, t) satises
on c the H older condition with exponent > 0 if the ratio
|v(P1)v(P2)|
[r(P1,P2)]

is bounded from above for any P


1
and P
2
belonging to c; here r(P
1
, P
2
) is the distance between the points P
1
and P
2
. The upper bound of this ratio is called
the H older coecient of the function v on c. A function v(x, t) belongs to the class C
m+
(c) if its derivatives
of order m are Holder continuous on c with exponent , 0 < < 1. A function v(x, t) belongs to the class
C
2m+,m+
(c) if on c it has continuous derivatives

l
u
x
l
1
1
...x
ln
n
t
l
0
with
n

i=0
l
i
= l and
n

i=1
l
i
+ 2l
0
2m, and these
derivatives are H older continuous with exponent .
A surface

S in the space R
n+1
belongs to the class A
m+
, 0 < < 1, (A
m
) if for every point P of

S there is
a sphere with center P such that inside the surface

S can be represented by the equation
x
i
= (x
1
, . . . , x
i1
, x
i+1
, . . . , x
n
, t)
for certain i n and the function is of class C
m+
(respectively C
m
). Likewise, an (n1)-dimensional surface
in E
n
belongs to the class A
m+
(A
m
) if any of its points has a neighborhood where the surface is represented by
the equation
x
i
= (x
1
, . . . , x
i1
, x
i+1
, . . . , x
n
) (1.2)
for a certain i n and the function is of class C
m+
(respectively of class C
m
).
437
2. Posing of Problems. Second-order linear parabolic equations retain many properties of the simplest
equation of this type the heat equation
n

i=1

2
u
x
2
i

u
t
= 0. (1.3)
One of the most important properties of Eq. (1.3) is the maximum principle. In this section, we shall investigate
the forms the maximum principle takes for dierent classes of second-order parabolic equations. Just as for the heat
equation, for general second-order parabolic equations the maximum principle implies the uniqueness of solutions
for the main boundary-value problems and the Cauchy problem. These problems for general parabolic equations
are posed in the same way as for the heat equation (see [19]).
Now we shall pose the main problems for the parabolic equation (1.1).
Consider the domain D described in subsection 1. The function u(x, t) is a solution of the rst boundary-
value problem for Eq. (1.1) in the domain D with conditions
u[
t=0
= (x), (1.4)
u[
S
= (x, t) (1.5)
if it is continuous in

D and satises (1.1) in

D and conditions (1.4) and (1.5) on ; here (x) and (x, t) are
given continuous functions.
Sometimes we shall consider also solutions of this problem having discontinuities at the points of (see
subsection 6).
The function u(x, t) is called a solution of the oblique derivative problem for (1.1) in the domain D with
conditions
u[
t=0
= (x),
_
u

+a(x, t)u
_
S
= (x, t) (1.6)
if it is continuous in

D and satises Eq. (1.1) in

D and conditions (1.6) on ; here (x), a(x, t), and (x, t) are
given continuous functions. It is supposed that S at every point has a tangent plane and is some direction not
parallel to the t axis and such that the angle between and the inner normal to S is acute.
If the domain D is a cylinder Q, coincides with the direction conormal to S
_
u

i,j=1
a
ij
(x, t) cos(, x
j
)
u
xi
_
,
and a(x, t) 0, the above-mentioned problem is called the second boundary-value problem for Eq. (1.1).
The Cauchy problem for Eq. (1.1) in the layer H = R
n
(0, T] with the condition
u[
t=0
= (x) (1.7)
has the function u(x, t) for solution if this function is continuous in

H and satises Eq. (1.1) in H and condition (1.7),
where (x) is a given continuous function.
Here we have given the denition of classical solutions of the main problems for Eq. (1.1). In Sects. 5 and 10
we shall dene the generalized solutions of the same problems.
3. Maximum Principle in a Bounded Domain. Now we shall study some properties of the solutions
u(x, t) of (1.1) in the domain D. These properties are generalizations of the well-known maximum principle for
the heat equation (see [19]).
Let us note that the transformation u = ve
t
, where > 0 is a constant, changes (1.1) to an equation also
having the form (1.1), and the coecient in v of the new equation is c(x, t) . If the coecient c(x, t) is bounded
from above (c(x, t) M), then for suciently large the coecient of the transformed equation is strictly negative.
In what follows, we shall repeatedly use this fact.
Theorem 1. Suppose the function u(x, t) is continuous in

D and has continuous derivatives occurring in
the operator L; suppose that L(u) 0 in

D and that c(x, t) < M, M = const. Then if u(x, t) 0 on , we
have u(x, t) 0 in

D.
Proof. Consider rst the case M < 0, that is, c(x, t) < 0. If there is a point (x, t) in

D such that u(x, t) is
negative, then the minimum of u(x, t), attained at the point (x
0
, t
0
), also is negative. According to the suppositions
of the theorem, the point (x
0
, t
0
) can be either inside D or inside the upper base t = T.
438
Therefore, at this point we have
u
xi
= 0 (i = 1, . . . , n),
u
t
0, cu > 0. Simultaneously at the point
(x
0
, t
0
) the inequality
n

i,j=1
a
ij

2
u
xi xj
0 is fullled. Indeed, after the transformation of the independent variables

i
=
n

j=1
k
ij
x
j
(i = 1, . . . , n) we obtain
n

i,j=1
a
ij

2
u
x
i
x
j
=
n

i,j=1
b
ij

2
u

j
,
where the matrices A = |a
ij
|, B = |b
ij
|, and K = |k
ij
| are connected as follows: B = KAK

. It is well-known
that we can choose a nondegenerate matrix K so that B will be diagonal at the point (x
0
, t
0
). Because we have

2
u

2
i
0 (i = 1, . . . , n), the inequality
n

i,j=1
b
ij

2
u

x=x
0
t=t
0
0
and therefore
n

i,j=1
a
ij

2
u
x
i
x
j

x=x
0
t=t
0
0.
As a result, we have L(u) 0 at the point (x
0
, t
0
); this is a contradiction proving the theorem for c(x, t) < 0.
If the coecient c(x, t) < M and M > 0 we can make the change of variables u(x, t) = v(x, t)e
Mt
. The
function v(x, t) satises an equation of the form (1.1) with a negative coecient in v and nonpositive right-hand
term. As we have already proved, v(x, t) 0 in

D, and the same is true for u(x, t) = v(x, t)e
Mt
. The proof of
the theorem is thereby completed.
Obviously, if L(u) 0 in

D and u(x, t) 0 on , and the remaining suppositions of Theorem 1 are
fullled, we have u(x, t) 0 everywhere in

D.
Remark. It follows easily from the proof of Theorem 1 that the following assertion is valid. If L(u) < 0 and
c 0 in D, then u(x, t) cannot assume a negative minimum in

D . We can also see that if L(u) > 0 and c 0
in D, u(x, t) cannot assume a positive maximum in

D .
Theorem 2. Suppose the function u(x, t) is continuous in

D and satises Eq. (1.1), where f(x, t) is a bounded
function ([f[ N) and the coecient c(x, t) 0. Suppose that [u(x, t)[

m. Then everywhere in

D we have
[u(x, t)[ Nt +m. (1.8)
Proof. Consider in

D the functions w

(x, t) = Nt +mu(x, t). These functions are nonnegative on , and


in

D , taking into account that c 0, we get
L(w

) = N +Nct +cmL(u) N +[f[ 0.


According to Theorem 1, both functions w

(x, t) are nonnegative in



D, whence (1.8) follows.
Corollary. Let the suppositions of Theorem 2 be fullled and f(x, t) 0. Then everywhere in

D
[u(x, t)[ max

[u(x, t)[.
Theorem 3. Suppose the function u(x, t) is continuous in

D and satises in

D Eq. (1.1) with [f(x, t)[ N.
Suppose that [u(x, t)[

m and c(x, t) c
0
< 0. Then everywhere in

D
[u(x, t)[ max
_
N
c
0
, m
_
.
439
Proof. Denote N
1
= max
_
N
c0
, m
_
and consider the functions w

(x, t) = N
1
u(x, t) in

D. These functions
are nonnegative on and in

D we have
L(w

) = N
1
c f N
1
c
0
+N 0.
According to Theorem 1, the functions w

(x, t) are nonnegative everywhere in



D and the assertion of the theorem
follows.
If c 0 and f 0, Theorem 2 leads to the following consequence, analogous to the well-known maximum
and minimum principle for the heat equation.
Theorem 4. Suppose that the hypotheses of Theorem 2 are valid, f 0, and c 0. Then the solution
u(x, t) attains its maximal and minimal values on , that is, everywhere in

D we have the inequalities
m
1
min

u(x, t) u(x, t) max

u(x, t) m
2
.
To prove this theorem we shall note that the functions u(x, t) m
1
and m
2
u(x, t) are nonnegative on
and satisfy Eq. (1.1). We can apply to them the corollary to Theorem 2 and get
u(x, t) m
1
max

(u(x, t) m
1
) = m
2
m
1
,
m
2
u(x, t) max

(m
2
u(x, t)) = m
2
m
1
.
Theorem 5. Let us change in Theorem 2 the condition c 0 to the condition c(x, t) M, where M > 0.
Then everywhere in

D
[u(x, t)[ e
Mt
(Nt +m). (1.9)
To prove this theorem we shall make the change of variables u = ve
Mt
and apply Theorem 2 to the function v.
From the estimate (1.9) we can deduce the uniqueness theorem for the rst boundary-value problem (1.1),
(1.4), (1.5) if the coecient c(x, t) is bounded from above.
From the same estimate it follows that the solution of the rst boundary-value problem depends continuously
on the right-hand term f(x, t), the initial function (x), and the boundary function (x, t). Indeed, it follows
from (1.9) that [u(x, t)[ e
MT
(T + 1) if [f[ < , [[ < , and [[ < , that is, [u(x, t)[ is arbitrarily small if [f[,
[[, and [[ are suciently small.
The following theorem is called the strict maximum principle.
Theorem 6. Suppose the function u(x, t) is continuous in

D and satises the equation L(u) = 0 in

D .
We shall also suppose that all the coecients of this equation are bounded in

D, c(x, t) 0 and
n

i,j=1
a
ij

j

n

i=1

2
i
for all real
i
; > 0 is some constant. Let us suppose that at some point (x
0
, t
0
)

D the function u(x, t)
attains a positive maximum: u(x
0
, t
0
) = max

D
u(x, t) = m > 0. Then u(x, t) = m at every point (x, t)

D such that
t < t
0
and that can be connected with the point (x
0
, t
0
) by a continuous curve of the form x = x(t), lying entirely
in

D .
Proof. Let us suppose the contrary. Suppose at some point P
1
(x
1
, t
1
) that can be connected with the point
P
0
(x
0
, t
0
), t
1
< t
0
, by a continuous curve x = x(t) lying in

D the value of u(x, t) is less than m. We can
connect the points (x
1
, t
1
) and (x
0
, t
0
) by a broken line contained in

D with vertices P
1
(x
1
, t
1
), P
2
(x
2
, t
2
),. . . ,
P
k
(x
k
, t
k
), P
0
, where t
1
< t
2
< . . . < t
k
< t
0
. If we prove that the inequality u(x
s
, t
s
) < m leads to the inequality
u(x
s+1
, t
s+1
) < m (s = 1, 2, . . . , k 1) then, passing from vertex P
s
(x
s
, t
s
) to the vertex P
s+1
(x
s+1
, t
s+1
) we shall
be able to establish the inequality u(x
0
, t
0
) < m. The contradiction to the supposition that u(x
0
, t
0
) = m will prove
the theorem.
To simplify the notations we shall suppose that s = 1. So we suppose that u(x
1
, t
1
) < m
1
< m, and we shall
prove that u(x
2
, t
2
) < m.
440
We can suppose that x
1
= x
2
= 0, because we can reduce the general case to this one by the linear
transformation of variables
= t,
i
= x
i
x
1
i
(x
2
i
x
1
i
)
t t
1
t
2
t
1
(i = 1, . . . , n).
The equation L(u) = 0 fullls the conditions of the theorem also in the new variables (, ).
Let us consider the cylinder Q
1
r , t
1
t t
2
, where r =
_
n

i=1
x
2
i
_
1/2
, and the constant < 1 is
chosen so that the cylinder Q
1
lies in the domain D and on the lower base of the cylinder we have u(x, t
1
) < m
1
.
Consider the function
w(x, t) = m(mm
1
)(
2
r
2
)
2
e
(tt
1
)
u(x, t)
in the cylinder Q
1
. The constant > 0 can be chosen so that the inequality L(w) < 0 holds in Q
1
. In fact,

x
i
(
2
r
2
)
2
= 4x
i
(
2
r
2
),

2
x
i
x
j
(
2
r
2
)
2
= 8x
i
x
j
4(
2
r
2
)
ij
,
L(w) = cm(mm
1
)e
(tt
1
)
_
8
n

i,j=1
a
ij
x
i
x
j
4(
2
r
2
)
n

i=1
a
ii
4(
2
r
2
)
n

i=1
b
i
x
i
+c(
2
r
2
)
2
+(
2
r
2
)
2
_
(m m
1
)e
(tt
1
)
8r
2
+ (
2
r
2
)f
1
(x, t) +(
2
r
2
)
2
,
where f
1
(x, t) is a bounded function. As we have > 0, the function f
2
(x, t) = 8r
2
+(
2
r
2
)f
1
(x, t) +(
2
r
2
)
2
is positive in some neighborhood
1
< r of the side of the cylinder Q
1
. In the rest of the cylinder, where
r
1
, the function (
2
r
2
) is bounded from below by some positive constant and, therefore, f
2
(x, t) is positive
for suciently large . For such , L(w) < 0 in Q
1
.
It is easy to verify that the function w(x, t) is nonnegative on the side part of the boundary of Q
1
and on its
lower base. Indeed, for r = we have w = mu 0. For t = t
1
and r we get
w(x, t) = m(mm
1
)(
2
r
2
)
2
u m(mm
1
)
4
m
1
m(mm
1
) m
1
= 0.
According to Theorem 1, we have w(x, t) 0 everywhere in the cylinder Q
1
. Therefore
w(0, t
2
) = m(mm
1
)
4
e
(t
2
t
1
)
u(0, t
2
) 0.
Hence
u(0, t
2
) m(mm
1
)
4
e
(t
2
t
1
)
< m.
The theorem is proved.
Another proof of this theorem and some of its generalizations can be found in L. Nirenbergs paper [20].
4. Oblique Derivative Problem. For the parabolic Eq. (1.1) with f 0 and c 0 there is a theorem
about the derivative of a solution at a maximum point on the boundary analogous to the corresponding theorem for
an elliptic equation (see [21, 22]). We shall use this theorem to prove the uniqueness of the solution for the oblique
derivative problem (1.1), (1.6).
Theorem 7. Suppose that the function u(x, t) is continuous in

D and satises the equation L(u) = 0
in

D . Suppose the coecients of the equation satisfy the suppositions of Theorem 6. We shall also suppose that
for any point P of the surface S there exists a ball A
P
containing P, such that all its points except P lie in

D
and its radius directed to P is not parallel to the t axis.
Suppose the function u(x, t) attains its maximal value at a point P
1
(x
1
, t
1
) S. Then either the function
u(x, t) is constant in a neighborhood of the point P
1
for t t
1
or
u(P1)

< 0, where is an arbitrary direction such


that the angle between this direction and the direction from the point P
1
to the center of the ball A
P1
is acute.
(We suppose that
u

exists at the point P


1
.)
441
Proof. Let us suppose that t
1
< T. If u(x, t) is not constant in a neighborhood of P
1
for t t
1
, then,
according to Theorem 6, we have
u(P) < u(P
1
) (1.10)
for all the inner points of the ball A
P1
belonging to a suciently small neighborhood of P
1
. We shall suppose that
the origin is at the center of the ball, at the point A
P1
, and that the radius R of the ball is so small that everywhere
in the ball the inequality (1.10) holds and A
P1


D. Consider the function
v(x, t) = e
(r
2
+t
2
)
e
R
2
, where r =
_
n

i=1
x
2
i
_
1/2
.
It is easy to see that
L(v) =
_
2
_
n

i=1
a
ii
t +
n

i=1
b
i
x
i
_
+ 4
2
n

i,j=1
a
ij
x
i
x
j
+c
_
e
(r
2
+t
2
)
ce
R
2
> 0 (1.11)
for r > and for suciently large > 0.
Let us consider the domain bounded by the part of the boundary of the ball A
P1
containing the point P
1
and by the plane perpendicular to the radius OP
1
and situated so near to the point P
1
as to make (1.11) valid
in . Consider the auxiliary function w(P) = u(P
1
) u(P) v(P), where > 0 is a constant; this function is
nonnegative on the surface of the ball A
P1
, because on this surface we have v(P) 0 and u(P) u(P
1
). On the rest
of the boundary of (1.10) holds; so for suciently small we have w(P) 0 on the whole boundary of . As we
have L(w) = cu(P
1
) L(v) < 0, it follows from Theorem 1 that w(P) 0 everywhere in . Moreover, w(P
1
) = 0.
Therefore
w(P1)

0, whence
u(P
1
)


v(P
1
)

= 2e
(r
2
+t
2
)
_
r
2
+t
2
cos(

OP
1
, ) < 0.
The proof given above will need practically no changes if t
1
= T: only in A
P1
the points with t T have to
be considered. The theorem is proved.
The uniqueness of the solution for the oblique derivative problem follows from Theorem 7 if the coecients
of Eq. (1.1) and the surface S satisfy the conditions of this theorem and a(x, t) 0. Indeed, it is easy to show that
under the above-mentioned conditions on the surface S any point of the domain D can be connected by the curve
of the form x = x(t) with any point (x, 0) D
0
. The dierence u(x, t) between two solutions of the oblique
derivative problem is equal to zero for t = 0 and satises the equation L(u) = 0 in

D , and the condition given
on the surface S is
u

+a(x, t)u = 0 (a(x, t) 0). (1.12)


According to Theorem 6, the function u(x, t) cannot attain its maximal positive or its minimal negative value
in

D . If the maximal positive value m of u(x, t) is attained at a point P
0
S, then, by virtue of Theorem 7
and the condition (1.12), u(x, t) m in a neighborhood of the point P
0
. Then Theorem 6 leads to u[
t=0
= m. The
case of the negative minimum at the point P
0
S can be reduced to that of the positive maximum by changing
the sign of u(x, t). The contradiction we obtain proves the theorem.
Remark. If the direction lies in a plane t = const, then it is possible to prove the uniqueness of the solution
for the oblique derivative problem also without the supposition c(x, t) 0. In fact, in this case the transformation
u = ve
t
, changing the equation to one with a negative coecient in v, does not change the boundary condi-
tion (1.12). What is more, when the direction lies in a plane t = const, for the uniqueness of the solution it
is sucient to require that the coecient a(x, t) in the condition (1.12) be bounded from above. In this case,
condition (1.12) can be reduced to one with a 0 by the transformation u = vz(x, t), where the smooth positive
function z is dened on the surface S and satises the inequality
z

+az < 0. It is easy to construct such a function,


for example, when S A
2
and
cos(, ) > > 0. (1.13)
442
Under the supposition that lies in a plane t = const and (1.13) is fullled, we shall prove that the solution
of the oblique derivative problem depends continuously on the right-hand term of the equation and on the initial
and boundary functions. We shall consider a domain D with side boundary S, belonging to the class A
2
.
Let the function u(x, t) satisfy Eq. (1.1) in

D and the condition (1.6) on . Suppose that [f(x, t)[ < ,
[(x)[ < , and [(x, t)[ < . As we have remarked before, one can suppose that c(x, t) c
0
< 1 in D and that
a(x, t) 0.
Let us construct a function v(x, t) C
2
such that v[
S
= 0 and
v

S
> 1. Suppose that [L(v)[ < M
1
and
[v[ < M
1
. Consider the auxiliary functions w

= v u. The function w
+
cannot attain a positive maximal value
on the surface S, because we have
w+

+ aw
+
=
v

+ av +
_
u

+ au
_
> + > 0 on S. If the function w
+
attains its positive maximal value at a point P
0
of

D , at this point we have
w
+
(P
0
)
max [L(w
+
)[
c
0

max [L(v)[ + max [f[


c
0
<
M
1
+ 1
c
0
.
This inequality follows from Eq. (1.1) at the point P
0
. For t = 0, we have w
+
[v[ + < (M
1
+ 1). As a result
we get w
+
< (M
1
+ 1) in the domain D. Similarly we get w

< (M
1
+ 1). Therefore
[u(x, t)[ < (2M
1
+ 1),
that is, [u(x, t)[ is arbitrarily small for suciently small .
Theorems similar to those proved here can be found in [22].
5. The Maximum Principle in an Unbounded Domain and the Cauchy Problem. Now we shall
prove that the solution of the Cauchy problem for Eq. (1.1) is unique and depends continuously on the right-hand
term of (1.1) and on the initial function. First we shall prove for the unbounded domain H theorems similar to
Theorem 1. To do this we have to impose some conditions on the growth of the coecients and on the admissible
growth of the solution as
n

i=1
x
2
i
.
Theorem 8. Suppose the function u(x, t) in

H is continuous and bounded from below: u(x, t) > m,
m > 0. Suppose the function u(x, t) in H has continuous derivatives, appearing in the operator L, and the inequality
L(u) 0 holds. Suppose that the coecients a
ij
, b
i
, and c satisfy the following conditions:
[a
ij
(x, t)[ < M(r
2
+ 1), [b
i
(x, t)[ < M
_
r
2
+ 1, c(x, t) < M,
where r =
_
n

i=1
x
2
i
_
1/2
and M is a positive constant.
Under these suppositions, if u 0 for t = 0, then u(x, t) 0 everywhere in

H.
Proof. Let us consider the auxiliary function
w(x, t) =
m
r
2
0
(r
2
+Kt)e
t
+u(x, t).
It is possible to choose the constants K > 0 and > 0 so that for all r
0
> 0 the value of L(w) will be negative.
Indeed,
L(w) =
m
r
2
0
e
t
_
2
n

i=1
a
ii
+ 2
n

i=1
b
i
x
i
+cr
2
+Kct K r
2
Kt
_
+L(u)
and for r 1 we have
L(w)
m
r
2
0
e
t
[M(4n + 1)(r
2
+ 1) r
2
+ (M )Kt] < 0
if > 2M(4n + 1). For r < 1, taking into account that [a
ij
[ < 2M and [b
i
[ < 2M, we get
L(w)
m
r
2
0
e
t
[M(8n + 1) K] < 0
if K > M(8n + 1).
443
Consider now w(x, t) in the cylinder Q
r0
r r
0
, 0 t T. For t = 0 we have w(x, 0) u(x, 0) and for
r = r
0
the function w(x, t) m +u > 0, so, according to Theorem 1, everywhere in Q
r0
the inequality w(x, t) 0
holds.
Any xed point from H is contained in Q
r0
for all suciently large r
0
. At this point, according to what was
proved above, we have w(x, t) =
m
r
2
0
(r
2
+Kt)e
t
+u(x, t) 0. Passing to the limit in this inequality as r
0
, we
obtain the assertion of the theorem.
Remark. Instead of the condition u(x, t) > m in Theorem 8 it is sucient to require u(x, t) > (r)r
2
m
with lim
r
(r) = 0. Under this condition the proof undergoes practically no changes, but as the auxiliary function
w(x, t) we have to take (r
0
)(r
2
+Kt)e
t
+u(x, t).
Moreover, it is possible to prove Theorem 8 under the supposition u(x, t) > m(r
q
+1), where m and q are
some real numbers. To do this, one has to take instead of
m
r
2
0
e
t
(r
2
+Kt) another auxiliary function
2m
r
2pq
0
(r
2
+Kt)
p
e
t
with 2p > q; then we only need to repeat the proof of Theorem 8.
Theorem 9. Suppose the coecients a
ij
and b
i
are bounded in

H:
[a
ij
[ < M, [b
i
[ < M, [c[ < M,
and the function u(x, t) is continuous in

H and satises the inequalities
L(u) 0, u(x, t) e
(r
2
+1)
in H; here is a positive constant. Then if u(x, 0) 0, then u(x, t) 0 everywhere in

H.
To prove this theorem, we have to take the auxiliary function
w(x, t) = e
2(r
2
+1)e
t
(r
2
0
+1)
+u(x, t)
and repeat the reasonings we used to prove Theorem 8 with respect to the new function. Doing this, we have to
verify the inequality
L(e
2(r
2
+1)e
t
) < 0
if > 0 is chosen in a correct way. We have
L(e
2(r
2
+1)e
t
) = e
2(r
2
+1)e
t
+t
_
4
n

i=1
a
ii
+ 16
2
e
t
n

i,j=1
a
ij
x
i
x
j
+ 4
n

i=1
b
i
x
i
+ce
t
2(r
2
+ 1)
_
< 0
for t
1

, if = M
_
1

+8n+48n
_
. So we get u(x, t) 0 for 0 t
1

. We can repeat this argument for the layer


1

t
2

, then for
2

t
3

and so on; in this way we establish that u(x, t) 0 everywhere in



H.
Theorem 10. Suppose that u(x, t) is continuous and bounded in

H and satises in H Eq. (1.1) and
that the inequalities [u(x, 0)[ M
1
, [f(x, t)[ M
2
, and c(x, t) M
3
hold. The coecients a
ij
and b
i
satisfy
the suppositions of Theorem 8. Then everywhere in

H
[u(x, t)[ e
M3t
(M
1
+M
2
t). (1.14)
To prove this theorem we shall consider the auxiliary functions w

= e
M3t
(M
1
+ M
2
t) u. According to
the suppositions of the theorem, w

(x, 0) 0. We have
L(w

) = e
M3t
[(c M
3
)(M
1
+M
2
t) M
2
] f M
2
e
M3t
f 0.
By virtue of Theorem 8, w

0 everywhere in

H and therefore (1.14) holds.
If only bounded solutions of (1.1) are considered and the coecients of (1.1) are subject to the suppositions
of Theorem 8, then the uniqueness of the solution of the Cauchy problem for (1.1) and the continuous dependence
of this solution on the initial function (x) and the right-hand term f(x, t) of the equation follow immediately from
Theorem 10. Moreover, under the above-mentioned conditions on the coecients of the equation it follows from
444
the remark to Theorem 8 that the solution of the Cauchy problem is unique in the class of functions growing not
faster than a degree of r as r .
When the coecients of Eq. (1.1) are bounded, it follows from Theorem 8 that the solution of the Cauchy
problem is unique in the class of functions growing not faster than e
r
2
, > 0, as r .
The results stated below show that there is no uniqueness for the solutions of the Cauchy problem in the wider
class of functions characterized by the inequality
[u(x, t)[ < e
(r
2+
+1)
, > 0. (1.15)
This means that there are solutions of Eq. (1.3) dierent from identical to zero and satisfying the condition u(x, 0)=0
and the inequality (1.15).
It is impossible to weaken the conditions imposed on the coecients of Eq. (1.1) in Theorem 8, remaining
in classes of estimates of the coecients characterized by degrees of r. Indeed, for any > 0 the function
u(x, t) =
_

_
+
_
F

(x,t)
e
y
2
dy for 0 < t T,
0 for t = 0,
where the function
F

(x, t) =
(

x
2
+ 1 +x)

t
is continuous and bounded in

H, vanishes for t = 0 and for t > 0 satises the equation
1

2
(x
2
+ 1)(
_
x
2
+ 1 x)
2

2
u
x
2
+
1

2
(x
_
x
2
+ 1)(
_
x
2
+ 1 x)
2
u
x

u
t
= 0.
In this equation, the coecient of

2
u
x
2
does not grow faster than M[x[
2+2
and the coecient of
u
x
does not grow
faster than M[x[
1+2
. Therefore, for coecients having such growth there is no uniqueness for solutions in the class
of bounded functions.
The questions concerning the uniqueness of solutions of the Cauchy problem for parabolic equations and for
systems parabolic according to I. G. Petrovskiy were studied in many papers (see [3]).
For the heat equation (1.3) there are stronger results than Theorem 9. These results are due to E. Holm-
gren [23], A. N. Tikhonov [24], and S. Tacklind [25]. It is proved that the solution of the Cauchy problem for
the heat equation is unique in the class of functions satisfying the condition
[u(x, t)[ < e
|x|h(|x|)
([x[ > 1), (1.16)
where is an arbitrary positive constant and h(r) is a positive nondecreasing function such that

_
1
dr
h(r)
= . (1.17)
One can construct examples to show that if the integral (1.17) converges, then a solution of the Cauchy problem
for the heat equation is not unique in the class of functions satisfying (1.16) (see [24, 25]).
For the parabolic systems the uniqueness theorems for the Cauchy problem in classes of growing functions
is studied in [26, 27] and some others. For parabolic equations and systems with growing coecients the questions
of the uniqueness were studied in [28].
Remark. For Eq. (1.1), one can also study boundary-value problems in unbounded domains. Let Q

(0, T), where

is a complement of a bounded domain in the space E


n
. The uniqueness of a bounded
solution for the rst and second boundary-value problems, as well as for the oblique derivative problem for Eq. (1.1)
in Q

can be proved similarly to the case of a bounded domain, but the choice of the auxiliary function, growing
as r , is dierent (see, for example, [29]).
445
6. Some Generalizations of Uniqueness Theorems. One can pose the rst boundary-value problem
and the Cauchy problem for Eq. (1.1) in more general terms than we did in Sect. 1. We can suppose that the initial
condition u(x, 0) = (x) holds for all x except a certain set c, lying in D
0
. The function (x) can have discontinuities
at the points of c.
The following theorems state the conditions under which the uniqueness theorems hold.
Theorem 11. Suppose the set c D
0
has the following property: there exists a positive number < n
such that for every given > 0 the set c can be covered by a nite number of n-dimensional balls K
1
, . . . , K
N
with
radii
1
, . . . ,
N
, respectively, and
N

s=1

s
< . (1.18)
Suppose that the function u(x, t) is dened, continuous, and bounded in

D c, satises in

D the equation
L(u) = 0, (1.19)
and on the boundary of D satises the conditions
u(x, t) = 0 on S, (1.20)
u(x, 0) = 0 for x D
0
c. (1.21)
The coecients b
i
and c of the Eq. (1.19) are assumed to be bounded in

D, and the coecients a
ij
are assumed to
be continuous in

D and satisfy the inequality
n

i,j=1
a
ij
(x, t)
i

j

n

i=1

2
i
, > 0. (1.22)
Under these suppositions, u(x, t) 0 everywhere in

D c.
Proof. We can choose a point ( x, 0) D
0
and consider the cylinder Q

x,
[x x[ < , 0 < t <
2
, where
[x x[ =
_
n

i=1
(x
i
x
i
)
2
_
1/2
and > 0 is suciently small. Suppose that Q
x,
= Q

x,
D. Consider the auxiliary
function
v(x, t; x, ) =
1
(t +
2
)
/2
e

i,j=1
a
ij
( x,0)(x
i
x
i
)(x
j
x
j
)
4(t+
2
)
. (1.23)
Here a
ij
(x, t) denotes the element of a matrix inverse to |a
ij
(x, t)|; is the same as in (1.18); is a xed number
satisfying the inequalities

n
< < 1. (1.24)
We shall show that in a domain Q
x,
, where

> 0 and does not depend on x and , the inequality


L(v(x, t; x, )) < 0
holds.
Dene a
kl
= a
kl
( x, 0), a
kl
= a
kl
( x, 0), and
k
= x
k
x
k
(k, l = 1, . . . , n). We have
v
x
k
=

i=1
a
ki

i
2(t +
2
)
v,

2
v
x
k
x
l
=
_

a
kl
2(t +
2
)
+

2
4(t +
2
)
2
n

i,j=1
a
ki
a
lj

j
_
v,
v
t
=
_


2(t +
2
)
+

4(t +
2
)
2
n

i,j=1
a
ij

j
_
v,
446
L(v) =
n

k,l=1
a
kl

2
v
x
k
x
l

v
t
+
n

k,l=1
(a
kl
a
kl
)

2
v
x
k
x
l
+
n

k=1
b
k
v
x
k
+cv
=
v
2(t +
2
)
_

k,l=1
a
kl
a
kl
+

+

2(t +
2
)
n

k,l,i,j=1
a
kl
a
ki
a
lj

j

1
2(t +
2
)
n

i,j=1
a
ij

j

n

k,l=1
(a
kl
a
kl
) a
kl
+

2(t +
2
)
n

i,j,k,l=1
(a
kl
a
kl
) a
ki
a
lj

j

n

i,k=1
b
k
a
ki

i
+
2c(t +
2
)

_
. (1.25)
Without loss of generality, we can suppose that c(x, t) 0. Since
n

k=1
a
kl
a
ki
=
li
and the quadratic form
n

i,j=1
a
ij

j
is positive (this follows from (1.22)), from (1.25) we deduce
L(v) =
v
2(t +
2
)
_

_
n

(1 )
2(t +
2
)
n

i,j=1
a
ij

j

n

k,l=1
(a
kl
a
kl
) a
kl
+

2(t +
2
)
n

i,j,k,l=1
(a
kl
a
kl
) a
ki
a
lj

j

n

i,k=1
b
k
a
ki

i
+
2c(t +
2
)

v
2(t +
2
)
_

_
n

1
(1 )
[x x[
2
t +
2
+M
1
(

) +M
2
(

)
[x x[
2
t +
2
+M
3

_
. (1.26)
Here (

) = max [a
kl
(x, t)a
kl
(x

, t

)[ for (x, t)

D, (x

, t

) D, [xx

, [tt

[
2
(k, l = 1, . . . , n),
and the constants
1
> 0 and M
i
> 0 (i = 1, 2, 3) do not depend on

, , and x. The inequalities (1.24) and (1.26)


show that L(v(x, t; x, )) < 0 in Q
x,
, where

depends on the modulus of continuity () of the coecients a


ij
and on the upper bounds of [b
i
[ in

D but does not depend on x and .
In D Q
x,
the function v(x, t; x, ) is bounded uniformly in x and together with all its derivatives
appearing in L(v). Therefore [L(v(x, t; x, ))[ < M
4
, M
4
> 0 for (x, t) D Q
x,
, and M
4
does not depend on x.
As a consequence, for the function
z(x, t; x, ) = v(x, t; x, ) +M
4
t (1.27)
we have everywhere in D the inequality
L(z(x, t; x, )) < 0. (1.28)
By virtue of (1.23) and (1.27), we can obtain an estimate from below for z(x, t; x, ) in the domain Q
x,
:
z(x, t; x, )
1
(2
2
)
/2
e

M
5
|x x|
2

. (1.29)
Here M
5
> 0 and
2
> 0 do not depend on x and .
Now suppose that M = sup

D
[u(x, t)[ and (x
0
, t
0
) is an arbitrary point in

D. Let us take an arbitrary small
number > 0 and prove that [u(x
0
, t
0
)[ < .
From (1.23) and (1.27) one can deduce that
z(x
0
, t
0
; x, ) M
6
, (1.30)
where M
6
does not depend on x and (the point (x
0
, t
0
) remains xed). Let us cover the set c by n-dimensional
balls K
1
, . . . , K
N
such that
N

s=1

s
<

2
MM
6
; (1.31)
here
s
is the radius of the ball K
s
and x
s
= (x
s
1
, . . . , x
s
n
) is its center, s = 1, . . . , N. We shall also assume that

s
<

t
0
(s = 1, . . . , N).
447
Set D

= D
N

s=1
Q
x
s
,s
;

will denote the set of boundary points of D

such that t < T. Compare on


the set D

the function u(x, t) with the function


w(x, t) =
M

2
N

s=1

s
z(x, t; x
s
,
s
),
where z(x, t; x
s
,
s
) (s = 1, . . . , N) are dened as in (1.27) and (1.23). Everywhere on

we have
[u(x, t)[ w(x, t). (1.32)
In fact, on S and on D
0
c, (1.32) follows from the conditions (1.20), (1.21). If (x, t)

Q
x
s
1,s
1
for a certain s
1
,
1 s
1
N, then according to (1.29) we get
w(x, t)
M

s1
z(x, t; x
s1
,
s1
) M [u(x, t)[.
By virtue of (1.28) we have L(wu) < 0 everywhere in D

. Applying Theorem 1 to w+u and wu, we can


see that the inequality (1.32) is valid everywhere in D

, and particularly in the point (x


0
, t
0
). Taking into account
(1.30) and (1.31), we get from (1.32)
[u(x
0
, t
0
)[ w(x
0
, t
0
)
M
6
M

2
N

s=1

s
< .
As the point (x
0
, t
0
) and the number are arbitrary, the statement of Theorem 11 follows.
Remark 1. If the boundary of the domain D
0
is smooth, Theorem 11 is also valid for c = .
Remark 2. If L(u) 0 in

D , u(x, t) 0 everywhere in c, and the other suppositions of Theorem 11
hold too, then u(x, t) 0 everywhere in

D c. The proof of this assertion is similar to that of Theorem 11.
Theorem 12. Suppose c is a set in the plane t = 0 having the following property: there exists a positive
number < n such that for any given > 0 any bounded domain c
1
c can be covered by a nite number
of n-dimensional balls K
1
, . . . , K
N
of radii
1
, . . . ,
N
, respectively, so that the inequality (1.18) holds. Suppose
the function u(x, t) is dened, continuous, and bounded in

H c and satises Eq. (1.19) in H and the initial
condition (1.21) for t = 0 outside the set c. Suppose also that all the coecients of Eq. (1.19) are bounded in

H,
the coecients a
ij
are uniformly continuous in

H, and the inequality (1.22) holds in

H.
Under these suppositions, u(x, t) 0 everywhere in

H c.
The proof of this theorem is similar to that of Theorem 11. As in Theorem 8, we only need to use, besides
the auxiliary function z(x, t; x, ), another auxiliary function, growing as r : w
1
(x, t) = (r
2
+Kt)e
t
.
2 A PRIORI ESTIMATES
1. S. N. Bernsteins a priori Estimates. In the study of boundary-value problems and the properties of
solutions of partial dierential equations, a priori estimates of solutions in various norms are of great importance.
An estimate is called an a priori estimate if it is obtained under the supposition of the existence of the solution
for the problem considered and depends on the data of the problem. That is, it depends on the coecients of
the equation and its right-hand term, on the domain D, and on the initial and boundary data. In Theorems 2 and 5
of Sect. 1 (theorems on the generalization of the well-known maximum principle for the heat equation), a priori
estimates of the absolute value of a solution of the rst boundary-value problem are established in terms of the data
of the problem considered. In this section, we shall establish the most important a priori estimates for the solutions
and their derivatives for the second-order parabolic equations.
S. N. Bernstein rst pointed out the important part of a priori estimates in the study of solutions for
boundary-value problems in his papers on the elliptic equations (see, e.g. [30, 31]). In these papers, he gave methods
of obtaining a priori estimates for the solutions. Here the method of auxiliary functions is very important. This
method became widespread recently in the study of second-order elliptic and parabolic equations. S. N. Bernsteins
methods were used in many works (see, e.g., [18, 3236]). The following theorems of Bernstein, giving an a priori
estimate for derivatives of a solution of Eq. (1.1), are proved by the method of auxiliary functions.
448
Theorem 1 (see [37]). Suppose the function u(x, t) satises Eq. (1.1) at the points of the cylinder Q =
(0, T), where is a bounded domain in the space E
n
with the boundary . Suppose that [u(x, t)[ M in

Q,
the coecients a
ij
, b
i
, and c and the right-hand term f of Eq. (1.1) are continuous in

Q together with their
derivatives of the rst order with respect to x
1
, . . . , x
n
, and
n

i,j=1
a
ij
(x, t)
i

j

n

i=1

2
i
, > 0. (2.1)
Suppose also that the function u(x, t) in

Q has continuous derivatives

3
u
xixjx
k
and

2
u
x
k
t
(i, j, k = 1, . . . , n).
Then in any cylinder

Q

= Q

t T, where > 0 is an arbitrarily small number, the inequality


n

k=1
_
u
x
k
_
2
M
1
(2.2)
holds. Here M
1
depends only on , M, and and on the maximum of the absolute values of the coecients of (1.1),
the right-hand term f, and their derivatives with respect to x
1
, . . . , x
n
in

Q
/2
.
Proof. Assume that c(x, t) 0. As we have remarked in subsection 3 of Sect. 1, this assumption can be
made without loss of generality. It is sucient to prove the inequality (2.2) for the cylinder K
/4
[, T], where
K
/4
is an n-dimensional ball of radius /4, lying together with its boundary inside
/2
, because the set

can
be covered by a nite number of such balls. Moreover, one can assume that the concentric to K
/4
ball K
/2
of
radius /2 also lies in
/2
.
For simplicity we shall assume that the center of the ball K
/2
is at the origin. Consider in the cylinder
G

= K
/2

2
, T
_
the auxiliary function
w(x, t) = (x)
n

k=1
_
u
x
k
_
2
+Nu
2
+e
N1(T)
, (2.3)
where
= t

2
, (x) =
_

2
4

n

i=1
x
2
i
_
2
.
Note the following property of (x):
_

x
i
_
2
M
2
for x K
/2
(i = 1, . . . , n); (2.4)
here M
2
> 0 is a constant.
We shall show that for a correct choice of the constants N and N
1
the function w(x, t) in

G

attains its
maximal value either on t =

2
or on
n

i=1
x
2
i
=

2
4
. Taking into account the remark to Theorem 1 of Sect. 1, we can
see that it is sucient to prove that everywhere in G

we have
L(w) > 0. (2.5)
To calculate L(w) we use the formula
L(v
1
v
2
) = v
1
L(v
2
) +v
2
L(v
1
) + 2
n

i,j=1
a
ij
v
1
x
i
v
2
x
j
cv
1
v
2
.
We have
L(w) = (x)
_
2
n

k=1
u
x
k
L
_
u
x
k
_
+ 2
n

i,j,k=1
a
ij

2
u
x
k
x
i

2
u
x
k
x
j
c
n

k=1
_
u
x
k
_
2
_
+ [L() ]
n

k=1
_
u
x
k
_
2
+ 4
n

i,j,k=1
a
ij

x
i
u
x
k

2
u
x
k
x
j
+ 2NuL(u)
+ 2N
n

i,j=1
a
ij
u
x
i
u
x
j
c
_
(x)
n

k=1
_
u
x
k
_
2
+Nu
2
_
+ (N
1
+c)e
N1(T)
. (2.6)
449
Dierentiating Eq. (1.1) with respect to x
k
, we get
L
_
u
x
k
_
=
n

i,j=1
a
ij
x
k

2
u
x
i
x
j

i=1
b
i
x
k
u
x
i

c
x
k
u +
f
x
k
(k = 1, . . . , n).
In the right-hand side we can omit members that are obviously positive and obtain
L(w) 2
n

i,j,k=1
a
ij
x
k
u
x
k

2
u
x
i
x
j
2
n

i,k=1
b
i
x
k
u
x
k
u
x
i
2
n

k=1
c
x
k
u
u
x
k
+ 2
n

k=1
f
x
k
+ 2
n

i,j,k=1
a
ij

2
u
x
k
x
i

2
u
x
k
x
j
+ [L() ]
n

k=1
_
u
x
k
_
2
+ 4
n

i,j,k=1
a
ij

x
i
u
x
k

2
u
x
k
x
j
+ 2Nuf + 2N
n

i,j=1
a
ij
u
x
i
u
x
j
+ (N
1
+c)e
N1(T)
. (2.7)
Then we use the elementary inequality
[ab[
1
2
_
a
2
+
1

b
2
_
, (2.8)
where > 0 is an arbitrary positive number. One can estimate the rst term in the right-hand side of (2.7) by
means of this inequality in the following way:
2
n

i,j,k=1
a
ij
x
k
u
x
k

2
u
x
i
x
j
2M
3

i,j,k=1

u
x
k

2
u
x
i
x
j

nM
3

i,j=1
_

2
u
x
i
x
j
_
2

n
2
M
3

k=1
_
u
x
k
_
2
.
We denote by M
i
, i = 3, . . ., the constants depending on the same data as M
1
.
In a similar way we can estimate the remaining terms in (2.7). Taking additionally into account the condi-
tion (2.1), we obtain the following inequality:
L(w) 2
_
M
4
M
5
1

k=1
_

x
k
_
2
+
_
n

i,j=1
_

2
u
x
i
x
j
_
2
+
_
2N
M
6

M
7
_
n

k=1
_
u
x
k
_
2
+N
1
+c M
8
N M
9
. (2.9)
Let us take > 0 so small that the coecient of
n

i,j=1
_

2
u
xi xj
_
2
in the right-hand side of (2.9) is positive
everywhere in G

; this can be achieved by virtue of (2.4). Having xed , we can choose N so large that the coecient
of
n

k=1
_
u
x
k
_
2
in (2.9) becomes positive. Then we choose N
1
so large as to make positive the term in braces in (2.9).
We have proved the inequality (2.5). As we have already noted, it follows that w(x, t) attains its maximum
in

G

either on t =

2
or on
n

l=1
x
2
l
=

2
4
. Therefore, w(x, t) NM
2
+ e
N1T
everywhere in

G

. Moreover, in

G

the inequality
_
t

2
_
(x)
n

k=1
_
u
x
k
_
2
NM
2
+e
N1T
holds. In particular, for the points of the cylinder K
/4
[, T] we obtain
n

k=1
_
u
x
k
_
2

NM
2
+e
N1T
_
t

2
_
(x)

NM
2
+e
N1T

2
_
3
2
16
_
2
= M
1
,
which was to be proved.
450
Remark. By the same method it is possible to estimate the higher-order derivatives of the function u(x, t).
We only have to suppose that the function u(x, t) and the coecients and the right-hand term of (1.1) have a higher
degree of smoothness. So to obtain estimates for the derivatives

2
u
xi xj
(i, j = 1, . . . , n) in the cylinder

Q

it is
sucient to suppose that in

Q dierential operators

2
x
k
x
l
(k, l = 1, . . . , n) can be applied to Eq. (1.1) and to
use instead of function (2.3) the auxiliary function
w
1
(x, t) = (x)
n

k,l=1
_

2
u
x
k
x
l
_
2
+N
2
n

k=1
_
u
x
k
_
2
+e
N3(T)
.
The rest of the reasoning undergoes no changes. To obtain estimates for the derivatives

p
u
x
p
1
1
...x
pn
n
it is sucient
to suppose that in

Q one can dierentiate (1.1) p times with respect to x
1
, . . . , x
n
.
After we have obtained in

Q

estimates for the derivatives of u(x, t) of the rst and second order with
respect to x
1
, . . . , x
n
, the estimate for
u
t
in

Q

follows from Eq. (1.1). Similarly, if we intend to get estimates


for the higher-order derivatives of u(x, t) containing the dierentiation with respect to t, we have to use equations
obtained from (1.1) by dierentiation. To obtain the estimates for the derivatives

p+q
u
x
p
1
1
...x
pn
n
t
q
, where p + 2q r
(r 2), it is sucient to suppose that the dierential operator

k+l
x
k
1
1
...x
kn
n
t
l
with k + 2l r, l
_
r
2

1, can be
applied to (1.1) in

Q .
Applying Bernsteins method, we can establish a priori estimates also in a closed domain.
Theorem 2. Suppose the function u(x, t) is a solution of the rst boundary-value problem for Eq. (1.1) in
the cylinder Q = (0, T) with the initial and boundary conditions
u(x, 0) = (x), u(x, t)[
S
= (x, t). (2.10)
The coecients and the right-hand side of (1.1) are supposed to be continuous in

Q together with their rst-order
derivatives with respect to x
1
, . . . , x
n
, and the inequality (2.1) holds. The boundary of the domain belongs to
the class A
2
. Suppose also that the derivatives
u
x
k
(k = 1, . . . , n) are continuous in

Q, and in

Q continuous
derivatives

3
u
xixjx
k
,

2
u
x
k
t
(i, j, k = 1, . . . , n) exist.
Under these suppositions, everywhere in

Q the inequality (2.2) holds with M
1
depending only on the do-
main Q, the constant , on M = max

Q
[u(x, t)[, and on the maximum of the absolute values of the following functions:
a
ij
, b
i
, c, f, , their rst-order derivatives with respect to x
1
, . . . , x
n
, the function (x, t), and its rst- and second-
order derivatives.
Proof. Without loss of generality, we can assume that c(x, t) 0 (as in Theorem 1). We shall estimate
n

k=1
_
u
x
k
_
2
for (x, t)

S. According to our supposition, the boundary of the domain can be covered by a nite
number of overlapping pieces
p
such that each of them can be represented in the form
x
i
=
i
(
1
, . . . ,
n1
) (i = 1, . . . , n) (2.11)
with
i
C
2
. Therefore, in
p
one can introduce a new coordinate system by the formulas x
i
=
i
(
1
, . . . ,
n
),

i
=

i
(x
1
, . . . , x
n
) (i = 1, . . . , n) so that
p
in these coordinates is dened by the equation
n
= 0, for the points
of the inequality
n
> 0 holds, and
i
C
2
and

i
C
2
.
We shall construct functions
p
(
1
, . . . ,
n1
), having the following properties:
1
2

p
1,
p
=
1
2
on
the piece
p
and
p
1 in a certain subdomain
p
of
p
. The subdomains
p
are chosen so that they still cover .
In the cylinder Q
p
_
(
1
, . . . ,
n1
)
p
, 0
n

1
N
, 0 t T
_
we shall consider an auxiliary function
z(, t) = z(
1
, . . . ,
n
, t) = u(
1
, . . . ,
n1
,
n
, t) (
1
, . . . ,
n1
, t) +N
1
e
Nn

p
(
1
, . . . ,
n1
), (2.12)
where the constants N > 0 and N
1
> 0 will be chosen later. Calculate L(z) in the variables
1
, . . . ,
n
, t:
L(z) =
n

i,j,k=1

ij
(, t)

2
z

j
+
n

i=1

i
(, t)
z

i
+c(, t)z
z
t
= L(u) L() +N
1
e
Nn
L(
p
) +L(N
1
e
Nn
)
p
2NN
1
e
Nn
n

j=1

nj

j
N
1
ce
Nn

p
.
451
From these formulas we can deduce that
L(z) = f L() +N
1
e
Nn
_
L(
p
) +N
2

nn

p
N
n

p
2N
n

j=1

nj

j
_
. (2.13)
According to (2.1) we have
nn

> 0 everywhere in Q
p
; moreover,
p

1
2
. Therefore, for suciently
large N > 0 the expression in square brackets on the right-hand term of (2.13) becomes positive. Having xed
such N, we can choose N
1
> 0 so large that the whole right-hand term of (2.13) becomes positive.
Because L(z) > 0, according to the remark to Theorem 1 of Sect. 1, the function z(, t) cannot attain its
maximal value in

Q
p
in

Q, neither inside of Q
p
nor on t = T. We shall show that for suciently large N
1
> 0
the function z(, t) takes its maximal value at points belonging to

S
p
=
p
[0, T].
It follows from (2.12) that z(, t) = N
1
at the points of

S
p
and z(, t) = N
1

p
N
1
at the points of S
p
=

p
[0, T]. If 0
n

1
N
and (
1
, . . . ,
n1
) belongs to the boundary of the piece
p
, then z(, t) 2M +
N1
2
< N
1
if N
1
> 4M. For
n
=
1
N
we have z(, t) 2M +
N1
e
< N
1
if N
1
> 4M. Finally, also for t = 0 the function z(, t)
does not exceed N
1
if N
1
is suciently large. In fact,
z(, 0)

n
=
()

n
NN
1
e
Nn

p
max

()

1
2e
NN
1
< 0
for N
1
> 0 suciently large.
So for a correct choice of N and N
1
the function z(, t) takes its maximal value on

S
p
, and there it is
identically equal to N
1
. Therefore, at all the points of

S
p
we have the inequality
z
n
0, and this leads to
u

Sp
N
1
N. (2.14)
Considering the auxiliary function
z
1
(, t) = u(
1
, . . . ,
n1
,
n
, t) (
1
, . . . ,
n1
, t) N
1
e
Nn

p
(
1
, . . . ,
n1
)
and reasoning in a similar way, we can establish the inequality
u

Sp
N
1
N. (2.15)
From (2.14) and (2.15) we can deduce that the estimate

N
1
N
holds on

S
p
. Since the derivatives of u with respect to
1
, . . . ,
n1
on

S
p
coincide with the corresponding derivatives
of the function , we have proved the validity of (2.2) on

S
p
and therefore everywhere on S.
Consider now in the cylinder

Q the function
w(x, t) =
n

k=1
_
u
x
k
_
2
e
N2t
+e
N2(Tt)
,
where the constant N
2
> 0 will be chosen later. Let us compute L(w):
L(w) = e
N2t
_
2
n

i,j,k=1
a
ij

2
u
x
k
x
i

2
u
x
k
x
j
2
n

i,j,k=1
a
ij
x
k
u
x
k

2
u
x
i
x
j
2
n

i,k=1
b
i
x
k
u
x
i
u
x
k
2u
n

k=1
c
x
k
u
x
k
+ 2
n

k=1
f
x
k
u
x
k
+ (N
2
c)
n

k=1
_
u
x
k
_
2
+ (N
2
+c)e
N2T
_
. (2.16)
452
Taking into account (2.8), from (2.16) we get
L(w) e
N2t
_
(2 M
2
)
n

i,k=1
_

2
u
x
i
x
k
_
2
+
_
N
2

M
3

M
4
_
n

k=1
_
u
x
k
_
2
+ (N
2
M
5
)e
N2T
M
5
_
. (2.17)
First we take <
2
M2
, then choose N
2
>
M3

+M
4
+M
5
. According to (2.17), we have L(w) > 0. Therefore w(x, t)
cannot have a maximum in

Q .
This means that the maximum is attained on t = 0 or on S, and according to what was already proved, we
have w(x, t) M

1
, where M

1
depends on the same quantities as M
1
in the conditions of the theorem.
Hence
n

k=1
_
u
x
k
_
2
M

1
e
N2T
= M
1
.
The theorem is proved.
Theorem 3. Suppose the function u(x, t) is a solution of the rst boundary-value problem (1.1), (2.10) in
the cylinder Q. The coecients a
ij
, b
i
, and c and the function f are continuous in

Q together with their derivatives
of the rst and second order with respect to x
1
, . . . , x
n
and the inequality (2.1) holds. The surface belongs to
the class A
3
. The derivatives
u
t
,
u
x
k
,

2
u
x
k
x
l
(k, l = 1, . . . , n) are continuous in

Q, and in

Q the derivatives

4
u
xixjx
k
x
l
,

3
u
x
k
x
l
t
(i, j, k, l = 1, . . . , n) exist.
Under these suppositions, everywhere in

Q we have
n

k,l=1
_

2
u
x
k
x
l
_
2


M
1
, (2.18)
where

M
1
depends only on the domain Q, the constant , and the maximum of absolute values of the following
functions: , a
ij
, b
i
, c, f, their rst and second derivatives with respect to x
1
, . . . , x
n
, , and its derivatives up to
the third order inclusively.
Proof. We shall use the notations
M
Q
= max

Q
_
n

k,l=1
_

2
u
x
k
x
l
_
2
_
1/2
, M

= max

_
n

k,l=1
_

2
u
x
k
x
l
_
2
_
1/2
.
Consider in

Q the following auxiliary function:
w(x, t) = e
Nt
n

k,l=1
_

2
u
x
k
x
l
_
2
+e
N(Tt)
.
For the correct choice of N > 0 this function attains its maximal value on . We have
L(w) = e
Nt
_
2
n

i,j,k,l=1
a
ij

3
u
x
k
x
l
x
i

3
u
x
k
x
l
x
j
+ 2
n

k,l=1

2
u
x
k
x
l
_
n

i,j=1
a
ij

4
u
x
i
x
j
x
k
x
l
+
n

i=1
b
i

3
u
x
i
x
k
x
l


3
u
x
k
x
l
t
+c
n

k,l=1

2
u
x
k
x
l
_
+N
n

k,l=1
_

2
u
x
k
x
l
_
2
+ (N + c)e
NT
_
.
The expression in square brackets can be transformed by means of the equation obtained from (1.1) by dierentiation
with respect to x
k
and x
l
. We shall also take into account the estimate we have obtained for
n

i=1
_
u
xi
_
2
in Theorem 2
and the inequality (2.8). We get
L(w) e
Nt
_
(2 M
2
)
n

i,k,l=1
_

3
u
x
k
x
l
x
i
_
2
+
_
N
M
3

M
4
_
n

k,l=1
_

2
u
x
k
x
l
_
2
+ (N M
5
)
_
.
453
It follows from this inequality that for <
2
M2
and N >
M3

+M
4
+M
5
the function w(x, t) cannot have a maximum
in

Q . Therefore max

Q
w M
2

+e
NT
; this leads to M
2
Q
e
NT
(M
2

+e
NT
), and hence
M
Q
M
6
M

+M
7
. (2.19)
Now we shall get an estimate for M

. As we did in the proof of Theorem 2, we can introduce in the space


(x
1
, . . . , x
n
) in the vicinity of the piece
p
a new coordinate system (
1
, . . . ,
n
), such that
p
in these coordinates
has the equation
n
= 0 and for the points of the inequality
n
> 0 holds. In the cylinder
Q
p
_
(
1
, . . . ,
n1
)
p
, 0
n

1
N
2
, 0 t T
_
we shall consider the auxiliary function
z
k
(, t) =
u(
1
, . . . ,
n1
,
n
, t)

(
1
, . . . ,
n1
, t)

k
+N
1
e
N2n

p
(
1
, . . . ,
n1
), k ,= n,
where the function
p
(
1
, . . . ,
n1
) is the same as in Theorem 2. Similarly to (2.13), we get
L(z
k
) = N
1
e
N2n
_
N
2
2

nn

p
N
2

p
2N
2
n

j=1

nj

j
+L(
p
)
_
L
_

k
_

i,j=1

ij

2
u

i=1

k
u

k
u +
f

k
,
hence
L(z
k
) > N
1
e
N2n
_

3
N
2
2
M
8
_
M
9
M
Q
M
10
. (2.20)
It follows from the inequality (2.20) that if
N
2
=
_
M
11
M
Q
+M
12
, (2.21)
then L(z
k
) > 0 everywhere inside Q
p
and also for t = T. According to the remark to Theorem 1 of Sect. 1,
the function z
k
(, t) cannot attain its maximal value in

Q
p
, neither inside Q
p
nor on t = T.
After that, as in Theorem 2, we can see that z
k
[

Sp
= N
1
,
z
k
[
Sp
N
1
, z
k
[
n=
1
N
2
2M
1
+
N
1
e
< N
1
for N
1
> 4M
1
,
and z
k
2M
1
+
N1
2
< N
1
if (
1
, . . . ,
n1
) belongs to the boundary of the piece
p
and N
1
> 4M
1
. For t = 0 we
have
z
k

t=0
=

2

k

n
N
1
N
2
e
N2n

p
max

k

n

N
1
N
2
2e
< 0
if
N
1
=
2e
N
2
max

+ 1 = M
13
. (2.22)
Therefore, the function z
k
(, t) attains its maximal value on

Q
p
at the points of

S
p
, where z
k
N
1
, and we
have
z
k

Sp
0,
hence

2
u

Sp
N
1
N
2
.
454
Similar arguments lead to

2
u

k

n

Sp
N
1
N
2
,
and we get

2
u

k

n

N
1
N
2
on

S
p
(k = 1, . . . , n 1). (2.23)
From Eq. (1.1) (in the variables
1
, . . . ,
n
, t) we obtain

2
u

2
n
=
1

nn
_

i+j<2n

ij

2
u

j
+
n

i=1

i
u

i
+cu
u
t
f
_
. (2.24)
On

S
p
we have

u
t

(, t)
t

M
14
. (2.25)
From (2.21)(2.25) one can obtain
_
n

i,j=1
_

2
u

j
_
2
_
1/2

Sp
M
15
+M
16
_
M
Q
.
Hence
_
n

i,j=1
_

2
u

j
_
2
_
1/2

S
M
17
+M
18
_
M
Q
.
Taking into account that
_
n

i,j=1
_

2
u
xi xj
_
2
_
1/2

t=0
=
_
n

i,j=1
_

2

xi xj
_
2
_
1/2
M
19
, we get
M

M
20
+M
21
_
M
Q
. (2.26)
Comparing (2.19) and (2.26) we arrive at the inequality
M
Q
M
22
_
M
Q
+M
23
,
which leads to
M
Q
M
24
.
The theorem is proved.
Corollary. Under the suppositions of Theorem 3, the estimate

u
t


M
1
holds in

Q; here

M
1
depends on
the same quantities as

M
1
.
In fact, by virtue of Eq. (1.1) the derivative
u
t
can be expressed by means of the derivatives
u
xi
and

2
u
xi xj
(i, j = 1, . . . , n) and for them we have already obtained an estimate.
By similar arguments one can obtain estimates in a closed domain

Q for the derivatives of higher order of
u(x, t) (assuming they exist); we only need to require a higher degree of smoothness of the functions a
ij
, b
i
, c, f,
, and . First, we have to estimate simultaneously the derivatives

3
u
x
k
x
l
xm
and

2
u
x
k
t
(k, l, m = 1, . . . , n); then
the derivatives

4
u
x
k
x
l
xmxq
and

3
u
x
k
x
l
t
(k, l, m, q = 1, . . . , n), and so on.
Note that from Bernsteins estimates one can deduce that a family of uniformly bounded solutions of (1.1)
is compact in the sense of uniform convergence in any closed domain contained in

Q (see Theorem 1).
Results similar to Theorems 1, 2, and 3 are valid also for elliptic second-order equations (see [31, 38]).
2. A priori Estimates in the Norms of H older Spaces. In the study of nonlinear equations, a priori
estimates independent of the smoothness of the coecients of equations play an important part (see, for exam-
ple, [18]). Such estimates for the solutions of linear equations were only recently established ([39, 40, 124]). Now we
shall formulate a theorem giving an a priori estimate for the Holder constant for the solution of a parabolic equation.
This theorem is a generalization of a theorem proved by J. Nash [39]. The proof can be found in [39] and [18].
455
Theorem 4. Suppose that u(x, t) is a solution of the equation
n

i,j=1

x
i
_
A
ij
(x, t)
u
x
j
_
+
n

i=1
B
i
(x, t)
u
x
i
+C(x, t)u
u
t
= F(x, t) (2.27)
in the domain Q, and in

Q the following inequalities hold:
[u[ M,
1
n

i=1

2
i

n

i,j=1
A
ij
(x, t)
i

j

2
n

i=1

2
i
(0 <
2
<
1
),
[B
i
(x, t)[ B (i = 1, . . . , n), [F(x, t)[ N, [C(x, t)[ C
1
, N +C
1
M = N
1
.
Then for (x
1
, t
1
) and (x
2
, t
2
) belonging to Q

, 0 < t
1
< t
2
, we have the following inequality:
[u(x
2
, t
2
) u(x
1
, t
1
)[
Amax
_
M+N
1

, (M+N
1
)B

,
M
min(

t
1
, 1)
_
[x
2
x
1
[

+Amax
_
M+N
1

2
, (M+N
1
)B
2
,
M
min(

t
1
, 1)
_
(t
2
t
1
)

;
here A, , and are constants, depending only on
1
,
2
, and n, and 0 < <
1
2
and 0 < <
1
4
.
For the solutions of linear second-order parabolic equations, a priori estimates similar to the well-known
estimates of J. Schauder for linear elliptic equations can be proved (see [41]). In the case of one space variable, such
estimates are due to C. Ciliberto [42]. For the equations with several space variables estimates of this type were
established by A. Friedman [4346]. These estimates can be useful in the study of the properties of solutions of linear
parabolic equations, such as the dependence of the smoothness of solutions on the smoothness of the coecients,
the compactness of families of solutions in various norms, and many other properties. One can use these estimates
to establish existence theorems for the solutions of the rst boundary-value problem for Eq. (1.1) (see [43]) and also
for quasilinear parabolic equations [18].
We shall use the following notations:
H(u) = sup
P,QD
[u(P) u(Q)[
d(P, Q)

, where d(P, Q) = ([x x

[
2
+[t t

[)
1/2
.
Let us introduce the norms
[u[

= sup
D
[u[ +H(u),
[u[
1+
= [u[

+
n

i=1

u
x
i

,
[u[
2+
= [u[

+
n

=1

u
x
i

+
n

i,j=1

2
u
x
i
x
j

u
t

.
The function u(x, t) belongs to the class H
j+
(D) if the norm [u[
j+
is nite, j = 0, 1, 2.
Theorem 5. Suppose the coecients of Eq. (1.1) in the domain

D satisfy the conditions
[a
ij
[

M
1
, [b
i
[

M
1
, [c[

M
1
,
n

i,j=1
a
ij

j

n

i=1

2
i
,
_

_
(2.28)
and a solution u(x, t) of Eq. (1.1) belongs to the class H
2+
(D). Then
[u[
2+
M([f[

+[[
2+
),
where the function is dened in

D and u[

= ; the constant M depends only on M


1
, , and the domain D. It
is also supposed that S belongs to the class A
2+
.
456
Let D

be a domain consisting of points of D such that their distance from the boundary is greater
than . The function u(x, t) belongs to the class K
j+
(D

) if u and its derivatives with respect to x


1
, . . . , x
n
up to
the order j belong to the class H

(D

). Dene
[u[
K
j+
(D

)
=
j

l=0

l1+...+ln=l

l
u
x
l1
1
. . . x
ln
n

.
Theorem 6. Suppose the function u(x, t) satises Eq. (1.1) and has in D Holder continuous (with expo-
nent ) derivatives with respect to x
i
of the rst and second order and also a continuous derivative with respect
to t. Let be a nonnegative integer. Then, if
[a
ij
[
K
+
(D
/2
)
M
1
, [b
i
[
K
+
(D
/2
)
M
1
, [c[
K
+
(D
/2
)
M
1
,
and (2.28) holds, we have
[u[
K
2++
(D

)
M([f[
K
+
(D
/2
)
+ max
D
/2
[u[),
where M depends only on M
1
, , and and on the size of the domain D.
The proof of Theorems 5 and 6 (in more general form) are given in [43, 46]. Note that in [45] there is
an estimate for [u[
1+
in

D in terms of max

D
[f[ when u[

= 0; the estimate depends only on the H

(D) norms of
the coecients a
ij
, on the smoothness of a
ij
on S, and on the maximal absolute values of the other coecients of
Eq. (1.1) in

D.
The proofs of these theorems are based upon the representation of the solutions of (1.1) by means of the fun-
damental solution (see Sect. 4).
3. Integral Estimates. When functional methods for solving boundary-value problems are applied, a priori
estimates in integral norms become very important. We shall use these estimates in Sects. 7 and 8. Theorems 7 and 8
can be considered as theorems on the continuous dependence for the solutions of the rst boundary-value problem
on the initial function and the right-hand term of the equation in corresponding integral norms.
Theorem 7. Suppose the function u(x, t) has continuous derivatives of the rst and second order in

Q and
satises Eq. (2.27) and the initial condition
u(x, 0) = (x).
Suppose u[
S
= 0, the coecients A
ij
C
1
(

Q) and B
i
, C, and F are bounded in Q. We also suppose that
n

i,j=1
A
ij

j

n

i=1

2
i
, > 0, and S A
1
. Then for any , 0 T, we have the inequalities
_

u
2
(x, ) dx +

_
0
_

i=1
_
u
x
i
_
2
dxdt M
1
_

_
0
_

F
2
dxdt +
_

2
(x) dx
_
, (2.29)
_

i=1
_
u
x
i
_
2

t=
dx +

_
0
_

_
u
t
_
2
dxdt M
2
_

_
0
_

F
2
dxdt +
_

2
+
n

i=1
_

x
i
_
2
_
dx
_
, (2.30)
where the constant M
1
depends only on the domain Q, on , and on the maximal absolute values of B
i
and C, and
the constant M
2
depends on the same quantities and on the maximal absolute values of A
ij
and
Aij
t
.
Proof. Multiplying both sides of Eq. (2.27) by u(x, t)e
t
and integrating over the cylinder Q

= [0, ],
we get
__
Q
e
t
u
n

i,j=1

x
i
_
A
ij
u
x
j
_
dxdt +
__
Q
e
t
u
n

i=1
B
i
u
x
i
dxdt
+
__
Q
e
t
Cu
2
dxdt
__
Q
e
t
u
u
t
dxdt =
__
Q
e
t
Fu dxdt. (2.31)
457
This equality needs some transformations. By M
3
we shall denote an expression depending on the same
quantities as M
1
, and by N
1
an expression such that [N
1
[ M
3
__
Q
_
F
2
+u
2
+[u[
n

i=1

u
xi

_
e
t
dxdt. The equal-
ity (2.31) can be rewritten in the form

__
Q
e
t
u
n

i,j=1

x
i
_
A
ij
u
x
j
_
dxdt +
__
Q
e
t
u
u
t
dxdt = N
1
. (2.32)
Let us transform the rst integral:

__
Q
e
t
u
n

i,j=1

x
i
_
A
ij
u
x
j
_
dxdt =
__
Q
e
t
n

i,j=1
A
ij
u
x
i
u
x
j
dxdt
__
Q
n

i=1
_
u
x
i
_
2
e
t
dxdt.
The second integral is equal to
__
Q
e
t
u
u
t
dxdt =
__
Q
1
2
e
t
(u
2
)
t
dxdt =
1
2
_

u
2
(x, )e

dx
1
2
_

u
2
(x, 0) dx +

2
__
Q
e
t
u
2
dxdt.
So we get from (2.32) the inequality

__
Q
n

i=1
_
u
x
i
_
2
e
t
dxdt +
e
T
2
_

u
2
(x, ) dx
1
2
_

2
(x) dx
+

2
__
Q
e
t
u
2
dxdt M
3
__
Q
F
2
e
t
dxdt + M
3
__
Q
u
2
e
t
dxdt +M
3
__
Q
[u[
n

i=1

u
x
i

e
t
dxdt.
The last integral can be estimated in the following way:
M
3
__
Q
[u[
n

i=1

u
x
i

e
t
dxdt

2
__
Q
n

i=1
_
u
x
i
_
2
e
t
dxdt +
nM
2
3
2
__
Q
u
2
e
t
dxdt.
Choose =
2nM
2
3

+ 4M
3
. Then

2
__
Q
n

i=1
_
u
x
i
_
2
e
t
dxdt +
e
T
2
_

u
2
(x, ) dx +

4
__
Q
e
t
u
2
dxdt
1
2
_

2
dx +M
3
__
Q
F
2
e
t
dxdt.
By changing e
t
in the left-hand side to e
T
and to the unity in the right-hand side, we obtain (2.29).
Let us multiply Eq. (2.26) by
u
t
and integrate over Q

. We get
__
Q
_
u
t
n

i,j=1

x
i
_
A
ij
u
x
j
_
+
n

i=1
B
i
u
x
i
u
t
+Cu
u
t

_
u
t
_
2
_
dxdt =
__
Q
F
u
t
dxdt. (2.33)
Since we have
__
Q
u
t
n

i,j=1

x
i
_
A
ij
u
x
j
_
dxdt =
__
Q
n

i,j=1
A
ij
u
x
j

2
u
x
i
t
dxdt
=
1
2
__
Q

t
_
n

i,j=1
A
ij
u
x
i
u
x
j
_
dxdt +
1
2
__
Q
n

i,j=1
A
ij
t
u
x
i
u
x
j
dxdt
=
1
2
_

_
n

i,j=1
A
ij
u
x
i
u
x
j
_
t=
dx +
1
2
_

i,j=1
A
ij
(x, 0)

x
i

x
j
dx +
1
2
__
Q
n

i,j=1
A
ij
t
u
x
i
u
x
j
dxdt,
458
we can obtain from (2.33) the inequality
1
2
_

_
n

i,j=1
A
u
x
i
u
x
j
_
t=
dx +
__
Q
_
u
t
_
2
dxdt
M
4
___
Q
_

u
t

i=1

u
x
i

u
u
t

F
u
t

+
n

i=1
_
u
x
i
_
2
_
dxdt +
_

i=1
_

x
i
_
2
dx
_

1
2
__
Q
_
u
t
_
2
dxdt +M
5
___
Q
_
u
2
+F
2
+
n

i=1
_
u
x
i
_
2
_
dxdt +
_

i=1
_

x
i
_
2
dx
_
,
where the constants M
4
and M
5
depend on the same quantities as M
2
. Therefore,

2
_

i=1
_
u
x
i
_
2

t=
dx +
1
2
__
Q
_
u
t
_
2
dxdt
M
5
___
Q
F
2
dxdt +
_

i=1
_

x
i
_
2
dx +
__
Q
_
u
2
+
n

i=1
_
u
x
i
_
2
_
dxdt
_
. (2.34)
To estimate the last integral in the right-hand side of (2.34) we shall use the inequality (2.29). Then from (2.34)
we get (2.30).
Theorem 8. Suppose that a function u(x, t) C
3
(

Q) satises Eq. (1.1) in

Q and the conditions
u(x, 0) = (x) and u[
S
= 0. Suppose also that the coecients a
ij
(x, t) are continuous, b
i
(x, t), c(x, t), and f(x, t)
are bounded in

Q, and the boundary of the domain belongs to the class A
2
. Then the following inequality holds:
__
Q
_
u
2
+
_
u
t
_
2
+
n

i=1
_
u
x
i
_
2
+
n

i,j=1
_

2
u
x
i
x
j
_
2
_
dxdt M
___
Q
f
2
dxdt +
_

2
+
n

i=1
_

x
i
_
2
_
dx
_
; (2.35)
here the constant M depends only on the modulus of continuity of the coecients a
ij
(x, t), on the least eigenvalue
of the matrix |a
ij
|, on the supremums of [a
ij
(x, t)[, [b
i
(x, t)[, and [c(x, t)[, and on the domain Q.
Proof. First we consider a particular case of this problem, when the coecients a
ij
are constant, b
i
= c = 0,
and u(x, t) vanishes in a neighborhood of S. In this case, Eq. (1.1) has the form
L
0
(u)
n

i,j=1
a
ij

2
u
x
i
x
j

u
t
= f.
We shall square both sides of the equation, multiply by e
t
, and integrate over the domain Q. The result is
__
Q
_
e
t
n

i,j,k,l=1
a
ij
a
kl

2
u
x
i
x
j

2
u
x
k
x
l
2e
t
u
t
n

i,j=1
a
ij

2
u
x
i
x
j
+e
t
_
u
t
_
2
_
dxdt =
__
Q
e
t
f
2
dxdt.
We transform the rst sum by twice integrating by parts and the second sum by integrating by parts once. Taking
into account that the function u(x, t) is zero in a neighborhood of S, we get
__
Q
e
t
_
n

i,j,k,l=1
a
ij
a
kl

2
u
x
i
x
k

2
u
x
j
x
l
+ 2
n

i,j=1
a
ij
u
x
i

2
u
x
j
t
+
_
u
t
_
2
_
dxdt =
__
Q
e
t
f
2
dxdt. (2.36)
459
The integral of the second sum can be transformed in the following way:
2
__
Q
e
t
n

i,j=1
a
ij
u
x
i

2
u
x
j
t
dxdt =
__
Q

t
_
e
t
n

i,j=1
a
ij
u
x
i
u
x
j
_
dxdt +
__
Q
e
t
n

i,j=1
a
ij
u
x
i
u
x
j
dxdt
=
_

e
t
n

i,j=1
a
ij
u
x
i
u
x
j

t=T
t=0
dx +
__
Q
e
t
n

i,j=1
a
ij
u
x
i
u
x
j
dxdt

_

e
t
n

i=1
_
u
x
i
_
2

t=T
dx
_

i,j=1
a
ij

x
i

x
j
dx +
__
Q
e
t
n

i=1
_
u
x
i
_
2
dxdt; (2.37)
here is the least eigenvalue of the matrix |a
ij
|. Note that
1
n

i,j,k,l=1
a
ij
a
kl

2
u
x
i
x
k

2
u
x
j
x
l

2
n

i,j=1
_

2
u
x
i
x
j
_
2
. (2.38)
From (2.36), (2.37), and (2.38) we deduce that

2
__
Q
n

i,j=1
_

2
u
x
i
x
j
_
2
e
t
dxdt +
__
Q
n

i=1
_
u
x
i
_
2
e
t
dxdt +
__
Q
_
u
t
_
2
e
t
dxdt

__
Q
f
2
e
t
dxdt +M
1
_

i=1
_

x
i
_
2
dx, (2.39)
which is valid for all > 0. By M
i
(i = 1, 2, . . .) here and below we denote constants depending on the quantities
cited for M in the statement of the theorem.
Now we shall prove inequality (2.39) in the case where the a
ij
are constant, b
i
= c = 0 and the function
u(x, t) vanishes outside a domain with boundary that can coincide with S only on a plane piece. Without loss of
generality, we can assume that this part of S is a part of the plane x
n
= 0.
To be denite we shall suppose that the domain where u(x, t) does not vanish lies in the half-space x
n
> 0.
As before, we have
__
Q
e
t
f
2
dxdt =
__
Q
e
t
_
u
t
_
2
dxdt +
_

e
t
n

i,j=1
a
ij
u
x
i
u
x
j

t=T
t=0
dx
+
__
Q
e
t
n

i,j=1
a
ij
u
x
i
u
x
j
dxdt +
__
Q
e
t
n

i,j,k,l=1
a
ij
a
kl

2
u
x
i
x
k

2
u
x
j
x
l
dxdt
+
T
_
0
e
t
n

i,j,k,l=1
a
ij
a
kl
_ _
xn=0
u
x
i

2
u
x
k
x
l
dx
dx
j

_
xn=0
u
x
i

2
u
x
l
x
j
dx
dx
k
_
dt. (2.40)
1
The inequality (2.38) follows from the fact that for any symmetric matrix A = a
ij
having the least eigenvalue > 0 and any
symmetric matrix B =
ij
the inequality
n

i,j,k,l=1
a
ij
a
kl

ik

jl

2
n

i,k=1

2
ik
()
holds. Inequality () is obvious for diagonal A. Note that the left-hand side of () is the trace of the matrix K = ABAB. Let C be
an orthogonal matrix such that CAC
1
is diagonal. Let sp(K) denote the trace of K. Since we have
sp(K) = sp(CKC
1
) = sp(CAC
1
CBC
1
CAC
1
CBC
1
)
and under an orthogonal transformation the right-hand side of () does not change, () is valid also in the general case.
460
It follows from u[
xn=0
= 0 that for i ,= n, j ,= n the derivatives
u
xi
and

2
u
xi xj
vanish on x
n
= 0. Moreover,
the integrals over the boundary x
n
= 0 are nonzero only for j = n and k = n. So the nonzero terms of the last sum
have the following form:
n

k,l=1
a
nn
a
kl
_
xn=0
u
x
n

2
u
x
k
x
l
dx
dx
n

j,l=1
a
nj
a
nl
_
xn=0
u
x
n

2
u
x
l
x
j
dx
dx
n
= 2
n

l=1
a
nn
a
nl
_
xn=0
u
x
n

2
u
x
n
x
l
dx
dx
n
2
n

l=1
a
nn
a
nl
_
xn=0
u
x
n

2
u
x
l
x
n
dx
dx
n
= 0.
The remaining integrals in (2.40) can be estimated as in the rst case, and we obtain (2.39) from (2.40).
Let us prove the inequality (2.35) for the general case. The domain can be covered by a nite number of
overlapping domains
l
(l = 1, . . . , N
1
) such that in any of them there exists a smooth transformation of independent
variables x
i
transforming the part of the boundary of
l
lying on to a plane. Under such a transformation, the form
of the equation and the smoothness of the coecients do not change. We shall call the new independent variables
i
,
and the coecients a
ij
(, t),

b
i
(, t), and c(, t) correspondingly. The least eigenvalue of the matrix | a
ij
(, t)| is
> 0, and the maximum modulus of all the coecients of the transformed equation is denoted by M
2
.
The idea of the proof of estimate (2.35) in the general case is as follows. The function u(x, t) is represented as
a sum of the functions u
k
(x, t); each of these functions is nonzero only in a small domain. The domain can be interior
or lying at the boundary. In a small domain, the part of the boundary of
l
belonging to S can be transformed to
a plane and the equation with variable coecients can be approximated by an equation with constant coecients.
For this approximation Eq. (2.35) is already obtained.
To realize this plan, we cover the domain Q by N overlapping small domains Q
k
such that in any two points
of Q
k
the dierence of the values of the coecients a
ij
(, t) does not exceed

2(n+1)
and each domain Q
k
belongs to
one of the domains
l
(0, T).
Take a so-called partition of unity (see [47, p. 162]). This means that the unity is represented in the form
1
N

k=1
h
k
(x, t), where h
k
(x, t) is a non-negative suciently smooth function that is nonzero only in the domain Q
k
(k = 1, . . . , N). We have
u(x, t) =
N

k=1
u(x, t)h
k
(x, t) =
n

k=1
u
k
(x, t).
Let L(u
k
) = f
k
(x, t). In the new coordinates , t, this equation can be rewritten as
n

i,j=1
a
ij
(
0k
, t
0k
)

2
u
k

u
k
t
=

f +
n

i,j=1
[ a
ij
(
0k
, t
0k
) a
ij
(, t)]

2
u
k

i=1

b
i
(, t)
u
k

i
c(, t)u
k
g
k
(, t),
where (
0k
, t
0k
) is some point in Q
k
. We shall estimate the right-hand side of the equation we obtained:
g
2
k
(n + 1)
2
_

f
2
k
+

2
4(n + 1)
2
n

i,j=1
_

2
u
k

j
_
2
+M
2
2
n

i=1
_
u
k

i
_
2
+M
2
2
u
2
k
_
,
since [ a
ij
(
0k
, t
0k
) a
ij
(, t)[

2(n+1)
. The function u
k
(, t) satises a parabolic equation with constant coecients
and the right-hand side g
k
and is nonzero only in one of the domains Q
k
. Therefore, according to what was proved
before, for the function u
k
(, t) the inequality (2.39) holds:

2
__
Q
n

i,j=1
_

2
u
k

j
_
2
e
t
d dt +
__
Q
n

i=1
_
u
k

i
_
2
e
t
d dt +
__
Q
_
u
k
t
_
2
e
t
d dt

__
Q
g
2
k
e
t
d dt +M
1
_

i=1
_
u
k
(, 0)

i
_
2
d (n + 1)
2
__
Q
_

f
2
k
+M
2
2
n

i=1
_
u
k

i
_
2
+M
2
2
u
2
k
_
e
t
d dt
+

2
4
__
Q
n

i,j=1
_

2
u
k

j
_
2
e
t
d dt +M
1
_

i=1
_
u
k
(, 0)

i
_
2
d.
461
Hence
__
Q
_
n

i,j=1
_

2
u
k

j
_
2
+
n

i=1
_
u
k

i
_
2
+
_
u
k
t
_
2
_
e
t
d dt
M
3
___
Q
_

f
2
k
+
n

i=1
_
u
k

i
_
2
+u
2
k
_
e
t
d dt +
_

i=1
_
u
k
(, 0)

i
_
2
d
_
. (2.41)
Making the inverse transformation from
i
to x
i
, we can obtain the estimate (2.41) also in the old coordinates
(perhaps, with another constant M
3
).
Because 0 h
k
(x, t) 1, we have
u
2
k
(x, t) = u
2
(x, t)h
2
k
(x, t) u
2
(x, t). (2.42)
For the derivative
u
k
xi
we obtain the following estimate:
_
u
k
x
i
_
2
=
_
(h
k
u)
x
i
_
2
=
_
h
k
u
x
i
+u
h
k
x
i
_
2
2h
2
k
_
u
x
i
_
2
+ M
4
u
2
2
_
u
x
i
_
2
+M
4
u
2
. (2.43)
We estimate also f
k
(x, t):
[f
k
(x, t)]
2
= [L(u
k
)]
2
= [L(h
k
u)]
2
=
_
h
k
L(u) +uL(h
k
) + 2
n

i,j=1
a
ij
u
x
i
h
k
x
j
cuh
k
_
2
M
5
_
[L(u)]
2
+u
2
+
n

i=1
_
u
x
i
_
2
_
= M
5
_
f
2
+u
2
+
n

i=1
_
u
x
i
_
2
_
. (2.44)
Using the inequality (2.41) in the old coordinates and the estimates (2.42), (2.43) and (2.44), we obtain
the following inequality:
__
Q
e
t
_
n

i,j=1
_

2
u
k
x
i
x
j
_
2
+
n

i=1
_
u
k
x
i
_
2
+
_
u
k
t
_
2
_
dxdt
M
6
___
Q
e
t
_
f
2
+
n

i=1
_
u
x
i
_
2
+u
2
_
dxdt +
_

_
n

i=1
_
u(x, 0)
x
i
_
2
+u
2
(x, 0)
_
dx
_
. (2.45)
We have u(x, t) =
N

k=1
u
k
(x, t); hence
u
2
(x, t) M
7
N

k=1
u
2
k
(x, t),
_
u
t
_
2
M
7
N

k=1
_
u
k
t
_
2
,
_
u
x
i
_
2
M
7
N

k=1
_
u
k
x
i
_
2
,
_

2
u
x
i
x
j
_
2
M
7
N

k=1
_

2
u
k
x
i
x
j
_
2
.
Using these estimates and summing the inequalities (2.45) with respect to k from 1 to N, we come to
__
Q
e
t
_
n

i,j=1
_

2
u
x
i
x
j
_
2
+
n

i=1
_
u
x
i
_
2
+
_
u
t
_
2
_
dxdt
M
8
___
Q
e
t
_
f
2
+
n

i=1
_
u
x
i
_
2
+u
2
_
dxdt +
_

_
n

i=1
_

x
i
_
2
+
2
_
dx
_
. (2.46)
462
Since u[
S
= 0, u(x
1
, x
2
, . . . , x
n
, t) =
x1
_
x
0
1
u
x1
(, x
2
, . . . , x
n
, t) d, where the point (x
0
1
, x
2
, . . . , x
n
, t) S. There-
fore,
u
2
(x, t) d
x1
_
x
0
1
_
u
x
1
_
2
d,
where d is the diameter of the domain . Integrating this inequality over the cylinder Q, we get
__
Q
u
2
(x, t) dxdt M
6
__
Q
n

i=1
_
u
x
i
_
2
dxdt. (2.47)
It follows from the inequalities (2.46) and (2.47) that
__
Q
e
t
_
u
2
+
n

i=1
_
u
x
i
_
2
+
n

i,j=1
_

2
u
x
i
x
j
_
2
+
_
u
t
_
2
_
dxdt
M
10
___
Q
e
t
_
f
2
+
n

i=1
_
u
x
i
_
2
_
dxdt +
_

2
+
n

i=1
_

x
i
_
2
_
dx
_
.
If we take = 2M
10
, we obtain the required estimate (2.35).
Remark. The estimate (2.35) holds not only in the cylinder Q but also in a noncylindrical domain D if
the boundary S belongs to the class A
2
.
An estimate similar to (2.35) was rst obtained by S. N. Bernstein in 1908 for the solution of the Dirichlet
problem for a general elliptic second-order equation with two independent variables ([48], see also [49]). Another
proof of (2.35) for a parabolic second-order equation is given in [50, 51]. In these papers, the authors use similar
estimates of the solution to the Dirichlet problem for second-order elliptic equations with several independent
variables; these last estimates are obtained in [52, 53].
Note that estimates similar to (2.35) in L
p
(p > 1) norms and also estimates for elliptic and parabolic
equations of higher orders require essentially new approaches. The approach connected with the parametrix
(that is, almost an inverse operator) of a problem, can be found in [5457].
In [58, 59], to obtain estimates for elliptic equations and systems the approach is based on the use of Fourier
transform for the equations with constant coecients; after that the equation with variable coecients is compared
to an equation with constant coecients in small domains, similarly to the reasoning we used in proving (2.35).
Estimates of the type (2.35) are widely used in proving the existence for boundary-value problems. We
shall use these estimates in Sects. 7 and 8. Similar estimates in L
p
norms are used also in the study of nonlinear
equations [55].
3 SOLVING OF BOUNDARY-VALUE PROBLEMS BY THE ROTHE METHOD. CAUCHY PROBLEM
In this section, we construct the solutions of the main problems for Eq. (1.1) with several space variables
by a method similar to the method of straight lines for the equation with one space variable. E. Rothe [60] used
this method to solve the rst boundary-value problem for a parabolic equation with two independent variables.
A review of works where Rothes method is used for two or more independent variables to prove the solvability of
boundary-value problems and to construct their approximate solutions can be found in [61]. Note that in [62, 63]
the solutions of the rst and second boundary-value problems for quasilinear parabolic equations with several space
variables are constructed by means of the Rothe method.
Using this method we shall construct solutions of the rst and second boundary-value problems for Eq. (1.1)
in the cylinder Q = (0, T). The construction is based on theorems of existence of solutions to the corresponding
problems for elliptic equations. In this section, we also study the smoothness of a solution of the rst boundary-
value problem in accordance with the smoothness of the data (coecients of the equation, the boundary of
the domain , the initial and boundary functions). The rst and second boundary-value problems are studied for
463
a cylindrical domain Q and a noncylindrical domain D under very weak assumptions on the boundaries of these
domains (existence of the barrier functions). The existence of the solution of the Cauchy problem for Eq. (1.1) in
a layer H will be proved as a corollary to the existence theorem for the solution of the rst boundary-value problem.
For the case of the Cauchy problem we shall suppose only that c(x, t) M and [f(x, t)[ M and any growth of
the coecients a
ij
(x, t) and b
i
(x, t) as
n

i=1
x
2
i
is admitted. Examples are given showing that the suppositions
about c and f are essential for the solvability of the Cauchy problem in the class of bounded functions.
1. Construction of a Solution of the First Boundary-Value Problem, Smooth up to the Bound-
ary of the Domain. Consider the rst boundary-value problem for Eq. (1.1) in the cylinder Q = (0, T) with
the following conditions:
u(x, 0) = (x), u[
S
= (x, t). (3.1)
Theorem 1. There exists in the cylinder Q a unique solution of the problem (1.1), (3.1), continuous in

Q
together with its derivatives, appearing in Eq. (1.1), if the following suppositions are fullled.
(1)
n

i,j=1
a
ij
(x, t)
i

j

n

i=1

2
i
for (x, t)

Q and any
1
, . . . ,
n
; here = const > 0.
(2) The coecients a
ij
, b
i
, and c of Eq. (1.1) and its right-hand side f in

Q belong to the class C
4+
with
respect to x
1
, . . . , x
n
and are Lipschitz continuous with respect to t together with their derivatives of the rst and
the second order in x
1
, . . . , x
n
.
(3) The boundary of the domain belongs to the class A
6+
.
(4) (x) C
6+
(

); C
6+
as a function of local coordinates
1
, . . . ,
n1
on the surface . The
derivatives

t
,

2

i t
(i = 1, . . . , n 1) exist; these derivatives and also the derivatives of up to the fourth order
with respect to
1
, . . . ,
n1
are Lipschitz continuous with respect to t.
(5) Consistency conditions are fullled:
(x, 0) = (x) for x , (3.2)
_
n

i,j=1
a
ij
(x, 0)

2

x
i
x
j
+
n

i=1
b
i
(x, 0)

x
i
+c(x, 0) f(x, 0)
_
x
=

t
(x, 0)

x
. (3.3)
Proof. Without loss of generality, one can suppose that c(x, t) < 0 in

Q (if this is not so, this can be achieved
by the transformation u = ve
t
). We can decompose the interval [0, T] into m equal parts and instead of Eq. (1.1)
we can take the following system of elliptic equations:
L
t
(u)
n

i,j=1
a
ij
(x, kt)

2
u(x, kt)
x
i
x
j
+
n

i=1
b
i
(x, kt)
u(x, kt)
x
i
+c(x, kt)u(x, kt)
u(x, kt) u(x, (k 1)t)
t
= f(x, kt) (3.4)
with the conditions
u(x, kt) = (x, kt) for x , (3.5)
where t =
T
m
, k = 1, . . . , m; when k = 0 we take u(x, 0) = (x). The functions u(x, kt) can be found
consecutively by solving the Dirichlet problem for a second-order elliptic equation. According to Schauders theorem
[2, p. 150], under the suppositions of Theorem 1 on the coecients of (1.1), the boundary , and the functions (x)
and (x, t), for all k there is a unique solution u(x, kt) of the problem (3.4), (3.5), belonging to the class C
6+
(

)
as a function of x. We shall prove that for suciently small t the families of functions
u(x, kt),
u(x, kt)
x
i
,
u
t

u(x, kt) u(x, (k 1)t)
t
,

2
u(x, kt)
x
i
x
j
,

t
_
u
x
i
_

1
t
_
u(x, kt)
x
i

u(x, (k 1)t)
x
i
_
,

3
(x, kt)
x
i
x
j
x
l
,

t
_

2
u
x
i
x
j
_

1
t
_

2
u(x, kt)
x
i
x
j


2
u(x, (k 1)t)
x
i
x
j
_
,

2
u
t
2

u(x, kt) 2u(x, (k 1)t) +u(x, (k 2)t)
t
2
(i, j, l = 1, . . . , n)
_

_
(3.6)
464
are uniformly bounded. For kt t (k + 1)t we linearly extend the functions u(x, kt); the resulting
functions are called u
t
(x, t). Because the functions (3.6) are uniformly bounded, we can select from the family
u
t
(x, t) a subsequence that is uniformly converging in

Q as t 0; the limit u(x, t) satises (1.1) in

Q and
the conditions (3.1) on .
First we shall show that in

Q the functions u(x, kt) are uniformly bounded with respect to t.
Let t be xed; if u(x, kt) attains its positive maximum (as a function of x and k) at a point lying in

Q,
then at this point we have, owing to (3.4),
c(x, kt)u(x, kt) +f(x, kt) 0,
whence
u(x, kt) max

f
c

= M
1
.
It is possible to establish a similar estimate for u(x, kt) at a point of its negative minimum. If the maximum of
[u(x, kt)[ is attained at a point of , then we have at this point
[u(x, kt)[ max

[(x)[, [(x, t)[ = M


2
.
Thus for any x and k we have the inequality
[u(x, kt)[ M
3
,
with the constant M
3
not depending on t.
Now we shall prove that
u(x,kt)
x
l
(l = 1, . . . , n) is uniformly bounded with respect to t in

Q.
The estimate for
n

l=1
_
u(x,kt)
x
l

2
on S is obtained in the same way as in Theorem 2 of Sect. 2. To nd
the estimate for this sum in

Q we take the following auxiliary function:
w(x, k) = e
Nkt
n

l=1
_
u(x, kt)
x
l
_
2
+e
N(Tkt)
;
N will be chosen later.
Set w = w(x, k) w(x, k 1); we have
L
t,N
(w)
n

i,j=1
a
ij

2
w
x
i
x
j
+
n

i=1
b
i
w
x
i
+cw e
Nt
w
t
.
It is easy to see (as in Theorem 1 of Sect. 1) that if L
t,N
(w) > 0 in

Q , the function w(x, k) cannot attain
a maximum in

Q .
To calculate L
t,N
(w) we shall use the identities
(v
1
v
2
)
t
= v
1
v
2
t
+v
2
(x, k 1)
v
1
t
,
(v
2
)
t
= 2v
v
t

(v)
2
t
.
We have
L
t,N
(w) = e
Nkt
_
2
n

i,j,l=1
a
ij

2
u
x
l
x
i

2
u
x
l
x
j
+ 2
n

l=1
u
x
l
_
n

i,j=1
a
ij

3
u
x
l
x
i
x
j
+
n

i=1
b
i

2
u
x
l
x
i

1
t

_
u
x
l
__
+
1
t
n

l=1
_

_
u
x
l
_
2
_
+c
n

l=1
_
u
x
l
_
2
e
Nt
e
Nkt
(e
Nkt
)
t
n

l=1
_
u
x
l
_
2
+
_
c
e
Nt
1
t
_
e
NT
_
e
Nkt
_
(2 M
4
)
n

l,i=1
_

2
u
x
l
x
i
_
2
+
_
1e
Nt
t

M
5

M
5
_
n

l=1
_
u
x
l
_
2
+
_
1e
Nt
t
M
6
_
e
NT
M
7
_
;
465
here the M
i
are constants independent of N and t. Here we have used equations obtained by dierentiation of
both sides of (3.4) with respect to x
l
(compare to (2.16) and (2.17)).
Take <
2
M4
, and after this N = 4
_
M
5
+
M5

+M
6
+M
7
_
. Then for all t <
1
N
everywhere in

Q
the inequality L
t,N
(w) > 0 is valid, owing to
1e
Nt
t
> Ne
1
>
N
4
.
Therefore,
w(x, k) max

w(x, k) max

l=1
_
u(x, kt)
x
l
_
2
+e
NT
M
8
.
Hence
n

l=1
_
u(x, kt)
x
l
_
2
M
8
e
NT
= M
9
,
where M
9
does not depend on t.
Now we shall prove that the functions

2
u(x,kt)
x
l
xp
(l, p = 1, . . . , n) are uniformly bounded in

Q with respect
to t.
Consider in

Q the auxiliary function
w
1
(x, k) = e
N1kt
n

l,p=1
_

2
u(x, kt)
x
l
x
p
_
2
+e
N1(Tkt)
.
Using the equations obtained by twice dierentiating both sides of (3.4) with respect to x
l
and x
p
, we obtain
L
t,N1
(w
1
) e
N1kt
_
(2 M
10
)
n

i,l,p=1
_

3
u
x
i
x
l
x
p
_
2
+
_
1 e
N1t
t

M
10

M
10
_
n

l,p=1
_

2
u
x
l
x
p
_
2
+
1 e
N1t
t
M
10
_
> 0,
if
<
2
M
10
, N
1
> 4
_
M
10

+M
10
_
, and t <
1
N
1
.
Therefore, if N
1
is chosen as above, we have in

Q the inequality
w
1
(x, k) max

w
1
.
Setting
M
Q
= max

Q
_
n

l,p=1
_

2
u(x, kt)
x
l
x
p
_
2
_
1/2
, M

= max

_
n

l,p=1
_

2
u(x, kt)
x
l
x
p
_
2
_
1/2
,
we get
M
Q
M
11
(M

+ 1). (3.7)
Then, as in Theorem 3 of Sect. 2, we can estimate
n

l,p=1
_

2
u
x
l
xp
_
2
on S. As a result we obtain
M

M
12
_
_
M
Q
+ 1
_
. (3.8)
It follows from (3.7) and (3.8) that M
Q
M
13
.
As a consequence of Eqs. (3.4), the dierence quotients
u
t
=
u(x, kt) u(x, (k 1)t)
t
are also uniformly bounded in

Q with respect to t.
466
Let us estimate in

Q the derivatives of the form

3
u(x,kt)
x
l
xpxq
and

x
l
_
u
t
_
(l, p, q = 1, . . . , n).
Set
M

Q
= max

Q
_
n

l,p,q=1
_

3
u
x
l
x
p
x
q
_
2
_
1/2
, M

Q
= max

Q
_
n

i=1
_

x
l
_
u
t
__
2
_
1/2
,
and denote by M

and M

the maximum values of these sums on correspondingly. Using the function


w
2
(x, k) = e
N2kt
n

l,p,q=1
_

3
u
x
l
x
p
x
q
_
2
+e
N2(Tkt)
,
with N
2
suciently large, and reasoning as we did above, we can obtain the estimate
M

Q
M
14
(M

+ 1). (3.9)
Then, using the functions
z

lq
(, k) =

2
u(
1
, . . . ,
n1
,
n
, kt)

l

q


2
(
1
, . . . ,
n1
, kt)

l

q
N
3
e
N4n

p
(
1
, . . . ,
n1
) (l, q = 1, . . . , n 1)
(the notations are the same as in Theorem 2 of Sect. 2), we show that on any piece

S
p
the inequality
_
n1

l,q=1
_

3
u

n
_
2
_
1/2

Sp
M
15
__
M

Q
+ 1
_
(3.10)
holds.
Dierentiating Eqs. (3.4) in the vicinity of S
p
with respect to
1
, . . . ,
n
, t, we get

nn

3
u

2
n
=

i+j<2n

ij

3
u

l
+

l
_
u
t
_
+R
l
(u); (3.11)
here R
l
(u) denotes the sum of terms containing derivatives of u with respect to
1
, . . . ,
n
of order not higher than
the second. From (3.10) and (3.11) it follows that
M

M
16
__
M

Q
+M

+ 1
_
. (3.12)
Comparing (3.9) and (3.12), we get
M

Q
M
12
(M

+ 1). (3.13)
Now let us estimate M

. We extend u(x, kt) to k = 1 and x



by setting
u(x, t) = (x) t
_
n

i,j=1
a
ij
(x, 0)

2
(x)
x
i
x
j
+
n

i=1
b
i
(x, 0)
(x)
x
i
+c(x, 0)(x) f(x, 0)
_
. (3.14)
In the vicinity of S
p
for 0 k m we consider the auxiliary functions
z

(, k) =
u(
1
, . . . ,
n1
,
n
, kt)
t

(
1
, . . . ,
n1
, kt)
t
N
5
e
N6n

p
(
1
, . . . ,
n
).
Because of the smoothness of (x, t) and the consistency conditions (3.2), (3.3) the value L
t
_

t
_
is uniformly
bounded with respect to t. Using the equations obtained by applying to both sides of (3.4) the operator

t
, we
arrive at L
t
(z
+
(, k)) > 0 and L
t
(z

(, k)) < 0 in the neighborhood of S


p
, if N
5
and N
6
are chosen in a correct
way. Then, as in Theorem 2 of Sect. 2, we can get

n
_
u
t
_

Sp
M
18
.
467
Because of the smoothness of (x, t) we also have

i
_
u
t
_

Sp
M
19
(i = 1, . . . , n 1).
Finally, the smoothness of (x) and equality (3.14) lead to

i
_
u
t
_

t=0
M
20
(i = 1, . . . , n).
Thus we have
M

M
21
. (3.15)
It follows from (3.13) and (3.15) that
M

Q
M
22
,
and this means that the derivatives

3
u
x
l
xpxq
are bounded uniformly in

Q with respect to t. It follows from
equations obtained by dierentiating (3.4) with respect to x
l
that the derivatives

x
l
_
u
t
_
also are uniformly
bounded in

Q.
Then by the same arguments we used to obtain bounds for

3
u(x,kt)
x
l
xpxq
and

x
l
_
u(x,kt)
t
_
we can prove
the uniform boundedness of

4
u(x,kt)
x
l
xpxqxr
and

2
x
l
xp
_
u(x,kt)
t
_
(l, p, q, r = 1, . . . , n). The uniform boundedness of
the dierence quotients

2
u
t
2
=

t
_
u(x,kt)
t
_
follows from equations obtained by applying the operator

t
to both
sides of (3.4). So we can prove in

Q the uniform boundedness of all functions (3.6) with respect to t.
Now we shall consider a family of functions u
t
(x, t); they are continuous in

Q. Because the functions
u(x, kt), their derivatives
u(x,kt)
xi
(i = 1, . . . , n), and dierence quotients
u(x,kt)
t
are uniformly bounded in

Q,
the functions u
t
(x, t) are equicontinuous in

Q. According to the Arcela theorem, one can single out from the family
u
t
(x, t) a subsequence u
m
(x, t), converging uniformly in

Q to some function u(x, t).
The families
ut(x,t)
xi
,

2
ut(x,t)
xi xj
, and
ut(x,t)
t
are also uniformly bounded and equicontinuous in

Q; this
follows from the uniform boundedness of functions (3.6). Therefore, a new subsequence can be chosen in the sequence
u
m
(x, t) (it will also be called u
m
(x, t)), such that the corresponding subsequences
u
m
(x,t)
xi
,

2
u
m
(x,t)
xi xj
, and
u
m
(x,t)
t
converge uniformly in

Q to some continuous functions u
i
(x, t), u
ij
(x, t), and u
0
(x, t) respectively.
It is easy to show that u
i
(x, t) =
u(x,t)
xi
, u
ij
(x, t) =

2
u(x,t)
xi xj
, and u
0
(x, t) =
u(x,t)
t
. (The proof is similar
to one given for the heat equation in [19, p. 361362].) Then for the chosen subsequence we pass to the limit as
m in Eqs. (3.4) and the result is that the function u(x, t) everywhere in

Q satises Eq. (1.1). Because of
the uniform convergence of u
m
(x, t), the function u(x, t) satises also the boundary conditions (3.1) on .
The uniqueness of the solution of the problem (1.1), (3.1) is proved in Sect. 1.
Remark. If greater smoothness is required from the data of problem (1.1), (3.1), then the function u(x, t)
constructed in Theorem 1 will also be more smooth than is stated in this theorem. Namely, to prove the existence
in

Q S of continuous derivatives of the form

p+q
u(x, t)
x
p1
1
. . . x
pn
n
t
q
, where p + 2q r (r > 2) (3.16)
it is sucient to suppose that (x) C
r+4+
() and that in any closed domain

Q



Q S there exist continuous
derivatives that are Holder continuous in x:

p+q
a
ij
(x, t)
x
p1
1
. . . x
pn
n
t
q
, where p + 2q r + 2, q
_
r
2
_
;
the same is supposed about the derivatives of b
i
(x, t), c(x, t), and f(x, t).
468
In fact, under such suppositions according to a theorem by E. Hopf [2, p. 152] the solution u(x, kt) of
problem (3.4), (3.5) for any t and for k = 1, . . . , m =
T
t
in any closed domain

Q

belongs to the class


C
r+4+
as a function of x
1
, . . . , x
n
. In the cylinder

_
n

i=1
(x
i
x
0
i
)
2

2
, 0 t T
_
,
consider the auxiliary function
w(x, k) =

(x)
n

l=1
_
u(x, kt)
x
l
_
2
+N
7
u
2
(x, kt) +e
N8(Tkt)
; (3.17)
here x
0
= (x
0
1
, . . . , x
0
n
) ,

(x) =
_

i=1
(x
i
x
0
i
)
2
_
2
, > 0 is suciently small, and the constants N
7
and N
8
are suciently large. As in Theorem 1 of Sect. 2, we can show that L
t
(w) > 0 in the cylinder R

and therefore
in

R
/2
the inequality
n

l=1
_
u(x,kt)
x
l
_
2
M
23
holds; M
23
does not depend on t. Similarly, one can prove that
the families of derivatives and dierence quotients

p
x
p1
1
. . . x
pn
n

q
u(x, t)
t
q
, where p + 2q r + 2, (3.18)
are uniformly bounded with respect to t in R
1
(
1
> 0) (see the remark to Theorem 1 of Sect. 2). It is easy to
deduce from the boundedness of these families that the limit function u(x, t) in

Q S has continuous derivatives of
the form (3.16).
To prove the existence of continuous derivatives (3.16) in the closed cylinder

Q we need only additional
suppositions that the boundary of the domain belongs to the class A
r+4+
, (x) C
r+4+
(

) and belongs
to the class C
r+4+
as a function of local coordinates
1
, . . . ,
n1
on and has on

S continuous derivatives of
the form

p+q

p1
1
. . .
pn1
n1
t
q
, where p + 2q r + 4, q
_
r
2
_
+ 1.
Besides this, the functions and on must satisfy some new consistency conditions. These conditions are
deduced from the equation obtained by applying dierential operators

p+q
x
p
1
1
...x
pn
n
t
q
with p + 2q r 2 to (1.1)
and then expressing the derivatives of u(x, t) in terms of and .
In fact, under such suppositions one can show that in

Q all the derivatives and dierence quotients of
the form (3.18) are uniformly bounded with respect to t. This is done by the methods used in Theorem 1. Then,
using the same arguments as above, we can prove the existence of continuous derivatives (3.16) in

Q.
2. First Boundary-Value Problem for Noncylindrical Domains. In this section, a solution to
the rst boundary-value problem for Eq. (1.1) is constructed under more general suppositions about the boundary
of the domain as well as about the initial and boundary functions.
Consider Eq. (1.1) in the domain D described in Sect. 1. A function v
P0
(x, t) is called a barrier at a point
P
0
(x
0
, t
0
) if it has the following properties:
(A) v
P0
(x, t) is dened and continuous in the intersection G
P0
of the closed domain

D with some neighborhood
of the point P
0
.
(B) v
P0
(x
0
, t
0
) = 0, v
P0
(x, t) > 0 at all the points of G
P0
dierent from P
0
.
(C)

L(v
P0
)
n

i,j=1
a
ij
(x, t)

2
vP
0
xi xj
+
n

i=1

b
i
(x, t)
vP
0
xi
+ c(x, t)v
P0

vP
0
t
< 0 everywhere inside G
P0
for all
continuous functions a
ij
(x, t),

b
i
(x, t), and c(x, t) such that the absolute values of the dierences between them and
a
ij
(x, t), b
i
(x, t), and c(x, t) respectively are suciently small in G
P0
.
The following theorem is an auxiliary one.
Theorem 2. In the cylinder Q = (0, T), there exists a unique solution u(x, t) of the problem (1.1), (1.3),
bounded and continuous in

Q and satisfying Eq. (1.1) in

Q and the boundary conditions (3.1) on , if
the following conditions are fullled:
469
(1) Supposition (1) of Theorem 1.
(2) The functions a
ij
, b
i
, c, and f are continuous in

Q, and in

Q the derivatives of the form

p+q
a
ij
x
p1
1
. . . x
pn
n
t
q
, where p + 2q 4, q 1, (3.19)
exist and are Holder continuous with respect to x in any closed domain contained in

Q . The same is supposed
about the derivatives of this form of b
i
, c, and f.
(3) The boundary of the domain has the following property: there exists a number , 0 < < n, such
that for any > 0 the boundary can be covered by a nite number N of n-dimensional balls of radii
1
, . . . ,
N
such that
N

l=1

l
< .
(4) The function (x) is bounded and continuous inside and the function (x, t) is bounded and continuous
on S.
(5) There exists a barrier for every point P
0
.
If at some point P
1
the consistency condition (P
1
) = (P
1
) is fullled, then the function u(x, t) is
continuous at the point P
1
and satises the condition u(P
1
) = (P
1
).
If in

Q there exist continuous derivatives of the form

p+q
a
ij
x
p1
1
. . . x
pn
n
t
q
, p + 2q r + 2, q
_
r
2
_
, (3.20)
and the same derivatives of functions b
i
, c, and f, then the function u(x, t) in

Q has continuous derivatives of
the form

p+q
u
x
p1
1
. . . x
pn
n
t
q
, p + 2q r. (3.21)
If besides this in a domain

contained inside the function (x) belongs to the class C


r+2
and the deriva-
tives of the form (3.20) of the functions a
ij
, b
i
, c, and f are continuous on the set t = 0, x

, then the deriva-


tives (3.21) also are continuous on this set.
Proof. We shall construct sequences of uniformly bounded and equicontinuous in

Q functions a
m
ij
, b
m
i
, c
m
,
and f
m
that converge uniformly to a
ij
, b
i
, c, and f, respectively, together with derivatives of the form (3.19) in any
closed domain contained in

Q . Functions f
m
are constructed so that f
m
(x, t) = 0 at all the points situated at
a distance not greater than
1
m
(m = 1, 2, . . .) from . We can extend the functions (x) and (x, t), dened on , to
all of the cylinder

Q so that the resulting function U(x, t) will be bounded and continuous in

Q. Then we construct
a sequence of uniformly bounded innitely dierentiable functions U
m
(x, t) that converge uniformly to U(x, t) in
any closed domain contained in

Q , and U
m
(x, t) = 0 at all the points situated at a distance from not greater
than
1
m
(m = 1, 2, . . .). Then we approximate the domain from inside by a sequence of expanding domains
m
having the following properties: the boundary of
m
belongs to the class A
6+
,
m

m+1
,

m=1

m
= , and all
the points of at a distance greater than
1
2m
from belong to the domain
m
(m = 1, 2, . . .).
Denote by
m
the boundary of
m
, by Q
m
the cylinder
m
(0, T), and by S
m
its side boundary
m
(0, T].
According to Theorem 1, for each m in

Q
m
there exists a unique solution u
m
(x, t) of the equation
L
m
(u)
n

i,j=1
a
m
ij
(x, t)

2
u
x
i
x
j
+
n

i=1
b
m
i
(x, t)
u
x
i
+c
m
(x, t)u
u
t
= f
m
(x, t), (3.22)
satisfying the conditions
u
m
(x, 0) = U
m
(x, 0) for x

m
, u
m
(x, t) = U
m
(x, t) on

S
m
; (3.23)
note that the consistency conditions (3.2), (3.3) are fullled because the functions U
m
and f
m
vanish in a neigh-
borhood of
m
.
According to Theorem 5 of Sect. 1, the functions u
m
are uniformly bounded. Applying S. N. Bernsteins
estimates (see Theorem 1 of Sect. 2 and the remark to this theorem), we can show that in any closed domain
470



Q the derivatives
u
m
xi
,

2
u
m
xi xj
,
u
m
t
,

2
u
m
xi t
,

3
u
m
xixjx
l
,

3
u
m
xixjt
, and

2
u
m
t
2
(i, j, l = 1, . . . , n; m = 1, 2, . . .)
are uniformly bounded. Therefore, one can choose in the sequence u
m
(x, t) a subsequence that converges uniformly
in

Q

together with the derivatives appearing in Eq. (1.1). Considering a sequence of domains

Q

1


Q

2
. . .

m
. . . such that

m=1

m
=

Q and using the diagonal process, we obtain a subsequence u
m
k
(x, t) that
converges uniformly together with
u
m
k
xi
,

2
u
m
k
xi xj
, and
u
m
k
t
in any closed subdomain of

Q . The limit function
u(x, t) = lim
k
u
m
k
(x, t) obviously satises Eq. (1.1) in

Q .
As was done for the heat equation (see [19, pp. 362365]), one can show that also for the general equation (1.1)
from the existence of a barrier for a point P
0
it follows that the boundary condition is satised at this point:
lim
PP0
u(P) = U(P
0
) if lim
m
U
m
(P) = U(P) uniformly in some neighborhood of the point P
0
. Therefore, according
to supposition (5) of this theorem, the function u(x, t) satises condition (3.1) everywhere on .
The uniqueness of the solution follows from Theorem 11 of Sect. 1. Note that because the solution is unique,
not only the subsequence u
m
k
(x, t) but the whole sequence u
m
(x, t) converges to u(x, t) in

Q .
Suppose now that P
1
, the functions and are continuous at the point P
1
, and (P
1
) = (P
1
) = . Let
us show that lim
PP1
u(P) = . First consider the case of = 0. Then we can dene the function U(x, t) so that it is
continuous at the point P
1
and construct the functions U
m
(x, t) so that they have the following additional property:
for every > 0, there exists a number m
0
() so large and a neighborhood G
P1
of the point P
1
so small that in G
P1
the inequality [U
m
(x, t)[ < holds for all m > m
0
(). From (3.23) it follows that [u
m
(x, t)[ < for m > m
0
() on
the part of the surface

S
m
belonging to G
P1
. Because it was supposed that a barrier exists (supposition 5), we come
to lim
PP1
u(P) = 0 in the same way as was done in [19]. This means that the boundary condition (3.1) is satised
at the point P
1
.
Now let (P
1
) = (P
1
) = ,= 0. We can change the variable u in (1.1) and (3.1): u = v + . The func-
tion v(x, t) in

Q satises the equation
L(v) = f(x, t) c(x, t)
and the boundary conditions v[
\
= U[
\
. As we have proved, the solution of this problem exists and is unique,
and the function v(x, t) is continuous at the point P
1
and v(P
1
) = 0. Obviously the function u(x, t) = v(x, t) +
is a solution of Eqs. (1.1), (3.1). Because of the uniqueness this function coincides with the solution constructed
above. Therefore, the latter at the point P
1
satises the boundary condition u(P
1
) = , which was to be proved.
Now suppose that in

Q the coecients a
ij
, b
i
, and c and the right-hand side f of Eq. (1.1) have continuous
derivatives of the form (3.20). In this case one can additionally require that the derivatives (3.20) of functions a
m
ij
,
b
m
i
, c
m
, and f
m
converge uniformly (as m ) in any closed domain

Q



Q to the corresponding derivatives
of a
ij
, b
i
, c, and f. It follows from S. N. Bernsteins estimates that in any closed domain

Q



Q the derivatives

p+q
u
m
x
p1
1
. . . x
pn
n
t
q
, p + 2q r + 2, (3.24)
are uniformly bounded. This leads to the compactness of the derivatives of the form (3.21) of u
m
(x, t), and we can
deduce that the derivatives (3.21) of u(x, t) exist in

Q



Q .
Finally, suppose that in a subdomain

the function (x) C


r+2
and the derivatives of the form (3.20)
of a
ij
, b
i
, c, and f are continuous on the set t = 0, x

. Then one can construct the sequence U


m
(x, t) so that
in any closed domain contained in

the derivatives

p
U
m
(x,0)
x
p
1
1
...x
pn
n
, p r +2, converge uniformly to the corresponding
derivatives of (x). We can also suppose that the derivatives of the form (3.20) of a
m
ij
, b
m
i
, c
m
, and f
m
converge as
m uniformly in all of the cylinder

[0, T]; here


is any closed subdomain of

. Let x
0

be an arbitrary
point and let > 0 be so small that the n-dimensional ball
n

i=1
(x
i
x
0
i
)
2

2
lies in

. As we did in the remark


to Theorem 1 of this section, we can show that in the cylinder

R

=
_
n

i=1
(x
i
x
0
i
)
2

2
, 0 t T
_
, where
0 <

< , the derivatives (3.24) of the functions u


m
(x, t) are uniformly bounded with respect to m. Therefore
the function u(x, t) has continuous derivatives (3.21) in

R

and in particular on the set


_
t = 0,
n

i=1
(x
i
x
0
i
)
2

2
_
.
As the point x
0

is arbitrary, this means that the derivatives (3.21) of u(x, t) are continuous everywhere on
the set t = 0, x

.
471
Remark 1. Supposition (5) of Theorem 1 is, for example, fullled when for every point x
0
= (x
0
1
, . . . , x
0
n
)
there exists an n-dimensional sphere with center y
0
= (y
0
1
, . . . , y
0
n
) that contains x
0
and has no more common points
with

. In this case the barrier at the point P
0
(x
0
, t
0
) S is the function
v
P0
(x, t) =
K
0

K
, (3.25)
with = (x, t) =
_
n

i=1
(x
i
y
0
i
)
2
+ (t t
0
)
2
_
1/2
,
0
= (x
0
, t
0
), and K > 0 a suciently large number. The re-
quirements (A) and (B) for a barrier are obviously fullled. To verify (C) we calculate

L(v
P0
):

L(v
P0
) = K
K4
_

2
_
n

i=1
a
ii
+
n

i=1

b
i
(x
i
y
0
i
) (t t
0
)
_
(K + 2)
n

i,j=1
a
ij
(x
i
y
0
i
)(x
j
y
0
j
)
_
+ cv
P0
< K
K4
_
M
1
(K + 2)

2
0
_
< 0,
if c(x, t) < 0 (this does not lead to loss of generality) and K is suciently large. As a barrier for the point P
1
(x
1
, 0),
where x
1
= (x
1
1
, . . . , x
1
n
)

, we can take the function
v
P1
(x, t) =
n

i=1
(x
i
x
1
i
)
2
+K
1
t (3.26)
with K
1
> 0 suciently large. This function also obviously fullls conditions (A) and (B). Let us verify (C):

L(v
P1
) = 2
n

i=1
[ a
ii
+

b
i
(x
i
x
1
i
)] + cv
P1
K
1
< 0,
if
n

i=1
(x
i
x
1
i
)
2
< 1 and K
1
> 0 is suciently large.
Remark 2. If in Theorem 2 we suppose in addition to (2) that in

Q the derivatives

p+q
a
ij
x
p1
1
. . . x
pn
n
t
q
, where p + 2q 6, q 2,
exist in any closed domain

Q



Q , and they are H older continuous with respect to x, and the same is supposed
on the derivatives of b
i
and c, then condition (5) can be replaced by a weaker one:
(5

) For any point P


0
, there exists a function V
P0
(x, t), having the properties (A) and (B) of a barrier
and the following property (C

):
(C

) L(V
P0
)
n

i,j=1
a
ij
(x, t)

2
V
P0
x
i
x
j
+
n

i=1
b
i
(x, t)
V
P0
x
i
+c(x, t)V
P0

V
P0
t
< 0 everywhere inside G
P0
.
The function V
P0
(x, t) is also called a barrier.
Under the suppositions we have made, the existence of the solution of problem (1.1), (3.1) in the cylinder Q
is proved as in Theorem 2; we only have to take a
m
ij
= a
ij
, b
m
i
= b
i
, and c
m
= c.
Remark 3. When the coecients of Eq. (1.1) do not depend on t, condition (5

) is fullled if all the points


of are regular with respect to the Laplace equation
n

i=1

2
u
x
2
i
= 0 (see [64] or [19, p. 266]). This means that
the boundary of the domain must be such that the Dirichlet problem for the Laplace equation is solvable for
any continuous boundary function dened on .
In this case, we can again take for the barrier at the point P
1
(x
1
, 0), x
1


, the function (3.26). To construct
a barrier for a point P
0
(x
0
, t
0
) S we rst note that for such a point it is sucient to make sure that (A), (B),
and (C

) are fullled only for t t


0
. This can be proved for Eq. (1.1) in the same way as for the heat equation
(see [19, pp. 364365]).
472
Let us suppose that in Eq. (1.1) we have c(x) < c
0
< 0. Let w(x) be the solution of the elliptic equation
n

i,j=1
a
ij
(x)

2
w
x
i
x
j
+
n

i=1
b
i
(x)
w
x
i
+c(x)w = 0
in the domain with the boundary condition
w(x)[

= F(x),
where F(x) is a continuous function such that F(x
0
) =
1
c0
and F(x) <
1
c0
for x ,= x
0
. The function w(x) exists
because of the regularity of the points of (see [65]). The maximum principle gives w(x) <
1
c0
everywhere in

except the point x = x


0
, where we have w =
1
c0
.
We shall demonstrate that at the point P
0
(x
0
, t
0
) we can take as a barrier the function
V
P0
(x, t) =
1
c
0
w(x) + (t
0
t).
In fact, for t t
0
we have
V
P0
(x, t) > 0 for
n

i=1
(x
i
x
0
i
)
2
+ (t
0
t) > 0,
V
P0
(x
0
, t
0
) = 0, L(V
P0
) =
c(x)
c
0
+c(x)(t
0
t) + 1 < 0,
and thus V
P0
is a barrier.
Consider now Eq. (1.1) with conditions (3.1) in the domain D bounded by the planes t = 0 and t = T and
in general not cylindrical. We shall suppose that the intersection D
0
of

D and the plane t = 0 is a closure of some
bounded n-dimensional domain with boundary .
Theorem 3. In the domain D there exists a unique solution of the problem (1.1), (3.1), continuous in

D
and satisfying (1.1) in

D if the following conditions are fullled:
(1) Supposition (1) of Theorem 1.
(2) Supposition (2) of Theorem 2 (with

Q changed to

D).
(3) (x) C(D
0
), (x, t) C(

S).
(4) The consistency condition (3.2) holds.
(5) For any point P
0
(x
0
, t
0
) S there exists an (n+1)-dimensional sphere having only one common point P
0
with

D, and the radius of this sphere directed to P
0
is not parallel to the t axis.
Proof. Let U(x, t) be an arbitrary function, continuous in

D and satisfying (3.1). We shall approximate
the coecients a
ij
, b
i
, c, and f of (1.1) by sequences of uniformly bounded innitely dierentiable functions a
m
ij
,
b
m
i
, c
m
, and f
m
converging uniformly in any closed domain

D



D together with the corresponding derivatives
of the form (3.19). Let us construct the sequence of step-solids D
m
(m = 1, 2, . . .), consisting of the cylinders
Q
m
l
=
m
l

_
(l1)T
2
m
,
lT
2
m

. Here the boundaries of n-dimensional domains


m
l
belong to the class A
6+
for all
m and l, l = 1, 2, . . . , 2
m
; we have D
m+1
D
m
and

m=1
D
m
= D.
Let us denote by
m
l
the part of the boundary of the cylinder Q
m
l
, consisting of its lower base t =
(l1)T
2
m
and
its side boundary. According to Theorem 2, for any m in the cylinder Q
m
1
there exists a unique solution u
m
1
(x, t) of
Eq. (3.22) satisfying the condition
u
m
1
(x, t)[

m
1
= U(x, t)[

m
1
.
The function u
m
1
(x, t) is innitely dierentiable and satises Eq. (3.22) everywhere in

Q
m
1

m
1
, in particular for
t =
T
2
m
, x
m
1
.
Now consider Eq. (3.22) in the cylinder Q
m
2
with conditions u
m
2
(x, t) = u
m
1
(x, t) on the part of
m
2
where
the function u
m
1
is dened and u
m
2
(x, t) = U(x, t) on the rest of the boundary
m
2
. According to Theorem 2, in Q
m
2
there exists a unique solution u
m
2
(x, t) of Eq. (3.22) satisfying these conditions. This function u
m
2
(x, t) is innitely
473
dierentiable; it satises Eq. (3.22) in

Q
m
2

m
2
, and also for t =
T
2
m
, x
m
2

m
1
. In this way the solutions
u
m
1
and u
m
2
are glued to each other continuously together with their derivatives of all orders with respect to x
for t =
T
2
m
. It follows from Eq. (3.22) that their derivatives containing dierentiation with respect to t are also
continuous for t =
T
2
m
.
In the same way we construct functions u
m
l
(x, t) in the cylinders Q
m
l
(l = 3, . . . , 2
m
). Set u
m
(x, t) = u
m
l
(x, t)
in Q
m
l
(l = 1, 2, . . . , 2
m
); we obtain a function u
m
(x, t), that is innitely dierentiable in D
m
, coincides with U(x, t)
at the points of the boundary of D
m
that belong to

m
l
, and satises Eq. (3.22) at all other points of

D
m
.
We can perform this for m = 1, 2, . . . Using S. N. Bernsteins estimates, we can establish the compactness in
the sense of uniform convergence of the sequence u
m
(x, t) and its derivatives
u
m
xi
,

2
u
m
xi xj
, and
u
m
t
(i, j = 1, . . . , n)
in any closed subdomain

D



D . Using the diagonal process, we can select a subsequence u
ms
(x, t) that
converges as m
s
to a function u(x, t) satisfying Eq. (1.1) in

D .
Using Remark 2 to Theorem 11 of Sect. 1 and the barrier functions, we can prove that u(x, t) satises
conditions (3.1) on . For a point P
1
(x
1
, 0) D
0
we can again use the barrier (3.26). If the point P
0
(x
0
, t
0
) S,
then, according to condition (5) of this theorem there exists a sphere with center (x

, t

), x

,= x
0
, having only
one common point P
0
with

D. Let us take a neighborhood of P
0
so small that for all points in this neighborhood
the inequality
n

i=1
(x

i
x
i
)
2
> 0 holds. Then in this neighborhood the function similar to (3.25)
v
P0
(x, t) =
K
0

K
, (3.27)
where = (x, t) =
_
n

i=1
(x
i
x

i
)
2
+(tt

)
2
_
1/2
,
0
= (x
0
, t
0
), and K > 0 is suciently large, has all the properties
of a barrier. In fact, for
n

i=1
(x

i
x
i
)
2
> 0 and c(x, t) < 0 (this does not restrict generality) we have

L(v
P0
) < K
K4
_
M
1
(K + 2)

2
_
< 0
for suciently large K. Conditions (A) and (B) are obviously fullled.
The uniqueness of the solution is proved in Sect. 1.
Remark 1. As in the previous theorem, in Theorem 3 one can omit the consistency condition (3.2), supposing
instead that the functions and are bounded and continuous in and on S respectively.
Remark 2. The suppositions of Theorems 2 and 3 concerning the smoothness of coecients a
ij
, b
i
, and c of
Eq. (1.1) and its right-hand side f are too strong. This is connected with the method of proof, namely with the use
of S. N. Bernsteins estimates for the derivatives of the solution to the parabolic equation. Using A. Friedmans
estimates (see Sect. 2) it is possible to prove the existence of a solution of problem (1.1), (3.1) under the supposition
that the functions a
ij
, b
i
, c, and f are continuous in

D and H older continuous in any closed domain

D



D
and the surface S has the same properties as in Theorem 3 (see [43]).
3. Cauchy problem. In this section, we shall consider the Cauchy problem in the layer H0<t T, xE
n

with the initial condition


u(x, 0) = (x), x E
n
. (3.28)
The solution of problem (1.1), (3.28) will be constructed as the limit of the solutions of the rst boundary-value
problems in a sequence of innitely expanding cylinders. We shall require that the absolute value of the right-hand
side f(x, t) of Eq. (1.1) be bounded in H and the coecient c(x, t) be bounded from above in H. The coecients
a
ij
(x, t) and b
i
(x, t) can grow in an arbitrary way as
n

i=1
x
2
i
. The uniqueness theorems for the Cauchy problem
are proved in subsection 5 of Sect. 1.
Theorem 4. The function u(x, t), continuous and bounded in

H, satisfying for t = 0 the initial condition
(3.28) and satisfying for t > 0 Eq. (1.1) exists under the following suppositions:
(1)
n

i,j=1
a
ij
(x, t)
i

j

n

i=1

2
i
for (x, t)

H and all
1
, . . . ,
n
( = const > 0).
(2) In the layer H there exist continuous derivatives of the form (3.19), and the same derivatives of func-
tions b
i
, c, and f; the coecients a
ij
, b
i
, c, and f are continuous in

H.
(3) The function f(x, t) is bounded in the layer

H.
(4) c(x, t) M < for (x, t)

H.
(5) The function (x) is continuous and bounded in the whole space E
n
.
474
Proof. We shall approximate the coecients a
ij
, b
i
, c, and f of Eq. (1.1) by the sequences of innitely dier-
entiable functions a
m
ij
, b
m
i
, c
m
, and f
m
, converging uniformly with the corresponding derivatives of the form (3.19)
in any closed bounded domain contained in H. We shall suppose that [f
m
(x, t)[ M
1
and c
m
(x, t) M
2
, where
M
1
and M
2
do not depend on m. Denote by
m
the ball
n

i=1
x
2
i
< m
2
; by
m
denote its boundary, and by Q
m
denote the cylinder
m
(0, T). According to Theorem 2, for all m = 1, 2, . . . there exists a solution u
m
(x, t) of
Eq. (3.22) in the cylinder Q
m
satisfying the conditions
u
m
(x, 0) = (x) for x
m
,
u
m
(x, t) = (x) for x
m
, 0 t T.
As follows from estimate (1.9) of Sect. 1, the absolute values [u
m
(x, t)[ are bounded by a constant not depending
on m.
Let us x an arbitrary R > 0 and consider in the cylinder Q
R
the functions u
m
(x, t) for m > R. According
to Theorem 1 of Sect. 2 and the remark to this theorem, the families u
m
,
_
u
m
xi
_
,
_

2
u
m
xi xj
_
, and
_
u
m
t
_
are
compact in the sense of uniform convergence in the cylinder

Q
R
t for all > 0. Let R tend to innity and
tend to zero. By the diagonal process we can select a subsequence u
m
k
(x, t) converging at any point (x, t) H to
a bounded function u(x, t) that satises Eq. (1.1) in H. Using barriers of type (3.26), we can show that the function
u(x, t) is continuous for t = 0 and satises the initial condition (3.28).
Remark. Under the suppositions of Theorem 4 the solution of problem (1.1), (3.28) that we have just
constructed is not in general unique even in the class of bounded functions. To have uniqueness in this class we
have to impose additional restrictions on the growth of the coecients of Eq. (1.1) as
n

i=1
x
2
i
(see Theorem 10
of Sect. 1).
Let us give examples showing that conditions (3) and (4) of Theorem 4 cannot be omitted.
Consider the equation

2
u
x
2

u
t
= x (3.29)
in the strip H
1
0 < t T
1
, < x < + with the initial condition
u(x, 0) = 0, < x < +. (3.30)
It is easy to verify that the unbounded function u(x, t) = xt is a solution of problem (3.29), (3.30). If we suppose
that there is another solution v(x, t) of this problem and it is bounded, then w(x, t) = u(x, t) v(x, t) will satisfy
the heat equation

2
w
x
2

w
t
= 0 in the strip H
1
with zero as the initial condition. This contradicts the uniqueness
theorem for the solution of the Cauchy problem in the class of functions growing not faster than a linear function
as [x[ (Theorem 9 of Sect. 1). So Theorem 4 does not hold without condition (3).
Consider now in H
1
the equation

2
u
x
2
+c(x)u
u
t
= 0 (3.31)
with the initial condition
u(x, 0) = 1, < x < +. (3.32)
The function c(x) will be constructed in the following way. Take an integer k
0
> 0 such that for all k > k
0
the inequality
4

T(2k + 1) < e
k
(e 1)
holds. Then for k > k
0
we shall have
e
k
+ 4

Tk < e
k+1
4

T(k + 1).
For e
k
4

Tk x e
k
+ 4

Tk, set c(x) = e


k
(k = k
0
+1, k
0
+ 2, . . .). For the remaining x dene c(x) so that it
is an innitely dierentiable positive even function, nondecreasing for x 0.
475
For e
k
[x[ e
k+1
(k = k
0
+ 1, k
0
+ 2, . . .) we have
0 c(x) e
k+1
e[x[. (3.33)
We shall show that problem (3.31), (3.32) has no bounded solution. Suppose the contrary: let u(x, t) be
a solution of (3.31), (3.32) such that [u(x, t)[ M. Setting u = ve
te
k
we deduce from (3.31) and (3.32) that
v
t


2
v
x
2
= [c(x) e
k
]v, v(x, 0) = 1. (3.34)
We shall consider Eq. (3.34) as a heat equation with the right-hand side [c(x) e
k
]v. Since u(x, t) is assumed
to be bounded and estimate (3.33) holds, the function [c(x) e
k
]v grows as [x[ not faster than a linear
function. Therefore, the solution of Cauchy problem (3.34) is given by the well-known formula
v(x, t) = 1 +
t
_
0
d
+
_

Z(x, t; , )[c() e
k
]v(, ) d,
where
Z(x, t; , ) = [4(t )]

1
2
e

(x)
2
4(t)
is a fundamental solution of the heat equation. Hence we have
u(x, t) = e
te
k
_
1 +
t
_
0
d
+
_

Z(x, t; , )[c() e
k
]e
e
k
u(, ) d
_
. (3.35)
Thus, the bounded solution of problem (3.31), (3.32) must satisfy (3.35). From (3.35) we can deduce
the inequality
u(x, t) e
te
k
_
1 M
t
_
0
d
+
_

Z(x, t; , )[c() e
k
[ d
_
. (3.36)
We have constructed c(x) so that for [ e
k
[ 4

Tk we have c() e
k
= 0. Setting x = e
k
in (3.36) and
using estimate (3.33), we obtain
u(e
k
, t) e
te
k
_
1 M
t
_
0
d
_
e
k
4

Tk
_

+
+
_
e
k
+4

Tk
_
Z(e
k
, t; , )(e[[ +e
k
) d
_
= e
te
k
_
1
M

t
_
0
d
_
2k
_
T
t
_

+
+
_
2k
_
T
t
_
(e[e
k
+ 2

t [ +e
k
)e

2
d
_
e
te
k
_
1
2M

t
_
0
d
+
_
2k
_
T
t
[e
k
(e + 1) + 2e

t ]e

2
d
_
. (3.37)
The quantity in braces in the right-hand side of the inequality (3.37) tends to unity as k + for any t
from the interval (0, T). In fact,
+
_
2k
_
T
t
e
k
(e + 1)e

2
d < e
k
(e + 1)
+
_
2k
e

d = (e + 1)e
k
0 (k +),
476
+
_
2k
_
T
t
2e

t e

2
d =
+
_
4k
2
T
t
e

t e
s
ds < e

T
+
_
2k
2
e
s
ds = e

Te
4k
2
0 (k +).
Therefore, for suciently large k we have
u(e
k
, t) >
1
2
e
te
k
,
and this contradicts the supposition that u(x, t) is bounded.
4. Second Boundary-Value Problem. One can use the Rothe method also to construct the solution
of the second boundary-value problem for Eq. (1.1) in the cylinder Q = (0, T). Boundary conditions for this
problem have the form
u(x, 0) = (x), (3.38)
n

i,j=1
a
ij
(x, t) cos(, x
i
)
u
x
j
+a(x, t)u = 0 on S, (3.39)
where is the direction of the inner normal to S.
Theorem 5. There exists in the cylinder Q a unique function u(x, t), continuous in

Q together with its
rst-order derivatives with respect to x
1
, . . . , x
n
, satisfying boundary conditions (3.38), (3.39) on having in

Q continuous second-order derivatives with respect to x


1
, . . . , x
n
, rst-order derivatives with respect to t, and
satisfying Eq. (1.1), if the following suppositions are fullled:
(1) Supposition (1) of Theorem 1.
(2) The functions a
ij
, b
i
, c, and f belong to the class C
8+,4+
(

Q) (see Sect. 1).
(3) The boundary of domain belongs to the class A
6+
.
(4) (x) C
6+
(

); a(x, t) C
8+,4+
as a function of local coordinates
1
, . . . ,
n1
, t on the surface

S.
(5) The consistency condition
_
n

i,j=1
a
ij
(x, 0) cos(, x
i
)
(x)
x
j
+a(x, 0)(x)
_
x
= 0 (3.40)
is fullled.
Proof. As we have already noted when we proved Theorem 1, one can suppose without loss of generality
that c(x, t) < 0 in

Q. We can also suppose that a(x, t) < 0 (see the remark to Theorem 7 of Sect. 1). Consider
the system of equations (3.4) with the boundary conditions
n

i,j=1
a
ij
(x, kt) cos(, x
i
)
u(x, kt)
x
j
+a(x, kt)u(x, kt) = 0 (k = 1, . . . , m). (3.41)
For k = 0 we set u(x, 0) = (x). It is well known that for every k there exists a unique solution u(x, kt) of
the elliptic equation (3.4) satisfying the condition (3.41), and the solution u(x, kt) C
6+
(

) (see, for example,


[66], [2] and Theorem 7 from [63]).
We shall show that the functions u(x, kt) are uniformly bounded in the cylinder

Q with respect to t.
As in Theorem 1, it is possible to estimate [u(x, kt)[ at an inner point where [u[ attains its maximum. It is
obvious that [u(x, kt)[ is uniformly bounded for k = 0. Finally it follows from (3.41) that u(x, kt) as a function
of x cannot attain its positive maximum or negative minimum on S. In fact, if the positive maximum or negative
minimum was attained on S, according to the well-known property of solutions of elliptic equations (see [21]) we
would have at the point where the positive maximum of u(x, kt) is attained
u(x, kt) > 0,
n

i,j=1
a
ij
(x, kt) cos(, x
i
)
u(x, kt)
x
j
< 0,
477
and this is impossible because of (3.41). Similarly, one can prove that u(x, kt) cannot have a negative minimum
on S.
Let us estimate the derivatives
u(x,kt)
x
l
(l = 1, . . . , n). By S. N. Bernsteins method one can prove the uni-
form boundedness of
n

l=1
_
u(x,kt)
x
l

2
in any closed domain

Q



Q S. To do this we have to use an auxiliary
function of the form (3.17) (see the remark to Theorem 1 of this section).
Now we shall prove that the derivatives
u(x,kt)
x
l
are uniformly bounded in a neighborhood of S. Let
p
be
a piece of allowing a parametric representation
x
i
= x
ip
(
1
, . . . ,
n1
) (i = 1, . . . , n), where x
ip
C
6+
.
In the cylinder Q
p
= G
p
(0, T), where G
p
is an intersection of a neighborhood of
p
with the domain , we shall
introduce a new coordinate system (
1
, . . . ,
n
, t) given by formulas
x
i
= x
ip
(
1
, . . . ,
n1
) +
n
h
i
(
1
, . . . ,
n1
, t) (i = 1, . . . , n);
here
h
i
(
1
, . . . ,
n1
, t) =
1
A(x, t)
n

j=1
a
ij
(x, t) cos(, x
j
),
A(x, t) =
_
n

i=1
_
n

j=1
a
ij
(x, t) cos(, x
j
)
_
2
_
1/2
.
The boundary condition (3.41) will have the following form on the piece S
p
=
p
[0, T]:
_
u(, kt)

n
+
a(, kt)
A(, kt)
u(, kt)
_
n=0
= 0.
Set u(, kt) = v(, kt)(, kt), where
(, t) = 1
n
_
a(, t)
A(, t)

n=0
+
_
( > 0)
for 0
n
( > 0 is suciently small); for
n
the function (, t) is extended in a smooth way. The functions
v(, kt) in the cylinder Q
p
satisfy the equations of the form

L
k
(v(, kt)) =
n

i,j=1
a
ij
(, kt)

2
v

j
+
n

i=1

b
i
(, kt)
v

i
+ c(, kt)v
v(, kt) v(, (k 1)t)
t
=

f(, kt) (3.42)
and the boundary condition
_
v(, kt)

n
v(, kt)
_
n=0
= 0. (3.43)
Let us take an arbitrary point
1
= (
1
1
, . . . ,
1
n1
, 0) and consider in the cylinder Q

p
_
0
n
,
n

i=1
(
i

1
i
)
2

2
, 0 t T
_
an auxiliary function
z(, k) =
1
()
_
n1

l=1
_
v

l
_
2
+
_
v

n
v
_
2
_
+Nv
2
+e
N1(Tkt)
+
n
,
478
where
1
() =
_

n1

i=1
(
i

1
i
)
2
_
2
, and N, N
1
, and are positive constants. Using Eq. (3.42) and equations
obtained by dierentiating (3.42) with respect to
l
, it is possible to prove (as was done in Theorem 1) that for
suciently large N and N
1
inside Q

p
(and for t = T) the inequality

L
k
(z(, k)) < 0 holds for any t and k. It
follows that the function z(, k) cannot attain its maximal value in Q

p
inside Q

p
(or for t = T).
So we have proved that the function z(, k) is uniformly bounded with respect to t for
n
= and for
n1

i=1
(
i

1
i
)
2
=
2
. Obviously z(, k) is uniformly bounded also for t = 0.
Suppose that z(, k) attains its maximal value at a point (
1
, . . . ,
n
, t), with
n
= 0. Then the function
z

(, k) =
1
()
n1

l=1
_
v

l
_
2
+Nv
2
+e
N1(Tkt)
+
n
also attains its maximal value at this point, because it follows from (3.43) that
v
n
v = 0 for
n
= 0. Therefore
at this point we have
z

n
0. On the other hand, by dierentiating (3.43) with respect to
l
, we obtain

2
v

n=0
=
v

n=0
(l = 1, . . . , n 1). (3.44)
Now, taking into account (3.42) and (3.44), we have
z

n=0
= 2
1
()
n1

l=1
_
v

l
_
2

n=0
+ 2Nv
2
[
n=0
+ > 0.
This contradiction proves that the function z(, k) cannot attain its maximum for
n
= 0.
So we have proved that z(, k) is uniformly bounded in the cylinder Q

p
. The boundedness of the derivatives
u(x,kt)
x1
,. . . ,
u(x,kt)
xn
in the cylinder Q

p
, where 0 <

< , follows. Since the piece


p
and the point
1
are
arbitrary, we have proved that these derivatives are uniformly bounded everywhere in

Q.
Now we have to estimate in

Q the derivatives

2
u(x,kt)
x
l
xq
(l, q = 1, . . . , n) and the dierence quotient
u(x,kt)
t
=
u(x,kt)u(x,(k1)t)
t
. Using S. N. Bernsteins method and the function
w(x, k) =
2
(x)
n

l,q=1
_

2
u(x, kt)
x
l
x
q
_
2
+N
2
n

l=1
_
u(x, kt)
x
l
_
2
+e
N3(Tkt)
,
where
2
(x) =
_

i=1
(x
i
x
0
i
)
2
_
2
, x
0
, we shall prove that
n

l,q=1
_

2
u(x,kt)
x
l
xq

2
is uniformly bounded in any
closed domain

Q



Q S. As follows from Eqs. (3.4), thus we have proved also the uniform boundedness in

Q

of
the dierence quotient
u(x,kt)
t
. Then we extend the function u(x, kt) to k = 1 by formula (3.14) and consider
in the cylinder Q

p
the auxiliary function
z
1
(, k) =
1
()
_
n1

l,q=1
_

2
v

l

q
_
2
+
n1

l=1
_

2
v

l
_
2
+
_
v
t
_
2
_
+N
4
_
n1

l=1
_
v

l
_
2
+
_
v

n
v
_
2
_
+e
N5(Tkt)
+
n
.
Using this function and taking into account the consistency condition (3.40), we shall show as we did before that
in Q

p
, 0 <

< , the derivatives



2
u
i j
(i = 1, . . . , n; j = 1, . . . , n1) and the dierence quotient
u
t
are uniformly
bounded. It follows from Eqs. (3.14) that the derivative

2
u

2
n
in Q

p
also is bounded uniformly with respect to t. In
this way we have proved that in

Q the functions

2
u(x,kt)
xi xj
(i, j = 1, . . . , n) and
u(x,kt)
t
are uniformly bounded.
479
Then, using the function
w
1
(x, k) =
2
(x)
n

l,q,r=1
_

3
u(x, kt)
x
l
x
q
x
r
_
2
+N
6
n

l,q=1
_

2
u(x, kt)
x
l
x
q
_
2
+e
N7(Tkt)
,
we shall estimate the derivatives

3
u(x,kt)
x
l
xqxr
(l, q, r = 1, . . . , n) in

Q



Q S. The equations obtained by dierenti-
ating (3.4) with respect to x
l
allow us to obtain an estimate in

Q

for the functions



x
l
_
u(x,kt)
t

(l = 1, . . . , n).
Using the auxiliary function
z
2
(, k) =
1
()
_
n1

l,q,r=1
_

3
v

r
_
2
+
n1

l,q=1
_

3
v


2
v

q
_
2
+
n1

l=1
_

l
_
v
t
__
2
+
_

n
_
v
t
_

v
t
_
2
_
+N
8
_
n1

l,q=1
_

2
v

q
_
2
+
n1

l=1
_

2
v

l
_
2
+
_
v
t
_
2
_
+e
N9(Tkt)
+
n
,
we can show in the same way as we did before that the functions

3
u(x,kt)
x
l
xqxr
and

x
l
_
u(x,kt)
t

(l, q, r = 1, . . . , n)
are uniformly bounded in

Q.
After that, using Bernsteins method, we can prove that the derivatives

4
u(x,kt)
x
l
xqxrxs
(l, q, r, s = 1, . . . , n) are
uniformly bounded in any closed domain

Q



QS. Equations obtained by twice dierentiating (3.4) with respect
to x
1
, . . . , x
n
give the uniform boundedness of functions

2
x
l
xq
_
u(x,kt)
t

(l, q = 1, . . . , n) in any closed domain



Q S.
Let us extend the functions u(x, kt) by linear interpolation for kt t (k + 1)t. As we have proved,
the families obtained in this way, u
t
(x, t) and
_
ut(x,t)
x
l
_
(l = 1, . . . , n), are compact in

Q and the families
_

2
ut(x,t)
x
l
xq
_
(l, q = 1, . . . , n) and
ut(x,t)
t
are compact in any closed domain contained in

Q S together with its
boundary. As in the proof of Theorem 1 we can see that the sequence u
tm
(x, t) uniformly converges for t
m
0
in

Q to a function u(x, t). This limit function satises Eq. (1.1) in

Q and conditions (3.38) and (3.39) on .
The uniqueness of the solution of problem (1.1), (3.38), (3,39) is proved in Sect. 1.
Note that one can obtain estimates for the solution of the second boundary-value problem (1.1), (3.38),
(3.39) in the same way as we did for the functions u(x, kt) and their derivatives and dierence quotients.
4 THE FUNDAMENTAL SOLUTION OF A LINEAR PARABOLIC EQUATION. GREENS FUNCTION.
METHOD OF INTEGRAL EQUATIONS FOR SOLVING BOUNDARY-VALUE PROBLEMS
In the study of linear partial dierential equations, the so-called fundamental solutions, that is, solutions
having singularities of a prescribed type, are of great importance. Many works are dedicated to the construction
of fundamental solutions for parabolic equations and systems and to their applications for solving boundary-value
problems and the Cauchy problem. A fundamental solution for a general second-order parabolic equation with many
independent variables and smooth coecients was rst constructed in [67, 68]. For parabolic equations of special
form, fundamental solutions were considered in [6971] and others. In W. Pogorzelskis papers [72,73] a fundamental
solution is constructed for a second-order parabolic equation with only H older continuous coecients. For systems
of equations parabolic in the sense of I. G. Petrovskiy, fundamental solutions are constructed in [27, 74, 75]. In all
these papers the classical method of E. E. Levi [76] is used for the construction of fundamental solutions.
In this section, we construct a fundamental solution for the equation L(u) = 0 and establish some of its
important properties. With the help of fundamental solutions we obtain solutions for the Cauchy problem and
the rst boundary-value problem for Eq. (1.1) under the supposition that the coecients and the function f(x, t)
are only H older continuous.
1. Construction of a Fundamental Solution. As is known, the function
w(x, t; , ) =
1
2

(t )
e

(x)
2
4(t)
480
satises the heat equation

2
u
x
2

u
t
= 0 (4.1)
for t > and has the following property: for any bounded and continuous function (x) dened for all x we have
lim
t+0
+
_

w(x, t; , )() d = (x), (4.2)


uniformly on any nite interval. For a more general equation
n

i,j=1
a
ij

2
u
x
i
x
j

u
t
= 0, (4.3)
with coecients a
ij
independent of x and t, a solution with similar properties also can be constructed explicitly.
In what follows, we shall suppose that the coecients a
ij
in Eq. (4.3) are functions of parameters and ,
where = (
1
, . . . ,
n
) E
n
, 0 T. Denote by a
ij
(, ) elements of the matrix inverse to |a
ij
(, )|; A(, )
will denote the determinant of the matrix |a
ij
(, )|. Then let us set
(, ; s) = (, ; s
1
, . . . , s
n
) =
n

i,j=1
a
ij
(, )s
i
s
j
,
W
,
(x, t; , ) = [4(t )]

n
2
[A(, )]

1
2
e

(,;x)
4(t)
. (4.4)
By direct calculation one can show that W
,
(x, t; , ) as a function of variables x, t satises Eq. (4.3) for t > .
Suppose the functions a
ij
(x, t) and (x, t) are continuous and bounded in the layer

H0 t T, x E
n
, and
n

i,j=1
a
ij

j

n

i=1

2
i
, > 0. (4.5)
We shall demonstrate that under these suppositions for W
,
(x, t; , ) a formula similar to (4.2) holds:
lim
t+0
_
En
W
,
(x, t; , )(, t) d = (x, ) (0 < T). (4.6)
First suppose that a
ij
and are constant. The change of variables

i
= x
i
+
n

j=1

ij

j
(i = 1, . . . , n) (4.7)
transforms the quadratic form (x ) =
n

i,j=1
a
ij
(x
i

i
)(x
j

j
) to the canonical form =
n

i=1

2
i
. The determi-
nant of transformation (4.7) obviously is (det |a
ij
|)

1
2
=

A. Hence
I(x, t, )
_
En
W
,
(x, t; , )d =
_
En
[4(t )]

n
2
e

i=1

2
i
4(t)
d.
Passing to polar coordinates, we obtain
I(x, t, ) = (4)

n
2

n
+
_
0
(t )

n
2
r
n1
e

r
2
4(t)
dr,
481
where
n
=
2
n
2
(
n
2
)
is the area of the n-dimensional space. Setting
r
2
4(t)
= q, we obtain
I(x, t, ) =
4

n
2

_
n
2
_
+
_
0
4
n
2
q
n
2
1
e
q
dq =
for all t and , t > . Thus we have proved (4.6) for constant a
ij
and . Now let a
ij
and not be constant. We
have
I(x, t, ) = (x, t)
_
En
W
x,
(x, t; , ) d +(x, t)
_
En
[W
,
(x, t; , ) W
x,
(x, t; , )] d
+
_
En
W
,
(x, t; , )[(, ) (x, t)] d = I
1
+I
2
+I
3
. (4.8)
As we have proved,
lim
t+0
I
1
= (x, );
in fact, according to (4.4), the coecients a
ij
appearing in the formula for W
x,
(x, t; , ) are taken at a point (x, ).
To estimate the function under the integral sign in I
2
we shall use the mean-value theorem. We obtain
[W
,
(x, t; , ) W
x,
(x, t; , )[ (t )

n
2
e

|x|
2
t

_
M
1
[x [
2
t
max
i,j
[a
ij
(, ) a
ij
(x, )[ +M
2
[A(, ) A(x, )[
_
,
where

> 0 and [x [ =
_
n

i=1
(x
i

i
)
2
_1
2
. Using the elementary inequality
q

q
K

e
1q
(0 q < +, 0 <
1
<

), (4.9)
which holds for all 0, we obtain
[W
,
(x, t; , )W
x,
(x, t; , )[ (t)

n
2
e

1
|x|
2
t
M
3
max
i,j
[a
ij
(, )a
ij
(x, )[+M
2
[A(, )A(x, )[. (4.10)
Let us represent the integral I
2
as a sum I
(1)
2
+I
(2)
2
, where I
(1)
2
is taken over the ball of radius and with center = x.
The integral I
(2)
2
is taken over the remaining part of the space
1
, . . . ,
n
. Taking into account the estimate (4.10)
and the continuity of the functions a
ij
(x, t) and A(x, t), one can see that the absolute value of the integral I
(1)
2
is
arbitrarily small for suciently small . If is xed and t > 0 is suciently small, the integral I
(2)
2
also can be
made arbitrarily small in absolute value, because according to (4.10) it is uniformly converging with respect to x
and the function under the integral sign uniformly tends to zero as t +0. Thus we have lim
t+0
I
2
= 0. Similarly
we obtain lim
t+0
I
3
= 0. This proves (4.6).
Using similar arguments it is also possible to prove that
lim
t0
_
En
W
,
(x, t; , )(, ) d = (x, t) (4.11)
under the same suppositions on a
ij
and .
Now we shall consider a general linear second-order homogeneous parabolic equation in the layer H:
L
x,t
u
n

i,j=1
a
ij
(x, t)

2
u
x
i
x
j
+
n

i=1
b
i
(x, t)
u
x
i
+c(x, t)u
u
t
= 0. (4.12)
482
A function Z(x, t; , ) is called a fundamental solution of Eq. (4.12) in the layer H if it has the following
properties:
(1) The function Z(x, t; , ) in the domain 0 < t T, x E
n
, E
n
is continuous jointly in x,
t, , and together with its derivatives

2
Z
xi xj
(i, j = 1, 2, . . . , n) and
Z
t
and satises Eq. (4.12) with respect to
variables x and t. The function Z(x, t; , ) is bounded in any domain t +[x [ , where > 0.
(2) For any continuous and bounded function () in E
n
and for all x E
n
, [0, T) the relation
lim
t+0
_
En
Z(x, t; , )() d = (x) (4.13)
holds, and the integral tends to its limit uniformly with respect to x in any bounded domain of the space E
n
.
Theorem 1. Suppose all the coecients of Eq. (4.12) are bounded and continuous in

H jointly in x and t
and are H older continuous with respect to x:
[a
ij
(x

, t) a
ij
(x, t)[ M
4
[x

x[

,
[b
i
(x

, t) b
i
(x, t)[ M
4
[x

x[

,
[c(x

, t) c(x, t)[ M
4
[x

x[

_
(i, j = 1, . . . , n; > 0). (4.14)
Suppose additionally that the coecients a
ij
in

H are Holder continuous with respect to t:
[a
ij
(x, t

) a
ij
(x, t)[ M
4
[t

t[

. (4.15)
Let us also suppose that the inequality (4.5) holds.
Then there exists a unique fundamental solution Z(x, t; , ) of Eq. (4.12) in the layer H. For Z(x, t; , )
the estimates
[Z(x, t; , )[ < M
5
(t )

n
2
e

1
|x|
2
t
,

Z(x, t; , )
x
i

< M
5
(t )

n+1
2
e

1
|x|
2
t
,

2
Z(x, t; , )
x
i
x
j

< M
5
(t )

n
2
1
e

1
|x|
2
t
,

Z(x, t; , )
t

< M
5
(t )

n
2
1
e

1
|x|
2
t
_

_
(4.16)
hold; here M
5
and
1
are positive constants. The function Z(x, t; , ) is positive everywhere for t > . If in

H there
exist bounded and continuous derivatives
aij
xj
,

2
aij
xi xj
, and
bi
xi
(i, j = 1, . . . , n), H older continuous with respect
to x, then for t > the function Z(x, t; , ) as a function of variables , satises the equation conjugate to (4.12):
L

,
(Z)
n

i,j=1

2
[a
ij
(, )Z]

i=1
[b
i
(, )Z]

i
+ c(, )Z +
Z

= 0. (4.17)
Proof. We shall try to nd the fundamental solution in the form
Z(x, t; , ) = W
,
(x, t; , ) +
t
_

d
_
En
W
,
(x, t; , )(, ; , ) d, (4.18)
where the function W
,
(x, t; , ) is given by (4.4) and the function (x, t; , ) is to be found.
First we shall study the properties of the improper integral
V (x, t; , ) =
t
_

d
_
En
W
,
(x, t; , )(, ; , ) d. (4.19)
483
According to (4.4), for the function W
,
(x, t; , ) we have the following estimate:
[W
,
(x, t; , )[ < M
6
(t )

n
2
e

1
|x|
2
t
,
1
> 0. (4.20)
Using the inequality (4.9) we obtain estimates for the derivatives of W
,
(x, t; , ):

W
,
(x, t; , )
x
i

< M
6
(t )

n+1
2
e

2
|x|
2
t
, (4.21)

2
W
,
(x, t; , )
x
i
x
j

< M
6
(t )

n
2
1
e

2
|x|
2
t
, (4.22)

W
,
(x, t; , )
t

< M
6
(t )

n
2
1
e

2
|x|
2
t
; (4.23)
here
1
>
2
> 0.
Suppose that the function (x, t; , ) is continuous jointly in x, t, , for t > and for any x, and can be
estimated in the following way:
[(x, t; , )[ < M
7
(t )

n
2
1+1
e

3
|x|
2
t
, (4.24)
where
1
> 0 and
3
> 0. Moreover, we shall suppose that for [x

x[
2
< a(t ), where a > 0 is a constant, we
have
[(x

, t; , ) (x, t; , )[ M
7
[x

x[
2
(t )

n
2
1+3
e

3
|x|
2
t
, (4.25)
where
2
> 0 and
3
> 0. Let us show that under such suppositions the function (4.19) is continuous and has
continuous derivatives
V
xi
,

2
V
xi xj
(i, j = 1, . . . , n), and
V
t
for t > and any x and .
According to our suppositions, the function
J(x, t, ; , ) =
_
En
W
,
(x, t; , )(, ; , ) d (4.26)
is continuous with respect to all its arguments and has derivatives of any order with respect to x and t for < < t
and any x and . Taking into account (4.20) and (4.24), we obtain
[J(x, t, ; , )[ < M
8
_
En
(t )

n
2
e

1
|x|
2
t
( )

n
2
1+1
e

3
||
2

d
< M
8
( )
11
n

i=1

[(t )( )]

1
2
e
4
_
(x
i

i
)
2
t
+
(
i

i
)
2

_
d
i
= M
8
( )
11
n

i=1
_

1
2
[
4
(t )]

1
2
e
4
(x
i

i
)
2
t
_
= M
9
( )
11
(t )

n
2
e

4
|x|
2
t
. (4.27)
Therefore, for t > 0, where > 0 is arbitrary, and for any x and the inequality
[J(x, t, ; , )[ < M
(1)

( )
11
(4.28)
holds. This allows us to establish the uniform convergence of the integral
t
_

J(x, t, ; , ) d with respect to x, t, ,


and . Thus, we have established that the function V (x, t; , ) is continuous jointly in all its arguments for t >
and any x and . According to (4.27), we have
[V (x, t; , )[
t
_

[J(x, t, ; , )[ d < M
10
(t )

n
2
+1
e

4
|x|
2
t
. (4.29)
484
Using (4.21) and (4.24), with the help of similar arguments we obtain the estimate

J(x, t, ; , )
x
i

_
En

W
,
(x, t; , )
x
i

[(, ; , )[ d < M
11
(t )

1
2
( )
11
(t )

n
2
e

4
|x|
2
t
. (4.30)
Hence

J(x, t, ; , )
x
i

< M
(2)

(t )

1
2
( )
11
for t > 0 and any x and . Therefore the integral
t
_

J(x, t, ; , )
x
i
d
converges uniformly with respect to x, t, , and in the above-mentioned domain.
According to the well-known theorem (see [77]), it follows that for t > and any x and the derivatives
V
x
i
=
t
_

d
_
En
W
,
(x, t; , )
x
i
(, ; , ) d (i = 1, . . . , n) (4.31)
exist and are continuous. According to the inequality (4.30), for the derivatives
V
xi
we have the estimate

V (x, t; , )
x
i

< M
12
(t )

n+1
2
+1
e

4
|x|
2
t
. (4.32)
To prove the existence of the derivatives

2
V
xi xj
(i, j = 1, . . . , n) consider for t > 0 the integral
F(x, t; , ) =
t
_

d
_
En

2
W
,
(x, t; , )
x
i
x
j
(, ; , ) d =
t
_

2
J(x, t, ; , )
x
i
x
j
d.
It can be represented as a sum:
F(x, t; , ) =
+

2
_

2
J(x, t, ; , )
x
i
x
j
d +
t
_
+

2
J(x, t, ; , )
x
i
x
j
d = F
1
(x, t; , ) +F
2
(x, t; , ).
Using the estimates (4.22) and (4.24), we get

2
J(x, t, ; , )
x
i
x
j

< M
13
_
En
(t )

n
2
1
( )

n
2
1+1
e

2
|x|
2
t

3
||
2

d
< M
14
(t )
1
( )
11
(t )

n
2
e

4
|x|
2
t
. (4.33)
Therefore, for < < +

2
the inequality

2
J(x, t, ; , )
x
i
x
j

< M
(3)

( )
11
holds, and it follows that the integral F
1
converges uniformly in the domain t > 0, x E
n
, E
n
.
485
To estimate the function under the integral sign in F
2
, we shall rst transform
J(x,t,;,)
xi
in the following
way:
J
x
i
=
_
En
W
,
(x, t; , )
x
i
(, ; , ) d = (y, ; , )
_
En
W
y,
(x, t; , )
x
i
d
+ (y, ; , )
_
En
_
W
,
(x, t; , )
x
i

W
y,
(x, t; , )
x
i
_
d +
_
En
[(, ; , ) (y, ; , )]
W
,
(x, t; , )
x
i
d;
(4.34)
here y is an arbitrary xed point. Let x be located inside the ball K of radius
1
2
with center at an arbitrary
point x
0
. In the rst term of the right-hand side of (4.34) we can single out the integral over a ball K
1
of radius 1,
concentric to the ball K, and apply to this integral the GaussOstrogradski formula, taking into account that
W
y,
(x,t;,)
xi
=
W
y,
(x,t;,)
i
. We obtain
J
x
i
= (y, ; , )
_
1
W
y,
(x, t; , ) cos(,
i
) dS

+ (y, ; , )
_
En\K1
W
y,
(x, t; , )
x
i
d
+ (y, ; , )
_
En
_
W
,
(x, t; , )
x
i

W
y,
(x, t; , )
x
i
_
d +
_
En
[(, ; , ) (y, ; , )]
W
,
(x, t; , )
x
i
d;
(4.35)
here
1
is the boundary of the ball K
1
and is the outer normal to
1
. Dierentiating (4.35) with respect to x
i
and then setting y = x, we get

2
J
x
i
x
j
= (x, ; , )
_
1
_
W
y,
(x, t; , )
x
j
_
y=x
cos(,
i
) dS

+ (x, ; , )
_
En\K1
_

2
W
y,
(x, t; , )
x
i
x
j
_
y=x
d
+ (x, ; , )
_
En
_

2
W
,
(x, t; , )
x
i
x
j

2
W
y,
(x, t; , )
x
i
x
j
_
y=x
_
d
+
_
En
[(, ; , ) (x, ; , )]

2
W
,
(x, t; , )
x
i
x
j
d = I
1
+I
2
+I
3
+I
4
. (4.36)
Since the inequality (4.21) holds, we have for the function under the integral sign in I
1
for x K,
1
the estimate

_
W
y,
(x, t; , )
x
j
_
y=x
cos(,
i
)

< M
15
(t )

n+1
2
e


2
4(t)
< M
16
.
Taking into account additionally (4.24), we get
[I
1
[ < M
16

n
M
7
( )

n
2
1+1
e

3
|x|
2

< M
17
( )

n
2
1+1
e

3
|x|
2

. (4.37)
(Here
n
is the area of the unit sphere in the space E
n
.) Hence
[I
1
[ < M
(4)

for +

2
< < t, x K, E
n
. (4.38)
Then with the help of the inequalities (4.22), (4.24), and (4.9), we can estimate [I
2
[:
486
[I
2
[ < M
18
( )

n
2
1+1
e

3
|x|
2

_
|x|
1
2
(t )

n
2
1
e

2
|x|
2
t
d
M
19
( )

n
2
1+1
e

3
|x|
2
t
_
|x|
1
2
(t )

n
2
[x [
2
e

2
2
|x|
2
t
d
4M
19

n
( )

n
2
1+1
e

3
|x|
2
t
+
_
1
2
(t )

n
2
r
n1
e

2
2
r
2
t
dr
< M
20
( )

n
2
1+1
e

3
|x|
2
t
+
_
0
q
n
2
1
e

2
2
q
dq = M
21
( )

n
2
1+1
e

3
|x|
2
t
. (4.39)
It follows that
[I
2
[ < M
(5)

for +

2
< < t, x K, E
n
. (4.40)
Applying the mean-value theorem and using the inequalities (4.9), (4.14), and (4.22), we come to the following
estimate for the function under the integral sign in I
3
:

2
W
,
(x, t; , )
x
i
x
j

2
W
y,
(x, t; , )
x
i
x
j
_
y=x

< M
22
(t )

n
2
1+

2
e

4
|x|
2
t
. (4.41)
From (4.24) and (4.41) we obtain
[I
3
[ < M
23
( )

n
2
1+1
(t )

2
1
e

3
|x|
2
t
. (4.42)
It follows that
[I
3
[ < M
(6)

(t )

2
1
for +

2
< < t, x K, E
n
. (4.43)
The integral I
4
can be represented as the sum I
4
= I
(1)
4
+I
(2)
4
, where the integral I
(1)
4
is taken over the domain
dened by the inequality [ x[
2
< a( ). Taking into account (4.22) and (4.25), we obtain
[I
(1)
4
[ < M
24
_
|x|
2
<a()
[ x[
2
( )

n
2
1+3
e

3
||
2

(t )

n
2
1
e

2
|x|
2
t
d
< M
25
_
En
(t )

n
2
1+

2
2
( )

n
2
1+3
e
4
_
|x|
2
t
+
||
2

_
d.
Using arguments similar to those used when we deduced the inequality (4.27), we get
[I
(1)
4
[ < M
26
(t )

2
2
1
( )
31
(t )

n
2
e

4
|x|
2
t
. (4.44)
Therefore, for +

2
< < t, x K, E
n
the estimate
[I
(1)
4
[ < M
(7)

(t )

2
2
1
(4.45)
is valid. The integral I
(2)
4
can be estimated with the help of inequalities (4.22) and (4.24):
487
[I
(2)
4
[ < M
27
_
|x|
2
a()
( )

n
2
1+1
e

3
||
2

(t )

n
2
1
e

2
|x|
2
t
d
< M
27
_
En
( )

n
2
1+

1
2
e

3
||
2

1
2
[ x[
1
(t )

n
2
1
e

2
|x|
2
t
d
< M
28
_
En
( )

n
2
1+

1
2
(t )

n
2
1+

1
2
e
4
_
|x|
2
t
+
||
2

_
d
< M
29
(t )

1
2
1
( )

1
2
1
(t )

n
2
e

4
|x|
2
t
. (4.46)
If +

2
< < t, x K, E
n
this leads to
[I
(2)
4
[ < M
(8)

(t )

1
2
1
. (4.47)
It follows from relations (4.38), (4.40), (4.43), (4.45), and (4.47) that the integral
F
2
(x, t; , ) =
t
_
+

2
J(x, t, ; , )
x
i
x
j
d
converges uniformly in the domain t , x K, E
n
. We have shown before that the integral
F
1
(x, t; , ) =
+

2
_

2
J(x, t, ; , )
x
i
x
j
d
also converges uniformly in this domain. Thus, we have proved the uniform convergence of the integral F(x, t; , ) =
F
1
(x, t; , ) +F
2
(x, t; , ) in this domain. Therefore in this domain there exist continuous derivatives

2
V (x, t; , )
x
i
x
j
=
t
_

d
_
En

2
W
,
(x, t; , )
x
i
x
j
(, ; , ) d (i, j = 1, . . . , n). (4.48)
Since the number > 0 and the point x
0
(the center of the ball K) are arbitrary, equality (4.48) holds everywhere
in the domain 0 < t T, x E
n
, E
n
.
Because of inequalities (4.33), (4.37), (4.39), (4.42), (4.44), and (4.46), we get

2
V (x, t; , )
x
i
x
j

2
_

2
J(x, t, ; , )
x
i
x
j

d +
t
_
+

2
J(x, t, ; , )
x
i
x
j

d < M
30
(t )

n
2
1+4
e

4
|x|
2
t
, (4.49)
where
4
> 0.
Now we shall show that for t > and any x and the function V (x, t; , ) has a continuous derivative
V
t
,
which can be calculated by the following formula:
V (x, t; , )
t
= (x, t; , ) +
t
_

d
_
En
W
,
(x, t; , )
t
(, ; , ) d = (x, t; , ) +
t
_

J(x, t, ; , )
t
d. (4.50)
According to Eq. (4.3), for W
,
(x, t; , ) we have
J(x, t, ; , )
t
=
_
En
n

i,j=1
a
ij
(, )

2
W
,
(x, t; , )
x
j
x
i
(, ; , ) d (4.51)
488
if < < t. Therefore the integral
t
_

J(x, t, ; , )
t
d (4.52)
is similar to the integral (4.48) we have studied before, and as before, we can see that the integral (4.52) converges
absolutely and uniformly in the domain t > 0, x K, E
n
.
Let us consider the dierence
V (x, t + t; , ) V (x, t; , )
t
(x, t; , )
t
_

J(x, t, ; , )
t
d
=
_

_
J(x, t + t, ; , ) J(x, t, ; , )
t

J(x, t, ; , )
t
_
d +
1
t
t+t
_
t
J(x, t + t, ; , ) d (x, t; , )
= J(x, t + t, t

; , ) (x, t; , )
_
+1
_

+
t
_
t2
_
J
t
(x, t, ; , ) d
+
_
+1
_

+
t
_
t2
_
J
t
(x, t

, ; , ) d +
t2
_
+1
_
J
t
(x, t

, ; , )
J
t
(x, t, ; , )
_
d, (4.53)
where t

and t

are some values between t and t + t. To be denite we shall suppose that t > 0. According to
relation (4.6) we get
lim
t0
J(x, t + t, t

; , ) (x, t; , ) = 0.
Since the integral (4.52) converges uniformly with respect to t, for suciently small
1
> 0 and
2
> 0 the rst
three integrals in the right-hand side of (4.53) can be made arbitrarily small in absolute value independently of t.
The fourth integral can be estimated in the following way:

t
_
t2
J
t
(x, t

, ; , ) d

_
t2

J
t
(x, t

, ; , )

d.
It follows that this integral can be made arbitrarily small for suciently small
2
> 0 and t. Finally for xed

1
and
2
the last integral in the right-hand side of (4.53) tends to zero as t 0, because the function under
the integral sign tends to zero and is continuous in and t

in the closed domain +


1
t
2
, t t

T.
Thus, the left-hand side of the equality (4.53) tends to zero as t +0. In a similar way we can consider
the case where t 0.
The equality (4.50) is proved.
Now we pass to the immediate construction of the fundamental solution Z(x, t; , ) in the form (4.18).
Suppose that the function (x, t; , ) satises the conditions (4.24) and (4.25). Using formulas (4.18), (4.31),
(4.48), and (4.50), we get
L
x,t
(Z) = L
x,t
(W
,
(x, t; , )) +
t
_

d
_
En
L
x,t
(W
,
(x, t; , ))(, ; , ) d (x, t; , ).
The requirement that L
x,t
(Z) be zero leads to the following equation for :
(x, t; , ) = L
x,t
(W
,
(x, t; , )) +
t
_

d
_
En
L
x,t
(W
,
(x, t; , ))(, ; , ) d. (4.54)
489
We seek a solution of the integral equation (4.54) as a series
(x, t; , ) =

m=1

m
(x, t; , ), (4.55)
where

1
(x, t; , ) = L
x,t
(W
,
(x, t; , )),

m+1
(x, t; , ) =
t
_

d
_
En
L
x,t
(W
,
(x, t; , ))
m
(, ; , ) d (m = 1, 2, . . .). (4.56)
Let us show that the series (4.55) converges uniformly for t > . Taking into account (4.3), we have for
W
,
(x, t; , )

1
(x, t; , ) =
n

i,j=1
[a
ij
(x, t) a
ij
(, )]

2
W
,
(x, t; , )
x
i
x
j
+
n

i=1
b
i
(x, t)
W
,
(x, t; , )
x
i
+c(x, t)W
,
(x, t; , ).
By inequalities (4.14), (4.15), and (4.20)(4.22) we get the estimate
[
1
(x, t; , )[ = [L
x,t
(W
,
(x, t; , ))[ < M
31
(t )

n
2
1+
e

5
|x|
2
t
, (4.57)
where
5
> 0. Then we have
[
2
(x, t; , )[ < M
2
31
t
_

d
_
En
(t )

n
2
1+
e

5
|x|
2
t
( )

n
2
1+
e

5
||
2

d
= M
2
31
t
_

(t )
1
( )
1
d
n

i=1
+
_

[(t )( )]

1
2
e
5
_
(x
i

i
)
2
t
+
(
i

i
)
2

_
d
i
= M
2
31

n
2
[
5
(t )]

n
2
e

5
|x|
2
t
t
_

(t )
1
( )
1
d
= M
2
31

n
2
[
5
(t )]

n
2
e

5
|x|
2
t
(t )
21
1
_
0
s
1
(1 s)
1
ds =

2
()
(2)
M
2
31
_

5
_n
2
(t )

n
2
1+2
e

5
|x|
2
t
,
where (h) is the gamma-function.
It is easy to show by induction with respect to m that
[
m
(x, t; , )[ <

m
()
(m)
M
m
31
_

5
_
(m1)n
2
(t )
m
n
2
1
e

5
|x|
2
t
(m = 1, 2, . . .). (4.58)
Since we have (m) [m 1]! for m > 2, it follows from (4.58) that the series (4.55) converges absolutely and
uniformly for t > and the inequality (4.24) holds for the function (x, t; , ). From the uniform convergence
of the series (4.55) it follows that the function (x, t; , ) is continuous jointly in all its arguments in the domain
0 < t T, x E
n
, E
n
.
Let us show that the function (x, t; , ) we have constructed satises the inequality (4.25) for [x

x[
2
<
a(t ). It is sucient to prove that both terms of the right-hand side of Eq. (4.54) satisfy the inequality (4.25).
Consider rst the dierence L(W
,
(x

, t; , )) L(W
,
(x, t; , )). We shall concentrate on nding bounds
only for the terms most dicult to estimate, that is, terms containing second derivatives of the function W
,
.
Using the inequalities (4.14), (4.15), and (4.22), we obtain
490

[a
ij
(x

, t) a
ij
(, )]

2
W
,
(x

, t; , )
x
i
x
j
[a
ij
(x, t) a
ij
(, )]

2
W
,
(x, t; , )
x
i
x
j

[a
ij
(x

, t) a
ij
(x, t)]

2
W
,
(x

, t; , )
x
i
x
j

[a
ij
(x, t) a
ij
(, )]
_

2
W
,
(x

, t; , )
x
i
x
j


2
W
,
(x, t; , )
x
i
x
j
_

< M
32
[x

x[

(t )

n
2
1
e

2
|x

|
2
t
+M
33
[x

x[[[x [

+ (t )

]([x [ +[x

[)(t )

n
2
2
_
e

2
|x|
2
t
+
_
1 +
[x [
2
t
_
e

4
| x|
2
t
_
,
(4.59)
where x belongs to the interval joining x and x

. We consider rst such that satisfy the condition [x[ 2[xx

[,
and then the remaining values of . Using the inequality (4.9) we can see that for [x

x[
2
< 2a(t) the right-hand
side of (4.59) is not greater than
M
34
[x

x[
2
(t )

n
2
1+3
e

5
|x|
2
t
,
where
2
> 0 and
3
> 0. Similarly we can estimate the remaining terms of the dierence L(W
,
(x

, t; , ))
L(W
,
(x, t; , )). Thus we have
[L(W
,
(x

, t; , )) L(W
,
(x, t; , ))[
< M
35
[x

x[
2
(t )

n
2
1+3
e

5
|x|
2
t
for [x

x[
2
< 2a(t ). (4.60)
Now we pass to the estimate of the term (x, t; , ) of the right-hand side of (4.54). We have
(x

, t; , ) (x, t; , )
=
t
_
t
1
2a
|x

x|
2
d
_
En
L(W
,
(x

, t; , ))(, ; , ) d
t
_
t
1
2a
|x

x|
2
d
_
En
L(W
,
(x, t; , ))(, ; , ) d
+
t
1
2a
|x

x|
2
_

d
_
En
L(W
,
(x

, t; , )) L(W
,
(x, t; , ))(, ; , ) d = J
1
+J
2
+J
3
. (4.61)
By (4.24) and (4.57), for [x

x[
2
< a(t ) we get
[J
2
[ < M
36
t
_
t
1
a
|x

x|
2
d
_
En
(t )

n
2
1+
( )

n
2
1+
e
5
_
|x|
2
t
+
||
2

_
d
= M
37
(t )

n
2
e

5
|x|
2
t
t
_
t
1
a
|x

x|
2
(t )
1
( )
1
d
< M
37
_
t
1
2a
[x

x[
2
_
1
(t )

n
2
e

5
|x|
2
t
t
_
t
1
a
|x

x|
2
(t )
1
d
< M
38
[x

x[
2
(t )

n
2
1+
e

5
|x|
2
t
. (4.62)
In a similar way we can estimate [J
1
[. Then, taking into account (4.60) and (4.24), we nd
491
[J
3
[ < M
39
[x

x[
2
t
1
2a
|x

x|
2
_

d
_
En
(t )

n
2
1+3
( )

n
2
1+
e
5
_
|x|
2
t
+
||
2

_
d
= M
40
[x

x[
2
(t )

n
2
e

5
|x|
2
t
t
1
2a
|x

x|
2
_

(t )
31
( )
1
d
< M
40
[x

x[
2
(t )

n
2
1++3
e

5
|x|
2
t
1
|x

x|
2
2a(t)
_
0
s
1
(1 s)
31
ds
< M
41
[x

x[
2
(t )

n
2
1++3
e

5
|x|
2
t
. (4.63)
Comparing the inequalities (4.60)(4.63), we obtain the required estimate (4.25) for the right-hand side of
Eq. (4.54) and, therefore, for the function (x, t; , ). Thus we have proved that the function Z(x, t; , ), given
by formula (4.18), for t > is continuous jointly in the variables x, t, , and , together with its derivatives
Z
xi
,

2
Z
xi xj
(i, j = 1, . . . , n), and
Z
t
. The function Z(x, t; , ) satises Eq. (4.12) with respect to the variables x and t.
It follows from estimate (4.20) for the function W
,
(x, t; , ) and estimate (4.29) for the function (4.19)
that Z(x, t; , ) is bounded for t +[x [ > 0.
Now let us show that for Z(x, t; , ) the relation (4.13) holds.
Let () be a continuous and bounded function in E
n
. Suppose that x is an arbitrary point of the space E
n
and 0 < T. Then we have
_
En
Z(x, t; , )() d =
_
En
W
,
(x, t; , )() d +
_
En
V (x, t; , )() d,
where the function V (x, t; , ) is given by (4.19). Taking into account (4.29), we can see that

_
En
V (x, t; , )() d

< M
42
sup[[(t )
1
;
therefore,
lim
t+0
_
En
V (x, t; , )() d = 0 (4.64)
uniformly with respect to x E
n
. From (4.6) and (4.64) the relation (4.13) follows.
Thus we have constructed the function Z(x, t; , ) that is a fundamental solution of Eq. (4.12) in the layer H.
Let us prove the uniqueness of the fundamental solution of Eq. (4.12).
From the properties of the fundamental solution it follows that the function
u(x, t) =
_
En
Z(x, t; , )() d (4.65)
is a bounded solution of the Cauchy problem for Eq. (4.12) in the layer H

0 < t T, x E
n
with the initial
condition
u(x, ) = (x), (4.66)
where (x) is an arbitrary continuous function vanishing outside of some bounded domain. In fact, let (x) = 0
for [x[ R; then
u(x, t) =
_
||R
Z(x, t; , )() d.
492
For t > , we have
L(u) =
_
||R
L
x,t
(Z(x, t; , ))() d = 0.
According to (4.13) the function (4.65) for t = satises the initial condition (4.66). From (4.13) we can also
deduce that the function (4.65) is bounded in the domain < t T, [x[ 2R. Finally in the domain
< t T, [x[ 2R this function also is bounded because in this domain we have
[u(x, t)[ sup
|x|R
[Z(x, t; , )[
_
||R
[()[ d M
R
;
this follows from the boundedness of Z(x, t; , ) for [x [ R.
According to Theorem 10 of Sect. 1, a bounded solution of the Cauchy problem (4.12), (4.66) is unique in
the layer H

.
Suppose now that besides the fundamental solution Z(x, t; , ) of Eq. (4.12) that we have constructed
above, there exists another fundamental solution Z
1
(x, t; , ) of the same equation. Then, taking into account
the uniqueness of the solution for the Cauchy problem (4.12), (4.66), we obtain
_
En
[Z(x, t; , ) Z
1
(x, t; , )]() d = 0
for any x E
n
and 0 < t T and for any continuous function () vanishing outside some bounded domain.
Since Z and Z
1
are continuous functions for t > , it follows that Z
1
(x, t; , ) Z(x, t; , ) for t > , x E
n
, and
E
n
.
Taking into account estimates (4.20)(4.22) for W
,
(x, t; , ), and also the inequalities (4.29), (4.32), and
(4.49) for the function (4.19), we can deduce that for the functions
Z(x, t; , ),
Z(x, t; , )
x
i
,

2
Z(x, t; , )
x
i
x
j
(i, j = 1, . . . , n)
the estimates (4.16) hold. It follows from Eq. (4.12) that the estimate (4.16) holds for
Z(x,t;,)
t
.
Now we shall prove that Z(x, t; , ) > 0 for t > .
Suppose rst that Z(x, t; , ) < 0 for some x
0
E
n
,
0
E
n
, t = t
0
, and , t
0
> . Then, because
Z(x
0
, t
0
; , ) is continuous, we have Z(x
0
, t
0
; , ) < 0 for [
0
[ < , where > 0. Consider Eq. (4.12) with
initial condition (4.66) in the layer H

, where (x) > 0 for [x


0
[ < and (x) 0 for [x
0
[ . According
to Theorem 8 of Sect. 1, the solution u(x, t) of the Cauchy problem (4.12), (4.66) must be nonnegative everywhere
in H

. On the other hand, we nd from (4.65) that


u(x
0
, t
0
) =
_
|
0
|<
Z(x
0
, t
0
; , )() d < 0.
This contradiction shows that Z(x, t; , ) 0 everywhere for t > .
Suppose now that Z(x, t; , ) = 0 for x = x
1
, =
1
, t = t
1
, and =
1
, where t
1
>
1
.
Consider Z(x, t;
1
,
1
) in the cylinder R
1
+ t t
1
, [x x
1
[ N, where 0 < < t
1

1
and
N > 0 is an arbitrary number. Since Z(x, t;
1
,
1
) 0 in R and Z(x
1
, t
1
;
1
,
1
) = 0, we have Z(x, t;
1
,
1
) = 0
everywhere in R. This follows from Theorem 6 of Sect. 1, which in the case of zero maximum or minimum holds
not only for c(x, t) 0 but also for c(x, t) M, where M > 0. If tends to zero and N tends to innity, we get
Z(x, t;
1
,
1
) = 0 everywhere in the layer
1
< t t
1
, x E
n
, and this is in contradiction with the formulas
(4.4), (4.18), and (4.19) and the estimate (4.29). Thus we can see that the function Z(x, t; , ) is strictly positive
in the domain 0 < t T, x E
n
, E
n
.
Now we shall prove that Z(x, t; , ) satises Eq. (4.17) for t > as a function of the variables and , if in H
there exist continuous and bounded derivatives
aij(x,t)
xj
,

2
aij(x,t)
xi xj
, and
bi(x,t)
xi
(i, j = 1, . . . , n), H older continuous
with respect to x.
493
Transforming the left-hand side of Eq. (4.17) and changing to x and to t, we get
L

x,t
(v) =
n

i,j=1
a
ij
(x, t)

2
v
x
i
x
j
+
n

i=1
_
2
n

j=1
a
ij
(x, t)
x
j
b
i
(x, t)
_
v
x
i
+
_
n

i,j=1

2
a
ij
(x, t)
x
i
x
j

i=1
b
i
(x, t)
x
i
+c(x, t)
_
v
v
(t)
= 0. (4.67)
According to the additional suppositions on the smoothness of a
ij
and b
i
, we can see that the coecients of Eq. (4.67)
satisfy conditions (4.14) and (4.15). As we have proved, in the domain 0 t < T, x E
n
, E
n
there
exists a unique fundamental solution Z

(x, t; , ) of Eq. (4.67). We shall show that Z(x, t; , ) = Z

(x, t; , ).
Consider the following easily veried identity:
vL(u) uL

(v) =
n

i=1

x
i
_
n

j=1
a
ij
_
v
u
x
j
u
v
x
j
_
+
_
b
i

j=1
a
ij
x
j
_
uv
_

(uv)
t
. (4.68)
Here we set u(x, t) = Z(x, t; , ) and v(x, t) = Z

(x, t;

), where E
n
,

E
n
, 0 <

T. For x E
n
and < t <

we have
Z

(x, t;

)L(Z(x, t; , )) Z(x, t; , )L

(Z

(x, t;

)) = 0. (4.69)
Let us integrate the identity (4.68) over the cylinder R

< t

t t

<

, [x[ N
1
, where N
1
> 0 is
an arbitrary number, and after that use the GaussOstrogradski formula. Taking into account (4.69), we get
0 =
_
S

i=1
_
n

j=1
a
ij
_
Z

Z
x
j
Z
Z

x
j
_
+
_
b
i

j=1
a
ij
x
j
_
ZZ

_
cos(, x
i
) dS

_
|x|N1
Z(x, t

; , )Z

(x, t

) dx +
_
|x|N1
Z(x, t

; , )Z

(x, t

) dx,
where S

is the side boundary ([x[ = N


1
) of the cylinder R

and is a normal to S

. Let N
1
tend to innity.
According to the estimates (4.16) for Z and similar estimates for Z

, the integral over S

tends to zero for N


1
,
and we arrive at the following equality:
_
En
Z(x, t

; , )Z

(x, t

) dx =
_
En
Z(x, t

; , )Z

(x, t

) dx.
Since the values t

and t

are arbitrary, this means that the integral


_
En
Z(x, t; , )Z

(x, t;

) dx does not depend


on t for < t <

:
_
En
Z(x, t; , )Z

(x, t;

) dx = (, ;

). (4.70)
In the same way as we have proved the relation (4.13) it is easy to prove that
lim
t+0
_
En
Z(x, t; , )(x, t) dx = (, ) (4.71)
for any function (x, t) continuous and bounded in the layer t + , x E
n
( > 0). Let us pass to
the limit in the equality (4.70) as t + 0; according to (4.71) we get
(, ;

) = Z

(, ;

). (4.72)
On the other hand, taking into account the property of the function Z

(x, t;

) similar to (4.71), we can deduce


from (4.70) passing to the limit as t

0 that
(, ;

) = Z(

; , ). (4.73)
From (4.72) and (4.73) it follows that Z(

; , ) = Z

(, ;

). Since

and

were chosen in an arbitrary


way, this means that Z(x, t; , ) = Z

(x, t; , ) everywhere in the domain 0 < t T, x E


n
, E
n
. It
follows that Z(x, t; , ) satises (4.17) everywhere in this domain with respect to the variables and .
Theorem 1 is proved.
494
Remark. We can get from relations (4.4), (4.18), and (4.19) the following estimate from below for Z(x, t; , ):
Z(x, t; , ) M
43
(t )

n
2
e
M44
|x|
2
t
M
45
(t )

n
2
+1
e
4
|x|
2
t
. (4.74)
In particular, for [x [
2
< a(t ), where a > 0, we obtain
Z(x, t; , ) M
46
(t )

n
2
. (4.75)
2. Cauchy Problem. The fundamental solution of Eq. (4.12) constructed above can be applied to the study
of the Cauchy problem for a linear second-order parabolic equation. We shall prove an existence theorem for
the solution of the Cauchy problem under very weak suppositions about the coecients of the equation and its
right-hand side. In Sect. 3 we have proved the existence theorem by another method and under much stronger
suppositions.
Theorem 2. In the layer H there exists a unique bounded solution of the Cauchy problem for Eq. (1.1)
with the boundary condition
u(x, 0) = (x), x E
n
, (4.76)
if the following suppositions are fullled:
(1) The coecients a
ij
, b
i
, and c in the layer

H are bounded, continuous, and satisfy the relations (4.14)
and (4.15), and the inequality (4.5) holds.
(2) The function f(x, t) in the layer

H is bounded, continuous, and H older continuous with respect to x.
(3) The function (x) is continuous and bounded in E
n
.
The solution of problem (1.1), (4.76) is given by the formula
u(x, t) =
t
_
0
d
_
En
Z(x, t; , )f(, ) d +
_
En
Z(x, t; , 0)() d = U
1
(x, t) +U
2
(x, t). (4.77)
Proof. It follows from the properties of the fundamental solution Z(x, t; , ) and from estimates (4.16) that
the function U
2
(x, t) in the layer

H is bounded and for t > 0 satises the homogeneous equation (4.12). According
to (4.13), this function satises also the initial condition (4.76).
The rst of the estimates (4.16) gives
[U
1
(x, t)[ M
47
sup[f[
t
_
0
d
_
En
(t )

n
2
e

1
|x|
2
t
d M
48
t;
hence the function U
1
(x, t) is bounded in

H and for t = 0 it satises the zero initial condition:
U
1
(x, 0) = 0.
It remains to prove that for t > 0 the function U
1
(x, t) satises Eq. (1.1). To do this, we represent U
1
(x, t)
as a sum:
U
1
(x, t) =
t
_
0
d
_
En
W
,
(x, t; , )f(, ) d
t
_
0
d
_
En
V (x, t; , )f(, ) d = U
(1)
1
(x, t) +U
(2)
1
(x, t), (4.78)
where
V (x, t; , ) =
t
_

d
_
En
W
,
(x, t; , )(, ; , ) d.
495
The integral U
(1)
1
(x, t) is similar to the integral (4.19) that we have studied when we constructed the funda-
mental solution. Taking into account condition (2) of this theorem, we deduce (as in the proof of Theorem 1) that
the function U
(1)
1
(x, t) for t > 0 satises the equation
L(U
(1)
1
(x, t)) = f(x, t)
t
_
0
d
_
En
L
x,t
(W
,
(x, t; , ))f(, ) d. (4.79)
In the integral U
(2)
1
(x, t) we can change the order of integration:
U
(2)
1
(x, t) =
t
_
0
d
_
En
W
,
(x, t; , )(, ) d, (4.80)
where
(, ) =

_
0
d
_
En
(, ; , )f(, ) d. (4.81)
We can do this because of estimates (4.20) and (4.24). By arguments similar to those used above we can deduce
that the function (4.80) for t > 0 satises the following equation:
L(U
(2)
1
(x, t)) = (x, t)
t
_
0
d
_
En
L
x,t
(W
,
(x, t; , ))(, ) d. (4.82)
Because of relations (4.14), (4.15), and (4.20)(4.32) we have
[L
x,t
(W
,
(x, t; , ))[ < M
49
(t )

n
2
1+

2
e

4
|x|
2
t
.
This inequality allows the change of integration order in the right-hand side of (4.82); we obtain
L(U
(2)
1
(x, t)) =
t
_
0
d
_
En
_
(x, t; , )
t
_

d
_
En
L
x,t
(W
,
(x, t; , ))(, ; , ) d
_
f(, ) d. (4.83)
Comparing (4.79) and (4.83), we get
L(U
1
(x, t)) = f(x, t) +
t
_
0
d
_
En
f(, )
_
(x, t; , ) L
x,t
(W
,
(x, t; , ))

t
_

d
_
En
L
x,t
(W
,
(x, t; , ))(, ; , ) d
_
d = f(x, t),
because the function (x, t; , ) for 0 < t T satises Eq. (4.54) and we deduce that the function U
1
(x, t) for
t > 0 satises Eq. (1.1).
The uniqueness of the solution of problem (1.1), (4.76) is proved in Sect. 1.
Remark 1. With the help of inequalities (4.16) it is easy to show that the problem (1.1), (4.76) has a solution
also in case where the function (x) is continuous for all x and satises the inequality
[(x)[ < M
50
e
M51|x|
2
, > 0. (4.84)
496
In this case, the solution is again given by formula (4.77). It follows from this formula that
sup
0tT
[u(x, t)[ < M
52
e
M53|x|
2
+ sup

H
[f(x, t)[ t. (4.85)
The uniqueness of the solutions of Cauchy problem in the class of functions satisfying (4.85) is proved in Sect. 1.
It is interesting to note that one cannot prove the existence of the solution for the Cauchy problem for all
t > 0 if one allows = 0 in the estimate (4.84). In fact, the function u(x, t) = (1 4At)

1
2
e
Ax
2
e
14At
is a solution of
the equation

2
u
x
2

u
t
= 0 with the initial condition u(x, 0) = e
Ax
2
(A > 0). This function is not dened for t
1
4A
,
and lim
t
1
4A
0
u(x, t) = . According to the uniqueness theorem, for t <
1
4A
there is no other solution of the Cauchy
problem growing not faster than e
Mx
2
.
Remark 2. If in Theorem 2 we suppose additionally that (x) C
2+
, then, using the methods of [43] we
can prove that the solution of problem (1.1), (4.76) belongs to the class C
2,1
in the closed domain

H.
3. First Boundary-Value Problem. Greens Function. Consider now for Eq. (1.1) in the cylinder
Q = (0, T) the rst boundary-value problem with conditions
u(x, 0) = (x) for x , u[
S
= 0. (4.86)
It is well known that the rst boundary-value problem for the heat equation can be solved with the help of the so-
called double-layer heat potentials, similar to the double-layer potentials for the Laplace equation [78]. For Eq. (1.1)
with coecients satisfying the suppositions (4.14) and (4.15) it is not possible in general to construct a double-layer
potential, because the coecients a
ij
(, ) and, consequently, also the function Z(x, t; , ) are not assumed to be
dierentiable in
1
, . . . ,
n
. Therefore we shall use another method to construct the solution of problem (1.1), (4.86);
it is due to W. Pogorzelski [79]. Note that we do not suppose the continuity of the solution of problem (1.1), (4.86)
for t = 0 on the boundary of the domain ; therefore (x) on may be dierent from zero or even not dened.
Theorem 3. There exists in the cylinder Q a unique solution of the problem (1.1), (4.86) if the following
suppositions are fullled:
(1) The coecients a
ij
, b
i
, and c in

Q satisfy the inequalities (4.5), (4.14), and (4.15).
(2) The function f(x, t) is continuous in

Q jointly in the variables x and t and is H older continuous with
respect to x.
(3) The domain is bounded and belongs to the class A
1+0
.
(4) The function (x) is continuous and bounded in .
Proof. A function G(x, t; , ) will be called a Greens function of the boundary-value problem (1.1), (4.86)
if it is dened and continuous jointly in the variables for x

, , 0 < t T and has the form
G(x, t; , ) = Z(x, t; , ) (x, t; , ),
where (x, t; , ) has the following properties
(1) L
x,t
((x, t; , )) = 0 for x , , 0 < t T; (4.87)
(2) (x, t; , ) = Z(x, t; , ) for , (x, t) S, 0 < t T; (4.88)
(3) lim
t+0
(x, t; , ) = 0 for x , . (4.89)
To construct a Greens function it is sucient to nd the function (x, t; , ), satisfying the relations (4.87)(4.89).
We shall construct this function for and any x E
n
.
Let us extend the coecients a
ij
, b
i
, c, and f of Eq. (1.1) to all of the layer

H0 t T, x E
n
so that
they remain bounded and satisfy the conditions of this theorem. For x E
n
we set
(x, t; , ) = Z(x, t; , ),
497
where Z(x, t; , ) is a fundamental solution of Eq. (4.12) in the layer H. For a xed , the function (x, t; , )
so dened is bounded in the domain R

0 < t T, x E
n
because of the estimate (4.16). Inside R

this
function satises Eq. (4.12) and the conditions
(x
1
, t
1
; , )

x
1
=
Z(x
1
, t; , )

x
1
on S,
lim
|x|
(x, t; , ) = 0, lim
t
(x, t; , ) = 0.
_

_
(4.90)
Here
F(x
1
, t)

x
1
lim
xx
1

xEn\

i,j=1
a
ij
(x, t) cos(
x
1 , x
i
)
F(x, t)
x
j
,
and
x
1 is the direction of the inside normal (inside with respect to the domain E
n
) to at the point x
1
.
According to the remark at the end of subsection 5 of Sect. 1, conditions (4.90) dene a unique bounded
solution of Eq. (4.12) in the domain R

.
On the other hand, the solution of problem (4.12), (4.90) in the domain R

can be found in the form of


an integral similar to the simple layer potential for the Laplace equation, that is, in the form

(x, t; , ) =
t
_

d
_

Z(x, t; , )(, ; , ) d

, (4.91)
where the function (x, t; , ) is to be dened. One can show (see [80]) that for integrals having the form (4.91)
with a continuous function (x, t; , ) the formula

(x, t; , )

x
=
1
2
(x, t; , ) +
t
_

d
_

Z(x, t; , )

x
(, ; , ) d

holds for (x, t) S and t > , similar to the jump formula for the normal derivative for ordinary potentials. By
the rst of conditions (4.90) we obtain for (x, t; , ) the integral equation
(x, t; , ) = 2
Z(x, t; , )

t
_

d
_

2
Z(x, t; , )

x
(, ; , ) d

, x , < t T. (4.92)
The solution (x, t; , ) of (4.92) can be found by successive approximations, as was done above for the so-
lution of Eq. (4.54). The solution (x, t; , ) is continuous for < t T and satises the inequality
[(x, t; , )[ < M
54
(t )

n+1
0
2
e

4
|x|
2
t
; (4.93)
here
0
> 0 is the same number as in condition (3) of the theorem. Setting (x, ; , ) = 0, we obtain a function
that is continuous everywhere for x and t T (recall that is a xed point).
From (4.16) and (4.93) we can deduce that the function (4.91) is continuous and bounded in the closed
domain

R

. Since the solution of the second boundary-value problem (4.12), (4.90) is unique, it follows that

(x, t; , ) = Z(x, t; , )
for x E
n
, 0 < t T.
The function (4.91) is dened and continuous not only outside Q but in all of the layer 0 < t T,
x E
n
. In particular, for x
1
and < t we have
lim
xx
1
x

(x, t; , ) = lim
xx
1
xEn\

(x, t; , ) = Z(x
1
, t; , );
498
therefore, the function

(x, t; , ) considered for x , satises the relation (4.88). The relations (4.87) and (4.89)
obviously hold.
Setting G(x, t; , ) = Z(x, t; , )

(x, t; , ), we obtain the required Greens function.
Taking into account inequalities (4.16) and (4.93), it is easy to show that for

(x, t; , ) the following
estimate holds:
[

(x, t; , )[ < M
55
(t )

n
0
2
e

4
|x|
2
t
. (4.94)
Therefore, for the Greens function G(x, t; , ) we have the same estimate as for Z(x, t; , ):
[G(x, t; , )[ < M
56
(t )

n
2
e

4
|x|
2
t
. (4.95)
Now let us prove that the solution of the rst boundary-value problem (1.1), (4.86) in the cylinder
Q = (0, T) is given by the formula
u(x, t) =
t
_
0
d
_

G(x, t; , )f(, ) d +
_

G(x, t; , 0)() d = u
1
(x, t) +u
2
(x, t). (4.96)
The boundedness of the function (4.96) follows from the estimate (4.95) for G(x, t; , ). According to the same
estimate, [u
1
(x, t)[ M
57
t; therefore u
1
(x, 0) = 0. The function u
2
(x, t) can be represented as a sum:
u
2
(x, t) =
_

Z(x, t; , 0)() d +
_

(x, t; , 0)() d = u
(1)
2
(x, t) +u
(2)
2
(x, t).
Similarly to (4.13) it is easy to show that if the function (x) is continuous and bounded in we have
lim
t+0
_

Z(x, t; , 0)() d = (x) for x ; (4.97)


here the convergence is uniform in any closed domain lying strictly inside . It follows that u
(1)
2
(x, 0) = (x) for
x . As a consequence of (4.94) we have [u
(2)
2
(x, t)[ M
58
t
0/2
, hence u
(2)
2
(x, 0) = 0. Thus, the function (4.96)
satises the rst of the conditions (4.86).
Let us show that the function (4.96) satises Eq. (1.1) in

Q . The properties of G(x, t; , ) allow us to
assert that the function u
2
(x, t) in

Q is a solution of the homogeneous equation (4.12). The function u
1
(x, t)
can be represented as follows:
u
1
(x, t) =
t
_
0
d
_

Z(x, t; , )f(, ) d +
t
_
0
d
_

Z(x, t; , )h(, ) d

= u
(1)
1
(x, t) +u
(2)
1
(x, t),
where
h(, ) =

_
0
d
_

(, ; , )f(, ) d;
the change of integration order eectuated in the second term is legitimate because of the estimate (4.93).
It is obvious that u
(2)
1
(x, t) in

Q satises Eq. (4.12). In the same way as in Theorem 2, we can show that
u
(1)
1
(x, t) in

Q is a solution of Eq. (1.1). Therefore, the function (4.96) in

Q satises Eq. (1.1).
It remains to verify that the function (4.96) satises the second of conditions (4.86).
Let x
1
. Denote by K

the intersection of the domain and the n-dimensional ball of radius with center
at the point x
1
. Consider the function u
2
(x, t) for x K

and 0 < t T, where > 0 is an arbitrarily small


number. The integral over can be represented as a sum of the integral over K
2
and the integral over K
2
.
In accordance with such a representation we get
u
2
(x, t) = u

2
(x, t) + u

2
(x, t).
499
Using the estimate (4.95) for G(x, t; , ), we nd
[u

2
(x, t)[ < M
59
sup[[
_
|x|3
[G(x, t; , 0)[ d < M
60
t

1
2

_
0
_
r

t
_
n1
e

4
r
2
t
dr < M
61

1
2
<

2
(4.98)
if 0 < <
0
(, ). Let us x such a and then choose
1
(, ) > 0 so small as to have the inequality
[u

2
[ <

2
(4.99)
for [x x
1
[ <
1
and 0 < t T. This is possible because of the following property of the function G(x, t; , ):
lim
xx
1
G(x, t; , ) = 0
uniformly with respect to K
2
and 0 < t T. From (4.98) and (4.99) we get
[u
2
(x, t)[ < for 0 < t T and [x x
1
[ <
1
(, ),
and it follows that u
2
(x, t)[
S
= 0 for 0 < t T.
Similarly one can prove that u
1
(x, t)[
S
= 0 for 0 < t T.
Thus, the function (4.96) is the required solution of the problem (1.1), (4.86) in the cylinder Q. The unique-
ness of the solution follows from Theorem 11 of Sect. 1.
Remark. The solution of the rst boundary-value problem (1.1), (4.86) constructed in Theorem 3 turns out
to be continuous everywhere in

Q (including the points of the set x , t = 0) if in addition to the suppositions
of Theorem 3 we suppose that the consistency condition
(x) = 0 for x
holds. In fact, let us set (x) = 0 for x lying outside of . Then the function (x) becomes continuous and bounded
in all of the space E
n
. According to (4.13) the relation (4.97) holds for any x E
n
, including x , and the tending
to the limit is uniform in any closed domain of the space E
n
, in particular, in

. Therefore
lim
t+0
xx
1

Z(x, t; , 0)() d = 0.
As was proved in Theorem 3, we have [u
1
(x, t)[ M
57
t and [u
(2)
2
(x, t)[ M
58
t
0/2
for t > 0, and it follows that
lim
t+0
xx
1

u(x, t) = 0.
4. For a second-order parabolic equation (1.1) it is possible to construct a potential theory similar to
the classic potential theory for second-order elliptic equations [2]. However, the suppositions on the smoothness of
the coecients of Eq. (1.1) have to be more strict than they were in previous sections. Here we give a sketch of
another possible approach to solving the rst boundary-value problem by means of potential theory (see [81]).
Suppose that the coecients a
ij
(x, t) in

Q have continuous second derivatives with respect to x
1
, . . . , x
n
,
the derivatives of b
i
(x, t) in x
1
, . . . , x
n
and the function c(x, t) are also continuous in

Q, and the boundary of
the domain belongs to the class A
1+
. In this case, one can construct a double-layer potential with the help of
the fundamental solution Z(x, t; , ):
v(x, t) =
t
_
0
d
_

P(Z(x, t; , ))(, ) d

,
where (, ) is a function continuous on

S that is called the density of the potential,
P(Z(x, t; , ))
n

j=1
_
n

i=1

i
[a
ij
(, )Z(x, t; , )] b
j
(, )Z(x, t; , )
_
cos(,
j
),
500
and is the direction of the inner normal to . We shall seek the solution of the rst boundary-value problem for
Eq. (4.12) with the conditions
u(x, 0) = 0, u[
S
= (x, t) (4.100)
as a potential with an unknown density (, ). There is a theorem about a jump of the double-layer potential
on S, that is, the relation
lim
xx
1

v(x, t) =
1
2
(x
1
, t) +v(x
1
, t), (4.101)
where
v(x
1
, t) =
t
_
0
d
_

P(Z(x
1
, t; , ))(, ) d

.
Taking into account conditions (4.100) and formula (4.101), we obtain an integral equation for the determination
of (, ):
1
2
(x, t) +
t
_
0
d
_

P(Z(x, t; , ))(, ) d

= (x, t). (4.102)


According to the suppositions on the smoothness of , we can see that the kernel of integral Eq. (4.102) has a weak
singularity
[P(Z(x, t; , ))[ < M
0
(t )

n+1
2
+0
e

0
|x|
2
t
,
where M
0
,
0
, and
0
are positive constants. Equation (4.102) can be solved by means of successive approximations.
5 GENERALIZED SOLUTIONS OF BOUNDARY-VALUE PROBLEMS.
UNIQUENESS THEOREMS. SOME AUXILIARY PROPOSITIONS
In modern investigations on partial dierential equations, the concept of generalized solution plays an im-
portant role. The most frequent way to introduce a generalized solution is to dene it either as a function satisfying
an integral identity that is in some sense equivalent to the dierential equation or as a limit of a sequence of smooth
solutions of the equation (they are also called classical). Depending on the class of functions the generalized solution
belongs to and on the integral identity that is taken to dene it, we get dierent ways of posing problems. To show
that a certain way to dene a generalized solution is reasonable it is necessary to prove the existence and uniqueness
of this solution under some suppositions on the data. The question of the smoothness of this generalized solution
also arises. That is, we have to investigate whether the generalized solution is suciently smooth if the coecients
of the equation, the boundary data, and the boundary itself have a certain degree of smoothness and whether it
will satisfy the equation and the boundary conditions in the ordinary sense, if it is smooth.
In this section, we shall consider as an example two dierent denitions of the generalized solution of the rst
boundary problem for a linear second-order parabolic equation and we shall prove the uniqueness of these solutions.
In Sects. 69 this problem will serve as a model to demonstrate various methods to prove the existence of generalized
solutions. In Sect. 11 we investigate the question of smoothness for generalized solutions of Eq. (1.1)
In what follows, we shall need some functional spaces and some auxiliary propositions.
1. Some Auxiliary Propositions and Notations. As before we shall denote by Q the cylinder (0, T),
where is a domain in the space E
n
. We shall suppose that A
1
.
Denote by

the points of that are situated at a distance greater than > 0 from the boundary of
the domain , and Q

will denote the cylinder

(0, T). The class of functions that are square integrable in Q


will be denoted by L
2
(Q). The scalar product of elements v
1
and v
2
from L
2
(Q) will be dened by the formula
(v
1
, v
2
) =
__
Q
v
1
v
2
dxdt.
The class of innitely dierentiable (k times continuously dierentiable) functions in

Q, vanishing in Q Q

for some > 0 will be denoted by


0
C

(Q) (respectively
0
C
k
(Q)). In a similar way, we dene classes
0
C

() and
0
C
k
().
501
A function w(x, t) integrable in Q is called a generalized derivative of the integrable in x
k
function v(x, t)
with respect to x
k
if for any (x, t)
0
C

(Q) the equality


__
Q
_
v

x
k
+w
_
dxdt = 0 holds. In this case we shall write
w =
v
x
k
. In a similar way, we can dene the generalized derivative with respect to t and higher-order generalized
derivatives (see [82, pp. 39, 40]). Note the following property: if a sequence of functions v
m
(x, t) converges weakly
as m to v(x, t) in the space L
2
(Q) and the norms of
vm
x
k
are bounded in L
2
(Q) uniformly with respect to m,
then v(x, t) has a generalized derivative
v
x
k
L
2
(Q) and
vm
x
k
converge weakly to
v
x
k
([82, pp. 42, 43]).
Denote by W
1
() the Hilbert space of functions (x) such that (x) L
2
() and

xi
L
2
() (i, . . . , n),
with the scalar product
(
1
,
2
)
1
=
_

1
(x)
2
(x) dx +
_

i=1

1
x
i

2
x
i
dx.
By W
1,1
(Q) we shall denote the Hilbert space of functions v(x, t) such that v(x, t) L
2
(Q),
v(x,t)
xi
L
2
(Q)
(i = 1, . . . , n), and
v(x,t)
t
L
2
(Q), with the scalar product
(v
1
, v
2
)
1,1
=
__
Q
v
1
(x, t)v
2
(x, t) dxdt +
__
Q
_
n

i=1
v
1
x
i
v
2
x
i
+
v
1
t
v
2
t
_
dxdt.
It is well known [82, p. 76] that the spaces W
1
() and W
1,1
(Q) are complete. The norms in spaces L
2
(Q), W
1
(),
and W
1,1
(Q) will be denoted by |v|
0
, |v|
1
, and |v|
1,1
respectively.
From S. L. Sobolevs embedding theorems it follows that a function from W
1,1
(Q) can be changed on a set
of measure zero so that it is square integrable on an intersection of the cylinder

Q and any n-dimensional plane
or n-dimensional surface of class A
1
. In particular, such a function is square integrable on the surface S and on
the intersection of the cylinder Q and any plane t = const.
What is more, since the embedding operator is completely continuous, the values of the function
v(x, t) W
1,1
(Q) on the n-dimensional planes or surfaces close to each other also are close in the sense of mean
values [82, pp. 94, 96]. In particular, if v(x, t) W
1,1
(Q) and v(x, 0) = (x), then
_

[v(x, t) (x)]
2
dx 0 for
t 0.
By
0
W
1
() denote the subset of functions (x) W
1
() vanishing on the surface . By
0
W
1,1
(Q) denote
the subspace of functions v(x, t) W
1,1
(Q) vanishing on the surface S.
An important auxiliary tool that will be used below is the averaging of functions. Let the function
v(x, t) L
2
(Q). We shall extend this function to the whole space by dening it as zero outside the cylinder Q.
The function
v
h
(x, t) =
+
_

_
En
v(, )
h
(x , t ) d d
will be called an average function of radius h, where
h
(x, t) is called the averaging kernel and has the following
properties:
h
is innitely dierentiable;
h
= 0 for [x[
2
+(t)
2
h
2
(h > 0);
h
> 0 for [x[
2
+(t)
2
< h
2
;
+
_

_
En

h
(x , t ) d d = 1.
It is easy to show [82, pp. 19, 20, 41] that v
h
(x, t) v(x, t) as h 0 in the norm of L
2
(Q), and for
v(x, t) W
1,1
(Q) the equalities
_
v
xi
_
h
=
v
h
xi
(i = 1, . . . , n) and
_
v
t
_
h
=
v
h
t
hold for suciently small h in any
closed domain lying in the interior of the cylinder Q.
Similarly it is possible to dene the averaging operator in the domain .
Lemma 1. The functions of class
0
C

() are everywhere dense in


0
W
1
(), and the functions of class
0
C

(Q)
are everywhere dense in
0
W
1,1
(Q).
In [82, p. 102] it is proved that for any function (x)
0
W
1
() there exists a sequence
m
(x)
0
C

() that
converges weakly in W
1
() to (x). The functions
m
(x) are obtained by multiplying (x) by smooth compactly
502
supported in functions
m
(x), equal to unity in
1/m
, and then averaging of this product (x)
m
(x). Carrying out
with some more precision the proof of a lemma from [82, p. 102], it is possible to show that
m
(x) converges strongly
to (x) in the norm of W
1
(). Besides that, the rst assertion of Lemma 1 can be obtained as a corollary to a well-
known theorem of functional analysis (see [83, p. 203]). According to this theorem there exist linear combinations

N
=
N

m=1
C
N
m

m
(x) of the functions
m
(x), weakly converging to (x), such that |
N
(x)(x)|
1
0 for N .
Obviously
N
(x) also belongs to
0
C

() (N = 1, 2, . . .).
The second assertion of the lemma can be proved in a similar way.
For the functions (x)
0
W
1
() the following inequality holds:
_

2
(x) dx M
_

i=1
_

x
i
_
dx, (5.1)
and for the functions v(x, t)
0
W
1,1
(Q) we have the inequality
__
Q
v
2
(x, t) dxdt M
__
Q
n

i=1
_
v
x
i
_
2
dxdt, (5.2)
where the constant M depends only on the size of the domain .
The inequalities (5.1) and (5.2) were established in Sect. 2 for smooth functions and in particular for the func-
tions belonging to classes
0
C

() and
0
C

(Q). Since the set


0
C

() is dense in
0
W
1
(), the inequality (5.1) holds
also for (x)
0
W
1
(). The inequality (5.2) for v(x, t)
0
W
1,1
(Q) can be proved in the same way.
Note the following important proposition.
Theorem 1. From any sequence of functions v
m
(x, t) that is uniformly bounded in the norm W
1,1
(Q) it is
possible to select a subsequence converging in the norm L
2
(Q) to a function v(x, t), belonging to the space W
1,1
(Q).
If v
m
[
S
= 0 and v
m
(x, 0) converge to the function (x) in the norm L
2
(), then the limit function v(x, t) belongs
to
0
W
1,1
(Q) and is equal to (x) for t = 0.
The proof of this theorem can be found in [82, pp. 8394].
Lemma 2. In the domain with boundary belonging to class A
2
, there exists an orthonormal in L
2
()
system of functions X
k
(x) having the following properties:
(1) X
k
(x) are twice continuously dierentiable in

and vanish on the boundary of the domain .
(2) For any function (x)
0
W
1
() and any number > 0 there exists a linear combination
N
(x) =
N

k=1
c
k
X
k
(x) of functions X
k
(x) such that
|
N
|
1
< . (5.3)
(3) For any function v(x, t)
0
C

(Q), vanishing in a neighborhood of the boundary Q, and any > 0 there


exists a linear combination v
N
(x, t) =
N

k=1
c
k
(t)X
k
(x) such that
__
Q
_
(v v
N
)
2
+
n

i=1
_
(v v
N
)
x
i
_
2
_
dxdt < .
Proof. It is sucient to prove the inequality (5.3) only for (x)
0
C

(), because the set of these functions


is dense in
0
W
1
(). Since A
2
, there exists a function (x) C
2
(

), vanishing on and positive in the domain .


Since
(x)
(x)

0
C
2
(), we have
_
(x)
(x)

0
C

() and
_
_

_
h
_
_
1
<

2
for a suciently small h > 0.
503
Without loss of generality, one can assume that lies inside the cube 0 x
i
(i = 1, . . . , n). Set

X
k
(x) = sink
1
x
1
sin k
2
x
2
. . . sin k
n
x
n
, where k
j
(j = 1, . . . , n) are some integers and k = (k
1
, k
2
, . . . , k
n
).
It is well known that a function from
0
C

() can be expanded in a Fourier series with respect to



X
k
(x) and that
this series converges uniformly in

together with its derivatives. Therefore, there exists a linear combination

1kj N
c
k

X
k
(x) of functions

X
k
(x) such that
_
_
_
_
_

_
h

1kjN
c
k

X
k
(x)
_
_
_
_
0
+
n

i=1
_
_
_
_
_

_
h
x
i

1kj N
c
k


X
k
(x)
x
i
_
_
_
_
_
0
<

2
.
Therefore,
_
_
_
_
(x)
(x)

1kjN
c
k

X
k
(x)
_
_
_
_
0
< (5.4)
and
_
_
_
_

x
i
_
(x)
(x)
_

1kj N
c
k


X
k
(x)
x
i
_
_
_
_
0
< (i = 1, . . . , n). (5.5)
Set X

k
(x) = (x)

X
k
(x). It follows from the inequality (5.4) that
_
_
_
_

1kjN
c
k
X

k
_
_
_
_
0
=
_
_
_
_

1kj N
c
k

X
k
__
_
_
_
0
< M
1
;
then
_
_
_
_

x
i

1kj N
c
k
X

k
x
i
_
_
_
_
0
=
_
_
_
_

_
1

x
i

1kj N
c
k


X
k
x
i

x
i

1kjN
c
k

X
k
__
_
_
_
0
=
_
_
_
_

_

x
i
_

_
+

x
i

1kj N
c
k


X
k
x
i

x
i

1kjN
c
k

X
k
__
_
_
_
0
M
2
_
_
_
_

x
i
_

1kjN
c
k


X
k
x
i
_
_
_
_
0
+M
2
_
_
_
_

1kj N
c
k

X
k
_
_
_
_
0
< M
3
,
where the constants M
1
and M
2
do not depend on . Therefore,
_
_
_
_

1kj N
c
k
X

k
_
_
_
_
1
< (M
1
+M
3
).
To complete the proof of assertion (2) of the lemma it is sucient to orthogonalize and normalize the functions
X

k
(x) in the space L
2
() and replace X
k
(x) by these newly obtained functions.
Let us prove assertion (3). Consider the Fourier-series expansion of v(x, t) in

Q with respect to the functions
sin
m
T
t (m = 1, 2, . . .):
v(x, t) =

m=1

m
(x) sin
m
T
t.
This series converges absolutely and uniformly in

Q together with its derivatives (see [84, p. 507]). Denote
v
l
(x, t) =
l

m=1

m
(x) sin
m
T
t.
For a suciently large l we have
__
Q
_
(v v
l
)
2
+
n

i=1
_
(v v
l
)
x
i
_
2
_
dxdt < . (5.6)
504
On the other hand, according to the assertion (2) that we have already proved, one can choose N such that
_
_
_
_

m
(x)
N

k=1
c
km
X
k
(x)
_
_
_
_
2
1
<

l
, (m = 1, . . . , l).
From this inequality and from (5.6) it follows that
__
Q
_
(v v
N
)
2
+
n

i=1
_
(v v
N
)
x
i
_
2
_
dxdt < M
4
,
where
v
N
(x, t) =
l

m=1
sin
m
T
t
N

k=1
c
km
X
k
(x) =
N

k=1
X
k
(x)
l

m=1
c
km
sin
m
T
t =
N

k=1
c
k
(t)X
k
(x),
and the constant M
4
does not depend on and N.
Lemma 2 is proved. It will be used in Sect. 9.
In what follows, besides the spaces introduced above we shall consider the space W
k
() of the functions
(x) such that their derivatives up to order k belong to L
2
(). By W
2k,k
(Q) we shall denote the space of functions
v(x, t) such that

l+m
v
x
l1
1
. . . x
ln
n
t
m
L
2
(Q), where l + 2m 2k.
The functions from W
2k,k
(Q) have in the cylinder Q generalized derivatives belonging to L
2
(Q) up to order 2k with
respect to the space variables x
1
, . . . , x
n
and up to order k with respect to the time t and the corresponding mixed
derivatives. The norm |v|
2k,k
of the function v(x, t) in the space W
2k,k
(Q) will be dened by the formula
|v|
2k,k
=
___
Q

0l+2m2k
_

l+m
v
x
l1
1
. . . x
ln
n
t
m
_
2
dxdt
_
1/2
.
Consider the space of functions obtained as a closure of the functions from the class C

(

Q) vanishing in
in the norm of W
2,1
(Q). Denote this space by

W
2,1
(Q). Obviously

W
2,1
(Q) is a subspace of the Hilbert space
W
2,1
(Q) with the scalar product
(v
1
, v
2
)
2,1
=
__
Q
_
v
1
v
2
+
n

i=1
v
1
x
i
v
2
x
i
+
v
1
t
v
2
t
+
n

i,j=1

2
v
1
x
i
x
j

2
v
2
x
i
x
j
_
dxdt
for any elements of v
1
and v
2
.
2. Generalized Solutions and Uniqueness Theorems. Here we consider generalized solutions of
the rst boundary-value problem for a parabolic equation. We shall give two denitions of the generalized solution
of this problem and prove the corresponding uniqueness theorems.
Denition 1. A function u(x, t) is called a weak solution of the rst boundary-value problem for the equation
L(u)
n

i,j=1

x
i
_
A
ij
(x, t)
u
x
j
_
+
n

i=1
B
i
(x, t)
u
x
i
+C(x, t)u
u
t
= F(x, t) (5.7)
in the cylinder Q with conditions
u(x, 0) = (x), u[
S
= 0, (5.8)
if u(x, t)
0
W
1,1
(Q), u(x, 0) = (x), and for any function (x, t)
0
W
1,1
(Q) the equality
__
Q
_
n

i,j=1
A
ij
u
x
j

x
i

_
n

i=1
B
i
u
x
i
+Cu
u
t
F
_

_
dxdt = 0 (5.9)
holds. We suppose that A
ij
(x, t), B
i
(x, t), and C(x, t) are bounded and measurable and F(x, t) L
2
(Q).
505
We note immediately that the validity of the identity (5.9) for all (x, t)
0
W
1,1
(Q) is equivalent to its
validity for all (x, t)
0
C

(Q). This follows from the density of the set of functions


0
C

(Q) in
0
W
1,1
(Q).
Theorem 2. A weak solution of the problem (5.7), (5.8) is unique.
Proof. Denote by u(x, t) a dierence between two solutions of the problem. This function obviously satises
the equality
__
Q
_
n

i,j=1
A
ij
u
x
j

x
i

_
n

i=1
B
i
u
x
i
+Cu
u
t
_

_
dxdt = 0 (5.10)
for any (x, t)
0
W
1,1
(Q) and the condition u(x, 0) = 0.
In the identity (5.10) we can set (x, t) = u(x, t)e
t
and then transform the equality thus obtained, taking
into account that u(x, t)
0
W
1,1
(Q) and u(x, 0) = 0. We have
0 =
__
Q
_
u
u
t
e
t
+e
t
_
n

i,j=1
A
ij
u
x
i
u
x
j

i=1
B
i
u
x
i
u Cu
2
__
dxdt
=
1
2
_
t=T
u
2
e
T
dx +
__
Q
e
t
_

2
u
2
+
n

i,j=1
A
ij
u
x
i
u
x
j

i=1
B
i
u
x
i
u Cu
2
_
dxdt.
Since the coecients B
i
(x, t) and C(x, t) are bounded and the matrix |A
ij
(x, t)| is positive denite, we
arrive at the inequality
0
__
Q
e
t
_

2
u
2
+
n

i=1
_
u
x
i
_
2
Mu
2
M
n

i=1

u
x
i

[u[
_
dxdt

__
Q
e
t
_

2
u
2
+
n

i=1
_
u
x
i
_
2
Mu
2

i=1
_
u
x
i
_
2
M
1
u
2
_
dxdt =
__
Q
e
t
_

2
M M
1
_
u
2
dxdt,
(5.11)
where M and M
1
are constants depending only on the coecients of Eq. (5.7) and on the dimension n of the space E
n
.
If we choose > 2(M +M
1
), then it follows from the inequality (5.11) that u(x, t) 0 in the cylinder Q.
Denition 2. Suppose that the functions u
m
(x, t) C
3
(

Q) satisfy in Q the equation
L(u
m
)
n

i,j=1
a
ij
(x, t)

2
u
m
x
i
x
j
+
n

i=1
b
i
(x, t)
u
m
x
i
+c(x, t)u
m

u
m
t
= f
m
(x, t).
Suppose that
u
m
(x, 0) =
m
(x), u
m
[
S
= 0.
If the f
m
(x, t) converge in the norm of L
2
(Q) to the function f(x, t) as m , the
m
(x) converge in the norm
of W
1
() to the function (x), and the u
m
(x, t) converge in the norm of L
2
(Q) to the function u(x, t), then u(x, t)
is called a strong solution of the equation
L(u)
n

i,j=1
a
ij
(x, t)

2
u
x
i
x
j
+
n

i=1
b
i
(x, t)
u
x
i
+c(x, t)u
u
t
= f(x, t) (5.12)
in the cylinder Q with the condition (5.8).
Note that it follows from the inequality (2.35) of Sect. 2 that the u
m
converge to u in the norm of W
2,1
(Q).
We shall suppose that a
ij
(x, t) are continuous in

Q, b
i
(x, t) and c(x, t) are bounded, and A
2
.
506
Theorem 3. A strong solution of the problem (5.12), (5.8) is unique.
Proof. Suppose there are two solutions of this problem: u(x, t) and u(x, t). This means that there exist
sequences f
m
,
m
, u
m
and

f
m
,
m
, u
m
, converging in the norm of L
2
respectively to f, , u and f, , u.
Set v
m
= u
m
u
m
. According to the a priori estimate (2.35) we have
|v
m
|
2,1
M
2
(|f
m


f
m
|
0
+|
m

m
|
1
). (5.13)
Since the right-hand side of inequality (5.13) tends to zero as m , we also have v
m
0 as m ; therefore,
u u.
Passing to the limit as m in the equation L(u
m
) = f
m
, we can see that the strong generalized solution
u(x, t) satises Eq. (5.12) almost everywhere in

Q and belongs to the space W
2,1
(Q). Condition (5.8) is satised in
the mean, in the sense that was dened above for the functions of class W
1,1
(Q).
The following proposition is also valid. Let the function v(x, t) W
2,1
(Q), v[
S
= 0, v(x, 0) = (x), and
let v(x, t) satisfy Eq. (5.12) almost everywhere in

Q. Let u(x, t) be a strong generalized solution of the problem
(5.12), (5.8), A
2
, and a
ij
(x, t) C
1
(Q). Then u(x, t) v(x, t).
In fact, multiplying (5.12) by a function (x, t)
0
C

(Q) and integrating over Q, we obtain


__
Q
_
n

i,j=1
a
ij
v
x
i

x
j

_
n

i=1
_
b
i

j=1
a
ij
x
j
_
v
x
i
+cv
v
t
f
_

_
dxdt = 0.
The same identity is valid for the function u(x, t). Therefore u(x, t) and v(x, t) are weak solutions of the rst
boundary-value problem for the equation
n

i,j=1

x
i
_
a
ij
u
x
j
_
+
n

i=1
_
b
i

j=1
a
ij
x
j
_
u
x
i
+cu
u
t
= f
in the cylinder Q with the conditions u(x, 0) = (x) and u[
S
= 0. Since the weak solution of the rst boundary-
value problem is unique, we have u(x, t) v(x, t).
6 METHOD OF FINITE DIFFERENCES
Consider the rst boundary-value problem for Eq. (5.7) in the cylinder Q = (0, T), where is a bounded
domain in the space E
n
with the boundary A
1
with the conditions (5.8). We suppose that the coecients of
Eq. (5.7) are continuous in

Q,
Aij(x,t)
t
are bounded, and the right-hand side F(x, t) L
2
(Q); the matrix |A
ij
(x, t)|
is assumed to be symmetric and positive denite:
n

i,j=1
A
ij

j

n

i=1

2
i
, > 0. We shall prove the existence
of a weak generalized solution of the boundary-value problem (5.7), (5.8) by the method of nite dierences (see
Sect. 5, Denition 1). We shall suppose that the function (x) belongs to the class
0
W
1
().
First we prove the existence of a generalized solution of the problem (5.7), (5.8) under the supposition that
(x)
0
C
1
() and F(x, t) C(

Q).
Let us cover the domain 0 t T, x E
n
by a rectangular grid consisting of the lines t = rh and
x
i
= p
i
h (i = 1, . . . , n), where the p
i
take all integer values and r takes all integer values on the interval
_
0,
T
h

. We
suppose that T is a multiple of h. The node points of the grid are points with coordinates that are multiples of h.
At the node points we shall dene a function u
h
. Denote by S
h
the node points belonging to

Q that are
neighbors to a node point lying outside

Q S. The remaining nodes of the grid lying in

Q will be called inner
nodes: the set of inner nodes is denoted by Q
h
. Dene the function u
h
as zero at the nodes S
h
and at the nodes lying
outside

Q S. On the base of the cylinder Q set at the node points u
h
(x, 0) = (x). The values of the function u
h
at the inner node points Q
h
are determined by a system of linear equations obtained by substituting the derivatives
in Eq. (5.7) by corresponding dierence quotients.
507
Let (x
1
, . . . , x
n
, t) be a node point of the grid. For simplicity we shall use the following notations:
v(x
1
, . . . , x
n
, t) = v, v(x
1
, . . . , x
k1
, x
k
h, x
k+1
, . . . , x
n
, t) = v(x
k
h), v(x
1
, . . . , x
n
, t h) = v(t h),
v(x
k
+h) v
h
=
v

x
k
,
v v(x
k
h)
h
=
v
x
k
,

_
v
x
k
_

x
i
=

2
v

x
i
x
k
;
similar notations will be used for dierence quotients with respect to t. To be concise we shall omit the index h of
the function u
h
.
For every point of Q
h
the following linear equation can be written:
n

i,j=1

x
i
_
A
ij
u
x
j
_
+
n

i=1
B
i
u
x
i
+Cu
u
t
= F. (6.1)
Besides the value of the function u at the chosen point this equation contains the values of u in the neighboring
node points. The values of the coecients and the right-hand side are taken at the chosen point. Thus we obtain
m linear equations with m unknowns, where m is the number of inner node points. The fact that u = 0 at the points
of S
h
and u[
t=0
= has to be taken into account. The solvability of this system will be shown below.
Suppose that a solution of the system (6.1) exists. Under this supposition we can obtain some estimates
for the solution and its dierence quotients, independent of h. To do this we multiply both sides of Eq. (6.1) by
u
t
e
t
and then transform the equality thus obtained. For u and its dierence quotients
u
xi
and
u
t
we can
obtain an estimate similar to the a priori estimate (2.30) for the solutions of Eq. (5.7). The calculations including
dierence quotients are similar to those carried out when we deduced estimate (2.30). We also use the obvious
identities
v
w

x
i
=
(vw)

x
i
w(x
i
+h)
v

x
i
=
(vw(x
i
+h))
x
i
w
v
x
i
,
v
w
x
i
=
(vw)
x
i
w(x
i
h)
v
x
i
,
v
w
t
=
(vw)
t
w(t h)
v
t
.
_

_
(6.2)
By N
i
we shall denote quantities not greater in absolute value than
M
_

u
t

_
[u[ +
n

i=1

u
x
i

_
+
n

i=1
_
u(t h)
x
i
_
2
_
e
t
,
where the constant M depends only on the coecients of Eq. (5.7).
Thus, multiplying (6.1) by
u
t
e
t
, we obtain
F
u
t
e
t
=
_
u
t
_
2
e
t

i,j=1
u
t

x
i
_
A
ij
u
x
j
_
e
t
+N
1
=
_
u
t
_
2
e
t
+
n

i,j=1
A
ij
u
x
j

2
u
x
i
t
e
t

i,j=1

x
i
_
u
t
A
ij
(x
i
+h)
u(x
i
+h)
x
j
e
t
_
+N
1
. (6.3)
This last relation is also valid for the points of S
h
, because at these points we have
u
t
= 0, and this means
that both sides of equality (6.3) vanish. Recall that in accordance with (6.2), we have the identity

t
_
A
ij
u
x
i
u
x
j
e
t
_
=
(A
ij
e
t
)
t
u(t h)
x
i
(t h)
x
j
+A
ij
e
t

t
_
u
x
i
u
x
j
_
=
_
A
ij
t
e
t
+A
ij
(t h)
e
t
t
_
u(t h)
x
j
u(t h)
x
j
+A
ij
e
t

t
_
u
x
i
u
x
j
_
.
508
We transform the rst sum in the right-hand side of (6.3):
n

i,j=1
A
ij

2
u
x
i
t
u
x
j
e
t
=
1
2
n

i,j=1

t
_
A
ij
u
x
i
u
x
j
e
t
_

1
2
n

i,j=1
A
ij
t
e
t
u(t h)
x
i
u(t h)
x
j

1
2
n

i,j=1
e
t
t
A
ij
(t h)
u(t h)
x
i
u(t h)
x
j

1
2
n

i,j=1
A
ij
e
t

t
_
u
x
i
u
x
j
_
+
n

i,j=1
A
ij

2
u
x
i
x
j
u
x
j
e
t
=
1
2
n

i,j=1

t
_
A
ij
u
x
i
u
x
j
e
t
_
+N
2

1
2
n

i,j=1
e
t
t
A
ij
(t h)
u(t h)
x
i
u(t h)
x
j
+
1
2
n

i,j=1
A
ij
_
2

2
u
x
i
t
u
x
j


2
u
x
i
t
_
u
x
j
+
u(t h)
x
j
__
e
t
. (6.4)
The last sum is equal to
1
2
n

i,j=1
A
ij
_
u
x
j

u(t h)
x
j
_

2
u
x
i
t
=
h
2
n

i,j=1
A
ij

2
u
x
i
t

2
u
x
j
t
0. (6.5)
Moreover, we have
e
t
t
=
e
t
e
(th)
h
= e
t
e
h
1
h
< e
t
,
and therefore

1
2
n

i,j=1
e
t
t
A
ij
(t h)
u(t h)
x
i
u(t h)
x
j

e
t
2
n

i=1
_
u(t h)
x
i
_
2
. (6.6)
From (6.3) and (6.4), taking into account estimates (6.5) and (6.6), we obtain the inequality
F
u
t
e
t

_
u
t
_
2
e
t
+
1
2
n

i,j=1

t
_
A
ij
u
x
i
u
x
j
e
t
_
+
e
t
2
n

i=1
_
u(t h)
x
i
_
2

i,j=1

x
i
_
u
t
A
ij
(x
i
+h)
u(x
i
+h)
x
i
e
t
_
+N
3
. (6.7)
Let us sum this inequality over all the points of S
h
and Q
h
. The sum over all these points will be denoted
by

Q
, and the sum over the points lying on the plane t = t
0
, where 0 t
0
T, will be denoted by

t=t0
. Since
u
t
= 0 at the points S
h
, the sum

xi
_
u
t
A
ij
(x
i
+h)
u(xi+h)
xj

, taken over the points Q


h
and S
h
, is equal to zero.
Moreover, it is obvious that

Q
n

i,j=1

t
_
A
ij
u
x
i
u
x
j
e
t
_
=
1
h

t=T
n

i,j=1
A
ij
u
x
i
u
x
j
e
T

1
h

t=0
n

i,j=1
A
ij
u
x
i
u
x
j
.
Thus, summing (6.7) over Q
h
and S
h
, we have

Q
__
u
t
_
2
+

2
n

i=1
_
u(t h)
x
i
_
2
_
e
t
+
1
2h

t=T
n

i,j=1
A
ij
u
x
i
u
x
j
e
T

1
2h

t=0
A
ij
u
x
i
u
x
j

Q
F
u
t
e
t
+M
1

Q
_

u
t

_
[u[ +
n

i=1

u
x
i

_
+
n

i=1
_
u(t h)
x
i
_
2
_
e
t

Q
_
F
2
+
1
4
_
u
t
_
2
_
e
t
+
1
4

Q
_
u
t
_
2
e
t
+M
2

Q
_
u
2
+
n

i=1
_
u
x
i
_
2
_
e
t
+M
1

t=0
n

i=1
_
u
x
i
_
2
.
Here and below the constants M
i
depend only on the coecients of Eq. (5.7) and on the domain Q.
509
Therefore,
1
2

Q
_
u
t
_
2
e
t
+

2
n

i=1
_
u(t h)
t
_
2
e
t
+
1
2h

t=T
n

i,j=1
A
ij
u
x
i
u
x
j
e
T

Q
F
2
e
t
+M
2

Q
_
u
2
+
n

i=1
_
u
x
i
_
2
_
e
t
+M
1

t=0
n

i=1
_
u
x
i
_
2
+
1
2h

t=0
n

i,j=1
A
ij
u
x
i
u
x
j
.
Since
n

i,j=1
A
ij
u
xi
u
xj

n

i=1
_
u
xi
_
2
and u(x, 0) = (x), we have

Q
_
u
t
_
2
e
t
+

Q
n

i=1
_
u(t h)
x
i
_
2
e
t
+

h

t=T
n

i=1
_
u
x
i
_
2
e
T
2

Q
F
2
e
t
+M
3
_

Q
_
u
2
+
n

i=1
_
u
x
i
_
2
_
e
t
+
1
h

i=1
_
(x)
x
i
_
2
_
, (6.8)
where

denotes the sum over all the node points of .


Estimate now the sum

Q
u
2
e
t
in terms of

Q
n

i=1
_
u
xi
_
2
e
t
. Let d
0
denote the diameter of the domain .
One can express the value of the function u at a chosen node point P(x
1
, . . . , x
n
, t) in terms of dierence quotients
of this function with respect to the space variables. Let the point
P
i
(x
1
, . . . , x
i1
, x
i
p
i
h, x
i+1
, . . . , x
n
, t)
belong to S
h
. Then
u(P) = u(P) u(P
i
) = h
_
u
x
i
+
u(x
i
h)
x
i
+. . . +
u(x
i
(p
i
1)h)
x
i
_
.
Therefore,
u
2
(P) p
i
h
2
pi1

l=0
_
u(x
i
lh)
x
i
_
2
d
0
h
n

i=1
pi1

l=0
_
u(x
i
lh)
x
i
_
2
.
Multiplying this inequality by e
t
and summing over all the points of Q
h
S
h
, we get

Q
u
2
e
t
d
2
0

Q
n

i=1
_
u
x
i
_
2
e
t
. (6.9)
Hence, taking into account the inequality (6.8), we have

Q
_
u
t
_
2
e
t
+

Q
n

i=1
_
u(t h)
x
i
_
2
e
t
+

h

t=T
n

i=1
_
u
x
i
_
2
e
t
2

Q
F
2
e
t
+M
4

Q
n

i=1
_
u
x
i
_
2
e
t
+
M
3
h

i=1
_
(x)
x
i
_
2
. (6.10)
Suppose that >
2M4

and h <
1

. Then from the inequality (6.10) it follows that

Q
_
u
t
_
2
e
t
+

2

Q
n

i=1
_
u
x
i
_
2
e
t
2

Q
F
2
e
t
+
M
3
h

i=1
_
(x)
x
i
_
2
.
510
Taking into account that 1 e
t
e
T
, we get

Q
__
u
t
_
2
+
n

i=1
_
u
x
i
_
2
_
M
5
_

Q
F
2
+
1
h

i=1
_
(x)
x
i
_
2
_
.
With the help of (6.9), we arrive at the nal inequality

Q
_
u
2
+
_
u
t
_
2
+
n

i=1
_
u
x
i
_
2
_
M
6
_

Q
F
2
+
1
h

i=1
_
(x)
x
i
_
2
_
. (6.11)
From inequality (6.11) follows the solvability of the system (6.1). In fact, the solution of the homogeneous
system (6.1) corresponds to the case where F 0 and 0. It follows from the inequality (6.11) that in this case
u
h
= 0 at all the points of Q
h
. The solvability of the system (6.1) follows from the uniqueness of its solution.
Let (x, t) be a point with coordinates satisfying the inequalities (p
i
1)h < x
i
p
i
h (i = 1, . . . , n),
(r 1)h < t rh. We set u
h
(x, t) = u
h
(p
1
h, . . . , p
n
h, rh). In this way we dene a piecewise-constant function
u
h
(x, t) (that in general is not continuous) everywhere in the layer

H0 t T, x E
n
. This function is equal
to zero at all the points outside

Q.
In this way we can extend to the whole layer

H also the functions
u
h
xi
and
u
h
t
, obtaining the piecewise-
constant functions u
h
i
(x, t) and u
h
0
(x, t) respectively. Multiplying (6.11) by h
n+1
, we come to the following relation:
__
Q
_
(u
h
)
2
+
n

i=1
(u
h
i
)
2
+ (u
h
0
)
2
_
dxdt M
6
_

Q
F
2
h
n+1
+

i=1
_
(x)
x
i
_
2
h
n
_
. (6.12)
Since
(x)
xi
=

xi
(x
1
, . . . , x
i1
,
i
, x
i+1
, . . . , x
n
), where x
i
h <
i
< x
i
, in the right-hand side of the in-
equality (6.12) we have integral sums for the functions F
2
(x, t) in Q and
n

i=1
_
(x)
xi
_
2
in . These sums are uniformly
bounded with respect to h. Therefore, the functions u
h
, u
h
i
, and u
h
0
are uniformly bounded in the L
2
(Q)-norm with
respect to h, and one can select a subsequence h
k
0 such that u
h
k
, u
h
k
i
, and u
h
k
0
converge weakly in L
2
(Q) to
the functions u, u
i
, and u
0
respectively. Since the norm of a weak limit of a sequence of functions does not exceed
the lower limit of their norms, we obtain from (6.12)
__
Q
_
u
2
+
n

i=1
u
2
i
+u
2
0
_
dxdt M
6
_ __
Q
F
2
dxdt +
_

i=1
_

x
i
_
2
dx
_
. (6.13)
Let us prove that u
i
=
u
xi
, u
0
=
u
t
, and that the function u satises conditions (5.8). Suppose that the func-
tion (x, t)
0
C

(Q) and vanishes for t = T. We have the following obvious relation:

Q
_
(t h)
u
h
t
+u
h

t
_
=

Q
(u
h
)
t
=
1
h

t=0
u
h
. (6.14)
Let us dene piecewise-constant functions
h
(x, t) and
h
0
(x, t), setting at the node points
h
(x, t) = (x, t h)
and
h
0
(x, t) =
(x,t)
t
and extending the functions
h
and
h
0
to the whole layer

H in the same way as was done
for the functions u
h
. We can multiply (6.14) by h
n+1
, and since u
h
(x, 0) = (x), we obtain
__
Q
(
h
u
h
0
+u
h

h
0
) dxdt =
_

h
(x, h)
h
(x) dx, (6.15)
where
h
(x) is a piecewise-constant function, equal to (x) at the nodes of the grid and extended to all x E
n
in the same way as u
h
. Since for h 0 the functions
h
(x, t) converge uniformly in Q to (x, t), the functions

h
(x, h) converge uniformly in to (x, 0), and the functions
h
0
(x, h) converge uniformly in Q to
(x,t)
t
, we can
511
pass to the limit in the relation (6.15) as h
k
0. Taking into account that u
h
k
and u
h
k
0
converge weakly to u and
u
0
respectively, we get
__
Q
_
u
0
+u

t
_
dxdt =
_

(x, 0)(x) dx. (6.16)


If the function (x, t) has compact support in Q, then the relation (6.16) means that u(x, t) has a generalized
derivative
u
t
in Q equal to u
0
(x, t) and that
u
t
L
2
(Q).
In the same way, one can show that
u
xi
= u
i
, and this means that the function u(x, t) belongs to W
1,1
(Q).
Therefore, for u(x, t) we have the relation
__
Q
_

u
t
+u

t
_
dxdt =
_

(x, 0)u(x, 0) dx. (6.17)


From (6.16) and (6.17) it follows that the initial condition u[
t=0
= (x) is satised.
Since the functions u
h
are equal to zero at the nodes of S
h
and outside

QS, we can assume that the integral
in the left-hand side of (6.13) is taken over the whole layer

H0 t T, x E
n
. Therefore, u(x, t) W
1,1
(H).
Since u(x, t) = 0 outside of

Q, the function u(x, t) vanishes on S and therefore belongs to
0
W
1,1
.
To show that u(x, t) is a generalized solution of the problem (5.7), (5.8), we only need to establish the rela-
tion (5.9).
Suppose that (x, t)
0
C

(Q). Let us multiply Eq. (6.1) at every inner node point (x, t) of the grid by
(x, t) and then transform the equality thus obtained. We get
F =
u
t
+
n

i,j=1

x
i
_
A
ij
u
x
j
_
+
n

i=1
B
i
u
x
i
+ Cu
=
u
t
+
n

i,j=1

x
i
_
A
ij
(x
i
+h)
u(x
i
+h)
x
j
_

i,j=1
A
ij

x
i
u
x
j
+
n

i=1
B
i
u
x
i
+ Cu. (6.18)
Since (x, t) = 0 at the points of S
h
(for suciently small h), the last relation holds also at these points.
Summing the equality (6.18) over all inner and boundary points, we obtain

Q
F =

Q
_

u
t

n

i,j=1
A
ij

x
i
u
x
j
+
n

i=1
B
i
u
x
i
+ Cu
_
. (6.19)
Dene the piecewise-constant functions F
h
(x, t),
h
(x, t),
h
i
(x, t), A
h
ij
(x, t), B
h
i
(x, t), and C
h
(x, t) as we
have dened u
h
(x, t),
h
(x, t), and
h
0
(x, t). Multiplying (6.19) by h
n+1
, we can bring it to the form
__
Q
F
h

h
dxdt =
__
Q
_

h
u
h
0

n

i,j=1
A
h
ij

h
i
u
h
j
+
n

i=1

h
B
h
i
u
h
i
+
h
C
h
u
h
_
dxdt. (6.20)
In (6.20) we can pass to the limit as h = h
k
0. Taking into account that the functions F
h
,
h
, A
h
ij

h
i
,

h
B
h
i
, and
h
C
h
converge in L
2
(Q) as h 0 to the functions F, , A
ij

xi
, B
i
, and C respectively, and
the functions u
h
, u
h
i
, and u
h
0
converge weakly as h = h
k
0 to u,
u
xi
, and
u
t
respectively, we obtain the iden-
tity (5.9).
Thus we have proved the existence of the generalized solution of the problem (5.7), (5.8) under the supposition
that (x)
0
C
1
() and F(x, t) C(

Q). The solution we have constructed by virtue of (6.13) satises the inequality
__
Q
_
u
2
+
n

i=1
_
u
x
i
_
2
+
_
u
t
_
2
_
dxdt M
6
_ __
Q
F
2
dxdt +
_

i=1
_

x
i
_
2
dx
_
. (6.21)
512
Suppose now that (x)
0
W
1
() and F(x, t) L
2
(Q). We can select subsequences
m
(x)
0
C

() and
F
m
(x, t) C(

Q) such that
m
in the norm of W
1
(), and F
m
F in the norm of L
2
(Q). As we have
proved, for any pair of functions
m
(x) and F
m
(x, t) in the cylinder Q there exists a generalized solution u
m
(x, t)
of the equation
n

i,j=1

x
i
_
A
ij
(x, t)
u
m
x
j
_
+
n

i=1
B
i
(x, t)
u
m
x
i
+C(x, t)u
m

u
m
t
= F
m
(x, t)
with the conditions
u
m
[
t=0
=
m
(x), u
m
[
S
= 0.
Taking into account the estimate (6.21), we can see that the functions u
m
(x, t) are bounded in the norm of W
1,1
(Q)
uniformly in m. Therefore, one can select a subsequence u
m
k
(x, t) converging weakly in W
1,1
(Q) to the function
u(x, t). In the same way as was done above, we can prove that u[
t=0
= (x) and u[
S
= 0.
Passing to the limit in the identity (5.9), written for the solutions u
m
(x, t), we can see that the limit function
u(x, t) also satises this identity.
The existence of a solution of the problem (5.7), (5.8) is proved.
Since the generalized solution of the problem (5.7). (5.8) is unique, the whole sequence u
h
converges to u(x, t)
as h 0. The functions u
h
are approximate solutions of this problem.
The method of nite dierences is widely used in the theory of partial dierential equations to prove the ex-
istence of solutions of boundary-value problems and also to nd approximate solutions of these problems. The ele-
ments of the method of nite dierences can be found in [19]. For the rst time the method of nite dierences was
used to prove the existence of the solution of the Dirichlet problem for the Laplace equation in L. A. Lusterniks
paper [85]. Then this method was used to prove the existence of solutions of problems for equations of various types
by R. Courant, C. Friedrichs, and H. Levy [86].
A review of the literature concerning the use of the method of nite dierences in proofs of the existence
theorems is given in [61]. Questions concerning numerical solution of parabolic equations by the method of nite
dierences are studied also in [87].
7 SOME FUNCTIONAL METHODS OF SOLVING BOUNDARY-VALUE PROBLEMS
1. Recently the so-called functional methods of solving boundary-value problems became widespread.
The essence of these methods is to consider a boundary-value problem as a problem of solving some operator
equation in a specially chosen functional space.
The origin of functional methods is connected with the names of S. L. Sobolev [82, 92] and K. Friedrichs [93].
Later these methods were used for the solution of boundary-value problems for elliptic and hyperbolic equations
and systems in [8891, 9597] and so on. Boundary-value problems for parabolic equations and systems were also
studied by functional methods (see, e.g., [91, 99101, 125128]).
In what follows, we shall consider the rst boundary-value problem as an example of the use of functional
methods for the proof of existence theorems for generalized solutions of boundary-value problems. This method
coincides in its most important features with the method used by M. I. Vishik in [90,99] for the proof of the solvability
of the boundary-value problems for parabolic and hyperbolic systems.
2. Consider in the cylinder Q = (0, T) Eq. (5.7) with conditions
u(x, 0) = 0, u[
S
= 0. (7.1)
Suppose that
n

i,j=1
A
ij
(x, t)
i

j

n

i=1

2
i
for all the points (x, t)

Q, where > 0. We shall construct a generalized
solution of the problem (5.7), (7.1) in the sense of Denition 1 of Sect. 5.
Denote by

W
1,1
(Q) the Hilbert space of functions belonging to W
1,1
(Q) and vanishing on . In the space

W
1,1
(Q) the scalar product u
1
, u
2
=
__
Q
_
u1
t
u2
t
+
n

i=1
u1
xi
u2
xi
_
dxdt is introduced. The corresponding norm will
be denoted by |u|
0
1,1
. As before, (u
1
, u
2
) =
__
Q
u
1
u
2
dxdt denotes the scalar product in L
2
(Q). By V (Q) we shall
513
denote the set of functions v(x, t) belonging to

W
1,1
(Q) and having mixed derivatives

2
v
xi t
(i = 1, . . . , n) square
integrable in Q. Obviously V (Q) is dense in

W
1,1
(Q).
According to Sect. 5, u(x, t) is a weak generalized solution of the problem (5.7), (7.1) if u(x, t)

W
1,1
(Q)
and satises the integral identity
_
u
t
,
_
+
n

i,j=1
_
A
ij
u
x
i
,

x
j
_

i=1
_
B
i
u
x
i
,
_
(Cu, ) = (F, ) (7.2)
for any function (x, t) of class
0
W
1,1
(Q). Obviously, to prove the existence of a weak generalized solution of
the problem (5.7), (7.1) it is sucient to show that there exists a function u(x, t)

W
1,1
(Q), satisfying the integral
identity
_
u
t
,
v
t
e
t
_
+
n

i,j=1
_
A
ij
u
x
i
,

2
v
x
j
t
e
t
_

i=1
_
B
i
u
x
i
,
v
t
e
t
_

_
Cu,
v
t
e
t
_
=
_
F,
v
t
e
t
_
(7.3)
for any function v(x, t) V (Q) and any > 0. In fact, if (x, t)
0
W
1,1
(Q), then v(x, t) =
t
_
0
(x, )e

d
belongs to V (Q). Substituting this v(x, t) in (7.3), we get (7.2). The uniqueness of the weak generalized solution
of the problem (5.7), (7.1) is proved in Sect. 5.
Theorem. If A
ij
, B
i
, C, and
Aij
t
are bounded and measurable in Q and F L
2
(Q), then there exists
a weak generalized solution of the problem (5.7), (7.1).
The proof of this theorem is based on a lemma about operators in a Hilbert space. Let A be an operator in
a Hilbert space H. Denote by D
A
(D
A
H) the domain of operator A and by R
A
its range.
Lemma. If D
A
is a dense set in H and the operator A has a bounded inverse A
1
, then the range R
A
of
the operator A

conjugate to A coincides with the whole space H.


To prove the lemma it is sucient to show that for any element h H an element y can be found such that
A

y = h.
Let us take an arbitrary element v D
A
. In view of the suppositions of the lemma, there exists an element w
of R
A
such that Av = w and v = A
1
w. The linear functional
l(w) = (h, v) = (h, A
1
w)
is bounded in R
A
, since
[l(w)[ |h| |A
1
w| |A
1
| |h| |w|.
Because the operator A
1
is bounded, the domain R
A
is closed, and this means that it is a subspace of H. According
to Rieszs theorem on the form of general linear functionals in H(see, e.g., [83, p. 180]) there exists a unique element y
such that l(w) = (y, w). This last equality can be rewritten in the following form: (h, v) = (y, Av) for any v D
A
.
This means that y D
A
and A

y = h, which was to be proved.


Proof of the theorem. We shall use the following notation:
l
1
(u, v) =
_
u
t
,
v
t
e
t
_
+
n

i,j=1
_
A
ij
u
x
i
,

2
v
x
j
t
e
t
_

i=1
_
B
i
u
x
i
,
v
t
e
t
_

_
Cu,
v
t
e
t
_
, (7.4)
l
2
(v) =
_
F,
v
t
e
t
_
.
It is easy to verify that if v V (Q), then l
1
(u, v) is a bounded linear functional with respect to u, dened in

W
1,1
(Q).
In fact,
[l
1
(u, v)[ M
1
_
|v|
0
1,1
+
n

j=1
_
_
_
_

2
v
x
j
t
_
_
_
_
0
_
|u|
0
1,1
,
514
where M
1
depends only on the upper bound of the absolute values of A
ij
, B
i
, C, and
Aij
t
in

Q; this upper
bound will be called M. According to Rieszs theorem, there exists an operator A, dened on V (Q) and such that
l
1
(u, v) = u, Av.
The functional l
2
(v) is dened for v

W
1,1
(Q) and it is bounded, because

_
F,
v
t
e
t
_

|F|
0
_
_
_
_
v
t
_
_
_
_
0
|F|
0
|v|
0
1,1
.
Therefore, there exists an element F
1


W
1,1
(Q) such that
l
2
(v) = F
1
, v. (7.5)
Now one can rewrite the equality (7.3) in the form u, Av = F
1
, v.
Let us prove that the operator A dened on V (Q)

W
1,1
(Q) has a bounded inverse A
1
. To do this we
shall show that
v, v M
2
v, Av,
where M
2
does not depend on v.
Now we shall estimate each term on the right-hand side of (7.4), having made the substitution u = v. We
have
_
v
t
,
v
t
e
t
_
=
_
_
_
_
v
t
e

2
t
_
_
_
_
2
0
, (7.6)

i=1
_
B
i
v
x
i
,
v
t
e
t
_

n

i=1

_
B
i
v
x
i
,
v
t
e
t
_

M
n

i=1
_
_
_
_
v
x
i
e

2
t
_
_
_
_
0
_
_
_
_
v
t
e

2
t
_
_
_
_
0

Mn
2
_
_
_
_
v
t
e

2
t
_
_
_
_
2
0

M
2
n

i=1
_
_
_
_
v
x
i
e

2
t
_
_
_
_
2
0
, (7.7)
where > 0 is any positive number. Then
_
Cv,
v
t
e
t
_

_
Cv,
v
t
e
t
_

M
_
_
_ve

2
t
_
_
_
0
_
_
_
_
v
t
e

2
t
_
_
_
_
0

M
2
_
_
_
_
v
t
e
2t
_
_
_
_
2
0

M
2
_
_
_ve

2
t

2
0

M
2
_
_
_
_
v
t
e

2
t
_
_
_
_
2
0

MM
3
2
n

i=1
_
_
_
_
v
x
i
e

2
t
_
_
_
_
2
0
. (7.8)
In the last estimate we have used the inequality
_
_
_ve

2
t
_
_
_
2
0
M
3
n

i=1
_
_
_
_
v
x
i
e

2
t
_
_
_
_
2
0
,
which is valid for functions belonging to W
1,1
(Q) vanishing on S (see Sect. 5, inequality (5.2)). Estimate now
n

i,j=1
_
A
ij
v
xi
,

2
v
xj t
e
t
_
:
n

i,j=1
_
A
ij
v
x
i
,

2
v
x
j
t
e
t
_
=
1
2
n

i,j=1
__
Q

t
_
A
ij
v
x
i
v
x
j
e
t
_
dxdt

1
2
n

i,j=1
_
A
ij
t
v
x
i
,
v
x
j
e
t
_
+

2
n

i,j=1
_
A
ij
v
x
i
,
v
x
j
e
t
_
=
1
2
_
t=T
n

i,j=1
A
ij
v
x
i
v
x
j
e
T
dx
1
2
n

i,j=1
_
A
ij
t
v
x
i
,
v
x
j
e
t
_
+

2
n

i,j=1
_
A
ij
v
x
i
,
v
x
j
e
t
_


2
n

i=1
_
_
_
_
v
x
i
e

2
t
_
_
_
_
2
0

M
2
n

i,j=1
_
_
_
_
v
x
i
e

2
t
_
_
_
_
0
_
_
_
_
v
x
j
e

2
t
_
_
_
_
0

2

Mn
2
_
n

i=1
_
_
_
_
v
x
i
e

2
t
_
_
_
_
2
0
. (7.9)
515
Taking into account (7.6)(7.9), we obtain
v, Av = l
1
(v, v)
_
1
Mn
2

M
2
__
_
_
_
v
t
e

2
t
_
_
_
_
2
0
+
_

2

Mn
2

M
2

MM
3
2
_
n

i=1
_
_
_
_
v
x
i
e

2
t
_
_
_
_
2
0
. (7.10)
Let us choose > 0 suciently small to have
1
Mn
2

M
2
>
1
2
,
and then choose suciently large to have

2

Mn
2

M
2

MM
3
2
>
1
2
.
It follows from the inequality (7.10) that
v, Av
1
2
e
T
_
_
_
_
_
v
t
_
_
_
_
2
0
+
n

i=1
_
_
_
_
v
x
i
_
_
_
_
2
0
_
=
1
2
e
T
v, v.
Thus we get |A
1
| 2e
T
. According to the lemma the operator A

is dened on the whole space



W
1,1
(Q).
Therefore, there exists an element u

W
1,1
(Q) such that
A

u = F
1
,
where F
1
is dened by formula (7.5).
So for any v V (Q) the equality
F
1
, v = A

u, v
holds, and, consequently,
F
1
, v = u, Av.
This means that u is a generalized solution of the problem (5.7), (7.1).
Note that for elliptic equations of any order the solution of the rst boundary-value problem is constructed
in a similar way by L. Garding in [88].
3. The method we shall describe below was used in its various forms for nding solutions of the boundary-
value problems for elliptic, hyperbolic, and parabolic equations and systems and also for solving the Cauchy problem
for higher-order hyperbolic equations and systems (see [89, 9597, 101]). The description of various applications of
this method can be found in [98]. It seems that the original idea of this method is due to K. Friedrichs. He applied
it to the rst boundary-value problem for Eq. (5.12) with conditions (7.1). The idea is as follows. The operator L
dened by the left-hand side of Eq. (5.12) is considered as dened on functions from the Hilbert space

W
2,1
(Q)
(see Sect. 5). This operator maps

W
2,1
(Q) into L
2
(Q). If we show that the set L(u), where u is any element of

W
2,1
(Q), coincides with the set L
2
(Q), then we shall have proved that the problem (5.12), (7.1) is solvable for any
f L
2
(Q). If the set L(u) is closed, it is sucient to show that in L
2
(Q) there is no nonzero element orthogonal
to all the elements of the form L(u). To prove this is the same as to prove the uniqueness of a generalized solution
(dened in a certain sense) of the problem conjugate to (5.12), (7.1).
Now consider the problem (5.12), (7.1) and let us nd a strong generalized solution of this problem (see
Sect. 5) for some function f L
2
(Q). As we have already noted, for such generalized solutions of the problem
(5.12), (7.1) the inequality (2.35) of Sect. 2 holds. It follows easily from this inequality that the range of opera-
tor L(u), dened for the functions u(x, t), belonging to

W
2,1
(Q), is closed in L
2
(Q). We shall show that this range
coincides with the whole space L
2
(Q) if we prove the following proposition: from the equality
(L(u), v) = 0, (7.11)
where u is any element of

W
2,1
(Q) and v is some element of L
2
(Q), it follows that v = 0. This proposition is true if
a
ij
, b
i
, c, and
aij
x
k
and their rst-order derivatives in t are bounded in Q and the boundary of the domain belongs
to the class A
2
. The proof of this proposition can be found in [101]. It is based on the substitution of a function
u

W
2,1
(Q), dened in a special way in terms of v, in the equality (7.11).
516
8 SOLUTION OF THE FIRST BOUNDARY-VALUE PROBLEM BY THE METHOD OF CONTINUATION
WITH RESPECT TO A PARAMETER
The method of solution of boundary-value problems by the method of continuation with respect to a param-
eter takes its origin in S. N. Bernsteins works on solutions of the Dirichlet problem for elliptic equations [48]. Then
this method was substantially developed in Leray and Schauders work [102]. Now this method is widely used to
prove the existence of solutions of boundary-value problems for various types of linear and nonlinear equations and
systems (see, e.g., [2, 18, 55, 103]).
For the demonstration of this method we shall take as an example the rst boundary-value problem for
Eq. (5.12). For simplicity we consider this equation in the cylinder Q with conditions (7.1). In a similar way, it is
possible to consider the case of a noncylindrical domain D and nonzero boundary conditions.
Consider the strong generalized solution of the problem (5.12), (7.1) in the sense of Denition 2 of Sect. 5.
In other words, u(x, t) is a solution of the problem (5.12), (7.1) if there exists a sequence of functions u
m
(x, t)
of class C
3
(

Q), converging in the mean to u(x, t) such that u
m
[

= 0 and the functions L(u


m
) = f
m
converge in
the mean to f. It follows from the estimate (2.35) that the u
m
converge to u in the norm of W
2,1
(Q). Note that
the inequality (2.35) holds also for a generalized solution of the problem (5.12), (7.1). We have already noted in
Sect. 5 that a strong generalized solution satises Eq. (5.12) almost everywhere and satises the condition (7.1) in
the mean. Thus, we have to nd a function u(x, t)

W
2,1
(Q) such that L(u) = f, where f L
2
(Q) is a given
function. We make the same suppositions on the coecients of (5.12) and the boundary of the domain as in
Theorem 8 of Sect. 2, when we obtained the inequality (2.35).
The following construction is based on the estimate (2.35) and on the existence in Q of the solution of
the problem for some equation of type (5.12), e.g., for the heat equation:
L
0
(u)
n

i=1

2
u
x
2
i

u
t
= f. (8.1)
The solution of Eq. (8.1) with conditions (7.1) can be obtained by many methods (it is possible to use
the Fourier method, the Laplace transform, or any of the methods considered in previous sections). It is sucient
to construct a solution only for functions f that are innitely dierentiable in Q and have compact support in this
cylinder (that is, vanish in a neighborhood of the boundary of Q). The set of these functions is dense in the space
L
2
(Q), and we can use the inequality (2.35) to obtain a solution of the problem (8.1), (7.1) in the space

W
2,1
(Q)
for any f L
2
(Q). The inequality (2.35) for the solution of problem (8.1), (7.1) means that there exists an inverse
operator L
1
0
for L
0
(u). This inverse operator is bounded and it maps the whole space L
2
(Q) into

W
2,1
(Q).
To prove the existence of a strong generalized solution of problem (5.12), (7.1) consider the set of parabolic
equations
L

(u) L(u) + (1 )L
0
(u) = f, (8.2)
depending on a parameter , 0 1. For = 0 the operator L

(u) coincides with L


0
(u), and for = 1 it
coincides with the operator L(u). The function u belongs to the space

W
2,1
(Q). Note that the estimate (2.35) is
valid for all the operators L

(u) with a constant M not depending on , since it is possible to estimate uniformly


in the moduli of continuity of the coecients in the higher derivatives, the least eigenvalues of the matrix of these
coecients, and the upper bound of absolute values of all the coecients of the operators L

(u), 0 1.
Rewrite equations (8.2) in the form
L

(u) L
0
(u) +(L L
0
)u = f (8.3)
and then apply to both sides of (8.3) the operator L
1
0
. We obtain
u +L
1
0
(L L
0
)u = L
1
0
f. (8.4)
Denote by A the operator L
1
0
(L L
0
) mapping the functions from

W
2,1
(Q) into functions of the same
space. The operator A is bounded, since |(L L
0
)u|
0
M
1
|u|
2,1
and |L
1
0
v|
2,1
M
2
|v|
0
. Equation (8.4) has
the form
u +Au = L
1
0
f, (8.5)
517
where |A| < 1 for suciently small , 0 <
1
(M). Therefore, Eq. (8.5) for 0
1
can be solved in
the space

W
2,1
(Q) for any function f L
2
(Q) by method of successive approximations (see [83, p. 47]). We can
apply the operator L
0
to both sides of (8.5) and obtain as a result that the solution of Eq. (8.5) belongs to

W
2,1
(Q)
and satises Eq. (8.3) for 0
1
almost everywhere in Q.
The fact that the solution u(x, t) of Eq. (8.3) for =
1
belongs to

W
2,1
(Q) means that there exists
a sequence of functions u
m
(x, t) C

(

Q) such that u
m
[

= 0 and u
m
u in the norm of W
2,1
(Q) for m .
Therefore, for all solutions u(x, t) of Eq. (8.3) for =
1
belonging to the class

W
2,1
(Q) the inequality (2.35) holds.
Hence it follows that there exists an operator L
1
1
and its norm is bounded by a constant depending only on M.
Rewrite Eq. (8.2) in the form
L

(u) L
1
(u) + (
1
)(L L
0
)u = f (8.6)
and carry out for (8.6) the same calculations we did for (8.3). The part of the operator L
0
in Eq. (8.6) will play
the operator L
1
. Thus we can show that Eq. (8.2) with condition (7.1) has a solution belonging to

W
2,1
(Q) for
0
1

1
, that is, for 2
1
, because the norm of the operator L
1
1
(L L
0
) in

W
2,1
(Q) is bounded
uniformly with respect to . Repeating this reasoning, after a nite number of steps equal to
_
1
1

+1 we establish
the solvability of Eq. (8.3) with condition (7.1) in the space

W
2,1
(Q) for = 1, and this was to be proved.
Note that we have obtained the solution of the problem (5.12), (7.1) under weak suppositions about the co-
ecients of the equation L(u) = f. Namely, we supposed only that the coecients a
ij
are continuous and b
i
and c
are bounded in

Q. A similar existence theorem can be proved for any domain D such that for this domain the esti-
mate (2.35) holds and there exists a smooth solution of the rst boundary-value problem for the heat equation (8.1).
9 APPLICATION OF GALERKINS METHOD
TO THE CONSTRUCTION OF A SOLUTION OF THE FIRST BOUNDARY-VALUE METHOD
In this section, we construct a weak (see Sect. 5) generalized solution of the rst boundary-value problem
in the cylinder Q = (0, T) for Eq. (5.7) with conditions (5.8). We shall suppose that the coecients A
ij
are
of class C
1
(

Q); B
i
, C, and F are continuous in

Q; the initial function (x)
0
W
1
(),
n

i,j=1
A
ij
(x, t)
i

j

n

i=1

2
i
( > 0), and the boundary of the domain belongs to the class A
2
. We shall use Galerkins method in the form
used by E. Hopf to construct a solution of the boundary-value problem for the NavierStokes system [104]. Later
this method was widely used for the construction of solutions of boundary-value problems for linear equations of
various types, and also for linear elliptic, parabolic, and hyperbolic systems (see [99, 105]).
According to Lemma 2 of Sect. 5, in the domain there exists a system of functions X
1
(x), X
2
(x),. . . ,
X
k
(x),. . . , with properties (1), (2), and (3), of Lemma 2. We shall seek the approximate solution u
N
(x, t) of
the problem (5.7), (5.8) as a sum
u
N
(x, t) =
N

k=1

N
k
(t)X
k
(x). (9.1)
The functions
N
k
(t) are dened by the conditions
_

[L(u
N
) f]X
k
(x) dx = 0 for 0 t T (k = 1, . . . , N). (9.2)
This means that
_

i,j=1
A
ij
u
N
x
j
X
k
x
i
+
_
n

i=1
B
i
u
N
x
i
+Cu
N

u
N
t
F
_
X
k
_
dx = 0 (k = 1, . . . , N). (9.3)
Taking into account the fact that the system X
k
(x) is orthogonal in L
2
(), we can obtain from equality (9.3)
a system of ordinary linear rst-order dierential equations with continuous coecients:
d
N
k
(t)
dt
=
N

s=1

A
ks
(t)
N
s
(t) +

B
k
(t) (k = 1, . . . , N); (9.4)
this system allows us to dene the functions
N
k
(t).
518
The initial function (x) belongs to
0
W
1
(). Therefore, according to Lemma 2 of Sect. 5 there exists
a sequence of functions
N
(x) =
N

k=1
c
N
k
X
k
(x), converging in the norm of W
1
() to the function (x). Dene

N
k
(0) from the condition
u
N
(x, 0)
N

k=1

N
k
(0)X
k
(x) =
N
(x)
N

k=1
c
N
k
X
k
(x). (9.5)
This means that

N
k
(0) = c
N
k
(k = 1, . . . , N). (9.6)
The solution of the system (9.4) with conditions (9.6) is constructed on the interval 0 t T.
We shall show that the function u(x, t) = lim
N
u
N
(x, t) is a generalized solution of the problem (5.7), (5.8)
in the sense of Denition 1 of Sect. 5.
First we show that u
N
(x, t) are bounded in the norm of W
1,1
(Q) uniformly in N. To do this we multiply
Eqs. (9.3) by
N
k
(t)e
t
and take the sum of the equalities thus obtained with respect to k from 1 to N. Then we
integrate with respect to t from 0 to T. We arrive at the following relation:
_
Q
e
t
_

i,j=1
A
ij
u
N
x
i
u
N
x
j
+
n

i=1
B
i
u
N
u
N
x
i
+ (Cu
N
F)u
N
u
N
u
N
t
_
dxdt = 0. (9.7)
We can apply the integration by parts to the last term in the left-hand side of (9.7). In the same way as we
did when we deduced the estimate (2.29) we shall obtain that for suciently large > 0 the inequality
1
2
__
Q
e
t
u
2
N
dxdt +
__
Q
e
t
n

i,j=1
A
ij
u
N
x
i
u
N
x
j
dxdt M
1
___
Q
F
2
dxdt +
_

2
N
dx
_
holds. Here and below, M
i
will be constants not depending on N. It follows from the last inequality that
__
Q
u
2
N
dxdt M
2
and
__
Q
n

i=1
_
u
N
x
i
_
2
dxdt M
2
.
Now multiply Eqs. (9.3) by
d
N
k
(t)
dt
, take the sum with respect to k from 1 to N, and integrate with respect
to t from 0 to T. We get
__
Q
_

i,j=1
A
ij
u
N
x
j

2
u
N
x
i
t
+
n

i=1
B
i
u
N
x
i
u
N
t
+ (Cu
N
F)
u
N
t

_
u
N
t
_
2
_
dxdt = 0.
Transform the rst sum under the integral sign using integration by parts, as we did when we deduced (2.30). We
obtain the inequality
_
t=T
n

i=1
_
u
N
x
i
_
2
dx +
__
Q
_
u
N
t
_
2
dxdt M
3
___
Q
F
2
dxdt +
_

2
N
+
n

i=1
_

N
x
i
_
2
_
dx
_
, (9.8)
and it follows that
__
Q
_
u
N
t
_
2
dxdt M
4
.
When we proved the inequality we used the fact that
_
t=0
n

i,j=1
A
ij
uN
xi
uN
xj
M
5
_

i=1
_
N
xi
_
2
dx and the last integral
is bounded uniformly in N because of the convergence of
N
xi
to

xi
in the norm of L
2
().
519
Thus we have |u
N
|
1,1
M
6
. According to Theorem 1 of Sect. 5, it is possible to select from the sequence
u
N
(x, t) a subsequence u
N
k
(x, t) converging in the mean to the function u(x, t) W
1,1
(Q). The derivatives
uN
k
xi
and
uN
k
t
will converge weakly to the corresponding generalized derivatives
u
xi
and
u
t
as N
k
. Conditions (5.8)
are satised for u(x, t); we have to take into account equalities (9.5), the condition u
N
[
S
= 0, and Theorem 1 of
Sect. 5.
Now we show that u(x, t) satises the integral identity (5.9).
Note that it is sucient to prove (5.9) for the functions (x, t) innitely dierentiable in

Q and vanishing
in a neighborhood of the boundary of Q. In fact, we can insert into (5.9) a function (x, t) =
m
(t)(x, t), where
(x, t)
0
C

(Q) and
m
(t) is an innitely dierentiable function with the following properties: 0
m
(t) 1;

m
(t) = 0 for 0 t
1
m
and for T
1
m
t T;
m
(t) = 1 for
2
m
t T
2
m
(m = 1, 2, . . .). Passing to
the limit in (5.9) as m , we see that the integral identity (5.9) is satised for any function (x, t)
0
C

(Q),
and therefore for any function (x, t)
0
W
1,1
(Q).
According to Lemma 2 of Sect. 5, there exists a sequence
m
(x, t) =
n

k=1
c
m
k
(t)X
k
(x) such that the
m
converge to in the mean together with their derivatives in x
1
, . . . , x
n
. First we shall prove that the integral
identity (5.9) is satised for the functions
m
(x, t). Multiply the equality (9.3) by c
m
k
(t), take the sum with respect
to k from 1 to m, and integrate with respect to t from 0 to T. We obtain
__
Q
_

i,j=1
A
ij
u
N
x
j

m
x
i
+
_
n

i=1
B
i
u
N
x
i
+Cu
N
F
u
N
t
_

m
_
dxdt = 0.
In this relation, we can pass to the limit with xed m as N tends to innity along the subsequence N
k
chosen above.
We get
__
Q
_

i,j=1
A
ij
u
x
j

m
x
i
+
_
n

i=1
B
i
u
x
i
+Cu F
u
t
_

m
_
dxdt = 0.
In this equality, we can pass to the limit as m and obtain the integral identity (5.9). Thus we have proved that
u(x, t) is a weak generalized solution of problem (5.7), (5.8). The uniqueness of this solution is proved in Sect. 5.
Because we have constructed a unique solution u(x, t), the whole sequence u
N
(x, t) converges to it. Therefore
the construction used to prove the existence of a solution for problem (5.7), (5.8) also gives an easy way to obtain
an approximate solution of this problem. Various modications of this method are often used in practice.
Note that without major changes it is possible to prove the convergence of the functions u
N
(x, t) to the so-
lution u(x, t) under weaker suppositions on the smoothness of the coecients of the equation and the boundary of
the domain.
10 GENERALIZED SOLUTION FOR THE CAUCHY PROBLEM
We shall consider a Cauchy problem in the domain H0 < t T
1
, x E
n
for Eq. (5.7) with the initial
condition
u(x, 0) = (x), x E
n
. (10.1)
The functions A
ij
(x, t),
Aij (x,t)
t
, B
i
(x, t), and C(x, t) are assumed to be continuous and bounded in

H. Let
the absolute values of these functions not exceed M and let the inequality
n

i,j=1
A
ij

j

n

i=1

2
i
, > 0 hold.
In this section, we shall dene a generalized solution of the Cauchy problem (5.7), (10.1) and prove the ex-
istence and uniqueness of such a solution in a class of functions that can grow in a certain manner as [x[
_
n

i=1
x
2
i
_1
2
. We shall suppose that in any bounded domain (x) W
1
, F(x, t) L
2
and for any > 0
the integrals
_
En
e
|x|
2
_

2
(x) +
n

k=1
_
(x)
x
k
_
2
_
dx,
__
H
e
|x|
2
F
2
(x, t) dxdt (10.2)
converge.
520
The function v(x, t) is said to belong to V

if in any bounded domain v(x, t) W


1,1
and the integral
_
En
e
|x|
2
v
2
(x, t) dx
converges uniformly in t for 0 t T
1
and, moreover, that
__
H
e
|x|
2
_
v
2
+
n

k=1
_
v
x
k
_
2
+
_
v
t
_
2
_
dxdt <
(for some > 0). If v(x, t) V

for all > 0, then we shall say that v(x, t)



V .
Denition. The function u(x, t) will be called a generalized solution of the problem (5.7), (10.1) if u(x, t)
belongs to the class V

for some > 0, is equal to (x) for t = 0, and satises the equality
T
_
0
_
En
_
n

i,j=1
A
ij
u
x
j

x
i

_
n

i=1
B
i
u
x
i
+Cu
du
dt
F
_

_
dxdt = 0, (10.3)
where 0 < T T
1
, and (x, t) is an arbitrary function with compact support with respect to x, belonging to
the class W
1,1
.
We shall prove the existence of a generalized solution of the Cauchy problem in the class

V and the uniqueness
of the generalized solution in a wider class of functions. That is, in proving the uniqueness of the solution of Cauchy
problem we shall assume that a solution u(x, t) belongs to the class V
1
for some
1
> 0, where the numbers
1
in
general are dierent for dierent solutions.
To prove the uniqueness of the generalized solution it is sucient to prove that the generalized solution
u(x, t) of class V
1
for a certain
1
> 0, is identically zero if (x) 0 and F(x, t) 0.
1. Proof of the Uniqueness Theorem. Denote by
R
(x) a smooth function equal to unity for [x[ R
and equal to zero for [x[ R + 2, such that 0
R
(x) 1 and
n

k=1
_
R(x)
x
k
_
2
1. Insert into the identity (10.3)
instead of (x, t) the function
e
21|x|
2
|x|
2
t

R
(x)u(x, t),
where the constants > 0 and R > 0 will be chosen later. We shall obtain
T
_
0
_
En
e
21|x|
2
|x|
2
t
_
n

i,j=1
A
ij
u
x
j
_
(u
R
)
x
i
2(2
1
+t)x
i

R
u
_
+
R
u
u
t
C
R
u
2

R
u
n

i=1
B
i
u
x
i
_
dxdt = 0.
(10.4)
It follows from the equality (10.4) that
T
_
0
_
|x|R
e
21|x|
2
|x|
2
t
_
n

i,j=1
_
A
ij
u
x
i
u
x
j
2(2
1
+t)x
i
A
ij
u
u
x
j
_
+u
u
t
_
dxdt
M
1
T
_
0
_
En
e
21|x|
2
|x|
2
t
_
u
2
+[u[
n

i=1

u
x
i

_
dxdt
+M
1
T
_
0
_
R|x|R+2
e
21|x|
2
|x|
2
t
_
u
2
+
n

i=1
_
u
x
i
_
2
+
_
u
t
_
2
_
(1 +[x[) dxdt, (10.5)
where M
1
depends only on M and n.
521
Let us transform the integral I
1
=
T
_
0
_
|x|R
e
21|x|
2
|x|
2
t
u
u
t
dxdt. It is obvious that
I
1
=
1
2
T
_
0
_
|x|R
e
21|x|
2
|x|
2
t
(u
2
)
t
dxdt
=
1
2
_
|x|R
u
2
(x, T)e
21|x|
2
T|x|
2
dx +

2
T
_
0
_
|x|R
[x[
2
u
2
e
21|x|
2
|x|
2
t
dxdt,
because u(x, 0) = 0.
Now we estimate the integral
I
2
=
T
_
0
_
|x|R
e
21|x|
2
|x|
2
t
n

i,j=1
2(2
1
+t)x
i
A
ij
u
u
x
j
dxdt.
We have
[I
2
[ M
2
(T + 2
1
)
T
_
0
_
|x|R
e
21|x|
2
|x|
2
t
[x[
n

i=1

u
x
i

[u[ dxdt


2
T
_
0
_
|x|R
e
21|x|
2
|x|
2
t
n

i=1
_
u
x
i
_
2
dxdt + (T + 2
1
)
2
M
3
T
_
0
_
|x|R
[x[
2
u
2
e
21|x|
2
|x|
2
t
dxdt,
where the constants M
2
and M
3
depend only on , M, and n.
These transformations bring the inequality (10.5) to the following form:
T
_
0
_
|x|R
e
21|x|
2
|x|
2
t
_
2
n

i,j=1
A
ij
u
x
i
u
x
j
+[x[
2
u
2
_
dxdt +
_
|x|R
e
21|x|
2
T|x|
2
u
2
(x, T)dx
2M
1
T
_
0
_
En
e
21|x|
2
|x|
2
t
_
u
2
+[u[
n

i=1

u
x
i

_
dxdt +
T
_
0
_
|x|R
e
21|x|
2
|x|
2
t
n

i=1
_
u
x
i
_
2
dxdt
+ 2(T + 2
1
)
2
M
3
T
_
0
_
|x|R
[x[
2
u
2
e
21|x|
2
|x|
2
t
dxdt
+ 2M
1
T
_
0
_
R|x|R+2
e
21|x|
2
|x|
2
t
(1 +[x[)
_
u
2
+
n

i=1
_
u
x
i
_
2
+
_
u
t
_
2
_
dxdt. (10.6)
The last of the integrals tends to zero as R , because e
21|x|
2
(1 + [x[) < e
1|x|
2
and u belongs to
the class V
1
. Taking into account that
n

i,j=1
A
ij
u
xi
u
xj

n

i=1
_
u
xi
_
2
and passing to the limit in (10.6) as R ,
we get
522
T
_
0
_
En
e
21|x|
2
|x|
2
t
_
2
n

i=1
_
u
x
i
_
2
+[x[
2
u
2
_
dxdt +
_
En
e
21|x|
2
T|x|
2
u
2
(x, T) dx
2M
1
T
_
0
_
En
e
21|x|
2
|x|
2
t
_
u
2
+[u[
n

i=1

u
x
i

_
dxdt
+
T
_
0
_
En
e
21|x|
2
|x|
2
t
n

i=1
_
u
x
i
_
2
dxdt + 2(T + 2
1
)
2
M
3
T
_
0
_
En
[x[
2
u
2
e
21|x|
2
|x|
2
t
dxdt
M
4
T
_
0
_
En
e
21|x|
2
|x|
2
t
u
2
dxdt + 2
T
_
0
_
En
e
21|x|
2
|x|
2
t
n

i=1
_
u
x
i
_
2
dxdt
+ 2(T + 2
1
)
2
M
3
T
_
0
_
En
e
21|x|
2
|x|
2
t
[x[
2
u
2
dxdt, (10.7)
where M
4
depends only on , M, and n.
Choose > 0 so large and T
0
> 0 so small as to have the inequality 2M
3
(T + 2
1
)
2
< for all T T
0
.
Then it follows from (10.7) that
_
En
e
21|x|
2
T|x|
2
u
2
(x, T) dx M
4
T
_
0
_
En
e
21|x|
2
t|x|
2
u
2
(x, t) dxdt for 0 < T T
0
. (10.8)
Let us integrate the inequality (10.8) with respect to T from 0 to T
0
:

_
0
_
En
e
21|x|
2
T|x|
2
u
2
(x, T) dxdT
M
4

_
0
_
T
_
0
_
En
e
21|x|
2
t|x|
2
u
2
(x, t) dxdt
_
dT M
4

_
0
_
En
e
21|x|
2
t|x|
2
u
2
(x, t) dxdt.
If we choose =
1
2M4
, then it follows from the last inequality that u(x, t) 0 for 0 t . Note that
depends only on , M, and n. Apply the same reasonings to the domain t 2, then to the domain 2 t 3,
and so on; after a nite number of steps we shall obtain u(x, t) 0 in the whole layer 0 t T. The uniqueness
of the generalized solution of the Cauchy problem (5.7), (10.1) is proved.
A similar uniqueness theorem is proved in [106].
2. Construction of the Generalized Solution. Now we shall show that under the conditions (10.2) on
the initial function (x) and the right-hand side F(x, t), there exists a generalized solution of the Cauchy problem
(5.7), (10.1) belonging to the class

V .
Consider the cylinders Q
N
having as bases balls of radius N with center at the origin; the height of the cylin-
ders is T
1
. Denote by u
N
(x, t) the generalized solution of the rst boundary-value problem for Eq. (5.7) in the cylin-
der Q
N
with the initial condition u
N
(x, 0) = (x)
N2
(x) and the boundary condition u
N
[
S
= 0. In Sect. 6, it is
proved that under the suppositions we have made on the coecients of (5.7) and the functions (x) and F(x, t)
such a solution u
N
(x, t)
0
W
1,1
(Q
N
) exists and by denition satises the relation
__
QN
_
n

i,j=1
A
ij
u
N
x
j

x
i

_
n

i=1
B
i
u
N
x
i
+Cu
N

u
N
t
F
_

_
dxdt = 0 (10.9)
for (x, t)
0
W
1,1
(Q
N
).
523
Let us show that for any > 0 the integrals
__
QN
e
|x|
2
_
n

i=1
_
u
N
x
i
_
2
+
_
u
N
t
_
2
_
dxdt,
_
|x|N
e
|x|
2
u
2
N
(x, T) dx
are bounded uniformly in N and in T T
1
. Insert into (10.9) the function (x, t) = e
|x|
2
(t+1)
u
N
(x, t). Repeating
the calculations we made above when we proved the uniqueness of the solution of the Cauchy problem (having now
= 2
1
= ), we obtain that for
0
and all T, 0 < T T
1
, the following inequality holds:
T
_
0
_
|x|N
e
|x|
2
(t+1)
n

i=1
_
u
N

x
i
_
2
dxdt +
_
|x|N
e
|x|
2
(T+1)
u
2
N
(x, T) dx
M
5
_
T
_
0
_
|x|N
e
|x|
2
(t+1)
(u
2
N
+F
2
) dxdt +
_
|x|N
e
|x|
2
u
2
N
(x, 0) dx
_
,
where
0
is suciently small (the constant
0
depends only on , M, and n). Hence, because of the suppositions
we made on the functions (x) and F(x, t) we can deduce that the integrals
T
_
0
_
|x|N
e
|x|
2
n

i=1
_
u
N
x
i
_
2
dxdt,
_
|x|N
e
|x|
2
u
2
N
(x, T) dx (10.10)
are bounded uniformly in N and T for
0
, and therefore for all positive .
The proof that the integral containing the derivative
uN
t
is also uniformly bounded is a little more
complicated. To prove this we insert the function (x, t) = e
|x|
2
(t+1)
u
N
(x, t) into the equality (10.9); here
v

(x, t)
v(x,t)v(x,t)

( is some xed positive number). We can suppose that the function u


N
(x, t) is extended
to the domain t < 0: u
N
(x, t) u
N
(x, 0) if t < 0. The equality (10.9) takes the following form:
__
QN
_
n

i,j=1
A
ij
u
N
x
j
(u
N
e
|x|
2
(t+1)
)
x
i
+e
|x|
2
(t+1)
u
N
_
u
N
t
Cu
N

n

i=1
B
i
u
N
x
i
+F
__
dxdt = 0. (10.11)
For simplicity we omit the index N in u
N
and we denote by K
i
quantities bounded in absolute value by
M
6
_
1 +
_ __
QN
e
|x|
2
(t+1)
u
2

dxdt
_
1/2
_
,
where the constant M
6
does not depend on N and . Doing this we take into account the uniform boundedness of
integrals (10.10) for all > 0, and, therefore, the uniform in N and T boundedness of integrals
__
QN
[x[
m
e
|x|
2
n

i=1
_
u
N
x
i
_
2
dxdt,
_
|x|N
[x[
m
e
|x|
2
u
2
N
(x, T)dx
(m > 0 is an arbitrary number). It follows from (10.11) that
__
QN
n

i,j=1
A
ij
u
x
j
u

x
i
e
|x|
2
(t+1)
dxdt +
__
QN
u

u
t
e
|x|
2
(t+1)
dxdt = K
1
. (10.12)
Let us estimate the rst integral in the left-hand side of the equality (10.12). To do this we use the relation
(vw)

= v

w(x, t ) +vw

.
524
We obtain
__
QN
_
n

i,j=1
A
ij
u
x
i
u
x
j
e
|x|
2
(t+1+)
_

dxdt
=
__
QN
__
n

i,j=1
A
ij
u
x
i
u
x
j
_

e
|x|
2
(t+1)
[x[
2
n

i,j=1
A
ij
u
x
i
u
x
j
e
|x|
2
(t+1+ )
_
dxdt
=
__
QN
_
n

i,j=1
A
ij
u
x
i
u
x
j
_

e
|x|
2
(t+1)
dxdt +K
2
=
__
QN
n

i,j=1
A
ij
_
u
x
i
u
x
j
_

e
|x|
2
(t+1)
dxdt +K
3
; (10.13)
here
0 < < .
The expression under the last integral sign can be transformed in the following way:
n

i,j=1
A
ij
_
u
x
i
u
x
j
_

=
n

i,j=1
A
ij
_
u
x
i
u

x
j
+
u(x, t )
x
i
u

x
j
_
= 2
n

i,j=1
A
ij
u
x
i
u

x
j

i,j=1
A
ij
_
u
x
i

u(x, t )
x
i
_
u

x
j
= 2
n

i,j=1
A
ij
u
x
i
u

x
j

i,j=1
A
ij
u

x
i
u

x
j
.
Hence and from (10.13) we obtain that
__
QN
_
n

i,j=1
A
ij
u
x
i
u
x
j
e
|x|
2
(t+1+)
_

dxdt
= 2
__
QN
n

i,j=1
A
ij
u
x
i
u

x
j
e
|x|
2
(t+1)
dxdt
__
QN
n

i,j=1
A
ij
u

x
i
u

x
j
e
|x|
2
(t+1)
dxdt + K
3
.
Therefore
__
QN
n

i,j=1
A
ij
u
x
i
u

x
j
e
|x|
2
(t+1)
dxdt =
1
2
__
QN
_
n

i,j=1
A
ij
u
x
i
u
x
j
e
|x|
2
(t+1+)
_

dxdt
+

2
__
QN
n

i,j=1
A
ij
u

x
i
u

x
j
e
|x|
2
(t+1)
dxdt +K
4

1
2
__
QN
_
n

i,j=1
A
ij
u
x
i
u
x
j
e
|x|
2
(t+1+)
_

dxdt +K
4
. (10.14)
It is obvious that
__
QN
v

dxdt =
1

T1
_
0
_
|x|N
[v(x, t) v(x, t )] dxdt
=
1

T1
_
0
_
|x|N
v(x, t) dxdt
1

T1
_

_
|x|N
v(x, ) dxd =
1

T1
_
T1
_
|x|N
v(x, t) dxdt
1

0
_

_
|x|N
v(x, t) dxdt.
Therefore the inequality (10.14) can be rewritten in the following form:
__
QN
n

i,j=1
A
ij
u
x
i
u

x
j
e
|x|
2
(t+1)
dxdt

1
2
T1
_
T1
_
|x|N
n

i,j=1
A
ij
u
x
i
u
x
j
e
|x|
2
(t+1+)
dxdt
1
2
_
|x|N
_
n

i,j=1
A
ij
u
x
i
u
x
j
_
t=0
e
|x|
2
(1+)
dx +K
4
K
5
.
525
From this inequality and from (10.12) it follows that
__
QN
u

u
t
e
|x|
2
(t+1)
dxdt M
7
_
1 +
_ __
QN
u
2

e
|x|
2
(t+1)
dxdt
_
1/2
_
.
Passing to the limit as 0 (M
7
does not depend on ), we obtain
__
QN
_
u
N
t
_
2
e
|x|
2
(t+1)
dxdt M
7
_
1 +
_ __
QN
_
u
N
t
_
2
e
|x|
2
(t+1)
dxdt
_
1/2
_
.
From the latter inequality it follows that the integral
__
QN
_
u
N
t
_
2
e
|x|
2
(t+1)
dxdt (10.15)
is uniformly bounded with respect to N.
Since in any bounded domain the functions u
N
(x, t) and their rst-order derivatives are uniformly bounded in
the L
2
-norm, we can select a subsequence u
N
k
(x, t) that in any bounded domain converges weakly to the function
u(x, t). The function u(x, t) in any bounded domain belongs to the class W
1,1
and for t = 0 satises the initial
condition (10.1). Since the integrals (10.10) and (10.15) are uniformly bounded, the integrals
T1
_
0
_
En
e
|x|
2
_
n

i=1
_
u
x
i
_
2
+
_
u
t
_
2
_
dxdt and
_
En
e
|x|
2
u
2
(x, T) dx (10.16)
converge for any >0. Moreover, since the integrals (10.15) and (10.16) are bounded, the integral
_
En
e
|x|
2
u
2
(x, t)dx
converges uniformly in t for 0 t T
1
. Therefore, u(x, t)

V .
Passing to the limit as N = N
k
in the equality (10.9), where the function (x, t) has compact support
with respect to x, we can see that the function u(x, t) satises the condition (10.3).
Thus we have proved that the function u(x, t) is a generalized solution of the Cauchy problem (5.7), (10.1).
From the estimates of the integrals (10.16) that we have established above it follows that there is a continuous
dependence of the solution on the right-hand side F(x, t) and on the initial function (x), that is, the integrals (10.16)
are small if the integrals
T1
_
0
_
En
F
2
(x, t)e
|x|
2
dxdt
and
_
En
e
|x|
2
_

2
(x) +
n

i=1
_
(x)
x
i
_
2
_
dx
are small.
Let us note that the suppositions on the growth of the initial function (x) and the right-hand side F(x, t)
that we made when we proved the existence of the generalized solution of the Cauchy problem (5.7), (10.1), are
fairly exact. In fact, we can consider as an example the Cauchy problem for the heat equation

2
u
x
2

u
t
= 0 (10.17)
with the initial condition
u(x, 0) = e
Ax
2
, A > 0 (10.18)
526
(see subsection 2 of Sect. 4). The function (10.18) does not fulll the condition of the convergence of the rst
integral of (10.2). The solution
u(x, t) = (1 4At)

1
2
e
Ax
2
14At
(10.19)
of the problem (10.17), (10.18) in the strip 0 < t T
1
<
1
4A
belongs to the class V

for some = (T
1
).
According to the uniqueness theorem, there is no other generalized solution of the problem (10.17), (10.18). But
the function (10.19) tends to innity as t
1
4A
. Therefore the problem (10.17), (10.18) has no generalized solution
dened for all t > 0. If we take as an initial function instead of (10.18) the function e
Ax
2
, where > 0, then
the integrals (10.2) turn out to be converging for all > 0, and therefore there exists a generalized solution of
the Cauchy problem for Eq. (10.17) with this initial function for all t > 0.
11 ON THE SMOOTHNESS OF GENERALIZED SOLUTIONS
In the previous sections we have proved, using various methods, the existence of the generalized solutions of
the rst boundary-value problem for Eqs. (5.7) or (5.12). The generalized solutions were understood as in Denitions
1 or 2 of Sect. 5.
In this section, we shall show that these generalized solutions have derivatives up to a certain order if
the right-hand side of the equation, its coecients, and the boundary of the domain are suciently smooth.
In Theorem 1 we show that a weak generalized solution of the problem (5.7), (5.8) under some additional
suppositions on the data belongs to
0
W
2,1
(Q) and satises the equation almost everywhere. We shall prove this
using nite dierences; the essence of the proof is as follows. Let us consider the dierence quotients
u
x
i
=
u(x
1
, . . . , x
n
, t) u(x
1
, . . . , x
i1
, x
i
h, x
i+1
, . . . , x
n
, t)
h
.
It is possible to establish that they satisfy an energy inequality similar to the inequality (2.29). Hence we
can see that the dierence quotients corresponding to the derivatives of the generalized solution with respect to
the space variables are bounded in L
2
(Q) uniformly in h. By Lemma 3, which is formulated below, we get that
the solution u(x, t) has generalized derivatives of the second order with respect to the space variables. One can
consider dierence quotients for the derivatives of the generalized solution and establish for them estimates that are
uniform in h. This allows us to prove the existence of higher-order derivatives. The estimates near S are obtained
in the same way as in Theorem 1. But these estimates are rather complicated technically. Therefore, in Theorem 2
we prove the smoothness of generalized solutions only inside the cylinder Q.
This theorem states that for suciently smooth coecients and the right-hand side F of Eq. (5.7)
u(x, t) W
2k,k
(Q

) in any interior subdomain Q

of the cylinder Q. For suciently large k by embedding theo-


rems [82] we obtain that u(x, t) has continuous derivatives of higher orders inside Q. The method described here
was used in [107] to prove the smoothness of generalized solutions for strongly elliptic systems of equations.
We have to note that the proof of Theorem 1 has as a principal goal the demonstration of the method. In
fact, under the supposition that the coecients in the higher derivatives in Eqs. (5.7) and (5.12) have continuous
derivatives, it is clear that the forms (5.7) and (5.12) are equivalent. It follows from Theorem 2 of Sect. 5 and
the existence theorem of Sect. 8 that the weak generalized solution belongs to
0
W
2,1
(Q).
First we prove three lemmas that establish the connection between the dierence quotients of a function and
its generalized partial derivatives.
Lemma 1. Let the function u(x, t) W
1,1
(Q) vanish in a neighborhood of S. Consider its dierence
quotients in the direction of some axis x
k
:
u
k
h
(x
1
, . . . , x
n
, t)
1
h
[u(x
1
, . . . , x
n
, t) u(x
1
, . . . , x
k1
, x
k
h, x
k+1
, . . . , x
n
, t)].
Then
|u
k
h
|
0

_
_
_
_
u
x
k
_
_
_
_
0
(11.1)
and
_
_
_
_
u
k
h

u
x
k
_
_
_
_
0
0 for h 0 (k = 1, . . . , n). (11.2)
527
Proof. First we establish the inequality (11.1) for an innitely dierentiable compact support function u(x)
of one variable. We have
u
h
(x) =
1
h
[u(x) u(x h)] =
1
h
x
_
xh
du
d
d.
Therefore,
[u
h
(x)]
2

1
h
2
_
x
_
xh

du
d

d
_
2

1
h
x
_
xh
_
du
d
_
2
d,
+
_

u
2
h
(x) dx
1
h
+
_

_
x
_
xh
_
du
d
_
2
d
_
dx =
1
h
+
_

d
+h
_

_
du
d
_
2
dx =
+
_

_
du
d
_
2
d.
We have proved inequality (11.1) for the case of one variable.
Now let u(x, t) be a continuously dierentiable function of (n+1) variables. For this function inequality (11.1)
can be written along any line parallel to the x
k
axis. Integrating by the remaining variables, we get the required
inequality (11.1). If u(x, t) W
1,1
(Q) vanishes in a neighborhood of S, we can consider its averaged functions (see
Sect. 5) u

(x, t). For these averaged functions (11.1) holds. We can pass to the limit in (11.1) as 0, and thus
establish the validity of the rst statement of the lemma.
Note that it follows from the estimate (11.1) that as h 0 the functions u
k
h
converge weakly to the derivative
u
x
k
(see [82]).
To prove the convergence of u
k
h
to
u
x
k
in the mean we shall rst estimate u
h
for the functions of one variable:
u
h
(x)
du(x)
dx
=
1
h
x
_
xh
du()
d
d
du(x)
dx
=
1
h
x
_
xh
_
du()
d

du(x)
dx
_
d.
Therefore,
+
_

_
u
h
(x)
du(x)
dx
_
2
dx
+
_

1
h
x
_
xh
_
du()
d

du(x)
dx
_
2
d dx
=
1
h
+
_

h
_
0
_
du(x)
dx

du(x )
dx
_
2
ddx =
1
h
h
_
0
d
+
_

_
du(x)
dx

du(x )
dx
_
2
dx.
Writing such an inequality for each line parallel to the x
k
axis and integrating with respect to the remaining
variables, we arrive at the following inequality:
__
Q
_
u
k
h
(x, t)
u(x, t)
x
k
_
2
dxdt
1
h
h
_
0
d
__
Q
_
u(x, t)
x
k

u(x , t)
x
k
_
2
dxdt. (11.3)
Since the estimate (11.3) is valid for any smooth functions with compact support in Q, it is also valid for functions
belonging to W
1,1
(Q) and vanishing in a neighborhood of S.
The right-hand side of (11.3) tends to zero as h 0, because
u
x
k
belongs to L
2
(Q) and the theorem on
the continuity in L
2
is valid ([82, p. 16]). Relation (11.2) is proved.
Similar assertions are true also for the dierences with respect to t. It is clear that both assertions remain
true if u
k
h
is changed to u
k
h

1
h
[u(x
1
, . . . , x
k1
, x
k
+h, x
k+1
, . . . , x
n
, t) u(x
1
, . . . , x
n
, t)], that is,
|u
k
h
|
0

_
_
_
_
u
x
k
_
_
_
_
0
and
_
_
_
_
u
k
h

u
x
k
_
_
_
_
0
0 for h 0. (11.4)
528
Lemma 2. Let u(x, t) W
1,1
(Q) vanish in a neighborhood of S S
1
, where S
1
is a plane piece of S lying
in the plane x
n
= 0. Then relations (11.1) and (11.2) of the previous lemma hold for u
k
h
if k ,= n.
Proof. For simplicity, let us suppose that the cylinder Q lies in the half-space x
n
> 0. Consider Q

that
is a part of the cylinder Q such that x
n
. Denote by u

(x, t) averaged functions for u(x, t). If the averaging


radius is suciently small, the averaged functions are dened in Q

. For the functions u

(x, t) in the domain Q

the inequality (11.1) holds if k ,= n. The functions u

(x, t) and their derivatives converge in the mean in Q

to
the function u(x, t) and its derivatives. Passing to the limit as 0 in the equality (11.1), written for u

and Q

(for k ,= n), and then making tend to zero, we get the estimate (11.1) for the function u(x, t). The relation (11.2)
can be proved in a similar way.
Lemma 3. Let the function u(x, t) belong to L
2
(Q) and either (1) vanish in a neighborhood of S or (2) vanish
in a neighborhood of S S
1
, where S
1
is a plane piece of the boundary S lying in the plane x
n
= 0. Consider
the dierences u
k
h
, where k = 1, . . . , n in the rst case and k ,= n in the second one. Suppose that |u
k
h
(x, t)|
0
M,
where the constant M does not depend on h.
Then there exist the corresponding derivatives
u(x,t)
x
k
L
2
(Q) and their norms are bounded by a constant M.
Proof. Since the functions u
k
h
are bounded in L
2
(Q), it is possible to select out of these sequences a weakly
converging subsequence. Let the subsequence u
k
hm
weakly converge to v(x, t) as m . Then |v(x, t)|
0
M. It
remains only to prove that v(x, t) =
u(x,t)
x
k
. To do this let us consider the function (x, t)
0
C

(Q). It is easy to
verify that for suciently small h
m
(u
k
hm
, ) = (u,
k
hm
). (11.5)
By Lemma 1, the
h
converge in L
2
(Q) as h 0 to the function

x
k
(see (11.4)). Passing to the limit as m
in (11.5), we get
(v, ) =
_
u,

x
k
_
.
Therefore, v =
u
x
k
, and the lemma is proved.
Theorem 1. Suppose u(x, t) is a weak generalized solution of the rst boundary-value problem for Eq. (5.7)
in the cylinder Q with conditions
u(x, 0) = 0, u[
S
= 0. (11.6)
Suppose the coecients A
ij
(x, t) C
1
(

Q), B
i
(x, t), and C(x, t) are bounded, F(x, t) L
2
(Q), and S A
2
. Then
the solution u(x, t) W
2,1
(Q) and satises Eq. (5.7) almost everywhere in Q.
Proof. By denition, the function u(x, t)
__
Q
_
n

i,j=1
A
ij

x
i
u
x
j

_
n

i=1
B
i
u
x
i
+Cu
u
t
_

_
dxdt =
__
Q
Fdxdt (11.7)
for any (x, t)
0
W
1,1
(Q).
We shall use the following notations:
(f, g) =
__
Q
f(x, t)g(x, t) dxdt,
Af, g =
__
Q
n

i,j=1
A
ij
(x, t)
f
x
i
g
x
j
dxdt.
In what follows, the expression Ru will denote a linear dierential operator of order not higher than one with
bounded coecients. The equality (11.7) can be rewritten in the following way:
Au, + (Ru, ) +
_
u
t
,
_
= (F, ). (11.8)
529
Let us consider the dierences
f
h
=
1
h
[f(x
1
, . . . , x
n
, t) f(x
1
, . . . , x
k1
, x
k
h, x
k+1
, . . . , x
n
, t)],
where x
k
is one of the space variables. Let (x) be a smooth function in , vanishing in a -neighborhood of
the boundary of the domain and equal to zero outside a 2-neighborhood of . Then for suciently small h
the function ()
h

0
W
1,1
(Q) if W
1,1
(Q). Therefore, one can insert = ()
h
into the equality (11.8) and
get
Au, ()
h
+ (Ru, ()
h
) +
_
u
t
, ()
h
_
= (F, ()
h
). (11.9)
Let us denote by M
i
constants not depending on h and the functions (x, t). Since F(x, t) L
2
(Q) and
u(x, t)
0
W
1,1
(Q), we have
[(F, ()
h
)[ +[(Ru, ()
h
)[ M
1
|()
h
|
0
.
By Lemma 1, |()
h
|
0
||
1
M
2
||
1
. Therefore
[(F, ()
h
)[ +[(Ru, ()
h
)[ M
3
||
1
. (11.10)
Taking into account the equality (f, g
h
) = (f
h
, g), which is valid for f and g from L
2
(Q) if one of these
functions vanishes in an h-neighborhood of S, we get
_
u
t
, ()
h
_
=
__
u
t
_
h
,
_
=
_

u
h
t
,
_
. (11.11)
Let us denote by N
i
quantities not exceeding in absolute value M
i
||
1
. Then by (11.9), (11.10), and (11.11)
we have

u
h
t
,
_
+Au, ()
h
= N
1
. (11.12)
Since (u)
h
= u
h
+u(x
k
+h)
h
, we have
_

u
h
t
,
_
=
_
(u)
h
t
,
_

h
u(x
k
+h)
t
,
_
=
_
(u)
h
t
,
_
+N
2
. (11.13)
In a similar way, we get
Au, ()
h
=
n

i,j=1
_
A
ij
u
x
i
,
[
h
+
h
(x
k
h)]
x
j
_
=
n

i,j=1
_
A
ij
u
x
i
,

n
x
j
_
+
n

i,j=1
_
A
ij
u
x
i
,
h

x
j
+

h
x
j
(x
k
h) +
h
(x
k
h)
x
j
_
. (11.14)
Since
_
_
_
_

x
j
+

h
x
j
(x
k
h) +
h
(x
k
h)
x
j
_
_
_
_
0
M
4
|
h
|
0
+M
5
||
0
+M
6
||
1
,
it follows from (11.14) that
Au, ()
h
=
n

i,j=1
_
A
ij

u
x
i
,

h
x
j
_
+N
3
.
Let us continue the transformation of the right-hand side of this equality:
Au, ()
h
=
n

i,j=1
_
A
ij

u
x
i
,
_

x
j
_
h
_
+N
3
=
n

i,j=1
__
A
ij
(u)
x
i
A
ij
u

x
i
_
h
,

x
j
_
+N
3
=
n

i,j=1
_
A
ij
(u)
h
x
i
,

x
j
_
+N
4
.
530
Finally we arrive at the relation
Au, ()
h
= A(u)
h
, +N
4
. (11.15)
Combining equalities (11.12), (11.13), and (11.15), we get
_
(u)
h
t
,
_
+A(u)
h
, = N
5
.
Therefore

_
(u)
h
t
,
_
+A(u)
h
,

M
7
||
1
.
Denote = (u)
h
. Then we have

_
(u)
h
t
, (u)
h
_
+A(u)
h
, (u)
h

M
7
|(u)
h
|. (11.16)
Since
_
f,
f
t
_
=
1
2
_

f
2
(x, T) dx 0 for functions f(x, t) vanishing for t = 0 and belonging to W
1,1
(Q) and
the quadratic form Ag, g |g|
2
1
( > 0) for the functions g(x, t)
0
W
1,1
(Q), it follows from (11.16) that
|(u)
h
|
2
1
M
7
|(u)
h
|
1
.
Since (x) 1 in the cylinder Q
2
and all the points of this cylinder are at a distance not less than 2 from S, we
have |u
h
|
1
M
8
in Q
2
. This means that the dierence quotients of the derivatives of u(x, t) with respect to
the space variables are bounded in the mean uniformly in h:
_
_
_
_
_
u
x
i
_
h
_
_
_
_
0
M
8
in Q
2
(i = 1, . . . , n). (11.17)
Therefore, by Lemma 3 applied to
u
xi
, there exist second-order generalized derivatives

x
k
_
u
xi
_
, square integrable
in Q
2
. Since the numbers k and > 0 can be chosen arbitrarily, it follows that in

Q S there exist second-order
generalized derivatives of the function u(x, t) with respect to the space variables.
Therefore, the relation (11.7), written for (x, t)
0
C

(Q), can be integrated by parts. The result will be


(L(u) F, ) = 0, and it follows that almost everywhere inside Q the function u(x, t) satises Eq. (5.7). We have
shown that

2
u
xi x
k
L
2
(Q
2
) for any > 0 and for i, k = 1, . . . , n. To complete the proof it is sucient to prove
that all these derivatives are integrable in the neighborhood of the boundary S.
Let us consider a subdomain of the base of the cylinder Q, adjacent to a part of the boundary , that can
be transformed to a piece of the plane x
n
= 0 by a change of the independent variables. Denote this subdomain
by

. Then the domain

(0, T) is also transformed to a cylinder with a plane piece of boundary S, and this plane
piece has the equation x
n
= 0. According to the suppositions made about the smoothness of S, the coecients of
the transformed equation will satisfy the same smoothness conditions as the coecients of the original equation (5.7).
In the neighborhood of the plane piece

S of the boundary S we shall estimate the dierence quotients of
u
xi
in the same way as was done for strictly interior domains Q

. To do this we shall consider dierence quotients only


along directions tangent to

S, that is, for k ,= n, and the cutting-o smooth function (x) we shall choose so that it
vanishes outside

(0, T) and is equal to unity in a neighborhood of some piece of S lying strictly inside

S. Then
in this neighborhood the estimates (11.17) hold, and by Lemma 2 all the second-order derivatives with respect to
the space variables, besides

2
u
x
2
n
, are square integrable in this domain.
By Eq. (5.7), which is satised almost everywhere, the derivative

2
u
x
2
n
also is square integrable. Since
the neighborhood of the boundary can be divided into a nite number of parts similar to the one considered above,
we have u(x, t) W
2,1
(Q). Theorem 1 is proved.
Theorem 2. Suppose that u(x, t) is a weak generalized solution of the rst boundary-value problem (5.7),
(11.6) in the cylinder Q. Suppose the coecients A
ij
(x, t), B
i
(x, t), and C(x, t) belong to the class C
2k2,k1
(

Q),
F(x, t) W
2k2,k1
(Q) and the coecients A
ij
(x, t) have in

Q bounded derivatives up to order 2k 1 with respect
to the space variables. Then u(x, t) W
2k,k
in the cylinder Q

, contained together with its boundary in



Q .
531
Proof. Let us denote by D
q
the operator of dierentiation of order q with respect to the space variables.
First we prove that D
2k
u and D
2k2 u
t
belong to L
2
in any cylinder Q

. The proof is conducted by induction.


Suppose that D
q
u L
2
in any cylinder Q

and D
q2 u
t
L
2
. (For q = 2 such as inclusion is proved in Theorem 1.)
Let us prove that in any Q

also D
q+1
u and D
q1 u
t
are square integrable if q + 1 2k.
Let us dierentiate Eq. (5.7) (q2) times with respect to the space variables (this is possible by the inductive
hypothesis). We have
n

i,j=1

x
i
_
A
ij

x
j
D
q2
u
_


t
D
q2
u
=
n

i,j=1
_

x
i
_
A
ij

x
j
D
q2
u
_
D
q2

x
i
_
A
ij
u
x
j
__
D
q2
_
n

i=1
B
i
u
x
i
_
D
q2
(Cu) +D
q2
F. (11.18)
Denote the right-hand side of (11.18) by F
1
. It contains the derivatives of u with respect to the space variables up
to order q 1, derivatives of A
ij
in x
1
, . . . , x
n
up to order q 1, and derivatives of B
i
, C, and F up to order q 2.
Therefore, by the inductive hypothesis and the suppositions on the smoothness of the coecients of Eq. (5.7), we
have DF
1
L
2
in any cylinder Q

. Let be an arbitrary suciently smooth function with compact support in Q.


Then, multiplying (11.18) by
h
, integrating over Q, and then integrating by parts, we obtain
AD
q2
u,
h
+
_

t
D
q2
u,
h
_
= (F
1
,
h
).
(The notations are the same as in the previous theorem.)
Since (F
1
,
h
) = (F
1h
, ), we have

AD
q2
u,
h
+
_

t
D
q2
u,
h
_

M
1
||
0
; (11.19)
by M
i
here and below are denoted constants not depending on h and on the function . Obviously (11.19) is valid
also for any function with compact support in Q belonging to W
1,1
(Q).
Let (x, t) be a smooth function vanishing in a neighborhood of and equal to unity in some cylinder
Q



Q . Insert into the inequality (11.19) = (
2
D
q2
u
hh1
)
h1
. Let us denote by N
j
quantities not exceeding
in absolute value M
_
1 +

q
|D
q
u
h
|
0
_
, where

q
|D
q
u
h
|
0
is the sum taken over all the derivatives with respect
to the space variables of order q and the constant M depends only on the coecients of the equation and the norms
D
m
u and D
m2 u
t
in L
2
(Q) for m q.
Let us transform the left-hand side of the inequality (11.19):
_

t
D
q2
u,
h
_
=
_

t
D
q2
u
h
,
_
=
_

t
D
q2
u
hh1
,
2
D
q2
u
hh1
_
=
_

t
(D
q2
u
hh1
), D
q2
u
hh1
_

t
D
q2
u
hh1
, D
q2
u
hh1
_
=
T
_
0

t
__

[D
q2
u
hh1
]
2
dx
_
dt +N
1
N
1
,
(11.20)
since |D
q2
u
hh1
|
0
|D
q
u|
0
by Lemma 1.
Transform in a similar way the rst term of the inequality (11.19):
AD
q2
u,
h
=
n

i,j=1
_
_
A
ij

x
i
(D
q2
u)
_
hh1
,

x
j
[
2
D
q2
u
hh1
]
_
=
n

i,j=1
_
A
ij

x
i
D
q2
u
hh1
,

x
j
[
2
D
q2
u
hh1
]
_
+N
2
=
n

i,j=1
_
A
ij

x
i
[D
q2
u
hh1
],

x
j
[D
q2
u
hh1
]
_
+N
3
= AD
q2
u
hh1
, D
q2
u
hh1
+N
3
. (11.21)
532
From (11.19), (11.20), and (11.21), we can deduce that
AD
q2
u
hh1
, D
q2
u
hh1
M
1
|(
2
D
q2
u
hh1
)
h1
|
0
+N
4
.
Passing to the limit in this inequality as h
1
0 and using Lemma 1, we get
AD
q1
u
h
, D
q1
u
h
M
1
|D(
2
D
q1
u
h
)|
0
+N
4
= N
5
.
Since this inequality is valid for any derivative D
q1
u
h
and Af, f
n

i=1
_
f
xi
,
f
xi
_
, we have

q
(D
q
u
h
, D
q
u
h
) M
2
_
1 +

q
|D
q
u
h
|
0
_
. (11.22)
As above, the sum is taken over all the derivatives with respect to the space variables of order q.
It follows from the inequality (11.22) that |D
q
u
h
|
0
M
3
in the domain where 1. Therefore, by
Lemma 3 in this domain there exist derivatives D
q+1
u. Dierentiating Eq. (5.7) (q 1) times with respect to
the space variables and expressing

t
(D
q1
u) from the resulting equality, we can see that the derivative

t
(D
q1
u)
also is square integrable in Q.
The induction statement is proved. This means that D
2k
u L
2
(Q

) and

t
(D
2k2
u) L
2
(Q

) for any
interior subdomain Q

. Dierentiating Eq. (5.7), it is easy to see that other derivatives of the solution u(x, t) also
exist and belong to L
2
(Q

). It follows that u(x, t) W


2k,k
(Q

). The theorem is proved.


In some papers, another method is used to study the smoothness of generalized solutions; it is based on
averaged functions. The main idea of this method is as follows: instead of the function (x, t) the derivative of its
average is inserted into the integral identity (5.9) (and not a dierence quotient, as was done above). In this way,
one can get an estimate for the derivatives of the averages of the generalized solution. This estimate is uniform with
respect to the radius of averaging. In the clearest form this method was presented by K. Friedrichs [108], where
he proved the smoothness of generalized solutions of strongly elliptic systems at the interior points of the domain.
The use of this method to prove the smoothness up to the boundary meets with considerable diculties.
12 BEHAVIOR OF SOLUTIONS WITH INFINITE INCREASING OF TIME
In this section, we prove some simple theorems on the behavior of solutions of the Cauchy problem and
boundary-value problems with innite increasing of time. We suppose that the coecients and the right-hand side
of Eq. (1.1) are bounded in the domains considered and c(x, t) 0. The solutions are also assumed to be bounded.
Theorem 1. Suppose the function u(x, t) is a solution of the Cauchy problem for Eq. (1.1) in the half-space
t > 0 or of a boundary-value problem for Eq. (1.1) in the cylinder Q = (0, +) with one of the boundary
conditions
u[
S
= 0 (12.1)
or
l(u)
_
u

+au
_
S
= 0, (12.2)
where a 0 and is a direction in the space (x
1
, . . . , x
n
), making an acute angle with the inner normal to
the boundary of the domain . Suppose that c(x, t) < c
0
< 0, where c
0
is a constant, and f(x, t) 0 uniformly
in x as t +. Then u(x, t) 0 uniformly in x as t +.
Proof. To prove this theorem we rst note that the solution u(x, t) is bounded for all x and t 0, that is,
[u(x, t)[ M. This follows from Theorem 3 of Sect. 1 for the problem (1.1), (12.1) and from similar estimates for
the remaining problems (see Theorems 7 and 8 of Sect. 1). Let us x an arbitrary > 0 and choose T such that
[f(x, t)[ < c
0
for t T.
Consider the functions w

(x, t) = Me
c0(tT)
+ u(x, t). These functions are nonnegative for t = T. If
condition (12.1) is fullled on S, then w

(x, t)[
S
> 0 for t T. If we have (12.2) on S, then l(w

) 0 for t T.
For t T we have L(w

) = ML(e
c0(tT)
)+cf Me
c0(Tt)
(c
0
+c)c
0
+[f[ < 0. It follows from Theorems 1,
7 or 8 of Sect. 1 that w

0 for t T, that is,


[u(x, t)[ +Me
c0(Tt)
< 2
for t > T
1
(). The theorem is proved.
533
Theorem 2. Suppose the function u(x, t) satises Eq. (1.1) in the cylinder Q, u[
S
= 0, and f(x, t) 0
uniformly in x as t . Then u(x, t) 0 uniformly in x as t .
Proof. We shall make the following change of the unknown function in (1.1): u(x, t) = v(x, t)(x), where
the function (x) > 0 will be chosen later.
For the new unknown function v(x, t) we obtain an equation where the coecient in v(x, t) is L(). Therefore,
if we choose the function (x) so as to have L() < c
0
< 0 for all values of x considered, we shall be able to apply
Theorem 1 that we have proved above to the function v(x, t). According to Theorem 1, v(x, t) 0 as t . This
means that also u(x, t) 0 uniformly with respect to x as t .
For example, we can take (x) = 1 e
x1
. Without loss of generality, we can suppose that x
1
< 0 for
the points of the domain . Then we have in
L() = a
11

2
e
x1
b
1
e
x1
+c(1 e
x1
) < 0
for suciently large > 0.
Remark 1. Theorem 1 is valid also for a noncylindrical domain D, bounded by the plane t = 0 and a surface S,
and Theorem 2 is valid for a noncylindrical domain D under the additional supposition that the projection of S
to the plane t = 0 is situated in a bounded part of this plane. The condition u[
S
= 0 can be changed to u[
S
0
uniformly in x as t .
Remark 2. It follows easily from the proof of Theorems 1 and 2 that [u(x, t)[ < Me
t
, > 0, if f(x, t) 0
and u[
S
= 0.
Theorem 3. Suppose that the function u(x, t) satises Eq. (1.1) in the cylinder Q,
_
u

+ a(x, t)u
_
S
= 0,
where a(x, t) < a
0
< 0, and f(x, t) 0 uniformly in x as t . Then u(x, t) 0 uniformly in x as t .
Proof. As above, we shall make a change of the unknown function in Eq. (1.1), u = v(x), where (x) > 0.
The boundary condition will take the following form:
_
v

+
a+

v
_
S
= 0. Therefore, if we intend to reduce
the proof of the theorem to the use of Theorem 1, it is sucient to construct a smooth function (x) > 0 such that
L() < 0 and
_

+a
_
S
0.
Set (x) = (x) +A, where (x) is a function constructed in the previous theorem and A > 0. The operator
L() = L() +Ac < 0. The constant A > 0 will be chosen so large that

+a =

+a +Aa < 0.
Remark 1. The remarks to Theorem 2 are also valid for Theorem 3. In Theorems 2 and 3, one can change
the supposition that u(x, t) satises Eq. (1.1) and conditions (12.1) or (12.2) can be changed to the supposition
that the following inequalities are valid: L(u) 0 in Q and u[
S
0 or
_
u

+au
_
S
0. Then u(x, t) > if t is
suciently large, that is, lim
t
inf
x
u(x, t) 0.
Remark 2. In Theorem 3, the condition a < a
0
< 0 cannot be omitted, because under the condition that
u

S
= 0 the solution does not necessarily tend to zero (e.g., u 1 for the equation

2
u
x
2

u
t
= 0). However,
if under the condition a 0 there exists a smooth function (x) such that
_

+ a
_
S
0 and L() < 0,
the statement of Theorem 3 remains valid.
Such a function is easy to construct, for example, in the case of one dimension (n = 1) for the interval [0, X]
under the boundary conditions
_
u
x
a
1
(t)u
_
x=0
= 0,
_
u
x
+a
2
(t)u
_
x=X
= 0, if a
1
(t) 0 and a
2
(t) a
0
> 0. In
this case, we can take (x) = A coshx for suciently large A > 0 and > 0.
Theorem 4. Suppose that u(x, t) is a solution of the Cauchy problem for the equation L(u) = 0 with
a bounded initial function u(x, 0). Suppose there exists a positive function v(x) such that L(v) 0 and v(x)
as [x[ , where [x[ =
_
x
2
i
_
1/2
. Then the solution u(x, t) 0 as t uniformly with respect to x in any
bounded domain of the space E
n
.
534
Proof. Let us take any > 0 and consider an n-dimensional ball of radius R with center at the origin. To
prove the theorem it is sucient to show that [u(x, t)[ if [x[ R and t T
1
, where T
1
is a suciently large
number.
According to Theorem 10 of Sect. 1 (for M
2
= M
3
= 0) the solution u(x, t) is bounded: [u(x, t)[ M
1
.
Choose R
1
such that v(x) >
2M1

V
R
for [x[ = R
1
> R, where V
R
= max
|x|R
v. Then w

(x, t) =
v(x)
2VR
u(x, t) 0
for [x[ = R
1
. The operator L(w

) 0 in the cylinder [x[ R


1
, 0 < t < . By Remark 1 to the previous
theorem, w

(x, t)

2
for t T
1
and for all x such that [x[ R
1
. Therefore, [u(x, t)[

2
+
v(x)
2VR
for [x[ R
and t T
1
, and this is what was to be proved.
Remark. In Theorem 1 of Sect. 1, the supposition L(u) 0 can be weakened. Namely, it is sucient to
suppose that the function u(x, t) satises the inequality L(u) 0 everywhere in the domain D except for a nite
number of smooth surfaces S
k
(k = 1, . . . , N). The function u(x, t) has to be continuous in the closed domain D
and have derivatives along the normal on both sides of the surfaces S
k
. On the surfaces S
k
, the inequalities
u

+
<
u

(12.3)
must hold; here
u
+
means a derivative in the normal direction taken from the side of the surface S
k
where
the normal is directed and
u

is the derivative in the normal direction taken from the other side of the surface.
The proof of Theorem 1 of Sect. 1 under these suppositions remains the same; we only need to note that by virtue
of (12.3) the function u(x, t) cannot attain its minimal value on the surfaces S
k
.
In a similar way, it is possible to weaken the suppositions of Theorems 8 and 9 of Sect. 1. In Theorem 4
of this section it is also possible to admit as v(x) a continuous function that satises condition (12.3) on a nite
number of smooth surfaces.
Theorem 5. Suppose that u(x, t) is a bounded solution of the Cauchy problem for the equation L(u) = 0.
Suppose that c(x, t) c
0
< 0 for all t 0 and x belonging to some domain in the space x
1
, . . . , x
n
. Suppose also
that outside some ball (that is, for [x[ R) there exists a positive function (r) such that (r) as r
and L() 0 (r = [x[). Then u(x, t) 0 as t uniformly in x in any bounded domain of the space x
1
, . . . , x
n
.
Proof. According to Theorem 4 and the remark to this theorem, it is sucient to construct a function v(r)
such that v(r) as r and L(v) 0 everywhere except for a nite number of surfaces where the rst
derivatives of v have discontinuities and the condition (12.3) is fullled:
v
+
<
v

.
Without loss of generality, we can assume that c(x, t) c
0
< 0 for [x[ , where < R. Set v(r) = e
r
2
for r ( > 0). We have
L(v) =
_
4
2
n

i,j=1
a
ij
x
i
x
j
+ 2
n

i=1
a
ii
+ 2
n

i=1
b
i
x
i
+c
_
v < 0
for r < and a suciently small > 0.
In the ring r R we set v(r) = A(1 e
r
2
) + B, where A, B, and are positive constants. Let us
calculate L(v) for < r < R:
L(v) = Ae
r
2
_
4
2
n

i,j=1
a
ij
x
i
x
j
+ 2
n

i=1
a
ii
+ 2
n

i=1
b
i
x
i
_
+cv < 0
for suciently large and any A and B. We can choose the constant A > 0 so small that the condition (12.3)
is fullled for r = : 2Ae

2
< 2e

2
. Then we choose the constant B > 0 so that the function v(r) is
continuous for r = .
Finally, for r R set v(r) = A
1
(r) + B
1
, where A
1
> 0 and B
1
> 0 are again chosen so that the condi-
tion (12.3) is fullled for r = R and the function v(r) is continuous.
The theorem is proved.
535
It is easy to see that the divergence of the integral

_
R
re

r
_
r
0
b()

d
dr, where
b() = sup
|x|=
n

i=1
(a
ii
+b
i
x
i
)
n

i,j=1
a
ij
xixj

2
is a sucient condition for the existence of the function (r). In the case where the integral diverges we can take
as (r) the function
r
_
R
e

_
r
0
b()

d
d.
Theorem 6. Suppose that u(x, t) is a solution of the Cauchy problem for the equation L(u) = 0 and
the initial function u(x, 0) tends to zero as [x[ . If
n

i=1
(a
ii
+b
i
x
i
) > > 0 for all x and t 0, (12.4)
then u(x, t) 0 as t uniformly in x.
Proof. It is sucient to prove this for the initial function u(x, 0) vanishing outside some ball. The function
v(x, t) = (t +1)

e
r
2
t+1
(r = [x[) tends to zero as t , and one can choose > 0 and > 0 so that the inequality
L(v) < 0 is fullled. In fact,
L(v) = v
_
1
(t + 1)
2
_
4
2
n

i,j=1
a
ij
x
i
x
j
r
2
_
+
1
t + 1
_
2
n

i=1
(a
ii
+b
i
x
i
)
_
+c
_
. (12.5)
First we take so small that the quantity in the rst square brackets of the right-hand side of (12.5) becomes non-
positive; then, using (12.4), we choose so small that the quantity in the second square brackets becomes negative.
If M is suciently large, the functions Mv u are positive for t = 0 and satisfy the inequality L(Mv u) < 0
for t > 0. According to Theorem 8 of Sect. 1, the functions Mv u are nonnegative for all t 0. Therefore,
[u(x, t)[ Mv(x, t) and u(x, t) 0 uniformly in x as t .
Remark 1. Theorem 6 is valid also when the condition (12.4) is fullled only outside some cylinder, that is,
for [x[ R and t 0. One can nd the proof in [109].
Remark 2. The theorems we have proved above concerned the cases where u(x, t) tends to zero with the in-
creasing of t under some suppositions on the coecients of Eq. (1.1). It is clear that these theorems can be
extended to the case where the coecients of Eq. (1.1) and the right-hand side f(x, t) tend uniformly in x to
functions depending only on x as t . Suppose that the limit elliptic equation
n

i,j=1
a
ij
(x)

2
v
x
i
x
j
+
n

i=1

b
i
(x)
v
x
i
+ c(x)v =

f(x) (12.6)
has a solution v(x), dened for all x from the domain considered. Suppose this solution is bounded together with its
derivatives occurring in the equation. To investigate the problem on when the solution u(x, t) of Eq. (1.1) tends to
the corresponding solution v(x) of Eq. (12.6) as t we have to apply the theorem proved above to the dierence
u v. We shall not give here the statements of theorems thus obtained.
If the coecients of Eq. (1.1) do not depend on t, the behavior of solutions as t is studied in detail
in [110]. Some extensions of these results to equations with coecients depending on x and t are given in [111].
In [112], the rate of decrease is studied for solutions of a second-order parabolic equation in the domain
0 < x < X(t); this rate depends on the behavior of the function X(t) as t . It is supposed that the solution
u(x, t) is equal to zero for x = 0 and for x = X(t).
536
The question on the stabilization of the solutions of boundary-value problems for parabolic systems of
the form
u
t
+ Au = f, where A is a strongly elliptic operator, is studied in [99]. The asymptotic behavior of
solutions of the Cauchy problem for parabolic systems is studied in [113, 114].
Conditions sucient to assure the fact that solutions of the Cauchy problem for parabolic equations and
systems tend to the mean value of the initial function u(x, 0) as t are described in [114116].
The asymptotic behavior of the solutions of boundary-value problems for parabolic equations and systems
when the equations considered degenerate as t is studied in [117, 118].
Questions on the asymptotic behavior of the solutions of the Cauchy problem and boundary-value problems
for parabolic second-order equations were studied also in [119123] and in other papers.
REFERENCES
1. E. B. Dynkin, Markov processes and related analysis problems, Usp. Mat. Nauk, 15, No. 2, 324 (1960).
2. C. Miranda, Equazioni a derivate parziali di tipo ellittico. Ergehbnisse f ur Mathematik und ihre Grenzgebiete,
Neue Folge, Heft 2, Springer-Verlag, BerlinG ottingenHeidelberg (1955).
3. I. G. Petrovskiy, On the Cauchy problem for linear partial dierential equations in nonanalytic functions,
Bull. MGU, sect. A 1, 7 (1938).
4. I. M. Gelfand and G. E. Shilov, Some Problems of the Theory of Dierential Equations (Generalized Functions,
Vol. 3) [in Russian], Gos. Izd. Fiz. Mat. Lit., Moscow (1958).
5. V. A. Ilyin, On the solvability of boundary-value problems for hyperbolic and parabolic equations, Usp.
Mat. Nauk, 15, No. 2, 97154 (1960).
6. T. Kato, Integration of the equation of evolution in a Banach space, J. Math. Soc. Jpn., 5, No. 2, 209234
(1953).
7. P. E. Sobolevskiy, On equations of parabolic type with nonbounded operators in Banach spaces, Trudy
MMO, 10, 297350 (1961).
8. M. A. Krasnoselskiy and S. G. Kreyn, On dierential equations with nonbounded operators in Banach
spaces, Dokl. Akad. Nauk SSSR, 111, No. 1, 1922 (1956).
9. A. M. Ilyin, Degenerate elliptic and parabolic equations, Mat. Sb., 50 (94), No. 4, 443498 (1960).
10. E. M. Landis, Some problems of the qualitative theory of elliptic and parabolic equations, Usp. Mat. Nauk,
14, No. 1, 2185 (1959).
11. O. A. Oleynik, Boundary-value problems for linear equations of elliptic and parabolic type, Izv. Akad. Nauk
SSSR, Ser. Mat., 25, No. 1, 320 (1961).
12. A. A. Samarskiy, Equations of the parabolic type with discontinuous coecients, Dokl. Akad. Nauk SSSR,
121, No. 2, 255228 (1958).
13. L. I. Kamynin and V. N. Maslennikov, On the maximum principle for a parabolic equation with discontinuous
coecients, Sib. Mat. Zh., 2, No. 3, 384399 (1961).
14. S. L. Kamennomostskaja (S. Kamin), On the Stephan problem, Mat. Sb., 53, No. 4, 489514 (1961).
15. O. A. Oleynik, On a method of solving the general Stephan problem, Dokl. Akad. Nauk SSSR, 135, No. 5,
10541057 (1960).
16. Lee De-yuan, On uniqueness of the solution of the Cauchy problem for an equation of parabolic type, Dokl.
Akad. Nauk SSSR, 129, No. 5, 979982 (1959).
17. R. Z. Khasminskiy, On some dierential equations arising in the study of oscillations with small random
perturbations, Dokl. Akad. Nauk SSSR, 142, No. 3, 560563 (1962).
18. O. A. Oleynik and S. N. Kruzhkov, Quasilinear second-order parabolic equations with several independent
variables, Usp. Mat. Nauk, 16, No. 5, 115155 (1961).
19. I. G. Petrovskiy, Lectures on Partial Dierential Equations [in Russian], Gos. Izd. Fiz. Mat. Lit., Moscow
(1961).
537
20. L. Nirenberg, A strong maximum principle for parabolic equations, Comm. Pure Appl. Math., 6, No. 2,
167177 (1953).
21. O. A. Oleynik, On the properties of solutions of some boundary-value problems for equations of elliptic type,
Mat. Sb., 30, No. 3, 695702 (1952).
22. R. Vyborny, On the properties of solutions of some boundary-value problems for equations of parabolic type,
Dokl. Akad. Nauk SSSR, 117, No. 4, 563565 (1957).
23. E. Holmgren, Sur les solutions quasianalytiques de lequation de la chaleur, Arkiv Math., 18 (1924).
24. A. N. Tikhonov, Uniqueness theorems for the heat equation, Mat. Sb., 42, No. 2, 199216 (1935).
25. S. T acklind, Sur les classes quasianalytiques des solutions des equationa aux derivees partielles du type
parabolique, Nord Acta Regial Soc. Sci. Uppsaliensis (4), 10 (1936).
26. G. N. Zolotarev, On the uniqueness of the solution of the Cauchy problem for systems parabolic in the sense
of I. G. Petrovskiy, Izv. Vyssh. Uchebn. Zaved., Mat., No. 2, 118135 (1958).
27. S. D. Eydelman, On fundamental solutions of parabolic systems. II, Mat. Sb., 53 (95), No. 1, 73136
(1961).
28. Ya. I. Zhitomirskiy, Cauchy problem for parabolic systems of linear partial dierential equations with growing
coecients, Izv. Vyssh. Uchebn. Zaved., Mat., No. 1, 5574 (1959).
29. M. Krzyzanski, Sur lunicite des solutions des second et troisi`eme probl`emes de Fourier relatifs a lequation
lineaire normale du type parabolique, Ann. Polon. Math., 7, No. 2, 201208 (1960).
30. S. Bernstein, Sur la generalisation du probl`eme de Dirichlet (Premi`ere partie), Math. Ann., 62, 253271
(1906).
31. S. Bernstein, Sur les equation du calcul des variations, Ann. Ec. Norm., 431485 (1912).
32. I. G. Petrovskiy, A new proof of the existence of solution for the Dirichlet problem by means of nite
dierences, Usp. Mat. Nauk, 8, 161170 (1940).
33. O. A. Oleynik and T. D. Wentzel, The Cauchy problem and the rst boundary-value problem for a quasilinear
equation of parabolic type, Dokl. Akad. Nauk SSSR, 97, No. 4, 605608 (1954).
34. O. A. Ladyzhenskaya and N. N. Uraltseva, Quasilinear elliptic equations and variational problems with several
independent variables, Usp. Mat. Nauk, 16, No. 1, 1990 (1961).
35. A. V. Pogorelov, Bending of Convex Surfaces [in Russian], Gostekhizdat, MoscowLeningrad (1951).
36. A. V. Pogorelov, On MongeAmp`ere Equations of Elliptic Type [in Russian], Izd. KhGU (1960).
37. S. N. Bernstein, Bounds for consecutive derivatives for the solutions of equations of parabolic type, Dokl.
Akad. Nauk SSSR, 18, No. 7, 385388 (1938).
38. S. Bernstein, Sur une generalisation des theor`emes de Liouville et de M. Picard, Compt. Rend. Acad. Sci.
(Paris), 151, 636639 (1910).
39. J. Nash, Continuity of solutions of parabolic and elliptic equations, Am. J. Math., 80, No. 4, 931954 (1958).
40. S. N. Kruzhkov, On an a priori estimate for solutions of linear parabolic equations and on solving boundary-
value problems for a class of quasilinear parabolic equations, Dokl. Akad. Nauk SSSR, 138, No. 5, 10051008
(1961).
41. J. Schauder,

Uber lineare elliptische Dierentialgleichungen zweiter Ordnung, Math. Zeitschr., 38, 257282
(1934).
42. C. Ciliberto, Formule di maggiorazione e teoremi di esistenza per soluzioni delle equazioni paraboliche in due
variabili, Ricerche Matem., 3, 4075 (1954).
43. A. Friedman, Boundary estimates for second order parabolic equations and their applications, J. Math.
Mech., 7, No. 5, 771791 (1958).
44. A. Friedman, On quasi-linear parabolic equations of the second order, J. Math. Mech., 7, No. 5, 793809
(1958).
45. A. Friedman, On quasi-linear parabolic equations of the second order. II, J. Math. Mech., 9, No. 4, 539556
(1960).
538
46. A. Friedman, Interior estimates for parabolic systems of partial dierential equations, J. Math. Mech., 7,
No. 3, 393417 (1958).
47. I. M. Gelfand and G. E. Shilov, Generalized Functions and Operations over Them (Generalized Functions,
Vol. 1) [in Russian], Gos. Izd. Fiz. Mat. Lit., Moscow (1958).
48. S. N. Bernstein, Study and solving of second order partial dierential equations of elliptic type, Soobshch.
Khark. Matem. Obshch., Ser. 2, 11, 1164 (1908).
49. S. N. Bernstein, On some a priori estimates for a generalized Dirichlet problem, Dokl. Akad. Nauk SSSR,
124, No. 4, 735738 (1959).
50. O. A. Ladyzhenskaya, On the solvability of principal boundary-value problems for equations of parabolic and
hyperbolic type, Dokl. Akad. Nauk SSSR, 97, No. 3, 395398 (1954).
51. E. Gagliardo, Teoremi di esistenza e di unicita per problemi al contorno relativi ad equazioni paraboliche
lineari e quasi lineari in n variabili, Ricerche Matem., 5, 239257 (1956).
52. O. A. Ladyzhenskaja, On the closure of an elliptic operator, Dokl. Akad. Nauk SSSR, 79, No. 5, 723725
(1951).
53. R. Caccioppoli, Limitazioni integrali per le soluzioni di unequazione lineare elliptica a derivate parziali,
Giorn. Matem. Battaglini, 80, 186212 (19501951).
54. O. V. Guseva, On boundary-value problems for strongly elliptic systems, Dokl. Akad. Nauk SSSR, 102,
No. 6, 10691072 (1955).
55. A. I. Koshelev, A priori estimates in L
p
and generalized solutions of elliptic equations and systems, Usp.
Mat. Nauk, 13, No. 4, 2988 (1958).
56. L. N. Slobodetskiy, Estimates for solutions of elliptic and parabolic systems, Dokl. Akad. Nauk SSSR, 120,
No. 3, 468471 (1958).
57. S. Agmon, A. Douglis, and L. Nirenberg, Estimates near the boundary for solutions of elliptic partial dif-
ferential equations satisfying general boundary conditions, Comm. Pure Appl. Math., 12, No. 4, 623727
(1959).
58. M. Schechter, General boundary-value problems for elliptic partial dierential equations, Comm. Pure Appl.
Math., 12, No. 3, 457486 (1959).
59. M. Schechter, Remarks on elliptic boundary-value problems, Comm. Pure Appl. Math., 12, No. 4, 561578
(1959).
60. E. Rothe, Zweidimensionale parabolische Randwertaufgabe als Grenzfall eindimensionaler Randwertauf-
gaben, Math. Ann., 102, 650670 (1930).
61. M. I. Vishik, A. D. Myshkis, and O. A. Oleynik, Partial dierential equations, Matematika v SSSR za sorok
let, 1, 563636 (1959).
62. T. D. Wentzel, The rst boundary-value problem for a quasilinear parabolic equation with several space
variables, Mat. Sb., 41 (83), No. 4, 499520 (1957).
63. Chzhou Yuy-Lin, Boundary-value problems for nonlinear parabolic equations, Mat. Sb., 47 (89), No. 4,
431484 (1959).
64. M. V. Keldysh, On the solvability and stability of Dirichlet problem, Usp. Mat. Nauk, 8, 171232 (1940).
65. O. A. Oleynik, On the Dirichlet problem for equations of elliptic type, Mat. Sb., 24 (66), No. 1, 314 (1949).
66. G. Giraud, Sur certaines probl`emes non lineaires de Neumann et sur certains probl`emes non lineaires mixtes,
Ann. Ec. Norm. Sup. (3), 49, 1104, 245308 (1932).
67. F. Dressel, The fundamental solution of the parabolic equation, Duke Math. J., 7, No. 4, 186203 (1940).
68. F. Dressel, The fundamental solution of the parabolic equation. II, Duke Math. J., 13, No. 1, 6170 (1946).
69. M. Gevrey, Sur les equations aux derivees partielles du type parabolique, J. Math. Pures Appl. (6), 10,
105148 (1914).
70. E. Rothe,

Uber die Grundl osung bei parabolischen Gleichungen, Math. Zeitschr., 33, 488504 (1931).
539
71. W. Feller, Zur Theorie der stochastischen Prozesse, Math. Ann., 113, 113160 (1936).
72. W. Pogorzelski,

Etude de la solution fondamentale de lequation parabolique, Richerce Matem., 5, 2557


(1956).
73. W. Pogorzelski, Study of the integrals of the parabolic equation and of the boundary value in an unbounded
domain, Mat. Sb., 47 (89), No. 4, 397430 (1959).
74. S. D. Eydelman, On fundamental solutions of parabolic systems, Mat. Sb., 38 (80), No. 1, 5192 (1956).
75. L. N. Slobodetskiy, On the fundamental solution and the Cauchy problem for a parabolic system, Mat. Sb.,
46 (88), No. 2, 229258 (1958).
76. E. E. Levi, Sulle equazioni lineare totalemente ellittiche alle derivate parziali, Rend. Circ. Matem. Palermo,
24, 275317 (1907).
77. S. L. Sobolev, Equations of Mathematyical Physics [in Russian], Gostekhizdat, Moscow (1954).
78. V. I. Smirnov, A Course in Higher Mathematics, Vol. IV [in Russian], Gos. Izd. Fiz. Mat. Lit., Moscow (1958).
79. W. Pogorzelski,

Etude dune fonction de Green et du probl`eme aux limites pour lequation parabolique
normale, Ann. Polon. Math., 4, No. 3, 288307 (1958).
80. W. Pogorzelski, Proprietes des integrales de lequation parabolique normale, Ann. Polon. Math., 4, No. 1,
6192 (1957).
81. L. N. Slobodetskiy, Potential theory for parabolic equations, Dokl. Akad. Nauk SSSR, 103, No. 1, 1922
(1955).
82. S. L. Sobolev, Some Applications of Functional Analysis in Mathematical Physics [in Russian], Izd. LGU
(1950).
83. L. A. Lusternik and V. I. Sobolev, Elements of Functional Analysis [in Russian], Gostekhizdat, Moscow
Leningrad (1951).
84. G. M. Fichtenholz, A Course in Dierential and Integral Calculus, Vol. III, Gos. Izd. Fiz. Mat. Lit., Moscow
(1960).
85. L. A. Lusternik,

Uber einige Anwendungen der direkten Methoden in Variationsrechnung, Mat. Sb., 33,
173202 (1926).
86. R. Courant, K. Friedrichs, and H. Lewy,

Uber die partiellen Dierenzengleichungen der mathematischen


Physik, Math. Ann., 100, 3274 (1928).
87. V. K. Saulev, Integrating of Parabolic Equations by the Method of Grids [in Russian], Gos. Izd. Fiz. Mat. Lit.,
Moscow (1960).
88. L. Garding, Dirichlets problem for linear elliptic partial dierential equations, Math. Scand., 1, 5572
(1953).
89. L. Garding, Cauchy Problem for Hyperbolic Equations [Russian translation], Inostr. Lit., Moscow (1961).
90. M. I. Vishik, On strongly elliptic systems of dierential equations, Mat. Sb., 29, No. 3, 615676 (1951).
91. M. I. Vishik and O. A. Ladyzhenskaya, Boundary-value problems for partial dierential equations and some
classes of operator equations, Usp. Mat. Nauk, 11, No. 6, 4197 (1956).
92. S. L. Sobolev, Methode nouvelle `a resoudre le probl`eme de Cauchy pour les equations lineaires hyperboliques
normales, Mat. Sb., 1 (43), 3972 (1936).
93. K. Friedrichs, On the dierential operators in Hilbert space, Amer. J. Math., 61, 523544 (1939).
94. L. Garding, Some directions and problems in the theory of linear partial derential equations, Usp. Mat.
Nauk, 15, No. 1, 137152 (1960).
95. K. Friedrichs, Symmetric hyperbolic linear dierential equations, Comm. Pure Appl. Math., 7, No. 2, 345392
(1954).
96. K. Friedrichs, Symmetric positive linear dierential equations, Comm. Pure Appl. Math., 11, No. 3, 333418
(1958).
540
97. P. Lax, On Cauchys problem for hyperbolic equations and the dierentiability of solutions of elliptic equa-
tions, Comm. Pure Appl. Math., 8, No. 4, 615633 (1955).
98. A. A. Dezin, Existence and uniqueness theorems for the solutions of boundary-value problems for partial
dierential equations in functional spaces, Usp. Mat. Nauk, 14, No. 3, 2173 (1959).
99. M. I. Vishik, Cauchy problem for equations with operational coecients, the mixed boundary problem for
systems of dierential equations and methods of solving them, Mat. Sb., 39 (81), No. 1, 51148 (1956).
100. O. A. Ladyzhenskaya, On solving nonstationary operator equations, Mat. Sb., 39 (81), No. 4, 491524
(1956).
101. O. A. Ladyzhenskaya, On nonstationary operator equations and their applications to linear problems of
mathematical physics, Mat. Sb., 45 (87), No. 2, 123158 (1958).
102. J. Leray and L. Schauder, Topologie et equations fonctionnelles, Ann. Ec. Norm. Sup., 51, 4578 (1934).
103. O. A. Ladyzhenskaya, A simple proof of the solvability of the principal boundary-value problems and eigen-
value problems for linear elliptic equations, Vestnik LGU, 11, 2329 (1955).
104. E. Hopf,

Uber die Anfangswertaufgabe f ur die hydrodinamischen Grundgleichungen, Math. Nachr., 4,


213231 (1950).
105. M. I. Vishik, Boundary-value problems for quasilinear strongly elliptic systems of equations in divergent
form, Dokl. Akad. Nauk SSSR, 138, No. 3, 518521 (1961).
106. S. D. Eydelman, Integral maximum principle for strongly parabolic systems and some of its applications,
Izv. Vyssh. Uchebn. Zaved., Mat., No. 2, 252258 (1959).
107. L. Nirenberg, Remarks on strongly elliptic partial dierential equations, Comm. Pure Appl. Math., 8, No. 4,
649675 (1955).
108. K. Friedrichs, On the dierentiability of the solutions of linear elliptic dierential equations, Comm. Pure
Appl. Math., 6, No. 3, 299325 (1953).
109. A. M. Ilyin, On the behavior of the solution of Cauchy problem for a parabolic equation with the innite
growing of time, Usp. Mat. Nauk, 16, No. 2, 115121 (1961).
110. R. Z. Khasminskiy, Ergodic properties of recursive diusion processes and the stabilization of solutions of
the Cauchy problem for parabolic equations, Teor. Veroyatn. Primen., 5, No. 2, 196214 (1960).
111. Ya. I. Kanel, Stabilization of solutions of the Cauchy problem for some parabolic equations, Usp. Mat.
Nauk, 17 (1962).
112. Yu. N. Cheremnykh, On the asymptotics of solutions of parabolic equations, Izv. Akad. Nauk SSSR, Ser.
Mat., 23, No. 6, 913924 (1959).
113. S. D. Eydelman, Liouville theorems and stability theorems for solutions of parabolic systems, Mat. Sb.,
44 (86), No. 4, 481508 (1958).
114. S. D. Eydelman and F. O. Porper, On stabilization of the solution of Cauchy problem for parabolic systems,
Izv. Vyssh. Uchebn. Zaved., Mat., No. 4, 210217 (1960).
115. M. Krzyzanski, Sur lallure asymptotique des potentiels de la chaleur et de lintegrale de FourierPoisson,
Ann. Polon. Math., 3, No. 2, 288299 (1956).
116. V. M. Polyakova, On the stabilization of the solution of the heat equation, Dokl. Akad. Nauk SSSR, 129,
No. 6, 12301233 (1959).
117. M. I. Vishik and L. A. Lusternik, Stabilization of solutions of some dierential equations in a Hilbert space,
Dokl. Akad. Nauk SSSR, 111, No. 1, 1215 (1956).
118. M. I. Vishik and L. A. Lusternik, Stabilization of solutions of parabolic equations, Dokl. Akad. Nauk SSSR,
111, No. 2, 273275 (1956).
119. M. Krzyzanski, Sur lallure asymptotique des solution dequation du type parabolique, Bull. Acad. Polon.
Sci. el., III, No. 5, 247251 (1956).
120. M. Krzyzanski, Sur lallure asymptotique des solution des probl`emes de Fourier relatifs a une equation lineaire
parabolique, Atti Acad. Naz. Lincei Rend., Cl. s. mat., Ser. 8, 28, No. 1, 3743 (1960).
541
121. A. Friedman, Asymptotic behavior of solutions of parabolic equations, J. Math. Mech., 8, No. 3, 372392
(1959).
122. A. Friedman, Convergence of solutions of parabolic equations to a steady state, J. Math. Mech., 8, No. 1,
5776 (1959).
123. A. Friedman, Asymptotic behavior of solutions of parabolic equations of any order, Acta Mathematica, 106,
No. 12, 143 (1961).
124. O. A. Ladyzhenskaya and N. N. Uraltseva, A boundary-value problem for linear and quasilinear parabolic
equations, Dokl. Akad. Nauk SSSR, 139, No. 3, 544547 (1961).
125. J. Lions, Sur les probl`emes mixtes pour certains syst`emes paraboliques dans des ouverts non cylindriques,
Ann. Inst. Fourier, 7, 143182 (1957).
126. P. Lax and A. Milgram, Parabolic equations, Ann. Math. Studies, 33, 167190 (1954).
127. V. P. Mikhaylov, On the Dirichlet problem and the rst boundary-value problem for a parabolic equation,
Dokl. Akad. Nauk SSSR, 140, No. 2, 303306 (1961).
128. J. Lions and E. Magenes, Remarques sur les probl`emes aux limites pour operateurs paraboliques, Compt.
Rend. Acad. Sci. (Paris), 251, No. 20, 21182120 (1960).
542

Vous aimerez peut-être aussi