v: 2019-02-11

REAL ANALYSIS LECTURE NOTES

RASUL SHAFIKOV

6. Differential Equations
6.1. Ordinary Differential Equations. Consider a system of d first order ordinary differential
equations (ODE for short) and an initial condition
(1) y 0 = f (t, y), y(t0 ) = y0 ,
where y = (y1 , . . . , yd ) is a vector of unknown functions of a real variable t, y 0 = (dy1 /dt, . . . , dyd /dt),
and f : Ω → Rn is a continuous map on a domain Ω ⊂ Rn+1 . We say that y = y(t) defined on a
t-interval J containing t0 is a solution of the initial value problem (1) if y(t0 ) = y0 , (t, y(t)) ∈ Ω,
y(t) is differentiable and y 0 (t) = f (t, y(t)) for t ∈ J. These requirements are equivalent to the
following: y(t0 ) = y0 , (t, y(t)) ∈ Ω, y(t) is continuous and
Z t
y(t) = y0 + f (s, y(s))ds, t ∈ J.
t0
Here integration should be understood component-wise.
An initial value problem involving a system of equations of m-th order
(2) z (m) = F (t, z, z (1) , . . . , z (m−1) ), z (j) (x0 ) = z0j for j = 0, . . . , m − 1,
where z (j) = dj z/dtj , z and F are n-dimensional vectors, and F is defined on an (mn + 1) di-
mensional domain Ω, can be considered as a special case of (1), where y is a d-dimensional vector
(2) (m−1)
symbolically, y = (z, z (1) , . . . , z (m−1) ), or more precisely, y = (z1 , . . . , zn , z10 , . . . zn0 , z1 , . . . , zn ).
Correspondingly,
f (t, y) = (z (1) , . . . , z (m−1) , F (t, y)), y0 = (z0 , . . . , z0m−1 ).
For example, if n = 1, then z is a scalar, and (2) becomes
y10 = y2 , . . . , ym−1
0 0
= ym , y m = F (t, y1 , . . . , ym ),
yj (t0 ) = z0j−1 for j = 1, . . . , m,
where y1 = z, y2 = z 0 , . . . , ym = z (m−1) .
The most fundamental question concerning ODE (1) is the existence and uniqueness of solutions.
Example 6.1. Consider the initial value problem given by y 0 = y 2 and y(0) = c > 0. It is easy to
c
see that y = 1−ct is a solution, but it exists only on the range −∞ < t < 1/c, which depends on
the initial condition. The initial value problem y 0 = |y|1/2 , y(0) = 0 has more than one solution, in
fact, it has a one-parameter family of solutions defined by y(t) = 0 for t ≤ c, y(t) = (t − c)2 /4 for
t ≥ c ≥ 0. 
The following result gives basic conditions when a local solution of an ODE exists and is unique.
Theorem 6.1. Let y, f ∈ Rd , f (t, y) be continuous on R = {|t−t0 | ≤ a, |y−y0 | ≤ b} and uniformly
Lipschitz continuous with respect to y. Let M be a bound for |f (t, y)| on R, α = min(a, b/M ). Then
the initial value problem (1) has a unique solution y = y(t) on [t0 − α, t0 + α].
1
2 RASUL SHAFIKOV

Proof. Let y 0 (t)≡ y0 . Suppose that y k (t)

has been defined on [t0 − α, t0 + α], continuous, and
satisfies |y k (t) − y0 | ≤ b for k = 0, 1, . . . , n. Put
Z t
n+1
(3) y (t) = y0 + f (s, y n (s)) ds.
t0

Then, since f (t, y n (t)) is defined and continuous on [t0 − α, t0 + α], the same holds for y n+1 (t). It
is clear that Z t
n+1
|y (t) − y0 | ≤ |f (s, y n (s)|ds ≤ M α ≤ b.
t0
Hence, y 0 (t), y 1 (t), . . .
are defined and continuous on [t0 − α, t0 + α] and |y n (t) − y0 | ≤ b.
It will now be verified by induction that
M K n |t − t0 |n+1
(4) |y n+1 (t) − y n (t)| ≤ , for t0 − α ≤ t ≤ t0 + α, n = 0, 1, . . . ,
(n + 1)!
where K is a Lipschitz constant for f . Clearly, (4) holds for n = 0. Assume that it holds for up to
n − 1. Then Z t
n+1 n
f (s, y n (s)) − f (s, y n−1 (s) ds, n ≥ 1.
 
y (t) − y (t) =
t0
Thus, the definition of K implies that
Z t
n+1 n
|y (t) − y (t)| ≤ K |y n (s) − y n−1 (s)| ds
t0
and so, by (4),
t
M Kn M K n |t − t0 |n+1
Z
n+1 n
|y (t) − y (t)| ≤ |s − t0 |n ds = .
n! t0 (n + 1)!
This proves (4) for general n. It follows from this inequality that
X∞
y0 + [yn+1 (t) − yn (t)] =: y(t)
n=0
is uniformly convergent on [t0 − α, t0 + α], i.e., we have a uniform limit
(5) y(t) = lim yn (t).
n→∞
Since f (t, y) is uniformly continuous on R, f (t, yn (t)) → f (t, y(t)) as n → ∞ on [t0 − α, t0 + α].
Thus, by taking the limit in the integral in (3) gives
Z t
(6) y(t) = y0 + f (s, y(s))ds.
t0
Hence, (5) is a solution of (1).
In order to prove uniqueness, let y = z(t) be any solution of (1) on [t0 − α, t0 + α]. Then
Z t
z(t) = y0 + f (s, z(s))ds.
t0
An induction similar to that used above gives, using (3),
M K n |t − t0 |n+1
|y n (t) − z(t)| ≤ for t0 − α ≤ t ≤ t0 + α, n = 0, 1, . . . .
(n + 1)!
If n → ∞, it follows from (5) that |y(t) − z(t)| ≤ 0, i.e., y(t) ≡ z(t). This proves the theorem. 
 REAL ANALYSIS LECTURE NOTES 3

One can show that if the function f in (1) is merely continuous, then the initial value problem
always has a solution, but it may not be unique. This is Peano’s Existence theorem.
Consider now a homogeneous linear system of differential equations of the form
(7) y 0 = A(t)y
and the corresponding inhomogeneous system
(8) y 0 = A(t)y + f (t),
where A(t) is a d × d matrix of continuous functions in t, and f (t) is a continuous vector function
of size d. It is a consequence of the theorem on existence and uniqueness of solutions of ODEs
that (7) has unique solution given the initial condition y(t0 ) = y0 . Further, if y(t) is a solution
of (7) and y(t0 ) = 0 for some t0 , then y(t) ≡ 0. The following is immediate
Proposition 6.2 (Principle of Superposition). Let y = y1 (t), y2 (t) be solutions of (7), then any
linear combination y = c1 y1 (t) + c2 y2 (t) with constants c1 , c2 is also a solution. If y = y1 (t) and
y = y0 (t) are solutions of (7) and (8) respectively, then y = y0 (t) + y1 (t) is a solution of (8).
By a fundamental matrix Y (t) of (7) we mean a d × d matrix such that its columns are solutions
of (7) and det Y (t) 6= 0. If Y = Y0 (t) is a fundamental matrix of solutions and C is a constant d × d
matrix, then Y (t) = Y0 (t)C is also a solution, in fact, any solution of (7) can be obtained this way
for some suitable C.
Example 6.2. Let R be a constant d × d matrix with real coefficients. Consider the system of
differential equations
(9) y 0 = R y.
Let y1 6= 0 be a constant vector, and λ be a (complex) number. By substituting y = y1 eλt into the
equation we see that a necessary and sufficient condition for y to be a solution of (9) is
Ry1 = λy1 ,
i.e., that λ is an eigenvalue and y1 6= 0 be a corresponding eigenvector of R. Thus to each eigenvalue
λ of R there corresponds at least one solution of (9). If R has distinct eigenvalues λ1 , λ2 , . . . , λd
with linearly independent eigenvectors y1 , . . . , yd , then
 
Y = y1 eλ1 t , . . . , y1 eλ1 t
is a fundamental matrix for (9). 
Finally, linear differential equations of higher order can be reduced to a system of first order.
Indeed, let pj (t), j = 0, . . . , d−1, and h(t) be continuous functions. Consider the linear homogeneous
equation of order d
(10) u(d) + pd−1 (t)u(d−1) + · · · + p1 (t)u0 + p0 (t)u = 0
and the corresponding inhomogeneous equation
(11) u(d) + pd−1 (t)u(d−1) + · · · + p1 (t)u0 + p0 (t)u = h(t).
We let y = (u, u(1) , . . . , u(d−1) ),
 
0 1 0 0 ··· 0
 0
 0 1 0 ··· 0 

A(t) = 
 · · · ,

 0 0 0 0 ··· 1 
−p0 −p1 −p2 −p3 ··· −pd−1
4 RASUL SHAFIKOV

and f (t) = (0, . . . , 0, h(t)). With this choice of u, A and f , equation (10) becomes equivalent to (7),
while (11) equivalent to (8). With this we can apply results available for first order systems, in
(d−1)
particular, given the initial conditions u(t0 ) = u0 , u0 (t0 ) = u00 , ..., u(d−1) (t0 ) = u0 , where uj0 are
arbitrary numbers, the corresponding initial value problem has a unique solution. The Principle of
Superposition also holds: let u = u1 (t), u2 (t) be two solutions of (10), then any linear combination
u(t) = c1 u1 (t) + c2 u2 (t) is also a solution, and u(t) = u1 (t) + u0 (t) is a solution of (11) if u0 (t) is.
6.2. Partial Differential Equations. A partial differential equation (PDE for short) is an equa-
tion that involves an unknown function of two or more independent variables and certain partial
derivatives of the unknown function. More precisely, let u denote a function of n independent
variables x1 , . . . , xn , n ≥ 2. Then a relation of the form
(12) F (x1 , . . . , xn , u, ux1 , . . . , uxn , ux1 x2 , . . . ) = 0,
where F is a function of its arguments, is a partial differential equation in u. The following equations
are some examples of PDEs on R2 with coordinates (x, y):
(13) xux + yuy − 2u = 0
(14) yux − xuy = x
(15) uxx − uy − u = 0
(16) uux + yuy − u = xy 2
(17) uxx + x(uy )2 + yu = y
A typical problem in the theory of PDEs is for a given equation to find on some domain of Rn
a solution satisfying certain additional initial or boundary conditions. Analogous to ODEs, the
highest-order derivative appearing in a PDE is called the order of the equation. Thus, (13), (14),
and (16) are all first-order PDEs, and the remaining two are second-order. If there exists a function
u defined in a domain under consideration, such that u and its derivatives identically satisfy (12),
then u is called a solution of the equation.
A PDE is called linear if it is at most of the first order in u and its derivatives. Equation (13), (14),
and (15) above are linear, while the other two are not.
Example 6.3. The first-order linear ODE of the form
du
+ u = f (x)
dx
has the general solution Z x
u(x) = e−(x−t) f (t)dt + Ce−x .
0
Now if we consider the first order PDE on R2 for the unknown function u = u(x, y),
∂u
(18) + u = f (x),
∂x
then its general solution is given by
Z x
u(x) = e−(x−t) f (t)dt + g(y)e−x ,
0
where g(x) is an arbitrary function of y. It is easy to see that for any choice of g, the function
u satisfies (18). Thus the general solution of a PDE may contain some arbitrary functions, not
necessarily constants. 
 REAL ANALYSIS LECTURE NOTES 5

A map between function spaces that involves differentiation is called a differential operator. For
∂2 ∂2
example, the map L : C ∞ (R2 ) → C ∞ (R2 ) given by L(u) = ∂t 2 − ∂x2 is a second-order differential
operator. A differential operator L is called linear if L(c1 u1 + c2 u2 ) = c1 L(u1 ) + c2 L(u2 ). It is
immediate that L is linear if and only if the PDE L(u) = 0 is linear, and that any finite sum of
linear differential operators is again linear. Given a linear operator L, the PDE Lu = 0 is called
a homogeneous PDE, while Lu = f is inhomogeneous for f 6≡ 0. As for ODEs, the Principle of
Superposition also holds: a linear combination of solutions of a homogeneous equation is again a
solution, and the sum of solutions of a homogeneous and inhomogeneous equation is a solution of
the inhomogeneous one.
We now consider three classical PDEs: the Wave equation, the Heat equation, and the Laplace
equation. For a function u = u(x, t), where t ∈ R, x = (x1 , . . . , xn ) ∈ Rn , the equation of the form
∂2u
 2
∂2u ∂2u

2 ∂ u
(19) − c + · · · + = − c2 ∆x u = 0 c = const.,
∂t2 ∂x21 ∂x2n ∂t2
is called the Wave equation. This equation describes many types of elastic and electromagnetic
waves. In many physical application the variable t represents time and x represents coordinates in
the Euclidean space where the physical experiment takes place. As an equation of second order, a
typical initial condition for the Wave equation is of the form
u(x, 0) = f (x),
ut (x, 0) = g(x).
When n = 1 the equation has surprisingly simple solution. We let ξ = x − ct and η = x + ct. After
this change of variable equation (19) has the form uξη = 0, which after elementary considerations
admits the solution u(ξ, η) = F (ξ) + G(η), or after returning to the origin variables,
u(x, t) = F (x − ct) + G(x + ct).
The general solution for an arbitrary n can be obtained using the theory of Fourier series.
Consider now the equation of the form
∂u
(20) − k∆x u = 0.
∂t
This is the so-called Heat equation. Again t ∈ R is the “time” variable, x = (x1 , . . . xn ), and
∆x is the Laplacian in variable x. As a primary physical application, equation (20) describes the
conduction of heat with the function u usually representing the temperature of a “point” with
coordinates x at time t, but more generally it governs a range of physical phenomena described as
diffusive. Typical initial conditions are
u(x, 0) = f (x),
u(0, t) = 0,
where the first condition can be interpreted as the initial temperature of the system, and the second
one declares that the temperature is fixed at the “end” point. Solution of (20) for n = 1 can be
found, for example, using the separation of variables method: assuming k = 1 we seek solution of
the form u(x, t) = X(x)T (t). Putting this into (20) gives
T 0 (t) X 0 (x)
= .
T (t) X(x)
6 RASUL SHAFIKOV

Since the right-hand side is independent of t and the left-hand side of x, each side must be a
constant. This gives
T 0 = λT, X 00 = λX.
Solving these equations gives
√ √
u(x, t) = eλt (Ae λx
+ Be− λx
).
Initial conditions will then specify the values of the constants.
Finally, for a function u = u(x1 , . . . , xn ) consider the equation
∂2u ∂2u
(21) ∆u = 2 + · · · + 2 = 0.
∂x1 ∂xn
This PDE, which is called the Laplace equation, has many applications in gravitation, elastic
membranes, electrostatics, fluid flow, etc, as well as numerous applications in other areas of pure
mathematics. There are two types of boundary conditions on a bounded domain Ω ⊂ Rn for the
Laplace equation that give a well-posed problem: the Dirichlet condition
u(x) = f (x), x ∈ ∂Ω,
and the Neumann condition
∂u
(x) = f (x), x ∈ ∂Ω,
∂~n
∂u
where ∂~ n (x) is the derivative in the normal direction to ∂Ω. Solutions of (21) are called harmonic
functions. For n = 2, solutions can be found again by separation of variables.
Unlike the theory of ODE, not every PDE has a solution. In 1957 H. Lewy found an example
of a first order linear PDE that has no solution. The corresponding (complex-valued) differential
operator on C ∞ (R3 ) is
Lu = −ux − iuy + 2i(x + iy)uz .
Then there exists a real valued function f (x, y, z) of class C ∞ (R3 ) such that the equation Lu =
f (x, y, z) has no solution of class C 1 (Ω) in any open subset Ω ⊂ R3 . While Lewy’s example is not
explicit, later explicit constructions were also found.
The situation is different, however, if the functions involved in a PDE are real-analytic. Recall
that for a domain Ω ⊂ Rn a function f (x) : Ω → R is real-analytic if near any point in Ω it can
be represented by a convergent power series. More precisely, if a = (a1 , . . . , an ) ∈ Ω, there exists a
polydisc U (a, r) = {x ∈ Rn : |xj − aj | < r, j = 1, . . . , n}, U ⊂ Ω such that
∞
X
f (x) = bk (x − a)k ,
|k|=0

where k = (k1 , k2 , . . . , kn ) is the multi-index, |k| = k1 + k2 + · · · + kn , ak ∈ R, and

(x − a)k = (x1 − a1 )k1 · ... · (xn − an )kn .
The coefficients ak of the power series are, in fact, the Taylor coefficients of f that can be computed
by the formula
1 k 1 ∂ |k| f
bk = D f (a) = (a).
k! k1 !k2 ! . . . kn ! ∂xk11 ∂xk22 · · · ∂xknn
 REAL ANALYSIS LECTURE NOTES 7

The space of real-analytic functions is denoted by C ω (Ω).

Every real-analytic function in infinitely
differentiable, but the converse is not true, for example the function h(x) on R given by
(
0, if x ≤ 0,
h(x) =
e−1/x , if x > 0
is of class C ∞ (R) as repeated application of the L’Hôspital rule shows, but it is not real analytic
at the origin. The reason is that all derivatives of h at zero vanish, but h is not identically zero
on any neighbourhood of the origin, and so h cannot be represented by its Taylor series. A map
f : Rn → Rm is real-analytic if every component of f is a real-analytic function.
Theorem 6.3 (Cauchy-Kovalevskaya theorem). Consider the system of partial differential equa-
tions
n−1 N
∂ui XX ∂uj
(22) = akij (x1 , . . . , xn−1 , u1 , . . . , uN ) +bi (x1 , . . . , xn−1 , u1 , . . . , uN ), i = 1, . . . , N
∂xn ∂xk
k=1 j=1
with the initial condition
(23) ui = 0 on xn = 0, i = 1, . . . , N.
Let the functions akij
and bi be real analytic at the origin of RN +n−1 . Then the system (22) with
the initial condition (23) has a unique (among real-analytic functions) system of solutions ui that
is real analytic at the origin.
The proof of the Cauchy-Kovalevskaya theorem can be found in comprehensive textbooks on
PDEs.