Université Paris-Dauphine

January 2020
MIDO - L3
Differential equations
Notes by Yannick Viossat, revisited by Daniela Tonon
2
Contents

1 Existence and uniqueness of solutions 5

1.1 A simple example and remarks on notation. . . . . . . . . . . . . . . . . . . . . . 5
1.2 First order differential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Existence and uniqueness of solutions . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 How can solutions cease to exist? . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Proof of Cauchy-Lipschitz theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 Autonomous equations 17
2.1 Generalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Equilibria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Autonomous equations in dimension 1 . . . . . . . . . . . . . . . . . . . . . . . . 20

3 Explicit solutions in dimension 1 23

3.1 Guessing ! . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Separation of variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 Linear differential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4 Comparison principles 31
4.1 Comparison principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Applications to blow-up and to perturbation analysis . . . . . . . . . . . . . . . . 34

5 Systems of linear differential equations 39

5.1 The equation X 0 (t) = AX(t) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.1.1 The case A diagonalizable in R . . . . . . . . . . . . . . . . . . . . . . . . 41
5.1.2 Phase portraits for A diagonalizable in R. . . . . . . . . . . . . . . . . . . 42
5.1.3 The case A diagonalizable in C . . . . . . . . . . . . . . . . . . . . . . . . 47
5.1.4 The general case through triangularization . . . . . . . . . . . . . . . . . 49
5.1.5 The equation X 0 (t) = AX(t) and the exponential of a matrix . . . . . . . 50
5.1.6 Computing the exponential of a matrix . . . . . . . . . . . . . . . . . . . 52
5.2 The equation X 0 (t) = AX(t) + B(t) . . . . . . . . . . . . . . . . . . . . . . . . . 53

6 Nth order equations 57

6.1 How to reduce a nth order equation to a first-order one? . . . . . . . . . . . . . . 57

3
4 CONTENTS

6.2 Nth order linear equations with constant coefficients . . . . . . . . . . . . . . . . 58

7 Stability and linearization 63

7.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.2 Stability of linear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.3 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

8 Geometric approach to differential equations 67

8.1 The slope field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
8.2 Idea of Euler method with a fixed step size . . . . . . . . . . . . . . . . . . . . . 68
8.3 The case of autonomous equations . . . . . . . . . . . . . . . . . . . . . . . . . . 68

9 Continuity of the flow 71

10 Lyapunov functions 75
10.1 Evolution of a quantity along trajectories . . . . . . . . . . . . . . . . . . . . . . 75
10.2 Global Lyapunov function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
10.3 Lyapunov function for an equilibrium . . . . . . . . . . . . . . . . . . . . . . . . 77
10.4 Gradient systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
10.5 Hamiltonian systems in R2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Chapter 1

Existence and uniqueness of

solutions

1.1 A simple example and remarks on notation.

A differential equation is an equation where the unknown is a function instead of a number.

One of the simplest differential equations is

x0 (t) = ax(t) (1.1)

where a, x0 (t), x(t) are real numbers and x0 (·) denotes the derivative with respect to t of the
function x : J → R, J ⊂ R an open interval. This is indeed a functional equation. That is, a
solution of (1.1) is not a number but a function. More precisely, a solution of (1.1) is described
by an open time interval J ⊂ R and a differentiable function x : J → R such that (1.1) holds
for all times t in J. The solutions of (1.1) can be computed explicitly: these are the functions
x : R → R of the form
x(t) = λeat (1.2)
where λ ∈ R, as well as the restrictions of these functions to any open time interval J ⊂ R.1
We immediately see that the choice of J, the time interval of definition, is not unique. If, as
we will soon do, we restrict our attention to solutions defined on a maximal time interval (i.e.
the largest time interval of definition, hence in this case the whole set R), for any initial time
t0 and initial position x0 , there exists a unique solution such that x(t0 ) = x0 . Indeed, imposing
the initial condition x(t0 ) = x0 , we uniquely determine the value of λ, since λeat0 = x0 gives
λ = x0 e−at0 . In this case we say that the initial value problem x0 (t) = ax(t) and x(t0 ) = x0 is
well posed. One of the main questions of this chapter is to determine conditions ensuring that
initial value problems are well posed for more general differential equations.
Before doing so, let us stress that notation does not matter. Depending on application fields
and authors, the derivative of the function x(·) may be denoted by ẋ or dx dt , so that (1.1) will
dx
be written: ẋ(t) = ax(t) or dt (t) = ax(t). For conciseness, we could also skip the indication of
time and write (1.1) in functional form: x0 = ax (or ẋ = ax, or dx
dt = ax). Though written in a
different way, this is the same equation, and the solutions are still given by (1.2). Similarly, in
books for first or second year students, the function is often denoted by y and the variable by
x, so that (1.1) becomes
y 0 (x) = ay(x).
and the solutions y(x) = λeax . Again, this is just a matter of notation: solutions are still
functions whose derivative is equal to a times the function. We prefer the notation (1.1) because
1
Indeed, if (J, x(·)) is a solution, then for all t in J, x0 (t)−ax(t) = 0, hence e−at (x0 (t)−ax(t)) = 0, dt
d
[e−at x(t)] =
−at at
0. Thus, there exists a constant λ such that for all t in J, e x(t) = λ, that is, x(t) = λe . Conversely, any
function of this form is a solution.

5
6 CHAPTER 1. EXISTENCE AND UNIQUENESS OF SOLUTIONS

we want to think of x(t) as describing the position at time t of a particle (body) moving through
space according to the law of motion (1.1). Here, x(t) is a real number, so our particle is moving
on the real line, but later, we will consider particles moving in R2 , R3 , or even Rd , where d is
a positive integer. We will then use the notation X(t) for the particle’s position. That is, our
convention is that x(t) ∈ R and X(t) ∈ Rd . Similarly, except if indicated otherwise, a function
denoted by f takes its values in R while F takes its values in Rd .2
We focus here on the mathematical theory rather than on applications, however differential
equations arise in many different fields: physics, chemistry, ecology, demography, epidemiology,
but also economics and finance. Equation (1.1), for instance, may be seen as a basic population
growth model: the Malthusian model.3 The same equation may be seen as an interest rate
model: at a yearly interest rate r such that ea = 1 + r, a deposit of 100 euros will yield a sum
of 100eat euros t years later. This is an instance of the unifying power of mathematics: a single
equation for several applications.

1.2 First order differential equations

Though we will occasionally consider more general equations, this course focuses on explicit
first-order differential equations, that is, equations of the form

X 0 (t) = F (t, X(t)) (1.3)

with X(t) ∈ Rd and F : Ω → Rd , where Ω is an open subset of R × Rd , d a positive integer. We

see R as a space of times and Rd as a space of possible positions of a particle moving through
Rd . Thus X(t) is the position at time t of our imaginary particle. Eq. (1.3) is of first-order
because second or higher derivatives of X(·) do not appear. It is explicit because it is of the
form (1.3) as opposed to the implicit form G(t, X(t), X 0 (t)) = 0. The equation is autonomous if
F does not depend on t, and nonautonomous otherwise (see Chapter 2).
Example 1.2.1. x0 (t) = x(t) is autonomous; x0 (t) = tx(t) is nonautonomous.
Definition 1.2.2. A solution of (1.3) is described by a nonempty open time-interval J and a
differentiable function X : J → Rd such that for all t in J, (t, X(t)) ∈ Ω and X 0 (t) = F (t, X(t)).4

So strictly speaking, a solution is a couple (interval, function defined on this interval), but
we will often abuse vocabulary and refer to the function as a solution. Though we only require
solutions of (1.3) to be differentiable, i.e. their derivative is defined for all t ∈ J (that’s why J
has to be open!), they are typically much more regular:
Proposition 1.2.3. If F is C k (Ω), then all solutions of (1.3) are at least C k+1 on their domain
of definition.
2
What we want to study are systems whose state at time t may be described by a finite number of parame-
ters x1 (t),..., xd (t), or equivalently by a vector X(t) = (x1 (t), ..., xd (t)), evolving according to some differential
equation. Such problems arise naturally in many different fields. In physics, a possible system would be the
position and velocity of a planet (here d = 2); in chemistry, the concentrations of some chemical substances; in
ecology, the population densities of d species constituting an ecosystem; in epidemiology, the population densities
of three subgroups: those affected by a disease, those who never got the disease, and those who recovered from
the disease and can no longer get it (here d = 3); in finance, the values of various assets; in macroeconomics, a
number of macroeconomic variables such as the unemployment rate, the accumulated capital, a measure of the
population know-how, etc., which together define the state of a simplified economy. Thus our imaginary particle
with position X(t) at time t actually represents the state of some system, and the space in which it evolves is the
set of all possible states for this system. But we will usually not specify what the system is, and just think of a
particle moving through space.
3
(1.2) may be seen as the continuous time version of a discrete time model assuming that at each generation,
the population is multiplied by a constant k > 0. Taking a time-scale such that generation time is T = 1 and
letting a = ln k and λ be the the initial population size, then leads to (1.1) with x(0) = λ. Note that if k > 1, or
equivalently a > 0, then the population grows without bounds, which was frightening Malthus.
4
The condition (t, X(t)) ∈ Ω ensures that F (t, X(t)) is well defined. This condition is void when Ω = R × Rd
and boils down to t ∈ I when Ω = I × Rd for some time interval I ⊂ R.
1.3. EXISTENCE AND UNIQUENESS OF SOLUTIONS 7

Proof. By contradiction, assume that F is C k (Ω) but X : J → Rd is not C k+1 (J). Note that
X(·) is differentiable since X 0 (·) is defined, hence at least C 0 (J). Let n ≤ k be the highest integer
such that X(·) is C n . Since n ≤ k and F is C k (Ω), F is also C n (Ω), hence t → F (t, X(t)) is
C n (J) as a composite function. Thus, X 0 (·) is C n (J) and X(·) is C n+1 (J), contradicting the
definition of n.

Exercise 1.2.4. Prove Proposition 1.2.3 by induction.

Letting (t0 , X0 ) ∈ Ω, a solution of the initial value problem, also called Cauchy problem,
(
X 0 (t) = F (t, X(t))
(1.4)
X(t0 ) = X0

is a solution (J, X(·)) of (1.3) such that t0 ∈ J and X(t0 ) = X0 .

For instance, a solution of (
x0 (t) = x(t)
(1.5)
x(0) = 1

is the exponential function, that is, the function x : R → R such that for all t in R, x(t) = et .
There are other solutions, indeed, an infinite number of them. This is because for any open
interval J containing 0, the restriction of the exponential function to J is still a solution. Clearly,
this multiplicity of solutions is spurious, since they are all defined by the same formula. To get
rid of it, we introduce the concept of maximal solution.

Definition 1.2.5. A solution (J, X(·)) of (1.3) is maximal if it cannot be extended to a larger
˜ X̃(·)) such that J is strictly contained in J˜
time-interval. That is, if there isn’t any solution (J,
and X̃(t) = X(t) for all t in J.

With this definition, we can now say that the initial value problem (1.5) has a unique maximal
solution (that is, exactly one): the exponential function, defined on the whole R. We say that
this problem is well posed. In the following section, we give well posedness sufficient conditions
for an initial value problem.

1.3 Existence and uniqueness of solutions

In this section, let d be a positive integer, Ω an open subset of R × Rd , F : Ω → Rd and

(t0 , X0 ) ∈ Ω.
When the function F is not continuous, a solution to (1.4) could not exist:

Exercise 1.3.1. Define f : R → R by f (x) = 1 if x ≤ 0 and f (x) = −1 if x > 0. Show that the
autonomous differential equation x0 (t) = f (x(t)) doesn’t have any solution such that x(0) = 0.

When the function F is continuous, it can be shown that (1.4) has at least one solution, but
this solution is not necessarily unique.

Theorem 1.3.2 (Cauchy-Peano-Arzela Global). Suppose F is continuous on Ω. Then the initial

value problem (1.4) has at least a maximal solution (J, X(·)), where J is an open interval.

The proof of this theorem is beyond the scope of this course.

Example 1.3.3. Consider the initial value problem x0 (t) = 3|x(t)|2/3 and x(0) = 0 (here,
d = 1, Ω = R2 , and f (t, x) = 3|x|2/3 ; in particular, f is continuous). The functions x(·) and
y(·) defined on R by x(t) = t3 and y(t) = 0, respectively, are both maximal solutions. Thus, this
problem does not have a unique maximal solution.
8 CHAPTER 1. EXISTENCE AND UNIQUENESS OF SOLUTIONS

Exercise 1.3.4. Show that the above initial value problem has an infinite number of maximal
solutions!

Multiplicity of maximal solutions is a huge problem for applications. This means that the
knowledge of the law of motion and of the initial state of the system is not enough to predict
its evolution. Fortunately, if the function F is regular enough, for instance of class C 1 , then the
initial value problem (1.4) has a unique maximal solution. This result is known as the Cauchy-
Lipschitz theorem in France, and as the Picard or Picard-Lindelöf theorem in other parts of the
world. Funnily, both Cauchy and Picard were French, so this is not a case of French chauvinism!

Theorem 1.3.5 (Cauchy-Lipschitz / Picard-Lindelöf Global). Consider the initial value problem
(1.4).
1) If F is Lipschitz or continuously differentiable on Ω, then the initial value problem (1.4)
has a unique maximal solution. That is, a maximal solution exists and it is unique.
2) Solutions of (1.4) are restrictions of the unique maximal solution to smaller time-intervals:
if (Jmax , Xmax (·)) is the unique maximal solution, then (J, X(·)) is a solution if and only if J
is an open interval containing t0 and included in Jmax , and X(t) = Xmax (t) for all t in J.

Remark 1.3.6. The second point means that we don’t lose generality in focusing on maximal
solutions. So from now on, except if indicated otherwise, solution will always mean maximal
solution.

Remark 1.3.7. A stronger version of Theorem 1.3.5 is stated and proved in Section 1.5.

A fundamental consequence of Theorem 1.3.5 is that two (maximal) solutions cannot inter-
sect. That is, two solutions are either disjoint or equal.

Corollary 1.3.8. Assume F is C 1 (Ω) or Lipschitz. Let X : J → Rd and Y : J 0 → Rd be

maximal solutions of (1.3). If there exists t∗ ∈ J ∩ J 0 such that X(t∗ ) = Y (t∗ ), then J = J 0 and
X(·) = Y (·).

Proof. Let t∗ be a real number such that X(t∗ ) = Y (t∗ ). Let Z ∗ := X(t∗ ) = Y (t∗ ). Then
(J, X(·)) and (J 0 , Y (·)) are two maximal solutions of the initial value problem Z 0 (t) = F (t, Z(t))
and Z(t∗ ) = Z ∗ . Due to the Cauchy-Lipschitz theorem, this problem has a unique maximal
solution, hence (J, X(·)) = (J 0 , Y (·)).

In dimension 1, the theorem is even more powerful. This is because, on a given interval
where both are defined, if two continuous functions do not intersect, then one remains above
the other.

Corollary 1.3.9 (1D Comparison Principle). Let Ω be an open subset of R × R, and let
f : Ω → R be C 1 (Ω) or Lipschitz. Let x : J → R and y : J 0 → R be two maximal solutions of
x0 (t) = f (t, x(t)). If there exists t∗ ∈ J ∩ J 0 such that x(t∗ ) < y(t∗ ), then for all t ∈ J ∩ J 0 ,
x(t) < y(t).

Proof. Assume by contradiction that there there exists t∗ ∈ J ∩ J 0 such that x(t∗ ) < y(t∗ ) and
t1 in J ∩ J 0 such that x(t1 ) ≥ y(t1 ). Since x(·) and y(·) are differentiable, hence continuous, and
since J ∩ J 0 is an open interval, it follows from the intermediate value theorem that there exists
t ∈ J ∩ J 0 such that x(t) = y(t). Due to Corollary 1.3.8, we thus have x(·) = y(·) on J = J 0 .
This contradicts the assumption x(t∗ ) < y(t∗ ).

Exercise 1.3.10. Under the same assumptions of Corollary 1.3.9. Show that if there exists
t0 ∈ J ∩ J 0 such that x(t0 ) ≤ y(t0 ), then for all t ∈ J ∩ J 0 , x(t) ≤ y(t).
1.4. HOW CAN SOLUTIONS CEASE TO EXIST? 9

1.4 How can solutions cease to exist?

The previous section studied existence and uniqueness of solutions. This section studies a related
question: how can solutions cease to exist?
Consider the special case of Eq. (1.3) when the function F is defined on a set Ω of the form
Ω = I × Rd , where I is a nonempty open interval of R. We then say that a solution (J, X(·)) of
(1.3) is global if J = I. In the example below, I = R.
Example 1.4.1. a) The function x : R → R given by x(t) = 0 is a global solution of

x0 (t) = x2 (t) (1.6)

b) The function y :] − ∞; 1[→ R given by y(t) = 1/(1 − t) is also a solution, but it is not global
since it is not defined on the whole R.

It is immediate that a global solution is maximal (prove it!). However, the converse is not
true: a maximal solution need not be global. To see this, note that in the above example,
the function y(·) is not global, but it is nonetheless maximal. Indeed, y(·) cannot be extended
backward in time as it is already defined until −∞, and it cannot be extended forward in time
because it converges to +∞ as t → 1: a solution extending y(·) would have to be differentiable
hence continuous by definition of solutions of (1.3), but y(·) has no continuous extension.
It is instructive to compare what happens at both extremities of the interval ] − ∞, 1[. At
−∞, the solution y(·) cannot be extended because, so to speak, the time t reaches the boundary
of the space in which it is allowed to evolve. At t = 1, the solution cannot be extended because,
again speaking informally, the position X(t) of the particle reaches the boundary (+∞) of the
space in which it is allowed to evolve. In both cases, (t, X(t)) reaches - or more precisely,
converges to - the boundary of the space Ω in which it is allowed to evolve, implying that the
solution cannot be extended. This turns out to be the only way in which a maximal solution
may cease to exist. To make this precise, we need a definition of “convergence to the boundary”.
In the two definitions below, the closure J¯ of the interval J is taken in R̄ = R ∪ {±∞}.
Definition 1.4.2. Let Y (·) be a function from an interval J of R to Rm . Let t∗ ∈ J¯ and
Y ∗ ∈ Rm . Y ∗ is an accumulation point of Y (·) as t → t∗ if there exists a sequence (tn ) of
elements of J such that tn → t∗ and Y (tn ) → Y ∗ as n → +∞. This is equivalent to: for any
neighborhoods V of t∗ and W of Y ∗ , there exists t in V ∩ J such that Y (t) ∈ W .
Definition 1.4.3. Let Ω be an open subset of Rm and J an interval of R. Let Y : J → Ω and
¯ We say that Y (t) converges to the boundary of Ω as t → t∗ if Y (·) has no accumulation
t∗ ∈ J.
point in Ω as t → t∗ . We then write Y (t) → Bd(Ω) as t → t∗ .
Remark 1.4.4. To say that Y (t) converges to the boundary of Ω as t → t∗ is equivalent to say
that either the set of accumulation points of Y (·) as t → t∗ is empty or all the accumulation
points of Y (·) as t → t∗ belong to the boundary of Ω.
Remark 1.4.5. To say that Y (t) converges to the boundary of Ω as t → t∗ is equivalent to say
that for any compact subset K ⊂ Ω, it exists a neighborhood V of t∗ such that Y (t) ∈
/ K for all
t ∈ V ∩ J.
Example 1.4.6. Let I be an open interval, t∗ be in the closure of I and X : I → Rd be an
arbitrary function. If Ω = I × Rd and Y (t) = (t, X(t)), then Y (t) → Bd(Ω) as t → t∗ if and
only if (t∗ ∈ {sup I, inf I} or ||X(t)|| → +∞ as t → t∗ , or both). Here || · || is the euclidean
norm in Rd .

We can now state our theorem. Note that we require here that F is C 1 or Lipschitz in order
to use Cauchy-Lipschitz / Picard-Lindelöf Theorem 1.3.5 to show the existence of a solution.
However, since uniqueness is not required, continuity would be enough for the existence of a
solution, due to Cauchy-Peano-Arzela Theorem 1.3.2.
10 CHAPTER 1. EXISTENCE AND UNIQUENESS OF SOLUTIONS

Theorem 1.4.7 (Characterization of maximal solutions). Let d be a positive integer, and Ω

an open subset of R × Rd . Let (J, X(·)) be a solution (not necessarily maximal) of (1.3), with
F : Ω → Rd assumed to be C 1 or Lipschitz. Then (J, X(·)) is a maximal solution if and only if
(t, X(t)) converges to the boundary of Ω, when t → sup J and when t → inf J.

Remark 1.4.8. Note that in Definition 1.2.2 of solution of (1.3), we could also have taken J
as a general interval of R (not necessarily open). In this case Theorem 1.4.7 would have implied
that J has to be open if (J, X(·)) is a maximal solution. Indeed, suppose J =]α, β] (the case
J = [α, β[ is similar). Then, continuity of X on [β − ε, β] ⊂]α, β], implies X(t) → X(β) ∈ Rd
as t → β (hence (β, X(β)) is an accumulation point for (t, X(t)) as t → sup J = β) and since
β ∈ J, (β, X(β)) ∈ Ω and not on its boundary. A contradiction.

With the additional assumption Ω = I × Rd , the theorem takes the form:

Theorem 1.4.9 (Explosion alternative/ Blow-up in finite time). Let (J, X(·)) be a solution of
(1.3). Then (J, X(·)) is a maximal solution if and only if the following two conditions are both
satisfied:
1) sup J = sup I or (sup J < sup I and ||X(t)|| → +∞ as t → sup J).
2) inf J = inf I or (inf J > inf I and ||X(t)|| → +∞ as t → inf J).

Proof. This follows immediately from Theorem 1.4.7 and Example 1.4.6. Note that in the case
I = R, when the maximal solution ceases to be defined either t or ||X(t)||, or both, go to
infinity, hence the name “explosion alternative” for this result. When I = R, finite time blow-up
is equivalent to say that the solution is not global.

Before proving Theorem 1.4.7, a few remarks are in order.

¯ By continuity of x(·),
Remark 1.4.10. Let x(·) be continuous and x : J ⊆ R → R. Let t∗ ∈ J.
∗
|x(t)| → +∞ as t → t if and only if (x(t) → +∞ or x(t) → −∞). So when d = 1, we may
replace in Theorem 1.4.9 the conclusion |x(t)| → ∞ by (x(t) → +∞ or x(t) → −∞).
By contrast, if X(t) = (x1 (t), x2 (t)) ∈ R2 then it can be that ||X(t)|| → +∞ but neither
|x1 (t)|, nor |x2 (t)| goes to infinity. This is because X(t) can be spiral. For instance, let X(t) =
(et cos t, et sin t). As t → +∞, ||X(t)|| → +∞ but et cos t and et sin t keep taking the value 0 so
their absolute value does not go to +∞. This phenomenon may also arise when sup J < +∞,
e.g., for X̃(t) = X(tan t), with X(t) defined as above.

Remark 1.4.11. Conditions 1) and 2) in Theorem 1.4.9 can be also stated as follows:
1bis) If sup J < sup I, then ||X(t)|| → +∞ as t → sup J.
2bis) If inf J > inf I, then ||X(t)|| → +∞ as t → inf J.
Important: When, for instance, sup J = sup I, it may be that we also have ||X(t)|| → +∞ as
t → sup J, but it is not necessary. For instance, this is the case for the solution t → et of
x0 (t) = x(t), but not for the solution t → 0.

Theorem 1.4.9 implies that if the norm of the solution is bounded by a continuous function
defined on the whole interval I, then the solution is global:

Corollary 1.4.12 (Non explosion criterion). Let Ω = I × Rd . Let (J, X(·)) be a maximal
solution of (1.3). Let g : I → R be continuous. Let t0 ∈ J.
1) If ||X(t)|| ≤ g(t) on [t0 , sup J[, then sup J = sup I.
2) If ||X(t)|| ≤ g(t) on ] inf J, t0 ], then inf J = inf I.
3) If ||X(t)|| ≤ g(t) on J, then X(·) is a global solution.
1.4. HOW CAN SOLUTIONS CEASE TO EXIST? 11

Proof. Let us prove 1). Assume by contradiction that ||X(t)|| ≤ g(t) on [t0 , sup J[ and sup J <
sup I. We have that g is defined at sup J. Since the interval [t0 , sup J] is compact and g(·) is
continuous, there exists a constant K such that g(t) ≤ K on [t0 , sup J]. Therefore ||X(t)|| ≤
g(t) ≤ K on [t0 , sup J[, hence ||X(t)|| cannot go to +∞ as t → sup J. Since neither sup J = sup I
nor ||X(t)|| → +∞ as t → sup J, this contradicts Theorem 1.4.9.
The proof of 2) is similar, and 3) follows from 1) and 2).

Exercise 1.4.13. Let J be a nonempty open interval and f : R × R → R a function of class

C 1 . Let x : R → R, y : R → R and z : J → R be solutions of x0 (t) = f (t, x(t)). Assume that
there exists τ ∈ J such that x(τ ) < z(τ ) < y(τ ). Show that for all t ∈ J, x(t) < z(t) < y(t).
Conclude that since x(·) and y(·) are global then so is z(·).

Proof of Theorem 1.4.7 (characterization of maximal solutions)

In this proof, “solution” refers to a solution which is not necessarily maximal. First note that
if X(t) can be extended to a solution of (1.3) after sup J then by continuity of this extension
(remember Proposition 1.2.3), it must be that (t, X(t)) has a limit (hence an accumulation point)
in Ω as t → sup J, and similarly in inf J. It follows that if (t, X(t)) → Bd(Ω) both in sup J and
in inf J then X(·) is maximal, in other words, these are sufficient conditions of maximality. The
difficulty is to show that these conditions are also necessary.
The idea of the proof is as follows. Assume that (t, X(t)) does not converge to the boundary
of Ω as t → sup J. Then (t, X(t)) has an accumulation point (t∗ , X ∗ ) in Ω as t → sup J, and
necessarily sup J = t∗ < +∞. We have to show that this accumulation point is also a limit.
Since Ω is open, we can find a compact neighborhood of (t∗ , X ∗ ) included in Ω. Since F is
continuous, it follows that F is bounded on this neighborhood. Since X 0 (t) = F (t, X(t)), this
implies that X 0 (·) is bounded when (t, X(t)) is in this neighborhood. For this reason, X(·)
cannot evolve too wildly, and it can be shown that it converges to X ∗ as t → t∗ . We can then
extend X(·) to a solution defined after sup J = t∗ by sticking to X(·) a solution of the initial
value problem: X 0 (t) = F (t, X(t)) and X(t∗ ) = X ∗ . This contradicts the maximimality of X(·).
Note that we only need existence of a solution of X 0 (t) = F (t, X(t)) such that X(t∗ ) = X ∗ , not
uniqueness. For this reason, the theorem is still true if we only assume F to be continuous.
Here is the full proof. We just prove that if X(·) is maximal, then (t, X(t)) converges to the
boundary of Ω as t → sup J. The proof of the same result for inf J is similar.
Let J =]α, β[, with α and β in R̄. Assume that (t, X(t)) does not converge to the boundary
of Ω as t → β, hence Y (·) := (·, X(·)) has an accumulation point in Ω as t → β. By the definition
of accumulation point, there exists (t∗ , X ∗ ) in Ω and an increasing sequence (tn ) converging to
β such that (tn , X(tn )) converges to (t∗ , X ∗ ) as n → +∞. This implies that t∗ = β, hence
β < +∞, and that X(tn ) → X ∗ as n → +∞. We have to show that X ∗ is indeed a limit as
t → β (and not only for the sequence tn !):

Lemma 1.4.14. X(t) → X ∗ as t → β.

Suppose this is true. Let X̃(·) be a solution of X 0 (t) = F (t, X(t)) such that X̃(β) = X ∗ ,
defined on ]β − ε, β + ε[ for some ε > 0 (such a solution exists due to Cauchy-Lipschitz theorem).
Define Z : ]α, β + ε[→ Rd by Z(t) = X(t) if t ∈]α, β[ and Z(t) = X̃(t) if t ∈ [β, β + ε[.
The function Z(·) is continuous, thanks to Lemma 1.4.14, and it is differentiable with derivative
Z 0 (t) = F (t, Z(t)) on ]α, β[ and on ]β, β + ε[. Moreover, Z 0 (t) = F (t, Z(t)) → F (β, Z(β)) as
t → β ± (with t 6= β). By a standard result from calculus, this implies that Z(·) is differentiable
in β with derivative Z 0 (β) = F (β, Z(β)). Thus, Z(·) is a solution of (1.3). Since Z(·) is a strict
extension of X(·), this contradicts the maximality of X(·).
It remains to prove Lemma 1.4.14. Since Ω is open, there exists ε > 0 such that the closed
ball B of center (β, X ∗ ) and radius ε is included in Ω. Since B is compact, the continuous
function F is bounded on B by some constant K. It follows that if t ∈ J and (t, X(t)) ∈ B,
12 CHAPTER 1. EXISTENCE AND UNIQUENESS OF SOLUTIONS

then
||X 0 (t)|| = ||F (t, X(t))|| ≤ K (1.7)
Now recall that tn → β and X(tn ) → X ∗ . Therefore, we can find an integer N such that
tN ≥ β − ε, K(β − tN ) < ε/2 and ||X(tN ) − X ∗ || < ε/2.
We claim that:
Claim 1.4.15. For all t in [tN , β[, ||X(t) − X ∗ || < ε.

Indeed, otherwise, let τ be the smallest time greater than tN such that ||X(t)−X ∗ || ≥ ε (this
smallest time exists by continuity of X(·) and of ||X(·) − X ∗ ||). By definition, ||X(τ ) − X ∗ || ≥ ε
and ||X(t) − X ∗ || < ε for all t ∈ [tN , τ [. The latter implies that for all t ∈ [tN , τ [, (t, X(t)) ∈ B,
hence ||X 0 (t)|| ≤ K. Therefore
Z τ
||X(τ ) − X(tN )|| ≤ ||X 0 (t)||dt ≤ K(τ − tN ) < K(β − tN ) < ε/2.
tN

Thus,
||X(τ ) − X ∗ || ≤ ||X(τ ) − X(tN )|| + ||X(tN ) − X ∗ || < ε/2 + ε/2 = ε,
contradicting the definition of τ .
We now conclude: due to (1.7), Claim 1.4.15 and the fact that tN ≥ β − ε, (t, X(t)) ∈ B
hence ||X 0 (t)|| ≤ K for all t in [tN , β[. Therefore
Z β
||X 0 (t)||dt ≤ K(β − tN ) < +∞.
tN

Therefore, the integral tβN X 0 (t)dt is convergent, which implies that X(·) has a limit when t → β.
R

Finally, since X ∗ is an accumulation point, the only possible limit is X ∗ . This concludes the
proof of Lemma 1.4.14, hence of Theorem 1.4.7.

1.5 Proof of Cauchy-Lipschitz theorem

This Section might be useful for later courses on dynamical systems.

Theorem 1.3.5 is a simplified version of the standard Cauchy-Lipschitz theorem. Before
stating and proving the standard theorem, we need to introduce the notion of locally Lipschitz
functions.
Definition 1.5.1. A function F defined on a subset Ω of Rn is locally Lipschitz if every X in
Ω has a neighborhood V such that F is Lispschitz on V ∩ Ω,
Remark 1.5.2. A function F defined on a subset Ω of Rn is locally Lipschitz if and only if for
all compact sets C ⊂ Ω, the restriction of F to C is Lipschitz.

Of course, any Lipschitz function is locally Lipschitz, but the converse is not true. This is
because in the definition, the Lipschitz constant of the restriction of F to C may depend on C,
and may go to infinity as C becomes larger. For instance, the function f : R → R defined by
f (x) = x2 is locally Lipschitz but not Lipschitz. Indeed, since f 0 (x) → +∞ as x → +∞, it
follows that f is not Lipschitz on R. However, on every compact subset C of R, f 0 (x) is bounded,
which implies that the restriction of f to C is Lipschitz. Generalizing this reasoning, it is easy
to see that any C 1 function is locally Lipschitz. Thus, locally Lipschitz functions generalize
both Lipschitz and C 1 functions. This generalization is strict in the sense that there are locally
Lipschitz functions that are neither Lipschitz nor C 1 . For instance, the function f : R → R
defined by f (x) = x if x > 0 and f (x) = x2 if x ≤ 0. Moreover, being locally Lipschitz is stronger
√
than being merely continuous. For instance, the function f : R+ → R defined by f (x) = x is
continuous but not locally Lipschitz.
1.5. PROOF OF CAUCHY-LIPSCHITZ THEOREM 13

Exercise 1.5.3. Being differentiable and being locally Lipschitz are two conditions that are
stronger than continuity but weaker than being C 1 . Show that these conditions cannot be com-
pared: give an example of a function which is locally Lipschitz but not differentiable, and an ex-
ample of a function which is differentiable but not locally Lipschitz (hint: try f (x) = xn sinm ( x1 )
for x 6= 0, and f (0) = 0, for a good choice of n and m).

The assumption of the standard Cauchy-Lipschitz theorem is not exactly that the function F
in (1.3) is Lipschitz or locally Lipschitz, but that it is (jointly) continuous, and locally Lipschitz
with respect to the space variable X.5 Being locally Lipschitz in X means that for all compact
sets C ⊂ Ω, there exists a constant KC such that, for all (t, X, Y ) ∈ R × Rd × Rd :

if (t, X) and (t, Y ) are in C, then ||F (t, X) − F (t, Y )|| ≤ KC ||X − Y ||.

We can now state the standard Cauchy-Lipschitz theorem:

Theorem 1.5.4 (Cauchy-Lipschitz / Picard-Lindelöf Global, Strong form). Let d be a positive

integer, and Ω an open subset of R × Rd . Let F : Ω → Rd . If F (t, X) is jointly continuous
in (t, X) and locally Lipschitz with respect to X, then the initial value problem (1.4) is well
posed, that is, it has one and only one maximal solution. Moreover, all solutions of (1.4) are
restrictions of the maximal solution.

The proof of the theorem proceeds in two steps. We first prove existence and uniqueness of
a local solution, that is, defined on some smaller closed interval [t0 − a, t0 + a]. We then show
how to extend this solution to a maximal time interval.

Lemma 1.5.5 (Cauchy-Lipschitz / Picard-Lindelöf Local, Strong form). Under the same as-
sumptions of Theorem 1.5.4, if a > 0 is small enough, then the initial value problem (1.4) has
a solution defined on [t0 − a, t0 + a] and this solution is unique.

Remark 1.5.6. The use of the term solution in the above theorem is improper according to
Definition 1.2.2. What Lemma 1.5.5 means, is that there exists a function X : [t0 −a, t0 +a] → Rd
such that for all t in [t0 − a, t0 + a], (t, X(t)) ∈ Ω and X 0 (t) = F (t, X(t)). Of course, in order
that differentiability makes sense for X(·) up to the boundary of [t0 − a, t0 + a], it is better to see
the solution as defined on a bigger nonempty open time-interval J containing [t0 −a, t0 +a]. This
will not be a problem since we are not saying that the solution given by Lemma 1.5.5 is maximal.
See also Remark 1.4.8 which ensures that maximal solutions are defined on open intervals.

Proof. To simplify the exposition, assume that t0 = 0. We begin with a fundamental remark:
let Y : [−a, a] → Rd . Then Y (·) is a solution of (1.4) if and only if Y (·) is continuous and
satisfies the integral equation
Z t
∀t ∈ [−a, a], Y (t) = X0 + F (s, Y (s)) ds. (1.8)
0

Indeed, if Y (·) is a solution of (1.4) we can find (1.8) just integrating (1.4) over 0 and t. On the
other hand, if Y , given by (1.8), is continuous on [−a, a], then F (·, Y (·)) is continuous on [−a, a].
Hence Y (·) is continuously differentiable on [−a, a] and Y 0 (t) = F (t, Y (t)) for all t ∈ [−a, a].
Moreover Y (0) = X0 . Hence Y (·) is a solution of (1.4).
Thus, it suffices to prove that the integral equation (1.8) has a unique (continuous) solution.
To do so, the idea is to define a function Φ which associates to Y (·) the right-hand side of (1.8)
and to show that this function has a unique fixed point, this will prove that (1.8) has a unique
solution. The assumptions we will make during the proof ensure that this auxiliary function is
5
If F is continuous in t and locally Lipschitz in X, then F is jointly continuous, so we may replace the
assumption‘jointly continuous in (t, X)” by “continuous in t”.
14 CHAPTER 1. EXISTENCE AND UNIQUENESS OF SOLUTIONS

well defined and contracting, so that the existence and uniqueness of a fixed point follows from
Banach fixed point Theorem 6 (also called contraction mapping Theorem).
Let B̄r denote the closed ball with center (0, X0 ) and radius r. Since Ω is open:
Claim 1: there exists r > 0 such that B̄r ⊂ Ω.
Since F is locally Lipschitz with respect to X:
Claim 2: there exists Kr > 0 such that F is Kr -Lipschitz with respect to X on B̄r .
Since F is jointly continuous and B̄r compact:
Claim 3: F is bounded on B̄r by some constant M .
Finally, we may choose a > 0 such that:
Condition 4: aKr < 1, aM < r and a < r.
Let Ea denote the set of continuous functions X(·) : [−a, a] → Rd such that X(0) = X0 and

||X(·) − X0 ||∞ ≤ r (1.9)

Note that using the supremum norm, defined as ||X(·)||∞ = supt∈[a,a] ||X(t)||, this is a complete
metric space (e.g., as a closed subset of the set of functions from [−a, a] to Rd ). Indeed C 0 ([a, a])
is a complete metric space with respect to the supremum norm (note that C 0 ([a, a]) is even more
than metric, it is a complete normed space, i.e. a Banach space), and Ea is a closed subset of
C 0 ([a, a]).
Step 1: The auxiliary mapping. If X(·) ∈ Ea , define X̃ := Φ(X) by
Z t
∀t ∈ [−a, a], X̃(t) = Φ(X)(t) := X0 + F (s, X(s)) ds
0

Clearly, X̃(·) is continuous and such that X̃(0) = X0 . Moreover, for all t in [−a, a],
Z t
||X̃(t) − X0 || = || F (s, X(s)) ds || ≤ |t|M ≤ aM < r (1.10)
0

where the inequalities follow from Claim 3 and Condition 4. Thus X̃(·) ∈ Ea and we may see
Φ : Ea → Ea as a mapping from Ea to itself.
Step 2: This mapping is contracting. Indeed, if X(·) and Y (·) are in Ea , then due to Claim
2 and to a < r,

∀s ∈ [−a, a], ||F (s, X(s)) − F (s, Y (s))|| ≤ Kr ||X(s) − Y (s)|| ≤ Kr ||X(·) − Y (·)||∞ .

Thus letting X̃ = Φ(X) and Ỹ = Φ(Y ), for all t ∈ [−a, a]

Z t
||X̃(t)− Ỹ (t)|| = || [F (s, X(s))−F (s, Y (s))] ds || ≤ |t|Kr ||X(·)−Y (·)||∞ ≤ aKr ||X(·)−Y (·)||∞
0

so ||Φ(X) − Φ(Y )||∞ ≤ λ||X(·) − Y (·)||∞ with λ = aKr < 1 thanks to Condition 4.
Therefore, Banach fixed point Theorem ensures that the mapping Φ has a unique fixed point,
that is, there exists exactly one function in Ea that solves (1.8). Thus we have the existence of
a solution, and its uniqueness within Ea .
In order to prove that (1.8) has only one solution, we still need to show that all solutions are
in Ea (there could be solutions outside this set, for example solutions for which ||X(τ )−X0 || > r
6
Banach fixed point Theorem Let (X, d) be a non-empty complete metric space with a contraction
mapping T : X → X, i.e. there exists δ ∈ [0, 1) such that

d(T (x), T (y)) ≤ δd(x, y)

for all x, y in X. Then T admits a unique fixed-point x∗ in X (i.e. T (x∗ ) = x∗ ). Furthermore, x∗ can be found as
follows: start with an arbitrary element x0 in X and define a sequence (xn )n∈N by xn = T (xn−1 ), then xn → x∗ .
1.5. PROOF OF CAUCHY-LIPSCHITZ THEOREM 15

for a τ ∈ [−a, a]). To do so, consider a continuous solution Y : [−a, a] → Rd of (1.8). Necessarily,
Y (0) = X0 , so if Y (·) ∈/ Ea , this is because (1.9) is not satisfied. This implies that there exists
a time τ ∈ [−a, a] such that ||Y (τ ) − X0 || > r. To fix the ideas, assume τ > 0. By continuity
of Y (·), there exists a first (i.e. minimal) time t1 > 0 such that ||Y (t1 ) − X0 || > r. Since
||Y (t) − X0 || ≤ r for all t ∈ [0, t1 [, by the same computation as in (1.10), ||Y (t1 ) − X0 || ≤ t1 M ≤
aM < r, a contradiction. Therefore, all solutions are in Ea , this concludes the proof of Lemma
1.5.5.

Remark 1.5.7. Under the same assumptions as in the above lemma, we can construct a sequence
of functions, called Picard’s sequence, X n : [−a, a] → R, fixing X 0 ≡ X0 and for all n ≥ 0,
Z t
∀t ∈ [−a, a], X n+1 (t) = X0 + F (s, X n (s)) ds. (1.11)
0
The reader can show (see the exercise below) that ∀n ∈ N, X n is a sequence of functions in
Ea , where Ea is defined as in the proof of the above Lemma. Hence by the Banach fixed point
Theorem, X n uniformly converges to the unique solution X. Thus, this sequence can be used to
compute the solution numerically.

Exercise 1.5.8 (Proof of Lemma 1.5.5 using Picard’s method). Under the same assumptions of
Theorem 1.5.4, let (X n )n∈N be the Picard’s sequence of functions defined above. Let Ea , r, Kr , M
be defined as in the proof of Lemma 1.5.5.

1) Prove by induction that ∀n ∈ N, X n ∈ Ea , i.e. that for all n ∈ N, X n is continuous on

[−a, a],
X n (0) = X0 , and ∀t ∈ [−a, a]||X n (t) − X0 || ≤ r.

2) Prove by induction that ∀n ≥ 1, ∀t ∈ [−a, a],

(Kr )n−1 |t|n

||X n (t) − X n−1 (t)|| ≤ M .
n!

3) Prove that the series of functions +∞ n n−1 normally converges on [−a, a], hence


n=1 X − X
P

the convergence is uniform on [−a, a].

4) Deduce that (Xn ) uniform converges on [−a, a], to a function X : [−a, a] → R, that is
continuous and such that X(0) = X0 , ||X(·) − X0 ||∞ < r.

5) Using Claim 2, show that the sequence of functions F (·, X n (·)) converges uniformly to
F (·, X(·)) on [−a, a].

6) Conclude that
Z t
X(t) = X0 + F (s, X(s))ds
0
and hence X(·) is a solution of the initial value problem (1.4).

Lemma 1.5.9. Let (JY , Y ) and (JZ , Z) be two solutions of the initial value problem (1.4). Then
Y (t) = Z(t) for all t in J = JY ∩ JZ .

Proof. The only set where it makes sense to consider both solutions is J = JY ∩ JZ . Let J ∗ ⊂ J
be the set of points on which Y and Z agree. Note that t0 ∈ J ∗ , hence J ∗ is not empty. Since Y
and Z are continuous, J ∗ is closed in J (being the set of t such that Y (t)−Z(t) = 0). Let t1 ∈ J ∗ ,
hence Y (t1 ) = Z(t1 ). Then, fixing X1 = Y (t1 ), Y and Z are two solutions of X 0 (t) = F (t, X(t))
with the same initial condition X(t1 ) = X1 . Therefore, by Lemma 1.5.5 (uniqueness part) Y
and Z agree on ]t1 − a, t1 + a[, for a small enough. Hence J ∗ is open in J (for all t1 ∈ J ∗ we
have found a neighborhood entirely contained in J ∗ ). Since the only open and closed set in J
are either J or ∅, and J ∗ is non empty, we have that J ∗ is equal to J.
16 CHAPTER 1. EXISTENCE AND UNIQUENESS OF SOLUTIONS

We can now prove Theorem 1.5.4. Consider the set S of all solutions (J, Y ) of the initial value
problem (1.4). This set is nonempty by Lemma 1.5.5, indeed this Lemma gives in particular a
solution defined on the open interval ]t0 − a, t0 + a[⊂ [t0 − a, t0 + a]. Define the “maximal” time
interval [
Jmax := J. (1.12)
(J,Y )∈S

That is, Jmax is the union of all (open) intervals on which the initial value problem (1.4) has a
solution. Note that Jmax is open.7 Define the function Ymax : Jmax → Rd by

∀ t ∈ Jmax , Ymax (t) = Z(t) for all solutions (J, Z) such that t ∈ J.

The function Ymax is well defined: indeed, if t ∈ Jmax , then there exists a solution (JY , Y ) such
that t ∈ JY , so we may give a value to Ymax (t), and if (JZ , Z) is another solution such that
t ∈ JZ , then due to Lemma 1.5.9, Y (t) = Z(t), so there is no ambiguity on the value of Ymax (t).
The fact that around each t ∈ Jmax , Ymax agrees with a solution of the initial value problem
(1.4) implies that Ymax is a solution of this problem. Finally, the definition of Jmax and Lemma
1.5.9 imply that this solution is maximal and that all other solutions are restrictions of this
maximal solution. This concludes the proof.
Adapting the proof of Lemma 1.5.5 we can prove

Theorem 1.5.10. Let d be a positive integer, and I an open subset of R. Let F : I × Rd → Rd .

If F (t, X) is jointly continuous in (t, X), locally Lipschitz with respect to X, and there exist
A, B ≥ 0 such that
||F (t, X)|| ≤ A||X|| + B ∀(t, X) ∈ I × Rd ,
then the unique solution of the initial value problem (1.4) is defined on the whole I.

Remark 1.5.11. This theorem is a particular case of Corollary 4.2.1.

7
Since our definition of solutions requires that they be defined on an open interval, it follows from (1.12) that
Jmax is open, as an uncountable union of open sets. Alternatively, if we had not required J open in the definition
of a solution, we could have proved that Jmax is open as in Remark 1.4.8.
Chapter 2

Autonomous equations

This chapter shows that the theory of Chapter 1 allows us, without any computation, to under-
stand the behavior of autonomous equations in dimension 1. We begin with some generalities.

2.1 Generalities

An autonomous equation is an equation of the form1

X 0 (t) = G(X(t)) (2.1)

with G : Ωd → Rd , where Ωd is an open subset of Rd . It may be seen as a nonautonomous

equation (1.3) with F independent of t by letting Ω = R × Ωd and F (t, X) = G(X) for all t ∈ R.
Thus, any result on nonautonomous equations has an equivalent for autonomous equations. For
instance, it follows from Cauchy-Lipschitz theorem that if G is Lipschitz or C 1 , then for any
t0 ∈ R and X0 ∈ Ωd , (2.1) has a unique maximal solution (J, X(·)) such that X(t0 ) = X0 .
Moreover, by the Explosion alternative, sup J = +∞ or ||X(t)|| → +∞ as t → sup J, or both,
and similarly for inf J.
Autonomous equations are important for three reasons: first, they often arise in applications.2
Second, they have special properties that make them easier to study. Third, any nonautonomous
equation may be seen as an autonomous equation. Indeed, consider a nonautonomous equation

X 0 (t) = F (t, X(t)) (2.2)

with F : Ω → Rd , where Ω ⊂ Rd+1 . Define G : Ω → Rd+1 by G(Y ) = (1, F (Y )) so that if

Y = (t, X), then
G(Y ) = G(t, X) = (1, F (t, X)).
Associate, now, to X : J → Rd the function Y : J → Rd+1 defined by Y (t) = (t, X(t)).

Claim 2.1.1. (J, X(·)) is a solution of (2.2) such that X(t0 ) = X0 if and only if (J, Y (·)) is a
solution of the autonomous equation Y 0 (t) = G(Y (t)) such that Y (t0 ) = (t0 , X0 ).

Proof. X(t0 ) = X0 ⇔ Y (t0 ) = (t0 , X0 ), and X 0 (t) = F (t, X(t)) ⇔ Y 0 (t) = (1, F (t, X(t)) ⇔
Y 0 (t) = G(Y (t)).

The reader can check that if we start from a maximal solution of one equation, we end up
with a maximal solution of the other. We conclude that solving Y 0 (t) = G(Y (t)) is equivalent to
solving (2.2). In practice, this is not a good method to solve nonautonomous equations because
1
More precisely, (2.1) is a first-order autonomous differential equation in implicit form.
2
The term autonomous comes from the fact that, in physics, a system which is “autonomous”, in the sense
that it is isolated from the rest of the world, should be described by autonomous equations.

17
18 CHAPTER 2. AUTONOMOUS EQUATIONS

the dimension increases: instead of having a variable X(t) in Rd , we end up with a variable Y (t)
in Rd+1 . But the fact that we can use this method, in theory means, that for many theoretical
results, it suffices to prove the result for autonomous equation, and then to apply the above
trick to show that the result still holds for nonautonomous equations. This is one of the reasons
why we will mostly focus on autonomous equations.
Exercise 2.1.2. Write the following equations in their equivalent autonomous form:
a) x0 (t) = t + x(t)
b) X 0 (t) = (x0 (t), y 0 (t)) = (t + x(t), t − y(t)), with X(t) = (x(t), y(t)) ∈ R2 .
c) x0 (t) = 3x2 (t) (why is this a silly example?)

Invariance by translation through time.

Imagine a particle whose position X(t) at time t evolves according to a first order differential
equation. If this equation is autonomous, that is, of the form X 0 (t) = G(X(t)), then X 0 (t)
depends only on X(t), and not explicitly on t. That is, the direction and speed of the particle
depend only on its current position.3 It follows that solutions of autonomous equations are
invariant by translation through time. That is, if X : J → Rd is solution of (2.1), then for any
∆ ∈ R, the function defined on JY =] inf J + ∆, sup J + ∆[ by Y (t) = X(t − ∆), Y : JY → Rd ,
is also a solution. Indeed,
Y 0 (t) = X 0 (t − ∆) = G(X(t − ∆)) = G(Y (t))
Moreover, it is easy to see that if X(·) is maximal, then so is Y (·). It follows that if X(·) is the
maximal solution taking the value X0 at time τ0 , then Y (·) is the maximal solution taking the
value X0 at time τ1 = τ0 + ∆. For future reference, we write this as a proposition:
Proposition 2.1.3. Consider an autonomous equation (2.1) with G of class C 1 or Lipschitz.
Let X0 ∈ Rd , τ0 ∈ R and ∆ ∈ R. Let (JX , X(·)) and (JY , Y (·)) be the maximal solutions taking
the value X0 at times τ0 and τ0 + ∆, respectively. Then JY = {t + ∆, t ∈ JX } and:
∀t ∈ JX , Y (t + ∆) = X(t). 4

In particular (take τ0 = 0 and ∆ = t0 ), for any t0 in R, we may recover from the solution
X(·) taking the value X0 at time 0 the solution taking the value X0 at time t0 : this is the
function Y (·) defined by Y (t) = X(t − t0 ). Thus, for autonomous equations, we may assume
without loss of generality that the initial condition is given at time 0.
Exercise 2.1.4. What is the relation between the graph of x : J → R and the graph of the
function y(·) defined by y(t) = x(t − t0 )? Recall that the solution from x0 (t) = x(t) such that
x(0) = 1 is the exponential function. Without computing it, deduce from this fact the graph of
the solution of x0 (t) = x(t) such that x(3) = 1?

Consequences for trajectories. The trajectory 5 (or orbit) generated by a solution (J, X(·)) of
(2.1) is the set {X(t), t ∈ J} ⊂ Rd of positions/values taken by X(t) for t in J.
Proposition 2.1.5. Consider an autonomous equation as in (2.1) and assume G of class C 1
or Lipschitz. Let X0 ∈ Rd .
a) If two solutions take the value X0 , then they follow the same trajectory.
b) The trajectories generated by two solutions are either equal or disjoint.
3
If the equation is nonautonomous, then the direction and speed of the particle also depend on time. Consider
for instance the equation x0 (t) = x(t) sin t. If x(t0 ) = x0 = 1 and t0 = π/2, then x0 (t0 ) = 1, and the particle goes
forward; if x(t0 ) = 1 and t0 = −π/2, then x0 (t0 ) = −1, and the particle goes backward. So indeed, x0 (t0 ) does
not depend only on x(t0 ).
4
The interpretation is that, if two particles with law of motion (2.1) start from the same position X0 , the
second ∆ units of time later than the first one, then they will visit the same points in space, the second particle
always ∆ units of time after the first one.
5
The concept of trajectory is used also for nonautonomous equations
2.2. EQUILIBRIA 19

Proof. a) Let (JX , X(·)) and (JY , Y (·)) be solutions of (2.1) such that, respectively, X(t0 ) =
Y (t1 ) = X0 , for t0 ∈ JX , t1 ∈ JY . Then t1 = t0 + ∆, for a ∆ ∈ R. By Proposition 2.1.3,
JY = {t + ∆, t ∈ JX } and Y (t + ∆) = X(t) for all t in JX . Thus, letting TX and TY be the
trajectories generated by X(·) and Y (·), respectively:

TY = {Y (t), t ∈ JY } = {Y (t + ∆), t ∈ JX } = {X(t), t ∈ JX } = TX .

b) If the trajectories generated by two solutions are not disjoint, then there exists X0 in Rd such
that both solutions take the value X0 , possibly at different times. Then by a), these solutions
generate the same trajectory.

Remark 2.1.6. Note that, for nonautonomous equations, trajectories can intersect. Assume F
is C 1 (Ω) or Lipschitz. Let X : J → Rd and Y : J 0 → Rd be maximal solutions of (1.3). Then,
there could exist J 3 t1 6= t2 ∈ J 0 such that X(t1 ) = Y (t2 ). (Find an example!) This is not in
contradiction with Corollary 1.3.8, which is about solutions and not about trajectories!

2.2 Equilibria

A solution (J, X(·)) of (2.1) is stationary if J = R and X(·) is constant. A point X ∗ ∈ Ωd is an

equilibrium of (2.1) if G(X ∗ ) = 0. The link between these notions is that:

Claim 2.2.1. If X ∗ is an equilibrium, then X(t) = X ∗ for all t ∈ R defines a stationary

solution. Moreover, if G is C 1 or Lipschitz, then for any t0 ∈ R, the stationary solution is the
unique maximal solution such that X(t0 ) = X ∗ .

Indeed, if X ∗ is an equilibrium, we then have that X(t) = X ∗ for all t ∈ R satisfies X 0 (t) =
0 = G(X ∗ ) = G(X(t)) for all t ∈ R, and uniqueness follows from Cauchy-Lipschitz theorem.
Assuming that G is C 1 or Lipschitz is important: otherwise, as in Example 1.3.3, solutions
starting at an equilibrium could go away from it.

Remark 2.2.2. Equilibria are only defined for autonomous equations. Stationary (= constant)
solutions are also defined for nonautonomous equations. For instance, x(t) = 0 is a stationary
solution of x0 (t) = tx(t).

Equilibria are very important, because they are the only possible (finite) limits of solutions.

Proposition 2.2.3 (Importance of equilibria). Let (J, X(·)) be a solution of X 0 (t) = G(X(t))
with G continuous. Let X ∗ ∈ Rd . If X(t) → X ∗ as t → sup J (or as t → inf J), then G(X ∗ ) = 0.
That is, X ∗ is an equilibrium.

Proof. Assume that X(t) → X ∗ as t → sup J. By the Explosion alternative, sup J = +∞.
Moreover, since G is continuous, G(X(t)) → G(X ∗ ) as t → sup J = +∞. Since X 0 (t) = G(X(t)),
this implies that X 0 (t) → G(X ∗ ), hence, letting Y = G(X ∗ ), Xi0 (t) → Yi for all i ∈ {1, .., d}.
If Y 6= 0, then there exists i ∈ {1, ..., d} such that Yi 6= 0. Assume Yi > 0 (the case Yi < 0
is similar). Since Xi0 (t) → Yi as t → +∞, there exists T < +∞ such that for all t ≥ T ,
Xi0 (t) ≥ Yi /2. For t ≥ T ,
Z t Z t
Xi (t) = Xi (T ) + Xi0 (s)ds ≥ Xi (T ) + Yi /2 ds = Xi (T ) + (t − T )Yi /2.
T T

Thus, Xi (t) → +∞ as t → +∞, contradicting our assumption X(t) → X ∗ . Therefore Y = 0.

Remark 2.2.4. Proposition 2.2.3 shows that if a solution of an autonomous equation converges
to a point in inf J or sup J, then this point is an equilibrium. But solutions of an autonomous
equation do not necessarily converge: they may grow to infinity, cycle (when d ≥ 2), or have a
more complicated behavior.
20 CHAPTER 2. AUTONOMOUS EQUATIONS

Example 2.2.5. a) The exponential function is solution on R of the autonomous equation

x0 (t) = x(t). As t → −∞, it converges to 0 which, as implied by Proposition 2.2.3, is an
equilibrium. But as t → +∞, it grows to infinity.
b) The function X : R2 → R defined by X(t) = (cos t, sin t) is solution of the autonomous
equation X 0 (t) = F (X(t)) with F (x, y) = (−y, x). It cycles endlessly, and in particular, does
not converge to an equilibrium.

There are different kinds of equilibria. In particular, an equilibrium is attracting if nearby

solutions move towards it, and repelling if nearby solutions go away from it. Formally:

Definition 2.2.6. Let X ∗ be an equilibrium of (2.1). Then the equilibrium is said to be attract-
ing if it has a neighborhood V such that for any solution (J, X(·)) for which X(t0 ) ∈ V for a
t0 ∈ J, then sup J = +∞ and X(t) → X ∗ as t → +∞. The equilibrium is said to be repelling
if it has a neighborhood V such that for any solution (J, X(·)) with X(t0 ) ∈ V \{X ∗ } for some
t0 ∈ J, there exists T ∈ J, T > t0 , such that X(t) ∈
/ V for all t ∈ J ∩ [T, +∞[.

Other notions of equilibrium stability will be studied in later chapters.

2.3 Autonomous equations in dimension 1

Recall that if the law of motion of a particle is described by an autonomous equation (2.1),
then the direction and speed of this particle depends only on its position. It follows that we
may define, and draw, the direction of movement of the particle when the particle is placed at
X0 . Consider for instance a particle moving through the real line according to the equation
x0 (t) = x(t). If x(t) = x0 = 0, then its speed is zero: the particle is at rest. If x(t) = x0 > 0,
then x0 (t) > 0: the particle goes forward. If x(t) = x0 < 0, then x0 (t) < 0: the particle goes
backward. We can indicate this information on the phase line, that is, the real line seen as
the set of possible positions of the particle, by dots for equilibria and arrows for directions of
movement.
Example of such drawings will be given in the course. The next proposition shows that, for
autonomous equations in dimension 1, they tell us almost everything on the qualitative behavior
of solutions. In the following proposition, saying that x : J → R increases from x1 ∈ R̄ to x2 ∈ R̄
means that x(·) is increasing, x(t) → x1 as t → inf J and x(t) → x2 as t → sup J.

Proposition 2.3.1. Consider an autonomous equation in dimension 1: x0 (t) = g(x(t)), with

g : R → R assumed C 1 or Lipschitz. Let (t0 , x0 ) ∈ R × R. Let (J, x(·)) be the solution such
that x(t0 ) = x0 , (J 3 t0 ). Let x+ be the smallest equilibrium greater than x0 if there is one,
and x+ = +∞ otherwise. Let x− be the largest equilibrium smaller than x0 if there is one, and
x− = −∞ otherwise.

1. If g(x0 ) = 0, that is, if x0 is an equilibrium, then J = R and x(t) = x0 for all t ∈ R.

2. If g(x0 ) > 0, then x(t) increases from x− to x+ .

3. If g(x0 ) < 0, then x(t) decreases from x+ to x− .

4. If x(t) has a finite limit in sup J (resp. inf J), then sup J = +∞ (resp. inf J = −∞).

Proof. 1) follows from Claim 2.2.1. For 2), let us first show that x(t) < x+ for all t ∈ J. If
x+ = +∞, this is obvious. If x+ < +∞, then x+ is an equilibrium, and the result follows from
Claim 2.2.1 and the fact that two distinct solutions cannot cross. Similarly, x(t) > x− for all
t ∈ J. By definition of x+ and x− , there are no equilibria in ]x− , x+ [, so g(x), being continuous,
has a constant sign on this interval. Therefore g(x(t)), hence x0 (t), has a constant sign. Since
2.3. AUTONOMOUS EQUATIONS IN DIMENSION 1 21

g(x0 ) > 0, this sign is positive. Therefore, x(·) is increasing and converges as t → sup J to some
finite or infinite limit x∗ , such that x0 ≤ x∗ ≤ x+ since x(·) is increasing and x(t) < x+ for all t.
Case 1: If x+ < +∞, then x(·) is bounded from above, so x∗ is finite. Therefore, x∗ is an
equilibrium, due to Proposition 2.2.3; due to the fact that x0 ≤ x∗ ≤ x+ and that the only
equilibrium in this interval is x+ , it follows that x∗ = x+ .
Case 2: If x+ = +∞, then there are no equilibria greater than x0 . Therefore, x∗ cannot be
finite, because as a finite limit x∗ should be an equilibrium. Therefore x∗ = +∞ = x+ . Thus in
both cases, x(t) → x+ as t → sup J.
The proof of x(t) → x− as t → inf J is similar, and so is the proof of 3). Finally, point 4)
follows from the explosion alternative.

Exercise 2.3.2. Show that if x− and x+ are finite, that is, if there are equilibria smaller than
x0 and equilibria greater than x0 , then x(·) is bounded and global.

Example 2.3.3. Assume g(x) = x (so x0 (t) = x(t)). The only equilibrium is 0. If x0 = 0,
then by 1), J = R and x(t) = x0 for all t. If x0 > 0, then g(x0 ) > 0, so by 2), x(·) is strictly
increasing, converges to +∞ as t → sup J and to 0 as t → inf J, so that inf J = −∞ by 4).
Theorem 1.5.10 can be applied to prove that sup J = +∞. If x0 < 0, then g(x0 ) < 0, so by
3), x(·) is strictly decreasing, converges to −∞ as t → sup J, and to 0 as t → inf J, so that
inf J = −∞ by 4). Theorem 1.5.10 can be applied to prove that sup J = +∞.

Example 2.3.4. Consider the logistic growth equation x0 (t) = rx(t)(1 − x(t)).6 Then g(x) =
rx(1−x). There are two equilibria: 0 and 1. The function g is negative below 0, positive between
0 and 1, and negative again above 1. If x0 = 0 or x0 = 1, then x(·) is stationary. If x0 < 0,
then x(·) decreases, converges to −∞ as t → sup J, and to 0 as t → inf J (so inf J = −∞).
If x0 ∈]0, 1[, then x(·) is bounded and global, and increases from 0 to 1. If x0 > 1, then x(·)
decreases from +∞ to 1 (hence sup J = +∞).

Proposition 2.3.1 shows that, up to velocity and blow-up considerations, the qualitative
behavior of x0 (t) = g(x(t)) only depends on the sign of g. The method to use this proposition is
to study this sign, then draw the phase line accordingly, and “read” it. Note that the phase line
does not give information on a single solution, but on all solutions. Different arrows (or dots)
correspond to different solutions.

Exercise 2.3.5. Let x∗ be an equilibrium of x0 (t) = g(x(t)). Show that if there exists ε > 0 such
that g is positive on ]x∗ − ε, x∗ [ and negative on ]x∗ , x∗ + ε[, then x∗ is attracting. Conclude that,
whenever g 0 (x∗ ) is well defined, if g 0 (x∗ ) < 0, then x∗ is attracting. Show that if g 0 (x∗ ) > 0,
then x∗ is repelling.

Exercise 2.3.6. For x0 (t) = x2 (t), show that the equilibrium 0 is neither attracting nor repelling.

Exercise 2.3.7. Sketch the phase line for x0 = sin x and show that all solutions are global.

Exercise 2.3.8. For x0 = x2 (x − 1)(x + 1), which equilibria are attracting? repelling?

Exercise 2.3.9. Draw the phase line of x0 = x(1 − x) − p for p > 1/4, p = 1/4 and p < 1/4.

6
In ecology, this models a population whose growth depends on a limited resource. When the population is
small, it grows with a growth rate close to r. As it gets larger, the resource becomes limiting and the growth-rate
diminishes. When its density is higher than the carrying capacity K, that is, the maximal capacity that the
environment may sustain, then its growth rate becomes negative. We chose units so that K = 1.
22 CHAPTER 2. AUTONOMOUS EQUATIONS
Chapter 3

Explicit solutions in dimension 1

In this chapter we will see some techniques to be used in order to find solutions of differential
equations in dimension 1.

3.1 Guessing !

Guessing the solution is a perfectly rigorous way to solve a differential equation. Consider for
instance the initial value problem

x0 (t) = −x2 (t) ; x(0) = 1. (3.1)

Since the function x → x2 is C 1 , this problem has a unique maximal solution. Assume that for
whatever reason you correctly guess that this solution is the function u :] − 1, +∞[→ R defined
by u(t) = 1/(t + 1). Differentiating shows that, indeed, u0 (t) = −1/(t + 1)2 = −u2 (t). Checking
that 0 ∈] − 1, +∞[ and u(0) = 1/(0 + 1) = 1 then proves that u(·) is a solution of (3.1). Is this
solution maximal? Noting that u(·) is defined till +∞ and that u(t) → +∞ as t → −1 shows
that this solution cannot be extended, hence is maximal. Thus, u(·) is the maximal solution of
(3.1).
Note that the proof is perfectly rigorous. The only problem is that you first need to guess the
solution, and this is not necessarily easy! This being said, if in a test you encounter a question
such as “prove that the solution to this initial value problem is this function”, that is, if the
teacher provides the correct guess, do not venture in complicated methods: just check that the
guess is a true solution!

3.2 Separation of variables

Separation of variables in practice. Consider an equation of the form

x0 (t) = g(x(t))h(t) (3.2)

where g and h are real valued functions. Equations that take this form are called “equation with
separated variables”. (Note that they are not autonomous in general.) They may be solved as
follows: without worrying about what this means, write the equation as
dx
= h(t)g(x)
dt
then formally “separate variables” to get:
dx
= h(t)dt
g(x)

23
24 CHAPTER 3. EXPLICIT SOLUTIONS IN DIMENSION 1

Now integrate
dx
Z Z
= h(t)dt
g(x)
to obtain Ψ(x) = H(t) + K where Ψ is a primitive of 1/g, H a primitive of h, and K a constant.
Inverting Ψ and letting x = x(t) leads to the guess that the general solution is

x(t) = Ψ−1 (H(t) + K) (3.3)

which is defined on the maximal interval containing x0 and such that this expression is well
defined. If we want to guess the solution such that x(t0 ) = x0 , then we just need to choose
the constant K appropriately, that is, K = Ψ(x0 ) − H(t0 ). We then check that our guesses are
correct. There are two delicate steps that have to be justified:

• we divided by g(x) and this is not possible for an x such that g(x) = 0,

• we invert Ψ and this is not always possible.

Example 3.2.1. Assume that we want to solve the initial value problem x0 (t) = −x2 (t) and
x(0) = 1 (in this example g(x) = −x2 and h(t) = 1). We know that our solution (J, x(·)) does
not cross the stationary solution (R, 0), since its initial value is x(0) = 1. Therefore x(t) 6= 0
for all t ∈ J. We formally write
dx
= −x2 (t),
dt
then we integrate
dx
Z Z
− = dt,
x2
obtaining
1 1
= t + K ⇒ x(t) = .
x (t + K)
Since we want x(0) = 1, we choose K such that 1/(0 + K) = 1, that is, K = 1. We thus guess
that the solution is given by
x(t) = 1/(t + 1)
and it is defined on the maximal time interval containing t0 = 0 and such that this expression is
well defined, that is, J =] − 1, +∞[. We then check that the function x :] − 1, +∞[→ R defined by
x(t) = 1/(t + 1) is indeed solution, is maximal, and the unique maximal solution, as in Section
3.1.
Example 3.2.2. Assume that we want to find all maximal solutions of the equation x0 (t) = tx(t).
We know that (R, 0) is the unique maximal stationary solution such that x̃(0) = 0 and that,
solutions (J, x(·)) such that x(0) = x0 6= 0 do not cross the stationary solution (R, 0).
We formally write:

dx dx t2
Z Z
2 /2
= tx ⇒ = tdt ⇒ ln |x| = +K ⇒ |x| = eK et
dt x 2
2
hence x(t) = Cet /2 for some constant C. Since this expression is defined on the whole R, we
2
guess that solutions are the functions x : R → R defined by x(t) = Cet /2 , with C ∈ R. Checking
that these are indeed solutions is a simple computation which is left to the reader. Moreover,
since these solutions are global, they are maximal. It remains to check that there are no other
maximal solutions. To do this, let (t0 , x0 ) ∈ R2 , and note that taking C = x0 exp(−t20 /2) gives
a maximal solution such that x(t0 ) = x0 . Since any maximal solution must take some value x0
at some time t0 , this shows that all maximal solutions are indeed of the above form.1
1
Indeed, let (J, u(·)) be a maximal solution, let t0 ∈ J and let x0 = u(t0 ). Then for C0 = x0 exp(−t20 /2),
2
(J, u(·)) and (R, t → C0 et /2 ) are two maximal solutions with the same initial condition, hence they are equal,
since the function (t, x) → tx is C 1 . Therefore, any maximal solution is one of those we had found.
3.2. SEPARATION OF VARIABLES 25

Why does it work? The method may seem weird, but the following proposition shows that
when g and h are regular enough, and when we take care of the case g(x0 ) = 0, the method
always works.

Proposition 3.2.3. Consider the initial value problem x0 (t) = g(x(t))h(t) and x(t0 ) = x0 ,
with g : R → R and h : I → R, where I is an open interval of R. Assume g is of class C 1
or Lipschitz, and h continuous so that, by Theorem 1.5.4, this problem has a unique maximal
solution (J, x̄(·)).
1) if g(x0 ) = 0, the solution is stationary: J = I and x̄(t) = x0 for all t in I.
2) if g(x0 ) 6= 0, then for all t in J, g(x̄(t)) 6= 0 and
Z x̄(t) Z t
du
= h(s)ds (3.4)
x0 g(u) t0

Proof. 1) is a variant of Claim 2.2.1. For 2), assume by contradiction that g(x̄(t1 )) = 0 for some
t1 ∈ J, and let x1 = x̄(t1 ). Then g(x1 ) = 0 and (J, x̄(·)) is a solution of x0 (t) = g(x(t))h(t)
and x(t1 ) = x1 . By Case 1) applied to this initial value problem, this implies that for all t in
J, x̄(t) = x1 , hence g(x0 ) = g(x̄(t0 )) = g(x1 ) = 0. This contradicts our assumption g(x0 ) 6= 0.
Therefore g(x(t)) does not vanish, and we can divide our equation by g(x(t)). We obtain that
for all t in J,
x0 (t)
= h(t)
g(x(t))
so that:
x0 (s)
Z t Z t
ds = h(s)ds
t0 g(x(s)) t0

The change of variable u = x(s) in the first integral and x(t0 ) = x0 then lead to (3.4).

Some remarks are in order to let the procedure done in the previous examples be rigorous.
(Please always check them in the exercises!)
First, we see from Proposition 3.2.3 that the case g(x0 ) = 0 is special and should in principle
be separated from the case g(x0 ) 6= 0. To see what can go wrong, imagine that we want to find
the solution of x(t) = −x2 (t) such that x(0) = 0. If we proceed as in Example 3.2.1, then once
we obtain x(t) = 1/(t+K) we have a problem. Indeed, there is no value of K such that x(0) = 0.
We may overcome the problem in two ways: by noting from the beginning that the solution is
stationary, as follows from point 1) of Proposition 3.2.3, or by noting that to get x(0) = 0, we
would need to have K = ∞. Giving to K an infinite value in x(t) = 1/(t + K) leads to the
correct guess that the solution is x(t) = 0.2 . The origin of this difficulty is that when writing
dx/g(x) = h(t)dt, we implicitly assume that g(x) is nonzero. In Example 3.2.2, we are
2
lucky, as we recover the stationary solution x(t) = 0 from the formula x(t) = Cet /2 by letting
C = 0. While, in Example 3.2.1, we must be careful.
Second, let Ψ(x) = xx0 g(u)du
R
. From the fact that g is continuous and that g(x(t)) does not
vanish, we know that for all t in J, g(u) has a constant sign between x0 and x(t). It follows
that on the domain of interest, Ψ is strictly monotone, hence invertible. So (3.4) indeed leads
to (3.3). More precisely, it leads to (3.3) with K = 0 because we chose specific primitives Ψ and
H: more precisely, such that Ψ(x0 ) = H(t0 ). If we do not take this precaution, then we still get
(3.3) but then we have to fix K using the initial condition.
Finally, if g is not C 1 or Lipschitz but just continuous, we may still apply the method of
separation of variables, but with the following modifications: first, if g(x0 ) = 0, then there is a
stationary solution equal to x0 , but there might also be other solutions. Second, if g(x0 ) 6= 0, the
2
We will see later that solutions of differential equations depend continuously on the initial condition. This
explains why we may recover the solutions corresponding to the case g(x0 ) = 0 as limits of solutions corresponding
to the case g(x0 ) 6= 0.
26 CHAPTER 3. EXPLICIT SOLUTIONS IN DIMENSION 1

formula (3.4) is only valid as long as g(x(t)) does not vanish. See Example 1.3.3 and Exercise
1.3.4.
Exercise 3.2.4. Solve the initial value problem x0 (t) = tx(t) and x(1) = 1.
Remark 3.2.5. An autonomous equation x0 (t) = f (x(t)) with f : I → R of class C 1 is a
particular case of equations with separated variables: the case g(x) = f (x) and h(t) = 1. Thus,
the method of separation of variables allows us to reduce the problem of computing solutions of
these equations to a problem of computation of primitives (however this can still be hard to do
in some case!).
Exercise 3.2.6. Solve (= find explicitly all solutions of) the equation x0 (t) = −x2 (t). Similarly,
solve the logistic equation x0 (t) = x(t)(1−x(t)). Compared to the qualitative resolution of Chapter
2, what additional information do we get?

We conclude with a remark which will be used in Chapter 4.

Remark 3.2.7. Consider the initial value problem x0 (t) = K|x|α and x(t0 ) = x0 , with K > 0,
α > 0. The method of separation of variables allows us to compute explicitly the solutions. If
0 < α ≤ 1, then solutions are global, while if α > 1, solutions explodes: forward in time if
x0 > 0 (they are defined on an interval ] − ∞, T [ with T < +∞, and go to +∞ as t → T ), and
backward in time if x0 < 0 (they are defined on an interval ]T, +∞[ with T > −∞, and go to
−∞ as t → T ).

3.3 Linear differential equations

In this section, I is a nonempty open interval, t0 ∈ I and x0 are real numbers, a(·) and b(·) are
continuous functions from I to R, and A(·) is a primitive of a(·).
We want to solve explicitly the nonhomogeneous linear equation
x0 (t) = a(t)x(t) + b(t). (NH)
To do so, we first solve the homogeneous equation associated to (NH): 3

x0 (t) = a(t)x(t). (H)

Proposition 3.3.1. 1) The maximal solutions of (H) are the functions x : I → R given by
x(t) = λeA(t) (3.5)
with λ ∈ R. Thus, they are global and the set of these solutions is a real vector space of dimension
1.
2) The initial value problem x0 (t) = a(t)x(t) and x(t0 ) = x0 has a unique maximal solution:
Z t
x(t) = x0 exp( a(s)ds).
t0

Proof. 1) Let (J, x(·)) be a potential solution and let z(t) = e−A(t) x(t). Then, recalling a = A0 ,
x0 = ax ⇔ x0 − A0 x = 0 ⇔ e−A (x0 − A0 x) = 0 ⇔ z 0 = 0.
Thus, x(·) is solution if and only if there exists λ ∈ R such that, for all t in J, z(t) = λ, that
is, x(t) = λeA(t) , and since this defines a solution on the whole of I, maximal solutions are
global. Moerover the set of these solutions is a real vector space of dimension 1 generated by
the nonzero function eA(t) .
2) Let A(t) = tt0 a(s)ds. By 1), maximal solutions of (H) are global and given by (3.5), that
R

is x(t) = λeA(t) . Moreover, x(t0 ) = x0 if and only if λ = x0 .

Eq. (H) may be written as x0 (t) − a(t)x(t) = 0 and may be called in French “equation linéaire sans second
3

membre”. By contrast, Eq (NH) may be written as x0 (t) − a(t)x(t) = b(t) and is thus “avec second membre”.
3.3. LINEAR DIFFERENTIAL EQUATIONS 27

Remark 3.3.2. Eq. (H) may also be seen (and solved) as an equation with separated variables.
Proposition 3.3.3. 1) Maximal solutions of (NH) are global and they are given by Duhamel’s
formula:
 Z t  Z t Z t 
x(t) = λ + b(s)e−A(s) ds eA(t) = λeA(t) + b(s) exp a(u)du ds (DF)
t0 t0 s

where λ is a real number.

2) The initial value problem x0 (t) = a(t)x(t) + b(t) and x(t0 ) = x0 has a unique solution,
given by Duhamel’s formula with λ = x0 and A(t) = tt0 a(s)ds.
R

3) Let xp (·) be a solution of (NH). Then the solutions of (NH) are the functions x : I → R
of the form
x(t) = xp (t) + λeA(t)
That is, x(·) is solution of (NH) if and only if there exists a solution y(·) of (H) such that
x(·) = xp (·) + y(·). We say that the general solution of (NH) is the sum of a particular solution
of (NH) and of the general solution of the homogenous equation (H) and we write:

SN H = xp (·) + SH

where SN H and SH are the sets of solutions of (NH) and (H), respectively.
4) The solutions of (NH) form an affine subspace of dimension 1 whose direction is the set
of solutions of the homogeneous equation (H).

Proof. To prove 1), we start from the general solution of the associated homogeneous equation:
x(t) = λeA(t) and we let the constant λ vary! That is, letting J be a subinterval of I and
x : J → R be differentiable, we define λ(t) = x(t)e−A(t) so that

x(t) = λ(t)eA(t) (3.6)

and we look for conditions on λ(·) so that (3.6) is a solution of (NH): an example of change of
variables. We find that:

x0 = ax + b ⇔ λ0 eA + λaeA = aλeA + b ⇔ λ0 eA = b ⇔ λ0 = be−A .

Thus, letting t0 ∈ J, x(·) is solution if and only if

Z t
λ(t) = λ(t0 ) + b(s)e−A(s) ds ∀t ∈ J.
t0

Multiplying by eA(t) implies that x(·) is solution if and only if it is of the form (DF). Finally,
since (DF) is defined on the whole I, maximal solutions are global. Note that the constant λ
appearing in the statement of Duhamel’s formula corresponds to λ(t0 ) in the proof.
Proof of 2). Let A(t) = tt0 a(s)ds. Then by 1), solutions of (NH) are given by (DF).
R

Moreover, x(t0 ) = x0 if and only if λ = x0 .

Proof of 3) and 4). Let xp : I → R be solution of (NH), let x : I → R be differentiable and
let y(·) = x(·) − xp (·) so that x(·) = xp (·) + y(·). Then x0 = x0p + y 0 = axp + b + y 0 . Therefore:

x0 = ax + b ⇔ axp + b + y 0 = a(xp + y) + b ⇔ y 0 = ay.

This proves 3), and 3) implies 4).

Remark 3.3.4.
a) If we let x(·) be given by (DF) and add a solution y(·) = µeA(·) of (H), then z(·) = x(·)+y(·)
is still of the form (DF), with λ replaced by λ̃ = λ + µ. Thus Duhamel’s formula is coherent
with 3).
28 CHAPTER 3. EXPLICIT SOLUTIONS IN DIMENSION 1

b) You are allowed to used Duhamel’s formula directly, but you do not have to know it by heart
(the author of these lines does not know it). However, you should be able to find it by searching
solutions of the form x(t) = λ(t)eA(t) , as we did: a method called “variation of constants”.
c) The fact that (NH) has a unique maximal solution such that x(t0 ) = x0 also follows from
Theorem 1.5.4 (the strong version of Cauchy-Lipschitz).
d) The above results are valid for linear equations. An equation such as x0 (t) = a(t)x2 (t)+b(t)
is not linear, because of the term x2 (t). Thus, the above results do not say anything on this
equation.

Remark 3.3.5. To solve an initial value problem, it may be useful to proceed in two steps: first,
find all solutions of the underlying differential equation (without taking into account the initial
condition). This solution depends on one or several constants. Second, determine the value of
these constants in order to satisfy the initial condition.

Superposition principle A first way to solve (NH) is to apply (DF) or, equivalently, to
recover it from (3.6). A second way is to guess a particular solution xp (·) and then to add to
this solution the general solution of (H). We will see later how to guess particular solutions when
b(·) has a simple form (polynomial, exponential,...), however note that when b(·) is the sum of
two functions for which we know a particular solution, we may apply the following trick:

Proposition 3.3.6 (Superposition principle). If x1 (·) is solution of x0 = ax + b1 and x2 (·) of

x0 = ax + b2 , then x = x1 + x2 is solution of x0 = ax + (b1 + b2 ).

Proof. Under the assumptions of the proposition, x0 = x01 + x02 = (ax1 + b1 ) + (ax2 + b2 ) =
a(x1 + x2 ) + (b1 + b2 ) = ax + (b1 + b2 ).

We end this section with some worked-out exercises. Note that since we know that maximal
solutions of (NH) are global, we are only interested in the expression of x(t). If this were not
the case, we would need to specify the interval over which the solutions are defined.

Example 3.3.7. Solve (1) x0 (t) = −x(t)+t; (2) x0 (t) = −x(t)+et , and (3) x0 (t) = −x(t)+t+et .
Solution: The general solution of x0 (t) = −x(t) is x(t) = λe−t . Moreover, x1 (t) = αt + β is
solution of (1) if and only if α = −(αt + β) + t, hence if α = 1 and β = −1. Thus, a particular
solution of (1) is x1 (t) = t − 1 and the general solution of (1) is x(t) = t − 1 + λe−t .
Similarly, x2 (t) = λet is solution of (2) if and only if λ = −λ + 1, that is, λ = 1/2, thus
t t
x2 (t) = e2 is a particular solution of (2) and the general solution of (2) is x(t) = e2 + λe−t .
Finally, by the superposition principle, xp (t) = x1 (t) + x2 (t) is solution of (3), hence the
general solution of (3) is x(t) = t − 1 + 21 et + λe−t .

Example 3.3.8. Solve the initial value problem (IVP): x0 (t) = 2tx(t) and x(1) = 1.
Solution: a primitive of a(t) = 2t is A(t) = t2 , thus the general solution of x0 (t) = 2tx(t) is
2
x(t) = λet . It satisfies the initial condition x(1) = 1 if and only if λe = 1, that is, λ = 1/e.
2 2
Thus the solution of (IVP) is x(t) = 1e et = et −1 .

Example 3.3.9. Let a and b be real numbers (not functions!), with a 6= 0. Solve the initial
value problem (IVP): x0 (t) = ax(t) + b and x(t0 ) = x0 .
Solution 1: let us first solve x0 (t) = ax(t) +b, without the initial condition. A particular solution
is the stationary solution xp (t) = −b/a. The general solution of the associated homogeneous
equation is y(t) = λeat . Thus the general solution of x0 (t) = ax(t) + b is x(t) = − ab + λeat . It
satisfies x(t0 ) = x0 if and only if − ab + λeat0 = x0 , that is λ = (x0 + ab )e−at0 . Thus the solution
of (IVP) is
b b −at0 at b b a(t−t0 )
   
x(t) = − + x0 + e e = − + x0 + e .
a a a a
3.3. LINEAR DIFFERENTIAL EQUATIONS 29

Solution 2: the general solution of x0 (t) = ax(t) is x(t) = λeat , thus we look for a solution of
(IVP) of the form x(t) = λ(t)eat . We find that:

x0 (t) = ax(t) + b ⇔ λ0 (t)eat + aλ(t)eat = aλ(t)eat + b ⇔ λ0 (t) = be−at

Thus x(·) is solution of the equation if and only if there exists λ0 ∈ R such that
Z t
b  −as t b −at
λ(t) = λ0 + be−as ds = λ0 − e t0 = λ0 − a (e − e−at0 ) (3.7)
t0 a

Hence
b
 
x(t) = λ(t)e at
= λ0 − (e−at − e−at0 ) eat
a
Moreover it satisfies x(t0 ) = x0 if and only if λ0 eat0 = x0 , that is, λ0 = x0 e−at0 . Thus, we get
that the solution of (IVP) is

b b b b a(t−t0 )
 
x(t) = x0 ea(t−t0 ) − + ea(t−t0 ) = − + x0 + e .
a a a a

Solution 3: According to point 2) of Proposition 3.3.3, the solution is given by

 Z t 
x(t) = x0 + be−A(s) ds eA(t)
t0
Rt
with A(t) = t0 ads = a(t − t0 ). Thus,
 Z t  Z t 
x(t) = x0 + be−a(s−t0 ) ds ea(t−t0 ) = x0 ea(t−t0 ) + be−as ds eat .
t0 t0

Moreover, Z t
b  −as t b
be−as ds = − e t = − (e−at − e−at0 )
t0 a 0 a
Therefore
b −at b b a(t−t0 )
   
x(t) = x0 ea(t−t0 ) − (e − e−at0 ) eat = − + x0 + e .
a a a

Note that if a < 0, then for all solutions, x(t) → −b/a as t → +∞.
30 CHAPTER 3. EXPLICIT SOLUTIONS IN DIMENSION 1
Chapter 4

Comparison principles

Introduction. This chapter shows that we can learn a great deal by comparing the behavior
of solutions of equations, we do not know how to solve, to the behavior of solutions we can solve.
Let f : Ω → R, where Ω is a nonempty open in R × R, and consider the differential equation

x0 (t) = f (t, x(t)), (4.1)

A function u : J → R is a subsolution of (4.1) if, for all t in J, u0 (t) ≤ f (t, u(t)), and a
supersolution if, for all t in J, u0 (t) ≥ f (t, u(t)). Except for times t such that u(t) = v(t), the
fact that u is a subsolution and v a supersolution of (4.1) does not imply that u0 (t) ≤ v 0 (t).
Indeed:

Example 4.0.1. u(t) = et , x(t) = 0 and v(t) = −et are respectively subsolution, solution and
supersolution of x0 (t) = 2x(t). But at all times t in R, u0 (t) > x0 (t) > v 0 (t).

Nevertheless, we will prove that under some regularity conditions, if a subsolution starts below
a supersolution, then it remains below at all later times. We will use such comparison principles
as follows: assume that we do not know how to solve (4.1) but we know how to solve

x0 (t) = g(t, x(t)) (4.2)

for some function g ≥ f . Let u : J → R be solution of (4.1). Then for all t in J,

u0 (t) = f (t, u(t)) ≤ g(t, u(t)),

thus u is subsolution of (4.2). Therefore, under some regularity conditions, for all t ≥ t0 ,
u(t) ≤ v(t) where v is the solution of (4.2) such that u(t0 ) = v(t0 ). Since we assumed that we
can compute v explicitly, this provides an explicit upper bound for u, which allows to better
understand its qualitative behavior. Similarly, for t ≥ t0 , lower bounds for u may be obtained
by using that u is a supersolution of a differential equation.

4.1 Comparison principles

Conventions: I and J ⊂ I are nonempty open intervals of R, the function f : I × R → R is at

least continuous (we need at least existence of solutions!), t0 is an element of I, and all functions
are defined on I or a subinterval of I containing t0 . When we say that a function v is greater
than u for all t ≥ t0 , we mean that this holds for all times t ≥ t0 such that u(t) and v(t) are well
defined. Finally, when we consider subsolutions or supersolutions of a differential equation, we
implicitly assume that these functions are differentiable on the interior of the interval on which
they are defined.

31
32 CHAPTER 4. COMPARISON PRINCIPLES

Proposition 4.1.1 (Linear comparison principle). Let a(·) and b(·) be continuous. Let u be a
subsolution and v a supersolution of the linear equation x0 (t) = a(t)x(t) + b(t). If u(t0 ) ≤ v(t0 )
then u(t) ≤ v(t) for all t ≥ t0 .

Proof. Assume u(t0 ) ≤ v(t0 ). Then w = v − u satisfies w(t0 ) ≥ 0. Moreover:

w0 = v 0 − u0 ≥ (av + b) − (au + b) = a(v − u) = aw

hence w0 − aw ≥ 0. Therefore, if A(·) is a primitive of a(·) and z(t) = e−A(t) w(t):

z 0 (t) = e−A(t) (w0 (t) − a(t)w(t)) ≥ 0.

Thus, for all t ≥ t0 , e−A(t) w(t) ≥ e−A(t0 ) w(t0 ) ≥ 0, hence w(t) ≥ 0. That is, u(t) ≤ v(t).

Remark 4.1.2. A solution is both a subsolution and a supersolution, so Proposition 4.1.1

also allows to compare solutions to subsolutions (or solutions to supersolutions). Besides, the
condition u(t0 ) ≤ v(t0 ) is of course satisfied when u(t0 ) = v(t0 ).
Remark 4.1.3 (An important particular case, classically known as Gronwall’s Lemma). If
R t
a(s)ds
u0 (t) ≤
R a(t)u(t), then u(t) ≤ u(0)e
0 for all t ≥ 0. Indeed, the function defined by v(t) =
t
a(s)ds
u(0)e 0 is the unique solution (therefore also a supersolution) of x0 (t) = a(t)x(t), x(0) =
u(0). Since v(0) = u(0) ≥ u(0) we can apply the linear comparison principle to u(·) and v(·) to
obtain the thesis.
Exercise 4.1.4. Under the same assumptions of Proposition 4.1.1, show that if u(t0 ) = v(t0 ),
then u(t) ≤ v(t) for all t ≥ t0 , and u(t) ≥ v(t) for all t ≤ t0 . That is, in a weak sense, u and v
cross at t0 .
Proposition 4.1.5 (Comparison principle). Let f be C 1 . Let u be a subsolution and v a
supersolution of x0 (t) = f (t, x(t)). If u(t0 ) ≤ v(t0 ), then u(t) ≤ v(t) for all t ≥ t0 .

Proof. The proof relies on the linear comparison principle. Let w(t) = v(t) − u(t). Let
( f (t,v(t))−f (t,u(t))
v(t)−u(t) if u(t) 6= v(t)
a(t) = ∂f
(4.3)
∂x (t, u(t)) if u(t) = v(t).
Note that for all t, f (t, v(t)) − f (t, u(t)) = a(t)w(t). Therefore,

w0 (t) = v 0 (t) − u0 (t) ≥ f (t, v(t)) − f (t, u(t)) = a(t)w(t)

hence w is supersolution of x0 (t) = a(t)x(t). We claim that a(·) is continuous. Therefore, by the
linear comparison principle (with b(·) = 0), we get that for all t ≥ t0 , w(t) ≥ x(t) Rwhere x(·)
is the solution of x0 (t) = a(t)x(t) such that x(t0 ) = w(t0 ), that is, x(t) = w(t0 ) exp( tt0 a(s)ds).
Note that x(t0 ) ≥ 0 since w(t0 ) = v(t0 ) − u(t0 ) ≥ 0 by assumption. Therefore, for all t ≥ t0 ,
w(t) ≥ x(t) ≥ 0 hence v(t) ≥ u(t).
We still need to prove that a(·) is continuous at all times t∗ of its domain. If u(t∗ ) 6= v(t∗ ),
this is clear because in the neighborhood of t∗ , a(·) is given by (4.3). If u(t∗ ) = v(t∗ ), the proof
is as follows: for all t such that u(t) 6= v(t), a(t) is the average slope of f (t, ·) between u(t) and
v(t), thus there exists c(t) between u(t) and v(t) such that
∂f
a(t) = (t, c(t))
∂x
By (4.3), this is also the case if u(t) = v(t) (with c(t) = u(t)). Since c(t) is between u(t) and
v(t), and u(t∗ ) = v(t∗ ), it follows that c(t) → u(t∗ ) as t → t∗ . Since f is C 1 , this implies that:
∂f ∂f ∗
a(t) = (t, c(t)) →t→t∗ (t , u(t∗ )) = a(t∗ )
∂x ∂x
hence a(·) is continuous at t∗ . This concludes the proof.
4.1. COMPARISON PRINCIPLES 33

Remark 4.1.6. The assumption f of class C 1 is not fully needed: we only need that the partial
derivative with respect to the second variable (that is, x) is jointly continuous.
Exercise 4.1.7 (Variants). Under the same assumptions of Proposition 4.1.5, show that:
i) if u(t0 ) ≥ v(t0 ), then u(t) ≥ v(t) for all t ≤ t0 ;
ii) if u(t0 ) = v(t0 ), then u(t) ≤ v(t) for all t ≥ t0 and u(t) ≥ v(t) for all t ≤ t0 ;
iii) if u(t0 ) < v(t0 ), then u(t) < v(t) for all t ≥ t0 . Hint: introduce the solutions x(·) and
y(·) of x0 (t) = f (t, x(t)) such that, respectively, x(t0 ) = u(t0 ) and y(t0 ) = v(t0 ), and use both
Cauchy-Lipschitz Theorem and Proposition 4.1.5.

Proposition 4.1.5 implicitly requires that the functions u and v are differentiable1 , but often
we want to have information on functions that are only continuous. Thus, we would like to have
a version of the comparison principle for functions that are only continuous. To do so, recall
from the proof of the Cauchy-Lipschitz Theorem that the initial value problem
x0 (t) = f (t, x(t)) and x(t0 ) = x0 (4.4)
may be written in integral form:
Z t
x(t) = x0 + f (s, x(s)) ds (4.5)
t0

More precisely, the reader can check that (we already gave some details in the proof of Cauchy-
Lipschitz Theorem):
Claim 4.1.8. A function x : J → R is differentiable and solve (4.4) if and only if it is continuous
and a solution of (4.5).

The idea of the following result is to replace the assumption that u is subsolution of x0 (t) =
f (t, x(t)) by a similar condition on the integral form (4.5).
Proposition 4.1.9 (Integral form of the comparison principle). Let f : I × R → R be C 1 and
increasing in its second variable. Let u : J → R be continuous. If for all t ≥ t0 ,
Z t
u(t) ≤ K + f (s, u(s))ds (4.6)
t0

then for all t ≥ t0 , u(t) ≤ v(t) where v(·) is the solution of v 0 (t) = f (t, v(t)) such that v(t0 ) = K.
Rt
Proof. Let U (t) = K + t0 f (s, u(s)) ds. For all t ≥ t0 , U (t) ≥ u(t) and
U 0 (t) = f (t, u(t)) ≤ f (t, U (t))
since f is increasing in its second variable. Therefore, U is a subsolution of x0 (t) = f (t, x(t)).
Moreover, U (t0 ) = K = v(t0 ). Thus, by the comparison principle and by definition of v, for all
t ≥ t0 , U (t) ≤ v(t), hence u(t) ≤ v(t).
Remark 4.1.10. Taking t = t0 in (4.6) shows that, necessarily, K ≥ u(t0 ). In most applica-
tions, we actually take K = u(t0 ).

Note that, if u is subsolution of x0 (t) = f (t, x(t)), then u satisfies (4.6) for all K ≥ u(t0 ),
this proof is left to the reader. Thus, the assumptions of the integral form of the comparison
principle are stronger than those of Proposition 4.1.5 in two ways: first, for a differentiable
function, (4.6) is more demanding than u0 (t) ≤ f (t, u(t)); second, and more importantly, f must
be increasing in its second variable. The upside is that the integral form applies to continuous
functions u, while Proposition 4.1.5 does not.
Finally, a variant of the proof of Proposition 4.1.9 shows that the result still holds if f (t, x) =
a(t)x + b(t), with a(·) and b(·) continuous, and a(·) nonnegative. We thus get:2
1
Indeed, the definition of u subsolution and v supersolution requires that these functions are differentiable.
2
The reason why this is not a direct corollary of Proposition 4.1.9 is that f is not C 1 .
34 CHAPTER 4. COMPARISON PRINCIPLES

Proposition 4.1.11 (Integral form of the linear comparison principle). Let a(·) and b(·) be
continuous, with a(·) nonnegative. Let u : J → R be continuous. If for all t ≥ t0 ,
Z t
u(t) ≤ K + (a(s)u(s) + b(s))ds (4.7)
t0

then for all t ≥ t0 , u(t) ≤ v(t) where v is the solution of v 0 = av + b such that v(t0 ) = K.

Remark 4.1.12 (Integral form of the Rclassical Gronwall’s Lemma). In many applications, b(·) =
0. We then get: u(t) ≤ v(t) = K exp( tt0 a(s)ds).

Vocabulary: comparisons principles are also known as Gronwall’s Inequalities (or Gronwall’s
Lemmas), in honor of the Swedish mathematician Thomas Hakon Grönwall (1877-1932).

4.2 Applications to blow-up and to perturbation analysis

Explosion and non-explosion

A corollary of the integral form of the linear comparison principle is that if, asymptotically,
||F (t, X)|| grows less than linearly in ||X||, then all solutions of X 0 (t) = F (t, X(t)) are global:

Corollary 4.2.1. Let a(·) and b(·) be continuous functions from I to R, with a(·) nonnegative.
a) Let X : J → Rd be C 1 . If for all t ≥ t0 ,

||X 0 (t)|| ≤ a(t)||X(t)|| + b(t)

then for all t ≥ t0 , ||X(t)|| ≤ v(t) where v is solution of v 0 (t) = a(t)v(t) + b(t) and v(t0 ) =
||X(t0 )||.
b) Let F : I × Rd → Rd be C 1 or Lipschitz, and such that for all (t, X) ∈ I × Rd ,

||F (t, X)|| ≤ a(t)||X|| + b(t).

Then all solutions of X 0 (t) = F (t, X(t)) are global.

Rt
Proof. a) Since X(t) = X(t0 ) + t0 X 0 (s)ds, for all t ≥ t0 ,
Z t Z t
0
||X(t)|| ≤ ||X(t0 )|| + ||X (s)||ds ≤ ||X(t0 )|| + [a(s)||X(s)|| + b(s)]ds
t0 t0

Thus, the result follows from Proposition 4.1.11 applied to u(t) = ||X(t)|| and K = ||X(t0 )||.
b) Let X : J → Rd be solution of X 0 (t) = F (t, X(t)). For all t, ||X 0 (t)|| = ||F (t, X(t))|| ≤
a(t)||X(t)|| + b(t). Therefore, by a), for all t ≥ t0 , ||X(t)|| ≤ v(t), where v is solution of
v 0 (t) = a(t)v(t) + b(t) and v(t0 ) = ||X(t0 )||. But v(·) is global, thus the non-explosion criterion
(with g(t) = v(t)) implies that sup J = sup I. The proof that inf J = inf I is similar and we
omit it (the idea is to first prove an analog of a) backward in times, see the exercise below, and
then to apply the same reasoning).

Remark 4.2.2. Note that, if for all (t, X) ∈ I × Rd we have ||F (t, X)|| ≤ a(t)||X|| + b(t),
then a(t) is necessarily nonnegative for all t ∈ I. Indeed, suppose there exists t ∈ I such that
a(t) < 0, then taking the limit for kXk → +∞ we should have 0 ≤ limkXk→+∞ ||F (t, X)|| ≤ −∞
a contraddiciton.

Remark 4.2.3. Theorem 1.5.10 is a particular case of the above corollary (where a(·), b(·) are
taken constant).
4.2. APPLICATIONS TO BLOW-UP AND TO PERTURBATION ANALYSIS 35

Exercise 4.2.4. Under the same assumptions of point a) of Corollary 4.2.1, but for t ≤ t0 ,
show that for all t ≤ t0 , ||X(t)|| ≤ w(t) where w is solution of w0 (t) = −a(t)w(t) − b(t) and
w(t0 ) = ||X(t0 )||. Hint: let Y (t) = X(−t) and τ0 = −t0 ; apply a) to Y (·) to show that for all
t ≥ τ0 , ||Y (t)|| ≤ v(t) where v is solution of v 0 (t) = a(−t)v(t) + b(t) and v(τ0 ) = ||Y (τ0 )||; then
let w(t) = v(−t) and find the differential equation satisfied by w.

In dimension 1, comparison principles may also be used to show that solutions of some
differential equations explode.

Corollary 4.2.5 (Blow-up in dimension 1). Let f : R → R be continuous. Let x : J → R be a

maximal solution of x0 (t) = f (x(t)) such that x(t) → +∞ as t → sup J. If there exists α > 1,
and x̄ > 0 such that f (x) ≥ xα for all x ≥ x̄, then sup J < +∞.

Proof. Since x(t) → +∞ as t → sup J, there exists t0 ∈ J such that for all t ∈ [t0 , sup J[,
x(t) ≥ x̄, hence x0 (t) = f (x(t)) ≥ xα (t). Therefore, by the comparison principle, for all t ≥ t0
such that these functions are well defined, x(t) ≥ v(t) where v(·) is solution of v 0 (t) = v α (t) and
v(t0 ) = x(t0 ) ≥ x̄ > 0. But it follows from Remark 3.2.7 that v(·) goes to +∞ in finite time.
Therefore, sup J < +∞.3

Exercise 4.2.6. Prove an analogue of this result for non autonomous differential equations.

Exercise 4.2.7. Prove an analogue of Corollary 4.2.5 backward in time.

Example 4.2.8 (Forward in time blow-up). Let x0 > 1. Let (J, v(·)) be a solution of the initial
value problem x0 = x2 − 1 and x(0) = x0 . We have: sup J < +∞.

Proof. Let f (x) = x2 − 1. Since f (x0 ) > 0 and since there are no equilibria greater than x0 ,
it follows from Proposition 2.3.1 (on autonomous equations in dimension 1) that v(t) → +∞
as t → sup J. Moreover, in +∞, f (x) ∼ x2 , therefore, for all x large enough, f (x) ≥ xα for
α = 3/2 > 1. By Corollary 4.2.5, this implies that sup J < +∞.

Example 4.2.9 (Backward in time blow-up). Let x0 > 1. Let (J, v(·)) be the solution of the
logistic equation x0 = x(1 − x) and x(0) = x0 . We have: inf J > −∞.

Proof. Write J =]a, b[. Since we want to use Corollary 4.2.5, which is a result forward in time,
define y : J˜ =] − b, −a[→ R by y(t) = x(−t). We have:

y 0 (t) = −x0 (−t) = −x(−t)(1 − x(−t)) = −y(t)(1 − y(t)) = y 2 (t) − y(t)

and y(0) = x0 > 1. The same reasoning as in Example 4.2.8 shows that sup J˜ < +∞. Therefore
−a < +∞, hence a > −∞, that is, inf J > −∞.

Perturbation analysis.

Comparison principles are useful to show that a perturbed equation behaves asymptotically as
the underlying unperturbed equation.

Example 4.2.10 (Vanishing perturbation). Let (J, x(·)) be a solution of x0 (t) = −x(t) + ε(t)
with the perturbation term ε : R → R continuous and such that ε(t) → 0 as t → +∞. Show that
x(t) → 0 as t → +∞, as in the unperturbed equation x0 (t) = −x(t).
3
Remark 3.2.7 considers the equation x0 (t) = |x(t)|α , not x0 (t) = xα (t), but since we are interested in a solution
that is positive for all large enough times, this is not an issue.
36 CHAPTER 4. COMPARISON PRINCIPLES

Proof. First note that sup J = +∞ since the equation is linear. Now let η > 0. Since ε(t) → 0
as t → +∞, there exists T ∈ R such that, for all t ≥ T , |ε(t)| ≤ η, hence:

∀t ≥ T, −x(t) − η ≤ x0 (t) ≤ −x(t) + η (4.8)

Therefore, on [T, +∞[, x(·) is subsolution of x0 = −x + η and supersolution of x0 = −x − η.

Hence, by the linear comparison principle, for all t ≥ T ,

y(t) ≤ x(t) ≤ z(t) (4.9)

where y(·) is the solution of x0 = −x−η such that y(T ) = x(T ) and z(·) the solution of x0 = −x+η
such that z(T ) = x(T ). By Proposition 2.3.1 (on autonomous equations in dimension 1), in +∞,
y(t) → −η and z(t) → η. Thus, taking liminf and limsup in (4.9), we get:

−η ≤ lim inf x(t) ≤ lim sup x(t) ≤ +η

t→+∞ t→+∞

Since this holds for all η > 0, it follows that x(t) → 0 as t → +∞.

Exercise 4.2.11 (Bounded perturbation). Let (J, x(·)) be a solution of x0 (t) = −x(t)+ε(t) with
ε : R → R continuous and such that lim supt→+∞ |ε(t)| ≤ K. Show that lim supt→+∞ |x(t)| ≤ K.
Exercise 4.2.12 (Vanishing perturbation for a nonlinear equation). (∗) Let (J, x(·)) be a solu-
tion of x0 (t) = −x3 (t) + ε(t) with ε : R → R continuous and such that ε(t) → 0 as t → +∞.
Show that sup J = +∞ and x(t) → 0 as t → +∞. Hint in this note.4

In the above examples and exercises, all solutions of the unperturbed equation converge to
the same limit, irrespective of the initial condition. Otherwise, the method is still useful but
does not allow to pinpoint precisely the asymptotic behavior of x(t).
Exercise 4.2.13 (unperturbed equation with several possible limit points). (∗)
Let (J, x(·)) be solution of x0 (t) = x(t)(1 − x(t)) + ε(t) with ε : R → R continuous and such
that ε(t) → 0 as t → +∞. Let t0 ∈ J. We want to show that as t → sup J, x(t) has a finite
or infinite limit, which is −∞, 0 or 1, and that moreover, sup J < +∞ in the first case, and
sup J = +∞ in the two latter cases.
1. Assume x(t0 ) > 0. Why can’t we be sure that x(t) ≥ 0 for all t ≥ t0 ? More generally, why
knowing x(t0 ) is not helpful?
2a. Show that for all t in J, x0 (t) ≤ x(t) + ε(t).
2b. Let (JV , v) denote the solution of v 0 (t) = v(t) + ε(t) and v(t0 ) = x(t0 ). Show that Jv = R
and that for all t in [t0 , sup J[, x(t) ≤ v(t).
2c. Show that if sup J < +∞, x(t) → −∞ as t → sup J.
3. In questions 3), 4) and 5), we assume sup J = +∞. Let δ ∈]0, 1/4[.
3a. Show that there exists Tδ ≥ t0 such that for all t ≥ Tδ ,

x(t)(1 − x(t)) − δ ≤ x0 (t) ≤ x(t)(1 − x(t)) + δ

3b. Denote by x− +
δ and xδ the smallest and largest roots of x(1 − x) + δ, and give the phase
0
portrait of w (t) = w(t)(1 − w(t)) + δ.
−
3c. Show that as t → +∞, lim sup x(t) ≤ x+
δ , and then that lim sup x(t) ≤ 1 (hint: xδ → 1 as
δ → 0).
3d. Show that if lim inf x(t) < x−
δ , then x(t) → −∞ as t → +∞. Deduce that if lim inf x(t) < 0,
then x(t) → −∞ as t → +∞ (hint: x− δ → 0 as δ → 0).

The difficulty is that the equation is no longer linear. Note however that when x(t) ≥ 0, x0 (t) ≤ ε(t) and if
4

x(t) ≤ 0, then x0 (t) ≥ ε(t). Use this to prove that sup J = +∞. Then replace the linear comparison principle
with the comparison principle and Remark 4.1.6.
4.2. APPLICATIONS TO BLOW-UP AND TO PERTURBATION ANALYSIS 37

4a. Denote by yδ− and yδ+ the smallest and largest roots of x(1 − x) + δ, and give the phase
portrait of u0 (t) = u(t)(1 − u(t)) − δ.
4b. Show that if lim inf x(t) > yδ− , then lim inf x(t) ≥ yδ+ . Deduce that if lim inf x(t) > 0, then
lim inf x(t) ≥ 1.
5. Show that as t → +∞, x(t) → 0 or x(t) → 1
6. Show that if x(t) → −∞ as t → sup J, then sup J < +∞ and conclude.
38 CHAPTER 4. COMPARISON PRINCIPLES
Chapter 5

Systems of linear differential

equations

In Chapter 3, we saw how to solve linear differential equations in dimension 1. This chapter
goes a step further and studies systems of linear differential equations such as:
(
x0 (t) = tx(t) + 3y(t) + t2
y 0 (t) = et x(t) − 2ty(t) + 3

or more generally
 0
 x1 (t)
 = a11 (t)x1 (t) + . . . + a1d (t)xd (t) + b1 (t)
.. ..
 . .
 0
xd (t) = ad1 (t)x1 (t) + . . . + add (t)xd (t) + bd (t)

which may be written in matrix form:

X 0 (t) = A(t)X(t) + B(t) (NH)

where X(t) and B(t) are column vectors in Rd , and A(t) is a d × d square matrix. This
corresponds to the particular case of the equation X 0 (t) = F (t, X(t)) where F (t, X) = A(t)X +
B(t). We will mostly focus on the case where A(t) is a constant matrix, but we begin with some
general properties of Eq. (NH). As usual, I denotes a nonempty open interval.

Proposition 5.0.1. Let A : I → Md (R) and B : I → Rd be continuous. Let t0 ∈ I, X0 ∈ Rd .

Let SN H and SH denote respectively the set of solutions of the nonhomogeneous equation (NH)
and of the associated homogeneous equation

X 0 (t) = A(t)X(t) (H)

1. Equations (H) and (NH) have a unique solution such that X(t0 ) = X0 , and it is global.

2. SH is a d-dimensional real vector space, and the function ϕ : SH → Rd that associates to

a solution of (H) its value in t0 is an isomorphism.

3. Let (X1 (·), ..., Xd (·)) be a family of solutions of (H). The following assertions are equiva-
lent:
(a) (X1 (·), ..., Xd (·)) is a basis of SH
(b) (X1 (t0 ), ..., Xd (t0 )) is a basis of Rd .
(c) For all times t in I, (X1 (t), ..., Xd (t)) is a basis of Rd .

39
40 CHAPTER 5. SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS

4. The general solution of (NH) is the sum of a particular solution of (NH) and the general
solution of (H). That is, if Xp (·) ∈ SN H , then
SN H = Xp (·) + SH = {X : I → Rd , ∃Y ∈ SH , X = Xp + Y }.
Thus, SN H is a d-dimensional real affine space with direction SH .

Proof. 1) The fact that these initial value problems have a unique solution follows from the
strong form of Cauchy-Lipschitz theorem (Theorem 1.5.4)1 , and it is global by Corollary 4.2.1
with a(t) = ||A(t)|| and b(t) = ||B(t)||.
2) SH is nonempty by 1), and if X(·) and Y (·) are elements of SH and λ is a real number,
then Z(·) = λX(·) + Y (·) satisfies
Z 0 = λX 0 + Y 0 = λAX + AY = A(λX + Y ) = AZ
hence Z(·) ∈ SH . Therefore, SH is a real vector space. Moreover, let ϕ : SH → Rd be
the function that associates to a solution of (H) its value in t0 . Then ϕ is obviously linear
(ϕ(λX + Y ) = λX(t0 ) + Y (t0 ) = λϕ(X) + ϕ(Y ) for all λ ∈ R and X, Y ∈ SH ), and it is one-to-
one - that is, bijective - by 1). Indeed, 1) implies that for any element X0 of Rd , there is exactly
one element of SH with image X0 . Therefore, ϕ is an isomorphism, hence SH has dimension d
as Rd .
3) The equivalence between (a) and (b) follows from the fact that ϕ is an isomorphism.
Moreover (c) implies (b), which implies (a), and (a) implies (c) since the implication (a) ⇒ (b)
is valid for all t0 in I.
4) The proof is the same as in dimension 1.

Remark 5.0.2. The fact that SH is a d-dimensional vector space, implies that every solution
X(·) of (H) is a linear combination of d linearly independent solutions X1 (·), . . . , Xd (·) of (H),
i.e.
X(·) = λ1 X1 (·) + · · · + λd Xd (·)
for some λ1 , . . . λd ∈ R. X1 (·), . . . , Xd (·) form a basis for SH .
Remark 5.0.3. From 3), it follows that d solutions X1 (·), . . . , Xd (·) of (H) are linearly inde-
pendent if and only if for all t ∈ I the vectors
X1 (t), . . . , Xd (t) ∈ Rd
are linearly independent.

As in dimension 1, the superposition principle may help us to find particular solutions. The
proof is the same as is dimension 1.
Proposition 5.0.4. Let A : I → Md (R), and let B1 , B2 , X and Y go from I to Rd . If
X 0 = AX + B1 and Y 0 = AY + B2 , then Z = X + Y satisfies Z 0 = AZ + (B1 + B2 ).

We now focus on the case of constant coefficients, that is, when the matrix A(t) does not
depend on t. We first study the homogeneous equation.

5.1 The equation X 0 (t) = AX(t)

We denote by (H) the equation

X 0 (t) = AX(t) (H)
1
Indeed, the function defined by F (t, X) = A(t)X + B(t) is continuous. Moreover, if C is a compact subset
of I × Rd , then its projection CI on I is also compact, and for any (t, X, Y ) such that (t, X) and (t, Y ) are in C,
||F (t, X) − F (t, Y )|| ≤ ||A(t)|| × ||X − Y || ≤ K||X − Y || where K = maxt∈CI ||A(t)||. Thus F is locally Lipschitz
with respect to X.
5.1. THE EQUATION X 0 (T ) = AX(T ) 41

5.1.1 The case A diagonalizable in R

Explicit solutions.

Assume first that d = 2. If A is diagonal:

!
λ1 0
A=
0 λ2

then denoting solutions by X(t) = (x(t), y(t))T , the system reads:

(
x0 (t) = λ1 x(t)
. (5.1)
y 0 (t) = λ2 y(t)

Hence the system decouples into two equations that can be solved independently. The solutions
are: (
x(t) = x0 eλ1 t
(5.2)
y(t) = y0 eλ2 t
which, thinking to what comes later, can also be written as
! !
λ1 t 1 λ2 t 0
X(t) = x0 e + y0 e .
0 1

Similarly, for a generic d, if A = diag(λ1 , ..., λd ), then the d equations may be solved indepen-
dently. The ith line leads to xi (t) = µi eλi t and the general solution of the system is:

µ 1 e λ1 t
 
d
X(t) = 
 ..  X
= µi eλi t Vi (5.3)
. 
i=1
µ d e λd t

where Vi is the ith vector of the canonical basis and µi the ith coordinate of X(0) in this basis.
When A is diagonalizable, we get exactly the same formula, but in the eigenvector basis.

Proposition 5.1.1. Let B = (W1 , ..., Wd ) be a basis of eigenvectors of A, associated to real

eigenvalues λ1 ,...,λd . Let Xi (t) = eλi t Wi . Then (X1 (·), ..., Xd (·)) is a basis for the set of
solutions of (H). Thus, the general solution is
d
X
X(t) = µB λi t
i e Wi , (µB B
1 , ..., µd ) ∈ R
d

i=1

Moreover, µB
i is the i
th coordinate of X(0) in the basis B.

Proof. The proof is based on the following fundamental Lemma:

Lemma 5.1.2. Let V be a real or complex eigenvector of A associated to the real or complex
eigenvalue λ. Then X(t) = eλt V satisfies (H) and X(0) = V .

Proof. X 0 (t) = λeλt V = eλt λV = eλt AV = Aeλt V = AX(t), and X(0) = V .2

We now prove the proposition. By Lemma 5.1.2, Xi (·) is solution. Moreover, (X1 (0), ..., Xd (0)) =
(W1 , ..., Wd ) is a basis of Rd . Therefore, by point 3) in Proposition 5.0.1, (X1 (·), ..., Xd (·)) is a
basis of the set SH of solutions of (H). Finally, to see that µB i is the i
th coordinate of X(0) in

the basis B, take t = 0 in the formula.

2
We let the reader check that when λ ∈ C\R, the derivative of the application C → C defined by t → eλt is
still given by t → λeλt , so that the previous computation is correct.
42 CHAPTER 5. SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS

We may also adopt a matricial point of view to get the same proof. If A is diagonalizable,
there exists an invertible matrix P such that A = P −1 DP and D = diag(λ1 , ..., λd ). Let
Y = P X. Then

X 0 = AX ⇔ X 0 = P −1 DP X ⇔ P X 0 = DP X ⇔ Y 0 = DY.
Pd λi t V ,
From (5.3) we get Y (t) = i=1 µi e i where Vi is the ith vector of the canonical basis, hence
d
−1
X
X(t) = P Y (t) = µ i e λi t W i
i=1

with Wi = P −1 Vi . Note that Wi is an eigenvector of A associated to λi . Indeed,

AWi = P −1 DP P −1 Vi = P −1 DVi = P −1 λi Vi = λi Wi

Thus, we obtained exactly the same formula as in Proposition 5.1.1. Note that the matricial
point of view is computationally more expensive: indeed, it requires, among other things, the
computation of P −1 , whose columns are the eigenvectors Wi . However, once we know these
eigenvectors, we can directly apply Proposition 5.1.1.

5.1.2 Phase portraits for A diagonalizable in R.

General Introduction: trajectories. Let X : J → Rd be a solution of a differential

equation, modeling a system whose state at time t is X(t). Recall that the trajectory (also
called orbit) associated to this solution is the set T = {X(t), t ∈ J} of successive states of
the system as t describes J (both forward and backward in times).3 The trajectory may also be
defined as the projection of the graph of X(·) on Rd . Indeed, the projection of (t, X(t)) ∈ R × Rd
on Rd is X(t). Thus the projection of the graph Γ = {(t, X(t)), t ∈ J} is {X(t), t ∈ J} = T .
We would like to have an idea of trajectories associated to solutions of (H). We first note
that, as seen in Proposition 2.1.5, for a C 1 or Lipschitz autonomous equation, the trajectory
associated to a solution satisfying an initial condition X(t0 ) = X0 depends only on the initial
position X0 , not on the initial time t0 . Thus, there are several solutions, but a unique trajectory
going through X0 . (As seen in Chapter 2, if X(·) and Y (·) are two solutions taking the same
value X0 at different times, then Y (·) is just a translation in time of X(·), so both solutions visit
exactly the same states, one following the other.4 )
It follows that, it is in principle possible to draw the trajectories associated to the solutions
of an autonomous equation (though for practical reasons, we will only do it in dimension 2!).
Such a drawing is called phase portrait - or phase line in dimension 1. For a non autonomous
equation, there are typically several trajectories going through the same point X0 , possibly an
infinity, so such a drawing would be a mess.
Let us give a few examples of trajectories in dimension 1 (not only in the linear case).
Consider the equation x0 = x. Since 0 is an equilibrium, any solution taking the value 0 is
stationary and its trajectory is simply {0}. The solution such that x(0) = 1 is x(t) = et , defined
on R. The associated trajectory is T = {et , t ∈ R} =]0, +∞[. The solution such that x(0) = −1
is −et . The associated trajectory is {−et , t ∈ R} =] − ∞, 0[. We let the reader check that
similarly, for any x0 > 0, the unique trajectory going through x0 is ]0, +∞[, and for any x0 < 0,
3
You may also see the trajectory as the set of positions occupied successively by a particle whose position at
time t is X(t).
4
Formally, if X :]a, b[→ R is the solution of X 0 (t) = G(X(t)) such that X(0) = X0 , then the solution such that
X(t0 ) = X0 is the function Y :]a + t0 , b + t0 [→ Rd defined by Y (t) = X(t − t0 ). The trajectory TY associated to
this solution is

TY = {Y (t), t ∈]a+t0 , b+t0 [} = {X(t−t0 ), t ∈]a+t0 , b+t0 [} = {X(t−t0 ), t−t0 ∈]a, b[} = {X(τ ), τ ∈]a, b[} = TX .

Thus this trajectory does not depend on t0

5.1. THE EQUATION X 0 (T ) = AX(T ) 43

the unique trajectory going through x0 is ] − ∞, 0[. Note that, as it should be, trajectories
associated to two different solutions are either equal or disjoint. Note also that for x0 6= 0,
solutions lead away from the equilibrium; that is, as time unfolds, the trajectories ]0, +∞[ and
] − ∞, 0[ are described from 0 to ±∞. This is what the arrows indicate when we draw the phase
line of this equation.

Exercise 5.1.3. Draw the graph of three representative solutions of x0 = x (in the plane with
an horizontal time axis and a vertical position axis), and draw arrows on these graphs indicating
the direction of time. Project these graphs - including the arrow - on the position axis. Check
that this gives you the phase line of x0 = x, but drawn vertically.

Exercise 5.1.4. Show that the trajectories for x0 = −x are the same as for x0 = x, but traveled
in opposite directions. Do the same exercise as above for this equation.

Now consider the logistic equation x0 = x(1−x). Since 0 and 1 are equilibria, the trajectories
going through 0 and 1 are respectively {0} and {1}. If a solution takes the value x0 ∈]0, 1[, then
in the sense of Proposition 2.3.1, it comes from 0 and goes to 1, taking all the intermediate
values by continuity, but not the values 0 and 1. So the associated trajectory is ]0, 1[. Similarly,
solutions starting at x0 > 1 come from +∞ and go to 1, and the associated trajectory is ]1, +∞[,
traveled from +∞ toward 1, as indicated by the arrow pointing towards 1 on the phase line.
Finally, the trajectory going through x0 < 0 is ] − ∞, 0[, traveled from 0 to −∞.
Here are examples in dimension 2. Consider first the function defined by X(t) = (x(t), y(t))
where x(t) = cos t and y(t) = sin t. It is solution of x0 = −y and y 0 = x. As t describes R, X(t)
describes the unit circle of R2 , that is, U = {(x, y) ∈ R2 , x2 + y 2 = 1}. Thus the trajectory
associated to X(·) is U . Note that the graph of X(t) would be an helicoid in R × R2 = R3 , but
its trajectory (projection of the graph on R2 ) is a circle.
Now consider the function X(t) = (et cos t, et sin t) = et (cos t, sin t). As t increases, X(t)
does not only rotate counter-clockwise, it also goes away from 0. Indeed, ||X(t)|| = et increases.
The graph would be an helicoid with circles of increasing length, and the trajectory is a spiral
leading away from (0, 0). Similarly, for X(t) = (e−t cos t, e−t sin t), the graph is a spiral but now
leading towards (0, 0).
As a last example, consider the function X(t) = (x(t), y(t)) where x(t) = et and y(t) = e−t .
Note that y(t) = 1/x(t), so that X(t) = (x(t), 1/x(t)). As t describes R, x(t) describes ]0, +∞[,
so the trajectory associated to X(·) is

T = {(x(t), y(t)), t ∈ R} = {(x(t), 1/x(t)), t ∈ R} = {(x, 1/x), x ∈]0, +∞[}

Thus, T is the graph of the function x → 1/x restricted to ]0, +∞[, hence a hyperbola.

Exercise 5.1.5. Show that the trajectory associated to X(t) = (e−t , et ) is also {(x, 1/x), x ∈
]0, +∞[}.

Phase portraits. We now come back to Eq. (H) in dimension 2. Recall that when A is
diagonal: !
λ1 0
A=
0 λ2
the solutions are given by (
x(t) = x0 eλ1 t
(5.4)
y(t) = y0 eλ2 t
For any value of the eigenvalues, there is an equilibrium at the origin, and solutions that start
on one of the axis remain on this axis. But for x0 6= 0, whether x(t) goes from 0 to ±∞, x(t)
is constant, or x(t) goes from ±∞ to 0, depends on the sign of λ1 . Similarly, the fate of y(t)
44 CHAPTER 5. SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS

depends on the sign of λ2 . We thus have to consider all the possible cases.

The case λ1 < 0 < λ2 .

A diagonal. As mentioned above, solutions that start on the x-axis (y0 = 0) remain on the
x-axis (y(t) = 0 for all t). Moreover, if x0 6= 0, then x(t) comes from +∞ or −∞ (depending on
the sign of x0 ) and goes to zero, since λ1 < 0. Thus, all solutions starting on the x-axis converge
to the equilibrium (0, 0). The x-axis is called the stable line. Similarly, solutions that start on
the y-axis remain on the y-axis, but, except if they start at the equilibrium, they tend away
from it: the y-axis is called the unstable line.
If x0 6= 0, y0 6= 0, then the sign of x(t) and y(t) remains constant; moreover, x(t) → 0 and
|y(t)| → +∞ as t → +∞, while |x(t)| → +∞ and y(t) → 0 as t → −∞. Thus all solutions
that do not start on the stable or unstable line come from the stable line and go towards the
unstable line. Their trajectories look like portions of hyperbolas. We say that the equilibrium
is a saddle.
We may be more precise (though we will never ask you to!). Assume for concreteness x0 > 0
and y0 > 0 (the other cases are similar). From x(t) = x0 eλ1 t , we get t = λ11 ln x(t)
x0 , hence

 λ2
 
λ2 x(t) x(t)
  
λ1
y(t) = y0 eλ2 t = y0 exp ln = y0 exp ln  = Cxγ (t) (5.5)
λ1 x0 x0

λ /λ1
with C = y0 /x0 2 > 0 and γ = λ2 /λ1 < 0. The trajectory associated to this solution is thus

T = {(x(t), y(t)), t ∈ R} = {(x(t), Cxγ (t), t ∈ R} = {(x, Cxγ ), x ∈]0, +∞[}

since x(t) describes ]0, +∞[ as t describes R. This trajectory is thus the graph of the function
x → Cxγ restricted to ]0, +∞[. Since γ < 0, it indeed looks like a branch of hyperbola (it is a
branch of hyperbola when γ = −1).
If y0 < 0, x0 > 0, we get the same formula but now C < 0. If x0 < 0, we get

T = {(x, C|x|γ ), x ∈] − ∞, 0[}

where C = y0 /|x0 |λ2 /λ1 has the sign of y0 . Thus, in all cases, the trajectory is a portion of the
graph of x → C|x|γ , the portion in ]0; +∞[ if x0 > 0, and the portion in ] − ∞; 0[ if x0 < 0, and
with C of the sign of y0 .
These observations allow to draw the phase portrait of the equation: that is, a drawing of
representative trajectories and the direction in which they are traveled. This drawing is made
in the phase plane R2 (for d ≥ 3, Rd is called the phase space). See Fig. 5.1.

A diagonalizable. Now consider the case where A has still two eigenvalues λ1 , λ2 such that
λ1 < 0 < λ2 but is not diagonal. Because A has two distinct real eigenvalues, it is diagonalizable
in R. Let W1 and W2 be eigenvectors associated to λ1 and λ2 , respectively. We know from
Proposition 5.1.1 that the general solution of (H) is

X(t) = xB
0e
λ1 t
W1 + y0B eλ2 t W2 (5.6)

where xB B
0 and y0 are the coordinates of X(0) in the eigenvector basis B = (W1 , W2 ):

X(0) = xB B
0 W 1 + y0 W 2

The coordinates of the solution in the eigenvector basis are thus (xB (t), y B (t)) = (xB
0e
λ1 t , y B eλ2 t ).
0
These are exactly the same equations as (5.4) but for the coordinates in the basis (W1 , W2 ).
We thus obtain the same kind of behavior as in the case A diagonal, except that the stable line
5.1. THE EQUATION X 0 (T ) = AX(T ) 45

Figure 5.1: A diagonal, λ1 < 0 < λ2 Figure 5.2: A diagonal, λ2 < 0 < λ1
   
−1 0 1 0
Two saddles with A diagonal, respectively A1 = and A2 = . Red arrows indicate
0 1 0 −1
the direction of movement at each point in space, blue curves are trajectories. Left, the stable line is
the x-axis and the unstable line the y-axis. Right, this is the opposite. There are actually 9 kinds of
trajectories (depending on the sign of x0 and y0 ), but the picture shows only the four cases corresponding
to x0 6= 0 and y0 6= 0. When drawing the picture by hand, you do not need to put the red arrows, but
you should indicate the other kinds of trajectories (the equilibrium and the trajectories on the axis, which
are opened half-lines), and put arrows on the blue curves. See drawings made in the course.

is now the span of W1 and the unstable line the span of W2 . In particular, all solutions that
do not start on one of these lines come from the stable line and go towards the unstable line.
Examples of phase portraits in this case will be given in the course.

Exercise 5.1.6. Sketch the phase portrait if λ1 < 0 < λ2 and: a) W1 = (0, 1)T , W2 = (1, 0)T ;
b) W1 = (1, 0)T and W2 = (1, 1)T .

The case λ2 < 0 < λ1 .

The phase portrait is the same as in the case λ1 < 0 < λ2 . The stable line is now the y-axis (or
more generally span(W2 ) and the unstable line the x-axis (more generally, span(W1 )). See Fig.
5.2 for the case A diagonal.

The case λ1 < 0, λ2 < 0.

A diagonal. If A is diagonal with eigenvalues λ1 < 0, λ2 < 0, then all solutions tend towards
the equilibrium (0, 0). We say that the equilibrium is a sink ( “puits” in French, though “sink”
means “évier”).
The way solutions tend toward the equilibrium depends on which eigenvalue is the strongest,
in the sense of having the largest absolute value. If the strongest eigenvalue is λ2 , that is, if
λ2 < λ1 < 0, then
y(t) y0 eλ2 t y0 (λ2 −λ1 )t
= λ t
= e →t→+∞ 0
x(t) x0 e 1 x0
and solutions tend toward (0, 0) tangentially to the x-axis. As t → −∞, |y(t)|/|x(t)| → +∞,
hence backward in time, solutions have asymptotically the direction of the y-axis, as in a
46 CHAPTER 5. SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS

Figure 5.3: λ2 > λ1 > 0 Figure 5.4: λ1 = λ2 > 0 Figure 5.5: λ1 > λ2 > 0
Three sources with A diagonal. Only trajectories with x0 6= 0, y0 6= 0 are shown. The matrices used are:
     
1 0 2 0 2 0
A3 = A4 = A5 =
0 2 0 2 0 1

parabola. By contrast, if λ1 < λ2 < 0, that is,pif λ1 is the strongest eigenvalue, then tra-
jectories look like portions of the graph of x → C |x|. If λ1 = λ2 , then typical trajectories are
opened half-lines.
To check that solutions really behave in that way, we may compute y(t) as a function of
x(t). In the case x0 > 0, y0 > 0, we find as in (5.5) that along the trajectory, y = Cxγ with
γ = λ2 /λ1 , but now γ > 0. If γ > 1, the trajectory looks like a portion of parabola; if γ = 1, this
√
is a portion of straight line, and if γ < 1, it looks like a portion of the graph of x → x. Similar
computations for the other possible signs of x0 and y0 lead to phase portraits as in Figures 5.3
to 5.5, but reversing all the arrows [see drawings made in the course].
A diagonalizable. If A is diagonalizable, then the solutions are given by (5.6). Since we assume
that the eigenvalues λ1 and λ2 are both negative, solutions behaves as we just discussed, except
that the x-axis is replaced by span(W1 ) and the y-axis by span(W2 ). Thus, if λ2 < λ1 < 0,
solutions tend towards the origin tangentially to the direction of W1 and in backward time, go
to infinity with a direction tending towards the direction of W2 . If λ1 < λ2 < 0, the reverse
thing happens [see drawings made in the course].

Exercise 5.1.7. Why didn’t we discuss the case A diagonalizable with λ1 = λ2 ? Answer here.5

The case λ1 > 0, λ2 > 0. Solutions now go away from the equilibria: we say that the
equilibrium is a source. Moreover, it may be seen (see this note6 ) that this case is the exact
“opposite” of the case λ1 < 0, λ2 < 0 in the following sense: if λ2 > λ1 > 0, the phase portrait is
exactly the same as in the case λ2 < λ1 < 0, except that we need to reverse the arrows, that is,
the trajectories are the same but travelled in opposite directions. Similarly, the case λ1 > λ2 > 0
is the opposite of the case λ1 < λ2 < 0, and the case λ1 = λ2 > 0 is the opposite of the case
λ1 = λ2 < 0. This leads to Fig. 5.3 to 5.5.

5
Because if A is diagonalizable with a single eigenvalue λ, then A is diagonal! Indeed, P −1 AP = λI gives
A = P λIP −1 = λI.
6
If X :]a, b[→ Rd is solution of X 0 = G(X), then the function Y :] − b, −a[→ Rd defined by Y (t) = X(−t)
satisfies for all t ∈] − b, −a[: Y 0 (t) = −X 0 (−t) = −G(X(−t)) = −G(Y (t)), hence Y is solution of Y 0 = −G(Y ),
and the reader will check that it is maximal if and only if X(·) is maximal. Since X(·) and Y (·) visit the same
points in space, but at opposite times, it is easily seen that they define the same trajectory:

TY = {Y (t), t ∈] − b, −a[} = {X(−t), t ∈] − b, −a[} = {X(−t), −t ∈] − b, −a[} = {X(τ ), τ ∈]a, b[} = TX

but traveled in opposite directions.

5.1. THE EQUATION X 0 (T ) = AX(T ) 47

The cases λ1 6= 0, λ2 = 0 and λ1 = λ2 = 0. If λ1 6= 0, λ2 = 0, then there is a line of

equilibria: the span of W2 . We let the reader check that if λ1 < 0, λ2 = 0, then all solutions
converge to this line of equilibria following trajectories parallel to the span of W1 , while if λ1 > 0,
λ2 = 0, solutions go away from the line of equilibria, again with trajectories parallel to the span
of W1 . In the even more special case λ1 = λ2 = 0, any point is an equilibrium, and solutions do
not move.

5.1.3 The case A diagonalizable in C

The case d = 2.

Proposition 5.1.8. Let A ∈ M2 (R) be diagonalizable in C. Let W be a complex eigenvector of

A associated to the eigenvalue λ. Let Z(t) = eλt W . Then the real and imaginary part of Z(t)
form a basis of SH . Thus, the general solution is

X(t) = µ1 Re(Z(t)) + µ2 Im(Z(t)), (µ1 , µ2 ) ∈ R2

Proof. From Lemma 5.1.2, we know that Z 0 = AZ. Denoting by Xz (t) and Yz (t) the real and
imaginary part of Z(t), this implies that Xz0 + iYz0 = A(Xz + iYz ) = AXz + iAYz . Therefore,
Xz0 = AXz and Yz0 = AYz . Thus, Xz (·) and Yz (·) are real solutions of (H). Moreover, we claim
that, (Xz (0), Yz (0)) is a basis of R2 . By point 3 in Proposition 5.0.1, it follows that (Xz (·), Yz (·))
is a basis of SH , proving the proposition.
It remains to prove the claim. In what follows, C2 is thought of as a complex vector space.
Let W = Xw + iYw . Since A is a real matrix, W̄ = Xw − iYw is another eigenvector of A, and
(W, W̄ ) is a basis of C2 (indeed, an eigenvector basis). Since both, W and W̄ are complex linear
combinations of Xw and Yw , it follows that (Xw , Yw ) generates C2 , hence is a basis of C2 . Thus,
Xw and Yw are independent in C2 : for all (µ1 , µ2 ) in C2 , µ1 Xw + µ2 Yw = 0 ⇒ µ1 = µ2 = 0.
This holds a fortiori for all (µ1 , µ2 ) in R2 . Therefore, Xw and Yw are independent in R2 , hence
(Xw , Yw ) is a basis of R2 . But Z(0) = W hence (Xz (0), Yz (0)) = (Xw , Yw ), hence (Xz (0), Yz (0))
is a basis of R2 .

Let us describe more precisely how solutions behave. Let W = Xw + iYw be an eigenvector
of A associated to λ = a + ib. Let Z(t) = e(a+ib)t W = eat (cos bt + i sin bt)(Xw + iYw ). We have:

Z(t) = eat (cos bt + i sin bt)(Xw + iYw ) = eat [(cos bt Xw − sin bt Yw ) + i(cos bt Yw + sin bt Xw )] .

Thus, the general solution is

X(t) = µ1 eat (cos bt Xw − sin bt Yw ) + µ2 eat (cos bt Yw + sin bt Xw ).

Let X B (t) = (X B (t), Y B (t))T denote the vector of coordinates of X(t) in the basis B = (Xw , Yw ).
Grouping the terms in Xw and in Yw in the previous equation, we obtain:
! ! !
B at µ1 cos bt + µ2 sin bt at cos bt sin bt µ1
X (t) = e =e
−µ1 sin bt + µ2 cos bt − sin bt cos bt µ2
! !
µ1 cos t − sin t
Note that X B (0) = . Therefore, letting M (t) = , we obtain:
µ2 sin t cos t

X B (t) = eat M (−bt)X B (0) (5.7)

As you may know, M (t) is the rotation matrix of angle t, hence M (−bt) the rotation matrix
of angle −bt. Therefore, the interpretation is as follows: in the basis B, the coordinates of the
solution are obtained from their value at time 0 by combining a dilatation of factor eat and a
rotation of angle −bt. When a = 0, they describe a circle. When a > 0, an expanding spiral.
48 CHAPTER 5. SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS

Figure 5.6: A spiral source. Figure 5.7: A center Figure 5.8: A spiral sink
The case A diagonalizable in C when B = (Xw , Yw ) is the canonical basis. The matrices used are:
     
1 −2 0 1 −1 −2
A6 = A7 = A8 =
2 1 −1 0 2 −1

Note that A8 is not the opposite of A6 , otherwise trajectories would be the same (with arrows reversed).

When a < 0, a contracting spiral. Whether the rotation is clockwise or counterclockwise de-
pends on the sign of b (note that b 6= 0, since λ = a + ib ∈
/ R).

Phase portraits. We can now draw the phase portraits. There are two cases.

Case 1 : if B = (Xw , Yw ) is the canonical basis of R2 (that is, if W = (1, 0)T + i(0, 1)T = (1, i)).
In that case, the coordinates in the basis B are just the standard coordinates, so if a = 0,
X(t) describes a circle, if a > 0, an expanding spiral, if a < 0 a contracting spiral. We then say
that the equilibrium (0, 0) is respectively a center, a spiral source and a spiral sink. See Fig. 5.6
to 5.8.
Case 2 : if B is not the canonical basis, then X(t) describes what could be called a “circle in
the basis B”, that is, a set defined by

C = {xXw + yYw , x2 + y 2 = r}.

If Xw and Yw are orthogonal, this is an ellipse whose axis have the directions of Xw and Yw .
Otherwise, this is a curve which is similar to an ellipse (whose axis no longer have the directions
of Xw and Yw ), but the author of these lines did not check whether it is precisely an ellipse or not
(the willing student can check!). Thus, when a = 0, solutions describes ellipses or similar curves.
When a > 0 and a < 0, solutions still describe expanding or contracting spirals, respectively,
but these spirals are “ellipsoidal”. Again, we say that the equilibrium is a center, a spiral source
or a spiral sink, respectively. The phase portraits will be drawn in the course.

Remark 5.1.9. Matricially speaking, letting

!
a b
Aref = , (5.8)
−b a

the first case corresponds to A = Aref and the second case to A = P −1 Aref P , where P is the
matrix whose ith column give the coordinates of ith vector in the canonical basis B = (Xw , Yw ).
Indeed, the real and imaginary parts of the equality A(Xw + iYw ) = (a + ib)(Xw + iYw ) imply
that the endomorphism with matrix A in the canonical basis has matrix Aref in the basis B.
5.1. THE EQUATION X 0 (T ) = AX(T ) 49

The case A diagonalizable in C in dimension d.

Summary: The case A diagonalizable in C in dimension d is similar to the case d = 2. Each
real eigenvector leads to a real solution of X 0 = AX and each complex eigenvector to a complex
solution. The real and imaginary parts of this complex solution are real solutions. If we group
the real solutions associated in this way to the real and complex eigenvectors, we again obtain
a basis of the set of real solutions of X 0 = AX, hence a formula for the general solution.
Let us be a bit more precise. If A is diagonalizable in C, then it has a basis of eigenvectors
consisting of q pairs of complex conjugate eigenvectors (Wi , W̄i ), and possibly d − 2q real eigen-
vectors W2q+1 , ..., Wd . Due to Lemma 5.1.2, we may associate to these eigenvectors a family of
d complex solutions, of which the last d − 2q are real.

(Z1 (·), Z̄1 (·), ..., Zq (·), Z̄q (·), Z2q+1 (·), ..., Zd (·))

Denoting by Xk (·) and Yk (·) the real and imaginary part of Zk (·), for 1 ≤ k ≤ q, we obtain as
in the case d = 2 that Xk (·) and Yk (·) are real solutions. Thus we obtain a family BH of d real
solutions:
BH = (X1 (·), Y1 (·), ..., Xq (·), Yq (·), Z2q+1 (·), ..., Zd (·))
It may be shown, as in the case d = 2, that taking the values of these solutions at t = 0 gives
a basis of Rd . This shows by Proposition 5.0.1 that BH is a basis of SH . Thus, solutions are
linear combinations of these basic solutions.
Matricially, A may be written in the form P −1 Mref P where Mref is a block-diagonal matrix
such that each block is either a 2 × 2 block of form (5.8) - corresponding to a pair λ = a + ib,
λ̄ = a−ib of complex eigenvalues - or a 1×1 block containing a real eigenvalue. Each of this block
corresponds to an invariant plane or line, on which the solutions behaves as we described for the
invariant planes, and as on the eigenaxis in the case A diagonalizable in R for the invariant lines.
This is somewhat difficult to visualize, except in dimension 3, where when A is diagonalizable
in C but not in R, Mref necessarily takes the form:
 
a b 0
 −b a 0 
 
0 0 λ

Exercise 5.1.10. (∗) Sketch the phase portrait for the equation X 0 = Mref X, when Mref
is the above 3 × 3 matrix. Note that there are 9 such phase portraits (three choices for the
sign of a, three choices for λ), or even 18 if we care about whether the rotation is clockwise or
counterclockwise (which depends on the sign of b). So just draw a couple of cases.

5.1.4 The general case through triangularization

Consider the following system, where A is a 2 × 2 triangular matrix:

(
x0 (t) = x(t) + y(t)
. (5.9)
y 0 (t) = y(t)

The last line leads to y(t) = y0 et , so the first line becomes x0 (t) = x(t) + y0 et . This is nonhomo-
geneous linear equation in dimension 1. We may solve it by the method of variation of constants,
that is, by looking for solutions of the form x(t) = λ(t)et . This leads to x(t) = (x0 + y0 t)et .
Thus the general solution of (5.9) is
!
x0 et + y0 tet
X(t) =
y0 e t

Any triangular system in Rd may be dealt with, in a similar fashion. We first find xd (·). This al-
lows to find xd−1 (·), which allows to find xd−2 (·), etc. If A is not triangular, but triangularizable,
50 CHAPTER 5. SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS

that is of the form A = P −1 T P with T triangular, we let Y = P X and note that:

X 0 = AX ⇔ X 0 = P −1 T P X ⇔ P X 0 = T P X ⇔ Y 0 = T Y.

The system Y 0 = T Y is triangular, so we may solve it, and deduce from the solutions of this
system the solutions of X 0 = AX. Note that in Md (C), every matrix is triangularizable. So
if we accept without proof that triangular systems in Md (C) may be dealt with as triangular
systems in Md (R) (just make the same formal computation), this gives a method to solve any
linear system of the form X 0 (t) = AX(t) (with A in Md (R), and even A in Md (C)!). We now
show that there is another way to solve such systems in the general case.

5.1.5 The equation X 0 (t) = AX(t) and the exponential of a matrix

If a is a real number, the solution of x0 (t) = ax(t) and x(0) = x0 is x(t) = eta x0 . This section
shows that this formula can be generalized to the equation X 0 (t) = AX(t), but replacing the
term eta with the exponential of the matrix tA, a notion we now define.

Definition 5.1.11. Let A ∈ Md (K), where K = R or K = C. The exponential of the matrix A

is defined by
+∞ n
X Ak X Ak A2 Ak
exp(A) = = lim =I +A+ + ... + + ....
k=0
k! n→+∞
k=0
k! 2! k!

Remark 5.1.12. As usual, we take the convention A0 = I where I is the d × d identity matrix.

Remark 5.1.13. The definition makes sense because the series is normally convergent. Indeed
take the norm on the space Md (K) such that ||Ak || ≤ ||A||k . Such a norm is obtained by taking
any norm || · || on Kd and taking on Md (K) the subordinate norm:

||AX||
||A|| = max = max ||AX||.
X6=0 ||X|| {X∈Kd : ||X||=1}

P+∞ ||A||k
Since ||Ak || ≤ ||A||k and k=0 k! = exp(||A||) < +∞, the series defining exp(A) converges.

Remark 5.1.14. In dimension 1, when writing the solution of x0 (t) = ax(t) and x(0) = x0 , we
can write indifferently ta or at and exp(ta)x0 or x0 exp(ta), because we are dealing with reals
numbers. But in dimension d, we should write tA and not At (by convention the scalar always
comes before the matrix), and more importantly, exp(tA)X0 and not X0 exp(tA): these are not
the same vectors!

Exponentials of matrices have properties similar to exponentials of real numbers, but the
property eA+B = eA eB is only valid if A and B commute, that is, if AB = BA.

Proposition 5.1.15 (basic properties). Let A and B be in Md (K). Let 0 and I denote the zero
and identity matrix of Md (K), respectively. Let t and s be real numbers. We have:

1. e0 = I.

2. If P is an invertible matrix in Md (K), then exp(P −1 AP ) = P −1 (exp A)P .

3. eA is a polynom in A (the polynom depending on A).

4. If AB = BA, then:

(a) P (A)Q(B) = Q(B)P (A) for any polynoms P and Q.

(b) P (A)eB = eB P (A) for any polynom P .
(c) exp(A + B) = exp(A) exp(B) = exp(B) exp(A)
5.1. THE EQUATION X 0 (T ) = AX(T ) 51

5. eA is invertible with inverse e−A .

6. e(t+s)A = etA esA .

Proof. 1) is an easy computation. To prove 2), recall that (P −1 AP )k = P −1 Ak P . Thus:

n n n
!
(P −1 AP )k P −1 Ak P Ak
= P −1
X X X
= P
k=0
k! k=0
k! k=0
k!
−1 AP
As n → +∞, the expression on the left goes to eP while the expression on the right goes
to P −1 eA P . This proves 2).
3) Let F = {P (A)|P ∈ K[X]} ⊂ Md (K) denote the set of polynoms in A. It is easily checked
that F is a vector subspace of Md (K), hence is closed in Md (K). Thus, as a limit of elements
of F , exp(A) is also in F .
4a) is a classical linear algebra property. We recall the proof for completeness. Assume that
AB = BA. Since A0 = I, we have Ak B = BAk for k = 0. Moreover, if Ak B = BAk , then
Ak+1 B = AAk B = ABAk = BAAk = BAk+1 , where we used successively Ak B = BAk and
AB = BA. So by induction Ak B = BAk for all k ∈ N. It follows immediately that for any
polynom P , P (A)B = BP (A). Now let à = P (A). Since ÃB = B Ã, the same reasoning as
above shows that à commutes with any polynom in B.
4b) follows from 3) and 4a).
4c) The proof that eA+B = eA eB is exactly the same as the proof that ex+y = ex ey . This
is because this proof involves only additions and multiplications (no division). Since when
AB = BA, addition and multiplication of the matrices A and B have the same properties as
addition and multiplication of real numbers, the proof for real numbers extends to matrices. We
could also have used this argument for 4a) and 4b).
5) Since A and −A commute, it follows that eA e−A = eA+(−A) = e0 = I.
6) tA and sA commute.

The following example shows that if AB 6= BA, then there is no reason to expect that
eA+B = eA eB .
! !
1 0 0 1
Example 5.1.16. If A = and B = , then eA+B , eA eB and eB eA all differ.
0 0 0 0
Indeed, let C != A + B. For !all k ≥ 2, Ak = !
A, C k = C, and B k
! = 0. It follows that: !
e 0 1 1 e e e e e 1
eA = , eB = , eA+B = , eA eB = and eB eA = .
0 1 0 1 0 0 0 1 0 1

Proposition 5.1.17 (Continuity and derivability). Let A ∈ Md (R). The application g : R →

Md (R) defined by g(t) = etA is of class C ∞ and

g 0 (t) = Ag(t) = AetA = etA A (5.10)

Proof. We first show that g is differentiable with derivative g 0 (t) = AetA . We begin by showing
that g 0 (0) = A. Let h(t) = g(t)−g(0)
t − A. We want to show that h(t) → 0 as t → 0. We have

1  tA  1 +∞
X tk Ak +∞
X tk−2 Ak
h(t) = e − I − tA = =t
t t k=2 k! k=2
k!

Therefore, for |t| < 1,

+∞ +∞
X |tk−2 | ||Ak || X ||A||k
||h(t)|| ≤ |t| ≤ |t| ≤ |t| exp(||A||) →t→0 0.
k=2
k! k=2
k!
52 CHAPTER 5. SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS

Therefore, g is differentiable at t = 0 and g 0 (0) = A. We now prove the general formula. We

have:
!
g(t0 + t) − g(t0 ) e(t0 +t)A − et0 A etA − I
= = et0 A →t→0 g 0 (0)et0 A = Aet0 A
t t t

due to the previous computation. Therefore, g is differentiable and g 0 (t) = AetA = Ag(t) for all
t. Since A and tA commute, we also have AetA = etA A, hence we proved (5.10).
It remains to check that g is C ∞ . To see this, note that g is differentiable, hence C 0 .
Moreover, if g is C k then, since g 0 (t) = Ag(t) and matrix multiplication is a C ∞ operation, g 0 is
C k ; therefore g is C k+1 . Thus, by induction, g is C k for all k. This concludes the proof.

Corollary 5.1.18. Let A ∈ Md (R) and let X0 ∈ Rd . The application X : R → Rd defined by

X(t) = etA X0 is differentiable with derivative X 0 (t) = AetA X0 = AX(t).

Proof. If M : R → Md (R) is differentiable and X0 ∈ Rd , then the application X : R → Rd

defined by X(t) = M (t)X0 is differentiable with derivative X 0 (t) = M 0 (t)X0 . Indeed,

X(t + h) − X(t) M (t + h) − M (t)

 
= X0 = (M 0 (t) + ε(h))X0 →h→0 M 0 (t)X0 ,
h h

where ε(h) is a matrix in Md (R) such that ε(h) → 0. This ensures that ε(h)X0 →h→0 0 because
taking a subordinate norm: ||ε(h)X0 || ≤ ||ε(h)|| ||X0 ||. The result then follows from Proposition
5.1.17.

Theorem 5.1.19. Let A ∈ Md (R) and X0 ∈ Rd . The initial value problem X 0 (t) = AX(t) and
X(0) = X0 has a unique maximal solution: the function X : R → Rd defined by

X(t) = etA X0

Proof. X(·) is solution by Corollary 5.1.18. It is global, hence maximal. Finally, the function
G : Rd → Rd defined by G(X) = AX is C 1 , hence there is a unique maximal solution. Uniqueness
may also be shown directly. Indeed, if (J, X̃(·)) is another solution (maximal or not), then letting
Y (t) = e−tA X̃(t) we get: Y 0 (t) = e−tA (X̃ 0 (t) − AX̃(t)) = 0 since X̃ 0 = AX̃. So for all t in J,
Y (t) = Y (0) = X̃(0) = X0 hence X̃(t) = etA X̃0 . That is, X̃(·) is just a restriction of the solution
X(·).

5.1.6 Computing the exponential of a matrix

To make a good use of Theorem 5.1.19, we need to learn how to compute the exponential of a
matrix. We begin with some simple cases.
! !
λ1 0 λk1 0
Case 1: A diagonal. If A = , then Ak = , so that
0 λ2 0 λk2
 
+∞
! P+∞ λk1 !
A
X1 λk1 0 k=0 k! 0 e λ1 0
e = = P+∞ λk2  =
k! 0 λk2 0 0 eλ2
k=0 k=0 k!

More generally, if A = diag(λ1 , ..., λd ), then eA = diag(eλ1 , ..., eλd ) and etA = diag(eλ1 t , ..., eλd t ).

Case 2: A diagonalizable (in C). If A = P −1 DP , with D diagonal, then eA = P −1 eD P .

Similarly, etA = P −1 etD P . Since computing etD is immediate, this allows to compute etA .
5.2. THE EQUATION X 0 (T ) = AX(T ) + B(T ) 53

Case 3: A nilpotent. If A is nilpotent, that is, if Aq = 0 for some integer q ≥ 1, then

X Ak q−1
X Ak
eA = =
k≥0
k! k=0
k!

since Ak = 0 for all k ≥ q. Thus the sum defining eA is finite and relatively easy to compute
when q is small. Note that by a standard linear algebra exercise, if a d × d matrix A is nilpotent,
then Ad = 0, thus, checking that a matrix is nilpotent, is a computation that takes a finite time.7

Case 4: A=D+N with D diagonal, N nilpotent, and DN=ND. Since D and N com-
mute, eD+N = eD eN , and since D is diagonal and N nilpotent, eD and eN are easy to compute.

Case 5: A = P−1 (D + N)P with D diagonal, N nilpotent, and DN = ND. Then

eA = P −1 eD eN P .

General case It may be shown that any matrix A in Md (C) may be put in the form

A = P −1 (D + N )P

with P invertible, D diagonal, N nilpotent, and DN = N D. The matrices P , D and N may

be computed explicitly. Thus, the previous case is the general case. Readers who want to know
how to compute the matrices P , D, N are invited to look up “décomposition de Dunford” on
Wikipedia or any good linear algebra textbook. See also “Jordan reduction”.
Note that for 2 × 2 matrices, the usual case in the exercises we will ask you to solve, the
decomposition is readily obtained. Indeed, either A is diagonalizable or it has a single eigenvalue
λ. But in the latter case, its characteristic polynomial is P (X) = (X − λ)2 so that by Cayley-
Hamilton Theorem (A−λI)2 = 0. Thus, letting D = λI and N = A−λI we get that A = D +N
with D diagonal, N nilpotent (since N 2 = 0), and DN = N D because D is not only diagonal,
but scalar.

5.2 The equation X 0 (t) = AX(t) + B(t)

The notion of the exponential of a matrix allows us to generalize Duhamel’s formula for nonho-
mogeneous systems of linear equations, when the matrix A is constant.

Proposition 5.2.1. Let A ∈ Md (R). Let B : R → Rd be continuous. Solutions of X 0 (t) =

AX(t) + B(t) are global and given by
 Z t 
tA −sA
Y (t) = e X0 + e B(s)ds . (DF2)
0

with X0 in Rd . This formula gives the solution with value X0 at t = 0.

Proof. First note that by Proposition 5.0.1, we know that solutions are defined on the whole
R (hence at t = 0) and that for each X0 in Rd , there is a unique solution with value X0 at
t = 0. Thus, it suffices to check that if Y (·) is given by (DF2), then it is a global solution with
7
To prove that if a d × d matrix A is nilpotent, then Ad = 0, a possibility is to show that the sequence KerAk ,
k = 0, 1, ... is increasing for the inclusion relationship, that is, KerAk ⊂ KerAk+1 , and that as soon as there is
equality for some k, the sequence becomes stationary. Since each time KerAk is strictly included in KerAk+1 , the
latter has a strictly higher dimension, and since these dimensions are at most equal to d, such a strict inclusion
may happen at most d times. Thus, after d steps (starting at k = 0), the sequence must be stationary, that is,
KerAd = KerAk for all k ≥ d. If KerAd = Rd , then Ad = 0 and the matrix is nilpotent. Otherwise, no power of
A is equal to 0, hence the matrix is not nilpotent.
54 CHAPTER 5. SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS

Y (0) = X0 . We’ll let you do it. Of course, this perfectly correct proof feels like cheating, since
this does not tell us where the formula comes from!
To get the formula, as in dimension 1, start with the general solution t → etA X0 of the
associated homogeneous equation. Then let the constant vector X0 vary, that is, look for
solutions of the form Y (t) = etA X(t). Formally, let Y : R → Rd be differentiable. Define
X : R → Rd by X(t) = e−tA Y (t). Then X(·) is differentiable and:

Y 0 (t) = AetA X(t) + etA X 0 (t) = etA (AX(t) + X 0 (t)) = AY (t) + etA X 0 (t)

Therefore, Y 0 (t) = AY (t) + B(t) if and only if etA X 0 (t) = B(t), that is, X 0 (t) = B̃(t), where
B̃(t) = e−tA B(t) is a column vector. Letting xi (t) and b̃i (t) denote the ith coordinate of X(t)
and B̃(t), respectively, this means that:
 
x01 (t) b̃1 (t)
 

 . 
..  =  
.   .. 


x0d (t) b̃d (t)
Rt
That is, for all i in {1, .., d}, x0i (t) = b̃i (t). This is equivalent to xi (t) = xi0 + 0 b̃i (s) ds, for some
constant xi0 . In matrix form, this reads
Z t
X(t) = X0 + B̃(s) ds.
0

where X0 = R(x10 , ..., xd0 )T . Thus, Y (·) is solution of Y 0 (t) = AY (t) + B(t) if and only if
X(t) = X0 + 0t e−sA B(s)ds for some X0 in Rd . This is equivalent to (DF2).
 R t −sA 
Exercise 5.2.2. Show that the solution with value X0 at t0 is Y (t) = etA e−t0 A X0 + t0 e B(s)ds .

Another approach to the equation

X 0 (t) = AX(t) + B(t) (NH)

is to start from a basis (X1 (·), ..., Xd (·)) of the set of solutions of the associated homogeneous
P
equation (H). The general solution of (H) is X(t) = i µi Xi (t). Now, let the constants µi vary.
That is, search for a solution of (NH) of the form
X
X(t) = µi (t)Xi (t)
i

with µi (·) : R → R differentiable. We then have

!
0
(µ0i Xi µi Xi0 ) (µ0i Xi µ0i Xi µ0i Xi .
X X X X X
X = + = + µi AXi ) = +A µi Xi = AX +
i i i i i

Thus, X(·) is solution of (NH) if and only if for all t in R:

µ0i (t)Xi (t) = AX(t) + B(t),

X
AX(t) +
i

that is, if and only if

µ0i (t)Xi (t) = B(t)
X

Letting M (t) denote the matrix with columns X1 (t), ...Xd (t), and µ(t) = (µ1 (t), ..., µd (t))T , this
condition may be written M (t)µ0 (t) = B(t). Moreover, the matrix M (t) is invertible since by
point 3 in Proposition 5.0.1, (X1 (t), ...Xd (t)) is a basis of XH for any t in R. Thus X(t) is
solution if and only if
µ0 (t) = M −1 (t)B(t).
5.2. THE EQUATION X 0 (T ) = AX(T ) + B(T ) 55

It then suffices to integrate in order to find µ(t) (up to d integration constants which depend on
the initial condition), hence X(t). As the method used to derive (DF2), this is a generalization
of the method of variation of constants in dimension 1. This method is natural when A is diag-
onalizable, so that we easily obtain a basis of SH . The method using the exponential is natural
when A is not diagonalizable. Note, however, that the exponential method is essentially a partic-
ular case of the basic method with Xi (t) = etA Vi , where Vi is the ith vector of the canonical basis.

We end this chapter by coming back to our initial equation: X 0 (t) = A(t)X(t) + B(t). The
difficulty in solving this equation is not the term B(t), but the fact that the matrix A(t) is not
constant. If for any times s and t, A(t) and A(s) commute (which boils down to the fact that
A(t) and A0 (t) commute when A(·) is differentiable), then Rit may be seen that Eq. (DF2) may
be generalized, replacing as in dimension 1 the term tA by 0t A(s)ds. If A(t) does not commute
with A(s) for all t, s, then life is harder and the equation cannot always be solved explicitly.
56 CHAPTER 5. SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
Chapter 6

Nth order equations

Many differential equations are second-order equations, in particular those coming from physics,
and higher order-equations may also arise. A way to study these equations is to show that they
are equivalent to an auxiliary first-order equation. This is the topic of the following section.
Some techniques however are specific to nth order equations. This is the topic of Section 6.2.

6.1 How to reduce a nth order equation to a first-order one?

An example. Let g denote the gravitational constant of the Earth and consider the equation

z 00 (t) = −g (6.1)

with z(t) ∈ R. This is a second-order differential equation: indeed it involves a second deriva-
tive. It models the evolution of the altitude of an object with unit mass subject to the earth
gravitational force. Maximal solutions are easily found by integrating twice. These are the
functions z : R → R such that z(t) = − 12 g(t − t0 )2 + v0 (t − t0 ) + z0 , where v0 and z0 correspond
respectively to the ascending speed and the altitude at time t0 . Note that there is an infinite
number of solutions such that z(t0 ) = z0 . This does not contradict Cauchy-Lipschitz theorem,
since the equation is not of first-order.
To apply Cauchy-Lipschitz theorem, we need to transform the equation into a first-order
one. To do so, let F (z1 , z2 ) = (z2 , −g) and Z(t) = (z1 (t), z2 (t)). We then have:
( (
0 z10 (t) = z2 (t) z2 (t) = z10 (t)
Z (t) = F (Z(t)) ⇔ ⇔
z20 (t) = −g z100 (t) = −g

Thus, Z(·) is solution of

Z 0 (t) = F (Z(t)) (6.2)
if and only if its second coordinate is the derivative of its first coordinate, and its first coordinate
is a solution of (6.1); that is, if and only if it is of the form Z(t) = (z(t), z 0 (t)) for some solution
z(·) of (6.1). Moreover, it is easily checked that Z(·) is then maximal if and only if z(·) is
maximal.

Exercise 6.1.1. Show that solving (6.1) is equivalent to solving (6.2) in the sense that the
mapping from the set of maximal solutions of (6.1) to the set of maximal solutions of (6.2)
which maps the real valued function z(·) to Z(·) = (z(·), z 0 (·)) is well defined and bijective.

Applying Cauchy-Lipschitz to the first order equation (6.2) and coming back to (6.1) yields:

Proposition 6.1.2. Let t0 ∈ R, let (z0 , v0 ) ∈ R2 . There is a unique maximal solution of (6.1)
such that z(t0 ) = z0 and z 0 (t0 ) = v0 .

57
58 CHAPTER 6. NTH ORDER EQUATIONS

Proof. Let (IPV1) denote the problem z 00 (t) = −g with (z(t0 ), z 0 (t0 )) = (z0 , v0 ) and (IPV2) the
problem Z 0 (t) = F (Z(t)) with Z(t0 ) = (z0 , v0 ). Since F is C 1 , Cauchy-Lipschitz Theorem tells
us that (IPV2) has a unique maximal solution. Denote it by (J, Z(·)) and let Z(·) = (z1 (·), z2 (·)).
Then, as we saw, z1 (·) is a maximal solution of (6.1) and z10 (·) = z2 (·) so that (z1 (t0 ), z10 (t0 )) =
ˆ z(·))
(z1 (t0 ), z2 (t0 )) = (z0 , v0 ). Therefore, z1 (·) is a maximal solution of (IPV1). Similarly, if (J,
0
is another maximal solution of (IPV1), then letting Ẑ(·) = (z(·), z (·)), we get that (J, ˆ Ẑ) is
ˆ
solution of (IPV2). By uniqueness of the maximal solution of (IPV2), (J, Ẑ(·)) = (J, Z(·))
hence (J, ˆ z(·)) = (J, z1 (·)). Thus, (IPV1) has a unique maximal solution.

The key-point is that the initial condition does not involve just the initial position z(t0 ), but
also the initial speed z 0 (t0 ). This shows that, even if we are only interested in the evolution of
the position of the object, we need to describe its state as a vector (position, speed) and study
the evolution of this vector in R2 , called the phase space of this equation.
General case. More generally, an explicit nth order equation is an equation of the form

x(n) (t) = f (t, x(t), x0 (t), ..., x(n−1) (t)) (6.3)

with x(t) ∈ Rd and f : Ω → Rd where Ω is an open subset of R × Rnd . If Ω = I × Rnd where

I is an open interval, then a solution of (6.3) is a couple (J, x(·)) where J is a nonempty open
subinterval of I and x : J → Rd a differentiable function satisfying (6.3) for all t in J. The
notions of maximal and global solutions are the same as for first-order equations. To simplify
notation, we assume in what follows that d = 1, that is, x(t) ∈ R. The case x(t) ∈ Rd is similar.
Assume thus that d = 1. The equation (6.3) may be reduced to a first-order equation in Rn
by letting
F (t, x1 , ..., xn ) = (x2 , ..., xn , f (t, x1 , ..., xn )).
The function defined by X(t) = (x0 (t), x1 (t), ..., xn−1 (t)) is the solution of

X 0 (t) = F (t, X(t)). (6.4)

if and only if it is of the form X(t) = (x(t), x0 (t), ..., x(n−1) (t)) for some function x(·) solution of
(6.3). Note that F is as regular as f . We let the reader check that applying Cauchy-Lipschitz
theorem to (6.4) and then coming back to the initial equation, we obtain:

Proposition 6.1.3. If f is C 1 or Lipschitz, then for any initial time t0 and vector (x0 , x1 , ..., xn−1 ) ∈
Rn , there is a unique maximal solution of (6.3) such that for all k ∈ {0, ..., n − 1}, x(k) (t0 ) = xk .

The phase space is then Rn . For the explosion alternative, we obtain:

Proposition 6.1.4. If f is C 1 or Lipschitz and z : J → R is a maximal solution of (6.3), then:

1) sup J = sup I or (sup J < sup I and ||(z(t), z 0 (t), ..., z (n−1) (t))|| → +∞ as t → sup J).
2) inf J = inf I or (inf J > inf I and ||(z(t), z 0 (t), ..., z (n−1) (t))|| → +∞ as t → inf J).

The fact that ||(z(t), z 0 (t), ..., z (n−1) (t))|| goes to infinity does not imply that |z(t)| goes to infinity.
E.g., it might be that z(·) is bounded but that the absolute value |z 0 (t)| of its derivative goes to
infinity in sup J. This is why sup J < sup I does not imply that |z(t)| → +∞ as t → sup J.

6.2 Nth order linear equations with constant coefficients

Consider the second order linear equation with constant coefficients:

x00 (t) + a1 x0 (t) + a0 x(t) = b(t) (NH2)

6.2. NT H ORDER LINEAR EQUATIONS WITH CONSTANT COEFFICIENTS 59
! !
0 1 0
with a0 , a1 , x(t), b(t) in R. Letting A = , B(t) = , we let the reader
−a0 −a1 b(t)
check that (NH2) is equivalent in the sense of Section 6.1 to the linear system of equations

X 0 (t) = AX(t) + B(t) (6.5)

More generally, the nth order linear differential equation with constant coefficients

x(n) (t) + an−1 x(n−1) (t) + · · · + a1 x0 (t) + a0 x(t) = b(t) (6.6)

may be solved through the system of n first-order linear differential equations (6.5) with
   
0 1 0 ··· 0 0
 .. .. .. .. ..   .. 

 . . . . . 


 . 

A=
 .. .. ..  and B(t) = 
  .. 
 . . . 0   . 

0 ··· ··· 0 1 0
   
   
−a0 −a1 · · · −an−2 −an−1 b(t)

There is however a more efficient method to solve (6.6), at least when b(t) has a simple form.
Let SN H and SH denote respectively the sets of solutions of (6.6) and of the associated
homogeneous equation

x(n) (t) + an−1 x(n−1) (t) + · · · + a1 x0 (t) + a0 x(t) = 0 (6.7)

The characteristic polynomial of this equation is P = X n + an−1 X n−1 + · · · + a1 X + a0 . We

now explain that:
1) SN H is an affine space with direction SH ;
2) a basis of SH may be found through the decomposition of the characteristic polynomial;
3) a particular solution of (6.6) may be easily found when b(t) has a simple form, which
together with 1) and 2), allows to solve (6.6) .
The next proposition proves point 1).

Proposition 6.2.1. Let X0 = (x0 , x1 , .., xn−1 ) ∈ Rn

1. There is a unique solution of (6.6) such that (x(t0 ), x0 (t0 ), ..., x(n−1) (t0 )) = X0 and it is
global. The same holds for (6.7) as a particular case.

2. SH is a real vector space of dimension n.

3. If xp (·) ∈ SN H , then SN H = xp (·) + SH .

Proof. 1. Existence and uniqueness of a solution follows from the version of Cauchy-Lipschitz
theorem for nth order differential equations. It is global because solutions of (6.5) are global.
2 & 3. The proof is the same as the proof of the corresponding results in Proposition
5.0.1.

The next proposition proves point 2). In its statement, we make an abuse of notation and
denote by tk eλt the function from R to R defined by x(t) = tk eλt .

Proposition 6.2.2.

1. Let λ ∈ R. If P (λ) = 0, then eλt is solution of (6.7). More generally, if λ is a root of P

with multiplicity m then tk eλt is solution of (6.7) for all k ∈ {0, 1, ..., m − 1}.
60 CHAPTER 6. NTH ORDER EQUATIONS

2. Let λ = α + iβ, with β 6= 0. If λ is a root of P with multiplicity m, then tk eαt cos(βt) and
tk eαt sin(βt) are solutions of (6.7) for all k ∈ {0, 1, ..., m − 1}.

3. The solutions obtained as above form a basis of SH .

Proof. For simplicity, we only prove the result in the case of second order equations (n = 2),
which is the one which typically arises in applications. The proof in the general case is similar
but requires more tedious computations.
1. Assume n = 2. Let x(t) = eλt . Then, the kth derivative of x is x(k) (t) = λk eλt , therefore
x00 (t) + a1 x0 (t) + a0 x(t) = λ2 eλt + a1 λeλt + a0 eλt = P (λ)eλt . Therefore, if P (λ) = 0 then t → eλt
is solution of (H). The proof is the same for a general n.
Now, let λ be a root of P with multiplicity 2, so that P (λ) = P 0 (λ) = 0. We already know
that t → eλt is a solution of (6.7). Moreover, letting x(t) = teλt we have x0 (t) = eλt (λt + 1) and
more generally, x(k) (t) = eλt (λk t + kλk−1 ). Thus,

x00 (t) + a1 x0 (t) + a0 x(t) = eλt [(λ2 t + 2λ) + a1 (λt + 1) + a0 t] = eλt (P (λ)t + P 0 (λ)) = 0

Thus, t → teλt is indeed solution. In the case n = 2, a root of P has multiplicity at most 2, so
this concludes the proof. For a general n, it may be shown as a corollary of Lemma 6.2.4 below
that if x(t) = teλt , then we still have x(n) (t) + ... + a1 x0 (t) + a0 x(t) = eλt (P (λ)t + P 0 (λ)), and
that more generally, if x(t) = tk eλt , then x(n) (t) + ... + a1 x0 (t) + a0 x(t) = eλt Q(t) where Q is a
polynomial whose coefficients are all 0 if P (λ) = ... = P (k) (λ) = 0. The result follows.
2. The same computation, but in C, shows that if λ is a root of P with multiplicity q, then
tk eλt is a complex solution of (6.7) for all k ∈ {1, ..., q − 1}; but it is easily seen that the real
and imaginary parts of a complex solution are real solutions. The result follows.
3. First note that a real root with multiplicity q gives q solutions. A complex root λ with
multiplicity q gives 2q solutions, but its conjugate root λ̄ gives the same solutions (or solutions
that are proportional), so together, λ and λ̄ give 2q potentially independent solutions. Thus,
the number of solutions we obtained from the roots of P is equal to the sum of the multiplicity
of these roots, that is, to n, which is the dimension of SH . To prove that these solutions form a
basis of SH , it remains to check that they are independent. For simplicity, we do it only in the
case n = 2. There are then three cases.
Case 1: If P has two distinct real roots λ1 , λ2 , then the two solutions we obtain are s1 (t) =
eλ1 t , s2 (t) = eλ2 t . Let (µ1 , µ2 ) ∈ R2 and assume that x(·) = µ1 s1 (·) + µ2 s2 (·) = 0, that is,
x(t) = µ1 eλ1 t + µ2 eλ2 t = 0 for all t ∈ R. Without loss of generality, assume λ2 > λ1 . Then as
t → +∞, x(t)e−λ2 t = µ1 e(λ1 −λ2 )t + µ2 →t→+∞ µ2 . But x(t) = 0 for all t, therefore, µ2 = 0, and
then µ1 = 0 as well. It follows that s1 (·) and s2 (·) are independent.1
Case 2: if P has a double real root λ, then the solutions we obtain are s1 (t) = eλt and
s2 (t) = teλt . Since s1 (t)/s2 (t) → 0 as t → +∞, an argument similar to the one of Case 1 shows
that s1 and s2 are independent.
Case 3: If P has two complex conjugate roots λ, λ̄ with λ = α + iβ, β 6= 0. The solutions we
obtain are then s1 (t) = eαt cos βt and s2 (t) = eαt sin βt. If µ1 s1 + µ2 s2 = 0, then taking t = 0
gives µ1 = 0, and then µ2 = 0. So s1 and s2 are independent.

To solve the nonhomogeneous equation (6.6), we still need to find a particular solution. In
1
This is an analytic proof. A more algebraic proof is as follows: if x(·) = µ1 s1 (·) + µ2 s2(·) = 0, then
 we
0 1 1
also have x (·) = λ1 µ1 s1 (·) + λ2 µ2 s2 (·) = 0. It follows that for all t, AX(t) = 0 where A = and
λ1 λ2
 
µ1 s1 (t)
X(t) = . Since det(A) = (λ2 − λ1 ) 6= 0, the matrix A is invertible, and AX(t) = 0 implies X(t) = 0.
µ2 s2 (t)
It follows that the functions µ1 s1 and µ2 s2 are both zero, hence that µ1 = µ2 = 0.
6.2. NT H ORDER LINEAR EQUATIONS WITH CONSTANT COEFFICIENTS 61

general, this may be done by variation of parameters (but be careful, see this note)2 . However,
when b(t) has a simple form, e.g., polynomial, then it is possible to find a particular solution
directly, thanks to the following result.

Proposition 6.2.3. Let Q be a polynomial of degree q and λ ∈ R. Let m be the multiplicity of

λ as a root of P (m = 0 if P (λ) 6= 0).
1a). If b(t) = eλt Q(t), then there is a solution of the form eλt S(t) where S is a polynomial
of degree q + m, and with valuation at least m (that is, all the terms of degree m − 1 or less are
zero). In particular, if P (λ) 6= 0, then there is a solution of the above form with deg S = q.
1b). If b(t) = Q(t), then there is a polynomial solution of degree q + m where m is the
smallest integer such that am 6= 0.
1c). If b(t) = eλt , then there is a solution of the form µtm eλt where m is the multiplicity of
λ as a root of P .
2) If b(t) = eαt cos(βt)Q(t) or b(t) = eαt sin(βt)Q(t), then there is a solution of the form
eαt [cos(βt)S
1 (t)+sin(βt)S2 (t)], where S1 and S2 are polynomials such that max(deg S1 , deg S2 ) =
q + m, where m is the multiplicity of λ = α + iβ as a root of P .

We first need a lemma.

Lemma 6.2.4. Let P = an X n + ... + a1 X + a0 and S be polynomials. Let λ ∈ C, and x(t) =

eλt S(t). Let ϕx (·) = an x(n) (·) + ... + a1 x0 (·) + a0 x(·). Then
n
X S (k) (t)
ϕx (t) = eλt P (k) (λ) (6.8)
k=0
k!

Proof. Since both ϕx and the right-hand-side of (6.8) depend linearly on P , it suffices to prove
q!
the formula when P = X q . We then have P (k) (λ) = (q−k)! λq−k and

q q
X q! X q!
ϕx (t) = x (q)
(t) = (eλt )(q−k) S (k) (t) = eλt λq−k S (k) (t)
k=0
k!(q − k)! k=0
k!(q − k)!

Using the previous expression for P (k) (λ) then yields (6.8).

We now prove the proposition

Proof. 1a). Let m denote the multiplicity of λ as a root of P , so that P (k) (λ) = 0 for all
k < m and P (m) (λ) 6= 0. Let f : Rq+m [X] → Rq [X] be the linear map defined by f (S) =
Pn (k) (λ) S (k) . Since P (k) (λ) = 0 for all k < m, f (S) = Pn (k) (λ) S (k) ; so if deg(S) ≤
k=0 P k! k=m P k!
q + m then deg(f (S)) ≤ q, hence f is well defined.
We claim that Kerf = Rm−1 [X]. Indeed, if S = 0 or deg S < m, then S (k) = 0 for all k ≥ m,
and f (S) = 0. Otherwise, since by definition of m, P (m) (λ) 6= 0, f (S) is of degree (deg S) − m ≥
0, hence f (S) 6= 0. It follows that the rank of f is (q + m + 1) − m = q + 1 = dim Rq [X], hence
that f is onto. Therefore, there exists S in Rq+m [X] such that f (S) = Q.
By Lemma 6.2.4, this implies that the function defined by x(t) = eλt S(t) is solution of
(6.6). Note that since deg(f (S)) = deg(S) − m, we have deg(S) = q + m. Moreover, if S̃ is
2
Though x(t) ∈ R, due to the fact that the equation is of order n, the equation is equivalent to a system of
n linear equations, not to a first order equation in dimension 1. Thus the appropriate method of variation of
constants is not to take a single solution x(·) of (6.7) and to search for solutions of the form y(t) = λi (t)x(t).
Rather,P it consists in taking a basis of solutions of (6.7) (x1 (·), ..., xn (·)), and to search for solutions of the form
y(t) = λ (t)xi (t) satisfying certain conditions. These conditions, which we do not write here, are those that
i i
arise from the method of variations of parameters for the equivalent system of linear equations. It is just as simple
as finding a particular solution of this system, and then saying that its first component is a particular solution of
(6.6).
62 CHAPTER 6. NTH ORDER EQUATIONS

the polynomial obtained from S by subtracting all the terms in S of degree m − 1 or less, so
that deg(S̃ − S) < m, we have: f (S̃) = f (S) + f (S̃ − S) = f (S) = Q, so that the function
t → x̃(t) = eλt S̃(t) is also solution. Since deg(S̃) = deg(S) = q + m and since S̃ has valuation
at least m, this proves 1a).
1b) and 1c) are particular cases of 1a), with respectively λ = 0, and Q = 1.
2) Consider the equation in C with b(t) = eλt Q(t) where λ = α + iβ. The same argument
as for 1a) but in C[X] shows that there is a complex solution of the form x(t) = eλt S(t),
with degS = q + m. Write S = Sr + iSim . Taking real and imaginary parts of the equality
x(n) (t) + ... + a0 x(t) = eλt Q(t) shows that the real and imaginary parts of t → eλt S(t) are
solutions of the equation with respectively b(t) = eαt cos(βt)Q(t), and b(t) = eαt sin(βt)Q(t),
and they are both of the form eαt [cos(βt)S1 (t) + sin(βt)S2 (t)] with max(degS1 , degS2 ) = degS.
The result follows.

Superposition principle. Finally, if b(t) is the sum of two terms with a nice form, then as for
first-order equations, we may use the superposition principle. That is, if x1 (·) is a solution of
the equation with right-hand side b1 (·) and x2 (·) of the equation with right-hand side b2 (·), then
x1 (·) + x2 (·) is a solution of the equation with right-hand side b1 (·) + b2 (·). The proof is the
same as for Proposition 3.3.6.
Chapter 7

Stability and linearization

7.1 Definitions

Consider an autonomous differential equation X 0 (t) = F (X(t)) with F : Rd → Rd of class C 1 .

Let X ∗ be an equilibrium of this equation. We say that:
- X ∗ is stable (or Lyapunov stable), if for any neighborhood V of X ∗ , there is a neighborhood
W of X ∗ such that any solution starting in W remains in V at all later times. That is, if
X(0) ∈ W , then X(t) ∈ V for all t ≥ 0.
- X ∗ is attracting, if it has a neighborhood V such that any solution starting in V converges
towards X ∗ as t → +∞. That is, X(0) ∈ V ⇒ X(t) →t→+∞ X ∗ .
- X ∗ is asymptotically stable, if it is both stable and attracting.
- X ∗ is unstable, if it is not stable. That is, X ∗ is unstable if there exists a neighborhood V
of X ∗ such that, for any neighborhood W of X ∗ , there exists X0 in W such that the solution
starting at X0 eventually leaves V (forward in time).
Of course, in the definitions of stable and attracting, we implicitly require that any solution
starting close enough to X ∗ be defined for all t in [0, +∞[.
Equivalent definitions may be given using basis of neighborhoods of X ∗ . For instance, X ∗
is stable if and only if for any ε > 0, there exists α > 0 such that, for any solution X(·):
||X(0) − X ∗ || ≤ α ⇒ ∀t ≥ 0, ||X(t) − X ∗ || ≤ ε.

Exercise 7.1.1. Give equivalent definitions of “attracting” and “unstable” in a similar way.

Exercise 7.1.2. Let f : R → R be C 1 . Show that for the differential equation x0 (t) = f (x(t)),
attracting implies stable, so that attracting is equivalent to asymptotically stable. This is no
longer the case in higher dimension.

7.2 Stability of linear systems

Before stating the next proposition, we need some vocabulary. Let A ∈ Md (R) and let λ
be a (real or complex) eigenvalue of A. The geometric multiplicity of λ is the dimension of
Ker(A − λI). The algebraic multiplicity of λ is its multiplicity as a root of the characteristic
polynomial of A.

Exercise 7.2.1. Show that the algebraic multiplicity is always weakly larger than the geometric
multiplicity.

Exercise 7.2.2. Show that A is diagonalizable if and only if the geometric multiplicity of each
eigenvalue of A is equal to its algebraic multiplicity.

63
64 CHAPTER 7. STABILITY AND LINEARIZATION

Exercise 7.2.3. What are the geometric and algebraic multiplicity of 2 as an eigenvalue of the
following matrix? !
2 1
0 2

Consider a system of linear autonomous differential equations X 0 (t) = AX(t), with A ∈

Md (R). Note that the origin (the zero of Rd ) is an equilibrium for this system. Let Sp(A)
denote the spectrum of A, that is, its set of eigenvalues. Let a = maxλ∈Sp(A) Re(λ) be the
maximal real part of the eigenvalues of A. The following result shows that when a 6= 0, the sign
of a determines the stability of the origin.

Theorem 7.2.4. (stability of linear systems)

1. If a < 0, then 0 is asymptotically stable.

2. If a > 0, then 0 is unstable.

3. If a = 0 then 0 is not asymptotically stable. It is stable if and only if the geometric

multiplicity of each eigenvalue of A with zero real part is equal to its algebraic multiplicity.
This is the case for instance if A is diagonalizable.

Proof. (sketch) The proof of point 2. is simple: just consider solutions starting on an unstable
eigenaxis (that is, the span of an eigenvector associated with a positive eigenvalue) or what
could be called an unstable eigenplane (the span of the real and imaginary part of a complex
eigenvector associated to a complex eigenvalue with a positive real part).
The proof of point 1. is based on the fact that any matrix A in Md (R) may be written
in the more convenient form A = P −1 (D + N )P with P invertible, D diagonal, N nilpotent,
DN = N D, moreover, N is triangular with zeros on the diagonal, which implies that D and A
have the same eigenvalues (e.g., using the Jordan reduction of a matrix). Using the subordinate
norm associated to the infinite norm on Rd , we then have:

||etA || ≤ C||etD || ||etN || with C = ||P −1 || ||P ||

Moreover, it is easily shown that ||etD || ≤ eat where a is the maximal real part of the eigenvalues
of D, hence of the eigenvalues of A. Besides, due to the fact that there is a finite number of
term in the sum defining etN , ||etN || ≤ Q(|t|) for some polynomial Q of degree at most d − 1. It
follows that:
||etA || ≤ eat Q(|t|) →t→+∞ 0
since a < 0. For each solution X(·), since X(t) = etA X(0) and since we use a subordinate
norm, we get ||X(t)|| ≤ ||etA || ||X(0)|| →t→+∞ 0. We say that 0 is globally attracting (it does
not only attract solutions starting in a neighborhood of itself, but all solutions; in other words,
we can take V = Rd in the definition of “attracting”). Moreover, since etA goes to zero as
t → +∞, it follows that etA is bounded by some constant K on [0; +∞[. Thus, for each ε > 0, if
||X(0)|| ≤ α := ε/K, then for all t ∈ [0, +∞[, ||X(t)|| ≤ K||X(0)|| ≤ ε. Thus 0 is stable. Since
it is also attracting, it is asymptotically stable.
The proof of point 3. is based on the fact that when a = 0, it may be shown that the norm
of etA is bounded under the condition of the theorem, but grows till infinity at a polynomial
speed if these conditions are not fulfilled, due to polynomial terms coming from etN .

Exercise 7.2.5. For the following matrices A, is the origin stable, asymptotically stable or
unstable for the system X 0 = AX?
! ! ! ! !
1 −1 0 −1 1 0 −1 0 0 0 −1
A= ;A = ;A = ;A = ; A= .
0 −1 0 −1 0 0 0 −1 −1 0
7.3. LINEARIZATION 65

Exercise 7.2.6. Let A and P be d × d real matrices, with P invertible. Show that the origin
has the same stability for X 0 = AX and for X 0 = P −1 AP X.

Exercise 7.2.7. Show that the following matrix is not diagonalizable but that for each of its
eigenvalues with zero real part, the geometric multiplicity is equal to the algebraic multiplicity:
 
0 0 0
A =  0 −1 1 
 
0 0 −1

Exercise 7.2.8. Consider the case of planar linear systems of differential equations. What is
the stability of a saddle? a source? a spiral source? a sink? a spiral sink? a center?

7.3 Linearization

Consider a nonlinear autonomous differential equation

X 0 (t) = F (X(t)) (7.1)

with F : Rd → Rd of class C 1 . Let X ∗ be an equilibrium and

!
∂Fi
AX ∗ = (X ∗ ).
∂xj 1≤i,j≤d

be the Jacobian matrix of F at X ∗ . Then by definition of the Jacobian matrix,

F (X ∗ + h) = F (X ∗ ) + AX ∗ h + o(h),

where h ∈ Rd . Since X ∗ is an equilibrium, F (X ∗ ) = 0, hence

F (X ∗ + h) = AX ∗ h + o(h) (7.2)

Thus, if X(·) is a solution of (7.1) and h(t) = X(t) − X ∗ , we have:

h0 (t) = X 0 (t) = F (X(t)) = F (X ∗ + h(t)) = AX ∗ h(t) + o(h(t)).

That is,
h0 (t) = AX ∗ h(t) + o(h(t)) (7.3)
The linearized system at X ∗ is the first-order approximation of (7.3), that is:

h0 (t) = AX ∗ h(t). (7.4)

The1 idea of introducing the linearized system is that close to X ∗ , the term o(h) in (7.3) is
small, so that the evolution of h(t) = X(t) − X ∗ is almost given by (7.4). This suggests that
the behavior of the linearized system close to 0 should provide information on the behavior of
the initial system close to X ∗ . This should be the case if a condition that ensures that the
first order approximation governs the local behavior of (7.3) is satisfied (like when, to use the
derivative of a nonlinear function from R to R to know whether this function is locally increasing
or decreasing, we need this derivative to be nonzero). This condition is that the equilibrium is
hyperbolic, in the following sense:

Definition 7.3.1. The equilibrium X ∗ is hyperbolic if the Jacobian matrix of F at X ∗ (that is,
AX ∗ ) has no eigenvalues with zero real part.
1
Of course, we could write the linearized system as X 0 (t) = AX(t): this is the same system and a more usual
notation. We do not do so here, because we want to insist on the fact that the variable in the linearized system
is meant to approximate the difference between the solution and the equilibrium in the initial, nonlinear system.
66 CHAPTER 7. STABILITY AND LINEARIZATION

In that case, either all eigenvalues of AX ∗ have negative real parts, and the origin is asymp-
totically stable under (7.4) by Theorem 7.2.4, or at least one eigenvalue of AX ∗ has a positive
real part and the origin is unstable under (7.4). Thus, for the linearized system associated to a
hyperbolic equilibrium, the origin cannot be stable without being asymptotically stable. This
explains why this case does not appear in the theorem below.

Theorem 7.3.2 (Linearization theorem). Assume that an equilibrium X ∗ of a nonlinear system

(7.1) is hyperbolic. Then it has the same stability as the origin in the linearized system at X ∗ .
That is, X ∗ is asymptotically stable for (7.1) if and only if 0 is asymptotically stable for (7.4),
and unstable for (7.1) if and only if 0 is unstable for (7.4).

Let us give some examples. Consider the equations in dimension 1: i) x0 = x3 , ii) x0 = −x3 ,
iii) x0 = 0. In all three cases, the unique equilibrium is 0 and the linearized system is h0 = 0.
However, the equilibrium is unstable in case i), asymptotically stable in case ii), and stable but
not asymptotically stable in case iii). This shows that we cannot deduce the stability of the
equilibrium of the nonlinear equation from the linearized system. This is because, in this case,
the Jacobian matrix at the equilibrium is a 1 × 1 matrix whose unique entry is zero, hence with
a zero eigenvalue. Thus, the equilibrium is not hyperbolic and we cannot apply Theorem 7.3.2.
Now consider the nonlinear system:
(
x0 = x2 + y
(7.5)
y0 = x − y

We let the reader check that there are two equilibria:

! (0, 0) and (−1, −1), and that the Jacobian
2x 1
matrix at X = (x, y) is A(x,y) = . Thus, the Jacobian matrix at X1∗ = (0, 0) is
1 −1
!
0 1
A(0,0) = and the linearized system at (0, 0) is h0 = A(0,0) h, that is:
1 −1
(
h01 = h2
(7.6)
h02 = h1 − h2

Since det AX ∗ = −1, the eigenvalues of AX ∗ are real and with opposite signs (in other words, the
origin is a saddle for the linearized system). It follows that: a) X ∗ is hyperbolic (all eigenvalues
of AX ∗ have nonzero real parts) ; b) the origin is unstable for the linearized system (since at
least one eigenvalue of AX ∗ has a positive real part). Therefore, X ∗ is unstable for (7.5).
!
−2 1
Similarly, at X2∗
= (−1, −1), the Jacobian matrix is A(−1,−1) = . Its determi-
1 −1
nant is equal to 1. This implies that either the eigenvalues are real and with the same sign,
or they are complex conjugates. In both cases, their real parts have the same sign, hence they
are negative since T r(A(−1,−1) ) = −3 < 0. It follows that X2∗ is hyperbolic, and that the origin
is asymptotically stable for the linearized system at X2∗ . Therefore, by Theorem 7.3.2, X2∗ is
asymptotically stable under (7.5). 2

Remark 7.3.3. The behavior of a smooth nonlinear system of differential equations near an
hyperbolic fixed point is much more related to the behavior of the linearized system than stated in
Theorem 7.3.2. This follows from the Hartman-Grobman Theorem, which is outside the scope of
this course. In essence, it states that oriented trajectories of the smooth nonlinear system near
an hyperbolic equilibrium are continuous deformations of oriented trajectories of the linearized
system near 0. For a first approach, see, e.g., Section 8.1-8.3 of Hirsch, Smale and Devaney’s
textbook: “Differential Equations, Dynamical Systems and An Introduction to Chaos”.

2
It is easily checked that the eigenvalues of A(−1,−1) are actually real, but this is not needed: knowing the sign
of their real parts suffices.
Chapter 8

Geometric approach to differential

equations

8.1 The slope field

Let (J, X(·) ) be a solution of a non autonomous equation

X 0 (t) = F (t, X(t)) (8.1)

with F : Rd+1 → Rd . Recall that, as usual, the graph of X(·) is the set Γ = {(t, X(t)), t ∈ J}.
The aim of this section is to explain how to have a rough idea of this graph without explicitly
solving the differential equation.
A vector field in Rn is a function ϕ : Rn → Rn seen as associating to each point X in Rn the
vector starting from X and with coordinates ϕ(X) (thus, going from X to X + ϕ(X)).
The velocity field associated to (8.1) is the vector field in Rd+1 associating to the point (t, X)
the vector starting from (t, X) and with coordinates (1, F (t, X)).
The slope field is a vector field colinear to the velocity field but where we consider normalized
vectors for a better visualization.
The slope field allows to get a rough idea of the graph of the solutions. For simplicity, the
result below is stated for d = 1, but it is actually general.

Proposition 8.1.1. (Geometric characterization of solutions) Consider a differential equation

x0 (t) = f (t, x(t)) (8.2)

with f : R2 → R. Let x : J → R be differentiable. The following assertions are equivalent:

a) (J, x(·)) is solution of (4.1); that is: ∀t ∈ J, x0 (t) = f (t, x(t)).
b) For all t in J, the slope of the graph at (t, x(t)) is f (t, x(t)).
c) For all t in J, the tangent to the graph of x(·) at (t, x(t)) has the direction of the vector
(1, f (t, x(t)).
d) For all t ∈ J, the graph of x(·) is tangent at (t, x(t)) to the velocity field (or equivalently,
to the slope field).

We leave the proof to the reader. In higher dimension, we similarly get that a differentiable
function X : J → Rd is solution of (8.1) if and only if for each t in J, the direction of the graph
at (t, X(t)) is the direction of the slope field. Thus, a way to get a rough idea of the graph of
solutions is to draw and “follow” the slope field. This is also the idea of numerical methods,
such as Euler method, which we roughly explain below.

67
68 CHAPTER 8. GEOMETRIC APPROACH TO DIFFERENTIAL EQUATIONS

8.2 Idea of Euler method with a fixed step size

Let X : J → Rd be a solution of (8.1) such that X(t0 ) = X0 . We would like to approximate

numerically this solution. To do so, fix a stepsize τ > 0 and let tn = t0 + nτ . Then consider the
function Yτ defined as follows:
a) Yτ (t0 ) = X0 (we start from X0 ).
b) Yτ (t1 ) = Yτ (t0 ) + τ F (t0 , X0 ), with Yτ (t) affine between t0 and t1 (from t0 to t1 , we follow
the slope field at (t0 , X0 ) for τ units of time).
c) Yτ (t2 ) = Yτ (t1 ) + τ F (t1 , Yτ (t1 )) (from t1 to t2 , we follow the slope field at (t1 , Yτ (t1 )) for
τ units of time).
d) More generally Yτ (tn+1 ) = Yτ (tn ) + τ f (tn , Yτ (tn )), with Yτ (t) affine between tn and tn+1 .

Remark 8.2.1. From t0 to t1 , we follow the direction of the slope field at (t0 , X0 ); that is, we
approximate the true solution by its tangent at t0 . From t1 to t2 , we follow the direction of the
slope field in (t1 , Yτ (t1 )), that is, the direction of the tangent to the solution of the equation with
value Yτ (t1 ) at t1 . Since Yτ (t1 ) 6= X(t1 ), this is not the solution that we try to follow. Thus there
are two mistakes: first, approximating a curve by its tangent; second, using a wrong tangent,
that is, a line with slope F (t1 , Yτ (t1 )) instead of F (t1 , X(t1 )).

It is clear that these mistakes may accumulate, and that the further we get from X0 , the
worse our guarantee (and our accuracy) on our numerical approximation will be. Nevertheless,
it may be shown that on any compact subinterval of J, Yτ (·) converges uniformly towards X(·)
as τ → 0, with an explicit bound on the approximation error.
Note that, if solutions of our differential equation are convex, then Euler method is bound to
underestimate them, because the tangent is always below a convex curve. Similarly, we expect
that Euler method will overestimate concave solutions. A way to deal with this problem is to
use more refined methods, that try to estimate and take into account higher derivatives of F ,
for instance, the Runge-Kutta method which will be studied in Master 1.

8.3 The case of autonomous equations

For an autonomous differential equation X 0 (t) = f (X(t)), the slope field is invariant by trans-
lation through time. Together with Proposition 8.1.1, this gives a geometric proof of the fact
that the set of solutions of an autonomous equation is invariant through time, a fact we already
proved analytically.
Besides, let us define the direction field of this equation as the vector field in Rd associating to
the point X the vector starting from X and with coordinates f (X). It may be seen that, except
for equilibria, trajectories of solutions are tangent to the direction field. Thus, for autonomous
equations, following the direction field allows to get an idea of the phase portrait.1
Of course, we cannot draw the direction field at each and every point. So we will try to
partition the phase space Rd (the space of possible states of the system) into several zones,
such that in each of this zone trajectories go, roughly, qualitatively, in the same direction. A
standard tool to do so are the nullclines. Given an autonomous equation X 0 (t) = F (X(t)),
with F = (F1 , ..., Fd ), the ith -nullcline is the set of points X in Rd such that Fi (X) = 0. In
one of the regions separated by the nullclines, all the components of F have a constant sign, so
that trajectories in this region go roughly in the same direction. This is best illustrated by an
example.
1
The expressions “velocity field”, “slope field” and “direction field” may be used with a somewhat different
meaning depending on authors, but the context should make the vocabulary clear.
8.3. THE CASE OF AUTONOMOUS EQUATIONS 69

Consider the linear system (

x0 = y
y 0 = −x
Here the nullclines are just the x-axis (on which y 0 = 0) and the y-axis (on which x0 = 0). They
separate R2 in four orthants. For any solution (x, y), as long as it lies in the open orthant x > 0,
y > 0, we have x0 > 0, y 0 < 0, so solutions move towards the “South-East” of R2 . Similarly,
solutions move South-West when x > 0, y < 0, North-West when x < 0, y < 0, and North-East
when x < 0, y > 0. This suggests that, unless solutions escape to infinity, they keep rotating
clockwise. Of course, we already knew that the system is linear and the origin a center, since it
is easily checked. However, the same kind of arguments apply even to the nonlinear system
(
x0 = y + y 3
y 0 = −x3

whose behavior cannot even be analyzed locally by linearization, since the unique equilibrium
(0, 0) is not hyperbolic (even if it were, this would only provide information on the behavior
close to the origin, while the above geometrical arguments provide information on the behavior
of the system globally, that is, in the whole R2 ). We will see in some exercises how to use these
ideas to make rigorous statements, and sometimes fully understand qualitatively, the behavior
of nonlinear systems.
For systems in R2 , it may also be useful to draw the isoclines. These are the curves defined
by F2 /F1 = k (or y 0 /x0 = k), that is, the set of points X in R2 such that the slope of the
trajectory at X is equal to k. The nullclines correspond to the particular case k = 0 (which
corresponds to y 0 = 0) and k = ±∞ (which corresponds to x0 = 0).
70 CHAPTER 8. GEOMETRIC APPROACH TO DIFFERENTIAL EQUATIONS
Chapter 9

Continuity of the flow

Consider an autonomous differential equation X 0 (t) = F (X(t)), with F : Rd → Rd of class C 1

or Lipschitz, and assume, to begin with, that all solutions are defined for all times. Let φ(t, X0 )
denote the value at time t of the solution starting at X0 (that is, satisfying X(0) = X0 ). This
defines a function φ : R × Rd → Rd called the flow of this differential equation. The function
φt : Rd → Rd defined by φt (X0 ) = φ(t, X0 ) is called the time t map of the flow. Applying the
function φt amounts to following solutions for t units of time. When t is positive, this makes us
travel into the future. When t is negative, into the past.
The flow is a fundamental tool to study differential equations, as it allows to consider solu-
tions globally, and not only one at a time. For instance, if A is a subset of Rd filled at time 0 with
imaginary particles whose law of motion is described by the differential equation, then φt (A) is
the position of this set of particles t units of time later. It may be that each of these particles
moves but that collectively they always occupy the same set of positions (think of a rotating
circle), so that for all t, φt (A) = A, in which case we say that A is globally invariant. It may be
that for each subset A of Rd , and for all times t, the volume of φt (A) is equal to the volume of A,
in which case we say that the flow (or the underlying differential equation) is volume-preserving.
These and other notions are very useful to study dynamical systems. Asking whether these
properties hold requires a tool allowing to consider the behavior of sets of solutions, and this
tool is the flow. Here, we will only study some of its basic properties. The most basic are the
following.

Proposition 9.0.1 (basic properties of the flow). 1) φ0 (X0 ) = X0 for all X0 ∈ Rd ; in other
words, φ0 = Id.
2) φs+t (X0 ) = φs (φt (X0 )) = φt (φs (X0 )) for all X0 in Rd , and t, s in R
3) For all t in R, the time t map of the flow φt is invertible with inverse φ−t .
∂φ
4) ∂t (t, X0 ) = F (φ(t, X0 ))

Proof. These properties are all obvious once one understands the definition of the flow.
Point 1) just says that if we follow a solution for 0 units of time, we do not move!
Point 2) says that following a solution for t + s units of time amounts to first following it for
t units of time, and then again for s units of time starting for the point we had reached (or first
s units of time, then t). The only subtle point is that this also holds if s, t or both are negative
(e.g., if we follow a solution for, say, 10 units of time, and then go back in time for 1 unit of
time, then it is as if we had followed the solution for 9 units of time.)
Point 3) follows from 1) and 2): by 2), φt ◦ φ−t = φ−t ◦ φt = φt+(−t) = φ0 and by 1), φ0 = Id.
The result follows.
For point 4), let X0 ∈ Rd , and denote by X(t) the value at time t of the solution with
initial condition X(0) = X0 . Thus, by definition of the flow, φ(t, X0 ) = X(t). Therefore,

71
72 CHAPTER 9. CONTINUITY OF THE FLOW

∂φ
∂t (t, X0 ) = X 0 (t) = F (X(t)) = F (φ(t, X)).

The following property implies that solutions of differential equations depend continuously
on their initial position. We do not assume anymore that all solutions are defined for all times;
thus, the time t map of the flow is now only defined on the set Ωt of initial positions X0 in Rd
such that the solution satisfying X(0) = X0 is defined until time t.
Proposition 9.0.2 (Continuity of solutions with respect to their initial position). Consider an
autonomous differential equation X 0 (t) = F (X(t)) with F : Rd → Rd of class C 1 or Lipschitz.
Let t ∈ R. Fix X0 in Ωt . There exists ε > 0 and a constant K such that for any Y0 in Rd , if
||Y0 − X0 || < ε, then Y0 ∈ Ωt and for all s between 0 and t:
||φs (Y0 ) − φs (X0 )|| ≤ ||Y0 − X0 ||eK|s| .
That is, if X(·) and Y (·) are the solutions such that X(0) = X0 and Y (0) = Y0 , respectively,
||Y (s) − X(s)|| ≤ ||Y0 − X0 ||eK|s| . (9.1)

Proof. We prove the result when F is K-Lipschitz. In this case, all solutions are defined for all
times (check it!), so Ωt = Rd for all t. Moreoever, the result is then true for any ε, so that ε will
not appear in the proof below. When F is C 1 but not Lipschitz, the proof uses the fact that
F is still locally Lipschitz, hence the need to focus on a neighborhood of the trajectory of X(·).
This is why we need to use ε in this case.
Assume thus that F is K-Lipschitz. Let X(·) and Y (·) be the solutions such that X(0) = X0
and Y (0) = Y0 respectively. We have:
Z t Z t
0
X(t) = X(0) + X (s)ds = X0 + F (X(s))ds.
0 0
Rt
and similarly, Y (t) = Y0 + 0 F (Y (s))ds. Thus,
Z t
Y (t) − X(t) = (Y0 − X0 ) + [F (Y (s)) − F (X(s))]ds.
0

Letting u(t) = ||Y (t) − X(t)||, and using that

||F (Y (s)) − F (X(s))|| ≤ K||Y (s) − X(s)||,
we obtain that for all t ≥ 0, Z t
u(t) ≤ u(0) + Ku(s) ds.
0
Therefore, it follows from Gronwall’s Lemma in the integral form, that for all t ≥ 0, u(t) ≤
u(0)eKt . A similar reasoning shows that for all t ≤ 0, u(t) ≤ u(0)eK|t| , so that in general,
||Y (t) − X(t)|| ≤ ||Y0 − X0 ||eK|t| .

Remark 9.0.3. Of course the above proposition holds even if we start from an initial time
t0 6= 0.
Corollary 9.0.4 (Continuously dependence from the initial data). Under the assumption of
Proposition 9.0.2, if the initial condition Y0 tends to X0 , then the solution Y (·) converges uni-
formly to X(·) over every bounded interval of time [t0 − T, t0 + T ].

Proof. Form 9.1, we have

||Y (·) − X(·)||∞ = sup ||Y (t) − X(t)|| ≤ ||Y0 − X0 ||eKT .
t∈[t0 −T,t0 +T ]
 73

We can also use Proposition 9.0.2 to prove uniqueness of the solution.

Corollary 9.0.5 (Uniqueness). Under the assumption of Proposition 9.0.2, if Y0 = X0 then

Y (·) = X(·).

Remark 9.0.6. There are many variants of Proposition 9.0.2: the value at a fixed time t of the
solution of X 0 = F (X) such that X(t0 ) = X0 depends continuously not only on X0 , but also on
t0 and on the law of motion F , and on all of this jointly. This also holds for non autonomous
equations.
74 CHAPTER 9. CONTINUITY OF THE FLOW
Chapter 10

Lyapunov functions

Consider an autonomous differential equation

X 0 (t) = F (X(t)) (10.1)

with F : Rd → Rd of class C 1 . We have seen in Chapter 7 that linearizing this system allows to
get information on the stability of hyperbolic equilibria. Analyzing nonlinear equations, via the
method of linearization close to equilibria, has however several drawbacks. First, the method
does not apply to all equilibria. Second, it does not allow to detect setwise attractors (e.g., if a
solution spirals and eventually approaches a whole circle instead of a single point). Third, it does
not provide any information on the size of basins of attraction, where the basin of attraction of
an equilibrium X ∗ is defined as the set of points X0 in Rd such that the solution starting at X0
is defined for all t ≥ 0 and converges to X ∗ as t → +∞. The aim of this chapter is to introduce
a method that allows, at least in some cases, to solve these problems. The method consists in
studying quantities that decrease along trajectories. A good candidate in applications to physics
is the total energy of the system which, in the absence of an exterior energy source, is constant
or decreases, due to friction terms. If we can find such a quantity, we then know that solutions
travel from areas of the phase plane where this quantity is large, to areas where it is small.
Drawing the landscape obtained by taking this quantity as the altitude allows then to get a
good understanding of the global behavior of the system.

10.1 Evolution of a quantity along trajectories

Let U ⊂ Rd . Consider a solution (J, X(·)) and a subinterval J˜ of J such that X(t) ∈ U on J.
˜
d d
Let L : R → R be differentiable and for all X ∈ R , let

L̇(X) := ∇L(X) · F (X).

Finally define l : J˜ → R by l(t) = L(X(t)). In this way, the evolution of l corresponds to the
evolution of L as we follow the solution X(·).

Proposition 10.1.1. If L̇(X) ≤ 0 on U , then the function l is decreasing. That is, L decreases
along pieces of trajectories in U .
Similarly, if L̇(X) < 0 (resp. = 0) then l is strictly decreasing (resp. is constant). That is,
L decreases strictly (resp. is constant) along pieces of trajectories in U .

˜ l0 (t) = ∇L(X(t)) · X 0 (t) = ∇L(X(t)) · F (X(t)) = L̇(X(t)) with X(t) ∈ U

Proof. For all t in J,
˜ Thus, if L̇(X) ≤ 0 (resp. < 0, = 0) on U , then l0 (t) ≤ 0 (resp. < 0, = 0), for
by definition of J.
˜
all t in J.

75
76 CHAPTER 10. LYAPUNOV FUNCTIONS

Remark 10.1.2. Geometrically, the condition ∇L(X) · F (X) ≤ 0 means that the angle between
the gradient of L and the vector field defined by F is obtuse (that is, of at least 90 degrees). Since
trajectories are tangent to this vector field, this means that trajectories go from high level curves
of L to low level curves (or are tangent to level curves of L if ∇L(X) · F (X) = 0, remember
that ∇L(X) gives the direction of maximal growth of L at X).

10.2 Global Lyapunov function

Definition 10.2.1. L is a global Lyapunov function for (10.1) if L̇(X) ≤ 0 on Rd .

By Proposition 10.1.1, a global Lyapunov function is a function that decreases along all
trajectories. Moreover, if L̇(X) = 0 on Rd , then L is actually constant along trajectories. We
then say that L is a constant of movement.

Proposition 10.2.2. Let L be a global Lyapunov function for (10.1).

1. If L is coercive (i.e. L(X) → +∞ as ||X|| → +∞), then all solutions of (10.1) defined at
t = 0 are defined and bounded on [0, +∞[.

2. If X(·) is a solution of (10.1) defined till +∞, then any accumulation point X ∗ of X(t)
as t → +∞ satisfies L̇(X ∗ ) = 0. In other words, if there exists a sequence (tn ) such that
tn → +∞ and X(tn ) → X ∗ as n → +∞, then L̇(X ∗ ) = 0.

3. Let E = {X ∈ Rd |L̇(X) = 0}. Let X(·) be a solution defined and bounded on [0, +∞[.
Then d(X(t), E) → 0 as t → +∞, where d(X, E) = inf Y ∈E ||X − Y ||. Moreover, if E
consists of isolated points, then there exists X ∗ ∈ E such that X(t) → X ∗ as t → +∞.

4. If L is coercive and E consists of isolated points, then for all solutions X(·), there exists
X ∗ in E such that X(t) → X ∗ as t → +∞ (this limit X ∗ may depend on the solution.).

Proof. 1) Assume that L is coercive. We claim that for all λ, Gλ = {X ∈ Rd |L(X) ≤ λ} is

bounded. Indeed, otherwise, there exists a sequence of points Xn in Gλ such that ||Xn || → +∞
as n → +∞. Since L is coercive, this implies that L(Xn ) → +∞ as n → +∞. This contradicts
the fact that L(Xn ) ≤ λ for all n.
This being seen, let (J, X(·)) be a solution of (10.1). Since L decreases along trajectories, it
follows that for all t ∈ J ∩ R+ , L(X(t)) ≤ L(X(0)). By the previous claim with λ = L(X(0)),
this implies that X(t) is bounded on J ∩ R+ . Thus, sup J = +∞ by the explosion alternative.
The result follows.
2) Let X ∗ in Rd . Assume that there exists a sequence (tn ) going to +∞ such that X(tn ) → X ∗
as n → +∞. By contradiction, assume that L̇(X ∗ ) 6= 0, hence L̇(X ∗ ) < 0 (since L̇ ≤ 0). Recall
that if we follow a solution X(·) going through X∗, then L̇(X ∗ ) is the derivative of the function
l(t) = L(X(t)) when X(t) = X ∗ . Thus, letting φt (X0 ) denote the value at time t of the solution
such that X(0) = X0 and l(t) = L(φt (X ∗ )) the value of L along the solution starting at X ∗ , we
have:
l0 (0) = L̇(X ∗ ) < 0.
Therefore, there exists t > 0 such that

l(t) = L(φt (X ∗ )) < L(X ∗ ).

Let ε = L(X ∗ ) − L(φt (X ∗ )) > 0. Since L and the time t map of the flow φt are continuous (see
Proposition 9.0.2), the function L ◦ φt is also continuous. Therefore, there exists a neighborhood
V of X ∗ such that for all X in V ,

L(φt (X)) ≤ L(φt (X ∗ )) + ε/2 = L(X ∗ ) − ε/2.

10.3. LYAPUNOV FUNCTION FOR AN EQUILIBRIUM 77

Since X(tn ) → X ∗ , there exists N ∈ N such that X(tN ) ∈ V (indeed, this holds as soon as N is
large enough). Therefore, L(φt (X(tN ))) ≤ L(X ∗ ) − ε/2, that is,

L(X(tN + t)) ≤ L(X ∗ ) − ε/2.

Since L decreases along trajectories, for all s ≥ tN + t, L(X(s)) ≤ L(X(tN + t)) ≤ L(X ∗ ) − ε/2.
Therefore,
lim sup L(X(s)) ≤ L(X ∗ ) − ε/2.
s→+∞

But since X(tn ) → L(X ∗ ) as n → +∞, it follows that

lim L(X(tn )) = L(X ∗ ) ≤ L(X ∗ ) − ε/2 < L(X ∗ ),

n→+∞

a contradiction. Thus, L̇(X ∗ ) = 0.

3) Let X(·) be a solution defined and bounded on [0, +∞[. Assume by contradiction that
d(X(t), E) does not go to 0 as t → +∞. Then there exists ε ≥ 0 and a sequence (tn ) going
to +∞ such that d(X(tn ), E) ≥ ε for all n, hence ||X(tn ) − Y || ≥ ε for all n and for all
Y ∈ E. Since the function X(·) is bounded on [0, +∞[, it follows that the sequence X(tn ) is also
bounded. Therefore, up to considering a subsequence,1 we may assume that X(tn ) converges:
∃X ∗ ∈ Rd , X(tn ) →n→+∞ X ∗ . But then we both have ||X ∗ − Y || ≥ ε for all Y ∈ E, by
taking limits in ||X(tn ) − Y || ≥ ε, and, due to 2), X ∗ ∈ E, a contradiction. This shows that
d(X(t), E) → 0 as t → +∞.
Assume now that E consists of isolated points and let X ∗ be an accumulation point of X(t)
as t → +∞ (such an accumulation point exists since X(t) is bounded in the neighborhood of
+∞). Recall that by 2), X ∗ ∈ E. We will show that X(t) → X ∗ . Indeed, otherwise there exists
ε > 0 such that:
∀T ∈ R, ∃t ≥ T, ||X(t) − X ∗ || ≥ ε. (10.2)
But since E consists of isolated points, there exists α > 0 such that:

∀X ∈ E, ||X − X ∗ || ≤ α ⇒ X = X ∗ . (10.3)

Let η = min(ε, α) and let T ∈ R. Due to (10.2), there exists s1 ≥ T such that ||X(s1 )−X ∗ || ≥ η.
But since X ∗ is an accumulation point of X(·) as t → +∞, there also exists s2 ≥ T such
that ||X(s2 ) − X ∗ || ≤ η, hence, by continuity of X(·), a time t between s1 and s2 such that
||X(t) − X ∗ || = η. Note that t ≥ T . We conclude that:

∀T ∈ R, ∃t ≥ T, ||X(t) − X ∗ || = η.

It follows that there exists a sequence (tn ) going to +∞ such that ||X(tn ) − X ∗ || = η for all n.
Since X(tn ) is bounded, up to considering a subsequence, we may assume that X(tn ) converges:
∃Y ∈ Rd , X(tn ) →n→+∞ Y . Moreover, ||Y − X ∗ || = η ≤ α and Y ∈ E. But by (10.3), this
implies Y = X ∗ , contradicting the fact that ||Y − X ∗ || = η.
4) Simply put 1) and 3) together.

10.3 Lyapunov function for an equilibrium

Let U be an open subset of Rd . Let X ∗ ∈ U be an equilibrium of (10.1). Let L : U → R be

differentiable.
1
The standard expression “up to considering a subsequence” should be understood as follows: there is a
subsequence (τn ) of (tn ) such that X(τn ) converges. This is to this subsequence that what follows applies. But
since we want to avoid to introduce new notation, we will forget about the initial sequence X(tn ) and use the
same piece of notation to denote its converging subsequence.
78 CHAPTER 10. LYAPUNOV FUNCTIONS

Definition 10.3.1. L is a Lyapunov function for X ∗ if:

(a) L(X ∗ ) = 0 and L(X) > 0 on U \{X ∗ } ; and2
(b) L̇(X) ≤ 0 on U .
It is a strict Lyapunov function for X ∗ if moreover:
(c) L̇(X) < 0 on U \{X ∗ }.

In the next proposition, Bd(K) denotes the boundary of the set K, that is the difference
between the closure and the interior of K.

Proposition 10.3.2. 1) If L is a Lyapunov function for X ∗ , then X ∗ is stable.

2) If L is a strict Lyapunov function for X ∗ , then X ∗ is asymptotically stable. Moreover,
if K ⊂ U is a compact set containing X ∗ in its interior and λ = min{X∈Bd(K)} L(X), then the
basin of attraction of X ∗ contains {X ∈ K | L(X) < λ}.
3) If L is a Lyapunov function for X ∗ and L̇ = 0 on U , then X ∗ is stable but not asymptot-
ically stable.

Proof. Let us start with a lemma:

Lemma 10.3.3. Assume that L is a Lyapunov function for X ∗ . Let K ⊂ U be a nonempty
compact set and let λ = min{X∈Bd(K)} L(X). Let A = {X ∈ K | L(X) < λ}. For any solution
(J, X(·)) such that X(0) ∈ A, we have: sup J = ∞ and X(t) ∈ K for all t ≥ 0.

The proof of the lemma is as follows: since K is bounded, by the explosion alternative, it suffices
to show that X(t) never leaves K forward in time. Assume by contradiction that this is not the
case. Then, since X(·) is continuous and K is closed, it can be shown that there is a time T ≥ 0
such that X(t) ∈ K on [0, T ] and X(T ) ∈ Bd(K). 3 Since X(t) ∈ K ⊂ U on [0, T ] and L̇ ≤ 0
on U , it follows that L(X(t)) decreases between 0 and T . Thus, L(X(T )) ≤ L(X(0)) < λ. Since
X(T ) ∈ Bd(K), this contradicts the definition of λ.
We now prove the proposition.
1) Assume that L is a Lyapunov function for X ∗ and let V be a neighborhood of X ∗ . We
want to show that there exists a neighborhood W of X ∗ such that if a solution starts in W , then
it remains in V forever. We first note that there exists ε > 0 such that K := B̄(X ∗ , ε) ⊂ V ∩ U
(where B̄(X ∗ , ε) is the closed ball of center X ∗ and radius ε). We have Bd(K) ⊂ K ⊂ U and
X∗ ∈ / Bd(K) (since X ∗ ∈ int(K)). By definition L is positive on Bd(K). Since Bd(K) is
compact and L is continuous, this implies that λ := min{L(X), X ∈ Bd(K)} is positive. Since
L(X ∗ ) = 0 and L is continuous, it follows that there exists α > 0 such that L(X) < λ for all
X ∈ B̄(X ∗ , α). Letting W = B̄(X ∗ , α), the lemma shows that if X(0) ∈ W , then for all t ≥ 0,
X(t) ∈ K hence X(t) ∈ V . Therefore, X ∗ is stable.
2) X ∗ is stable by 1) so it suffices to show that X ∗ is attracting. For ε > 0 small enough,
K := B̄(X ∗ , ε) is a subset of U . Fix such ε and let λ := min{X∈Bd(K)} L(X). For the same
reasons as in the proof of 1), λ > 0 and there exists α > 0 such that L(X) < λ for all
X ∈ V := B̄(X ∗ , α). Note that by definition of λ, this implies α < ε. Now let X(·) be a solution
such that X(0) ∈ V . It follows from the Lemma that X(t) ∈ K for all t ≥ 0. Thus, since K is
closed, any accumulation point of X(t) as t → +∞ is in K. Moreover, since K ⊂ U , the same
2
Assuming L(X ∗ ) = 0 is just a convention, to fix ideas, what is important is that X ∗ is a strict local minimum
of the function L; but we stick to this convention because this is not a loss of generality (we can always replace L
by L − L(X ∗ ), without affecting b) and c)), and because it is sometimes nicer to check that a quantity is positive
rather than greater than a fixed level L(X ∗ ).
3
Indeed, let T = sup à where à = {t ≥ 0, ∀s ∈ [0, T ], X(s) ∈ K}; note that à is nonempty (contains 0) and
upper bounded (by assumption), so that T is well defined; moreover à = [0, +∞[∩X −1 (K) hence à is closed
because K is closed and X(·) is continuous, so T ∈ Ã hence X(t) ∈ K on [0, T ], and in particular, X(T ) ∈ K.
Finally, X(T ) ∈ / int(K). Otherwise, by continuity of X(·), we would have X(s) ∈ K for all s in a neighborhood
of T , contradicting the definition of T . Thus X(T ) ∈ K\int(K) = cl(K)\int(K) = Bd(K).
10.4. GRADIENT SYSTEMS 79

arguments as in the proof of Proposition 10.2.2 show that X(t) converges to the set of points of
K such that L̇(X) = 0. Since L is a strict Lyapunov function for X ∗ and K ⊂ U , the only point
of K such that L̇(X) = 0 is X ∗ , therefore X(t) converges to X ∗ . It follows that X ∗ is attracting,
hence asymptotically stable. The above argument applied to any compact set K ⊂ U proves
the end of 2).
3) X ∗ is stable by 1). Thus, to show that it is not asymptotically stable, we must show
that it is not attracting. Let V be any neighborhood of X ∗ . We need to show that there
are solutions starting in V that do not converge to X ∗ . The proof of 1) shows that there
exists X0 ∈ V ∩ U \{X ∗ } (in fact a whole open set of such points) such that the solution
starting at X0 is defined and in U for all t ≥ 0. Since X0 ∈ U \{X ∗ }, L(X0 ) > 0, hence
L(X0 ) 6= L(X ∗ ). Moreover, since L̇ = 0 on U and X(t) ∈ U for all t ≥ 0, it follows that for all
t ≥ 0, L(X(t)) = L(X(0)) = L(X0 ) 6= L(X ∗ ). Therefore, as t → +∞, L(X(t)) cannot converge
towards L(X ∗ ), hence by continuity of L, X(t) cannot converge towards X ∗ . This concludes the
proof.

10.4 Gradient systems

A gradient system is a system of differential equations of the form X 0 (t) = −∇V (X(t)) for some
function V which we will assume to be C 2 so that the above system of differential equations is
C 1 . Note that V is defined only up to an additive constant, in the sense that for any constant
C, the function Ṽ = V + C generates the same gradient system.
Proposition 10.4.1. Let V : Rd → R be of class C 2 . For the gradient system X 0 (t) =
−∇V (X(t)):

1. V is a global Lyapunov function. Moreover, V̇ (X) = 0 if and only if X is an equilibrium.

2. If X ∗ is an isolated equilibrium and a strict local minimum of V , then X ∗ is asymptotically
stable.

Proof. 1) Let F (X) = −∇V (X), so that the equation reads X 0 (t) = F (X(t)). We have:
V̇ (X) = ∇V (X) · F (X) = −F (X) · F (X) = −||F (X)||2 , hence V̇ (X) ≤ 0 with equality if
and only if F (X) = 0, that is, if X is an equilibrium.
2) Assume first that V (X ∗ ) = 0. Since X ∗ is an isolated equilibrium, there exists a neighbor-
hood U of X ∗ in which X ∗ is the only equilibrium. Thus, it follows from 1) that for every X in
U , we have V̇ (X) ≤ 0, and V̇ (X) < 0 if X 6= X ∗ . Moreover, since X ∗ is a strict local minimum
of V , we may also assume, up to considering a smaller neighborhood, that V (X) > V (X ∗ ) = 0
on U \{X ∗ }. Thus, V is a strict Lyapunov function for X ∗ and the result follows from point 2)
in Proposition 10.3.2. If V (X ∗ ) 6= 0, then apply the above reasoning to Ṽ = V − V (X ∗ ), which
generates the same gradient system and satisfies Ṽ (X ∗ ) = 0.

10.5 Hamiltonian systems in R2

An Hamiltonian system in R2 is a system of the form:

∂H

 x0 (t) = (x(t), y(t))


∂y
 y 0 (t) = − ∂H (x(t), y(t))


∂x
with H : R2 → R of class C 2 . The function H is called the Hamiltonian (or Hamiltonian
function) of the system. It is defined up to an additive constant.
80 CHAPTER 10. LYAPUNOV FUNCTIONS

Proposition 10.5.1. Consider an Hamiltonian system in R2 with Hamiltonian function H.

1. H is a constant of motion, i.e. a quantity that is conserved throughout the motion.

2. If X ∗ is a strict local minimum of H, then X ∗ is stable but not asymptotically stable.

 
Proof. 1) The system may be written X 0 (t) = F (X(t)) with F (x, y) = ∂H ∂H
∂y (x, y), − ∂x (x, y) .
Thus,
∂H ∂H ∂H ∂H
   
Ḣ = ∇H · F = , · ,− = 0.
∂x ∂y ∂y ∂x
2) Assume first H(X ∗ ) = 0. Since X ∗ is a strict local minimum of H, there exists a neighborhood
U of X ∗ such that H(X) > 0 on U \{X ∗ }. Moreover, Ḣ = 0 on Rd , hence on U . The result
then follows from point 3 of Proposition 10.3.2. If H(X ∗ ) 6= 0, apply the same reasoning to
H̃ = H − H(X ∗ ), which generates the same Hamiltonian system and satisfies H̃(X ∗ ) = 0.