Notes 4190

Download as pdf or txt
Download as pdf or txt
You are on page 1of 73

Math 5190: Lecture notes.

1
Math 5190: lecture notes. 2

1. Geometric Preliminaries
1.1. Differential equations and their solutions. We begin by re-
viewing the concepts of an ordinary differential equation and initial
value problem. Vector fields, integral curves, and flows furnish us with
a geometric formulation of these concepts.
Let U ⊂ Rn+1 be an open domain and V : U → Rn a C 1 (continu-
ously differentiable) vector valued function.
Definition 1.1. An ordinary differential equation is a constraint of the
form
φ′ (t) = V(t, φ(t)), (1)
where φ : I → U is a continuously differentiable function called the
solution of the equation. An initial value problem (IVP) is an ODE
together with a constraint of the form
φ(0) = x0 , (2)

where x0 ∈ U is a given initial position.


Definition 1.2. We call the ODE (1) autonomous if V does not de-
pend on t. The autonomization of (1) is the autonomous ODE in n + 1
dependent variables having the form
ξ ′ (t) = 1, (3)

φ (t) = V(ξ(t), φ(t)) (4)
Using autonomization no generality is lost if we develop the general
aspects of ODE theory strictly in terms of autonomous equations.
Intuitively, an ODE is an assignment of a velocity to every point in
space and time. A solution curve represents the motion of a particle
whose velocity is determined by the time and the particle’s position.
An autonomous ODE simply describes a velocity field that depends
only on position and is independent of time. Therefore, it is useful to
consider an autonomous ODE as a vector (velocity) field.
In our discussion, it is important to make a distinction between
points and vectors. The former should be considered as elements of
some domain U ⊂ Rn , whereas the latter are elements of the vector
space Rn . Geometrically, it makes sense to take linear combinations of
vectors, but there is little profit in trying to “add” points. Likewise,
in analysis of ODEs, the zero vector plays a very special role. How-
ever, there is nothing special about the point (0, . . . , 0); it is just an
arbitrarily chosen origin, and may not even be included in U.
Math 5190: lecture notes. 3

Thus, a vector field is a function V : U → Rn that assigns to every


point of U a vector in Rn . We will also write a vector field as
n
X
V= V i (x1 , . . . , xn )ei , x = (x1 , . . . , xn ) ∈ U;
i=1

here e1 , . . . , en is the standard basis of Rn and V 1 , . . . , V n are the scalar


components of a vector in Rn . If we assume that the components V i are
continuously differentiable functions, we will speak of a C 1 vector field.
If the components are analytic (expressible as power series) functions
we will speak of an analytic vector field.
For f : U → R a differentiable function, we define V[f ], the direc-
tional derivative of f with respect to V , to be the function
n
X
V[f ] = V i Di f (5)
i=1

For this reason, we identify vector fields with first order differential
operators, and write
n
X ∂
V= V i (x1 , . . . , xn ) .
i=1
∂xi
In this formulation the standard vectors e1 , . . . , en correspond to the
partial derivative operators ∂/∂x1 , . . . , ∂/∂xn .
As we already mentioned a vector field is exactly the same as an
autonomous ODE. An integral curve is a parameterized curve φ : I →
U, where I ⊂ R is an open interval, such that
φ′ (t) = V(φ(t)), t ∈ I. (6)
In other words, an integral curve is a particular solution of the above au-
tonomous ODE. The general solution of an autonomous, n-dimensional
ODE depends on n constants of integration. If we judiciously include
the constants of integration as parameters in a general solution, we
obtain something called a flow.
Let U ⊂ Rn be an n-dimensional domain, Û ⊂ R × U an (n + 1)-
dimensional domain, and Φ : Û → U a C 1 function.
Definition 1.3. We say that Φ is a flow if it satisfies the following
conditions.
(A) For all x ∈ U, the quantity Φ(0, x) is defined and equal to x;
(B) We have
Φ(s, Φ(t, x)) = Φ(s + t, x); (7)
Math 5190: lecture notes. 4

(C) For a fixed x ∈ U the set


Ux = {t ∈ R : (t, x) ∈ Û }; (8)
is an open interval in R.
If in addition to the above 3 conditions,
(D) the RHS of (10) is defined whenever the LHS is,
we call Φ(t, x) a maximal flow.
A flow is a special way to specify a general solution of an ODE, one
where the constants of integration represent initial position. What is
the relationship between vector fields and flows?
Definition 1.4. Let V : U → Rn be a vector field and Φ(t, x), x ∈ U
a flow. We say that V generates the flow if
V(x) = (D1 Φ)(0, x). (9)
Note: above D1 Φ denotes the partial derivative with respect to the
first, “time” variable.
For each t ∈ R let us set Ut = {x ∈ Rn : (t, x) ∈ Û} and define the
transformation Φt : Ut → U by
Φt (x) = Φ(t, x), (t, x) ∈ Û.
Condition (A) can be restated by saying that Φ0 is the identity trans-
formation on U. Condition (B) can be restated as the 1-parameter
group law
Φs+t = Φs ◦ Φt . (10)
For this reason, we often refer to a flow as a 1-parameter group gener-
ated by a vector field
Theorem 1.5 (Existence). Let V : U → Rn be a C 1 vector field.
Then, there exist a C 1 flow generated by V.
Theorem 1.6 (Uniqueness). Let V : U → Rn be a C 1 vector field.
Any two flows generated by V agree on some open neighbourhood of
U ⊂ R × U. There exists a unique maximal flow generated by V.
We will present the proof later. Here, we present some examples.
Example 1.7. Consider the differential equation
1
ẋ = , x > 0.
x
The general solution is √
x = K + 2t.
Math 5190: lecture notes. 5

The corresponding vector field is


1 ∂
V = , x > 0.
x ∂x
The domain of the vector field is U = {x ∈ R : x > 0}. The corre-
sponding flow is given by

Φ(t, x) = x2 + 2t.
The domain of the flow is
1
Û = {(t, x) : x > 0 and − x2 < t < +∞}.
2
Note that the limit of Φ(t, x) as t → +∞ does not exist, while the limit
of Φ(t, x) as t → −x2 /2 is equal to 0, a value outside the domain U.
Example 1.8. Let’s consider the non-autonomous ODE
dy
= x + y.
dx
The general solution is obtained by rewriting the ODE as
de−x y
ex = x,
dx
separating variables and integrating. We obtain
y = −x − 1 + Cex .
where C is a constant of integration.
Let us rewrite the general solution as a flow. First, we autonomize
the equation by rewriting it as
ẋ = 1,
ẏ = x + y,
and writing the solution as
x = x0 + t,
y = −(x0 + t) − 1 + Cex0 +t .
We set t = 0 in the second equation and obtain
y0 = −x0 − 1 + Cex0 ,
or equivalently,
C = e−x0 (y0 + x0 + 1).
Therefore, the general solution, written as a flow, is
x = x0 + t
y = (y0 + x0 + 1)et − (x0 + t) − 1.
Math 5190: lecture notes. 6

The domain of the flow is all real (t, x, y). A straightforward calculation
shows that condition (10) is satisfied.

Finally, let us discuss the relationship between the general solution to


an ODE and a flow. We would like to say that a flow is the same thing
as a general solution where the initial values serve as the constants of
integration. In one direction, we have the following
Proposition 1.9. Let Φ(t, x) be a flow generated by a C 1 vector field
V : U → Rn . Then,
∂Φ(t, x)
= V(Φ(t, x)).
∂t
Proof. Homework. 

For the other direction, we will make use of the following fundamental
Theorem regarding solutions of initial value problems.

Theorem 1.10. Let V : U → Rn be a C 1 vector field and x0 ∈ U a


fixed position. Then, there exists a C 2 curve φ(t) such that

φ′ (t) = V(φ(t)),
φ(0) = x0 .

The curve in question is locally unique. This means that all such curves
agree in some neighborhood of t = 0.

Proposition 1.11. Let V(x) be a vector field, and let Φ(t, x) be a


general solution of the ODE ẋ = V(x); i.e.;
∂Φ(t, x)
= V(Φ(t, x)).
∂t
If in addition, Φ(0, x) = x, then Φ(t, x) is a flow.

Proof. Let us establish that Φ satisfies the 1-parameter group law (B).
Fix a position x0 and a time t1 set

x1 = Φ(t, x0 ),

and define curves

φ(t) = Φ(t + t1 , x0 ),
ψ(t) = Φ(t, x1 ).
Math 5190: lecture notes. 7

Observe that
φ̇(t) = Φ̇(t + t1 , x0 ) = V(Φ(t + t1 , x0 )) = V(φ(t)),
ψ(0) = x1 ,
ψ̇(t) = Φ̇(t, x1 ) = V(Φ(t, x1 )) = V(ψ(t)),
ψ(0) = x1 .
Therefore, by the uniqueness of solutions to initial value problems, we
must have
Φ(t + t1 , x0 ) = φ(t) = ψ(t) = Φ(t, Φ(t1 , x0 )),
as was to be shown. 
Math 5190: lecture notes. 8

2. The reparameterization theorem.


A vector can be informally defined as a quantity having the proper-
ties of magnitude and direction. When the motion of a hypothetical
particle is governed by a field of velocity vectors, two distinct aspects
of the motion are determined. First, is the trajectory of the particle,
that is to say the geometric shape traced out by the particle’s motion.
Second, is the speed with which the particle traverses that trajectory.
If we modify a field of velocities by changing the magnitude, but not
the direction of the vectors, it is reasonable to expect that the corre-
sponding trajectories retain their shape, but that the hypothetical test
particles move along these trajectories with a different speed.
In this regard, it is important to distinguish between a curve and
its parameterization. For example, consider the unit semicircle in R2
given by
x2 + y 2 = 1, y > 0.
Here are two parameterizations of this curve:

x = s, y = 1 − s2 , −1 < s < 1,
and
x = cos t, y = sin t, 0 < t < π.
The two parameterizations are related by an invertible reparameteri-
zation function, namely
s = cos t, 0 < t < π.
In general, let φ(t) ∈ Rn be a parameterization of a curve in n-
dimensional space. We say that ψ(s) ∈ Rn is a reparameterization
of the same curve if there exists a continuously differentiable function
t = τ (s) such that τ ′ (s) 6= 0 for all s and such that ψ(s) = φ(τ (s)).
Let A : U → Rn be a C 1 vector field. Let g : U → R be a C 1 ,
nowhere vanishing function: g(a) 6= 0, a ∈ U. Scaling A by g gives us
a new vector field, namely B = gA.
Theorem 2.1. The integral curves of A and B are related by a repa-
rameterization.
Proof. Suppose that B = gA. Consider the integral curves φ(t), ψ(t) of
A, B, respectively, both originating at some fixed initial point x ∈ U;
φ(0) = ψ(0) = x.
We need to construct a reparameterization function τ (s) such that
ψ(s) = φ(τ (s)).
Math 5190: lecture notes. 9

We will do this in a way that incorporates dependence on the initial


point x explicitly.
Let Φ(t, x) be the flow corresponding to A(x). For each t, x in the
domain of Φ(t, x), set
Z t
du
σ(t, x) = .
0 g(Φ(u, x))

For x ∈ U and t1 , t2 ≥ 0 set


y = Φ(t1 , x), s1 = σ(t1 , x), s2 = σ(t2 , y).
We have
t1 +t2
du
Z
σ(t1 + t2 , x) = (11)
0 g(Φ(u, x))
Z t1 Z t1 +t2
du du
= + (12)
0 g(Φ(u, x)) t1 g(Φ(u, x))
Z t2
dv
= σ(t1 , x) + (13)
0 g(Φ(v, y))
= s1 + s2 , (14)
The step from (12) to (13) is justified by making a change of variables
u = v + t1 and by noting that
Φ(u, x) = Φ(v + t1 , x) = Φ(v, y), 0 ≤ v ≤ t2 .
By assumption,
1
σ̇(t, x) = 6= 0.
g(Φ(t, x))
Hence, by the inverse function theorem there exists an inverse function
τ (s, x) such that
τ (σ(t, x), x) = t.
We now define
Ψ(s, x) = Φ(τ (s, x), x), (15)
and claim that Ψ(s, x) is the flow determined by B. Since σ(0, x) = 0,
we have τ (0, x) = 0, whence
Ψ(0, x) = Φ(0, x) = x.
By (14) we have
s1 + s2 = σ(t1 + t2 , x)
Hence, by (11), we have
τ (s1 + s2 , x) = t1 + t2 = τ (s1 , x) + τ (s2 , x),
Math 5190: lecture notes. 10

and
y = Φ(τ (s1 , x), x) = Ψ(s1 , x).
Therefore,
Ψ(s1 + s2 , x) = Φ(τ (s1 + s2 , x), x)
= Φ(τ (s2 , y) + τ (s1 , x), x)
= Φ(τ (s2 , y), Φ(τ (s1 , x), x))
= Φ(τ (s2 , y), y)
= Ψ(s2 , y)
= Ψ(s2 , Ψ(s1 , x)).
Finally, observe that
Ψ̇(0, x) = τ̇ (0, x)Φ̇(0, x)
1
= A(x)
σ̇(0, x)
= g(x)A(x)
= B(x).
Therefore, Ψ is the flow generated by B. However, every integral curve
of A, B, respectively, has the form
φ(t) = Φ(t, x), ψ(t) = Ψ(t, x), x ∈ U.
By construction, all such curves are related by the reparameterization
shown in (15). 
Example 2.2. Consider the non-autonomous scalar differential equa-
tion
dy x
= − , y > 0.
dx y
The general solution is √
y = K − x2 .
Autonomizing the above gives the planar ODE
x
ẏ = − , ẋ = 1,
y
which corresponds to the vector field
∂ x ∂
A= − .
∂x y ∂y
This vector field determines the following flow:
 p  p
Φ(t, x, y) = x + t, x2 + y 2 − (t + x)2 , |t + x| < x2 + y 2 , y > 0
Math 5190: lecture notes. 11

Multiplying A by y yields the vector field


∂ ∂
B=y −x .
∂x ∂y
The corresponding ODE is linear, with a linear flow, namely
  
cos(s) sin(s) x
Ψ(s, x, y) = .
− sin(s) cos(s) y
Let us construct the reparameterization function, as per the proof of
the above theorem. We have
p
g(x, y) = y, g(Φ(t, x, y)) = x2 + y 2 − (x + t)2 .
Hence,
t
du
Z
σ(t) = p
0 x + y − (x + u)2
2 2

t+x
dv
Z
= p
x x2 + y2 − v2
 
t+x
= sin−1 − s0 ,
r
where
p x
r= x2 + y 2 , s0 = sin−1
.
r
This is the position dependent reparameterization function from Ψ to
Φ. The reparameterization function from Φ to Ψ is given by the inverse,
namely
τ (s, x, y) = r sin(s + s0 ) − x,
= r sin(s) cos(s0 ) + r cos(s) sin(s0 ) − x,
= y sin(s) + x cos(s) − x.
Composing with the reparameterization function with the flow gives
 p 
Φ(τ (s, x, y), x, y) = x + τ (s, x, y), x2 + y 2 − (x + τ (s, x, y))2
p
= (x cos(s) + y sin(s), x2 + y 2 − (x cos(s) + y sin(s))2 )
p
= (x cos(s) + y sin(s), (y cos(s) − x sin(s))2 )
= (x cos(s) + y sin(s), y cos(s) − x sin(s)),
= Ψ(s, x, y).
Next, let us check the domains of the reparameterizations. The time
domains for the first flow Φ are given by −r − x < t < r − x. The
Math 5190: lecture notes. 12

domain of t cannot be extended further because Φ(t, x, y) tends to (r, 0)


as t → r − x and tends to (−r, 0) as t → −r − x. However these limits
are outside the domain of our vector field.
What about the time domains of the second flow Ψ? Here we must
ask the question: what is the domain of B? If the domain is all of
R2 , then Ψ(s, x, y) is defined for all s, x, y. However, in order for us to
compare A and B, we must restrict the domain of B to the domain
of A, namely the upper half plane {(x, y) ∈ R2 : y > 0}. Hence, the
constraint on s becomes

−x sin(s) + y cos(s) > 0, y > 0,

or equivalently,

π π
− − s0 < s < − s0 , s0 = tan−1 (x/y).
2 2

The domain of
 
−1 t+x
σ(t, x, y) = sin − s0
r

is the interval −r − x < t < r − x. The image of s = σ(t, x, y) is


the interval − π2 − s0 < s < π2 − s0 . Hence the latter interval becomes
the domain of the inverse function τ (s, x, y), while the former interval
becomes the image of τ (s, x, y).
¶ x ¶
€€€€€€€€ - €€€€ €€€€€€€€
¶x y ¶y
2

1.75

1.5

1.25

0.75

0.5

0.25

-1.5 -1 -0.5 0 0.5 1 1.5 2


Math 5190: lecture notes. 13

¶ ¶
y €€€€€€€€ -x €€€€€€€€
¶x ¶y
2

1.5

0.5

-0.5

-1

-1.5

-1.5 -1 -0.5 0 0.5 1 1.5 2


Math 5190: lecture notes. 14

3. Diffeomorphisms and coordinate systems.


Definition 3.1. Let U, V ⊂ Rn be domains. A diffeomorphism F :
U → V is an invertible, continuously differentiable transformation such
that the inverse transformation is also continuously differentiable.
Equivalently, we can consider n functions of n variables, F i (x1 , . . . , xn ),
i = 1, . . . , n so that equations
y i = F i (x1 , . . . , xn ), i = 1, . . . , n (16)
have the property that if we restrict y ∈ V , then a solution x ∈ U
exists and is unique. The transformation of the givens y 1, . . . , y n to the
solutions x1 , . . . , xn defines inverse functions
xi = Gi (y 1 , . . . , y n), i = 1, . . . , n. (17)
To have a diffeomorphism we require that all partial derivatives
∂y j
Di F j (x1 , . . . , xn ) = , (18)
∂xi
∂xi
Dj Gi (y 1, . . . , y n ) = j , i, j = 1, . . . , n (19)
∂y
exist and are continuous.
Definition 3.2. Let F : U → Rn be a continuously differentiable trans-
formation. The Jacobian transformation J F : U → Matn R is defined
to be the matrix of the partial derivatives of F ; to wit,
D1 F 1 . . . Dn F 1
 

J F =  ... ..
.
..  .
. (20)
D1 F n . . . Dn F n
Note: here Matn R denotes the vector space of n × n real matrices.
Proposition 3.3 (Linear approximation). Let F : U → Rn be a C 1
transformation. Then,
F (x) = F (x0 ) + J F (x0 )(x − x0 ) + o(kx − x0 k), x, x0 ∈ U. (21)
Here, the little o notation means that the difference of the LHS and the
RHS is a remainder function R(x, x0 ) such that
R(x, x0 )
→ 0 as x → x0 .
kx − x0 k
We can also restate the above result by saying that the linear approx-
imation to y = F (x) at x0 is the affine function
ŷ = y 0 + M(x − x0 ),
Math 5190: lecture notes. 15

where M = J F (x0 ) is the indicated n × n constant matrix and where


y 0 = F (x0 ). The assertion that F is a C 1 function is equivalent to the
condition that
y − ŷ = o(kx − x0 k).

Proposition 3.4. Let F : U → V be a diffeomorphism. Let x0 ∈ U


be given and set y 0 = F (x0 ). Then J F (x0 ) and J G(y 0 ) are inverse
matrices. Indeed, the linear approximation to the inverse function x =
G(y) at y 0 = F (x0 ) is given by

x̂ = x0 + M −1 (y − y 0 ),

where M = J F (x0 ), as before.


We can summarize the above result by saying that linear approxima-
tions of inverse functions are inverses of each other. In other words,
the affine function x 7→ ŷ is the inverse of the affine function y 7→ x̂.

Theorem 3.5 (Inverse Function Theorem). If F : U → V is a diffeo-


morphism, then
det J F (a) 6= 0 (22)
for all a ∈ U. Conversely, if F is continuously differentiable and if
(22) holds for a particular a ∈ U, then F is a diffeomorphism provided
we restrict the domain to a sufficiently small neighborhood of a.

Example 3.6. Consider the function F (x) = x3 . This function is in-


vertible, with inverse G(x) = x1/3 . It is continuously differentiable:
F ′ (x) = 3x2 . However, the inverse isn’t: G′ (x) = (1/3)x−2/3 is unde-
fined at x = 0. Therefore F (x) is not a diffeomorphism of the real line.
However, if we restrict the domain to, for example to U = {a : a > 0},
we obtain a diffeomorphism from the half-line U to U.

Example 3.7. The following example illustrates that there exist dif-
feomorphisms belonging strictly to class C 1 . Consider the function

y = F (x) = sgn(x)x2 + 2x.

We claim that this is a diffeomorphism of the real line R. We have


F ′ (x) = 2|x| + 2; the derivative exists and is continuous at all x ∈ R.
The inverse function
(√
y + 1 − 1 if y ≥ 0;
x = G(y) = √ (23)
1 − 1 − y if y < 0
Math 5190: lecture notes. 16

is also continuously differentiable for all y ∈ R. Indeed,


1

 2√y + 1 if y ≥ 0;


G′ (y) = (24)
1
 √ if y < 0.


2 1−y
However, quite clearly F (x) is does not belong to class C 2 ; the second-
order derivative does not exist at x = 0.
Example 3.8. Consider the continuously differentiable function F (x) =
x2 . Observe that F ′ (x) = 2x vanishes at x = 0. Hence, F (x) does not
define a diffeomorphism of the real line. However, F ′ (x) 6= 0 for x > 0.
Therefore,√ we can define a continuously differentiable inverse, namely
G(x) = x, x > 0 by restricting the domain. The resulting function
is a diffeomorphism of the half-line {a ∈ R : a > 0}.
Example 3.9. Consider the transformation of R2 ∼ = C given by
F (x, y) = exp(x + i y), x, y ∈ R,
or equivalently by
u = ex cos y, v = ex sin y.
We calculate
∂u ∂u
 
 x x

 ∂x ∂y  e cos y −e sin y
 ∂v ∂v  = ex sin y ex cos y ,
J F (x, y) =  

∂x ∂y
2x
det J F (x, y) = e 6= 0.
Hence, by the inverse function theorem, an inverse exists, but only
locally. For example we could take
1
x = log(u2 + v 2 ), y = tan−1 (v/u), u > 0.
2
In this way we obtain a diffeomorphism F : U → V of the planar
regions
U = {a ∈ R2 : −π/2 < a2 < π/2},
V = F (U) = {a ∈ R2 : a1 > 0}.
Speaking informally, a system of n-dimensional coordinates is a way
of assigning n numbers to points in n-dimensional space in such a way
that (1) every point has a unique “address” and so that (2) we can
take partial derivatives of functions with respect to the coordinates in
question.
Math 5190: lecture notes. 17

Thus, we require n functions y i = F i (x1 , . . . , xn ), i = 1, . . . , n where


the domain U is restricted in such a way that the mapping
a 7→ F (a), a∈U
is one-to-one. Condition (2) requires that the n × n partial derivatives
i j n

∂x /∂y ij=1 be defined and continuous. Since the n × n matrix in
question is the inverse of the matrix ∂xi /∂y j ij , our requirement is


that the Jacobian matrix J F be invertible for all points in U.


In other words, a coordinate system requires the same information
as a diffeomorphism, but this information is interpreted differently. We
regard a diffeomorphism as a transformation of position. However, a
change of coordinates does not alter position, but rather transforms
the address of the position.
In this regard, Rn is equipped with a special coordinate system that
corresponds to the identity diffeomorphism. These are the standard
coordinates, which we will denote by x1 , . . . , xn . Thus x1 is a function
that maps an element of Rn to its first component, etc. However, in
deference to tradition we will use x = x1 , y = x2 to denote the standard
coordinates of R2 and use x = x1 , y = x2 , z = x3 to denote the standard
coordinates of R3 .
Example 3.10. Consider the usual polar coordinates r, θ, where
x = r cos θ, y = r sin θ. (25)
If we write (x, y) = G(r, θ), we define a transformation G : R → R2 , 2

but that transformation fails to be a diffeomorphism; it is not one-


to-one. In order to assign a unique polar coordinate address to every
point we must solve for r, θ. This requires us to restrict the domain,
somehow. One way to do this is to set
p
r = x2 + y 2, θ = Arg(x + iy). (26)
Here Arg is the principal argument function whose domain is the com-
plex plane minus the non-positive x-axis; i.e. the set
U = {(a1 , a2 ) ∈ R2 : a1 > 0 or a2 6= 0}. (27)
If we now write (r, θ) = F (x, y), we define a diffeomorphism F : U → V
where
V = {(a1 , a2 ) ∈ R2 : a1 > 0, −π < a2 < π}. (28)
Other choices of domain are possible, of course. Each of these choices
corresponds to a distinct diffeomorphism.
Math 5190: lecture notes. 18

4. Transformation of vector fields.


Let F : U → V be a diffeomorphism of domains U, V ⊂ Rn . Let
y j = F j (x1 , . . . , xn ) be the corresponding change of coordinates. Let’s
consider the meaning of the symbols ∂y j /∂xi and ∂xi /∂y j . For the
former, we write

∂y j ∂F j (x1 , . . . , xn )
i
= i
= (Di F j )(x1 , . . . , xn ).
∂x ∂x

Thus, ∂y j /∂xi is the same thing as the function Di F j : U → R. Let


G : V → U be the inverse transformation. Since xi = Gi (y 1 , . . . , y n),
we would like to be able to write

∂xi ∂Gi (y 1 , . . . , y n )
=
∂y j ∂y j
= (Dj Gi )(y 1 , . . . , y n )
= (Dj Gi ◦ F )(x1 , . . . , xn ). (29)

Definition 4.1. To that end, we define ∂/∂y j , j = 1, . . . , n to be the


vector field on the domain U whose component in the ei direction is
the function Dj Gi ◦ F : U → R; in other words,

n
X ∂xi ∂

= , j = 1, . . . , n, (30)
∂y j i=1
∂y j ∂xi

where ∂xi /∂y j is the function shown in (29). We call these the funda-
mental vector fields corresponding to the coordinates y 1 , . . . , y n. Given
a function u = f (x1 , . . . , xn ), we define ∂u/∂y j to be the directional
derivative of f with respect to the j th fundamental vector field.

Using the above definitions we can express the multi-variable chain rule
in the familiar fashion, namely

n
X ∂u ∂xi
∂u
= , j = 1, . . . , n. (31)
∂y j i=1
∂xi ∂y j

Example 4.2. Let (r, θ) = F (x, y) be the transformation (26) from


Cartesian to polar coordinates. Let us calculate the fundamental vector
Math 5190: lecture notes. 19

fields ∂/∂r and ∂/∂θ. We calculate

∂x x x
= cos θ = = p ,
∂r r x2 + y 2
∂y y y
= sin θ = = p ,
∂r r x + y2
2

∂ ∂x ∂ ∂y ∂
= +
∂r ∂r ∂x ∂r ∂y
x ∂ y ∂
=p +p .
x2 + y 2 ∂x x2 + y 2 ∂y

Similarly,

∂x
= −r sin θ = −y,
∂θ
∂y
= r cos θ = x,
∂θ
∂ ∂x ∂ ∂y ∂
= +
∂θ ∂θ ∂x ∂θ ∂y
∂ ∂
= −y +x .
∂x ∂y

Proposition 4.3. Let F : U → V be a diffeomorphism and y j =


F j (x1 , . . . , xn ) the corresponding change of coordinates. Then,
n
X ∂y j ∂

= , i = 1, . . . , n. (32)
∂xi j=1
∂xi ∂y j

Proof. Let G : V → U be the inverse transformation. By definition,


G ◦ F is the identity transformation. The Jacobian matrix of the
identity transformation is the identity matrix. Hence, by the chain
rule
 1
∂x ∂x1
 1  
∂y ∂y 1
 ∂x1 . . . ∂xn   ∂y 1 . . . ∂y n 
 . ..   . .. 
.. ..
JF =   .. . .  and J G ◦ F =  .. . . 
  
 ∂y n ∂y n   ∂xn ∂xn 
. . . . . .
∂x1 ∂xn ∂y 1 ∂y n

are inverse matrices. We obtain (32) by inverting the relations (30). 


Math 5190: lecture notes. 20

Proposition 4.4. The fundamental vector field ∂/∂y j , j = 1, . . . , n is


the unique vector field A : U → Rn that satisfies
(
1 k = j,
A[y k ] = δjk = (33)
0 k 6= j

Proof. By the definition (30), we have


n
∂y k X ∂xi ∂y k
= .
∂y j i=1
∂y j ∂xi

Above, we established that the matrix of partials ∂xi /∂y j is inverse to


the matrix ∂y k /∂xi . This establishes that the matrix formed by the
partials ∂y k /∂y j is the n × n identity matrix.
To prove the converse consider a vector field
n
X ∂
A= Ai (x1 , . . . , xn ) ,
i=1
∂xi

such that (33) holds. Hence, the functions Ai , i = 1, . . . , n satisfies the


system of equations
 1 
∂y ∂y 1  
 ∂x1 . . . ∂xn  A1
 .
.. ..  .
.   ..  = ej .
 . .
 .

 ∂y n ∂y n  An
. . .
∂x1 ∂xn
Using the inverse matrix of partials to solve the above gives
∂xi
Ai (x1 , . . . , xn ) = .
∂y j

Note: δji is known as the Kronecker symbol.
Example 4.5. Returning to the example of polar coordinates, we have
∂r x ∂r y
=p =p
∂x x2 + y 2 ∂y x2 + y 2
∂θ −y ∂θ x
= 2 , = 2 .
∂x x + y2 ∂y x + y2
Note: here we used the fact that
θ = Arg(x + iy) = tan−1 (y/x) + const.
Math 5190: lecture notes. 21

Observe that the following matrices are inverses:


∂r ∂r
  x y 
p p
 ∂x ∂y  x2 + y 2 x2 + y 2 
J F (x, y) = 
 ∂θ
=
−y x ,
∂θ 

∂x ∂y x2 + y 2 x2 + y 2

∂x ∂x
  x 
p −y
 ∂r ∂θ   x2 + y 2 
(J G ◦ F )(x, y) =  =
 ∂y ∂y   y
,
x

p
2
x +y 2
∂r ∂θ

Previously, we showed that

 x 
    p −y
∂ ∂ ∂ ∂  x2 + y 2 
= 
y

∂r ∂θ ∂x ∂y 
p x

x2 + y 2

Right-multiplying the above relation by the inverse matrix gives

 x y 
    p p
∂ ∂ ∂ ∂ x2 + y 2 x2 + y 2 
= .

∂x ∂y ∂r ∂θ
 −y x
x + y2
2 x2 + y 2

Consequently, we have

∂r ∂ p 2 2

= x +y
∂r ∂r !
x ∂ y ∂ p 2 
= p +p x + y2
x2 + y 2 ∂x x2 + y 2 ∂y
= 1.
Math 5190: lecture notes. 22

Similar calculations show that


!
∂θ x ∂ y ∂
tan−1 (y/x)

= p +p
∂r x2 + y 2 ∂x x2 + y 2 ∂y
= 0,
∂r ∂ p 2 2

= x +y
∂θ ∂θ
 
∂ ∂ p 2 
= −y +x x + y2
∂x ∂y
= 0,
 
∂θ ∂ ∂
tan−1 (y/x)

= −y +x
∂θ ∂x ∂y
= 1.
Let A : U → Rn be a vector field on a domain U ⊂ Rn . The standard
vector fields ∂/∂xi = ei , i = 1, . . . , n are a basis, and hence A has a
unique expression of the form
n
X ∂
A= Ai (x1 , . . . , xn ) .
i=1
∂xi

Now let y j = F j (x1 , . . . , xn ) be another system of coordinates on the


same domain. Since the fundamental vector fields ∂/∂y j , j = 1, . . . , n
are also a basis, we can also express the given vector field as
n
X ∂
A= B j (y 1, . . . , y n ) .
j=1
∂y j

Definition 4.6. We call the functions B j : V → R the components of


A relative to coordinates y 1 , . . . , y n .
In order to understand the geometry underlying the transformation
from components Ai : U → R to the components B j : V → R, we need
to introduce the push-forward transformation.
Definition 4.7. Let F : U → V be a C 2 diffeomorphism, and let
A : U → Rn be a C 1 vector field. We call the C 1 vector field F ∗ A :
V → Rn , defined by
F ∗ A = (J F · A) ◦ F −1 . (34)
the push-forward of A.
Note: the dot in J F · A refers to matrix vector multiplication.
Math 5190: lecture notes. 23

Proposition 4.8. Let F , A be as above, and let B = F ∗ A. Let Ai :


U → R and B j : V → R be the component functions of the vector fields
A and B. Letting y j = F j (x1 , . . . , xn ), we have
n n
n ∂ ∂
X X
i 1
A= A (x , . . . , x ) i = B j (y 1 , . . . , y n) j . (35)
i=1
∂x j=1
∂y
In particular, if G : V → U denotes the inverse diffeomorphism and
∂/∂ x̂j denotes to the constant vector field in direction ej on the domain
V , then
∂ ∂
j
= G∗ j , j = 1, . . . , n,
∂y ∂ x̂
Example 4.9. Let (r, θ) = F (x, y), where F : U → V as before.
Above, we showed that
∂ ∂r ∂ ∂θ ∂
= +
∂x ∂x ∂r ∂x ∂θ
x ∂ y ∂
=p − 2 2
x + y ∂r x + y ∂θ
2 2

∂ sin θ ∂
= cos θ − .
∂r r ∂θ
Letting x̂, ŷ denote the standard coordinates on the domain V , we have
∂ ∂ sin ŷ ∂
F∗ = cos ŷ − .
∂x ∂ x̂ x̂ ∂ ŷ
Let us calculate this result directly. Using our previous calculations,
 x 
p
2 2
(J F · e1 )(x, y) =  x−y+ y  .

x2 + y 2
Hence,
((J F · e1 ) ◦ G)(x̂, ŷ) = (J F · e1 )(x̂ cos(ŷ), x̂ sin(ŷ))
cos(ŷ)
!
= sin(ŷ)


sin(ŷ)
= cos(ŷ) e1 − e2 ,

as was to be shown.
Thus, the push-forward of a vector field, or what is the same a change
of coordinates, requires the knowledge of both the forward and the
inverse diffeomorphism. Below we prove that only the knowledge of the
Math 5190: lecture notes. 24

inverse diffeomorphism is required, provided we are willing to invert a


matrix.
Proposition 4.10. Let F : U → V be a diffeomorphism and G =
F −1 : V → U its inverse. If A : U → Rn is a vector field, then
F ∗ A = (J G)−1 · (A ◦ G). (36)
Proof. The above formula follows from the definition of F ∗ A and from
fact that J G and J F ◦ G are inverse matrices. 
Example 4.11. Let us introduce two-dimensional parabolic coordi-
nates (u, v) by setting
x = u2 − v 2 , y = 2uv. (37)
Let us call the forward transformation (u, v) = F (x, y) and the inverse
2

1.5

0.5

-0.5

-1

-1.5

-1.5 -1 -0.5 0 0.5 1 1.5 2

Figure 1. Parabolic coordinates


(x, y) = G(u, v). We have explicit formulas for G, but explicit formulas
for F would be “messy” and involve radicals. In any case, G is more
convenient for generating the coordinate grid for the u, v coordinates.
Half of the grid lines in the above figure were generated by setting
u = u0 treating v as a parameter, and then plotting the corresponding
Math 5190: lecture notes. 25

curve (x, y) = G(u0 , v). The other grid lines were generated by plotting
(x, y) = G(u, v0 ) for various constant values v = v0 .
In order to express a vector field using parabolic coordinates it would
seem that we find explicit expressions for F . However, we can avoid this
∂ ∂
by using (36). For example, let us express x +y using parabolic
∂x ∂y
coordinates. We have
∂x ∂x
 
 
 ∂u ∂v  u −v
J G(u, v) = 
  =2 .
∂y ∂y  v u
∂u ∂v
Taking the inverse of the above matrix, we have

∂u ∂v
  u v 
 ∂x ∂x  1  u2 + v 2 u2 + v 2 
 ∂v ∂v  = 2  −v
J F (x, y) =   (38)
u 
∂x ∂y u2 + v 2 u2 + v 2
Hence,
∂u ∂v
 
   
∂ ∂ ∂ ∂   ∂x ∂x 

, = ,
∂x ∂y ∂u ∂v  ∂v ∂v 
∂x
∂y
 u v 
2(u + v 2 )
2 2(u2 + v 2 ) 
 
∂ ∂ 
= , 
−v u

∂u ∂v  
2(u + v 2 )
2 2(u2 + v 2 )
We can now switch to parabolic coordinates.
  
∂ ∂ ∂ ∂ x
x +y = ,
∂x ∂y ∂x ∂y y
∂u ∂v
 
 u2 − v 2
   
∂ ∂  ∂x ∂x
= ,  
∂u ∂v  ∂v ∂v  2uv
∂x ∂y
 
1 ∂ ∂
= u +v .
2 ∂u ∂v
Note that we didn’t write J F in terms of x, y, because we don’t have
formulas for u and v in terms of x and y. We therefore express all our
calculations in terms of u and v.
Math 5190: lecture notes. 26

5. Covariance
The principle of covariance states that any operation defined in terms
of intrinsic geometry can be carried out in any system of coordinates
with suitable transformation laws relating the operation and the an-
swers. Let’s consider how the principle of covariance manifests in the
case of directional derivatives.
First, let us consider how functions transform under a change of
coordinates.
Definition 5.1. Let F : U → V be a diffeomorphism and let g : V → R
be a function of n variables. The function f : U → R defined by
f = g ◦ F. (39)

is called the pull-back of g by F and is denoted by f = F g.
Setting y j = F j (x1 , . . . , xn ) allows us to express (39) succinctly as
f (x1 , . . . , xn ) = g(y 1, . . . , y n ).
Therefore, the pull-back transformation is the geometric description of
the process of transforming a function by making a change of variables.
Recall that the directional derivative operation combines a C 1 vec-
tor field and a C 2 function to produce a C 1 function. For example,
given vector field A : U → Rn and function f : U → R the directional
derivative gives us the function A[f ] : U → R. The push-forward and
the pull-back are adjoint operations relative to the directional deriva-
tive bracket; pushing forward a vector field is, in a sense, equivalent to
pulling back a function.
Proposition 5.2. Let A : U → Rn be a C 1 vector field, F : U → V a
C 2 diffeomorphism, and g : V → R a C 2 function. Letting B = F ∗ A
and f = F ∗ g, we have
F ∗ (B[g]) = A[f ]. (40)
Proof. Setting y = F (x), the chain rule gives us
∂f (x1 , . . . , xn ) ∂g(y 1 , . . . , y n )
=
∂xi ∂xi
n
X ∂g(y 1 , . . . , y n ) ∂y j
= , i = 1, . . . , n.
j=1
∂y j ∂xi
Equivalently,
n
X
Di f = (Dj g ◦ F ) Di F j , i = 1, . . . , n
j=1
Math 5190: lecture notes. 27

Hence,
n
X
A[f ] = Ai Di f
i=1
n
X
= Ai (Dj g ◦ F ) Di F j
i,j=1
X n
= A[F j ](Dj g ◦ F )
j=1
n
X
= (B j ◦ F )(Dj g ◦ F )
j=1

= B[g] ◦ F,
as was to be shown. 
The upshot of the above result is that a directional derivative calcu-
lated in a different coordinate system agrees with the directional deriva-
tive calculated relative to standard coordinates. Indeed, let A, B, F , f, g
be as in the Proposition. By definition,
n n
X ∂ X ∂
A= Ai (x1 , . . . , xn ) i = B j (y 1 , . . . , y n) j ,
i=1
∂x j=1
∂y

where Ai : U → R are the components of A and where B j : V → R


are the components of B = F ∗ A. Since f = F ∗ g,
f (x1 , . . . , xn ) = g(y 1, . . . , y n )
is the same scalar function expressed in two different ways. By defini-
tion of directional derivative,
n
X
A[f ] = Ai Di f,
i=1
n
X
B[g] = B j Dj g.
j=1

The Proposition asserts that


A[f ](x1 , . . . , xn ) = B[g](y 1, . . . , y n ),
or what is equivalent
n 1 n n 1 n
n ∂f (x , . . . , x ) n ∂g(y , . . . , y )
X X
i 1 i 1
A (x , . . . , x ) = B (y , . . . , y ) .
i=1
∂xi j=1
∂y j
Math 5190: lecture notes. 28

Example 5.3. Next, let us illustrate the principle of covariance. Let


f (x, y) be a function and set
g(u, v) = f (u2 − v 2 , 2uv) = (f ◦ G)(u, v).
In other words, g gives the form of the expression f (x, y) in parabolic
coordinates. Geometrically, f : U → R and g = G∗ f : V → R is the
pullback function (We will choose suitable domains U, V below). Let
us calculate the directional derivative
 
∂ ∂
x +y [f (x, y)]
∂x ∂y
in both coordinate systems and compare the answers. We obtain
 
1 ∂ ∂
u +v [g(u, v)]
2 ∂u ∂v
∂f (u2 − v 2 , 2uv) ∂f (u2 − v 2 , 2uv)
 
1
= u +v
2 ∂u ∂v
= u2 (D1 f )(x, y) + uv(D2 f )(x, y) − v 2 (D1 f )(x, y) + uv(D2 f )(x, y)

 
∂ ∂
= x +y [f (x, y)],
∂x ∂y
as was to be shown.
What happens if there are multiple systems of coordinates, say stan-
dard coordinates S, coordinates A and coordinates B? Does a direc-
tional derivative calculated in coordinates A agree with the same cal-
culation using coordinates B? The answer must be “yes”, because both
calculation must agree with the answer obtained using standard coor-
dinates. In this type of situation there are 6 different transformations
we must consider: standard coordinates to A, standard coordinates to
B, A to B, and all the inverse transformations. Of course a transforma-
tion from S to A, followed a transformation from A to B, should be the
same thing as the transformation from S to B. This kind of coherence
property is our next subject.
Proposition 5.4. Let F : U → V and G : V → W be diffeomor-
phisms. Let A : U → Rn be a vector field. Then,
G∗ (F ∗ A) = (G ◦ F )∗ A. (41)
Similarly, let h : W → R be a function. Then,
F ∗ (G∗ h) = (G ◦ F )∗ h. (42)
Proof. Homework. 
Math 5190: lecture notes. 29

Example 5.5. Let’s consider three coordinate system in the plane:


the standard (x, y) coordinates, the polar (r, θ) coordinates and the
parabolic (u, v) coordinates. These are related by the following diffeo-
morphisms:
(x, y) = G1 (r, θ) = (r cos(θ), r sin(θ)),
√ √
(u, v) = G2 (r, θ) = ( r cos(θ/2), r sin(θ/2)),
(x, y) = G3 (u, v) = (u2 − v 2 , 2uv).
Here G1 : V → U, G2 : W → V , and G2 : W → U are diffeomorphisms
of the following domains:
U = {(a1 , a2 ) : a1 > 0 or a2 6= 0},
V = {(a1 , a2 ) : a1 > 0 and − π < a2 < π},
W = {(a1 , a2 ) : a1 > 0}.
In order for the above definitions to be sound we must have
G1 = G3 ◦ G2 .
The following calculation verifies this. Throughout, it is convenient to
identify vectors in R2 with complex numbers. This allows us to write

G2 (r, θ) = r exp(i θ/2),
G3 (u, v) = (u + i v)2 .
Hence,

(G3 ◦ G2 )(r, θ) = G3 ( r exp(i θ/2))
= r exp(i θ)
= G1 (r, θ).
Next, consider a function f (x, y). The same function can be ex-
pressed using polar coordinates as
f (x, y) = g1 (r, θ)
where g1 = G∗1 f , because
f (x, y) = f (r cos θ, r sin θ) = (f ◦ G1 )(r, θ) = g1 (r, θ).
Similarly, to express f (x, y) in parabolic coordinates we use g2 = G∗3 f
because
g2 (u, v) = (f ◦ G2 )(u, v) = f (u2 − v 2 , 2uv) = f (x, y).
However, we can also express g2 (u, v) as a function of r, θ by setting
g2 (u, v) = g3 (r, θ),
Math 5190: lecture notes. 30

where g3 = G∗2 g2 because



g3 (r, θ) = (g2 ◦ G3 )(r, θ) = g2 ( r exp(i θ/2)) = g2 (u, v).

However all of the above are equal to f (x, y) and so g3 = g1 , or to put


it another way
G∗3 (G∗2 f ) = G∗1 f.

That’s exactly what (42) asserts.


Similarly, if we start with a given vector field

∂ ∂
A = A1 (x, y) + A2 (x, y) ,
∂x ∂y

we can write it using polar coordinates as

∂ ∂
A = B 1 (r, θ) + B 2 (r, θ) ,
∂r ∂θ
or using parabolic coordinates:

∂ ∂
A = C 1 (u, v) + C 2 (u, v) .
∂u ∂v

Here B = (F 1 )∗ A and C = (F 2 )∗ A, where F 1 = (G1 )−1 and F 2 =


(G2 )−1 are the inverse diffeomorphisms. However, a vector field ex-
pressed in parabolic coordinates can be rewritten using polar coor-
dinates by making use of the F 3 = (G3 )−1 transformation; in other
words,we also have B = (F 3 )∗ C. Thus, we have two different ways of
calculating B 1 (r, θ) and B 2 (r, θ). The above Proposition asserts that
both calculations yield the same answer.
Let’s do a specific calculation of this type. Take A = ∂/∂x. Earlier,
we established that
∂ ∂u ∂ ∂v ∂
= +
∂x ∂x ∂u ∂x ∂v
u ∂ v ∂
= − ,
2(u2 + v 2 ) ∂u 2(u2 + v 2 ) ∂v
 
 u 
∂ ∂  2(u2 + v 2 ) 
= , v .
∂u ∂v


2(u2 + v 2 )
Math 5190: lecture notes. 31

We also have
∂u ∂u
 
   
∂ ∂ ∂ ∂  ∂r ∂θ 
, = ,  
∂r ∂θ ∂u ∂v  ∂v ∂v 
∂r ∂θ √ !
cos(θ/2) r
− sin(θ/2)
 
∂ ∂ √
2 r 2
= , sin(θ/2)

r
∂u ∂v √
2 r 2
cos(θ/2)
Taking inverses gives
√ √ !
2 r cos(θ/2) 2 r sin(θ/2)
   
∂ ∂ ∂ ∂
, = , √ √
∂u ∂v ∂r ∂θ −2 sin(θ/2)/ r 2 cos(θ/2)/ r
Combining the above relations and switching to the r, θ variables gives
  √ 
∂ 1 ∂ ∂ r cos(θ/2)
= , √
∂x 2r ∂u ∂v − r sin(θ/2)
√ √ ! √
2 r cos(θ/2) 2 r sin(θ/2)
  
1 ∂ ∂ r cos(θ/2)
= , √ √ √
2r ∂r ∂θ −2 sin(θ/2)/ r 2 cos(θ/2)/ r − r sin(θ/2)
  
∂ ∂ cos(θ)
= , .
∂r ∂θ − sin(θ)/r
However, we could have calculated the above directly by using the
transformation from polar to Cartesian coordinates:
∂ ∂r ∂ ∂θ ∂
= +
∂x ∂x ∂r ∂x ∂θ
∂ sin(θ) ∂
= cos(θ) −
∂r r ∂θ
The two answers agree.
Math 5190: lecture notes. 32

6. Coordinate transformations and ODEs


Consider a system of autonomous ODEs
ẋi = Ai (x1 , . . . , xn ), x ∈ U. (43)
defined on some domain U ⊂ Rn . Let F : U → V be a diffeomorphism,
and let y j = F j (x1 , . . . , xn ) be the corresponding system of coordinates
on U. We wish to write a system
ẏ j = B j (y 1, . . . , y n ), y∈V (44)
that is in some sense equivalent to the given system (43). Since
n
dy j X ∂y j dxi
=
dt i=1
∂xi dt
n
X ∂y j
= Ai (x1 , . . . , xn )
i=1
∂xi
= B (y 1 , . . . , y n )
j

we recognize that ODEs are governed by the same transformation laws


as vector fields. Expressing the system (43) relative to the y j coordi-
nates is equivalent to expressing the vector field
n n
X
i 1 ∂ n
X ∂
A= A (x , . . . , x ) i = B j (y 1, . . . , y n ) j
i=1
∂x j=1
∂y

relative to the y j coordinate system; the calculation the push-forward


B = F ∗ A is an equivalent process.
However we interpret the transformation from (43) to (44), it can
only be meaningful if the solutions of the two systems are in some
sense equivalent. We need a transformation law for solutions.
Let Φ(t, x) denote the flow generated by A and let us write the
solution to (43) as
x̂i = Φi (t, x1 , . . . , xn ), i = 1, . . . , n.
Thus, x̂1 , . . . , x̂n are functions on the flow domain Û ⊂ R × U that
satisfy
∂ x̂i
= Ai (x̂1 , . . . , x̂n ), i = 1, . . . , n. (45)
∂t
Similarly, let Ψ(t, y) denote the flow generated by B and let
ŷ j = Ψj (t, y 1, . . . , y n ), j = 1, . . . , n (46)
Math 5190: lecture notes. 33

express the general solution to (44). Since y j = F j (x1 , . . . , xn ) it makes


sense that compatibility of the two solution be the condition

ŷ j = F j (x̂1 , . . . , x̂n ). (47)

The above condition can be expressed in a component-free language as

F ◦ Φt = Ψt ◦ F , (48)

where, as before, Φt (x) = Φ(t, x) and Ψt (y) = Ψ(t, y).

Definition 6.1. Let F : U → V be a diffeomorphism and let Φ(t, x)


be a flow defined on U. We define the push-forward Ψ = F ∗ Φ to be the
mapping defined by
Ψt = F ◦ Φt ◦ F −1 . (49)
The domain of Ψ is the set

V̂ = {(a0 , a1 , . . . , an ) ∈ Rn+1 : (a0 , F −1 (a1 , . . . , an )) ∈ Û}, (50)

where Û is the domain of Φ.

Note that with the above definition, the push-forward flow satisfies
condition (48).

Proposition 6.2. Let F , Φ, Ψ be as above. With the above definition,


Ψ is a flow.

Proof. Let G = F −1 denote the inverse transformation. We have

Ψ0 (y) = F (Φ0 (G(y))) = F (G(y)) = y, y ∈ V.

As well,

Ψt1 +t2 = F ◦ Φt1 +t2 ◦ G


= F ◦ Φt1 ◦ Φt2 ◦ G
= F ◦ Φt1 ◦ G ◦ F ◦ Φt2 ◦ G
= Ψt1 ◦ Ψt2 .

The various domain requirements of a flow are left as exercises for the
interested reader. 

Proposition 6.3. Let A : U → Rn be a vector field and let Φ denote


the flow generated by A. Let F : U → V be a diffeomorphism, and set
B = F ∗ A and Ψ = F ∗ Φ. Then, Ψ is the flow generated by B.
Math 5190: lecture notes. 34

Proof. By the chain rule and by Proposition 1.11,


∂Ψ(t, y) ∂F (Φ(t, x))
=
∂t ∂t
∂Φ(t, x)
= J F (Φ(t, x)) ·
∂t
= J F (Φ(t, x)) · A(Φ(t, x))
By definition of the push-forward of a vector field,
B(y) = J F (x) · A(x).
Hence,
(D1 Ψ)(0, y) = J F (x) · A(x) = B(y),
as was to be shown. 
Example 6.4. Let r, θ be the usual polar coordinates. Earlier, we
showed that
∂ ∂ sin(θ) ∂
= cos(θ) −
∂x ∂r r ∂θ
In other words, the system
ṙ = cos(θ) (51)
θ̇ = − sin(θ)/r
is equivalent to the system
ẋ = 1,
ẏ = 0.
The general solution, in the form of a flow, of the latter system is
Φ(t, x, y) = (x + t, y).
To obtain the solution of the system in polar coordinates, call it Ψ(t, r, θ)
we have to calculate the push-forward of Φ relative to the transforma-
tion (r, θ) = F (x, y). We do so by writing
Ψ(t, r, θ) = F (Φ(t, r cos θ, r sin θ))
= F (r cos θ + t, r sin θ)

  
−1 r sin(θ)
= r 2 + 2tr cos θ + t2 , tan
r cos(θ) + t
A straight-forward calculation verifies that the above is, indeed, the
general solution to (51). Figure 2 shows the push forward F ∗ A, while
Figure 3 displays the push-forward of the Cartesian coordinate grid via
F ; i.e. the grid consisting of the curves x cos(y) = C1 and the curves
x sin(y) = C2 .
Math 5190: lecture notes. 35

¶ Sin@yD ¶
Cos@yD €€€€€€€€ - €€€€€€€€€€€€€€€€€€€ €€€€€€€€
¶x x ¶y
3

-1

-2

-3
0.25 0.5 0.75 1 1.25 1.5 1.75 2

Figure 2. Push-forward of ∂/∂x.

7. First integrals
We have already mentioned the close connection between autonomous
ODEs and vector fields. On the other hand, we identify vector fields
with first-order differential operators. In this section we discuss first
integrals, a geometric idea that ties together solutions of an ODE and
the directional derivative operation.
Let us introduce a system of autonomous ODEs
ẋi = Ai (x1 , . . . , xn ), x ∈ U.
Consider the variation of a scalar function u = f (x1 , . . . , xn ) along a
particular integral curve φ(t) ∈ U. Setting x̂i = φi (t), we have
u̇ = (f ◦ φ)′ (t)
Xn
= Di f (x̂1 , . . . , x̂n )Ai (x̂1 , . . . , x̂n )
i=1
= (A[u] ◦ φ)(t).
Math 5190: lecture notes. 36

-1

-2

-3
0.2 0.4 0.6 0.8 1

Figure 3. Push-forward of Cartesian coordinate grid.

This then is the geometric meaning of the directional derivative: it


names the rate of change of a given scalar function along an integral
curve of a given vector field.
Definition 7.1. Let A : U → Rn be a C 1 vector field and f : U → R
a C 2 function. We call f a first integral whenever A[f ] = 0.
Proposition 7.2. A function f : U → R is a first integral of a vector
field A : U → Rn if and only if it is constant along all integral curves
of A. More precisely, if Φ(t, x) is the flow generated by A, then f (x)
is a first integral if and only if
(f ◦ Φ)(t, x) = f (x).
Example 7.3. Consider the vector field A = −y∂/∂x + x∂/∂y. The
function f (x, y) = x2 + y 2 is a first integral because A[x2 + y 2 ] = 0
( a simple calculation). Equivalently, the flow generated by A is the
function
Φ(t, x, y) = (x cos t − y sin t, x sin t + y cos t).
Math 5190: lecture notes. 37

Observe that
(f ◦ Φ)(t, x, y) = (x cos t − y sin t)2 + (x sin t + y cos t)2
= x2 + y 2 = f (x, y),
as was to be shown.
Analytic geometry in the plane is based on the notion that a curve
is defined by an equation in 2 variables. Analogously, a curve in n
dimensions should require n − 1 independent equation. The following
definition and propositions make this more precise.
Definition 7.4. Let U ⊂ Rn be a domain. We say that continuously
differentiable functions f 1 , . . . , f j : U → R are functionally indepen-
dent if the j × n matrix of partial derivatives
D1 f 1 . . . Dn f 1
 
 ... ..
.
.. 
. (52)
D1f j . . . Dnf j
has rank j at all points of U.
Proposition 7.5. Let f 1 , . . . , f j−1 : U → R be C 1 functions and set
f j = F (f 1 , . . . , f j−1) where F is a C 1 function of j − 1 variables. Then
f 1 , . . . , f j−1, f j are functionally dependent. Conversely, if f 1 , . . . , f j
are C 1 and functionally dependent, then after rearranging the order of
f 1 , . . . , f j and restricting the domain, we have f j = F (f 1 , . . . , f j−1)
for some C 1 function F of j − 1 variables.
Example 7.6. By way of example, let us consider the functions u =
(x + y)/x, x > 0 and v = tan−1 (y/x), x > 0. The matrix of partial
derivatives
 ∂u ∂u  
−y/x2

∂x ∂y 1/x
= (53)
∂v
∂x
∂v
∂y
−y/(x2 + y 2) x/(x2 + y 2)
is singular; the rank is < 2 everywhere. We can therefore express one
function in terms of the other. Indeed,
u = tan(v) + 1.
Proposition 7.7. Let f 1 , . . . , f n−1 : U → R be functionally indepen-
dent, continuously differentiable functions. Fix a point (a1 , . . . , an ) ∈ U
and set C i = f i (a1 , . . . , an ), i = 1, . . . , n − 1. Then, there exists a C 1
curve x̂i = φi (t) that passes through (a1 , . . . , an ) and satisfies the equa-
tions
f i (x̂1 , . . . , x̂n ) = C i , i = 1, . . . , n − 1.
Every other such curve is related to φ(t) by a reparameterization.
Math 5190: lecture notes. 38

The proof of the above Propositions relies on the implicit function the-
orem; we will not present it here. However, using the inverse function
theorem, we can prove the following.
Proposition 7.8. Let U ⊂ Rn be an n-dimensional domain and F 1 , . . . , F n :
U → R functionally independent, C 1 functions. Then, after suitably
restricting the domain to U1 ⊂ U the mapping F : U1 → V , where
V = F (U1 ) ⊂ Rn , is a diffeomorphism.
The upshot is that n functionally independent functions define a system
of coordinates (provided the domain is suitably restricted).
In light of Proposition 7.7, the knowledge of n − 1 first integrals of
an n-dimensional vector field A : U → Rn determines the trajectories
of the integral curves of the corresponding ODE ẋ = A(x). To obtain
the flow/general solution one needs only to determine the correct pa-
rameterization of these integral curves. This will be explored carefully
in the next section. Here we limit our discussion to some examples.
Example 7.9. We have already established that I1 = x2 + y 2 is a first
integral of the vector field −y∂/∂x+x∂/∂y. Since this is 2-dimensional
vector field, a single first integral suffices to determine the integral
curves. In this case, for values I1 > 0, the integral curves are circles
centered at the origin. The actual solutions of the ODE, namely
x = x0 cos t − y0 sin t, y = x0 sin t + y0 cos t
describe a parameterization of these circles.
Example 7.10. Consider the following 3-dimensional ODE:
ẋ = 0
ẏ = x
ż = y.
The following functions are first integrals:
I1 = x (54)
I2 = 2xz − y 2. (55)
It’s quite evident that I1 = x is a first integral. As for I2 , observe that
I˙2 = 2z ẋ + 2xż − 2y ẏ = 2xy − 2xy = 0.
Therefore, the integral curves are parabolas described by equations (54)
where we restrict I1 , I2 to specific values.
A solution of the ODE is given by a certain parameterization of
these parabolas. We can solve the system in question in a step-by-step
Math 5190: lecture notes. 39

fashion, taking anti-derivatives at each step. We obtain the following


general solution
x = x0 ,
y = x0 t + y0 ,
z = x0 t2 /2 + y0 t + z0 .
This is a parameterization of the curve (54). Having the general solu-
tion, an explicit calculation verifies that
x = x0 = I1 ,
2xz − y 2 = 2x0 z0 − y02 = I2 .
Math 5190: lecture notes. 40

8. Rectification
Definition 8.1. Let A : U → Rn be a C 1 vector field and F : U → V
a C 1 diffeomorphism. We say that F rectifies A if

F ∗A = ,
∂ x̂1
where x̂1 , . . . , x̂n are the standard coordinates on V . Equivalently, set-
ting y j = F j (x1 , . . . , xn ) we can say that F rectifies A if

A= .
∂y 1
We can easily describe the flow generated by a unit vector. Therefore,
if we can rectify a vector field, we can solve the corresponding ODE.
Example 8.2. Earlier, we saw that
∂ ∂ ∂
A = −y +x = ,
∂x ∂y ∂θ
where r, θ are the usual polar coordinates. Therefore, the transforma-
tion p
θ = tan−1 (y/x), r = x2 + y 2
rectifies the given vector field A. Hence, the flow generated by A can
be given by
θ = θ0 + t, r = r0 .
Rewriting the above using x, y coordinates yields the usual expression
for the flow, namely
x = r cos θ = r0 cos(θ0 + t) = r0 cos(θ0 ) cos(t) − r0 sin(θ0 ) sin(t)
= x0 cos(t) − y0 sin(t),
y = r sin θ = r0 sin(θ0 + t)
= x0 sin(t) + y0 cos(t).
Also, note that r is a first integral.
Proposition 8.3 (Rectification gives flows). Let A : U → Rn be a
vector field and F : U → V a diffeomorphism that rectifies A. Then,
the flow generated by A is given by
Φ(t, x1 , . . . , xn ) = G(y 1 + t, y 2 , . . . , y n ),
where y j = F j (x1 , . . . , xn ), and where G = F −1 is the inverse diffeo-
morphism.
Math 5190: lecture notes. 41

Proof. Let
Ψ(t, x̂1 , . . . , x̂n ) = (x̂1 + t, x̂2 , . . . , x̂n )
denote the flow generated by ∂/∂ x̂1 . Our definition of Ψ is equivalent
to Φ = G∗ Ψ. Since

A = G∗ 1 ,
∂ x̂
the conclusion follows by Proposition 6.3. 
Proposition 8.4 (Rectification gives first integrals.). Let A : U → Rn
be a vector field and F : U → V a diffeomorphism that rectifies A.
Then
y j = F j (x1 , . . . , xn ), j = 2, . . . , n
are first integrals of A.
Proof. By definition,

.
A=
∂y 1
Hence, by the covariance of the directional derivative, we have
A[y 1 ] = 1,
A[y j ] = 0, j = 2, . . . , n.

Next, suppose that A : U → Rn be a C 1 vector field such that
A (0, x2 , . . . , xn ) 6= 0 for all x ∈ U0 where
1

U0 = {(0, x2 , . . . , xn ) ∈ U}
is the intersection of U and the x1 = 0 hyperplane. Let
Φ(t, x) = (Φ1 (t, x1 , . . . , xn ), . . . , Φn (t, x1 , . . . , xn )
be the flow generated by A. Consider the equation
Φ1 (t̂, x1 , . . . , xn ) = 0, (56)
where t̂ is the unknown. Observe that Φ1 (0, 0, x2 , . . . , xn ) = 0 and that
Φ̇1 (0, 0, x2 , . . . , xn ) = A1 (0, x2 , . . . , xn ) 6= 0.
Hence, by the implicit function theorem, there exists a C 1 function t̂ =
τ (x1 , . . . , xn ) defined in some neighborhood Û ⊂ U of the hyperplane
slice U0 . With t̂ as above, set
I j = Φj (t̂, x1 , . . . , xn ).
Proposition 8.5 (Flows give first integrals). With the above defini-
tions, the functions I 2 , . . . , I n are first integrals of A.
Math 5190: lecture notes. 42

Proof. By the 1-parameter group property, we have


Φ1 (τ (x) − t, Φ(t, x)) = Φ1 (τ (x), x).
Since locally, t̂ = τ (x) is the unique solution of the equation system
we must have
τ (Φ(t, x)) = τ (x) − t.
Let us define functions Qj (x1 , . . . , xn ), j = 2, . . . , n by writing
I j = Qj (x1 , . . . , xn ) = Φj (τ (x), x).
We have
Qj (Φ(t, x)) = Φj (τ (Φ(t, x)), Φ(t, x))
= Φj (τ (x) − t, Φ(t, x))
= Qj (x).
This establishes that I 2 , . . . , I n are first integrals. 
Next, let us discuss how to solve an autonomous ODE using first
integrals. As we already indicated, knowledge of n − 1 functionally
independent first integrals serves as a description of the unparameter-
ized integral curves. We can obtained the desired parameterization by
employing a special coordinate system built using the first integrals,
and by means of a reparameterization akin to the one shown in the
proof to Theorem 2.1.
The next two Proposition indicate how to obtain a rectification by
means of first integrals.
Proposition 8.6. Let A : U → Rn be a C 1 vector field such that
A1 (x1 , . . . , xn ) 6= 0 and let
ξ j = F j (x1 , . . . , xn ), j = 2, . . . , n (57)
be n − 1 functionally independent, C 1 , first integrals. Then, setting
ξ 1 = F 1 (x1 , . . . , xn ) := x1 , (58)
we obtain a diffeomorphism F : U1 → V , where U1 ⊂ U is a suitably
restricted subdomain U1 ⊂ U. Furthermore,

A = B 1 (ξ 1 , . . . , ξ n ) , (59)
∂ξ 1
where the function B 1 satisfies and is defined by the relation
B 1 (ξ 1 , . . . , ξ n ) = A1 (x1 , . . . , xn ). (60)
Math 5190: lecture notes. 43

Proof. Our first claim is that ξ 1 , ξ 2 , . . . , ξ n are functionally indepen-


dent. Suppose not. By assumption, ξ 2, . . . , ξ n are functionally inde-
pendent, and so x1 = ξ 1 = F (ξ 2 , . . . , ξ n ) for some C 1 function F of
n − 1 variables. However, since ξ 2 , . . . , xn are first integrals, this would
imply that x1 is also a first integral. This would necessitate
A[ξ 1] = A1 (x1 , . . . , xn ) = 0,
a contradiction.
Therefore ξ 1 , . . . , xn are functionally independent. Hence, by taking
F 1 (x1 , . . . , xn ) = x1
and F j as in (57), and by restricting the domain to some U1 ⊂ U we
obtain a diffeomorphism F : U1 → V .
Finally, let us prove (59). Let B = F ∗ A. By definition,
n
X ∂
A= B j (ξ 1, . . . , ξ n ) j .
j=1
∂ξ
By assumption,
A[ξ 1 ] = A1 (x1 , . . . , xn ),
A[ξ j ] = 0, j = 2, . . . , n.
Therefore, B j (ξ 1 , . . . , ξ n ) = 0 for j = 2, . . . , n and
B 1 (ξ 1 , . . . , ξ n ) = A1 (x1 , . . . , xn ),
as was to be shown. 
Proposition 8.7. Let A1 : U → R be a non-zero, C 1 function and let
A = A1 (x1 , . . . , xn ) ∂x∂ 1 be the indicated vector field. Define
Z x1
1 1 1 2 n du
ξ = F (x , x , . . . , x ) := , (61)
0 A (u, x2 , . . . , xn )
1

ξ j = F j (x1 , . . . , xn ) := xj , j = 2, . . . , n. (62)
Then,

A= ; (63)
∂ξ 1
i.e., F rectifies A.
Proof. Observe that
A[ξ 1 ] = (D1 F 1 )(x1 , . . . , xn )A1 (x1 , . . . , xn )
A1 (x1 , . . . , xn )
= = 1,
A1 (x1 , . . . , xn )
A[ξ j ] = 0, j = 2, . . . , n.
Math 5190: lecture notes. 44


Example 8.8. Consider the generator of the 2-dimensional rotation
group:
∂ ∂
A = −y +x .
∂x ∂y
p
We have already established that r = x2 + y 2 is a first integral. Let
us therefore take p
ξ = x, η = x2 + y 2
as a new system of coordinates. As our domain, we can take U1 =
{(a1 , a2 ) ∈ R2 : a2 > 0}. Letting (ξ, η) = F (x, y) denote the corre-
sponding transformation, we have
V = F (U1 ) = {(a1 , a2 ) ∈ R2 : a2 > |a1 |}.
The inverse diffeomorphism is given by
p
x = ξ, y = η 2 − ξ 2 , η > |ξ|.
Therefore,
∂ξ ∂ξ
 
   
∂ ∂ ∂ ∂  ∂x ∂y 
, = ,  ∂η ∂η 
∂x ∂y ∂ξ ∂η
∂x ∂y
 

∂ ∂
 1 0
= ,  x y ,
∂ξ ∂η p
x2 + y2
p
x2 + y2
∂ ∂ p ∂
−y +x = − η2 − ξ 2
∂x ∂y ∂ξ
The next step is to reparameterizep the simple translational flow for
∂/∂ξ so as to obtain the flow for − η 2 − ξ 2 ∂/∂ξ. According to Propo-
sition 8.7 we can do this by making the change of coordinates
Z ξ p
du −1 η 2 − (ξ)2
θ= p = tan − π/2 = tan−1 (y/x) − π/2,
0 − η −ξ
2 2 ξ
p
r = η = x2 + y 2 .
Example 8.9. Let’s return to the differential equation
ẋ = 0, ẏ = x, ż = y.
Earlier we demonstrated that
ξ 2 = x, ξ 3 = 2xz − y 2
Math 5190: lecture notes. 45

are first integrals. Let us use this information to rectify the correspond-
ing vector field. We begin by introducing coordinates ξ 1 = y with ξ 2 , ξ 3
as above, and restricting the domain to y > 0. We calculate
 
    1 0 0
∂ ∂ ∂ ∂ ∂ ∂
, , = 2
, 1, 3  0 1 0 ,
∂x ∂y ∂z ∂ξ ∂ξ ∂ξ −2y 2z 2x
∂ ∂ ∂
x +y = ξ2 1 .
∂y ∂z ∂ξ
Next, we follow Proposition 8.7 and introduce coordinates
Z ξ1
1 du ξ1
η = = ,
0 ξ2 ξ2
= y/x,
η 2 = ξ 2 = x,
η 3 = ξ 3 = 2xz − y 2 .
Relative to these coordinates, we have
∂ ∂ ∂ ∂
x +y = ξ2 1 = 1 .
∂y ∂z ∂ξ ∂η
We can use this rectification to integrate the given ODE. The inverse
transformation is given by
1 3 2
x = η2, y = η1η2, z = η /η + (η 1 )2 η 2 .

2
Therefore, the flow is given by
η 1 = η01 + t, η 2 = η02 , η 3 = η03 .
x = η 2 = x0 ,
y = η 1 η 2 = (y0 /x0 + t)x0 = x0 t + y0 ,
1
(2x0 z0 − y02)/x0 + (y0 /x0 + t)2 x0

z=
2
1 2
= x0 t + 2ty0 + z0 .
2
Math 5190: lecture notes. 46

9. Symmetries
Definition 9.1. Let A : U → Rn be a C 1 vector field, U1 , V1 ⊂ U
subdomains, and F : U1 → V1 a diffeomorphism. We say that F is a
symmetry of A if F ∗ A = A. Equivalently, setting y j = F j (x1 , . . . , xn )
we say that F is a symmetry if
n n
X ∂ X ∂
A= Ai (x1 , . . . , xn ) i
= Ai (y 1 , . . . , y n) i . (64)
i=1
∂x i=1
∂y
We say that F is a conformal symmetry if F ∗ A = f A where f : V1 →
R is a non-vanishing function.
Example 9.2. Consider the rotation generator −y∂/∂x + x∂/∂y and
the following transformations.
(1) The transformation x̂ = −y, ŷ = x is a symmetry because
∂ ∂ ∂ ∂
−y +x = −ŷ + x̂ . (65)
∂x ∂y ∂ x̂ ∂ ŷ
(2) The transformation
√ √
x = u2 + v 2 cos u, y= u2 + v 2 sin u
is a conformal symmetry because
  
∂ ∂ ∂ u ∂ 1 ∂ ∂
−y +x = − = − −v +u .
∂x ∂y ∂u v ∂v v ∂u ∂v
(3) Let k be a real number. The homothety transformations
x̂ = ek x, ŷ = ek y
form a 1-parameter group of symmetries, because for every k
relation (65) holds.
Definition 9.3. Let A : U → Rn be a vector field and Φ(t, x) the flow
generated by A. We say that a diffeomorphism F : U1 → V1 preserves
the solutions of the corresponding ODE ẋ = A(x) if Φ if F ∗ Φ = Φ.
Similarly, we say F preserves the integral curves of A if there exists a
reparameterization function τ (s, x) such that
(F ∗ Φ)(s, x) = Φ(τ (s, x), x). (66)
Note: recall that a flow reparameterization function τ (s, x) satisfies
τ (0, x) = 0
τ̇ (0, x) 6= 0.
Math 5190: lecture notes. 47

Proposition 9.4. A diffeomorphism F is a symmetry of a vector field


A if and only if F preserves the solutions of A. Similarly, F is a
conformal symmetry if and only it preserves the integral curves.
Proof. Earlier we demonstrated the covariance of flows and vector fields.
Thus, F ∗ Φ is the flow for F ∗ A. Therefore F ∗ A = A if and only if
F ∗ Φ = Φ.
Next, let us suppose that F is a conformal symmetry of A, with
F ∗ A = f A. Let Φ(t, x) be the flow generated by A and Ψ(t, x) the
flow generated by f A. By the reparameterization theorem, there exists
a reparameterization function τ (s, x) such that
Ψ(s, x) = Φ(τ (s, x), x). (67)
Conversely, suppose that (66) holds. It follows that Ψ(s, x) as defined
by (67) is the flow generated by F ∗ A. Observe that
Ψ̇(0, x) = Φ̇(τ (0, x), x)τ̇ (0, x),
= f (x)Φ̇(0, x),
where
f (x) = τ̇ (0, x). (68)

Example 9.5. Let us return to the symmetries of the rotation gener-
ator. The rotation flow is given by
x = x0 cos t − y0 sin t, y = x0 sin t + y0 cos t.
We now consider the effects of the 3 transformations given above on
this flow.
(1) We have
x̂ = −y = −x0 sin t − y0 cos t
= x̂0 cos t − ŷ0 sin t,
ŷ = x = x̂0 sin t + ŷ0 cos t.
The form of the flow is preserved by the transformation.
(2) Recall that
x2 + y 2 = x20 + y02 .
is a first integral, and hence the integral curves have the form
x2 + y 2 = const.. Next, consider the same function, but in the
new coordinates:
u2 + v 2 = x2 + y 2
Therefore, integral curves are preserved by the transformation.
Math 5190: lecture notes. 48

As for the reparameterization function, the flow in the u, v


coordinates is given by
q q
u = u0 + t, v = u20 + v02 − u2 = v02 − 2u0 t − t2 .
If we employ the reparameterization function
t = u0 cos s − v0 sin s − u0 ,
we obtain
u = u0 cos s − v0 sin s,
q
v = u20 + v02 − u20 cos2 s − v02 sin2 s + 2u0 v0 sin s cos s
q
= u20 sin2 s + v02 cos2 s + 2u0 v0 sin s cos s
= u0 sin s + v0 cos s
provided we restrict our flows and vector fields to the domain
v > 0.
Math 5190: lecture notes. 49

10. The Lie bracket


Definition 10.1. Let A, B : U → Rn be C 2 vector fields. We define
the Lie bracket of A and B to be the vector field
[A, B] = J B · A − J A · B (69)
Xn
A[B i ] − B[Ai ] ∂xi

= (70)
i=1
Xn
Aj Dj B i − B j Di Ai ∂xi ,

= (71)
i,j=1

where for convenience we are using


∂ ∂
∂xi = i , ∂r = , etc.
∂x ∂r
Proposition 10.2. Let f : U → Rn be a C 2 function. Then
[A, B][f ] = A[B[f ]] − B[A[f ]] (72)
Indeed, [A, B] is the unique vector field that satisfies the above property.
Proof. The first assertion follows from a straight-forward calculations.
Note that all the 2nd order derivative parts cancel. As for the second
assertion, observe that A[xi ] = Ai , and similarly for B. Hence, by (72)
we have
[A, B][xi ] = A[B i ] − B[Ai ],
in full agreement with (69). 
Proposition 10.3. The Lie bracket has the following properties. Through-
out, A, B, C are vector fields and f : U → R is a function.
[A, B] = −[B, A], (73)
[A + B, C] = [A, C] + [B, C], (74)
[A, [B, C]] + [B, [C, A]] + [C, [A, B]] = 0, (75)
[A, f B] = A[f ]B + f [A, B] (76)
As a consequence of the above properties, we have
[A, A] = 0,
[f A, B] = −B[f ]A + f [A, B].
Also, if c is a constant, then
[cA, B] = [A, cB] = c[A, B].
Math 5190: lecture notes. 50

Proposition 10.4. Let F : U → V be a diffeomorphism and y j =


F j (x1 , . . . , xn ) the corresponding change of coordinates. For A, B :
U → Rn , vector fields, we have
[F ∗ A, F ∗ B] = F ∗ [A, B]. (77)
Proof. This follow directly from Proposition 10.2 and the covariance of
the directional derivative. 
Example 10.5. Observe that, by definition of the Lie bracket,
[∂xi , ∂xj ] = 0.
We can use this fact together with properties (74) (75) to calculate the
Lie bracket. For example,
[x∂x + y∂y , −y∂x + x∂y ] = [x∂x , −y∂x ] + [x∂x , x∂y ]
+ [y∂y , −y∂x ] + [y∂y , x∂y ]
= 0.
Switching to polar coordinates, we have
x∂x + y∂y = r∂r ,
−y∂x + x∂y = ∂θ .
Another straightforward calculation confirms the principle of covari-
ance:
[r∂r , ∂θ ] = 0.
Math 5190: lecture notes. 51

11. Infinitesimal symmetries


Definition 11.1. Let A, B : U → Rn be vector fields. We say that
A is an infinitesimal symmetry of B if [A, B] = 0. We say that A
is an infinitesimal conformal symmetry of B if [A, B] = f B, where
f : U → Rn is a function.
Note: if A is a symmetry of B, then necessarily B is a symmetry
of A. However, the analogous statement does not hold for conformal
symmetries.
Proposition 11.2. Let A, B : U → Rn be vector fields. If [A, B] = 0,
then for all functions g : U → Rn , the vector field A is a conformal
symmetry of gB. Conversely, if A is an infinitesimal conformal sym-
metry of B, then there exists a subdomain U1 ⊂ U and a non-zero
function g : U1 → R such that [A, gB] = 0.
Proof. The forward implication follows from property (76) of the Lie
bracket. Let us prove the converse. Rectify A so that A = ∂y1 relative
to some coordinates y j = F j (x1 , . . . , xn ), where F : U ⊃ U1 → V is a
diffeomorphism. By assumption [A, B] = f B. Let h = G∗ f , where
G = F −1 . Equivalently,
h(y 1 , . . . , y n) = f (x1 , . . . , xn ).
Set  Z y1 
1 n 2 n
g(x , . . . , x ) = exp − h(u, y , . . . , y )du .

Hence, by construction,
∂g
A[g] = = −f g,
∂y 1
[A, gB] = g[A, B] + A[g]B
= f gB − f gB = 0

Proposition 11.3. Let A, B : U → Rn be C 2 vector fields and Φ the
C 2 flow generated by A. Let
C(t, x) = C t (x) = ((Φt )∗ B)(x)
be the indicated time-dependent vector field. Then,
Ċ t = −[A, C t ]. (78)
In particular,
Ċ 0 = −[A, B].
Math 5190: lecture notes. 52

Proof. By definition of the push-forward, we have


(C t ◦ Φt )(x) = C(t, Φ(t, x)) = J Φt (x) · B(x).
Taking the derivative with respect to t and using the chain rule gives
Ċ t ◦ Φt + (J C t ◦ Φt )Φ̇t = (J Φt )˙· B.
By definition, Φ̇t = A ◦ Φt . Hence, By Clairaut’s theorem (interchange
of partial derivatives with respect to time and position),
(J Φt )˙ = J Φ̇t
= J (A ◦ Φt )
= (J A ◦ Φt ) · J Φt
Hence,
(J A ◦ Φt ) · J Φt · B = Ċ t ◦ Φt + (J C t ◦ Φt ) · (A ◦ Φt )
= (Ċ t + J C t · A) ◦ Φt .
However, since
J Φt · B = C t ◦ Φt ,
we have
J A · C t = Ċ t + J C t · A,
thereby establishing (78). 
Proposition 11.4. Let A : U → Rn be a vector field and Φt the flow
generated by A. Then Φt is a symmetry of A for all t, while A is its
own infinitesimal symmetry.
More generally, we have
Proposition 11.5. Let A, B : U → Rn be vector fields and Φt the
flow generated by A. Then A is an infinitesimal symmetry of B if and
only if for all t, Φt is a symmetry of B.
Proof. Set C t = (Φt )∗ B and suppose that C t = B for all t. Then,
since C t is, by assumption, constant with respect to t, we have, by
Proposition 11.3, [A, B] = 0.
Let us now prove the converse. Suppose that [A, B] = 0. Hence, by
Proposition 10.4, and using the fact (Φt )∗ A = A we have
(Φt )∗ [A, B] = [A, C t ] = 0. (79)
By Proposition (11.3), we have
Ċ t = −[C t , A] = 0,
Since C 0 = B we conclude that C t = B for all t. 
Math 5190: lecture notes. 53

Proposition 11.6. Let A, B : U → Rn be vector fields and Φs , Ψt the


respective flows. Then [A, B] = 0 if and only if
Φs ◦ Ψt = Ψt ◦ Φs
for all s, t for which the above transformations are defined.
Proof. Above, we showed that [A, B] = 0 if and only if Φt∗ B = B for
all t. By Proposition 9.4, the latter is true if and only if
Φt∗ Ψs = Φt ◦ Ψs ◦ Φ−t == Ψs
for all s, as was to be shown. 
Example 11.7. Let A = x∂x + y∂y and B = −y∂x + x∂y . Let
Φt (x, y) = (et x, et y)
Ψt (x, y) = (cos(t)x − sin(t)y, sin(t)x + cos(t)y)
be the corresponding flows. Consider the transformation
(u, v) = Φt (x, y).
We have
 t 
e 0
(∂x , ∂y ) = (∂u , ∂v ) ,
0 et
 t   −t 
e 0 e u
x∂x + y∂y = (∂u , ∂v ) t
0 e e−t v
= u∂u + v∂v ,
 t   −t 
e 0 −e v
−y∂x + x∂y = (∂u , ∂v )
0 et e−t u
= −v∂u + u∂v
Similarly, consider the transformation
(u, v) = Ψt (x, y).
Equivalently,
   
u x
= Rt ,
v y

where
 
cos t − sin t
Rt =
sin t cos t
Math 5190: lecture notes. 54

is the indicated rotation matrix. We have


(∂x , ∂y ) = (∂u , ∂v )Rt ,
 
x
x∂x + y∂y = (∂u , ∂v )Rt ,
y
 
−1 u
= (∂u , ∂v )Rt Rt ,
v
= u∂u + v∂v .
As well,
 
−y
−y∂x + x∂y = (∂u , ∂v )Rt
x
 
x
= (∂u , ∂v )Rt R π2 ,
y
 
u
= (∂u , ∂v )Rt R π2 R−t ,
v
 
u
= (∂u , ∂v )R π2 ,
v
= −v∂u + u∂v
Proposition 11.8. Let F : U → V be a diffeomorphism and G : V →
U the inverse transformation. For A : U → Rn , a vector field, and
f : U → R, a function, we have
F ∗ (gB) = G∗ (g)F ∗ (B). (80)
Proof. Homework. 
Proposition 11.9. Let A, B : U → Rn be vector fields and Φt the
flow generated by A. Then A is an infinitesimal conformal symmetry
of B if and only if for all t, Φt is a conformal symmetry of B.
Proof. Suppose that A is a conformal symmetry of B. Then, there
exists a non-zero function g : U → R such that [A, gB] = 0. Hence,
Φt∗ (gB) = gB.
By the definition of push-forward and the above Proposition,
Φt∗ (gB) = Φ∗−t (g)Φt∗ (B).
Hence,
Φt∗ (B) = hB,
Math 5190: lecture notes. 55

where
h = g/Φ∗−t (g).
Conversely, suppose that Φt is a conformal symmetry for all t. Hence,
there exists a function ht : U → R such that
Φt∗ B = ht B.
Taking d/dt, setting t = 0, and using Proposition 11.5 gives
[A, B] = ḣ0 B,
as was to be shown. 
Example 11.10. The radial vector field A = x∂x + y∂y is an infini-
tesimal conformal symmetry of the vector field B = ∂x − (x/y)∂y . Let
Φt be the flow generated by A. Fix t and consider the transformation
(u, v) = Φt (x, y) = (et x, et y).
We have
 t  
e 0 1
∂x − (x/y)∂y = (∂u , ∂v )
0 et −u/v
= et (∂u − (u/v)∂v )
Therefore, Φt is a conformal symmetry of B.
Proposition 11.11. Let A, B : U → Rn be vector fields and Φt the
flow generated by A. If A is an infinitesimal conformal symmetry of
B, then Φt preserves the integral curves of B.
Proof. See Proposition 9.4. 
Example 11.12. Let us return to the vector fields of the preceding
examples. The integral curves of B are the semicircles about the origin,
x2 + y 2 = R2 in the upper half-plane, y > 0. Changing the variables
to u, v the equation of the circle becomes u2 + v 2 = (et R)2 . Therefore
the Φt transforms a circle of radius R to a circle of radius et R.
Math 5190: lecture notes. 56

12. Symmetry Integration


Definition 12.1. Let A : U → R2 where U ⊂ R2 be a C 1 vector field
and
dy/dx = ω(x, y) (81)
a non-autonomous ODE, where ω : U → R is a C 1 function. We call
A an infinitesimal symmetry of such an ODE if A is an infinitesimal
conformal symmetry of ∂x + ω(x, y)∂y .
Proposition 12.2. A vector field A = ξ ∂x + η ∂y is an infinitesimal
symmetry of (81) if and only if
ω 2 ξy + ω(ξx − ηy ) − ηx + ξωx + ηωy = 0 (82)
Proof. We have
[ξ∂x + η∂y , ∂x + ω∂y ] = (−ωξy − ξx )∂x + (−ωηy − ηx + ξωx + ηωy )∂y
The RHS is proportional to ∂x − ω∂y if and only if (82) holds. 
Equation (82) is called the symmetry determining equation.
Proposition 12.3. Let u = F 1 (x, y), v = F 2 (x, y) be a change of
coordinates. An ODE (81) in standard coordinates is equivalent to an
ODE
dv
= ω̂(u, v),
du
where
vx + ω(x, y)vy
ω̂(u, v) = , (83)
ux + ω(x, y)uy
where ux = ∂x (u) = ∂u/∂x, etc.
Proof. By the usual change of coordinate formula for vector fields, we
have
∂x = ux ∂u + vx ∂v
∂y = uy ∂u + vy ∂v
∂x + ω∂y = (ux + ωuy )∂u + (vx + ωvy )∂v
= (ux + ωuy )(∂u + ω̂(u, v)∂v ).

Definition 12.4. Let U1 , V1 ⊂ U be subdomains. We call a diffeomor-
phism F : U1 → V1 a symmetry of the ODE (81) if
ω̂(u, v) = ω(u, v)
where (u, v) = F (x, y) and where ω̂ is defined as per (83).
Math 5190: lecture notes. 57

Proposition 12.5. A vector field A : U → R2 is an infinitesimal


symmetry of the ODE (81) if and only if the corresponding flow Φt is
a symmetry for all t.
Proposition 12.6. The unit vector field ∂y is a symmetry of (81) if
and only if ω(x, y) = f (x) is independent of y. In this case, the general
solution of (81) is given by
Z x
y= f (s) ds

In light of the above proposition, one method of integration is to rec-


tify the symmetry and then find the general solution by means of a
quadrature.
Proposition 12.7. The vector field g(y)∂y is a symmetry of an ODE
(81) if and only if there exists a function f (x) such that
ω(x, y) = f (x)g(y).
In this case, the general solution is given by
Z y Z x
ds
= f (s)ds,
g(s)
i.e, by separation of variables.
Proof. Let us introduce coordinates
y
ds
Z
u = x, v= ,
g(s)
and thereby rectify
g(y)∂y = ∂v .
By the preceding Proposition, in these new coordinates the ODE has
the form
dv
= f (u),
du
for some function f (u) = f (x). Since
∂u + f (u)∂v = ∂x + f (x)g(y)∂y
the form of the ODE in standard coordinates is
dy
= f (x)g(y),
dx
a separable equation. Again, by the preceding proposition, the general
solution has the form Z u
v= f (s),
as was to be shown. 
Math 5190: lecture notes. 58

Proposition 12.8. The vector field A = x∂x + y∂y is a symmetry of


an ODE (81) if and only if
ω(x, y) = f (y/x),
for some function f of one variable. In this case, the general solution
is given by
Z y/x
ds
x = exp (84)
f (s) − s
Proof. We rectify A by introducing coordinates u = y/x, v = log(x).
Now
x∂x = ∂v − u∂u ,
y∂y = u∂u ,
x∂x + y∂y = ∂v .
By (83) the ode dv/du = f1 (u) is equivalent to (81) where
1/x
f1 (u) = ,
−y/x2 + ω(x, y)/x
or equivalently,
1
ω(x, y) = f (y/x), where f (u) = − u.
f1 (u)
Therefore, the general solution is given by
Z u
v= f1 (s) ds,

or equivalently, by
y/x
ds
Z
log(x) = .
f (s) − s

Math 5190: lecture notes. 59

13. Differential 1-forms.


Definition 13.1. Let V be an n-dimensional vector space. A linear
form is a linear mapping V → R. The set of all such forms the n-
dimensional vector space V ∗ , called the dual of V . Let v1 , . . . , vn be a
basis of V . The unique basis ǫ1 , . . . , ǫn ∈ V ∗ such that
ǫi (vj ) = δji , i, j = 1, . . . , n
is called the dual basis of v1 , . . . , vn .
Note: it’s customary to represent elements of Rn as column vectors and
elements of (Rn )∗ as row vectors.
Definition 13.2. Let U ⊂ Rn be a domain. A differential 1-form, or
simply a differential, is a mapping α : U → (Rn )∗ . For a vector field
A : U → Rn , we define α · A : U → R to be the function
(α · A)(x) = α(x)(A(x)).
Expressing a differential as a row vector of functions

α = α1 , . . . , αn
and a vector field as a column vector of functions
 1
A
A =  ...  ,
An
we can express the product of a differential 1-form and a vector field as
n
X
α·A = αi Ai .
i=1
1 2
We say that α is of class C , C , analytic, etc if its components α1 , . . . , αn
are C 1 , C 2 , analytic, etc.
Definition 13.3. Let f : U → R be a differentiable function. We call
the 1-form
 
∂f ∂f
df := ,..., n
∂x1 ∂x
the differential of f . We call the constant differentials dx1 , . . . , dxn the
basic 1-forms and write
n
X ∂f
df = i
dxi .
i=1
∂x
Math 5190: lecture notes. 60

Proposition 13.4. Let F : U → V be a diffeomorphism and y j =


F j (x1 , . . . , xn ) the corresponding change of coordinates. Then,
n
j
X ∂y j
dy = dxi (85)
i=1
∂xi
are the unique differential 1-forms having the property that

dy j · = δij . (86)
∂yi
Furthermore, every differential 1-form α : U → (Rn )∗ has the unique
expression
Xn
α= α̃j (y 1 , . . . , y n )dy j , (87)
j=1

where α̃j : V → R are functions defined by


 

α̃ (y , . . . , y ) = α · j (x1 , . . . , xn )
j 1 n
∂y
n
X ∂xi
= j
αi (x1 , . . . , xn )
i=1
∂y

Proposition 13.5. Let f : U → R be a C 1 function and y j = F j (x1 , . . . , xn )


a coordinates system. Then,
n
X ∂f
df = j
dy j . (88)
j=1
∂y

Indeed, df is the unique differential 1-form characterized by the condi-


tion that for every C 0 vector field A : U → Rn we have
df · A = A[f ]. (89)
A differential can also be regarded as the integrand of a line integral.
Proposition 13.6. Let α : U → R be a continuous differential and
γ : I → U a C 1 curve, where I = [t0 , t1 ]. The line integral
Z Z Z t1 X n
α = (α ◦ γ)(γ̇) = αi (γ(t))γ i′ (t) dt
γ I t0 i=1

is invariant under orientation preserving reparameterization.


Definition 13.7. If α is a 1-form such that α = df for some function
f : U → R, then we call α an exact differential.
Math 5190: lecture notes. 61

Proposition 13.8. Let α = df be an exact differential on a domain


U. Let γ : I → U be a parameterized curve. Then,
Z
α = f (x1 ) − f (x0 ),
γ

where xi = γ(ti ), i = 0, 1. Conversely, suppose that U is a connected


domain, and that for all curves γ 1 , γ 2 with endpoints x0 , x1 ∈ U we
have Z Z
α= α.
γ1 γ2
Then, α = df is an exact differential with
Z x
f (x) = α, x ∈ U
x0
where the above integral is taken along an arbitrary curve that connects
x0 to x.
Definition 13.9. Say that α : U → (Rn )∗ is a closed differential if
∂αi (x1 , . . . , xn ) ∂αj (x1 , . . . , xn )
− = 0, (90)
∂xj ∂xi
for all 1 ≤ i < j ≤ n.
Proposition 13.10. An exact differential is closed. Conversely, if U
is a simply connected domain, then every closed differential is exact.
Math 5190: lecture notes. 62

14. Symmetries and integrating factors.


In this section, we resrict our attention to 2-dimensional, simply
connected domains U ⊂ Rn .
Proposition 14.1. Let α : U → (R2 )∗ be a C 1 differential. There
after a suitable domain restriction U1 ⊂ U there exists a C 1 function
µ : U1 → R such that µα is exact.
Proof. Write α = M dx+N dy. Consider the vector field A = −N ∂x +
M ∂y . By rectifying this vector field, we can find a non-constant first
integral ξ = f (x, y) such that
dξ · A = A[ξ] = 0.
Hence,
M N
= Mξy − Nξx = 0.
ξx ξy
We now define
µ = ξx /M
on a suitably restricted domain on which ξ is defined and M 6= 0. In
this way,
µα = ξx dx + ξx (N/M) dy = ξx dx + ξy dy = dξ.
If M is identically zero, we simply define µ = 1/N. 
Definition 14.2. Let α = Mdx + Ndy be a C 1 differential defined on
a domain U ⊂ R2 . We call γ : I → U, where I ⊂ R is an interval, an
integral curve of α if the function
f (t) = (α ◦ γ)(t) · γ̇(t)
is identically zero for all t ∈ I.
Proposition 14.3. Let α, γ be as above. Then γ is an integral curve
of α if and only if Z
α=0
γ|I1
for all subintervals I1 ⊂ I.
Proposition 14.4. Let α, γ be as above. Then γ is an integral curve
of α if and only if γ describes a solution of the ODE dy/dy = −M/N.
Definition 14.5. We call a function µ : U → R such that µα is closed
an integrating factor of a differential α.
To solve an ODE dy/dx = −M/N it suffices to find an integrating
factor for the differential α = Mdx + Ndy and restrict the domain.
Math 5190: lecture notes. 63

Proposition 14.6. Let A = ξ∂x + η∂y be a vector field and α =


M dx + N dy a differential. If α · A = 0, then A is an infinitesimal
symmetry of dy/dx = −M/N.
Proof. The vector fields A and −N∂x + M∂y are proportional if and
only if A · α = 0. Observe that if f is any function, then
[A, f A] = A[f ]A.
Hence, A is a conformal symmetry of f A for all functions f . 
If α · A = 0, we call A the trivial symmetry of dy/dx = −M/N.
Proposition 14.7. Let U ⊂ R2 be a simply connected domain, A :
U → Rn a vector field and α = M dx + N dy a differential such that
α · A 6= 0. Then A is a symmetry of the ODE dy/dx = −M/N if and
only if
1
µ= (91)
α · A,
is an integrating factor of α.
Proof. Write A = ξ(x, y)∂x +η(x, y)∂y . Let µ be as above, and consider
the differential
M N
µα = dx + dy,
Mξ + Nη Mξ + Nη
= 1/(η − ω(x, y)ξ) (dy − ω(x, y)dx)
where
ω(x, y) = −M/N.
By (90) and by Proposition 13.10, the above differential is closed if and
only if
∂x (η − ξω)−1 + ∂y (ω/(η − ξω)) = 0.
Taking derivatives and clearing denominators, the above condition is
equivalent to
−∂x (η − ξω) + (η − ξω)ωy − ω∂y (η − ξω) = 0
The above, in turn, is equivalent to the determining equation (82). 
Example 14.8. Let’s revisit the integration method for a homogeneous
differential equations
dy
= f (y/x).
dx
Earlier, we showed that x∂x + y∂y is an infinitesimal symmetry. Hence,
µ = ((dy − f (y/x)dx) · (x∂x + y∂y ))−1 = (y − xf (y/x))−1
Math 5190: lecture notes. 64

is an integrating factor. Hence the solution of the homogeneous ODE


can be obtained by integrating
Z  
−f (y/x) 1
dx + dy .
y − xf (y/x) y − xf (y/x)
The above differential is closed (check!), hence locally exact. To obtain
solution, let us integrate the dy term (i.e., we treate x as a constant.
We obtain
dy 1 dy
Z Z
=
y − xf (y/x) y/x − f (y/x) x
Z y/x
du
= F (y/x) =
u − f (u)
Observe that
1
∂x F (y/x) = F ′ (y/x)(−y/x2) = (−y/x2 )
y/x − f (y/x)
 
1 y 1 f (y/x)
= − =− −
x y − xf (y/x) x y − xf (y/x)
Therefore,
Z  
−f (y/x) 1
dx + dy =
y − xf (y/x) y − xf (y/x)
− log(x) + F (y/x) + C.
Thus, the same general solution can be obtained by integration. The
necessary changes of variables arise naturally in the course of this in-
tegration.
Proposition 14.9. The vector field f (x)∂y is a symmetry of an ODE
(81) if and only if only if
ω(x, y) = p(x)y + q(x),
where p(x) = f ′ (x)/f (x).
Proof. From the determining equation (82), f (x)∂y is a symmetry if
and only if
ωy = f ′ (x)/f (x).
This is true if and only if ω has the form shown above. 
Proposition 14.10. The general solution of a linear ODE
dy
= p(x)y + q(x)
dx
Math 5190: lecture notes. 65

is given by
x
q(s)
Z
y = f (x) ds,
f (s)
Rx
where f (x) = exp( p(s) ds).
Proof. With the above definition, we have f ′ (x)/f (x) = p(x). Hence,
f (x)∂y is a symmetry, and hence 1/f (x) is an integrating factor. We
have
Z x
f ′ (x)y
 
dy q(x) y q(s)
− 2
dx − dx = d −d ds.
f (x) f (x) f (x) f (x) f (s)
Therefore,
1
Z
(dy − (p(x)y + q(x))dx) =
f (x)
Z x
y/f (x) − (q(s)/f (s))ds,

where the constant of integration is incorportated in the RHS integral.



Math 5190: lecture notes. 66

15. Reduction of order


Consider a general second-order ordinary differential equations
d2 y
 
dy
= ω x, y, . (92)
dx2 dx
We convert the above ODE to a 3D autonomous system by introducing
the variable y1 = dy/dx and writing
ẋ = 1
ẏ = y1 ,
y˙1 = ω(x, y, y1).
In other words, solving a 2nd order ODE is equivalent to rectifying the
vector field
A = ∂x + y1 ∂y + ω(x, y, y1)∂y1 (93)
on a 3-dimensional domain with coordinates (x, y, y1). We will show
that a conformal infinitesimal symmetry of A allows us to reduce the
order of (92). This means that we can formulate a first order ODE
dv
= φ(u, v) (94)
du
such that the solutions of (92) are obtained from the solutions of (94)
by a quadrature (the old-fashioned name for an anti-derivative).
Example 15.1. Let us consider a 1-dimensional conservative mechan-
ical system
d2 x
= a(x), (95)
dt2
where a(x) is a position-dependent acceleration function. We introduce
a velocity variable v and seek a function v = f (x) that gives velocity
as a function of position. Since v = dx/dt, the function f must satisfy
dv
= f ′ (x(t))f (x(t)) = a(x(t)).
dt
In other words, v = f (x) is a solution of the following 1st order ODE
dv a(x)
= . (96)
dx v
If we know a solution v = f (x) of (96), we can obtain a solution
x = p(t) of (95) by integrating the ODE
dx
= f (x).
dt
Math 5190: lecture notes. 67

This can be done using a quadrature, namely


Z x
du
t= . (97)
f (u)
The general solution of (95) should depend on 2 constants of integra-
tion: the initial position and the initial velocity. One of these constants
arises in (97), the other arises when (and if) we solve (96).
Now for 1-dimensional mechanical systems it is always possible to
solve the reduced equations by means of a quadrature (separate vari-
ables). Indeed the general solution of (96) is given by
1 2
v = E − U(x), (98)
2
where Z x
U(x) = − a(s)ds (99)
x0
is called a potential function and x0 is some fixed position. The E is a
constant of integration known as the energy of the mechanical system.
Note that if we change x0 , we change E. Thus, energy is relative
in exactly the same way that the potential function is relative. The
relationship between time and position can now be given implictly as
Z x
ds
t=± p . (100)
2E − 2U(s)
We use the + branch on the portions of the trajectory where the ve-
locity is positive and − on the portions where it is negative.
From the point of view of geometry, we are trying to describe the
integral curves of the 3-dimensional vector field
∂t + v∂x + a(x)∂v . (101)
on a 3-dimensional domain (called phase space) with coordinates (t, x, v).
These integral curves are describe by means of 2 first-integrals:
1
E = v 2 + U(x),
2
x
ds
Z
τ = t − sgn(v) p ,
2E − U(s)
Definition 15.2. We define an autonomous 2nd order ODE to be an
ODE of the form
d2 y
= ω(y, dy/dx). (102)
dx2
Math 5190: lecture notes. 68

The solutions of such an ODE correspond the integral curves of the


vector field
 
ω(y, y1)
∂x + y1 ∂y + ω(y, y1)∂y1 = ∂x + y1 ∂y + ∂y1 .
y1
A 1st integral can be obtained by integrating the 1st order ODE
dy1 ω(y, y1)
= (103)
dy y1
Indeed, if y1 = f (y) is a solution of (103), we can obtain a general
solution of (102) by using a quadrature to solve dy/dx = f (y); i.e. by
taking
Z y
ds
x= .
f (s)
The case of a conservative mechanical system presented above is an
example of an 2nd order autonomous ODE. Here is another example.
Example 15.3. Consider the 2nd order ODE
d2 y dy
2
= .
dx dx
The above is an autonomous equation, so we can reduce order and
consider the equation
dy1 y1
= = 1.
dy y1
The general solution is y1 = y − C1 . We now integrate
dy
= y − C1
dx
using a quadrature to obtain
x = log(y − C1 ) − C̃,

or equivalently, after setting C2 = eC̃ ,


y = C2 ex + C1 .
Thus, a strategy for integrating a 2nd order ODE is to find and
rectify an infinitesimal symmetry. Doing so will, in effect, transform
the given 2nd order ODE into an autonomous one. In order to put this
strategy into practice we have to inquire about the transformation law
for 2nd order ODEs.
Math 5190: lecture notes. 69

Definition 15.4. Let (u, v) = F (x, y) be a diffeomorphic transforma-


tion. The first prolongation of this diffeomorphism is the transforma-
tion (u, v, v1 ) = F (1) (x, y, y1) where, c.f., (83)
vx + y1 vy
v1 = (104)
ux + y1 uy
The second prolognation is the transformation (u, v, v1, v2 ) = F (2) (x, y, y1, y2)
where
ux vy − uy vx (∂x + y1 ∂y )[v1 ]
v2 = 3
y2 + . (105)
(ux + y1 uy ) ux + y1 uy
Earlier, we proved the following.
Proposition 15.5. Let (u, v) = F (x, y) be a diffeomorphic transfor-
mation. Then the solutions of a 1st order ODE
dy
= ω(x, y)
dx
are equivalent to the solutions of
dv
= ω̃(u, v)
dy
where where v1 = ω̃(u, v) is related to y1 = ω(x, y) by the 1st prolonga-
tion formula (104).
Generalizing to 2nd order, we have the following.
Proposition 15.6. Let (u, v) = F (x, y) be a diffeomorphic transfor-
mation. Then the solutions of a 2nd order ODE (92) are equivalent to
the solutions of
d2 v
 
dv
= ω̃ u, v, ,
du2 du
where v2 = ω̃(u, v, v1 ) is related to y2 = ω(x, y, y1) by the 2nd prolon-
gation formula (105).
Proof. Observe that
∂x + y1 ∂y = (ux + y1 uy )(∂u + v1 ∂v ),

where v1 , y1 are related as above. Similarly,

∂x + y1 ∂y + y2 ∂y1 = (ux + y1 uy )(∂u + v1 ∂v + v2 ∂y2 )


where y2 , v2 are related by the above formula. 
Math 5190: lecture notes. 70

Definition 15.7. Let A = ξ(x, y)∂x +η(x, y)∂y be a vector field defined
on a 2-dimensional domain. Setting
υ(x, y, y1) = (∂x + y1 ∂y )[η(x, y) − y1 ξ(x, y)] (106)
= ηx + y1 (ηy − ξx ) − y12ξy , (107)
we call the 3-dimensional vector field
A(1) = ξ(x, y)∂x + η(x, y)∂y + υ(x, y, y1)∂y1 (108)
the first prolongation of A.
Proposition 15.8. Let Φt be the flow generated by a 2-dimensional
(1)
vector field A = ξ(x, y)∂x + η(x, y)∂y . Then Φt is the flow generated
by the 3-dimensional vector field A(1) .
Proof. Homework. 
Definition 15.9. Let us say that a vector field A = ξ∂x + η∂y is an
infintesimal symmetry of a 2nd order ODE (92) if the prolongation
A(1) is an infinitesimal conformal symmetry of the vector field ∂x +
y1 ∂y + ω(x, y, y1)∂y1 .
Proposition 15.10. A 2nd order ODE (92) admits the translation
generator ∂x as an infinitesimal symmetry if and only if it is an au-
tonomous ODE of the form (102).
Proof. Note that the prolongation of ∂x is equal to ∂x + 0∂y + 0∂y1 .
Calculating the Lie bracket, we obtain
[∂x , ∂x + y1 ∂y + ω(x, y, y1)∂y1 ] = ∂x (ω(x, y, y1))∂y1
Thus in order for ∂x to be a conformal symmetry, ωx must vanish. 
Next, let us consider 2nd order equations that admit x∂x + y∂y as
a symmetry. Since we know how to rectify this vector field, we will
obtain a class of equations that admit a reduction of order. We will
call these types of equations, homogeneous 2nd order ODEs.
Proposition 15.11. A 2nd order ODE (81) admits x∂x + y∂y as an
infinitesimal symmetry if and only if it has the form
d2 y
= ω(y/x, dy/dx)/x (109)
dx2
The general solution of such an equation is given by
Z y/x !
du
x = exp , (110)
g(u)
Math 5190: lecture notes. 71

where v1 = g(v) is a solution of the reduced ODE


dv1
= ω(v, v1 + v)/v1 − 1. (111)
dv
Proof. Let us rectify
x∂x + y∂y = ∂u
by taking
u = log(x), v = y/x
as a change of coordinates. Applying the prolongation formulas (104)
and (105) gives
−y/x2 + y1 /x
v1 = = −y/x + y1 ,
1/x
1/x2
v2 = y2 + (∂x + y1 ∂y )[y1 − y/x]
1/x3
= xy2 + (y/x2 − y1 /x)/(1/x)
= xy2 + y/x − y1
Now in the u, v coordinates, an autonomous ODE (one that admits ∂u
as a symmetry) has the form
d2 v
= ω̃(v, v1 ).
du2
Setting v2 = ω̃(v, v1 ) and applying the above formula gives
y2 = v2 /x
Writing
v2 = ω̃(v, v1 ), y2 = (v2 + v1 )/x = ω(v, y1)/x,
we obtain the transformation formula
ω̃(v, v1 ) = ω(v, v1 + v) − v1 .
The reduction method now follows from doing the reduction of an au-
tonomous ODE in (u, v) coordinates, and then re-expressing the solu-
tion in (x, y) coordinates. 
We expect the flow generated by an infinitesimal symmetry of a 2nd
order equation to be a symmetry of that equation. We will not take the
time to prove this in full generality, but will make note of the following.
Recall that the vector field x∂x + y∂y generates the flow
Φt (x, y) = (et x, et y).
Proposition 15.12. Let k > 0 be a constant. The transformation
(u, v) = (kx, ky) preserves a homogeneous 2nd order ODE (109).
Math 5190: lecture notes. 72

Proof. Apply the transformation formulas (104) (105). 


Example 15.13. Consider the following homogeneous 2nd order ODE
d2 y 1 dy/dx
2
= + .
dx y x
The above is a homogeneous equation with
ω(y/x, y1) = (y/x)−1 + y1 .
The reduced equation is therefore,
dv1
= (1/v + v1 + v)/v1 − 1 = v + 1/v.
dv
The general solution of the reduced equation is
p
v1 = ± v 2 + 2 log(v) + C1 .
Since y1 = v1 + v, the general solution of the given 2nd order equation
is !
Z y/x
ds
x = exp ± p (112)
C2 s2 + 2 log(s) + C1
One could also say that this homogeneous system is equivalent to a
mechanical system in u, v coordinates with potential −v 2 /2 − log(v).
Let’s derive the reduced equation directly. Set
v = y/x
dv dy
v1 = x = −v
dx dx
dv1 d2 y
v2 = x = x 2 − v1
dx dx
d2 y dy
=x 2 − +v
dx dx
The given ODE can now be expressed as
1 dy dy
v2 = + − +v
v dx dx
1
=v+
v
We seek a function v1 = g(v) that is compatible with the solution of
the above equation. This implies that
1
v2 = g ′(v)g(v) = v +
v
Math 5190: lecture notes. 73

and hence that


g(v)2 = v 2 + 2 log(v) + C
Having obtained g(v) we can now solve for v as a function of x by
noting that
dv g(v)
=
dx x
Hence,
Z v
ds
log(x) = ;
g(s)
another way to write the solution above.

You might also like