Sophus Lie’s Approach to Differential Equations

IAP lecture 2006 (S. Helgason)

1 Groups

Let X be a set. A one-to-one mapping ϕ of X onto X is called a bijection.

Let B(X) denote the set of all bijections of X onto X.
Let ϕ, ψ ∈ B(X). We define a product ϕψ by (ϕψ)(x) = ϕ(ψ(x)) x ∈
X. Then ϕψ ∈ B(X). The identity map I : X → X belongs to B(X).
If ϕ ∈ B(X) we define ϕ−1 by ϕ−1 (ϕ(x)) = x. Then ϕ−1 ∈ B(X) and
ϕϕ−1 = ϕ−1 ϕ=I.Also, the product ϕψ is associative, i.e.,

(ϕψ)τ = ϕ(ψτ ) .

In fact,

(ϕψ)τ (x) = (ϕψ)(τ (x)) = ϕ(ψ(τ (x))) = ϕ(ψτ (x)) .

In general, ϕψ 6= ψϕ. For this consider the example X = {1, 2, 3}. Then
B(X) consists of the permutation of 1, 2, 3. Let ϕ : X → X fix 1 but
exchange 2 and 3. Let ψ : X → X fix 3 but exchange 1 and 2. Then
ϕψ(1) = 3 but ψϕ(1) = 2 so ϕψ 6= ψϕ.
The set B(X) is a prototype of a group. A group is a set G such that
for any a, b, ∈ G is an associated new element ab ∈ G such that

a(bc) = (ab)c .

One also assumes the existence of an element e ∈ G such that ea = ae

for all a ∈ G. One also assumes that for each a ∈ G there exists an
element a−1 ∈ G such that aa−1 = a−1 a = e.
Example. B(X) is a group.
If G is a group a subset H ⊂ G is called a subgroup if h ∈ H , k ∈ H
implies hk, h−1 ∈ H. Then H is a group. H is called a normal subgroup
if gHg −1 ⊂ H for all g ∈ G. If H is a normal subgroup the family of
cosets gH = {gh : h ∈ H} can be made into a group, denoted G/H,
(the factor group) by the product definition

g1 H · g 2 H = g 1 g2 H (well-defined) .

G is said to be abelian if xy = yx for all x, y ∈ G. Let G 0 denote the

subgroup of G generated by all commutators xyx −1 y −1 (x, y ∈ G). The
1
G0 is a normal subgroup and G/G0 is abelian. We can form the sequence

G ⊂ G0 ⊃ (G0 )0 ⊃ · · · .

G is said to be solvable if this ends with e.

As mentioned, B(X) is a group and its subgroups are called transfor-
mation groups of X.

2 Polynomials

It is well known from algebra that each polynomial equation p(x) = 0 of

degree 1, 2, 3, 4 is solvable by radicals. Through Abel’s work one knows
that this is no longer true for p(x) of degree 5.
In Galois’ theory the theory becomes deeper. Here to each polynomial
p(x) is associated a certain finite group, a subgroup of the group of
permutations of the roots. The solvablity of this group is equivalent to
p(x) = 0 being solvable by radicals. The equation x 5 − x − 13 = 0 can be
shown to have Galois group, the full permutation group S 5 of five letters
which can be shown not to be solvable (in contrast to S i (i < 5)). Thus
the equation is not solvable by radicals.

3 Lie’s Program.

Inspired by Galois’ theory, Lie got the idea of doing something in this
spirit for differential equations. First let us look at a first order equation
dy
(3.1) = F (x, y) ,
dx
a solution being by definition a function y = u(x) such that u 0 (x) =
F (x, u(x)).
Thus a solution is a curve in R2 (an integral curve). A transformation
T ∈ B(X) (X = R2 ) is said to leave the equation (3.1) stable if it
permutes the integral curves.
Special case. F (x, y) = g(x). Then (3.1) becomes
dy
= g(x) ,
dx
with the solutions
Zx
u(x) = g(t) dt + C C = constant.
0
2
The integral curves are all parallel so each translation T t : (x, y) →
(x, y + t) leaves the equation stable. Note also that

Ts+t = Ts Tt ,

that is, t → Tt is a 1-parameter group. More generally, consider the

differential equation

dy Y (x, y)
(3.2) = (Lie’s notation)
dx X(x, y)

and assume we have a 1-parameter group ϕ t (t ∈ R) of differentiable

bijections of R2 leaving (3.2) stable.
Consider the vector field on R2

n
d(ϕt ·p)
o 't •p
Φp = dt p
t=0

Here Φp is the tangent vector to the orbit ϕt · p at p ∈ R2 . Thinking of

a vector at p = (x, y) as a directional derivative we write the vector field
in the form
∂ ∂
(3.3) Φp = Φx,y = ξ(x, y) + η(x, y) .
∂x ∂y

Theorem 3.1 (Lie (1874)). Assume ϕt (t ∈ R2 ) is a 1-parameter group

leaving (3.2) stable. The ∃ function U (x, y) such that

∂U −Y ∂U X
(3.4) = , =
∂x Xη − Y ξ ∂y Xη − Y ξ

and U (x, y) = const is the solution to (3.2).

A proof will be indicated later.

The statement is equivalent to the statement that
X dy − Y dx
Xη − Y ξ
3
is an exact differential
∂U ∂U
dx + dy
∂x ∂y
or equivalently that (Xη−Y ξ)−1 is an integrating factor for the equation

X dy − Y dx = 0 .

The theorem means that the solution to (3.2) can be found by inte-
grating the equations (3.4) which is done in Calculus. 1
We sketch the method. We write (3.4) in the form

∂U ∂U
= M (x, y) , = N (x, y) .
∂x ∂y
Assuming U a solution we have
Zy
U (x, y) = N (x, z) dz + g(x)
a

for some function g(x). Then

Zy
∂
M (x, y) − N (x, z) dz = g 0 (x) .
∂x
a

Since ∂/∂y of the left hand side equals

My − N x = 0

the function g(x) does exist and the formula for U (x, y) gives a solution.
In the old days this was expressed: Under the assumption of the theorem,
equation (3.2) can be solved by quadratures. This was to emphasize the
analogy with solving an algebraic equation in radicals when the Galois
group was abelian. Later we discuss the solvable case.
Lie’s remarkable theorem seems unfortunately ignored in recent books
on ordinary differential equations. Before a proof let us consider a few
examples. A favorite of mine is the equation

dy y + x(x2 + y 2 )
(3.5) =
dx x − y(x2 + y 2 )

1
At least in the good old Thomas.
4
which at first looks rather formidable. However, we can write it

ȕ-Į
dy y
− • (x,y)
dx x = x2 + y 2
(3.6) y dy (0,0)
1+
x dx

The slope of the ray from (0, 0) to (x, y) is xy = tan α and the slope of
dy
the tangent to the integral curve through (x, y) is dx = tan β. Since

tan β − tan α
tan(β − α) =
1 + tan α tan β

(3.6) states that

tan(β − α) = x2 + y 2 .
This means that the angle β − α is constant as (x, y) varies on a circle
with center (0, 0). Thus each rotation

(3.7) ϕt : (x, y) → (x cos t − y sin t , x sin t + y cos t)

maps each integral curve into another integral curve, in other words
leaves the equation (3.5) stable. Also the rotations ϕ t form a group
(ϕt+s = ϕt ϕs ). Here we have from (3.7)
 
d(ϕt · p) ∂ ∂
Φp = = −y +x
dt t=0 ∂x ∂y

so by Lie’s theorem
−1
(x − y(x2 + y 2 ))x − (y + x(x2 + y 2 ))(−y) = (x2 + y 2 )−1


is an integrating factor. Also

   
X dy − Y dx x y
= − y dy − + x dx
x2 + y 2 x2 + y 2 x2 + y 2

and by inspection we see that this is

∂U ∂U
dx + dy
∂x ∂y
5
for
y x2 + y 2
U (x, y) = Arctan ( ) − .
x 2
The solution U (x, y) = c can be written
1
y = x tan( (x2 + y 2 ) + c) .
2
Exercise. (an example from Lie’s paper)
Consider the differential equation
dy y
(3.8) =f .
dx x

•(e x,e y)
t t

On each ray • (x,y)

•0

consider the tangent to the integral curve through (x, y). At the point
(et x, et y) the slope of the tangent is the same. Thus the map

ϕt : (x, y) → (et x, et y)

leaves the equation stable. Also, ϕ t+s = ϕt ϕs . Here

 
d(ϕt · p) ∂ ∂
Φp = =x +y .
dt t=0 ∂x ∂y

By the theorem −f ( xy ) dx + dy has integrating factor (y − f ( xy )x)−1 , in

other words
!
!
∂ 1 ∂ −f xy
=
∂x y − f xy x ∂y y − f xy x

and y

−f x
1 ∂U ∂U
y dx + y
dy = dx + dy .
y−f x x y−f x x ∂x ∂y
The solution is U (x, y) = c.

6
 To find U we know from the last equation that
Zy
dz
U (x, y) =
+ g(x)
z − f xz x
1

and as explained before the relation

y

∂U −f x

= y
∂x y−f x x

gives a formula for g 0 (x) in terms of x alone. Thus the solution to (3.8)
is given by two quadratures.
x2
Exercise. Carry this out for f (z) = z 2 + 2z. Solution is y = c−x .
We still do not have a differential equation analog to the algebraic the-
orem for solvable Galois group. For this we consider groups depending
on more parameters. Let X = R2 and let G be the subgroup of B(X)
preserving distances and orientation. If σ ∈ G let t be the translation
such that σ · 0 = t · 0. Then t−1 σ · 0 = 0 so t−1 σ is a rotation k around
the origin. Let θ(σ) be the angle between the x-axis ` and k · ` and let
t = (x(σ) , y(σ)). Then
      
x x(σ) cos θ(σ) sin θ(σ) x
σ: → + .
y y(σ) − sin θ(σ) cos θ(σ) y

A simple computation shows

x(στ −1 ) = x(σ) − x(τ ) cos(θ(σ) − θ(τ )) + y(τ )(sin θ(σ) − θ(τ ))

y(στ −1 ) = y(σ) − x(τ ) sin(θ(σ) − θ(τ )) − y(τ ) cos(θ(σ) − θ(τ ))
θ(στ −1 ) = θ(σ) − θ(τ ) mod (2π) .

Thus the elements of G are parameterized by three parameters such

that the parameters of product and inverse are smooth functions of the
parameters of the factors. This suggests the definition of a Lie group.
What is a Lie algebra?

Consider a transformation group of R n depending effectively on r

parameters:

T : (x1 , . . . , xn ) → (x01 , . . . , x0n ) where

x0i = fi (x1 , . . . , xn ; t1 , . . . , tr ) ,
T = I for (t1 , . . . , tr ) = (0, . . . , 0) .
7
 Assume that if S is the transformation corresponding to the parame-
ters (s1 , . . . , sr ) then
T S −1 : x0i = fi (x1 , . . . , xn ; u1 , . . . , ur )
and the ui are smooth functions of the ti , sj above. Generalizing Φp
above Lie defined the vector fields on R n :
n  
X ∂fi ∂
(3.9) Tk = .
∂tk t=0 ∂xi
i=1

Now if X and Y are any vector fields on R n they map functions into
functions so X ◦ Y and Y ◦ X are well defined maps of functions. So is
the bracket
[X, Y ] = X ◦ Y − Y ◦ X
and it is easily seen to involve only first order derivatives (the second
derivatives cancel) so [X, Y ] is another vector field.
Lie proved the fundamental fact that the vector fields (3.9) satisfy
r
cpk` Tp ,
X
(3.10) [Tk , T` ] =
p=1

where the coefficients cpk` are constants, satisfying

r
cpk` = −cp`k , (cpkq cq`m + cpmq cqk` + cp`q cqmk ) = 0 .
X
(3.11)
q=1
Pr
If we put g = {X = 1 ap Tp |ap ∈ R}, relations (3.11) can be stated:
For X, Y, Z ∈ g we have
(3.12) [X, Y ] = −[Y, X]

(3.13) [X, [Y, Z]] + [Y, [Z, X]] + [[Z, X], Y ] = 0 .

A vector space with a rule of composition (X, Y ) → [X, Y ] satisfying
(3.12) – (3.13) is called a Lie algebra. Thus g is a Lie algebra.
Lie also proved a converse that to a finite-dimensional Lie algebra
there exists (locally) a transformation group corresponding to it.
A Lie algebra g is abelian if [X, Y ] = 0 for all X, Y ∈ g. Putting
Dg = [g, g] and DS g = D(DS−1 g), g is said to be solvable if DS g = 0
for some S.
We can now state an analog to the solvablity of p(x) = 0 in radicals
in terms of the Galois group.
8
Theorem 3.2. Suppose the system of differential equations
dyj
(3.14) = f j (x, y1 , . . . , yr ) 1≤j≤r
ax
has a solvable r-dimensional stability group in (x, y 1 , . . . , yr ) with r-dimensional
orbits. Then the solution can be found by repeated quadratures, explicitly
given by the group.
For the proof a good reference is P. Olver, Applications of Lie Groups
to Differential Equations (Springer 1986).
The case r = 1 is the one in Theorem 3.1. Let us verify the theorem
for an example (from Olver’s book).
Example. Consider the differential equation

d2 y dy
(3.15) x2 2
= f (x − y) .
dx dx
dy
Putting z = dx we can put the equation in the form (3.14)

dz dy
(3.16) x2 = f (xz − y) , = z.
dx dx
Here the transformations
z t
Ts,t : (x, y, z) → (sx, y + tx, + ) s > 0,t ∈ R
s s
leave the system (3.16) stable. Also

Tσ,τ ◦ Ts,t = Tσs,t+τ s .

The one-parameter subgroups Ts,0 and T1,t generate the vector fields
∂ ∂ ∂ ∂
X1 = x −z , X2 = x + .
∂x ∂z ∂y ∂z
Then [X1 , X2 ] = X2 so the Lie algebra RX1 +RX2 is solvable. According
to the theorem, (3.15) is solvable by quadratures. Let us verify this. We
put w = xz − y. Then dw dz
dx = x dx so

dw
x = f (w)
dx
giving
dw
Z
= log |x| + C .
f (w)
9
Writing this in the form w = g(x) we arrive at the equation
dy y g(x)
− = .
dx x x
∂
This has the symmetry group (x, y) → (x, y + tx) with vector field x ∂y .
−1
By Theorem 3.1 x is an integrating factor for
y + g(x)
dy − dx = 0
x
that is
dy y + g(x) ∂U ∂U
− 2
dx = dx + dy .
x x ∂x ∂y
Thus
∂U 1 y
= so U= + h(x)
∂y x x
∂U y y g(x)
= − 2 + h0 (x) = − 2 − 2 .
∂x x x x
Thus
Zx
g(t)
h(x) = − dt
t2
1
and the solution to (3.15) takes the form
Zx
g(t)
y = −x dt + cx .
t2
1

4 The Heat Equation

This is the equation

∂2u ∂u
(4.1) 2
= .
∂x ∂t
Here a transformation(x, t, u) → (x1 , t1 , u1 ) leaves (4.1) stable if it per-
mutes the solutions. Lie determined the stability group (cf. Olver’s book
above). It is a six-dimensional group times an (uninteresting) infinite-
dimensional group. One 1-parameter subgroup is quite interesting. It
gives the result that if f (x, t) is a solution to (4.1) then so is the function
 
1 −sx2 x t
u(x, t) = e 1+4st f , .
(1 + 4st)1/2 1 + 4st 1 + 4st
10
5 Proof of Lie’s Theorem

So far as I know this interesting theorem does not occur in most recent
books on ordinary differential equations. Older proofs seem a bit obscure
(but take a look at Lie’s original proof (Collected Works, Vol. 3)). The
proof in Olver’s book is clean and rigorous but contained in a longer
theory of prolongations.
Below is a short proof. Suppose ϕt is a 1-parameter group leaving the
equation

dy Y (x, y)
(5.1) =
dx X(x, y)

stable. If U (x, y) = c is a solution we have with ϕ t (x, y) = (xt , yt ),

U (xt , yt ) = c(t) (all t)

so
∂U dxt ∂U dyt
+ = c0 (t)
∂x dt ∂y dt
and by (3.3)

∂U ∂U
(5.2) ξ+ η = c0 (0) .
∂x ∂y
Secondly
∂U ∂U dy
+ =0
∂x ∂y dx
so
∂U ∂U
(5.3) X+ Y = 0.
∂x ∂y

If c0 (0) 6= 0 we can normalize U such that c0 (0) = 1. Then (5.2) and

(5.3) imply
∂U −Y ∂U X
= , =
∂x Xη − Y ξ ∂y Xη − Y ξ
so (Xη − Y ξ)−1 is an integrating factor for X dy − Y dx = 0 as claimed.
On the other hand, if c0 (0) = 0, (5.2)–(5.3) imply dy/dx = η/ξ so the
integral curves are just the orbits of ϕ t .

11
6 Other Contacts with Group Theory. An Example.

Many of you are familiar with the mean-value theorem for harmonic
functions, that is solutions to Laplace’s equation
 2
∂2

∂
Lu = + u = 0.
∂x21 ∂x22
The theorem states that the solutions are characterized by
Z2π
1
(6.1) u(x1 , x2 ) = u(x1 + r cos θ , x2 + r sin θ) dθ
2π
0

for all x1 , x2 and all r ≥ 0. Geometrically this means that for each
p ∈ R2 , r > 0, we have
(6.2) u(p) = (M r u)(p) ,
the mean value of u on a circle with center p and radius r. (Exercise:
What is the 1-dimensional version of this result?)
Invoking the above group G of isometries of R 2 let x denote the trans-
lation by (x1 , x2 ), let k = k(θ) the rotation by θ and y the translation
by (r, 0). Then (6.1) can be stated
Z2π
1
(6.3) u(x · 0) = u(xky · 0) dk .
2π
0

A Riemannian manifold X with distance function d has an analog to

L above, the Laplace-Beltrami operator. We assume X is two-point ho-
mogeneous, that is given any two pairs (p, q), (p 0 , q 0 ) in X with d(p, q) =
d(p0 , q 0 ) there exists an element g from the isometry group I(X) such
that g · p = p0 , g · q = q 0 . Fix o ∈ X and let K be the subgroup of I(X)
fixing o. Then the solutions to Lu = 0 are characterized by
Z
(6.4) u(x · o) = u(xky · o) dk x, y ∈ I(X) ,
K

dk being the Haar measures on K (Godement).

Since the set {xky · o : k ∈ K} is the sphere in X with center x and
radius d(o, y) formula (6.4) can also be written
u(p) = (M r u)(p) p∈X, r≥0
in exact analogy with (6.2).
12