Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Flows and Lie Derivatives

Luigi T. Sousa
June 2021

1
Contents
1 Integral Curves and Differential Equations 3

2 A downstream Flow of Vectors 4


2.1 One-parameter group of transformations . . . . . . . . . . . . . . 5

3 Lie derivatives 7
3.1 Induced Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.1.1 Pushforward . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.1.2 Pullback . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.1.3 Induced maps on mixed tensors . . . . . . . . . . . . . . . 10
3.2 Lie Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

4 References 14

2
1 Integral Curves and Differential Equations
In past discussions we used a curve c : R → M to derive our tangent vectors
X|p at a point p = c(0), but one could ask the opposite: given a vector field X,
is there a curve that “follows” the vector field, i.e, passes through M such that
the elements of the vector field X are always tangent to c(t)? Given a chart
(U, x), our question would be, in component form

dxµ (c(t))
= X µ (c(t)) , (1)
dt
which is just a system of n first order ordinary differential equations (ODEs),
and thus do have a solution since the X µ are all smooth functions of x, and
better yet, if we are given a set of initial conditions xµ0 = xµ (c(0)) then the
Existence and Uniqueness theorem from ODEs guarantees there is, at least
locally, just one solution to equation (1). The solution to equation (1) is what is
known as an integral curve, since finding such a curve is equivalent to solving
a set of ODEs.

Figure 1: Integral curves of a embedded vector field

A heuristic way to think about integral curves would be if your vector field
describes the velocity field of a ideal fluid, then the integral curve starting at
some point would be the trajectory of a single particle of the fluid passing
through that point. Another heuristic example would be the usual magnetic
field lines usually depicted coming from a magnet, which are the integral curves
of the magnetic field.

3
2 A downstream Flow of Vectors
Since a integral curve is defined by the differential equation depending on
the parameter t and the initial condition x0 , one could then denote one such
curve as σ(t, x0 ) and eq(1) would then become
d µ
σ (t, x0 ) = X µ (σ(t, x0 )) (2)
dt
with initial condition
σ µ (0, x0 ) = xµ0 .
This is nice since we now can denote all possible integral curves on M induced
by a vector field X by just σ(t, x) : R × M → M , and we call σ(t, x) the flow
generated by X ∈ X(M ). Mathematically, a flow satisfies

σ(t, σ(s, x0 )) = σ(t + s, x0 ) (3)

for any t, s ∈ R such that both sides of eq(3) make sense (i.e, t, s, t + s are all in
the domain of σ). This can be proven by noting that we have
d µ
σ (t, σ(s, x0 )) = X µ (σ(t, σ(s, x0 )))
dt
σ µ (0, σ(s, x0 )) = σ µ (s, x0 ),
and, on the other hand,
d d
σ µ (t + s, x0 ) = σ µ (t + s, x0 ) = X µ (σ(t + s, x0 ))
d(t + s) dt
σ µ (0 + s, x0 ) = σ µ (s, x0 ),
and since both σ(t, σ(s, x0 )) and σ(t+s, x0 ) satisfy the same ODE with the same
initial condition, then by the Existence and Uniqueness Theorem they must be
the same.
In a more intuitive way, what this expression says is that each initial con-
dition x generates only one “flow line” (integral curve), since plugin in a given
point x0 will give the same curve at “time” t = 0 or at a later “time” t 6= 0,
even after a arbitrary time step to t + s.
∂ ∂
-Example: let M = R2 and X = y ∂x + x ∂y be a vector field. To find the
flow generated by X, we must look at the equation that locally defines the flow:

d
σ µ (t, (x, y)) = X µ σ(0, (x, y)) = X µ (x, y) =⇒

dt t=0
 !  
d x y
=⇒ σ t, = ,
dt t=0 y x
 !  
x x
σ 0, = ,
y y

4
giving us the set of equations:
   
x(0) x
=
y(0) y
      
d x(t) y 0 1 x
= = =⇒
dt t=0 y(t) x 1 0 y
 
0 1 
   t    
x(t) 1 0 x(0) cosh(t) sinh(t) x
=⇒ =e = ,
y(t) y(0) sinh(t) cosh(t) y
and finally, we get that our flow is
 !   
x cosh(t) sinh(t) x
σ t, = .
y sinh(t) cosh(t) y

2.1 One-parameter group of transformations


Given a flow σ(t, x), if t is fixed then σt (x) becomes a diffeomorphism σ :
M −→ M , and we can turn this family of diffeomorphisms into a commutative
group by the rules:
i) (σt ◦ σs )(x) = σt+s (x) = σs+t (x) = (σs ◦ σt )(x) (this follows from (3));
ii) σ0 = unity/identity element (σ0 (x) = x);
iii) σ−t = (σt )−1 (since σt ◦ σ−t = σt−t = σ0 ).
This group is called the one-parameter group of transformations of M
and while it may look like (R, +) locally, overall globally it usually will be a bit
different. Take for instance the result obtained in our last example
 
cosh(t) sinh(t)
σt = ,
sinh(t) cosh(t)

which is basically a member of SO+ (1, 1), the Lorentz group in 1+1 dimensions.
If we consider the action of σε , with ε infinitesimally small, then from (2)
we find that the coordinates of a point x ∈ M is mapped to

σεµ (x) = σ µ (ε, x) = xµ + εX µ (x). (4)

In this context we call X the infinitesimal generator of σt .


One can also refer to σ as the exponentiation of X, denoted as

σ µ (t, x) = etX xµ .

The reason for this denomination is as follows: if we consider the Taylor series
expansion for our flow σ(t, x), we would get
X tn dn X n n 
t d

σ µ (t, x) = µ
 µ 
σ (s, x) = σ (s, x) =
n! dsn
s=0 n! dsn
s=0
n≥0 n≥0

5
d
= et ds σ µ (s, x) s=0 ,

(5)
d dxν ∂
thus recalling that σ µ (0, x) = xµ and noting that ds = ds ∂xν = X by the
chain-rule, we have that eq(5) thus becomes

σ µ (t, x) = etX xµ .

The flow follows some nice exponential rules:


i)σ µ (0, x) = e0X xµ = xµ ;
d d tX µ 
ii) σ µ (t, x) = e x = XetX xµ ;
dt dt
iii)σ µ (t, σ(s, x)) = σ µ (t, esX xν ) = etX esX xµ = e(t+s)X xµ = σ µ (t + s, x) .

6
3 Lie derivatives
Given two vector fields X and Y and their respective flows σ(s, x) and τ (t, x),

d µ
σ (s, x) = X µ (σ(s, x))
ds
d µ
τ (t, x) = Y µ (τ (t, x)) ,
dt
the Lie derivative of the vector field Y along the flow of X would be defined as
1 
LX Y = lim Y |σε (x) − Y |x , (6)
ε7→0 ε

but there’s a problem with this definition: the difference between Y |σε (x) and
Y |x is ill-defined, since they live in different vector spaces, i.e, Y |σε (x) ∈ Tσε (x) M
and Y |x ∈ Tx M , so we need a sensible way to map Y |σε (x) into Tx M . Indeed
there’s a natural way to do this mapping, as we shall discuss next:

3.1 Induced Maps


A smooth map ϕ : M −→ N between two manifolds M and N (N might be
M itself) induces two types of linear maps between their tangent and cotangent
spaces:

3.1.1 Pushforward
-Definition(3.1.1): given a smooth map ϕ : M −→ N , the pushfor-
ward (or sometimes differential map) of ϕ, denoted ϕ∗ (dϕ if talking about
dif f erential maps) is the linear map

ϕ∗ : Tp M −→ Tϕ(p) N

that pushes vectors from the tangent space of M, Tp M , f orward into the tan-
gent space of N , Tϕ(p) N .

ϕ
ϕ∗ V
Tϕ(p) N
V ϕ(p)

p Tp M
N
M ϕ∗

Figure 2: ϕ and it’s pushforward ϕ∗

7
One can get a explicit form for ϕ∗ by recalling the definition of a tangent
vector as a directional derivative along a given curve. Let f ∈ F(N ), then
f ◦ ϕ ∈ F(M ). Recall that a vector V ∈ Tp M acts on a scalar field g ∈ F(M )
to produce a number, V [g], in particular, V can act on f ◦ ϕ, and so we define
ϕ∗ V ∈ Tϕ(p) N by the relation
(ϕ∗ V )[f ] ≡ V [f ◦ ϕ] . (7)
Given charts (U, x) on M and (V, y) on N , eq(7) then becomes
(ϕ∗ V )[f ◦ y −1 (y0 )] ≡ V [f ◦ ϕ ◦ x−1 (x0 )] , (8)
where x0 = x(p) and y0 = y(ϕ(p)). Expanding V into it’s components and local
basis, V = V µ ∂x∂ µ and ϕ∗ V = Ve ν ∂y∂ ν , we have that eq(8) turns into

∂ ∂
Ve ν ν [f ◦ y −1 (y0 )] = V µ µ [f ◦ ϕ ◦ x−1 (x0 )] ,
∂y ∂x

thus choosing f = y α we are able to find the explicit relation between Ve ν and
V µ as follows:
∂ ∂
Ve ν ν [f ◦ y −1 (y0 )] = Ve ν ν [y α ◦ y −1 ◦ y ◦ ϕ(p)] =
∂y ∂y
∂y α
= Ve ν(ϕ(p)) = Ve ν δνα = Ve α , (9)
∂y ν
and, from the right-hand side:
∂ ∂
Vµ [f ◦ ϕ ◦ x−1 (x0 )] = V µ µ [y α ◦ ϕ ◦ x−1 ◦ x(p)] =
∂xµ ∂x
∂y α
=Vµ (ϕ(p)) , (10)
∂xµ
finally, piecing together (9) and (10), we get that
α
∂y
Ve α = V µ µ (ϕ(p)) , (11)
∂x
α
where ∂y
∂xµ is just the Jacobian matrix of the map ϕ. We can naturally extend
ϕ∗ to any (q, 0)-tensor as in ϕ∗ : T q0,p (M ) −→ T q0,ϕ(p) (N ) by using the same
relation.
-Proposition: let ϕ : M −→ N and ψ : N −→ P . Then, the pushforward of
the composite map ψ ◦ ϕ : M −→ P is given by
(ψ ◦ ϕ)∗ = ψ∗ ◦ ϕ∗ .
-Proof: let xµ , y ν and z ρ be coordinates in M , N and P , respectively,
X ∈ Tx M be a vector in M and f ∈ F(P ) be a scalar field in P . Then, if we
analyse the action of (ψ ◦ ϕ)∗ on X we have that
(ψ ◦ ϕ)∗ X[f ] = X[f ◦ ψ ◦ ϕ] ,

8
which, in component form becomes
ρ ρ ν
e ρ = X µ ∂z = X µ ∂z ∂y
X
∂xµ ∂y ∂xµ
ν

ρ ν
∂z ∂y
where we made use of the chain-rule as usual. Note that ∂y ν ∂xµ is the product

of two Jacobian matrices, in particular, the matrices of the transformations


ψ and ϕ, a.k.a, the composition of ψ∗ and ϕ∗ , since matrix multiplication is
the same as composition of linear transformations, and thus, reverting back to
abstract form,

(ψ ◦ ϕ)∗ X[f ] = (ψ∗ ◦ ϕ∗ )X[f ] ⇐⇒ (ψ ◦ ϕ)∗ = (ψ∗ ◦ ϕ∗ )


-Example: let σt : M −→ M be a one-parameter transformation in M , then
σt (x) naturally induces a isomorphism between Tx M and Tσt (x) M ,

(σt )∗ : Tx M −→ Tσt (x) M


X(x) 7→ X(σt (x)) .

3.1.2 Pullback
-Definition(3.1.2): the pullback of a (smooth) map ϕ : M −→ N , de-
noted ϕ∗ , is the linear map between the respective cotangent spaces,

ϕ∗ : Tϕ(p)

N −→ Tp∗ M .

One thing to note here is that the pullback actually goes in the opposite direction

of ϕ, hence it’s name: it pulls covectors from Tϕ(p) N back into Tp∗ M . The

explicit form of ϕ is given by the relation

hϕ∗ ω, V i = hω, ϕ∗ V i =⇒ (12)


∂y µ ∂y µ ∂y α
eµ V µ = ωµ Ve µ = ωµ V α
=⇒ ω α
= ωµ α V α = ωα µ V µ ⇐⇒
∂x ∂x ∂x
∂y α
⇐⇒ ω
e µ = ωα µ ,
∂x
α
where once again we see the Jacobian ∂y ∂xµ of ϕ. As before, one can extend the
pullback to arbitrary (0, r)-tensors ϕ∗ : T 0r,ϕ(p) (N ) −→ T 0r,p (M ) by the same
construction.
-Proposition: let ϕ : M −→ N and ψ : N −→ P , then the pullback of
ψ ◦ ϕ : M −→ P is given by

(ψ ◦ ϕ)∗ = ϕ∗ ◦ ψ ∗ .

-Proof: let ω ∈ Tψ(ϕ(x)) P be a covector in P and X ∈ Tx M be a vector in
M . Then, we know by eq(12) that

h(ψ ◦ ϕ)∗ ω, Xi = hω, (ψ ◦ ϕ)∗ Xi = hω, (ψ∗ ◦ ϕ∗ )Xi = hψ ∗ ω, ϕ∗ Xi =

9
= h(ϕ∗ ◦ ψ ∗ )ω, Xi
∴ h(ψ ◦ ϕ)∗ ω, Xi = h(ϕ∗ ◦ ψ ∗ )ω, Xi ⇐⇒ (ψ ◦ ϕ)∗ = (ϕ∗ ◦ ψ ∗ )


3.1.3 Induced maps on mixed tensors


Up to now, we worked with maps which didn’t necessarily have an inverse,
but that’s a bit problematic when we try to investigate what the induced maps
would look like on mixed tensors, or even what the inverse of a pullback or
pushforward would be like, so now, let’s assume the map ϕ : M −→ N is a
dif f eomorphism and define some new stuff:
The inverse pushf orward of ϕ is the (linear) map (ϕ∗ )−1 : Tϕ(p) N −→ Tp M
µ
given explicitly by the Jacobian of ϕ−1 , ∂x
∂y ν , which itself is the inverse of the
∂y µ
Jacobian of ϕ, ∂xν . This follows from eq(11) since now the Jacobian is invertible:

∂y α ∂xν ∂y α ∂xν ∂xν


Ve α = V µ µ =⇒ Ve α α = V µ µ α = V µ µ = V µ δµν = V ν
∂x ∂y ∂x ∂y ∂x
ν
∂x
∴ V ν = Ve α α .
∂y
This is also known as just the pullback of a vector Ve ∈ Tϕ(p) N
-Proposition: let ϕ = σt : M −→ M , then ((σt )∗ )−1 = (σ−t )∗ .
-Proof: we know that

σt ◦ σ−t = σt−t = σ0 = idM

and also
(ψ ◦ ϕ)∗ = ψ∗ ◦ ϕ∗ ,
so putting both facts together we get

(σt ◦ σ−t )∗ = (σt )∗ ◦ (σ−t )∗ .

From the left-hand side, it’s easy:

(σt ◦ σ−t )∗ = (σ0 )∗ = (idM )∗ = 1 ,

thus, if we apply ((σt )∗ )−1 on both sides, we get that

((σt )∗ )−1 ◦ 1 = ((σt )∗ )−1 ◦ (σt )∗ ◦ (σ−t )∗

((σt )∗ )−1 = 1 ◦ (σ−t )∗


((σt )∗ )−1 = (σ−t )∗ .


10
Similarly, the inverse pullback of ϕ is the map (ϕ∗ )−1 : Tp∗ M −→ Tϕ(p)

N,
−1
given again by the Jacobian of ϕ , with the same relation as before,
∂xα
ωµ = ω
eα ,
∂y µ

but it’s good to keep in mind that here ω ∈ Tϕ(p) N and ω e ∈ Tp∗ M , contrary to
the vectors where V ∈ Tp M and Ve ∈ Tϕ(p) N . This is also known as just the
pushf orward of a covector ω e ∈ Tp∗ M . Also in a similar way as before, we get
∗ −1 ∗
that ((σt ) ) = (σ−t ) by the same reasoning.
Naturally, with ϕ being a diffeomorphism, it’s now easy to talk about push-
forwards and pullbacks of mixed tensors, since now we know what the push-
forward does to covariant objects and what the pullback does to contravariant
stuff. The pushf orward of a general tensor of type (q, r) is given by

α ...α ∂y µ1 ∂y µq ∂xβ1 ∂xβr


Teµ1 ...µqν1 ...νr = T 1 qβ1 ...βr α1 ... αq ... ,
∂x ∂x ∂y ν1 ∂y νr
while it’s pullback is given by

α1 ...αq ∂xµ1 ∂xµq ∂y β1 ∂y βr


T µ1 ...µqν1 ...νr = Te β1 ...βr ... ... .
∂y α1 ∂y αq ∂xν1 ∂xνr

3.2 Lie Derivatives


Now that we’ve explored induced maps, we have a natural way to map
a vector Y |σ (x) ∈ Tσε (x) M into Tx M , giving us the tool needed to properly
define the Lie derivative.
-Definition(3.2): given two vector fields X and Y in M and their respective
flows,
d µ
σ (s, x) = X µ (σ(s, x))
ds
d ν
τ (t, x) = Y ν (τ (t, x)) ,
dt
the Lie derivative of the vector field Y along the flow of X is
1 
LX Y = lim (σ−ε )∗ Y |σε (x) − Y |x =
ε7→0 ε

1  d
= lim (σ−ε )∗ Y |σε (x) − (σ0 )∗ Y |σ0 (x) = (σ−s )∗ Y |σs (x)
ε7→0 ε ds s=0
where we used the pullback of Y by σ−ε to make sure both vectors are always
in the same vector space such that their difference makes sense.
Let’s analyse the above definition and get an actual way to compute the Lie
derivative: first, we evaluate the former term inside the brackets as

(σ−ε )∗ Y |σε (x0 ) [f ] ≡ Y [f ◦ σ]

11
Ye µ ∂µ f (x) x0 = Y µ ∂µ f (σ(−ε, x0 )) σ(ε,x0 ) ,

where f ∈ F(M ) and Ye µ ≡ (σ−ε )∗ Y µ . Since that’s true for any function f , in
particular we can choose f = xν ,

Ye µ ∂µ xν x = Y µ ∂µ σ ν (−ε, x0 ) σ(ε,x )

0 0

Ye µ δµν = Y µ ∂µ (xν − εX ν ) σ(ε,x0 )


Ye ν = Y µ (σ(ε, x0 ))[δµν − ε∂µ X ν (σ(ε, x0 ))]

Ye ν = Y µ (xα α ν ν α α
0 + εX (x0 ))[δµ − ε∂µ X (x0 + εX (x0 ))]

Ye ν = [Y µ (x0 ) + εX α (x0 )∂α Y µ (x0 )][δµν − ε∂µ (X ν (x0 ) + εX α (x0 )∂α X ν (x0 ))] , 1
which, after multiplying out and omitting any term with ε2 (since ε is taken
infinitesimally small, in the limit they would go away on definition (3.2)) gives
us
Ye ν = Y ν (x0 ) + εX α (x0 )∂α Y ν (x0 ) − εY α (x0 )∂α X ν (x0 ) ,
and putting this result back into definition (3.2) yields

LX Y = (X α ∂α Y ν − Y α ∂α X ν )eν .

Let’s now consider the commutator (or Lie bracket) of two vectors X and
Y (let f ∈ F(M )):

[X, Y ]f = X[Y [f ]] − Y [X[f ]]. (13)


Given a local chart, eq (13) becomes, in component form:

[X, Y ]f = X µ ∂µ [Y ν ∂ν [f ]] − Y ν ∂ν [X µ ∂µ [f ]] =

= X µ (∂µ Y ν )∂ν f + X µ Y ν ∂µ ∂ν f − (Y ν (∂ν X µ )∂µ f + Y ν X µ ∂ν ∂µ f ).


Since f, X µ and Y ν are all smooth functions and summation indices are dummy
indices, we can freely rewrite the indices and reorder some terms:

[X, Y ]f = X µ (∂µ Y ν )∂ν f − Y µ (∂µ X ν )∂ν f + X µ Y ν ∂µ ∂ν f − X µ Y ν ∂µ ∂ν f =

= (X µ ∂µ Y ν − Y µ ∂µ X ν )∂ν f . (14)
Equation (14) shows us that the commutator of two vector fields X and Y
is another vector field given by

[X, Y ] = XY − Y X = (X µ ∂µ Y ν − Y µ ∂µ X ν )eν ,

and from that we can see that

LX Y ≡ [X, Y ] ,
1 Here d
we expanded the Taylor series for Y and X around x0 and took X α ∂α ≡ ds

12
which is the reason why some authors define the Lie derivative to be the com-
mutator [X, Y ].
Since the Lie derivative measures how much two vectors fail to commute,
it’s geometrical interpretation is that of the vector that closes the parallelogram
made from X and Y : if we start at a point x0 ∈ M and go along the flow of
X followed by the flow of Y , and compare that to start going through the flow
of Y and then X we have that (let σ(s, x) and τ (t, x) be the flows of X and Y ,
respectively):

τ µ (δ, σ(ε, x)) ' τ µ (δ, xν + εX ν (x)) '


' xµ + εX µ (x) + δY µ (xν + εX ν (x)) '
' xµ + εX µ (x) + δY µ (x) + εδX ν ∂ν Y µ (x) , (15)

σ µ (ε, τ (δ, x)) ' σ µ (ε, xν + δY ν (x)) '


' xµ + δY mu (x) + εX µ (xν + δY ν (x)) '
' xµ + δY µ (x) + εX µ (x) + εδY ν ∂ν X µ (x) . (16)
If we take the difference of (15) and (16) we see that it’s proportional to the
Lie Derivative of Y along X:

τ µ (δ, σ(ε, x)) − σ µ (ε, τ (δ, x)) ' εδ(X ν ∂ν Y µ − Y ν ∂ν X µ )(x) =

= εδ[X, Y ]µ = εδLX Y µ .

τ (t, x)
σ(ε, τ (δ, x))
ε
τ (δ, x)

εδ[X,Y]
τ (δ, σ(ε, x))
δ
δ
σ(s, x)
x0 ε σ(ε, x)

Figure 3: Failure of closure of the parallelogram of X and Y .

13
4 References
1. Geometry, Topology and Physics, Second Edition, by Mikio Nakahara
2. INTRODUCTION TO DIFFERENTIAL GEOMETRY, by Joel W. Rob-
bin and Dietmar A. Salamon
3. An Introduction to Manifolds, Second Edition, by Loring W. Tu

4. https://en.wikipedia.org/wiki/Pushforward_(differential)

14

You might also like