Classnotes: Geometry & Control of Dynamical Systems, M. Kawski.

April 21, 2009 138

4.7 The Levi-Civita connection and parallel transport

In the earlier investigation, characterizing the shortest curves between two points was cast as a
variational problem. The resulting necessary condition has the form of a system of second order
differential equations. As a consequence, geodesics, as solutions of smooth initial value problems
automatically have substantial regularity properties.
This section takes a closer look at the coefficients that appeared in the Euler Lagrange equation of
the variational problem: These are linear combinations of partial derivatives of the Riemannian
metric. Intuitively they measure how much the metric varies from point to point. Or even more
markedly, how much the difference between the metric on the manifold from the flat Euclidean
manifold varies from point to point. Clearly such rates of change of the Riemannian metric are
predestined to lead to an intrinsic curvature of the manifold.
The plan for the upcoming discussions is to revisit the Christoffel symbols as derivatives of the
metric (??)Chris1, and to attribute rates of change of the metric to rates of change of vector
fields. As a first step, we explore the rates of change (directional derivatives) of coordinate
vector fields. This development shall be guided by a close inspection of derivatives in/of curvi-
linear coordinates in Euclidean spaces, and exploration of derivatives of tangent vector fields
on hypersurfaces. Eventually this leads to an axiomatic definition of covariant derivatives or
connections (or connexions). The last part of this section is devoted to the discussion of some
technical properties of this derivative. Subsequent sections shall take this notion of connection
to develop elegant characterizations of curvature.

The previous section introduced the Christoffel symbols in equation (216)

1  ∂gik ∂gjk ∂gij 
[ij, k] = + − (286)
2 ∂uj ∂ui ∂uk
basically as abbreviations for certain linear combinations of partial derivatives of the Riemannian
metric. Using the symmetry of the Riemannian metric it is straightforward to solve backwards to
express the partial derivatives of the Riemannian metric as linear combinations of the Christoffel
symbols:
1  ∂gik ∂gjk ∂gij ∂gji ∂gki ∂gjk  ∂gik
[ij, k] + [jk, i] = ( + − ) + ( + − ) = (287)
2 ∂uj ∂ui ∂uk ∂uk ∂uj ∂ui ∂uj
What matters to us is that the coefficients in the geodesic equation are indeed, essentially, the
rates of change of the Riemannian metric: If (in a suitable coordinate chart) the metric is
2
constant, then d (u◦σ)
dt2
≡const. This says that if a manifold is flat, then the shortest curves are
(straight) lines.
∂ ∂
Recall that the components gij = h ∂u i , ∂uj i of the metric are the inner products of the coordi-
nate vector fields. One goal of this section is to attribute the rate of change of the Riemannian
∂
metric, in a geometrically consistent way, to the rates of change of the basic vector fields ∂u i
themselves. Given the characteristic features of any notion of derivative, linearity and being a
derivation (i.e., satisfying Leibniz’ rule or the product rule), it is not far-fetched to ask that such
a notion of a derivative ∇Z (in the direction of a vector field Z) satisfy (compare theorem (??))
   
∂ ∂ ∂ ∂ ∂ ∂
Zgij = Zh ∂ui , ∂uj i = h∇Z ∂ui , ∂uj i + h ∂ui , ∇Z ∂uj i (288)

In the end, this compatibility condition connecting notions of rates of change of vector fields to
the rate of change of the metric plays a critical role in singling out a special covariant derivative
among many possible others.
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 21, 2009 139

This leads to the question about what one might mean by the rate of change of a vector field. In
some sense, this corresponds to, instead of focusing on shortest, to study notions of straight and
parallel. Before reading on, it is recommended to spend some effort on the following exploration:

Exercise 4.65 Explore/develop a notion of the rate of change of the coordinate vector fields
∂ ∂
∂r = cos θD1 + sin θD2 and ∂θ = −r sin θD1 + r cos θD2 with respect to (“in the directions of”)
each other, and in the directions of D1 and D2 . [[This is much simpler than it looks – very
simple and intuitive ideas for directional derivatives are the right thing to do!]].

One basic observation is that the standard coordinate vector fields Di in Euclidean space serve
as models for straightness, for constancy, or parallelism. Consequently, for basically any notion
of derivative, the derivative of these fields ought to be zero (although not for Lie derivatives, see
below).
∂ ∂
In terms of the preceding exercise, compare how ∂θ changes in the direction of D1 = ∂x with how
∂ ∂ ∂
D1 changes in the direction ∂θ . However, note that the Lie derivative satisfies L ∂ ∂x = −L ∂ ∂θ ,
∂θ ∂x
∂
and thus does not have the desired asymmetry that reflects that D1 as straight while ∂θ as not
straight. In the Euclidean context it is natural to define

Definition 4.16 The Euclidean connection (or Euclidean

m covariant derivative) is the map
m

∇ : Γ∞ (Rm ) × Γ∞ (Rm ) 7→ Γ∞ (Rm ) defined by ∇Z X i Di = Z(X i )Di .
P P
i=1 i=1

Looking ahead to general manifolds where there is no reason to expect that a typical (“any”)
vector field be constant in any reasonable sense, the defining equation may be written in a longer,
more suggestive form as:
m m m
!
X X X
i
∇Z (X) = ∇Z X Di = Z(X i )Di + X i (∇Z (Di )) = Z(X i )Di (289)
| {z }
i=1 i=1 =0 i=1

The next objective is to develop similar notions for derivatives on manifolds. These should be
intrinsic, i.e. not utilize an imbedding into an ambient Euclidean space. Nonetheless, coordinate
changes and surfaces in R3 provide valuable guidance on how to define covariant derivatives on
general manifolds.

Exercise 4.66 Let (x1 , x2 ) denote the standard coordinates on R2 and suppose (u1 , u2 ) = (Φ1 ◦
(x1 , x2 ), Φ2 ◦ (x1 , x2 )) is another set of smooth coordinates. Expand the coordinate vector fields
∂
∂ui
as linear combinations of Dj with partial derivatives of Φk as coefficients. Calculate the
matrix gij representing the standard Riemannian, i.e. Euclidean, metric w.r.t. the coordinates
(u1 , u2 ). Use the identities ∇Di Dj ≡ 0 and equation (??) to calculate ∇ ∂ ∂u∂ j .
∂ui
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 21, 2009 140

Next explore vector fields on hypersurfaces in Euclidean. spaces. Working locally, assume
w.l.o.g. that the manifold is the graph of a function xm+1 = f (x1 , x2 , . . . , xm ). For the sake of
clarity consider the special case m = 2 (note that if m = 2 there necessarily is some repetition
among the indices of Γkij ). Suppose M 2 ⊆ R3 is the graph of a smooth function x3 = f (x1 , x2 ).
Using the local coordinates u(x) = (x1 , x2 ) the coordinate vector fields and induced metric are
∂
!
1 + (D f )2 (D f ) · (D f )
∂u 1 = D 1 + (D 1 f )D3 1 1 2
∂
and (gij )i,j=1,2 = (290)
∂u2
= D2 + (D2 f )D3 (D1 f ) · (D2 f ) 1 + (D2 f )2
∂
Using the Euclidean connection (??), and identifying the vector fields ∂u i on M with the natural
∂ 3
extensions of the vector fields ı∗ ∂ui on ı∗ M to vector fields on R , calculate
∂
∇ ∂ ∂u1 = ∇D1 +(D1 f )D3 (D1 + (D1 f )D3 ) = (D12 f )D3
∂u1

∂
∇ ∂ ∂u1 = ∇D2 +(D2 f )D3 (D1 + (D1 f )D3 ) = (D2 D1 f )D3
∂u2
(291)
∂
∇ ∂ ∂u2 = ∇D1 +(D1 f )D3 (D2 + (D2 f )D3 ) = (D1 D2 f )D3
∂u1
∂
∇ ∂ ∂u2 = ∇D2 +(D2 f )D3 (D2 + (D2 f )D3 ) = (D22 f )D3 .
∂u2

In general there is no reason for these vector fields on Rm+1 to restrict to vector fields on M m .
However, in the case of smooth hypersurfaces M m ⊆ Rm+1 it is natural to extend the set of
coordinate vector fields on M m to a frame on Rm+1 by adjoining a normal vector field ν. The
existence of such normal vector field depends crucially on the Riemannian metric on the ambient
space. In the case of m = 2 this yields
∂
∂u1
= D1 + (D1 f )D3
∂
∂u2
= D2 + (D2 f )D3 (292)
ν = − (D1 f )D1 − (D2 f )D2 + D3 .
∂
It is straightforward to invert this system, to solve for Di in terms of ∂ui
and ν
(1+D2 f )2 ∂ (D1 f )(D2 f ) ∂ (D1 f )
D1 = 1+(D1 f )2 +(D2 f )2
· ∂u1
− 1+(D1 f )2 +(D2 f )2
· ∂u2
− 1+(D1 f )2 +(D2 f )2
·ν
(D1 f )(D2 f ) ∂ (1+D2 f )2 ∂ (D2 f )
D2 = − 1+(D1 f )2 +(D2 f )2
· ∂u1
+ 1+(D1 f )2 +(D2 f )2
· ∂u2
− 1+(D1 f )2 +(D2 f )2
·ν (293)
(D1 f ) ∂ (D2 f ) ∂ 1
D3 = 1+(D1 f )2 +(D2 f )2
· ∂u1
+ 1+(D1 f )2 +(D2 f )2
· ∂u2
− 1+(D1 f )2 +(D2 f )2
· ν.
∂
Use these to rewrite the derivatives (??) as linear combinations of ∂ui
and ν

∂ (D12 f )(D1 f ) ∂ (D12 f )(D2 f ) ∂ (D12 f )

∇ ∂ ∂u1 = 1+(D1 f )2 +(D2 f )2
· ∂u1
+ 1+(D1 f )2 +(D2 f )2
· ∂u2
+ 1+(D1 f )2 +(D2 f )2
·ν
∂u1

∂ (D1 f )(D1 D2 f ) ∂ (D2 f )(D1 D2 f ) ∂ (D1 D2 f )

∇ ∂ ∂u1 = 1+(D1 f )2 +(D2 f )2
· ∂u1
+ 1+(D1 f )2 +(D2 f )2
· ∂u2
+ 1+(D1 f )2 +(D2 f )2
·ν
∂u2
(294)
∂ (D1 f )(D1 D2 f ) ∂ (D2 f )(D1 D2 f ) ∂ (D1 D2 f )
∇ ∂ ∂u2 = 1+(D1 f )2 +(D2 f )2
· ∂u1
+ 1+(D1 f )2 +(D2 f )2
· ∂u2
+ 1+(D1 f )2 +(D2 f )2
·ν
∂u1
∂ (D1 f )(D22 f ) ∂ (D2 f )(D22 f ) ∂ (D22 f )
∇ ∂ ∂u2 = 1+(D1 f )2 +(D2 f )2
· ∂u1
+ 1+(D1 f )2 +(D2 f )2
· ∂u2
+ 1+(D1 f )2 +(D2 f )2
· ν.
∂u2
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 21, 2009 141

∂
One recognizes the coefficients of the vector fields ∂u i in these expansions as the Christoffel
symbols of the second kind, i.e. for suitable functions cij
X
∇ ∂ ∂u∂ j = Γkij ∂u∂ k + cij ν (295)
∂ui k

This observation suggests to define an induced connection on hypersurfaces by

X
∇ ∂ ∂u∂ j = Γkij ∂u∂ k . (296)
∂ui k

Of course it needs to be checked that this is indeed well-defined, i.e. independent of the co-
ordinate chart. We will postpone this until after the definition of connections on any smooth
manifold, not only on hypersurfaces.

Exercise 4.67 Verify that with this definition the following identity holds:
   
∂ ∂ ∂ ∂ ∂
(g
∂ui j`
) = h∇ ∂ ∂uj
, ∂u`
i + h ∂uj
, ∇ ∂ ∂u`
i (297)
∂ui ∂ui

These formulas also lead to an in intuitive interpretation of the role of the Christoffel symbols
in the geodesic equations: In the case of hypersurfaces M m ⊆ Rm+1 one may consider geodesics
σ : [a, b] 7→ M m as curves in Rm+1 . In general their acceleration σ 00 will not be zero – i.e. they
will not be straight lines (traversed at constant speed) in Rm+1 – but the acceleration σ 00 (t) is
always perpendicular to the tangent plane Tσ(t) M m .

Exercise 4.68 Prove this assertion, i.e. suppose that σ : [a, b] 7→ M m ⊆ Rm+1 is a geodesic
on a hypersurface. Show that the “acceleration” (ı ◦ σ)00 (t), considered as a curve in T Rm+1 , is
always orthogonal to Tσ(t) M m ⊆ Tσ(t) Rm+1 . [[Compare also the MAPLE worksheets.]]

It is time for a formal definition – the following invariant definition is usually associated with
the name Koszul connection. Note, that this definition does not require a metric – but at this
time we shall only consider it in the setting of Riemannian manifolds.

Definition 4.17 Suppose M is a smooth manifold. An (affine) connection (or “covariant deriva-
tive” on M is a map ∇ : Γ∞ (M ) × Γ∞ (M ) 7→ Γ∞ (M ) that satisfies for all X, Y, Z ∈ Γ∞ (M ),
all f ∈ C ∞ (M ) and all c ∈ R
(i) ∇f Y +Z X = f ∇Y X + ∇Z X,
(ii) ∇Z (cX + Y ) = c∇Z X + ∇Z Y ,
(iii) ∇Z (f X) = (Zf )X + f ∇Z X .

Definition 4.18 An affine connection ∇ on a Riemannian manifold M is called a Riemannian

connectionif it in addition to (i)-(iii), it also satisfies, for all X, Y, Z ∈ Γ∞ (M ),
(iv) ∇X Y − ∇Y X = [X, Y ], and
(v) ZhX , Y i = h∇Z X , Y i + hX , ∇Z Y i.

We note on the side that these definitions immediately generalize to covariant derivatives of any
tensor. Moreover, covariant derivatives commute with contractions of tensors (for details on
both compare Spivak II, 2nd ed., p.244).
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 21, 2009 142

Theorem 4.42 A Riemannian manifold admits exactly one Riemannian connection.

Proof. This is an immediate consequence from the following calculation which expresses the
covariant derivative entirely in terms of the Riemannian metric. For vector fields X, Y, Z ∈
Γ∞ (M ) use property (v):
XhY , Zi = h∇X Y , Zi + hY , ∇X Zi
Y hZ , Xi = h∇Y X , Xi + hZ , ∇Y Xi (298)
ZhX , Y i = h∇Z X , Y i + hZ , ∇Z Y i
Suitably adding and subtracting the right and left hand sides and using property (iv) yields
XhY , Zi − Y hZ , Xi − ZhX , Y i
= h[X, Y ] , Zi + h[X, Z] , Y i − h∇Y Z + ∇Z Y , Xi (299)
= h[X, Y ] , Zi + h[X, Z] , Y i + h[Y, Z] , Xi − 2h∇Y Z , Xi
and hence 
1
h∇Y Z , Xi = 2 h[X, Y ] , Zi + h[X, Z] , Y i + h[Y, Z] , Xi
 (300)
− XhY , Zi + Y hZ , Xi + ZhX , Y i
Since the inner product is nondegenerate this completely determines ∇Y Z.

Conversely on a Riemannian manifold (R, h· , ·i) define in any chart (u, U ) with Christoffel
symbols Γkij as defined in equation (216)
m
X
∇∂ ∂
∂uj
= Γkij ∂u∂ k (301)
∂ui k=1

and extend to all vector fields by demanding the properties (i) through (iii). Due to the symmetry
∂
Γkij = Γkji the left hand side of (iv) vanishes in the case of coordinate vector fields X = ∂u i and
∂
Y = ∂uj , as does the right hand side. Finally, (v) follows from
m
∂ ∂ ∂ ∂
Γ`ki · g`j + Γ`kj · gi`
P
h∇ ∂ ∂ui
, ∂uj
i + h ∂u i , ∇∂ ∂uj
i =
∂uk ∂uk `=1
m
[ki, s]g `s · g`j + [kj, s]g `s · gi`
P
=
`,s=1
m
[ki, s]δjs + [kj, s]δis
P
= (302)
s=1

= [ki, j] + [kj, i]
 
∂g ∂g ∂gki ∂gki ∂gji ∂gkj
= 12 ∂ukji + ∂uijk − ∂uj
+ ∂uj
+ ∂uk
− ∂ui
∂
= h ∂
∂uk ∂ui
, ∂
∂uj
i .

Exercise 4.69 Establish the classical transformation formula for the Christoffel symbols of the
second kind: Suppose that (u, U ) and (v, V ) are charts with metrics gij and g̃αβ , and Christoffel
symbols Γkij and Γ̃γαβ respectively. Then:
m m
∂ui ∂uj ∂v γ X ∂ 2 us ∂v γ
Γ̃γαβ =
X
Γkij · + . (303)
∂v α ∂v β ∂uk ∂v α ∂v β ∂us
i,j,k=1 s=1
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 21, 2009 143

m m
∂
(Xui ) ∂u (Y uj ) ∂u∂ j
P P
In a chart (u, U ) calculate for general vector fields X = i and Y =
i=1 j=1

m
 P 
∇X Y = ∇X (Y uj ) ∂u∂ j
j=1
m  m 
(XY uj ) ∂u∂ j + (Xui )(Y uj )∇ ∂
P P
= ∂ ∂uj (304)
j=1 i=1 ∂ui
m  m 
∂
(XY uk ) + Γkij (Xui )(Y uj )
P P
= ∂uk
k=1 i=1

This expression looks remarkably similar to the geodesic equations – but there are some technical
issues that need to be addressed: If σ : [a, b] 7→ M is a smooth curve then σ 0 = σ∗ dt
d
is defined
only along the curve, i.e. it is not immediately clear in which sense one may take X = Y = σ 0 in
above formula. The following exercise suggests that one can indeed make sense of a derivative
of a vector field defined only along a curve.

Exercise 4.70 Suppose p ∈ M and X, Y ∈ Γ∞ (p) are smooth vector fields defined in a neigh-
bourhood of p. Show that (∇X Y )p is completely determined by Xp and the values of Y along any
curve σ : (−ε, ε) 7→ M such that σ 0 (0) = Xp . [[Suppose that σ is a curve as described above,
0
and X 0 , Y 0 are vector fields such that Xp0 = Xp and for all t ∈ (−ε, ε), Yσ(t) = Yσ(t) . Use these
0
to show that (∇X 0 Y )p = (∇X Y )p , and that (∇X Y )p = (∇X Y )p . Analyze the differences and
use the axioms (??), e.g. (iii) with f such Xp f = 0]]

Definition 4.19 Suppose σ : [a, b] → 7 M is a curve. A vector field along the curve σ is a
function X : [a, b] 7→ T M such that π ◦ X = σ.

It is critical that X is defined on [a, b] as opposed to being defined on the subset σ([a, b]) ⊆ M .
In particular, if the curve is not one-to-one, e.g. if σ(t1 ) = σ(t2 ) for some t1 = t2 it is permissible
to have Xt1 6= Xt2 . In general our interest lies in smooth curves, and smooth vector fields along
a curve – defined as smooth maps between manifolds [a, b] ⊆ R and T M .
In analogy to vector fields along a curve one may define vector fields along a hypersurfaces and
similar objects – but we shall not have any immediate need for these. [[Some discussions of
curvature make use of vector fields defined on rectangles X : (−ε, ε) × [0, 1] 7→ M in conjunction
with maps H : (−ε, ε) × [0, 1] 7→ M , i.e., homotopies between curves H(·, 0) and H(·, 1).]]

Exercise 4.71 Suppose that X : [a, b] 7→ T M is a smooth vector field along a smooth curve
σ : [a, b] 7→ M such that for all t ∈ [a, b], σ 0 (t) 6= 0. Show that for any t ∈ [a, b] there exists
ε > 0, an open neighborhood U of σ([t − ε, t + ε]) ⊆ M and a vector field X : U 7→ T M such that
X ◦ σ = X on [t − ε, t + ε]. [[Use the immersion theorem (2.13).]]
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 21, 2009 144

Proposition 4.43 For every smooth curve σ : [a, b] 7→ M there exists exactly one operation,
D
denoted dt , from C ∞ -vector fields along σ to C ∞ -vector fields along σ that satisfies:
(i) for all C ∞ vector fields along σ, D
dt (Y + Z) = D
dt Y + D
dt Z
(ii) for all f ∈ C ∞ [a, b] and all C ∞ vector fields Y along σ, D
dt (f Y ) = f 0Y + f · D
dt Y
(iii) If t0 ∈ [a, b] and Y is a smooth vector field defined on a neighborhood of σ(t0 ) ∈ M then
D

dt t0 Y = ∇σ 0 (t0 ) Y . (305)

D
Definition 4.20 The unique operation dt characterized in proposition (??) is called the covari-
ant derivative along the curve σ.

Proof (of proposition (??)). The easiest way is to verify the proposition is to use a coordinate
chart to obtain an extension Y of Y (which depends on the coordinates).
Thus suppose σ : [a, b] 7→ M is a smooth curve, Y : [a, b] 7→ T M is a smooth vector field along
σ and p = σ(t) for some t ∈ [a, b]. Pick any chart (u, U ) aboutP p. Expand Y in terms of the
coordinate vector fields evaluated along the curve, i.e. Y (t) = m i ∂
j=1 (Yt u ) ∂uj |σ(t) . Note that
each map t 7→ ∂u∂ j |σ(t) is a smooth vector field along (a segment of) the curve σ. Now use the
properties (i), (ii), and (iii) to calculate:
m   
DY  D ∂ 
 
(Yt uj )
P
dt t = dt t ∂uj σ(t)
j=1
m  
(Y uj )0 (t) ∂  D ∂ 
  
+ (Yt uj )
P
= ∂uj σ(t) dt t ∂uj σ(t)
j=1
m  
(Y uj )0 (t) ∂ 
+ (Yt uj )∇σt0 ∂u∂ j
P 
= ∂uj σ(t) (306)
j=1
m  m 
(Y uj )0 (t) ∂ 
(σt0 ui )∇∂ ∂

+ (Yt uj )
P P
= ∂uj σ(t)

 ∂uj
j=1 i=1 ∂ui σ(t)


m  m 
(Yt uk )0 (t) + Γkij (σ(t)) · (σt0 ui ) · (Yt uj ) ∂ 
P P 
= ∂uk σ(t)
k=1 i,j=1

This calculation shows that there is no choice, i.e. given a connection ∇, the conditions (i)
through (iii) uniquely determine the value of DY D
dt . We leave it as an exercise that the dt is
indeed well-defined, i.e. does not depend on the choice of the chart (u, U ).

Exercise 4.72 Suppose p, σ and Y are as in the proof of the proposition and suppose that
DY 

(u, U ) and (v, V ) are charts about p. Verify that dt t is well-defined, i.e. its value does not
depend on the choice of the chart in the preceding calculation.
Moreover, verify that the definition made does in fact have the properties (i), (ii) and (iii).

Note that even when σ 0 (t) = 0 for some t ∈ [a, b] then DY


dt t need not be zero! Indeed the double

sum in the preceding calculation vanishes at such points, but the first term need not. In the
case that σ(t) = q is constant over an interval [t1 , t2 ] then the restriction Y : [t1 , t2 ] 7→ Tq M is
just a curve in the single tangent space at the point q, and the covariant derivative reduces to
the derivative of a curve in Tq M ∼= Rm .
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 21, 2009 145

Summarizing, a connection provides a notion of rate of change of one vector fields in the direction
on another. An important special case is the implied notion of the rate of change of a vector
field along a curve. This provides a means of comparing tangent vectors at different points, and
suggests a notion of a constant vector field, and even provides for a natural notion of geodesics
that does not require a Riemannian structure!
D
Definition 4.21 A vector field Y along a curve σ is called parallel along σ if dt Y ≡ 0.
D
Definition 4.22 Suppose ∇ is a connection with associated covariant derivative dt along smooth
D 0
curves. A smooth curve σ : [a, b] 7→ M is called a geodesic if it satisfies dt σ ≡ 0

Exercise 4.73 Verify that in the case of a Riemannian manifold with Levi-Civita connection
∇ this notion of geodesic agrees with that as a critical point of the energy functional.

Proposition 4.44 Suppose σ : [a, b] 7→ M is a smooth curve and Ya ∈ Tσ(a) M . Then there
exists a unique vector field Y : [a, b] 7→ T M along σ that is parallel along σ.

Proof. Consider (??) with DY dt ≡ 0 as a differential equations for Y . In each chart (u, U ) this
is a linear equation, and hence the initial value problem has a unique, globally defined solution,
i.e. on a maximal time interval (t1 , t2 ) such that the image σ((t1 , t2 )) ⊆ U is a connected subset
of U . Since σ([a, b]) ⊆ M is compact it suffices to consider a finite sequence of charts, and it is
obvious that Y extends to a unique vector field along σ defined on all of [a, b].

Definition 4.23 The vector field Y defined in proposition (??) is called the parallel transport
of Ya ∈ Tσ(a) M along σ. It is convenient to write τt (Ys ) ∈ Tσ(t+s) M for the value of the parallel
transport of Ys ∈ Tσ(s) M along σ, i.e., technically along the restriction of σ to [s, s + t] or
[s + t, s]). (This makes sense as long as s, t, s + t ∈ [a, b].)
S
Note that with this definition the domain of τt is the union s Tσ(s) M taken over all those
s ∈ [a, b] such that s + t ∈ [a, b]. Since in most cases confusion is unlikely, one usually omits
precise descriptions of the domain . . .
Proposition 4.45 Suppose that I ⊆ R is an interval, σ : I 7→ M is a smooth curve with
associated parallel transport τ . Then the following hold for all t, s ∈ I such that t + s ∈ I
(i) On the respective domains τs ◦ τt = τs+t
(ii) For each t the restriction of τt to Tσ(s) M is an isomorphism onto Tσ(s+t) M .
(iii) If ∇ is the Levi-Civita connection on a Riemannian manifold M , then the restriction of
τt to any Tσ(s) M is an isometry.
Proof. The first property is obvious since τ is defined as the flow of a linear differential equation
(on a subset of T M ). Taking s = −t shows that τt ◦ τ−t is the identity (on the appropriate union
of tangent spaces). Hence each τt is a bijection.
Now suppose M is a Riemannian manifold and ∇ the Levi-Civita connection on M . In view of
the polarization identity ha , bi = 41 (ka+bk2 −ka−bk2 ) it suffices to show that parallel transport is
norm-preserving. Thus suppose σ : [a, b] 7→ M a smooth curve and Y a parallel vector field along
σ. Note that DY 2
dt is again a smooth vector field along a curve, and consequently kY k : [a, b] 7→ R
is a continuous function. As a consequence of the subsequent lemma it is clear that kY k2 is
constant on any interval (t1 , t2 ) on which σ 0 is nonzero. By continuity it follows that kY k2 is
also constant on any interval on which σ 0 vanishes only at isolated points.
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 21, 2009 146

If σ 0 (t) = 0 for all t in some interval [t1 , t2 ] then the restriction of σ and Y to this subinterval
are a constant curve σ(t) ≡ q and a curve in Tq M . But as is clear from equation (??), Y must
be constant: since Y is parallel along σ the left hand side vanishes, and since σ 0 = 0 the double
sum is zero. Again invoking continuity, since kY k2 is constant on either kind of subinterval, it
is constant throughout.

Lemma 4.46 Suppose Y and Z are smooth vector fields along a smooth curve σ : (−ε, ε) 7→ M
with p = σ(0) and Xp = σ 0 (0) 6= 0. Then for all t0 sufficiently close to t = 0
d D D
  
dt t0 hY , Zi = h dt t0 Y , Zt0 i + hYt0 , dt t0 Zi. (307)

Proof. Suppose σ, Y, Z and Xp are as in the statement of the lemma. Since σ 0 (0) 6= 0 the
curve is locally an immersion, and thus there exist smooth vector fields Y , Z, ∈ C ∞ (p) such that
Y = Y ◦ σ and Z = Z ◦ σ (in some interval about 0). From the definitions it follows that
d d
 
dt t0 hY , Zi = dt t0 hY ◦ σ , Z ◦ σi

= σ 0 (t0 )hY , Zi
(308)
= h∇Xp Y , Z p i + hY p , ∇Xp Zi
D D
 
= h dt t0
Y , Zt0 i + hYt0 , dt t0
Zi.

The foregoing has defined the parallel transport along a curve in terms of the covariant derivative
along a curve, i.e. in terms of a connection ∇. But one may proceed in the opposite way,
starting with a notion of parallel transport to define a connection. As a step in that direction
the following proposition shows that from the parallel transport as defined above one can recover
the connection:
Proposition 4.47 Suppose σ : (−ε, ε) 7→ M is a smooth curve with σ(0) = p and σ 0 (0) = Xp ,
and Y ∈ Γ∞ (p) is a smooth vector field defined in a neighborhood of p containing σ(−ε, ε)).
Then
∇Xp Y = lim h1 τ−h Yσ(t) − Yp .


(309)
h→0

Proof. Recall from exercise (??) that for a vector field Y ∈ C ∞ (p) the value of (∇X Y )p is
completely determined by the restriction of Y to any curve σ such that σ 0 (0) = Xp . Thus the
proof reduces to a simple calculation when the given vector field Y along the curve σ is expanded
as a linear combination of parallel vector fields. Choose a basis {W10 , W20 , . . . , Wm0 } ⊆ Tp M
and let W1 , W2 , . . . Wm : (−ε, ε) 7→ T M denote their parallel transports along σ. Since each τt is
an isometry, it is clear that {W1t , W2t , . . . , Wmt } ⊆ Tσ(t)
Pm M form a basis for each t. Consequently,
i
there exist functions a : (−ε, ε) 7→ R that Y = i=1 ai Wi . Since each Wi is parallel along
Psuch
m
σ it follows that τ−t Y (t) − Y (0) = i=1 (ai (t) − ai (0))Wi0 and hence
m
X m
X
i
1 d D
ai · Wi = ∇Xp Y.



lim τ−t Yσ(t) − Yp . = dt t=0 a · Wi (0) = (310)
h→0 t dt t=0
i=1 i=1

This construction is very similar to that of the Lie derivative which used the tangent map of
the flow in place of the parallel transport employed here. The similarities suggest that one may
in analogy define covariant derivatives of tensor fields in terms of the parallel transport. This is
indeed possible – for details see e.g. Spivak II, pp.252–253.
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 21, 2009 147

Exercise 4.74 Explore defining the covariant derivative ∇X ω of a differential one-form ω in

terms of the parallel transport along a curve. Use the result to compare X(ω(Y )) with ω(∇X Y )
and (∇X ω)(Y ). For a function f ∈ C ∞ (M ), how does ∇X (df ) compare to d(∇X f ) = d(Xf )?

Exercise 4.75 Make a table comparing and contrasting Lie derivatives and the tangent maps
of the flow on one side, and connections and parallel transport on the other side.
Pay special attention to the (lack of) linearity over functions f ∈ C ∞ (M ), to which vector fields
may be replaced by tangent vectors, to extensibility to tensor fields, to invariant definitions and
descriptions in local coordinates, to the special case of (coordinate) fields in Euclidean spaces,
. . . As the next section shows, the integrability conditions on one side will be contrasted by
conditions for flatness and characterizations of curvature.