LECTURE 5: THE RIEMANNIAN CONNECTION

1. Linear connections on tensor fields

Now let M be a smooth manifold, and ∇ a linear connection on (vector fields
of) M . We will extend ∇ to a linear connection on all tensor fields. This is very
easy for (0, 0)-tensor fields (= functions), since we already have a nice one,
∇ : Γ(T M ) × C ∞ (M ) → C ∞ (M ), (X, f ) 7→ ∇X f := Xf = df (X),
which obviously satisfies the two conditions in the definition of linear connections.
According to the remark in the previous lecture, a linear connection on (r, s)-
tensor fields is a bi-linear map
∇ : Γ(T M ) × Γ(⊗r,s T M ) → Γ(⊗r,s T M ), (X, T ) 7→ ∇X T,
that satisfies
(1) ∇f X T = f ∇X T ,
(2) ∇X (f T ) = f ∇X T + (Xf )T .
Again there are too much choices of linear connections in general, and most of
them are not interesting. However, if we impose extra assumptions that all these
connections ∇ on (r, s)-tensor fields are related in the following natural way:
(3) ∇ coincide with the given connections on Γ(T M ) and C ∞ (M ),
(4) ∇X (T1 ⊗ T2 ) = (∇X T1 ) ⊗ T2 + T1 ⊗ ∇X T2 ,
(5) C(∇X T ) = ∇X C(T ), where
C : Γ(⊗r,s T M ) → Γ(⊗r−1,s−1 T M )
is the contraction map that pairs the first vector with the first covector.
Then one can prove
Theorem 1.1. Given any linear connection ∇ (on vector fields), there is a unique
linear connection on all tensor fields that satisfies conditions (1)-(5) above.

Sketch of proof. First we use the conditions (3)-(5) to derive the formula of ∇ on
1-forms. Let ω ∈ Ω1 (M ) = Γ(T ∗ M ) be any 1-form, then by (3) and (5) we must
have
X(ω(Y )) = ∇X (ω(Y )) = ∇X (C(ω ⊗ Y )) = C(∇X (ω ⊗ Y )).
Now use (4), we get
C(∇X (ω ⊗ Y )) = C(∇X ω ⊗ Y + ω ⊗ ∇X Y ) = (∇X ω)(Y ) + ω(∇X Y ).
1
2 LECTURE 5: THE RIEMANNIAN CONNECTION

So we conclude
(1.1) (∇X ω)(Y ) = X(ω(Y )) − ω(∇X Y ),

Second we can use (4) iteratively to show that for any (r, s)-tensor field T ,
(∇X T )(ω1 , · · · , ωr , Y1 , · · · , Ys ) =X(T (ω1 , · · · , ωr , Y1 , · · · , Ys ))
X
− T (ω1 , · · · , ∇X ωi , · · · , ωr , Y1 , · · · , Ys )
(1.2) i
X
− T (ω1 , · · · , ωr , Y1 , · · · , ∇X Yj , · · · , Ys ).
j

This can be done by induction. In particular, this shows the uniqueness.

To prove the existence, one just need to check that the connections defined by
equations (1.1) and (1.2) satisfies all conditions (1)-(5). 

In particular, since a Riemannian metric g is a (0, 2)-tensor field on M , we get

(∇X g)(Y, Z) = XhY, Zi − h∇X Y, Zi − hY, ∇X Zi.
Definition 1.2. A tensor field T is called parallel if ∇X T = 0 for all X ∈ Γ(T M ).
Example. One can view the identity map
I = Id : Γ(T M ) → Γ(T M )
as a (1, 1)-tensor via
I(ω, Y ) = ω(Y ).
Then it is parallel since according to (1.1),
(∇X I)(ω, Y ) = X(ω(Y )) − (∇X ω)(Y ) − ω(∇X Y ) = 0.

2. The Levi-Civita connection

Now let (M, g) be a Riemannian manifold, and ∇ a linear connection on M .
Definition 2.1. We say ∇ is compatible with g if the Riemannian metric g is
parallel. In other words, ∇ is compatible with g if for all X, Y, Z ∈ Γ(T M ),
X(hY, Zi) = h∇X Y, Zi + hY, ∇X Zi.
Definition 2.2. A connection ∇ is on (M, g) is called a Levi-Civita connection (also
called a Riemannian connection) if it is torsion-free and is compatible with g.

For example, if we let M = Rn with the canonical Riemannian metric g0 , then

the canonical linear connection (i.e. the one with all Christoffel symbols Γlij = 0) is
a Levi-Civita connection. An nontrivial example is
 LECTURE 5: THE RIEMANNIAN CONNECTION 3

Example. Let M = S n equipped with the round metric g = ground , i.e. the induced
metric from the canonical metric in Rn+1 . We denote by ∇ the canonical (Levi-
Civita) connection in Rn+1 . For any X, Y ∈ Γ(T ∗ S n ), one can extend X, Y to
smooth vector fields X̄ and Ȳ on Rn+1 , at least near S n . By localities we proved
last time, the vector
∇X̄ Ȳ
n
at any point p ∈ S depends only on the vector X̄(p) = X(p) and the vectors
X̄(q) = X(q) for q ∈ S n . In other words, it is indepedent of the choice of the
extension we chose. So for simplicy we will write ∇X Y instead of ∇X̄ Ȳ for points
on S n . It is a vector that is not necessary tangent to S n . We define ∇X Y be the
“orthogonal projection” of ∇X Y onto the tangent space of S n , i.e.
∇X Y := ∇X Y − h∇X Y, ~ni~n,
where ~n (=~x) is the unit out normal vector on S n . I claim that it is a (=the)
Levi-Civita connection of (M, g).
To prove this, first notice that ∇ is bilinear, and ∇f X Y = f ∇X Y . Also
∇X (f Y ) = ∇X (f Y ) − h∇X (f Y ), ~ni~n
= f ∇X (Y ) − f h∇X (Y ), ~ni~n + (Xf )Y − h(Xf )Y, ~ni~n
= (Xf )Y + f ∇X Y,
where we used the fact that Y is a tangent vector field of S n and thus h(Xf )Y, ni = 0.
So ∇ is a linear connection on S n .
This connection is torsion free because (we use [X, Y ] ⊥ ~n here!)
∇X Y − ∇Y X = ∇X Y − ∇Y X − h∇X Y − ∇Y X, ~ni~n
= [X, Y ] − h[X, Y ], ~ni~n
= [X, Y ].
Finally this connection is compatible with the metric g, since
XhY, Zi = h∇X Y, Zi + hY, ∇X Zi = h∇X Y, Zi + hY, ∇X Zi,
where we used the fact that the difference between ∇X Y and ∇X Y is a vector in
the normal direction, and thus is perpendicular to Z.
Remark. By the same argument, one can prove that if (X, g) is a Riemannian mani-
fold, with a Levi-Civita connection ∇M , and if (N, ι∗ g) is a Riemannian submanifold
of (M, g), then the “orthogonal projection” of ∇M onto T N ,
∇N M T
X Y := (∇X̄ Ȳ ) ,

defines a Levi-Civita connection on (N, ι∗ g).

Remark. Since any Riemannian manifold can be embedded to the standard Euclidian
space isometrically, the arguments in the previous remark immediately implies that
on any Riemannian manifold, there exists Levi-Civita connection!
4 LECTURE 5: THE RIEMANNIAN CONNECTION

Our main theorem is to prove

Theorem 2.3 (The fundamental theorem of Riemannian geometry). On any Rie-
mannian manifold (M, g), there is a unique Levi-Civita connection.
Remark. Roughly speaking, smooth manifolds are the underlying space, and in ge-
ometry we are interested in various extra geometric structures defined on manifold.
Given any smooth manifold, one has
• infinitely many different distance function (all compatible with the underly-
ing topology),
• infinitely many different measures,
• infinitely many differnt Riemannian structures,
• infinitely many linear connections etc.
However, for the first five lectures in this course, we proved that if you fix a Rie-
mannian metric, then you will get
• a canonical distance function (the Riemannian distance)
• a canonical measures (the Riemannian measure),
• a canonical linear connections (the Levi-Civita connection).

First proof (coordinate free). Assume the Levi-Civita connection exists. Then
h∇X Y, Zi =X(hY, Zi) − hY, ∇X Zi
=X(hY, Zi) − hY, ∇Z Xi − hY, [X, Z]i
=X(hY, Zi) − Z(hY, Xi) + h∇Z Y, Xi − hY, [X, Z]i
=X(hY, Zi) − Z(hY, Xi) + h∇Y Z, Xi + h[Z, Y ], Xi − hY, [X, Z]i
=X(hY, Zi) − Z(hY, Xi) + Y (hZ, Xi) − hZ, ∇Y Xi
+ h[Z, Y ], Xi − hY, [X, Z]i
=X(hY, Zi) − Z(hY, Xi) + Y (hZ, Xi) − hZ, ∇X Y i − hZ, [Y, X]i
+ h[Z, Y ], Xi − hY, [X, Z]i
It follows that ∇X Y must be the vector satisfying
2h∇X Y, Zi =X(hY, Zi) − Z(hY, Xi) + Y (hZ, Xi)
(2.1)
− hZ, [Y, X]i + h[Z, Y ], Xi − hY, [X, Z]i.
The right hand side is determined by the metric. So the uniqueness is proved. [The
last formula is called the Koszul formula.]
To prove the existence, one only need to check that the ∇X Y defined by the
above formula satisfies all conditions of Levi-Civita connections. 

Second proof (local coordinate). Again we first prove uniqueness. Let ∇ be a Levi-
Civita connection. Pick a coordinate neighborhood and let Γkij be the functions
 LECTURE 5: THE RIEMANNIAN CONNECTION 5

∇∂i ∂j = Γkij ∂k . Then it is enough to prove that the Γkij ’s are determined by the
metric g. First we note that by torsion free property,

Γkij = Γkji .

Second we calculate
∂i gjk = ∂i (g(∂j , ∂k )) = g(∇∂i ∂j , ∂k ) + g(∂j , ∇∂i ∂k )
= g(Γlij ∂l , ∂k ) + g(∂j , Γlik ∂l ) = Γlij glk + Γlik gjl .

Similarly one can prove

∂j gki = Γljk gli + Γlji gkl and ∂k gij = Γlki glj + Γlkj gil .

So we get
∂j gki + ∂i gjk − ∂k gij = 2glk Γlij .
It follows

(2.2) 2Γlij = g lk (∂j gki + ∂i gjk − ∂k gij ).

This proves the uniqueness.

For the existence, we can define locally (for X = X i ∂i and Y = Y j ∂j )

∇X Y = X i ∂i Y j ∂j + X i Y j Γlij ∂l ,

where Γlij is the function given by (2.2). By tedious compotations one can check
that this give a Levi-Civita connection whose Christophel symbols are exactly the
Γlij ’s. [In particular we immediately see that the connection is torsion free.] 

The local expression (2.2) for Γlij in terms of gij ’s is very useful in computations.
For example, we have

Proposition 2.4. Let ∇ be the Levi-Civita connection on (M, g). Then

1 √
Γjji = √ ∂i G.
G
[Note: for the left hand side we are using the Einstein summation convection!]

Proof. We first use the formula (2.2) to get

2Γjji = g jk (∂i gkj + ∂j gik − ∂k gij ) = g jk ∂i gkj = Tr((g rs )∂i (gkj )).

We need
6 LECTURE 5: THE RIEMANNIAN CONNECTION

Lemma 2.5. Let A = A(t) be a family of nonsingular matrices

that depends smoothly on t, then
d d
(2.3) Tr(A−1 A) = ln det A.
dt dt
Proof. By the standard perturbation trick, it is enough to prove the
theorem for diagonalizable matrices. We write A = P −1 DP , where
D is the diagonal matrix whose entries are the eigenvalues of A.
Then the left hand side becomes [We use A0 to represent dA dt
]
Tr(A−1 A0 ) = Tr P −1 D−1 P [(P −1 )0 DP + P −1 D0 P + P −1 DP 0 ]

= Tr P (P −1 )0 + D−1 D0 + P −1 P 0

= Tr D−1 D0 ,

where in the last step we used the fact

P (P −1 )0 + P −1 P 0 = (P −1 P )0 = 0.
For the right hand side we have
(ln det A)0 = (ln det D)0 .
So the problem is converted to prove (2.3) for the diagonal ma-
trix D (whose entries depends on t), which is trivially true after
straightforward computations. 
Applying this to A = (gij ), we get
√ 1 √
2Γjji = ∂i ln det(gkj ) = 2∂i ln G = 2 √ ∂i G,
G
and the conclusion follows. 

3. The Hessian
Now let (M, g) be a Riemannian manifold, and ∇ the Levi-Civita connection on
M . For any vector field X ∈ Γ(T M ), we can define a linear map
∇X : Γ(T M ) → Γ(T M ), Y 7→ ∇Y X.
According to locality 2, at each point p, ∇X is just a map from Tp M to Tp M . In
particular, it makes sense to talk about the trace of ∇X at each p, which gives us
a function on M .
Lemma 3.1. div(X) = Tr(∇X).

Proof. Both sides are functions on M , so one only need to prove it at one point p.
We pick a local coordinate system near p. Then
∇∂i X = (∇∂i X j )∂j + X j ∇∂i ∂j = ∂i (X j )∂j + X j Γkij ∂k
 LECTURE 5: THE RIEMANNIAN CONNECTION 7

implies
1 √ 1 √
Tr(∇X) = ∂i (X i ) + X i Γjji = ∂i (X i ) + X i √ ∂i G = √ ∂i (X i G) = div(X).
G G

Recall that ∆f = −div∇f . So by the proposition above, we get another formula
for the Laplace-Beltrami operator:
∆f = −Tr(∇2 f ).
Definition 3.2. For any f ∈ C ∞ (M ), we will call ∇2 f = ∇(∇f ) the Hessian of f .
Remark. The ∇ in ∇f here represents the gradient, not the connection. The con-
nection on functions is ∇X f = Xf = df (X). In other words, the connection ∇ is
∇f = df .
So ∇2 f is a map
∇2 f : Γ(T M ) → Γ(T M ),
which can be identified with a (1, 1)-tensor
∇2 f (X, ω) = ω(∇X ∇f ).
Using the metric g, one can also convert this (1, 1)-tensor ∇2 f into a (0, 2)-tensor
∇2 f (X, Y ) = ∇2 f (X, [Y ) = ([Y )(∇X ∇f ) = h∇X ∇f, Y i.
Proposition 3.3. ∇2 f is a symmetric (0, 2)-tensor.
Proof. By metric compatibility,
∇2 f (X, Y ) = h∇X ∇f, Y i = ∇X (h∇f, Y i) − h∇f, ∇X Y i = X(Y f ) − (∇X Y )f.
On the other hand, by the torsion-free property,
X(Y f ) − (∇X Y )f = Y (Xf ) − (∇Y X)f.
So we conclude
∇2 f (X, Y ) = ∇2 f (Y, X).

Remark. One could define the Hessian of f with respect to any linear connection
(and without using the metric structure), by setting
∇2 f (X, Y ) := X(Y f ) − (∇X Y )f.
Then the proof above shows that the Hessian is symmetric if and only if the con-
nection is torsion free. This gives another explaination of the torsion tensor.