Math 108b: Notes on the Spectral Theorem

From section 6.3, we know that every linear operator T on a finite dimensional inner prod-
uct space V has an adjoint. (T ∗ is defined as the unique linear operator on V such that
hT (x), yi = hx, T ∗ (y)i for every x, y ∈ V – see Theroems 6.8 and 6.9.) When V is infinite
dimensional, the adjoint T ∗ may or may not exist.

One useful fact (Theorem 6.10) is that if β is an orthonormal basis for a finite dimen-
sional inner product space V , then [T ∗ ]β = [T ]∗β . That is, the matrix representation of the
operator T ∗ is equal to the complex conjugate of the matrix representation for T .

For a general vector space V , and a linear operator T , we have already asked the ques-
tion “when is there a basis of V consisting only of eigenvectors of T ?” – this is exactly when
T is diagonalizable. Now, for an inner product space V , we know how to check whether vec-
tors are orthogonal, and we know how to define the norms of vectors, so we can ask “when
is there an orthonormal basis of V consisting only of eigenvectors of T ?” Clearly, if there is
such a basis, T is diagonalizable – and moreover, eigenvectors with distinct eigenvalues must
be orthogonal.

Definitions Let V be an inner product space. Let T ∈ L(V ).

(a) T is normal if T ∗ T = T T ∗

(b) T is self-adjoint if T ∗ = T

For the next two definitions, assume V is finite-dimensional: Then,

(c) T is unitary if F = C and kT (x)k = kxk for every x ∈ V

(d) T is orthogonal if F = R and kT (x)k = kxk for every x ∈ V

Notes
1. Self-adjoint ⇒ Normal (Clearly, if T ∗ = T , then T ∗ and T commute!)

2. Unitary or Orthogonal ⇒ Invertible (Show that N (T ) = {0}.)

3. Unitary or Orthogonal ⇒ Normal (This is because, when T is unitary (or orthogonal),

we must have T ∗ T = T T ∗ = I, the identity operator on V – see Theorem 6.18.)
Example A simple example of an orthogonal operator is the rotation operator Tθ on R2 . In
the standard basis,  
cos θ − sin θ
Aθ = [Tθ ]β = .
sin θ cos θ
Notice that A is orthogonal. Also, A∗ = AT 6= A, so A is not self-adjoint. However, it is
easy to check that AA∗ = A∗ A = I, so A is normal. If the matrix A were over the complex
field, it would have a orthonormal basis of eigenvectors by the spectral theorem (see below)!
However, over the reals, we see that the characteristic polynomial of A does not split (unless
sin θ = 0) – therefore, there are no real eigenvalues and therefore no eigenvectors at all for
the operator Tθ on R2 .

Spectral Theorem 1 Let T be a linear operator on a finite dimensional complex in-

ner product space V . Then, T is normal if and only if there exists an orthonormal basis (for
V ) consisting of eigenvectors of T .

Spectral Theorem 2 Let T be a linear operator on a finite dimensional real inner prod-
uct space V . Then, T is self-adjoint if and only if there exists an orthonormal basis (for V )
consisting of eigenvectors of T .

We’ll prove the simple direction first: Assume that there is an orthonormal basis β con-
sisting of eigenvectors of T . Then, we know that [T ]β is diagonal, and also that [T ∗ ]β = [T ]∗β
is diagonal. Since diagonal matrices commute, we have that

[T ∗ T ]β = [T ∗ ]β [T ]β = [T ]β [T ∗ ]β = [T T ∗ ]β .

Hence, T ∗ T = T T ∗ , so T is normal. This holds whether F = C or F = R.

We’ll see in the lemmas below that T normal implies that every eigenvector of T with
eigenvalue λ is also an eigenvector of T with eigenvalue λ̄. Since β is a basis of eigenvectors
of either T or T ∗ , we have that both [T ]β and [T ∗ ]β are diagonal matrices with their eigen-
values on the diagonal. In the case F = R, every eigenvalue of T is real, so λ = λ̄, and we
must have that [T ∗ ]β = [T ]β . Therefore, in the case of a real inner product space, we know
that T is not only normal, but also self-adjoint!

Lemma (pg. 269) Let V be a finite-dimensional inner product space and T ∈ L(V ). If
T has an eigenvector with eigenvalue λ, then T ∗ has an eigenvector with eigenvalue λ̄

See the book for the proof. Note that the above lemma is true for any linear operator T
– but we do not know that the eigenvectors for T and T ∗ are related! The next theorem
states that when T is normal, we know the eigenvectors for T and T ∗ are the same – this is
the fact we used in the proof of the spectral theorem above.
Theorem 6.15 (pg. 371) Let V be an inner product space and T ∈ L(V ) be a normal
operator. Then,
(c) If T (x) = λx (for some x ∈ V and some λ ∈ F ), then T ∗ (x) = λ̄x.
(d) If λ1 and λ2 are distinct eigenvalues of T with eigenvectors x1 and x2 , then hx1 , x2 i = 0.

Sketch of proof: First prove that, since T is normal, kT (x)k = kT ∗ (x)k for every x ∈ V . Then,
since U = T − λI is normal, and U ∗ = T ∗ − λ̄I, we know that U (x) = 0 if and only if
U ∗ (x) = 0. In other words, T (x) = λx if and only if T ∗ (x) = λ̄I.

For (d), Simply compute λ1 hx1 , x2 i = λ2 hx1 , x2 i. Since λ1 6= λ2 , we must have hx1 , x2 i = 0.

We’re almost ready to finish the proof of the first spectral theorem, but first, we’ll need
the following theorem:

Schur’s Theorem Let T be a linear operator on a finite-dimensional inner product space

V . Suppose the characteristic polynomial of T splits. Then, there exists an orthonormal
basis β for V such that the matrix [T ]β is upper triangular.

Proof: Induct on n = dim V . Of course, if n = 1, take any basis β = {x} with kxk = 1. Then,
[T ]β = (a) (where a = T (x)) is upper-triangular. Now, assume the result for n − 1: that is,
suppose that if V is any finite-dimensional inner product space with dimension n − 1 and if
T is any linear operator such that the characteristic polynomial splits, then there exists an
orthonormal basis for V such that [T ]β is upper triangular. We wish to show this result for n.

Fix an n−dimensional inner product space V and a T ∈ L(V ) such that the character-
istic polynomial of T splits. This implies that T has an eigenvalue λ. Therefore, T ∗ has
an eigenvalue λ̄ (by Lemma, pg. 369). Let z be a unit eigenvector of T ∗ , so T ∗ z = λ̄z.
Consider the space W = span({z}). Clearly, W ⊥ is n − 1 dimensional, and TW ⊥ is a linear
operator on W ⊥ , so by the induction hypothesis, there exists a basis γ such that [TW ⊥ ]γ is
upper-triangular. (Notice that we know that TW ⊥ splits since W ⊥ is T −invariant! Check
this fact below:)

Let y ∈ W ⊥ . We want to show that T (y) is in W ⊥ . For every , compute

Finally β = γ ∪ {z} is orthonormal, and [T ]β is upper-triangular: Letting [T z]β = a ∈ F n ,

 
[TW ⊥ ]γ |
[T ]β =  a . 
0 |
Proof of Spectral Theorem 1 Assume T is normal. Since F = C, we know (from the fun-
damental theorem of algebra) that the characteristic polynomial of T splits. By Schur’s
theorem, we can find an orthonormal basis β = {v1 , v2 , ..., vn } such that A = [T ]β is upper-
triangular. Clearly v1 is an eigenvector of T (since T (v1 ) = A11 v1 .) The fact that T is normal
will imply that v2 , ..., vn are also eigenvectors!

Assume v1 , v2 , ..., vk−1 are all eigenvectors Let λ1 , λ2 , ..., λk−1 be the corresponding eigen-
values. Then, since A is upper-triangular, look at the k th column of A and note Amk = 0 if
m > k: This lets us compute

T (vk ) = A1k v1 + A2k v2 + ...Akk vk

From Theorem 6.15, the fact that T is normal implies that T ∗ (vj ) = λ̄j vj . Since the basis
is orthonormal, we have the formula Ajk = hT (vk ), vj i for the coefficients in the equation
above. Computing,

Ajk = hT (vk ), vj i = hvk , T ∗ (vj )i = vk , λ̄j vj = λj hvk , vj i = 0.

Therefore, we have simply that T (vk ) = Akk vk , so vk is an eigenvector for T .

By induction, we have shown that v1 , v2 , ..., vn are all eigenvectors of T , so β is an or-

thonormal basis consisting of only eigenvectors of T , and the spectral theorem is proven. 

Before we can prove the second version of the spectral theorem, for F = R, we need the
following lemma:

Lemma (pg. 373) Let T be a self-adjoint operator on a finite-dimensional inner product

space V . Then the following two facts hold (whether we have F = R or F = C)

(a) Every eigenvalue of T is real.

(b) The characteristic polynomial of T splits.

Proof of (a): From Theorem 6.15, if x is an eigenvalue of T , we have both T (x) = λx

for some λ ∈ F and T ∗ (x) = λ̄x. However, since T is self-adjoint, T (x) = T ∗ (x). Therefore,
λx = λ̄x. Since x 6= 0, we have that λ = λ̄; i.e., λ is real.

Proof of (b): When F = C, we already know the characteristic polynomial splits. Let β
be any orthonormal basis for V ; we know that A = [T ]β satisfies A∗ = A. Define, for every
x ∈ Cn ,
S(x) = Ax.
S has the same characteristic polynomial as T , but since S is a linear operator on a com-
plex inner product space, we know the characteristic polynomial of S splits. Moreover,
since S is self-adjoint, each eigenvalue λ1 , ..., λn is real, and the polynomial pT (t) = pS (t) =
(λ1 − t)(λ2 − t)...(λn − t) splits over the field F = R. 
Proof of Spectral Theorem 2: Assume F = R and T is a self-adjoint linear operator on
V . Then, the characteristic polynomial of T splits, and Schur’s theorem implies that there
exists an orthonormal basis β for V such that A = [T ]β is upper-triangular. The same proof
as for the first spectral theorem now works since T is normal, but it is easier to note that
since T ∗ = T , we know that both A and AT = A∗ = A are upper triangular. Therefore, A
is diagonal.

Finally, we will show the next theorem, which includes the fact that unitary (and orthog-
onal) operators must be normal!

Theorem 6.18 Let V be a finite dimensional inner product space and T ∈ L(V ). The
following are equivalent:

(a) T ∗ T = T T ∗ = I

(b) hT (x), T (y)i = hx, yi for every x, y ∈ V

(e) kT (x)k = kxk for every x ∈ V

Proof:

(a) ⇒ (b) Assume (a) T ∗ T = T T ∗ = I.

Given x, y ∈ V , since T ∗ T (x) = I(x) = x, we have hx, yi = hT ∗ T (x), yi = hT (x), T (y)i.

(b) ⇒ (e) Assume (b) hx, yi = hT (x), T (y)i for every x, y ∈ V

Let β = {v1 , v2 , ..., vn } be an orthnormal basis for V . Then T (β) = {T (v1 ), T (v2 ), ..., T (vn )}
is an orthonormal basis for V (since, from (b), hT (vi ), T (vj )i = hvi , vj i = δij ).
Pn Pn
Given any x ∈ V , write x = i=1 ai vi , we can easily compute kxk2 = i=1 |ai |2 .

Pn Pn
Since T is linear, T (x) = i=1 ai T (vi ); the same computation shows kT (x)k2 = i=1 |ai |2
(e) ⇒ (a) Assume (e) kT (x)k2 = kxk2 for every x ∈ V .

For all x ∈ V ,
hx, xi = kxk2 = kT (x)k2 = hT (x), T (x)i = hx, T ∗ T (x)i
This implies, for all x ∈ V , that

(?) hx, (I − T ∗ T )(x)i = 0.

Let U = I − T ∗ T . Check that U is self-adjoint: U ∗ = I ∗ − (T ∗ T )∗ = I − T ∗ T . Then, by the

spectral theorem there exists an orthonormal basis β consisting of eigenvectors of U .

For all x ∈ β, U (x) = λx. By (?), 0 = hx, U (x)i = λ̄ hx, xi. Since x 6= 0, we must
have λ = 0. This implies that [U ]β is the zero matrix! Hence, U = I − T ∗ T = To (the zero
transformation on V ), so T ∗ T = I. Since V is finite dimensional, this proves that T ∗ must
be the inverse of T and therefore T T ∗ = I.