Lecture 28: The Spectral Theorem

28 The Spectral Theorem

28.1 Review: Hermitian Spaces
Last time, we discussed Hermitian spaces which were just complex vector spaces with a positive defnite Hermitian
form. Given a linear operator T , we defned the adjoint T ∗ , which had the property that
⟨v, T ∗ w⟩ = ⟨T v, w⟩.
We called a linear operator T normal if T T ∗ = T ∗ T . We then were able to state the Spectral Theorem.

28.2 The Spectral Theorem

The Spectral Theorem demonstrates the special properties of normal and real symmetric matrices.

Theorem 28.1
Given a Hermitian space V , for any normal linear operator T , there exists an orthonormal eigenbasis of V :
{u1 , · · · , un }. In matrix form, for any normal matrix M ∈ GLn (C), there exists a unitary matrix P such
that P −1 M P is diagonal.
The real version states that for a Euclidean vector space V and a symmetric linear operator T , there exists
an orthonormal eigenbasis; equivalently, for any symmetric matrix M ∈ GLn (R), there exists an orthogonal
matrix P such that P −1 M P is diagonal. All eigenvalues of real symmetric matrices are real.

Example 28.2    
3 −1 1
Say we take the symmetric matrix M = . One eigenvector is 2
√1
with eigenvalue λ1 = 2. The
−1 3   1
1
other eigenvector should be orthogonal, so it is √12 with eigenvalue λ2 = 4.
−1
In 2 dimensions, the fact that our change of basis is coming from an orthogonal basis tells us that the
change in coordinates is just a rotation. This rotation makes the operator diagonalizable.

This is called the Spectral Theorem because the eigenvalues are often referred to as the spectrum of a matrix.
Any theorem that talks about diagonalizing operators is often called a spectral theorem.
Now we will state some lemmas in order to prove the Spectral Theorem.

Lemma 28.3 (Lemma 1)

First, for a linear operator T : V −→ V where V is Hermitian, and a subspace W ⊂ V such that T (W ) ⊂ W,
then T ∗ (W ⊥ ) ⊂ W ⊥ .

Proof. Take some u ∈ W ⊥ . For any w ∈ W, then

⟨w, T ∗ u⟩ = ⟨T w, u⟩ = 0,
by the defnition of the adjoint linear operator, so T ∗ u ∈ W ⊥ .

Lemma 28.4 (Lemma 2)

If T v = λv, then T ∗ v = λv. This just means that T and T ∗ have the same eigenvectors, and eigenvalues
related by complex conjugation.

Proof. We can frst consider a special case.

• Special Case: λ = 0. We want to show that T v = 0 implies that T ∗ v = 0.
⟨T ∗ v, T ∗ v⟩ = ⟨T v, T v⟩ = 0
by normality (this was a property shown last time). Since T is positive defnite, this implies that T ∗ v = 0.
This also implies that if a vector v ∈ ker T , then v ∈ ker T ∗ as well.

138
Lecture 28: The Spectral Theorem

• General Case: λ ∈ C. We want to show that T v = λv implies that T ∗ v = λv. In this case, create a
new linear operator S = T − λI. Then, Sv = 0. Also, the adjoint is S ∗ = T ∗ − λI and SS ∗ = S ∗ S, since
T T ∗ = T ∗ T. By the frst case, since v ∈ ker S, we also have that v ∈ ker S ∗ so T ∗ = λv as needed.

Proof of the Spectral Theorem. Now we can prove the Spectral Theorem by induction on the dimension of V.
The general process will be breaking V up into a direct sum and fnding orthonormal eigenbases over each part.
Since the feld is F = C, it is always possible to fnd an eigenvector w ∈ V such that T w = λw. By normalizing
w, assume that ⟨w, w⟩ = 1. Then, let W = Span(w); then V = W ⊕ W ⊥ since all subspaces of Hermitian spaces
are non-degnerate. In addition, since w is an eigenvector, T W ⊂ W.
If we can show that T W ⊥ preserves W ⊥ , then we can induct. By Lemma 2, w is also an eigenvector of T ∗ , so
T ∗ (W ) ⊂ W, and by Lemma 1, (T ∗ )∗ (W ⊥ ) = T (W ⊥ ) ⊂ W ⊥ . Then V = W ⊕ W ⊥ , and T maps each part of the
splitting to itself, so by induction there exists u2 , · · · , un an orthonormal eigenvasis for W ⊥ , and {w, u2 , · · · , un }
is an orthonormal eigenbasis for V.

Student Question. Why is the equivalent of the theorem not true over the real numbers?
Answer. Most of this argument works, except in the very frst step, where we found an eigenvector and eigenvalue.
We cannot guarantee this will happen with normal linear operators over the real numbers. However, as we found
last week, for symmetric (and Hermitian) matrices, the eigenvalues are all real, and in particular it is always
possible to fnd one eigenvector w ∈ V with real eigenvalue to allow the induction to occur.
As an application, suppose
 we have
 a quadratic function f (x, y) = ax + bxy + cy . Let this function be
2 2

a b/2
represented by a matrix ; then
b/2 c
 
x
f (x, y) = (x, y)M = ⟨v, T v⟩
y

for a linearoperatorT given by M. By the Spectral Theorem, there  exists an orthogonal change of coordinates
x′
 
λ1 0 x
T
P MP = , where P is an orthogonal matrix. It takes =P . Then
0 λ2 y y′
     ′
x ′ ′ λ1 x
f (x, y) = (x, y)M = (x , y ) = λ1 (x′ )2 + λ2 (y ′ )2 .
y λ2 y′

Example 28.5  
3 −1
If f (x, y) = 3x − 2xy + 3y , the matrix would be M =
2 2
. Then
−1 3

3x2 − 2xy + 3y 2 = 2(x′ )2 + 4(y ′ )2 ,

for some change of coordinates taking x and y to x′ and y ′ , since we saw earlier that the eigenvalues were 2
and 4.

In n variables, let f (x1 , · · · , xn ) = a11 x21 + · · · + ann x2n + 2aij xi xj be represented by the matrix M =
P
  i<j
a11 · · · aij
 .. ..
 . .


(aij ) =  .
.

..

 aji 
ann
  ′ 
x1 x1
 ..   .. 
Applying the spectral theorem gives P and λi , where P is a change of coordinates taking  .  = P  .  .
xn x′n
In this new coordinate system,

f (x1 , · · · , xn ) = λ1 (x′1 )2 + · · · + λn (x′n )2 .

139
Lecture 28: The Spectral Theorem

In two and three dimensions, this has a geometric interpretation. Consider a curve in R2 with an equation of
the form
ax2 + bxy + cy 2 + dx + ey + f = 0.

There are several options; the frst three are called conics, and the rest are degenerate conics of some form.
• Ellipse. ax2 + by 2 = 1
• Hyperbola. ax2 − by 2 = 1
• Parabola. ax2 − y = 0
• Two crossing lines. (a1 x + b1 y)(a2 x + b2 y) = 0
• Two parallel lines. x2 = a, a > 0
• One line. x2 = 0
• One point. x2 + y 2 = 0
• The empty set ∅. x2 + y 2 = −1

Theorem 28.6
After an isometry, all curves of the form ax2 + bxy + cy 2 + dx + ey + f = 0 look like one of these options.

Proof. The #»T #» #»

 curve be
 rewritten as v A v + B v + f = 0. After an orthogonal change of basis, we can assume
λ1 0
that A = Then λ1 x2 + λ2 y 2 + b1 x + b2 y + f = 0.
0 λ2 .
• Nonzero eigenvalues. If λ1 , λ2 =
̸ 0, then we can complete the square, taking x ⇝ x+ 2λ
b1
1
and y ⇝ y+ 2λ
b2
2
.
Then
λ1 x2 + λ2 y 2 = c.
If c = 0, it is a single point or two intersecting lines, and if c ̸= 0, it is an ellipse or a hyperbola.
• Zero eigenvalues. The other cases.
The details of this aren’t that important, and we glossed over much of it at the end. The main idea to take away
is just the power of the Spectral Theorem to take a complicated quadratic form and simplify into something
more readily analyzed.

One idea where the Spectral Theorem is at play is in multivariable calculus with the second derivative test.
Hidden behind the test is that we have a symmetric matrix and we are trying to fgure out what the signs of
the eigenvalues are, as that will tell us whether our critical point is a minimum, maximum, or saddle point.

140
MIT OpenCourseWare
https://ocw.mit.edu

Resource: Algebra I Student Notes

Fall 2021
Instructor: Davesh Maulik
Notes taken by Jakin Ng, Sanjana Das, and Ethan Yang

For information about citing these materials or our Terms of Use, visit: https://ocw.mit.edu/terms.