A short proof of Perrons theorem.

Hannah Cairns, April 25, 2014.

A matrix A or a vector is said to be positive if every component

is a positive real number. A B means that every component of A
is greater than or equal to the corresponding component of B. In the
same way, A > B means that every component of A is greater than
the corresponding component of B.
Definition. The spectral radius (A) of a square matrix A is the
maximum of the absolute values of the eigenvalues of A.
Theorem. (Perrons Theorem.) Let A be a positive square matrix.
Then:
a) (A) is an eigenvalue, and it has a positive eigenvector.
b) (A) is the only eigenvalue on the disc || = (A).
c) (A) has geometric multiplicity 1.
d) (A) has algebraic multiplicity 1.
Preliminaries. The proof of the Perron-Frobenius theorem will
rely on the following simple positivity trick.
Trick. Let A be a positive square matrix. If and are two
nonequal column vectors with , then A > A. There is some
positive > 0 with A > (1 + )A.
P
Proof. The vector A( ) = Aij (j j ), so
X X
(A[ ])i min(Aij )(j j ) = min(Aij ) (j j ) > 0.
ij ij
j

So (A)i (A)i > 0, and by definition A > A. This proves the

first statement. If we change the vector A() by a small amount then
it will still be positive, so there is some > 0 with A( ) A > 0,
or A > (1 + )A. That proves the second statement. 
We also need the idea of the spectral radius of a matrix, together
with a theorem from linear algebra.
Theorem. (Gelfands formula) The spectral radius of a matrix A
can be written in terms of the norms of its powers:
(A) = lim(||An ||)1/n .

We can now begin the proof of the Perron-Frobenius theorem.

1
2

Proof.

(a) A has as an eigenvalue, and has a positive eigenvector.

There is some eigenvalue with || = (A). Let be an eigenvector.

Let be the vector with j = |j |. Then
X X
(A)i = Aij |j |  Aij j | = |i | = (A)i ,

so A (A).
If they are not equal, then by the positivity trick, A2 > (A)A,
and there is some positive > 0 with A2 (1 + )(A)A.
The matrix A is nonnegative, and so are all its powers An , so multi-
plying both sides of the equation by An preserves the inequality:
An+2 (1 + )(A)An+1 .
This is true for every n, so
An+1 (1 + )(A)An
((1 + )(A))2 An1
..
.
((1 + )(A))n A.
Gelfands formula tells us that
(A) = lim ||An ||1/n (1 + )(A)
n

which is a contradiction, so our assumption must have been mistaken.

This means that the two vectors were equal: A = (A).
So || is also an eigenvector with eigenvalue (A). It is not only
nonnegative but positive, because A|| = (A)|| is positive.

(b) The only eigenvalue on the circle || = (A) is (A).

Suppose 6= (A) is some other eigenvalue on the circle || = (A).

We will reach a contradiction. Repeat the reasoning from part (a)
again: if = ||, then we showed that A = (A), or
 
X X 
Aij |j | =  Aij j 

for every 1 j n.
 3

Lemma. Let A be a positive n n matrix. Let be a vector in Cn .

Choose an index j. If
X X 
 
Aij |j | = 
 Aij j ,

then there is some c C with c 6= 0 so that the product c is a

nonnegative vector.
Proof. If = 0, then
P c = 1 works. P Suppose
P is not zero. Square
both sides and write | Aij j |2 = ( Aij j )( Ak` ` ).
We get the equality
X X
Aij Ak` |j ||` | = Aij Ak` j ` .
Subtract the right-hand side from the left:
X

Aij Ak` |j ||` | j ` = 0
All of the coefficients Aij Ak` are positive real numbers, and the term
in parentheses has nonnegative real part, and positive real part un-
less j ` 0 for every pair j, `.
Then we can take c = ` for some index ` where ` 6= 0. 
The conditions of the lemma hold, so c > 0 for some complex
number c 6= 0. But this means that
(c) = c() = c(A) = A(c) 0
is positive, and c > 0, so must be nonnegative.
But (A) is the only nonnegative number on the circle || = (A).
(c) (A) has geometric multiplicity 1.
Suppose is the positive eigenvector above and 0 is a linearly in-
dependent eigenvector of the eigenvalue (A).
We can assume that 0 is real; otherwise we take real and imaginary
parts, and the parts are still eigenvectors, because A and (A) are real.
One of them must be linearly independent of .
Let c > 0 be chosen so that c 0 is nonnegative and at least
one entry is zero. It is not the zero vector, because , 0 are linearly
independent. But
A[ c 0 ]
c 0 = > 0.
(A)
We chose c so that at least one entry was zero, so this is a contradic-
tion. Therefore, there cannot be two linearly independent eigenvectors,
so (A) has geometric multiplicity 1.
4

(d) (A) has algebraic multiplicity 1.

As in previous parts, let be a right eigenvector of (A). We now

know that there is only one eigenvector and that it is positive.
Let be a positive left eigenvector of (A). We can get such an
eigenvector by applying (a) to AT .
This pair of eigenvectors allows us to decompose Rn as a direct sum.
Lemma. The space = {x Rn : x = 0} is invariant under A.
Proof. If x = 0, then Ax = (A)x = x = 0. 

This space has dimension n 1, and is a vector that is not in ,
because X
= j j > 0.
Therefore, Rn is the direct sum of span{} and .
Let 2 , . . . , n be a basis for the space = {x Rn : x = 0}.
Let X be the matrix
 
X = 2 n .

Then XAX 1 leaves invariant the spaces X 1 span{} = span{e1 }

and X 1 = span{e2 , . . . , en }, so this change of basis turns A into a
block diagonal matrix:
" #
1
(A) 0
X AX =
0 Y
for some matrix Y .
Suppose has algebraic multiplicity more than one. Then it must
also be an eigenvalue of Y . But then Y must have an eigenvector
for (A) too, which makes two eigenvectors, which contradicts (b).
Therefore, (A) has algebraic multiplicity one. This proves part (d).

Finally, we can easily extend the theorem to the case where A is

nonnegative and has a positive power Am .

Theorem. (Perron-Frobenius theorem.) The statements (a), (b),

(c), (d) are also true for nonnegative matrices A so that some power
Am is positive.

Proof. Let 1 , . . . , n be the eigenvalues of A, counted with alge-

braic multiplicity. Suppose 1 has the largest absolute value.
Then the eigenvalues of Am are m m m
1 , . . . , n , where 1 is the largest in
absolute value. Perrons Theorem tells us that m 1 is positive, and it has
 5

a positive eigenvector . It also tells us that all the other eigenvalues

of Am are smaller.
But if |1 |m > |j |m for j 2, then |1 | > |j | for j 2. There-
fore, 1 is the largest eigenvalue in absolute value. It has algebraic
multiplicity one. Its eigenvector must be , because no other eigenvec-
tor of A has the right eigenvalue.
So we know (b), (c), (d), and (a) except for the positivity of 1 . This
follows from the fact that is positive, and 1 = A 0.
It turns out that this is equivalent to a condition on the directed
graph G(A) on {1, . . . , n} with an edge from i to j if Aij > 0.
There is some m with Am > 0 if and only if
a) every vertex in G(A) can be reached from every other vertex
b) the gcd of the lengths of the loops in G(A) is one.
This can be proved by the methods we saw in class. We sketch a
proof.
Sketch. We observe that there is a path of length m from i to j
in G(A) if and only if (Am )ij > 0. The first condition guarantees that
for every i, j there are infinitely many m with (Am )ij > 0, and the
second condition tells us that there exist loops at any vertex of any
sufficiently long length (which we can obtain by chaining the shorter
loops together).
Suppose that we can get from any vertex i to any other vertex j in
time at most s, and that there are loops at every vertex of every length
at least r. Let m = s + r. Then for every pair i, j, we can go from i to
j in time t s and then pick a loop at j which has length (m t) r.
This is a path of length m from i to j, so (Am )ij > 0.
References:
Roger A. Horn and Charles R. Johnson. Matrix analysis. Cambridge, 1985. (See
chapter 8.)

Laurent Saloff-Coste. Lectures on finite Markov chains. In Lectures on Proba-

bility Theory and Statistics, volume 1665 of Lecture Notes in Mathematics, pages
301413. http://www.math.cornell.edu/~lsc/math778-only/stf.pdf