The Perron-Frobenius Theorem and its

Application to Population Dynamics

by

Jerika Baldin

A project submitted in partial fulfillment

of the requirements for the degree of
Mathematics (April 2017)

Department of Mathematical Sciences

Lakehead University
2017
 ACKNOWLEDGEMENTS

First and foremost, I wish to thank my supervisor, Dr. Christopher Chlebovec,

for his guidance and support throughout the research and write up phases of my
thesis. His suggestions and assistance were greatly appreciated. I wish to also extend
my gratitude to the faculty, staff, and students with whom I’ve had the pleasure of
coming into contact with at Lakehead University.

2
 TABLE OF CONTENTS

ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . 2

CHAPTER

I. Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Positive and Non-negative Matrices . . . . . . . . . . . . . . . 1

1.2 Jordan Normal Form . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Graphs and Matrices . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

II. The Perron-Frobenius Theorem for Positive Matrices . . . . 10

2.1 The Perron-Frobenius Theorem Stated . . . . . . . . . . . . . 10

2.2 Positive Eigenpair . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Index of ρ(A) . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 Multiplicities of ρ(A) . . . . . . . . . . . . . . . . . . . . . . 17
2.5 Collaltz-Wielandt Formula . . . . . . . . . . . . . . . . . . . 18

III. Perron Frobenius Theorem for Non-negative Matrices . . . . 20

3.1 Non-negative Eigenpair . . . . . . . . . . . . . . . . . . . . . 20

3.2 Reducibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Primitive Matrices . . . . . . . . . . . . . . . . . . . . . . . . 26

IV. The Leslie Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4.1 The Rabbit Problem . . . . . . . . . . . . . . . . . . . . . . . 30

4.2 Leslie Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3
 CHAPTER I

Preliminaries

1.1 Positive and Non-negative Matrices

Let A = (aij ) and B = (bij ) be matrices of size n × n. A matrix A is said to

positive if every aij > 0 and is denoted by A > 0. Also, A is said to be non-negative
if every aij ≥ 0 and is denoted by A ≥ 0. A matrix A > B if aij > bij for all i, j.
Listed below are some useful properties about matrices.

Proposition I.1. [3, p.663] Let P and N be matrices of size n × n with x, u, v, and
z vectors.

(i) P > 0, x ≥ 0, x 6= 0 =⇒ Px > 0

(ii) N ≥ 0, u ≥ v ≥ 0 =⇒ Nu ≥ Nv

(iii) N ≥ 0, z > 0, Nz = 0 =⇒ N = 0

(iv) N > 0, u > v > 0 =⇒ Nu > Nv

Proof.

(i) Let P = (aij ) and let Pi = (ai1 · · · aij · · · ain ) be the i th row of P. Since
x ≥ 0, x 6= 0 there is some j with xj > 0. Then the i th entry of Px is Pi x =
ai1 x1 + · · · + aij xj + · · · + ain xn . Also, since aij xj > 0 we have that (Px)i =

1
 ai1 x1 + · · · + aij xj + · · · + ain xn ≥ aij xj > 0 for each i = 1, 2, . . . n.

(ii) Let N = (bij ). Let u = (ui )i and v = (vi )i . Since u ≥ v we have ui ≥ vi for each
i = 1, 2, . . . , n. Then, bij ui ≥ bij vi which implies that bi1 u1 + bi2 u2 + · · · + bin un ≥
bi1 v1 + bi2 v2 + · · · + bin vn . It follows that (Nu)i ≥ (Nv)i for each i and therefore,
Nu ≥ Nv.

(iii) Let N = (bij ) and z = (zi )i . Since Nz = 0 we have that (Nz)i = 0 for each i . We
know (Nz)i = bi1 z1 + bi2 z2 + · · · + bin zn = 0. Since z > 0 and N ≥ 0 we know that
bij = 0 for each j = 1, 2, . . . , n. Therefore, since bij = 0 we have that N = 0.

(iv) Let N = (bij ). Let u = (ui )i and v = (vi )i . If N > 0 then bij > 0. Since
ui > vi we have that bi1 u1 + bi2 u2 + · · · + bin un > bi1 v1 + bi2 v2 + · · · + bin vn . Therefore,
Nu > Nv.

1.2 Jordan Normal Form

Theorem I.2. Jordan Normal Form [3, p.590]

A Jordan matrix or a matrix in Jordan Normal Form is a block matrix that has Jordan
blocks down its diagonal and is zero everyone else. Every matrix A in Mn (C) with
distinct eigenvalues σ(A) = {λ1 , λ2 , . . . , λn } is similar to a matrix in Jordan Normal
Form. That is, there exists an invertible matrix P such that

 
J(λ1 ) 0 · · · 0 

 0 ... 
0 
P−1 AP = J =  . .
 
 .. .. .. 
 . .  
 
0 0 · · · J(λn )

2
For each eigenvalue λj ∈ σ(A) there exists one Jordan segment J(λj ) that is made
up of tj Jordan blocks where tj = dimN (A − λj I). Indeed,

 
J1 (λj ) 
 
 J2 (λj ) 
J(λj ) =  . (1.1)
 
..

 . 

 
Jtj (λj )

Then, J∗ (λj ) represents an arbitrary block of J(λj ). That is,

 
λj 1 
 .. .. 
 . . 
J∗ (λj ) =  . (1.2)
 
..

 . 1 
 
λj

Also, Ak = PJk P−1 and

 k  
k k k
k

k−1
k−2
k−m+1
λ 1  λ 1
λ 2
λ ··· m−1
λ 
 .
.. ... 
 
k k

k−1 .. 
k
  λ 1
λ . 
J∗ (λ) =   = .
   
. .. ..

 . . 1  . . k

k−1
λ 
   1 
   
λ λk
m×m

(1.3)

Let A be a square matrix. The algorithm below is used to find the Jordan Normal
Form of matrix A.

(1) Find the distinct eigenvalues {λ1 , . . . , λn } of A such that λ ∈ σ(A). For each
eigenvalue λj we have the segment J(λj ) which is made up of tj = dimN ul(A −
λj I) Jordan blocks, J∗ (λj ). We obtain J(λj ) as seen in (1.1) where J∗ (λj ) as
seen in (1.2).

3
 (2) Find the rank for each Jordan segment J(λ1 ), . . . , J(λn ) where i = 1, . . . , n and
ri (λj ) = rank((A − λj I)i ).Then,

r1 (λi ) = r1 (A − λi I)

r2 (λi ) = r2 (A − λi I)2

..
.

rn (λi ) = rn (A − λi I)n

Stop computing the rank value once the value begins to repeat. The index(λi ) =
the smallest positive integer k such that rank((A−λi I)k ) = rank((A−λi I)k+1 ).
This k value gives the size of the largest Jordan block for J(λi ).

(3) The number of i × i Jordan blocks in J(λj ) is given by the formula,

vi (λj ) = ri−1 (λj ) − 2ri (λj ) + ri+1 (λj ).

This computed value gives the size of all of the individual Jordan blocks. Once
you have established how many of each block size are needed matrix J can be
constructed.

Remark I.3. [2, p.683] Recall that the algebraic multiplicity is the number of times
λ appears as a root of the characteristic polynomial and the geometric multiplicity is
the number of linearly independent eigenvectors associated with λ or in other words,
geomultA (λ) = dimN (A − λI). Also, it is important to note that we can use Jordan
Normal Form to find the algebraic and geometric multiplicities directly.
For some eigenvalue λ, the algebraic multiplicity is equal to the sum of the sizes
of all Jordan blocks in J(λ). The geometric multiplicity is equal to the number of
Jordan blocks associated with λ. When index(λ) = 1 we have that the largest Jordan

4
 block is of the size 1 × 1. In this case, from Theorem I.2 (1 .1 ) it is evident that
algmultA (λ) = geomultA (λ).

1.3 Norm

Definition I.4. [3, p.280] A matrix norm is a function ||∗|| from the set of all complex
matrices into R satisfies the following properties:

(i) ||A|| ≥ 0 and ||A|| = 0 ⇐⇒ A = 0

(ii) ||αA|| = |α|||A|| for all scalars α

(iii) ||A + B|| ≤ ||A|| + ||B||

(iv) ||AB|| ≤ ||A||||B||

An example of a general matrix norm is the infinity norm. The infinity norm of
a square matrix is the maximum
( n ) the absolute row sums and is denoted by ||A||∞ .
of
X
That is, ||A||∞ = max |aij | .
1≤i≤n
j=1

1.4 Graphs and Matrices

Definition I.5. [4, p.3] A directed graph G = {G0 , G1 , r, s} consists of a finite set
G0 of vertices such that G0 = {v1 , v2 , . . . , vm }, a finite set G1 of edges such that
G1 = {e1 , e2 , . . . , en } and maps r, s : G1 −→ G0 where r(ei ) is the range of the
edge ei and s(ei ) is the source of the edge ei . A path in G is a sequence of edges
e = e1 e2 · · · en with r(ei ) = s(ei+1 ) for 1 ≤ i < n where e has length |e| = n. A cycle
is a path e = e1 · · · en with r(en ) = s(e1 ). A simple cycle occurs when there is no
repetition of edges or vertices along the path except maybe your starting and ending
vertex.

5
Example I.6.

e1 e3

e2
v1 v2
(1.4)

e2
e1

e3
v1 v2

e4 (1.5)

In Figure (1.4) we have that G0 = {v1 , v2 } and G1 = {e1 , e2 , e3 } where r(e1 ) =

v1 , s(e1 ) = v1 , r(e2 ) = v1 , s(e2 ) = v2 , and r(e3 ) = v2 , s(e3 ) = v2 . In Figure (1.5) we
have that G0 = {v1 , v2 } and G1 = {e1 , e2 , e3 , e4 } where r(e1 ) = v1 , s(e1 ) = v1 , r(e2 ) =
v1 , s(e2 ) = v1 , r(e3 ) = v2 , s(e3 ) = v1 , and r(e4 ) = v1 , s(e4 ) = v2 .

A graph is said to be strongly connected if between any two vertices v and w

there exists a path from v to w and from w to v. In Figure (1.4) the graph is not
strongly connected since there is no path from v1 to v2 . On the other hand, the graph
in Figure (1.5) is strongly connected since there exists a path from v1 to v2 and vice
versa.
A graph associated with a non-negative matrix A is denoted by G(A) where the
directed edge (vi , vj ) is in E exactly when aij 6= 0. A non-negative matrix A in which
each non-zero entry is replaced by 1 is called the adjacency matrix for G(A) and is
denoted by AG . Given a graph G, the adjacency matrix AG is an |E 0 | × |E 0 | matrix
where AG (i, j) is 1 when there is a path of length 1 from vi to vj and 0 otherwise.
 
0.7 0 
Example I.7. Consider A =   . Then, G(A) is the graph from Figure
0.2 0.9

6
(1.4) with adjacency matrix

v1 v2
 
v1  1 0
AG =  .
v2 1 1

Given the graph in Figure (1.5) its adjacency matrix is

v1 v2
 
v1  1 1
AG =  .
v2 1 0

Lemma I.8. [3, p.672] Let A ≥ 0. Then (Ak )ij > 0 if and only if there exists a path
of length k from vi to vj in G(A).

(k) (k)
Proof. Let A = (aij ) and let aij denote the (i, j)-entry in Ak . Then, aij =
X
aih1 ah1 h2 · · · ahk−1 j > 0 if and only if there exists a set of indices h1 , h2 , . . . , hk−1
h1 ,...,hk−1
such that aih1 > 0 and ah1 h2 > 0 and · · · and ahk−1 j > 0. Since ast > 0 for some
arbitrary s and t, this is equivalent to saying there is a path of length 1 from vs to
vt in G(A). The previous sentence is equivalent to finding a path of length k from vi
to vj . Therefore, there exists a path of length k in G(A) from vi to vj if and only if
(k)
aij > 0.

1.5 Spectrum

The spectrum of a square matrix A is denoted by σ(A) and is equal to the set
of all eigenvalues of A. The spectral radius of A is the maximum eigenvalue of A in
absolute value; that is, ρ(A) = max {|λ| : λ ∈ σ(A)}. The following are important
facts to note about spectral radius.

7
 Proposition I.9. [3]

(i) A > 0 =⇒ ρ(A) > 0

(ii) A > 0 ⇔ A/ρ(A) > 0

(iii) If c is a constant then, ρ(cA) = |c|ρ(A)

(iv) Let r > 0 then, ρ(A) = r ⇔ ρ(A/r) = 1

(v) If ρ(A) < 1 then, lim Ak = 0. That is, lim (Ak )ij = 0 for each i, j.
k→∞ k−→∞

Proof.

(i) Let A = (aij ) with aij > 0. We know that ρ(A) ≥ 0. Suppose ρ(A) = 0. Then
it follows that σ(A) = {0}. If this were the case we would have the Jordan Normal
Form of A would imply Ak = 0 for some k ≥ 0. If A > 0 then Ak 6= 0 for all k ≥ 0.
This poses as contradiction that ρ(A) = 0. Therefore, ρ(A) > 0.

(ii) (=⇒) If A > 0 from (i) we know that ρ(A) > 0. From the definition of spectral
radius and positive matrices we know that any entry aij > 0 of A divided by ρ(A) > 0
will also be positive.
(⇐=) Suppose A/ρ(A) > 0. From (i) we have that ρ(A) > 0. If ρ(A) > 0 then for
A/ρ(A) > 0 we must have A > 0.

(iii) Let c be a constant. We know that ρ(A) = max {|λ| : λ ∈ σ(A)}. If λ is an

eigenvalue of A with eigenvector x then multiplying A by some constant c we get,

Ax = λx

cAx = cλx

(cA)x = (cλ)x

8
 This shows the eigenvalue of cA is cλ. Let {λ1 , . . . , λp } be the distinct eigenvalues of
A then it follows that, ρ(A) = max {|λ1 |, |λ2 |, . . . , |λp |}. We know ρ(A) = |λM | for
some λM . Since |λi | ≤ |λM | for each i we have that |c||λi | ≤ |c||λM |. Indeed,

ρ(cA) = max {|cλ1 |, |cλ2 |, . . . , |cλp |}

= max {|c||λ1 |, |c||λ2 |, . . . , |c||λp |}

= |c||λM |

= |c|ρ(A)

Therefore, if c is constant the eigenvalues of cA are {cλ1 , . . . , cλp }.

(iv) Let ρ(A) = r where r > 0. !

A
(=⇒) We want to show that if ρ(A) = r then ρ = 1. From (iii) we have that,
r
!
A 1 1
ρ =  ρ(A) = (r) = 1 as required.
 
r r r
! !
A A
(⇐=) We want to show that if ρ = 1 then ρ(A) = r. If ρ = 1 from (iii)
r r
1
it follows that  ρ(A) = 1 and since r > 0 we have that ρ(A) = 1(r) = r.
 
r

(v) Since ρ(A) < 1, then |λ| < 1 for all λ ∈ σ(A). We know from Theorem I.2 (1.3)
kj
   
k k k−j k k(k − 1) · · · (k − j + 1)
that each entry of J∗ is of the form λ . Since = ≤
j j j!   j!
j
 
k k−j k  k k−j 
 
this implies that λ ≤ λk−j . Taking the absolute value we get,  λ ≤
j j! j
k j k−j k j /j! ∞
|λ| −→ 0. Taking lim j−k
= so, L’Hopital’s rule can be applied. Re-
j! k−→∞ |λ| ∞
peatedly applying L’Hopital’s rule it is evident that k j tends to 1 and |λ|j−k will
k j /j!
eventually approach ∞ so it follows that lim −→ 0. Since Jk∗ −→ 0 we have
k−→∞ |λ|j−k

that Ak −→ 0.

9
 CHAPTER II

The Perron-Frobenius Theorem for Positive

Matrices

2.1 The Perron-Frobenius Theorem Stated

Theorem II.1. [3, p.667] If A is our n × n matrix, with A > 0 and r = ρ(A), then
the following statements are true.

(i) r > 0

(ii) r ∈ σ(A) (r is called the Perron-Frobenius eigenvalue)

(iii) algmult A (r) = 1 and geomult A (r) = 1

(iv) There exists an eigenvector x > 0 such that Ax = r x

(v) The Perron- Frobenius eigenvector is the unique vector defined by

Ap = r p, p > 0 and ||p||1 = 1

and, except for positive multiples of p, there are no other non-negative eigen-
vectors for A, regardless of the eigenvalue.

(vi) r is the only eigenvalue on the spectral circle of A.

10
(vii) r = max f (x) (the Collatz-Wielandt formula), where
x∈N

[Ax]i
f (x) = min and N = {x : x ≥ 0 with x 6= 0}.
1≤i≤n xi
xi 6=0

We are going to prove all these statements in the following sections.

2.2 Positive Eigenpair

Remark II.2. |Ax| ≤ |A||x|. Indeed, using the triangle inequality we get,

(|Ax|)i = |ai1 x1 + ai2 x2 + · · · + ain xn |

≤ |ai1 x1 | + |ai2 x2 | + · · · + |ain xn |

≤ |ai1 ||x1 | + |ai2 ||x2 | + · · · + |ain ||xn |

= |A|i |x|

where |A|i represents the ith row of |A|.

 
1 0 0
 
Let A = 
0 −5 0 . The eigenvalues of A are λ1 = 1, λ2 = −5, λ3 = 2. That
 
0 0 2
is, σ(A) = {1, 2, −5}. This implies that ρ(A) = | − 5| = 5, but 5 ∈
/ σ(A). Below we
will show that if A > 0, then the spectral radius of A is in fact an eigenvalue of A.

Theorem II.3. Positive Eigenpair [3, p.663]

Let A > 0 be an n × n matrix then the following hold true,

(i) ρ(A) ∈ σ(A)

(ii) If Ax = ρ(A)x then A|x| = ρ(A)|x| and |x| > 0

11
Proof. We will first prove this for the case when ρ(A) = 1. Since ρ(A) = |λ| = 1 and
Ax = λx we have that (λ, x) is an eigenpair of A. Then,

|x| = |λ||x| = |λx| = |Ax|

≤ |A||x|

= A|x|

From Remark II.2 we have that |x| ≤ A|x|.

Let z = A|x| and y = z − |x|. We know that 0 < |x| ≤ A|x|, so it follows that
z > 0. Since, |x| ≤ A|x| we have that y ≥ ! 0. We want to show y = 0. Suppose on
(Ay)i
the contrary that y 6= 0. Let ci = . Since (Ay)i > 0 and zi > 0 we have
zi
ci > 0 for each i = 1, 2, . . . , n. Let c = min{ci } and take  < c. Then  ≤ ci for each
i. Therefore,
!
(Ay)i
>
zi

(Ay)i > zi

(Ay)i > (z)i

Ay > z

We know that Ay > z or equivalently,

A(z − |x|) > (A|x|)

Az − A|x| > (A|x|)

Az > A|x| + A|x|

12
 Az > A|x|(1 + )
!
A
z > A|x|
1+
!
A
z>z
1+

!
A
Let B = which implies Bz > z. Since A > 0 we have that B > 0 so we get
1+
the following result,

Bz > z

B2 z > Bz > z

B3 z > Bz > z

..
.

Bk z > z

! !
A 1 1
We know that ρ(B) = ρ = ρ(A) = < 1 so, from Proposition
1+ 1+ 1+
I.9 (v) we have that lim Bk = 0 for each k . Then 0 ≥ z which is a contradiction,
k−→∞

so y = 0. Since y = 0 = A|x| − |x| we have that x is an eigenvector for A with

corresponding eigenvalue ρ(A) = 1.
In the general case, suppose ρ(A) = k. By Proposition I.9 (i), we know k > 0.
A
Since ρ(A) = k from Proposition I.9 (iii ) we have ρ = 1 so 1 is an eigenvalue
k
A A
of . We want to show ρ(A) ∈ σ(A). Since 1 is an eigenvalue of we have that
k k
A
x = x which implies that Ax = kx = ρ(A)x. Taking the absolute value we get,
k
|Ax| = |ρ(A)x| and it follows that A|x| = ρ(A)|x|.

13
 The above proof of positive eigenpair yielded the following important result: ρ(A)
is an eigenvalue for A > 0.

Remark II.4. In the previous proofs we have used ρ(A) = 1. In light of Proposition
I.9 (iii), in further proofs we can assume without loss of generality that this is true.

2.3 Index of ρ(A)

Definition II.5. [3, p.510] If A is a square matrix then λ is a semi-simple eigen-

value if and only if algmultA (λ) = geomultA (λ). An eigenvalue is called simple if
algmultA (λ) = 1.

P
Lemma II.6. [3, p.664] For nonzero vectors {z1 , . . . , zn } it follows that || j zj ||2 =
P
j ||zj ||2 ⇐⇒ zj = αj z1 where αj > 0.

Theorem II.7. [3, p.664] Let A > 0 be an n × n matrix. Then, the following
statements are true.

(i) ρ(A) is the only eigenvalue of A on the spectral circle. That is, |λ| = ρ(A) then
λ = ρ(A).

(ii) index(ρ(A)) = 1. In other words, ρ(A) is a semi-simple eigenvalue.

Proof.

(i) Assume without loss of generality that ρ(A) = 1. Then (λ, x) is an eigenpair for
A such that |λ| = 1. From Theorem II.3 (ii), we know that 0 < |x| ≤ A|x|. Consider
the k th entry such that 0 < |xk | = (A|x|)k . Let j = 1, 2, . . . , n. We know that
Xn
(A|x|)k = akj |xj |. Also,
j=1

|xk | = |λ||xk |

= |(λx)k |

14
 = |(Ax)k |
 
Xn 
= akj xj 
 
 
j=1

Therefore,
 
Xn  X n n
X
akj xj  = akj |xj | = |akj xj |.
 

 
j=1 j=1 j=1

From Lemma II.6 we have that, akj xj = αj (ak1 x1 ) for each j. We know akj > 0 so it
follows that,

akj xj = αj (ak1 x1 )

αj (ak1 x1 )
xj =
akj

αj ak1
xj = x1
akj

αj (ak1 )
Let πj = > 0 then we have that xj = πj x1 . Let x = x1 p where p =
akj
(1, π2 , . . . , πn )T > 0. Then, Ax = λx =⇒ A(x1 p) = λ(x1 p) =⇒ Ap = λp. Then,
λp = Ap = |Ap| = |λp| = |λ|p = p. So we have that λp = p. Setting λp − p = 0
we get that p(λ − 1) = 0. We know that p > 0 so λ − 1 = 0 which implies λ = 1.
Therefore, λ = 1 is the only eigenvalue of A on the spectral circle.

(ii) We want to show that index(1) = 1. Suppose on the contrary, index(1) = m > 1.
That is, the largest Jordan block for J∗ (1) is of m x m size. We know that λ = 1
so from Theorem I.2 (1 .3 ) we have that ||Jk ||∞ ≥ ||Jk∗ ||∞ so each individual Jordan
block, ||Jk∗ ||∞ −→ ∞. Since ||Jk ||∞ ≥ k for all k we have that ||Jk ||∞ −→ ∞ as
k −→ ∞. From Theorem I.2 we know that Jk = P−1 Ak P and since ||Jk ||∞ −→ ∞

15
it directly follows that ||Ak ||∞ −→ ∞. Indeed,

Jk = P−1 Ak P

||Jk ||∞ = ||P−1 Ak P||∞

||Jk ||∞ ≤ ||P−1 ||∞ ||Ak ||∞ ||P||∞

Since P is invertible P 6= 0 it follows that, ||P||∞ 6= 0 and ||P−1 ||∞ 6= 0. Therefore,

||Jk ||∞
||Ak ||∞ ≤ −→ ∞.
||P−1 ||∞ ||P||∞

(k) (k)
X (k)
Let Ak = [aij ] where [aij ] represents (i, j) entries of Ak . Then, ||Ak ||∞ = |aik j |
j
for some 1 ≤ ik ≤ n. From above, we know that there exists a vector p > 0 such
that p = Ap. Continually multiplying both sides by A we eventually get p = Ak p.
Then,

||p||∞ ≥ pik

X (k)
= ai k j p j
j

X 
(k)
≥ aik j (min pi )
i
j

= ||Ak ||∞ (min pi ) −→ ∞

i

Since p was a constant vector it cannot approach ∞ so, we have a contradiction.

Therefore, index(1) = 1. If the index(1) = 1 we want to show that λ = 1 is a
semi-simple eigenvalue. Since index(1) = 1 from Remark I.3 it directly follows that
algmultA (1) = geomultA (1) = 1 as desired.

16
2.4 Multiplicities of ρ(A)

Theorem II.8. [3, p.664] If A > 0, then algmultA (ρ(A)) = 1. In other words,
the spectral radius of A is a simple eigenvalue of A.Therefore, dimN (A − ρ(A)I) =
geomultA ρ(A) = algmultA ρ(A) = 1.

Proof. Assume without loss of generality that ρ(A) = 1. We want to show that
algmultA (1) = 1. Suppose on the contrary, algmultA (1) = m > 1. From Theorem
II.7 (ii ) we know λ = 1 is a semi-simple eigenvalue so it follows that algmultA (1) =
geomultA (1). Since geomultA (1) = m > 1, we must have m ≥ 2 which means
dimN (A − 1I) ≥ 2. Thus, there exists at least two linearly independent eigenvectors
say, (x, y) such that x 6= αy for all α ∈ C. Select a non-zero component from y,
x 
i
say yi 6= 0 and let z = x − y. Since x and y are linearly independent eigen-
yi x  x 
i i
vectors we have that Ax = x and Ay = y so, Az = Ax − Ay = x − y
yi yi
which implies that Az = z. From Theorem II.3 we know that A|z| = |z| > 0. Then
x 
i
zi = xi − yi = xi − xi = 0. This shows zi = 0, but |z| > 0 so we have a contra-
yi
diction. Therefore, m = 1 and algmultA (1) = algmultA (ρ(A)) = 1.
Now we want to show that dimN (A − 1I) = geomultA (1) = algmultA (1) = 1. In
Theorem II.7 (ii ) it was shown that geomultA (λ) = algmultA (λ). From above, we
found that that the algmultA (1) = 1 so it directly follows that geomultA (1) = 1.

Since dimN (A − ρ(A)I) is a one-dimensional space that can be spanned by some

X
v > 0 there exists a unique eigenvector p ∈ N (A − ρ(A)I) such that pj = 1.
j
This unique eigenvector is called the P erron-F robenius eigenvector for A > 0 and
the eigenvalue r = ρ(A) is called the P erron-F robenius eigenvalue for A. Since
A > 0 ⇐⇒ AT > 0 and ρ(A) = ρ(AT ) then in addition to the eigenpair (r, p) for A
there exists an eigenpair (r, q) for AT . Since qT A = rqT the vector, qT > 0, is called
the lef t-hand P erron-F robenius eigenvector for A.

17
Proposition II.9. [3, p.666] Although eigenvalues of A > 0 other than ρ(A) may or
may not be positive there are no non-negative eigenvectors for A > 0 other than the
Perron vector p and its positive multiples.

Proof. Let r = ρ(A) and (λ, y) be an eigenpair for A such that y ≥ 0. Given (r, q)
where q is the left-hand Perron-Frobenius eigenvector for A we know that rq = AT q.
Since qT y > 0 we get,

(rq)T = (AT q)T

rqT = qT A

rqT y = (qT A)y

rqT y = qT (Ay)

rqT y = λqT y

r=λ

Therefore, r = ρ(A) = λ. From Theorem II.8 we know the eigenspace for the Perron-
Frobenius eigenvalue of r is one-dimensional so p can be written as p = αy for some
1
α. Since p > 0 and y ≥ 0 we must have α > 0. Then, y = p and therefore, y is a
α
positive multiple of the Perron-Frobenius one.

2.5 Collaltz-Wielandt Formula

Theorem II.10. [3, p.666] The Perron eigenvalue of A > 0 is given by r = max f (x),
x∈N

where
[Ax]i
f (x) = min and N = {x : x ≥ 0 with x 6= 0}.
1≤i≤n xi
xi 6=0

18
 [Ax]i
Proof. Let β = f (x) for some arbitrary x ∈ N . Since f (x) ≤ for each i we
xi
[Ax]i
have that β ≤ for each i so, βxi ≤ [Ax]i which implies that 0 ≤ βx ≤ Ax. Let
xi
p and qT be the respective right-hand and left-hand Perron-Frobenius eigenvectors
for A associated with the Perron-Frobenius eigenvalue r . Then,

βx ≤ Ax

βqT x ≤ (qT A)x

βqT x ≤ rqT x

β≤r

Therefore, f (x) ≤ r for all x ∈ N . We know max f (x) ≤ r. We want to show that
x∈N

in fact r = max f (x). Indeed,

x∈N

[Ap]i
f (p) = min
1≤i≤n pi

[rp]i
= min
1≤i≤n pi

pi
= r min
1≤i≤n pi

=r

Since f (p) = r for p ∈ N it follows that, r = max f (x).

x∈N

19
 CHAPTER III

Perron Frobenius Theorem for Non-negative

Matrices

3.1 Non-negative Eigenpair

A standard fact to note about spectral radius is that ρ(A) = lim ||Ak ||1/k [3,
k−→∞

p.619]. We will use this result to prove the following Lemma.

Lemma III.1. [3, p.619] If |A| ≤ B then ρ(A) ≤ ρ(|A|) ≤ ρ(B), where |A| denotes
the matrix having entries |aij | and define B ≤ C to mean bij ≤ cij for each i and j.

Proof. Using the triangle inequality we get that, |Ak | ≤ |A|k for every positive integer
k. So, |A| ≤ B implies that |A|k ≤ Bk . Using our result from above we get,
||Ak ||∞ = |||Ak |||∞ ≤ |||A|k ||∞ ≤ ||Bk ||∞ . Then,

||Ak ||1/k k 1/k k 1/k

∞ ≤ |||A| ||∞ ≤ ||B ||∞

lim ||Ak ||1/k k 1/k k 1/k

∞ ≤ lim |||A| ||∞ ≤ lim ||B ||∞
k−→∞ k−→∞ k−→∞

ρ(A) ≤ ρ(|A|) ≤ ρ(B)

20
Theorem III.2. [3, p.670] Let A ≥ 0 be an n×n matrix with r = ρ(A), the following
statements are true.

(i) r ∈ σ(A), but r = 0 is possible.

(ii) Az = rz for some z ∈ N = {x|x ≥ 0 with x 6= 0}.

[Ax]i
(iii) r = max f (x), wheref (x) = min (ie.the Collatz-Wieldant Formula
x∈N 1≤i≤n xi
xi 6=0
remains valid).
1
Proof. Let Ak = A + E > 0 where E is a matrix containing all 1’s and k =
k
1, 2, . . . , n. Then it follows that Ak > 0 since A ≥ 0, k > 0 and E > 0. Let
rk > 0 and pk > 0 be the Perron-Frobenius eigenvalue and the Perron-Frobenius
eigenvector corresponding to Ak , respectively. We know {pk }∞
k=1 is bounded since

{pk }∞
k=1 ⊆ {x ∈ R : ||x||1 = 1}. The Bolanzo-Wierstrass Theorem states that every

bounded sequence has a convergent subsequence. Thus, since {pk }∞

k=1 is bounded,

a convergent subsequence exists. Let z be an arbitrary point in Rn . We have that

{pki }∞
i=1 −→ z for some increasing sequence ki where z ≥ 0 and z 6= 0 since pki > 0
n
X
and ||pki ||1 = |pki | = 1. Hence, since A1 > A2 > . . . > A from Lemma III.1
i=1
it is clear that r1 ≥ r2 ≥ . . . ≥ r so {rk }∞
k=1 is a decreasing sequence of positive

numbers bounded below by r. In other words, lim rk = r∗ exists and r∗ ≥ r. In

k−→∞
∗
particular, lim rki = r ≥ r. But, from our definition of Ak we have lim Ak = A
i−→∞ k−→∞

which implies that lim Aki −→ A so it is also true that,

i−→∞

Az = lim Aki lim pki

i−→∞ i−→∞

= lim Aki pki

i−→∞

= lim rki pki

i−→∞

= lim rki lim pki

i−→∞ i−→∞

21
 = r∗ z

Since Az = r∗ z we have that r∗ ∈ σ(A) and therefore, r∗ ≤ r. Since r = ρ(A), we

know r∗ will always be less than r as it is not the maximum eigenvalue so therefore,
r∗ = r.
Let qk T be the left-hand Perron-Frobenius eigenvector for Ak . From Proposition
II.9 we know that for every x ∈ N and k > 0 we have that qk T > 0. Then it follows
that, 0 ≤ f (x)x ≤ Ax ≤ Ak x. Indeed,

qk T [f (x)x] ≤ qk T (Ak x)

f (x)qk T x ≤ (qk T Ak )x

f (x)qk T x = rk qk T x

f (x) ≤ rk

Since rk −→ r∗ = r we have that f (x) ≤ r. Since f (z) = r and z ∈ N it directly

follows that max f (x) = r. Therefore, the Collatz-Wieldant formula is satisfied.
x∈N

3.2 Reducibility

Definition III.3. [3, p.671] A matrix A ≥ 0 is said to be a reducible matrix when

there exists a permutation matrix P such that,
 
X Y 
PT AP =  
0 Z

where X and Z are both square. Otherwise, A is said to be an irreducible matrix.

To determine whether a matrix is irreducible it is useful to use the following fact: a

22
matrix A is irreducible if and only if G(A) is strongly connected [3, p. 671].
 
1 0 
Example III.4. The matrix A =   is reducible because we are able to find
1 1
   
X Y  0 1 
a permutation matrix P such that PT AP =   . Namely, P =  , and
0 Z 1 0
 
1 1 
PT AP =  . We notice that the matrix A gives the graph G(A) as seen in
0 1
Figure (1.4) on page 6. We also notice that the graph is not strongly connected since
there does not exist a path from v1 to v2 . Hence, this matrix is not irreducible and
therefore, it must be reducible. On the other hand, the graph in Figure (1.5) on page
6 is strongly connected which implies its corresponding adjacency matrix is in fact
irreducible.
 
∗ ∗ 
Remark III.5. A matrix A =   where the ∗’s are non-negative and non-
∗ 0
zero numbers, is reducible since G(A) is the same graph as in Figure (1.4) and the
discussion in Example III.4 above will follow.

The following Lemma shows a simple process that allows us to convert non-
negative irreducible matrices into a positive matrix.

Lemma III.6. [1, p.192] Converting Non-negativity and Irreducibility to Positivity

Let A ≥ 0 be an n × n irreducible matrix then, (I + A)n−1 > 0 where I is the identity
matrix.

Proof. From Lemma I.8, it follows that if (Ak )ij > 0 if and only if there exists a path
of length k from vi to vj in G(A). If there exists a path of length at most k between
Xk
every pair of vertices, then AL > 0.
L=0
Since G(A) has n vertices, it follows from a simple induction proof with respect
to n that between any two pairs of vertices there exists a path of length at most

23
 n−1
X
n − 1. This implies that AL > 0, by Lemma I.8. Then, for any positive integers
L=0
n−1
X n−1
X
L
c0 , c1 , . . . , cn−1 it follows that A > 0 if and only if cL AL > 0. Let I be the
L=0 L=0
identity matrix. Since A and I commute the Binomial Theorem holds. Indeed,

n−1   n−1  
n−1
X n−1 L (n−1)−L
X n−1
(A + I) = A I = AL .
L=0
L L=0
L

  n−1
n−1 X
Choose cL = . Since AL > 0, we have that (A + I)n−1 > 0.
L L=0

Definition III.7. [3, p.600] Function of Jordan Blocks

For a k × k Jordan block, J∗ , with eigenvalue λ, and for a function f : C −→ C is
said to be defined (or to exist) at A when f (λ), f 0 (λ), . . . , f (k−1) (λ) exist such that,
f (J∗ ) is defined to be

 
0 f 00 (λ) f (k−1) (λ)
  f (λ) f (λ) 2!
··· (k−1)! 
λ 1 · · · 0  

... ..
f (λ) f 0 (λ)

. . .
. . . . . ...  
  
 .. 
f (J∗ ) = f  .
 
= .. .. f 00 (λ)  .

. .

 .. · · · . . . 1   2! 
   
  f (λ) f 0 (λ) 
 

0 ··· ··· λ  
f (λ)

Definition III.8. [3, p.601] Function of 

a Matrix 
..
 . 
 
Suppose that A = PJP−1 , where J =   J∗ (λj )  is in Jordan form. If f

...
 

exists at A then the value of f at A is defined to be

 
..
 . 
−1
  −1
f (A) = Pf (J)P = P
 f (J∗ (λj )) P .

...
 

24
Remark III.9. [3, p.601] The matrix function f (A) produces a uniquely defined
matrix.

Based on the definitions above, if f exists at A, then f (A) = Pf (J)P−1 . Thus,

if (λ, x) is an eigenpair for A, it follows that (f (λ), x) is an eigenpair for f (A). This
result will be used in the following proof.

Theorem III.10. [3, p.673] Let A ≥ 0 be an n × n irreducible matrix, then the

following statements are true.

(i) r = ρ(A) ∈ σ(A) and r > 0.

(ii) algmult A (r) = 1.

(iii) There exists an eigenvector x > 0 such that Ax = r x.

(iv) The unique vector defined by

Ap = r p, p > 0, and ||p||1 = 1

is called the Perron vector. There are no other non-negative eigenvectors for
A except for positive multiples of p, regardless of the eigenvalue.

(v) The Collatz-Wieldant formula r = max f (x), where

x∈N

[Ax]i
f (x) = min and N = {x : x ≥ 0 with x 6= 0}.
1≤i≤n xi
xi 6=0

Proof. From Theorem III.2 we know that r = ρ(A) ∈ σ(A) which shows state-
ment (i) is true. Now we want to show that the algmultA (r) = 1 is also true.
Since A is irreducible we will let B = (I + A)n−1 > 0 be our matrix of size
n × n. Let f (A) = B. From Theorem III.8 since f is differentiable it follows that

25
f (x) = (1 + x)n−1 . Then, λ ∈ σ(A) if and only if (1 + λ)n−1 ∈ σ(B). From Re-
mark I.3 we know that in terms of Jordan blocks the algmultA (λ) is equal to the
sum of the sizes of all Jordan blocks associated with λ. Then the size of the Jordan
blocks corresponding to f (λ) will be the same size as our matrix function f (A) so
we have that algmultA (λ) = algmultB ((1 + λ)n−1 ). If u = ρ(B) then by defini-
 n−1
tion, u = max |u| = max |(1 + λ)|n−1 = max |(1 + λ)| = (1 + r)n−1 since
λ∈σ(B) λ∈σ(A) λ∈σ(A)
|1 + λ| ≤ 1 + |λ| ≤ 1 + r where r is an eigenvalue of A. Since r is an eigenvalue of A
from Theorem II.8 we have thatalgmultA (r) = 1 as desired.
From Theorem III.2 we know there exists a non-negative eigenvector x ≥ 0 asso-
ciated with our eigenvalue r. Then we can find a corresponding eigenvector for our
eigenvalue r. So if (λ, x) is an eigenpair for A then it follows that (f (λ), x) is an
eigenpair for f (A). In our case, since (r, x) is an eigenpair for A implies that (u, x)
is an eigenpair for B. The Perron-Frobenius Theorem for Positive Matrices ensures
that x must be a positive multiple of the Perron vector B which implies that x > 0.
We want to show that Ax = rx. If r = 0 then Ax = 0 which is not possible since
A 6= 0 and is non-negative and x > 0. Therefore, r > 0 and we obtain our result for
(iii).
x
To show Ap = rp we simply scale our eigenvector x by to get our result for
||x||1
(iv). Part (v) was proven in Theorem III.2 (iii).

3.3 Primitive Matrices

Definition III.11. [3, p.674] A non-negative irreducible matrix A having only one
eigenvalue, r = ρ(A), on its spectral circle is said to be a primitive matrix. Equiva-
lently, a matrix is called primitive if there exists a positive integer k such that Ak is
a positive matrix [1, p.198].

26
  
1 1 
Example III.12. The matrix A =   is primitive since for any k > 0 it
1 0
 
2 1 
follows that Ak > 0. For instance, A2 =   > 0. On the contrary, the matrix
1 1
 
0 1
A=  is irreducible, but not primitive since its eigenvalues are ±1 which both
1 0
lie on the unit circle.

Remark III.13. [1, p.202] Every primitive matrix is irreducible, but not every irre-
ducible matrix is primitive. We can tell whether a matrix is primitive by looking at
the greatest common divisor of the cycles within the graph.

Theorem III.14. [1, p.202] A non-negative matrix A is primitive if and only if G(A)
is strongly connected and has two relatively prime cycle lengths.

Corollary III.15. [1, p.203] A non-negative matrix is primitive if and only if the
corresponding graph is strongly connected and the gcd of its simple cycles is 1. We
call such a graph a primitive graph.
 
0 2 
Example III.16. The matrix A =   is primitive with eigenvalues 2 and −1,
1 1
 
0 4 
but the matrix B =   with eigenvalues 2, and −2 is not primitive. If you were
1 0
to graph matrices A and B it would be clear that matrix A is the only non-negative
primitive matrix. The graph of matrix A is shown below.

e2

e3
v1 v2
e4
e1 (3.1)

27
From Theorem III.14, for matrix A to be primitive its corresponding graph must
be strongly connected and have two relatively prime cycle lengths. It is evident the
graph below is strongly connected as there exists a path from v1 to v2 and vise versa.
In addition, the graph has two relatively prime cycles; the first cycle starting at v2
and travelling along e1 to e3 back to v2 and the second cycle starting at v1 travelling
along e3 to e2 to e1 and then back to v1 . The first cycle is of length 2, the second is
of length 3 which are two relatively prime cycles.

Definition III.17. Projection [3, p.385]

Let X and Y be vector spaces over C. Suppose V = X ⊕ Y . Then, for each v ∈ V
there are unique vectors x ∈ X and y ∈ Y such that V = X + Y .

(i) The vector x is called the projection of v onto X along Y .

(ii) The vector y is called the projection of v onto Y along X.

Theorem III.18. Projector [3, p.386]

Let X and Y be vector spaces over C and let v be in vector space V . The unique
linear operator P defined by Pv = x is called the projector onto X along Y and P
has the following properties:

(i) P2 = P

(ii) R(P) = {x|Px = x}

(iii) R(P) = N (I − P) = X and R(I − P) = N (P) = Y

Theorem III.19. [3, p.339] A linear projector P on V is a projector if and only if

P2 = P.

We showed in Proposition I.9 that if ρ(A) < 1 then lim Ak exists. The next
k−→∞

result gives necessary and sufficient conditions for the existence of lim Ak .
k−→∞

28
Lemma III.20. [3, p.630] For A ∈ Cn×n , lim Ak exists if and only if ρ(A) < 1 or
k−→∞

else ρ(A) = 1, where λ = 1 is the only eigenvalue on the unit circle, and λ = 1 is
semi-simple. When it exists, lim Ak is equal to the projector onto N (I − A) along
k−→∞

R(I − A).

Theorem III.21. Primitive Matrices [3, p.674]

Let A be a non-negative irreducible matrix and r = ρ(A). Then A is primitive if
and only if
!k
A pqT
lim =G= >0
k−→∞ r qT p

where p and q are the respective Perron vectors for A and AT . Also, G is the
projector onto N (A − rI) along R(A − rI).
A
Proof. By the Perron-Frobenius Theorem for Irreducible Matrices III.10, 1 = ρ
r
A
is a simple eigenvalue. We know A is primitive if and only if is primitive. Matrix
r  A k 1
k
A is primitive if and only if there is a k > 0 such that A > 0. Then, = Ak .
r r
k A
Since A > 0 and r > 0 it follows that is also primitive. By definition III.11,
r
A  A 
is primitive if and only if 1 = ρ is the only eigenvalue on its spectral circle.
r r  
A k A
By Lemma III.20, we have that lim exists since 1 = ρ is a simple
k−→∞ r r  A k
eigenvalue and the only eigenvalue on its spectral circle. In addition, lim is
k−→∞ r
 A  A
the projector onto N I − along R I − . Then it follows from Lemma III.20
r r
that the lim Ak exists and is equal to the projector onto N (I − A) along R(I − A).
k−→∞
!k
A pqT
Also, G = lim = T > 0 is a projector by Theorem III.19.
k−→∞ r q p

29
 CHAPTER IV

The Leslie Model

In this chapter, we will take all of our previous knowledge and apply it to real
world applications. We will begin with a simple population model developed through
the Rabbit Problem.

4.1 The Rabbit Problem

In this problem, let At represent the number of adult pairs of rabbits at the end
of month t and let Yt be the number of youth pairs of rabbits at the end of month
t. To begin, we will start with one pair of youth rabbits. Each youth pair takes two
months to mature into adulthood. In this particular model, both adults and youth
give birth to a pair at the end of every month, but once a youth pair matures to
adulthood and reproduces, it then becomes extinct. The procedure goes as follows:

(1) At the end of Month 0, At = 0 and Yt = 1.

(2) At the end of Month 1, At = 1 and Yt = 1.

(3) At the end of Month 2, At = 1 and Yt = 2. This gives us a total of 3 pairs of

rabbits at the end of Month 2.

(4) At the end of Month 3, At = 2 and Yt = 3. Hence, we have a total of 5 pairs.

30
 (5) At the end of Month 4, At = 3 and Yt = 5, giving us a total of 8 pairs at the
end of Month 4.

The table below is a simple illustration of the above procedure where P1 , P2 , . . . , Pn

represent the first pair of rabbits, the second pair of rabbits, and so on. It is also
important to note the underline represents the pair(s) that reproduce a new pair
during that specific month.

Month(t) Pairs At Yt Total Number of Pairs

0 P1 0 1 1
1 P1 P2 1 1 2
2 P1 P2 P3 1 2 3
3 P1 P2 P3 P4 P5 2 3 5
4 P1 P2 P3 P4 P5 P6 P7 P8 3 5 8
.. .. .. .. ..
. . . . .

Remark IV.1. Thus, the total number of rabbit pairs at the end of the nth month is
equal to the sum of the number of pairs at the end of the previous two months.

We can make this model more reasonable by utilizing the age classes through a math-
ematical model. Indeed,
    
 Yt+1  1 1  Yt 
   
Y0 1
=    where = .
A0 0

At+1 1 0 At

We can write this as,

 
1 1 
 
1
f (t + 1) =   f (t), where f (0) = .
1 0 0

Hence, the number of youth pairs at the end of month t + 1 should equal the number

31
of youth pairs at the end of month t plus the number of adult pairs at the end of
month t. Also, the number of adults pairs at the end of month t + 1 is equal to the
number of youth pairs at the end of month t.
This model has some unrealistic features such as, every young surviving to adult-
hood or the same number of offspring being produced by both the adult pair and
young pair. To avoid these setbacks we can generalize this model using survival and
fertility rates. Since the 1’s in the first row represent the number of offspring pro-
duced so we can replace these 1’s with birth rates b1 and b2 where b1 , b2 ≥ 0. Since
the lower 1 in our matrix represents a youth surviving into adulthood we will replace
it by s, which is called the survival rate, with 0 < s ≤ 1. Indeed,
 
 b1 b2 
f (t + 1) =   f (t).
s 0

It is evident that this simple model has some features of population such as, the
population size may increase, decrease or die out, but it can be improved even further.
The Fibonacci Model with two age classes can be generalized to k age classes and we
call this the Leslie Model.

4.2 Leslie Model

The Leslie Model is stated as f (t+1) = Lf (t) where f (t) and f (t+1) are population
vectors and L is our Leslie Matrix. We define L by,

 
b b ··· bn
 1 2 
 
s 1 0 · · · 0
 
 
L=
 s2 .

... .. 
 
.


 
sn−1 0

32
Each component is the number of individuals of a particular age class such that
f (t) = (f1 (t), f2 (t), . . . , fn (t))T . Then at time t, fi (t) is the number of individuals of
age class i. That is, the number of individuals of age class a where i − 1 ≤ a < i for
i = 1, 2, . . . , n. Let bk and sk denote the birth rate and survival rate of L, respectively.
That is, the first row of matrix L consists of birth rates where bk is the number of
off-spring produced by an individual of age class i at time t. The diagonal entries of
L consists of survival rates, where sk is the probability that an individual in age class
i will survive to age class i + 1. All other entries of our Leslie matrix are zeroes. For
each k, we assume 0 < si ≤ 1 and bi ≥ 0 and we will assume the width of a particular
age class is exactly one time unit.
Suppose you were to divide a population into specific age classes say, G1 , G2 , . . . , Gn
such that each age class covers the same number of years. If fk (t) is the number of
individuals in Gk at time t, then it follows that

f1 (t + 1) = f1 (t)b1 + f2 (t)b2 + · · · fn (t)bn (4.1)

and

fk (t + 1) = fk−1 (t)sk−1 for k = 2, 3, . . . , n. (4.2)

Matrix equations (4.1) and (4.2) become f (t + 1) = Lf (t), where L is our Leslie
matrix. We also notice that,

fk (t)
Fk (t) =
f1 (t) + f2 (t) + · · · + fn (t)

is the percentage of population in Gk at time t. The vector F(t) = (F1 (t), F2 (t), . . . , Fn (t))T
represents the age distribution at time t.
We will now show that F∗ = lim F(t) exists and will determine its value. Let
t−→∞

33
f (t + 1) = Lf (t) where L is our Leslie Matrix stated above. The graph of G(L)
is strongly connected so from Theorem III.14 we know that L is primitive. Since
G(L) is strongly connected it also follows that L is irreducible. Then we have that,
r = ρ(L) > 0 is an eigenvalue for our matrix, by Theorem 2.1. Also, since L is prim-
!t
L pqT
itive we have that lim = G = T > 0, by Theorem III.21. By a simple
k−→∞ r q p
t
induction proof it is true that f (t) = L f (0). From here if f (0) = 0 we get that,

f (t) Lt f (0)  L t pqT  qT f (0) 

lim = lim = lim f (0) = Gf (0) = f (0) = p .
t−→∞ r t t−→∞ rt t−→∞ r qT p qT p

fk (t)
We know that, Fk (t) = is the percentage of population that is in Gk at time
||f (t)||1
t. Indeed,

f (t) f (t)/rt limt−→∞ f (t)/rt

F∗ = lim F(t) = lim = lim = = p.
t−→∞ t−→∞ ||f (t)||1 t−→∞ ||f (t)||1 /r t limt−→∞ ||f (t)||1 /rt

It is evident! that f (t) approaches 0 if r < 1. If r = 1 then f (t) approaches Gf (0) =

T
q f (0)
p > 0. Finally, if r < 1 the results from our Leslie analysis imply that
qT p
fk (t) approaches ∞ for each k.

34
 REFERENCES

[1] Flahive M. Robson R. Cull, P., Difference equations: from rabbits to chaos,
Springer, 2005.

[2] Leslie Hogben, Handbook of linear algebra, CRC Press, 2006.

[3] Carl D. Meyer, Matrix analysis and applied linear algebra, vol. 2, Siam, 2000.

[4] M. Tomforde, The structure of graph C ∗ -algebras and their generalizations, vol. 46,
Mathematics Subject Classification, 2006.

35