Math 228: Solving linear recurrence with eigenvectors

Mary Radcliffe

1 Example

I’ll begin these notes with an example of the eigenvalue-eigenvector technique used for solving linear
recurrence we outlined in class. Since all the recurrences in class had only two terms, I’ll do a three-term
recurrence here so you can see the similarity.
Let us consider the recurrence an = 2an−1 + an−2 − 2an−3 , subject to a0 = 2, a1 = 2, a2 = 4.

As in class, define the vector vn as  

an
vn =  an−1  .
an−2

Notice that
      
an+1 2an + an−1 − 2an−2 2 1 −2 an
vn+1 =  an  =  an = 1 0 0   an−1  .
an−1 an−1 0 1 0 an−2
 
2 1 −2
By defining A =  1 0 0 , we thus have that vn+1 = Avn . Therefore, since v2 is the first vector,
0 1 0
we have vn+1 = An−1 v2
Hence, we would like to represent vn in terms of the eigenvectors of A. If we are able to do so, then
we can calculate a formula for vn , and hence for an .

First, let’s calculate eigenvectors of A. Note that

 
2 − λ 1 −2
 1 −λ 0  = (2 − λ)(λ2 ) − 1(−λ) − 2(1) = −λ3 + 2λ2 + λ − 2 = (λ2 − 1)(2 − λ).
0 1 −λ

Hence, the eigenvalues of A are λ1 = 1, λ2 = −1, and λ3 = 2.

Next we calculate eigenvectors of A using standard row reduction techniques:

   
1 1 −2 1 1 −2
R2−R1
λ1 = 1 : A − λI =  1 −1 0  −−−−−→  0 −2 2 
0 1 −1 0 1 −1
 
R3− 1 R2,R2/2
1 1 −2
−−−−2−−−−−→  0 1 −1 
0 0 0
 
1 0 −1
R1−R2
−−−−−→  0 1 −1 
0 0 0

1
 Examining coefficients, we thus obtain that the first coordinate is the negative of the third,and the
1
second coordinate is also the negative of the third. Hence, the desired eigenvector is x1 =  1 .
1
Without detailing the computations for the remaining eigenvectors, we obtain
   
1 2
for λ2 = −1, x2 =  −1  ; for λ3 = 2, x3 =  2  .
1 1

We therefore wish to write the initial vector v2 in terms of x1 , x2 , x3 , as follows:

         
a2 4 1 1 2
v2 =  a1  =  2  = c1 x1 + c2 x2 + c3 x3 = c1  1  + c2  −1  + c3  2  .
a0 2 1 1 1

Hence, we consider the linear system c1 + c2 + 2c3 = 4, c1 − c2 + 2c3 = 2, c1 + c2 + c3 = 2. Without

detailing the algebra, we obtain the solution to this system is given by c1 = −1, c2 = 1, c3 = 2.
Therefore, we have

 
an
vn =  an−1  = An−2 v2
an−2
= An−2 (−x1 + x2 + 2x3 )
= −An−2 x1 + An−2 x2 + 2An−2 x3
= −(1)n−2 x1 + (−1)n−2 x2 + 2(2)n−2 x3
     
1 1 2
= −  1  + (−1)n  −1  + 2n−1  2 
1 1 1

Noting that the first term of the vector is indeed an , we thus obtain

an = −1 + (−1)n + 2n .

2 Theory

In general, this technique will work with any recurrence relation that takes the form

an = α1 an−1 + α2 an−2 + · · · + αk an−k + p(n),

where p(n) is a polynomial in n. We here sketch the theoretical underpinnings of the technique, in the
case that p(n) = 0.
Imagine a recurrence relation takin the form an = α1 an−1 + α2 an−2 + · · · + αk an−k , where the αi are
constants and the first k values of the sequence (an ) are known.
 
an
 an−1 
Write vn =  . Note that the first vector defined in this way will always be vk−1 =
 
..
 . 
an−k+1

2
 
ak−1
 ak−2 
, which is known by the initial conditions.
 
 ..
 . 
a0
As in the example above, our goal is to construct a matrix A so that Avn = vn+1 . We note that

   
an+1 αn an + α2 an−1 + · · · + αk an−k+1
 an   an 
vn+1 =  = ,
   
.. ..
 .   . 
an−k+2 an−k+2
 
α1 α2 α3 ··· αk−1 αk

 1 0 0 ... 0 0 

and hence we may take A = 
 0 1 0 ... 0 0 .

 .. 
 . 
0 0 0 ··· 1 0

Suppose that A is diagonalizable, and has the eigenvalue-eigenvector pairs (λ1 , x1 ), (λ2 , x2 ), . . . , (λk , xk ).
Write vk−1 , our known constant vector, as a linear combination of x1 , . . . , xk , in the following form:

vk−1 = c1 x1 + c2 x2 + · · · + ck xk .

Note that since A is diagonalizable, it is necessarily true that there exists constants c1 , . . . , ck such that
this equation holds. Moreover, as in the above example, we have that vn = An−k+1 vk−1 , and hence

vn = An−k+1 vk−1 = An−k+1 (c1 x1 + c2 x2 + · · · + ck xk ) = c1 λn−k+1

1 x1 + c2 λn−k+1
2 x2 + · · · + ck λn−k+1
k xk .

Noting that an is the first coordinate of vn , we can then read off the first coordinate of the vector to
obtain a formula for an .