Fibonacci and Related Sequences in Periodic Tridiagonal Matrices
Fibonacci and Related Sequences in Periodic Tridiagonal Matrices
Fibonacci and Related Sequences in Periodic Tridiagonal Matrices
1. INTRODUCTION
Tridiagonal matrices are matrices like
b1
a2
0
0
0
0
C1
b2
C2
33
h3
H
0
0
0
0
0
0
0
35
0
0
0
c4
h5
C5
36
b6 J
03
b4
0
0
0
and are made up of three diagonal sequences \a,-}, \bj\, \c,'f of real or complex numbers. They are of much use
in the numerical analysis of matrices. They also have interesting arithmetical properties being connected with the
theories of continued fractions, recurring sequences of the second order, and, in special cases, permutations, graph
theory, and partitions. We shall be considering two functions of such matrices, the determinant and the permanent.
By the permanent of the matrix
A=
Mx*
n,n(n)
( 1, 2,-,
n \
n: \ ir(1), n(2), - , irfn) )
Thus the definition of the permanent is simpler than the corresponding definition of the determinant in that no
distinction is made between odd and even permutations. In spite of this apparent simplicity, permanents are usually
much more difficult than determinants in their computation and manipulation. For tridiagonal matrices, however,
determinants and permanents are not very different. In fact we see that
per
\b1
c, "1
hih2+32Ci
and
per
b1
cj
32
0
b2
C2
33
h3
and, in general, the permanent of the tridiagonal matrix based on | a, I, < b; \, | c, I \s equal to the determinant
of the matrix based on { - a / f , \bj\f
| Cj\. Thus it is sufficient and simpler to consider the permanent function
of tridiagonal matrices. In fact we shall need only the method of expansion by minors in developing what follows.
2. STANDARDIZATION OF TRIDIAGONAL MATRICES
For our present purposes we make the assumption that the elements/? on the main diagonal are all different from
zero. It is therefore possible to divide the elements in each row by its main diagonal element. Thus we obtain a
matrix of the form
150
APR. 1975
r ;A
Ci
1
A3
0
0
0
0
0
0
(2)
Lo
0
0
0
0
0
c2
7
A4
0
0
c3
7
A5
0
151
0
0
0
0
c4
1
A6
c5
1
whose permanent (or determinant) is related to that of the original matrix (1) by the factor hfb2 OQ. Our next
step towards standardization is to observe that the permanent of (2) is not a function of A2 and C? but only of their
product A2C1. To see this, we expand the permanent by minors in the first column obtaining
per
7
A3
0
0
0
c2
1
A4
0
0
0
03
1
A5
0
0 0
0 0
c4 0
1 c5
A6 1
/ C3 0 0
A4 7 c 4 0
0 A5 7 c 5
0
0 A6 1
+ A2C1 per
By induction, therefore, the permanent of such a matrix as (2) will depend only on
.
A2C1,A3C2,-~,AnCn1
Hence, without loss of generality, we may assume that the C's are all equal to 1 and by an obvious change in notation define the standard tridiagonal matrix by
r / /
a
M = Mn = Mn(a1f
a2/ - ,
i i
O a2
0 0
an^)
/
a3
0
0
0
0
0
0
0
0
0 0 -
loo
10
0 .
0
/
7 .
/
7
0 ... &n~1 1
A = An =
We also adopt the conventions
(3)
An(a1ra2/-,an^)) = per
A0 = 1
and
Mn(a1/a2f
A.y = 0 .
1 BASIC PROPERTIES
We begin with the basic recurrence for
A n.
Theorem 1. If n > 7,
A n f c / / - / a f l . / j = An^(av-fan^2)
an^An.2(a1t-fa^3l
Proof. This follows at once by expanding An by minors of the elements of the last column of Mn(aj, , an^l
This recurrence is an efficient way of calculating successive A's when the a'% are given. It is clear from (4) that A,,
is linear in each of its independent variables 0,7, , a _ ; . For future use we give Table 1 of A,,. We observe from
this table that An is unaltered when its arguments are reversed. In general we have
Theorem2.
Proof.
An(an.1f
an-2/ -, a-f).
The theorem holds trivially for n = 0, 7, 2. If true for n - 7 amd n - 2, (4) becomes
An(a1,a2,-,an-i)
= A _ / (an-2,
~-.*i)
+ <*n-i&n-2(a>n-3,
But the right-hand side is the result of expanding the permanent of Mn(an-i,
of its first row. Hence the theorem is true for n and the induction is complete.
an-2/
-,
i>-
152
[APR.
Table 1
An(a7,a2,~',an1)
~~
-1
0
1
/
/
1 + CL-]
"
'
~ ~
7+
df+a2
5
a-, + Of 0,3 + &2CL4 + CL3&5 + &1GL4 + a<2a5
f=1
Since An is linear in each variable ay one can ask what are the functions Gj and Hj in
(5)
An(a7/
a^f)
= Gj + Hjdj
&n-l(ah"'*a>n-2h
n-1 =
an-3>-
n-2^p
Theorems. In (5),
Gj = An-j(aj+i,
- , ani)Aj(ai,
-,
- , an^)AH(a1t
Hj = An-H(ai+1,
aH)
- , ah2)
Proof. This can be proved by expanding An by minors of the elements of \Xsjth column and using Laplacian development of these minors. However, a simpler proof is afforded by the introduction of the following generalized
permanents A ^ r defined for K < r by
(6)
AK/r
= AKj(a1t
a2t -)
= AK(ar-.K+1f
arK+2,
- , aM)
= AK(ar-i,
ar2,
-,
ar^K+1).
In particular we have
Theorem 1 applied to these two equivalent definitions gives us the following useful relations.
(7)
&K,r = &K-1,r
(8)
AKtr
= A/c-//r-7
r-K+1&K-2,r
+ar-1&K-2,r-2-
A = AK/n AnK
+ anKAK
hn
An_K
In fact this is trivial when K = 0 by (3) and! (6) and when K= 1 it is a restatement of Theorem 1. To proceed inductively for K to K + 1 we note that
&n-K = &n-(K+1) + a>n~(K+1)&n~1-(K+1)
by Theorem 1. Substituting this into our induction hypothesis (9) we obtain
A n = &n-(K+1)\&Ktn+<ln-K&K-1,n\
But by (7) the quantity in the braces in A^+^n.
in (6) and (9) the theorem follows.
As a corollary we have
+an-(K+7)AK,nAn-7-(K+1)
1975]
da-
= A
H(al>
Theorem 4.
1+*^!}+ ?*) +...+ an--i\ An(aj,a2,-,an1)
1 \ J
\ 1
An,1(a2, a3,
-,an-i)
Proof. By Theorems 2 and 1 we may write
&n(ai,~',a>n-l) =
&(a-i,-*ai)
&n-l(a<2> "'* 0>n-l)
__ An^(an^1f
l
/An2
e = xn+y/5)= z + ^ l + jiU...
whose successive convergents
Hence
An(1, 1, / , - V 1) = Fn+1
a fact which follows at once from (4). Conversely as soon as we have developed other formulas like (10) we can
evaluate other continued fractions of Ramanujan type given in Theorem 4.
Uh = U(P,Q) =
where
P = a + b,
and
(11)
and
(12)
U0 = 0,
Q = ab
U1 = 1,
U2 = P
Uh = PU^-dUn
= Fn+1 .
154
[APR.
6. THE CASEp= 1
In this simple case we have
Theorem 5.
(13)
Proof.
An(ai)
Un+1(i,-ai).
By (12),
Un+1(1,-ai)
= Un(l,-a1)
+ a1Un..1(l,-a1)
But by (4),
An(a7)
= An-i(ai)
aiAn-2(cij)
An^(k)
= | ft + jT+~4o,)n - (1 - sJT+lof]
^ -
^TTla).
= 1
sin fan/3)
= \2n
An-r(2)
/ (2n
-(-!)"}/3.
as is easily verified.
7. THE CASEp = 2
This case is also relatively simple. We have
Theoremd
Proof.
An(ai,a2>
= (1 +a1 +a2>An.2(^u^2)"
a a
i 2^n-4(^v^2)
But
An.3
= An_2 - a2An4
Elimination of A _ 3 gives the theorem for n odd. If n is even, we simply interchange the roles of a / and 0,2.
The counterpart of Theorem 5 for/7 = 2 is
Theorem!.
A2n(aj,a2)
Proof
= Un+i(1
+ai+a2,aia2).
-aia2y\ln-2
with
W0 = 1,
Wi = A2(ai/a2)
= 1 + 01 .
But
Un+i(l + di +02, 0102) - 02Un(1 + 0,1+02,
0102)
enjoys the same recurrence and the same initial conditions. This proves the first part of the Theorem. The second
part is proved in the same way.
We note that, unlike A 2/7 fa/, 02), the function A2n+i(oi,
02) is symmetric in a ; and 02.
Examples of Theorem 7 are
1975]
= 2",
-V=
Fn+1,
A2n(j,
A2n-id,-i)
1) =
= a2*i/3
A4n(u,u )
Fn.i
(-1)"
= 1 - nu>
^-sm(*n/3)
A2n(i2)
2\j2
w
A2n(-j,
A2n(-Co,-oo2)
- <2^y-(2-^r_i
Here
= 2n
= %i (1 + (-1) ),
,co2) = n + 1,
A2n+1(-co
&2n(1,0)
-1) = Fn+2,
A2n+1(u,u )
A2n_l(i2>
= 2n~1,
A2n(b,1)
155
- iiijtr+u-jtr
*
-JLtJlL
^M*$*~*-
Inspection of the above examples shows them to behave exponentially, linearly or periodically as n -> .. This is a
general fact, true of periodic a's of any period length p.
8. THE GEWERAL PERIODIC CASE
We now take up the complicated general case of p > 3, although the theorems we are about to obtain hold for/7 =
1 and 2. For this purpose we enlarge the definition (6) of A^r to include the cases K > r. That is, we define for the
periodic case
AK/r(a1f
<r-K+2> - ,
r-lh
A5(a2,
A4j(ah
A3to(a1f
as) .
A4f1+a1A3ro
which for K = 5 and r = 2 is a particular case of (7). Formulas (7) and (8) are still true in general by Theorem 1.
let
A (p,s) = A P /s + as A p 2 ,s-;
B(p,s) = as(APfSAp2s-l
Then \in^s
&p-1,s&P-1,s-l)-
(modp),
An+p
= A(p,s)An - B(p,s)Anp
let n=ph +s. If in (9) we set K = p and use the fact that an+; = as+,- we get
Aph+S
= APfSAp(h^1)+s
+ asAp.1fSAp(h.1)+s.1
&p(h-1)+s-1
= &p-1,s-1&p(h-2)+s
+ aS&p-2,s-1Ap(h-2)+s-1
Beginning with (14) and continually applying (15) gives the following for A,,
] C a*M I V 2 ; - / )
11=1
&p(h-ii-l)+s
+&p-ua!1s\
&p-2j-l
\ h~1&s~l
150
[APR.
h-1
(16)
&p-1,s&p-1,s-1 J2
^{^p-Zs-^^&pfh-Ml+s
\i=1
= &ph+s - A p s &p (h- 1)+s ~ a-s^p- 1,s A s - / 1 Ap-2,5- 7 f
Next we multiply both sides of (16) by ds&p-2,s1
ana
"
to both sides. Sf we subtract this result from (16) when h is replaced by h + 1 we get
&p(h+1)+s~ &p,s&ph+s = as \&p-2,s-1Aph+s
~&p,s&p-2,s-1
+ &p-l,s-1&p-1,s&p(h-1)+s
&p(h-1)+s
B(p,s) = (-1)pa1a2"-ap
Theorem 9.
Proof.
(17)
where the subscripts on the a's are to be taken modulo/?, because then, by definition of B(p,s) we
B(p,s) = (-1)pasas-i
-as-p+1
(-1)pa1d2-ap.
To prove (17) we note that it holds for/? = 1 since the left member is 1 and the product of a's is vacuous. Assuming the result holds for/? and noting that (7) gives
and
1/S
Ap ;/S_ 7
&p-i,s&p-i,s-il
- , ap-2>
Theorem 11.
An+P
= PAn - QAnp
Theorem 12.
(20)
Ahp+S
= AsUh+1(PfQ)
+ (Ap+S-
PAs)Uh(P,Q).
19751
157
Proof. This relation holds for h = 0 and, since (JJ2(P,Q) - P for h - 1. By Theorems 11 and 12 both sides enjoy
the same recurrence. Hence they coincide.
9. MORE QM THE FUWCTIOWP
The function
P = Pp(ai,d2,
'".ap)
P1 = 1
P2 = 1+a-j + 0*2.
~',ap)
T 1
2
/ + a 1 + 0,2
1 + a 1 + 0,2 + a3
1 + a-i
+ a
+ a
5
5
+ a
+ a
1'a3
1+
^1
i=1
a a
i i+3
1=1
6
6
24
/ * i C aiai+2+^2
1=1
a a
1+
a +
aa
aa
i<j<6
1=1
j+2ai+4
i=1
Further entries in this table are left to the curiosity of the reader. It will be observed that the entries cease to be
symmetric functions of the a's with p = 4.
Um(l-D
= Fm.
-0-3 = ai(L2(l3
= 1.
This means that the three a's are the roots any cubic equation of the form
x3 + cx-
(21)
1 = 0.
0L2 = co,
a j = to
or some other permutation of these. For this case Theorem 12 gives the examples
&3h(l
&3h+i(l
co, co 2 j = Fh+1 2
co, co. j =
&3h+2(i^,<^2)
co2Fn
Fh+1+u2Fn
= 2Fh+1
Another special case is that of c = -2 in which the roots of (21) are - 1 and the two Fibonacci irrationals, for
example
a1
For this choice we get
= Q,
_
<i2 = 6,
CL3 =
-1.
158
A3h(9,9,-1)
APR. 1975
= Fh+2
A3h+1(Qj,-i)
Fh+1-0Fh
A3h+2(d,d,-1)
= (1 + MFh+1.
The reader may wish to write such formulas for other permutations of 6,0, - 1 .
For/7 = 4 our requirement becomes
(22)
Examples are
a ; = i,
a3 = -I,
a.4 = 1,
a 1CL2CI3CL4 = 1.
_
0*2 = 6, a3 = co ,
0,4 = 6.
a / = GO,
a2 = %(t + y/t2
a4 = 1/2(t -
a3 = /2(-t - y/W^Te),
+4e),
^/FTJe),
where t is any real or complex parameter and e = 1. In any case there are eight permutations of the four a's that
maintain (22). These are, in cycle notation
(1)(2)(3)(4),
(1)(3) (24)f
(13X2X41
(1234),
(1432).
With any one of these choices we have for An = An (a-j, 0*2, a3, 04)
A4h = Fh+1-~a4(1
A4h+1
=
A4h+2
= F
h+1
+a
+ a2)Fh
-a1a4Fh
~^1^2a4Fh
(1 l) h+1
A4h+3 = (1 + a1+a2)Fh+1
Instead of forcing An to involve the Fibonacci numbers we can make it a linear function of n by choosing/7 = 2 and
Q = 1 because Un (2,1) = n.
For/7 = 3 the conditions become
(23)
a<f +a2 + a3 = 1,
aja2a3
= -1.
One obvious solution is to choose two of the a's equal to 1 and the third - 1 . Thus we find
A3h(1,
1, -1) = 2h + 1,
A3h+1(1,-1,1h2h
A3h+1(j,
+ 1, A3h+2(J,-1,1)
1, - / ; = 7,
= 2h+2,
A3h+2(1,
1, -1) = 2h +2,
A3h(-1,1,j)=1,
A3h+J
A3h(1,
-1, 1) = 1,
(-1,1,1)= 1, A3h+2(-i,1,1)
= 0.
-2 cos (4-n/7),
-2 cos (6TT/7).
+ cx+ 1 = 0
= As+ (As+3 -
As)h.
The reader may have observed in the above that, of all the formulas for Ahp+S, the simplest is that for s = p - 1.
The reason for this phenomenon is to be seen by substituting s = -1 in Theorem 12. We obtain simply
Ahp^
Ap^Up(P,Q).