The Fibonacci-King Q Matrix at 50 (50)

Stuart Eugene Thiel

March 2010
1 Introduction
This year, 2010, marks the 50
th
birthday of the Fibonacci Q matrix, rst
discovered in 1960 by Charles King [3]. Although some, e.g. [2], aver that
it has been underappreciated, the Q matrix has certainly spawned a lively
interest in the application of matrix methods to Fibonacci research.
The present paper seeks to generalize Q beyond its original two dimen-
sions to n dimensions. One diculty presents itself immediately: the term
generalized Fibonacci matrix has already been assigned to describe any
matrix whose entries are all Fibonacci numbers (e.g., [5]). By contrast, this
article seeks to characterize a class of matrices whose entries can be any real
numbers, but which perform the essential functions, mutatis mutandis, of
the original two-dimensional Q. Comity makes it best simply to avoid the
claim that this constitutes a generalization of the Fibonacci Q matrix, and
dene by Q
n
an n n real matrix satisfying our desiderata, to be detailed
below. Within the text, such a matrix will often be referred to as Kings
Q, a term as yet unclaimed except perhaps by chess players and therefore
unambiguous.
To do the job thoroughly will require a side trip into the general theory
of linear Fibonacci identities, i.e., expressions with real coecients a
j
, j =
1, 2, . . . , n, satisfying

n
j=1
a
j
F
kn+j
= 0 for all Fibonacci numbers F
j
. The
excursion yields a simple recipe for generating linear Fibonacci identities,
one which turns out to be useful in characterizing the class of n n Kings
Q matrices.

M.S. Candidate, Northeastern Illinois University

1
1.1 Notation
Dene the k
th
Fibonacci number for any integer k by F
k
= F
k1
+F
k2
, F
1
=
F
2
= 1, to yield the famous Fibonacci sequence
. . . , 8, 5, 3, 2, 1, 1, 0, 1, 1, 2, 3, 5, 8, 13, 21, 34, . . .
Dene the n-vector of Fibonacci numbers
v
n,k
= (F
k
, F
k1
, F
k2
, . . . , F
kn+2
, F
kn+1
)
T
.
The superscript
T
denotes transpose. All vectors default to column vectors.
Note that the terms of v
n,k
are given in decreasing order. This convention
conforms to the 2 2 case analyzed by King.
2 Kings Q Matrices
2.1 Singular Kings Q Matrices
Kings Fibonacci Q matrix has the form
Q
2
=
_
1 1
1 0
_
, Q
1
2
=
_
1 0
1 1
_
Its essential property, easily veried, is Q
2
v
2,k
= v
2,k+1
and Q
1
2
v
2,k
= v
2,k1
for all integers k. Simple induction yields Q
m
2
v
2,k
= v
2,k+m
for all integers k
and m. Hence the denition:
Denition. An nn matrix is a Kings Q matrix of order n, denoted Q
n
,
if and only if it is nonsingular and satises Q
m
n
v
n,k
= v
n,k+m
for all integers
k and m, or it is either of the two singular matrices dened as
Q
n
=
_
a
n1
0
I
n1
Z
n1,1
_
, Q
n
=
_
Z
n1,1
I
n1
0 b
n1
_
,
where I
n
is the n n identity matrix, Z
i,j
is a matrix of zeroes, a
n
is a row
vector of two 1s followed by (n 2) zeroes, and b
n
is a row vector of (n 2)
2
zeroes followed by (1, 1). For example:
Q
7
=
_

_
1 1 0 0 0 0 0
1 0 0 0 0 0 0
0 1 0 0 0 0 0
0 0 1 0 0 0 0
0 0 0 1 0 0 0
0 0 0 0 1 0 0
0 0 0 0 0 1 0
_

_
, Q
7
=
_

_
0 1 0 0 0 0 0
0 0 1 0 0 0 0
0 0 0 1 0 0 0
0 0 0 0 1 0 0
0 0 0 0 0 1 0
0 0 0 0 0 0 1
0 0 0 0 0 1 1
_

_
Q
m
n
v
n,k
= v
n,k+m
and Q
m
n
v
n,k
= v
n,km
, but only for positive integers m,
because the matrices are singular. By inspection, each has (n 1) linearly
independent columns, and a zero column, so the rank of each is (n 1).
However, although the inverse matrices do not exist, it is easy to see how Q
n
and Q
n
complement one another with respect to the most signicant property
of Kings Q matrices: for any positive integers m
1
and m
2
, Q
m
1
n
Q
m
2
n
v
n,k
=
Q
m
1
n
v
n,km
2
= v
n,k+m
1
m
2
. For that reason, we bend the denition to admit
these matrices as Kings Q matrices. They are sui generis in this respect.
2.2 Nonsingular Kings Q Matrices
Denition. Any real n-vector
n
= (a
n
, a
n1
, . . . , a
1
)
T
, n 3, is a linear gen-
eralized Fibonacci identity (LGFI) if and only if a
T
n
v
n,k
=

n
j=1
a
j
F
kn+j
=
0 for all integers k.
(It is a generalized Fibonacci identity because it works equally well on
generalized Fibonacci sequences as on standard ones. A generalized Fi-
bonacci sequence {G
i
} obeys the Fibonacci recurrence relation G
k
= G
k1
+
G
k2
, but the initial conditions G
1
and G
2
can take any values. In partic-
ular, for present purposes, the generalized Fibonacci numbers need not be
integers. The well-known relation G
k
= G
1
F
k2
+G
2
F
k1
(e.g., [4, at 109]),
guarantees that if
T
n
v
n,k
= 0 for all k,
T
n
(G
n
, G
n1
, . . . , G
kn+1
)
T
= 0 as
well, and vice versa.)
Comment. No LGFI can be shorter than the 3-vector dening the Fi-
bonacci relation,
3
= (1, 1, 1)
T
(i.e., F
k
F
k1
F
k2
= 0). It will be
shown below that in a well-dened sense all LGFIs are linear combinations
of that dening relation.
3
Consider any nn matrix Y
n
, each of whose rows is an LGFI. Y
n
v
n,k
= 0
for all k. Let Q
n
= Q
n
+Y
n
. Then Q
n
v
n,k
= Q
n
v
n,k
+Y
n
v
n,k
= v
n,k+1
+ 0 =
v
n,k+1
. Conversely, suppose Q
n
v
n,k
= v
n,k+1
. Then (Q
n
Q
n
)v
n,k
= Q
n
v
n,k

Qv
n,k
= v
n,k+1
v
n,k+1
= 0, for all k, forcing the conclusion that that rows
of (Q
n
Q
n
) are all LGFIs. We have proved
Proposition. Every nonsingular Kings Q matrix of order n 3 admits
the decomposition Q
n
= Q
n
+ Y
n
for some nonzero matrix Y
n
whose rows
are all LGFIs. The matrix Q
n
+ Y
n
is a Kings Q matrix whenever it is
nonsingular.
Comment. Simple induction suces to show Q
m
n
v
n,k
= v
n,k+m
.
Now consider Y
n
, another n n matrix whose rows are LGFIs. (Q
n
+
Y
n
)v
n,k
= v
n,k1
. If (Q
n
+ Y
n
) is nonsingular, it is equal to

Q
1
n
for some
Kings Q matrix

Q
n
.
Consider any nonsingular nn matrix X
n
. The following section exhibits
and discusses an nn projection matrix, P
n
, having the property that P
n
X
n
is a matrix of LGFIs. Upon our return, we use that apparatus to prove that
every such matrix X
n
generates a Kings Q matrix.
3 Excursion: Linear Generalized Fibonacci
Identities
This section applies the algebra of orthogonal projection to the study of Fi-
bonacci identities to show, for n 3: every real n-vector can be decomposed
into an LGFI and a length-n segment of a generalized Fibonacci sequence;
there is a one-to-one correspondence between nonzero real vectors of length
n2 and LGFIs of length n; all LGFIs of length n take the form C
n
, where
is a real (n 2)-vector and the n (n 2) matrix C
n
is given by, e.g.,
C
7
=
_
_
_
_
_
_
_
_
_
_
1 0 0 0 0
1 1 0 0 0
1 1 1 0 0
0 1 1 1 0
0 0 1 1 1
0 0 0 1 1
0 0 0 0 1
_
_
_
_
_
_
_
_
_
_
4
The columns of C
n
are built by padding the fundamental LGFI (1, 1, 1)
T
with zeroes.
By denition,
n
is an LGFI if and only if
T
n
v
n,k
= 0 for all integers k,
and it clearly works just as easily for a segment of a generalized Fibonacci
sequence G
1
v
n,k2
+G
2
v
n,k1
. To restate, an LGFI
n
will map an n-vector
into zero if and only if the terms of that vector follow the Fibonacci recurrence
identity G
k
= G
k1
+G
k2
(by convention, in descending order). Such vectors
are referred to, generically, as GF segments, i.e., segments of length n from
some generalized Fibonacci sequence.
The strategy of proof is to apply the elementary algebra of orthogonal
projection so as to map any real n-vector onto the (n 2)-dimensional sub-
space spanned by the columns of C
n
. The algebra is thoroughly explained in
any of several excellent texts, including [1], [6], [7]. Proofs of results common
to most or all such texts are omitted. Throughout this discussion, n 3 is
a nite integer.
As an initial step, note by inspection that the rank of C
n
is n 2, so C
n
has full column rank. Therefore,
_
C
T
n
C
n
_
1
is well-dened and exists.
Dene the projection matrix P
n
= C
n
_
C
T
n
C
n
_
1
C
T
n
, and dene M
n
=
I
n
P
n
. Direct calculation conrms that P
n
and M
n
are symmetric matrices,
orthogonal to one another. They are idempotent matrices of rank n 2
and 2, respectively. Further, any real n-vector
n
can be decomposed into

n
= P
n

n
+M
n

n
, with (P
n

n
)
T
(M
n

n
) = 0.
Proposition. For any nonzero real
n
, P
n

n
is either an LGFI or the zero
vector.
proof. By construction, v
T
n,k
C
n
= 0, because the columns of C
n
are all
LGFIs. Thus
v
T
n,k
P
n

n
= v
T
n,k
C
n
_
C
T
n
C
n
_
1
C
T
n

n
=
_
v
T
n,k
C
n
_ _
C
T
n
C
n
_
1
C
T
n

n
= 0.
Proposition. For any nonzero real
n
, M
n

n
is either a GF segment or the
zero vector.
proof. Follows directly from (P
n

n
)
T
(M
n

n
) = 0.
Comment. Because
n
= P
n

n
+ M
n

n
, another way to state these two
propositions is, every nonzero real vector is either an LGFI or a GF segment,
or can be decomposed into the sum of an LGFI and a GF segment.
5
Proposition. For any nonzero real LGFI
n
, let

=
_
C
T
n
C
n
_
1
C
T
n

n
.
Then
n
= C
n

; i.e., there is a one-to-one correspondence between every

nonzero length-n LGFI
n
and every nonzero length-(n 2) real vector .
proof. Assume
n
is a nonzero LGFI. Apply the projection to obtain

n
= P
n

n
+ M
n

n
. P
n

n
= C
n

cannot be zero because

n
is not a GF
segment, so P
n

n
is an LGFI. It suces therefore to show that M
n

n
= 0.
Suppose, to the contrary, that M
n

n
= 0. If so, then M
n

n
must be a GF
segment, and because
n
and P
n

n
are both LGFIs, (M
n

n
)
T
(
n
P
n

n
) =
0. But, by hypothesis M
n

n
=
n
P
n

n
= 0, so (M
n

n
)
T
(
n
P
n

n
) is a
real sum of squares and must be strictly positive. To relieve the contradiction,
we must abandon the hypothesis that M
n

n
= 0. QED.
Corollary. There is a one-to-one correspondence between integer vectors
of length (n 2) and LGFIs of length n having integer coecients.
proof. Inspection of C
n
shows that if all the terms of are integers,
all the terms of
n
will be integers as well. The converse follows from the
identity

=
_
C
T
n
C
n
_
1
C
T
n

n
.
Comment. To eliminate the redundancy caused by scalar multiples, re-
strict attention to integer vectors for whose nonzero terms the greatest
common divisor is 1. To eliminate the redundancy caused when the rst
or last term of
n
is zero, e.g.,
T
n
v
n,k
= (0,
T
n
)v
n+1,k+1
=

n
j=1
a
j
F
kn+j
,
restrict attention to whose rst and last terms are nonzero.
Example. For n 4, let be the (n 2)-vector (1, 1, . . . , 1)
T
. C
n

has the form

n
= (1, 0, 1, 1, . . . , 1, 1, 2, 1)
T
(i.e., 1, 0, (n 4) consecutive
1s, 2, 1). By setting m = n2, we have proved Lucass summation identity

m
j=1
F
k+j
= F
k+m+2
F
k+2
(e.g., [4, at 69-71]), and shown its equivalence
to a particularly simple form of .
Comment. The immediate signicance of this proposition, beyond its
implications for Kings Q, is to show that LGFIs are ubiquitous, even though
only a relative few are catalogued.
4 Return to Kings Q
4.1 Existence of Kings Q Matrices: Proof
Proposition. Let C
n
and P
n
= C
n
_
C
T
n
C
n
_
1
C
T
n
be dened as in the pre-
vious section. Then, for any real n n matrix X
n
with at least one col-
6
umn not residing in the nullspace of P
n
, the rows of Y
n
= (P
n
X
n
)
T
are
LGFIs and
_
Q
n
+Y
n
_
v
n,k
= v
n,k+1
. Moreover, if X
n
is also nonsingular,
Q
n
=
_
Q
n
+Y
n
_
thus dened is nonsingular and is therefore a Kings Q
matrix.
proof. The rst assertion is obvious. The strategy of proof of the latter
assertion is to show, by contradiction, that the nullspace of Q
n
=
_
Q
n
+Y
n
_
consists only of the zero vector.
Suppose Q
n
w = 0 for some nonzero real n-vector w. Then
0 = Q
n
w = Q
n
w + (P
n
X
n
)
T
w = Q
n
w +X
T
n
P
n
w,
because P
n
is symmetric. Multiply through by
_
X
T
n
_
1
to obtain
_
X
T
n
_
1
Qw+
P
n
w = 0. But, for nonzero w, P
n
w is always an LGFI, so for any k,
v
T
n,k
_
X
T
n
_
1
Q
n
w +v
T
n,k
P
n
w = v
T
n,k
_
X
T
n
_
1
Q
n
w = 0.
It remains to show that v
T
n,k
_
X
T
n
_
1
Q
n
w cannot in fact equal zero for all
k, i.e., that
_
X
T
n
_
1
Q
n
w is not an LGFI.
_
X
T
n
_
1
Q
n
w can be an LGFI only
if
_
X
T
n
_
1
Q
n
is a projection matrix equivalent to P
n
. But that is impossible.
The rank of P
n
is n 2, while that of
_
X
T
n
_
1
Q
n
is n 1 because it is the
product of a nonsingular matrix and a matrix of rank n 1. QED.
4.2 Existence of Kings Q Matrices: by Example
Let X
n
be the identity matrix, so Y
n
= (P
n
X
n
)
T
= P
n
. For n = 5, the
resulting nonsingular matrix Q
n
= Q
n
+Y
n
is:
Q
5
=
_

_
11
8
5
8

1
4

1
8

1
8
5
8
17
24

1
12

5
24
1
8

1
4
11
12
5
6
1
12

1
4

1
8

5
24
13
12
17
24
3
8

1
8
1
8

1
4
11
8
3
8
_

_
Further, while the theorem shows that all nonsingular X
n
generate Kings
Q matrices, that characterization is incomplete. Singular X
n
can also gener-
7
ate nonsingular Kings Q, as shown by this example:
Q
5
=
_

_
1 1 0 0 0
0 1 1 0 0
0 0 1 1 0
0 0 0 1 1
0 0 0 1 0
_

_
, Y
5
= Q
5
Q
5
=
_

_
0 0 0 0 0
1 1 1 0 0
0 1 1 1 0
0 0 1 1 1
0 0 0 0 0
_

_
In this example, X
5
= Y
T
5
, which is singular, but Q
5
is nonsingular.
5 Conclusion
This focus of this work has been to broach the idea of n-dimensional Kings
Q matrices. It has succeeded, as usual, in opening more questions than it has
answered, in particular questions relating to the higher algebraic properties
of the set of such matrices, and to its symbiosis with the theory of Fibonacci
identities, linear and otherwise.
As ever, this is a work in progress.
References
[1] Goldberger, Arthur S. 1964. Econometric Theory.
[2] Johnson, R.C. 2007. Matrix Methods for Fibonacci and Related Se-
quences, www.dur.ac.uk/bob.johnson/bonacci/ (November).
[3] King, Charles. 1960. Some Properties of Fibonacci Numbers, M.S. The-
sis, San Jose State College, San Jose, CA.
[4] Koshy, Thomas. 2001. Fibonacci and Lucas Numbers with Applications.
[5] Lee, G.-Y. and J.-S. Kim. 2003. The Linear Algebra of the Generalized
Fibonacci Matrix. Fibonacci Quarterly (41:5), 451-465.
[6] Searle, Shayle. 1982. Matrix Algebra Useful for Statistics.
[7] Theil, Henri. 1971. Principles of Econometrics.
8