DELVING

deeper
Arthur T. Benjamin and Jennifer J. Quinn

Revisiting Fibonacci and

Related Sequences

W
e fully concur with Richard Askey’s Febru- dence between them. To apply this method, we
ary 2004 “Delving Deeper” column. Discov- need to know what the Fibonacci numbers count.
ering and proving identities containing Fi- We define the Fibonacci numbers by f0 = 1, f1 = 1,
bonacci numbers can be satisfying for students and and for n ≥ 2, fn = fn – 1 + fn – 2. Then the nth Fibonacci
teachers alike. His article touched on multiple strate- number counts the ways to tile a 1 × n board with
gies including induction, linear algebra, and a hefty squares and dominoes. Letting s represent a square,
dose of algebraic manipulation to derive many interest- which has length 1, and d represent a domino,
ing identities. However, a single method can be em- which has length 2, f3 = 3 counts the tilings
ployed to explain all of these identities more con-
cretely, leading to deeper understanding and intuition. sss, sd, ds,
We are referring to the method of combinatorial proof.
A combinatorial proof explains an identity by and f4 = 5 counts the tilings
counting—by counting a set in two different ways
or by counting two sets and providing a correspon- ssss, ssd, sds, dss, dd.

This department focuses on mathematics content that appeals to secondary school A tiling of length 5 (called a 5-tiling) can be created
teachers. It provides a forum that allows classroom teachers to share their mathematics by appending a square to a 4-tiling or a domino to a
from their work with students, their classroom investigations and projects, and their
3-tiling. So there are
other experiences. We encourage submissions that pose and solve a novel or interesting
mathematics problem, expand on connections among different mathematical topics,
present a general method for describing a mathematical notion or solving a class of prob- f5 = f4 + f3 = 8
lems, elaborate on new insights into familiar secondary school mathematics, or leave the
reader with a mathematical idea to expand. Send submissions to “Delving Deeper” by 5-tilings. In general an n-tiling can be created by ap-
accessing mt.msubmit.net.
pending a square to an (n – 1)-tiling or a domino to
Edited by Al Cuoco, [email protected] an (n – 2)-tiling. Hence the number of tilings will
Center for Mathematics Education, Newton, MA 02458 continue to grow like the Fibonacci numbers.
Now consider the Fibonacci numbers and the
E. Paul Goldenberg, [email protected]
squares of Fibonacci numbers given in table 1.
Education Development Center, Newton, MA 02458
Interestingly, the sum of the squares of two

Vol. 99, No. 5 • December 2005/January 2006 | MATHEMATICS TEACHER 357

Copyright © 2006 The National Council of Teachers of Mathematics, Inc. www.nctm.org. All rights reserved.
This material may not be copied or distributed electronically or in any other format without written permission from NCTM.
TABLE 1 (2) fm + n = fm fn + fm − 1 fn − 1 ,
n 0 1 2 3 4 5 6 7 8 9 10
fn 1 1 2 3 5 8 13 21 34 55 89 generalizing the Fibonacci pattern given in equa-
fn2 1 1 4 9 25 64 169 441 1156 3025 7921 tion (1).
Now consider the partial sum of the squares of
the Fibonacci numbers. For example:

Breakable in the middle 12 + 12 = 2 = 1• 2

1 +1 +2
2 2 2
= = 2•3
6
1 +1 +2 +3
2 2 2 2
= 15 = 3 • 5
f5 f5
12 + 12 + 22 + 32 + 52 = 40 = 5 • 8
Unbreakable in the middle
The pattern suggests that
f4 f4 f02 + f12 + L + fn2 = fn • fn + 1 .
Fig. 1 A 10-tiling that is breakable in the middle can be cut
into two tilings of length 5. A 10-tiling that is unbreakable To “see” why this is true, we count the ways to
in the middle must have a domino on cells 5 and 6; it can tile a board of length n and another board of length
be cut into a 4-tiling, a domino, and another 4-tiling. n + 1. Naturally, this can be done in fn fn + 1 ways. How
many of these tilings have their last common break,
called a fault, at cell k? From figure 2 we see the an-
swer is fk2 since after the last fault, there is just one
way to tile the two boards in a fault-free way. The
n
f4 k index k can be as small as 0, since both tilings have a
common fault before the first cell, and k can be as
large as n, when the second tiling ends with a square.
1 2 n+1
Summing over all values of k shows that there are
n
Fig. 2 There are fk2 tilings of an n-board and an (n + 1)-
board with the last common break occuring at cell k.
∑f k
2
tilings,
k=0

as desired.
consecutive Fibonacci numbers appears to be an- As an exercise, we invite the reader to show that
other Fibonacci number. More experimentation
suggests that for all n ≥ 1, f0 + f1 + L + fn = fn + 2 − 1

(1) fn2− 1 + fn2 = f2n . by counting the tilings of a board of length n + 2

with at least one domino, and considering the loca-
This can be seen directly by a combinatorial argu- tion of the last domino. Additionally, by tiling
ment that counts the set of tilings of length 2n in two boards of length 2n or 2n + 1 and considering the
different ways. On one hand, there are f2n such location of the last square, derive
tilings. On the other hand, either a 2n-tiling is break-
able in the middle or it is not, as illustrated in figure f0 + f2 + f 4 + L + f2n = f2n + 1
1. A breakable tiling can be created in f n2 ways. An
2
unbreakable tiling can be created in f n– 1 ways. Hence and
the set of 2n-tilings is also counted by the sum
f1 + f3 + f5 + L + f2n − 1 = f2n − 1.
fn2 + fn2− 1 .
Returning to table 1, we see that the square of a Fi-
However, there is nothing special about the middle bonacci number and the product of its neighboring Fi-
of an even length board. We can take this idea further bonacci numbers always differs by 1. For example, 52 –
by investigating the breakability after any cell. This 3 • 8 = 1, 82 – 5 • 13 = –1, and in general it appears that
time let’s count the tilings of length m + n in two dif-
ferent ways. There are certainly fm + n such tilings. But (3) fn2 − fn − 1 fn + 1 = ( −1)n .
now ask whether the tiling is breakable after cell m.
Here, a breakable tiling can be created in fm fn ways. To understand this identity, we illustrate that every
An unbreakable tiling has a domino covering cells m pair of n-tilings can be easily transformed into two
and m + 1; it can be created in fm –1 fn – 1 ways. So tilings of length n + 1 and length n – 1.

358 MATHEMATICS TEACHER | Vol. 99, No. 5 • December 2005/January 2006

 Suppose A and B are n-tilings where A covers named Tile 1, Tile 2, . . . , Tile n – k. Since there are
cells 1 through n and B covers cells 2 through n + 1,
as shown in figure 3. Suppose A and B have their ⎛ n − k⎞
last fault at cell k. Then A can be decomposed into a ⎜⎝ k ⎟⎠
k-tiling A1, followed by an n – k tiling A2. Similarly
B can be decomposed into a (k – 1)-tiling B1, fol- ways to choose k of these tiles to be dominoes, the
lowed by an (n – k + 1)-tiling B2. We exchange A2 identity follows.
and B2 (a process called tailswapping) to obtain the But it may come as a surprise to see other Fi-
tilings A+ = A1B2 and B – = B1A2, shown in figure 4, bonacci identities proved this way. Recall the Eu-
with length n + 1 and n – 1, respectively. Notice clidean algorithm for finding the greatest common
that tailswapping A+ and B – in figure 4 brings us divisor of two numbers. It is based on the fact that
back to A and B. Hence there are almost as many if n = qm + r, then
tiling pairs (A, B) where A and B have length n as
there are tiling pairs (C, D) where C has length n + gcd(n, m) = gcd(m, r),
1 and D has length n – 1. We say almost because
the tiling that consists of all dominoes is the only
A1 A2
one that is fault-free. Since (A, B) can be fault-free
only when n is even and (C, D) can be fault-free A:
only when n is odd, we understand why 1 2 n
n+1
fn2 − fn − 1 fn + 1 = ( −1)n .
B:

Delving deeper into the technique of tailswap- B1 k B2

ping easily allows us to discover and prove more
general identities. For example, if A and B are n- Fig. 3 Two n-tilings A and B, offset by one cell, with their
tilings with an offset of r (instead of 1) cells, then last fault occurring at cell k
tailswapping produces A+ and B – with respective
lengths n + r and n – r. Thus fn2 is almost fn – r fn + r . A1 B2
Their difference is the number of fault-free tilings.
The pair (A+, B –) can only be fault-free when n – r A: +

is even, B – has all dominoes, and A+ has all domi- 1 2 n+1

noes on cells r through n + 1. The remaining cells n
of A+ (1 through r – 1 and n + 2 through n + r) can B –:
be tiled f r2– 1 ways. On the other hand (A, B) has f r2– 1
B1 k A2
fault-free tilings precisely when n – r is odd. Can
you find them? Consequently
Fig. 4 Swapping the tails of the two n-tilings in figure 3,
fn2 − fn − r fn +rr = ( −1)n + 1− r fr2− 1 . i.e., the subtilings A2 and B2, creates an (n + 1)-tiling A+ and
an (n – 1)-tiling B–. This provides a bijection between offset
By changing the picture slightly and allowing tiling pairs with faults.
nonsymmetrical offsets as in figure 5 we can easily
“see” the following identity from Askey (2004): n-r odd:
fr – 1

fn + s fn − r − fn fn + s − r = ( −1)n − r fr − 1 f s − 1 . A:
1 2
Of course it is no surprise that combinatorial r r+1 n n+1 n+s
proofs can be given to identities that involve bino- B:
mial coefficients. For example, the “diagonal sum of n-r odd: fs – 1
Pascal’s Triangle” identity, fr – 1

n/2⎛ n − k⎞ A+:
∑ ⎜⎝ k ⎟⎠ = fn , 1 2 r r+1 n n+1
n+s
k=0

B –:
counts length n tilings by considering how many
dominoes are used. A length n tiling with k domi- Fig. 5 An n-tiling and an (n + s — r)-tiling with asymmetric offsets can use tail swap-
noes must have n – 2k squares and therefore has ping to establish the identity fn + sfn – r – fnfn + s – r = (–1)n – rfr – 1fs – 1. The tilings that cannot
k + (n – 2k) = n – k tiles, which from left to right are be swapped are pictured here.

Vol. 99, No. 5 • December 2005/January 2006 | MATHEMATICS TEACHER 359

 since gcd(n, m) is unchanged by subtracting a mul- FEuclidean FAlgorithm.
tiple of m from n. For example, if we always choose Consequently, to find gcd(F978, F96), FEuclid’s
q to be as large as possible, then we can compute FAlgorithm finds

gcd(978, 96) = gcd(96, 18) = gcd(18, 6) = gcd(6, 0) = 6. gcd( F978 , F96 ) = gcd( F96 , F18 ) = gcd( F18 , F6 )
= gcd( F6 , F0 ) = F6 ,
One of our favorite Fibonacci fun facts is best
described by the traditional definition of Fibonacci since F0 = 0. In general, if gcd(n, m) = g, then Eu-
numbers whereby F0 = 0, F1 = 1 and in general, for clid’s algorithm starts with gcd(n, m) and eventu-
n ≥ 0, Fn = fn– 1. Behold, ally reduces it to gcd( g, 0) = g. Thus FEuclid’s FAl-
gorithm begins with gcd(Fn, Fm) and eventually
(4) gcd(Fn, Fm) = Fgcd(n, m). reduces it to gcd(Fg, F0) = Fg. Thus for any n and m,
gcd(Fn, Fm) = Fgcd(n, m) as promised.
We have already proved special cases of this equa- We have only barely scratched the surface of
tion. For instance, equation (3) demonstrates that con- what is possible with combinatorial proofs. Going
secutive Fibonacci numbers are relatively prime since further, if we allow c1 colors for squares and c2 col-
any number that divides Fn and Fn + 1 must also divide ors for dominoes, we can combinatorially explain
±1. (This fact can also be proved by induction.) Thus: identities defined by initial conditions uj = 0 for
j < 0, u0 = 1 and by the recurrence un = c1un – 1 +
gcd(Fn + 1, Fn) = 1 = F1 = Fgcd(n + 1, n) c2un – 2, for n ≥ 1. Identities for kth order recurrences
un = c1un – 1 + … + ckun – k with the same initial condi-
Also observe that if n is a multiple of m, say n = tions are also easily explained by allowing colored
qm, then by equation (2): tiles with length at most k. Recurrences with other
initial conditions can be modeled by the same tiling
Fn = Fqm = fqm − 1 = f( m − 1)+( q − 1)m problems, but restrictions are placed on the initial
= fm − 1 f( q − 1)m + fm − 2 f( q − 1)m − 1 tile. For details, see Benjamin and Quinn (June
2003) or Benjamin and Quinn (2003). Finally, by
= Fm F( q − 1)m + 1 + Fm − 1 F( q − 1)m
looking at random tilings, even identities that in-
volve the golden ratio
Thus if F(q – 1)m is a multiple of Fm, then so is Fqm.
1+ 5
So by induction on q, we have that Fn = Fqm is a φ=
2
multiple of Fm. (For an inductionless proof, see
Benjamin and Quinn, June 2003, or Benjamin and such as Binet’s formula
Quinn 2003.) Therefore, if n = qm, then
1 ⎡ ⎤
Fn = ⎢
⎣
φ n − ( −1 / φ )n ⎥⎦
gcd(Fn, Fm) = gcd(Fqm, Fm) = Fm = Fgcd(n, m). 5

Now suppose that n = qm + r where r > 0. Then can be given a combinatorial explanation.
by equation (3), We end this article with a comment from
Richard Askey, who was quoted in the preface of
gcd( Fn , Fm ) = gcd( Fm , Fqm + r ) = gcd( Fm , fqm − 1+ r ) Boros and Moll (2004) as saying,
= gcd( Fm , fqm − 1 fr + fqm − 2 fr − 1 )
= gcd( Fm , Fqm Fr + 1 + Fqm − 1 Fr ) If things are nice, there is probably a good reason
why they are nice: and if you do not know at least
= gcd( Fm , Fqm − 1 Fr ) one reason for this good fortune, then you still have
work to do.
where the last step follows, since Fqm is a multiple
of Fm and thus Fqm Fr + 1 can be ignored when com- So the next time you see a nice identity like
puting the greatest common divisor with Fm. Also,
⎛ n − i⎞ ⎛ n − j⎞
we know that Fm has no factors in common with
Fqm – 1 (since Fm divides Fqm, which is relatively prime
∑ ∑ ⎜⎝ j ⎟⎠ ⎜⎝ i ⎟⎠ = f2 n + 1
i ≥0
0 j ≥0

to Fqm – 1) and therefore

or
gcd(Fn, Fm) = gcd(Fm, Fr) n

∑ ( n − k) f k
= fn +3 − (n + 3),
k=0
But wait. This is just the Euclidean algorithm with
Fs inserted on top of everything. Let’s call this the we hope you will ask yourself, “What is the under-

360 MATHEMATICS TEACHER | Vol. 99, No. 5 • December 2005/January 2006

lying combinatorial explanation?” REFERENCES ARTHUR BENJAMIN,
At least that’s what we are counting on. Askey, Richard A. “Fibonacci and Related [email protected], is
Sequences.” Mathematics Teacher 97 professor of mathematics
Editors’ notes: Arthur Benjamin and (February 2004): 116–19. at Harvey Mudd College,
Jennifer Quinn add one more approach Benjamin, Arthur T., and Jennifer J. Claremont, CA 91711.
to the growing storehouse of methods Quinn, “The Fibonacci Numbers— JENNIFER QUINN,
we have seen in “Delving Deeper” for Exposed More Discretely.” [email protected], is profes-
solving recurrence relations. This ap- Mathematics Magazine 76 (June 2003): sor of mathematics at
proach involves combinatorial proof, 182–92. Occidental College, Los Angeles, CA
what they call Proofs That Really Count ———. Proofs That Really Count: The Art 90041. Together, they are the editors
(Benjamin and Quinn 2003). We plan of Combinatorial Proof. Washington DC: of Math Horizons, published by the
an installment sometime next year to Mathematical Association of America, Mathematical Association of
summarize and catalogue the methods 2003. America. Their book Proofs That
that have appeared so far. Send us your Boros, George, and Victor H. Moll. Really Count: The Art of
favorites. Irresistible Integrals: Symbolics, Combinatorial Proof (MAA 2003)
Benjamin and Quinn include some Analysis, and Experiments in the received the Choice Award for
provocative questions. We have a few Evaluation of Integrals. Cambridge: Outstanding Academic Title.
more to add. The context of tiling a Cambridge University Press, 2004.
wall leads to many interesting counting Parker, Ruth. “What Will the Implementa-
problems. tion Take?” Mathematics Teacher 84
(September 1991): 442–58. ∞
• The authors describe how, if “triomi-
noes’’ (tiles of length 3) are allowed
in the tiling, the number of ways
there are to tile a length n wall satis-
fies the recurrence

un = un – 1 + un – 2 + un – 3

What are the initial conditions? What

properties do these numbers have?

• What if tiles of any length are al-

lowed in the tiling? How many ways
are there to tile a length n wall? Such
tilings are known as compositions of
n (see Benjamin and Quinn 2003 for
more details). Many teachers know
this as the Trains problem (Parker
1991): Count the number of trains
of length n that can be made from
Cuisinaire-like rods. There are many
related questions. For example:

(1) How many Cusineaire trains of

length n have exactly k cars?
(2) Suppose you lay out all the
Cusineaire trains of length n on
your desk. How many rods of
length k will there be on the desk?
(Connie Vann, a teacher in Dan-
vers, Massachusetts, came up with
this problem.)

There are many other questions one

could ask. Why not pick one and work
on it?

Vol. 99, No. 5 • December 2005/January 2006 | MATHEMATICS TEACHER 361