9-12 Solutions PDF

125
Chapter 9: Probability
SOLUTIONS FOR PART III

9. PROBABILITY
9.1. If A B , then P(A) P(B)TRUE. P(B) is the sum of the probabilities assigned to points in A plus the sum of the proabilities assigned to
points in B A.
9.2. If P(A) and P(B) are not zero, and P(A|B) = P(B|A), then P(A) =
P(B)TRUE. The two quantities equal P(A B)/P(B) and P(A B)/P(A),
so they are equal if and only if the denominators are equal.
9.3. If P(A) and P(B) are not zero, and P(A|B) = P(B|A), then A and B
are independentFALSE. If A and B are the same event, then P(A B) =
P(A) > P(A)P(B), and they are not independent.
9.4. If P(A) > 1/2 and P(B) > 1/2, then P(A B) > 0TRUE. If P(A
B) = 0, then A and B have no common points with positive probability, and
P(A B) = P(A) + P(B) > 1, which is impossible.
9.5. If A, B are independent, then A and B c are independentTRUE. P(A

B c ) = P(A) P(A B) = P(A) P(A)P(B) = P(A)(1 P(B)) = P(A)P(B c ).
9.6. If A, B are independent, then A c and B c are independentTRUE.

P(Ac B c ) = P(Ac ) P(Ac B) = 1 P(A) (P(B) P(A B)) =
1 P(A) P(B) + P(A)P(B) = (1 P(A))(1 P(B)) = P(A c )P(B c ).
9.7. An event and its complement are independent if and only if their probabilities are 1 and 0. The probability of the intersection of A and A c is 0. If
they are independent, then 0 = P(A A c ) = P(A)P(Ac ) = P(A)(1 P(A)).
This requires that P(A) or 1 P(A) is 0.
9.8. Restaurant menu items. Let S denote the event that what the man
orders is out of stock. Let A and F denote the events of ordering pasta
and fish, respectively. We are given that P(A) = P(F) = .5 and P(S|A) =
P(S|F) = .5. By direct computation, P(S) = P(S A) + P(S F) =
P(S|A)P(A) + P(S|F)P(F) = .5(.5) + .5(.5) = .5.
The answer makes sense, since the probability that the dish is out
of stock is .5 every day, no matter what he orders. Arbitrary values may
arise for the values P(A), P(F), P(S|A), and P(F|A), in which case the
conclusion is obtained by the computation above.
9.9. Conditional probability for drawing marbles. The contents of the containers are R R , B B , and R B . In choosing a random marble from a random
container, there are six equally likely outcomes, three R and three B . The
126
number of outcomes in which the selected ball and the other in its container are both black is two. The condition probability of the other ball
being black (event O ), given that the selected ball is black (event B ), is
computed as P(O|B) = P(O B)/P(B) = (1/3)/(1/2) = 2/3.
9.10. Rolling two dice, one red and one green. There are 36 equally likely
possible outcomes.
a) Given that the red die shows a six, the probability of double-sixes is
1/6. Given that the red die shows a six, the roll is double-sixes if and only
if the green die shows a six, which occurs with probability 1/6.
b) Given that at least one die shows a six, the probability of doublesixes may be 1/6 or 1/11, depending on how the information was obtained.
If we saw one die and observed that it showed a six, then the probability
that the other die also shows a six is 1/6.
If someone tells us after seeing both dice that at least one shows a six,
then we can compute the conditional probability. The event "at least one
six" ( B ) consists of 11 equally likely rolls, of which exactly one is "doublesix" ( A). The formal computation of conditional probability is P(A|B) =
P(A B)/P(B) = (1/36)/(11/36) = 1/11.
9.11. The TV game show problem. The prize must be behind the contestants door (event C ) or the other unopened door (event D ). We seek the
conditional probabilities of these two events, given what has happened so
far. Let A be the door opened by the host. For each door, the probability of
being correct in the game is 1/3. Given event C , the conditional probability
of opening A is 1/2. Given event D , the conditional probability of opening A is 1. Given that the prize is behind A, the conditional probability of
opening it is 0.
Thus the conditional probability of opening A, given the door chosen
by the contestant, is (1/3)(1/2) + (1/3)1 + (1/3)0 = 1/2. The probability
that C occurs and A is opened is (1/3)(1/2) = 1/6. The probability that
D occurs and A is opened is (1/3)(1) = 1/3. The probability that C occurs
given that A is opened is thus (1/6)/(1/2) = 1/3, and the probability that
D occurs given that A is opened is (1/3)/(1/2) = 2/3.
9.12. Bertrands Paradox. In the following probability models, let p be the
probability that
the length ofthe randomaly generated chord of the unit
circle exceeds 3. Note that 3 is the length of the sides of an equilateral
triangle inscribed in the circle.
a) If the endpoints of the chord are generated by two random spins on
the circumference of the circle, then p = 1/3. After spinning once, draw
the inscribed triangle
with a corner at the chosen point. The length of the
chord will exceed 3 if and only if the second chosen point lies between the
127
other two corners of the triangle on the circle. Since probability of lying in
an arc is proportional to its length, the probability of this is 1/3.
b) If the midpoint of the chord is generated by throwing a dart at the
circle, then p = 1/4. The chord is generated from its midpoint P by drawing
the chord through P that is perpendicular to the segment
from P to the
center of the circle. The midpoint of a chord of length 3 has distance 1/2
from the center of the circle. Thus the length exceeds 3 if and only if the
generated midpoint has distance less than 1/2 from the center. Since area
is proportional to the square of the radius and probability is proportional
to area, the circle of radius 1/2 has 1/4 of the probability.
c) A model where p = 1/2. We first pick the distance of the chord
from the center; then we choose a random chord
with that distance from
the center. The length of the chord exceeds 3 if and only if its distance
from the center is less than 1/2. We pick this distance at random from the
interval [0, 1].
first vote, the probability that A never trails in the remaining portion of
the election is the solution to the Ballot Problem for an election ending at
(a 1, b). We have computed this to be (a b)/a . Since both these events
a ab
must occur, the probability that A is always ahead is a+b
= ab
.
a
a+b
9.13.
Triangles on triples among n equally spaced points on a circle. The

n
triples
are equally like. Thus we count how many yield three equal
3
lengths, two equal lengths and one different length, and three different
lengths.
When 3|n , there are n/3 triples that form equilateral triangles; otherwise there are none. To generate two equal segments, we choose their
common point in n ways and then their common length in b(n 1)/2c ways.
When 3|n , we must subtract from n b(n 1)/2c the n/3 equilateral triangles. The resulting probabilities are listed below by the congruence class
of n modulo 6.
0
equilateral
isosceles
other
2
(n1)(n2)
3n8
(n1)(n2)
n4
n1
3
n2
n5
n2
3
n1
n4
n1
2
(n1)(n2)
3n5
(n1)(n2)
n5
n2
3
n1
n4
n1
3
n2
n5
n2
9.14. In Bertrands Ballot Problem with outcome (a, b) where a > b, the
probability that A is always ahead of B is ab
, and the probability that the
a+b
2b
score is tied at some point is a+b . The two probabilities must sum to 1 when
a > b. A fails to be always ahead if and only if B has at least as much
at some point. Since A winds up ahead, B has at least as much at some
point if and only if there is a tie at some point. Thus it suffices to find the
probability that A is always ahead.
If A is always ahead of B, then A must get the first vote, and A must
get at least as many votes as B in every initial segment of the remainder.
The probability that A gets the first vote is a/(a + b). Given that A gets the
128
9.15. a) If m 0s and n 1s are placed in some order around a circle, then

there are exactly m n positions such that every arc of the circle starting at
that position and moving clockwise contains more 0s than 1s. This holds
trivially when n = 1. For m n > 0, there is some pair of 1s with at
least one 0 between them. Let S be a set of two positions consisting of a 0
followed immediately by a 1.
Neither position in S is good; the 1 comes too soon. A position outside
S is good if and only if it is good in the smaller arrangement obtained by
deleting S . The number of good starting places thus equals the number of
good starting places in the smaller arrangement, which by the induction
hypothesis is m 1 (n
1) = m n .
1 2n
b) There are n+1
ballot lists of length 2n . By part (a), every arrangen
ment of n 1s and n + 1 0s can be cut in exactly one place to obtain a list
in which all initial segments have more 0s than 1s. The first element (just
after the cut) must be a 0. Deleting this 0 yields a ballot list, and the process is reversible. Thus the ballot lists and the cyclic arrangements of n
1s and n + 1 0s are equinumerous. Since n and n + 1 are relatively prime,
the number of these cyclic arrangements
(and ballot lists of length 2n ) is

2n+1
1
1 2n
exactly the Catalan number 2n+1
=
= Cn .
n+1 n
n+1
9.16. A conditional probability computation. We are given three independent random variables X 1 , X 2 , X 3 , each defined on [n ] and equally likely to
take on any value in [n ]. The variable X 1 + X 2 is less than 4 only for the
outcomes (1, 2), (1, 1), (2, 1), which together have probability 3/n 2 . Thus
P(X 1 + X 2 4) = 1 3/n 2 . The events X 1 + X 2 + X 3 6 and X 1 + X 2 4
both occur if and only if X 3 = 2 and (X 1 , X 2 ) {(3, 1), (2, 2), (1, 3)} or
X 3 = 1 and (X 1 , X 2 ) {(3, 1), (2, 2), (1, 3), (4, 1), (3, 2), (2, 3), (1, 4)}. Thus
the probability that both occur is 1n n32 + 1n n72 = 10
. The desired conditional
n3
probability is
10 n 2 3
/ n2
n3
10
.
n 2 3
The computation is valid only for n 4.
9.17. Application of Bayes Theorem. Against her four opponents, the tennis player wins with probability .6, .5, .45, .4, respectively. These are the
conditional probabilities P(A|Bi ), where A is the event that she wins and Bi
is the event that she plays the i th opponent. We want to compute P(Bi |A)
for each i . Since she plays 30 matches against each of the first two and 20
matches against each of the last two opponents, we have P(Bi ) = .3, .3, .2,
.2, respectively. Thus we have all the data to compute The denominator,
which equals P(A), is .6 .3 + .5 .3 + .45 .2 + .4 .2 = .5. Thus the answers
are P(A|Bi )P(Bi ) 2, which equal .36, .30, .18, .16, respectively.
129
9.18. If half the females and one-third of the males in a class are smokers, and two-thirds of the students are male, then of the smokers are female.
Consider choosing a random student in the class. Let A be the event of
being a smoker. Let B1 , B2 be the events of being female or male, respectively. We seek P(B1 |A). We are given P(A|B1 ) = 1/2, P(A|B2 ) = 1/3, and
P(B2 ) = 2/3. By Bayes Formula, we compute
3
ai bi
(1/2)(1/3)
P(B1 + A) =
=
= .
6a j b j
(1/2)(1/3) + (1/3)(2/3)
7
arbitrary permutation of the others, decide which of A and B comes first,

and insert A and B so they will be separated by k . There are n + 1 k ways
to make the insertion. Thus the probability is n!2(n+1k)
, which simplifies
(n+2)!
9.19. Simpsons paradox - it is possible for A to have a higher batting average than B in day games and night games but for B to have a higher
average overall. If the daytime hits/attempts are a/b and c/d , respectively,
and the nighttime hits/attempts are w/x and y/z , respectively, then inequalities characterizing the situation described are a/b > c/d , w/x > y/z ,
and (c + y)/(d + z) > (a + w)/(b + x). The situation described occurs if and
only if these three inequalities are satisfied by eight integers also satisfying
0 a b, 0 c d , 0 w x , 0 y z . Equivalently, the three inequalities can be written as ad > bc , wz > x y , and (c + y)(b + x) > (a + w)(d + z).
Note that if b = d and x = z , then the inequalities lead to a contradiction and the paradox cannot occur. This suggests that one way in which
the paradox can occur is if the poor daytime performance of B and the good
nighttime performance of A are unimportant to their season averages, as
would occur if B hardly ever plays in daylight and A hardly ever plays at
night. For example, the paradox occurs with the numbers (a, b) = (3, 10)
and (c, d) = (0, 1) for daytime and (w, x) = (1, 1) and (y, z) = (7, 10) for
nighttime. For the season, the averages are 4/11 and 7/11.
9.20. Simpsons Paradox with university faculties of equal size.
Asst. Profs.
Assoc. Profs.
Full Profs.
Total
Univ. H
Total
40
30
30
100
Univ. Y
Women
25
5
5
35
Fraction
.6
.17
.17
.35
Total
80
10
10
100
Women
40
0
0
40
Fraction
.5
0
0
.4
9.21. A perfect game in bowling. If on each roll the probability of a strike

is p , then the probability that 12 consecutive rolls are strikes is p 12 . The
value of p such that p 12 = .01 is approximately .60. This suggests that
perfect games are quite rare but not unheard of among very good bowlers.
9.22. Probabilities for people in a line. We have A, B , and n other people
in random order, meaning that all (n + 2)! permuations are equally likely.
To form a permutation with exactly k people between A and B , form an
to
130
2(n+1k)
.
(n+1)(n+2)
To check that these sum to 1, we compute

Pn
(n+1)n
2
2k
2
(n+1)(n+2)
=
= 2(n+1)
k=0 n+2
(n+1)(n+2)
n+2
2
2n+2n
n+2
=1
9.23. Probability of first player getting first Head. The probability of heads
is p on each flip. If x is the probability that the first player wins, then x is
the sum of the probability of winning on the first flip and the probability
of winning later. Winning later requires tails from each player on the first
round, but then the rest of the game is the same as the original game.
Thus x = p + (1 p)2 x , which yields x = p/[1 (1 p)2 ] = 1/(2 p).
When the coin is fair, the first player wins with probability 2/3. (Note: the
probability
that the first player wins on the i th try is p(1 p)2 (i 1). Thus
P
also x = i=1 p(1 p)2 (i 1).
9.24. A spinner with regions 1, 2 . . . , n and payoff 2k for region k .

a) the expected payoff per spin of the dial is (2n+1 2)/n
P . The payoff
k
is 2 with probability 1/n , so the expected payoff is (1/n) nk=1 2k , which
equals (2n+1 2)/n .
b) Given a coin flip to move to a neighboring region, the gambler should
accept the gamble unless the current region is n . If the gambler already has
the maximum payoff, then switching is a guaranteed loss. If the current
payoff is the minimum, then switching is a guaranteed gain, with expected
payoff 12 2n + 12 4. When 2 k n 1, the expected payoff for switching is
1 k+1
2
+ 12 2k1 . This equals 2k + 2k2 and exceeds 2k .
2
The expected payoff under the optimal strategy is
!
n1
X

1 n+1
1
2n1 + 2 + 2n +
(2k + 2k2 ) =
2
+ 2n1 + 2n2 4
n
n
k=2
9.25. Another gambler. The amounts in envelopes are a1 a2 an . For

1 i n 1 the probability is pi that the envelopes contain ai and ai+1
dollars. He opens one envelope and has the option to switch. If he sees a 1
dollars, then he switches for a guaranteed gain. If he sees an dollars, then
switching would be a guaranteed loss.
For 2 k n 1, when he sees ak the other envelope may contain ak1
or ak+1 . Since he opened one of the two envelopes at random, the probability that he sees ak and the other is ak1 is pk1 /2, and the probability that
he sees ak and the other is ak+1 is pk /2. The probability that he sees ak is
131
the sum of these, ( pk1 + pk )/2. The conditional probability that the other
is ak1 is pk1 /( pk1 + pk ), and the conditional probability that the other is
ak+1 is pk /( pk1 + pk ).
Thus the expected payoff for switching when he sees ak is
and E(X i ) = P(X i, j = 1) =
ak1 pk1 + ak+1 pk

.
pk1 + pk
He should accept the gamble to switch if this quantity exceeds a k .
9.26.
Pn If X is a random variable that takes values only in [n ], then E(X ) =
k=1 P(X k). Let pi = P(X = i). We compute
E(X ) =
n
X
i=1
i pi =
n X
i
X
i=1 k=1
pi =
n X
n
X
k=1 i=k
pi =
n
X
k=1
P(X k).
Essentially, we add up in two ways the probability pi for all pairs (k, i) such
that 1 k i n .
9.27. Expected number of selections to find the desired key.

a) Under random order without replacement, the expectation is (n +
1)/2. All permutations of the keys are equally likely for the selection order.
Thus the right key is equally likely to be in each position. With probability
P
1/n that the key is in position k , the expected position is (1/n) n1=i k =
(n + 1)/2.
b) Under random selection with replacement, the expectation is n . On
each selection, the probability of success is 1/n . Thus we seek the expectation for the geometric distribution with success probability 1/n . By
Proposition 9.29, the expectation is n .
9.28. The expected numberof matching pairs among k socks pulled from a
pile of n pairs of socks is 2k /(2n 1). In each proof below, we express the
number X of matching pairs as a sum of indicator variables, and then we
use linearity of the expectation. In proof 1, we extract a random k -set of
socks. In proofs 2 and 3, we assume that the k socks are extracted in some
order. This produces every k -set in k ! ways, so the results are the same
when we make the ordered selections equally likely as when we make the
k -sets equally likely.

Proof 1. Let X i be 1 if pair i appears; otherwise X i = 0. Of the 2n
k

equally likely choices for the k socks, 2n2
include pair i . Hence
k2
2n 2
P(X i = 1) =
k2
,
2n
k
k(k 1)
,
2n(2n 1)
132
Since X =
Pn
X i , and since the vari

ables X 1 , . . . , X n have the same distribution, E(X ) = n E(X i ) = 2k /(2n 1).
Proof 2. Let X i, j be 1 if the i th and j th socks match; otherwise X i, j =

0. Any two socks are equally likely to be chosen i th and j th. Of the 2n
2

choices of two socks, n are matched pairs. Hence P(X i, j = 1) = n/ 2n
=
2
1/(2n 1), and E(X i, j ) = P(X i, j = 1) = 1/(2n 1). Since X is the sum of

k
such variables, E(X ) = 2k /(2n 1).
2
Proof 3. Let X j be 1 if the j th sock extracted completes a pair with
P
an earlier sock, so kj=1 X j counts the completed pairs of socks. The probability that the mate of the j th sock appears among the first j 1 socks is
( j 1)/(2n 1). Thus E(X j ) = ( j 1)/(2n 1), and

k
k
X
k
1 X
E(X ) =
E(X j ) =
( j 1) =
/(2n 1).
2n 1
2
j=1
k(k1)
.
2n(2n1)
i=1
j=1
9.29. Given a set of n men and n women, the expected number of malefemale couples in a random pairing is n 2 /(2n 1). Let X be the random
variable giving the number of male-female couples. We express X as the
sum of n 2 indicator variables. Define X i, j by X i, j = 1 when the i th man
is paired with the j th woman, X i, j = 0 otherwise. By the linearity of
expectation, E(X ) = n 2 E(X i, j ).
By symmetry, an individual is equally likely to be paired with any one
of the other 2n 1. Thus E(X i, j ) = 1/(2n 1), and E(X ) = n 2 /(2n 1).
9.30. There are 34,650 arrangements of the letters in M I SS I SS I P P I .
Among the 11 letters,
there are four Is, four Ss, two Ps, and one M. Thus

11
11!
we compute 4,4,2,1
, which equals 4!4!2!1!
9.31. Solution 9.40 overestimates the probability of hitting for the cycle.
Not all the appearances of a batter count as an at-bat. Other outcomes are
possible; walks and sacrifices are not counted as at-bats.
9.32. A polynomial p such that p(n) = 3n for n = 0, 1, 2, 3, 4. The multinomial expansion of (1 + 1 + 1)n is a sum of multinomial coefficients, since
always 1ki = 1. We write the coefficients as
n
k1 , k2 , k3
n(n 1) (n k1 k2 + 1)
n!
=
k1 !k2 !(n k1 k2 )!
k 1 !k 2 !
For each choice of k1 and k2 , this expression is a polynomial in n . If

we sum all such expressions that are nonzero when 0 n 4, then the
value at such n will be 3n . There are 15 nonnegative integer solutions to
k1 + k2 4. Coefficients with k 1 + k + 2 = r can be expressed as multiples
133
of n(n 1) (n r + 1). Summing these 15 multinomial coefficients yields

the desired polynomial as
9.36. Bijection from ballot paths to nondecreasing functions f : [n ] [n ]

such that f (i) i for all i . Given a ballot path P , define f by letting f (i)
be one more than the vertical coordinate where P steps from horizontal
coordinate i 1 to horizontal coordinate i . Since P cannot step down and
cannot be above height i 1 when the horizontal coordinate is i 1, f is
nondecreasing and satisfies f (i) i for all i . Given such a function f ,
the only path P that yields f by this map is the path that includes the
horizontal steps from (i 1, f (i) 1) to (i, f (i) 1) and the vertical edges
from (i, f (i) 1) to (i, f (i + 1) 1) for all i . Hence this map is a bijection.
1 + 2n + 2n(n 1) + 43 n(n 1)(n 2) + 23 n(n 1)(n 2)(n 3).

Note that evaluating at n = 0, 1, 2, 3, 4 yields 1, 2, 3, 9, 27, 81, respectively.
9.33. Random monomials in k variables with total degree n . The possible
lists of exponents for the k monomials are the nonnegative integer solutions

to e1 + + ek = n . The number of these (see Theorem 5.23) is n+k1
.
k1
a) The probability that all k variables have positive exponent in the
Qk1 ni
chosen monomial is i=1
. The outcomes in which all k variables have
n+ki
positive exponent correspond
to
the positive integer solutions
to e1 + +

n+k1
of
these.
Dividng
by
and
canceling
like
ek = n . There are n1
k1
k1
factors yields the probability.
b) When (n, k) = (10, 4), the probability that the exponents are different
60
is 143
. Four distinct values summing to 10 can be {4, 3, 2, 1} or {5, 3, 2, 0}
or {5, 4, 1, 0} or {6, 3, 1, 0} or {7, 2, 1, 0}. Each such set can be assigned
to

variables in 24 ways. Hence the probability of this event is 120/ 13
,
which
3
60
. (Here 0 is allowed as an exponent.)
simplifies to 143
9.34. Rolling six dice, each having three red faces, two green faces, and one
blue face: the probability of getting the same totals over six dice is 5/36.
The six physical dice correspond to six positions in an arrangement. The
number of waysof getting three red, two green, one blue is the multinomial
6
6!
= 621
= 60. On each die the probabilities of red, green,
coefficient 3,2,1
blue are 1/2, 1/3, 1/6, respectively. Thus the probability of each of these arrangements is (1/2)3 (1/3)2 (1/6)1 , and the probability of the specified event
60
5
= 36
.
is 896
9.35. The coefficient of x 4 y 16 in the expansion of (x + x y + y)16 is 1820. Each
x + x y + y factor contributes 1 to the exponent of x or 1 to the exponent of y
or 1 to each exponent. Since the total degree of x 4 y 16 is 20 and we have 16
factors, we must choose the term x y four times. Hence we choose the term
x not at all and the term y 12 times. The coefficient is the number of ways
16
we can arrange these choices, 0,4,12
. This equals 16 15 14 13/(4 3 2 1).
The coefficient of x 4 y 8 in the expansion of (x 2 + x y + y 2 )6 is 90. Each

x + x y + y factor contributes 2 to the exponent of x or 2 to the exponent of y
or 1 to each exponent. We can obtain contributions to the coefficient of x 4 y 8
by choosing x y from none, two, or four factors (taking x y an odd number
of times yields odd total exponents). The contributions in the three cases

6
6
6
are the numbers of ways to arrange the choices: 2,0,4
, 1,2,3
, and 0,4,2
,
respectively. These have values 15, 60, and 15, respectively.
134
9.37. The number

of lattice paths of length 2n that never step above the

diagonal is 2n
.
Each
such lattice path ends at a point (k, 2n k) with k
n
that do not step
n . By Solution 9.10, the number

of paths ending at (a, b)

P2n 2n
a+b
2n
above the diagonal is a+b

.
Hence
the
answer
is
k=n [ k k+1 ].
a
a+1

The sum telescopes, leaving only + 2n
at the beginning and 0 at the end.
n
9.38. Bijections involving ballot lists of length 2n .

a) From the set S of nonnegative integer lists a1 , . . . , a2n+1 such that consecutive entries differ by 1 and a1 = a2n+1 = 0. Given a ballot list b1 , . . . , b2n ,
let ai be the excess of the number of ones over the number of zeros among
the first i 1 terms of b. The resulting 2n + 1-tuple a is in S . The map is
easy to invert. Given a S , let bi = 1 if ai+1 > ai , and bi = 0 if ai+1 < ai .
b) From the set of arrangements of 2n people in 2 rows of length n so that
heights are increasing in each row and column. Given a ballot list with n As
and n B s, let the positions correspond to the 2n people in increasing order
of height. If the i th position in the list is the j th A, place the i th person in
the j th position in the first row. If the i th position in the list is the j th B ,
place the i th person in the j th position in the second row. Since the people
are in increasing order in the original list, the rows of the new arrangement
are increasing in height. Since the j th A in a ballot list occurs before the
j th B , the columns of the new arrangement are increasing in height. To
obtain the unique ballot list from which a particular arrangement comes,
consider the people in increasing order of height, and let position i of the
list be an A if and only if the i th person is in the first row.
9.39. Bijection from ballot lists of length 2n to non-crossing pairings of 2n
points on a circle. Given a non-crossing pairing of 2n points p1 , . . . , p2n in
order on the circle, record in position i a 1 if pi is paired with a higher point,
a 0 if pi is paired with a lower point. For each pair of points, the resulting
1 appears in the list before the resulting 0. Hence every initial segment
has at least as many 1s as 0s, and the list is a ballot list.
135
Chapter 10: Two Principles of Counting
Given any ballot list of length 2n , we can invert the map as follows:
When we reach a 1 in position i , we leave pi unmatched. When we reach a
0 in position i , we match pi to the point corresponding to the most recent
unmatched 1, which exists since we have seen at least as many 1s as 0s.
No crossing occurs, because when the second point of the first pair in a
crossing is reached, it would have been matched instead to the first point
in the second pair, since that corresponded to a more recent unmatched 1.
in the interval [0, 1]. Hence the row player guarantees the best return by
choosing x to make these equal; this maximizes the minimum. He chooses
x = (c d)/(b d a + c); the denominator is non-zero. Using this in the
bcad
expression for the expected return in either column yields bda+c
. This is
the maximum that the row player can guarantee. Similarly, the column
player can guarantee losing no more than this.
9.40. The Finger Game (see Example 8.20).

a) values of x that guarantee a positive expectation for B no matter what
A does. If B plays one finger with probability x , then the expected payoff
to B when A plays one finger is 3 5x dollars per game. When A plays
two fingers, the expectation is 7x 4. To guarantee positive expectation,
B chooses x so that 3 5x > 0 and 7x 4 > 0. This requires x < 3/5 and
x > 4/7, so the solution is the interval 4/7 < x < 3/5 (not very big).
b) Probability of winning. In the solution that optimizes the amount
won, each player plays one finger with probability 7/12. Since B wins
when the fingers differ, the probability that B wins in this situation is
(7/12)(5/12) + (5/12)(7/12) = 70/144. In other words, B wins slightly less
than half of the games, but loses a small enough amount when losing that
B comes out ahead in money in the long run.
10. TWO PRINCIPLES OF COUNTING
9.41. Paying each option with equalprobability is optimal for both players
in the matrix game with payoffs ac db to the row player if and only if a = d
and b = c. If the column player (B) plays each column with probability
1/2, then the expected payoff to the row player (A) is half the row sum in
each row. Thus playing the options with equal probability can be optimal
for both players only if the rows have the same sum. By the analogous
argument, the columns must also have the same sum.
Thus d must equal both a + b c and a + c b. This yields b = c.
Similarly, c = b + a d = b + d a yields a = c. For every such matrix,
the equal strategy guarantees that A has expected payoff at least (a + b)/2
and that B has expected loss at most (a + b)/2.

9.42. The value of the game with payoff matrix ac db . If any entry e is a row
minimum and a column maximum (a saddlepoint), then the row player
will play that row, guaranteeing winning at least e, and the column player
will play that column, guaranteeing losing at most e. Since the column
player cannot guarantee losing less than the row player can guarantee
winning, these strategies are optimal, and the value of the game is e.
By playing row 1 with probability x , the row player guarantees receiving the minimum of ax + c(1 x) = c + (a c)x and bx + d(1 x) =
d + (b d)x . If the matrix has no saddlepoint, then a c and b d have
opposite signs and the graphs of the two linear functions cross at some x
136
10.1. During a major league baseball season in which there are 140,000
at-bats and 35,000 hits,
a) Some player hits exactly .250FALSE. It may happen that there
are 1000 players, each with 140 at-bats, half of them having 28 hits each
(batting .200) and the other half having 42 hits each (batting .300).
b) Some player hits at least .250TRUE. If each hitter has more than
four times as many at-bats as hits, then the total number of at-bats would
be more than four times the total number of hits, which would contradict
the given data.
c) Some player hits at most .250TRUE. If each hitter has less than
four times as many at-bats as hits, then the total number of at-bats would
be less than four times the total number of hits, which would contradict
the given data.
10.2. Distinct three-person
committees over 11 years. If a set of size n has
11 distinct triples, i.e. 3n 11, then n 6.
10.3. If S is a (2n + 1)-subset of {1, 2, . . . , 3n}, then S contains three consecutive numbers, and this is not forced for a 2n -subset. Partition [3n ] into
n classes, with the i th class consisting of the numbers {3i 2, 3i 1, 3i}.
Since S has 2n + 1 distinct numbers, by the pigeonhole principle it must
have 3 numbers from some class. These three numbers are consecutive.
On the other hand, the set that omits all the numbers divisible by three has
2n numbers but no three consecutive numbers (it has the first two numbers
from each class).
10.4. Every set of n + 1 numbers in [2n ] contains a pair of relatively prime
numbers. Any pair of consecutive numbers is relatively prime, since an
integer greater than 1 cannot divide two consecutive integers. Hence it
suffices to partition the set [2n ] into the n pairs of the form (2i 1, 2i).
Since there are only n of these pairs, the pigeonhole principle guarantees
that a set of n + 1 numbers in [2n ] must use two from some pair, and these
are relatively prime.
137
The result is best possible, because in the set of n even numbers in

[2n ], every pair has a common factor. Note that since each pair of even
numbers is not relatively prime, a solution to the problem by partitioning
[2n ] into n classes and applying the pigeonhole principle must put the n
even numbers into n different classes.
10.8. Given five points in a square

with sides of length 1, the distance between some pair of them is at most 2/2. Express the square as the union
of four subsquares with sides of length 1/2. The distance
between any
two points in a square with sides of length is at most 2/2. Therefore,
every configuration that has two points in one subsquare has the desired
property. Since we have only four subsquares (classes), the pigeonhole
principle guarantees that the five given
points have two in some class, and
the distance between these
is
at
most
2/2.
The upper bound 2/2 for the minimum distance is best possible, because the configuration with points at the four corners
and one point in the
center has no pair of points closer together than 2/2.
10.5. Every set of seven distinct integers contains a pair whose sum or difference is a multiple of 10. If two numbers in our set of seven are in the
same congruence class modulo 10, then we have a pair whose difference is
a multiple of 10. If one is congruent to i and another is congruence to 10 i
mod 10, then we have a pair whose sum is a multiple of 10.
We therefore define six sets, each of which is a congruence class or a
pair of congruence classes that sum to 0. These sets are 0, 1 9, 2 8, 3
7, 4 6, 5. These sets partition the set of integers. Whenever we choose
seven integers, the pigeonhole principle guarantees that two must lie in
the same set, and then their sum or difference is a multiple of 10.
10.6. If 1 through 10 appear in some order around a circle, then some three
consecutive numbers sum to at least 17. Each number appears in three
P10
such triples, so the total of the sums of triples is 3 i=1
= 165.
i = 3 11
2
When 165 units fall into ten triples, at least one triple receives at least 17.
10.7. Numbers 1 through 12 on a clock face.
For every arrangement of the numbers 1 through 12 in a circle, there
are three consecutive numbers with sum at least 20. Given an arbitrary
arrangement, partition the 12 numbers into 4 consecutive triples. The 78
P12
units in the sum i=1
i = 13 12/2 are distributed into these four classes,
so some class receives at least the average of 78/4 = 19.5 units. Since each
class receives an integer amount, some class receives at least 20.
There are five consecutive numbers with sum at least 33. We cannot
partition 12 objects into sets of size five. Nevertheless, each number belongs to five sets of five consecutive numbers. Thus the twelve sets of
P12
five consecutive numbers have sums that total to 5 i=1
i = 390, and the
largest of these sums must be at least d390/12e = 33.
The sets of three consecutive elements cannot all have sum 19 or 20. If
there is such an arrangement, then successive triples (overlapping in two
positions) cannot have the same total, because this puts the same number
at the beginning of the first triple and the end of the second. Thus the
triples alternate between sum 19 and sum 20. We may assume that the
triples starting in odd positions have sum 19. Since x 1 + x 2 + x 3 = x 5 + x 6 +
x 7 = 19 and x 2 + x 3 + x 4 = x 4 + x 5 + x 6 = 20, we have x 1 = x 4 1 = x 7 . Thus
the diametrically opposite numbers are the same, which is forbidden.
138
10.9. Pigeonhole
generalization: If p1 , . . . , pk are natural numbers, then
P
1k +
pi is the minimum n such that every way of distributing n objects
into classes 1, . . . , k yields at least pi objects in class i for some i . If the i th
class
P has fewer than
P pi objects, then the total number of objects is at most
( pi 1P
) = k +
pi . Thus some threshold
must be met when there are
P
pi objects. A total of k +
pi does not suffice, since then we
1k+
can put pi 1 objects in class i for each i .
10.10. If ten people each mark off a 100-yard field on a 400-yard segment,
then some point lies in at least four fields. Let x 1 , . . . , x 10 be the starting
points of the ten fields, with 0 x 1 x 10 300. Consider the three
disjoint intervals [0, 100), [100, 200), and [200, 300]. With ten numbers in
these three intervals, by the pigeonhole principle at least one of the three
contains at least four of the starting points. The right endpoint of that
interval is in the four fields having those starting points.
10.11. Multiples of x near an integer. Let S = {x, 2x, . . . , (n 1)x}.
a) For n N, if two numbers in S have fractional parts differing by at
most 1/n , then some number in S differs by at most 1/n from an integer.
Let i x and j x have fractional parts differing by at most 1/n , where j > i .
That is, i x = k + y and j x = l + z , where k, l Z and |y z| 1/n . Thus
( j i)x = (l k) + z y . Since l k is an integer, we conclude that ( j i)x
is within 1/n of an integer. Also ( j i)x S , since 0 < j i < j .
b) Some number in S is within 1/n of an integer. All fractional
S parts
lie in the interval [0, 1). Let Ir = [(r 1)/n, r/n), so [0, 1) = rn=1 Ir . If
some number in S has fractional part in I1 or In , then it is within 1/n of an
integer. Otherwise, the fractional parts of the n 1 numbers in S lie in the
n 2 intervals I2 , . . . , In1 . By the pigeonhole principle, some two of them
lie in one of these intervals. Since the intervals have length 1/n , we have
two numbers in S with fractional parts that differ by at most 1/n . Part (a)
now completes the proof.
139
10.12. Every set of n integers has a nonempty subset whose sum is divisible
by n . Let S = {b1 , . . . , bn }. We prove the stronger result that
P j some consecutive set {bi+1 , . . . , b j } has sum is divisible by n . The sum k=i+1 bk is divisP
Pj
ible by n if and only if ik=1 bk and k=1 bk have the same remainder upon
Pj
division by n . This suggests introducing the partial sums a j = k=1 bk . We
have n such sums, and we place them in the remainder classes 0, . . . , n 1.
If any a j is divisible by n , we are done, and {b1 , . . . , b j } is the desired set.
Otherwise, we are placing n partial sums in the n 1 remaining classes,
and by the pigeonhole principle a pair of these must go in the same class.
As noted above, if ai and a j have the same remainder under division by n ,
then {bi+1 , . . . , b j } is the desired set.
The integers 1, . . . , 1 (or any list of n 1 integers congruent to 1 modulo
n ) are integers with no nonempty subset summing to a multiple of n .
10.15. Every set of 55 numbers in [100] contains two numbers differing by

9, and this is best possible. If there are 55 numbers, then some congruence class mod 9 contains at least seven of the numbers, by the pigeonhole
principle. Each congruence class mod 9 contains exactly 11 numbers in
[99]. If 11 numbers are placed in a row and seven are chosen, then some
consecutive pair must be chosen (partition them into six classes consisting
of a pair of consecutive numbers and one singleton class in the case of 11
the pigeonhole principle implies two numbers are chosen from one such
pair). This completes the proof, because when two consecutive numbers
are taken from a congruence class modulo 9, their difference is exactly 9.
The proof suggests the example with 54 numbers. To avoid having two
numbers differing by 9, we must take the six numbers in odd positions in
each congruence class of size 11, and we can take the evens or the odds
from the class of size 12. The resulting set is 19, 1927, 3745, 5563,
7381, 9199. This is the only example.
10.13. If S is a list of n + 1 positive integers with sum k , for k 2n + 1, then

for each i k , some subset of S has sum i . Proof 1. Suppose that S has
largest element r and has m copies of the number 1. If m r 1, then it is
possible to add 1s successively, increasing the subset sum by one each time
until reaching r 1. The next set is the number r alone. Then the 1s can be
included one-by-one again until the next largest number in the collection
can be substituted for them, and so on. To see that m r 1, observe that
the bound on the sum yields 1 r + (n m) 2 + m 1 k 2n + 1, which
simplifies to m r 1. When k = 2n + 2, the property fails, violated by a
collection of n + 1 2s.
Proof 2. By induction on n ; for n = 0, the only example is {1}, which
works. Now suppose n 1. If S contains only 1s then all sums can be
achieved, so assume some integer a S exceeds 1. Since a exceeds 1,
the sum of the remaining n elements is less than 2n , so we can apply the
induction hypothesis to obtain subsets of {a} summing to all integers from
0 to k a . For k a + 1 i k , adding the element a to such a subset
summing to i a will construct a subset of S summing to i , if i a 0.
This requires a (1 + k)/2. This holds, since 2a > k + 1 (the bad condition)
and a + n k (the sum of S ) would imply a > n + 1 and then k 2n + 2.
10.14. Any six people contain three that pairwise know each other or three
that pairwise dont know each other. Consider an arbitrary person x among
the six. Among the five other people, the pigeonhole principle guarantees
that there must be at least three who know x or at least three who dont
know x . By symmetry (between acquaintance and nonacquaintance), we
may assume that x knows at least three others. If two who know x know
each other, then we have (at least) three people who all know each other.
Otherwise, we have (at least) three people (all knowing x ) none of whom
know each other.
140
10.16. For n = qk + r with 1 r k , the largest subset of [n ] having no two

numbers differing by k has size (n + r )/2 if q is even and (n + k r )/2 if q is
odd. Each congruence class modulo k has q or q + 1 elements in [n ], and r
of them have size q + 1. Numbers differing by k are consecutive elements
of a congruence class modulo k .
If a congruence class C has s elements in [n ], and more than ds/2e
elements of C lie in S , then by the pigeonhole principle two consecutive
elements of C lie in S . (Explicitly, consider bs/2c disjoint classes of two
consecutive elements of C , plus the last element by itself if s is odd. These
sets partition C . Choosing more elements than sets forces two element into
a single set, and two elements from one set are consecutive.) This bound is
best possible, because the ds/2e odd-indexed elements are non-consecutive.
With this argument, we must have at most dq/2e elements in the congruence classes of size q and at most d(q + 1)/2e elements in the congruence
classes of size q + 1. Since there are r classes of size q , our set is limited to
size r d(q + 1)/2e+(k r ) dq/2e. In comparison to r (q + 1)/2 +(k r )(q/2) =
n/2, we gain r/2 if q is even and (k r )/2 if q is odd. The bound is achieved
by taking the numbers 1, . . . , k , then 2k + 1, . . . , 3k , and so on, up to n .
10.17. The Erdos-Szekeres result on monotone sublists is best possible. The
list of mn integers below has no increasing sublist of length m + 1 and no
decreasing sublist of length n + 1. Partition the values 1, . . . , mn into m sets,
with the j th set being {( j 1)n + 1, . . . , jn}. Record each set in succession
in the list, but write each set down in reverse order. The list is thus
n, n 1, . . . , 1, 2n, 2n 1, . . . , n + 1, . . . , mn, mn 1, . . . , (m 1)n + 1.
141
This list has the desired property because a decreasing sublist has elements only in a single segment, but the segments have length n , and an
increasing sublist has at most one element from each segment, but there
are only m segments.
10.20. Generalization of the Chess Player Problem. The chess player plays
on d consecutive days for a total of at most b games. If k is small enough and
b 2d 1, then there must be a period of consecutive days during which
a total of exactly k games are played. The range for which the argument of
Example 10.8 works is k 2d b.
Let ai be the total number of games played on days 1 through i , and
set a0 = 0. Then a j ai is the total number of games played on days i + 1
through j . We seek i and j such that ai + k = a j . Since there is at least one
game each day, the numbers in {a j : 1 j d} are distinct, as are those
in {ai + k : 0 i d 1}. Hence a duplication among these 2d numbers
implies the desired result. Since ad b, and ad1 + k b 1 + k , the
pigeonhole principle yields the desired result whenever 2d > b + k 1.
Thus the argument works whenever k 2d b.
10.18. Multiple exam solutions. In an exam with three true/false questions,

in which every student answers each question, there are eight possible answer sheets: TTT, TTF, TFT, FTT, TFF, FTF, FFT, FFF.
a) At least nine students are needed to guarantee that no matter how
they answer the questions, some two students agree on every question. If
there are eight students, the list of eight answer sheets above has no two
agreeing on every question. If there are nine or more students, then because all answer sheets are of these eight types, the pigeonhole principle
implies that some two students have the same answers.
b) At least five students are needed to guarantee that no matter how they
answer the questions, some two students agree on at least two questions. If
two answer sheets agree in exactly two places, then they differ in the third
place, and hence the parity of the number of Ts differs. The four distinct
answer sheets FFF, TFT, FTT, TTF all have an even number of Ts, so no
two of them agree in two places.
There are four possible answer pairs for the first two questions. With
five students, some pair must agree on the first two questions.
Alternative proof. There are four possible sheets with even number of
Ts and four with an odd number. Five students having the same parity of
Ts would yield two identical sheets. Otherwise, two have opposite parity.
They agree in two places unless they disagree in all three, but then every
remaining answer sheet agrees in at least two places with one of them.
10.19. Private club needing 990 keys.
990 keys permit every set of 90 members to be housed. Suppose 90 members receive one key apiece, each to a different room, and the remaining 10
members receive keys to all 90 rooms. Each set of 90 members that might
arrive consists of k members of the first type and 90 k members of the second type. When the k members of the first type go to the rooms for which
they have keys, there are 90 k rooms remaining, and the 90 k members
of the second type that are present have keys to those rooms.
No scheme with fewer keys works. If the number of keys is less than
990, then by the pigeonhole principle (every set of numbers has one that is
at most the average) there is a room for which there are fewer than 11 keys.
Since the number of keys to each room is an integer, there are at most 10
keys to this room. Hence there is a set of 90 of the 100 members that has
no one with a key to this room. When this set of 90 members arrives, they
have keys to at most 89 rooms among them and cannot all be housed.
142
10.21. An extension of Example 10.8. We use the ideas of Exercise 10.16.

The chess player plays at most 132 games over 77 days. If there is no
consecutive period of days with a total of k games, then the number of
partial sums congruent to r modulo k is at most ds(r )/2e, where s(r ) is
the number of integers in [132] that are congruent to r . Note that s(r ) =
1 + b(132 r )/kc for 1 r k .
Whenever 23 k 25 and 1 r k , the value of s(r ) is 5 or 6.
Thus every congruence class contains at most 3 of the partial sums if no
two partial sums differ by k . This allows at most 3k partial sums. Since
there is at least one game each day over 77 days, there are 78 partial sums.
These exceeds 3k , so there must be two that differ by k .
When k = 26, we can avoid this as follows: play one game each on the
first 25 days, then 27 games on day 26, then one game each day for the next
25 days, then 27 games on day 52, then one game each day for the last 25
days. The total is 129 games, and no two partial sums differ by 26.
10.22. If m 2n and S is a set of m points on a circle with no two diametrically opposite, then S has at most n free points, where x S is free if fewer
than n points of S x lie in the semicircle clockwise from x . If m > 2n ,
consider a diameter of the circle not intersecting S . One of the resulting
semicircles has more than n points. The first point x in this semicircle is
not free, and any other point of S is free only if it is free in S x . Hence
S x has as many free points as S . When the remaining set has size 2n ,
partition it into pairs such that between a point and its mate there are
n 1 points in each direction on the circle. Then exactly one point in each
pair is free, and the original set S has at most n free points.
10.23. Every two-colored square grid of order at least five has a rectangle
with corner of the same color, but the grid of order four has no such rectangle, so the answer is five. Consider a subgrid of order five. Each column
143
has a majority in some color. By the pigeonhole principle, some color is the
majority in at least three of the columns. Suppose this is black. Each pair
of these three columns is black in at least one common row. If they are all
black in one common row, then they have at least 6 blacks in the other four
rows; another row receiving two blacks completes the rectangle.
If they are not black in a common row, then the three pairs of columns
have common blacks in three distinct rows. There remain at least three
blacks among these columns in the remaining two rows. Thus one of these
rows has at least two blacks, and this completes a black-cornered rectangle
with one of the three rows described initially.
hands in Ai A j Ak , and no hands belonging to all four sets. There are

4
ways to choose r of these sets. By the inclusion-exclusion formula, there
r

4 39
+ 6 26
4 13
+ 0 hands with no voids. Dividing by 52
are 52
13
13
13
13
13
yields the probability. To obtainthe probability
of

having at least one void,
26
52
subtract this from 1 to get [4 39
+
6
4]
/
.
13
13
13
10.24. There are 310 ways to assign 10 distinct people to three distinct
rooms, of which 310 3 210 + 3 ways leave no room empty. For each person, an independent choice among three rooms must be made, so there are
310 total ways to assign them (corresponding to ternary lists of length 10).
To count assignments with no room empty, apply inclusion-exclusion
to the universe of all 310 assignments. TLet Ai be the set of assignments
in which room i is empty. Note that | iS Ai | depends only on |S|. The
inclusion-exclusion formula counts the assignments with no rooms empty
(in none of A1 , A2 , A3 ) as 310 3 210 + 3 110 010 , because the number
of assignments with a specified k rooms empty is (3 k)10 . The answer is
310 3 210 + 3.
10.25. Decimal n -tuples with at least one each of {1, 2, 3}. We must avoid
n -tuples that lack any of these digits. The number of n -tuples lacking k
specified digits is (10 k)n . The inclusion-exclusion formula for the answer
is 10n 3 9n + 3 8n 7n .
10.26. The number of decimal m -tuples with no digit missing. We must

avoid the ten sets corresponding to the ten possible missing digits. The
number of m -tuples lacking k specified digits
k) m . Hence the
Pmis (10
k m
inclusion-exclusion formula for the answer is k=0 (1) k (10 k)m .
10.27. The probability that a bridge hand has no voids. Among the possible
hands, let Ai be the set of hands having a void in the i th suit. We want to
count the hands outside
all of A 1 , A2 , A3 , A4 . The total number of hands

39
.
There
are
hands
in Ai (one specified void guaranteed). There
is 52
13
13

26
are 13 hands in Ai A j (two specified voids guaranteed). There are 13
13
144
10.28. There are 16 natural numbers less than 252 that are relatively prime
to 252. Out of a universe of 252 = 22 32 7 numbers, we want to count those
that do not have 2 or 3 or 7 as a prime factor. Let A 1 , A2 , A3 be the numbers
in [252] that are divisible by 2,3,5, respectively. Since [n ] has exactly bn/kc
numbers divisible by k , the inclusion-exclusion formula allows us to count
the desired numbers by

252
252
252
252
252
252
252
252
+
+
+
= 72.
2
3
7
6
14
21
42

10.29. There are natural numbers less than 200 that have no divisor in
{6, 10, 15}. Since 10 divides 200, we can include 200 in the universe without changing the answer. Out of a universe of 200 numbers, we want to
count those that do not have 6 or 10 or 15 as a factor. Let A 1 , A2 , A3 be
the numbers in [200] that are divisible by 6,10,15, respectively. Since [n ]
has exactly bn/kc numbers divisible by k , the inclusion-exclusion formula
allows us to count the desired numbers by
200

200
200
200
200
200
200
200
+
+
+
= 140.
6
10
15
60
90
150
900
10.30. Euler totient formula.

If p, q are prime and (n) = #{i : 1 i < n and gcd(i, n) = 1}, then
( pq) = ( p)(q). If p is prime, then p has no common factor with any
natural number less than p , so ( p) = p 1. Hence the right side of the
equation is ( p 1)(q 1). If p, q are both prime, then a number less than
pq fails to be relatively prime to pq if and only if it is a multiple of p or q .
There are q 1 numbers less than pq that are multiples of p , also p 1
numbers less than pq that are multiples of q , and no numbers less than
pq that are multiples of both. Subtracting off the bad numbers, we have
( pq) = pq 1 ( p 1) (q 1) = pq p q + 1, which is the same as
( p 1)(q 1), as needed.
Q
, where P(m) is the set of
The general formula is (m) = m pP(m) p1
p
prime factors of m . The formula counts the elements of [m ] that are not
divisible by any element of P(m). For a given subset S P(m), there are
145
Qm
pS

p
elements of [m ] that are divisible by all the elements of S . By the
inclusion-exclusion formula,
(m) =
SP(m)
(1)|S| Q
m
pS
=m
X
SP(m) pP(m)S
Y 1
pS
=m
Y
pP(m)
1
p
10.31. If A1 , . . . , An are sets in a universe U , and N (T ) is the number of

elements of U that belong to the sets indexed by T but to no others among
A1 , . . . , An , then
X
\
N (T ) =
(1)|S||T | |
Ai |.
T S[n]
iS
Restrict U to the set U of elements that have all properties indexed by T .

Consider the n |T | properties not indexed by T . The desired elements are
those in U 0 that contain none of these n |T | properties, and the formula
above is precisely the usual inclusion-exclusion formula for counting this
subset of U 0 .
0
10.32. The number of permutations of [n ] that leave no odd number un

Pdn/2e
moved is k=0 (1)k dn/2e
(n k)!. Consider the universe U of all n ! perk
mutations of [n ]. Let the set Ai be those permutations that put 2i 1 in
position 2i 1, for 1 i dn/2e. The desired answer is the number of
permutations that are outside all of the T
sets A 1 , . . . , An . For any set S consisting of k indices, the permutations in iS Ai are those that permute the
n k elements not specified by S among themselves, arbitrarily. There are
(n k)! of these. Hence the inclusion-exclusion
formula
to count the items

Pdn/2e
outside all the Ai gives N () = k=0 (1)k dn/2e
(n
k)
!.
k
10.33. Assigning 2n courses to n professors. The courses are distinct, and

the professors are distinct. If each professor in turn is assigned two courses
from those that remain, then the number of selections for each successive
professor does not depend on how the previous choices were made. Thus
Qn1 2n2i
the product rule applies, and the courses can be assigned in i=0
.
2
This equals (2n)!/2n , which can be obtained directly by thinking of the list
of courses as 2n positions in a row; the answer is the number of ways to
arrange two copies each of n types of markers representing
the professors.

2n
Thus the answer is the multinomial coefficient 2,...,2
= (2n)!/2n .
In the spring, the possible assignments are the same, but we exclude
those where some professor has the same assignment as in the fall. Let A i
be the set of assignments in which the i th professor has the same assignment as in the fall. When k particular professors have repeat assignments,
this preassigns 2k courses, and the ways to complete the assignment are
146
the ways to assign

T2n 2k courses to n k professors without further restrictions. Thus iS Ai = (2n 2k)!/2nk whenever |S| = k . There are

n
sets of k professors. By the inclusion-exclusion principle, the number of
k

P
.
assignments that belong to none of the sets A i is thus nk=0 (1)k nk (2n2k)!
2nk
The probability of this event, when the courses are assigned at random, is
this divided by (2n 2k)!/2nk .
10.34. Selection of n objects

from 5 types with no type selected more than 4

times. There are n+51
ways
to select n objects from 5 types with repetition
51
allowed. These selections are our universe U . We want to eliminate the
selections in which a type is picked more than 4 times, so set A i is the
subset of U where type i is picked more than 4 times. Since the objects of
each type are indistinguishable, we may guarantee being in A i by picking 5
objects of type i initially and then completing the selection.
In other words,

(n10)+51
|Ai | = (n5)+51
.
Similarly,
|A
A
|
=
,
since
we
distribute 5 to
i
j
51
51
each violated type before considering arbitrary selections with repetition
for the remainder. The
sum by the inclusion-exclusion formula is
resulting

P
N () = 5k=0 (1)k nk n5k+4
.
4
10.35. Pairings of n boys and n girls as lab partners. For each i , the i th
tallest boy must not be matched to the i th tallest girl. Let A i be the set of
pairings that violate this condition.
pairs are allowed, the number of acceptable pairings

Pa) When same-sex
(2n)!
is nk=0 (1)k nk 2(2n2k)!
nk (nk)! . The universe is the set of all 2n n! pairings of 2n
people (with the people in order, arrange labels 1, 1, 2, 2, . . . , n, n to make
pairs, and then divide by n ! becauses the pairs are not labeled). Within
this universe, we count those pairings that avoid A 1 , . . . , An . Pairings that
lie in Ai for all i S in effect exclude 2 |S| people from participation by predetermining their pairs, so the number of ways to complete the pairing is
(2n2k)!
, where k = |S|. This depends only on |S|, so the Inclusion-Exclusion
2nk (nk)!
Principle yields the formula claimed.
b) When
pairs are not allowed, the number of acceptable pairP same-sex
ings is nk=0 (1)k nk (n k)!. The universe is the set of all n ! pairings of
boys with girls. Within this universe, we count those pairings that avoid
A1 , . . . , An . The number of pairings that lie in A i for all i S is (n |S|)!.
This depends only on |S|, so the Inclusion-Exclusion Principle yields the
formula claimed.
10.36. Permutations of two copies each of n types of letters with no consecutive letters the same. Let U be the universe of permutations of these
letters. Let Ai be the set of permutations in which the two letters of the
i th type appear consecutively. We want to count the permutations in none
of these sets. The number of arrangements having k i letters of type i is
147
Chapter 11: Graph Theory
P
Q
( ki )!/( ki !). Hence |U | = (2n)!/2n . For permutations in Ai , the two
copies of the i th letter must be next to each other (and indistinguishable),
so we can view them as a single letter, and |A i | = (2n 1)!/2n1 . More generally, when we consider the intersection of k of these sets and force k of the
pairs to appear consecutively, the number of arrangements is (2nk)!/2 nk .
This holds for any collection S of k indices from n , and there are nk such subsets
Pof indices. Hence the answer given by the inclusion-exclusion formula
is nk=0 (1)k nk (2n k)!/2nk .
k sets does not depend on which k sets are chosen, always equaling f (k).
For k = 0, the contribution is 2n , which suggests that our universe should
correspond to all subsets of [n ].
The resulting proof is that each side of the equality counts the subsets
of [n ] that contain none of the elements. On the right, we know that there
is exactly one such subset, the empty set. On the left, we define sets A i =
{T T
[n ]: i T } within the universe of all subsets of [n ]. Then the subsets
in iS Ai consist of all subsets of [n ] that contain the elements in S . There
are exactly 2n|S| of these, so the left-side is the inclusion-exclusion sum to
count the subsets of [n ] belonging to none of the sets A i .

Pn
r 1
k n nk+r 1
= n1
. The factor (1)k nk in the summand
10.41.
k=0 (1) k
r

1
suggests an inclusion-exclusion computation. The quantity nk+r
is the
r
number of ways to choose r elements from n k types. Here r remains fixed.
The term k = 0 suggests making the universe be the set of selections of r
elements from n types, with repetition allowed. When we select r elements
from n k types, we are leaving some types unused.
The right side is the number of selections of r elements from n types
in which every type is used at least once. The left side counts the same set
using the Inclusion-Exclusion Principle.
10.37. Spouse-avoiding circular permutations of n couples. Let U be the

set of circular permutations of the people in n couples. Let A i be the set of
circular permutations in which
T the i th couple occurs together. For S [n ],
the circular permutations in iS Ai are counted by replacing the i th couple
with a single token, for all i S , using an arbitrary circular permutation
of the resulting set, and multiplying by 2|S| to account for the orientations
of the special couples. By inclusion-exclusion,
the number
of circular per
P
mutations with no adjacent spouses is nk=0 (1)k nk 2k (2n 1 k)!.
10.38. Counting permutations. Let E nk be the number of permutations of

[n ] with k fixed points and let Dn = E n0 be the number of permutations of
[n ] with no fixed
points.
a) E nk = nk Dnk , We can partition the permutations with k fixed points
according to which positions are fixed. The number of permutations with a
particular set of k fixed points is Dnk , because the remaining n k elements
must be permuted among themselves without fixed points. This number
Dnk is the same no matter whichset of k points was chosen to be fixed,
and
these k points can be chosen in nk ways, so the total value of E nk is nk Dnk .

P
b) n ! = nj=0 nj Dn j . Every permutation has some number of fixed
points between 0 and n ; this partitions the set of permutations into disjoint
sets that we count separately. In part (a) we determined the number of
permutations that have exactly j fixed points, and summing this over all
values of j counts all possible permutations.

Pn
k n
10.39.
k=0 (1) k = 0 when n > 0, by inclusion-exclusion. This is the
inclusion-exclusion formula to count the elements belonging to none of n
sets A1 , . . . , An , where | iS Ai | = 1 for all S . In particular, this formula
holds when S = , so the universe has size 1, and when |S| = 1, so each
Ai has one element. Thus the sum counts the elements in a universe of
size 1 that belong to none of n sets equal to the universe. There is no such
element when n > 0, but when n = 0 the answer is 1.

P
10.40. Proof of nk=0 (1)k nk 2nk = 1 by inclusion-exclusion. Since the

Pn
sum has the form k=0 (1)k nk f (k), this suggests an inclusion-exclusion
argument with n sets in a universe, where the size of the intersection of
148
11. GRAPH THEORY

11.1. The graph with vertex set [12] in which vertices are adjacent if and
only if the numbers are relatively prime has 45 edges. We count edges by
summing the vertex degrees and dividing by 2. All of {1, 7, 11} have degree
11. All of {2, 4, 8} are adjacent precisely to the odd numbers, for degree 6.
Vertices 3 and 9 are adjacent to the 8 numbers not divisible by 3. Vertex
5 is adjacent to all but itself and 10. Vertices 6 and 12 are adjacent to
{1, 5, 7, 11}, and 10 is adjacent to {1, 3, 7, 9, 11}. The degrees (in order for
vertices 1 through 12) are thus 11, 6, 8, 6, 10, 4, 11, 6, 8, 5, 11, 4. They sum
to 90, so there are 45 edges.
11.2. In the graph G with vertex set Zn in which vertices u and v are adjacent if and only if they differ by 6, the number of components is gcd(n, 6)
when n > 12. Arrange the vertices in a circle. Each is adjacent to the vertices six positions away in each direction. Since n > 12, G is regular of
degree 2 and the components are cycles.
When n is divisible by 6, G consists of six cycles with n/6 vertices each;
each cycle goes once around the circle. When n is divisible by 3 but not 2,
G consists of three cycles with n/3 vertices each, going twice around the
circle. When n is divisible by 2 but not 3, G consists of two cycles with n/2
149
vertices each, going thrice around the circle. When n is relatively prime to
6, there is a single cycle through all the vertices.
When n = 12, the graph consists of six disjoint edges. When n < 12,
no pairs differ by 6, and the graph consists of n isolated vertices.
having an additional trail. Altogether, we have used the desired number

of trails to partition E(G).
Proof 4 (induction on the number of edges). With only one edge, G =
K 2 , and we have one trail. With more edges, if G has a vertex x of even
degree adjacent to a vertex y of odd degree, then G 0 = G x y has the
same number of vertices of odd degree as G . The trail decomposition of
G 0 guaranteed by the induction hypothesis has one trail ending at x and
no trail ending at y . Add x y to the trail ending at x to obtain the desired
decomposition of G . If G has no vertex of even degree adjacent to a vertex
of odd degree, then every vertex of G has odd degree. In this case, deleting
an edge x y reduces k , and we can add x y as a trail of length one to the
decomposition of G x y guaranteed by the induction hypothesis.
11.3. Eulerian with even order and odd size. An Eulerian graph with an
even number of vertices and an odd number of edges can be formed as the
union of an even cycle and an odd cycle that share one vertex.
11.4. If G is a connected non-Eulerian graph, then the minimum number of
trails that together traverse each edge of G exactly once is half the number
of vertices having odd degree. The assumption of connectedness is necessary, because the conclusion is not true for a graph with two components
in which one component is Eulerian.
Suppose G is a connected non-Eulerian graph, so G has 2k vertices
of odd degree for some positive integer k . A trail contributes even degree
to every vertex, except odd degree to its endpoints when they are not the
same vertex. Therefore, a partition of the edges into trails must have some
trail ending at each vertex of odd degree. Since each trail has (at most) two
ends, at least k trails are required. There are several ways to prove that k
trails are always enough.
Proof 1 (using Eulerian circuits). Pair up the vertices of odd degree
in G arbitrarily and form G 0 by adding a copy of each pair as an edge.
The resulting graph G 0 is connected and has even degree at every vertex.
Therefore it has an Eulerian circuit, since the characterization of Eulerian
graphs allows multiple edges. As we traverse the circuit, we start a new
trail in G after each time we traverse an edge of G 0 E(G). This yields k
pairwise edge-disjoint trails partitioning E(G).
Proof 2 (using Eulerian circuits). Form G 0 by adding a new vertex w
and making it adjacent to all vertices of G that have odd degree. The resulting graph G 0 is connected and has even degree at every vertex. Therefore
it has an Eulerian circuit. As we traverse the circuit, we start a new trail
in G after each time we leave w and arrive at a vertex of G . Since w has
degree 2k , the circuit visits it k times, so we obtain k pairwise edge-disjoint
trails partitioning E(G).
Proof 3 (induction on k ). If k = 1, we add an edge between the two odd
vertices, obtain an Eulerian circuit, and delete the added edge to obtain a
single trail covering all edges of G . If k > 1, let P be a path between two
vertices of odd degree. The graph G 0 = G E(P) has 2k 2 vertices with
odd degree since we have changed degree parity only at the ends of P . We
apply the induction hypothesis to each component of G 0 that has vertices
with odd degree. Every component not having vertices of odd degree has
an Eulerian circuit that contains a vertex of P ; we splice it into P to avoid
150
11.5. The vertices of a simple graph cannot have distinct degrees. In

a simple graph with n vertices, every vertex degree belongs to the set
{0, . . . , n 1}. If fewer than n values occur, then the pigeonhole principle
yields the claim. Otherwise, both n 1 and 0 occur as vertex degrees. This
is impossible; if one vertex is adjacent to all others, then there can be no
isolated vertex.
11.6. In a league with two divisions of 11 teams each, there is no schedule
with each team playing 7 games against teams within its division and 4
games against teams in the other division. In the language of graph theory,
the scheduled games can be viewed as edges in a graph with 22 vertices. We
are asking for a graph that is regular of degree 11, but we are also asking
for the subgraph induced by the 11 teams in one division to be regular of
degree 7. Since 7 and 11 are both odd, this is impossible, because every
graph has an even number of vertices of odd degree.
11.7. A connected graph in which every vertex degree is even has no edge
whose deletion disconnects it.
Proof 1 (degree-sum). Let G be a connected graph with all vertices
having even degree, and suppose G has a disconnecting edge. Deleting this
edge leaves a graph with vertices of odd degree in distinct components. A
component containing one of these vertices is a subgraph having an odd
number of vertices of odd degree, which cannot exist.
Proof 2 (Eulerian circuits). A connected graph with even degrees is
Eulerian. For any pair of vertices u, v , we can start at a visit to u on an
Eulerian circuit and traverse the circuit in two opposite directions until it
first reaches v . This gives a pair of edge-disjoint u, v -trails. Deleting an
single edge leaves one of these trails intact, and it contains a u, v -path.
Since u, v were chosen as arbitrary vertices, no deletion of a single edge
can make any vertex unreachable from any other.
151
11.8. If l, m, n are nonnegative integers with l + m = n , then there exists

a connected simple n -vertex graph with l vertices of even degree and m vertices of odd degree if and only if m is even, except for (l, m, n) = (2, 0, 2).
Since every graph has an even number of vertices of odd degree, and the
only simple connected graph with two vertices has both degrees odd, the
condition is necessary. To prove sufficiency, we construct such a graph G .
If m = 0, let G = Cl (except G = K 1 if also l = 0). If m > 0, begin with
K 1,m1 , which has m vertices of odd degree, and then add a path of length l
beyond one of the leaves. (Many other constructions also work.)
the edges x 0 x 1 and y0 y1 yields a spanning cycle of Q d+1 . It traverses all the
vertices of G 0 from x 0 to y0 , then crosses to G 1 , then traverses all vertices
of G 1 ending at x 1 , then returns to x 0 to complete the cycle.
11.9. Deleting a vertex of maximum degree cannot increase the average

degree, but deleting a vertex of minimum degree can reduce the average
degree. Deleting any vertex of a regular graph reduces the average degree.
For the first statement, suppose G has n vertices and m edges, and a, a 0 are
the average degrees of G and G x . Since G x has m d(x) edges and
(n2)a
< a if d(x) a .
degree sum 2m 2d(x), we have a 0 = na2d(x)
n1
n1
11.10. Inductive aspects of the d -dimensional cube Q d .

Inductively, Q d is constructed from the disjoint union of two copies of
Q d1 by adding a matching that pairs up corresponding vertices. The vertex
set of Q d is the set of d -tuples with entries in {0, 1}. Two vertices are
adjacent if they differ in exactly one place. Suppose d 1. The subgraph
of Q d consisting of all the vertices with 0 in the last place and all the edges
of Q d joining vertices in this set is isomorphic to Q d1 ; call this G 0 . The
subgraph of Q d consisting of all the vertices with 1 in the last place and all
the edges of Q d joining vertices in this set is isomorphic to Q d1 ; call this
G 1 . The edges of Q d that do not appear in these two subgraphs are those
between a vertex of G 0 and a vertex of G 1 differing only in the last place.
These three sets of edges are pairwise disjoint and account for all edges.
Q d has d 2d1 edges. Let ad be the number of edges in Q d . By the
inductive construction, ad = 2ad1 + 2d1 for d 1. Also a0 = 0. We prove
by induction on d that ad = d 2d1 . For d = 0, we have a0 = 0 = 0 21 . For
d > 0, we use the recurrence relation and then the induction hypothesis to
obtain ad = 2ad1 + 2d1 = 2((d 1)2d2 ) + 2d1 = d 2d1 .
Q d has a spanning cycle, for all d 2. Proof by induction on d . For
d = 2, the graph Q 2 is itself a 4-cycle. The inductive step is illustrated by
the picture below (going from Q 3 to Q 4 ). For the inductive step, suppose
the claim holds for Q d . Let C be a spanning cycle in Q d , and let x y be an
edge in C . Let G i be the copy of Q d in Q d+1 in which all the vertex labels
end with i {0, 1}, with x i , yi being the vertices obtaining by appending i to
the labels of x and y . Let C i be the copy of C in G i obtained by appending i
to the label of each vertex in C , listing the vertices in the same order as on
C . Then deleting x 0 y0 from C 0 and x 1 y1 from C 1 and replacing them with
152
x0
x1
y0
y1

11.11. The d -dimensional cube has d2 2d2 4-cycles and d3 2d+1 6-cycles.
Traversing a 4-cycle must change two coordinates and then unchange them.
Thus a 4-cycle is determined by choosing two coordinates to change and
choosing a fixed value in each of the other d 2 coordinates.
Traversing a 6-cycle similarly must change and unchange exactly three
coordinates, since only 4 vertices are available if d 2 coordinates remain
fixed. For each of the d3 2d3 choices for the d 3 fixed coordinates, the
number of ways to complete a 6-cycle is the number of 6-cycles in Q 3 . For
each pair of adjacent vertices and for each pair of complementary vertices
in Q 3 , there is exactly one 6-cycle through the remaining six vertices. The
other types of nonadjacent pairs leave no 6-cycle when deleted. Thus there
are 12+4=16 6-cycles in Q 3 , which completes the computation.
11.12. Isomorphism. Using the correspondence indicated below, the first
two graphs are isomorphic; the graphs are bipartite, with u i v j if and
only if i 6= j . The third graph contains odd cycles and hence is not isomorphic to the others.
u 1
u 2
v v
1 2
v4v3
u 4
u
3
u 1
v3
u 2
v4
u 3
v1
u 4
v2
These are the 3-dimensional cube and the graph obtained by deleting
four disjoint edges from K 4,4 . The vertices of the cube are the binary 3tuples; the parity of a vertex is the number of 1s in the 3-tuple. Each
vertex of the cube is adjacent to every vertex of the opposite parite except
its complement. Since the complement of the complement of a binary list is
the original list, this describes Q 3 as formed by deleting four disjoint edges
from K 4,4 .
153
11.13. G
= H if and only if G
= H . If f is an isomorphism from G to H ,
then f is a vertex bijection preserving adjacency and nonadjacency, and
hence f preserves non-adjacency and adjacency in G and is an isomorphism from G to H . The same argument applies for the converse, since the
complement of G is G .
11.17. Isomorphisms for the Petersen graph. Isomorphism is proved by giving an adjacency-preserving bijection between the vertex sets. For pictorial
representations of graphs, this is equivalent to labeling the two graphs with
the same vertex labels so that the adjacency relation is the same in both
pictures; the labels correspond to a permutation of the rows and columns
of the adjacency matrices to make them identical. There various drawings of the Petersen graph below illustrate its symmetries; the labelings
indicate that these are all the same (unlabeled) graph. The number of
isomorphisms from one graph to another is the same as the number of
isomorphisms from the graph to itself.
11.14. The smallest number of vertices for which nonisomorphic simple

graphs exist with the same vertex degrees is 5. For n 3, there is only
one isomorphism class with each number of edges. As listed in Example
11.26, there are 11 isomorphism classes of simple graphs with four vertices.
When the number of edges is in {0, 1, 5, 6}, there is only one isomorphism
class. With two edges, there are two isomorphism classes, but they have
different degree lists (similarly for four edges). With three edges, the three
isomorphism classes have degree lists 3111, 2220, and 2211, all different.
Hence a nonisomorphic pair of simple graphs with the same vertex degrees
must have at least five vertices. With five vertices, the 5-vertex path and
the disjoint union of a triangle and an edge are nonisomorphic graphs with
degree lists 22211.
11.15. There are two isomorphism classes of 4-regular simple graphs with 7
vertices. Simple graphs G and H are isomorphic if and only if their complements G and H are isomorphic, because an isomorphism : V (G) V (H )
is also an isomorphism from G to H , and vice versa. Hence it suffices to
count the isomorphism classes of 2-regular simple graphs with 7 vertices.
Every component of a finite 2-regular graph is a cycle. In a simple graph,
each cycle has at least three vertices. Hence each class it determined by
partitioning 7 into integers of size at least 3 to be the sizes of the cycles.
The only two graphs that result are C 7 (one cycle) and C 3 + C4 (two cycles).
11.16. The number of vertices in a graph isomorphic to its complement is
congruent to 0 or 1 (mod 4). If G and G are isomorphic,
then they have
the same number of edges, but together they have 2n edges (with none
repeated), so the number of edges in each must be n(n 1)/4. Since this
number must be an integer and the factors in the numerator cannot both
be divisible by 2, one factor in the numerator must be divisible by 4. Hence
the number of vertices is a multiple of 4 or one more than a multiple of 4.
Below are planar graphs with 4, 5, and 8 vertices that are isomorphic
to their complements.
154
5 10 6
2
7 1 3
7
8
7
4 5 10
9
8
8
9
6
3
10
4
8
3
9
4
10
5
7
With vertices named as above, the adjacency relation is the same in
via the correeach graph. This represents the disjointness relation on [5]
2
spondence below.
1
2
3
4
5
6
7
8
9
10
{1, 2} {3, 4} {5, 1} {2, 3} {4, 5} {3, 5} {5, 2} {2, 4} {4, 1} {2, 4}
Alternatively, one can name the vertices using four disjoint sets of labels and establish isomorphisms from each graph to the next. Transitivity
of the isomorphism relation then implies that each two graphs chosen from
this set are isomorphic.
n1
11.18. There are exactly 2( 2 ) simple graphs with vertex set v1 , . . . , vn in
which every vertex has even degree.
If we do not care about the degrees of

the vertices, then there are 2n simple graphs, since each pair of vertices
may form an edge or not. Similarly, there are 2C H (n1,2) simple graphs
with vertex set {v1 , . . . , vn1 }. We establish a bijection to that set of graphs.
Given a simple graph G with vertex set {v1 , . . . , vn1 }, we form a new graph
G 0 by adding a vertex vn and making it adjacent to each vertex that has odd
degree in G , as illustrated below. The vertices with odd degree in G have
even degree in G 0 . Also, vn itself has even degree because the number of
vertices of odd degree in G is even. Conversely, deleting the vertex v n from
a graph with vertex set {v1 , . . . , vn } having even vertex degrees produces
a graph with vertex set {v1 , . . . , vn1 }, and this is the inverse of the first
155
procedure. We have established a one-to-one correspondence between the

sets; hence they have the same size.
11.21. A graph G is connected if and only if for every partition of V (G) into
non-empty sets S, T , there is an edge x y with x S and y T . When G is
connected and S, T is a nontrivial partition, we choose u S and v T .
Since G is connected, G has a u, v -path, and this path must leave S along
some edge x y . When G is not connected, we let S be the vertex set of a
component of G and let T be the remainder of the vertex set; since S is a
component, there is no edge from S to T .
vn
G0
11.19. Combinatorial proofs

with

graphs.

a) For 0 k n , 2n = 2k + k(n k) + nk
. Consider the complete
2

n
graph K n , which has 2 edges. If we partition the vertices of K n into a k set and an n k -set, then we can count the edges as those within one block
of the partition and those choosing
a vertex from each. Hence the total

number of edges is 2k + nk
+
k(n
k)
2

.
P
P
ni
n
. Again consider the edges of K n ,

n i = n , then
b) If
2
2
and partition the vertices into sets Si with sizes {n i }. The left side of the
inequality counts the edges in K n having both ends in the same Si , which
is at most all of E(K n ).
11.20. a) If x and y are nonadjacent vertices of degree at least (n + k 2)/2
in a simple graph G with n vertices, then x and y have at least k common
neighbors. Let X and Y be the sets of neighbors of x and y . Since x and y
are nonadjacent, |X Y | n 2. Using the degree bound and identities
for set operations, we have
|X Y | = |X | + |Y | |X Y |
n+k2 n+k2
+
(n 2) = k.
2
2
b) Every simple n -vertex graph G with minimum degree at least bn/2c

is connected. It suffices to prove the stronger result that every two nonadjacent vertices have a common neighbor. Let N (x) and N (y) denote
the neighbor sets of x and y ; each has size at least bn/2c, so the sizes
sum to at least n 1. If x and y are nonadjacent, then N (x) and N (y)
are contained in V (G) {x, y}, and hence |N (x) N (y)| n 2. Since
|N (x) N (y)| + |N (x) N (y)| = |N (x)| + |N (y)|, we have
|N (x) N (y)| = |N (x)| + |N (y)| |N (x) N (y)| (n 1) (n 2) = 1.
Thus every two nonadjacent vertices have a common neighbor.

To show that the requirement on minimum degree is best possible,
consider the n -vertex graph consisting of two disjoint cliques with dn/2e
and bn/2c vertices and no additional edges. For n 2, this is a disconnected
graph with minimum degree bn/2c 1.
156
11.22. Measuring k gallons using buckets with integer capacities l > m > n
in gallons. We create vertices for the possible states of our system. Each
vertex is a triple (a, b, c) listing the integer numbers of gallons in the three
buckets. Our initial vertex is (l, 0, 0). Since we are only allowed to pour
from one gallon into another, the sum of the three integers is always l .
The operation of pouring from one bucket into another that already
has some water is not reversible; thus we need directed edges, like those
used in functional digraphs. Directed edges would also allow us to model
the operation of spilling out a bucket, if that were allowed.
Given this directed graph as a model, the problem is to determine
whether it has a path (following directed edges in the forward direction)
from the vertex (l, 0, 0) to a vertex in which k is one of the integers.
11.23. If G is a finite simple graph with minimum degree k , and k 2,
then G contains a path of length at least k and a cycle of length at least
k + 1. Because V (G) is finite, every path in G is finite, so G has a maximal
path P . Let u be an endpoint of P . Because P cannot be extended, every
neighbor of u belongs to P . Since u has at least k neighbors, its neighbor
v that is farthest along P from u is the other end of a path P 0 of length at
least k , and uv completes a cycle of length at least k + 1 with P 0 .
11.24. If G is a connected graph, then any two paths of maximum length in

G have a common vertex. Let P and Q be the paths, with Q at least as long
as P . The proof is by contradiction; if V (P) and V (Q) are disjoint, then
we find another path that is longer than P . Since G is connected, G have
a path from V (P) to V (Q). Let R be a shortest such path, with endpoints
x V (P) and y V (Q). By the choice of R , the internal vertices of R (if
there are any) are neither in P nor Q . Hence we can form a new path S
by traveling from one end of P to x along P , then to y along R , then to an
end of Q along Q . We use the end of P farther from x and the end of Q
farther from y , choosing arbitrarily if x or y is in the middle of its path.
If p, q, r are the lengths of P, Q, R , respectively, then the length of S is at
least p/2 + r + q/2. Since q p and r > 0, S is longer than P .
11.25. The
number of edges in an n -vertex triangle-free simple
maximum

graph is n 2 /4 . Suppose G is an n -vertex triangle-free simple graph. Let
157
k be the maximum of the vertex degrees, and let x be a vertex of degree k

in G . Since G has no triangles, there are no edges between neighbors of x .
Hence summing the degrees of x and vertices in the set S of non-neighbors
of x counts at least one endpoint of every edge, and the total is at least e(G).
Since we are summing over n k vertices, each having degree at most k ,
we obtain e(G) k(n k). By the Arithmetic-Geometric Mean Inequality
(Proposition 1.1), the product of two numbers is at most the square of their
average, so e(G) k(n k) n 2 /4. To prove that the bound
is best possible,
we exhibit K bn/2c,dn/2e as a triangle-free graph with n 2 /4 edges.
Proof 2 (induction on n ). A graph with 1 vertex and no edges is connected and has no cycles, so the statement holds for nP
= 1. For the induction
step, suppose n > 1. By the degree-sum formula, vV (G) d(v) = 2n 2,
which is less than 2n . Hence some vertex v has degree less than 2, by the
pigeonhole principle. If G is connected, then this vertex v has degree 1.
Deleting v and its incident edge leaves a graph G 0 with n 1 vertices and
n 2 edges. Since a vertex of degree 1 cannot be an internal vertex on a
path, all paths in G connecting vertices of G 0 remain in G 0 , and hence G 0
is also connected. By the induction hypothesis, G 0 has no cycles. Since a
vertex of degree 1 cannot belong to a cycle, G also has no cycles.
Conversely, suppose G has no cycles. Since G has fewer edges than
vertices, some component H of G has fewer edges than vertices. By the
degree-sum formula, the pigeonhole principle, and the connectedness of H ,
there is a vertex v in H that has degree 1 (in G ). Deleting v and its incident
edge leaves a graph G 0 with n 1 vertices and n 2 edges. Deleting a vertex
cannot create a cycle, so G 0 has no cycles. By the induction hypothesis, G 0
is connected. Since paths to or from the neighbor of v in G can be extended
to become paths to or from v , G is connected. Since G has n vertices and
n 1, the sum of the degrees of
bn/2c
dn/2e
11.26. Every n -vertex graph G with n k edges has at least k components.

Proof 1 (properties of spanning trees). Let l be the number of components of G , and let n i be the number of vertices in the i th component. The
i th component has a spanning
tree with n i 1 edges. Hence the total numP
ber of edges is at least li=1 (n i 1), which equals n l . The total number
of edges equals n k , and n k n l yields l k .
Proof 2 (properties of cut-edges). Let m = n k . We begin with n
isolated vertices and no edges, and we add the edges of the graph one by
one. At first there are n components. Each addition of an edge of the
graph reduces the number of components of the current graph by at most
one. Thus when m edges have been added there are at least n m = k
components.
11.27. If G is a graph with n vertices and n 1 edges, then G is connected
if and only if G has no cycles.
Proof 1 (properties of trees and connected graphs). If G is connected,
then we can delete edges from cycles, one by one, until we obtain a graph
G 0 with no cycles. Since at each step we deleted an edge of a cycle, G 0 is
connected (Lemma 11.22). Since G 0 is connected and has no cycles, it is
a tree and therefore has n 1 edges (Theorem 11.25). Since G has n 1
edges, we did not delete any, and G = G 0 .
Conversely, if G has no cycles, then every component of G is a tree.
Suppose the components are G 1 , . . . , G k , with G i having n i vertices, for
each i . By Theorem 11.25, G i hasP
n i 1 edges. Summing
the number of
P
k
edges from each component yields i=1
(n i 1) = ( n i ) k = n k edges
for G . Since G has n 1 edges, k = 1, and G is connected.
158
11.28. A graph G is a tree if and only if for all x, y V (G), there is exactly
one x, y -path in G . Necessity: If G is a tree, then G is connected and has
an x, y -path. If G has two x, y -paths, let H be their union, and let e be an
edge belonging to only of the paths. By the transitivity of the connection
relation, H e is connected. By Lemma 11.36, e thus belongs to a cycle in
H , which is contained in G . This contradicts the absence of cycles in G .
Sufficiency: The existence of x, y -paths makes G connected, and the
uniqueness of paths forbids cycles, since every pair u, v in a cycle is connected by two paths along the cycle. Thus G must be a tree.
11.29. If G is a tree with maximum degree k , then G has at least k leaves.

Proof 1a (maximal paths). Deleting a vertex x of degree k produces a
forest of k subtrees, and x has one neighbor wi in the i th subtree G i . Let
Pi be a maximal path starting at x along the edge xwi . The other end of Pi
must be a leaf of G and must belong to G i , so these k leaves are distinct.
(Note: choosing a maximal path is cleaner than follow paths from the
maximum degree vertex through each of its neighbors until you get stuck,
which must be at distinct leaves since there are no cycles.)
Proof 1b (leaves in subtrees). Deleting a vertex x of degree k produces
a forest of k subtrees. Each subtree is a single vertex, in which case the
vertex is a leaf of G , or it has at least two leaves, of which at least one is
not a neighbor of x . In either case we obtain a leaf of the original tree in
each subtree.
159
Proof 2 (counting two ways). Count the degree sum by edges and by
vertices. By edges, it is 2n 2. Let k be the maximum degree and l the
number of leaves. The remaining vertices must have degree at least two
each, so the degree sum when counted by vertices is at least k + 2(n l
1) + l . The inequality 2n 2 k + 2(n l 1) + 1 simplifies to l k . (Note:
The same idea is used in noting that degree 2(n 1) k remains for the
vertices other than the one of maximum degree. Since all degrees are 1 or
at least 2, there must be at least k vertices of degree 1.)
Proof 3 (induction on the number of vertices). For n 3, we inspect
the unique tree with n vertices. For n > 3, delete a leaf u . If 1(T u) =
1(T ), then by induction T u has at least k vertices, and replacing u adds
a leaf at the expense of at most one leaf from T u . Otherwise 1(T u) =
1(T ) 1, which happens only if the neighbor of u is the only vertex of
maximum degree in T . Now induction guarantees at least k 1 leaves in
T u , and adding u gives another, since the vertex of maximum degree in
T cannot be a leaf in T u (this is the reason for n = 3 in the basis).
a list of n natural numbers satisfying the condition. Since the sum is only
2n 2, the numbers cannot all be at least 2, so there is at least one 1. Also
2n 2 > n , so there must be at least one number greater than 1. Suppose
di = 1 and d j > 1. Form a new list of numbers by deleting di and reducing
d j by 1. This reduces the sum by 2 and produces a set of n 1 natural
numbers summing to 2(n 1) 2. By the induction hypothesis, there is
a tree with these degrees. Find a vertex x with degree d j 1 in such a
tree, and add a new vertex joined only to x . The resulting graph is a tree
(connected and acyclic) with degrees d1 , . . . , dn .
11.30. A connected graph with n vertices has exactly one cycle if and only
if it has exactly n edges. Let G be a connected graph with n vertices.
Necessity: If G has one cycle, then deleting an edge e of that cycle
leaves a connected graph with no cycle. By definition, G e is a tree, and
every tree with n vertices has n 1 edges. Hence G e has n 1 edges and
G has n edges.
Sufficiency: Let G have exactly n edges. If G has no cycles, then by
definition G is a tree and has n 1 edges. Therefore G has at least one cycle.
If G has more than one cycle, then we can find edges e, f such that each
belongs to a cycle that does not contain the other (otherwise, all the cycles
have the same edges). Hence G e f is connected. Continue deleting
edges from cycles until all are gone; at most n 2 edges remain. We now
have an acyclic connected graph with no cycles, since deleting edges on
cycles never disconnects a graph. Such a graph is a tree and hence has
n 1 edges, which is more edges than remain. Hence our assumption that
G has more than one cycle is impossible, and G must have exactly one cycle.
11.31. If d1 , . . . , dn are n natural numbers (with nP 2), then there exists

a tree with vertex degrees d1 , . . . , dn if and only if
di = 2n 2. First we
prove the condition is necessary. If aP
tree T has these degrees, then each
di 1 because T is connected. Also,
di must equal twice the number of
edges in T , and every n -vertex tree has n 1 edges.
We prove sufficiency by induction on n . If n = 2, then the only list
of two natural numbers summing to 2 is (1, 1), and this is realized as the
degrees of the tree with two vertices. For n > 2, we assume that this
condition is sufficient for a list of n 1 natural numbers. Let d1 , . . . , dn be
160
11.32. If T is a tree with m edges and G is a simple graph with (G) m ,

then T is a subgraph of G . We use induction on m . Basis step (m = 0):
every simple graph contains K 1 . Induction step (m > 0): suppose that the
claim holds for trees with fewer than m edges. Since m > 0, Lemma 11.39
allows us to choose a leaf v with neighbor u in T and consider the smaller
tree T 0 = T v . By the induction hypothesis, there is a subgraph H of G
isomorphic to T 0 , since (G) m > m 1. Let x be the vertex in H that
represents u in T 0 . Since T 0 has only m 1 vertices other than u , x has a
neighbor y in G that does not appear in H . Adding the edge x y to represent
uv expands this copy of T 0 in G into a copy of T .
11.33. Inductive proof that K n is a union of k bipartite graphs if and only
if n 2k . For k = 1, one airline can handle at most two cities, because K 3
contains an odd cycle; this is the basis case. For k > 1, suppose n > 2k , and
consider a schedule without odd cycles. If the subgraph colored by the k th
airline is bipartite, then we can partition the cities into two sets with no
edges of color k within them. One of these sets has more than 2k1 cities.
Since the pairs within this set are covered by k 1 subgraphs, the induction
hypothesis guarantees that some airline among the first k 1 can offer an
odd cycle. For the construction when n = 2k , partition the cities into two
sets S, T of equal size. Use copies of the construction for k 1 on each of
S, T , and give all edges between S and T to airline k . If n < 2k , discard
cities from this construction.
11.34. If a graph G has no cycles of even length, then every edge of G
appears in at most one cycle. If G has three pairwise internally-disjoint
w, z -paths, then the lengths of two of them must have the same parity (by
the pigeonhole principle), and the union of these two paths forms a cycle
through w, z of even length. Hence we can prove the claim by proving that
if G has an edge in two cycles, then G has a pair of points joined by three
pairwise internally-disjoint paths.
Let uv be an edge in distinct cycles C 1 and C 2 . Continuing along C 2
from v , let x y be the first edge of C 2 that does not appear in C 1 ; note that
x V (C 1 ). Let z be the first vertex of C 2 after x that appears in C 1 (this
161
may be y . Now C 1 C2 has three pairwise internally-disjoint x, z -paths,

two along C 1 and the third on C 2 .
11.38. The wheel with n vertices has chromatic number 3 when n is odd, 4
when n is even. The cycle with n 1 vertices has chromatic number 2 when
n 1 is even, 3 when n 1 is odd. In every proper coloring, the coloring
used on the remaining vertex of the wheel must be different from all colors
used on the cycle, since it is adjacent to all those vertices, so one additional
color is needed.
11.35. Every tree T has at most one perfect matching.

Proof 1 (contradiction). Suppose M and M 0 are two complete matchings in the tree. Form the symmetric difference of the edge sets, M 4 M 0 .
Since the matchings are complete, each vertex has degree 0 or 2 in the
symmetric difference, so every component is an isolated vertex or a cycle.
Since the tree has no cycle, every vertex must have degree 0 in the symmetric difference, which means that the two matchings are the same. (The
symmetric difference idea is quite powerful.)
Proof 2 (induction). For the basis step, a tree with one vertex has no
perfect matching; a tree with two vertices has one. For the induction step,
consider an arbitrary tree T on n > 2 vertices, and consider a leaf v . In any
perfect matching, v must be matched to its neighbor u . The remainder of
any matching is a matching in T {u, v}. Since each perfect matching in T
must contain the edge uv , the number of perfect matchings in T equals the
number of perfect matchings in T {u, v}. Each component of T {u, v} is a
tree; by the induction hypothesis, each component has at most one perfect
matching. The number of perfect matchings in a graph is the product of
the number of perfect matchings in each component, so the original T has
0 or 1 perfect matching.
11.36. a) If k N, then a k -regular bipartite graph has the same number of
vertices in each partite set. We count the edges in two ways. If the partite
sets are X, Y , then counting the edges by their endpoints in X gives k |X |,
and counting them by their endpoints in Y gives k |Y |. Hence k |X | = k |Y |,
which implies |X | = |Y | since k > 0.
b) For k 1, every k -regular bipartite graph has a perfect matching.
Suppose the bipartition is X, Y . Counting the edges by endpoints in X and
endpoints in Y shows that k |X | = k |Y |, so |X | = |Y |. Hence it suffices to
verify Halls condition; a matching saturating X will be a perfect matching.
Consider S X , and suppose there are m edges between S and N (S). Since
G is k -regular, we have m = k |S|. Since these m edges are incident to N (S),
we have m k |N (S)|. Hence k |S| k |N (S)|, which yields |N (S)| |S|.
Having chosen S X arbitrarily, we have established Halls condition.
11.37. K n,n has n ! perfect matchings and n !(n 1)!/2 cycles of length 2n .
Matchings are pairings of x 1 , . . . , x n with y1 , . . . , yn in some order; thus the
number of matchings is the number of permutations of [n ].
In K n,n , we can list the vertices in order on a cycle (alternating between
the partite sets), in 2(n !)2 ways. Each cycle arises exactly 2n times in this
way, since we obtain the same subgraph no matter where we start the cycle
and no matter which direction we follow.
162
11.39. If G does not have two disjoint odd cycles, then (G) 5. Proof 1
(direct). If G has no odd cycle, then (G) 2, so we may assume that G
has an odd cycle C . If (G V (C)) 3, we have an odd cycle disjoint from
C . Hence (G V (C)) 2, and we can combine this with a 3-coloring of C
to obtain a 5-coloring of G .
Proof 2 (contrapositive). If (G) 6, consider an optimal coloring.
The subgraph induced by vertices colored 1,2,3 coloring must have an odd
cycle, else it would be bipartite and we could replace these three colors
by two. Similarly, the subgraph induced by vertices colored 4,5,6 in the
optimal coloring has an odd cycle, and these two odd cycles are disjoint.
11.40. A graph with maximum degree k has chromatic number at most
k + 1. Such a graph can be colored with k + 1 colors using a greedy
algorithm. Place the vertices in an arbitrary order v1 , . . . , vn . Taking each
vertex in turn, assign it the least-indexed color in {1, . . . , k + 1} that has not
already been used on one of its neighbors. Since it has at most k neighbors,
there is some color left in {1, . . . , k + 1} available to use on it.
Equality holds for the complete graph with k + 1 vertices. Comment:
among connected graphs, complete graphs and odd cycles are the only
graphs where the chromatic number exceeds the maximum degree.
11.41. The segment graph of a collection of lines in the plane with no three
intersecting at a point is 3-colorable. The vertices of G are the points of
intersection of a family of lines; the edges are the segments on the lines
joining two points of intersection. By tilting the picture slightly, we can
insure that no pair of vertices has the same x -coordinate.
Of the neighbors that v has on a single line, one has larger x -coordinate
and one has smaller x -coordinate. Hence if we index V (G) in increasing
order of x -coordinates, each vertex will have at most two earlier neighbors.
Applying the greedy algorithm to this ordering produces a coloring using at
most three colors. Alternatively, if H G , the vertex of H with largest x coordinate has degree at most 2 in H , so (G) 1 + max H G (H ) 3. The
configuration below illustrates that the bound does not hold when more
than two lines are allowed to meet at a point; this configuration can arise
using six lines, and it yields a subgraph with chromatic number 4.
163
164
10 vertices having a planar complement, but the easy counting argument

here is not strong enough to prove that.
11.42. If n = k(k 1) 1, then (G n,k ) > k + 1, where G n,k is the generalized

cycle graph of Example 11.52. If only k + 1 colors are available, then by the
pigeonhole principle some color must be used k 1 times, since (k 2)(k +
1) = k(k 1) 2. Following the n steps around the circle, the minimum
separation between consecutive appearances among the k 1 appearances
of this color is less than k , since the total distance is k(k 1) 1, again by
the pigeonhole principle. Since vertices at most k 1 apart are adjacent,
this prohibits a proper k + 1-coloring.
11.43. The sum of the coefficients of (G; k) is 0 unless G has no edges. The
sum of the coefficients of a polynomial is its value when the argument is
1. The value of (G; 1) is the number of proper 1-colorings of G . This is 0
unless G has no edges.
11.44. A plane graph that is a tree has one face. We use induction on the
number of vertices. With one vertex, every two points other than the point
for the vertex are joined by a segment, except that pairs collinear with
the vertex need a path of two segments. For a tree T with more vertices,
consider two points x, y not in the embedding of T . Deleting a leaf of T and
its incident edge produces a tree T 0 . By the induction hypothesis, there is
a polygonal x, y -path avoiding T 0 . This path also avoids T unless it crosses
the extra edge. Since the path is formed of finitely many segments, it
crosses this edge finitely often. If it crosses the edge more than once, we
can shortcut between two such crossings to reduce the number of crossings
until we have a polygonal x, y -path that crosses T once on this edge and
nowhere else. Now we can cut the path just before and after this crossing
and replace the crossing by a path around the leaf, as suggested below.
11.45. The complement of each simple planar graph with at least 11 vertices is nonplanar. A planar graph with n vertices has at most 3n 6 edges.
Hence each planar graph with 11 vertices has at most 27 edges. Since
K 11 has 55 edges, the complement of each planar subgraph has at least
28 edges and is non-planar. In fact, there is also no planar graph on 9 or
11.46. Every simple planar graph has a vertex of degree at most 5. Theorem
11.65 states that every simple planar graph G with n vertices has at most
3n 6 edges. By the degree-sum formula, the degrees of our graph G thus
sum to at most 6n 12. By the pigeonhole principle, some vertex has degree
at most the average, which equals 6 12/n . Thus some vertex has degree
at most 5.
Every planar graph has chromatic number at most 6. Since multiple
edges do not affect chromatic number, we need only consider simple planar
graphs. We use induction on the number of vertices; every graph with at
most 6 vertices has chromatic number at most 6. When n > 6, we know
that a simple planar graph G with n vertices has a vertex x of degree at
most 5. Let G 0 = G x . Since G 0 is planar, the induction hypothesis implies
that G 0 has chromatic number at most 6. Since at most five of the colors
are used on the neighbors of x , there is some color available to use on x to
extend this to a proper 6-coloring of G .
11.47. A simple n -vertex planar graph with no cycle of length less than k
has at most (n 2)k/(k 2) edges. We may assume that the graph is connected by adding edges if it is not. Thus Eulers Formula applies. Consider
any embedding of the graph in the plane. Each face has at least k edges
on its boundary, and each edge contributes twice to boundaries of faces.
Therefore, counting the appearances of edges in faces grouped according to
the e edges or according to the f faces yields 2e k f .
Since the graph is connected, we can apply Eulers formula, n e + f =
2. Substituting for f in the inequality yields 2e k(2 n + e), which
rearranges to e k(n 2)/(k 2). The formula makes sense also when
k = 2, since the availability of multiple edges means there is no limit on
the total number of edges.
The Petersen graph has 10 vertices, 15 edges, and no cycle of length
less than 5. Since 15 exceeds 5(8)/3, this inequality implies that the Petersen graph is not planar.
11.48. The maximum number of edges in a simple outerplanar graph of
order n is 2n 3. For the lower bound, we provide a construction. A simple
cycle on n vertices together with the chords from one vertex to the n 3
vertices not adjacent to it on the cycle forms an outerplanar graph with
2n 3 edges.
For the upper bound, consider an arbitrary simple outerplanar graph
G . We may assume G is connected, else we can add an edge to join two
components while still having every vertex on the unbounded face F . Now,
since every vertex appears on F , it takes at least n edges to traverse the
165
Chapter 12: Recurrence Relations
boundary of F (note that if G is a tree, then F has length 2n 2). Since

the face lengths
P sum to 2e and every bounded face has length at least 3,
we have 2e =
l(Fi ) n + 3( f 1). Substituting for f in Eulers formula
yields e 2n 3.
Comments: 1) if G has no 3-cycles, the bound becomes e (3n 4)/2.
2) The upper bound can also be proved by a graph transformation, adding a
new vertex inside F and an edge from it to each vertex of G . This produces
an n + 1-vertex planar graph G 0 with n more edges than G . Since e(G 0 )
3(n + 1) 6 edges, we have e(G) 3(n + 1) 6 n = 2n 3.
We choose A and B to satisfy the initial conditions at n = 0 and n = 1.

This requires 1 = A + B and 3 = 3 A + 2 B . Thus A = 1 and B = 0, which
yields the solution claimed.
11.49. There are art galleries with n walls that require bn/3c guards. Three
vertices are used in forming each alcove in the art gallery below. No guard
at one point can see into more than one of them, so at least bn/3c guards
are needed.
12. RECURRENCE RELATIONS

12.1. If an = 3an1 2 for n 1, with a0 = 1, then an = 1 for n 0.
Iterating the recurrence suggests this answer, which is easily proved by
induction. The basis step is given in the initial condition. The induction
step assumes that an1 = 1 and computes an = 3an1 2 = 3 1 2 = 1.
12.2. If an = an1 + 2an2 for n 2, with a0 = 1 and a1 = 8, then an =

3 2n 2(1)n for n 0. The characteristic equation is x 2 x 2 = 0, with
solutions x = 2 and x = 1. The general solution is an = A2n + B(1)n .
This requires 1 = A + B and 8 = 2 A B . Thus A = 3 and B = 2, which
yields the solution claimed.
12.3. If an = 2an1 + 3an2 for n 2, with a0 = a1 = 1, then an = [3n +
(1)n ]/2 for n 0. The characteristic equation is x 2 2x 3 = 0, with
solutions x = 3 and x = 1. The general solution is an = A3n + B(1)n .
This requires 1 = A + B and 1 = 3 A B . Thus A = 1/2 and B = 1/2,
which yields the solution claimed.
12.4. If an = 5an1 6an2 for n 2, with a0 = 1 and a1 = 3, then an = 3n
for n 0. The characteristic equation is x 2 5x + 6 = 0, with solutions
x = 3 and x = 2. The general solution is an = A3n + B 2n .
166
12.5. If an = 3an1 1 for n 1, with a0 = 1, then an = (3n + 1)/2 for n 0.

The characteristic root is 3 and the inhomogenous term is a constant, so
the solution has the form A3n + c. For a particular solution involving the
inhomogeneous term, we require c = 3c 1 to obtain c = 1/2. Now the
initial condition yields 1 = A 30 + 1/2, so A = 1/2. Thus the solution is
an = (3n + 1)/2.
12.6. Analysis of the recurrence an = an1 + n . The characteristic root

of the homogeneous part is 1. When 6= 1, a particular solution bn =
1
n+1 is found by solving Cn = Cn1 + n for C . The general solution
+1
1
is then an = A(1)n + +1
n+1 . This is unbounded for || > 1.
When = 1, a particular solution bn = nn is found by solving Cnn =
C(n 1)n1 + n for C after dividing by and equating corresponding
coefficients. The linear term confirms that = 1, and the constant term
yields C = = 1. The general solution is an = A(1)n + n(1)n for
constant A, and again this is unbounded.
12.7. Recurrence relations with solution formula an = n 3 . Since the solution has no nontrivial exponential part, we seek a first-order relation of the
form an = an1 + f (n). It suffices to set f (n) = n 3 (n 1)3 = 3n 2 3n + 1.
Every homogeneous constant-coefficient first-order linear recurrence
has the form an = can1 , with general solution Ac n . The constant A cannot
be chosen to make Acn be n 3 .
12.8. Recurrence relation to count the pairings of 2n people. The last person can be paired with 2n 1 others, leaving a smaller instance of the
same problem to complete the arrangement, so an = (2n 1)an1 , is the
recurrence, with a0 = 1 (or a1 = 1).
12.9. If n > 0 and no three circles meet at a point, then n pairwise intersecting circles cut the plane into n 2 n + 2 regions. Let an be the number
of regions in such a configuration with n circles. Suppose n > 1, and consider adding the n th circle to the configuration consisting of the first n 1
circles. The n th circle intersects each earlier circle exactly twice. The
2(n 1) intersection points partition the new circle into 2(n 1) arcs, each
of which cuts an old region into two new regions. Other regions remain
unchanged. Hence an = an1 + 2(n 1) for n 2. This recurrence is not
valid when n = 1; we have the initial condition a1 =P2. The solution of a
n
f (i), so we have
recurrence of the form an = an1 + f (n) is an = a1 + i=2
Pn
Pn1
2
an = 2 + i=1 2(i 1) = 2 + 2 i=0 i = n n + 2.
167
12.10. Regions determined by a configuration of n pairwise non-parallel

lines in the plane, when no three line meet at a point. Add a circle that
encloses all the points of intersection, thereby placing a bound on the unbounded regions. Discarding the unbounded portions of the lines yields a
planar graphwith one more region than the original configuration. The
graph has 2n + 2n vertices, one for each pair of lines and two added
on the outer circle. It has n segments from each line, plus 2n arcs on
the
making a total of n 2 + 2n edges. By Eulers Formula,
outer circle,
n
2
+ 2n n 2n + f = 2; thus the graph has n(n + 1)/2 + 2 regions and
2
the original configuration has n(n + 1)/2 + 1 regions, which agrees with
Example 12.11.
which
the Summation
Identity

can be applied. By

Pn
Pnthe SummationnIdentity,
k1
n
n
n
)
=
+
(k
1
+
.
Thus
a
=
a
+
n
0
k=1
k=1 f (n) = 1 + 2 + 4 .
3
2
4
12.11. Recurrence for savings account. To obtain the next amount, add
$100 at the start of the year and then 5 percent at the end, so an =
1.05(an1 + 100).
To solve the recurrence when a0 = 0, we rewrite it as an = 1.05an1 +
105. The form of the solution is A(1.05)n + b0 , where b0 is a constant. The
recurrence yields b0 = 1.05b0 + 105, and hence b0 = 2100. Satisfying
the initial condition a0 = 0 now requires 0 = A(1.05)0 2100, which yields
A = 2100. Thus the sequence is an = 2100(1.05)n 2100.
12.12. Mortgage: Recurrence for $50,000 mortgage with interest per year
at 5 percent of the unpaid amount, followed by payment of $5,000 to end
the year. If an is the amount outstanding at the end of the n th year, then
an = (1.05)an1 5000, with a0 = 50000 (add 5 percent to the previous
amount and then subtract 5000). Iterating this recurrence by calculator
shows that the payment is completed in the 15th year.
12.13. Bounded regions formed by chords among n points on a circle. Let
an be the number of regions.
Pn1
a) an = an1 + f (n) for n 1, where f (n) = n 1 + i=1
(i 1)(n 1 i),
with initial condition a0 = 1. With no points, there is one region. When
we add an n th point, its chords to the n 1 other points add one region for
each region they cut. As they leave the new point, each is cutting a region
(this can be done successively for the n 1 chords), which accounts for the
term n 1. The other term counts the crossings made by the chords as
they continue; after each crossing, a new region is cut. The chord to the i th
vertex separates i 1 vertices from n 1 i vertices, and thus it cuts the
(i 1)(n 1 i) old chords crossing it.
b) Solution to the recurrence. The summation in the formula for f (n)
is an instance of both Exercise 9.11 and Exercise 5.30 (where the value

of the summand should be (i 1)(n i); the value of the sum is n1
.
3
This can also be computed by rearranging (i 1)(n 1 i) into terms to
168
12.14. Solution of an = can1 + f (n) n , where f is a polynomial of degree d

and is a constant. Define bn by setting an = n bn . Substituting into the
recurrence for hai and canceling n yields bn = (c/)bn1 + f (n). If c 6= ,
then Theorem 12.16 yields bn = Acn n + p(n), where p is a polynomial
of degree d . If c = , then Theorem 12.16 yields bn = p(n), where p is a
polynomial of degree d + 1. Multiplying by n yields an = Acn + p(n) n and
an = p(n) n in the two cases.
12.15. Recurrence for moving n spaces. With the options 1,2,3 at each
move, the recurrence is an = an1 + an2 + an3 .
12.16. The number of ways to park the three types of cars in the lot of length
n satisfies an = an1 + 2an2 , with a0 = 1 and a1 = 1. The last car may be
Type 1 (length 1) or Type 2 (length 2) or Type 3 (length 2). In the first Type,
there are an1 ways to complete the list. In the other two Types, there are
an2 ways to complete the list.
12.17. Domino tilings of two-by-n checkerboard are counted by the Fibonacci numbers. When the n th column is a vertical domino, there are
an1 ways to complete the tiling. Otherwise, it must be filled by two horizontal dominos, and then there are an2 ways to complete the tiling. Thus
an = an1 + an2 . For the initial conditions, a0 = a1 = 1. Combinatorially,
every such tiling consists of units of 1 vertical tile or two horizontal tiles,
which establishes a bijection to the 1,2-lists with sum n .
12.18. Shopkeeper making change. The number of ways that the shopkeeper can make change for n cents is the number of lists of 1s, 5s, and
10s that sum to n . If n > 0, every such list has a last element. The last element can be 1, 5, or 10, and in these three cases the number of ways to
complete the rest of the list is an1 , an5 , or an10 , respectively. Hence we
obtain the recurrence an = an1 + an5 + an10 , with the initial conditions
a0 = 1 (there is one such list that sums to 0, the empty list) and an = 0 for
n < 0. Alternatively, one can compute 10 values in a row to start the list;
starting with n = 0, the values of an are 1, 1, 1, 1, 1, 2, 3, 4, 5, 6.
12.19. Identities for Fibonacci numbers. Let Fn denote the n th Fibonacci
number,
F0 = F1 = 1 and Fn = Fn1 + Fn2 for n 2.
Pwhere
n
a)P i=0
Fi2 = Fn Fn+1 for n 0. Proof by induction on n . For n = 0, we
n
have i=0
Fi2 = F02 = 1 = F0 F1 . For the inductive step, suppose the claim
is true for n , where n 0. Using the induction hypothesis after isolating
the last term of the sum, we have
169

n+1
X
i=0
2
Fi2 = Fn+1
+
n
X
i=0
2
Fi2 = Fn+1
+ Fn Fn+1
= Fn+1 (Fn+1 + Fn ) = Fn+1 Fn+2 ,
where the
Pn last step uses the recurrence Fn+2 = Fn+1 + Fn for n 0.
b) i=0
F2i = F2n+1 . For n = 0, both sides equal 1. For the induction
Pn1
Pn
F2i = i=0
F2i . Note that the inducstep, F2n+1 = F2n + F2n1 = F2n + i=0
tion hypothesis is used only for one preceding value.
P2n1
(1)i F2ni = F2n1 .
c) i=0
12.20. Bijection from the set of 1,2-lists that sum to n to the set of 0,1-lists
of length n 1 that have no consecutive 1s. Since the n units of the sum
become n 1 binary positions, we want each 1 or 2 in the sum to contribute
0 or 1 bits, except that one bit will be lost. Convert each 1 to a 0 and
each 2 to a 10, and drop the last bit, which is always a 0 and provides no
information. The 0 following each 1 ensures that there are no consecutive
1s, the length is as desired, and the process is reversible.
Pn
12.21. For the Fibonacci numbers, 1 + i=0
Fi = Fn+2 .
Proof
1
(induction
on
n
).
Since
F
=
1 and F2 = 2, we have 1 +
0
P0
F
=
1
+
1
=
2
=
F
,
so
the
claim
holds
for n = 0. For the induction
i
2
i=0
Pn
Pn1
Fi ) + Fn = 1 +
step, suppose n > 0. We compute 1 + i=0 Fi = 1 + ( i=0
Fn+1 + Fn = 1 + Fn+2 . In the induction step, we only use the induction
Pn1
hypothesis i=0
Fi = Fn+1 for the one most recent value of the induction
parameter. Also the Fibonacci recurrence Fn+1 + Fn is valid for all n 0.
Hence the proof of the induction step is valid for n 1, and we need only
check one value in the basis step.
Proof 2 (combinatorial argument). The right side of the identity
counts lists of 1s and 2s that sum to n + 2. One such list has no 2s;
let that be the list counted by the isolated 1 on the left side. The other lists
have at most n 1s. The number of lists having n i 1s after their last 2 is
the number of lists that sum to i in the portion before that 2, which is Fi .
Summing over all choices for i counts the lists that have a 2.
Pn ni
.
12.22. For the Fibonacci numbers, Fn = i=0
i
i
>
n
i
,
the binomial coefficient
Proof 1 (induction on n ). When

P
is 0, so we can write the sum as i0 ni
without
changing it. When
i
n is 0 or 1, the sum has one nonzero term equal to 1 and the formula
holds. For the induction step, suppose n > 1. We use the recurrence for
Fn , the induction hypothesis
giving
the formula for Fn1 and Fn2 , and the

m1
recurrence relation mk = m1
+
for the binomial coefficients. The
k
k1
resulting computation is
Fn = Fn1 + Fn2 =
170
X n 1 i
i
i0
X n 2 i
i
i0
X n 1 i
n1
n1i
=
+
+
i
i 1
0
i1
i1
X
X

ni
ni
n
+
=
=
.
0
i
i

i1
X
i0
Proof 2 (combinatorial argument). Recall that Fn is the number of

lists of 1s and 2s that add up to n . In a list of 1s and 2s summing to n , if
there are i 2s, then there must be n 2i 1s. Hence there are n i terms all
together, and choosing the i positions for the 2s specifies the list. Summing
over all the possibilities for the number of 2s counts all the lists.
12.23. Combinatorial proof of Fn+m = Fn Fm + Fn1 Fm1 . Consider the 1, 2lists summing to n + m . These are formed by concatenating a sum to n with
a sum to m (in Fn Fm ways) or by concatenating a sum to n 1 with a sum to
m 1 separated by a 2 (in Fn1 Fm1 ways). Since every sum doesnt or does
have a 2 crossing from the first n to the last m , this counts all the sums.
Fn1 divides Fkn1 . Certainly this holds for k = 1. For k > 1, we write
Fkn1 = F(k1)n1+n = F(k1)n1 Fn + F(k1)n2 Fn1 .
By the induction hypothesis, both terms on the right are divisible by Fn1 ,
so Fn1 also divides their sum.
12.24. Every natural number can be written as a sum of distinct Fibonacci
numbers. Every natural number can be expressed as a sum of distinct
powers of 2. More generally, if {an } is a increasing sequence of natural
numbers such that a1 = 1 and an 2an1 for n > 1, then every natural number can be expressed as a sum of distinct terms. The Fibonacci
numbers for n 1 have this property, since Fn1 < Fn1 + Fn2 = Fn and
Fn = Fn1 + Fn2 < 2 Fn1 .
To prove the general statement, note first that an increasing sequence
of natural numbers is unbounded, since any potential bound M N is
exceeded after at most M elements of the sequence. Hence for any natural
number n there is a largest index k such that ak n . We now prove by
strong induction on n that any n N can be expressed as a sum of distinct
ai s. Since a1 = 1, the number 1 can be expressed. Now assume that n > 1
and numbers less than n can be expressed as a sum of distinct ai s. Let k be
the largest index such that ak n . Since n < ak+1 2ak , we have 2ak > n .
This means n ak < ak n . By the strong induction hypothesis, we know
that n ak can be expressed as a sum of distinct values in the sequence.
171
Since n ak < ak , it cannot be that ak is one of the values used. Hence we

can add ak to the expression for n ak to obtain an expression for n as a
sum of distinct ai s. This completes the induction step.
only one step. The pairs taking two steps are (3, 2), (5, 2), and (4, 3), with
(3, 2) (2, 1) (1, 0), (5, 2) (2, 1) (1, 0), and (4, 3) (3, 1) (1, 0).
Induction step: we prove the claim for k = n + 1, assuming that it holds
for k = n and k = n 1. If the algorithm takes n + 1 steps on the pair (a0 , a1 ),
then it takes n steps on (a1 , a2 ) and n 1 steps on (a2 , a3 ). By the induction
hypothesis, a1 + a2 Fn+2 and a2 + a3 Fn+1 . Hence (a1 + a2 ) + (a2 + a3 )
Fn+3 . It suffices to prove that a0 a1 + a2 and a1 a2 + a3 . The reason for
this is that when the Euclidean algorithm replaces (ai1 , ai ) with (ai , ai+1 ),
the numbers satisfy ai1 = kai + ai+1 , where ai+1 < ai . This requires k 1,
and hence ai1 ai + ai+1 . This completes the proof by induction.
To prove that the bound on a0 + a1 is best possible, it suffices to construct an example (a0 , a1 ) taking k steps whose sum is only Fk+2 . Let
(a0 , a1 ) = (Fk+1 , Fk ). These numbers have the desired sum. To show that
the pair takes the proper number of steps, note that (Fk+1 , Fk ) moves to
(Fk , Fk1 ) in the first step. Hence we need only verify that (F3 , F2 ) = (3, 2)
takes two steps, which we checked in the basis step.
12.25. Solution of the Fibonacci recurrence: Fn = Fn1 + Fn2 for n 2,

with F0 = F1 = 1.
a) By the characteristic equation method. The characteristic polynomial
of this recurrence is
x 2 x 1. Its roots, by
the quadratic formula, are
(1 5)/2. Let = (1 + 5)/2 and let = (1 5)/2. For every A, B R,
the sequence with terms A n + B n is a solution of the recurrence, and we
must select A, B to match the initial conditions. From F0 = 1, we require
A + B = 1. From F1 = 1, we require A + B = 1. Using
the first equation in the second and simplifying
5. With the first
leads to A B = 1/
equation,
this yields A =
(1 + 1/ 5)/2 and
B = (1 1/ 5)/2. Factoring
out 1/ 5 yields A = 1 / 5 and B = 2 / 5. Thus
1 1 + 5 n+1
1 1 5 n+1
Fn = (
)
(
)
2
2
5
5
b) By the generating function method. Multiplying the recurrence
by
P
x n and summing over the region of validity (n 2) yields n2 Fn x n =
P
P
P
x n2 Fn1 x n1 + x 2 n2 Fn2 x n2 . Letting F(x) = n0 Fn x n , we rewrite
the equation as F(x) 1 x = x(F(x) 1) + x 2 F(x). Solving for F(x) yields
F(x) = 1/(1 x x 2 ).
2
The
denominator 1 x x factors as (1 x)(1 x), where =
(1 + 5)/2 and = (1 5)/2. Using partial fractions, we write F(x) =
A
B
+ 1x
. Clearing fractions yields 1 = A(1 x) + B(1 x). Equating
1x
yields 1 = A + B and 0 = A B . Thus
corresponding
coefficients of x
A = 15 ( 1+2 5 ) and B = 15 ( 12 5 ), and formula for Fn is as in part (a).
12.26. Efficiency of the Euclidean algorithm. If the Euclidean algorithm

takes k steps on the pair (a0 , a1 ), (where k 2), then it takes k 1 steps on
the pair (a1 , a2 ) that comes next in the sequence and k 2 steps on the pair
(a2 , a3 ). We prove by induction on k that when the algorithm takes k 2
steps, the sum of the numbers in the input pair is at least Fk+2 .
Basis step: we verify the claim for k = 2 and k = 3, because we will
use in the induction step the hypothesis that the claim is true for k = n
and k = n 1. We have F4 = 5 and F5 = 8. We verify (the contrapositive)
that any pair with sum less than 5 can be done in fewer than 2 steps, and
any pair with sum less than 8 can be done in fewer than 3 steps. Here the
contrapositive converts the verification to checking finitely many cases. A
pair involving 1 moves to (1, 0) in one step. There are six other pairs of
natural numbers with sum less than 8. Of these, (2, 2), (4, 2), (3, 3) take
172
12.27. Deterministic prefix reversals end after at most Fn 1 flips. Our

pile is a permutation of 1 through n . When the top card is m , we reverse
the order of the first m cards. The largest card that ever appears will go
to its proper level, and nothing larger can make it come up again. The
largest card that appears later likewise can only appear once later. Thus
the process always stops at some step with 1 at the top.
Let ak be the maximum number of flips given that at most k distinct
cards appear at the top during the process. Let T be the number of the flip
on which the largest card m that appears reaches the top. Next it moves
to position m and never returns. Afterwards, at most k 1 different cards
can appear, so there are at most 1 + ak1 flips after T .
Before flip T , neither m nor 1 reaches the top, and neither does the
card that starts in position m ; call it l . Switch cards 1 and m in the original
stack S to form a stack S 0 . The moves before T will be the same in S and S 0 ,
meaning that the card reaching the top on corresponding flips is the same
for both. Thus 1 reaches the top in T flips for S 0 . If 1 6= l , then l and m
never reach the top in flipping S 0 ; thus T ak2 and ak ak2 + ak1 + 1 in
this case. If 1 = l , then k 1 values may reach the top in flipping S 0 , but
now the flip after flip T in stack S brings 1 to the top, and in total there are
at most ak1 + 1 flips. Thus in each case ak ak2 + ak1 + 1. Since a1 = 0
and a2 = 1, we obtain ak Fk 1 by induction. Finally, at most n different
cards appear at the top when processing a stack with n cards.
12.28. If hai satisfies a k th-order homogeneous linear recurrence with constant coefficients, and the characteristic polynomial of the recurrence factors
173
Q
as p(x) = ri=1 (x i )di for P
distinct 1 , . . . , r , then the solution of the recurrence has the form an = ri=1 qi (n)in , where each qi is a polynomial of
degree less than di . The k coefficients of these polynomials are determined
by the initial values.
A direct proof without the generating function method uses differentiation of the characteristic polynomial (which we do not introduce until
Chapter 16) and linear independence of sequences. To avoid these, we
use the generating function method discussed in this chapter. Indeed, this
solution is sketched in ApplicationP12.37.
k
Let the recurrence be an =
that the chari=1 ci ani for n k , so
Pk
k
acteristic
polynomial is also expressed as p(x) = x i=1 ci x ki . Let
P
A(x) = n=0 ; this is the (formal power series) generating function for the
sequence hai. We show first that A(x) = r (x)/ p(x) for some polynomial r of
P
j
degree less than k . Let C i (x) = k1i
j=0 a j x for 0 i k (with C k (x) = 0).
n
When the recurrence
P is multiplied by x and summed over n k , the term
ci ani becomes ci nk ani x n . After extracting x i , the summation is the
expansion of A(x), missing the terms before x ki that form Ci (x). Thus
c1 x n2 + c2 x n3 + c n1 . Subtracting times the second equation from the

first yields the desired third-order recurrence. Hence every solution to the
original recurrence also satisfies the new recurrence.
For x n = C n , we show that the sequence satisfies the desired recurrence by explicit computation.
A(x) C 0 (x) =
k
X
i=1
ci x i [ A(x) Ci (x)].
We obtain p(x)A(x) = r (x), where r (x) is a polynomial in x of degree less

than k . The coefficients in r (x) depend on the initial conditions a 0 , . . . , ak1
and the constants c1 , . . . , ck .
Q
Using the factorization p(x) = ri=1 (x i )di , P
we rewrite r (x)/ p(x)
using partial fraction expansion. We obtain A(x) = ri=1 qi (x)/(1 i x)di ,
where qi is a polynomial of degree less than di . The polynomials qi are
determined by the initial values
for hai.

P
n+d1 n n
Because (1 x)d =
x (Theorem 5.23), and n+d1
is
n=0
d1
d1
a polynomial in n with degree less than d , we obtain a formula for a n of the
form claimed.
12.29. If hbi and hdi are
to the inhomogeneous
Psolutions
Pk linear k th-order
k
h i (n)x ni and x n = g(n) + i=1
h (n)x ni , then
recurrences x n = f (n) + i=1
Pki
hbi + hdi is a solution to the recurrence x n = f (n) + g(n) + i=1
h i (n)x ni ).
Set x n = bn in the first recurrence and x n = dn in the second. Summing the
two resulting equations yields hbi + hdi as a solution to the third equation.
12.30. If hai is a solution of x n = c1 x n1 + c2 x n2 + c n , where c1 , c2 , c, R,

then hai and C n are solutions of the homogeneous third-order recurrence
x n = (c1 + )x n1 + (c2 c1 )x n2 c2 x n3 .
Every solution to the recurrence satisfies x n = c1 x n1 + c2 x n2 + c n
for all n (beyond the initial conditions), and hence it also satisfies x n1 =
174
(c1 + )C n1 + (c2 c1 )C n2 c2 C n3

= C n + C c1 n1 c1 n2 + c2 n2 c2 n3 = C n
12.31. Consider the recurrence an = an1 + an2 + an3 for n > 3.

If ai = 1 for i {1, 2, 3}, then an 2n2 for n 2. We use induction on
n . Basis step: 2 n 4. For n = 2 and n = 3, we cite the initial conditions.
Since the claim fails for n = 1, instead we use a4 = 1 + 1 + 1 = 3 22 .
Induction step: n > 4. We have an = an1 + an2 + an3 2n3 + 2n4 +
n5
2
= (4 + 2 + 1)2n5 < 2n2 .
If ai = i for i {1, 2, 3}, then an < 2n for n 1. We use induction on n .
Basis step: 1 n 3. This is given in the initial conditions.
Induction step: n > 4. We have an = an1 + an2 + an3 < 2n1 + 2n2 +
n3
2
= (4 + 2 + 1)2n3 < 2n .
12.32. The solution of an = 23 (1 + 3n2+1 )an1 for n 1, with a0 = 1, is an =

2n+1 /(3n + 1). Letting bn = (3n + 1)an , we substitute into the recurrence to
obtain 3n1+1 bn = 23 (1 + 3n2+1 ) 3n11+1 bn1 . Multiplying by 3n + 1 and simplifying
yields bn = 2bn1 . Note that b0 = 2a0 = 2. Thus the solution is bn = 2n+1 ,
which yields an = 2n+1 /(3n + 1).
12.33. Algorithm for finding the largest and smallest in a set of n numbers.
If n = 2, compare the two numbers. If n > 2, find the largest and smallest
in sets of sizes bn/2c and dn/2e (recursively), and compare the largests and
the smallests in these sets.
Letting an denote the number of comparisons made for a set of size n ,
we have an = adn/2e + abn/2c + 2 for n 4, with a1 = 0, a2 = 1, and a3 = 3.
When n = 2k , let bk = an . We obtain bk = 2bk1 + 2 for k 2, with
b1 = 1. To solve this recurrence, the particular solution C is found by C =
2C + 2, yielding C = 2, and the general solution is bk = A2k 2. Since
b1 = 1, we have bk = (3/2)2k 2. In terms of n , this yields an = (3/2)n 2.
12.34. Alternative solution of the recurrence Dn = (n 1)(Dn1 + Dn2 )

(with D0 = 1). Subtracting n Dn1 from both sides yields Dn n Dn1 =
(Dn1 (n 1)Dn2 . Sustituting f n = Dn n Dn1 yields f n = f (n 1).
Since D0 = 1 and D1 = 0, we have f 1 = 1. Thus the solution of the
recurrence for f is f (n) = (1)n .
175
Substituting this yields the first-order recurrenceP

Dn = n Dn1 + (1)n .
We use this to prove by induction on n that Dn = n ! nk=0 (1)k /k !. Basis
step (n = 0): Dn = 1 = 0!(1)0 /0!.
Induction step (n > 0):
P
(1)k
n
Dn = n Dn1 + (1)n = n(n 1)! n1
k=0 k! + (1)
12.38. A bijection from the set of parenthesizations of n + 1 distinct elements to the set of triangulations of a convex n + 2-gon, as suggested by the
figure below. Choose one edge of the (n + 2)-gon as a root, and associate
the n + 1 distinct elements with the other edges, in order. From a parenthesization, we obtain a triangulation. Write the parenthesization on the
root edge. The outermost pairing combines some parenthesization of the
first k elements with some parenthesization of the last n + 1 k , for some
k . Create a triangle using edges from the root edge to the vertex at the
end of the edge for the k th element. Write the parenthesization of the first
k elements on the chord cutting those edges off, and write the parenthesization of the last n + 1 k elements on the other chord. We now have a
(k + 1)-gon for the first group and an (n + 2 k)-gon for the second group.
Proceed recursively by the same method to produce the triangulations in
those polygons.
= n!
Pn1
k=0
(1)k
k!
+ n ! (1)
= n!
n!
Pn
k=0
(1)k
k!
12.35. The number

equivalence
relations on n elements satisfies the

PnBn of
n1
recurrence Bn =
k=1 k1 Bnk , with the initial condition B0 = 1. The
combinatorial argument to prove the recurrence is as follows. Consider
the element x ; it appears in an equivalence class of size k , for some
k with

ways,
and
1 k n . The elements equivalent to x can be chosen in n1
k1
then the partition is completed by using some partition of the remaining
n k elements. By summing over k , we obtain all the partitions.
12.36. Congruence classes of triangles with perimeter n and sides with integer length. Let an be the number of sets {x, y, z} N such that x, y, z are
the lengths of the sides of a triangle with perimeter n .
Option 1. If y + z x + 2, we can subtract 1 from each to get a triangle
on n 3. For any triangle on n 3, we can add 1 to each side to get a triangle
on n satisfying y + z x + 2. If n is even, every triangle on n satisfies
this, so we get an = an3 . If n is odd, there are bdn/2e /2c triangles with
y + z = dn/2e and x = bn/2c that are not counted by an3 . By considering
the possible congruences classes, we have an = an3 + b(n + 1)/4c when n
is odd. The initial conditions are an = 0 for n 0. This can be written as a
single recurrence as an = an3 + 1/2(1 (1)n ) b(n + 1)/4c.
Option 2. Triangles with y 6= z arise by adding 1 to the two smaller
parts of a triangle on n 2. There are bn/3c bn/4c triangles with y = z ,
since there is one such triangle for each y with n/4 < y n/3. Hence
an = an2 + bn/3c bn/4c.
12.37. Parenthesizations of a product of n + 1 elements. Let a0 be the number of parenthesizations. Observe that a0 = 1. The final product in the
parenthesization combines the initial k numbers with the final n + 1 k
numbers, for some k . There are ak1 ways to parenthesize the initial k elements, and there are ank ways to parenthesize the final n + 1 k elements.
Each way to do the first part can combine with each way to
P do the final
part. summing over k yields all parenthesizations, so an = nk=1 ak1 ank .
This is the same recurrence and initial condition satisfied by the Catalan
sequence, so an = Cn for all n .
176
We prove inductively that the function is a bijection. When n = 0, the

root edge and the edge for the element are the same; both get this element
as label. This is the only parenthesization, and the only triangulation is
the empty region. For larger n , the map is a bijection because the final
pairing can be retrieved by knowing the triangle that contains the root
edge, and then the induction hypothesis provides the inverse to retrieve the
parenthesizations of the two groups that are combined in the final pairing.
b bc
(bc)(de)
de
a((bc)(de))
12.39. Catalan numbers count noncrossing pairings of 2n people around a

circular table. A particular individual must pair with someone, and this
pair must leave an even number of intervening people on both sides. If
2k 2 is the number of intervening people onPthe left, then the pairing
can be completed in ak1 ank ways. Thus an = nk=1 ak1 ank . There is one
pairing when n = 0. Thus hai satisfies the Catalan recurrence, and the
solution is the sequence of Catalan numbers. (Exercise 9.26 requests a
bijective proof.)
177
12.40. Catalan numbers count piles of pennies. When the base has length
n , let k be the largest initial number of pennies above which the next row
is occupied; k = 4 in the illustration. the illustration. the illustration. the
illustration. the illustration. To complete such a pile, we complete above
the inital base of length k 1 in row 2 and above the remaining base of
length
Pn k in the base row. Thus the recurrence for the number of piles is
an = nk=1 ak1 ank . There is one pile with a base of no pennies. Thus the
solution is the Catalan numbers.
Pn1
a) an = an1 + i=0
ai for n 2, where a0 = a1 = 1. Let x 1 , . . . , x n be the
vertices of the path in order, and let z be the vertex off the path. There are
an1 spanning trees not using the edge zx n ; they combine the edge x n1 x n
with a spanning tree of K 1 Pn1 . Among trees containing zx n , let i be the
highest index such that all of the path x i+1 , . . . , x n appears in the tree. For
each i , there are ai such trees, since the specified edges are combined with
a spanning tree of K 1 Pi . The term 1 corresponds to i = 0; here the entire
tree is Pn zx n . This exhausts all spanning trees.
b) an = 3an1 an2 for n 3. We dispose of the summation by subtracting one instance of the recurrence formula from the next. For n 3,
Pn2
Pn1
we subtract an1 = an2 + i=0
ai from an = an1 + i=0
ai to obtain
an an1 = an1 an2 an1 , which yields the desired recurrence.
12.41. The n th difference of a polynomial of degree n is the constant n ! times

the leading coefficient of the polynomial. We prove that the first difference
of a polynomial of degree d with leading coefficient a is a polynomial of
degree d 1 with leading coefficient da . Since the n th difference is obtained
by applying the first difference n times, this yields the claimed result.
By the definition of first difference, the first difference of a sum of
polynomials is the sum of their first differences, and the first difference of
c times a polynomial f is c1 f . Thus it suffices to prove the claim for the
pure powers.
We have 1x d = (x + 1)d x d . Expanding (x + 1)d by the Binomial
Theorem shows that the result is a polynomial of degree d 1 with leading
coefficient d , as desired.
12.42. The number s(n, k) of surjective functions from an n -element set A to
a k -element set B satisfies the recurrence s(n, k) = k(s(n 1, k 1)+s(n 1, k)
when n > 1, with initial conditions s(1, 1) = 1 and s(1, k) = 0 for k > 1.
If |A| = 1, then the function covers all of B if and only if |B| = 1; this
verifies the initial conditions. Now suppose n > 1, A = {a1 , . . . , an }, B =
{b1 , . . . , bk }, and f : A B is surjective. If f (an ) = bi , then we can complete
a surjective function in two types of way; either no other element of A has
image bi under f , or some other element of A has image bi under f . In
the first case, we can complete the function in s(n 1, k 1) ways, by
combining f (an ) = bi with a surjective function from A {an } to B {bi }.
In the second case, we can complete the function in s(n 1, k) ways, by
combining f (an ) = bi with a surjective function from A {an } to B . We
have s(n 1, k 1) + s(n 1, k) ways to build a surjective function for
each of the k possible choices for f (an ), so the total number of surjective
functions is s(n, k) = k(s(n 1, k 1) + s(n 1, k).
12.43. For the graph G n sketched below, consisting of a path with n vertices
and one additional vertex adjacent to every vertex of the path, let a n be the
number of spanning trees.
178
12.44. For the graph G n with 2n vertices and 3n 2 edges pictured below,
the chromatic polynomial is
(G n ; k) = (k 2 3k + 3)n1 k(k 1).
Proof 1 (induction on n ). Since G 1 is a 2-vertex tree, (G 1 ; k) = k(k
1). For n > 1, let u n , vn be the two rightmost vertices of G n . The proper
colorings of G n are obtained from proper colorings of G n1 by assigning
colors also to u n and vn . Each proper coloring f of G n1 satisfies f (u n1 ) 6=
f (vn1 ). Thus each such f extends to the same number of colorings of G n .
There are (k 1)2 ways to specify f (u n ) and f (vn ) so that f (u n ) 6=
f (u n1 ) and f (vn ) 6= f (vn1 ). Of these extensions, k 2 give u n and vn the
same color, and we delete them. Since (k 1)2 (k 2) = k 2 3k + 3, the
induction hypothesis yields
(G n ; k) = (k 2 3k + 3) (G n1 ; k) = (k 2 3k + 3)n1 k(k 1).

Proof 2 (induction plus chromatic recurrence). Again (G 1 ; k) = k(k
1). Let e = u n vn . For n > 1, observe that (G n e; k) = (G n1 ; k)(k 1)2
and (G n e; k) = (G n1 ; k)(k 2), by counting the ways to extend each
coloring of G n1 to the last column. Thus
(G n ; k) = (G n u n vn ; k) (G n u n vn ; k)
= (G n1 ; k)[(k 1)2 (k 2)] = (k 2 3k + 3)n1 k(k 2).
179
12.45. Gamblers Ruin recurrence. Let an (r, s) be the probability that A

goes broke on the n th flip. If all of n, r, s are positive, then an (r, s) =
1
a (r + 1, s 1) + 12 an1 (r 1, s + 1). If at least one of n, r, s is 0, then
2 n1
an (r, s) = 0, except when n = r = 0, in which case a0 (0, s) = 1.
P
Letting A(x) = n0 an x n , we obtain A(x) a0 a1 x = c1 x [ A(x) a0 ] +
1 c1 a0 )x+a0
c2 x 2 A(x). Solving for A(x) yields A(x) = (a1c
2 .
1 xc2 x
12.46. Recurrence for partitions of integer n with k parts. The partitions

of n with k parts consist of those that have 1 as a part and those that do
not. Those with 1 as a part correspond to partitions of n 1 with k 1
parts. Those without 1 as a part still have k positive parts when we reduce
each part by 1, so these correspond to partitions of n k with k parts. Thus
pn,k = pn1,k1 + pnk,k . This argument requires k 1 and n k . To enable
computation of values, we need initial conditions p0,0 = 1, pn,0 = 0 for
n > 0, and pn,k = 0 for n < k .
12.47. Solution of the Catalan recurrence. Suppose that a0 = 1 and that

P
P
an = nk=1 ak1 ank for n 1. Rewrite the recurrence as an = n1
k=0 ak an1k
for n 1. The generating
function
method
yields
the
equation
A(x) 1 =
P
x [ A(x)]2 , where A(x) = k0 ak x k . By the quadratic formula, A(x) = (1
1 4x)/(2x). We choose the negative sign on the square root because A

is a formal power series in x , and thus the coefficient
of x 1 must be zero.

1/2
k
The extended binomial theorem yields k (4) as the coefficient of
x k in 1 4x . The Catalan number for n 1 is the coefficient of x n in

12 (1 4x)1/2 /x . Thus

n
Y
1 1/2
1
2n
n+1
n+1
(4)
Cn =
=
(2)
( i)
2 n+1
(n + 1)!
2
i=0
Qn
Qn

n
1
1
2n
2 n Y
i=1 (2i 1)
i=1 (2i)
Q
=
(2i 1) =
=
n
(n + 1)!
n+1
n!
n+1 n
i=1 i
i=0
12.48. Direct combinatorial deriviation of the Fibonacci generating function. We enumerate the 1,2-lists, with the coefficient of x n counting the
lists with sum n . At each position in the list, we can choose 1 or 2, with
the exponent recording the contribution to the sum. Hence our factor for
each position is x 1 + x 2 . When we form lists of length k , the generating
2 k
function is (x
we allow lists of all lengths, the full generating
P+ x ) . Since
function is k0 (x + x 2 )k . This is a formal geometric series, and the sum
is 1/(1 (x + x 2 )), which is 1/(1 x x 2 ).
12.49. If hai satisfies the recurrence an = c1 an1 +c2 an2 , with initial values
1 c1 a0 )x+a0
a0 , a1 , then the generating function for hai is (a1c
2 .
1 xc2 x
n
Multiplying the recurrence
by
x
and
then
summing
over
P
P
Pthe region of
validity (n 2) yields n2 an x n = c1 x n2 an1 x n1 + c2 x 2 n2 an2 x n2 .
180
12.50. Generating function for selections from five types of coins with between 2 and 6 coins of each type. For each type of coin, we make a choice,
independently. We model this choice with x 2 + x 3 + x 4 + x 5 + x 6 . The generating function for the full problem is the fifth power of this, one factor
for each type of coin. The number of such selections with n coins is the
coefficient of x n in (x 2 + x 6 )5 .

Pn
n 2
= 2n
by generating functions. Consider
12.51. Proof of
k=0
k
n

Pn
n
n
.
This
equals
the
left
side
and is the coefficient of x n in the
k=0 k nk

P
P
n k
n l
product of k0 k x and l0 l x . Thus the sum is the coefficient of x n

in (1 + x)n (1 + x)n . This coefficient is 2n
.
n
12.52. Proofs of
convolution identities using generating functions.
Pk m n
a) i=0
= m+n
. The left side is the coefficient of x k in the
i ki
k

P
P
product of the generating functions r 0 mr x r and r 0 nr x r . Since these
generating functions are (1 + x)m and (1 + x)n , the product is (1 + x)m+n .
The right side is indeed
coefficient
of x k in the product.

them+n+1

Pn
m+k nk
b) k=m r
= r +s+1 . Shifting the index of summation on the
s

Pm+n k m+nk
. This is now the coefficient of x m+n in the product
left yields k=0
r
s

P
P
of the generating functions j0 rj x j and j0 sj x j . We discard zeros and

j
j
P
P
P
j
j
j+r
j
r
x .
then shift the index to obtain
j0 r x =
jr r x = x
j0
r
r
(r +1)
as this generating function. SimTheorem
now yields x (1 x)
P 12.35
ilarly, j0 sj x j = x s (1 x)(s+1) . Hence the product is x r +s (1 x)(r +s+2) .
j+r +s12.35, the expansion of thism+ngenerating function is
P Byj+rTheorem
+s+1
x
. To obtain the coefficient of x
, we set j = m+nr s
j0
r +s+1

m+n+1
in the summand and extract the coefficient r +s+1 , as desired.

Pk m+ki1 n+i1
= m+n+k1
. We rewrite the identity as
c)
i=0
ki
i
k

Pk i+n1 ki+m1
k+m+n1
= m+n1 . The left side is the coefficient of x k in
i=0 n1
m1
r
r
P
P
x and r 0 r +n1
x .
the product of the generating functions r 0 r +m1
m1
n1
By Theorem 12.35, these generating functions are (1 x)m and (1 x)n .
The product is (1 x) (m+n) . By Theorem 12.35, the coefficient of x k in its
expansion is k+m+n1
, as desired.
m+n1
12.53.
Generating function for summingP
initial terms. The expression b n =
Pn
n
we
have
b
=
n
k=0 ak is a convolution;
k=0 ak 1. Hence the generating
P
function B(x) = n0 bn x n is the product of the generating functions A(x) =
P
P
n
1
= 1 x n . Thus B(x) = A(x)/(1 x).
n0 an x and (1 x)
181
12.54. The generating function for repeated rolls of a six-sided die by total
xx 7
outcome is 12x+x
7 . When the die is rolled r times, the generating function
P6
i r
is ( i=1 x ) . Since r may be any natural number, we sum over these possibilities. The result is a geometric series missing its initial term. This
simplifies as below:
X 1 x 6 r
x
=
1x
r 1
1
1x
1x 6
1x
1=
1x
1
1 x x(1 x 6 )
(1 x) (1 2x + x 7 )
x x7
=
1 2x + x 7
1 2x + x 7
12.55. Generating function for partitions into parts of size at most k . We

have a factor in the generating for the parts of each size; for each size, we
choose the number of parts. When the size i is as most k , we are allowed to
use any number of parts and have the factor 1 + x i + x 2i + . When i exmust choose none and have the factor 1. Hence the generating
ceeds k , we Q
k
function is i=1
(1 x i )1 .
12.56. Generating functions for partitions

The generating funcQof integers.
(1 + x i ), since each size i can be
tion for partitions into distinct parts is i=1
used 0 times or one time. Each choice of distinct positive integers summing
to n contributes 1 to the coefficient of x n in the product.
Q
The generating function for partitions into odd parts is
i=1 (1
x 2i1 )1 , since each odd size 2i 1 can be used any number of times. Using k copies of 2i 1 corresponds to choosing (x 2i1 )k in the expansion of
1/(1 x 2i1 ). Each choice yielding exponents that sum to n counts one partition of n into odd parts and contributes one to the coefficient of x n in the
formal power series.
12.57. The number of partitions of n into distinct parts equals the number
of partitions of n into odd parts. (Thus the generating functions of Exercise
12.56 are equal.) From a partition into odd parts, we form a partition into
distinct parts by iteratively combining two identical parts until no identical
parts remain. From a partition into distinct parts, we retrieve the unique
partition into odd parts that maps to it by splitting each part i = ji 2ri into
2ri copies of the odd number ji . This works because every positive integer
has a unique expression as an odd number times a power of 2.
182

9-12 Solutions PDF

Uploaded by

Copyright:

Available Formats

9-12 Solutions PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

9-12 Solutions PDF

Uploaded by

Copyright:

Available Formats

125

SOLUTIONS FOR PART III

9.5. If A, B are independent, then A and B c are independentTRUE. P(A

9.6. If A, B are independent, then A c and B c are independentTRUE.

9.15. a) If m 0s and n 1s are placed in some order around a circle, then

The computation is valid only for n 4.

arbitrary permutation of the others, decide which of A and B comes first,

9.21. A perfect game in bowling. If on each roll the probability of a strike

To check that these sum to 1, we compute

9.24. A spinner with regions 1, 2 . . . , n and payoff 2k for region k .

9.25. Another gambler. The amounts in envelopes are a1 a2 an . For

and E(X i ) = P(X i, j = 1) =

ak1 pk1 + ak+1 pk

9.27. Expected number of selections to find the desired key.

X i , and since the vari

For each choice of k1 and k2 , this expression is a polynomial in n . If

of n(n 1) (n r + 1). Summing these 15 multinomial coefficients yields

9.36. Bijection from ballot paths to nondecreasing functions f : [n ] [n ]

1 + 2n + 2n(n 1) + 43 n(n 1)(n 2) + 23 n(n 1)(n 2)(n 3).

The coefficient of x 4 y 8 in the expansion of (x 2 + x y + y 2 )6 is 90. Each

9.37. The number

above the diagonal is a+b

9.38. Bijections involving ballot lists of length 2n .

Chapter 10: Two Principles of Counting

9.40. The Finger Game (see Example 8.20).

10. TWO PRINCIPLES OF COUNTING

Chapter 10: Two Principles of Counting

The result is best possible, because in the set of n even numbers in

10.8. Given five points in a square

center has no pair of points closer together than 2/2.

Chapter 10: Two Principles of Counting

10.15. Every set of 55 numbers in [100] contains two numbers differing by

10.13. If S is a list of n + 1 positive integers with sum k , for k 2n + 1, then

10.16. For n = qk + r with 1 r k , the largest subset of [n ] having no two

n, n 1, . . . , 1, 2n, 2n 1, . . . , n + 1, . . . , mn, mn 1, . . . , (m 1)n + 1.

Chapter 10: Two Principles of Counting

10.18. Multiple exam solutions. In an exam with three true/false questions,

10.21. An extension of Example 10.8. We use the ideas of Exercise 10.16.

Chapter 10: Two Principles of Counting

hands in Ai A j Ak , and no hands belonging to all four sets. There are

10.26. The number of decimal m -tuples with no digit missing. We must

10.30. Euler totient formula.

Chapter 10: Two Principles of Counting

elements of [m ] that are divisible by all the elements of S . By the

10.31. If A1 , . . . , An are sets in a universe U , and N (T ) is the number of

Restrict U to the set U of elements that have all properties indexed by T .

10.32. The number of permutations of [n ] that leave no odd number un

10.33. Assigning 2n courses to n professors. The courses are distinct, and

the ways to assign

10.34. Selection of n objects

Chapter 11: Graph Theory

10.37. Spouse-avoiding circular permutations of n couples. Let U be the

10.38. Counting permutations. Let E nk be the number of permutations of

11. GRAPH THEORY

Chapter 11: Graph Theory

having an additional trail. Altogether, we have used the desired number

11.5. The vertices of a simple graph cannot have distinct degrees. In

Chapter 11: Graph Theory

11.8. If l, m, n are nonnegative integers with l + m = n , then there exists