THE UNIVERSITY OF TORONTO

UNDERGRADUATE MATHEMATICS COMPETITION

March, 2005
Time: 3 12 hours
No aids or calculators permitted.
It is not necessary to do all the problems. Complete solutions to fewer problems are preferred to partial solutions
to many.

1. Show that, if −π/2 < θ < π/2, then

Z θ
log(1 + tan θ tan x)dx = θ log sec θ .
0

R1
2. Suppose that f is continuously differentiable on [0, 1] and that 0
f (x)dx = 0. Prove that
Z 1 Z 1 Z 1
2 f (x)2 dx ≤ |f 0 (x)|dx · |f (x)|dx .
0 0 0

3. How many n × n invertible matrices A are there for which all the entries of both A and A−1 are either 0 or 1?
4. Let a be a nonzero real and u and v be real 3-vectors. Solve the equation

2ax + (v × x) + u = O

for the vector x.

5. Let f (x) be a polynomial with real coefficients, evenly many of which are nonzero, P which is palindromic. This
n
means that the coefficients read the same in either direction, i.e. ak = an−k if f (x) = k=0 ak xk , or, alternatively
f (x) = xn f (1/x), where n is the degree of the polynomial. Prove that f (x) has at least one root of absolute value
1.
6. Let G be a subgroup of index 2 contained in Sn , the group of all permutations of n elements. Prove that G = An ,
the alternating group of all even permutations.
7. Let f (x) be a nonconstant polynomial that takes only integer values when x is an integer, and let P be the set of
all primes that divide f (m) for at least one integer m. Prove that P is an infinite set.
8. Let AX = B represent a system of m linear equations in n unknowns, where A = (aij ) is an m × n matrix,
X = (x1 , · · · , xn )t is an n × 1 vector and B = (b1 , · · · , bm )t is an m × 1 vector. Suppose that there exists at least one
solution for AX = B. Given 1 ≤ j ≤ n, prove that the value of the jth component is the same for every solution X
of AX = B if and only if the rank of A is decreased if the jth column of A is removed.
9. Let S be the set of all real-valued functions that are defined, positive and twice continuously differentiable on a
neighbourhood of 0. Suppose that a and b are real parameters with ab 6= 0, b < 0. Define operators from S to R as
follows:
A(f ) = f (0) + af 0 (0) + bf 00 (0) ;
G(f ) = exp A(log f ) .
(a) Prove that A(f ) ≤ G(f ) for f ∈ S;
(b) Prove that G(f + g) ≤ G(f ) + G(g) for f, g ∈ S;
(c) Suppose that H is the set of functions in S for which G(f ) ≤ f (0). Give examples of nonconstant functions, one
in H and one not in H. Prove that, if λ > 0 and f, g ∈ H, then λf , f + g and f g all belong to H.
10. Let n be a positive integer exceeding 1. Prove that, if a graph with 2n + 1 vertices has at least 3n + 1 edges, then the
graph contains a circuit (i.e., a closed non-self-intersecting chain of edges whose terminal point is its initial point)
with an even number of edges. Prove that this statement does not hold if the number of edges is only 3n.

1
 Solutions.
1. Show that, if −π/2 < θ < π/2, then
Z θ
log(1 + tan θ tan x)dx = θ log sec θ .
0

Solution 1. [D. Han]

Z θ Z θ
log(1 + tan θ tan x)dx = log(1 + tan θ tan(θ − u)du
0 0
θ   
tan θ − tan u
Z
= log 1 + tan θ du
0 1 + tan θ tan u
Z θ
= [log(1 + tan2 θ) − log(1 + tan θ tan u)]du
0
Z θ Z θ
= θ log sec2 θ − log(1 + tan θ tan u)du = 2θ log sec θ − log(1 + tan θ tan x)dx .
0 0

The desired result follows.

Solution 2.
Z θ Z θ
log(1 + tan θ tan x)dx = [log(cos θ cos x + sin θ sin x) − log cos θ − log cos x]dx
0 0
Z θ
= [log cos(θ − x) + log sec θ − log cos x]dx
0
Z θ Z θ
= log cos(θ − x)dx + θ log sec θ − log cos xdx
0 0
Z θ Z θ
= log cos xdx + θ log sec θ − log cos xdx = θ log sec θ .♠
0 0

Rθ
Solution 3. Let F (θ) = 0
log(1 + tan θ tan x)dx. Then, using the substitution u = tan x, we find that
θ
sec2 θ tan x
Z
F 0 (θ) = log(1 + tan2 θ) + dx
0 1 + tan θ tan x)
Z tan θ  
tan θ u tan θ
= log sec2 θ + − + 2 + 2 du
0 1 + (tan θ)u u + 1 u + 1
 tan θ
1 2
= +2 log sec θ + log(1 + (tan θ)u + log(1 + u ) + tan θ[arctan u]
2 0
2 1 2
= 2 log sec θ − log(1 + tan θ) + log sec θ + θ tan θ
2
= log sec θ + θ tan θ .

Also, the derivative of G(θ) ≡ θ log sec θ is equal to log sec θ + θ tan θ. Since F 0 (θ) = G0 (θ) and F (0) = G(0) = 0, it
follows that F (θ) = G(θ) for − π2 < θ < π2 , as desired.
Comment. It is interesting to observe that
θ
log 1 + log sec2 θ
Z
log(1 + tan θ tan x)dx = θ · ,
0 2

the length of the interval times the average of the integrand values at the endpoint.
R1
2. Suppose that f is continuously differentiable on [0, 1] and that 0
f (x)dx = 0. Prove that
Z 1 Z 1 Z 1
2 f (x)2 dx ≤ |f 0 (x)|dx · |f (x)|dx .
0 0 0

2
 Solution 1. Z 1 Z 1
2 f (x)2 dx = f (x)[2f (x) − f (0) − f (1)]dx
0 0
Z 1
= f (x)[(f (x) − f (0)) − (f (1) − f (x)]dx
0
Z 1 Z x Z 1 
= f (x) f 0 (t)dt − f 0 (t)dt dx
0 0 x
Z 1 Z x Z 1
≤ |f (x)| f 0 (t)dt − f 0 (t)dt dx
0 0 x
Z 1 Z x Z 1 
≤ |f (x)| |f 0 (t)|dt + |f 0 (t)|dt dx
0 0 x
Z 1 Z 1 
= |f (x)| |f 0 (t)|dt dx
0 0
Z 1 Z 1
0
= |f (x)|dx · |f (x)|dx .♠
0 0

Rx
Solution 2. Let F (x) = 0
f (t)dt. Then, integrating by parts, we find that
Z 1  1 Z 1 Z 1
f 2 (x)dx = f (x)F (x) − f 0 (x)F (x)dx = − f 0 (x)F (x)dx .
0 0 0 0

Hence Z 1 Z 1 Z 1 
2 0 0
f (x)dx ≤ |f (x)||F (x)|dx ≤ |f (x)|dx · sup{|F (x)| : 0 ≤ x ≤ 1} .
0 0 0

Now Z 1 Z 1 Z 1
2|F (x)| = |F (x) − F (0)| + |F (1) − F (x)| = | f (t)dt| + | f (t)dt| ≤ |f (t)|dt ,
0 x 0

from which the result follows. ♠

3. How many n × n invertible matrices A are there for which all the entries of both A and A−1 are either 0 or 1?
Solution 1. Let A = (aij ) and A−1 = (bij ). Since A and A−1 are both invertible,
Pnthere is at least one 1 in each row
and in each column of either of these matrices. Suppose that 1 ≤ k ≤ n. Since 1 = j=1 akj bjk , there is a unique index
Pn Pn
m for which akm = bmk = 1. If l 6= k, then 0 = j=1 alj bjk = j=1 akj bjl , so that alm = bml = 0. Thus, the mapping
k −→ m is one-one and each row and each column of each matrix contains exactly one 1. Thus A is a permutation
matrix (as is A−1 ), and there are n! possibilities. ♠

Solution 2. Let e = (1, 1, 1, · · · , 1)t . P

Then each component of Ae is a positive integer. The kth component of e is
n
the kth component of A−1 (Ae), namely j=1 bkj (Ae)j, where bij is the (i, j)th component of A−1 . Thus

n
X
1= bkj (Ae)j
j=1

so that there is at most one value of i for which bki 6= 0 and bki = (Ae)i = 1. Since A−1 is invertible, there is at least
one 1 in each row, so there must be exactly one in each row. Similarly, there is at least one 1 in each column, so that
A−1 must be a permutation matrix, as must be A. Therefore, there are exactly n! possibilities. ♠
Pn
Solution 3. [E. Redelmeier] Let A = (aij ) and A−1 = (bij ). Since 1 = j=1 aij bji , there is a unique value of j,
Pn
say j = σ(i) for which aij = bji = 1. Suppose, for i 6= l, that σ(i) = σ(l). Then 0 = j=1 aij bjl ≥ aiσ(i) bσ(l)l = 1, a
contradiction. Hence σ is one-one and so a permutation. Suppose, for k 6= σ(i), aik = aiσ(i) . Then k = σ(l), for some
index l distinct from i and so aij bjl ≥ aiσ(l) bσ(l)l = 1, a contradiction. Hence each row contains exactly one 1 and,
moreover, A is a permutation matrix. So also is A−1 . Conversely, each permutation matrix has the desired property,
and so the number of matrices is n!. ♠

4. Let a be a nonzero real and u and v be real 3-vectors. Solve the equation

2ax + (v × x) + u = O

3
 for the vector x.
Solution 1. Take the dot and cross products by v to get

2a(v.x) + v.u = 0

and
2a(v × x) + v × (v × x) + v × u = O .
The second equation can be manipulated to

2a(v × x) + (v.x)v − (v.v)x + v × u = O

whence
1
v×x=− [(v.x)v − (v.v)x + (v × u)]
2a
1 1 1
= + 2 (v.u)v + (v.v)x − (v × u) .
4a 2a 2a
Therefore
1 1 1
2ax + u = − (v.u)v − (v.v)x + (v × u)
4a2 2a 2a
which can be rearranged to give

[8a3 + 2a(v.v)]x = −(v.u)v + 2a(v × u) − 4a2 u .

Observe that, since v · v > 0, the coefficient of x is nonzero and the same sign as a. ♠
Solution 2. Suppose that v = λu for some real λ. Then, since v × x is orthogonal to both v and x, we must have
that x = (2a)−1 u.
Suppose that v is not a multiple of u. Then we can write u = λv + w where w 6= O and v ⊥ w. Then {v, w, v × w}
is an orthogonal basis for 3-space. Define the operator

T (x) = 2ax + (v × x) .

Then T (v) = 2av, T (w) = 2aw + (v × w), T (v × w) = 2a(v × w) + v × (v × w) = 2a(v × w) − |v|2 w. Then, with
respect to this basis, we obtain the equation
    
2a 0 0 x λ
2
 0 2a −|v|   y  = −  1  .
0 1 2a z 0

Hence, the components of x are given by

λ −2a 1
x=− y= z= .♠
2a 4a2 + |v|2 4a2 + |v|2

Solution 3. Denote the three components of the vectors by subscripts in the usual way. Since v × x = (v2 x3 −
v3 x2 , v3 x1 − v1 x3 , v1 x2 − v2 x1 ), the given equation is equivalent to the system of scalar equations:

2ax1 − v3 x2 + v2 x3 = −u1

v3 x1 + 2ax2 − v1 x3 = −u2
−v2 x1 + v1 x2 + 2ax3 = −u3 .
The determinant D of the coefficients is equal to 2a(4a2 + |v|2 ), and so is nonzero. By Cramer’s Rule, we obtain the
solution:
Dx1 = −(u1 v1 + u2 v2 + u3 v3 )v1 + 2a(v2 u3 − v3 u2 ) − 4a2 u1 ,
Dx2 = −(u1 v1 + u2 v2 + u3 v3 )v2 + 2a(v3 u1 − v1 u3 ) − 4a2 u2 ,
Dx3 = −(u1 v1 + u2 v2 + u3 v3 )v3 + 2a(v1 u2 − v2 u1 ) − 4a2 u3 .♠

5. Let f (x) be a polynomial with real coefficients, evenly many of which are nonzero, P which is palendromic. This
n
means that the coefficients read the same in either direction, i.e. ak = an−k if f (x) = k=0 ak xk , or, alternatively

4
 f (x) = xn f (1/x), where n is the degree of the polynomial. Prove that f (x) has at least one root of absolute value
1.
Solution. If n = deg f (x) is odd, then k and n − k have opposite parity and
X
f (−1) = {ak ((−1)k + (−1)n−k ) : 0 ≤ k < n/2} = 0 .

Suppose that n = 2m. Then f (x) has 2m + 1 coefficients: m pairs (ak , an−k ) of equal coefficients (0 ≤ k ≤ m − 1)
and am . Since evenly many coefficients are nonzero, we must have that am = 0. Hence
m−1
X
f (x) = ak (xk + x2m−k )
k=0
m−1
X
= xm ak (x−(m−k) + xm−k )
k=0
m
X
= xm am−j (x−j + xj ) .
j=1
Pm Pm
Then f (eiθ ) = 2emiθ j=1 am−j cos jθ. Let P (θ) = j=1 am−j cos jθ. Then eiθ is a root of f if and only if P (θ) = 0.
R 2π
Now 0 P (θ)dθ = 0 and P (θ) is not constant (otherwise f would be identically 0), so P (θ) is a continuous real
function that assumes both positive and negative values. Hence, by the Intermediate Value Theorem, P must vanish
somewhere on the interval [0, 2π], and the result follows. ♠

6. Let G be a subgroup of index 2 contained in Sn , the group of all permutations of n elements. Prove that G = An ,
the alternating group of all even permutations.
Solution 1. Let u ∈ Sn \G. Since G is of index 2, uG = Gu whence u−1 Gu = G. Thus, x−1 gx ∈ G for each element
g in G. If t is a transposition, than x−1 tx runs through all transpositions as x runs through G. Therefore, if G contains
a transposition, then G must contain every transposition, and so must be all of Sn , giving a contradiction. Thus, every
transposition must belong to uG. Hence, if t1 and t2 are two transposition, then t1 = g1 u and t2 = t−1 2 = g2 u for g1
and g2 in G. Hence, t1 t2 = (g1 u)(u−1 g2−1 ) = g1 g2−1 ∈ G. Thus, G contains every even permutation, and so contains An .
The result follows.
Solution 2. As in Solution 1, we see that G must be normal, and so must be the kernel of a homomorphism φ from
Sn into the multiplicative group {1, −1}. Suppose that x ∈ Sn and x has order m, then 1 = φ(xm ) = (φ(x))m , whence
m is even or x ∈ G. It follows that every element of odd order must belong to G, and so in particular, G must contain all
3−cycles. Since (a, b)(a, c) = (a, b, c) and (a, b)(c, d) = (a, b, c)(a, d, c) (multiplying from the left), every product of a pair
of transpositions and hence every even permutation is a product of 3−cycles. Hence G contains all even permutations
and so must coincide with An .

7. Let f (x) be a nonconstant polynomial that takes only integer values when x is an integer, and let P be the set of
all primes that divide f (m) for at least one integer m. Prove that P is an infinite set.
Solution 1. Suppose that pk (x) is a polynomial of degree k assuming integer values at x = n, n + 1. · · · , n + k. Then,
there are integers ci for which      
x x x
pk (x) = ck,0 + ck,1 + · · · + ck,k .
k k−1 0
To see this, first observe that xk , k−1 x
, · · ·, x0 constitute a basis for the vector space of polynomials of degree not

exceeding k. So there exist real ck,i as specified. We prove by induction on k that the ck,i must in fact be integers. The
result is trivial when k = 0. Assume its truth for k ≥ 0. Suppose that
 
x
pk+1 (x) = ck+1,0 + · · · + ck+1,k+1
k+1
takes integer values at x = n, n + 1, · · · , n + k + 1. Then
 
x
pk+1 (x + 1) − pk+1 (x) = ck+1,0 + · · · + ck+1,k
k
is a polynomial of degree k which taken integer values at n, n + 1, · · · , n + k, and so ck+1,0 , · · · , ck+1,k are all integers.
Hence,    
n n
ck+1,k+1 = pk+1 (n) − ck+1,0 − · · · ck+1,k
k+1 1

5
is also an integer. ♣ (This is more than we need; we just need to know that the coefficients of f (x) are all rational.)
Let f (x) be multiplied by a suitable factorial to obtain a polynomial g(x) with integer coefficients. The set of
primes dividing values of g(m) at integers m is the union of the set of primes for f and a finite set, so it is enough to
obtain the result for g. Note that g assumes the values 0 and 1 only finitely often. Suppose that g(a) = b 6= 0 and let
P = {p1 , p2 , · · · , pr } be a finite set of primes. Define
g(a + bp1 p2 · · · pr x)
h(x) = .
b
Then h(x) has integer coefficients and h(x) ≡ 1 (mod p1 p2 · · · pr ). There exists an integer u for which h(u) is divisible
by a prime p, and this prime must be distinct from p1 , p2 , · · · , pr . The result follows. ♠
Pn
Solution 2. Let f (x) = k ak xn . The number a0 = f (0) is rational. Indeed, each of the numbers f (0), f (1), · · ·,
f (n) is an integer; writing these conditions out yields a system of n + 1 linear equations with integer coefficients for
the coefficients a0 , a1 , · · ·, an whose determinant is nonzero. The solution of this equation consists of rational values.
Hence all the coefficients of f (x) are rational. Multiply f (x) by the least common multiple of its denominators to get
a polynomial g(x) which takes integer values whenever x is an integer. Suppose, if possible, that values of f (x) for
integral x are divisible only by primes p from a finite set Q. Then the same is true of g(x) for primes from a finite set P
consisting of the primes in Q along with the prime divisors of the Q least common multiple. For each prime p ∈ P , select
a positive integer ap such that pap does not Q bdivide g(0). Let N = {pap : p ∈ P }. Then, for each integer u, g(N u) 6≡ 0
(mod N ). However, for all u, g(N u) = p p , where 0 ≤ bp ≤ ap . Since there are only finitely many numbers of this
type, some number must be assumed by g infinitely often, yielding a contradiction. (Alternatively: one could deduce
that g(N u) ≤ N for all u and get a contradition of the fact that |g(N u)| tends to infinity with u.) ♠
Solution 3. [R. Barrington Leigh] Let n be the degree of f . Lemma. Let p be a prime and k a positive integer. Then
f (x) ≡ f (x + pnk ) (mod pk ). Proof by induction on the degree. The result holds for n = 0. Assume that it holds
for n = m − 1 and f (x) have degree m. Let g(x) = f (x) − f (x − 1), so that the degree of g(x) is m − 1. Then
nk
p
X
nk
f (x + p ) − f (x) = g(x + i)
i=1
p(n−1)k
X
= (g(x + i) + g(x + i + p(n−1)k ) + . . . + g(x + i + (pk − 1)p(n−1)k )
i=1
p(n−1)k
X
≡ pk g(x + i) ≡ 0 ,
i=1

(mod pk ). [Note that this does not require the coefficients to be integers.] ♣
Suppose, if possible, that the set P of primes p that divide at least one value of f (x) forQinteger x is finite, and
that, for each p ∈ P , the positive integer a is chosen so that pa does not divide f (0). Let q = {pa : p ∈ P }. Then pa
does not divide f (0), nor any of the values f (q n ) for positive integer n, as these are all congruent modulo pa . Since any
prime divisor of f (q n ) belongs to P , it must be that f (q n ) is a divisor of q. But this contradicts the fact that |f (q n )|
becomes arbitrarily large with n. ♠

8. Let AX = B represent a system of m linear equations in n unknowns, where A = (aij ) is an m × n matrix,

X = (x1 , · · · , xn )t is an n × 1 vector and B = (b1 , · · · , bm )t is an m × 1 vector. Suppose that there exists at least one
solution for AX = B. Given 1 ≤ j ≤ n, prove that the value of the jth component is the same for every solution X
of AX = B if and only if the rank of A is decreased if the jth column of A is removed.
Solution 1. Let C1 , C2 , · · · , Cn be the n columns of A. Suppose that the dimension of the span of these columns
is reduced when Cj is removed. Then Cj is not a linear combination of the other Ci . Thus, if AX = AY = B and
Z = X − Y , then AZ = O. This means that z1 C1 + z2 C2 + · · · + zj Cj + · · · + zn Cn = 0. Because Cj is not a linear
combination of the other Ci , we must have that zj = 0, i.e. that xj = yj .
On the other hand, if the dimension of the span is the same whether or not Cj is present, then Cj must be
a linear combination of the remaining Ci . In this case, there is a vector Z with a nonzero value of zj for which
AZ = z1 C1 + · · · + zj Cj + · · · + zn Cn = O. Thus if X is a solution of AX = B, then X + Z is a second solution with a
different value of the jth component. ♠
Solution 2. As in the first solution, we observe that every solution of AX = B has the same jth entry if and only if
every solution of AX = O has zero in the jth position.
Suppose that every solution of AX = O has 0 in the jth position. Let Aj the the matrix A with the jth column
removed and Xj be the vector X with the jth entry removed. Then there is a linear isomorphism between the solutions
of AX = 0 and Aj Xj = 0, where X → Xj is defined by suppressing the jth entry. Since the target space of Aj has
dimension n − 1. we find that
n − rank A = nullity A = nullity Aj = (n − 1) − rank Aj ,

6
whence rank Aj = rank A − 1.
On the other hand, suppose that there are solutions of AX = O for which the jth component fails to vanish. Each
solution of Aj Y = 0 lifts to a solution of AX = 0 by inserting the entry 0 after the (j − 1)th position of Y . However,
this lifting does not get all the solutions, so that

n − rank A = nullity A > nullity Aj = (n − 1) − rank Aj

whence rank Aj > rank A − 1. Since, always, the rank of Aj does not exceed that of A, we must have that the ranks of
Aj and A are equal. The desired result follows. ♠

9. Let S be the set of all real-valued functions that are defined, positive and twice continuously differentiable on a
neighbourhood of 0. Suppose that a and b are real parameters with ab 6= 0, b ≤ 0. Define operators from S to R as
follows:
A(f ) = f (0) + af 0 (0) + bf 00 (0) ;
G(f ) = exp A(log f ) .
(a) Prove that A(f ) ≤ G(f ) for f ∈ S;
(b) Prove that G(f + g) ≤ G(f ) + G(g) for f, g ∈ S;
(c) Suppose that H is the set of functions in S for which G(f ) ≤ f (0). Give examples of nonconstant functions, one
in H and one not in H. Prove that, if λ > 0 and f, g ∈ H, then λf , f + g and f g all belong to H.
Solution. (a) The result if clear if A(f ) ≤ 0. Suppose that A(f ) > 0. We have that (log f )0 = f 0 /f and
(log f )00 = (f 00 /f ) − (f 0 /f )2 , so that
 0 2
af 0 (0) bf 00 (0) f (0)
A(log f ) = log f (0) + + −b .
f (0) f (0) f (0)

We have that
af 0 (0) bf 00 (0)
 
log A(f ) = log f (0) + log 1 + +
f (0) f (0)
af 0 (0) bf 00 (0)
≤ log f (0) + +
f (0) f (0)
 0 2
f (0)
= A(log f ) + b ≤ A(log f ) .
f (0)

(b) Note that, for any functions u, v ∈ S,

G(uv) = exp A(log u + log v) = exp[A(log u) + A(log v)] = exp A(log u) · exp A(log v) = G(u)G(v) .

Similarly G(u/v) = G(u)/G(v). From (a), we have the inequalities

       
f f g g
A ≤G and A ≤G .
f +g f +g f +g f +g

Adding these two inequalities yields

G(f ) + G(g)
1≤
G(f + g)
from which the desired result follows.
(c) Suppose that λ > 0 and f, g ∈ H. Then

G(λf ) = G(λ)G(f ) = λG(f ) ≤ λf (0) ;

G(f g) = G(f )G(g) ≤ f (0)g(0) = (f g)(0) ;

G(f + g) ≤ G(f ) + G(g) ≤ f (0) + g(0) = (f + g)(0) .

Suppose that a 6= 0, and let f (x) = ekx . Then G(f ) = eka . If ka > 0, then f 6∈ H; if ka < 0, then f ∈ H. Suppose
2
that a = 0, and let f (x) = ekx . Then G(f ) = e2kb . If k < 0, then f 6∈ H; if k > 0, then f ∈ H. ♠

10. Let n be a positive integer exceeding 1. Prove that, if a graph with 2n + 1 vertices has at least 3n + 1 edges, then the
graph contains a circuit (i.e., a closed non-self-intersecting chain of edges whose terminal point is its initial point)
with an even number of edges. Prove that this statement does not hold if the number of edges is only 3n.

7
 Solution 1. If there are two vertices joined by two separate edges, then the two edges together constitute a chain
with two edges. If there are two vertices joined by three distinct chains of edges, then the number of edges in two of the
chains have the same parity, and these two chains together constitute a circuit with evenly many edges. We establish
the general result by induction.
The result can be checked for n = 2. Suppose that it holds for 2 ≤ n ≤ m − 1. We may assume that we have a graph
G that contains no instances where two separate edges join the same pair of vertices and no two vertices are connected
by more than two chains. Since 3n + 1 > 2n, the graph is not a tree, and therefore must contain at least one circuit.
Consider one of these circuits, L. If it has evenly many edges, the result holds. Suppose that it has oddly many edges,
say 2k + 1 with k ≥ 1. Since any two vertices in the circuit are joined by at most two chains (the two chains that make
up the circuit), there are exactly 2k + 1 edges joining pairs of vertices in the circuit. Apart from the circuit, there are
(2m + 1) − (2k + 1) = 2(m − k) vertices and (3m + 1) − (2k + 1) = 3(m − k) + k ≥ 3(m − k) + 1 edges.
We now create a new graph G0 , by coalescing all the vertices and edges of L into a single vertex v and retaining all
the other edges and vertices of G. This graph G0 contains 2(m − k) + 1 vertices and at least 3(m − k) + 1 edges, and so
by the induction hypothesis, it contains a circuit M with an even number of edges. If this circuit does not contain v,
then it is a circuit in the original graph G, which thus has a circuit with evenly many edges. If the circuit does contain
v, it can be lifted to a chain in G joining two vertices of L by a chain of edges in G0 . But these two vertices of L must
coincide, for otherwise there would be three chains joining these vertices. Hence we get a circuit, all of whose edges lie
in G0 ; this circuit has evenly many edges. The result now follows by induction.
Here is a counterexample with 3n edges. Consider 2n + 1 vertices partitioned into a singleton and n pairs. Join
each pair with an edge and join the singleton to each of the other vertices with a single edge to obtain a graph with
2n + 1 vertices, 3n edges whose only circuits are triangles. ♠
Solution 2. [J. Tsimerman] For any graph H, let k(H) be the number of circuits minus the number of components
(two vertices being in the same component if and only if they are connected by a chain of edges). Let G0 be the
graph with 2n + 1 vertices and no edges. Then k(G0 ) = −(2n + 1). Suppose that edges are added one at a time to
obtain a succession Gi of graphs culminating in the graph G = G3n+1 with 2n + 1 vertices and 3n + 1 edges. Since
adding an edge either reduces the number of components (when it connects two vertices of separate components) or
increases the number of circuits (when it connects two vertices in the same component), k(Gi+1 ) ≥ k(Gi ) + 1. Hence
k(G) ≥ −(2n + 1) + (3n + 1) = n. Thus, the number of circuits in G is at least equal to the number of components in
G plus n, which is at least n + 1. Thus, G has at least n + 1 circuits.
Since 3(n + 1) > 3n + 1, there must be two circuits that share an edge.
A counterexample can be obtained by taking a graph with vertices A1 , · · ·, An , B0 , B1 , · · ·, Bn , with edges joining
the vertex pairs (Ai , Bi−1 ), (Ai , Bi ) and (Bi−1 , Bi ) for 1 ≤ i ≤ n. ♠