Berkeley Math Circle 2000-2001

Monthly Contest #6 — Solutions

1. In triangle ABC, let D be the midpoint of side BC. Let E and F be the feet of the perpendiculars to
AD from B and C, respectively. Prove that BE = CF .
Solution: We have ∠DF C = π/2 = ∠DEB. Also, ∠CDF = ∠BDE since they are vertical angles.
(It seems possible that E and F could lie on the same side of D, so that ∠CDF and ∠BDE would be
supplementary rather than equal; however, if they were supplementary and unequal, then one of them,
say ∠BDE, would be > π/2, so the sum of the angles of 4BDE would be > π, a contradiction.)
These equal angles imply 4DF C ∼ 4DEB. But CD = BC/2 = BD, so in fact 4DF C ∼ = 4DEB,
giving CF = BE.
Alternate Solution: (Thanks to Inna Zakharevich) Let H be the foot of the perpendicular from
A to BC. Then, using the bh/2 formula and the fact that D is the midpoint of BC, we get

AD · BE BD · AH CD · AH AD · CF
= Area(4ABD) = = = Area(4ACD) = .
2 2 2 2
Multiplying by 2/AD now gives BE = CF .
2. Let ABC be an equilateral triangle, and let P be a point on minor arc BC of the circumcircle of ABC.
Prove that P A = P B + P C.
Solution: Extend line P C through C to point D such that CD = BP . Note that ∠ACD =
π − ∠P CA = ∠ABP (since quadrilateral ABP C is cyclic), and AC = AB since 4ABC is equilateral.
Consequently, 4ACD ∼ = 4ABP . In particular, we have ∠P DA = ∠CDA = ∠BP A = ∠BCA (again
by cyclicity) = π/3. But also ∠AP D = ∠AP C = ∠ABC (cyclicity) = π/3. We conclude that triangle
AP D is equilateral. So, P A = P D = P C + CD = P C + BP .
Remark: This is a special case of Ptolemy’s Theorem: if RST U is any convex cyclic quadrilateral,
then RS · T U + ST · RU = RT · SU . The proof of the theorem is similar to the above.
3. Determine all triples (x, y, n) of integers such that x2 + 2y 2 = 2n .
Solution: It is easy to check that (±2r , 0, 2r) and (0, ±2r , 2r + 1) satisfy this equation for any
nonnegative integer r. We will show that these are all the solutions by an infinite descent method.
So suppose we have some solution (x0 , y0 , n0 ). If x0 is odd, then 2n0 is odd, which forces n0 = 0 and then
x20 + 2y02 = 1, so y0 = 0 (or else 2y02 > 1 already) and then x0 = ±1. On the other hand, if x0 is even,
we can let x0 = 2x00 and then 4x002 + 2y02 = 2n0 ⇒ y02 + 2x002 = 2n0 −1 , so (x1 , y1 , n1 ) = (y0 , x0 /2, n0 − 1)
is another solution to our equation, where n0 has been replaced by n0 − 1. Now if x1 is even, we can
repeat this construction to get another new solution (x2 , y2 , n2 ) with n0 − 1 replaced by n0 − 2, and
so on. These integers n cannot go on decreasing forever, since there does not exist an integral solution
where n < 0. Thus, eventually our process terminates, which means we get to a solution (xk , yk , nk )
with xk odd. By the above, this is possible only if xk = ±1, yk = 0, nk = 0.
On the other hand, the above construction can be performed in reverse: we have xi = 2yi+1 , yi =
xi+1 , ni = ni+1 + 1 for each value of i ≥ 0. Now we claim that (xi , yi , ni ) = (±2(k−i)/2 , 0, k − i) when
k − i is even, and (0, ±2(k−i−1)/2 , k − i) when k − i is odd. The proof is by downward induction: the
base case i = k is certainly true; given that the statement holds for some i > 0, it is simple algebra
to check that it holds for i − 1 by applying our reverse construction. Thus, the claim is true for each
i ≥ 0. In particular, (x0 , y0 , n0 ) = (±2k/2 , 0, k) or (0, ±2(k−1)/2 , k), which fits the form above. So,
every solution is of this form.
Remark: For those who like heavy √ machinery, this problem can also√be solved quite
√ rapidly by
using unique factorization in the ring Z[ −2], factoring x2 + 2y 2 as (x + −2y)(x − −2y).

1
4. Suppose that S is a set of 2001 positive integers, and n different subsets of S are chosen so that their
sums are pairwise relatively prime. Find the maximum possible value of n. (Here the “sum” of a finite
set of numbers means the sum of its elements; the empty set has sum 0.)
Solution: The answer is 22000 + 1. To see that we cannot do better than this, note that at least
half of the 22001 possible subsets of S have even sums. Indeed, if all elements of S are even, then all
subsets have even sums; on the other hand, if there exists some odd s ∈ S, we can divide the subsets
of S into pairs of the form {T, T ∪ {s}} for each subset T not containing s. Since the sum of T and
that of T ∪ {s} are of opposite parity, each pair contains exactly 1 subset with an even sum. So, in
this case, half the subsets of S have even sums. The upshot is that, in either case, there are at most
22000 subsets of S with odd sums. Since our chosen subsets can include at most one subset whose sum
is even (because no two sums can have a common factor of 2), we cannot choose more than 22000 + 1
subsets altogether.
Now, we must construct an example to show that we can have n = 22000 +1. To do this, let k = (22000 )!,
and let S = {k, 2k, 4k, 8k, . . . , 21999 k, 1}. We consider the 22000 subsets containing the element 1, plus
the one subset {k}. It is evident that k, the sum of the last subset, is relatively prime to the sum of
any subset containing 1, since this latter sum is of the form ak + 1 for some a. So now we just need
to prove that any two distinct subsets containing 1 have relatively prime sums. Well, any such set
consists of several distinct powers of 2, multiplied by k, plus 1. The sum of these powers of 2 is some
number a, 0 ≤ a < 22000 . Thus the subset’s sum is ak + 1. However, it follows from the uniqueness of
binary representation that, for each possible value of a, there is only one subset whose sum is ak + 1.
Consequently, if we choose another, different subset (also containing 1), its sum is bk + 1 for some
b, 0 ≤ b < 22000 with a 6= b. Now suppose ak + 1 and bk + 1 are not relatively prime; then they have
some common prime factor p. So p | ak + 1 and p | bk + 1, hence p | (ak + 1) − (bk + 1) = (a − b)k.
Then, p | a − b or p | k. But a − b is nonzero and has absolute value < 22000 , so a − b is one of the
factors in the product 1 · 2 · 3 · · · 22000 = k, and we get a − b | k. Thus, we are guaranteed that p divides
k. But then p cannot divide ak + 1, so we have a contradiction. We conclude that our subset sums
are, in fact, pairwise relatively prime, completing the proof.
5. Let x1 , x2 , . . . , x1000 , y1 , y2 , . . . , y1000 be 2000 different real numbers, and form the 1000 × 1000 matrix
whose (i, j)-entry is xi + yj . If the product of the numbers in each row is 1, show that the product of
the numbers in each column is −1.
Solution: The given says that (xi + y1 )(xi + y2 ) · · · (xi + y1000 ) = 1 for each i = 1, 2, . . . , 1000. So, if
we let P (x) be the polynomial (x+y1 )(x+y2 ) · · · (x+y1000 )−1, the numbers xi are all roots of P . These
numbers are all distinct, and there are 1000 of them. But P , being of degree 1000, can only have 1000
roots, so these are all the roots of P and the polynomial factors as P (x) = c(x−x1 )(x−x2 ) · · · (x−x1000 )
for some constant c. Since the leading coefficient of P is 1, we conclude that c = 1. Thus,

(x + y1 )(x + y2 ) · · · (x + y1000 ) − 1 = (x − x1 )(x − x2 ) · · · (x − x1000 )

is a polynomial identity, valid for all x.

Now choose any j (1 ≤ j ≤ 1000); we wish to show that the product of the numbers in the jth column
of the matrix is −1. Letting x = −yj in the above equation, we get

(−yj + y1 )(−yj + y2 ) · · · (−yj + y1000 ) − 1 = (−yj − x1 )(−yj − x2 ) · · · (−yj − x1000 )

= (−1)1000 (x1 + yj )(x2 + yj ) · · · (x1000 + yj ).

However, the product on the left-hand side is 0, since the jth factor is −yj + yj = 0; also, (−1)1000 = 1.
Thus, we obtain −1 = (x1 + yj )(x2 + yj ) · · · (x1000 + yj ), which is what we wanted to prove.

Solutions
c 2001, Berkeley Math Circle.

2
 Berkeley Math Circle 2000-2001
Monthly Contest #7 — Solutions

1. Show that there exist infinitely many natural numbers n with the following property: the sum of all
the positive divisors of n, excluding n itself, equals n + 12.
Solution: Let p be any prime number greater than 3; we show that n = 6p has the desired property.
The positive divisors of 6p = 2 · 3 · p are 1, 2, 3, p, 2 · 3, 2 · p, 3 · p, and 2 · 3 · p. The sum of all the factors
other than 6p is equal to 6p + 12, as needed.
It is well known that there are infinitely many prime numbers. Since each value of p gives a different
value for n = 6p, we obtain infinitely many values for n.
2. 5 married couples gather at a party. As they come in and greet each other, various people exchange
handshakes — but, of course, people never shake hands with themselves or with their own respective
spouses. At the end of the party, one woman goes around asking people how many hands they shook,
and she gets nine different answers. How many hands did she herself shake?
Solution: Suppose that there were n couples, and the woman asked all 2n − 1 other attendees how
many hands they shook and received 2n − 1 different answers. We will show that she herself shook
n − 1 hands; hence, in our particular case, the answer is 4.
We work by induction. When n = 1, there is one couple, and no handshakes can occur, proving the
base case. Now suppose the result holds for n couples; we will prove it is valid for n + 1 couples. With
n + 1 couples present, the woman receives 2n + 1 different answers to her question. But no person
P can shake more than 2n hands (for 2n + 2 people, minus P and P ’s spouse); hence, these 2n + 1
numbers must be exactly 0, 1, 2, . . . , 2n in some order. In particular, one of these people, A, shook
everyone else’s hand except A’s own spouse (that accounts for the “2n” answer), and another, B, shook
no hands (the “0” answer). Because B did not shake A’s hand, A and B must be married to each
other. The remaining 2n people include the woman who asked the question, together with those who
answered 1, 2, . . . , 2n − 1 to her question. Now pretend that A and B had not attended the party,
so we are left with n couples. Each of these people shook hands with A and not with B; therefore,
when A and B are removed, their handshake counts become 0, 1, 2, . . . , 2n − 2. Hence, by the induction
hypothesis, the questioner shook n − 1 hands. But now, if we put A and B back in, we note that the
woman shook A’s hand as well (and not B’s). So, altogether, she shook n hands. This completes the
induction step, and now the proof is done.
3. Let ABCD be a square and E a point on side CD. The circle inscribed in triangle ADE touches
DE at F , and the circle inside quadrilateral ABCE, tangent to sides AB, BC, EA, touches AB at G.
Prove that lines AE, BD, and F G meet in a point.
Solution: Extend lines BC and AE to intersect at H. Then the circle inside quadrilateral ABCE,
tangent to AB, BC, and EA, is really the inscribed circle of 4HBA. (Actually, this is only true if
the circle lies inside 4HBA rather than outside it. However, the fact that CE is parallel to AB with
CE < CD = AB readily implies that E lies between H and A, and C lies between H and B, so that
the whole quadrilateral ABCE lies within 4HBA, so the circle drawn inside it does too.) Now let P
be the intersection point of lines BD and AE. Consider the homothety (scaling) about P that sends
point D to point B. Since homotheties preserve directions of lines, this map takes line AD to the line
through B and parallel to AD, namely line HB. Similarly, it takes line DE to line BA. And line EA
passes through P , the center of the homothety, so it goes to itself.
Thus, our homothety takes lines AD, DE, EA to lines HB, BA, AH(= AE), respectively, so it takes
4ADE to 4HBA. Consequently, the incircle of 4ADE is mapped to the incircle of 4HBA, and the
map also matches corresponding tangency points: F goes to G. But if a homothety about P takes F
to G, then P, F, G must be collinear. We now know that P lies on lines AE, BD, and F G, which is
what we need.

1
4. There are 3, 999, 999 cities in Antarctica, and some pairs of them are connected by roads. It is known
that, given any two cities, there is a sequence of roads leading from one to the other. Prove that the
cities can be divided into 1999 groups (of 2001 cities each) such that, given any two cities in the same
group, it is possible to get from one to the other using at most 4000 roads.
Solution: First, we provide some relevant graph-theoretic background. Any finite, connected graph
can be turned into a tree (a connected graph without cycles) by removing some edges. Proof: If our
graph has a cycle, any edge of that cycle can be removed without disconnecting the graph. So remove
this edge, leaving a new graph. If it has a cycle, we can again remove an edge; continuing in this
manner, we must eventually stop, since there are only finitely many edges to remove. We then have a
graph with no cycles; since no edge removal ever disconnected the graph, it must still be connected.
Also, given a tree, we can choose a root vertex r. Then, for any vertex v, there is a unique path from v
to r, never repeating a vertex (uniqueness follows from the absence of cycles). We call v a descendant
of w if this path goes through w. Every vertex is considered to be a descendant of itself and of r.
Suppose v is a descendant of w; then the path from v to r consists of the path from v to w followed
by the path from w to r. It follows that descent is transitive: if w in turn is a descendant of u, then v
is a descendant of u. It also follows that d(v, r) = d(v, w) + d(w, r), where d(x, y) denotes the distance
(i.e. number of edges in the path) from x to y. Finally, a vertex v with no descendants can be removed
and the graph will remain connected. Proof: every other vertex is connected to r by a path that does
not pass through v, so these vertices will remain connected to r — and hence to each other — when v
is removed.
Now we can solve our original problem. We state the graph-theoretic translation: given a connected
graph G on kn vertices (k ≥ 0, n ≥ 1), these vertices can be partitioned into k sets of size n such that
d(v, w) ≤ 2n − 2 whenever v, w are in the same set. (In our case, k = 1999, n = 2001.) We prove
this by induction on k. If k = 0, we form no vertex sets, and the statement is vacuously true. Now
suppose the statement holds for k − 1, where k ≥ 1, and we have a graph G on kn vertices. It suffices
to prove the result when G is a tree, since otherwise we can make it a tree by removing some edges
and partition the vertices of this tree appropriately. The same partition will then work for the original
graph G, since the distance between two vertices cannot increase when we put the deleted edges back.
So suppose G is a tree, and arbitrarily choose a root r. Now let a be a vertex for which d(a, r) is
maximal. Let a = v1 , v2 , . . . , vq = r be the path from a to r, and choose the smallest i such that
vi has at least n descendants. (Some such i surely exists, since r has kn descendants.) Let S be
the set of descendants of vi ; note that if v ∈ S, then every descendant of v is in S, by transitivity.
Notice that v1 , v2 , . . . , vi−1 are all descendants of vi−1 , so the minimality of i implies i − 1 < n. Thus,
d(a, vi ) = i − 1 ≤ n − 1. Now we claim the distance between any two elements of S is at most 2n − 2.
Indeed, suppose b, c ∈ S. We have d(b, vi ) = d(b, r) − d(vi , r) ≤ d(a, r) − d(vi , r) (by choice of a)
= d(a, vi ) ≤ n − 1. Similarly, d(c, vi ) ≤ n − 1, and so d(b, c) ≤ d(b, vi ) + d(vi , c) ≤ 2(n − 1), as claimed.
Now let a1 be an element of S at maximal distance from r. (For example, take a1 = a.) Then
a1 can have no descendants (except itself) in G, since if b were a descendant of a1 , we would have
d(b, r) = d(a1 , b) + d(a1 , r), contradicting maximality. Thus, we can remove a1 from G to leave a graph
G1 , which is still connected — in fact, still a tree with root r. Now a similar argument shows that,
if a2 ∈ S ∩ G1 is chosen to have maximal distance from r, then a2 can have no descendants in G1 :
any descendant would lie in S, by transitivity, and it would also be farther from r than a2 , violating
maximality. So, deleting a2 from G1 gives another rooted tree, G2 . Then, we can choose a3 ∈ S ∩ G2
to be maximally distant from r, and so forth. We continue removing vertices in this manner; since
S has at least n elements, we can remove n vertices. Thus, we choose distinct vertices a1 , a2 , . . . , an ,
all of which lie in S; this means that any two of these vertices are at distance ≤ 2n − 2 from each
other, and the remaining graph, G − {a1 , . . . , an }, is still a tree. Now using the induction hypothesis,
this remaining graph can be partitioned to form the remaining k − 1 sets of vertices, and the desired
partition of G is accomplished.
Remark: In fact, given an arbitrary connected graph with a vertex r selected, we can define v
to be a descendant of w whenever d(v, r) = d(v, w) + d(w, r), and the same solution works, without
having to assume the graph is a tree. However, the case of a tree helps to motivate the definition.

2
5. Let a1 , a2 , a3 , . . . be a sequence of positive integers with the following property: if S is any nonempty
set of positive integers, there exists s ∈ S with as ≤ gcd(S). Prove that n! is divisible by a1 a2 · · · an
for every positive integer n.
Solution: Fix n. Arrange the integers a1 , . . . , an in nonincreasing order, ai1 ≥ ai2 ≥ · · · ≥ ain .
We claim there exists a bijective function f : {1, 2, . . . , n} → {1, 2, . . . , n} such that aik | f (k) for
each k = 1, 2, . . . , n. To demonstrate this, we construct f inductively. Suppose that f (k) has been
defined for all values of k less than some j, and we wish to define f (j). Let d be the greatest common
divisor of i1 , i2 , . . . , ij . By the hypothesis, there exists some i ∈ {i1 , i2 , . . . , ij } such that ai ≤ d; since
aij = min{ai1 , ai2 , . . . , aij }, we have aij ≤ ai ≤ d. Now, we know of j distinct multiples of d lying in
the set {1, 2, . . . , n} (namely, i1 , i2 , . . . , ij ); this many multiples can only exist if jd ≤ n. Then, the
numbers aij , 2aij , . . . , jaij are also all in {1, 2, . . . , n}, since jaij ≤ jd ≤ n. At most j − 1 of these can
have been used up by the previously defined values f (1), f (2), . . . , f (j − 1), so some value is left over;
we define f (j) to be such a value. Then aij | f (j), as required.
So we can recursively define f (1), f (2), . . . , f (n) by the above method, and our construction ensures
that f is injective. Since it maps the finite set {1, 2, . . . , n} to itself, it must actually be bijective. Since
aik | f (k) for each k, we have

a1 a2 · · · an = ai1 ai2 · · · ain | f (1)f (2) · · · f (n) = 1 · 2 · · · n = n!.

Remark: In fact, n! is the smallest positive integer that necessarily satisfies this condition. Indeed,
if p is any prime, then we can define an to be the largest power of p dividing n, for each n, and this
produces a sequence meeting the condition of the problem statement. Then a1 a2 · · · an is the highest
power of p dividing n!, so taking the least common multiple of these values over all choices of p gives
us n!.

Solutions
c 2001, Berkeley Math Circle.

3
 Berkeley Math Circle 2000-2001
Monthly Contest #8 — Solutions

1. Every point of the plane is colored either red or blue. Prove that there exists an equilateral triangle
all of whose vertices are the same color.
Solution: Suppose that no such triangle exists; we will obtain a contradiction. Let ABC be any
equilateral triangle of side 1; then, by assumption, two vertices of 4ABC are one color and the other
vertex is the second color. Without loss of generality, we may suppose A and B are red, and C is blue.
Construct equilateral 4ABD (D 6= C); then D must also be blue. Extend ray √ AB to point E such
that BE = 1, and note that 4CDE is equilateral, with CD = CE = DE = 3. Since C and D are
both blue, E must be red.
But if we draw equilateral 4BEF , with F and C on the same side of AB, then F must be blue (since
B, E are red). Now complete the equilateral triangle CF G (G 6= B). We see that C and F are blue,
so G is red. However, 4AEG is also equilateral with A and E red, so G should be blue. This is our
contradiction, so our assumption — that no monochrome equilateral triangles existed — must be false.

2. The UC Berkeley math department is about to move into a new, one-story building consisting of a
2001 × 2001 square grid of rooms. They would like to install doors between adjacent rooms so that
each room has exactly two doors. Prove that this cannot be done.
Solution: Suppose that it can be done. Color the building in checkerboard fashion, and suppose
that we obtain a white rooms and b black rooms. Each door connects a white room with a black room.
So, if we consider, for each white room, the number of doors adjoining it, we will count each door
exactly once. Since every room is to have 2 doors, the total number of doors will be 2a by this count.
But similarly, if we consider, for each black room, the number of adjoining doors, each door will be
counted once, so that the total number of doors is also equal to 2b. So, 2a = 2b, or a = b. It follows
that the total number of rooms is a + b = 2a, an even number. But we also know the number of rooms
is 20012 , an odd number — contradiction. Hence, the desired condition cannot be met.
3. The manager of Chez Gastropod wants to write a menu consisting of 15 dishes. A “meal” is defined to
be a subset of this menu (possibly empty), but some meals are legal and others are not. The manager
may choose which meals are legal, but there is a requirement that the intersection of any two legal
meals should still be legal. He wants there to be exactly 2001 legal meals. Can he do it?
Solution: The answer is yes. Take any arbitrary 15-element set (menu), and call a collection of
subsets (meals) “valid” if the intersection of any two sets in the collection is again in the collection.
Thus, the objective is to show that there exists a valid collection containing exactly 2001 sets. We
will show, by downward induction, that there exists a valid collection with exactly n sets for each
n, 0 ≤ n ≤ 215 = 32768.
The base case n = 215 is clear: the collection of all subsets of the menu is certainly valid. For the
induction step, suppose that there is a valid collection C of n subsets (1 ≤ n ≤ 32768); we will show
that there is a valid collection of n − 1 subsets. Choose a subset in C of maximum possible size, and
remove it; let C 0 denote the remaining collection, so that it consists of n − 1 subsets. We claim C 0
is still valid. Indeed, if S, T ∈ C 0 , then the removed subset cannot be contained within either S or T
(because of maximality), hence it certainly cannot equal their intersection. But S ∩ T was in C; hence,
it is also in C 0 , as needed. Thus, the claim holds. Now, let n = 2001, and the problem is solved.
4. Given a line segment AB, construct a segment half as long as AB using only a compass. Construct a
segment one-third as long as AB using only a compass.
Solution: First, we provide (part of) an algorithm for circular inversion. Suppose we are given
point O and a circle centered at O of some radius r. If P is a point outside the circle, we wish to
construct point Q on ray OP , satisfying OP · OQ = r2 . Let the circle centered at P , with radius
OP , intersect the given circle centered at O at points X and Y . Then let the circle with center X,

1
 radius r (which is constructible since r = OX), and the circle with center Y , radius r, intersect at O
and Q. We claim this point Q is what we want. Indeed, it is clear from symmetry that OP is the
perpendicular bisector of XY . Since XQ = r = Y Q, Q lies on this bisector as well — that is, on
line OP . Moreover, let H denote the intersection of XY with OP ; then P H < P X = P O ⇒ H lies
on ray OP , and, since Q is the reflection of O across XY , Q will lie on the opposite side of H from
O. This shows that Q lies on ray OP , not just on line OP . Finally, observe that P X = P O and
XO = XQ ⇒ ∠OXP = ∠P OX = ∠QOX = ∠XQO, so, by equal angles, 4P OX ∼ 4XQO. Thus,
OP/OX = QX/QO ⇒ OP · OQ = OX · QX = r2 , as needed.
Now that this is done, we return to the original problem. By scaling, assume AB = 1. By drawing
circles centered at A and B of radius 1, and letting C be one of their intersection points, we obtain
an equilateral 4ABC. Similarly, we successively construct equilateral triangles BCD, BDE (with
D 6= A, E 6= C). We have ∠ABE = ∠ABC + ∠CBD + ∠DBE = 3(π/3) = π, so A, B, E are collinear,
and AE = AB + BE = 2. Then, since we have drawn the circle with center A and radius 1, we can
invert E across it according to the paragraph above, obtaining F such that AF = 1/2. In fact, F lies
on the given segment AB, so the segment AF is fully drawn.
Similarly, to construct a segment of length 1/3, let AB be given; extend AB to E as above so that
BE = 1, and then repeat the process, extending BE to G so that EG = 1. Then AG = 3, and
inverting G across our circle (center A, radius 1) will give what we need.
5. Let a1 = 3 and define an+1 = (3a2n + 1)/2 − an for n ≥ 1. If n is a power of 3, prove that an is divisible
by n.
Solution: The main trick is finding a closed-form expression for an , which requires some experi-
n
mentation. We will show that an = (22 +1 + 1)/3 for all n by induction. It is easy to check that the
formula holds for n = 1. And if it holds for some n, then

3a2n + 1
an+1 = − an
2
n  22n +1 + 1
1  h 22 +1 + 1 i2
= 3 +1 −
2 3 3
n n  22n +1 + 1
1  (22 +1 )2 + 2(22 +1 ) + 1
= +1 −
2 3 3
n+1 n n
1  22 +2 + 2(22 +1 ) + 4  22 +1 + 1
= −
2 3 3
n+1 n n
2 +1 2 +1 2 +1
2 +2 +2 2 +1
= −
3 3
n+1
22 +1 + 1
= ,
3
giving the induction step.
k−1
Now, by Euler’s theorem, 3k divides 22·3 − 1, for any nonnegative integer k, since φ(3k ) (i.e. the
k−1
number of integers in {1, 2, . . . , 3 } relatively prime to 3k ) equals 2 · 3k−1 . But notice that 22·3 − 1 =
k
k−1 k−1 k−1 k−1
(23 − 1)(23 + 1), and 3k−1 is odd ⇒ 23 ≡ 2 (mod 3) ⇒ 23 − 1 is relatively prime to 3k , so,
k 3k−1
in fact, 3 divides 2 + 1. Also, for any integers c, d with d odd, 2 + 1 | 2cd + 1. We conclude that
c
k a k−1
3 | 2 + 1 whenever 3 | a and a is odd.
k 3k
Applying this result twice in succession, we find that 3k | 23 + 1 and then that 3k+1 | 22 +1
+ 1, so
3k
k 2 +1
that 3 | (2 + 1)/3 = a3k for any integer k ≥ 0, and this is what we wanted to prove.

Solutions
c 2001, Berkeley Math Circle.