Solutions to the 75th William Lowell Putnam Mathematical Competition

Saturday, December 6, 2014

Kiran Kedlaya and Lenny Ng

A1 The coefficient of xn in the Taylor series of (1 − x + For example, for n = 5,

x2 )ex for n = 0, 1, 2 is 1, 0, 12 , respectively. For n ≥ 3,
the coefficient of xn is −1 2 0 0 0
 
2 −8 6 0 0 
1 1 1 1 − n + n(n − 1) B= 0

6 −18 12 0 .

− + =
n! (n − 1)! (n − 2)! n! 0 0 12 −32 20 
n−1 0 0 0 20 −20
= .
n(n − 2)!
Let C denote the matrix obtained from B by replacing
If n − 1 is prime, then the lowest-terms numerator is the bottom-right entry with −2n2 (for consistency with
clearly either 1 or the prime n − 1 (and in fact the latter, the rest of the diagonal). Expanding in minors along
since n−1 is relatively prime to n and to (n−2)!). If n− the bottom row produces a second-order recursion for
1 is composite, either it can be written as ab for some det(C) solving to det(C) = (−1)n (n!)2 ; a similar ex-
a 6= b, in which case both a and b appear separately in pansion then yields det(B) = (−1)n−1 n!(n − 1)!.
(n − 2)! and so the numerator is 1, or n − 1 = p2 for Remark: This problem and solution are due to one of
some prime p, in which case p appears in (n − 2)! and us (Kedlaya). The statement appears in the comments
so the numerator is either 1 or p. (In the latter case, on sequence A010790 (i.e., the sequence (n − 1)!n!) in
the numerator is actually 1 unless p = 2, as in all other the On-Line Encyclopedia of Integer Sequences (http:
cases both p and 2p appear in (n − 2)!.) //oeis.org), attributed to Benoit Cloitre in 2002.
A2 Let v1 , . . . , vn denote the rows of A. The determinant is A3 First solution: Using the identity
unchanged if we replace vn by vn − vn−1 , and then vn−1
by vn−1 − vn−2 , and so forth, eventually replacing vk by (x + x−1 )2 − 2 = x2 + x−2 ,
vk − vk−1 for k ≥ 2. Since vk−1 and vk agree in their first
k k
k − 1 entries, and the k-th entry of vk − vk−1 is 1k − k−1 1
, we may check by induction on k that ak = 22 + 2−2 ; in
the result of these row operations is an upper triangular particular, the product is absolutely convergent. Using
matrix with diagonal entries 1, 21 −1, 31 − 12 , . . . , n1 − n−1
1
. the identities
The determinant is then
x2 + 1 + x−2
n 
1 1
 n 
−1
 = x − 1 + x−1 ,
− = x + 1 + x−1
∏ k−1 ∏
k=2 k k=2 k(k − 1) x2 − x−2
= x + x−1 ,
(−1)n−1 x − x−1
= .
(n − 1)!n! we may telescope the product to obtain
Note that a similar calculation can be made whenever k k
22 − 1 + 2−2
∞   ∞
A has the form Ai j = min{ai , a j } for any monotone se- 1
∏ 1 − = ∏ 2k −2k
quence a1 , . . . , an . Note also that the standard Gaussian k=0 ak k=0 2 + 2
elimination algorithm leads to the same upper triangu- k+1 k+1 k k
∞
22 + 1 + 2−2 22 − 2−2
lar matrix, but the nonstandard order of operations used =∏ · −k−1
2k −2k k+1
22 − 22
here makes the computations somewhat easier. k=0 2 + 1 + 2
0 0
Remark: The inverse of A can be identified explicitly: 22 − 2−2 3
for n ≥ 2, it is the matrix B given by = 0 0 = .
2
2 +1+2 −2 7


 −1 i= j=1 Second solution: (by Catalin Zara) In this solution, we
2
−2i


 1<i= j<n do not use the explicit formula for ak . We instead note
Bi j = −(n − 1)n i = j = n first that the ak form an increasing sequence which can-



 ij |i − j| = 1 not approach a finite limit (since the equation L = L2 −2
 has no real solution L > 2), and is thus unbounded. Us-
0 otherwise.

ing the identity

ak+1 + 1 = (ak − 1)(ak + 1),

 2

one checks by induction on n that note crucially that these equations also hold for n ∈
{0, 1}. Therefore, the function f : [0, +∞) → R given
n  
1 2 an+1 + 1 by
∏ 1 − ak = 7 a0 a1 · · · an .
k=0
f (t) = |w|2t − t 2 |w|2 + 2t(t − 1)Re(w) − (t − 1)2
Using the identity
satisfies f (t) = 0 for t ∈ {0, 1, i + 1, j + 1}. On the other
a2n+2 − 4 = a4n+1 − 4a2n+1 , hand, for all t ≥ 0 we have

one also checks by induction on n that f 000 (t) = (2 log |w|)3 |w|2t > 0,
2 so by Rolle’s theorem, the equation f (3−k) (t) = 0 has at
q
a0 a1 · · · an = a2n+1 − 4.
3 most k distinct solutions for k = 0, 1, 2, 3. This yields
the desired contradiction.
Hence
Remark: By similar reasoning, an equation of the form
n
 
1 3 an+1 + 1 ex = P(x) in which P is a real polynomial of degree
∏ 1 − = q
k=0 ak 7 a2 − 4 d has at most d + 1 real solutions. This turns out to
n+1
be closely related to a concept in mathematical logic
3 known as o-minimality, which in turn has deep conse-
tends to 7 as an+1 tends to infinity, hence as n tends to quences for the solution of Diophantine equations.
infinity.
Second solution: (by Noam Elkies) We recall a result
A4 The answer is 13 . commonly known as the Eneström-Kakeya theorem.
First solution: Let an = P(X = n); we want the min- Lemma 1. Let
imum value for a0 . If we write Sk = ∑∞ k
n=1 n an , then
the given expectation values imply that S1 = 1, S2 = 2, f (x) = a0 + a1 x + · · · + an xn
S3 = 5. Now define f (n) = 11n − 6n2 + n3 , and note
that f (0) = 0, f (1) = f (2) = f (3) = 6, and f (n) > 6 be a polynomial with real coefficients such that 0 < a0 ≤ a1 ≤
for n ≥ 4; thus 4 = 11S1 − 6S2 + S3 = ∑∞ n=1 f (n)an ≥ · · · ≤ an . Then every root z ∈ C of f satisfies |z| ≤ 1.
6 ∑n=1 an . Since ∑n=0 an = 1, it follows that a0 ≥ 31 .
∞ ∞

Equality is achieved when a0 = 31 , a1 = 12 , a3 = 16 , and Proof. If f (z) = 0, then we may rearrange the equality 0 =
an = 0 for all other n, and so the answer is 31 . f (z)(z − 1) to obtain
Second solution: (by Tony Qiao) Define the probabil- an zn+1 = (an − an−1 )zn + · · · + (a1 − a0 )z + a0 .
ity generating function of P as the power series
∞ But if |z| > 1, then
n
G(z) = ∑ P(x = n)z .
n=0 |an zn+1 | ≤ (|an − an−1 | + · · · + |a1 − a0 |)|z|n ≤ |an zn |,

We compute that G(1) = G0 (1) = G00 (1) = G000 (1) = 1. contradiction.

By Taylor’s theorem with remainder, for any x ∈ [0, 1],
there exists c ∈ [x, 1] such that Corollary 2. Let

(x − 1)2 (x − 1)3 G0000 (c) f (x) = a0 + a1 x + · · · + an xn

G(x) = 1 + (x − 1) + + + (x − 1)4 .
2! 3! 4! be a polynomial with positive real coefficients. Then every
In particular, G(0) = + 1 1 0000
for some c ∈ [0, 1]. root z ∈ C of f satisfies r ≤ |z| ≤ R for
3 24 G (c)
However, since G has nonnegative coefficients and c ≥
r = min{a0 /a1 , . . . , an−1 /an }
0, we must have G0000 (c) ≥ 0, and so G(0) ≥ 31 . As in the
first solution, we see that this bound is best possible. R = max{a0 /a1 , . . . , an−1 /an }.

A5 First solution: Suppose to the contrary that there exist Proof. The bound |z| ≤ R follows by applying the lemma to
positive integers i 6= j and a complex number z such that the polynomial f (x/R). The bound |z| ≥ r follows by applying
Pi (z) = Pj (z) = 0. Note that z cannot be a nonnegative the lemma to the reverse of the polynomial f (x/r).
real number or else Pi (z), Pj (z) > 0; we may put w =
z−1 6= 0, 1. For n ∈ {i + 1, j + 1} we compute that Suppose now that Pi (z) = Pj (z) = 0 for some z ∈ C and
some integers i < j. We clearly cannot have j = i+1, as
wn = nw − n + 1, wn = nw − n + 1; then Pi (0) 6= 0 and so Pj (z) − Pi (z) = (i + 1)zi 6= 0; we
thus have j − i ≥ 2. By applying Corollary 2 to Pi (x),
we see that |z| ≤ 1 − 1i . On the other hand, by applying
 3

Corollary 2 to (Pj (x) − Pi (x))/xi−1 , we see that |z| ≥ Remark: The reader may notice a strong similarity be-
1 tween this solution and the first solution. The primary
1 − i+2 , contradiction.
difference is we compute that Ez0 (0) ≥ 0 instead of dis-
Remark: Elkies also reports that this problem is his
covering that Ez (−1) = 0.
submission, dating back to 2005 and arising from work
of Joe Harris. It dates back further to Example 3.7 in: Remark: It is also possible to solve this prob-
Hajime Kaji, On the tangentially degenerate curves, J. lem using a p-adic valuation on the field of alge-
London Math. Soc. (2) 33 (1986), 430–440, in which braic numbers in place of the complex absolute value;
the second solution is given. however, this leads to a substantially more compli-
cated solution. In lieu of including such a solution
Remark: Elkies points out a mild generalization which
here, we refer to the approach described by Victor
may be treated using the first solution but not the sec-
Wang here: http://www.artofproblemsolving.
ond: for integers a < b < c < d and z ∈ C which is
com/Forum/viewtopic.php?f=80&t=616731.
neither zero nor a root of unity, the matrix
  A6 The largest such k is nn . We first show that this
1 1 1 1 value can be achieved by an explicit construction. Let
a b c d
e1 , . . . , en be the standard basis of Rn . For i1 , . . . , in ∈
za zb zc zd {1, . . . , n}, let Mi1 ,...,in be the matrix with row vectors
ei1 , . . . , ein , and let Ni1 ,...,in be the transpose of Mi1 ,...,in .
has rank 3 (the problem at hand being the case a = Then Mi1 ,...,in N j1 ,..., jn has k-th diagonal entry eik · e jk ,
0, b = 1, c = i + 1, d = j + 1). proving the claim.
Remark: It seems likely that the individual polynomi- We next show that for any families of matrices Mi , N j as
als Pk (x) are all irreducible, but this appears difficult to described, we must have k ≤ nn . Let V be the n-fold ten-
prove. sor product of Rn , i.e., the vector space with orthonor-
Third solution: (by David Feldman) Note that mal basis ei1 ⊗ · · · ⊗ ein for i1 , . . . , in ∈ {1, . . . , n}. Let mi
be the tensor product of the rows of Mi ; that is,
Pn (x)(1 − x) = 1 + x + · · · + xn−1 − nxn .
n
If |z| ≥ 1, then mi = ∑ (Mi )1,i1 · · · (Mi )n,in ei1 ⊗ · · · ⊗ ein .
i1 ,...,in =1

n|z|n ≥ |z|n−1 + · · · + 1 ≥ |zn−1 + · · · + 1|,

Similarly, let n j be the tensor product of the columns of
with the first equality occurring only if |z| = 1 and the N j . One computes easily that mi · n j equals the product
second equality occurring only if z is a positive real of the diagonal entries of Mi N j , and so vanishes if and
number. Hence the equation Pn (z)(1 − z) = 0 has no so- only if i 6= j. For any ci ∈ R such that ∑i ci mi = 0, for
lutions with |z| ≥ 1 other than the trivial solution z = 1. each j we have
Since !
0= ∑ ci mi · n j = ∑ ci (mi · n j ) = c j .
Pn (x)(1 − x)2 = 1 − (n + 1)xn + nxn+1 , i i

it now suffices to check that the curves Therefore the vectors m1 , . . . , mk in V are linearly inde-
pendent, implying k ≤ nn as desired.
Cn = {z ∈ C : 0 < |z| < 1, |z|n |n + 1 − zn| = 1}
Remark: Noam Elkies points out that similar argument
are pairwise disjoint as n varies over positive integers. may be made in the case that the Mi are m × n matrices
and the N j are n × m matrices.
Write z = u + iv; we may assume without loss of gener-
ality that v ≥ 0. Define the function B1 These are the integers with no 0’s in their usual base
10 expansion. If the usual base 10 expansion of N is
Ez (n) = n log |z| + log |n + 1 − zn|. dk 10k + · · · + d0 100 and one of the digits is 0, then there
exists an i ≤ k − 1 such that di = 0 and di+1 > 0; then
One computes that for n ∈ R, Ez00 (n) < 0 if and only if we can replace di+1 10i+1 + (0)10i by (di+1 − 1)10i+1 +
(10)10i to obtain a second base 10 over-expansion.
u−v−1 u+v−1
<n< . We claim conversely that if N has no 0’s in its usual
(1 − u)2 + v2 (1 − u)2 + v2
base 10 expansion, then this standard form is the unique
In addition, Ez (0) = 0 and base 10 over-expansion for N. This holds by induc-
tion on the number of digits of N: if 1 ≤ N ≤ 9,
1 then the result is clear. Otherwise, any base 10 over-
Ez0 (0) = log(u2 + v2 ) + (1 − u) ≥ log(u) + 1 − u ≥ 0
2 expansion N = dk 10k + · · · + d1 10 + d0 100 must have
d0 ≡ N (mod 10), which uniquely determines d0 since
since log(u) is concave. From this, it follows that the
equation Ez (n) = 0 can have at most one solution with
n > 0.
 4

N is not a multiple of 10; then (N − d0 )/10 inherits the 1 ≤ x ≤ 3; then F(1) = F(3) = 0 and F(x) ≤ min{x −
base 10 over-expansion dk 10k−1 + · · · + d1 100 , which 1, 3 − x}. Using integration by parts, we obtain
must be unique by the induction hypothesis. Z 3 Z 3
f (x) F(x)
Remark: Karl Mahlburg suggests an alternate proof of dx = dx
uniqueness (due to Shawn Williams): write the usual 1 x 1 x2
Z 2 Z 3
expansion N = dk 10k + · · · + d0 100 and suppose di 6= 0 x−1 3−x
≤ dx + dx
for all i. Let M = cl 10l + · · · + c0 100 be an over- 1 x2 2 x2
expansion with at least one 10. To have M = N, we 4
must have l ≤ k; we may pad the expansion of M with = log .
3
zeroes to force l = k. Now define ei = ci − di ; since
1 ≤ di ≤ 9 and 0 ≤ ci ≤ 10, we have 0 ≤ |ei | ≤ 9. More- (Some minor adjustment is needed to make this com-
over, there exists at least one index i with ei 6= 0, since pletely rigorous, e.g., approximating f uniformly by
any index for which ci = 10 has this property. But if i is continuous functions.)
the largest such index, we have
B3 First solution: Assume by way of contradiction that A
i−1 has rank at most 1; in this case, we can find rational
10i ≤ ei 10i = − ∑ ei 10i numbers a1 , . . . , am , b1 , . . . , bn such that Ai j = ai b j for
j=0 all i, j. By deleting rows or columns, we may reduce to
i−1 the case where the ai ’s and b j ’s are all nonzero.
≤ ∑ ei |10i ≤ 9 · 10i−1 + · · · + 9 · 100 , Recall that any nonzero rational number q has a unique
j=0
prime factorization
a contradiction. q = ±2c1 3c2 5c3 · · ·
B2 In all solutions, we assume that the function f is inte-
grable. with exponents in Z. Set

First solution: Let g(x) be 1 for 1 ≤ x ≤ 2 and −1 c(q) = (c1 , c2 , c3 , . . . ).

for 2 < x ≤ 3, and define h(x) = g(x) − f (x). Then
R3
1 h(x) dx = 0 and h(x) ≥ 0 for 1 ≤ x ≤ 2, h(x) ≤ 0
Note that |ai b j | is prime if and only if c(ai ) + c(b j ) has
for 2 < x ≤ 3. Now one entry equal to 1 and all others equal to 0. The con-
Z 3 Z 2 Z 3
dition that m + n distinct primes appear in the matrix
h(x) |h(x)| |h(x)| implies that the vector space
dx = dx − dx
1 x 1 x 2 x ( )
Z 2 Z 3
|h(x)| |h(x)|
≥
1 2
dx −
2 2
dx = 0, ∑ xi c(ai ) + ∑ yi c(b j ) : xi , y j ∈ R, ∑ xi = ∑ y j
i j i j
R 3 f (x) R 3 g(x)
and thus 1 x dx ≤ 1 x dx = 2 log 2−log 3 = log 43 . contains a linearly independent set of size m + n. But
Since g(x) achieves the upper bound, the answer is that space evidently has dimension at most m + n − 1,
log 34 . contradiction.
Reformulation: (by Karl Mahlburg and Karthik Second solution: In this solution, we use standard
Adimurthi) Since f is integrable, it can be expressed terminology of graph theory, considering only sim-
in terms of the Hadamard basis ple undirected graphs (with no self-loops or multiple
 edges). We first recall the quick induction proof that
1
 x ∈ [1, 2) that a graph on k vertices with no cycles contains at
H0 (x) = −1 x ∈ [2, 3] most k − 1 edges: for k = 1, the claim is trivially true

0 x∈/ [1, 3] because there can be no edges. For k > 1, choose any
vertex v and let d be its degree. Removing the vertex v
Hn+1 (x) = Hn (2(x − 1) + 1) + Hn (2(x − 2) + 1). and the edges incident to it leaves a disjoint union of d
different graphs, each having no cycles. If the numbers
More precisely, we have f (x) = ∑n cn Hn (x) for some cn
of vertices in these graphs are k1 , . . . , kd , by induction
with |c0 | + |c1 | + | · · · | ≤ 1. Let gn = 13 (Hn (x)/x) dx; it
R
the total number of edges in the original graph is at most
is easy to show that the gn are strictly decreasing in n. (k1 − 1) + · · · + (kd − 1) + d = k − 1.
Thus
Returning to the original problem, suppose that A has
Z 3
4 rank at most 1. Draw a bipartite graph whose vertices
( f (x)/x)dx = c0 g0 + c1 g1 + · · · ≤ 1 · g0 = log . correspond to the rows and columns of A, with an edge
1 3
joining a particular row and column if the entry where
Second solution: (Art of Problem Solving, user they intersect has prime absolute value. By the previ-
libra_gold) Define the function F(x) = 1x f (t) dt for ous paragraph, this graph must contain a cycle. Since
R
 5

the graph is bipartite, this cycle must be of length 2k for so fn (−2−n+2 j ) has overall sign (−1) j , proving the
some integer k ≥ 2 (we cannot have k = 1 because the claimed alternation.
graph has no repeated edges). Without loss of gener- Remark: Karl Mahlburg suggests an alternate interpre-
ality, we may assume that the cycle consists of row 1, 2
tation of the preceding algebra: write 2− j fn (2−n+2 j ) as
column 1, row 2, column 2, and so on. There must then
exist distinct prime numbers p1 , . . . , p2k such that 2 2
2− j − 2−( j−1) + · · · + (−1) j−1 2−1 + (−1) j 2−1
|A11 | = p1 , |A21 | = p2 , . . . , |Akk | = p2k−1 , |A1k | = p2k . +(−1) j 2−1 + (−1) j+1 2−1 + (−1) j+2 2−2 + · · · ,
However, since A has rank 1, the 2 × 2 minor A11 Ai j − where the two central terms (−1) j 2−1 arise from split-
Ai1 A1 j must vanish for all i, j. If we put ri = |Ai1 | and ting the term arising from x j . Then each row is an alter-
c j = Ai j /A11 , we have nating series whose sum carries the sign of (−1) j unless
it has only two terms. Since n ≥ 3, one of the two sums
p1 · · · p2k = (r1 c1 )(r2 c1 ) · · · (rk ck )(r1 ck ) is forced to be nonzero.
= (r1 c1 · · · rk ck )2 , Remark: One of us (Kedlaya) received this problem
and solution from David Speyer in 2009 and submitted
which contradicts the existence of unique prime factor- it to the problem committee.
izations for positive rational numbers: the prime p1 oc-
curs with exponent 1 on the left, but with some even B5 We show that Patniss wins if p = 2 and Keeta wins if
exponent on the right. This contradiction completes the p > 2 (for all n). We first analyze the analogous game
proof. played using an arbitrary finite group G. Recall that for
any subset S of G, the set of elements g ∈ G which com-
B4 Define the polynomial fn (x) = ∑nk=0 2k(n−k) xk . Since mute with all elements of S forms a subgroup Z(S) of
G, called the centralizer (or commutant) of S. At any
f1 (x) = 1 + x, f2 (x) = 1 + 2x + x2 = (1 + x)2 , given point in the game, the set S of previously cho-
sen elements is contained in Z(S). Initially S = 0/ and
the claim holds for for n = 1, 2. For n ≥ 3, we show that
Z(S) = G; after each turn, S is increased by one ele-
the quantities
ment and Z(S) is replaced by a subgroup. In particular,
fn (−2−n ), fn (−2−n+2 ), . . . , fn (−2n ) if the order of Z(S) is odd at some point, it remains
odd thereafter; conversely, if S contains an element of
alternate in sign; by the intermediate value theorem, this even order, then the order of Z(S) remains even there-
will imply that fn has a root in each of the n intervals after. Therefore, any element g ∈ G for which Z({g})
(−2−n , −2−n+2 ), . . . , (−2n−2 , −2n ), forcing fn to have has odd order is a winning first move for Patniss, while
as many distinct real roots as its degree. any other first move by Patniss loses if Keeta responds
with some h ∈ Z({g}) of even order (e.g., an element of
For j ∈ {0, . . . , n}, group the terms of fn (x) as a 2-Sylow subgroup of Z({g})). In both cases, the win
is guaranteed no matter what moves follow.
···
Now let G be the group of invertible n × n matrices with
+ 2( j−5)(n− j+5) x j−5 + 2( j−4)(n− j+4) x j−4 entries in Z/pZ. If p > 2, then Z(S) will always contain
+ 2( j−3)(n− j+3) x j−3 + 2( j−2)(n− j+2) x j−2 the scalar matrix −1 of order 2, so the win for Keeta is
guaranteed. (An explicit winning strategy is to answer
+ 2( j−1)(n− j+1) x j−1 + 2 j(n− j) x j + 2( j+1)(n− j−1) x j+1 any move g with the move −g.)
+ 2( j+2)(n− j−2) x j+2 + 2( j+3)(n− j−3) x j+3 If p = 2, we establish the existence of g ∈ G such that
+ 2( j+4)(n− j−4) x j+4 + 2( j+5)(n− j−5) x j+5 Z({g}) has odd order using the existence of an irre-
ducible polynomial P(x) of degree n over Z/pZ (see
··· .
remark). We construct an n × n matrix over Z/pZ with
characteristic polynomial P(x) by taking the companion
Depending on the parity of j and of n − j, there may matrix of P(x): write P(x) = xn + Pn−1 xn−1 + · · · + P0
be a single monomial left on each end. When evaluat- and set
ing at x = −2−n+2 j , the trinomial evaluates to 0. In the
binomials preceding the trinomial, the right-hand term 0 0 · · · 0 −P0
 
dominates, so each of these binomials contributes with 1 0 · · · 0 −P1 
the sign of x j−2k , which is (−1) j . In the binomials fol-
0 1 · · · 0 −P 
g= 2 .
. . .
lowing the trinomial, the left-hand term dominates, so  .. .. . . ... .. 
. 
again the contribution has sign (−1) j .
0 0 · · · 1 −Pn−1
Any monomials which are left over on the ends also
contribute with sign (−1) j . Since n ≥ 3, there ex- In particular, det(g) = (−1)n P0 6= 0, so g ∈ G. Over
ists at least one contribution other than the trinomial, an algebraic closure of Z/pZ, g becomes diagonaliz-
 6

able with distinct eigenvalues, so any matrix commut- contradiction

ing with g must also be diagonalizable, and hence of
odd order. In particular, Z({g}) is of odd order, so Pat- n = (ms − nr)s0 + (r0 n − ms0 ) ≥ s + s0 > n;
niss has a winning strategy. 0
hence rs , rs0 remain consecutive in Fn .
Remark: It can be shown that in the case p = 2, the
only elements g ∈ G for which Z({g}) has odd order Let fn : [0, 1] → R be the piecewise linear function
are those for which g has distinct eigenvalues: in any which agrees with f at each element of Fn and is linear
other case, Z({g}) contains a subgroup isomorphic to between any two consecutive elements of Fn . Between
0
the group of k × k invertible matrices over Z/2Z for any two consecutive elements rs , rs0 of Fn , fn coincides
some k > 1, and this group has order (2k − 1)(2k − 0
with some linear function a + bx. Since s f ( rs ), s0 f ( rs0 ) ∈
2) · · · (2k − 2k−1 ). Z, we deduce first that
Remark: We sketch two ways to verify the existence
of an irreducible polynomial of degree n over Z/pZ for r0 r
b = ss0 ( f ( 0 ) − f ( ))
any positive integer n and any prime number p. One s s
is to use Möbius inversion to count the number of ir-
is an integer of absolute value at most K, and second
reducible polynomials of degree n over Z/pZ and then 0
give a positive lower bound for this count. The other that both as = s f ( rs ) − br and as0 = s0 f ( rs0 ) − br0 are
is to first establish the existence of a finite field F of integral. It follows that fn is integral.
cardinality pn , e.g., as the set of roots of the polynomial We now check that if n > 2K, then fn = fn−1 . For this,
n
x p −1 inside a splitting field, and then take the minimal it suffices to check that for any consecutive elements
polynomial of a nonzero element of F over Z/pZ which r m r0
s , n , s0 in Fn , the linear function a0 +b0 x matching f n−1
is a primitive (pn − 1)-st root of unity in F (which exist 0
from rs to rs0 has the property that f ( mn ) = a0 + b0 mn .
because the multiplicative group of F contains at most Define the integer t = n f ( mn ) − a0 n − b0 m. We then
one cyclic subgroup of any given order). One might be compute that the slope of fn from rs to mn is b0 + st,
tempted to apply the primitive element theorem for F 0
over Z/pZ, but in fact one of the preceding techniques while the slope of fn from mn to rs0 is at most b0 − s0t.
0
In order to have |b0 + st| , |b0 − s t| ≤ K, we must have
is needed in order to verify this result for finite fields,
as the standard argument that “most” elements of the (s + s0 ) |t| ≤ 2K; since s + s0 = n > 2K, this is only pos-
upper field are primitive breaks down for finite fields. sible if t = 0. Hence fn = fn−1 , as claimed.

One may also describe the preceding analysis in terms It follows that for any n > 2K, we must have fn =
of an identification of F as a Z/pZ-vector space with fn+1 = · · · . Since the condition on f and K implies that
the space of column vectors of length n. Under such f is continuous, we must also have fn = f , completing
an identification, if we take g to be an element of F − the proof.
{0} generating this group, then any element of Z({g}) Remark: The condition on f and K is called Lipschitz
commutes with all of F − {0} and hence must define an continuity.
F-linear endomorphism of F. Any such endomorphism Remark: An alternate approach is to prove that for
is itself multiplication by an element of F, so Z({g}) each x ∈ [0, 1), there exists ε ∈ (0, 1 − x) such that the
is identified with the multiplicative group of F, whose restriction of f to [x, x + ε) is linear; one may then de-
order is the odd number 2n − 1. duce the claim using the compactness of [0, 1]. In this
B6 Let us say that a linear function g on an interval is inte- approach, the role of the Farey sequence may also be
gral if it has the form g(x) = a + bx for some a, b ∈ Z, played by the convergents of the continued fraction of x
and that a piecewise linear function is integral if on ev- (at least in the case where x is irrational).
ery interval where it is linear, it is also integral. Remark: This problem and solution are due to one
For each positive integer n, define the n-th Farey se- of us (Kedlaya). Some related results can be proved
quence Fn as the sequence of rational numbers in [0, 1] with the Lipschitz continuity condition replaced by suit-
with denominators at most n. It is easily shown by in- able convexity conditions. See for example: Kiran S.
0 Kedlaya and Philip Tynan, Detecting integral polyhe-
duction on n that any two consecutive elements rs , rs0
dral functions, Confluentes Mathematici 1 (2009), 87–
of Fn , written in lowest terms, satisfy gcd(s, s0 ) = 1,
109. Such results arise in the theory of p-adic differen-
s + s0 > n, and r0 s − rs0 = 1. Namely, this is obvious
tial equations; see for example: Kiran S. Kedlaya and
for n = 1 because F1 = 01 , 11 . To deduce the claim for
0 Liang Xiao, Differential modules on p-adic polyannuli,
Fn from the claim for Fn−1 , let rs , rs0 be consecutive el- J. Inst. Math. Jusssieu 9 (2010), 155–201 (errata, ibid.,
ements of Fn−1 . If s + s = n, then for m = r + r0 we
0
669–671).
0 0
have rs < mn < rs0 and the pairs rs , mn and mn , rs0 satisfy the
0
desired conditions. Conversely, if s + s > n, then we
0
cannot have rs < mn < rs0 for a ∈ Z, as this yields the