Interesting arguments, ideas

Victor Wang
October 23, 2013

1. Hmm this is pretty tricky and I wish I actually knew what it meant but here is an application anyway.
(Adapted from Gabriel Dospinescu, 2010 MR U160). Let p be a prime and let n, s be positive integers.
Prove that
 
X n
vp (1)k k s vp (n!).
k
p|k,0kn

hf dkal; f

Solution: Let the sum
be S. Then by a roots of unity filter and the Stirling number representation
of k s (define 00 = 00 = 1),
p1 X
n  
X n s
pS = k ( j )k
j=0
k
k=0
p1 X
n   s  
X n j k
X k
= ( ) S(s, `) `!
j=0 k=0
k `
`=0
p1 min(s,n) n   
X X X k n
= S(s, `)`! ( j )k
j=0
` k
`=0 k=0
p1 min(s,n) n 
 X 
X X n n`
= S(s, `)`! ( j )k
j=0
` k `
`=0 k=`
p1 min(s,n)  
X X n
= S(s, `)`! ( j )` (1 j )n` .
j=0
`
`=0

If j = 0 and ` < n, then we get a zero term. Also, if s n, then from the ` = n terms we get
p1
X
n! ( j )n = pn!(1)n [p|n],
j=0

which is a multiple of p1+vp (n!) . Thus it suffices to show that for 1 j p 1,

   
n j n`
vp `! (1 ) > vp (n!)
`
(i.e.
for all 0 ` min(s, n 1). But using the extension of vp to the ring of algebraic numbers Q
1
vp (x) = d vp (f (0)), where f is the minimal polynomial of x), the LHS is just

n ` sp (n `) n `
vp (n!) + > vp (n!),
p1 p1

1
Victor Wang Interesting arguments, ideas

so were done (note that n ` 1 = sp (n `) > 0 in this case).[/hide]

Edit: Two more applications (also from PFTB)!
(Gabriel Dospinescu). Let p > 2 be a prime number and P let m and n be multiples of p, with n odd. For
m
any function f : {1, 2, . . . , m} {1, 2, . . . , n} satisfying k=1 f (k)  0 (mod p), consider the product
Qm m
n
k=1 f (k). Prove that the sum of these products is divisible by p . (Yes, we can strengthen it
easily.)
(St. Petersburg 2003). Let p be a prime and let n p and a1 , a2 , . . . , an be integers. Define
Pf0 = 1
and fk the number of subsets B {1, 2, . . . , n} having k elements and such that p divides iB ai .
Show that f0 f1 + f2 + (1)n fn is a multiple of p.
2. (MathOverflow, July 2012) By Newton interpolation (or other standard methods), one can easily
determine the number and structure (the latter to a lesser extent) of sequences (a1 , a2 , . . . , an ) (mod n)
represented by integer polynomials given the prime factorization of n (note that we can reduce to the
prime power case by the Chinese remainder theorem). For instance, see this AoPS thread. However,
for such f Z[x] we have strong restrictions like u v | f (u) f (v) for integers u, v.
IMO its then natural to wonder about (e.g. the new structure of valid f ) for rational polynomials in
general, where standard interpolation methods (in certain mods) arent as clean. (Mostly copied from
my comment below.)
To this end, consider a pair of positive integers (n, m) with n, m > 1 such that for every sequence
(a1 , a2 , . . . , an ) (Z/mZ)n , there exists an integer-valued polynomial f Q[x] satisifying f (x) ai
(mod m) whenever x i (mod n).
First, it is easy to show that n, m must be powers of the same prime. Indeed, if there exist distinct
primes p, q such that p | n and q | m, then for some sufficiently large integer `, we have q | f (x+n)f (x)
and q | f (x + q ` ) f (x) for every x. But gcd(n, q ` ) | n/p, so by Bzouts identity, q | f (x + n/p) f (x)
for all x and thus the ai(n/p) must all be congruent (mod q) in order for f to exist.
On the other hand, if n = pi and m = pj for some prime P p and positive integers i, j, then given
n k
a sequence (a1 , a2 , . . . , an ), consider the polynomial g(x) = k=1 ak x . By a simple induction on
d 0, we can show (using finite differences) that a degree d polynomial f (for convenience, say
deg 0 = 1) satisfying the desired properties exists iff d is the smallest number such that the division
of (x 1)d+1 g(x) by xn 1 gives a remainder r(x) with coefficients all divisible by m (call d + 1 the
*order* of the sequence (a1 , a2 , . . . , an ), so the all-zero sequence has order 0). Since (x 1)n = xn 1
in Fp , we have (x 1)jn = (xn 1)u(x) + pj v(x) for some polynomials u, v with integer coefficients, so
the order of any sequence is finite and at most jn.
However, this leads to the following two questions:
1. For fixed p, i, j, what is the maximum possible order M of a sequence (a1 , a2 , . . . , an ) (Z/mZ)n ?
(Resolved in the update.) 2. How many sequences are there of order r, where r [0, M ] is a fixed
integer? 3. Is there a reasonably nice way to describe the sequences of a fixed order r (possibly in
terms of one of the corresponding polynomials f, g)?
Looking at small cases, it seems that the answer to 1 should be pi + (j 1)(pi ), where denotes
Eulers totient function.
**Update:** OK, I think I have a (messy) proof that the answer to question 1 is indeed pi +(j 1)(pi ),
but it doesnt seem to lend itself to 2 or 3 in any way.
k k i
Fix p, i. First note that since (x 1)p = xp 1 (in Fp ) for k = i and k = i 1, we have (x 1)(p ) =
i
pi (x) in Fp as well, so (x 1)(p ) = pi (x) + pT (x) for some integer polynomial T of degree at most
(pi ) 1, where t denotes the tth cyclotomic polynomial. Observe that T (1) = 1, so 1 is not a root
of T in Fp .
Using this key fact, we will induct on j 1 to construct a sequence of integer polynomials Pj , Qj such
that i i i
(x 1)p +(j1)(p ) = (xp 1)Pj (x) + pj (x 1)Qj (x),

2
Victor Wang Interesting arguments, ideas

i1
1
vp (Qj (1)) = i 1, and in Fp , (x 1)p kQj (x) (i.e. 1 is a root of multiplicity pi1 1).
i i
(x1)p (xp 1)
For j = 1, we simply take P1 (x) = 1 and Q1 (x) = p(x1) , where clearly Q1 (1) = pi1 =
i1
vp (Q1 (1)) = i 1. Showing (x 1)p 1 kQj (x) is slightly harder, but not too bad. Its easy to show
i

by counting prime factors that pk is divisible by p for all 1 k pi 1 and not divisible by p2 iff
i

pi1 | k. Furthermore, by Babbages theorem we have kppi1 kp (mod p2 ) for 1 k p 1. Hence


i1 i1
for p = 2, we just need to show that (x 1)p kxp 1 in F2 , which is obvious; for p > 2 odd, we
need to show
p1 kpi1
! i1
p1 k p
pi1
X x X x
(x 1) k =
k k
k=1 k=1
Pp1 k
in Fp (note that k p = k by Fermats little theorem). But if h(x) = k=1 xk , then h(1) 0 (mod p)
while h0 (1) p 1 (mod p), so 1 is a simple root of h and were done with the base case.
i1
xp 1
Now assuming the result for some j 1 (so that x1 | Qj (x) in Fp ), we can write Qj (x) =
i1
xp 1
x1 R(x) + pS(x) for two integer polynomials R, S with deg S < pi1 1. (*) Then
i
+j(pi ) i
+(j1)(pi ) i
(x 1)p = (x 1)p (x 1)(p )

can be written as
i i
(xp 1)Pj (x)(x 1)(p ) + pj+1 (x 1)T (x)Qj (x) + pj pi (x)(x 1)Qj (x)

or equivalently after substitution,

i i
(xp 1)(Pj (x)(x 1)(p ) + pj R(x)) + pj+1 (S(x)pi (x) + T (x)Qj (x)),

so we can take i
Pj+1 (x) = Pj (x)(x 1)(p ) + pj R(x)
and
Qj+1 (x) = S(x)pi (x) + T (x)Qj (x).
As i1 i
(x 1)p | (x 1)(p ) = pi (x)
i1
1
in Fp and T (1) = 1, we see that (x 1)p kQj+1 (x).
It remains to show that vp (Qj+1 (1)) = i 1. By (*) and the definition of Qj+1 , we find Qj+1 (1) =
pi (1)S(1)+T (1)Qj (1) = pS(1)Qj (1) = pi1 R(1), so vp (Qj+1 (1)) i1. However, if pi | Qj+1 (1),
i1 i1
then p | R(1), so writing (*) in Fp we have Qj (x) = (x 1)p 1 R(x). But then (x 1)p | Qj (x),
contradicting our inductive hypothesis.
Thus our induction is complete.
Clearly this construction shows that the order of any sequence is at most M = pi + (j 1)(pi ). On
the other hand, it is easy to show that the order of (1, 0, . . . , 0) is M . Indeed, note that g(x) = x for
i
this sequence, and suppose x(x 1)M 1 /(xp 1) leaves a remainder with coefficients all divisible by
j
p . From the induction statement, we have
i
M 1 xp 1
(x 1) = Pj (x) + pj Qj (x),
x1

so writing Pj (x) = (x 1)U (x) + V for an integer V , we get pj | V . But then plugging in 1 to this
equation, 0 = (0)U (1) + (pi )V + pj Qj (1), whence pi | Qj (1), contradiction.

3
Victor Wang Interesting arguments, ideas

3. (Classical umbral calculus) Using linear operators, one can deal with recurrences more easily. This
is a powerful idea, for instance, when computing recurrences
Pn mod prime powers. For instance, we
can
Pn apply this to the Touchard polynomials Tn (x) = k=1 S(n, k)xk , using the identity Tn+1 (x) =
n

n
x k=0 k Tk (x) and defining the linear operator L(x ) = Tn (x). We can, of course, set particular
values of x, e.g. 1 for the Bell and Uppuluri-Carpenter numbers, respectively. Then using the
Frobenius endomorphism we can get some nice recurrences.
This also applies to linear recurrences, e.g. for any starting values, an = an1 + anp has period
dividing p2 1. However, this can also be done with generating functions, noting that the sequence
must be purely periodic.
4. Solve

xn 1 = (3 + 3x + + 3x9 + 2x10 + 2x11 + + 2x1209 + x1210 + x1211 + + x146409 )f (x) + 11 g(x).