Sie sind auf Seite 1von 10

SUMSETS MOD p

YSTEIN J. RDSETH

Abstract. In this paper we present some basic addition theorems modulo a prime p
and look at various proof techniques. We open with the Cauchy-Davenport theorem and
a proof using the Davenport transform. We continue with a result, due to Davenport,
on strongly connected digraphs. We include a brief application of Philip Halls theorem
on distinct representatives to sumsets mod p. Then we give a simple proof of Vospers
theorem using the Davenport transform. Finally, we look at sums of distinct residue
classes mod p. Complete results for this type of restricted sumsets were first obtained
using representation theory and multilinear algebra. These results do not seem to lend
themselves to transformation proofs. However, now we have the beautiful technique
called the polynomial method, which gives simple proofs of such results. We demon-
strate the polynomial method by using it in a second proof of the Cauchy-Davenport
theorem.

1. Introduction
Let A be a set of k 1 distinct integers a1 , a2 , . . . , ak , and let B be a set of 1
distinct integers b1 , b2 , . . . , b . The sumset A + B is defined as
A + B = {a + b | a A, b B}. (1)
Thus the sumset A + B consists of the distinct integers appearing in the rectangular array
a1 + b1 a1 + b2 . . . a1 + b
a2 + b1 a2 + b2 . . . a2 + b
.. .. .. (2)
. . .
ak + b1 ak + b2 . . . ak + b
What can be said about the number |A + B| of elements in A + B? That is, what
can be said about the number of distinct integers in the array (2)? Trivially, we have
|A + B| k, and it might very well happen that all the integers in the array (2) actually
are distinct. What about a lower bound for |A + B|?
We arrange the notation such that a1 < a2 < < ak and b1 < b2 < < b . In the
array (2) we start at the upper left hand corner, move along the first row, and continue
down the last column. Then we have a strictly increasing sequence of integers in A + B,
a1 + b1 , a1 + b2 , . . . , a1 + b , a2 + b , . . . , ak + b . (3)
2000 Mathematics Subject Classification. 11A07.
Key words and phrases. Sumsets, congruences, residue classes mod p.
1
2 . J. RDSETH

This sequence consists of k + 1 distinct integers. Hence,

|A + B| k + 1. (4)

By choosing A and B to be arithmetic progressions with a common difference, we see that


this result is best possible.
Let p be a prime. What happens to the considerations above if we replace the integers
by residue classes modulo p? Thus A = {a1 , a2 , . . . , ak } and B = {b1 , b2 , . . . , b } are now
subsets of Z/pZ, the finite field of p elements. We continue to write k for the number of
elements in A, and the number of elements in B is . The sumset A + B is defined by
(1). Consequently, A + B consists of the distinct residue classes among those listed in the
array (2). We continue to write |A + B| for the number of elements in A + B. Still we
have, of course, that |A + B| k, or rather |A + B| min(p, k). The question is if (4)
still holds. The answer is not unconditionally yes. We always have |A + B| p, so if
p < k + 1, then the answer is no.
But let us assume that k + 1 p. Is (4) then correct? The proof above for ordinary
integers, where we looked at the sequence (3), fails for residue classes. It is true that the
residue classes in the first row in (2) are all distinct, and the same goes for the residue
classes in the last column. But more than one of the elements in the first row may very
well also appear in the last column. Nevertheless, the inequality (4) does hold.

Theorem 1 (Cauchy-Davenport). |A + B| min(p, k + 1).

This is a basic result in the part of additive/combinatorial number theory known as


addition theorems mod p (and more generally as addition theorems in groups). The
theorem was proved by Cauchy [4] in 1813 and rediscovered by Davenport [6, 7] in 1935.
As we have already indicated, the proof is more complicated than that for integers.
Both Cauchy and Davenport employed what are known as transformation proofs. In
the literature it is mainly the so-called (Dyson) e-transform that is used. This is essentially
the same transform as the one employed by Cauchy. The Davenport transform is more or
less forgotten. In this paper we use the Davenport transform to prove Theorem 1. This
does not offer special advantages over the use of the e-transform in the proof of Theorem
1. But in Section 7 we shall see that the Davenport transform gives a rather nice proof
of Vospers theorem.
This paper is organized in 8 sections. In Section 2 we introduce some notation. In
Section 3 we describe the Davenport transform. Then, in a brief Section 4 we prove
Theorem 1. In Section 5 we show a result about directed graphs, a result which, according
to Davenport, is equivalent to Theorem 1. In Section 6 we employ Philip Halls theorem
on distinct representatives to say a bit more about where in the array (2) we can find
k + 1 distinct residue classes when k + 1 p. In Section 7 we consider an inverse
theorem mod p, and in Section 8 we close by mentioning a type of result which does not
seem to lend itself to any transformation proof.
SUMSETS MOD p 3

2. Notation
As before, let A and B be non-empty subsets of Z/pZ. The sumset A + B is defined
by (1). This set operation is commutative and associative. Also the sum of more than
two sets is uniquely defined. In particular, we write hA for A + + A (h addends). For
c Z/pZ we write c + A for {c} + A. We set B = {b | b B}, and A B = A + (B).
The number of elements in A is denoted by |A|. We write A B when A is a subset of B.
We reserve the symbol for proper inclusion. Thus A B means that A is contained in
B, but is not equal to B. We denote the relative complement of B in A by A \ B, while
B denotes the complement of B in Z/pZ. If there are residue classes a, d Z/pZ such
that A consists of the distinct residue classes a, a + d, a + 2d, . . . , a + (k 1)d, then A is
an arithmetic progression with difference d. For an integer r we will on some occasions
also write r for the residue class modulo p represented by r.

3. The Davenport Transform


We now introduce the Davenport transform (in modern notation), and give its most
important properties.
Suppose that 0 B, |B| 2, and A + B 6= Z/pZ. Then A + B A + 2B. If
A + B = A + 2B, then A + B = A + 2B = A + 3B = . . . For an a A and 0 6= b B,
we then have a + nb A + B for n = 0, 1, 2, . . . Thus we have A + B = Z/pZ, contrary
to assumptions. Therefore A + B A + 2B. Putting

X = (A + 2B) \ (A + B), (5)


we thus have X 6= . For x X we set

Bx = {b B | x b A + B}, Bx = B \ Bx .
We shall call Bx for a Davenport transform of B.
Now, we have 0 6 Bx 6= , and 0 Bx B. Clearly,

(A + Bx ) (x Bx ) A + B. (6)
We also have
(A + Bx ) (x Bx ) = , (7)
since we (in self-explanatory notation) have that if a + bx = x bx , then x bx = a + bx
A + B, so that bx Bx , a contradiction. By (6) and (7) we thus have

|A + B| |A + Bx | + |x Bx |,
that is,
|A + B| |A + Bx | + |B| |Bx |. (8)
4 . J. RDSETH

4. Proof of Theorem 1
Assume the theorem false. Then there are pairs of sets A, B such that A + B 6= Z/pZ,
and
|A + B| |A| + |B| 2. (9)
These properties remain valid if we replace B by b+B for some b B. We may therefore
assume that 0 B. Choose such a pair A, B for which |B| is minimal. Then |B| 2.
By Section 3, there exists a Davenport transform Bx such that (8) holds. We have
1 |Bx | < |B| and |A + Bx | < |A + B| < p. By (8) and (9),
|A + Bx | |A + B| |B| + |Bx | |A| + |Bx | 2,
which contradicts the minimality of |B|.
5. Digraphs
The Cauchy-Davenport theorem is given as statement A in Davenports 1935 paper.
What seems to be less known, is that Davenports paper also contains a statement B,
which he describes as equivalent to statement A. By this, he obviously means that
statement B follows easily from statement A, and vice versa. Davenport did not use
modern graph-theoretic language. Today we can formulate Davenports statement B as
follows:
Let B, V Z/pZ, where 0 6 B. Set
E = {(x, y) V V | y x B}.
Then D = (V, E) is a digraph with vertex set V and edge set E.
For a moment, think of V as fixed, while B is allowed to vary. If |B| is small, then
there are few edges, and the digraph D falls apart into several pieces (the graph D is
disconnected). On the other side, if |B| is large, then we probably have many edges, and
the chances should be good of D being in one piece (the graph D is connected). If |B| is
large, the chances should even be good of D being strongly connected, that is, there is a
directed path from any vertex v V to any other vertex w V . Therefore the following
theorem, which is Davenports statement B, appears quite natural.
Theorem 2 (Davenport). If |B| > p |V |, then D is strongly connected.
Let us deduce this theorem from the Cauchy-Davenport theorem. Let v V , and
A = {a V | v = a or there is a path from v to a}.
By putting B0 = {0} B, we have
(A + B0 ) V A.
To see this, let a + b0 = w (A + B0 ) V . If b0 = 0, then w = a A. If b0 B, we have
w a = b0 B, such that there is an edge from a A to w. Hence w A.
So, we have A (A + B0 ) V A, that is,
(A + B0 ) V = A.
SUMSETS MOD p 5

By inclusion-exclusion, we get
p |(A + B0 ) V | = |A + B0 | + |V | |(A + B0 ) V |,
that is,
p |V | |A + B0 | |A|.
We now assume that |B| > p |V |. Then |B| > |A + B0 | |A|, or since |B0 | = |B| + 1,
|A + B0 | < |A| + |B0 | 1.
The Cauchy-Davenport Theorem now gives |A + B0 | = p, that is A + B0 = Z/pZ. Hence,
A = (A + B0 ) V = (Z/pZ) V = V,
which means that there is a path from an arbitrary vertex v to any other vertex. This
completes the proof of Theorem 2.
6. Distinct Representatives
Davenports paper [6] begins on page 30 in the 1935 volume of Journal of the London
Mathematical Society. On this very page ends a paper by Philip Hall [11]. Halls paper
contains his famous theorem about distinct representatives.
Theorem 3 (P. Hall). Let C1 , C2 , . . . , Ck be sets. Suppose that the union of every selection
of s of these sets contains at least s elements for all s = 1, 2, . . . , k. Then we can find a
set of k distinct elements, one from each of the k sets.
There are many proofs of this result in the literature. Our two favourites are the ones
presented in [1, p. 28] (due to D. J. Shoesmith) and [19, p. 116] (due to R. Rado).
As before, let A = {a1 , . . . , ak }, B = {b1 , . . . , b } Z/pZ, |A| = k 1, |B| = 1.
Suppose that A + B 6= Z/pZ. By the Cauchy-Davenport theorem we have
|A + B| k + 1.
All ai + bj are given by (2). Is it possible to say something about where in the rectangular
array (2) we can find k + 1 distinct residue classes?
Let C be an ( 1)-subset of A + B. (For example, C = a1 + {b1 , . . . , bl1 }.) Then we
may choose one element from each row in (2), such that C and these elements constitute
k + 1 distinct residue classes in (2). (We may, for instance, choose all elements in the
first row in (2) and one element from each of the other rows.)
This is seen in the following way. The set of elements in row i in the rectangular array
(2) is equal to ai + B. Let
Ci = (ai + B) \ C.
For 1 i1 < i2 < < is k we have, by the Cauchy-Davenport theorem,
|Ci1 . . . Cis | = |({ai1 , . . . , ais } + B) \ C|
|{ai1 , . . . , ais } + B| |C|
s + l 1 (l 1) = s,
6 . J. RDSETH

and by Halls theorem we can find k distinct elements, one from each Ci . Along with the
1 elements of C, we thus have k + 1 distinct elements of A + B.
7. An Inverse Theorem mod p
The Cauchy-Davenport Theorem is a direct addition theorem mod p. Given sets A
and B, such a theorem usually gives a result about A + B; in our case a lower bound for
|A + B|. The corresponding inverse problem is to describe the structure of the sets A, B
for which |A + B| is small. The first non-trivial inverse theorem mod p is due to Vosper
[17].
Theorem 4 (Vosper). We have
|A + B| = min(p, |A| + |B| 1)
if and only if one of the following conditions is satisfied:
(i) |A| + |B| > p,
(ii) |A| = 1 or |B| = 1,
(iii) B = c A for a c Z/pZ,
(iv) A and B are arithmetic progressions with the same difference.
The principal step in the proof of Vospers theorem is the proof of the following result.
Theorem 5 (Vosper). Suppose that |B| 2, and that
|A + B| = |A| + |B| 1 < p 1. (10)
Then A is an arithmetic progression.
Vosper first used the Davenport transform to prove this result. Later he gave in [18] a
simpler proof by employing the e-transform. Another transform was employed by Chowla,
Mann, and Straus in [5], where they also gave a beautiful application of Vospers theorem
to diagonal forms over Z/pZ; cf. [15, p. 57], [14, Ch. 2]. Now we shall show how the
Davenport transform gives us an even simpler proof of Theorem 5 than Vospers second.
Proof of Theorem 5. Assume that the theorem is false. Let A, B be a pair of sets such
that A does not form an arithmetic progression, (10) holds, and |B| 2 is minimal. We
may also assume that 0 B.
Let X be given by (5). For an x X, let Bx be the corresponding Davenport transform
of B. By (8) and (10), we then have
|A + Bx | |A + B| |B| + |Bx |
|A| + |Bx | 1 < p 2.
By the minimality of |B| we have Bx = {0}. Setting B = B \ {0}, we thus have
Bx = B for all x X,
so that x B A + B for all x X, and we see that
A (X B ) A + B and A (X B ) = .
SUMSETS MOD p 7

It follows by (10) that


|A| + |B| 1 = |A + B| |A| + |X B |,
and application of Theorem 1 gives
|B| 1 |X B | |X| + | B | 1 = |X| + |B| 2.
Thus we have |X| 1, that is, |X| = 1. Since |A + B| < p 1, it follows that
|A + 2B| = |A + B| + 1 < p.
Applying Theorem 1 once more, we obtain
1 = |A + 2B| |A + B| |A + B| + |B| 1 |A + B| = |B| 1.
Hence |B| = 2. If 0 6= b B, the conditions upon A and B also hold for the sets
{ab1 | a A} and {0, 1}. Therefore it is no restriction to assume that B = {0, 1}. Thus
we have by (10),
|A + {0, 1}| = |A| + 1. (11)
Consider the residue classes 0, 1, . . . , p 1 as consecutive points on a circle. By (11), we
have exactly one element a A with a + 1 6 A. Hence, the elements of A form a set of
consecutive points on the circle, that is, A is an arithmetic progression with difference 1.
This completes the proof of Theorem 5. 
If we in Theorem 5 also assume that |A| 2, it is now easy to see that B is an arithmetic
progression with the same difference as A. For we may assume that A = {0, 1, . . . , |A|1}.
Then A = {0, 1} + A for A = {0, 1, . . . , |A| 2}. By (10) we have |A + B| < p, and using
Theorem 1 we get
|A| + |B| 1 = |A + B|
= |{0, 1} + A + B|
|{0, 1} + B| + |A | 1
= |{0, 1} + B| + |A| 2;
that is,
|{0, 1} + B| |B| + 1. (12)
If we once more regard the residue classes mod p as points on a circle, we see that (12) holds
if and only if B is an arithmetic progression with difference 1. The proof of Theorem 4 is
now easily completed.

In [12] the authors go one step beyond Vosper, and in [13] another step is taken. But we
are far away from existing conjectures generalizing Vospers theorem. On the other hand,
related to Vospers theorem, there is a strong inverse theorem by Freiman [10, Theorem
2.1], [15, Theorem 2.11]. This theorem of Freiman is the subject of [16].
8 . J. RDSETH

8. The Polynomial Method


We define another type of addition of sets of residue classes mod p by putting
b = {a + b | a A, b B, a 6= b}.
A+B (13)
As before, we let |A| = k and |B| = .
By the Cauchy-Davenport theorem we have
|A + A| min(p, 2k 1).
More than 40 years ago, Erdos and Heilbronn conjectured that we have the similar result
b min(p, 2k 3).
|A+A| (14)
If (14) is true, it is best possible. This is seen by taking A as an arithmetic progression.
If we want to prove (14) by transformation, we actually have to prove a more general
result for the set (13), since one or both sets are altered by transformation. But this plan
does not work, since the condition a 6= b in (13) is destroyed by all (known) transforma-
tions. So a non-transformation proof is what we need. Then it is natural first to try to
find such a proof of the Cauchy-Davenport theorem.
And the first non-transformation proof of the Cauchy-Davenport theorem was given in
1990 by Dias da Silva and Hamidoune [8] using multilinear algebra. In 1994, more than
30 years after Erdos and Heilbronn stated their conjecture, Dias da Silva and Hamidoune
[9] proved the conjecture using results from representation theory and multilinear algebra.
Soon after, Alon, Nathanson, and Ruzsa [2, 3] found a beautiful elementary proof of
(14). They even proved that
b min(p, k + l 2} if k 6= l.
|A+B| (15)
Alon, Nathanson, and Ruzsas new method of proof is called the polynomial method.
To demonstrate this method, we now use this technique to give a second proof of the
Cauchy-Davenport theorem.
We consider polynomials over the field Z/pZ. We put
Y
f (x, y) = (x + y c),
cA+B

and define g(x, y) by


Y Y
g(x, y) f (x, y) (mod (x a), (y b)),
aA bB

where g(x, y) is of degree at most k 1 in x, and of degree at most 1 in y. We have


g(a, b) = f (a, b) = 0 for all (a, b) A B.
A well-known theorem of Lagrange states that if a polynomial of degree at most n 1 in
one variable over Z/pZ has n zeros, then the polynomial is identically zero. It is now a
simple exercise to show that g(x, y) is the zero polynomial.
SUMSETS MOD p 9

We have
|A+B|  
X |A + B| j |A+B|j
f (x, y) = xy + terms of lower degree.
j=0
j

Suppose that |A + B| =6 p. Then the monomial xk1 y |A+B|k+1 has non-zero


Q coefficient,
and the exponent of y must be reducible when the monomial is taken mod bB (y b).
Thus |A + B| k + 1 , and the proof is complete.
To prove (15), consider the polynomial
Y
h(x, y) = (x y) (x + y c)
b
cA+B

instead of f (x, y).

While it is straightforward to generalize the Cauchy-Davenport theorem to the sum of


more than two sets, this is not the case with (14) or (15). Dias da Silva and Hamidoune
nevertheless showed how (14) generalizes to h equal addends, while Alon, Nathanson,
and Ruzsa showed how (15) generalizes to h not necessarily equal sets.
The polynomial method has its obvious advantages to transformation methods. But
the opposite is also true. It is an unsolved problem to produce inverse results by the
polynomial method. Nobody has succeeded in determining the structure of those A for
which equality holds in (14), nor the structure of those A, B for which equality holds in
(15). Nor do we know if the polynomial method can be used to prove Vospers theorem.
Nevertheless, problems of the type touched upon in this paper constitute a very active
field of research.

References
[1] I. Anderson, A First Course in Combinatorial Mathematics, Second ed., Oxford University Press,
New York, 1989.
[2] N. Alon, M. B. Nathanson, and I. Z. Ruzsa, Adding distinct congruence classes modulo a prime,
American Math. Monthly 102 (1995), 250255.
[3] N. Alon, M. B. Nathanson, and I. Z. Ruzsa, The polynomial method and restricted sums of residue
classes, J. Number Theory 56 (1996), 404417.
[4] A. L. Cauchy, Recherches sur les nombres, J. Ecole Polytech. 9 (1813), 99123; also in Oevres,
Serie 2, Tome 1, 3963.
[5] S. Chowla, H. B. Mann, and E. G. Straus, Some applications of the Cauchy-Davenport theorem,
Det Kongelige Norske Videnskabers Selskabs Skrifter 32 (1959), 7480.
[6] H. Davenport, On the addition of residue classes, J. London Math. Soc. 10 (1935), 3032.
[7] H. Davenport, A historical note, J. London Math. Soc. 22 (1947), 100-101.
[8] J. A. Dias da Silva and Y. O. Hamidoune, A note on the minimal polynomial of the Kronecker sum
of two linear operators, Linear Algebra and its Applications 141 (1990), 283287.
[9] J. A. Dias da Silva and Y. O. Hamidoune, Cyclic spaces for Grassmann derivatives and additive
theory, Bull. London Math. Soc. 26 (1994), 140146.
[10] G. A. Freiman, Foundations of a Structural Theory of Set Addition, Translations of Mathematical
Monographs, Vol. 37, American Math. Soc., Providence, R. I. (1973).
10 . J. RDSETH

[11] P. Hall, On representatives of subsets, J. London Math. Soc. 10 (1935), 2630.


[12] Y. O. Hamidoune and . J. Rdseth, An inverse theorem mod p, Acta Arith., 92 (3) (2000), 251262.
[13] Y. O. Hamidoune, O. Serra, and G. Zemor, On the critical pair theory in Z/pZ, Acta Arith. 121 (2)
(2006), 99115.
[14] H. B. Mann, Addition Theorems: The Addition Theorems of Group Theory and Number Theory,
Interscience Publishers, New York 1965.
[15] M. B. Nathanson, Additive Number Theory: Inverse Problems and the Geometry of Sumsets,
Springer-Verlag, New York 1996.
[16] . J. Rdseth, On Freimans 2.4-theorem, Trans. R. Norw. Soc. Sci. Lett. (to appear).
[17] A. G. Vosper, The critical pairs of subsets of a group of prime order, J. London Math. Soc. 31
(1956), 200205.
[18] A. G. Vosper, Addendum to The critical pairs of subsets of a group of prime order, J. London
Math. Soc. 31, (1956), 280282.
[19] R. J. Wilson, Introduction to Graph Theory, Forth ed., Longman, Essex 1996.

Department of Mathematics, University of Bergen, Johs. Brunsgt. 12, N-5008 Bergen,


Norway
E-mail address: Oystein.Rodseth@uib.no
URL: http://www.uib.no/People/nmaoy

Das könnte Ihnen auch gefallen