A LINEAR DIOPHANTINE PROBLEM

S. M. JOHNSON

1. I n t r o d u c t i o n . L e t #i, a 2 , . . . , at be a set of groupwise relatively prime

positive integers. Several a u t h o r s , (2; 3 ; 5; 6), h a v e determined b o u n d s for
t h e function F(di, . . . , at) defined b y t h e p r o p e r t y t h a t t h e equation
(1) n = aiXi + a 2 x 2 + . . . + atxt
has a solution in positive integers Xi, . . . , xt for n > F(a,\, . . . , at). If
F(ai, . . . , at) is a function of this t y p e , it is easy to see t h a t
(2) G(au • • • ,at) = F(au . . . , at) - ax - a2 - ... - at
is t h e corresponding function for the solvability of (1) in non-negative x's.
I t is well known t h a t a x a 2 is t h e best bound for F(a,i, a 2 ) and aia2 — a,\ — a 2
for G(di, a 2 ). Otherwise only in very special cases h a v e the best b o u n d s been
found, even for t = 3.
In the present paper a s y m m e t r i c expression is developed for the best b o u n d
for F(di, a 2 , a3) which solves t h a t problem and gives insight on the general
problem for larger values of t. In addition, some relations are developed which
m a y be of interest in themselves.

2. A G e n e r a l P r o p e r t y . For t > 2, let B(a\, a 2 , . . . , at) be t h e best b o u n d

for F(di, a 2 , . . . , dt), t h a t is, B is t h e m a x i m u m n u m b e r N where
t
%
(3) N 9^ 2 2 i°"i for a n y Xj > 0.

T h e n note t h a t B is the m a x i m u m N from a restricted set of n u m b e r s N

satisfying both (3) and
t

(4) N+dt = 2 2 yijdj, yij > 0 for each i.

since t h e definition of B implies B satisfies (4). T h u s , in particular,

N = (yn - ! ) # ! + 3,i2«2 + . . . + yi&u yij > 0.
B u t b y (3), 3/11 — 1 < 0 so t h a t yu = 1 since yu > 0. B y s y m m e t r y we h a v e
T H E O R E M 1. For every N sdtisfying (3) dnd (4) there dre representdtions of N
for each i = 1, 2, . . . , t of the form
t

(5) N = Yl yaah ytj > 0'

.7=1

dnd B is the mdximum such N.

Received October 21, 1957; in revised form March 9, 1959.
390

https://doi.org/10.4153/CJM-1960-033-6 Published online by Cambridge University Press

 A LINEAR DIOPHANTINE PROBLEM 391

3. The Case t = 3. A reduction formula. We seek an expression for

B = B(ah a2l a3) having the property that (1) is satisfied for n > B but is not
satisfied for n = B. Let us first reduce the problem to the case of pairwise
relatively prime as.
Let dij = (auaj)y at = b4ijdik, so that (h, b2) = (b2, 63) = (#3, &i) = 1.
Then we have
THEOREM 2.
(6) B(aly a2, a3) = di2d2SduB(bu b2, £3).
Proof. First we show that if we write d = dx2, b\ = dubh b2 = d2^b2 so that
(d, a3) = (6i,6 2 ) = 1, then
(7) B(dbi, db2, a3) = dB(bu b2, a 3 ).
Suppose that dB(bi, b2, a3) = dbix + db2y + a3s, x, y, z > 0. Then since (d,
a3) = 1, we must have z = wd, w > 0, so that B(bu b2, a3) = b±x + b2y
dzW, x, y, w > 0, a contradiction to the definition of B(bi, b2, a*). In addition,
for any positive integer m > 0, we show that
(8) dB(bi, b2, a3) + m = dbix + ^6 2 ^ + a3s, x, 3/, s > 0.
We apply a result from (2).
LEMMA 1 (Brauer). Let a and b be relatively prime positive integers. Then every
positive integer m divisible neither by a nor by b is representable either in the form
(9) m = au + bv, u > 0, v > 0,
or
(10) m = ab — au — bv, b>u>0,a>v>0.
Letting d = a and a 3 = b in Lemma 1, if (9) holds, we have
(11) dB(bu b2, a3) + m = d(B(bh 52, a3) + w) + z>a3
= J61X + ^6 2 ^ + a3(efe + z/)
by the definition of -B(5i, 52, #3), giving (8).
If (10) holds, we have 0 < u < a3, and 0 < v < d, so that
(12) d(B(bi, b2, a3) + a 3 — w) — z>a3 = dbix + db2y + (dz — v)az,
for x, y, and (dz — v) > 0, giving (8).
Finally, if m = ud, then (8) follows directly. If m = va%, write m = da%
+ (v — d)a% giving (8). Thus (7) holds. Applying the method of obtaining
(7) twice more gives (6) and Theorem 2.
We have thus reduced the problem to where the a's are pairwise relatively
prime. For the moment let a\ > a2 > a3. If
(13) a\ = ua2 + va%, u,v>0,
then B(ai, a2, a 3) = a2a% + a\ as Brauer showed in (2). Otherwise
(14) B(au a2, a3) < ataj + ak.

https://doi.org/10.4153/CJM-1960-033-6 Published online by Cambridge University Press

 392 S. M. JOHNSON

4. An expression for B(ai, a2, a 3 ). We develop a symmetric expression for

B(ai, #2, 03) for the case of pairwise relatively prime a's where each a* ^ xdj
+ yak, x > 0, y > 0. Later we show that this same form of expression gives
the general solution for / = 3.
DEFINITION. Let Lt = the minimum positive Kt satisfying
(15) Kifit = Vijdj + vikak, vtJ > 0, vik > 0, i = 1, 2, 3.
Such a number exists since B{ah ak) = djdk < Kat for large if.
THEOREM 3. Given
(16) (ai, a2) = («2, a8) = (as, ai) = 1
awd
(17) i*>l, i = 1, 2, 3
and
(15 ) JUjCLi =• Xijdj -f- X{kdk}

then the xtj are uniquely defined and

(18) x^ > 0.
Since Lt > 1, it follows from (10) and (16) that
(19) at = a/z* — ^,-a,- — »waA
where 0 < vi3 < ak, 0 < p a < a^. Thus vikak + a* = (a* — fl^)a^ > Z,/^ and
so by symmetry
(20) Lj < akj for each j 9e k.
If Xji = 0, then Ltdi = x^a^ and by (16) Lj = mdkl a contradiction to
(20). This gives (18). Also the xtj are uniquely determined since if Ltdt = x^a^
+xÎÀ;afc = Zijdj+Zijcdjc, then by (16) we have xtj = Zij+mak and x t t = zik — mdj.
If m > 0, x^- > a*. But then for some d > 0, I^a* = (a* + d)dj + x # f c and
by (19) we get (Lt — \)at = (d + vij)dj + (xik + vik)dk, contradicting the
definition of Lt. Similarly, for m < 0.
For / = 3 and (16) and (17) we show that there are just two numbers N
with properties (3) and (4) so that B is the larger of these numbers. From
(5) such a number N has representations of the form
(50 N = yijdj + yikdk i = 1, 2, 3.
Next observe that from (18) we have
(21) ytJ < Lj
since otherwise for some dj > 0 we would have N = (Lj + dj)dj + ytkak
= Xjidi + djdj + (xjk + yik)dk, contradicting (3). From (20) and (21) we
have
(22) ykj < ak, ykj < at.

https://doi.org/10.4153/CJM-1960-033-6 Published online by Cambridge University Press

 A LINEAR DIOPHANTINE PROBLEM 393

Next we show that the representations (5') for N are unique for each i.
For otherwise ykiaf + ykiaj = zkiai + zkja,j and from (16) and (22), ykj — zkj
= mau m < 0, and yki — zki = ma, m > 0, so that m = 0 and 3 ^ = s^, etc.
From (5') and Theorem 1 we now have unique representations of N of the
form
N = Jkidi + ykjdj = ytjds + yikak = yjkak + y^a*.
If 3 ^ = ytjj then y ^ = mak, contradicting (22). Thus either ykj < ytJ or
yki > ytj-

Case 1. If
(23) ykj < yi3
then ykiat = {yi3 — ^ ; ) a y + yikak so that y t< > L*. Thus by (21) we have
(24) yki = Lt.
Then by (24) and (5')
N = Ltai + ykjaj = y^at + yjkak
or (L t - yji)ai + ykja*= 3 ^ * , where L* > ?,< by (21). If Lt = y,, then
3>&;- = ma*, contradicting (22), so that Li — yjt > 0 and yjk > Z,# by the
definition of Lk. But then yjk = Lk by (21). Thus (23) implies that yki = Lu
yjk = Lk, and cyclically, ytj =JLj. But then by (15')
N = (xij + ykj)aj + xikak = Ljaj + yikak
and by the uniqueness of these representations and by cyclic permutation of
subscripts, we have
(25) yik = xik
and
(26) Lj = Xij -\- xkj.
Thus if ykJ < yih we get a unique number N where
(27) N = Li/at + xkjaj
with cyclic permutations of subscripts.
Case 2. If
(28) ykj > ytj,
we get another number where by symmetry
(29) N' = Lifii + xjkak
with cyclic permutations of subscripts. N j£ Nf since otherwise % jkak xkjaj
which implies xjk > a}j which by (25) contradicts (22). Note that these two

https://doi.org/10.4153/CJM-1960-033-6 Published online by Cambridge University Press

 394 S. M. JOHNSOH

numbers are the only numbers with properties (3) and (4) for (16), (17),
and t = 3. Since B is the largest number with property (3), it satisfies (4) so
that B is the maximum of N and N' and we have
THEOREM 4. Given (16) and (17), then for cyclic permutations of subscripts
(30) B{ax, a2, a3) = L&t + max (xkjaj, xJkak)
and (26) holds.
Also it is easy to verify that C, the corresponding best bound for G(a\, a2, a 3 ),
satisfies
(31) C(ai, a2, az) + ax + a2 + a 3 = 5 ( a i , a2, a 3 ).

5. A computing algorithm for Lt and x^-. Thus we have shown that

finding B is equivalent to finding the set of positive integers Lt and Xtj ex-
hibited in the form of a matrix of detached coefficients of the three equations
(15r) a s follows:

fll a2 az

- i l *12 xn
*21 -L2 X23
Xzi x32 -is

In order to develop a simple computing algorithm for these numbers, we

need the following result.
LEMMA 2. Given (#i, a2) = (a2, a3) = (a3, ai) = 1, £/zew aw^ system of integers
Kt > 1 and A*, > 0 (w^ necessarily Lt and xi}) satisfying (15) awd (26)
i^z = Vji + ^z-, implies that
(32) X ^ - » j ^ i = VjtvkJ + vkiKj = \ak > ak
for some positive integer X.
If we write
VjkiKtai — Vijdj) = vikvJkak = vik(Kjaj - vjial),
then
(vjkKt + vikvjl)ai = (vikKj + vJkVij)aj
and (32) follows by (16) and (26).
Furthermore, we have
THEOREM?5. / / (16) and (17) hold, then the Li and Theorem 4 are
characterized by the equations (15') and (26), and
{So) Lihj -\- XijXji = ak ,

for cyclic permutations of subscripts. That is, X = 1 in (32).

https://doi.org/10.4153/CJM-1960-033-6 Published online by Cambridge University Press

 A LINEAR DIOPHANTINE PROBLEM 395

Proof. Suppose a system of Kt and vtj satisfy (15), (26), and (33) where
at least one Kt > Lu the minimum positive integer satisfying (15).
Case 1. If Kx = Llt K2 = L2, then i£ 3 = U by (26) and Theorem 3.
Case 2. Suppose K\ = Lu but K2 > L2, Kz > Z 3 .
Then x12 = v12 and xu = vu by Theorem 3 and by (15), (26), and (33)
ax = K2KZ — vZ2v2z = K2KZ — (K2 — x12)(Kz — xu) = x12K^ + xuK2 —
XuXn > XuLz + XnL2 — xi2Xu = L2L% — x32x23 > a± by (32), a contradiction
to the assumption that K2 > L2j K% > L3.
CaseS. If K± > Liy L2 > K2,K^ > L%, then first observe that either vtj > xtj
or vik > xik} but not both. For suppose vtj > xtj and vik > xik. By (33)
VijVjjc + KjVik = at < XijXjk + LjXtk by (32). Thus vjk < xjk. Similarly
VijKk + vikvkj = at < XijLk + x^x^- so that vk] < xkj. But then at < L^L*
— xjkxkj < KjKk — vjkvkj = au a contradiction.
In addition either vn > xjt or vki > xki but not both. For suppose v}i > xjt
and vki > Xjt. By the previous remark vjk < xjk, vkj < xkj, leading to the
same contradiction obtained above. Thus either vu, ^23, ^31, or v2\, v-i2, Vn are
larger than the corresponding x's. That is vti > xtj for cyclic permutations
of subscripts.
Suppose z/21, v%2, Vu are larger than x 2 i, X32, x i 3 respectively. Then by (26)
(K2 — L2)a2 + (x23 - v2z)az = (v2i — x2X)ax > Lxa^
by the definition of L\. Thus v2i > Li and by cyclic permutation of subscripts
^32 > L2, Z/13 > L 3 .
Finally a 3 < LXL2 — Xi2x2i < LXL2 < v21v^2 < v2iv^2 + K2v-S1 = 03, a c o n -
tradiction.
Thus X = 1 in (32) implies that Kt = Lu vi3 = xi3.
Conversely, X = 1 in (32), for Kx — Lu vtj = xtj etc. By the following
computing algorithm we can always find sets of Kt and vtj with X = 1 in
(32). Thus they are the desired Lt and x^-. Moreover since the Xtj are unique
by Theorem 3, X is unique and must equal 1.
The usefulness of Theorem 5 is apparent since it will be easier to find X's
and v's satisfying (15), (26), and (33) rather than find minimal solutions to
(15).
The algorithm follows. First we solve for any ak in terms of at and a/, for
instance, for k = 3, giving
(34) z>2iai — K2a2 + a 3 = 0
with 0 < v2i < a2, 0 < K2 < ai by (10), easily done for example as in (4).
Next construct
(35) — Kidi + Z>12#2 + ^13^3 = 0
where

Vn = jT-J , Ki = a2 - »2i»is, ai = K2vn + vu

https://doi.org/10.4153/CJM-1960-033-6 Published online by Cambridge University Press

 396 S. M. JOHNSON

so that X = 1 in (32). If Ki > v2ï, then K± = Lly K2 = L2j and L 3 can be

found by (26). Then apply Theorem 4 for B(ah a2, a 3 ). If Ki < v2i, note that
Ki^v2i. For if K^Vn, then since Kx = a2 — VnViz, K\\a<2> But then in (34)
Ki\a%. Thus (&2, a3) > i£i > 1, by (17) a contradiction.
Therefore if K\ < v2\ we can construct another equation
(36) (v2l - ^ i ) a i - (i£2 - pvl2)a2 + (1 + ^ 1 3 ) a 3 - 0
with

Since v2i — £i£i > 0, K2 = K2 — pv\2 forms a smaller value of K2 in (34).

Note that the pair of equations giving the smallest values of Ki and K2
will still give X = 1 in (32). At each stage we repeat the above generating
of a smaller Ki or K2 until eventually Ki = Li, K2 = L 2 . By Theorem 5 this
will come about when we obtain equations of the type (34) and (35) with
Ki > v2\ and K2 > V\2.
To illustrate we find 5(137, 251, 256). First calculate that
ax - 75a2 + 73a 3 = 0.
Then by the algorithm we obtain
3ai + 31a2 - 32a 3 = 0,
7ai - 13a2 + 9a 3 = 0,
17ai + 5a2 - 14a3 = 0.
Thus the matrix of detached coefficients is

01 a2 «3

-24 8 5
7 -13 9
17 5 -14

and B = 24ai + 9a 3 = 5,592.

It should be pointed out that solving for (34) is not always necessary.
Many computational short cuts become apparent after some practice. Note
that the suggested algorithm is not merely numerical but gives algebraic
relations as well, enabling one to solve all previously solved special cases for
t = 3 by a unified approach. For example, see the end of the next section.

6. Extensions and restatement of basic theorem. Even if L 3 = 1, the

statement of Theorems 4 and 5 still holds, dropping the minimality con-
dition on the Lt. In this case, B = a\a2 + a3, see (2). But the matrix of co-
efficients is

https://doi.org/10.4153/CJM-1960-033-6 Published online by Cambridge University Press

 A LINEAR DIOPHANTINE PROBLEM 397

dl a2 a3
— a2 dl 0
a2 — Xzi — & i — x32 1
#31 #32 -1
with x3i < a2. Then X = 1, so that Theorem 5 gives the same result a\a2 + a3.
Next we show that Theorems 4 and 5 hold even though the a / s are not
reduced to a pairwise relatively prime set bu b2, #3.
We compare the L's and x*/s associated with ax, a2, a 3 with those L"s and
# i / ' s associated with 61, b2j bs. From (15r), Ltai = xtiaj + xikaky we see that
dJk\Lu dij\xik, dik\xij. Thus, setting Lt = dûkL(, xik = dijXik, we have
(35) LjLfc — x^x^- = flf if and only if L)Lk — x^4y = bu
since dijdik\JLj JLk Xjk xkj ) = dijdikbi = a^
Finally, all these results can be collected in the following form :
THEOREM 6. For (ai, a2, #3) = 1, define B to be the largest number not of the
form xa\ + ya2 + sa3, x, y, z > 0. Then for cyclic permutation of subscripts
B = L&i + max (xjkakl xkjaj),
where
Ltiai = xijaj -j- Xikak, JLi j> u, Xfj ^ (J, Xik ^ (J, Lu = x ^ -r x&i

-Z-/ î-i-/1' x ï jOC j i — ak .

The Z/s and x's can be found either by the computing algorithm discussed in
§5, modified to solve first for dif,k in terms of at and ah or by first applying
Theorem 2.
In conclusion, observe that the special cases previously obtained for t = 3
can be derived directly from the results of this paper.
Example. We can extend the results stated in (5) for B{a, a + 1, a + z).
Write a = kz — u, 0 < u < z, & > 1, 2 > 2. Then for u < k + 1 the co-
efficient matrix is

a = ai = kz — u a2 = kz — u + 1 a 3 = kz — u + z

— (z + k — u) z —u k — 1
2 - 1 - 2 1
& + 1— U U —k

If u < 1, then

£ = L3#3 + xua2 = ( ) (a + 2) + (z — «)(a + 1).

https://doi.org/10.4153/CJM-1960-033-6 Published online by Cambridge University Press

 398 S. M. JOHNSON

To correspond to the notation of (5), we solve for C + 1 = B + 1 — X ai-

Then

-{*¥)•
C + 1 = [ ---^--- ) a + (z - 2 - M )a.

If u > 1, then £ = L 3 a 3 + X2i#i = k(a + z) + (z — l)a, and

C+ 1 = + *) + ( * - 3)a

since
a + u a+ 1
+ 1.
s J
For / > 3, Theorem 1 holds and the author has verified that relations
analogous to Theorem 4 hold in many cases. However, this will be the subject
of a later paper.

REFERENCES

1. P. T. Bateman, Remark on a recent note on linear forms, Amer. Math. Monthly, 65 (1958).
517-518.
2. A. T. Brauer, On a problem of partitions—I, Amer. J. Math., 64 (1942), 299-312.
3. A. T. Brauer and B. M. Seelbinder, On a problem of partitions—•//, Amer. J. Math., 76
(1954), 343-346.
4. R. J. Levit, A minimum solution for a diophantine equation, Amer. Math. Monthly, 63
(1956), 646-651.
5. J. B. Roberts, Note on linear forms, Proc. Amer. Math. Soc. (1956), 465-469.
6. • On a diophantine problem, Can. J. Math., 9 (1957), 219-223.

The Rand Corporation

Santa Monica, California

https://doi.org/10.4153/CJM-1960-033-6 Published online by Cambridge University Press