On $A^4 + B^4 + C^4 = D^4$

Author(s): Noam D. Elkies

Source: Mathematics of Computation, Vol. 51, No. 184 (Oct., 1988), pp. 825-835
Published by: American Mathematical Society
Stable URL: http://www.jstor.org/stable/2008781
Accessed: 22-10-2016 22:02 UTC
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

American Mathematical Society is collaborating with JSTOR to digitize, preserve and extend access to
Mathematics of Computation

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms

MATHEMATICS OF COMPUTATION
VOLUME 51, NUMBER 184

OCTOBER 1988, PAGES 825-835

On A4 + B4 + C4 =D4
By Noam D. Elkies
Abstract. We use elliptic curves to find infinitely many solutions to A4 + B4 +C4 = D4
in coprime natural numbers A, B, C, and D, starting with

26824404 + 153656394 + 187967604 = 206156734.

We thus disprove the n = 4 case of Euler's conjectured generalization of Fermat's
Last Theorem. We further show that the corresponding rational points (?A/D, ?B/D,

?C/D) on the surface r4 + s4 + t4 = 1 are dense in the real locus. We also discuss the
smallest solution, found subsequently by Roger Frye.

1. Introduction. Euler conjectured in 1769 that the Diophantine equation

A4 + B4 + C4 = D4, or more generally

AN + AN + + AN-1 = AN (N ) 4),
has no solution in positive integers. (See [4, pp. 648ff.] for the early history of

this and related problems, and [5, Problem D1] for more recent research.) Nearly
two centuries later, a computer search [7] found the first and hitherto only known
counterexample to the general conjecture,

275 + 845 + 1105 + 1335 = 1445

but direct computer searches found no counterexample for N = 4, even though this
first case of the conjecture could not be proved.

In this paper we disprove this conjecture, exhibiting several counterexamples and

giving a recursive construction of infinitely many solutions for A4 + B4 + C4 = D4

in relatively prime natural numbers A, B, C, D. Since that Diophantine equation
is homogeneous, solving it is equivalent to finding a point

(r, s, t) = B D
on the surface r4 + s4 + t4 = 1 with rational coordinates r, s, t. In Section 2, we
start with an analysis of a parametrization of the simpler equation r4 + s4 + t2 = 1
as a pencil of conics. This yields a parametrization of r4 + 84 + t4 = 1 as a pencil
of curves of genus one. We consider this parametrization in Section 3, and find
the simplest curve in the pencil which could possibly have a rational point that
would disprove Euler's conjecture. It happens that there is such a rational point
of sufficiently small height to have been found by a direct computer search; this
produced our first solution

26824404 + 153656394 + 187967604 = 206156734.

Received October 14, 1987; revised January 21, 1988.

1980 Mathermiatics Sub)ject (CIlssiication (1985 Reviision). Primary 11D25, 11G05.

?1988 American Mathematical Society

0025-5718/88 $1.00 + $.25 per page

825

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms

826

NOAM

D.

ELKIES

(This solution was beyond the range of earlier exhaustive searches. We could only
find it by restricting the variables to lie on an appropriate curve. The transformation from the natural coordinates on that curve to the values of r, s, t required
rational functions sufficiently complicated (see (6) below) to get from the 2- and

3-digit numerator and denominator of the X-coordinate on that curve to our solution's 7- and 8-digit numbers.) We then show, in Section 4, how to use the theory
of elliptic curves to recursively generate arbitrarily many other solutions from our
first one. Throughout this paper we shall assume and use basic definitions and facts
about elliptic curves, for which a good reference is [9].

The author gratefully acknowledges the support of Harvard University's Society

of Fellows and the Communications Research Division of the Institute for Defense
Analyses for the research reported in this paper. The numerical and symbolic
computations needed for this research were greatly facilitated by the computer
program MACSYMA.

2. The Surface r4 + s4 + t2 = 1. In [3, p. 135] Demjanenko expresses the

surface

(1)

r4

+s4

+t2

as a pencil of conics parametrized by u:

(2a)

x+y,

x-y;

(2b) (U2 + 2)y2 = -(3U2 - 8u + 6)x2 - 2(u2 - 2)x - 2u,

(2c) (U2 + 2)t = 4(U2 - 2)x2 + 8ux + (2 - u2).
[We also have the conic y2 = -2x - 3x2, t = 4x2 _ 1 corresponding to the limit
u -* ox; this special case is equivalent to a parametrization given by Escott in

the late 19th century (see [4, p. 658]) we shall say more about this, and the
similar parametrization with u = 0, in Section 3. The parametrization (2) has been
independently rediscovered at least three times, by Andrew Bremner, Don Zagier
and the present author. Bremner wrote (1) as

2(1 + r2)(1 + 82) = (1 + r2 + s2)2 + t2

and factored both sides over Q(VZT) [1]. Zagier, generalizing from several special
cases of (2) communicated to him by de Vogelaere, observed that Escott's parame-

trization is equivalent to the identity 1-- s4 = Po- 2QoRo with

PO = 4X2-1, Qo=y+3x2+2x, Ro=y+3x2 -2x

(x,y given by (2a)), and, applying to this identity the automorphism gro
morphic to PSL2(Q)) of the ternary quadratic form p2 - 2QR, obtained in
many representations of 1 - r-4 as p2 - 2QR, and so infinitely many conics
Q = 0 on which 1 - r4- 4 is a perfect square ([10], which also finds the conditions

of Lemmas 1 and 2 below). The present author looked directly for an ellipse in the
rs-plane tangent to the Fermat quartic r4 + 84 = 1 at four points.]

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms

ON

A4

+BC4

D4

827

Solving (2b) for u, we find

-1 +4x2 1 (2x4+12x2y2+2y4j

U 3x2 + y2 + 2x

-1 + (r +s)2+ -r4-s4

r2 + rs + S2 + r + s

-1 +(r+8)2? it
r2 + rs + S2 + r + s'
and (2c) then selects the plus sign. Thus every rational solution of (1) lies on
the conic (2) for some rational or infinite value of u. Furthermore, the involution

u t-4 2/u merely replaces (r, s, t, x, y) by (-s, -r, -t, -x, y) in the parametrization
(2). Thus we may take u of the form 2m/n with m and n relatively prime integers,
m > 0 and n odd, because otherwise 2/u is of that form and the two corresponding
conics (2) are essentially the same.
We may now write (2b,c) in the form

(3b) (2m2 + n2)y2 = -(6m2 - 8mn + 3n2)x2 - 2(2m2 - n2)X - 2mn,

(3c) (2m2 + n2)t = 4(2m2 - n2)x2 + 8mnx + (n2 - 2m2).
If rational numbers x and y satisfy (3b) then we may recover t from (3c) and r, s

from (2a) and thus find a rational solution to (1); so we need only consider the conic
(3b). It will be convenient to define the functions S(k), R(k) of a nonzero integer k
by S(k) = the largest positive integer whose square divides k and R(k) =kS2(k);
for instance, for k = +23, +24, ?25 we have S(k) = 1, 2, 5 and R(k) = +23, +6, ?1.
Then we have

LEMMA 1. The conic (3b) has infinitely many rational points (x, y) if
R(2m2 + n2), R(2m2 - 4mn + n2)

are both products of primes congruent to 1 mod 8, and none otherwise.

Remarks. i) In particular, 2m2 - 4mn + n2 must be positive, that is, u2 -4u +

2 > 0 sO ju - 21 > v'-, else even the real locus of (3b) and a fortiori also the
rational locus is empty. This is an example of a "local condition at infinity";

the congruence condition on the prime factors of 2m2 + n2 and 2m2 - 4mn + n2
comes from local conditions at these finite primes. Since the curve (3b) has genus

zero, these necessary local conditions for the existence of a rational point are also
sufficient ("Hasse principle"). Compare this with the situation for curves of genus
one to be encountered in the next Section.

ii) By Quadratic Reciprocity, the prime factors of R(2m2 + n2) are already all
congruent to 1 or 3 mod 8, and those of R(2m2 -4mn + n2) to +1 mod 8, so the
1 mod 8 condition is not as stringent as it may first appear.

Proof. First reduce (3b) to the standard form

(4) X2 + ay2 + bZ2 = 0

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms

828

NOAM

D.

ELKIES

with a and b squarefree integers, by "completing the square":

[2mn + (2m2 -n2)X]2

= 2mn[2mn + 2(2m2 - a2)x + (6m2 - 8mn + 3n2)x2]

+ (4m4 - 12m3n + 12m2n2 - 6mna3 + n4)X2

= -2ma(2m2 + n2)y2 + (2m2 - 2mn + a2)(2m2 - 4mn + n2)x2,

and taking X = 2mn+(2m2 -n2)x, Y = ay, Z = 3x, where a = S (2mn(2m2+n2))

and 3 - S((2m2 - 2mn + n2)(2m2 - 4mn + n2)); then a = R(2mn(2m2 + n2)) and

b = R((2m2 - 2mn + n2)(2m2 - 4mn + n2)) in (4). Note that from a nontrivial
rational solution to (4) we may recover a rational solution to (3b) by multiplying

X, Y, and Z by a constant factor to make X = 2mn + (2m2 - n2)x and Z = Ox

consistent.

But it is known that, in general, an equation (4) has a rational solution -indeed,
infinitely many rational solutions if and only if at least one of a and b is negative
and -a and -b are congruent to squares modulo b and a respectively (see [6, pp.

272-275] for an effective algorithmic proof). Since n is odd and m, n are relatively
prime, we easily show that m, n, 2mi2+n2, 2m2-2mn+na2, and 2mi2-4mn+n 2 are
relatively prime in pairs, and thus that we need only ask that -2mn(2mi2 + n2) be

a square modulo each prime dividing R(2mi2-2mn + a2) and R(2M2 - 4mn + nr2),

and that (2mi2 - 2mn + n2)(2m2 - 4mn + n2) be a square modulo each prime

dividing R(m), R(n) and R(2m2 + a2). Three of these five conditions always hold:

(2M2 -2mn + 2)(2m2 - 4mn + n2) is congruent modulo m and n to the squares
n4, 4mi4 respectively, and
-2mn(2mi2 + n2) _ (2m2 -- n2)2 mod 2M2 - 2mn + n2
The remaining two conditions yield the lemma's constraints, for

(2m2 - 2mn + na2)(2m2 - 4mn + a2) = 2(2ma)2 mod 2M2 + n2,

so each prime factor of R(2m2 + n2) must have not only -2 but also +2 as a
quadratic residue; likewise

-2mn(2m2 + n2) - -2(2mn)2 mod 2m2 - 4mn + n2 2(m - n)2 - n2,

so each prime factor of R(2m2 - 4mn+n 2) must have both -2 and +2 as quadratic
residues; thus all these primes must be congruent to 1 mod 8. 'Note that, since n

is odd, so are 2i2 + n2 and 2m2 - 4mn + n2, whence 2 cannot occur as a prime
factor.) Finally, if R(2m2 -4mn+n 2) = -b is a product of positive primes then b
is negative in (4), so we are done. oI

For instance, we may take u = 4, when (in) = (2,1) satisfies the hypotheses
of Lemma 1; then (3b) becomes 9y2 - -1kX2 - 14x - 4, for which we find by
inspection the rational solution (x, y) = (- , 6) and recover from (2) the solution

(:,St) = (3,2, ) to (1). Furthermore, projecting from the known point (x, y) =
(- 2, 6) we find the parametrization

( k2+2k+17 k2+6k- 11'

(2k2+2 22 6k2 +66 /

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms

ON

A4

+B4

+C4

D4

829

of the conic 9y2 = -11X2 -14x - 4 (this is the other point on the intersection of

the conic with the line of slope k/3 through (- ', 6) ), from which we recover a
parametric solution

= 2k2?+6k+20 k2 + 31 4(2k4-3k3 + 28k2 75k + 80)\

,r, s, _ 3k2 + 33 '3k2 + 33' (3k2 + 33)2 j

to (1). In general, whenever u satisfies the hypotheses of Lemma 1 we will find a

similar parametric solution of (1) with r and s of degree 2 and t of degree 4 with
square denominator.

3. The Surface r4 + s4 + t4 = 1. To find a rational solution (r, s, t) of

(5)

r4

+s4

+t4

we must solve (1) with the additional restriction that ?t be a square. Reasoning

as before, we see that such a solution must necessarily have

(6a)

r=x+y,

s=x-y;

(6b) (2M2 + n 2)y2 =-(6m2 - 8mn + 3n2)X2 - 2(2m2 - n2)x - 2mn,

(6c) ?(2m2 + n2)t2 = 4(2m2 - n2)x2 + 8mnx + (n2 - 2m2)
for some relatively prime integers m, n with n odd. For example, take (m, n) =

(0,1) to get y2 = -3x2 + 2x, ?t2 = 1 - 4x2. The first conic has the obvious point
(x, y) = (0, 0), from which we find the parametrization
2

k2?3 i y=kx,
so

= -1 - 4x2 - k4 + 6k2 - 7

?t =1- 4x - (k2 +3)2

or, with the new variable z = (k2 + 3)t,

?Z2 = k4 + 6k2 - 7.
These are two curves of genus one with rational points (k, z) = (1,0) and are
thus elliptic curves. To bring them into Weierstrass form, perform the change of

coordinates k = 1 - 4/(1 p X), z = 8Y/(1 T X)2 to obtain the elliptic curves

y2 = X3 + XpTF2.
These curves are listed as #112A and #56C in [2, pp. 96 and 87], where we find
that they have only two and four rational points respectively: the point at infinity,

the 2-torsion point (X,Y) = (?1,0), and (for the curve y2 = X3 + X + 2) the
4-torsion points (X, Y) = (1, ?2). These correspond to the trivial solutions of (5)
which are permutations of (?1, 0, 0).
To find nontrivial solutions we must choose different m and n. We then obtain

other curves of genus one (the values of u for which that curve degenerates to
genus zero are not rational); these curves will necessarily be principal homogeneous
spaces for some elliptic curves, but need not be globally trivial homogeneous spaces

like the two we obtained for (m, n) = (0, 1) that is, they might not contain any
rational points. To narrow down our choices of (m, n), we first use Lemma 1 and

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms

830

NOAM

D.

ELKIES

the analogous

LEMMA 2. The conic (6c) has infinitely many rational points (x, t) if
R(2m2 - 2mn + n2), R(2m2 + n2) and R(2m2 + 2mn + n2)

are all products of primes congruent to 1 mod 8, and none otherwise.

Proof. Again we complete the square to find

[4mnx + (n2-2m 2)= (n2 - 2m2)[(n 2- 2m2) + 8mnx + 4(2m2 - n2)x2] + (16m4 + n4)x2

= F-(2m2 - n2) (2m2 + n 2)t2

+ 4(2m2 -2mn + n 2)(2m2 + 2mn + n 2)x2.
As in the proof of Lemma 1, we find that 2m2 + 2mn + n2 and 2M2 ? n2 are

relatively prime in pairs and we need only ask that qF(4m4 - n4) be a square
modulo R(2m 2-2mn+n 2) and R(2m2 + 2mn + n2) and that 4m4 +n4 be a square
modulo R(2m2 - n2) and R(2m2 + n2). (Since 16m4 + 4n4 > 0 the negativity
condition is automatically satisfied.) We find that

4mn4 _ n4 8m4 _-2n4 mod 2m2 + 2mn + n2,

so -2 and 2 must both be quadratic residues of each prime factor of

R(2m2 - 2mn + n2);

also

4m4 + n4 _ 2n4 -(2mn)2 mod 2m2 + n2,

so -1 and 2 must both be quadratic residues of each prime factor of R(2m2 + n2);
thus all these prime factors are congruent to 1 mod 8. The remaining condition is
always satisfied since
4m4 + n4- (2mn)2 mod 2m2 - n2,
so we are done. 0

In particular, m must be divisible by 4. The first few (m, n) which satisfy the

conditions of both Lemma 1 and Lemma 2 are the pair (0,1) already encountered,

(4,-7), (8,-5), (8,-15), (12,5), (20,-1), and (20,-9). Taking (m,n) = (4,-7)
fails, because then (6b) and (6c) become

81y2 = -467x2 + 34x + 56, ?81t2 = -68X2 - 224x + 17

which cannot simultaneously hold: the second equation forces x to have denominator not divisible by 5 with x 4 mod 5, and the first equation then refines this to

x _ 14 mod 25, when

(-68X2 - 224x + 17)/25 3 mod 5

cannot be of the form +81t2. We next try (m, n) = (8, -5) and obtain
153y2 = -779x2 - 206x + 80, ?153t2 = 412X2 - 320x - 103.

On the first conic we find by trial and error (or by applying the algorithm in [6]
the small rational solution (x, y) = (3/14,1/42) and thus the parametrization
51k2 - 34k - 5221 17k2 + 7558k - 779

(7) X 14(17k2+779) y= 42(17k2+779)

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms

ON

A4+B4+C4=D4

Substituting

this

831

value

of

in

+ 212(17k2 + 779)2t2

(8) =-4(31790k4 - 4267k3 + 1963180k2 - 974003k - 63237532).

The right side of this reduces modulo 3 to (x2 _ x - 1)2, from which we easily show
that for (8) to hold with rational x and t, the plus sign must be chosen. Using new
coordinates X = (k + 2)/7, Y = 3(17k2 + 779)t/14, we then further simplify (8) to

(9) y2 = -31790X4 + 36941X3 - 56158X2 + 28849X + 22030.

We easily verify that there is no local obstruction to a rational solution for
(9). This does not yet guarantee that such a solution exists, because (9) could
still represent a nontrivial element of the Tate-9afarevi6 group of the elliptic curve
which is its Jacobian. But it does encourage us to look for small rational solutions,
and in this case a few hours' computer (VAX) search for rational X such that the
right-hand side of (9) is a perfect square revealed the solution
/ 31 30731278'\

(467' 4672];
retracing our changes of variable we then recover the rational solutions

(10) (r~t)t(18796760 2682440 15365639\

(10) (r~s~t) ' 20615673' 20615673' 20615673]

for (5). Clearing denominators we obtain our first counterexample

26824404 + 153656394 + 187967604 = 206156734

to Euler's conjecture.

4. More Rational Solutions of (5). From our single solution to (9) we may
now compute arbitrarily many others:

PROPOSITION. There are infinitely many rational X that make the right-hand
side

-31790X4 + 36941X3 - 56158X2 + 28849X + 22030

of (9) a square. These yield infinitely many rational solutions (r, s, t) of (5).
Proof. We know two rational points

P?: (XY)= ( 31 30731278

467' 4672/

on the elliptic curve (8), so we need only show that the difference Q = P -P_
between them is of infinite order in the group of the Jacobian of that curve, i.e., is
not a torsion point. By [8, Theorem 2], there are only finitely many groups that

can occur as the rational torsion subgroup of an elliptic curve, and in particular
no torsion point can have index greater than 12; this reduces the proof to a finite,
if tedious, computation to show n Q : 0 for n = 2,3,... ,12. (Actually, we
need not invoke Mazur's deep theorem here; we may instead compute the NeronTate canonical height of Q and find it positive, or show directly that Q is not a

torsion point in the p-adic completion of (9) for some p. But Mazur's theorem
later simplifies considerably our proof that the rational points on (5) are dense

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms

832

NOAM

D.

ELKIES

in its real locus, so it seems natural to use that theorem here as well.) We can
reduce the tedium by noting that the Jacobian of (9) has a rational point of order

2, corresponding to (x, y, t) (x, -y, -t) in (6b,c) this does not depend on our
choice u =-16/5 or

(X, Y) i~(323 - 535X 291716

535 + 17X ' 17X + 535Y/

in (9). Thus by Mazur's theorem we need only check that nr Q gives neither 0 nor
a 2-torsion point for n = 2,... ,6; and this indeed turns out to be the case. El

A note on the addition law on (9), or generally an elliptic curve E given in the

form y2 = quartic(X) with a known pair of rational points P? = (X0, ?YO): Take

the point P_ to be the origin of our addition law, so P+ is identified with Q = P+ P_. It is then possible to find coordinate functions on E which put it in Weierstrass
form, and then compute the addition law in the usual way by "chords and tangents",

but these coordinate functions tend to have monstrously large coefficients even
when the coefficients of E's defining quartic are only moderately large as in (9).
It is more convenient to compute the addition law directly in terms of the given

coordinates X, Y, using secant and tangent parabolas Y = aX2 + bX + c. Indeed,

if such a parabola meets E in four points P1, P2, P3, P4 counting multiplicity, then
P1 + P2 + P3 + P4 = 2Q in E's group law, because

(X-XO)-2(aX2 +bX+c-Y)
is a rational function on E with divisor P1 + P2 + P3 + P4 -2 (P+ + PF). Given P1,
P2 and P3, we can then solve the linear equations for a, b, c to make the parabola go
through P1, P2, P3, and find the X-coordinate of P4 as the fourth zero of a quartic

with three known roots, and the Y-coordinate as a known quadratic in X. So, for
instance, to compute the coordinates of -Q, find a, b, c such that the parabola

Y = aX2 + bX + C has a point of triple contact with E at P+ (i.e., take

aX2+bX+c= Yo+a(X-Xo)+ O(X-Xo)2
to be the beginning of the Taylor expansion of Y at X = X0), then substitute

aX2 + bX + c for Y in the equation of E to obtain a quartic in X with a known

triple root at X0, so the remaining zero is easily computed. For instance, for our

curve (9), we compute

937766474523

467 15365639'
2096569897386251210893331

2 153656393
2096569897386251210893331

a =- 2 .153656393

b 334937219677623362815466
b= -153656393
1076124066222818157529571

c= 2 153656393
from which we find that -Q has X-coordinate

127473934493966820221865642313563283

129759559485872431282952710668698569'

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms

ON

A4

B4

C4

D4

833

and eventually recover our second solution

A = 1439965710648954492268506771833175267850201426615300442218292336336633,
B = 4417264698994538496943597489754952845854672497179047898864124209346920,
C = 9033964577482532388059482429398457291004947925005743028147465732645880,
D = 9161781830035436847832452398267266038227002962257243662070370888722169

to A4 + B4 + C4 = D4 in coprime integers A, B, C, D. Likewise, we can compute

n Q for n = 2,..., 6 to verify that Q is not a torsion point of (9); fortunately,
this requires only the coordinates of these points as single-precision real numbers,

since the corresponding integer solutions to A4 + B4 + C4 = D4 are too huge to be

profitably displayed even in a Math. Comp. article!

While we now have infinitely many rational points on the surface (5), they all
lie on the same curve. We proceed to show how to produce rational points off that

curve, and indeed enough rational points to comprise a dense subset of the real

locus of (5). It will be convenient to henceforth drop the requirement that u be of

the form 2m/n; instead of choosing between u and 2/u, we may choose the sign
in (6c), and will use the minus sign (the reason for this choice will appear later).
Our curve (9) then corresponds to u = 2/(-16/5) = -5/8. We now combine two

observations: we have already seen that any rational solution of (5) necessarily
comes from a rational

(11) ~~~~~~~-1 + (r + 8)2 - t

r2 + rs + 82 + r + s
in the parametrization (6), and we have implicitly used the 48 linear symmetries of the surface (5) generated by replacing each of r, s, t by their negatives
and permuting them. Of these, only four (generated by (r, s, t) H-+ (r, s, -t) and

(r, s, t) - (s, r, t)) preserve the value of u in (11). This leaves us with 48/4 = 12
different curves for each of our solutions to (5), only one of which corresponds to

u = -5/8. Thus, from each of our infinitude of solutions to (5) with u = -5/8
we obtain a few new values of u for which we know a rational solution to (6), and
for each such u we can expect to find infinitely more as in our proof of the above
Proposition. With a little more work we find that these solutions suffice to prove:

THEOREM. The rational solutions of (5) are dense in the set of real solutions.
Remark. In particular, it follows that there are infinitely many admissible u,
so infinitely many pairs (m, n) of relatively prime integers with n odd and

R(2m2 - nmn + n2) a product of primes congruent to 1 mod 8 for ,c = -2, 0, 2, 4;

and in fact u = 2m/n can be taken arbitrarily close to any real number outside

(2 - V's, 2 + V's). This result is hardly surprising, for on probabilistic grounds one

expects the number of admissible (m, n) with Iml, ll < N to be asymptotically

proportional to (N/ log N)2 for large N, but I cannot see how to simply prove the
infinitude of admissible u, let alone obtain their asymptotic distribution.

Proof. First note that, for any real solution (r, s, t) of (5) with rst :A 0, at least
one of the twelve possible values of u is negative or infinite: by replacing r or s by
their negatives if necessary, we can make r > 0 and s < 0 with r + s ) 0, when the
denominator
r2+ rs + s2 + r + s = r2 + (1 + s)(r + s)

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms

834

NOAM

D.

ELKIES

in (11) is positive and the numerator

-1 + (r + s)2 -t2 <t2

is negative (this is why we chose the minus sign in (6c)!). Next we show that, by
replacing s by -s in the family of rational solutions of (5) described in our proof
of the Proposition, we can obtain values of u arbitrarily close to any given negative
real number. Indeed, the real locus of (9) is connected, because the right-hand side

is 'positive only for X between its two real roots, -.3828... and .9987..., so its
infinite rational subgroup is dense; and the new value

7480X2 - 18500X + 6068

357X2 + 11286X + 1605

of u ranges from -oo at X = -7+ to a positive value at X = 2, and so attains

every negative value for some real solution of (9). Given a real point on (5), we
thus obtain from the rational points Pm = m Q on (9) points Q'(Pm) on elliptic
curves E'(Pm) that pass arbitrarily close to it. It remains to show that only all but
a finite number of these have infinite order and that the real loci of these curves are

connected. The former is easy: by Mazur's Theorem there are only eleven possible

orders n = 1, 2,..., 10, 12 for a rational torsion point of an elliptic curve over Q;
for each of these, either n Qm = 0 on Em for finitely many m, or nr Q(P) = 0

on E(P) identically for all P in (9), rational or not. Now for n = 1, the equality
n Q(P) = 0 means t = 0 which certainly does not hold identically; however, it

does hold for some (complex) P, and for P' sufficiently near P, the point Q(P) is
too close to zero to be an n-torsion point for any n < 12. Finally, to verify that the

real locus of (6) is connected, we need only check that the values of the right-hand

side of (6b) at the two (necessarily real) roots of the right-hand side of (6c) are of
opposite sign, or equivalently that their product,

- [4(<212)1 (7u4 - 48u3 + 10Ou2 - 96u + 48),

is negative; but that is clearly true for all u < 0. 0
Postscript. While our first counterexample

(A, B, C; D) = (2682440,15365639,18796760; 20615673)

to Euler's conjecture still seems beyond the range of reasonable exhaustive computer
search, there remained the possibility that smaller solutions may be found by such a
search. Shortly after hearing of the first solution, Roger Frye of Thinking Machines

Corporation asked whether it was minimal; I did not know, but suggested how one
might exhaustively search for smaller solutions: eliminating common factors and
permuting A, B, C if necessary, we may take D odd and not divisible by 5, and

C < D such that D4 - C4 is divisible by 625 and satisfies several other congruence
and divisibility properties, and for each such D and C look for a representation of

D4 - C4 as A4 + B4 with A, B divisible by 5. Frye translated this into a computer

program and ran it on various Connection Machines for about 100 hours to find

the minimal counterexample to Euler's conjecture:

958004 + 2175194 + 4145604 = 4224814.

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms

ON

A4

+B4

+C4

D4

835

He continued the search and found that this solution is unique in the range D < 106.
This solution appears on the parametrization (6) with (m, n) = (20, -9). We
include Frye's result with his permission.

Department of Mathematics
Harvard University

Cambridge, Massachusetts 02138

1. A. BREMNER, personal communication, Aug. 1987.

2. B. J. BIRCH & W. KuYK, Editors, Modular Functions on One Variable IV, Lecture Notes in
Math., vol. 476, Springer-Verlag, New York, 1975.

3. V. A. DEMJANENKO, "L. Euler's conjecture," Acta Arith., v. 25, 1973-74, pp. 127-135.
(Russian)

4. L. E. DICKSON, History of the Theory of Numbers, Vol. II: Diophantine Analysis, G. E. Stechert
& Co., New York, 1934.

5. R. K. GUY, Unsolved Problems in Number Theory, Springer-Verlag, New York, 1981.

6. K. IRELAND & M. ROSEN, A Classical Introduction to Modern Number Theory, Graduate

Texts in Mathematics, vol. 84, Springer-Verlag, New York, 1982.
7. L. J. LANDER & T. R. PARKIN, "Counterexamples to Euler's conjecture on sums of like
powers," Bull. Amer. Math. Soc., v. 72, 1966, p. 1079.

8. B. MAZUR, "Rational isogenies of prime degree," Invent. Math., v. 44, 1978, pp. 129-162.

9. J. SILVERMAN, The Arithmetic of Elliptic Curves, Graduate Texts in Mathematics, vol. 106,
Springer-Verlag, New York, 1986.
10. D. ZAGIER, "On the equation w4 + X4 + y4 = Z4," unpublished note, 1987.

This content downloaded from 98.113.130.60 on Sat, 22 Oct 2016 22:02:12 UTC
All use subject to http://about.jstor.org/terms