Taxicab Numbers and the Fermat Surface
ak + bk = ck + dk for
k = 4, 6, 8
Keywords: diophantine equations, equal sums of like powers, taxicab numbers, Fermat-Euler surface, Golden Ratio, Plastic Constant, Weber functions, Dedekind eta function.
I. Introduction
II. For k = 4 and x15 + x25 + x35 + x45 = x55 + x65 + x75 + x85
III. For k = 6 and Multi-grade Identities
IV. For k = 8, the Golden Ratio, and the Plastic Constant
V. Conclusion
The incident regarding Ramanujan and the taxicab number 1729 is an oft-told story. To recall, Ramanujan was hospitalized in Putney and Hardy came for a visit. In Hardy’s own words,
“I had ridden in taxi-cab no. 1729, and remarked that the number
seemed to me rather a dull one, and that I hoped it was not an ill omen. ‘No,’ he replied, ‘it is a very
interesting number; it is the smallest number expressible as a sum of two cubes
in two different ways.’”
Upon hearing this incident, J.E. Littlewood would remark that, “Every positive integer is one of Ramanujan’s personal friends.” While the anecdote gives the impression that Ramanujan then and there made a mental calculation and determined that 1729 had this property, something very much consistent with his reputation as a genius, the behind-the-scenes story was that it seems he already knew 1729 was 13 + 123 = 93 +103 (eq.1) and in fact had it written down in his early notebooks. (See Berndt and Bhargava’s article.) Furthermore, as noticed by this author, he also knew (with a slight rearrangement) the quadratic form identity,
(x2+9xy-y2)3 + (12x2-4xy+2y2)3 = (9x2-7xy-y2)3 + (10x2+2y2)3 (eq.2)
the first instance of which gives the above equation.
Thus, as a personal opinion, it seems the remarkable aspect of this anecdote was three-fold: first, that Hardy remembered at all his taxicab number (when was the last time you did the same?), second, by sheer coincidence, Ramanujan had done previous work related to this particular number, and third, that he remembered a particular property of this number which is just one result in a very extensive body of mathematical work. (And perhaps not that he instantly or intuitively realized it had an interesting property.)
This is not to detract from Ramanujan’s genius—far from it. To illustrate, another aspect of his work is tied up to 1729. This has to do with the curious Ramanujan Sum Identity given by the generating functions,
S anxn = (1 + 53x + 9x2)/R = {1, 135, 11161, 926271…}
S bnxn = (2 - 26x -12x2)/R = {2, 138, 11468, 951690…}
S cnxn = (2 + 8x -10x2)/R = {2, 172, 14258, 1183258…}
where the sum S is from n = 0 to ¥, R is the cubic polynomial 1- 82x - 82x2 + x3, and which yields the surprising result,
an3 + bn3 = cn3 + (-1)n
More of the terms can be found in Sloane’s “On-Line Enclopedia of Integer Sequences”. As an example, note that 1353 + 1383 = 1723 -1 and 111613 + 114683 = 142583 + 1, and so on, thus providing an infinite number of near-misses to the Fermat curve a3 + b3 = c3! One can notice that the leading coefficients of the polynomials involved, including R, are 1, 9, 10, 12, the very same numbers in 13 + 123 = 93 +103. And as pointed out by Hirschhorn, the solutions given by these generating functions in fact are contained in eq.2. As a case in point, let {x,y = 1,9} in eq.2 and we get,
13 + 1383 = (-135)3 + 1723
the same as the first non-trivial result of the Ramanujan Sum Identity. And to think this can somehow be connected to a taxicab once plying the streets of London! (Author’s Note: It might be a good suggestion, not to mention a good tourist attraction, that this cab, if it has been de-commissioned, should be located and displayed on the grounds outside a math department somewhere. Or at least the license plate. )
Ramanujan however wasn’t the first to find eq.1. In the History of the Theory of Numbers, Dickson mentions that in 1657, Fermat raised the problem of finding two cubes expressible as the sum of two other cubes. Without indicating his method, Frenicle de Bessy gave five solutions, one of which was eq.1. (It can be pointed out that Fermat, like Ramanujan, would pose problems the answers to which they already knew. Considering the small numbers involved, it is possible Fermat in fact may have known eq.1 prior to de Bessy!)
This question has already been generalized though: find two kth powers of integers expressible as the sum of two other kth powers, or,
x1k + x2k = x3k + x4k (eq.3)
For k = 2, 3, 4, the smallest solutions are,
12 + 72 = 52 + 52, 13 + 123 = 93 +103, 594 + 1584 = 1334 + 1344
To recall, while the equation x1k + x2k = x3k defines an algebraic curve (the Fermat curve), eq.3 defines an algebraic surface, usually also ascribed as the Fermat surface, though it is sometimes referred to as the Euler surface since Euler found the complete rational parametrization for k = 3 and the first for k = 4.
It is still an open problem whether there are solutions for k > 4. There is a conjecture, Euler’s Extended Conjecture (EEC), formalized by Ekl that contains this as a special case and maintains that for k > 4, there are no non-trivial integral solutions. For k = 5, Ekl (1998) has established that any solution must have a common sum greater than 4.01 x 1030, implying it should have a term on each side > 1,149,000. If there is a first solution, the numbers would be much greater than the ones known for k < 5. In “Euler’s Extended Conjecture and ak + bk + ck = dk for k > 4” [1] we started with k = 3, gave Euler’s complete parametrization and extended his results to the next odd degree, k = 5, though it involved radicals. For this one we’ll start with k = 4 and go to k = 6, 8 again using radicals, then describe a method applicable for any k.
In this section, we’ll give a version of Euler’s parametrization of eq. 3 for k = 4. Adopting the convention established by Lander et al of describing m kth powers equal to n kth powers as (k,m,n), then this is a (4,2,2). While (4,2,2) has been extensively studied by a lot of authors, the novel thing we’ll do here is that to find our solution, we’ll have to go through fifth powers. The basic method is due to Lander, L.J as described in his “Geometric Aspects of Diophantine Equations Involving Equal Sums of Like Powers” [2] and the original objective was to find a solution to the quintic octuple,
x15 + x25 + x35 + x45 = x55 + x65 + x75 + x85
However, he does not seem to have realized, or at least did not point out, that it had the curious consequence of providing a solution to (4,2,2). The method is described in [2] and as the title suggests, involves a geometric approach and derives a solution from an asymptote. We’ll “reverse-engineer” this and start with his result which, if stripped to its basics, ends up with finding a solution to the particular quintic octuple,
(ax2-v1xy+w1y2)5 + (ax2+v1xy+w2y2)5 + (bx2-v1xy+w3y2)5 + (bx2+v1xy+w4y2)5 =
(ax2+v1xy+w1y2)5 + (ax2-v1xy+w2y2)5 + (bx2+v1xy+w3y2)5 + (bx2-v1xy+w4y2)5 (eq.4)
with unknowns w1, w2, w3, w4 and variables a,b,x,y,v1. We’ll part ways with Lander and solve this equation using the same approach in “Ramanujan and the Cubic Equation 33 + 43 + 53 = 63” [3], namely by using the “internal structure” of the original equation to find a system of equations in the unknowns. We do this by subtracting one side from the other, expanding and collecting powers of x and y, equate them to zero, and we get the system,
a3(w1-w2)-b3(w3-w4) = 0
2av12(w1-w2)+3a2(w12-w22)-2bv12(w3-w4)-3b2(w32-w42) = 0
2a(w13-w23)-2b(w33-w43)+v12(w12-w22-w32+w42) = 0
w14-w24-w34+w44 = 0
One can easily see the (4,2,2) equation in the last line. Since we have four equations in four unknowns (v1 is just a free parameter) we can then resolve this into a single equation in a single unknown and hence derive a solution to (4,2,2) that is imbedded in one for (5,4,4)! While with modern computer algebra systems this can be solved by eliminating one unknown at a time, we can make this easier for us by using the same trick Euler used for (4,2,2). This involves doing a linear substitution slightly different from the one he used for k = 3. (Note that it is w1 and w4 in the same side.) Let,
w1 = p+q, w4 = r-s, w2 = p-q, w3 = r+s,
and we get,
a3q-b3s = 0
3a2pq-3b2rs+(aq-bs)v12 = 0
aq(3p2+q2)-bs(3r2+s2)+(pq-rs)v12 = 0
pq(p2+q2)-rs(r2+s2) = 0
Solving for and eliminating s,r,p one at a time and in that order conveniently involves only linear equations, and we end up with a final equation in q as,
(3a2q-bv12)(3a2q+bv12) = 0
We can now solve for q and the others with v1 as a free parameter. To prevent denominators in all of the unknowns, we set v1 = 3a2b2. Choosing the positive solution for q, we find the rest as,
p = a(a6+a4b2-2a2b4+b6), q = 3a2b5
r = b(a6-2a4b2+a2b4+b6), s = 3a5b2
and we have all our four unknowns p,q,r,s which solves,
(p+q)4 + (r-s)4 = (p-q)4 + (r+s)4
This is essentially the same solution Euler that found. And surprisingly, as the wi this can be used to solve the fifth power octuple (5,4,4) given by eq.4.
Two questions remain though: First, are there other forms of (5,4,4) that involve other parametrizations of (4,2,2)? Euler’s is not a complete parametrization nor is it the only one known, as in the same paper [2] Lander derived two more of degree 13 and 19 and other authors found still other formulas. Second, can the method be extended to higher degrees? In general, given an analogous (2m+1,2m,2m), by expanding and collecting powers of x and y, then a (2m,m,m) can be found imbedded in a system of 2m equations in 2m unknowns. For example, it is easy to use a (7,6,6) of similar form, with variables a,b,c, and one ends up with six equations in six unknowns, the last being a (6,3,3). Unfortunately, this is harder to resolve and tentative attempts by the author using similar substitutions indicate that the final equation may not have a non-trivial linear factor, though it may have for other substitutions and for certain a,b,c.
III. For k = 6 and Multi-grade
Identities
In [1] as was mentioned in the Introduction, we extended Euler’s method for (3,2,2) to find a radical parametrization of (5,2,2). In this section, we complete the generalization to make it applicable to (k,2,2) for any k > 2. To recall, given the equation x1k + x2k = x3k + x4k with k = 3, Euler defined,
x1 = p+q, x2 = p-q, x3 = r+s, x4 = r-s,
to get,
p(p2+3q2) = r(r2+3s2) (eq.5)
The general form is then,
Poly(p,q) = Poly(r,s)
with Poly(p,q) and Poly(r,s) as polynomials of the same degree n having the same coefficients ci. By the Fundamental Theorem of Algebra, these two obviously factor as,
(p+h1q)(p+h2q)… (p+hnq) = (r+h1s)(r+h2s)… (r+hns)
for complex numbers hi. Like Euler, we set one factor of Poly(p,q) and Poly(r,s) as sharing a common factor of like form, say u+h1v,
p+h1q = (a+h1b)(u+h1v), r+h1s = (c+h1d)(u+h1v),
Likewise for a second factor, with u+h2v,
p+h2q = (a+h2b)(u+h2v), r+h2s = (c+h2d)(u+h2v),
Using these four equations we can solve for the unknown p,q,r,s, and express them in the new variables a,b,c,d and u,v, giving,
p = au-(h1h2)bv, q = av+bu+(h1+h2)bv,
r = cu-(h1h2)dv, s = cv+du+(h1+h2)dv,
These are the “generic p,q,r,s” and the objective is that by substituting these into Poly(p,q) and Poly(r,s) then both sides obviously will have the two common factors,
(u+h1v)(u+h2v)
which will “fall out” in the equation and hence the new polynomial in u,v is only of degree n-2. One can then solve for u,v and express them in the four free parameters a,b,c,d! We can call this “Euler degree reduction”. For k = 3 given by eq.5 since the total degree for p (or r) is 3, then the resulting polynomial in u,v should only be linear, which enabled Euler to find the complete parametrization of cubic quadruples. For k = 5, in [1] we changed the xi to,
(Öp + Öq)k + (Öp - Öq)k = (Ör + Ös)k + (Ör - Ös)k (eq.6)
call this equation as Ek. E5 is then,
(Öp)(p2+10pq+5q2) = (Ör)(r2+10rs+5s2)
By squaring, the total degree is 5, but since the common factors are squared as well, then final degree is 5-2(2) = 1 so u,v is again linear and the solution is given in [1] . Likewise for k = 6, E6 is the cubic diophantine equation,
(p+q)(p2+14pq+q2) = (r+s)(r2+14rs+s2) (eq.7)
with final degree as 3-2 = 1 so u,v should be expected to be only linear and we’ll give its explicit 4-parameter solution. Using the generic expressions for p,q,r,s we choose the values h1+ h2 = 14 and h1h2 = 1 so,
p = au-bv, q = av+bu+14bv, r = cu-dv, s = cv+du+14dv,
and substituting them into eq.7, after removing common factors, it simplifies to,
((a+13b)(a2+14ab+b2)-(c+13d)(c2+14cd+d2))u = -((a+b)(a2+14ab+b2)-((c+d)(c2+14cd+d2))v
Letting v = (a+13b)(a2+14ab+b2)-(c+13d)(c2+14cd+d2) to avoid denominators, we immediately find u. Explicitly, p,q,r,s are then the beautifully symmetrical expressions,
p = -(a2+14ab+b2)2 + (ac+bc+13ad+bd)(c2+14cd+d2)
q = (a2+14ab+b2)2 - (ac+13bc+ad+bd)(c2+14cd+d2)
r = (c2+14cd+d2)2 - (ac+13bc+ad+bd)(a2+14ab+b2)
s = -(c2+14cd+d2)2 + (ac+bc+13ad+bd)(a2+14ab+b2)
which solves eq.7 or equivalently,
(Öp + Öq)6 + (Öp - Öq)6 = (Ör + Ös)6 + (Ör - Ös)6
The parametrization can be also be true for multi-grade k = 2,4,6 for a,b,c,d satisfying a condition. Still using eq.6, then E2 and E4 are,
E2 = p+q-r-s, E4 = p2+6pq+q2-(r2+6rs+s2)
with E6 as eq.7. Using the explicit formulas for p,q,r,s given above, we find that,
E2 = 12(ad-bc)(a2+14ab+b2-c2-14cd-d2)
E4 = 4(a2+14ab+b2-c2-14cd-d2)*P(z)
where P(z) is a sextic polynomial. Since E6 = 0, then to set E2 = E4 = E6 = 0 one simply has to find non-trivial a,b,c,d such that,
a2+14ab+b2 = c2+14cd+d2
IV. For k = 8, the Golden Ratio, and the Plastic
Constant
It seems there is this intriguing relationship between the Golden Ratio f and its little-known cousin, the Plastic Constant P which involve diverse mathematical concepts such as recurrence relations, Pisot numbers, Weber and Dedekind eta functions, etc and which are far from the topic of this paper. We’ll justify including them so as to better understand the context of our last parametrization. Expanding Ek for k = 8, we get,
p4+28p3q+70p2q2+28pq3+q4 = r4+28r3s+70r2s2+28rs3+s4 (eq.8)
This is a different situation from the others as we are now dealing with an irreducible polynomial of degree > 2. Fortunately, polynomials formed on either side of Ek (eq.6) are related to the tangent function and as such their linear factorizations are expressible in radicals. So the left hand side (LHS) of eq.8 factors over the extension Ö2 as,
(p2+2(7+4Ö2)pq+q2) (p2+2(7-4Ö2)pq+q2)
and similarly for the other side. The generic expressions for p,q,r,s involve the values (h1+ h2) and (h1h2) which indicate that given the polynomials formed by Ek it is enough to factor them down into quadratics, if we want “neat” solutions. (For certain k like k = 7 which involves cubics, any two of its linear factors will do but now it will be a bit messy.) For this one, we choose the positive case (or the negative, as the case may be),
h1+ h2 = 2(7+4Ö2), h1h2 = 1
Using these values for the generic expressions of the p,q,r,s, we then substitute them into eq.8, and we get the new equation LHS = RHS where,
LHS = (a2+2(7+4Ö2)ab+b2)(u2+2(7+4Ö2)uv+v2)*Poly(a,b,u,v)
RHS = (c2+2(7+4Ö2)cd+d2)(u2+2(7+4Ö2)uv+v2)*Poly(c,d,u,v)
with Poly(a,b,u,v) and Poly(c,d,u,v) as 2nd degree polynomials in all variables. Since both sides have the common factor (u2+2(7+4Ö2)uv+v2) which can be removed, then we have reduced eq.8 from being a fourth degree in p,q,r,s to just a 2nd degree in u,v, which is easily solved. One can see that the method can be extrapolated for all k > 2 as was stated earlier, with a non-trivial solution in radicals for all even k < 14.
Because we ended up with an equation in u,v of degree 2, to solve for either u or v inevitably a square root of an expression in a,b,c,d will be involved. But it turns out eq.8 has small solutions in the integers, one being p = 1, q = 15, r = 5, s = 7. With further inspection, a parametric family was found, given by,
p = (x-1)(x2-x-1), q = (x+1)(x2+x-1)
r = x3-x-1, s = x3-x+1
which solves eq.8 or equivalently,
(Öp + Öq)8 + (Öp - Öq)8 = (Ör + Ös)8 + (Ör - Ös)8
This is where f and P comes in. It’s hard to miss that the cubic polynomial (x-1)(x2-x-1), if equated to zero and solved, then one of its roots gives the Golden Ratio f = (1+Ö5)/2 = 1.6180339…. The unexpected thing is that the other polynomial x3-x-1 is interesting in its own right, since if also equated to zero and solved, its unique real root gives the Plastic Constant P = ((3+Ö(23/3))1/3 + (3-Ö(23/3))1/3)/61/3 = 1.3247179… For more on P, see Ian Stewart’s article, “Tales of A Neglected Number” or Weisstein’s “Mathworld”, with a link given in the references.
These two numbers, f and P, have many things in common, five of which will be given here: First, both are the limiting ratios of successive numbers of a sequence defined by a certain recurrence relation. For f, it is for the famous Fibonacci numbers Fn = Fn-1 + Fn-2 (with initial conditions F0 = 0, F1 = F2 = 1.). First few terms are 1,1,2,3,5,8,13… For P, it is for the Padovan sequence Pn = Pn-2 + Pn-3 (with initial conditions P0 = P1 = P2 = 1. Different starting values give the Perrin sequence.). First few terms are 1,1,2,2,3,4,5,7,9… Looking at the recurrence relations again, note that as will be shown later, f is involved with a discriminant of class number two while P is for one with class number three.
Second, both satisfies the algebraic identities x+1 = xm and x-1 = x-n for certain natural numbers (zero and positive integers) m and n. P satisfies P+1 = P3 and P-1 = P-4 while for f it is x+1 = x2 and x-1= x-1. It was proven by Aarts et al that P and f are the only such numbers.
Third, both are Pisot numbers, with P as the smallest. (These are positive algebraic integers greater than 1 all of whose conjugate elements have absolute value less than 1.)
Fourth, both their defining polynomials are Weber class polynomials. As such, their powers can be expressed in terms of powers of the Weber function or the Dedekind eta function. To elaborate more on this, the Dedekind eta function, h(t), is usually defined in terms of the infinite product,
h(t) = q1/24 P (1-qn)
while the three Weber functions f, f1, f2, are,
f(t) = q-1/48 P (1+qn-1/2), f1(t) = q-1/48 P (1-qn-1/2),
f2(t) = (Ö2)q1/24 P (1+qn)
where the product P is from n = 1 to ¥, q = exp(2pit), and t is the half-period ratio. Expressed as Dedekind eta quotients,
f(t) = e-pi/24h((t+1)/2)/h(t), f1(t) = h(t/2)/h(t),
f2(t) = (Ö2)h(2t)/h(t)
The Weber functions also just so happen to satisfy the nice identity f18 + f28 = f 8, which, interestingly enough in the context of this section, involves an eighth power.
As was mentioned earlier, f is the positive root of the Weber class polynomial for two negative even fundamental discriminants, -(4*5) and -(4*10), both with class number 2. It can therefore be expressed in terms of the 24th powers of the first two Weber functions. Given the first, f(t), and using the half-period ratio t = Ö-5,
26(f)6 = f(t)24 = (e-pi/24h((t+1)/2)/h(t))24 = 1148.4334022…
Or the second, f1(t), now with t = Ö-10,
26(f)12 = f1(t)24 = (h(t/2)/h(t))24 = 20607.8012403…
Incidentally these also provide relatively good approximations to the transcendental constant epÖd, with d not surprisingly as,
epÖ5 » 26(f)6 - 24, epÖ10 » 26(f)12 + 24
On the other hand, P is also the real root of the Weber class polynomial for the smallest negative odd fundamental discriminant with class number 3, namely -23. Its 24th power, call it Q, can then be expressed as the negative of the inverse 24th power of the third Weber function f2(t). Let P be the real root of x3-x-1 = 0, so,
Q = P24 = 853.025791919196…
or equivalently, where t = (1+Ö-23)/2,
Q = -f2(t)-24 = -((Ö2)h(2t)/h(t))-24 = 853.025791919196…
Also, a result similar to the above, epÖ23 » 212(P)24 - 24.
Fifth, as shown in this section both their defining polynomials are radical parametric solutions to the Fermat surface x18 + x28 = x38 + x48. In light of the first four connections, perhaps this may be more than just a coincidence.
We’ll end this paper with a few questions, some of which have been asked earlier:
and lastly,
For the second question, we have given solutions for all k up to k = 8, with the exception of k = 7. It might be a bit odd if there are really no solutions for this degree. If there are, then what is the highest possible k with a non-trivial solution?
With the last question, if indeed eq.3 can be proven to have no non-trivial integral solutions for k > 4, regardless of whether one term is zero or not, then this automatically proves Fermat’s Last Theorem (FLT) other than k = 3, 4 which anyway historically have individual and elementary proofs (by Euler and Fermat, respectively) before the much complicated general proof found by Wiles. In this paper, we showed that by taking advantage of the symmetry of the number of terms on each side of eq. 3, it can be susceptible to an algebraic “trick”, a generalization of what Euler used, to reduce the degree k.
A similar approach in fact was taken by Lamé in 1847 when he thought that he found a proof of FLT by using “…the factorization of xn + yn = zn into linear factors over the complex numbers”. However, Liouville pointed out the problem with this approach was that uniqueness of factorization into primes was needed for these complex numbers. For eq.3 since we are dealing with a more symmetrical equation, it should be interesting to know what bearing this symmetry ultimately has on eq.3, especially with regards to whether it means the conjecture about it is easier to solve than FLT or to what extent it can be amenable to an elementary analysis. (Not to mention if Sophie Germain’s partial proof for FLT(5) can be modified to apply to eq.3 for k = 5 and other similar primes.)
It seems one get a lot of returns from contemplating a taxicab number.
--End --
© 2005
Titus Piezas III
Nov. 29, 2005
http://www.oocities.org/titus_piezas/ramanujan.html ® Click here for an index
tpiezasIII@uap.edu.ph ® Remove “III” for email
References: