A proof of Sylvester's conjecture

Abstract: We prove Sylvester's conjecture, originating in his 1879 study of ternary cubic equations, that every prime $p\equiv4,7,8\pmod9$ is a sum of two rational cubes. Elkies announced a proof for the classes $4$ and $7$ in 1994, and Yin recently supplied a complete proof. For the remaining class $p\equiv8\pmod9$, we prove that the elliptic curve $E_p:y<sup>2=x<sup>3+p<sup>2/4$ has analytic rank one, as predicted by the Birch and Swinnerton-Dyer conjecture, and so $p$ is a sum of two rational cubes. The proof begins by adapting the auxiliary Rankin--Selberg construction from the authors' work on the rank one converse for CM elliptic curves. The Rankin--Selberg $L$-function factors as the $L$-function of $E_p$ times a complementary $L$-function. Chan's $3$-isogeny descent and the rank zero converse show that the complementary central $L$-value is non-zero, and so it suffices to prove that a cubic component of the associated Heegner point is non-torsion. A basic difficulty is that the unweighted Hecke trace of the underlying CM orbit vanishes. Our decisive idea is to take $λ$-division before taking the trace, where $λ=1-ω$ and $ω$ is a primitive cube root of unity. We prove that the resulting division boundary is non-zero by analysing Frobenius at $p$. The Galois action on the CM orbit and ramification theory then transfer this non-vanishing to the cubic component.

• The paper proves that every prime p ≡ 8 (mod 9) is a sum of two non-zero rational cubes, addressing the last unsolved case of Sylvester’s conjecture by leveraging $3$-isogeny descent and Rankin–Selberg convolutions.
• The authors use a novel division-boundary method to show that the relevant elliptic curves have rank one, relying on Frobenius obstruction and integral cubic projectors to extend the proof to larger primes.
• A comprehensive analysis of the auxiliary Rankin–Selberg factorization and the vanishing trace verifies that the analytic rank one condition holds, culminating in an elegant proof of Sylvester's conjecture

Main theorem and arithmetic setting

The paper proves the remaining case of Sylvester’s conjecture: every prime $p \equiv 4,7,8 \pmod 9$ is a sum of two nonzero rational cubes. The cases $p\equiv4,7\pmod9$ had been announced by Elkies and subsequently established by Hongbo Yin (Yin, 25 May 2026, Yin, 2 Jul 2026). Burungale and Tian address the inert case $p\equiv8\pmod9$, where the standard split-CM constructions used for the other congruence classes do not apply (2609.14893).

For $m\in\mathbf Q^\times$, the relevant elliptic curve is

$E_m:\qquad y^2=x^3+\frac{m^2}{4}.$

The diagonal cubic $X^3+Y^3=mZ^3$ is $3$-isogenous to $E_m$, and classical $3$-descent gives

$$m \text{ is a sum of two nonzero rational cubes} \quad\Longleftrightarrow\quad \operatorname{rank} E_m(\mathbf Q)>0$$

for cube-free integers $p\equiv4,7\pmod9$0. The central arithmetic result is therefore the following analytic-rank theorem:

$p\equiv4,7\pmod9$1

for every prime $p\equiv4,7\pmod9$2. Gross–Zagier and Kolyvagin then imply

$p\equiv4,7\pmod9$3

together with finiteness of the corresponding Tate–Shafarevich groups. The positive rank of $p\equiv4,7\pmod9$4 yields the desired rational cube representation of $p\equiv4,7\pmod9$5.

The congruence classes arise naturally from the $p\equiv4,7\pmod9$6-isogeny descent and the root number. For $p\equiv4,7\pmod9$7, the curves $p\equiv4,7\pmod9$8 and $p\equiv4,7\pmod9$9 have descent rank bound one and root number $p\equiv8\pmod9$0, so the Birch–Swinnerton-Dyer prediction is analytic rank one. The substantive problem is to prove that the forced central zero is simple.

Auxiliary Rankin–Selberg factorization

The proof introduces an auxiliary prime $p\equiv8\pmod9$1 satisfying

$p\equiv8\pmod9$2

where $p\equiv8\pmod9$3 in $p\equiv8\pmod9$4, $p\equiv8\pmod9$5, and $p\equiv8\pmod9$6 is the cubic residue symbol. Chebotarev’s theorem supplies infinitely many such primes; in fact, the paper proves that the admissible primes have natural density $p\equiv8\pmod9$7.

Let $p\equiv8\pmod9$8 and $p\equiv8\pmod9$9 denote the CM Hecke characters associated with the relevant elliptic curves, and set

$m\in\mathbf Q^\times$0

The associated Rankin–Selberg convolutions factor as

$m\in\mathbf Q^\times$1

and

$m\in\mathbf Q^\times$2

Thus the desired rank-one assertions reduce to two tasks: prove that the complementary factors $m\in\mathbf Q^\times$3 and $m\in\mathbf Q^\times$4 are nonzero, and prove that the two Rankin–Selberg derivatives at the central point are nonzero.

The complementary factors are handled algebraically. Chan’s explicit $m\in\mathbf Q^\times$5-isogeny descent produces a cubic-residue matrix whose determinant is nonzero precisely because $m\in\mathbf Q^\times$6. This gives a $m\in\mathbf Q^\times$7-Selmer group generated by the rational $m\in\mathbf Q^\times$8-torsion point and implies rank zero for the complementary elliptic curves. Burungale and Tian then apply their rank-zero converse theorem for CM elliptic curves to obtain

$m\in\mathbf Q^\times$9

This is a particularly efficient use of the auxiliary prime: the same cubic residue condition controls both the complementary Selmer calculation and the later Frobenius obstruction.

The CM construction and the vanishing trace

The principal geometric difficulty is specific to the inert condition $E_m:\qquad y^2=x^3+\frac{m^2}{4}.$0. The prime $E_m:\qquad y^2=x^3+\frac{m^2}{4}.$1 is inert in $E_m:\qquad y^2=x^3+\frac{m^2}{4}.$2, and the relevant CM orbit has conductor $E_m:\qquad y^2=x^3+\frac{m^2}{4}.$3. The natural $E_m:\qquad y^2=x^3+\frac{m^2}{4}.$4-Hecke orbit is indexed by

$E_m:\qquad y^2=x^3+\frac{m^2}{4}.$5

Let $E_m:\qquad y^2=x^3+\frac{m^2}{4}.$6 denote the images of the orbit points under a normalized modular parametrization to $E_m:\qquad y^2=x^3+\frac{m^2}{4}.$7. Because the CM modular form has Hecke eigenvalue $E_m:\qquad y^2=x^3+\frac{m^2}{4}.$8, the unweighted trace satisfies

$E_m:\qquad y^2=x^3+\frac{m^2}{4}.$9

This vanishing eliminates the usual direct argument based on the non-torsion of a distinguished Heegner point. The paper’s central device is to divide first by the CM endomorphism

$X^3+Y^3=mZ^3$0

and only then take the trace.

The construction begins with a base CM point defined over

$X^3+Y^3=mZ^3$1

A modular cube-root identity produces a point $X^3+Y^3=mZ^3$2 on $X^3+Y^3=mZ^3$3 whose $X^3+Y^3=mZ^3$4-descent class is exactly $X^3+Y^3=mZ^3$5. Equivalently, the authors construct a first $X^3+Y^3=mZ^3$6-division point on the Fermat cubic

$X^3+Y^3=mZ^3$7

with one affine coordinate in $X^3+Y^3=mZ^3$8 and the other in $X^3+Y^3=mZ^3$9. This asymmetry is essential: it allows Frobenius at $3$0 to detect the cubic residue symbol $3$1.

The degree-three Fermat isogeny is identified over $3$2 with multiplication by $3$3 on $3$4. Comparing the $3$5-power Frobenius action on the two coordinate lines with the CM Frobenius endomorphism yields

$3$6

where $3$7 is a nonzero $3$8-torsion point and $3$9 is determined by the sign in the base CM point and by $E_m$0. The nontriviality is equivalent to the condition $E_m$1.

This calculation is the local input from which the global division boundary is extracted.

The division boundary

Choose a $E_m$2-division $E_m$3 of one orbit point, so that

$E_m$4

The division boundary is defined by

$E_m$5

Because the unweighted trace of the $E_m$6 vanishes,

$E_m$7

so the boundary lies in $E_m$8. It is independent of the choice of division point because changing $E_m$9 by a $3$0-torsion point changes the trace by $3$1 times that point, which is zero.

The conductor-$3$2 ring-class extension is totally ramified at $3$3, while the residue field remains fixed. Consequently, all conjugate divisions have the same reduction. The preceding Frobenius calculation then gives the explicit identity

$3$4

where

$3$5

This is the paper’s decisive invariant. It records first-order information that disappears under the ordinary Hecke trace. In Tate-cohomological terms, the trace-zero orbit determines a class in

$3$6

and the connecting homomorphism associated with

$3$7

maps that class to $3$8. The boundary is therefore not an auxiliary computational artifact; it is the cohomological obstruction surviving the vanishing norm relation.

From the boundary to non-torsion cubic components

The boundary alone is a $3$9-torsion point. The next step is to show that it forces the nontrivial cubic character components of the CM orbit to be nonzero and nontorsion.

Let $$m \text{ is a sum of two nonzero rational cubes} \quad\Longleftrightarrow\quad \operatorname{rank} E_m(\mathbf Q)>0$$0 be the maximal $$m \text{ is a sum of two nonzero rational cubes} \quad\Longleftrightarrow\quad \operatorname{rank} E_m(\mathbf Q)>0$$1-power subextension of the conductor-$$m \text{ is a sum of two nonzero rational cubes} \quad\Longleftrightarrow\quad \operatorname{rank} E_m(\mathbf Q)>0$$2 ring-class field, and set

$$m \text{ is a sum of two nonzero rational cubes} \quad\Longleftrightarrow\quad \operatorname{rank} E_m(\mathbf Q)>0$$3

After tracing away the prime-to-$$m \text{ is a sum of two nonzero rational cubes} \quad\Longleftrightarrow\quad \operatorname{rank} E_m(\mathbf Q)>0$$4 part of the orbit, the authors obtain a point $$m \text{ is a sum of two nonzero rational cubes} \quad\Longleftrightarrow\quad \operatorname{rank} E_m(\mathbf Q)>0$$5. Its norm

$$m \text{ is a sum of two nonzero rational cubes} \quad\Longleftrightarrow\quad \operatorname{rank} E_m(\mathbf Q)>0$$6

satisfies

$$m \text{ is a sum of two nonzero rational cubes} \quad\Longleftrightarrow\quad \operatorname{rank} E_m(\mathbf Q)>0$$7

where $$m \text{ is a sum of two nonzero rational cubes} \quad\Longleftrightarrow\quad \operatorname{rank} E_m(\mathbf Q)>0$$8 and $$m \text{ is a sum of two nonzero rational cubes} \quad\Longleftrightarrow\quad \operatorname{rank} E_m(\mathbf Q)>0$$9. Hence $p\equiv4,7\pmod9$00 lies in $p\equiv4,7\pmod9$01.

The cubic projectors are

$p\equiv4,7\pmod9$02

where $p\equiv4,7\pmod9$03 generates the cyclic $p\equiv4,7\pmod9$04-power Galois group. Since the trivial and cubic characters are congruent modulo $p\equiv4,7\pmod9$05, the divided operators

$p\equiv4,7\pmod9$06

are integral in $p\equiv4,7\pmod9$07. They satisfy

$p\equiv4,7\pmod9$08

and a further identity showing that $p\equiv4,7\pmod9$09 is fixed by $p\equiv4,7\pmod9$10 if $p\equiv4,7\pmod9$11.

Assuming a cubic component vanished would therefore produce a point in $p\equiv4,7\pmod9$12 whose $p\equiv4,7\pmod9$13-multiple is $p\equiv4,7\pmod9$14. This is impossible: the $p\equiv4,7\pmod9$15-descent classes of points in $p\equiv4,7\pmod9$16 are $p\equiv4,7\pmod9$17 or $p\equiv4,7\pmod9$18, and neither is trivial in $p\equiv4,7\pmod9$19. The obstruction follows from the distinct ramification of $p\equiv4,7\pmod9$20 and $p\equiv4,7\pmod9$21.

Thus both cubic components are nonzero. They reduce to the identity at $p\equiv4,7\pmod9$22, because the conductor-$p\equiv4,7\pmod9$23 orbit has common reduction and the nontrivial cubic character sums vanish. A formal-group argument at the supersingular prime $p\equiv4,7\pmod9$24 excludes nonzero torsion specializing to the identity. Consequently both components have infinite order.

This establishes, for $p\equiv4,7\pmod9$25,

$p\equiv4,7\pmod9$26

as nontorsion points.

Heegner points and the Gross–Zagier argument

The geometric cubic components are identified with projections of Yuan–Zhang–Zhang Heegner points on the abelian surface $p\equiv4,7\pmod9$27 attached to the pair of coefficient-conjugate CM forms $p\equiv4,7\pmod9$28 and $p\equiv4,7\pmod9$29. The relevant modular parametrizations

$p\equiv4,7\pmod9$30

are realized as the two $p\equiv4,7\pmod9$31-projections of an $p\equiv4,7\pmod9$32-valued modular morphism.

After evaluating the Heegner functional and expanding the adelic integral into a finite sum, the local reciprocity factors reproduce precisely the cubic character weights. The comparison is explicit:

$p\equiv4,7\pmod9$33

Since $p\equiv4,7\pmod9$34 is nontorsion, the corresponding projected Heegner point is nontorsion.

The local toric data select the split quaternion algebra at every finite place and the definite algebra at infinity, so the coherent Shimura curve appearing in the Yuan–Zhang–Zhang formula is the modular curve used in the construction. The Rankin–Selberg root number is $p\equiv4,7\pmod9$35. The Gross–Zagier formula therefore gives

$p\equiv4,7\pmod9$36

Combining this with the factorization and the complementary nonvanishing yields

$p\equiv4,7\pmod9$37

The argument is logically complete: the nonzero Heegner height proves the Rankin–Selberg derivative is nonzero; the nonzero complementary factor transfers simple vanishing to the target elliptic curve.

Explicit computations

The construction is computationally effective. For a fixed $p\equiv4,7\pmod9$38, one selects an auxiliary prime $p\equiv4,7\pmod9$39, evaluates the cubic residue symbol, computes the division boundary, determines the relevant $p\equiv4,7\pmod9$40-power class-group quotient, and numerically evaluates the CM component.

For example, with $p\equiv4,7\pmod9$41 and $p\equiv4,7\pmod9$42, the paper gives an explicit base point and division point:

$p\equiv4,7\pmod9$43

The descent factors are

$p\equiv4,7\pmod9$44

For $p\equiv4,7\pmod9$45, one has $p\equiv4,7\pmod9$46 and

$p\equiv4,7\pmod9$47

The associated rational cube representation is

$p\equiv4,7\pmod9$48

The appendix treats all fourteen primes $p\equiv4,7\pmod9$49 below $p\equiv4,7\pmod9$50. The resulting representations of $p\equiv4,7\pmod9$51 can have very different heights. For instance,

$p\equiv4,7\pmod9$52

whereas the selected representation for $p\equiv4,7\pmod9$53 has a denominator with $p\equiv4,7\pmod9$54 digits, and the representation recorded for $p\equiv4,7\pmod9$55 has numerators and denominator with more than $p\equiv4,7\pmod9$56 digits. The CM component need not be a generator of the Mordell–Weil group: for $p\equiv4,7\pmod9$57, the computed component is identified with $p\equiv4,7\pmod9$58, while for $p\equiv4,7\pmod9$59 the corresponding multiplier is $p\equiv4,7\pmod9$60 before incorporating the auxiliary factor $p\equiv4,7\pmod9$61.

These computations are not used to establish the theorem. The numerical identifications are checked by exact rational arithmetic after recognition, while the non-torsion statements follow independently from the division-boundary argument.

Limitations and open questions

The proof depends on several specialized inputs. The complementary central nonvanishing uses the authors’ rank-zero converse theorem for CM elliptic curves, while the final derivative calculation uses the Yuan–Zhang–Zhang Gross–Zagier formula. The explicit construction also requires selecting an auxiliary prime $p\equiv4,7\pmod9$62 satisfying a cubic-residue condition; Chebotarev guarantees infinitely many choices, but the proof does not produce a uniform canonical choice optimized for arithmetic complexity.

The argument is specific to the CM family $p\equiv4,7\pmod9$63 and to the cubic character structure of the inert case. The paper identifies possible extensions to composite cube-sum parameters and to higher $p\equiv4,7\pmod9$64-divisions, but does not establish such extensions. In particular, it remains open within the paper whether a systematic higher division-boundary theory can yield information beyond analytic rank one or contribute to a fuller BSD formula for these curves.

The numerical appendix also illustrates a practical limitation: the constructive CM procedure can produce rational points of substantially larger height than the smallest known generators, especially for $p\equiv4,7\pmod9$65. The proof establishes non-torsion and rank one, but it does not provide an efficient general algorithm for recovering minimal-height rational cube representations.

Conclusion

The paper resolves Sylvester’s conjecture by proving analytic rank one for both $p\equiv4,7\pmod9$66 and $p\equiv4,7\pmod9$67 when $p\equiv4,7\pmod9$68. Its central contribution is the division-boundary method: when the unweighted CM Hecke trace vanishes, a first $p\equiv4,7\pmod9$69-division retains a nonzero torsion obstruction detected by Frobenius at $p\equiv4,7\pmod9$70. Integral cubic projectors then transfer this obstruction to nontorsion Heegner components. Combined with Rankin–Selberg factorization, complementary rank-zero nonvanishing, and the Gross–Zagier–Kolyvagin machinery, this yields the required rational representations of every prime $p\equiv4,7\pmod9$71 as a sum of two rational cubes (2609.14893).