What is the significance of the paper published by Levent Alpöge regarding the construction of a complex structure on S⁶?
Author: SUNNY99
Robotics Engineering
Link: https://www.zhihu.com/question/2 ... 2075467194410988234
On August 23, Anthropic researcher Levent Alpöge posted a 108-page paper on X.
The title itself states the conclusion: a family of two-dimensional complex tori over the (3, 4, ∞) modular curve,
upon completion at three special points, forms a complex structure on S⁶.
The question of whether S⁶ admits a complex structure is commonly known as the Hopf problem.
While the name honors Hopf (whose paper appeared in 1948),
Kirchhoff may have been the first in the literature to explicitly pose the question:
in 1947, he constructed an almost complex structure on S⁶ but acknowledged at the end of his paper
that he did not know whether S⁶ was a complex manifold. Counting from then,
this problem has remained open for nearly eighty years.
I. The Author
Alpöge’s background: undergraduate studies at Harvard, PhD from Princeton under
Fields Medalist Bhargava, specializing in number theory and arithmetic statistics.
His notable works include resolving Hilbert's Tenth Problem for rings of integers
over arbitrary number fields, in collaboration with Bhargava, Wei Ho, and Ari Shnidman.
After completing his PhD, he served as a Junior Fellow at the Harvard Society of Fellows
for several years before leaving academia in 2025 to join Anthropic.
This summer, his name has been closely linked with Claude, and he has released four major results in rapid succession.
On July 20, he tweeted a three-dimensional counterexample to the Jacobian conjecture.
The conjecture, originally proposed by Keller in 1939, posits that any polynomial map
with a non-zero constant Jacobian determinant must be invertible.
The counterexample he provided is brief enough to be verified by hand—
a polynomial map from ℂ³ to ℂ³ with a determinant identically equal to -2,
yet mapping three distinct points to the same image. After being independently verified by numerous people within a single day, Terence Tao wrote a comprehensive analysis the following day, and others even formalized the proof in Lean. Akhil Mathew had suggested the problem, while Tao credited the construction itself to Claude (specifically, Claude Fable 5).
Two other developments occurred around August 13. Tao, three human collaborators, and Claude constructed a Hadamard matrix of order 668, thereby resolving the remaining 12 open cases for orders below 2000—a set of problems that had long appeared on the FrontierMath list; Epoch AI subsequently marked this item as "solved." That same week, Anthropic announced that an unreleased research version of Claude had raised the lower bound for the proportion of zeros on the critical line of the Riemann zeta function from 41.6% to 67.2%; the mathematicians responsible for internally verifying this result were Alpöge and Ralph Furman.
Then came the case of $S^6$. The first three results have withstood rigorous public scrutiny, though the nature of that scrutiny differed: the Jacobian counterexample and the Hadamard matrix involved objects that provided finite certificates, allowing for repeated verification within days; the zeta function result was accompanied by a Lean formalization and preliminary review by number theory experts. However, the scale of the current problem is entirely different.
II. The Problem Itself
Before a complex structure, there is the almost-complex structure: one equips the tangent bundle with an operator $J$ such that $J^2 = -1$, making the tangent space at each point locally resemble multiplication by $i$. In 1953, Borel and Serre proved that among all spheres, only $S^2$ and $S^6$ admit an almost-complex structure. $S^2$ is the Riemann sphere, which naturally possesses a complex structure. The structure on $S^6$ arises from octonion multiplication; unfortunately, in 1951, Eckmann–Frölicher and Ehresmann–Libermann independently proved that it is non-integrable. The problem thus reached a subtle impasse: $S^6$ admits an almost-complex structure, yet no one knew whether a *different*, integrable complex structure existed on it.
For nearly eighty years, the mathematical community has primarily focused on narrowing the possibilities from the opposite direction. As early as 1953, Blanchard ruled out complex structures compatible with the round metric—such a structure would force $S^6$ to be a Kähler manifold, conflicting with the fact that $b_2 = 0$; LeBrun provided a famous alternative proof of this in 1987 using twistor methods.
Huckleberry, Kebekus, and Peternell proved that a hypothetical complex $S^6$ would not be almost homogeneous, that $h^0(TX) \leq 2$, and that the Picard group is determined by $H^1(X, \mathcal{O}_X)$. Campana, Demailly, and Peternell claimed in 1998 that the algebraic dimension could be reduced to 0; however, a lemma in their argument was later found to be incorrect. A corrected proof was published in 2020, and following this line of reasoning—further refined by Lehn, Rollenske, and Schinko—the algebraic dimension of a hypothetical complex $S^6$ was shown to be at most 1. This group of researchers will appear again later in the text.
Since 2005, Etesi has repeatedly announced a successful construction—most recently in 2024—but none have been accepted.
In 2016, Atiyah announced a proof of non-existence, which was likewise not accepted.
As of January 2026, a preprint by Jun Ling claiming non-existence remains available.
Alpöge himself acknowledged in a tweet that the literature contains numerous claims pointing in both directions regarding this problem.
As for its significance:
An almost complex structure is an algebraic-topological condition at the level of the tangent bundle, whereas a complex structure is an analytic condition involving integrability; the gap between the two encompasses much of modern geometry,
and $S^6$ is the only sphere caught in this divide. If this construction holds, the result would be a compact complex 3-manifold with $b_2 = 0$, non-Kähler status, and an Euler characteristic of 2—whereas known examples of this type, such as $S^3 \times S^3$ or Hopf manifolds, all have an Euler characteristic of 0.
III. The Construction Itself
Alpöge did not write down complex coordinates directly on $S^6$. Instead, he first constructed a space known to possess a complex structure and then proved that this space is indeed $S^6$. This space is a family of two-dimensional complex tori over $\mathbb{P}^1$, with a base defined by the triangle group $\Delta(3,4,\infty)$; the space is completed at three special points using three classical methods. At the cusp, Mumford's torus degeneration is employed; the central fiber is a del Pezzo surface $dP_6$ with its hexagonal boundary's three pairs of opposite sides glued together pairwise, resulting in two triple points.
At the elliptic points of orders 3 and 4, Kodaira's logarithmic transformation is used to produce two multiple fibers. The gluing process involves three integers $(\ell_0, \ell_1, \ell_2)$ that record the twisting; the fundamental group is calculated to be $\mathbb{Z}/|p|$, where $p = 12\ell_0 - 4\ell_1 - 3\ell_2$. The paper selects $(0, 1, -1)$, yielding $p = -1$, so the fundamental group becomes trivial. The homology groups are calculated degree by degree to match those of $S^6$, and the Euler characteristic is 2. Being simply connected with the integral homology of $S^6$ implies it is a homotopy 6-sphere; since there are no exotic 6-spheres, $X$ is diffeomorphic to $S^6$.
The paper also calculates the complete set of invariants for $X$, verifying them against necessary conditions established over the past two decades:
$h^0(TX) = 1 \leq 2$, $\mathrm{Pic}(X) \cong \mathbb{C}$, $h^{0,1} = 1 = h^{0,2} + 1$, $K_X$ is non-torsion, and the algebraic dimension is 1.
The most technically demanding data in the construction is deliberately condensed into the first two pages: two $4 \times 4$ integer matrices, the transformation laws for three periodic functions, and two twist vectors. These serve as a certificate for verifiers—whether human or machine—allowing them to fully reconstruct the main calculations starting from just these two pages.
IV. Two Independent Model-Assisted Audits
The author submitted the paper for separate audits by GPT and Kimi. Although the two models were independently verified in parallel, they remained susceptible to correlated errors—and since they relied on the same proof, this process cannot be equated to independent refereeing by two experts in complex geometry. The value of this type of audit lies in re-executing all computations amenable to finite verification, specifically hunting for low-level yet fatal errors involving matrices, symbols, integer lattices, and the interfaces between sections. The methodologies aligned, and the conclusions matched: everything that could be cross-checked did indeed match.
The matching elements, listed in the order they appear in the paper, are as follows:
The matrices $T_1$ and $T_2$ indeed have orders 3 and 4, respectively, and both have a determinant of 1; $T_0 = (T_1T_2)^{-1}$ is indeed unipotent, with $N = T_0 - I$ satisfying $N^2 = 0$ and having rank 2. Every matrix entry on this page was recomputed, and all were correct.
The vectors $\varepsilon$ and $\varepsilon'$ on the dual lattice are indeed fixed by their respective dual matrices; the determinant of the induced map $B_0$ is 1, a value that accounts for the irreducibility of the central fiber and the fact that its normalization is a single copy of $dP_6$.
Setting $(\ell_0, \ell_1, \ell_2) = (0, 1, -1)$ yields $p = -1$. The paper provides a control case: negating $v_2$ changes $\pi_1$ to $\mathbb{Z}/7$, and recomputation confirms the value is indeed 7. Once these two torsion vectors are fixed, standard fiber-direction regluing can only alter $p$ by a multiple of 12; thus, $p = -1$ is not a fragile value easily flipped by coordinate choices or orientation conventions.
The transformation laws satisfied by the periodic functions $\tau$, $\mu$, and $\beta$ are well-defined only if they close under the relations $g_1^3 = g_2^4 = 1$. Re-deriving every term along the orbit using symbolic computation yields results that match the intermediate expressions printed in the paper term-for-term; the sum of the cocycles associated with $\beta$ is exactly zero. Euler characteristic of the central fiber: for $dP_6$ it is 6, for the hexagonal ring it is 6, and for the glued image it is 2; $6 - 6 + 2 = 2$, which aligns with the paper's $e(W) = 2$.
The paper calculates integral homology via two independent paths: one using the Mayer–Vietoris sequence, and the other using the Leray spectral sequence combined with the "nearby cycles" appendix. Both paths converge at $H^2 \cong H^3 \cong \mathbb{Z}/|p|$. The fact that the same integer emerges at the end of both paths demonstrates a carefully designed cross-verification.
The audit did not stop at the first two pages. Following the paper's logic, the analysis traced the chain of existence arguments for Theorem 3.4, the freeness and proper discontinuity of the quotient construction in Section 4, the global gluing in Section 6, the relative signs and van Kampen calculation for $\pi_1$ in Section 7, the Smith normal form within the Mayer–Vietoris sequence, and the local Milnor fibers and specialization maps in Appendix B. It continued all the way to the path in Section 8—spanning the Hurewicz and Whitehead theorems to the identification $\Theta_6 = 0$—and examined the interfaces between sections one by one. GPT’s overall assessment was:
No fatal errors were found; parts amenable to recalculation were reduced to standard integer matrix operations; two vastly different topological calculations yielded the same integer $p = -1$, and this value of $-1$ is precisely what causes all intermediate topological obstructions to vanish; remaining uncertainties lie not in specific calculations, but in a few advanced, standard tools that require confirmation by experts in complex geometry and sheaf theory.
Kimi’s independent review reached the same conclusion.
However, this audit did not involve re-proving every sentence of the 108-page paper, nor did it independently formalize standard advanced tools such as nearby cycles, duality on complex analytic spaces, modular forms, or line bundle descent. Thus, these results should be viewed with that context in mind.
The models checked for consistency regarding the applicability conditions of these tools, the finite calculations newly introduced in the paper, sign conventions, and the interfaces between sections. To date, no fatal errors have been found.
V. The Real Point of Conflict
The paper's main theorem directly conflicts with published literature. Corollary 2.3 of the 2020 paper by Campana, Demailly, and Peternell states clearly:
A complex three-dimensional manifold homeomorphic to $S^6$ must have an algebraic dimension of 0. However, the manifold $X$ constructed in this paper has an algebraic dimension of 1. Both cannot be correct.
The diagnosis Alpöge offers in Section 10 is far more substantial than a mere "they overlooked non-regular fibers"; it is a concrete counter-claim that can be falsified:
Proposition 2.4 in CDP20 requires the higher direct image sheaf $R^2 f_*(T_X \otimes L)$ to vanish, yet for his manifold $X$, this sheaf does not vanish for any line bundle $L$;
the obstruction is concentrated on the non-regular central fiber $W$. The paper also distinguishes between two issues: the problem regarding monodromy counting in their proof is potentially fixable for this $X$;
what cannot be salvaged is the step of pulling back from the normalization of $W$ to the original fiber. Experts now need only determine who is correct regarding this specific calculation.
Given that the original 1998 CDP paper contained an error itself, and considering that proofs in this field are notoriously difficult to verify, expert scrutiny will focus heavily on this section.
This is precisely why Alpöge has explicitly written out every matrix and coordinate chart.
VI. A Few Expectations
Finally, a brief digression. A month ago, the Jacobian counterexample consisted of a single line of polynomials—something the global community could verify by hand in a day; this time, the $S^6$ result spans hundreds of pages.
Although no verifiable Lean certificate was provided for this work in modern geometry, the verification process was deliberately designed with a structure that facilitates external checking. This shift itself
is perhaps more noteworthy than the individual result, though I believe it is still best to include verifiable certificates when releasing findings; after all, the ability to verify conclusions will become increasingly critical.
I am quite curious to see how far mathematics will advance when the next generation of models arrives. Many major unsolved problems are now seeing candidate solutions emerge on a monthly basis.
Posted on 2026-08-24 16:18 • Guangdong
|
楼主