Leo's log

数学

八十七年与五分钟:雅可比猜想之死

Eighty-Seven Years and Five Minutes: The Death of the Jacobian Conjecture

2026 年 7 月 19 日是世界杯决赛日。第二天,一位数学家在社交媒体上发了一条随手写就的帖子,开头是”hello there”,结尾是一句拼写随意的”thanx”。帖子中间夹着三行多项式。

就是这三行多项式,杀死了一个悬置八十七年的数学猜想。

没有论文,没有发布会,没有几百页的证明。任何一个会用数学软件的人,花五分钟就能亲手确认它是对的。一个让几代数学家耗尽心力的问题,以一条一周后就会沉底的推文的形式,画上了句号。

我想讲讲这个故事。它不需要你懂数学——恰恰相反,这个故事最好的部分,都和数学无关。

一张完美的地图

先说这个猜想在问什么。

想象一张地图。一张好地图应该做到:地图上任何一小块区域,都清晰、比例准确、不重叠、不撕裂。你盯着地图上的任何一个局部看,都挑不出毛病。

现在问一个问题:如果一张地图的每一个局部都完美无缺,那么整张地图是不是也一定可靠——也就是说,地图上的每个点,是不是都只对应现实中的一个地方?

直觉说是。但想想揉面团。揉面的时候,面团的体积始终不变,任何一小块面都没有被压扁或撕开——每个”局部”都规规矩矩。可是折叠几次之后,原本相距很远的两粒面粉,会紧紧贴在一起。局部处处完好,整体却发生了折叠。

雅可比猜想问的就是这件事,只不过它把范围限定在一类特别”规矩”的变换上——用多项式(就是中学里那种由加减乘构成的式子)定义的变换。猜想断言:对这类变换来说,只要每个局部都完美(用行话说,“雅可比行列式是非零常数”),整体就绝不会折叠,一定可以完整还原回去。

换句话说,数学家们赌的是:多项式这种东西足够刚硬,硬到揉不出褶子。

这个赌注押了八十七年。2026 年 7 月,赌局揭晓:会折。那三行多项式定义的变换,每个局部都完美,却把三个不同的点压到了同一个位置上——拿着这张”完美地图”的人站在那个点上,永远无法知道自己来自哪里。

一个生下来就是错的猜想

这个猜想的名字来自德国数学家雅可比(Carl Jacobi)。他 1851 年就去世了,从没听说过这个猜想——只是猜想里用到的那个衡量”局部是否完美”的工具以他命名,猜想就这么姓了雅可比。这是数学界的常规操作:问题跟谁姓,往往和谁提出它没什么关系。

真正的起源,直到几个月前才被翻出来。2026 年初,有研究者在文献数据库里检索一个古老的德文数学词,意外挖出一篇 1884 年的论文:布拉格数学家路德维希·克劳斯(Ludwig Kraus)在里面把这个猜想当作定理”证明”了。他的证明最后一步是错的。此后一百四十年,没人记得他。

也就是说,这个猜想来到世上的第一天,就是以一个错误证明的形式出现的。这像一个预言。

1939 年,德国数学家凯勒(Ott-Heinrich Keller)正式提出了这个问题,猜想从此以他的工作为纪年起点。此后的几十年里,它慢慢积累起数学界一个独一无二的名声:错误证明特别多。至少有五个错误证明正式发表在学术期刊上,没发表就被发现有错的更是不计其数。2004 年还有一次著名的乌龙:一个新证明的消息被知名数学家群发邮件宣布,学界短暂沸腾,随后证明被发现有洞。一位专门研究这个问题的数学家甚至在专著里整理过历年错误证明的清单,像法医整理档案。

为什么这么多聪明人前赴后继地栽在它身上?因为它看起来太简单了。陈述它只需要本科一年级的微积分,任何数学系学生都能在十分钟内听懂题目。一道人人看得懂的题,悬赏八十七年无人能解——这种问题对数学家有致命的吸引力,也就成了最高效的职业生涯粉碎机。

一个被它耽误的人

说到被这个猜想改变命运的人,中文世界的读者都认识其中一位:张益唐。

1985 年,张益唐从北大来到美国普渡大学读博士,一读六年半,博士论文的题目就是《雅可比猜想与域扩张的次数》。论文做完了,猜想当然没有解决——但更糟的是,他和导师不欢而散,拿不到推荐信,毕业后多年找不到教职。此后的故事很多人都知道了:他在朋友的快餐店帮忙记账,在汽车旅馆打工,在学术界的视野之外沉寂了二十多年。直到 2013 年,五十八岁的他证明了素数间隔的有界性,一夜之间震动整个数学界。

张益唐后来的成就和雅可比猜想没有关系。但他青年时代最好的六年半,是交给这个猜想的。他不是唯一一个——他只是那些被这个问题消耗了岁月的人里,唯一在别处翻了盘、因此被我们记住名字的人。

回头看,这个猜想八十七年的历史里堆满了这样的账目:错误的证明、耗尽的年头、不欢而散的师生。而它的死法,是三行多项式加一句”thanx”。

死于一个下午

需要说清楚的是,八十七年里数学家并非一无所获。他们证明了:如果变换的式子足够简单(次数很低),猜想成立;在二维的情形下,直到相当复杂的程度都找不到反例;他们还发现这个问题和量子力学的代数结构存在深层联系,把它列入了二十一世纪最重要的数学问题清单。所有这些扎实的工作拼出了一幅心理地图:反例如果存在,一定藏在极其复杂、极其遥远的地方。

问题恰恰出在这幅心理地图上。二维情形的理论结果证明了”那里不可能有简单的反例”,于是所有人默认更高维也一样,搜索的探照灯从来没有认真照向三维里那片其实很浅的水域。这次找到的反例小得令人难堪:三个变量,次数不超过七,系数是一位数和个位分数。它不在深渊里,它在门口。

这就是路灯下找钥匙的老笑话,只不过这次钥匙真的掉在黑处,而且一伸手就能够到。八十七年”没人找到反例”这件事,原来证明的不是反例不存在,而是没人往正确的方向看过一眼。

然后是这次事件本身。据宣布者——一位在 AI 公司 Anthropic 工作的数学家——所说,这个反例是 AI 模型在他的引导下生成的,起因是一位同行朋友问起这个问题。反例随后又通过了另一家公司的自动证明系统的形式化检验。正式的同行评审还没有完成,但这类结果和普通的数学论文有本质区别:它不是几百页需要逐行检查的论证,而是一个五分钟的计算,全世界任何人都可以、也已经在自己的电脑上重复验证。找到它花了八十七年,确认它只要一个下午。

有个细节很能说明这个时刻的荒诞感。有人把这个结果加进了维基百科词条,引用来源一栏填的是那条推文——随后编辑被撤销,因为按照规则,推文不算可靠来源。规则没有错。只是规则写下的时候,没人想过一个八十七年的问题会以这种方式终结。

尾声:凯勒的两个猜想

最后说回凯勒。

他 1906 年生于法兰克福,1990 年去世。他一生提出过两个著名的猜想。一个是 1930 年的立方体铺砌猜想,源自他的博士论文,问的是用完全相同的方块铺满空间时是否总有两块完全贴合。另一个就是 1939 年的雅可比猜想。

立方体铺砌猜想的结局是:1992 年起,数学家借助图论转化和计算机搜索,逐步确定它在七维及以下成立、八维及以上不成立——最后一块拼图是靠自动求解器完成的。而雅可比猜想的结局,就是这个夏天:一个 AI 模型写出的三行多项式。

一个人,两个猜想,全部活得比他长,全部由他从未见过的机器给出了答案。

我不知道该把这叫作什么。它不算悲伤——问题被解决从来不是坏事,何况数学家八十七年积累的理解并没有作废,只是从此指向新的方向:为什么二维仍然可能是对的?加上什么条件能把定理救回来?但它也不只是新闻。找答案难、验答案易,这曾经是数学最深的鸿沟之一,无数人的一生就填在这道沟里。现在,机器开始承担”找”的那一半,而”验”的那一半便宜得像自来水。张益唐们用六年半没有换到的东西,一个下午就摆在了所有人面前。

八十七年与五分钟。这两个数字之间的落差,就是我们此刻正在穿过的那道门。


注:文中反例于 2026 年 7 月 20 日公布,已通过符号计算和自动证明系统的独立验证,正式同行评审仍在进行中。二维情形的雅可比猜想不受此反例影响,至今仍是悬案。

主要资料来源

  • Wikipedia: Jacobian conjecture / Ott-Heinrich Keller / Keller’s conjecture / Yitang Zhang
  • L. O. Rodríguez Díaz, On the origin of the Jacobian conjecture(关于 Kraus 1884 年论文的历史研究), arXiv:2512.23614
  • Wolfram MathWorld: Jacobian Conjecture(错误证明史,含 2004 年事件)
  • Alec Wilkinson, The Pursuit of Beauty, The New Yorker, 2015(张益唐特写)
  • Yitang Zhang, The Jacobian Conjecture and the Degree of Field Extension, Purdue University PhD thesis, 1991
  • Hacker News 讨论串及事件报道(OfficeChai, ForkLog 等),2026 年 7 月 20–21 日

Mathematics

Eighty-Seven Years and Five Minutes: The Death of the Jacobian Conjecture

八十七年与五分钟:雅可比猜想之死

July 19, 2026 was the day of the World Cup final. The following day, a mathematician posted something on social media. It opened with “hello there,” closed with a casually misspelled “thanx,” and in between sat three lines of polynomials.

Those three lines killed a mathematical conjecture that had stood open for eighty-seven years.

There was no paper. No press conference. No proof running to hundreds of pages. Anyone with math software on their laptop could confirm the result in five minutes. A problem that had consumed generations of mathematicians ended its life as a tweet — a format that sinks from view within a week.

I want to tell this story. You don’t need any mathematics to follow it. In fact, the best parts of this story have nothing to do with mathematics at all.

A Perfect Map

First, what the conjecture actually asks.

Picture a map. A good map should have this property: any small patch of it is clear, accurately scaled, with nothing overlapping and nothing torn. Zoom in anywhere you like, and you can’t find a flaw.

Now ask a question: if every local patch of a map is perfect, does that guarantee the map as a whole is trustworthy — that every point on the map corresponds to exactly one place in the real world?

Intuition says yes. But think about kneading dough. As you knead, the dough’s volume never changes, and no small piece of it ever gets crushed or torn — every “local patch” behaves impeccably. Yet after a few folds, two specks of flour that started far apart end up pressed together. Locally flawless everywhere; globally, folded.

That is what the Jacobian conjecture asks about — restricted to an especially well-behaved class of transformations, the ones defined by polynomials (the kind of expressions you met in school, built from adding and multiplying). The conjecture asserted: for these transformations, local perfection everywhere (in the jargon, “the Jacobian determinant is a nonzero constant”) guarantees the whole thing never folds. You can always recover exactly where you came from.

In other words, mathematicians were betting that polynomials are rigid — too stiff to knead a crease into.

The bet ran for eighty-seven years. In July 2026, the table showed its cards: they fold. The transformation defined by those three lines of polynomials is locally perfect at every single point, yet it presses three different points onto the same spot. Anyone standing at that spot, holding this “perfect map,” can never know where they came from.

A Conjecture Born Wrong

The conjecture is named for the German mathematician Carl Jacobi, who died in 1851 and never heard of it. The tool it uses to measure “local perfection” bears his name, and the conjecture inherited it — standard practice in mathematics, where the person a problem is named after often has little to do with the person who posed it.

The true origin surfaced only months ago. In early 2026, researchers searching a literature database for an archaic German mathematical term dug up a paper from 1884, in which the Prague mathematician Ludwig Kraus had “proved” the conjecture as a theorem. The final step of his proof was wrong. For the next hundred and forty years, no one remembered him.

Which is to say: this conjecture entered the world in the form of a flawed proof. It reads like a prophecy.

In 1939 the German mathematician Ott-Heinrich Keller posed the problem formally, and the conjecture has dated its life from his paper ever since. Over the following decades it accumulated a reputation unique in mathematics: it attracted wrong proofs. At least five incorrect proofs made it into peer-reviewed journals; the ones caught before publication are beyond counting. In 2004 came a famous fiasco — news of a new proof went out in a mass email from a prominent mathematician, the community briefly lit up, and then a hole was found. One specialist devoted part of a monograph to cataloguing the failed proofs over the years, the way a coroner keeps files.

Why did so many brilliant people walk into the same wall? Because the problem looks so simple. Stating it requires nothing beyond first-year calculus; any undergraduate can understand the question in ten minutes. A problem anyone can read, standing unclaimed for eighty-seven years — that combination is irresistible to mathematicians, which made it one of the most efficient career-shredders in the field.

A Man It Cost

Among the people whose lives this conjecture bent, one is famous far beyond mathematics: Yitang Zhang.

In 1985, Zhang came from Peking University to Purdue for his PhD. He spent six and a half years there, and his doctoral thesis was titled The Jacobian Conjecture and the Degree of Field Extension. The thesis was finished; the conjecture, of course, was not. Worse, he and his advisor parted on bad terms, he had no letters of recommendation, and for years afterward he could not find an academic position. The rest of the story many people know: he kept the books at a friend’s fast-food restaurant, worked at a motel, and lived for over two decades outside the sightlines of academia. Then in 2013, at fifty-eight, he proved that the gaps between primes are bounded — and overnight shook the entire mathematical world.

Zhang’s eventual triumph had nothing to do with the Jacobian conjecture. But the best six and a half years of his youth were paid to it. He was not alone — he is simply the only one among all the people this problem consumed who won so spectacularly elsewhere that we remember his name.

Looking back, the ledger of those eighty-seven years is full of entries like this: wrong proofs, spent years, students and advisors who stopped speaking. And the manner of the conjecture’s death: three lines of polynomials and a “thanx.”

Death in an Afternoon

To be fair, mathematicians did not come away from those eighty-seven years empty-handed. They proved the conjecture holds when the polynomials are simple enough (low degree). They verified that in two dimensions, no counterexample exists up to a remarkable level of complexity. They discovered that the problem is deeply linked to the algebraic structure of quantum mechanics, and it earned a place on a famous list of the most important mathematical problems for the twenty-first century. All of this solid work assembled a mental map: if a counterexample existed at all, it had to be hiding somewhere immensely complicated and far away.

The trouble was precisely that mental map. Theory had ruled out simple counterexamples in two dimensions, so everyone quietly assumed the same held in higher dimensions — and the searchlight was never seriously aimed at the shallow waters of dimension three. The counterexample just found is embarrassingly small: three variables, degree no more than seven, coefficients you can write with single digits and simple fractions. It was not lurking in the abyss. It was standing by the door.

This is the old joke about searching for your keys under the streetlight — except this time the keys really were lying in the dark, and within arm’s reach. Eighty-seven years of “no one has found a counterexample” turned out to prove not that no counterexample existed, but that no one had ever looked in the right direction.

Then there is the event itself. According to the mathematician who announced it — a researcher at the AI company Anthropic — the counterexample was produced by an AI model working under his direction, prompted by a question from a fellow mathematician. The example was subsequently checked by another company’s automated proof system. Formal peer review is still underway, but a result of this kind differs fundamentally from an ordinary mathematics paper: it is not hundreds of pages of argument to be checked line by line, but a five-minute computation that anyone in the world can — and many already have — reproduced on their own machine. Finding it took eighty-seven years. Confirming it took an afternoon.

One small detail captures the absurdity of the moment. Someone added the result to the Wikipedia article, citing the tweet as a source — and the edit was promptly reverted, because under the rules, a tweet is not a reliable source. The rules are not wrong. It’s just that when the rules were written, nobody imagined an eighty-seven-year-old problem would end this way.

Coda: Keller’s Two Conjectures

Which brings us back to Keller.

He was born in Frankfurt in 1906 and died in 1990. In his lifetime he posed two famous conjectures. One, from 1930, grew out of his doctoral thesis on tiling space with identical cubes: it asked whether such a tiling must always contain two cubes that meet face-to-face. The other, from 1939, was the Jacobian conjecture.

The cube-tiling conjecture met its end beginning in 1992, when mathematicians — using a graph-theoretic reformulation and computer search — established, step by step, that it is true in seven dimensions and below and false in every dimension above; the final piece was closed by an automated solver. And the Jacobian conjecture met its end this summer, in three lines of polynomials written by an AI model.

One man, two conjectures, both of which outlived him — and both answered by machines he never saw.

I don’t quite know what to call this. It isn’t sad — a problem getting solved is never a bad thing, and the understanding mathematicians built over eighty-seven years is not wasted; it now points somewhere new: Why might two dimensions still be true? What extra conditions would rescue a usable theorem? But it isn’t merely news, either. The gap between finding an answer and checking one was among the deepest chasms in mathematics, and entire lifetimes were spent filling it. Now machines are taking over the finding half, while the checking half has become as cheap as tap water. What Yitang Zhang could not buy with six and a half years was laid out in front of everyone in a single afternoon.

Eighty-seven years, and five minutes. The distance between those two numbers is the doorway we are all walking through right now.


Note: The counterexample was announced on July 20, 2026, and has been independently verified by symbolic computation and by an automated proof system; formal peer review is still in progress. The two-dimensional case of the Jacobian conjecture is untouched by this counterexample and remains open.

Principal Sources

  • Wikipedia: Jacobian conjecture / Ott-Heinrich Keller / Keller’s conjecture / Yitang Zhang
  • L. O. Rodríguez Díaz, On the Origin of the Jacobian Conjecture (historical study of Kraus’s 1884 paper), arXiv:2512.23614
  • Wolfram MathWorld: Jacobian Conjecture (on the history of erroneous proofs, including the 2004 episode)
  • Alec Wilkinson, The Pursuit of Beauty, The New Yorker, 2015 (profile of Yitang Zhang)
  • Yitang Zhang, The Jacobian Conjecture and the Degree of Field Extension, PhD thesis, Purdue University, 1991
  • Hacker News discussion thread and event coverage (OfficeChai, ForkLog, et al.), July 20–21, 2026