菲尔兹奖得主王虹,也发过Neurips

marsbitPublished on 2026-07-24Last updated on 2026-07-24

Abstract

菲尔兹奖得主王虹教授(2019年时)曾以共同一作者身份在NeurIPS上发表论文,研究机器学习中的低秩矩阵近似问题。该论文针对列子集选择(CSS)算法,利用调和分析中的经典工具——Riesz–Thorin插值定理,首次给出了关于近似误差更紧的理论上界,并在p≥2时证明其结果精确到常数。这项工作展示了将纯数学工具引入计算机科学以解决理论难题的跨学科价值。尽管王虹的个人主页未直接列出该论文链接,但它在当时即因扎实的论证而获得认可。文章指出,这与NeurIPS 2026所强调的理论贡献导向相符,表明顶会不仅关注应用突破,严谨的理论推进同样重要。

嗯?新晋菲尔兹奖得主王虹,也曾跨界做AI?

NeurIPS 2026出分在即,有网友扒出,王虹教授发过一篇NeurIPS 2019

而且还不是挂名,是实打实的共同一作。

那么问题来了,一个纯数方向的顶尖数学家,为什么会在AI顶会上发论文?

通读下来,我们的结论是,这是一篇数学理论+机器学习的最佳范例。

不过有意思的是,在王虹个人主页列出的将近40多篇论文和预印本中,几乎每篇都附有完整链接。

唯独这篇,是个例外。

王虹跨界挑战AI

这篇论文研究的,是机器学习和数据分析中的一项基础任务:低秩矩阵近似

简单来说,现实中的数据通常可以整理成一个矩阵,但这些矩阵往往非常大,直接存储和处理成本极高。

低秩近似就是用一个结构更简单、秩更低的矩阵,尽可能准确地还原原始矩阵。

近来常用的近似算法是列子集选择(Column Subset Selection,CSS)。

它的思路其实非常直观。

在面对一个包含大量列的数据矩阵,不再直接求出一个全新的低秩矩阵,而是从原矩阵中挑出具有代表性的若干列,再用它们张成的空间去近似整个矩阵。

由于被CSS选出来的列直接来自原始数据,因此会比普通矩阵分解得到的抽象向量更容易解释,同时它也能降低存储和计算成本,适合处理大规模数据。

此前的研究证明,其中对于一般的低秩近似,CSS算法的近似比上界大约是O(k+1)

这里的k指的是目标矩阵的秩,k越大,理论上允许的最坏误差就越大。

王虹等人的工作,则是把这个界进一步推进:

当1≤p≤2时,近似比为(k+1)^(1/p);

当p≥2时,近似比为(k+1)^(1−1/p)。

相较此前统一的O(k+1)结果,这一界明显更紧,算法能够被严格限制,最坏结果也只会比最优解差一点点。

另外,对于p≥2的情况,论文还构造了对应的下界,证明其结果精确到常数1

换句话说,这篇论文给出了近乎封顶的理论答案。

而这篇论文中最关键、也最能体现王虹数学背景的部分,是他们使用了来自调和分析的经典工具Riesz–Thorin插值定理

通常情况下,想证明一套算法在所有p值下都成立,需要针对不同的p分别展开复杂分析。

对于某些端点情况,例如p=1、p=2和p=∞,则相对容易处理。

接着Riesz–Thorin插值定理可以在掌握这些端点结果后,把结论“插值”到中间的所有p值。

具体来说,论文先证明p=1、2、∞三个特殊情况,再通过插值理论推出整个范围内的近似界。

事实上,这套工具在调和分析和算子理论中属于经典方法,却并不是当时理论计算机科学研究者最常使用的技术。

NeurIPS当年的审稿人也注意到了这一点。

审稿人最终认可这篇论文最主要的技术创新,就是往计算机领域引入Riesz–Thorin定理,最后的Meta Review则将它评价为一篇论证相当扎实的论文。

放到今天看,这篇论文其实也提供了一个非常典型的跨学科案例,机器学习的难题或许可以在纯数学中找到突破口。

NeurIPS 2026出分在即

时间拨回到现在,NeurIPS的审稿机制正在发生一次明显调整。

NeurIPS 2026要求作者在投稿时,从五种贡献类型中选择出最符合论文定位的一类:General、Theory、Use-Inspired、Concept & Feasibility以及Negative Results。

毫无疑问,王虹这篇2019年的论文属于其中的Theory类型。

按照NeurIPS 2026最新审稿指南,理论论文首先考察的是数学严谨性和正确性。证明、引理和整体逻辑必须成立,所以理论论文并不需要因为缺少实验而受到忽视。

同时NeurIPS 2026明确指出,理论贡献可以独立成立,设计新算法的目的也不一定是击败最新应用模型或最大数据集上的SOTA。

在原创性方面,也可以从其他学科引入新的证明工具,或对既有工具进行创新性综合。

而这几乎就是王虹这篇论文的精准描述。

他们没有提出一个今天意义上的神经网络架构,也没有训练参数量巨大的模型,而是把调和分析中的插值定理引入低秩近似,解决近似算法界限问题。

放到2026年的评审框架中,它依然是一篇非常标准的NeurIPS理论论文。

它也反向说明,NeurIPS并不等于神经网络模型发布会,对已有方法给出新的理解、发现新的性质、建立更紧的理论界限,同样属于有价值的原创贡献。

至于数学和AI之间的边界,也从来没有想象中那么清晰。

菲尔兹奖得主能发NeurIPS,数学语言也能为AI找到它真正的通解。

参考链接:

[1]https://proceedings.neurips.cc/paper_files/paper/2019/file/80a8155eb153025ea1d513d0b2c4b675-Paper.pdf

[2]https://neurips.cc/Conferences/2026/ReviewerGuidelines

[3]https://sites.google.com/view/hongwang/home

[4]http://xhslink.cn/o/8oQ3gm7qaxu

本文来自微信公众号“量子位”,作者:鹭羽

Trending Cryptos

Related Questions

Q菲尔兹奖得主王虹的NeurIPS 2019论文属于什么研究类型?

A属于理论(Theory)类型。这篇论文没有进行大规模实验或提出新模型,而是运用调和分析中的Riesz–Thorin插值定理来解决低秩矩阵近似算法的理论界限问题,是数学理论与机器学习结合的典范。

Q王虹等人的NeurIPS 2019论文主要解决了什么核心问题?

A论文研究了低秩矩阵近似中的一个基础任务,特别是针对“列子集选择”(CSS)算法,他们通过引入新的数学工具,为算法的近似比(近似误差的上界)建立了比之前更紧的理论界限,并证明对于p≥2的情况,其结果在常数因子内是最优的。

Q王虹的论文中,最关键的数学工具是什么?它在证明中起到了什么作用?

A最关键的数学工具是调和分析中的Riesz–Thorin插值定理。该定理允许研究者在证明了算法在几个特殊参数值(如p=1, 2, ∞)下的性能后,通过“插值”的方式,将结论推广到整个参数范围(所有p值),从而极大地简化了证明过程并得到了统一的紧致界。

Q根据NeurIPS 2026的新审稿指南,对理论(Theory)类论文的首要评审标准是什么?

A首要评审标准是数学的严谨性和正确性。证明、引理和整体逻辑链必须成立。理论贡献可以独立存在,其价值在于提供新的深刻理解或更优的理论界限,而不需要像应用型论文那样追求在大型数据集上超越现有最佳模型。

Q文章中提到,王虹的个人主页如何处理这篇NeurIPS 2019论文?这反映了什么?

A在王虹个人主页列出的近40篇论文中,几乎每篇都附有完整链接,唯独这篇NeurIPS 2019论文是个例外,没有提供链接。这可能反映出,作为一个纯数学方向的顶尖学者,她对这篇跨界到人工智能/机器学习领域的早期工作持有一种特殊或相对低调的态度,或许与其主要研究方向有所区分。

Related Reads

The Verdict in Choi Tae-won's Divorce Case: Revealing the Inheritance Undercurrent Behind SK Hynix's Trillion-Won Empire

SK Group Chairman Chey Tae-won's high-profile divorce case, involving a record 1.38 trillion won settlement, has drawn attention to the succession plans for Korea's second-largest conglomerate, especially its crown jewel, SK hynix. Unlike traditional chaebol scripts centered on the eldest son, Chey's three children from his marriage to former President Roh Tae-woo's daughter, Roh Soh-yeong, are carving distinct, non-traditional paths. Eldest daughter Chey Yun-jung (b. 1989) is seen as the most evident successor. With a scientific and consulting background, she holds executive roles at SK bioscience and SK Inc.'s growth support department, focusing on future strategy and biopharma. Her marriage is to an AI infrastructure entrepreneur, not a traditional business alliance. Second daughter Chey Min-jung (b. 1991) took a unique route, voluntarily serving as a South Korean naval officer, including an anti-piracy deployment. She later worked on policy and strategy for SK hynix in Washington D.C. before co-founding an AI-driven healthcare startup. She married a former U.S. Marine Corps officer, connecting her to U.S. defense and policy circles—networks crucial for a global semiconductor giant. The only son, Chey In-geun (b. 1995), who studied physics like his father, worked briefly at SK E&S before joining McKinsey. Despite fitting the traditional "heir" profile as the eldest son, he remains silent and holds no public position or shares in SK, suggesting the old succession playbook is obsolete. As SK hynix's valuation soars, becoming a geopolitical asset in the AI era, the heirs' legitimacy is no longer automatic. They must prove themselves in fields like AI biotech, global policy, and strategic consulting. Their marriages also reflect new elite networks in tech and defense, not old political alliances. Their inheritance is the complex challenge of navigating a globalized, tech-driven world, not just a corporate throne.

marsbit2 days ago 09:06

The Verdict in Choi Tae-won's Divorce Case: Revealing the Inheritance Undercurrent Behind SK Hynix's Trillion-Won Empire

marsbit2 days ago 09:06

From OpenSea to OpenRouter: Is Alex Atallah Repeating His 'Exit at the Peak' Playbook?

From OpenSea to OpenRouter: Is Alex Atallah Repeating His "Exit at the Peak" Playbook? According to the Wall Street Journal, payments giant Stripe is in talks to acquire the AI model aggregation platform OpenRouter in a potential deal valuing the company near $100 billion. This would mark founder Alex Atallah's second creation of a company reaching a $100 billion valuation, following his co-founding of NFT marketplace OpenSea. OpenRouter, founded just over three years ago, has grown rapidly by acting as a unified gateway for developers to access over 400 AI models. It currently has about 10 million users and processes over 200 trillion tokens monthly. While the platform's annualized revenue is around $50 million, its valuation has skyrocketed from $1.3 billion in March 2026. The potential acquisition by Stripe, a company OpenRouter's founder once likened it to, represents a major expansion into AI infrastructure for the payments leader. This move echoes Atallah's previous timing with OpenSea, where he departed before the NFT market's significant downturn. For OpenRouter, selling now may be strategic. Despite its scale, its business model—charging a 5-5.5% fee on AI inference calls—faces pressure from competition, open-source models, and potential price wars among model providers, limiting its profitability narrative for an IPO. A key asset for potential acquirers like Stripe is OpenRouter's vast repository of real-world AI usage data, which offers unique insights into model performance and developer preferences that are difficult to replicate. Whether this potential deal signifies a new valuation benchmark for AI infrastructure or another market peak signal remains to be seen.

链捕手2 days ago 08:42

From OpenSea to OpenRouter: Is Alex Atallah Repeating His 'Exit at the Peak' Playbook?

链捕手2 days ago 08:42

Trading

Spot

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of AI (AI) are presented below.

活动图片