塔纳卡重构

塔纳卡重构
Tannakian Reconstruction

原始链接: https://bartoszmilewski.com/2026/07/14/tannakian-reconstruction/

$[C, Set]$

塔纳卡重构（Tannakian reconstruction）是一个从所有函子提供的集体信息中恢复范畴结构的数学过程。用一个比喻来说，如果单个函子如同场景的快照，那么单张照片无法揭示整体结构。然而，通过“末端”（end）——一种范畴构造——将所有可能的“照片”（函子）叠加起来，我们就可以重构出原始的态射。该过程依赖于“纤维函子”，它们将对象映射到集合，将态射映射到函数。当我们考虑这些函子之间的所有自然变换集合时，本质上是在同时从每一个可能的视角观察该范畴。由于函子保留了范畴结构（将态射映射为函数），这些变换的总和包含了识别原始态射所需的充分数据。在技术层面，这是通过米田引理（Yoneda lemma）和“米田约化”（Yoneda reduction）来实现的，它们允许我们通过对所有函子进行积分来恢复源范畴的同态集（hom-sets）。最终，塔纳卡重构证明了一个范畴可以通过其表示来完全理解；正如幺半群可以通过其等变映射的总和来恢复一样，复杂的范畴也由它们映射到集合范畴的方式所定义。

Hacker News 最新 | 过往 | 评论 | 提问 | 展示 | 招聘 | 提交登录 Tannakian 重构 (bartoszmilewski.com) 6 分，由 ibobev 发布于 2 小时前 | 隐藏 | 过往 | 收藏 | 讨论 | 帮助考虑申请 YC 2026 年秋季班！申请截止日期为 7 月 27 日。指南 | 常见问题 | 列表 | API | 安全 | 法律 | 申请 YC | 联系搜索：

原文

Two friends, Alice and Bob, live in the same city, but on the opposite sides of a wide river. Every night, Bob looks at the lights on the other side and tries to guess, which one belongs to Alice. They come up with a clever arrangement: Alice will turn on her lights for 10 minutes every night at 10 p.m. Every night Bob will take a long-exposure photo at the pre-arranged time. At the end of the year, Bob will superimpose all the photos, and hopefully the only bright spot will be Alice’s window. This is Tannakian reconstruction in a nutshell.

A functor produces a picture of one category inside another. It’s a potentially lossy encoding, but it always preserves the structure of the source. If there is a connection (morphism) between two objects in the source category, there will always be a connection between their images in the target category.

In general it’s impossible to recover the structure of the source category by looking at only one such picture. But if the target category has enough resolution, and we superimpose all available pictures, we can recover the morphisms of the source category.

Fiber functors

A category with just the right resolution is the category of sets. Therefore we’ll be looking at functors in $[C, Set]$ (for historical reasons, these are called co-presheaves). Such a functor maps objects to sets, and morphisms to functions.

When dealing with functors, we usually imagine varying objects and morphisms while keeping the functor constant. Here, we are interested in using the totality of all functors while keeping the objects constant. To every object $c \colon C$ we will associate a mapping from functors to sets, simply by applying each functor to this object $c$ :

$\Phi_c \colon [C, Set] \to Set$

$\Phi_c \hat F = \hat F c$

This mapping is functorial. Indeed, a natural transformation between $\hat F$ and $\hat G$ is a family of functions $\alpha_a \colon \hat F a \to \hat G a$ . We define the action of $\Phi_c$ on a natural transformation by taking its component $\alpha_c$ .

$\Phi_a$ is called a fiber functor. You may think of it as probing an object and, through morphisms, its immediate neighborhood.

Tannakian reconstruction

To probe a hom-set $C(a, b)$ we’ll be looking at the set of functions $\hat F a \to \hat F b$ under all possible functors $\hat F \colon C \to Set$ .

This set happens to be the set of natural transformations between two fiber functors $\Phi_a$ and $\Phi_b$ , a hom-set in the functor category:

$[[C, Set], Set] (\Phi_a, \Phi_b)$

A set of natural transformations can be written as an end:

$\int_{\hat F \colon [C, Set]} Set (\Phi_a \hat F, \Phi_b \hat F) = \int_{\hat F \colon [C, Set]} Set (\hat F a, \hat F b)$

An end is like a gigantic product. In our analogy, it correspond to the superposition of all photos. Like with any product, if any of its components is empty, the whole end is empty. Since the end goes over every possible functor, what prevents us from hand-picking a functor that’s non-empty at $a$ and empty at $b$ ? Such a single bad apple would spoil the whole batch (there is no function from a non-empty set to an empty set).

What makes this end non-trivial is functoriality. Any time there is a morphism $f \colon a \to b$ , we automatically have a function $\hat F a \to \hat F b$ , for any functor $\hat F$ . In fact, because of the Yoneda embedding being full and faithful, we have as many such functions as there are morphisms in $C(a, b)$ . We have the isomorphism:

$\int_{\hat F} Set (\hat F a, \hat F b) \cong C(a, b)$

By superimposing the images of all functors, we recover the original hom-set. This is the categorical version of Tannakian reconstruction.

We’ll use the double-Yoneda trick to prove it. First, we use the Yoneda lemma:

$[C, Set] (C(a, -), \hat F) \cong \hat F a$

to explode the functors under the end:

$\int_{\hat F} Set ([C, Set] (C(a,-), \hat F), [C, Set] (C(b, -), \hat F)$

We can now apply the Yoneda reduction to “integrate” over $\hat F$ . The result is:

$[C, Set] (C(b, -), C(a, -))$

which, again by Yoneda, is equivalent to:

$C(a, b)$

As an example, let’s apply Tannakian reconstruction to a simple one-object category. Such a category has a single hom set, which forms a monoid under composition. A set-valued functor maps the single object of $C$ to a set — a representation of this monoid. Natural transformations between such functors are called equivariant maps. Tannakian reconstruction lets us recover the monoid from the totality of its representations. Notice that naturality/equivariance is baked into the definition of an end.