Professional Documents
Culture Documents
本科生学术杂志
2021
第 15 期
荷 思
2021
第 15 期
主办 《荷思》编辑部
主编 卜辰璟
编委 (按姓氏拼音排序)
卜辰璟 陈起渊 段哲凡 贺宇昕 黄虞来
孔繁浩 秦珺辉 尚鉴桥 肖咏涵 谢雨潇
张锐翀 张睿桐
排印 陈起渊
目录
专题介绍
研究讨论
数学小品
221 变号矩阵与可积系统 徐凯
𝑝‐adic Cohomology Theories
Chen Qiyuan1
ABSTRACT
In this article, we define two 𝑝-adic cohomology theories: crystalline co-
homology and rigid cohomology as such kinds of de Rham cohomology and
prove their desired properties as Weil cohomology theories, such as finiteness,
Künneth formula and Poincaré duality. We shall prove the Katz conjecture as
an application of 𝑝-adic cohomology theories.
Contents
1 Introduction 1
2 Crystalline cohomology 3
Divided powers and crystalline topoi . . . . . . . . . . . . . . . . . . . 3
Calculus with divided powers . . . . . . . . . . . . . . . . . . . . . . . 6
Verification of the desired properties . . . . . . . . . . . . . . . . . . . 9
Frobenius action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Rigid cohomology 16
Basic definitions and comparison results . . . . . . . . . . . . . . . . . 16
Verification of the desired properties . . . . . . . . . . . . . . . . . . . 21
1 Introduction
Let us fix the notations: 𝑝 a prime, ℓ a prime different from 𝑝, 𝑘 a field of charac-
teristic 𝑝 and 𝐾 a field of characteristic 0.
The motivation of a “good” cohomology theory (more precisely, Weil coho-
mology theory, [Stacks]) for algebraic varieties is from Weil conjecture, which is
about the zeta function of algebraic variety 𝑋 of characteristic 𝑝,
∞
𝑡𝑛
𝜁(𝑋, 𝑡) = ∑ |𝑋(𝔽𝑞𝑛 )| .
𝑛=1 𝑛
1
陈起渊,清华大学数学系数 80 班.
2 𝑝‐adic Cohomology Theories
• (Purity) The inverse of roots of 𝑃𝑖 (𝑡) are algebraic integers with absolute
𝑘
value 𝑞 2 , where 𝑘 is an integer between 0 and 2𝑖. When 𝑋 is proper and
smooth, 𝑘 = 𝑖. (It is an analog of purity of mixed Hodge structures.)
To prove the theorem one needs a Weil cohomology theory, that is, contravari-
ant functors 𝐻 • (𝑋), 𝐻𝑐• (𝑋), 𝐻𝑍• (𝑋) (the usual cohomology, cohomology with
compact support and cohomology supported on a closed set) from varieties over a
field 𝑘 of characteristic 𝑝 to vector spaces over some characteristic 0 field, satisfy-
ing:
• (Finiteness) 𝐻 • (𝑋), 𝐻𝑍• (𝑋), 𝐻𝑐• (𝑋) are finite dimensional with 𝐻 𝑖 (𝑋) = 0
for 𝑖 < 0 or 𝑖 > 2𝑑.
• (Künneth formula) 𝐻 • (𝑋 × 𝑌 ) = 𝐻 • (𝑋) ⊗ 𝐻 • (𝑌 ), similarly for 𝐻𝑐• and
𝐻𝑍• .
• (Poincaré duality) For smooth 𝑋, there is a trace map 𝐻 2𝑑 (𝑋) → 𝐾 satisfy-
ing 𝐻 • (𝑋) ⊗ 𝐻𝑐2𝑑−• (𝑋) → 𝐻 2𝑑 (𝑋) → 𝐾 being a perfect pairing.
• (Cycle class) There is a natural tranformation of functors 𝐴• (𝑋) → 𝐻 2• (𝑋).
One deduces then the Lefschetz fixed point theorem and apply it to the Frobenius
𝐹 𝑛 ∶ 𝑋 → 𝑋 to gain information about the rational points.
The “Riemann hypothesis” part gives information about the archimedean ab-
solute value of the roots of 𝑃𝑘 (𝑡) and then the number of rational points. It is then
natural to ask about 𝑝-adic absolute value of those roots (|𝛼|ℓ = 1). However, 𝑝
-adic étale cohomology behaves badly for characteristic 𝑝 varieties. Thus, alterna-
tive cohomology theories are necessary.
In characteristic zero, the de Rham cohomology theory 𝐻 𝑖 (𝑋) = ℍ𝑖 (Ω•𝑋∕𝐾 )
is a good cohomology theory [Stacks]. Therefore, a rough idea is to lift 𝑋∕𝑘 to
something over 𝐾 = Frac(𝑊 (𝑘)), and working out the de Rham cohomology. Two
ways to realize the idea are the crystalline cohomology and the rigid cohomology.
Another benefit of 𝑝-adic cohomology theories is that the Frobenius action on
étale cohomology can not be calculated down simply. However, in 𝑝 -adic coho-
mology theory, one can represent it as a action of a de Rham complex. There is
then a convenient algorithm of counting rational points ([Ked01]).
2 Crystalline cohomology 3
2 Crystalline cohomology
Divided powers and crystalline topoi
A few basic facts about divided power rings are enumerated here for further appli-
cations. For details, see [Stacks], [BO78], [Ber74].
Definition 2.1. A divided power ring is a triple (𝐴, 𝐼, 𝛾), where 𝐴 is a ring, 𝐼 an
ideal of 𝐴 and 𝛾𝑛 , 𝑛 ∈ ℕ a collection of maps 𝐼 → 𝐴 such that those desired
𝑛
properties of “ 𝑥𝑛! ” hold ( 𝑥, 𝑦 ∈ 𝐼, 𝑎 ∈ 𝐴 in the following):
• 𝛾0 (𝐼) = 1, 𝛾1 (𝑥) = 𝑥, 𝛾𝑛 (𝐼) ⊂ 𝐼, ∀𝑛 > 0.
(𝑛+𝑚)!
• 𝛾𝑛 (𝑥)𝛾𝑚 (𝑥) = 𝛾 (𝑥).
𝑛!𝑚! 𝑛+𝑚
A morphism between two divided power rings, (𝐴, 𝐼, 𝛾), (𝐵, 𝐽 , 𝛿), is a ring homo-
morphism 𝑓 ∶ (𝐴, 𝐼) → (𝐵, 𝐽 ) such that 𝑓 ∘ 𝛾𝑛 = 𝛿𝑛 ∘ 𝑓 . Let (𝐴, 𝐼, 𝛾) be a divided
power ring, a divided power algebra over 𝐴 is a divided power ring (𝐵, 𝐽 , 𝛿) with
a morphism (𝐴, 𝐼, 𝛾) → (𝐵, 𝐽 , 𝛿).
All divided power rings form a category with initial object (ℤ, 0, 0).
Definition 2.2. Let (𝐴, 𝐼, 𝛾) be a divided power ring, 𝐵 a ring and 𝐽 an ideal of 𝐵.
The divided power envelope of 𝐵, (𝐷𝐵,𝛾 (𝐽 ), 𝐽 ,̄ 𝛾) is defined by the representable
functor:
(𝐶, 𝐾, 𝛿) ↦ 𝐻𝑜𝑚𝐴 ((𝐵, 𝐽 ), (𝐶, 𝐾)).
The crystalline cohomology is defined via the crystalline sites and topoi. In
the above discussion, we may replace “rings” by “ringed topoi” to define similar
notions for divided power schemes. For example:
Like étale cohomology, crystalline cohomology works well only in “finite” case
(or “profinite” case). That is, we will assume 𝑚𝒪𝑆 = 0 for some integer 𝑚 without
mentioned.
𝑈 𝑇
𝑋 𝑆,
with 𝑈 → 𝑇 a divided power thickening and the divided power structures of 𝑇 and
𝑆 are compatible and morphisms defined by morphisms of divided power thick-
enings with compatibility conditions. A collection of morphisms {(𝑈𝑖 → 𝑇𝑖 ) →
(𝑈 → 𝑇 )} is a covering if {𝑇𝑖 } forms a open covering of 𝑇 and for each 𝑖, the
commutative diagram below is cartesian.
𝑈1 𝑇1
𝑈2 𝑇2
The small crystalline site Cris(𝑋∕𝑆) contains those 𝑈 → 𝑇 such that 𝑈 is an open
subscheme of 𝑋 with morphisms and coverings defined the same.
2 Crystalline cohomology 5
It can be checked that CRIS(𝑋∕𝑆) and Cris(𝑋∕𝑆) are actually sites. Then the
big crystalline topos (𝑋∕𝑆)CRIS is the topos corresponding to the big crystalline
site and similarly for the small crystalline topos (𝑋∕𝑆)Cris . From the definition,
one notices that a sheaf on the crystalline site is a compatible collection of Zariski
sheaves.
The inclusion of sites Cris(𝑋∕𝑆) → CRIS(𝑋∕𝑆) is continuous and co-
continuous. Hence we may define functors 𝑖∗ , 𝑖! ∶ Cris(𝑋∕𝑆) → CRIS(𝑋∕𝑆)
and 𝑖−1 ∶ CRIS(𝑋∕𝑆) → Cris(𝑋∕𝑆) respectively. Moreover, for a morphism
𝑓 ∶ 𝑋∕𝑆 → 𝑌 ∕𝑆 ′ , the natural map CRIS(𝑋∕𝑆) → CRIS(𝑌 ∕𝑆 ′ ) defines 𝑓 −1 , 𝑓∗
respectively.
Then we are able to define the push forward and pull back of crystalline topoi.
For a morphism 𝑓 ∶ 𝑋∕𝑆 → 𝑌 ∕𝑆 ′ , we have arrows:
𝑖∗ 𝑓∗ 𝑖−1
𝑖−1 𝑓 −1 𝑖!
Then 𝑖−1 ∘ 𝑓∗ ∘ 𝑖∗ and 𝑖−1 ∘ 𝑓 −1 ∘ 𝑖! are morphisms of topoi (𝑋∕𝑆)Cris and (𝑌 ∕𝑆)Cris ,
also denoted by 𝑓∗ and 𝑓 −1 . From now on, we will consider the small crystalline
site only.
The structure sheaf 𝒪𝑋∕𝑆 associates each divided power thickening 𝑈 → 𝑇
with the ring 𝒪(𝑇 ). For morphism 𝑋∕𝑆 → 𝑌 ∕𝑆 ′ , there exists a canonical mor-
phism 𝒪𝑌 ∕𝑆 ′ → 𝒪𝑋∕𝑆 . Then we may define module push forward and pull back,
denoted 𝑓∗ and 𝑓 ∗ respectively.
The crystalline site and Zariski sites are linked. There is a morphism of sites
Moreover, for an object 𝑈 → 𝑇 in the crystalline site, those objects lying over
it form a subsite, which is equivalent to the Zariski site of 𝑇 . For a sheaf ℱ ∈
(𝑋∕𝑆)Cris , the corresponding Zariski sheaf is denoted by ℱ𝑇 .
As in the Zariski case, quasi-coherent modules are the mainly objects for study.
Those are, “crystals” for some intuitive reasons.
Definition 2.10. A crystal on Cris(𝑋∕𝑆) is a sheaf of 𝒪𝑋∕𝑆 -modules such that for
any morphism 𝑢 ∶ 𝑇 ′ → 𝑇 , the map induced by 𝑢, 𝑢∗ ℱ𝑇 → ℱ𝑇 ′ is an isomorphism
and for each 𝑇 , ℱ𝑇 is a quasi-coherent 𝒪𝑇 module.
From definitions, a crystal is indeed a quasi-coherent sheaf of 𝒪𝑋∕𝑆 -modules.
Example 2.11. The sheaf 𝒪𝑋 defined by 𝒪𝑋 (𝑈 → 𝑇 ) = 𝒪(𝑈 ) is a crystal. There
is a canonical morphism: 𝒪𝑋∕𝑆 → 𝒪𝑋 , which is surjective.
Remark 2.12. In the definition of the crystalline site, we require 𝑈 to become an
Zariski open set in the object 𝑈 → 𝑇 . We can substitute the condition by requiring
𝑈 → 𝑋 to be an étale map and all the theory still preserved.
6 𝑝‐adic Cohomology Theories
We prove the fundamental comparison theorem in the section and derive some
corollaries. We shall use the notations in the above sections.
Lemma 2.14 (Poincaré lemma). Let (𝐴, 𝐼, 𝛾) be a divided power ring and 𝑃 =
𝐴⟨𝑡1 , ⋯ , 𝑡𝑛 ⟩ with the natural divided power structure, then there is an exact se-
quence:
0 → 𝑃 → Ω1𝑃 ∕𝐴 → ⋯ Ω𝑛𝑃 ∕𝐴 → 0.
𝑌′
𝑝
𝑋 𝑌
𝑆,
𝑛
with 𝑋, 𝑌 , 𝑆 as above and 𝑌 ′ = 𝔸𝑌 . For a crystal ℰ ∈ (𝑋∕𝑆)Cris , there is a
quasi-isomorphism of complexes:
𝑅𝑝∗ (ℰ𝐷𝑋 (𝑌 ′ ) ⊗ Ω•𝑌 ′ ∕𝑆 ) ≃ ℰ𝐷𝑋 (𝑌 ) ⊗ Ω•𝑌 ∕𝑆 .
with filtrants
𝑗 𝑗−𝑘
𝑔𝑟𝑘 Ω𝑌 ′ ∕𝑆 ≃ Ω𝑘𝑌 ′ ∕𝑌 ⊗ 𝑝∗ Ω𝑌 ∕𝑆 .
by the crystal condition. The result then follows from the Poincaré lemma 2.14. ◻
Lemma 2.16. For 𝑖 ∶ 𝑋 → 𝑌 a closed immersion with 𝑌 ∕𝑆 smooth, ℰ a quasi-
coherent 𝐷𝑋,𝛾 (𝑌 )-module with a connection. 𝑅𝑢𝑋∕𝑆∗ ℰ is represented by the
cosimplicial module (or a complex by Dold–Kan correspondence)
ℰ𝐷(0) → ℰ𝐷(1) → ℰ𝐷(2) → ⋯ .
8 𝑝‐adic Cohomology Theories
is given by
𝑒𝑖 𝑖 < 𝑗
ℎ𝑛 (𝑒𝑖 )(𝛼𝑗𝑛 ) =
{ 0 𝑖 ≥ 𝑗,
0 𝑖 < 𝑗,
𝛼𝑗𝑛 ∶ {0, 1, ⋯ , 𝑛} → {0, 1}, 𝛼𝑗𝑛 (𝑖) =
{ 1 𝑖 ≥ 𝑗.
◻
𝑛
Lemma 2.18. Notations as above. When 𝑌 = 𝔸𝑆 , the cosimplicial module
is contractible.
In particular,
𝑅Γ((𝑋∕𝑆)Cris , ℰ ) ≃ 𝑅Γ(𝑋Zar , ℰ ⊗ Ω•𝑌 ∕𝑆 ).
2 Crystalline cohomology 9
⋮ ⋮,
where the rows are de Rham complexes and the horizontal arrows are given by
𝑝∗1 − 𝑝∗2 + ⋯ + (−1)𝑛 𝑝∗𝑛+1 .
𝑛
The problem is étale local. Thus we may assume that 𝑌 = 𝔸𝑆 . The left vertical
arrow of the diagram represents the 𝑅𝑢𝑋∕𝑆∗ ℰ by 2.16. Other vertical arrows have
zero cohomology for they are contractible. One applies the spectral sequence to
conclude the cohomology of the total complex is the cohomology of 𝑅𝑢𝑋∕𝑆∗ ℰ .
We apply the spectral sequence to the other direction. Applying 2.15 for
𝑝𝑖 ∶ 𝑌 𝑛+1 → 𝑌 𝑛 , one deduces that the 𝐸1 package of the spectral sequence
becomes
ℋ0 ℋ1 ⋯
0 0
0
ℋ ℋ1 ⋯
∼
⋮ ⋮.
Here ℋ 𝑖 means the cohomology of the first row. Thus, all arrows except the first
row are cancelled in the 𝐸2 package of the spectral sequence. Thus the cohomology
of the total complex is the cohomology of the first arrow in the original commuta-
tive diagram, i.e., the de Rham complex. The theorem thus follows by comparing
the spectral sequence of two directions. ◻
Proof. Using the Čech-to-derived spectral sequence, we reduce to the local case.
That is, 𝑋 admits a lifting 𝑌 , such that 𝑌 ∕𝑆 smooth. Then there is an isomorphism:
The result follows from the vanishing result of cohomology of (Zariski) quasi-
coherent sheaves. ◻
Theorem 2.21 (Base change thoerem). Suppose there is a base change diagram
for 𝑓 ∶ 𝑋 → 𝑆 and 𝑓 ′ ∶ 𝑋 ′ → 𝑆 ′ ,
𝑢′
𝑋′ 𝑋
𝑆0′ 𝑆0
𝑢
𝑆′ 𝑆.
Proof. For the case that there is a smooth lifting 𝑌 ∕𝑆 of 𝑋, it follows from the base
change theorem for Zariski sheaves. The general case follows from cohomological
descent. ◻
Theorem 2.22 (Finiteness theorem). Let 𝑋∕𝑆0 be proper and smooth and 𝑆 be
noetherian. Then for any coherent crystal ℱ flat over 𝒪𝑋∕𝑆 , the complex 𝑅𝑓𝑋∕𝑆∗ ℱ
is quasi-isomorphic to a perfect complex, i.e. a complex whose terms are locally
free sheaves of finite rank and the number of non-zero terms of which is finite.
Proof. By Čech resolution and the computation in the local case 𝑅𝑓𝑋∕𝑆∗ ℱ has
finite tor-dimension. Hence it suffices prove that 𝑅𝑓𝑋∕𝑆∗ ℱ has coherent cohomol-
ogy. For 𝑆 is noetherian, ℐ 𝑛 = 0 for sufficiently large 𝑛. By dévissage arguments,
it suffices to prove for coherent 𝒪𝑋∕𝑆0 -modules. Then
Then the result follows from the finiteness result of cohomology of coherent
sheaves in Zariski topology. ◻
We may then define the “completed” version. That is, 𝑆 is a 𝑝 -adically com-
plete formal scheme with ℐ = (𝑝), 𝑆𝑛 = 𝑆∕ℐ 𝑛 . The completed crystalline site
Cris(𝑋∕𝑆)̂ is the combination of all Cris(𝑋∕𝑆𝑛 ) and crystals similarly defined. One
defines the completed crystalline topos then.
The proof is given in [BO78]. Then completions of the comparison results hold.
For example,
𝑅𝑢𝑋∕𝑆∗ ℰ ≃ ℰ ⊗ Ω•𝐷∕𝑆̂ ,
Then the vanishing theorem, the base change theorem and the finiteness theorem
also hold in the completed case for coherent crystals, since 𝑅 lim works well then.
If moreover 𝑋∕𝑆0 is proper, ℰ is coherent and 𝑋 lifts to 𝑌 , then 𝑌 is also
proper. By Grothendieck’s existence theorem, there exists a coherent 𝒪𝑌 -module,
also denoted by ℰ such that
In particular,
𝑅Γ((𝑋∕𝑆)Cris , 𝒪𝑋∕𝑆 ) ≃ 𝑅Γ(𝑋Zar , Ω•𝑌 ∕𝑆 ).
•
We will denote the left hand side by 𝐻cris (𝑋∕𝑆). We restrict to the case 𝑆 =
Spec(W(𝑘)) and 𝑆0 = Spec(𝑘) for some perfect field 𝑘 of characteristic 𝑝, which is
the crystalline cohomology for characteristic 𝑝 varieties.
We firstly discuss the case that 𝑋∕𝑘 is proper and 𝑋 admits a lifting. Then
there exists a cartesian diagram (if there exists an embedding W(𝑘) → ℂ, which is
the case for 𝑘 finite):
𝑋 𝒳 𝑋 an
with 𝒳 a proper smooth scheme over W(𝑘). Then we have the following compari-
son theorem between crystalline cohomology and Betti cohomology.
Theorem 2.24. There is a canonical isomorphism
𝑛
𝐻cris (𝑋∕W(𝑘)) ⊗ ℂ ≃ 𝐻 𝑛 (𝑋 an , ℂ),
All the modules in the left hand side is locally free. After tensoring ℂ, it becomes
the spectral sequence
𝑝𝑞 𝑝
𝐸1 = ℍ𝑞 (Ω𝑋 an ∕ℂ ) ⇒ ℍ𝑝+𝑞 (Ω•𝑋 an ∕ℂ ).
By (complex) Hodge theory, the spectral sequence degenerates and the right hand
side is isomorphic to 𝐻 𝑝+𝑞 (𝑋 an , ℂ). Hence the original spectral sequence degen-
erates and the result follows. ◻
For example, ℙ𝑛 , and hypersurfaces in ℙ𝑛 admits a lifting. Moreover, by
obstruction theory, such a lifting exists if and only if the obstruction class
𝜉 ∈ 𝐻 2 (𝑋, 𝒯𝑋 ) vanishes. In particular, this is the case when 𝑋 is a curve.
Theorem 2.25 (Künneth formula). For a crystal ℰ1 on 𝑋1 and a crystal ℰ2 on 𝑋2 ,
ℰ1 ⊠ ℰ2 defines a crystal on 𝑋1 × 𝑋2 and
Proof. The map is natural and we may check it locally. Then there are closed
immersions 𝑋1 → 𝑌1 and 𝑋2 → 𝑌2 with 𝑌1 , 𝑌2 ∕𝑆 smooth. The result follows
from the fact that the tensor product of the de Rham complex of ℰ1 and ℰ2 is
isomorphic to the de Rham complex of ℰ1 ⊠ ℰ2 . ◻
Theorem 2.26 (Poincaré duality). For 𝑋∕𝑘 proper smooth of dimension 𝑑, there
2𝑑
is a trace map tr ∶ 𝐻cris (𝑋) ≃ W(𝑘), such that for all coherent crystal ℰ over
𝑋∕W(𝑘), the pairing
is perfect.
Proof. We define the trace map firstly. Locally, 𝑋 lifts to a smooth scheme 𝑋̃
over W(𝑘) with 𝑝-adic completion by 𝒳 (though it is no longer compact) . Take a
compactification of 𝑋,̃ 𝑋.̄ 𝑋̄ → Spec(W(𝑘)) has a dualizing complex 𝜔. Then
𝜔|𝑋̃ ≃ Ω𝑑𝑋∕W(𝑘)
̃ .
The trace map is defined as (by the usual procedure, one checks the map is inde-
pendent of 𝑋)̄
2𝑑
𝐻cris (𝑋) ≃ 𝐻 𝑑 (𝒳 , Ω𝑑𝒳 ) → 𝐻 𝑑 (𝑋,̄ 𝜔) ≃ W(𝑘).
By derived Nakayama lemma and the finiteness theorem, it suffices to prove af-
ter tensoring 𝑘. By base change theorem, it suffices to prove similar results on
(𝑋∕𝑘)Cris . Then the derived global sections is represented by de Rham complexes
and the result follows from the Serre duality. ◻
Remark 2.27. The cycle class map can be defined: defining it locally by de Rham
cohomology and then applying cohomological descent.
2 Crystalline cohomology 13
Remark 2.28. Crystalline cohomology does not work well for non-proper or non-
smooth schemes. See 3.6 and tag 07LI in [Stacks].
Remark 2.29. For varieties with normal crossing singularities and becoming an
open set in a proper variety with complement a normal crossing divisor, there is a
canonical log structure on it. One defines their log-crystalline sites (similar to the
crystalline site, but requiring all objects having a log structure) and the correspond-
ing cohomology, called the log crystalline cohomology.
The log crystalline cohomology has similar comparison results to de Rham
cohomology and one can apply this to show that it is a good cohomology theory.
In fact, it can be proved that there is a comparison between log crystalline co-
homology and rigid cohomology. However, the log crystalline cohomology is in-
dispensable for there is an operator 𝑁, the log monodromy operator on it. Those
operators are required in 𝑝-adic Hodge theory.
Frobenius action
For this section, let 𝑆 = Spec(W(𝑘)) and 𝑆0 = Spec(𝑘), where 𝑘 is a characteristic
𝑝 perfect field. 𝑋 be a smooth scheme over 𝑘. The (absolute) Frobenius actions
𝐹𝑋 and 𝐹𝑆 induce a commutative diagram, where the square is cartesian:
𝐹𝑋
𝑋
𝐹𝑋∕𝑆
𝐹𝑆′
𝑋′ 𝑋
𝐹𝑆
𝑆 𝑆.
Proposition 2.30. The image of the map above is exactly the largest subcomplex
of 𝑝𝑖 𝐹𝑌 ∕𝑆∗ Ω𝑖𝑌 ∕𝑆 . Moreover, the map is an isomorphism to its image.
𝑐 −1
Ω𝑖𝑋∕𝑘 ∼ ℋ 𝑖 (Ω•𝑋∕𝑘 ).
with 𝑐 −1 (𝑥) = 𝑥, 𝑐 −1 (𝑑𝑥) = 𝑥𝑝−1 𝑑𝑥 and 𝑐 −1 is compatible with the wedge product.
Proof. Note that both sides are compatible with étale base change. Hence it suffices
to construct a canonical isomorphism when 𝑋 = Spec(𝑘[𝑡1 , ⋯ , 𝑡𝑛 ]). Then the
result follows from a direct calculation. ◻
(Ω•𝑌 ′ ∕𝑆 )𝜖𝑙 ∕(Ω•𝑌 ′ ∕𝑆 )𝜖𝑙+1 (𝐹𝑌 ∕𝑆 Ω•𝑌 ∕𝑆 )𝜖𝑙 ∕(𝐹𝑌 ∕𝑆 Ω•𝑌 ∕𝑆 )𝜖𝑙+1
•𝑝−𝜖𝑙 (𝑗)
∼
∼
𝑐 −1
Ω𝑋 ′ ∕𝑘 [−𝑗] ∼ ℋ 𝑗 (𝐹𝑋∕𝑘∗ Ω𝑌 ∕𝑆0 )[−𝑗].
Definition 2.34. For a cogauge 𝜂, the functor 𝐿𝜂 in the derived category of abelian
sheaves 𝒟 (𝒜 𝑏(𝑋)) is defined as follows: for a complex 𝐾, take a flat (as a ℤ -
module) resolution and then apply (−)𝜂 on it. For two cogauges 𝜂, 𝜂 ′ , with 𝜂 ′ ≤ 𝜂,
the functor 𝐿𝜂 ′ ∕𝜂 is defined by the distinguished triangle:
𝐿𝜂 → 𝐿𝜂 ′ → 𝐿𝜂 ′ ∕𝜂.
Proof. Choose such a lifting locally and glue the above local isomorphisms to-
gether by cohomological descent. ◻
Definition 2.36. For an 𝐹 -crystal 𝑀 (i.e. a finite rank free W(𝑘) -module with
a semilinear Frobenius action), choose a basis of it and then the Frobenius action
is represented by a matrix. The slopes of the action are the absolute values of the
eigenvalues of the matrix (the slopes are independent of the choice of bases). Let
𝛼1 ≤ ⋯ ≤ 𝛼𝑚 be the slopes counted by multiplicity. Then the Newton polygon
of 𝑋 is defined by the polygon with domain [0, 𝑚] with slope 𝛼𝑖 in the interval
𝑙
[𝑖 − 1, 𝑖]. If 𝑀 ′ = im(𝐹 ) = ⨁𝑖=0 𝑝𝑖 𝑀𝑖 , the Hodge number of 𝑀 is (𝑒0 , ⋯ , 𝑒𝑙 ),
the ranks of 𝑀0 , ⋯ , 𝑀𝑙 .
Definition 2.37. The Hodge polygon with parameter (𝑎0 , ⋯) is the polygon with
slope 𝑖 in the interval [𝑎0 + ⋯ + 𝑎𝑖−1 , 𝑎0 + ⋯ + 𝑎𝑖 ].
Lemma 2.38. The following basic result holds:
• The Newton polygon of an 𝐹 -crystal lies on the Hodge polygon of it.
• The Hodge polygon with parameter (𝑎0 , ⋯) lies on the Hodge polygon with
parameter (𝑏0 , ⋯) if and only if for any 𝑖 > 1, 𝑖𝑎0 + (𝑖 − 1)𝑎1 + ⋯ + 𝑎𝑖−1 ≤
𝑖𝑏0 + (𝑖 − 1)𝑏1 + ⋯ 𝑏𝑖−1 holds for any 𝑖.
• ℓ(𝑀 ′ ∕𝑀 ′ ∩ 𝑝𝑖 𝑀) = 𝑖𝑒0 + ⋯ + 𝑒𝑖−1 , where ℓ means length.
We are then able to prove the Katz conjecture:
𝑛
Theorem 2.39 (Katz, Mazur-Ogus). The Newton polygon of 𝑀 = 𝐻cris (𝑋∕W(𝑘))
𝑛−𝑞 𝑞
lies on the Hodge polygon with parameters 𝑎𝑞 = ℎ (𝑋, Ω𝑋∕𝑘 ).
Proof. From the above facts, we are reduced to prove the following:
where 𝜁𝑖 (𝑛) = max{0, 𝑛 − 𝑖}. After modulo torsion, the image of 𝜑 is 𝑀 ′ and the
image of 𝜓 is 𝑝𝑖 𝑀. Then 𝑀 ′ ∕𝑀 ′ ∩ 𝑝𝑖 𝑀 is a quotient of im 𝑞. Moreover, there is
a distinguished triangle:
ℍ𝑛 (𝑋, 𝐿(𝜁𝑗−1 +𝑗 −1)∕(𝜁𝑗 +𝑗)) → ℍ𝑛 (𝑋, 𝐿𝜁0 ∕(𝜁𝑗 +𝑗)) → ℍ𝑛 (𝑋, 𝐿𝜁0 ∕(𝜁𝑗−1 +𝑗 −1))
16 𝑝‐adic Cohomology Theories
where 𝜏 means the truncation. Then there is a spectral sequence, where the iso-
morphism is given by the Cartier isomorphism:
𝑝𝑞 𝑞
𝐸2 (𝑞 ≤ 𝑗 − 1) = 𝐻 𝑝 (𝑋, Ω𝑋∕𝑘 ) ≃ 𝐻 𝑝 (𝑋, ℋ 𝑞 (Ω•𝑋∕𝑘 )) ⇒ ℍ𝑝+𝑞 (𝑋, 𝜏≤𝑗−1 Ω•𝑋∕𝑘 ).
Thus, dim ℍ𝑝+𝑞 (𝑋, 𝜏≤𝑗−1 Ω•𝑋∕𝑘 ) ≤ ℎ0 + ⋯ + ℎ𝑗−1 . Combining all those 𝑗 gives the
result. ◻
3 Rigid cohomology
where Ω•]𝑋[ ∕𝐾
is the de Rham complex.
𝒫
𝑋 𝒫
𝑆,
𝑅𝑝1∗ Ω•]𝑋[ ∕𝐾
≃ Ω•]𝑋[ ∕𝐾
]𝑋[𝒫1 .
𝒫 𝒫
Remark 3.5. Such an immersion exists locally: for affine 𝑋, 𝑋 can be embedded
𝑛 ̂𝑛 . We may use the cohomological descent method to define
into 𝔸𝑘 . Then into 𝔸 𝒪
the convergent cohomology in general.
𝑑
𝐾⟨𝑇 ⟩ 𝐾⟨𝑇 ⟩𝑑𝑇 .
However, 𝐻 1 is not finite dimensional for there are infinitely many linearly inde-
pendent convergent series that no longer converge after integration.
such that 𝑗 is an open immersion and 𝑖 a closed immersion. A proper smooth frame
is a frame such that 𝑌 is proper and ]𝑌 [𝒫 is smooth.
†
Definition 3.8. Let ℱ be an abelian sheaf on ]𝑌 [𝒫 , the overconvergent sheaf 𝑗𝑋 ℱ
is
†
𝑗𝑋 ℱ = colim 𝑗𝑉 ∗ 𝑗𝑉∗ ℱ ,
where 𝑉 runs through all the strict neighbourhood of ]𝑋[𝒫 in ]𝑌 [𝒫 and 𝑗𝑉 the
inclusion 𝑉 →]𝑌 [𝒫 .
Definition 3.9. If there exists a proper smooth frame of 𝑋, the rigid cohomology
•
𝐻rig (𝑋) is
• †
𝐻rig (𝑋) = 𝑅Γ(𝑗𝑋 Ω•]𝑌 [ ∕𝐾
),
𝒫
where Ω•]𝑌 [ ∕𝐾
is the de Rham complex.
𝒫
18 𝑝‐adic Cohomology Theories
𝐾⟨𝑇 ⟩† 𝐾⟨𝑇 ⟩† 𝑑𝑇 ,
where 𝐾⟨𝑇 ⟩† contains all the overconvergent series (i.e. converges in some ra-
dius 𝜇 > 1). Then 𝐻 1 = 0 as the integration of overconvergent series remains
overconvergent.
In general, for an affine smooth variety 𝑋 = Spec(𝐴), 𝐴 can be lifted to
a finitely presented formally smooth algebra 𝒜 = 𝒪⟨𝑇1 , ⋯ , 𝑇𝑛 ⟩∕(𝑓1 , ⋯ , 𝑓𝑚 ).
Then the rigid cohomology of 𝑋 becomes the cohomology of the de Rham
†
complex Ω• † , where 𝒜𝐾 ≃ 𝐾⟨𝑇1 , ⋯ , 𝑇𝑛 ⟩† ∕(𝑓1 , ⋯ , 𝑓𝑚 ). That is, the Monsky–
𝒜𝐾 ∕𝐾
Washnitzer cohomology.
Remark 3.13. The way considering overconvergent series can be justified by the
language of adic spaces. Indeed, ]𝑋[𝒫 is not closed in ]𝑌 [𝒫 when viewing them
as adic spaces and we should not expect that a non-closed thing can have a good
cohomology theory. Thus we should consider the closure of ]𝑋[𝒫 in ]𝑌 [𝒫 and
work out its cohomology. Indeed, the strict neighbourhoods are neighbourhoods
of the closure of ]𝑋[𝒫 , ]𝑋[𝒫 . Thus, for 𝑖 ∶ ]𝑋[𝒫 →]𝑌 [𝒫 ,
•
𝐻rig (𝑋) = 𝑅Γ]𝑌 [𝒫 𝑖∗ 𝑖∗ Ω•]𝑌 [ ∕𝐾
.
𝒫
Proof. The proposition is trivial for projective 𝑋, for there exists a closed immer-
𝑛
sion 𝑋 → ℙ𝑘 , then an immersion 𝑋 → ℙ ̂𝑛 and a frame 𝑋 → 𝑋 → ℙ ̂𝑛 . In general,
𝒪 𝒪
one checks that both convergent cohomology and rigid cohomology functors are
3 Rigid cohomology 19
right Kan extensions from the category of projective varieties to the category of
proper varieties (for a proper variety, it has a simplicial resolution by projective
varieties). See [Tsu03] for details. ◻
We are now able to prove the comparison theorem between crystalline coho-
mology and convergent/rigid cohomology.
Ω•𝒳 ∕𝒪 ⊗ 𝐾 ≃ Ω•𝒳 .
𝐾 ∕𝐾
𝑋′ 𝑌′ 𝒫′
𝑋 𝑌 𝒫
Moreover, those categories are ”stacks” in some sense. That is, taking the fifth
for example, there is an equivalence (𝒰 = {𝑌𝑖 → 𝑌 } be an open covering):
Isoc† (𝑋 → 𝑌 ) → lim Isoc† (𝑋• → 𝑌• )
𝒰
where 𝑋• , 𝑌• means the simplicial scheme with respect to the open covering 𝒰.
Then we can define those concepts for the case 𝑋 does not admit a proper smooth
frame by descent.
Then for an overconvergent isocrystal ℰ (with the corresponding coherent sheaf
with integrable connection also denoted ℰ ), we may define the rigid cohomology
with coefficients
𝑛
𝐻rig (𝑋, ℰ ) ∶= ℍ𝑛 (ℰ ⊗ Ω•]𝑌 [ ∕𝐾 ).
𝒫
Then we have exact sequences for 𝑋 = 𝑈 ∪ 𝑍, with 𝑈 open and 𝑍 closed (we
shall omit the subscript “rig” if there is no confusion):
and
⋯ → 𝐻𝑐• (𝑈 ) → 𝐻𝑐• (𝑋) → 𝐻𝑐• (𝑍) → 𝐻𝑐•+1 (𝑈 ) ⋯ ,
and some excision theorems. For example, for 𝑍 ⊂ 𝑌 ⊂ 𝑋 being closed varieties,
there is an exact sequence
Using similar methods, one proves that the cohomology with compact support
is of finite dimension.
𝐻 • (𝑋, ℰ1 ) ⊗ 𝐻 • (𝑌 , ℰ2 ) ≃ 𝐻 • (𝑋 × 𝑌 , ℰ1 ⊠ ℰ2 ).
Proof. It is almost the same as the case for crystalline cohomology. One can check
it by first working locally via the de Rham complex and then noticing that the map
is canonical. ◻
is perfect.
Proof. The trace map is defined almost the same as the case for crystalline coho-
mology. We shall prove the case that ℰ is trivial. We consider two propositions:
(a)𝑛 Poincaré duality holds for all smooth varieties of dimension no greater than
𝑛.
(the trace map can be similarly defined) is perfect for 𝑍 (may not be smooth)
of dimension no greater than 𝑛 and 𝑋 smooth.
References 23
The case (a)0 is trivial. (b)0 is from the Gysin isomorphism. (b)𝑛 − 1 ⇒ (a)𝑛 is
done by de Jong’s alternation theorem and the Poincaré duality for crystalline co-
homology (note that the trace map for finite morphisms are compatible with all
those constructions). (b)𝑛 − 1 and (a)𝑛 ⇒ (b)𝑛 is done by excision and Gysin iso-
morphism. The whole process is similar to the proof of finiteness and we shall omit
it. ◻
Remark 3.25. The case for general overconvergent 𝐹 -isocrystals ℰ can be proved
by “𝑝-adic local monodromy theorem”, which claims that after a suitable base
change, the isocrystal is unipotent, i.e., has a filtration with each subquotient be-
coming trivial. See [Ked06] for details.
Remark 3.26. The cycle class map can be defined firstly by de Rham cohomology
locally and then applying cohomological descent.
All requirements for a good cohomology theory are therefore proved. We end
this article by remarking an example.
Example 3.27. If a family of varieties over 𝑘 comes from reduction of a proper
smooth family, the cohomology of all varieties in the family has the same dimen-
sion. The family of hypersurfaces in the projective space is the case. For example,
all plane cubic curves have ℎ0 = 1, ℎ1 = 2, ℎ2 = 1 and ℎ𝑘 = 0 for 𝑘 > 2. Here
ℎ𝑘 ∶= dim 𝐻 𝑘 .
Proof. The first statement is from the fact that Betti numbers are invariant over
a family and the comparison theorem between de Rham cohomology and Betti
cohomology.
The second statement is from the fact that for any homogenous polynomial
𝑓 ∈ 𝑘[𝑥0 , … , 𝑥𝑛 ], there is a element 𝑓 ̃ ∈ W(𝑘)[𝑥0 , … , 𝑥𝑛 ] defining a smooth
𝑛
hypersurface of ℙ𝐾 , as the set of singular hypersurfaces is Zariski closed in the
parameter space and all 𝑓 ̃ is Zariski dense in the the parameter space.
The last statement follows from computations of Betti numbers of cubic curves.
◻
The example illustrates that for singular varieties, rigid cohomology may not
be the “intuitive” one. For example, the first Betti number of a cuspidal cubic curve
over ℂ is 0 and the first rigid cohomology of a cuspidal cubic curve over 𝑘 is of
dimension 2.
References
[Ber74] Pierre Berthelot (1974). Cohomologie cristalline des schémas de caractéristique
𝑝 > 0. Lecture Notes in Mathematics 406. Springer.
[Ber86] Pierre Berthelot (1986). “Géométrie rigide et cohomologie des variétés al-
gébriques de caractéristique 𝑝”. Mémoires de la Société Mathématique de
France (23), 7–32.
24 𝑝‐adic Cohomology Theories
He Yuxin1
ABSTRACT
We may come across many cohomology theorems when we study geome-
try and topology. Theorems like de Rham theorm and Dolbeault theorem tell
us the connection between different cohomology theories. The passage will
introduce the axiomatic cohomology theory to give a more general view of the
phenomenon. The connection do not rely on a particular cohomology theory,
but is a natural conclusion of the propositions in homological algebra.
Contents
1 Preliminaries 25
Sheaves and presheaves . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Tensors and exact sequences . . . . . . . . . . . . . . . . . . . . . . . 28
3 Examples 38
Alexander–Spanier cohomology . . . . . . . . . . . . . . . . . . . . . 38
de Rham cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4 Multiplicative structure 41
1 Preliminaries
𝐽= 𝑆𝑥′ ∕𝑆𝑥 ,
⋃
𝑥∈𝑀
and equip it with the quotient topology. 𝐽 is called the quotient sheaf of 𝑆 ′ modulo
𝑆.
Definition 1.6. An exact sequence of sheaves is a sequence of sheaves and homo-
morphisms:
𝜑𝑖−1 𝜑𝑖 𝜑𝑖+1
⋯ / 𝑆 𝑖−1 / 𝑆𝑖 / 𝑆 𝑖+1 / ⋯,
0 /𝑅 /𝑆 /𝑇 / 0,
with 0 the constant sheaf of the form 𝑀 × 𝐺, where 𝐺 is the trivial 𝐾-module.
The sequence is exact if and only if for each 𝑥 ∈ 𝑀, the sequence
𝜑𝑉 𝜑𝑈
𝜌′𝑈 ,𝑉
𝑆𝑉′ / 𝑆′
𝑈
1. if 𝑈 = ⋃𝑖∈𝐼 𝑈𝑖 and 𝑓𝑖 ∈ 𝑆𝑈𝑖 with 𝑓𝑖 |𝑈𝑖 ∩𝑈𝑗 = 𝑓𝑗 |𝑈𝑖 ∩𝑈𝑗 , then there exists
𝑓 ∈ 𝑆𝑈 satisfying 𝑓𝑈𝑖 = 𝑓𝑖 , here if 𝑈 ∩ 𝑉 and 𝑓 ∈ 𝑆𝑉 , use 𝑓 |𝑈 to denote
𝜌𝑈 ,𝑉 𝑓 .
If a presheaf 𝑃 = {𝑆𝑈 ; 𝜌𝑈 ,𝑉 } satisfies the two conditions, then say the presheaf
is complete.
Notice a sheaf homomorphism 𝜑 ∶ 𝑆 → 𝑆 ′ induce presheaf homomorphism
𝜑 ∶ 𝛼(𝑆) → 𝛼(𝑆 ∗ ), and (𝜓 ∘ 𝜑)∗ = 𝜓 ∗ ∘ 𝜑∗ . Besides, a presheaf homomorphism
∗
𝛽(𝛼(𝑆)) → 𝑆
𝜉 = 𝜌𝑝,𝑈 𝑓 ↦ 𝑓 (𝑝), 𝑤ℎ𝑒𝑟𝑒 𝑓 ∈ Γ(𝑈 , 𝑆)
𝑆𝑈 → Γ(𝑈 , 𝛽(𝑃 ))
(𝑓 ∈ 𝑆𝑈 ) ↦ (𝑝 ∈ 𝑈 ↦ 𝜌𝑝,𝑈 𝑓 ).
0 / 𝑆′ /𝑆 / 𝑆 ′′ /0
is not always exact. In fact, the sequence may not be exact at Γ(𝑆 ′′ ). We only have
the following exact sequence
How does the sequence differ from exact sequence will be seen by extending the
sequence to a long exact sequence.
Theorem 1.11. If there is a short exact sequence
0 /𝑅 /𝑆 /𝑇 / 0,
where 𝑅 = 𝑘𝑒𝑟(𝑆 → 𝑇 ) is a fine sheaf, then it will induce the short exact sequence
𝜑
0 / Γ(𝑅) 𝑖 / Γ(𝑆) / Γ(𝑇 ) / 0.
Proof. The exactness at Γ(𝑅) is easy to see pointwisely. The exactness at Γ(𝑆) is
just by the definition of the kernel sheaf. So it suffices to check the sufficiency at
Γ(𝑇 ). For 𝑡 ∈ Γ(𝑇 ), since sheaves are locally homeomorphic to 𝑀, for an open
covering {𝑈𝑖 } of M, there exist 𝑠𝑖 ∈ Γ(𝑈𝑖 , 𝑆) such that 𝜑 ∘ 𝑠𝑖 = 𝑡|𝑈𝑖 , and {𝑈𝑖 } is
locally finite. Set 𝑠𝑖𝑗 = 𝑠𝑖 −𝑠𝑗 over 𝑈𝑖 ∩𝑈𝑗 , and let 𝑙𝑖 be the corresponding partition
of unity, then set
𝑠′𝑖 = ∑ 𝑙𝑗 ∘ 𝑠𝑖𝑗 ,
𝑗
30 Sheaf Cohomology
holds, then by the gluing lemma, there exists some 𝑠 ∈ Γ(𝑆) with 𝑠|𝑈𝑖 = 𝑠𝑖 −𝑠′𝑖 ,
then 𝜑 ∘ 𝑠 = 𝑡. ◻
For the tensor product, consider the tensor functor.
Proposition 1.12. If there is a short exact sequence
0 / 𝑆′ /𝑆 / 𝑆 ′′ / 0,
0 / 𝑆′ ⊗ 𝑇 / 𝑆 ⊗𝑇 / 𝑆 ′′ ⊗ 𝑇 / 0.
𝑆′ ⊗ 𝑇 / 𝑆 ⊗𝑇 / 𝑆 ′′ ⊗ 𝑇 / 0.
𝐴
𝑢 /𝐵 𝑣 /𝐶 /0
is exact for all 𝐾-module 𝑁 and the fact that 𝐻𝑜𝑚(𝐴⊗𝐵, 𝑁) ≃ 𝐻𝑜𝑚(𝐴, 𝐻𝑜𝑚(𝐵, 𝑁)).
While for the injectivity of 𝑆 ′ ⊗ 𝑇 → 𝑆 ⊗ 𝑇 ,since tensor product commutes
with direct limit, it suffices to consider the case of finitely generated torsionfree
𝐾-modules, and all such modules have the form 𝐾 ⊕𝑛 , for which the injectivity is
clear.
Theorem 1.13. If
0 / 𝑆′ /𝑆 / 𝑆 ′′ /0
0 / 𝑆′ ⊗ 𝑇 / 𝑆 ⊗𝑇 / 𝑆 ′′ ⊗ 𝑇 / 0.
At the same time, if either 𝑇 or 𝑆 ′ is a fine sheaf, then there is an exact sequence
𝐻 0 (𝑀, 𝑆)
≃ / Γ(𝑆)
𝐻 0 (𝑀, 𝑆 ′ )
≃ / Γ(𝑆 ′ ).
𝑆′ B /𝑆
BB
BB
BB
B!
𝑆 ′′
induces the commutative diagram
0 / 𝑆′ /𝑆 / 𝑆 ′′ /0
0 / 𝑇′ /𝑇 / 𝑇 ′′ / 0,
𝐻 𝑞 (𝑀, 𝑇 ′′ ) / 𝐻 𝑞+1 (𝑀, 𝑇 ′ ).
2.
𝐻 𝑞 (𝑀, 𝑆) → 𝐻 𝑞 (𝑀, 𝑆 ′ )
for 𝑆 → 𝑆 ′ is induced by 𝐻 𝑞 (Γ(𝐶 ∗ ⊗ 𝑆)) → 𝐻 𝑞 (Γ(𝐶 ∗ ⊗ 𝑆 ′ )),
3. 𝐻 𝑞 (𝑀, 𝑆 ′′ ) → 𝐻 𝑞+1 (𝑀, 𝑆 ′ ) for 0 → 𝑆 ′ → 𝑆 → 𝑆 ′′ → 0 is induced by
applying the snake lemma to the sequence of cochain maps
0 → Γ(𝐶 ∗ ⊗ 𝑆 ′ ) → Γ(𝐶 ∗ ⊗ 𝑆) → Γ(𝐶 ∗ ⊗ 𝑆 ′′ ) → 0,
which is exact since the resolution is torsionless.
2 Axiomatic sheaf cohomology 33
Uniqueness
After proving the existence of the cohomology theory using the resolution of the
constant sheaf. It left to prove that it is unique under isomorphisms. Firstly it is
required to make clear how to describe the homomorphism between cohomology
theories.
Definition 2.3. For two sheaf cohomology theories ℋ , ℋ ̃ , a homomorphism be-
tween them consists of a homomorphism 𝐻 𝑞 (𝑀, 𝑆) → 𝐻̃ 𝑞 (𝑀, 𝑆) for each 𝑆 and
𝑞, which should satisfy the following commutative conditions.
1.
𝐻 0 (𝑀, 𝑆)
≃ / Γ(𝑆)
𝐻̃ 0 (𝑀, 𝑆)
≃ / Γ(𝑆).
2. For homomorphism 𝑆 → 𝑇 ,
𝐻 𝑞 (𝑀, 𝑆) / 𝐻 𝑞 (𝑀, 𝑇 )
𝐻̃ 𝑞 (𝑀, 𝑆) / 𝐻̃ 𝑞 (𝑀, 𝑇 ).
𝐻̃ 𝑞 (𝑀, 𝑆 ′′ ) / 𝐻̃ 𝑞+1 (𝑀, 𝑆 ′ ).
ℋA
𝑖𝑑 /ℋ ,
AA }}
AA𝜑 𝜓 }}
AA }}
A ~}}
ℋ̃
and
ℋ̃ A
𝑖𝑑 / ℋ̃ .
AA }}
AA𝜓 𝜑
}}
AA }}
A ~}
}
ℋ
Lemma 2.4. For each sheaf 𝑆, there exist a fine sheaf 𝑆0 and a sheaf 𝑆 ̄ with the
short exact sequence
0 → 𝑆 → 𝑆0 → 𝑆 ̄ → 0.
Proof. Set 𝑆0 to be the sheaf of germs of discontinuous sections of 𝑆, and set 𝑆 ̄ to
be the quotient sheaf 𝑆0 ∕𝑆. To construct 𝑆0 , set 𝑆𝑈 to be the set of discontinuous
maps over 𝑈 , that is map 𝑓 ∶ 𝑈 → 𝑆 with 𝜋 ∘ 𝑓 = 𝑖𝑑. This forms a presheaf, and
let 𝑆0 to be the associated sheaf.
It suffices to prove that 𝑆0 is a fine sheaf. Set {𝑈𝑖 } to be a locally finite open
covering, then there exists a refined open covering {𝑉𝑖 } with 𝑉𝑖̄ ⊂ 𝑈𝑖 . Set the func-
tion 𝜑𝑖 = 𝜒𝑉𝑖 , and arrange every 𝑥 ∈ 𝑀 a 𝑉𝑖 , denoted by 𝑉𝑖𝑥 containing it. Set
𝜓𝑖 (𝑥) = 𝜒𝑉𝑖 (𝑥)𝛿𝑖𝑥 ,𝑖 , then ∑𝑖 𝜓𝑖 = 1. Now set the endomorphism of presheaf to be
𝑙𝑖̃ (𝑆)(𝑚) = 𝜓𝑖 (𝑚)𝑠(𝑚), and the presheaf endomorphism induces sheaf endomor-
phism 𝑙𝑖 of 𝑆. That’s the desired partition of unity.
The injection 𝑆 → 𝑆0 is given by the local homeomorphism at every 𝑠 ∈ 𝑆.
This is obviously injective and a sheaf homomorphism. ◻
2 Axiomatic sheaf cohomology 35
Theorem 2.5. For two cohomology theories ℋ and ℋ ̃ , there exists a unique ho-
momorphism ℋ → ℋ ̃ .
𝐻 0 (𝑀, 𝑆)
≃ / Γ(𝑆)
𝐻̃ 0 (𝑀, 𝑆)
≃ / Γ(𝑆).
𝑖𝑑 𝑖𝑑
Γ(𝑆0 ) / Γ(𝑆 ) / 𝐻̃ 1 (𝑀, 𝑆) /0
0
0 / 𝐻̃ 𝑞−1 (𝑀, 𝑆)̄ / 𝐻̃ 𝑞 (𝑀, 𝑆) /0
𝐻 𝑞−1 (𝑀, O𝑇 ̄ ) / 𝐻 𝑞 (𝑀, 𝑇 ) /0
OOO MMM
OOO MMM
OOO MMM
O' MM&
𝐻 𝑞−1 (𝑀, 𝑇 ̄ ) / 𝐻̃ 𝑞 (𝑀, 𝑇 ) /0
By the assumption all the other five faces of the cube are commutative, then the
desired right face of this cubic is also commutative. Notice that 𝑆 → 𝑇 will induce
the commutative diagram
0 /𝑆 /𝑆 / 𝑆̄ /0
0
0 /𝑇 /𝑇 / 𝑇̄ / 0.
0
The first square is easy to see the commutativity, and then map 𝑆 ̄ → 𝑇 ̄ can be
canonical constructed.
While for the third commutative diagram, the short exact sequence 0 → 𝑅 →
𝑆 → 𝑇 → 0 can be extended to the commutative diagram
0 /𝑅 /𝑆 /𝑇 /0
𝑖𝑑
0 /𝑅 /𝑆 /𝐺 /0
O O0
𝑖𝑑
0 /𝑅 / 𝑅0 / 𝑅̄ / 0,
here 𝐺 = 𝑆0 ∕𝑅.
The diagram can be extended to
0 /𝑅 /𝑆 /𝑇 /0
𝑖𝑑
0 /𝑅 /𝑆 /𝐺 /0
O O0 O
𝑖𝑑
0 /𝑅 / 𝑅0 / 𝑅̄ / 0,
2 Axiomatic sheaf cohomology 37
𝑖𝑑
/ Γ(𝑆 ) / Γ(𝐺) / 𝐻 1 (𝑀, 𝑅) /0
O 0 O O
𝑖𝑑
𝐻 0 (𝑀, 𝑇 )
≃ / Γ(𝑇 ) / Γ(𝐺)∕Γ(𝑆0 ) o ≃ ̄
Γ(𝑅)∕Γ(𝑅 0)
≃ / 𝐻 1 (𝑀, 𝑅),
and by the construction of the the map ℋ → ℋ ̃ , the following diagram is commu-
tative
𝐻 0 (𝑀, 𝑇 )
≃ / Γ(𝑇 ) / Γ(𝐺)∕Γ(𝑆 ) o ≃ ̄
Γ(𝑅)∕Γ(𝑅
≃ / 𝐻 1 (𝑀, 𝑅)
0 0)
𝑖𝑑 𝑖𝑑 𝑖𝑑
𝐻̃ 0 (𝑀, 𝑇 )
≃ / Γ(𝑇 ) / Γ(𝐺)∕Γ(𝑆0 ) o ≃ ̄
Γ(𝑅)∕Γ(𝑅 0)
≃ / 𝐻̃ 1 (𝑀, 𝑅).
The left and right commutative squares are just by the construction.
While for 𝑞 ≥ 1 and 𝐻 𝑞 (𝑀, 𝑇 ) → 𝐻 𝑞+1 (𝑀, 𝑅), consider the commutative
diagram
𝑖𝑑
0 / 𝐻 𝑞 (𝑀, 𝐺) ≃ / 𝐻 𝑞+1 (𝑀, 𝑅) /0
O O
𝑖𝑑
𝐻 𝑞 (𝑀, 𝑇 ) / 𝐻 𝑞 (𝑀, 𝐺) o ≃
𝐻 𝑞 (𝑀, 𝑅)̄
≃ / 𝐻 𝑞+1 (𝑀, 𝑅).
𝐻 𝑞 (𝑀, 𝑇 ) / 𝐻 𝑞 (𝑀, 𝐺) o ≃
𝐻 𝑞 (𝑀, 𝑅)̄
≃ / 𝐻 𝑞+1 (𝑀, 𝑅)
𝐻̃ 𝑞 (𝑀, 𝑇 ) / 𝐻̃ 𝑞 (𝑀, 𝐺) o ≃
𝐻̃ 𝑞 (𝑀, 𝑅)̄
≃ / 𝐻̃ 𝑞+1 (𝑀, 𝑅).
38 Sheaf Cohomology
The first two squares have been proved to be commutative, and the third square is
commutative by the construction. This complete the proof. ◻
Now it’s convinced that there is a unique cohomology theory under isomor-
phism, given by the fine torsion free resolution of the constant sheaf, assuming the
existence of such resolution. That’s 𝐻 𝑞 (𝑀, 𝑆) ≃ 𝐻 𝑞 (Γ(𝐶 ∗ ⊗ 𝑆)), where 𝐶 ∗ is the
fine torsion free resolution for the constant sheaf.
Theorem 2.6. For the fine resolution of S ( not the constant sheaf)
0 → 𝑆 → 𝐶0 → 𝐶1 → ⋯ ,
0 → 𝑆 → 𝐶 0 → 𝐻 1 → 0,
and
0 → 𝐻 𝑞 → 𝐶 𝑞 → 𝐻 𝑞+1 → 0,
Remark 2.7. All the above claims are based on the assumption of the existence of
fine torsionless resolution of the constant sheaf on M. There will be an example of
the resolution in lemma 3.1 to introduce Alexander–Spanier cohomology.
3 Examples
Alexander–Spanier cohomology
Assume 𝑀 is a paracompact compact Hausdorff space here. For open set 𝑈 ⊂ 𝑀,
use 𝐴𝑝 (𝑈 , 𝐾) to denote the 𝐾 modules of pointwise functions 𝑈 𝑝+1 → 𝐾 for each
𝑝 ≥ 0. The coboundary operator 𝑑 ∶ 𝐴𝑝 (𝑈 , 𝐾) → 𝐴𝑝+1 (𝑈 , 𝐾) is defined by
𝑝+1
𝑑𝑓 (𝑥0 , 𝑥1 , ⋯ , 𝑥𝑝+1 ) = ∑ (−1)𝑖 𝑓 (𝑥0 , ⋯ , 𝑥𝑖̂ , ⋯ , 𝑥𝑝+1 )
𝑖=0
3 Examples 39
0 /𝒦 / 𝒜 1 (𝑀, 𝐾) 𝑑 / 𝒜 2 (𝑀, 𝐾) 𝑑 /⋯
0 /𝐾 / 𝐴1 (𝑈 , 𝐾) 𝑑 / 𝐴2 (𝑈 , 𝐾) 𝑑 / ⋯.
Notice that 𝐴𝑝 (𝑀, 𝐾) is not a complete sheaf,and 𝐴𝑝 (𝑀, 𝐺) as well. The operator
𝑝 𝑝+1
d on 𝐴0 (𝑀, 𝐺) has image in 𝐴0 (𝑀, 𝐺), so it induces a map
𝑝 𝑝+1
𝐴𝑝 (𝑀, 𝐺)∕𝐴0 (𝑀, 𝐺) → 𝐴𝑝+1 (𝑀, 𝐺)∕𝐴0 (𝑀, 𝐺),
and induces a cochain complex denoted by 𝐴∗ (𝑀, 𝐺)∕𝐴∗0 (𝑀, 𝐺). The 𝑝-th
Alexander–Spanier cohomology modules for 𝑀 with coefficients in 𝐺 is the 𝑝-th
quotient module with respect to this sequence, or
𝑝
𝐻𝐴−𝑆 (𝑀; 𝐺) = 𝐻 𝑝 (𝐴∗ (𝑀, 𝐺)∕𝐴∗0 (𝑀, 𝐺)).
Lemma 3.2. If 𝑃 = {𝑃𝑈 , 𝜌𝑉 ,𝑈 } is a presheaf and 𝑆 = 𝛽(𝑃 ), set 𝑆0 = {𝑠 ∈
𝑃𝑀 , 𝜌𝑥,𝑀 𝑠 = 0, ∀𝑥 ∈ 𝑀}, then there is an exact sequence
0 → 𝑆0 → 𝑃𝑀 → Γ(𝑆) → 0,
where the morphism 𝑃𝑀 → Γ(𝑆) send an element 𝑠 ∈ 𝑃𝑀 to a map 𝑀 → 𝑆,
which takes value 𝜌𝑥,𝑀 𝑠 at 𝑥.
By theorem 2.6, 𝐻 𝑞 (𝑀, 𝒢 ) ≃ 𝐻 𝑞 (Γ(𝒜 ∗ (𝑀, 𝐺))), and 𝐴∗ (𝑀, 𝐺)∕𝐴∗0 (𝑀, 𝐺) ≃
𝑝
Γ(𝒜 ∗ (𝑀, 𝐺)), so 𝐻 𝑞 (𝑀, 𝒢 ) ≃ 𝐻𝐴−𝑆 (𝑀; 𝐺), a desired conclusion.
de Rham cohomology
De Rham cohomology can also be viewed as a sheaf cohomology theory. Set 𝐾 =
𝑅, and ℛ = 𝑀 × 𝑅 is the constant sheaf.
{𝐸 𝑝 (𝑈 ), 𝜌𝑉 ,𝑈 } forms a presheaf, where 𝐸 𝑝 (𝑈 ) is the set of differential 𝑝-forms
on 𝑈 . The exterior differential acts as the coboundary operator. The associated
sheaf can be denoted by ℰ 𝑝 (𝑀), and 𝑑 represents the induced coboundary operator.
There is a fine torsionless resolution of ℛ as
since Poincare lemma shows locally closed forms are exact forms. Define the co-
homology theory for 𝑀 with ceofficients in sheaves of real vector spaces by setting
𝐻 𝑞 (𝑀, 𝒥 ) = 𝐻 𝑞 (Γ(ℰ ∗ (𝑀) ⊗ 𝒥 )).
If 𝒥 = ℛ, then
𝐻 𝑞 (𝑀, ℛ) = 𝐻 𝑞 (Γ(ℰ ∗ (𝑀) ⊗ ℛ)) ≃ 𝐻 𝑞 (Γ(ℰ ∗ (𝑀))).
and the isomorphism is commutative with the coboundary operator.
Define the 𝑞-th de Rham cohomology group by 𝐻 𝑞 (𝐸 ∗ (𝑀)) as usual, then
notice that {𝐸 𝑝 (𝑈 ), 𝜌𝑉 ,𝑈 } is complete, so 𝐸 𝑝 (𝑀) ≃ Γ(ℰ 𝑝 (𝑀)) and it induces an
isomorphism between the cochain complexs Γ(ℰ ∗ (𝑀)) and 𝐸 ∗ (𝑀), so
𝑞
𝐻 𝑞 (𝑀, ℛ) ≃ 𝐻𝑑𝑒𝑅 (𝑀),
a desired conclusion.
4 Multiplicative structure 41
4 Multiplicative structure
Definition 4.1. For cochain complexes 𝐶 ∗ and 𝐶 ̂∗ , define their tensor by
(𝐶 ∗ ⊗ 𝐶 ∗̂ )𝑟 = 𝐶 𝑝 ⊗ 𝐶 ̂𝑞 ,
⨁
𝑝+𝑞=𝑟
𝐻 𝑝 (𝐶 ∗ ) ⊗ 𝐻 𝑞 (𝐶 ∗̂ ) → 𝐻 𝑝+𝑞 (𝐶 ∗ ⊗ 𝐶 ∗̂ ).
0 → 𝒦 → 𝐶0 → 𝐶 1 → ⋯ ,
(𝑑𝑝 ⊗ 𝑖𝑑 𝑝 + (−1)𝑝 𝑖𝑑 𝑝 ⊗ 𝑑𝑞 ),
⨁
𝑝+𝑞=𝑟
This is a fine torsionless resolution of the constant sheaf 𝒦 . The sequence is fine
since tensor products and direct sums preserve fineness. Lemma 4.2 shows that
each term is torsionless. To show the exactness at 𝑟-th place(𝑟 ≥ 1), apply the
Kunneth formula at each stalk. The conditions of Kunneth formula is satisfied
automatically since 𝐶 ′∗ is torsionless, implying that the tensor product with it keeps
exactness. The exactness at 𝒦 and 𝐶0 ⊗ 𝐶0 can be checked by computation.
Theorem 4.3 (Kunneth formula). If there are two cochain complexes 𝐶 ∗ and 𝐶 ′∗
such that 𝐶 ∗ ∗ 𝐶 ′∗ is an acyclic (i.e. exact) cochain complex, there is a functional
split short exact sequence
0 → (𝐻 ∗ (𝐶 ∗ ) ⊗ 𝐻 ∗ (𝐶 ′∗ ))𝑞 → 𝐻 𝑞 (𝐶 ∗ ⊗ 𝐶 ′∗ ) → (𝐻 ∗ (𝐶 ∗ ) ∗ 𝐻 ∗ (𝐶 ′∗ ))𝑞+1 → 0.
Superpose it through the pair (𝑝, 𝑞) such that 𝑝 + 𝑞 = 𝑟 and get a cochain map
and a homomorphism
0 → 𝒦 → 𝐶0 ⊗ 𝐶0̃ → ⋯ → 𝐶𝑝 ⊗ 𝐶𝑞̃ → ⋯ ,
⨁
𝑝+𝑞=𝑟
which will play the role of the bridge between two tensor products 𝐶 ∗ ⊗ 𝐶 ∗ and
𝐶 ∗̃ ⊗ 𝐶 ∗̃ .
Consider the injection
𝐶𝑝 ≃ 𝐶𝑝 ⊗ 𝒦 → 𝐶𝑝 ⊗ 𝐶0̃ → ∑ 𝐶𝑟 ⊗ 𝐶𝑠̃ ,
𝑟+𝑠=𝑝
𝐶𝑝 / 𝐶𝑝 ⊗ 𝒦 / 𝐶𝑝 ⊗ 𝐶0̃ /∑ ̃
𝑟+𝑠=𝑝 𝐶𝑟 ⊗ 𝐶𝑠
𝐶𝑝 / 𝐶 ⊗𝒦 / 𝐶 ⊗ 𝐶̃ /∑ ̃
𝑝+1 𝑝+1 0 𝑟+𝑠=𝑝+1 𝐶𝑟 ⊗ 𝐶𝑠 .
Notice that the arrow 𝐶𝑝 ⊗ 𝐶0̃ → 𝐶𝑝+1 ⊗ 𝐶0̃ can’t be added, since for element
𝑐𝑝 ⊗ 𝑐0̃ , it maps to 𝑑𝑝 (𝑐𝑝 ) ⊗ 𝑐0̃ + (−1)𝑝 𝑐𝑝 ⊗ 𝑑0̃ (𝑐0̃ ) = 𝑑𝑝 (𝑐𝑝 ) ⊗ 𝑐0̃ , if and only if 𝑐0̃
is the image from 𝒦 , suppose 𝑐𝑝 ≠ 0.
The diagram induce the cochain map Γ(𝐶 ∗ ⊗ 𝑆) → Γ(𝐶 ∗ ⊗ 𝐶 ∗̃ ⊗ 𝑆), and will
induce the identity map 𝐻 𝑞 (𝑀, 𝑆) → 𝐻 𝑞 (𝑀, 𝑆), since this holds for 𝑞 = 0.
References 43
Γ(𝐶 ∗ ⊗ 𝐶 ∗̃ ⊗ 𝑆) ⊗ Γ(𝐶 ∗ ⊗ 𝐶 ∗̃ ⊗ 𝑆) / Γ(𝐶 ∗ ⊗ 𝐶 ∗̃ ⊗ 𝐶 ∗ ⊗ 𝐶 ∗̃ ⊗ 𝑆),
By the two diagrams, the multiplication structure is independent of the fine torsion
free resolution [Spa66].
References
[AM69] M. F. Atiyah and I. G. Macdonald (1969). Introduction to Commutative Algebra.
Addison-Wesley Publishing Company.
[GH94] P. A. Griffiths and J. D. Harris (1994). Principles of Algebraic Geometry. Wiley
Classics Library.
[Spa66] E. H. Spanier (1966). Algebraic Topology. Mcgraw-Hill Book Company.
[War83] F. W. Warner (1983). Foundations of Differentiable manifolds and Lie Groups.
Springer-Verlag.
An Introduction to Factorization Algebras and
Factorization Homology
Kong Fanhao1
ABSTRACT
We introduce the theory of factorization algebras and factorization homol-
ogy. They are based on the theory on higher operads and algebras over them.
We also give two applications of factorization algebras and factorization ho-
mology: the Deligne conjecture and the Bar constructions.
Contents
2 Factorization algebras 58
Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3 Factorization homology 66
Homology theory for manifolds . . . . . . . . . . . . . . . . . . . . . . 66
Factorization homology . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Relation to factorization algebra and applications to 𝐸𝑛 -algebras . . . . 75
Stratified spaces and applications to 𝐸𝑛 -modules . . . . . . . . . . . . . 80
4 Further applications 84
Centralizers and (higher) Deligne conjecture . . . . . . . . . . . . . . . 84
An introduction to the Bar constructions . . . . . . . . . . . . . . . . . 88
1
孔繁浩,清华大学数学系数 72 班.
46 An Introduction to Factorization Algebras and Factorization Homology
It has been many years since people try to study quantum field theories using the
tool of mathematics. People introduce a number of physical ideas into mathemat-
ics, including fields, phase spaces, quantization, Feynman diagrams, observables,
etc. Roughly speaking, we know that observables on small sets induce observables
on large sets. The study of this property of observables yields a new structure,
called the “factorization algebra”. Since the start of the 21st century, where the
notion of factorization algebras was introduced, it has been studied by a number of
mathematicians, and many notable results have been developed.
On the other hand, there exists another mathematical structure which is closely
related to the factorization algebra, called “factorization homology”. It is a struc-
ture to describe homology theories specific to manifolds of a specific dimension,
with some specific structures. The specific structures can be smooth structures,
framings, or simply no structure at all. They are introduced by axioms similar
to the Eilenberg-Steenrod axioms for homology theory for spaces, including the
monoidal axiom and the excision axiom.
In order to examine factorization homology and factorization algebras, a
number of preliminaries are required; they include the theory of operads and 𝐸𝑛 -
algebras. The theory of operads is used to study the structures of multimorphisms
and the relations that the multimorphisms must satisfy, while the 𝐸𝑛 -algebras are
introduced to study iterated loop spaces and configuration spaces, and have been
heavily studied in algebraic topology ever since the seventies.
It has been proved that factorization homology and factorization algebras are
highly related. This leads to more studies on 𝐸𝑛 -algebras, together with related
concepts in algebraic topology related to quantum mechanics, including deforma-
tions of 𝐸𝑛 -algebras, the (higher) Deligne conjecture, (higher) string topology and
Bar constructions of iterated loop spaces. It is also related to Hoschchild homology
and cohomology, which arise as specific examples of factorization homology.
In this article, we will present the preliminaries we need for defining factoriza-
tion algebras and factorization homology in Section 1. In Section 2, we will give
the definition of factorization algebras, and present some basic operations, such
as pullbacks, pushforwards, extending from a basis and gluing. The definition of
homology theories for manifolds, factorization homology, together with their prop-
erties, relationships with factorization algebras and basic examples, will be given
in Section 3. In the final section, two applications will be introduced, including the
Bar constructions and the Deligne conjecture.
• The definitions and basic properties of model categories are from [Hov99]
as well. Note that the definitions are slightly different from other references.
• For any commutative ring with unit 𝑅, we use Ch(𝑅) to denote the (ordinary)
category of chain complexes over 𝑅, and we use ℂ𝕙(𝑅) to denote the ∞-
category of chain complexes over 𝑅. For a minimal supply to ∞-category of
chain complexes, the first chapter of [Lur16] may be helpful.
Basic definitions
Operads are first introduced by Peter May in his book [May72]. We first introduce
its definition and give some evident generalizations.
i) The unital identity: The following maps are the identity maps:
≅ 𝑒⊗𝟙 𝑐
𝐹 (𝑛) −
→ 𝟙 ⊗ 𝐹 (𝑛) −−−→ 𝐹 (1) ⊗ 𝐹 (𝑛) →
− 𝐹 (𝑛),
≅ 𝟙⊗𝑒⊗𝑛 𝑐
→ 𝐹 (𝑛) ⊗ 𝟙⊗𝑛 −−−−−→ 𝐹 (𝑛) ⊗ 𝐹 (1)⊗𝑛 →
𝐹 (𝑛) − − 𝐹 (𝑛);
𝑘 𝑘 𝑛𝑖 𝑘 𝑛
𝐹 (𝑘) ⊗ ⨂ 𝐹 (𝑛𝑖 ) ⊗ ⨂ ⨂ 𝐹 (𝑚𝑖𝑗 )
≅
/ 𝐹 (𝑘) ⊗ ⨂ 𝐹 (𝑛𝑖 ) ⊗ ⨂𝑖 𝐹 (𝑚𝑖𝑗 )
(𝑖=1 ) (𝑖=1 𝑗=1 ) 𝑖=1 ( 𝑗=1 )
𝑘
𝟙⊗ ⨂ 𝑐
𝑖=1
𝑘 𝑛𝑖 𝑘 𝑛𝑖
𝑐⊗ ⨂ ⨂ 𝟙
(𝑖=1 𝑗=1 )
𝐹 (𝑘) ⊗ ⨂ 𝐹 ∑ 𝑚𝑖𝑗
𝑖=1 (𝑗=1 )
𝑐
𝑘 𝑘 𝑛𝑖 𝑘 𝑛
𝐹 ∑ 𝑛𝑖 ⊗ ⨂ ⨂ 𝐹 (𝑚𝑖𝑗 )
𝑐
/ 𝐹 ∑ ∑𝑖 𝑚𝑖𝑗 ;
(𝑖=1 ) (𝑖=1 𝑗=1 ) (𝑖=1 𝑗=1 )
𝑘 𝑘
𝐹 (𝑘) ⊗ ⨂ 𝐹 (𝑛𝑖 )
≅ / 𝐹 (𝑘) ⊗ ⨂ 𝐹 (𝑛 −1 )
𝜎 (𝑖)
𝑖=1 𝑖=1
𝜌𝜎 ⊗𝟙 𝑐
𝑘
𝐹 (𝑘) ⊗ ⨂ 𝐹 (𝑛𝑖 )
𝑐 / 𝐹 (𝑛 + ⋯ + 𝑛 )
1 𝑘
𝑖=1
𝑘
𝐹 (𝑘) ⊗ ⨂ 𝐹 (𝑛𝑖 )
𝑐 / 𝐹 (𝑛 + ⋯ + 𝑛 )
1 𝑘
𝑖=1
𝑘
𝟙 ⊗ ⨂ 𝜌 𝜎𝑖 𝜌(𝜎1 ,⋯,𝜎𝑟 )
𝑖=1
𝑘
𝐹 (𝑘) ⊗ ⨂ 𝐹 (𝑛𝑖 )
𝑐 / 𝐹 (𝑛 + ⋯ + 𝑛 )
1 𝑘
𝑖=1
for any 𝜎𝑖 ∈ 𝑆𝑛𝑖 , where (𝜎1 , ⋯ , 𝜎𝑟 ) denotes its image under the evident inclu-
sion 𝑆𝑛1 × ⋯ × 𝑆𝑛𝑘 → 𝑆𝑛1 +⋯+𝑛𝑘 .
𝟙𝑋 ∶ 𝟙C → Map(𝑋, 𝑋)
(regarded as an object in Map(𝑋, 𝑋)), and the 𝑆𝑛 -actions and compositions are
given in the obvious way. This operad indeed describes multimorphisms on the
object 𝑋. We denote this operad by End(𝑋), and call it the endomorphism operad
of 𝑋.
Definition 1.3. Suppose 𝐶 is a set, called the set of colors. A colored operad with
values taken in a symmetric monoidal category C consists of:
for 𝜎 ∈ 𝑆𝑛 and 𝑐1 , ⋯ , 𝑐𝑛 , 𝑐 ∈ 𝐶;
iv) A composition law
𝑘
Hom(𝑐1 , ⋯ , 𝑐𝑘 ; 𝑐) ⊗ Hom(𝑐𝑖1 , ⋯ , 𝑐𝑖𝑛𝑖 ; 𝑐𝑖 ) → Hom(𝑐11 , ⋯ , 𝑐𝑘𝑛𝑘 ; 𝑐);
⨂
𝑖=1
such that the composition is associative, unital and compatible with the 𝑆𝑛 -action.
If C = SSet, we call such a colored operad with values in SSet a simplicial
colored operad.
From this definition, we see that an operad is a special case of a colored operad:
it is a colored operad with only one color. In this case, we have
𝐹 (𝑛) = Hom(∗,⏟
⋯ , ∗; ∗).
𝑛 copies
For this reason, unless focusing on specific examples, we aim at colored operads
instead of single-colored operads.
The endomorphism operad can also be generalized to the colored case, as fol-
lows:
Example 1.4. Suppose C is a symmetric monoidal category, D is a symmetric
monoidal C-enriched category, 𝑋 = {𝑋𝑐 }𝑐∈𝐶 is a set of objects in D, where 𝐶 is
the set of colors. Consider the colored operad given by
where the identity maps, 𝑆𝑛 -actions and compositions are given in the obvious way.
We denote this operad by End(𝑋), and call it the endomorphism operad of 𝑋. If
𝑋 is the set of all objects in D, we also write it as End(D).
Next we discuss morphisms between (colored) operads.
Definition 1.5. Suppose O1 , O2 are colored operads, with colorings 𝐶1 , 𝐶2 , respec-
tively. A map 𝐹 from O1 to O2 consists of a map 𝐹 ∶ 𝐶1 → 𝐶2 of colors, and a
map
𝐹 ∶ Hom(𝑐1 , ⋯ , 𝑐𝑛 ; 𝑐) → Hom(𝐹 (𝑐1 ), ⋯ , 𝐹 (𝑐𝑛 ); 𝐹 (𝑐))
Hom(𝑐1 , ⋯ , 𝑐𝑛 ; 𝑐)
𝜃 / Hom(𝑋 , ⋯ , 𝑋 ; 𝑋 )
𝑐1 𝑐𝑛 𝑐
𝜉 𝑓∗
𝑓∗
Hom(𝑌𝑐1 , ⋯ , 𝑌𝑐𝑛 ; 𝑌𝑐 ) / Hom(𝑋𝑐 , ⋯ , 𝑋𝑐 ; 𝑌𝑐 ).
1 𝑛
i) Comm(𝑛) = 𝟙C ;
ii) The unit map and composition map is the identity map of the unit object;
Higher operads
In practice we often have to change the underlying category of an operad. For
example, we sometimes face the problem of changing the underlying category of
an operad O from the category of topological spaces to some other category (for
example Ch(𝑅) for a ring 𝑅). If there exists a symmetric monoidal functor between
the underlying categories (in the case Ch(𝑅) we have the singular chain complex
functor), then we may take the pushforward of O along the symmetric monoidal
functor; however this is not always the case. In this section, we will introduce the
notion of a higher operad, which solves this problem nicely for topological operads.
To do this, we first review how we get a symmetric monoidal ∞-category from
an ordinal symmetric monoidal category.
𝑓 ∶ [𝐶1 , ⋯ , 𝐶𝑛 ] → [𝐷1 , ⋯ , 𝐷𝑚 ]
1 Preliminaries: Operads, higher operads and algebras over them 53
𝑓𝑖 ∶ 𝐶𝑗 → 𝐷 𝑖
⨂
𝑗∈𝑓 −1 (𝑖)
for any 1 ≤ 𝑖 ≤ 𝑚;
iii) Suppose
𝑔 ∘ (𝑓 |𝑆 ′ ) ∶ 𝑆 ′ → {1, ⋯ , 𝑙},
(𝑔 ∘ 𝑓 )𝑖 ∶ 𝐶𝑘 ≅ 𝐶𝑘 → 𝐷𝑗 → 𝐸𝑖
⨂ ⨂ ⨂ ⨂
𝑘∈(𝑔∘(𝑓 |𝑆 ′ ))−1 (𝑖) 𝑗∈𝑔 −1 (𝑖) 𝑘∈𝑓 −1 (𝑗) 𝑗∈𝑔 −1 (𝑖)
for every 1 ≤ 𝑖 ≤ 𝑙.
Next, take the category Fin∗ , the category of pointed finite sets, where the ob-
jects are ⟨𝑛⟩ ∶= {∗, 1, ⋯ , 𝑛} for all nonnegative integers 𝑛, and the morphisms are
all maps that preserve the base point. Then we have a forgetful functor 𝑝 ∶ C⊗ →
Fin∗ , that carries an object [𝐶1 , ⋯ , 𝐶𝑛 ] to the set ⟨𝑛⟩. Moreover, 𝑝 satisfies the
following properties:
Proposition 1.11. i) 𝑝 is a coCartesian fibration, which means for any
lifting 𝛼 that is coCartesian in the following sense: for any 𝐸 ∈ C⊗ the map
Construction 1.13. For any colored operad O over Set, we may define a category
O⊗ as follows:
𝑓 ∶ [𝐶1 , ⋯ , 𝐶𝑛 ] → [𝐷1 , ⋯ , 𝐷𝑚 ]
consists of a map 𝑓 ∶ ⟨𝑛⟩ → ⟨𝑚⟩, and a map 𝑓𝑖 ∈ Hom((𝐶𝑗 )𝑗∈𝑓 −1 (𝑖) ; 𝐷𝑖 ) for
any 1 ≤ 𝑖 ≤ 𝑚;
i) For any inert morphism 𝑓 ∶ ⟨𝑛⟩ → ⟨𝑚⟩, which means that 𝑓 −1 ({1, ⋯ , 𝑚}) →
{1, ⋯ , 𝑚} is an isomorphism, and any 𝐶 ∈ O⊗
⟨\⟩
, there exists a 𝑝-coCartesian
morphism 𝐶 → 𝐷 ∈ O⊗ lifting 𝑓 ;
is an isomorphism in H.
Remark 1.15. i) By this definition, for any Set-valued colored operad O, N(O⊗ )
is an ∞-operad. Thus this definition generalizes the definition of a (Set-valued
colored) operad to the higher categorial sense.
ii) If O⊗ is an ∞-operad, we define the underlying category of O⊗ to be O ∶=
O⊗⟨1⟩
. This is an ∞-category, since 𝑝 must be an inner fibration by [Lur09],
Proposition 2.3.1.5.
iii) By definition, for every 𝑛 ≥ 0 the map O⊗ ⟨𝑛⟩
→ O𝑛 , induced by the 𝑛 maps
𝜌𝑖 ∶ ⟨𝑛⟩ → ⟨1⟩ for every 1 ≤ 𝑖 ≤ 𝑛, is an equivalence of ∞-categories. Thus,
as in the ordinary case, we may regard objects in O⊗⟨𝑛⟩
as finite sequences of
objects in O of length 𝑛.
iv) Any symmetric monoidal ∞-category C⊗ is an ∞-operad. This is because
the condition that 𝑝 is a coCartesian fibration gaurantees i) and ii), and the
equivalence between C⊗ ⟨𝑛⟩
and C𝑛 gaurantees iii).
We now show how these definitions solve the problem mentioned in the begin-
ning of the section. We first notice that the construction of O⊗ also makes sense
in the case of a simplicial colored operad, except that we need to change ii) of
Construction 1.13 into the following construction:
Proposition 1.18. Suppose all the Hom-spaces of O are Kan complexes. Then
N⊗ (O) is an ∞-operad.
i) For any 𝑟 ≥ 0, the morphism space E𝑛 (𝑟) is the configuration space of 𝑟 ordered
disjoint open rectangles whose edges are parallel to the coordinate axes (each
called a little 𝒏-cube) lying in the unit open cube; in other words, E𝑛 (𝑟) is
the space of rectilinear embeddings from ∐𝑟𝑖=1 (0, 1)𝑛 to (0, 1)𝑛 , which will be
denoted by
𝑟
Rect ∐(0, 1)𝑛 , (0, 1)𝑛 ;
( 𝑖=1 )
Example 1.23. Iterated loop spaces. This is one of the first examples of 𝐸𝑛 -
algebras that have been studied. Suppose (𝑋, ∗) is a pointed topological space.
We define a map
𝜃 ∶ E𝑛 (𝑟) × (Ω𝑛 𝑋)𝑟 → Ω𝑛 𝑋
2 Factorization algebras
Factorization algebras are a way to describe multilinear structures that satisfy cer-
tain homotopy coherent properties. It is a bit similar to an algebra over the little
cubes operad, but this time the operad is replaced by a colored operad, with the
cubes (0, 1)𝑛 replaces by manifolds. We shall now present the definitions and the
basic properties, which are mostly based on [CG16] and [Gin13].
Basic definitions
Definition 2.1. Suppose 𝑋 is a topological space (which in the case of a factoriza-
tion algebra is often taken to be a manifold). Define the colored operad Fact(𝑋) to
be the following operad:
iii) The unit morphism, 𝑆𝑛 -action, and the composition are given in the obvious
way.
to be the map
𝐴⊗𝑛 → 𝐴, (𝑥1 , ⋯ , 𝑥𝑛 ) ↦ 𝑥1 ⋯ 𝑥𝑛 .
𝐴(𝛼1 , ⋯ , 𝛼𝑛 ) = 𝐴 ∐ 𝑈 𝑖 ∩ ⋯ ∩ 𝑈 𝑖𝑛 ,
(𝑖1 ∈𝛼1 ,⋯,𝑖𝑛 ∈𝛼𝑛 1 )
𝑠𝑖 ∶ 𝐴(𝛼1 , ⋯ , 𝛼𝑛 ) → 𝐴(𝛼1 , ⋯ , 𝛼𝑖 , 𝛼𝑖 , ⋯ , 𝛼𝑛 )
for every integer 1 ≤ 𝑖 ≤ 𝑛. Together this defines a simplicial object 𝐶•̌ (𝐼, 𝐴) in C,
where
𝐶𝑛̌ (𝐼, 𝐴) = ∐ 𝐴(𝛼),
𝛼∈𝑃 𝐼 𝑛+1
and the face and degeneracy maps are given above. We call this the simplicial Čech
complex of 𝐼. Note that there exists a map of simplicial objects 𝐶•̌ (𝐼, 𝐴) → 𝐴(𝑈 ),
where we regard 𝐴(𝑈 ) as the constant simplicial object.
We now assume that C is a stable ∞-category, whose tensor product is com-
patible with colimits over simplicial diagrams (say, for example, ℂ𝕙(𝕜)). Then by
the ∞-categorical Dold-Kan correspondence and taking the direct colimit, we ob-
̌ 𝐴) from the simplicial Čech complex. We call this the Čech
tain an object 𝐶(𝐼,
complex of 𝐼. Moreover we have a canonical map 𝐶(𝐼, ̌ 𝐴) → 𝐴(𝑈 ).
We are now ready to state the definition of a factorization algebra:
Note that in the case that C = ℂ𝕙(𝕜) where 𝕜 is a field, the above definition is
equivalent to that the chain map 𝐶• (𝐼, 𝐴) → 𝐴(𝑈 ) being a quasi-isomorphism; a
morphism between factorization algebras is an equivalence if and only if it induces
a quasi-isomorphism on every open subset.
Naively speaking, homotopy factorization algebras are to prefactorization alge-
bras as homotopy cosheaves are to precosheaves, where we change the underlying
site to the category of open subsets of 𝑋 with open coverings being factorization
coverings.
In the case that 𝑋 is a manifold, there exists a notion of a locally constant
factorization algebra, which we will give as follows:
Remark 2.8. In Example 3.24, we will prove that a locally constant prefactor-
ization algebra can be reformed to a factorization algebra, using factorization ho-
mology. We denote the full subcategory of FA(𝑋, C) spanned by locally constant
factorization algebras by FA𝑙𝑐 (𝑋, C).
2 Factorization algebras 61
Basic properties
We now discuss basic operations and properties of factorization algebras. Most of
them will be useful in the next section.
Firstly, we have the following proposition.
Proposition 2.9. The categories PreFA(𝑋, C), FA(𝑋, C), FA𝑙𝑐 (𝑋, C) are symmetric
monoidal ∞-categories.
𝑓∗ (𝐴)(𝑉 ) = 𝐴(𝑓 −1 (𝑉 ))
𝑓 −1 𝐴
Fact(𝑌 ) − → C⊗ ;
−−→ Fact(𝑋) −
The proof to this is really technical, so we refer the interested readers to Ap-
pendix B, [CG16]. We note that the uniqueness directly follows from the definition
of a factorization algebra.
This proposition tells us that a factorization algebra is “in some sense” defined
by its values on a few of its open subsets. This proposition is in particular very
useful in the case of a smooth manifold, since open disks form a factorizing basis,
and the value of a factorization algebra on a disk is usually easy to determine.
Next, we discuss what happens if we take the product of two spaces. Suppose
𝑋, 𝑌 are topological spaces and 𝜋 ∶ 𝑋 × 𝑌 → 𝑋 is the projection. By Proposition
2.10, 𝜋 induces a functor 𝜋∗ ∶ PreFA(𝑋 × 𝑌 ) → PreFA(𝑋). We make a stronger
claim:
Proposition 2.13. There exists a canonical functor
and
𝜋 ∗ ∶ FA𝑙𝑐 (𝑋 × 𝑌 ) → FA𝑙𝑐 (𝑋, FA𝑙𝑐 (𝑌 ))
≅
𝒰𝑓 = {𝑈 ⊆ 𝑋 ∣ 𝑈 open, 𝑓 ∶ 𝑈 −
→ 𝑓 (𝑈 )}.
𝒰 ∶= {𝑈 ⊆ 𝑋 ∣ 𝑈 ∈ 𝒰𝑓 , 𝑈 ∈ 𝒰𝑔∘𝑓 , 𝑓 (𝑈 ) ∈ 𝒰𝑔 }
are equal, since 𝒰 is a factorizing basis, (𝑔 ∘ 𝑓 )∗ (𝐶) and 𝑓 ∗ 𝑔 ∗ (𝐶) are equivalent.
Since the above construction is functorial, the functors (𝑔 ∘ 𝑓 )∗ and 𝑓 ∗ 𝑔 ∗ (𝐶) are
equivalent.
Remark 2.15. We note that the functors 𝑓∗ and 𝑓 ∗ are not adjoint functors. To
see this, take 𝑋 = {𝑐, 𝑑} to be the discrete space with two points, and 𝑌 = ∗. A
factorization algebra 𝐵 on 𝑌 in C is an object 𝐵 ∈ C and a map 𝑏 ∶ 𝟙 → 𝐵 in
C; while a factorization algebra 𝐴 on 𝑋 in C consists of two objects 𝐶, 𝐷 ∈ C,
two maps 𝑐 ∶ 𝟙 → 𝐶, 𝑑 ∶ 𝟙 → 𝐷 in C, corresponding to the objects 𝐴(𝑐) and
𝐴(𝑑). By direct computation, the factorization algebra 𝑓∗ (𝑓 ∗ (𝐵)) is 𝐵 ⊗ 𝐵, and
the factorization algebra 𝑓 ∗ (𝑓∗ (𝐴)) is 𝐴 ⊗ 𝐴. But there does not exist a natural
map 𝐴 ⊗ 𝐴 → 𝐴, nor does there exist a natural map 𝐵 ⊗ 𝐵 → 𝐵, thus 𝑓∗ and 𝑓 ∗
are not adjoint functors.
Nevertheless, if 𝑓 ∶ 𝑋 → 𝑌 is an open immersion of smooth manifolds, 𝐴 is a
factorization algebra on 𝐴, then for any 𝑈 ∈ 𝒰𝑓 there exists a map
induced by 𝑈 ↪𝑓 −1 (𝑓 (𝑈 ))
𝐴(𝑈 ) −−−−−−−−−−−−−−−−−→ 𝐴(𝑓 −1 (𝑓 (𝑈 ))) ≅ (𝑓 ∗ (𝑓∗ (𝐴)))(𝑈 ),
𝑟𝐽 ,𝑘 𝑟𝐽 −{𝑗},𝑘
𝑟𝐽 −{𝑘},𝑗
(𝐴𝐽 −{𝑘} )|𝑉𝐽 / 𝐴𝐽 −{𝑗,𝑘} .
3 Factorization homology
Factorization homology is a type of homology theory on manifolds. They produce
invariants of manifolds, just as classical homology theories produce invariants of
topological spaces. In some specific manifolds, factorization homology reduces to
some well-known structures, such as Hoschchild homology. In this section, we will
introduce homology theories on manifolds and factorization homology, and study
their properties, as well as their relationships with factorization algebras, especially
locally constant factorization algebras.
Mfld(𝑋,𝑒)
𝑛
∶= Mfld𝑛 × S∕𝑋 ,
S∕𝐵 Aut(ℝ𝑛 )
ii) If 𝑋 = 𝐵𝑂(𝑛), and 𝑒 is the canonical map 𝐵𝑂(𝑛) → 𝐵 Aut(ℝ𝑛 ), then Mfld(𝑋,𝑒) 𝑛
consists of smooth manifolds. This is essentially because the map from 𝑂(𝑛)
to the topological group of all diffeomorphisms on ℝ𝑛 and the characterization
of smooth manifolds in terms of their microbundle structure, which is shown
in [KS77]. We define Mfldun 𝑛
to be Mfld(𝐵𝑂(𝑛),𝑒)
𝑛
, and call it the ∞-category of
smooth 𝒏-dimensional manifolds. A morphism in Mfldun 𝑛
is equivalent to a
smooth embedding.
Proof. This is by Proposition 1.18, since it is the nerve of a topological single col-
ored operad, with morphism space Disk(𝑋,𝑒)
𝑛
(𝑟) the subspace of Hom(∐𝑟 ℝ𝑛 , ℝ𝑛 )
consisting of all (𝑋, 𝑒)-structured embeddings ∐𝑟 ℝ𝑛 → ℝ𝑛 . ◻
Thus, the ∞-category of algebras over the ∞-operad Diskfr𝑛 is equivalent to the
∞-category of 𝐸𝑛 -algebras.
We are now able to present what a homology theory for (𝑛-dimensional) man-
ifolds is. We interprete this in a way similar to the Eilenberg–Steenrod axioms.
First, we consider the excision axiom.
Proof. i) Notice that for any finite sets 𝐼, 𝐽 , any framed embedding ∐𝐼 ℝ𝑠 →
∐𝐽 ℝ𝑠 induces a (𝑋, 𝑒)-embedding ∐𝐼 (𝑁 × ℝ𝑠 ) → ∐𝐽 (𝑁 × ℝ𝑠 ), thus it in-
𝑁
duces a map 𝛾𝐼,𝐽 from the space of framed embeddings from ∐𝐼 ℝ𝑠 to ∐𝐽 ℝ𝑠
to the space of (𝑋, 𝑒)-embeddings from ∐𝐼 (𝑁 × ℝ𝑠 ) to ∐𝐽 (𝑁 × ℝ𝑠 ). There-
fore, 𝑁 × ℝ𝑠 is an 𝐸𝑠 -algebra object in the ∞-category Mfld(𝑋,𝑒)
𝑛
. Since 𝐻 is
symmetric monoidal, 𝐻(𝑁 × ℝ𝑠 ) is an 𝐸𝑠 -algebra in C.
where Embfr0 is the subspace of Embfr consisting of maps that map 0 in the last
component to 0. Using i), the map is given by
With this proposition, we are able to express the excision axiom for homology
theories for manifolds, which is again due to [Gin13]:
3
We assume once and for all that all modules are pointed.
70 An Introduction to Factorization Algebras and Factorization Homology
Definition 3.8. A homology theory for (𝑋, 𝑒)-manifolds with values in a symmet-
ric monoidal stable ∞-catgeory C is a functor
𝐻 ∶ Mfld(𝑋,𝑒)
𝑛
→ C, 𝑀 ↦ 𝐻(𝑀),
such that 𝐻 is symmetric monoidal, preserves sequential colimits, and satisfies the
excision axiom:
• For any (𝑋, 𝑒)-manifold 𝑀, if there exists a codimension 1 submanifold 𝑁
of 𝑀 and a trivialization 𝑁 × ℝ of its neighborhood such that 𝑀 is decom-
posible as 𝑅⨿𝑁×ℝ 𝐿, where 𝑅, 𝐿 are submanifolds of 𝑀 glued along 𝑁 × ℝ,
then the natural map
is an equivalence.
The ∞-category of homology theories is defined to be the full subcategory of
Alg(Mfld(𝑋,𝑒)
𝑛
) spanned by homology theories, denoted HT(Mfld(𝑋,𝑒)
𝑛
).
Factorization homology
Now we will introduce a specific homology theory, called factorization homology.
We will discuss the properties of factorization homology, and we will prove the
following result, which is from [AF12]:
Theorem 3.9. For any Disk(𝑋,𝑒)
𝑛
-algebra 𝐴, there exists, up to contractible choices,
a unique homology theory for (𝑋, 𝑒)-manifolds, satisfying the dimension axiom:
• The value of the homology theory on ℝ𝑛 is naturally equivalent to 𝐴;
which is exactly the factorization homology with coeffecients in 𝐴 (whose definition
will be given below).
We start with the definition of the factorization homology.
Definition 3.10. Any Disk(𝑋,𝑒)
𝑛
-algebra 𝐴 defines a functor Disk(𝑋,𝑒)
𝑛
→ C. On the
other hand, if 𝑀 is in Mfld𝑛 , then 𝑀 defines a functor 𝐸𝑀 ∶ (Disk(𝑋,𝑒)
(𝑋,𝑒)
𝑛
)op →
(𝑋,𝑒)
K, which maps an object in Disk𝑛 to the space of (𝑋, 𝑒)-embeddings from that
object to 𝑀. Together 𝑀 and 𝐴 defines a functor:
⊗
𝐸𝑀 ⊗ 𝐴 ∶ (Disk(𝑋,𝑒)
𝑛
)𝑜𝑝 × Disk(𝑋,𝑒)
𝑛
→ K × C −→ C.
∫ 𝐴 ∶= 𝐸𝑀 ⊗ 𝐴.
(𝑋,𝑒)
𝑀 Disk𝑛
3 Factorization homology 71
Note that by the definition of a coend, factorization homology can also be in-
terpreted as
𝐴
∫ 𝐴 = colim((Disk(𝑋,𝑒)
𝑛
)∕𝑀 → Disk(𝑋,𝑒)
𝑛
→ C).
−
𝑀
𝑖∗ ∶ Alg(Mfld(𝑋,𝑒)
𝑛
, C) → Disk(𝑋,𝑒)
𝑛
-Alg(C).
to 𝑖∗ , which is given by
𝑖! (𝐴)(𝑀) = ∫ 𝐴.
𝑀
Proof. From the theory of ∞-categories, we know that there exists a functor
𝑖! ∶ Fun(Disk(𝑋,𝑒)
𝑛
, C) → Fun(Mfld(𝑋,𝑒)
𝑛
, C),
𝑖∗ ∶ Fun(Mfld(𝑋,𝑒)
𝑛
, C) → Fun(Disk(𝑋,𝑒)
𝑛
, C).
(𝑖! (𝐴))𝑛
(Mfld(𝑋,𝑒) )𝑛 / C𝑛
𝑛
𝑓∗ 𝑓∗
(𝑖! (𝐴))𝑚
(Mfld(𝑋,𝑒) )𝑚 / C𝑚 .
𝑛
Since every map 𝑓 can be factored to be the composition of a surjective active map
followed by an injective active map followed by an inert map (where an active map
72 An Introduction to Factorization Algebras and Factorization Homology
is a map such that the inverse image of ∗ is ∗), it suffices to prove this in these three
cases.
If 𝑓 is inert, then 𝑓∗ is a projection, the commutativity is obvious.
If 𝑓 is injective and active, since 𝐴 is symmetric monoidal, 𝐴 preserves the
monoidal unit, thus 𝑖! (𝐴) also preserves the monoidal unit, therefore 𝑓∗ is the
canonical inclusion, the commutativity follows.
If 𝑓 is surjective and active, by taking all fibers of the map 𝑓 , we may assume
that 𝑚 = 1. Then the map 𝑓∗ is the 𝑛-fold tensor product. For any (𝑀1 , ⋯ , 𝑀𝑛 ) ∈
(Mfld(𝑋,𝑒)
𝑛
)𝑛 , we consider the following diagram:
𝑛 ∐
∏(Disk(𝑋,𝑒) )∕𝑀𝑖 / (Disk(𝑋,𝑒) ) 𝑛
𝑛 𝑛
𝑖=1 ∕ ∐ 𝑀𝑖
𝑖=1
∐
(Disk(𝑋,𝑒) )𝑛 / Disk(𝑋,𝑒)
𝑛 𝑛
𝐴𝑛 𝐴
⨂
C𝑛 / C.
This diagram is commutative. Moreover the top row is fully faithful and essen-
tially surjective, which means that it is an equivalence. Therefore, we obtain the
following natural equivalence
⎛ ⎞
(𝑋,𝑒) 𝐴
colim ⎜(Disk(𝑋,𝑒)
𝑛
) 𝑛 → Disk𝑛
−
→ C⎟
⎜ ∕ ∐ 𝑀𝑖 ⎟
⎝ 𝑖=1 ⎠
𝑛
𝐴𝑛 ⨂
≅ colim ∏(Disk(𝑋,𝑒)
𝑛
)∕𝑀𝑖 → (Disk(𝑋,𝑒)
𝑛
)𝑛 −−→ C𝑛 −−→ C .
( 𝑖=1 )
On the other hand, since ⊗ and colim commute in C, there exists a natural equiva-
lence
𝑛
𝐴𝑛 ⨂
colim ∏(Disk(𝑋,𝑒)
𝑛
)∕𝑀𝑖 → (Disk(𝑋,𝑒)
𝑛
)𝑛 −−→ C𝑛 −−→ C
( 𝑖=1 )
𝑛
𝐴
≅ colim (Disk(𝑋,𝑒) )∕𝑀𝑖 → Disk(𝑋,𝑒) →C .
−
⨂ ( 𝑛 𝑛 )
𝑖=1
The composition of the two equivalences yields the commutativity of the diagram,
as desired. ◻
The above proposition shows that factorization homology can be expressed as
symmetric monoidal left Kan extension, at least when C is ⊗-presentable. In par-
ticular, factorization homology is symmetric monoidal on 𝑀.
Next we show the independence of factorization homology with respect to 𝑋.
3 Factorization homology 73
∫𝐴 ⊗ ∫𝐴 → ∫ 𝐴
𝐿 ∫𝑁×ℝ 𝐴 𝑅 𝑀
is an equivalence.
Proof. First, we write the definition of a factorization algebra in another way. The
objects in (Disk(𝑋,𝑒)
𝑛
)∕𝑀 are those open subsets in 𝑀, that is homeomorphic to a
disjoint union of disks. Then if we define Ball(𝑀) to be the collection of the open
subsets in 𝑀 that are homeomorphic to a disk, then
𝑙
∫ 𝐴= colim ∫ 𝐴.
𝑀 𝑈1 ,⋯,𝑈𝑙 ∈Ball(𝑀), pairwise disjoint ⨂ 𝑈
𝑖=1 𝑖
∫𝐴 ⊗ ∫𝐴
𝐿 ∫𝑁×ℝ 𝐴 𝑅
can be interpreted as
colim ∫ 𝐴 ⊗ ∫ 𝐴 ⊗ ∫ 𝐴 ⇉ ∫ 𝐴 ⊗ ∫ 𝐴 ,
( 𝐿 𝑁×ℝ 𝑅 𝐿 𝑅
)
74 An Introduction to Factorization Algebras and Factorization Homology
where the two maps are the two module structures. Define 𝐿0 and 𝑅0 to be 𝐿−𝑁 ×
(−∞, 0] and 𝑅 − 𝑁 × [0, ∞), respectively. Then we have canonical equivalences
∫ 𝐴 ≅ ∫ 𝐴, ∫ 𝐴 ≅ ∫ 𝐴.
𝐿 𝐿0 𝑅 𝑅0
Under this equivalence, using the definition of the module structure, the colimit
above is equivalent to
𝑙
colim ∫ 𝐴,
̃ pairwise disjoint ⨂ 𝑈
𝑈1 ,⋯,𝑈𝑙 ∈Ball, 𝑖=1 𝑖
∫ 𝐴 ≅ ∫𝐴 ⊗ ∫ 𝐴 = 𝐴 ⊗ 𝐴,
𝑀 ℝ ∫{1}×ℝ 𝐴⊗∫{−1}×ℝ 𝐴 ℝ 𝐴⊗𝐴𝑜𝑝
∫ ∶ Disk(𝑋,𝑒)
𝑛
-Alg(C) ⇄ HT(Mfld(𝑋,𝑒)
𝑛
, C) ∶ evℝ𝑛
between Disk(𝑋,𝑒)
𝑛
-algebras and homology theories for (𝑋, 𝑒)-manifolds.
3 Factorization homology 75
Proof. By Proposition 3.11, ∫ is a left adjoint to the functor evℝ𝑛 . The unit of the
adjunction is an equivalence, because Disk(𝑋,𝑒)
𝑛
→ Mfld(𝑋,𝑒)
𝑛
is fully faithful, and
the Kan extension along a fully faithful functor restricts as the original functor. The
counit of the adjuntion evaluates on a symmetric monoidal functor 𝐹 as a morphism
∫ 𝐹 (ℝ𝑛 ) → 𝐹 . It suffices to show that this map is an equivalence, which means
we only have to prove that the map ∫𝑀 𝐹 (ℝ𝑛 ) → 𝐹 (𝑀) is an equivalence for any
(𝑋, 𝑒)-manifold 𝑀. We denote 𝐴 ∶= 𝐹 (ℝ𝑛 ).
If 𝑀 is the disjoint union of some copies of ℝ𝑛 , this is obvious. Suppose that
it is true for 𝑀 = 𝑆 𝑖 × ℝ𝑛−𝑖 for some 0 ≤ 𝑖 < 𝑛. Then for the case 𝑆 𝑖+1 × ℝ𝑛−𝑖−1 ,
it can be decomposed as ℝ𝑖+1 − ×ℝ
𝑛−𝑖−1
∐𝑆 𝑖+1 ×ℝ𝑛−𝑖 ℝ𝑖+ × ℝ𝑛−𝑖−1 , since both 𝐹 and
∫− 𝐴 have the excision property, we have
∫ 𝐴≅∫ 𝐴 ⊗ ∫ 𝐴
𝑆 𝑖+1 ×ℝ𝑛−𝑖−1 ℝ𝑖+1
− ×ℝ
𝑛−𝑖−1 ∫𝑆 𝑖 ×ℝ𝑛−𝑖 𝐴 ℝ𝑖+1
+ ×ℝ
𝑛−𝑖−1
≅ 𝐹 (ℝ𝑖+1
− ×ℝ
𝑛−𝑖−1
) ⊗ 𝐹 (ℝ𝑖+1
+ ×ℝ
𝑛−𝑖−1
)
𝐹 (𝑆 𝑖 ×ℝ𝑛−𝑖 )
≅ 𝐹 (𝑆 𝑖+1 × ℝ𝑛−𝑖−1 ).
Thus the map is an equivalence for 𝑀 = 𝑆 𝑖+1 × ℝ𝑛−𝑖−1 . By induction, for any
0 ≤ 𝑖 ≤ 𝑛, the above map is an equivalence for 𝑀 = 𝑆 𝑖 × ℝ𝑛−𝑖 .
Now, for any manifold 𝑀 that is obtained from a manifold 𝑁 making the counit
map an equivalence by adding a handle of index 𝑖 + 1, 𝑀 can be expressed as the
coproduct 𝑁 ∐𝑆 𝑖 ×ℝ𝑛−𝑖 ℝ𝑛 . Similarly as above, since 𝑁, 𝑆 𝑖 × ℝ𝑛−𝑖 , ℝ𝑛 all make the
counit map an equivalence, so does 𝑀. Therefore, by induction, the counit map is
an equivalence for all handlebodies.
We finish the proof by noticing that all smooth manifolds can be regarded as the
direct colimit of inclusions of handlebodies, and the functors ∫ 𝐴 and 𝐹 preserve
sequential colimits. ◻
Proof. By the proof of Proposition 3.7, to prove that 𝐴(𝑅) and 𝐴(𝐿) are right
and left 𝐸1 -modules over 𝐴(𝑁 × ℝ), it suffices to show that the map 𝐴(𝑅 − 𝑁 ×
[𝑡, +∞)) → 𝐴(𝑅) is an equivalence for all 𝑡 ∈ ℝ. This is since by Proposition 2.10,
the map 𝐴(𝑁 × (𝑎, 𝑏)) → 𝐴(𝑁 × (𝑎′ , 𝑏′ )) is an equivalence for all 𝑎′ ≤ 𝑎 < 𝑏 ≤ 𝑏′ ,
thus the Čech complexes of the covers
and
{𝑅 − 𝑁 × [𝑡 − 1, +∞)} ∪ {𝑁 × (𝑎, 𝑏) ∣ 𝑡 − 2 ≤ 𝑎 < 𝑏 ≤ 𝑡}
Then we know that 𝐴(𝑀) is equivalent to the Čech complex 𝐶(𝒰, ̌ 𝐴). But by direct
computation, the Čech complex is given by the classical two-sided Bar construction
𝐵(𝐴(𝑅), 𝐴(𝑁 × ℝ), 𝐴(𝐿)), which is equivalent to the tensor product 𝐴(𝑅)⊗𝐴(𝑁×ℝ)
𝐴(𝐿) (by [CG16]). Thus 𝐴(𝑀) is equivalent to 𝐴(𝑅) ⊗𝐴(𝑁×ℝ) 𝐴(𝐿), completing
the proof. ◻
Proof. Since the inclusion of open sets of 𝑀 are (𝑋, 𝑒)-embeddings, the map 𝑈 ↦
∫𝑈 𝐴 is a prefactorization algebra. To prove that it is a factorization algebra, we
construct a factorization algebra that is locally constant, and show that it is naturally
equivalent to 𝑖(𝐴). We deduce this by a path similar to the proof of Theorem 3.9.
We begin with the case 𝑀 ≅ ℝ𝑛 . Then 𝐴 is a (Disk(𝑋,𝑒) 𝑛
)∕𝑀 -algebra. But by
(𝑋,𝑒) (𝑀,𝑇 𝑀)
definition (Disk𝑛 )∕𝑀 is naturally isomorphic to Disk𝑛 , thus if 𝑀 = ℝ𝑛 ,
𝑛 𝑛
then 𝐴 is naturally a Disk(𝑛ℝ ,𝑇 ℝ ) -algebra, which is equivalent to a Diskfr𝑛 -algebra.
3 Factorization homology 77
ii) For any locally constant factorization algebra 𝐵, 𝑖(𝑗(𝐵)) sends every subset 𝑈
of 𝑀 to ∫𝑈 𝐵(ℝ𝑛 ). Now, take a metric on 𝑀, and take 𝒰 to be the collection
of all strictly convex subset of 𝑀. Then 𝒰 is stable under intersection, forms
a basis of 𝑀, and is a factorization basis of 𝑀. Moreover the values of 𝑖(𝑗(𝐵))
and 𝐵 on 𝒰 are naturally equivalent, since ∫ℝ𝑛 𝐵(ℝ𝑛 ) and 𝐵(ℝ𝑛 ) are naturally
equivalent. By Proposition 2.12, 𝑖(𝑗(𝐵)) and 𝐵 are naturally equivalent.
The above discussion shows that 𝑗 is an inverse to 𝑖, which completes the proof.
◻
𝑝∗ (𝑖(𝐴))(∗) = 𝑖(𝐴)(𝑀) = ∫ 𝑀.
𝐴
If we composite the diagram with (𝑝1 )∗ , we obtain that for any locally constant
factorization algebra 𝐴 on 𝑋 × 𝑌 , there is an equivalence
∫ 𝐴 = (𝑝1 × 𝑝2 )∗ (𝐴)(∗)
𝑋×𝑌
Remark 3.33. In the case of pointed manifolds and manifolds with or without
boundaries, all (stratified) disks are homeomorphic to the euclidean plane or half
plane.
Some results on locally constant factorization algebras in the usual case remain
true in the stratified case. We state some useful ones; the proofs are similar.
ii) For sub-disks 𝑇 ⊆ 𝑈 which are good neighborhoods (in 𝑉 ) with the same
index, 𝑇 × 𝐷 is a good neighborhood in 𝑋 of same index as 𝑈 × 𝐷.
Then pushforward along 𝑓 preserves local constantness.
Proposition 3.35 (c.f. Proposition 2.14). Suppose 𝑋, 𝑌 are stratified spaces with
finitely many strata. Then the projections 𝑋 × 𝑌 → 𝑋 and 𝑋 × 𝑌 → 𝑌 are
adequately stratified, and the functor 𝜋 ∗ ∶ FA(𝑋 × 𝑌 ) → FA(𝑋, FA(𝑌 )) restricts to
a functor 𝜋 ∗ ∶ FA𝑙𝑐 (𝑋 × 𝑌 ) → FA𝑙𝑐 (𝑋, FA𝑙𝑐 (𝑌 )), which is an equivalence.
Proposition 3.36 (c.f. Proposition 2.18). Suppose 𝑋 is a stratified space, 𝐴 is a
factorization algebra on 𝑋. If there exists an open cover 𝒰 of 𝑋 such that 𝐴|𝑈 is
locally constant for every 𝑈 ∈ 𝒰, then 𝐴 is locally constant.
We furthermore have the following proposition:
Proposition 3.37. Suppose 𝑖 ∶ 𝑋 → 𝑌 is a stratified embedding of stratified
spaces such that 𝑖(𝑋) is a union of strata of 𝑌 . Then taking the pushforward along
𝑖 preserves local constantness.
We will now focus on some examples.
Example 3.38. The half line, 𝑋 = [0, +∞), with the dimension-0 stratum given
by {0}. Then all connected open subsets of 𝑋 form a factorization basis ℐ of 𝑋.
An example of stratified locally constant prefactorization algebra on ℐ arises as
follows: Choose an 𝐸1 -algebra 𝐸, a right 𝐸-module 𝑀, and define the factorization
algebra 𝐴𝐸,𝑀 as follows:
i) For any 0 < 𝑎 < 𝑏, define 𝐴𝐸,𝑀 ((𝑎, 𝑏)) = 𝐸;
ii) For any 0 < 𝑎, define 𝐴𝐸,𝑀 ([0, 𝑎)) = 𝑀;
iii) The structure maps are given by the multiplication on 𝐸 and the structure of
𝑀 being a right 𝐸-module.
It is easy to verify that this forms an ℐ -prefactorization algebra, and satisfies
the locally constant condition. Furthermore, we have:
Proposition 3.39. i) The ℐ -prefactorization algebra 𝐴𝐸,𝑀 given above is an
ℐ -factorization algebra, and hence extends uniquely into a factorization al-
gebra on 𝑋, which we will still denote by 𝐴𝐸,𝑀 ;
ii) 𝐴𝐸,𝑀 is locally constant on 𝑋;
iii) Moreover, all locally constant factorization algebras on 𝑋 is equivalent to
some 𝐴𝐸,𝑀 ;
iv) Finally, there exists an equivalence between FA𝑙𝑐 ([0, ∞)) and the ∞-category
E1 -RMod of (pointed) right modules over 𝐸1 -algebras (whose objects are or-
dered pairs (𝐸, 𝑀) such that 𝐸 is an 𝐸1 -algebra and 𝑀 is a right 𝐸1 -module);
82 An Introduction to Factorization Algebras and Factorization Homology
FA𝑙𝑐 ((0, ∞))
≅ / E1 -Alg.
Example 3.40. The unit interval, 𝑋 = [0, 1], with the dimension-0 stratum given
by {0, 1}. Proposition 3.36 shows that
FA𝑙𝑐 (𝑋) ≅ FA𝑙𝑐 ([0, 1)) × FA𝑙𝑐 ((0, 1]) ≅ E1 -RMod × E1 -LMod.
FA𝑙𝑐 ((0,1)) E1 -Alg
∫ 𝐴 ≅ 𝑀 𝑟 ⊗ 𝑀𝑙 .
[0,1] 𝐸
Example 3.41. The pointed euclidean space, 𝑋 = ℝ𝑛∗ , with the dimension-0 stra-
tum given by {0}. Similar to the case above, we will discuss the relationship be-
tween locally constant factorization algebras on 𝑋 and 𝐸𝑛 -modules.
To do this, we notice that there exists a factorization basis of 𝑋, the open convex
sets, which we will denote by 𝒞 . Suppose 𝑀 is a module over an 𝐸𝑛 -algebra 𝐸.
We define a 𝒞 -prefactorization algebra 𝐴𝐸,𝑀 as follows:
iii) The structure maps are given by the multiplication on 𝐸 and the structure of
𝑀 being an 𝐸-module.
iii) We obtain a functor 𝜓 ∶ E𝑛 -Mod → FA𝑙𝑐 (𝑋), which fits into the following
commutative diagram:
E𝑛 -Alg
≅ / FA𝑙𝑐 (ℝ𝑛 )
_
𝜓
E𝑛 -Mod / FA𝑙𝑐 (𝑋);
and
E𝑛 -Mod → E𝑛 -Alg ≅ FA𝑙𝑐 (ℝ𝑛 ) → FA𝑙𝑐 (ℝ𝑛 − {0})
are equivalent, and identify E𝑛 -Mod with the pullback FA𝑙𝑐 (𝑋) ×FA𝑙𝑐 (ℝ𝑛 −{0})
FA𝑙𝑐 (ℝ𝑛 ).
Hence this proposition can be used to characterize 𝐸𝑛 -modules, which may
be regarded as a generalization of Theorem 3.23 in the 𝑛-dimensional case. In
particular, by taking the fiber we have the following corollary:
Corollary 3.43. The functor
E𝑛 -Mod(𝐴) → FA𝑙𝑐 (ℝ𝑛∗ ) × {𝐴}
FA𝑙𝑐 (ℝ𝑛 −{0})
Furthermore, we have
we obtain an equivalence
4 Further applications
In this section, we will briefly talk about two applications of the above results on
factorization algebras and factorization homology: the Deligne conjecture and the
Bar constructions. We will solve the Deligne conjecture by the tool of centralizers,
and extend the study of Bar constructions on associative algebras to arbitrary 𝐸𝑛 -
algebras. We will mostly present ideas and sketches of proofs; the details can be
found in [Fra11] and [GTZ12].
𝐴 ⊗ 𝔷𝑛 (𝑓 ) ⊗ 𝔷𝑛 (𝑔)
6 QQQ
𝟙⊗1𝔷𝑛 (𝑔) mmmm QQQ𝑒𝑓 ⊗𝟙
mm QQQ
mm mm QQQ
mm (
𝐴 ⊗ 𝔷𝑛 (𝑓 ) 𝐵 ⊗ 𝔷𝑛 (𝑔)
𝟙⊗1𝔷𝑛 (𝑓 ) vv
: QQQ 𝟙⊗1𝔷𝑛 (𝑔) mmmm
6 II
v QQQ𝑒𝑓 m II 𝑒𝑔
v Q QQQ mm m II
vvv QQ m mm II
I$
vv 𝑓 QQ(/ mmm 𝑔
𝐴 𝐵 / 𝐶,
𝔷𝑛 (∘) ∶ 𝔷𝑛 (𝑓 ) ⊗ 𝔷𝑛 (𝑔) → 𝔷𝑛 (𝑔 ∘ 𝑓 )
The following theorem, also known as the “relative Deligne conjecture”, claims
that the centralizer is computed by 𝐸𝑛 -Hochschild cohomology:
86 An Introduction to Factorization Algebras and Factorization Homology
𝔷𝑛 (∘)
𝔷𝑛 (𝑓 ) ⊗ 𝔷𝑛 (𝑔) / 𝔷 (𝑔 ∘ 𝑓 )
𝑛
≅ ≅
𝐸 𝐸
Map𝐴𝑛 (𝐴, 𝐵𝑓 ) ⊗ Map𝐵𝑛 (𝐵, 𝐶𝑔 )
∘ / Map𝐸𝑛 (𝐴, 𝐶 ),
𝐴 𝑔∘𝑓
For any 𝑔𝑖 ∶ 𝐴|𝑈𝑖 → 𝐵|𝑈𝑖 ∈ 𝑍(𝑈𝑖 ) for all 𝑖, we define 𝜌𝑈1 ,⋯,𝑈𝑟 ,𝑉 (𝑔1 , ⋯ , 𝑔𝑟 ) to be
a map of factorization algebras from 𝐴 to 𝐵 on a factorizing basis on 𝑉 . We take
the following basis:
and define the following map 𝜌𝑈1 ,⋯,𝑈𝑟 ,𝑉 (𝑔1 , ⋯ , 𝑔𝑟 ) of factorization algebras:
Proof of Theorem 4.1. We first notice that for any object 𝑋 in C, the composi-
tion map ∘ ∶ Map(𝑋, 𝑋) ⊗ Map(𝑋, 𝑋) → Map(𝑋, 𝑋) makes Map(𝑋, 𝑋) an 𝐸1 -
algebra. Now, by the relative Deligne conjecture, HH𝐸𝑛 (𝐴, 𝐴) is an 𝐸1 -algebra
in the ∞-category E𝑛 -Alg, which means it is an 𝐸𝑛+1 -algebra, by Dunn’s Theo-
rem. ◻
It can be directly verified that the right hand side of the definition is equivalent
to 𝑝∗ (𝜔1 (𝐴)), where 𝑝 ∶ 𝐼 → ∗ is the unique map. Thus the definition is just the
factorization homology of the associated factorization algebra on 𝐷1 . In particular,
this definition agrees with the definition for augmented associative algebras.
Hence, the above discussion shows that 𝐵(𝐴) has a natural structure of an 𝐸𝑛−1 -
algebra. We can therefore iterate this construction at most 𝑛 times:
4
This is often called the “Bar complex”, since the underlying category is often taken to be the derived
∞-category of chain complexes, ℂ𝕙(𝕜).
4 Further applications 89
𝐵 𝑖 (𝐴)(𝑉 ) ∶= ∫ 𝜔′ (𝐴|𝑉 ).
𝑉 ∪{∞}
The following theorem from [GTZ12] proved that this indeed gives the 𝐸𝑖 -
coalgebra structure on 𝐵 𝑖 (𝐴):
90 An Introduction to Factorization Algebras and Factorization Homology
Theorem 4.16. The maps ∇𝑈1 ,⋯,𝑈𝑠 ,𝑉 form the structure maps of a locally constant
factorization coalgebra, making 𝐵 𝑖 (𝐴) into an 𝐸𝑖 -coalgebra. Thus 𝐵 𝑖 lifts to a
functor
𝐵 𝑖 ∶ E𝑛 -Alg𝑎𝑢𝑔 → E𝑖 -coAlg(E𝑛−𝑖 -Alg𝑎𝑢𝑔 ).
In particular, taking 𝑖 = 1, we obtain the 𝐸1 -coalgebra structure on the Bar
object 𝐵(𝐴) of an augmented 𝐸𝑛 -algebra, which restricts to the standard coalgebra
structure on the Bar object if 𝐴 is an associative algebra.
References
[AF12] D. Ayala and J. Francis (2012). Factorization Homology of Topological
Manifolds. arXiv: 1206.5522 [math.RT].
[CG16] K. Costello and O. Gwilliam. (2016). Factorization Algebras in Quantum
Field Theory. Cambridge University Press.
[Dun88] G. Dunn (1988). “Tensor Product of Operads and Iterated Loop Spaces”.
Journal of Pure and Applied Algebra 50, 237–258.
[FN62] E. Fadell and L. Neuwirth (1962). “Configuration Spaces”. Mathematica
Scandinavica 10, 111–118.
[Fra11] J. Francis (2011). The Tangent Complex and Hochschild Cohomology of
𝐸𝑛 -rings. arXiv: 1104.0181 [math.RT].
[Gin13] G. Ginot (2013). Notes on Factorization Algebras, Factorization Homol-
ogy and Applications. arXiv: 1307.5213 [math.RT].
[GTZ10] G. Ginot, T. Tradler, and M. Zeinalian (2010). Derived Higher
Hochschild Homology, Topological Chiral Homology and Factorization
algebras. arXiv: 1011.6483 [math.RT].
[GTZ12] G. Ginot, T. Tradler, and M. Zeinalian (2012). Higher Hochschild Co-
homology, Brane Topology and Centralizers of 𝐸𝑛 -algebra Maps. arXiv:
1205.7056 [math.RT].
[Hov99] M. Hovey (1999). Model Categories. Vol. 63. Mathematical Surveys and
Monographs. American Methematical Society.
[KS77] R. C. Kirby and L. C. Siebenmann (1977). Foundational Essays on Topo-
logical Manifolds, Smoothings, and Triangulations. Princeton University
Press.
[Lur09] J. Lurie (2009). Higher Topos Theory. Vol. 170. Annals of Mathematics
Studies. Princeton University Press.
[Lur16] J. Lurie (2016). Higher Algebra. URL: http://www.math.harvard.
edu/l̃urie/papers/HA.pdf.
[MathOverflow] 𝑘-Disk algebras versus 𝐸𝑘 algebras. https://mathoverflow.net/
questions/181828/k-disk-algebras-versus-e-k-algebras.
[May72] J. P. May (1972). The Geometry of Iterated Loop Spaces. Vol. 271. Lec-
ture Notes in Mathematics. Springer.
[nLab] The nLab. https://ncatlab.org.
Stable Irrationality of Varieties
Bu Chenjing1
ABSTRACT
We present several methods of showing the existence of stably irrational
varieties within a given family, and we show the infinitude of stable birational
classes in the family of quartic threefolds.
Contents
1 Introduction 92
1 Introduction
For a projective variety, there are various notions of rationality, describing how
close a variety is to the projective space.
Definition 1.1. Let k be a field, and let 𝑋 be a projective k-variety.
• 𝑋 is rational, if there exists 𝑛 ∈ N, such that
𝑋 is birational to P𝑛k .
𝑋 × P𝑚
k is birational to P𝑛k .
𝑓∶ 𝑈 →𝑉, 𝑔 ∶ 𝑉 → 𝑈,
P𝑛k ⇢ 𝑋.
𝑓 ∶ P1K → 𝑋K
rational. This method was then modified by Colliot-Thélène and Pirutka [CTP16]
to show that over C, a very general quartic hypersurface in P4 is not retract rational.
They also developed the specialisation method, with which they can provide more
general examples of smooth varieties that are not retract rational over number fields
and local fields.
In this article, we give an exposition of the methods and results mentioned
above.
Chow groups
In this subsection, we recall the definition and some basic properties of the Chow
groups of a variety. A general reference is [Ful98].
In the following, let k be a field, and let 𝑋 be a k-variety.
Definition 2.1. Let 𝑑 be an non-negative integer. The free abelian group
Z𝑑 (𝑋) = ⨁ Z ⋅ [𝑍],
𝑍⊂𝑋
An element of the Chow group is thus a class of cycles. We say that the cycles
in the same class are rationally equivalent.
Here we explain four operations of the Chow group: proper pushforward, flat
pullback, the intersection product, and the Gysin map.
as a 𝑑-cycle of 𝑌 , where
𝑓∗ ∶ Z𝑑 (𝑋) → Z𝑑 (𝑌 ).
as a (𝑑 + 𝑟)-cycle of 𝑌 , where
• [𝑓 −1 (𝑍)] denotes the sum ∑ 𝑚𝑖 [𝑍𝑖 ], where the 𝑍𝑖 are the irreducible com-
ponents of 𝑓 −1 (𝑍), and 𝑚𝑖 is the geometric multiplicity of 𝑍𝑖 in 𝑓 −1 (𝑍),
defined as the length of the local ring 𝒪𝑓 −1 (𝑍),𝑍𝑖 .
𝑓 ∗ ∶ Z𝑑 (𝑌 ) → Z𝑑+𝑟 (𝑋).
2 Criteria for rationality 95
It turns out again that flat pullback preserves rational equivalence [Ful98, §1.7].
Switching to the cohomological indexing notation, we get a pullback map of Chow
groups
𝑓 ∗ ∶ CH𝑑 (𝑌 ) → CH𝑑 (𝑋).
𝑍1 ∩ 𝑍2 = 𝑍1 ×𝑋 𝑍2 .
However, this does not always produce cycles of the expected dimension, as the
subvarieties may not be in a general position to intersect. To work around this
difficulty, we use the Gysin map of the diagonal map, which acts as a pullback
along a closed embedding.
The Gysin map is defined for vector bundles as follows.
Theorem 2.5. Let 𝑝 ∶ 𝐸 → 𝑋 be a vector bundle of rank 𝑟. Then the flat pullback
is an isomorphism for all 𝑑. Its inverse is called the Gysin map, and denoted by 𝑖! ,
where 𝑖 is the zero section map 𝑋 → 𝐸.
See [Ful98, §3.3].
Recall that a regular embedding is a closed embedding of schemes, such that
the ideal sheaf is locally generated by regular sequences. For example, a closed
embedding of smooth varieties is always a regular embedding.
For a regular embedding, the normal cone is a vector bundle. We can use this
property to extend the definition of the Gysin map to this case.
Definition 2.6. Let 𝑖 ∶ 𝑍 → 𝑋 be a regular embedding of constant codimension 𝑒.
The Gysin map
𝑖! ∶ CH𝑑 (𝑋) → CH𝑑−𝑒 (𝑍)
𝜎 ∶ Z𝑑 (𝑋) → Z𝑑 (𝑁𝑍 𝑋)
We then compose this map with the Gysin map defined in Theorem 2.5, giving the
desired map
𝑖! ∶ CH𝑑 (𝑋) → CH𝑑−𝑒 (𝑍).
96 Stable Irrationality of Varieties
Using the Gysin map as a pullback along the diagonal map, we can define the
intersection product of cycles.
defined as follows.
× Δ!
CH𝑑1 (𝑋) ⊗ CH𝑑2 (𝑋) ⟶ CH𝑑1 +𝑑2 (𝑋 × 𝑋) ⟶ CH𝑑1 +𝑑2 −𝑛 (𝑋),
where × denotes the cross product map sending [𝑍1 ] ⊗ [𝑍2 ] to [𝑍1 × 𝑍2 ],
and Δ ∶ 𝑋 → 𝑋 × 𝑋 is the diagonal map, which is a regular embedding.
This product equips the Chow groups with the structure of a graded ring, called
the Chow ring.
• 𝑓 is proper.
is an isomorphism.
• 𝑋 is complete.
is an isomorphism.
We will show that retract rationality implies universal CH0 triviality. The proof
will need a few lemmas. First of all, we prove a moving lemma for 0-cycles.
2 Criteria for rationality 97
Lemma 2.9. Let 𝑋 be a smooth projective k-variety, with k infinite and perfect,
and let 𝑈 ⊂ 𝑋 be a dense open set. Then every 0-cycle of 𝑋 is rationally equivalent
to one supported in 𝑈 .
Remark 2.10. This is a special case of [EGA-II, Proposition 7.4.9], but the proof
given here is more elementary.
Lemma 2.11. Let k be a finite field, and let 𝑈 ⊂ P𝑛k be a non-empty open set.
Then for any extension 𝐹 ∕k of sufficiently large degree 𝑑, the open set 𝑈𝐹 ⊂ P𝑛𝐹
contains an 𝐹 -rational point.
Proof. First, we suppose that the base field k is infinite. Since retract rationality is
preserved under change of base field, it suffices to prove that 𝑋 is CH0 -trivial over
k, i.e. degk ∶ CH0 (𝑋) → Z is an isomorphism.
By definition, there exist non-empty open sets 𝑈 ⊂ 𝑋, 𝑉 ⊂ P𝑛k , and maps
𝑓 𝑔
𝑈 ⟶ 𝑉 ⟶ 𝑈,
There exists 𝑅 ∈ 𝑝−1 (𝑄) such that 𝜅(𝑅) ≃ 𝐹 . This is because 𝑝−1 (𝑄) ≃ Spec(𝐹 ⊗k
𝐹 ), and we can take 𝑅 to be the point defined by the maximal ideal which is the
kernel of the multiplication map 𝐹 ⊗k 𝐹 → 𝐹 .
Let 𝐴 ∈ 𝑉 ⊂ P𝑛k be a k-rational point, which exists since k is infinite. Let
𝐿 ≃ P1𝐹 ⊂ P𝑛𝐹 be a line connecting 𝑅 and 𝐴𝐹 . The map 𝑔𝐹 sends 𝐿 to a rational
line 𝐿′ in 𝑈𝐹 , which is a rational 𝐹 -map P1𝐹 ⇢ 𝑈𝐹 . This map extends to a map
𝐿′ ∶ P1𝐹 → 𝑋𝐹 ,
which is proper since P1𝐹 is complete [EGA-II, Corollary 5.4.3]. This is a line
connecting the points 𝑔𝐹 (𝑅) and 𝑔𝐹 (𝐴𝐹 ).
As 𝑅 is an 𝐹 -rational point, we have 𝜅(𝑔𝐹 (𝑅)) ≃ 𝐹 and 𝜅(𝑔𝐹 (𝐴𝐹 )) ≃ 𝐹 .
Hence, the pushforward map of 𝐿′ on CH0 gives
We also mention the following criterion for universal CH0 -triviality of a mor-
phism, which will be useful later.
̃ → 𝑋 be a proper mor-
Theorem 2.13 (Colliot-Thélène and Pirutka). Let 𝑓 ∶ 𝑋
phism of k-varieties. Suppose that
̃𝑀 is universally
• For every point 𝑀 ∈ 𝑋, not necessarily closed, the fibre 𝑋
CH0 -trivial as a 𝜅(𝑀)-variety.
̃𝜂 contains a 0-cycle
Let 𝜂𝑖 be the generic point of 𝐶𝑖 . By hypothesis, each fibre 𝑋 𝑖
∑𝑗 𝑛𝑖𝑗 [𝐷𝑖𝑗 ]
of degree 1, where 𝑛𝑖𝑗 ∈ Z. We regard each 𝐷𝑖𝑗 as a curve in 𝑋.̃ Then each
function 𝑔𝑖 ∘ 𝑓 defines a rational function 𝑔𝑖𝑗 on 𝐷𝑖𝑗 . Write
where
• [Δ𝑋 ] is the pushforward of [𝑋] ∈ CH𝑛 (𝑋) along the diagonal map 𝑋 →
𝑋 × 𝑋.
• 𝐷 is an 𝑛-cycle of 𝑋 × 𝑋, supported in 𝑍 × 𝑋 for some closed subvariety
𝑍 ⊂ 𝑋 of codimension at least 1.
• 𝑥0 is a 0-cycle of 𝑋 of degree 1.
In order to show that this property is equivalent to CH0 -triviality, we introduce
the notion of a correspondence.
Definition 2.15. Let 𝑋 and 𝑌 be complete k-varieties, of dimensions 𝑚 and 𝑛,
respectively. A correspondence from 𝑋 to 𝑌 is an element of the set
Corr(𝑋, 𝑌 ) = CH𝑚 (𝑋 × 𝑌 ).
The pullback of the diagonal class is the class of the generic point of 𝑋, which is a
0-cycle of 𝑋k(𝑋) of degree 1.
2 Criteria for rationality 101
Proof. We may assume 𝑋 = Spec 𝐴 is affine. Then the pullback of the diagonal
class is the closed point defined by the maximal ideal, which is the kernel of the
multiplication map
k(𝑋) ⊗ 𝐴 → k(𝑋).
On the other hand, by the moving lemma 2.9, every 0-cycle of 𝑋 can be moved
out of 𝑍. This shows that the action of 𝐷 is 0, as in the defining equation (2.15.1),
we are taking the intersection product of two disjoint cycles. But for any 0-cycle 𝛾
of 𝑋, the action of [𝑋] × 𝑥0 sends it to
For a field k, the Brauer group Br(k) = Br(Spec k) coincides with the classical
notion defined as the group of equivalence classes of central simple algebras. A
classical reference is [Gro68].
The Kummer exact sequence of étale sheaves
(−)𝑛
0 → μ𝑛 → Gm −−−→ Gm → 0
where the left hand side denotes the 𝑛-torsion subgroup of Br(𝑋).
Definition 2.20. Let 𝑀 be a contravariant functor from schemes to abelian groups
(e.g. étale cohomology), and let k ⊂ K be two fields. Let
where 𝐴 runs through all discrete valuation rings with fraction field K.
This is called the unramified version of the functor 𝑀. For example, one has the
𝑞
unramified Brauer group Brnr (K∕k), and unramified cohomology 𝐻nr (K∕k, μ𝑛 ).
A deep result on the cohomological purity of the Brauer group gives rise to the
following theorem.
3 The deformation method 103
Theorem 2.21. Let 𝑋 be a regular, complete, integral k-variety. Then the natural
map Br(𝑋) → Br(k(𝑋)) induces an isomorphism
Br(𝑋) ≃ Brnr (k(𝑋)∕k).
See [CTS19, Proposition 5.2.2].
The discussion above immediately implies the following.
Corollary 2.22. Let 𝑋 be a regular, proper, integral k-variety. Then
2
Br(𝑋) [𝑛] ≃ 𝐻nr (k(𝑋)∕k, μ𝑛 ). ◻
Since the Brauer group is a torsion group [CTS19, Proposition 1.3.6], it can
thus be computed by the unramified cohomology groups.
On the other hand, the second cohomology of μ𝑛 is a part of the cycle module
(in the sense of [Ros96])
⊗(𝑖−1)
⨁𝑖 𝐻 𝑖 (−, μ𝑛 ),
which we do not give a precise definition here. This allows it to be regarded as a
“coefficient group” for Chow groups. As a result, there is an action of correspon-
dences
2 2
Corr(𝑋, 𝑌 ) ⊗ 𝐻nr (k(𝑋)∕k, μ𝑛 ) ⟶ 𝐻nr (k(𝑌 )∕k, μ𝑛 )
for k-varieties 𝑋, 𝑌 . This action will relate the Brauer group with the decomposi-
tion of the diagonal.
Theorem 2.23. Let 𝑋 be a smooth projective k-variety. If 𝑋 admits a decomposi-
tion of the diagonal, then the natural map induces an isomorphism
Br(k) ⥲ Br(𝑋).
In particular, if k is separably closed, then Br(𝑋) = 0.
Sketch of proof. The decomposition of the diagonal implies that the identity map
and the constant map (to be precise, a sum of constant maps) induce the same action
on
2
𝐻nr (k(𝑋)∕k, μ𝑛 ).
2
But the action of a constant map factors through 𝐻nr (k∕k, μ𝑛 ), via the corestriction
map, whence the result follows. ◻
Families of cycles
Definition 3.1 (Kollár [Kol96, Definition 3.10]). Suppose that
• k is an algebraically closed field of characteristic 0.
• 𝑆 is a k-scheme.
• 𝑋∕𝑆 is a projective 𝑆-scheme, with a chosen relatively ample line bundle.
• 𝐵∕𝑆 is a reduced normal 𝑆-scheme.
• 𝑑 and 𝑑 ′ are non-negative integers.
A well-defined family of 𝑑-cycles of 𝑋 of degree 𝑑 ′ parametrised by 𝐵 is a cycle
𝐶 = ∑𝑖 𝑚𝑖 [𝐶𝑖 ] of 𝑋 ×𝑆 𝐵,
such that
• Each 𝐶𝑖 is an integral closed subscheme of 𝑋 ×𝑆 𝐵.
• For each 𝑖, the image of the projection map 𝑔𝑖 ∶ 𝐶𝑖 → 𝐵 is an irreducible
component of 𝐵. In particular, 𝑔𝑖 is flat over a dense open subset of 𝐵.
• Each fibre of 𝑔𝑖 defines a 𝑑-cycle of 𝑋 of degree 𝑑 ′ . This means that the
fibre is either empty or of dimension 𝑑, and that 𝑔𝑖 is flat over a dense open
subset of 𝐵.
The deep theorem below shows the existence of a universal family of cycles, in
that every family of cycles is realised as its pullback.
Theorem 3.2. Under the assumptions of Definition 3.1, for an 𝑆-scheme 𝑍, define
′ well-defined families of non-negative 𝑑-cycles
Chow𝑑,𝑑 (𝑍) = { }.
𝑋∕𝑆 of 𝑋 of degree 𝑑 ′ parametrised by 𝑍
Then
′
• Chow𝑑,𝑑
𝑋∕𝑆
is a contravariant functor from the semi-normal 𝑆-schemes to sets.
• 𝐵 is a smooth k-scheme.
• 𝑋 → 𝐵 is a projective morphism.
Lemma 3.5. In Situation 3.4, for any non-negative integer 𝑑, there exists
such that
Proof. By [EGA-IV3, Theorem 9.7.7], the set of points in the Hilbert scheme
Hilb𝑑+1
𝑋∕𝐵
corresponding to the geometrically integral subvarieties is locally con-
structible. Let 𝐺 be an irreducible component of this set, equipped with the re-
duced scheme structure. Then 𝐺 is quasi-projective over 𝑆, as the components of
Hilb𝑑+1
𝑋∕𝐵
are projective. Let
𝑊 ⊂ 𝐺 ×𝐵 𝑋
̃1 → 𝑊1
𝑊
Ab ∶ Div𝑊 ̃𝑗 → Pic𝑊
̃𝑗 ∕𝐺 ̃𝑗 ,
̃𝑗 ∕𝐺
where Div𝑊 ̃𝑗 is the scheme parametrising the effective Cartier divisors [FAG,
̃𝑗 ∕𝐺
̃𝑗 , and hence over 𝐵, and Pic ̃ ̃
Theorem 9.3.7], which is quasi-projective over 𝐺 𝑊𝑗 ∕𝐺𝑗
is the Picard scheme [FAG, Theorem 9.4.8]. Let
Δ𝑗 ⊂ Div𝑊 ̃𝑗 × Div𝑊
̃𝑗 ∕𝐺 ̃𝑗
̃𝑗 ∕𝐺
Of course, a rational equivalence of two 𝑑-cycles may involve more than one
(𝑑 + 1)-dimensional subvariety. The next lemma deals with this situation.
For simplicity, if 𝑉 ⊂ 𝑋𝑏 is a subvariety, we will say “the desingularisation”
of 𝑉 when we refer to the variety 𝑉̃ given by the previous lemma, and we simply
add a tilde to indicate this desingularisation.
Lemma 3.6. In Situation 3.4, for any non-negative integer 𝑑, there exists
such that
• For any 𝑏 ∈ 𝐵(k), and any data (𝑉𝑗 , 𝐷𝑗,1 , 𝐷𝑗,2 )𝑛𝑗=1 , where each 𝑉𝑗 is an
integral subscheme of 𝑋𝑏 of dimension 𝑑 +1, and 𝐷𝑗,1 , 𝐷𝑗,2 are two effective
Weil divisors on the desingularisation 𝑉̃𝑗 of 𝑉𝑗 , such that 𝐷𝑗,1 − 𝐷𝑗,2 is a
principal divisor on 𝑉𝑗 , there exists 𝑖 and 𝑡 ∈ (𝐻𝑖 )𝑏 (k), such that the fibre
((𝑊𝑖 )𝑡 , (𝐸𝑖,𝑗,1 )𝑡 , (𝐸𝑖,𝑗,2 )𝑡 ) is identical to the given data.
Proof. For each 𝑛-tuple (𝑇1 , … , 𝑇𝑛 ) as given by the previous lemma, we consider
the normalisation 𝐻 of the product 𝑇1 ×⋯×𝑇𝑛 , equipped with the data of 𝑛 triples as
given by the previous lemma. The collection of all such 𝐻 satisfies the requirement
of this lemma. ◻
𝑍1 , 𝑍2 ∈ Chow𝑑𝑋∕𝐵 (𝐵)
such that
• For any 𝑏 ∈ 𝐵(k), and any data (𝑉𝑖 , 𝐷𝑖,1 , 𝐷𝑖,2 )𝑛𝑖=1 as in Lemma 3.6, such
that
𝑍1,𝑏 + ∑𝑛𝑖=1 [𝐷𝑖,1 ] = 𝑍2,𝑏 + ∑𝑛𝑖=1 [𝐷𝑖,2 ] in Z• (𝑋𝑏 ),
there exists 𝑖 and a point 𝑡 ∈ (𝑀𝑖 )𝑏 (k), such that the fibre ((𝑊𝑖 )𝑡 , (𝐸𝑖,𝑗,1 )𝑡 ,
(𝐸𝑖,𝑗,2 )𝑡 ) is identical to the given data.
𝑓 ∶ 𝐻𝑖 → Chow𝑋∕𝐵 × Chow𝑋∕𝐵 ,
𝑖 𝑛 𝑖 𝑛
𝑡 ↦ (𝑍1 + ∑𝑗=1 (𝐸𝑖,𝑗,1 )𝑡 , 𝑍2 + ∑𝑗=1 (𝐸𝑖,𝑗,2 )𝑡 ),
where (𝐸𝑖,𝑗,1 or 2 )𝑡 is regarded as a cycle of 𝑋 via the pushforward along the desin-
gularisation map (onto a closed subvariety of 𝑋). Now let 𝑀𝑖 be the inverse image
of the diagonal along 𝑓 , with the reduced scheme structure, equipped with the data
of 𝑛𝑖 triples given by that of 𝐻𝑖 . This proves the second statement.
For the first statement, write 𝑍 = 𝑍1 − 𝑍2 . Let 𝑏 ∈ 𝐵 be a point where 𝑍𝑏
is rationally equivalent to zero. Then, there exist subvarieties 𝑉𝑗 ⊂ 𝑋𝑏 , where
𝑗 = 1, … , 𝑛, and rational functions 𝑔𝑗 on 𝑉𝑗 , which give rise to rational functions
𝑔̃𝑗 on a desingularisation 𝑉̃𝑗 , such that
where 𝑓𝑗 denotes the map 𝑉̃𝑗 → 𝑋𝑏 . Conversely, the existence of this data implies
that 𝑍𝑏 is rationally equivalent to zero, since 𝑀𝑗 is taken to be the inverse image
of the diagonal. Therefore, the locus where 𝑍𝑏 is equivalent to zero is exactly the
union of the images of the 𝑀𝑖 . ◻
We are now getting close to the main theorem, which states that the locus where
[𝑍1,𝑏 ] = [𝑍2,𝑏 ] is a countable union of closed sets. There is one further lemma
needed.
Then
Proof. Let 𝑡 ∈ 𝑀(k) ⧵ 𝑀 ∘ (k) be a point. As in the proof of Lemma 2.9, we can
find a curve 𝐶 in 𝑀, passing through 𝑡, and not contained in 𝑀 ⧵ 𝑀 ∘ . Taking the
normalisation of this curve, we may thus assume that 𝑀 is a smooth curve.
3 The deformation method 109
𝑍1 , 𝑍2 ∈ Chow𝑑𝑋∕𝐵 (𝐵)
be two well-defined families of cycles. Then there exists a countable family {𝐵𝑖 } of
closed subschemes of 𝐵, such that
𝐻 ⊂ Hilb𝑋∕𝐵 ,
𝐶 ⊂ Chow𝑈 ′ ∕𝐻 .
𝐶 ×𝐻 𝑈 ′ ⊂ 𝐶 ×𝐻 𝐻 ×𝐵 𝑋 ×𝐵 𝑋 ≃ 𝐶 ×𝐵 𝑋 ×𝐵 𝑋,
• 𝐵 is a smooth k-scheme.
• 𝑋 → 𝐵 is a projective morphism.
Then there exists a countable family {𝐵𝑖 } of closed subschemes of 𝐵, such that
Proof. Let 𝐹𝑖 , 𝐹𝑖′ be two of the schemes as in Lemma 3.10, with 𝑑𝑖 = 𝑑𝑖′ =
dim(𝑋∕𝐵), and let 𝐶𝑖 , 𝐶𝑖′ be the universal cycles, lying in 𝑋 ×𝐵 𝑋 ×𝐵 𝐹𝑖 or 𝑖′ .
0,𝑑 0,𝑑+1
Let 𝐺𝑗 , 𝐺𝑗 ′ be irreducible components of Chow𝑋∕𝐵 and Chow𝑋∕𝐵 , respec-
tively, where 𝑑 is arbitrary, and let 𝐷𝑗 , 𝐷𝑗 ′ be the universal cycles lying in 𝑋 ×𝐵
𝐺𝑗 or 𝑗 ′ .
We define two cycles of 𝑌 = 𝐹𝑖 ×𝐵 𝐺𝑗 ×𝐵 𝑋 ×𝐵 𝑋 ×𝐵 𝐺𝑗 ′ ×𝐵 𝐹𝑖′ by
𝐹𝑖 ×𝐵 𝐹𝑖′ ×𝐵 𝐺𝑗 ×𝐵 𝐺𝑗 ′ .
At the point
the cycle 𝑍1 gives [Δ𝑋𝑏 ]+𝑧1 +[𝑋𝑏 ]×𝑥1 , where 𝑧1 is a non-negative cycle supported
in 𝑍 × 𝑋𝑏 for 𝑍 ⊂ 𝑋𝑏 of codimension 1, and similarly, the cycle 𝑍2 gives [𝑋𝑏 ] ×
𝑥2 + 𝑧2 , with 𝑧2 likewise.
Therefore, Theorem 3.9 implies that the locus where the equation
• 𝐵 is a smooth k-scheme.
112 Stable Irrationality of Varieties
Stable equivalence
This subsection gives a variant of the above result, concerning stable equivalence
instead of retract rationality.
Definition 3.13. Two projective k-varieties are stably equivalent, if
𝑋 × P𝑚 is birational to 𝑌 × P𝑛
for some 𝑚, 𝑛 ∈ N.
Stable rationality is the same as stable equivalence to a point.
Lemma 3.14. Let 𝑋, 𝑌 be two k-varieties, such that there exist open sets 𝑈 ⊂ 𝑋,
𝑉 ⊂ 𝑌 × P𝑛 , and two morphisms 𝑝 ∶ 𝑈 → 𝑉 , 𝑞 ∶ 𝑉 → 𝑈 , such that 𝑞 ∘ 𝑝 = id𝑈 .
Then there exist two correspondences
such that for any field extension K∕k, the induced map
induces the identity map on CH0 . This is because every closed point of 𝑌 × P𝑛
is sent to another point that lives in the same slice of P𝑛 , and hence, is rationally
equivalent to it as a 0-cycle.
For the second part, we use an argument as in the proof of Theorem 2.18.
Namely, we change the base field to k(𝑋), to find that
where 𝛽 is the class of the generic point. The rest of the proof is analogous to the
proof of Theorem 2.18, (ii) ⇒ (iii). ◻
Note that the assumptions of this lemma is satisfied when 𝑋 and 𝑌 are stably
equivalent.
Using an argument as in the proof of Proposition 3.11, we obtain the following
result.
Theorem 3.15. Suppose
• k is an algebraically closed field of characteristic 0.
• 𝐵 is a smooth k-scheme.
• 𝑋 → 𝐵 and 𝑌 → 𝐵 are two projective morphisms.
Then the set of all points 𝑏 ∈ 𝐵(k) such that there exist correspondences
𝑓 ∈ Corr(𝑋𝑏 , 𝑌𝑏 ), 𝑔 ∈ Corr(𝑌𝑏 , 𝑋𝑏 ),
[Δ𝑋𝑏 ] = 𝐷 + 𝑔 ∘ 𝑓 ,
𝑋 ×𝐵 𝑋 ×𝐵 𝐹𝑖 ×𝐵 𝐹𝑖′ ×𝐵 𝐺𝑗 ×𝐵 𝐺𝑗 ′ ×𝐵 𝐻𝑘 ×𝐵 𝐻𝑘′ ,
where
• 𝐹𝑖 , 𝐹𝑖′ are given by Lemma 3.5.
• 𝐺𝑗 , 𝐺𝑗 ′ , 𝐻𝑘 , 𝐻𝑘′ are irreducible components of Chow𝑋×𝐵 𝑌 ∕𝐵 , parametrising
the correspondences from 𝑋 to 𝑌 for 𝐺𝑗 and 𝐺𝑗 ′ , and from 𝑌 to 𝑋 for 𝐻𝑘 ,
𝐻𝑘′ .
114 Stable Irrationality of Varieties
Corollary 3.16. Under the assumptions of Theorem 3.15, the set of all points 𝑏 ∈
𝐵(k) such that 𝑋𝑏 is stably equivalent to 𝑌𝑏 is contained in a countable union of
closed sets. Moreover, this union does not contain any point 𝑏 ∈ 𝐵(k) such that
𝑋𝑏 is smooth and has a decomposition of the diagonal, and 𝑌𝑏 does not have a
decomposition of the diagonal
Proof. The countable union of closed sets given by Theorem 3.15 satisfies this
requirement. Indeed, for those 𝑏 ∈ 𝐵(k) such that 𝑋𝑏 and 𝑌𝑏 are stably equivalent,
Lemma 3.14 shows that 𝑏 is in this union.
To prove the last statement, let 𝑏 ∈ 𝐵(k) be such a point. We show that such
correspondences 𝑓 , 𝑔 as in Theorem 3.15 do not exist between 𝑋𝑏 and 𝑌𝑏 . In fact,
if they exist, then 𝑔 ∘ 𝑓 acts on CH0 (𝑋) by the identity map. But since id𝑌 sends
every 0-cycle to its degree multiplied by a fixed 0-cycle of degree 1, so does the
correspondence 𝑔 ∘ 𝑓 = 𝑔 ∘ id𝑌 ∘ 𝑓 . Moreover, this holds over any field extension of
k. By Theorem 2.18, 𝑋 has a decomposition of the diagonal, a contradiction. ◻
• 𝐵 is a smooth k-scheme.
• There exist two k-points 𝑏0 , 𝑏1 ∈ 𝐵, such that the fibre 𝑋𝑏0 has a decompo-
sition of the diagonal, while the fibre 𝑋𝑏1 does not have a decomposition of
the diagonal.
Then there are uncountably many stable equivalence classes of varieties in this
family.
Proof. For those smooth fibres 𝑋𝑏 that do not have a decomposition of the diagonal,
we apply Corollary 3.16 to the constant family 𝐵 ×𝑋𝑏 → 𝐵 and the family 𝑋 → 𝐵.
It follows that the set of 𝑏′ ∈ 𝐵(k) such that 𝑋𝑏 is stably equivalent to 𝑋𝑏′ is
contained in a countable union of closed sets, which can not coincide with the
whole space.
By Theorem 3.12, the locus of smooth fibres with no decomposition of the
diagonal is the complement of a countable union of closed subsets of 𝐵. Therefore,
in order to cover this locus, there must be uncountably many stable equivalence
classes of fibres of 𝑋 → 𝐵, since each of these classes is contained in a countable
union of closed sets. ◻
4 The specialisation method 115
Situation 4.1. Let 𝐴 be a discrete valuation ring, with fraction field K and residue
field k. Let 𝒳 be an 𝐴-scheme. Suppose that
𝑖∗ 𝑗∗
CH1 (𝑋 s ) ⟶ CH1 (𝒳 ) ⟶ CH0 (𝑋) → 0,
where 𝑖, 𝑗 are the obvious inclusions. (The last term is CH0 instead of CH1 , since
Spec K is a 1-dimensional point in Spec 𝐴.)
By [Ful98, §2.6 and §20.1], there is a Gysin map
𝑖! ∶ CH1 (𝒳 ) → CH0 (𝑋 s ),
Lemma 4.3. In Situation 4.1, suppose that 𝐴 is henselian, and 𝒳 is proper and
s
flat over 𝐴. Let 𝑋sm ⊂ 𝑋 s be the open set where 𝑋 s is smooth. Then every 0-cycle
s s
of 𝑋 supported in 𝑋sm can be lifted along the specialisation map
Proof. Since the conditions of the lemma is preserved by a base change, we only
need to prove that degk ∶ CH0 (𝑋)̃ → Z is an isomorphism. By the first hypothesis,
this map is surjective. Thus, it suffices to show that for every 0-cycle 𝑥 of 𝑋 ̃ of
degree 0, 𝑥 is rationally equivalent to 0. But by the moving lemma 2.9, it is equiv-
alent to one supported in 𝑈̃ , which induces a cycle in 𝑈 , which is equivalent to 0 in
̃ ◻
𝑋 by hypothesis. Since 𝑓 is universally CH0 -trivial, 𝑥 is equivalent to 0 in 𝑋.
Lemma 4.6. Let 𝐴 be a discrete valuation ring with residue field k, and let 𝐹 ∕k be
an extension. Then there exists a complete discrete valuation ring 𝐵 with residue
field 𝐹 , together with a local map 𝐴 → 𝐵 inducing the field map k → 𝐹 .
See [Bou06, Chapter IX, Appendix, §2, Corollary, and Exercise 4].
̃ → 𝑋.
• The generic fibre 𝑋 has a desingularisation 𝑋
̃ is universally CH0 -trivial, so is 𝑋
Then if 𝑋 ̃s.
By Galois descent, there exists a finite extension 𝐹 ∕k over which everything in the
equation is defined, and we have
Therefore, there exists a finite extension 𝐸 such that the equation of the decompo-
sition of the diagonal holds over 𝐸. ◻
Proof. Our plan is to find a suitable base change in order to apply Theorem 4.7.
First, we replace 𝐴 by its completion.
Let 𝐹 be a finite extension of K, on which 𝑋 ̃ is defined. In other words, there
exists a smooth variety 𝑌 over 𝐹 , such that 𝑌K ≃ 𝑋,̃ and there is a desingularisation
map 𝑌 → 𝑋𝐹 , which coincides with the map 𝑋 ̃ → 𝑋 over K.
By Lemma 4.8, we may replace 𝐹 by a finite extension of it, so we may assume
that 𝑌 is universally CH0 -trivial.
Let 𝐵 be the integral closure of 𝐴 in 𝐹 . By [Ser79, Proposition I.3], 𝐵 is also a
discrete valuation ring. Since k is algebraically closed, the residue field of 𝐵 is also
k. We can thus do a base change along the map 𝐴 → 𝐵, and apply Theorem 4.7 to
complete the proof. ◻
In this case, there exists a k-variety 𝑌 such that 𝐹 is isomorphic to k(𝑌 ), and
there exist non-empty open sets 𝑈 ⊂ 𝑋 ×k 𝑌 and 𝑉 ⊂ P𝑛𝑌 , such that 𝑈 is a retract
of 𝑉 as a 𝑌 -scheme. There exists a closed point 𝑦 ∈ 𝑌 such that the fibres 𝑈𝑦 and
𝑉𝑦 are non-empty. This proves the lemma. ◻
Theorem 4.11 (Colliot-Thélène and Pirutka). Assume that the four assumptions of
̃𝐹 is retract rational for a field 𝐹 containing
Theorem 4.9 are satisfied. Then, if 𝑋
̃ s
K, then 𝑋 is universally CH0 -trivial. ◻
The example
• Let k be an algebraically closed field with char k ≠ 2.
• Let 𝐴 ⊂ P2 be a smooth conic, defined by the quadratic equation
𝛼(𝑧0 , 𝑧1 , 𝑧2 ) = 0.
𝛿1 = 0 and 𝛿2 = 0,
such that they each meet 𝐴 tangentially at 3 points, giving six tangent points
𝑄1 , … , 𝑄6 , and such that 𝐷1 ∩ 𝐷2 is nine distinct points 𝑃1 , … , 𝑃9 . These
nine points do not lie on 𝐴, since otherwise 𝐷1 and 𝐷2 will meet tangentially
at that point.
• Let 𝐵 ⊂ P2 be a cubic that intersects 𝐴 in the six points 𝑄1 , … , 𝑄6 . In fact,
for any nine given points on the plane, there exists a cubic curve passing
through all of them. We use these six points, and choose three other points
which are non-collinear and not on 𝐴. This ensures that the cubic does not
contain 𝐴, and can only intersect 𝐴 in these six points.
As cycles of 𝐴, we have
(𝐷1 + 𝐷2 ) ⋅ 𝐴 = 2𝐵 ⋅ 𝐴.
where
𝜕𝑔 𝜕𝑔 𝜕𝑔
𝑀 = { 𝑔 = 0, ( = 0 or = 0), = 0 },
𝜕𝑧1 𝜕𝑧2 𝜕𝑧3
We will see that 𝑇 has the property of being unirational but not having a de-
composition of the diagonal.
We next construct an explicit desingularisation of the threefold 𝑇 constructed
above, following [CTP16, Appendix A]. We do this in order to get a desingulari-
sation map which is universally CH0 -trivial, and such that the Brauer group of the
desingularisation is non-trivial.
𝐿 ∶ 𝑧0 = 𝑧 1 = 𝑧 2 = 0
Lemma 5.3. Let 𝑋 be a rationally connected complex variety. Then the Picard
number 𝜌(𝑋) is equal to the Betti number 𝑏2 (𝑋). ◻
then we denote by 𝑀 fin its finite part, which is the largest finite submodule of 𝑀
that is a direct summand.
Lemma 5.4. Let 𝑋 be a rationally connected complex variety, and let ℓ be a prime
number. Then
Br(𝑋) {ℓ} ≃ 𝐻 3 (𝑋, Zℓ (1)) {ℓ},
where the right hand side is the étale cohomology of the sheaf Zℓ (1) = ⟵
lim𝑛 μℓ𝑛 ,
which may be identified with Zℓ over C.
By Lemma 5.3, the “corank” of Br(𝑋) {ℓ} (i.e. number of summands Qℓ ∕Zℓ ) is
𝑏2 − 𝜌 = 0, so that
Br(𝑋) {ℓ}fin ≃ Br(𝑋) {ℓ}.
where [ℓ𝑛 ] indicates the subgroup of elements killed by ℓ𝑛 . Since the last term is
torsion-free, we have
Consequences
Following Colliot-Thélène and Pirutka [CTP16], we present some consequences
of the example given in the previous subsection.
The first result provides examples of smooth quartic threefolds over complex
numbers, which are not retract rational.
Theorem 5.5. Let 𝑃 → P𝑁 4
C be the family of all quartic hypersurfaces in PC . Let
𝑡 ∈ C ⧵ Q be a transcendental number. Then the set
is Zariski dense in P𝑁
C.
̃ ≠ 0.
such that 𝑓 is universally CH0 -trivial, and Br(𝑋)
𝑁
Let 𝑊 ⊂ P be the closed subset corresponding to the singular hypersurfaces.
Q
Let 𝑀 ∈ P𝑁 be the point corresponding to 𝑋. Choose a point 𝑀 ′ ∈ P𝑁 ⧵ 𝑊 , and
Q Q
let
𝐿 ≃ P1 ⊂ P𝑁
Q Q
be the straight line connecting 𝑀 and 𝑀 ′ , with generic point 𝜂. Then Theorem 4.9
implies that the quartic threefold 𝑋 ∘ defined by 𝜂, over the field Q(𝑥), is not geo-
metrically retract rational. Moreover, by Theorem 4.11, for any embedding
Q(𝑥) ↪ C,
the base change 𝑋C∘ is not retract rational, where a desingularisation of 𝑋 ∘ can be
obtained via Hironaka’s theorem, as is required by Theorem 4.11.
Let 𝑅 ∈ 𝐿(C) be a point whose coordinates are in Q(𝑡), but not in Q. Then the
quartic threefold 𝑃𝑅 is isomorphic to 𝑋C∘ , for some embedding Q(𝑥) ↪ C. Indeed,
we have a diagram of pull-back squares
𝑃𝑅 𝑃 𝐿C 𝑃 Q, 𝐿 𝑋∘
𝑅 𝜂
Spec C 𝐿C 𝐿Q Spec Q(𝑥) ,
Therefore, 𝑃𝑅 is not retract rational. This shows that every line passing through
𝑀 and a point not in 𝑊 contains infinitely many points where the hypersurface is
not retract rational. This implies Zariski density. ◻
Together with results from §3, this result allows us to obtain general statements
on the irrationality of quartic threefolds over complex numbers.
Theorem 5.7. There are uncountably many stable equivalence classes in the family
of quartic hypersurfaces in P4C .
𝑋 ∶ 𝑥0 𝑥1 (𝑥0 + 𝑥1 )(𝑥0 − 𝑥1 ) = 0
Remark 5.8. In these two theorems, the degree 4 can be replaced by any positive
multiple of 4, since one can consider quartic threefolds in P4 with multiplicity
𝑚 > 1, which will be a non-reduced hypersurface of degree 4𝑚. We use the fact that
the Chow group of a “non-reduced variety” is isomorphic to that of its reduction
[Ful98, Example 1.3.1].
Theorem 5.9. There exist smooth quartic hypersurfaces in P4Q that are not univer-
sally CH0 -trivial over any field containing Q, and hence not retract rational over
any field containing Q, and in particular, over C.
By Galois descent and by Lemma 4.8, we choose a finite extension 𝐾∕Q, over
which 𝑋, 𝑋 ̃ and 𝑓 are defined, such that 𝑓 is universally CH0 -trivial.
Let 𝒪𝐾 be the ring of integers of 𝐾. Let 𝑈 ⊂ Spec 𝒪𝐾 be an open set, such that
there exists a map of 𝑈 -schemes
𝒻 ∶ 𝒳̃ → 𝒳 ,
𝐻 3 (𝒳̃𝜅(𝑣) , Z2 ) ≃ 𝐻 3 (𝑋
̃Q , Z2 ).
𝒪𝐾,𝑣 ⊂ 𝐴,
such that the residue field of 𝐴 is 𝜅(𝑣). Let 𝐿 be the fraction field of 𝐴.
Finally, there exists a smooth 𝐴-scheme whose special fibre is 𝒳𝜅(𝑣) . Indeed,
in the projective space P𝑁 4
𝐿 parametrising the quartic hypersurfaces in P𝐿 , the set
𝑁
of points in P𝐿 with coordinates in 𝒪𝐿 lying over 𝒳𝜅(𝑣) is Zariski dense. But there
is an open set of P𝑁𝐿 whose points correspond to smooth hypersurfaces.
Now we apply Theorem 4.11 to complete the proof. ◻
𝑋 ∘ = 𝒳 ×𝐴 k
𝑋 = 𝒳 ×𝐴 F,
NormE∕F (𝑢 + 𝛽𝑣 + 𝛽 2 𝑤) + 𝑥𝑦(𝑥 − 𝑦) = 0.
𝑢𝑣𝑤 + 𝑥𝑦(𝑥 − 𝑦) = 0.
6 Example: Cubic threefolds 127
the inverse image of each of the 3 singular points is a union of two surfaces
P2E , which intersect in a line P1E , which contains 3 singular points.
𝑋2 → 𝑋
• a number field,
Then there exist smooth cubic hypersurfaces in P4k that is not universally CH0 -
trivial over k, and hence not retract rational over k.
Proof. By the above construction, and by Theorem 4.7, the generic fibre 𝑋 ∘ , as a
smooth k-variety, is not universally CH0 -trivial. ◻
Br(𝑌 ) ≃ Z∕3.
with the action of the Galois group 𝐺 = Gal(E∕F) ≃ Z∕3 by permuting the coor-
dinates 𝑢, 𝑣, 𝑤 cyclically.
Let 𝑈 ⊂ 𝑋 be the smooth locus, so that 𝑈E ⊂ 𝑋E is the complement of the
three singular points (6.0.2).
Let 𝑉 ⊂ 𝑈 be the open set given by 𝑥𝑦 ≠ 0. Then 𝑈E ⧵ 𝑉E consists of six
irreducible components Δ𝑢,𝑥 , Δ𝑣,𝑥 , Δ𝑤,𝑥 , Δ𝑢,𝑦 , Δ𝑣,𝑦 , Δ𝑤,𝑦 , where for example, Δ𝑢,𝑥
is defined by 𝑢 = 𝑥 = 0. The group 𝐺 acts on them by permuting 𝑢, 𝑣, 𝑤.
Let 𝑝 ∶ 𝑉E → A1E ⧵ {0} be the projection given by (𝑢, 𝑣, 𝑤, 𝑥, 𝑦) ↦ 𝑥∕𝑦. The
generic fibre is isomorphic to the surface 𝑢𝑣𝑤 = 1 in the affine space A3E(𝑥) , which
6 Example: Cubic threefolds 129
is isomorphic to the open subset 𝑢𝑣 ≠ 0 of A2E(𝑥) , so that its Picard group is zero.
Moreover, the only non-integral fibre is the fibre at 1, which consists of three irre-
ducible components, each being a principal divisor of 𝑉E , since they are defined in
𝑉E by the equations 𝑢 = 0, 𝑣 = 0 and 𝑤 = 0 respectively. Applying Lemma 6.3,
we obtain Pic(𝑉E ) = 0.
Let Div𝑈E ⧵𝑉E (𝑈E ) denote the group of divisors of 𝑈E supported in 𝑈E ⧵ 𝑉E . It
is a free abelian group of rank 6, generated by the divisors Δ𝑢,𝑥 , etc. The canonical
map
𝛽 ∶ Div𝑈E ⧵𝑉E (𝑈E ) → Pic(𝑈E )
is surjective, as its image is the kernel of the restriction map Pic(𝑈E ) → Pic(𝑉E ), the
latter group being zero. The kernel of 𝛽 consists of those divisors that are principal
in 𝑈E . We thus have an exact sequence of 𝐺-modules
𝛼 𝛽
0 → E[𝑉E ]× ∕E[𝑈E ]× ⟶ Div𝑈E ⧵𝑉E (𝑈E ) ⟶ Pic(𝑈E ) → 0. (6.4.1)
Let us take a closer look at the first term. Suppose 𝑓 ∈ E[𝑉E ]× . Using the
projection 𝑝 ∶ 𝑉E → A1E ⧵ {0} mentioned above, we can apply Lemma 6.2 to K =
E(𝑥), to conclude that 𝑓 has the form 𝑓 = 𝑡(𝑥∕𝑦) 𝑢𝑚 𝑣𝑛 , for 𝑡 a rational function,
and 𝑚, 𝑛 ∈ Z. Since 𝑓 has to be invertible on 𝑉E , we must have 𝑚 = 𝑛 = 0, and
𝑡(𝑥∕𝑦) = 𝑐 (𝑥∕𝑦)𝑘 for some 𝑐 ∈ E× and 𝑘 ∈ Z. It follows that E[𝑉E ]× ≃ E× ⊕ Z
and E[𝑈E ]× ≃ E× . The sequence (6.4.1) is thus
𝛼 𝛽
0 → Z ⟶ Z[𝐺] ⊕ Z[𝐺] ⟶ Pic(𝑈E ) → 0,
and the map 𝛼 sends 𝑘 ∈ Z to the divisor of the function (𝑥∕𝑦)𝑘 , which is the
element (𝑘𝜀, −𝑘𝜀) ∈ Z[𝐺] ⊕ Z[𝐺], where 𝜀 = ∑𝑔∈𝐺 𝑔 ∈ Z[𝐺].
Now is where the proof really begins. The exact sequence (6.4.1) will not be
used anywhere in this proof; what we use is the sequence (6.4.1) with 𝑌 in place of
𝑈 , where 𝑌 is a desingularisation of 𝑋 as in the statement of this theorem. Such
a sequence is obtained by a process as in the above argument. This gives an exact
sequence of Tate cohomology groups
≃Z
𝛿
̂ −1 (𝐺, Pic(𝑌E )) ⟶ 𝐻
0→𝐻 ⏞⏞⏞⏞⏞⏞⏞
̂ 0 (𝐺, E[𝑉 × ×
E ] ∕E ) ≃ Z∕3
𝛼′
̂ 0 (𝐺, Div𝑌 ⧵𝑉 (𝑌E )) → ⋯ ,
⟶𝐻 E E
where the 𝐻 ̂ −1 of Div𝑌 ⧵𝑉 (𝑌E ) vanishes, since the latter is a direct sum of copies
E E
of Z[𝐺] and Z, which both have zero 𝐻 ̂ −1 by direct computation.
The generator 𝑥∕𝑦 of the second non-zero term goes to a divisor of 𝑌E which is
the norm (in the 𝐺-module sense) of a divisor of 𝑌E . Indeed, we have seen that the
divisor of the rational function 𝑥∕𝑦 on 𝑈E is the norm of an element. But 𝑌E ⧵𝑈E is
the inverse image of the three singular points of 𝑋E , and the action of 𝐺 permutes
these three parts of 𝑌E ⧵ 𝑈E . Therefore, the divisor of 𝑥∕𝑦 on 𝑌E ⧵ 𝑈E is the norm
130 Stable Irrationality of Varieties
of the divisor of 𝑥∕𝑦 on one of these three parts. This shows that 𝛼 ′ (𝑥∕𝑦) = 0, so
̂ −1 (𝐺, Pic(𝑌E )) ≃ Z∕3.
that 𝑥∕𝑦 is in the image of 𝛿. It follows that 𝐻
By [CTS77, Lemma 15], there is a short exact sequence
where Br(𝑌 , E) = ker(Br(𝑌 ) → Br(𝑌E )), and similarly for Br(F, E). Since F and E
are finite, one has Br(F, E) = 0 by Wedderburn’s theorem. Since 𝐺 is cyclic, one
̂ −1 (𝐺, Pic(𝑌E )) ≃ Z∕3. It follows that the middle term is
has 𝐻 1 (𝐺, Pic(𝑌E )) ≃ 𝐻
Z∕3. In other words, there is a short exact sequence
Recall the projection 𝑝 ∶ 𝑉E → A1E defined above. We have seen that the inverse
image of the complement of {0, 1} is isomorphic to the subset of A3E defined by
𝑢𝑣𝑥(𝑥 − 1) ≠ 0. This shows that 𝑌E is rational, so that Br(𝑌E ) ≃ Br(E) ≃ 0 by
Theorem 2.23 and Wedderburn’s theorem. Therefore, Br(𝑌 ) ≃ Z∕3. ◻
Irrationality
We are interested in the case 𝑆 = P2 , and we consider the family of all hypersur-
faces of P2 × P3 of bidegree (2, 2). A general member of this family is a quadric
surface bundle over P2 .
As before, we use a specific example to establish the irrationality of a very
general member.
𝜕𝜈
𝛼 ′ ∈ 𝐻 2 (C(𝑋), Z∕2) 𝐻 1 (𝜅(𝜈), Z∕2) .
132 Stable Irrationality of Varieties
Applying Theorem 2.23, and applying Theorem 3.12 to the family of bidegree
(2, 2) hypersurfaces in P2 × P3 , we obtain the following.
Proof. 𝑋 is rational, if and only if the generic fibre 𝑋 ∘ is rational over the field
C(𝑆). Since 𝑋 ∘ is a smooth quadric surface, it is rational if and only if it has a
rational point, as the projection from a rational point will give a birational map
between 𝑋 ∘ and P2 . Thus it suffices to show that 𝑋 ∘ has a C(𝑆)-rational point.
By a theorem of Springer [Spr52], 𝑋 ∘ has a C(𝑆)-rational point, if and only
if 𝑋 ∘ has a K-rational point for some extension K∕C(𝑆) of odd degree. Thus, we
7 Example: Quadric surface bundles 133
only need to show that 𝑋 ∘ has a 0-cycle of odd degree, which will imply that 𝑋 ∘
has a closed point of odd degree.
But by hypothesis, 𝑋∕𝑆 has a multisection of odd degree, which gives rise to
a 0-cycle of 𝑋 ∘ of odd degree. ◻
For a quadric surface bundle 𝑋 → 𝑆, and an integral (2, 2)-class, that is, an
element 𝛼 ∈ 𝐻 2,2 (𝑋) ∩ 𝐻 4 (𝑋; Z), we say that 𝛼 meets the fibre 𝑋𝑠 in degree 𝑑,
where 𝑠 ∈ 𝑆, if the pairing of 𝛼 with the homology class of 𝑋𝑠 equals 𝑑.
Proof. Let 𝑆0 ⊂ 𝑆 be the locus where the rank of the quadratic form is ≥ 3 (the
full rank is 4), and let 𝑋0 = 𝑋 ×𝑆 𝑆0 .
Let 𝐹1 → 𝑆 be the relative variety of lines of 𝜋, i.e., the points of the fibre (𝐹1 )𝑠
correspond to straight lines contained in the fibre 𝑋𝑠 . When 𝑋𝑠 is non-degenerate,
it contains 2 families of lines, each parametrised by P1 . When the rank of the
quadratic form drops by 1, 𝑋𝑠 becomes a quadric cone, which contains 1 family of
lines parametrised by P1 .
This shows that 𝐹1 |𝑆0 → 𝑆0 factors as
𝑝
𝐹1 |𝑆0 ⟶ 𝑇0 ⟶ 𝑆0 ,
Next, we analyse the Hodge classes in the case 𝑆 = P2 , in order to verify the
assumption of this lemma. The key tool is the following technique of Voisin.
Lemma 7.6. Let 𝑌 → 𝐵 be a flat, projective family of complex varieties. Suppose
there exists 𝑏 ∈ 𝐵 and 𝜆 ∈ 𝐻 𝑝,𝑝 (𝑌𝑏 , R), such that the infinitesimal period map
is surjective, where 𝑇𝐵,𝑏 denotes the tangent space of 𝐵 at 𝑏. Then for any open
set 𝑈 ⊂ 𝐵 (in euclidean topology) containing 𝑏, such that 𝑌 |𝑈 → 𝑈 is a trivial
bundle, the map (notations are explained below)
𝑝,𝑝 2𝑝
𝜙 ∶ ℋR |𝑈 ↪ ℋR |𝑈 ≃ 𝐻 2𝑝 (𝑌𝑏 , R) × 𝑈 → 𝐻 2𝑝 (𝑌𝑏 , R) → 𝐹 𝑝−1 𝐻 2𝑝 (𝑌𝑏 , R)
is submersive at 𝜆.
𝑝,𝑞
We use the notation ℋ 𝑝,𝑞 , ℋR , etc., to refer to the vector bundles over 𝐵,
whose fibres are the cohomology of the fibres of 𝑌 → 𝐵. The notation 𝐹 𝑝−1 𝐻 2𝑝
refers to the Hodge filtration, and in this case, it is equal to 𝐻 𝑝−1,𝑝+1 ⊕ 𝐻 𝑝,𝑝 ⊕ ⋯ ⊕
𝐻 2𝑝,0 .
Proof. See, for example, [Voi07, §5.3.4]. ◻
In the following, we use the notation
𝑌 →𝐵
for the family of all smooth bidegree (2, 2) hypersurfaces in P2 ×P3 , and the notation
𝑌𝑏 refers to its fibres.
Proposition 7.7. The Hodge and Betti numbers of 𝑌𝑏 are given by
• 𝑏0 = 𝑏8 = 1.
• 𝑏1 = 𝑏3 = 𝑏5 = 𝑏7 = 0.
• 𝑏2 = ℎ1,1 = 𝑏6 = ℎ3,3 = 2.
• 𝑏4 = 46, ℎ0,4 = ℎ4,0 = 0, ℎ1,3 = ℎ3,1 = 3, ℎ2,2 = 40.
Proof. The Lefschetz hyperplane theorem shows that
for 𝑘 < 4 and 𝑝 + 𝑞 < 4. This, together with Poincaré/Serre duality, gives the first
three items.
We compute 𝑏4 by analysing the map 𝑌𝑏 → P2 . Let 𝐷 ⊂ P2 be the locus of
degenerate fibres. If 𝑌𝑏 is defined by the equation
2 3
∑𝑖,𝑗=0 ∑𝑘,𝑙=0 𝑎𝑖𝑗𝑘𝑙 𝑥𝑖 𝑥𝑗 𝑦𝑘 𝑦𝑙 = 0,
7 Example: Quadric surface bundles 135
where the coefficients 𝑎𝑖𝑗𝑘𝑙 are assumed to be symmetric with respect to 𝑖, 𝑗 and
𝑘, 𝑙, then 𝐷 is cut out by the equation
Therefore, 𝐷 is an octic curve, and hence has genus 21 and Euler number −40.
Recall that for a complex variety 𝑋 and a closed subvariety 𝑍 ⊂ 𝑋, we have
an additive formula 𝜒(𝑋) = 𝜒(𝑍) + 𝜒(𝑋 ⧵ 𝑍) of Euler numbers. Hence we have
is given by multiplication
is surjective. ◻
Theorem 7.9. The set of those 𝑏 ∈ 𝐵 such that 𝑌𝑏 has an integral (2, 2)-class
meeting the fibres of 𝑌𝑏 → P2 in odd degree is dense in 𝐵, in euclidean topology.
Proof. Instead of finding an integral class, we only need to find such a class with
the coefficient ring
𝑅 = {𝑚∕𝑛 ∣ 𝑚, 𝑛 ∈ Z, 2 ∤ 𝑛},
as we can multiply by an odd integer to turn such a class into an integral class. (We
have an obvious definition of an odd element in 𝑅.)
Let 𝑏0 ∈ 𝐵 be as in the previous corollary, and let 𝑈 ⊂ 𝐵 be an open set which
trivialises 𝑌 → 𝐵 near 𝑏0 . Such a trivialisation preserves the homology classes of
the fibres of 𝑌𝑏 → P2 .
We have shown that 𝐻 0,4 (𝑌𝑏0 ) = 0. Thus, Lemma 7.6 shows that the image of
the map
2,2
𝜙 ∶ ℋR |𝑈 → 𝐻 4 (𝑌𝑏0 , R)
contains an open set. Since the image consists of those classes that are of type (2, 2)
over some 𝑏 ∈ 𝑈 , it suffices to show that the elements of 𝐻 4 (𝑌𝑏0 , 𝑅) that meet the
fibres of 𝑌𝑏0 → P2 in odd degree are dense in 𝐻 4 (𝑌𝑏0 , R), so that one such element
lies in the image. We only need to prove this for 𝑏0 ∈ 𝐵, as the set of such 𝑏0 is
Zariski open in 𝐵.
The quadric surface bundle 𝑌𝑏0 has a constant section P2 → 𝑌𝑏0 given by 𝑠 =
𝑡 = 𝑢 = 0. This gives rise to an element 𝛼 ∈ 𝐻 4 (𝑌𝑏0 , Z), which intersects the fibres
of 𝑌𝑏0 → P2 in degree 1. For any 𝛽 ∈ 𝐻 4 (𝑌𝑏0 , 𝑅), the class 𝛼 + 2𝛽 also intersects
the fibres of 𝑌𝑏0 → P2 in odd degree. Such classes are dense in 𝐻 4 (𝑌𝑏0 , 𝑅). ◻
The results are summarised as follows.
Theorem 7.10. In the family of all bidegree (2, 2) hypersurfaces in P2 × P3 , a very
general member is not retract rational, while the rational members form a dense
subset in the family, in euclidean topology. ◻
References 137
References
[AM72] M. Artin and D. Mumford (1972). “Some elementary examples of unirational
varieties which are not rational”. Proceedings of the London Mathematical
Society 25 (3), 75–95.
[BC94] V. V. Batyrev and D. A. Cox (1994). “On the Hodge structure of projective
hypersurfaces in toric varieties”. Duke Mathematical Journal 75 (2), 293–
338.
[Blo10] S. Bloch (2010). Lectures on Algebraic Cycles. 2nd ed. New Mathematical
Monographs 16. Cambridge University Press.
[Bou06] N. Bourbaki (2006). Algèbre Commutative, Chapitres 8 et 9. Springer.
[CT05] J.-L. Colliot-Thélène (2005). “Un théorème de finitude pour le groupe de
Chow des zéro-cycles d’un groupe algébrique linéaire sur un corps 𝑝-adique”.
Inventiones mathematicae 159, 589–606.
[CT95] J.-L. Colliot-Thélène (1995). “Birational invariants, purity and the Gersten
conjecture”. 𝐾-Theory and Algebraic Geometry: Connections with Quadratic
Forms and Division Algebras, Part 1. Proceedings of Symposia in Pure Math-
ematics 58. AMS, 1–64.
[CTP16] J.-L. Colliot-Thélène and A. Pirutka (2016). “Hypersurfaces quartiques de di-
mension 3 : non-rationalité stable”. Annales Scientifiques de l’École Normale
Supérieure 49 (2), 371–397.
[CTS19] J.-L. Colliot-Thélène and A. N. Skorobogatov (2019). The Brauer–
Grothendieck group. Preprint.
[CTS77] J.-L. Colliot-Thélène and J.-J. Sansuc (1977). “La 𝑅-équivalence sur les
tores”. Annales Scientifiques de l’École Normale Supérieure 10 (2), 175–229.
[Deb03] O. Debarre (2003). “Variétés rationnellement connexes”. Astérisque 290,
243–266.
[EGA-II] A. Grothendieck (1961). “Éléments de géométrie algébrique II. Étude glob-
ale élémentaire de quelques classes de morphismes”. Publications mathéma-
tiques de l’I.H.É.S. 8, 5–222.
[EGA-IV3] A. Grothendieck (1966). “Éléments de géométrie algébrique IV. Étude locale
des schémas et des morphismes de schémas. 3e partie”. Publications mathé-
matiques de l’I.H.É.S. 28, 5–255.
[EKW16] H. Esnault, M. Kerz, and O. Wittenberg (2016). “A restriction isomorphism
for cycles of relative dimension zero”. Cambridge Journal of Mathematics
4 (2), 163–196.
[FAG] B. Fantechi et al. (2005). Fundamental Algebraic Geometry: Grothendieck’s
FGA Explained. Mathematical Surveys and Monographs 123. AMS.
[Ful98] W. Fulton (1998). Intersection Theory. 2nd ed. Ergebnisse der Mathematik
und ihrer Grenzgebiete 2. Springer.
[GH94] P. Griffiths and J. Harris (1994). Principles of Algebraic Geometry. Wiley
Classics Library. John Wiley & Sons.
138 Stable Irrationality of Varieties
[Gro68] A. Grothendieck (1968). “Le groupe de Brauer I, II, III”. Dix exposés sur la
cohomologie des schémas. Masson & North-Holland, 46–188.
[HPT16] B. Hassett, A. Pirutka, and Yu. Tschinkel. Stable rationality of quadric surface
bundles over surfaces. arXiv: 1603.09262.
[Kol96] J. Kollár (1996). Rational curves on algebraic varieties. Ergebnisse der Math-
ematik und ihrer Grenzgebiete 32. Springer.
[Mil80] J. S. Milne (1980). Étale Cohomology. Princeton Mathematical Series 33.
Princeton University Press.
[Ros96] M. Rost (1996). “Chow groups with coefficients”. Documenta Mathematica
1, 319–393.
[Sal76] G. Salmon (1876). Lessons Introductory to the Modern Higher Algebra.
3rd ed. Dublin.
[Ser79] J.-P. Serre (1979). Local Fields. Trans. by M. J. Greenberg. Graduate Texts in
Mathematics 67. Springer.
[Spr52] T. A. Springer (1952). “Sur les formes quadratiques d’indice zéro”. C. R.
Acad. Sci. Paris 234, 1517–1519.
[Voi07] C. Voisin (2007). Hodge Theory and Complex Algebraic Geometry II. Trans.
by L. Schneps. Cambridge Studies in Advanced Mathematics 77. Cambridge
University Press.
[Voi15] C. Voisin (2015). “Unirational threefolds with no universal codimension 2
cycle”. Inventiones mathematicae 201 (1), 207–237.
Applications of the Eells–Kuiper Invariant
to Exotic 7‐Spheres
Lan Qing1
ABSTRACT
We introduce the Eells–Kuiper invariant, and apply the Eells–
Kuiper invariant to find the 28 differentiable structures on 𝑆 7 . We
also apply it to circle bundles over homotopy ℂ𝑃 3 , to show that any
homotopy 7-sphere that admits smooth regular 𝑆 1 -actions is realized
as the total space of a principal 𝑆 1 -bundle over some homotopy ℂ𝑃 3 ,
with primitive Euler class.
Contents
1 Introduction 140
1 Introduction
In 1956, Milnor discovered that there exists a differentiable structure on
𝑆 7 that is different from the usual one in [Mil56], where he introduced a
differential invariant and applied it to a collection of 𝑆 3 -bundles over 𝑆 4 .
From the theory of h-cobordism discussed in [Mil59a] and [Sma61], there
are exactly 28 differentiable structures on 𝑆 7 . In 1962, Eells and Kuiper in-
troduced 𝜇 invariant which takes different values for different differentiable
structures on 𝑆 7 , completely classifying the 28 differentiable structures on
𝑆 7 . In [KM63], the group Θ𝑛 of homotopy spheres is discussed, where for
𝑛 ≥ 5, Θ𝑛 describes exotic spheres of dimension 𝑛.
In this thesis, we introduce the invariants of Milnor and Eells–Kuiper,
and apply the Eells–Kuiper invariant to the collection of 𝑆 3 -bundles over
𝑆 4 defined in [Mil56] to find 16 differentiable structures on 𝑆 7 , and then
using connected sum to find the 28 differentiable structures on 𝑆 7 .
We also apply a generalized Eells–Kuiper invariant to circle bundles
over homotopy ℂ𝑃 3 , to show that any homotopy 7-sphere that admits
smooth regular 𝑆 1 -actions is realized as the total space of a principal
𝑆 1 -bundle over some homotopy ℂ𝑃 3 , with primitive Euler class.
Throughout this thesis we will use ℚ as the coefficient ring, unless oth-
erwise stated.
The 𝜆 invariant
In [Mil56], Milnor introduced 𝜆 invariant for the discovery of exotic
7-spheres. For an oriented closed 7-manifold 𝑀 such that 𝐻 3 (𝑀) =
𝐻 4 (𝑀) = 0, since the seventh oriented cobordism ring Ω𝑆𝑂7 = 0, 𝑀
bounds a compact oriented 8-manifold 𝐵. The assumption on the cohomol-
ogy of 𝑀 implies that we have an isomorphism
𝑗 ∶ 𝐻 4 (𝐵, 𝑀) → 𝐻 4 (𝐵).
2 The invariants of Milnor 141
Let 𝜈 be the orientation class in 𝐻8 (𝐵, 𝑀), and let 𝜏(𝐵) be the signature
of 𝐵. Then we have a well-defined differential invariant
Thus
and then
becomes trivial. In essentially the same way as above we can show that
the 𝜆 invariant is well-defined, since two definitions will differ by some top
Pontrjagin number which is an integer.
In the special case where 𝑘 = 2,
45
𝜆(𝑀) = (𝜏(𝑊 ) − ⟨(7𝑗 −1 𝑝2 − 𝑗 −1 𝑝21 )∕45, [𝑊 , 𝑀]⟩)
7
3 1
≡ 𝜏(𝑊 ) + ⟨𝑗 −1 𝑝21 , [𝑊 , 𝑀]⟩ mod 1.
7 7
Trivially for integers 𝑎, 𝑏, 2(𝑎 + 3𝑏) ≡ 2𝑎 − 𝑏 mod 7 and 4(2𝑎 − 𝑏) ≡ 𝑎 +
3𝑏 mod 7. Consequently this 𝜆 invariant and the previous one defined for
7-manifolds indeed determine each other.
The 𝜆′ invariant
We say that a manifold 𝑊 is almost parallelizable if for some finite subset
𝐹 , 𝑊 − 𝐹 is parallelizable, i.e. 𝑀 − 𝐹 has trivial tangent bundle.
Let 𝜏𝑘 be the greatest common divisor of all 𝜏(𝑊 ) where 𝑊 ranges over
all almost parallelizable 4𝑘-manifolds without boundary.
Suppose 𝑀 is a smooth oriented homology (4𝑘 − 1)-sphere over ℤ and
𝑀 bounds a compact smooth oriented parallelizable manifold 𝑊 . Thus
𝜏(𝑊 ) mod 𝜏𝑘 is a well-defined invariant of 𝑀. But in fact, under the as-
sumption above 𝜏(𝑊 ) and 𝜏𝑘 are always divisible by 8.
The domain of 𝜇
We would like to define another invariant for certain manifolds. An oriented
manifold is said to be spin if its second Stiefel-Whitney class is zero.
Let 𝑀 be a closed smooth oriented (4𝑘 − 1)-manifold that bounds a
compact smooth oriented spin 4𝑘-manifold 𝑊 , such that
(a) over ℚ,
𝑖∗ ∶ 𝐻 1 (𝑊 ; ℤ2 ) → 𝐻 1 (𝑀; ℤ2 ).
3 The Eells–Kuiper 𝜇 invariant 143
Again condition (a) implies 𝐻 2𝑘−1 (𝑀) = 𝐻 4𝑖−1 (𝑀) = 0, ∀0 < 𝑖 <
𝑘, so 𝑗 ∶ 𝐻 4𝑖 (𝑊 , 𝑀) → 𝐻 4𝑖 (𝑊 ) and 𝑗 ∶ 𝐻 2𝑘 (𝑊 , 𝑀) → 𝐻 2𝑘 (𝑊 ) are
still isomorphisms. Indeed, condition (a) can be replaced by the slightly
weaker condition requiring them to be isomorphisms, so that one can pull
back Pontrjagin classes of 𝑊 . We call this condition (a’).
ℤF 0 0 0 ℤ
F FF
FF
FF
FF
0 0 FFF 0 0 0
FF
FF
FF
FF
FF
0 0 0 FF0 0
FF
FF
FF
FF
#
ℤ 0 0 0 ℤ
The following diagram is the 𝐸∞ page over ℚ, assuming the Euler class
does not vanish.
144 Applications of the Eells–Kuiper Invariant to Exotic 7‐Spheres
0 0 0 0 ℚ
0 0 0 0 0
0 0 0 0 0
ℚ 0 0 0 0
ℚ 0 0 0 ℚ
0 0 0 0 0
0 0 0 0 0
ℚ 0 0 0 ℚ
The definition of 𝜇
𝐴̂𝑘 [𝑋] = 𝐴̂𝑘 (𝑝1 , … , 𝑝𝑘−1 , 0)[𝑋] + 𝐴̂𝑘 (0, 0, … , 1)𝑝𝑘 [𝑋]
where
𝐴̂𝑘 (0, 0, … , 1)
𝑡𝑘 =
𝐿𝑘 (0, 0, … , 1)
𝑁𝑘 (𝑝1 , … , 𝑝𝑘−1 )[𝑋] = 𝐴̂𝑘 (𝑝1 , … , 𝑝𝑘−1 , 0)[𝑋] − 𝑡𝑘 𝐿𝑘 (𝑝1 , … , 𝑝𝑘−1 , 0)[𝑋].
Proof. This is essentially the same as in the previous section. Let (𝑊1 , 𝑀1 )
and (𝑊2 , 𝑀2 ) be two such choices for the space 𝑀. Reverse the orientation
of the latter and glue the two manifolds along the boundary to get 𝑋. Let 𝑟
be 2𝑘 or 4𝑖(0 < 𝑖 < 𝑘). We have a commutative diagram
zero
we find 𝐻 𝑟 (𝑋) → 𝐻 𝑟 (𝑊1 ) → 𝐻 𝑟 (𝑀) is the zero map. Thus
zero j
𝐻 𝑟 (𝑀) ← 𝐻 𝑟 (𝑋) ← 𝐻 𝑟 (𝑋, 𝑀)
implies that 𝑗 is epic.
It follows from easy diagram chasing that 𝑗 is an isomorphism and thus
𝑘 is an isomorphism.
If 𝛼 = 𝑗ℎ−1 (𝛼1 ⊕ 𝛼2 ) ∈ 𝐻 2𝑘 (𝑋), then
2 −1 2 2
⟨𝜈, 𝛼 ⟩ = ⟨𝜈, 𝑗ℎ (𝛼1 ⊕ 𝛼2 )⟩
= ⟨𝜈1 ⊕ (−𝜈2 ), 𝛼12 ⊕ 𝛼22 ⟩ = ⟨𝜈1 , 𝛼12 ⟩ − ⟨𝜈2 , 𝛼22 ⟩
where 𝜈 = [𝑋], 𝜈1 = [𝑊1 , 𝑀], 𝜈2 = [𝑊2 , 𝑀] are orientation classes. Thus
the intersection form of 𝑋 splits into the direct sum of those of 𝑊1 and 𝑊2 ,
and
𝜏(𝑋) = 𝜏(𝑊1 ) − 𝜏(𝑊2 ).
Clearly, the pullback of the tangent bundle of 𝑋 by inclusion 𝑖𝑙 ∶ 𝑊𝑙 →
𝑋 is the tangent bundle of 𝑊𝑙 , 𝑙 = 1, 2. By naturality, 𝐻 ∗ (𝑋) → 𝐻 ∗ (𝑊1 )⊕
𝐻 ∗ (𝑊2 ) → 𝐻 ∗ (𝑊𝑙 ), 𝑝𝑠 (𝑋) ↦ 𝑝𝑠 (𝑊𝑙 ) for any Pontrjagin class 𝑝𝑠 . So
𝑘(𝑝𝑠 (𝑋)) = 𝑝𝑠 (𝑊1 ) ⊕ 𝑝𝑠 (𝑊2 ). As above, for any monomial 𝑓 , 𝑓 (𝑝(𝑋))
becomes the direct sum of 𝑓 (𝑝(𝑊1 )) and 𝑓 (𝑝(𝑊2 )), so
𝑓 (𝑝1 , … , 𝑝𝑘−1 , 0)[𝑋] = ⟨𝑓 (𝑗1−1 𝑝1 (𝑊1 ), … , 𝑗1−1 𝑝𝑘−1 (𝑊1 ), 0), [𝑊1 , 𝑀]⟩
−⟨𝑓 (𝑗2−1 𝑝1 (𝑊2 ), … , 𝑗2−1 𝑝𝑘−1 (𝑊2 ), 0), [𝑊2 , 𝑀]⟩.
For simplicity, write
𝑓 (𝑝1 , … , 𝑝𝑘−1 , 0)[𝑊 ] = ⟨𝑓 (𝑗 −1 𝑝1 (𝑊 ), … , 𝑗 −1 𝑝𝑘−1 (𝑊 ), 0), [𝑊 , 𝑀]⟩.
The equality above becomes
𝑓 (𝑝1 , … , 𝑝𝑘−1 , 0)[𝑋] = 𝑓 (𝑝1 , … , 𝑝𝑘−1 , 0)[𝑊1 ] − 𝑓 (𝑝1 , … , 𝑝𝑘−1 , 0)[𝑊2 ].
Combining results above, with 𝑓 replaced by 𝑁𝑘 ,
̂
𝜇(𝑀1 ) − 𝜇(𝑀2 ) = 𝐴[𝑋]∕𝑎 𝑘
Rohlin invariant
This subsection is added to show that in 3-dimensional case, in the com-
mon domain of the Rohlin invariant and the Eells–Kuiper invariant, these
invariants are equivalent. This will not be used in the applications.
𝑒𝑖 , if 𝑖 = 𝑗
𝑎𝑖𝑗 =
{ lk(𝐿𝑖 , 𝐿𝑗 ), if 𝑖 ≠ 𝑗
𝜏(𝑊 )
𝜈(Σ) = mod 1.
16
This is well-defined because of the theorem of Rohlin:
Theorem 5. If 𝑀 is a simply-connected, closed, smooth, oriented 4-
manifold with even intersection form then 𝜏(𝑀) ≡ 0 mod 16.
Proposition 2. 𝜈(−𝑀) = −𝜈(𝑀), and 𝜈(𝑀1 #𝑀2 ) = 𝜈(𝑀1 ) + 𝜈(𝑀2 ).
There is another definition of Rohlin invariant, using the fact that ho-
mology 3-spheres have a unique spin structure, and the third spin cobordism
ring is zero. Define the Rohlin invariant of a homology 3-sphere to be
𝜏(𝑊 )
𝜈(Σ) = mod 1,
16
where 𝑊 is any smooth compact spin 4-manifold with spin boundary the
homology sphere.
To show that this is independent of 𝑊 , there is a corresponding version
of Rohlin’s theorem:
Theorem 6. The signature of a smooth closed spin 4-manifold 𝑋 is divisible
by 16.
This extends the previous definition. In fact, if the manifold 𝑊 is ob-
tained from 𝐷4 by gluing 2-handles according to an even surgery on a link in
𝑆 3 , then the canonical spin structure on 𝐷4 extends to 𝑊 . See Proposition
5.7.1 in [GS99] for a proof.
Comparison with 𝝁
Let Σ be a compact oriented integral homology 3-sphere. From universal
coefficient theorem, since for any 𝑘, 𝐻𝑘 (𝑆 3 ; ℤ) = 𝐻𝑘 (Σ; ℤ) is free over ℤ,
all Tor and Ext terms vanish and thus Σ is a compact oriented homology
3-sphere over any ℤ-algebra and in particular over ℚ and ℤ2 .
To show that Σ is in the domain of 𝜇, note that the third spin cobordism
ring is zero, so Σ always spin bounds a spin manifold. Since 𝐻 2 (Σ; ℚ) = 0,
condition (a) is satisfied. Since 𝐻 1 (Σ; ℤ2 ) = 0, condition (b) is satisfied.
From direct calculation one shows that for any 3-manifold 𝑀 in the
domain of 𝜇,
𝜏(𝑊 )
𝜇(𝑀) = − mod 1.
16
4 Applications to 𝑆 3 bundles over 𝑆 4 149
0 1
(ℎ + 𝑗 1)
Proof. Note that 𝜙𝑖+𝑗,𝑘+𝑙 (𝑞) = 𝜙𝑖,𝑘 (𝑞)𝜙𝑗,𝑙 (𝑞) for each 𝑞 ∈ 𝑆 3 , where the
multiplication on right hand side is the multiplication in 𝑆𝑂4 . But in a
Lie group 𝑆𝑂4 , 𝑞 ↦ 𝜙𝑖,𝑘 (𝑞)𝜙𝑗,𝑙 (𝑞) is homotopic to the sum in 𝜋3 (𝑆𝑂4 ) of
𝑞 ↦ 𝜙𝑖,𝑘 (𝑞) and 𝑞 ↦ 𝜙𝑗,𝑙 (𝑞). ◻
for some 𝛼, 𝛽, 𝛾, 𝛿.
for some 𝛼, 𝛾.
To find out the coefficients, it suffices to calculate for a concrete 𝐸𝑖𝑗 , say
𝐸01 .
Note that the clutching function of 𝐸01 is complex linear, which implies
that it can be identified with a complex vector bundle. The space 𝑆𝐸01 is
4 Applications to 𝑆 3 bundles over 𝑆 4 151
ℤF 0 0 0 ℤ
F FF
FF
FF
FF
0 0 FFF 0 0 0
FF
FF
FF
FF
FF
0 0 0 FF0 0
FF
FF
FF
FF
#
ℤ 0 0 0 ℤ
we see that the Euler class must give rise to an isomorphism. Thus
𝑒(𝐸𝑖𝑗 ) = ±(𝑖 + 𝑗) ∈ 𝐻 4 (𝑆 4 ; ℤ) ≅ ℤ
Consequently,
𝑝1 (𝐸𝑖𝑗 ) = ±2(𝑖 − 𝑗) ∈ 𝐻 4 (𝑆 4 ; ℤ) ≅ ℤ.
𝑒(𝐸𝑖𝑗 ) = 𝑖 + 𝑗 ∈ 𝐻 4 (𝑆 4 ; ℤ) ≅ ℤ
𝑝1 (𝐸𝑖𝑗 ) = 2(𝑖 − 𝑗) ∈ 𝐻 4 (𝑆 4 ; ℤ) ≅ ℤ.
Differentiable structures on 𝑆 7
If 𝑘 = 2, the formula for the Eells–Kuiper invariant becomes
(𝑝21 [𝑊 ] − 4𝜏[𝑊 ])
𝜇(𝑀 7 ) ≡ mod 1.
(27 × 7)
which is dual to
𝑄 ∶ 𝐻2𝑘 (𝑀) ⊗ 𝐻2𝑘 (𝑀) → ℤ,
𝑖∗
⋯ → 𝐻2𝑘+1 (𝑀, 𝜕𝑀) → 𝐻2𝑘 (𝜕𝑀) → 𝐻2𝑘 (𝑀)
→ 𝐻2𝑘 (𝑀, 𝜕𝑀) → 𝐻2𝑘−1 (𝜕𝑀) → ⋯
Consequently
𝑝21 [𝐵2ℎ−1 ] = 4(ℎ − 𝑗)2 = 4(2ℎ − 1)2 .
0, 1, 3, 6, 7, 8, 10, 13, 14, 15, 17, 20, 21, 22, 24, 27.
Proof. For each of the values in the lemma above, there are infinitely many
values of ℎ corresponding to it. ◻
Proposition 3. For any 7-dimensional manifold 𝑀 in the domain of 𝜇,
the underlying topological manifold has at least 28 different differentiable
structures.
Proof. From the cell structure we see that ℂ𝑃 3 is simply connected. The
second page 𝐸2 of the Serre spectual sequence associated to 𝑆𝐸𝑑 is
ℤ OO 0 ℤ OO 0 ℤ OO 0 ℤ
OOO OOO OOO
OOO𝑑 OOO𝑑 OOO𝑑
OOO OOO OOO
OOO OOO OOO
' ' '
ℤ 0 ℤ 0 ℤ 0 ℤ
Lemma 6.
⎧ ℤ, if 𝑑 = 0
⎪
𝜋1 (𝑆𝐸𝑑 ) = ⎨ 0, if 𝑑 = ±1
⎪
⎩ ℤ∕|𝑑|ℤ, otherwise.
𝐻 ∗ (ℂ𝑃 3 ; ℤ) = ℤ[𝑣]∕(𝑣4 ).
Proof.
Here we use the fact that when considering the real bundle underlying a
complex rank 𝑛 vector bundle 𝐸, 𝑤2 (𝐸) ≡ 𝑐1 (𝐸) mod 2. ◻
It is possible to apply the Eells–Kuiper invariant to find out the diffeo-
morphism classes of the homotopy spheres 𝑆𝐸1 and 𝑆𝐸−1 , which are spin.
Also since 𝐸1 is just the complex conjugate of 𝐸−1 , they are the same as
real vector bundles and thus 𝑆𝐸1 and 𝑆𝐸−1 are diffeomorphic, via an ori-
entation reversing diffeomorphism. It suffices to calculate the Eells–Kuiper
invariant for 𝑆𝐸−1 . However, 𝐷𝐸−1 is not spin. We cannot apply the for-
mula in previous sections directly to the pair (𝐷𝐸−1 , 𝑆𝐸−1 ).
̂ ), [𝑊 , 𝜕𝑊 ]
𝑆1 (𝑊 , 𝑧, 𝑐) = ⟨𝑒(𝑐+𝑧)∕2 𝐴(𝑊 ⟩
̂ ), [𝑊 , 𝜕𝑊 ]
𝑆2 (𝑊 , 𝑧, 𝑐) = ⟨ch(𝜆(𝑧) − 1)𝑒(𝑐+𝑧)∕2 𝐴(𝑊 ⟩
̂ ), [𝑊 , 𝜕𝑊 ] .
𝑆3 (𝑊 , 𝑧, 𝑐) = ⟨ch(𝜆2 (𝑧) − 1)𝑒(𝑐+𝑧)∕2 𝐴(𝑊 ⟩
Here 𝜆(𝑧) is the complex line bundle over 𝑊 with first Chern class 𝑧,
ch is the Chern character, and 𝐴(𝑊̂ ) is the 𝐴-polynomial
̂ of 𝑊 . Since we
4 2
are requiring 𝐻 (𝑀; ℚ) = 0, 𝑝1 and 𝑧 can be considered as elements of
𝐻 4 (𝑊 , 𝑀; ℚ) = 0 and thus 𝑝21 , 𝑝1 𝑧2 and 𝑧4 can be evaluated on the funda-
mental class [𝑊 , 𝜕𝑊 ] = [𝑊 , 𝑀]. However, 𝑝2 appears in 𝑒(𝑐+𝑧)∕2 𝐴(𝑊 ̂ ),
which cannot be considered as a relative class in general. Our strategy is
the same as for 𝜇 invariant, using the signature of 𝑊 to eliminate 𝑝2 from
the expression, i.e. replacing some constant multiple of 𝐿(𝑊 ) (evaluated on
the fundamental class) by some constant multiple of 𝜏(𝑊 ), so that 𝑝2 term
disappears.
These characteristic numbers for closed manifolds are integers. For
a proof, see Theorem 26.1.1. in [HBS13]. It follows that 𝑠𝑖 (𝑀, 𝑢) =
𝑆𝑖 (𝑊 , 𝑧, 𝑐) ∈ ℚ∕ℤ, 𝑖 = 1, 2, 3 are well-defined.
Note that, if 𝑊 is spin and 𝑀 has the induced spin structure, we may
take 𝑧 = 0, 𝑐 = 0 and thus
̂ ), [𝑊 , 𝜕𝑊 ]⟩ = 𝜇(𝑀).
𝑠1 (𝑀, 0) = ⟨𝐴(𝑊
So 𝑠1 is a generalization of 𝜇.
If 𝑐 = 0, explicit formulas are: for spin case
1 1
𝑆1 (𝑊 , 𝑧, 0) = − 𝜏(𝑊 ) + 7 𝑝21 ,
25 ⋅ 7 2 ⋅7
for nonspin case
1 1 1 1
𝑆1 (𝑊 , 𝑧, 0) = − 𝜏(𝑊 ) + 7 𝑝21 − 6 𝑧2 𝑝1 + 7 𝑧4 ,
25⋅7 2 ⋅7 2 ⋅3 2 ⋅3
where, by abuse of notation, everything is considered as a relative cohomol-
ogy class and is evaluated on [𝑊 , 𝜕𝑊 ] = [𝑊 , 𝑀]. If 𝑀 is spin but 𝑊 is
not, then
1 1 1 1
𝑆1 (𝑊 , 𝑧, 𝑐) = − 𝜏(𝑊 ) + 7 𝑝21 − 6 (𝑧 + 𝑐)2 𝑝1 + 7 (𝑧 + 𝑐)4 .
25⋅7 2 ⋅ 7 2 ⋅3 2 ⋅3
5 Applications to 𝑆 1 bundles over homotopy ℂ𝑃 3 157
𝑝1 ∶ 𝐻 2 (𝑀; ℤ) ≅ 𝐻4 (𝑀; ℤ) ≅ ℤ → ℤ
subject to
line bundle over 𝑋𝑘 with first Chern class 𝑑𝑣, and define 𝑆𝐸𝑑 , 𝐷𝐸𝑑 in the
same way. One verifies that only 𝑆𝐸1 and 𝑆𝐸−1 are homeomorphic to 𝑆 7
as in the previous sections, and it suffices to consider 𝑆𝐸−1 .
𝑋𝑘 is spin, so 𝑤2 (𝑇 𝐷𝐸𝑑 ) ≡ 𝑑 ⋅ 𝜋 ∗ 𝑣 mod 2 and it is zero iff 𝑑 is divisible
by 2, and in particular 𝐷𝐸−1 is not spin. 𝑆𝐸−1 is spin, being a homotopy
sphere.
Take 𝑢 = 0, 𝑧 = 0, 𝑐 = 𝜋 ∗ 𝑣. Since
The 𝑠1 invariant is
6 Further Discussions
For more information about 𝑆 1 -actions on homotopy spheres, see [Hsi66],
[MY68] and [Sch71].
The topology of sphere bundles over spheres has been discussed in
[Jam62a], [Jam62b]. A complete classification of 𝑆 3 -bundles over 𝑆 4 is
done in [CM03].
In [KM63], the group Θ𝑛 of homotopy spheres is discussed. The h-
cobordism classes of homotopy 𝑛-spheres form an abelian group Θ𝑛 under
connected sum operation. Actually, for 𝑛 ≥ 5 homotopy 𝑛-spheres are in-
deed homeomorphic to 𝑆 𝑛 , and two homotopy 𝑛-spheres are h-cobordant
if and only if they are diffeomorphic. Consequently Θ𝑛 describes exotic
spheres of dimension 𝑛, for 𝑛 ≥ 5.
A generalized version of the Eells–Kuiper invariant is the Kreck–Stolz
𝑠-invariant. See [TW15].
References
[BH58] A. Borel and F. Hirzebruch (1958). “Characteristic Classes and Homo-
geneous Spaces, I”. American Journal of Mathematics 80 (2), 458–538.
ISSN: 00029327, 10806377. URL: http://www.jstor.org/stable/
2372795.
[Bre13] G.E. Bredon (2013). Topology and Geometry. Graduate Texts in Math-
ematics. Springer New York. ISBN: 9781475768480. URL: https://
books.google.com.hk/books?id=wuUlBQAAQBAJ.
[BT13] R. Bott and L.W. Tu (2013). Differential Forms in Algebraic Topol-
ogy. Graduate Texts in Mathematics. Springer New York. ISBN:
9781475739510. URL: https://books.google.com.hk/books?
id=COuPBAAAQBAJ.
160 Applications of the Eells–Kuiper Invariant to Exotic 7‐Spheres
Xie Yuxiao1
ABSTRACT
Contents
1 Introduction 164
1
谢雨潇, 清华大学数学系数 80 班.
164 Transversality Theorems by Example
convention. Unless otherwise stated, the word “manifold” refers to one which has
no boundary but which is not necessarily smooth. Whenever we write 𝐶 𝑘 or 𝑊 𝑘,𝑝 ,
it is tacitly understood that 𝑘 ≥ 1, 1 < 𝑝 < ∞.
1 Introduction
A transversality theorem is a theorem stating that a certain desirable property can
be achieved by an arbitrarily small perturbation, or that the property in question is
generic. We begin by making this notion precise.
Let 𝑋 be a topological space, 𝑃 a property that points of 𝑋 may or may not
satisfy. We say 𝑃 holds for generic 𝑥 ∈ 𝑋 if the set
{𝑥 ∈ 𝑋 ∣ 𝑃 holds for 𝑥}
1. Transversality.
In the generic case, ℳ is a finite-dimensional smooth manifold. The present
paper concerns itself with how one proves such a result.
2. Orientation.
The moduli space ℳ is orientable. This seemingly innocent fact could be
highly nontrivial. Incidentally, an important tool for proving this is the de-
terminant line bundle on the space of Fredholm operators.
3. Compactness.
Any sequence in ℳ has a subsequence “converging” in a suitable sense.
Adding all possible limits to ℳ, we obtain a compactification ℳ. Of course,
such results are necessarily quite analytic in nature.
4. Gluing.
The compactified moduli space ℳ is a smooth manifold with an expected
boundary: Taking hint from the previous step, we expect the boundary to be
the space of another class of geometric objects. That all such objects arise
as boundary points is proved by a gluing procedure.
2 Transversality in general
In this section, we establish the theoretical foundations for proving transversality
theorems. Since we shall be mostly working with infinite-dimensional manifolds,
we first recall the relevant theory and fix our notation. After that, we prove the
Sard–Smale theorem and deduce from it a general transversality theorem, which
we shall apply in the sections that follow.
For a Banach space 𝒳 , we denote by ‖⋅‖𝒳 the norm defining 𝒳 .
Banach manifolds
Let 𝒳 , 𝒴 be Banach spaces, 𝑈 ⊂ 𝒳 an open set. A map 𝐹 ∶ 𝑈 → 𝒴 is (Fréchet)
differentiable at 𝑥 ∈ 𝑈 if there is a continuous linear map 𝐴 ∶ 𝒳 → 𝒴 such that
‖𝐹 (𝑥) − 𝐹 (𝑥0 ) + 𝐴(𝑥 − 𝑥0 )‖𝒴
lim = 0.
𝑥→𝑥0 ‖𝑥 − 𝑥0 ‖𝒳
In this case, 𝐴 is called the differential of 𝐹 at 𝑥0 and denoted by 𝐷𝐹𝑥0 . If 𝐹
is differentiable at each point in 𝑈 , we consider 𝐷𝐹 ∶ 𝑈 → ℒ (𝒳 , 𝒴 ), where
ℒ (𝒳 , 𝒴 ) is the Banach space of continuous linear maps 𝑋 → 𝑌 . We say 𝐹 is
continuously differentiable if 𝐷𝐹 is continuous. Iterating this definition, 𝐹 is 𝐶 𝑘
is 𝐷𝐹 is 𝐶 𝑘−1 .
166 Transversality Theorems by Example
For our applications, the maps are always sufficiently smooth, and we shall omit
the technical verifications of smoothness for reasons of space.
𝑘
‖𝑠‖𝐶 𝑘 ∶= ∑ sup |∇𝑖 𝑠|,
𝑖=0 𝑀
𝑘 1∕𝑝
𝑖 𝑝
‖𝑠‖𝑊 𝑘,𝑝 ∶= ∑ ∫ |∇ 𝑠| .
( 𝑖=0 𝑀 )
The completions of Γ(𝑀, 𝐸) with respect to these norms are Banach spaces which
we denote by 𝐶 𝑘 (𝑀, 𝐸) or 𝑊 𝑘,𝑝 (𝑀, 𝐸). We also have the space 𝐶 ∞ (𝑀, 𝐸), which
is a Fréchet space defined by the sequence (‖⋅‖𝐶 𝑘 )𝑘 of norms, i.e., a sequence 𝑠𝑛 →
𝑠 in 𝐶 ∞ (𝑀, 𝐸) if and only if ‖𝑠𝑛 − 𝑠‖𝐶 𝑘 (𝑀,𝐸) → 0 for all 𝑘. Since 𝑀 is compact,
it is easy to check that the topology on these spaces do not depend on the metrics
and connections chosen.
It is technically more difficult to define manifold structures on spaces of sections
of general fiber bundles, due to lack of linear structures. The case of trivial fiber
bundles, which amounts to spaces of maps between two manifolds, is discussed in,
e.g., [MS12, Remark B.1.24]. We shall not bother with this, since knowledge of
their tangent spaces already suffices in most applications.
2 Transversality in general 167
Fredholm theory
The Fredholm property is a crucial technical condition for generalizing finite-
dimensional phenomena. Fortunately, Fredholm maps abound in geometry. For
example, elliptic differential operators are Fredholm. In this article, we omit
the analytical verification of the Fredholm property and the computation of the
Fredholm index for reasons of time. However, we do recall the definitions:
Let 𝒳 , 𝒴 be Banach spaces. A continuous linear map 𝐴 ∶ 𝒳 → 𝒴 is Fredholm
if it has closed image and finite-dimensional kernel and cokernel. In this case,
its (Fredholm) index is defined to be ind 𝐴 ∶= dim ker 𝐴 − dim coker 𝐴. It is
standard that the space of Fredholm linear maps is open in the norm topology, and
the Fredholm index is locally constant on this space.
Let 𝒳 , 𝒴 be Banach manifolds. A 𝐶 1 map 𝐹 ∶ 𝒳 → 𝒴 is Fredholm if
𝐷𝐹𝑥 ∶ 𝑇𝑥 𝒳 → 𝑇𝐹 (𝑥) 𝒴 is Fredholm for all 𝑥 ∈ 𝒳 . If 𝒳 is connected, then by
continuity, ind 𝐷𝐹𝑥 is independent of 𝑥 ∈ 𝒳 , so 𝐹 has a well-defined Fredholm
index.
0 0 0 0
𝐴= , 𝐵= ,
(0 𝐴1 ) (0 𝐵 1 )
In other words, after changing the chart on the domain, 𝐹 takes the form (𝐹0 , 𝐴)
with respect to the splitting 𝒴 ≅ 𝒴0 ⊕ 𝒴1 .
Step 1. Any point in 𝑋 has a neighborhood 𝑉 such that the set of regular values
of 𝐹 |𝑉 is dense in 𝒴 .
Proof. Working in local charts, we are in the situation of the lemma. Thus we may
assume 𝐹 = (𝐹0 , 𝐴) on 𝑉 with respect to 𝒴 = 𝒴0 ⊕ 𝒴1 , where we retain the
notation above. The equation 𝐹 (𝑥) = 𝑦 becomes 𝑦0 = 𝐹0 (𝑥0 , 𝑥1 ), 𝑦1 = 𝐴1 𝑥1 for
𝑥 = 𝑥0 +𝑥1 ∈ 𝒳0 ⊕𝒳1 , 𝑦 = 𝑦0 +𝑦1 ∈ 𝒴0 ⊕𝒴1 . Thus 𝑦 is a regular value for 𝐹 |𝑉 if
and only if 𝑦0 is a regular value of 𝑥0 ↦ 𝐹0 (𝑥0 , 𝐵1 𝑦1 ), which is a finite-dimensional
map. Then the claim follows from Sard’s theorem for 𝐶 𝑘 maps. ◻
Note that we cannot conclude that the set fo regular values of 𝐹 |𝑉 is residual,
since Sard’s theorem only implies that the set of regular values of 𝐹 |𝑉 is residual
on each section 𝒴0 ⊕ {𝑦1 }.
Step 2. With 𝑉 constructed in Step 1, the set of regular values of 𝐹 |𝐾 2 is open for
any closed set 𝐾 ⊂ 𝑉 .
Proof. By the implicit mapping theorem, (b) implies 𝒮 −1 (0) is a Banach manifold.
We apply the Sard–Smale theorem to the projection 𝜋 ∶ 𝒮 −1 (0) → 𝒴 . It suffices
to show 𝜋 is Fredholm with index 𝑚. For (𝑥, 𝑦) ∈ 𝒮 −1 (0), since 𝑇(𝑥,𝑦) 𝒮 −1 (0) =
{(𝑣, 𝑤) ∈ 𝑇𝑥 𝒳 × 𝑇𝑦 𝒴 ∣ 𝐷1 𝒮(𝑥,𝑦) (𝑣) + 𝐷2 𝒮(𝑥,𝑦) (𝑤) = 0}, 𝐷𝜋(𝑥,𝑦) ∶ (𝑣, 𝑤) ↦ 𝑤,
we see that ker 𝐷𝜋(𝑥,𝑦) = ker 𝐷1 𝒮(𝑥,𝑦) , im 𝐷𝜋(𝑥,𝑦) = (𝐷2 𝒮(𝑥,𝑦) )−1 (im 𝐷1 𝒮(𝑥,𝑦) ) is
closed, and 𝐷2 𝒮(𝑥,𝑦) induces an isomorphism coker 𝐷𝜋(𝑥,𝑦) → coker 𝐷1 𝒮(𝑥,𝑦) , so
𝐷𝜋(𝑥,𝑦) is Fredholm with ind 𝐷𝜋(𝑥,𝑦) = ind 𝐷1 𝒮(𝑥,𝑦) by (a). For the last conclusion,
𝑇𝑥 𝒳 + 𝑇(𝑥,𝑦) 𝒮 −1 (0) = 𝑇(𝑥,𝑦) (𝒳 × 𝒴 ) since 𝑦 is a regular value of 𝜋, but 𝐷𝒮(𝑥,𝑦)
vanishes on 𝑇(𝑥,𝑦) 𝒮 −1 (0), so im 𝐷1 𝒮(𝑥,𝑦) = im 𝐷𝒮(𝑥,𝑦) = ℰ(𝑥,𝑦) . ◻
Remark 2.6. Note that in the finite-dimensional case, this reduces to the familiar
parametric transversality theorem with 𝒴 the space of parameters that provides
room for perturbation.
2. Fredholm property.
Check that 𝐷1 𝒮(𝑥,𝑦) ∶ 𝑇𝑥 𝒳 → ℰ(𝑥,𝑦) is Fredholm.
3. Regular value.
Check that 𝐷𝒮(𝑥,𝑦) ∶ 𝑇𝑥 𝒳 × 𝑇𝑦 𝒴 → ℰ(𝑥,𝑦) is surjective.
4. Passage to 𝐶 ∞ .
See the next subsection.
Remarks on smoothness
So far we have been solely concerned with Banach manifolds. However, many
spaces in geometry do not admit a natural Banach structure. This includes the
space of smooth objects or objects on noncompact manifolds, which are merely
Fréchet manifolds. Fortunately, it is often possible to remedy this problem. We
shall presently describe two ways to pass to smooth objects, i.e., to prove transver-
sality theorems of the form “For a generic smooth ..., we have ...”
The first approach is due to Clifford H. Taubes. It is based on the observation
that the set of good 𝐶 ℓ objects is open. Sometimes further modification is needed.
Since we shall follow this approach here, the reader will see this in action.
The second approach is due to Andreas Floer. Instead of 𝐶 ℓ objects, one con-
siders for (𝜀ℓ )ℓ ⊂ ℝ>0 the Banach space 𝐶𝜀∞ of 𝐶 ∞ objects with norm
∞
‖⋅‖𝐶𝜀∞ ∶= ∑ 𝜀ℓ ‖ ⋅ ‖𝐶 ℓ .
ℓ=0
Theorem 3.1. Let 𝑀 be a compact smooth manifold. Then a generic smooth func-
tion on 𝑀 is Morse.
𝒳 : the manifold 𝑀;
Remark 3.2. Once we know it is a smooth manifold, it is not hard to show that
𝛾 ↦ 𝛾(0) gives a diffeomorphism between the moduli space consider here and the
usual moduli space defined as the intersection of stable and unstable manifolds.
0,2
ℰ : the bundle ℰ 0,2 → 𝒫 1,2 × 𝒢 ℓ with fiber ℰ(𝛾,𝑔) = 𝐿2 (𝛾 ∗ 𝑇 𝑀);
where ∇𝑔 denotes gradient with respect to 𝑔. To compute it, applying 𝜕𝑡𝜕 |𝑡=0 to
(𝑔 + 𝑡ℎ)(∇𝑔+𝑡ℎ 𝑓 , 𝜉) = 𝜉𝑓 gives ⟨𝐷ℎ ∇𝑓 , 𝜉⟩ = −ℎ(∇𝑓 , 𝜉) for 𝜉 vector field.
Fredholm property
Regular value
is surjective.
Proof. Since 𝐷1 𝒮(𝛾,𝑔) is Fredholm, im 𝐷1 𝒮(𝛾,𝑔) and hence im 𝐷𝒮(𝛾,𝑔) is closed and
has finite codimension. Thus it suffices to show (im 𝐷𝒮(𝛾,𝑔) )⟂ = 0. Suppose 𝜂 ∈
𝐿2 (𝛾 ∗ 𝑇 𝑀) and
By elliptic regularity for 𝐷𝛾∗ ,5 the first equation implies 𝜂 ∈ 𝑊 1,2 . Thus 𝜂 is con-
tinuous by the Sobolev embedding theorem. If 𝜂(𝑡) ≠ 0 for some 𝑡 ∈ ℝ, then by
the lemma, there exists ℎ ∈ Sym2 𝑇𝛾(𝑡) ∗
𝑀 with ⟨𝜂(𝑡), 𝐷ℎ ∇𝑓 (𝛾(𝑡))⟩ > 0. Extend ℎ to
ℓ 2 ∗
a section ∈ 𝐶 (Sym 𝑇 𝑀). By continuity, ⟨𝜂, 𝐷ℎ ∇𝑓 ⟩ ≠ 0 in some neighborhood
𝐼 of 𝑡. Since 𝛾 is a flow line between distinct critical points, it is an embedding. If
𝜙 is a smooth cutoff function supported in an open set 𝑈 ⊂ 𝑀 with 𝛾 −1 (𝑊 ) = 𝐼,
then ⟨𝜂, 𝐷𝜙ℎ ∇𝑓 ⟩ ≥ 0 on ℝ, > 0 near 𝑡. This contradicts the second equation. ◻
Passage to 𝑪 ∞
We omit this.
Background material
Pseudoholomorphic curves
Let (𝑀, 𝐽 ) be an almost complex manifold, (Σ, 𝑗) a Riemann surface. A pseudo-
holomorphic or 𝐽 -holomorphic curve on 𝑀 parametrized by Σ is a smooth map
𝑢 ∶ Σ → 𝑀 such that 𝑑𝑢 is complex linear, i.e., 𝐽 ∘ 𝑑𝑢 = 𝑑𝑢 ∘ 𝑗, or 𝜕𝐽̄ 𝑢 = 0, where
1
𝜕𝐽̄ 𝑢 ∶= (𝑑𝑢 + 𝐽 ∘ 𝑑𝑢 ∘ 𝑗) ∈ Ω0,1 (Σ, 𝑢∗ 𝑇 𝑀).
2
We shall always consider the case that Σ is compact.
A 𝐽 -holomorphic curve 𝑢 ∶ Σ → 𝑀 is multiply covered if it factors nontrivially
through another Riemann surface, i.e., if there is a compact Riemann surface Σ′ , a
holomorphic branched covering 𝜑 ∶ Σ → Σ′ with deg 𝜑 > 1, and a 𝐽 -holomorphic
curve 𝑢′ ∶ Σ′ → 𝑀 such that 𝑢 = 𝑢′ ∘ 𝜑. Otherwise 𝑢 is simple.
A smooth map 𝑢 ∶ Σ → 𝑀 is somewhere injective if it has an injective point, a
point 𝑧 ∈ Σ such that 𝑑𝑢(𝑧) ≠ 0, 𝑢−1 (𝑢(𝑧)) = {𝑧}.
𝑘−1,𝑝
ℰ : the bundle ℰ 𝑘−1,𝑝 → ℬ 𝑘,𝑝 × 𝒥 ℓ with fibers ℰ(𝑢,𝐽 ) = 𝑊 𝑘−1,𝑝 (Σ, Λ0,1 ⊗𝐽
𝑢∗ 𝑇 𝑀), where Λ0,1 ⊗𝐽 𝑢∗ 𝑇 𝑀 is the bundle of 𝐽 -antilinear 1-forms;
Here
𝑘 ≥ 2, 𝑝 > 2, ℓ ≥ 𝑘 + 1 + max(1, 𝑛𝜒(Σ) + 2𝑐1 (𝐴) + 1).
Clearly
𝐷𝒮(𝑢,𝐽 ) (𝜉, 𝑌 ) = 𝐷𝑢 𝜉 + 21 𝑌 (𝑢) ∘ 𝑑𝑢 ∘ 𝑗.
Fredholm property
This follows from the generalized Riemann–Roch theorem for the Cauchy–
Riemann operator 𝐷𝑢 , cf. [MS12, Theorem C.1.10].
176 Transversality Theorems by Example
Regular value
Lemma 4.4. Let 𝐽0 , 𝜔0 be the standard linear almost complex and symplectic
structures on ℝ2𝑛 . Then End(ℝ2𝑛 , 𝐽0 , 𝜔0 ) acts transitively on ℝ2𝑛 ⧵ {0}. In other
words, for any 𝜉, 𝜂 ∈ ℝ2𝑛 ⧵{0}, there exists 𝑌 ∈ ℝ2𝑛×2𝑛 such that 𝑌 = 𝑌 ⊺ = 𝐽0 𝑌 𝐽0 ,
𝑌 𝜉 = 𝜂.
Proof. Simply take
1 ⊺ ⊺ ⊺ ⊺
𝑌 ∶= (𝜂𝜉 + 𝜉𝜂 + 𝐽0 (𝜂𝜉 + 𝜉𝜂 )𝐽0 )
|𝜉|2
1
− 4 (⟨𝜂, 𝜉⟩(𝜉𝜉 ⊺ + 𝐽0 𝜉𝜉 ⊺ 𝐽0 ) + ⟨𝜂, 𝐽0 𝜉⟩(𝐽0 𝜉𝜉 ⊺ − 𝜉𝜉 ⊺ 𝐽0 )). ◻
|𝜉|
𝐷𝒮(𝑢,𝐽 ) ∶ 𝑊 𝑘,𝑝 (Σ, 𝑢∗ 𝑇 𝑀)×𝐶 ℓ (𝑀, End(𝑇 𝑀, 𝐽 , 𝜔)) → 𝑊 𝑘−1,𝑝 (Σ, Λ0,1 ⊗𝐽 𝑢∗ 𝑇 𝑀)
is surjective.
Proof for 𝑘 = 1. Since 𝐷1 𝒮(𝑢,𝐽 ) is Fredholm, im 𝐷1 𝒮(𝑢,𝐽 ) and hence im 𝐷𝒮(𝑢,𝐽 ) is
closed and has finite codimension. Thus by the Hahn–Banach theorem, it suffices
to show (im 𝐷𝒮(𝑢,𝐽 ) )⟂ = 0, where ⟂ denotes its annihilator in the dual space.
′
Suppose 𝜂 ∈ 𝐿𝑝 (Σ, Λ0,1 ⊗𝐽 𝑇 𝑀) and
By elliptic regularity for 𝐷𝑢∗ ([MS12, Proposition 3.1.11 and Theorem C.2.3]), the
first equation implies 𝜂 ∈ 𝑊 1,𝑝 . Thus 𝜂 is continuous by the Sobolev embedding
theorem. We claim that 𝜂 vanishes at injective point of 𝑢. Since 𝑢 is simple, the set
of such points is open and dense, so by continuity, the claim implies 𝜂 = 0.
Let 𝑧 be an injective point of 𝑢. If 𝜂(𝑧) ≠ 0, then by the lemma, there exists
𝑌 ∈ End(𝑇𝑢(𝑧) 𝑀, 𝐽𝑢(𝑧) , 𝜔𝑢(𝑧) ) with ⟨𝜂(𝑧), 𝑌 ∘ 𝑑𝑢(𝑧) ∘ 𝑗(𝑧)⟩ > 0. Extend 𝑌 to a
section ∈ 𝐶 ℓ (𝑀, End(𝑇 𝑀, 𝐽 , 𝜔)). By continuity, ⟨𝜂, 𝑌 ∘ 𝑑𝑢 ∘ 𝑗⟩ > 0 in some
neighborhood 𝑉 of 𝑧 in Σ. By injectivity and compactness, take a neighborhood
𝑈 of 𝑧 in 𝑀 disjoint from 𝑢(Σ ⧵ 𝑉 ). If 𝜙 is a smooth cutoff function supported
in 𝑈 , then ⟨𝜂, (𝜙𝑌 ) ∘ 𝑑𝑢 ∘ 𝑗⟩ ≥ 0 on 𝑀, > 0 near 𝑧. This contradicts the second
equation. ◻
Proof for general 𝑘. Let 𝜂 ∈ 𝑊 𝑘−1,𝑝 (Σ, Λ0,1 ⊗𝐽 𝑢∗ 𝑇 𝑀). By the case 𝑘 = 1, there
exist (𝜉, 𝑌 ) ∈ 𝑊 1,𝑝 (Σ, 𝑢∗ 𝑇 𝑀) × 𝐶 ℓ (𝑀, End(𝑇 𝑀, 𝐽 , 𝜔)) with 𝐷𝒮(𝑢,𝐽 ) (𝜉, 𝑌 ) =
𝐷𝑢 𝜉 + 21 𝑌 (𝑢) ∘ 𝑑𝑢 ∘ 𝑗 = 𝜂. Then 𝐷𝑢 𝜉 = 𝜂 − 12 𝑌 (𝑢) ∘ 𝑑𝑢 ∘ 𝑗 ∈ 𝑊 𝑘−1,𝑝 , since 𝑢 ∈ 𝑊 ℓ,𝑝
by elliptic regularity for 𝜕𝐽̄ ([MS12, Theorem B.4.1]). By elliptic regularity for 𝐷𝑢
([MS12, Theorem C.2.3]), this implies 𝜉 ∈ 𝑊 𝑘,𝑝 . Thus 𝐷𝒮(𝑢,𝐽 ) is surjective. ◻
5 Example: Yang–Mills connections 177
Passage to 𝑪 ∞
We shall use the argument of Taubes to pass to 𝐶 ∞ almost complex structures.
Proof of main theorem. From the proof of the general transversality theorem, we
have ℳ(𝐴, Σ; 𝐽 ) is a manifold if 𝐷𝑢 is surjective for any 𝐽 -holomorphic curve
ℓ
𝑢 ∶ Σ → 𝑀. Let 𝒥reg be the set of such 𝐶 ℓ almost complex structures. We
ℓ
have proved 𝒥reg is dense in 𝒥 ℓ . For 𝐾 > 0, let 𝒥reg,𝐾
ℓ
be the set of 𝐶 ℓ al-
most complex structures 𝐽 such that 𝐷𝑢 is surjective for any 𝐽 -holomorphic curve
𝑢 ∶ Σ → 𝑀 representing 𝐴 such that (1) ‖𝑑𝑢‖𝐿∞ ≤ 𝐾, and (2) there exists 𝑧 ∈ Σ
with inf𝑤≠𝑧 𝑑(𝑢(𝑤),𝑢(𝑧))
𝑑(𝑤,𝑧)
≥ 1∕𝐾. Note that (2) implies 𝑢 is injective at 𝑧 and hence
ℓ ℓ ∞
simple, so 𝒥reg ⊂ 𝒥reg,𝐾 . We claim that 𝒥reg,𝐾 is open and dense in 𝒥 ∞ , so that
∞ ∞
𝒥reg = ⋂𝐾>0 𝒥reg,𝐾 is residual.
∞
To see that 𝒥reg,𝐾 is open in 𝒥 ∞ , let (𝐽𝑛 )𝑛 , (𝑢𝑛 )𝑛 , (𝑧𝑛 )𝑛 be sequences where
𝑑(𝑢 (𝑤),𝑢 (𝑧 ))
𝑢𝑛 ∶ Σ → 𝑀 is 𝐽𝑛 -holomorphic with ‖𝑑𝑢𝑛 ‖𝐿∞ ≤ 𝐾, inf𝑤≠𝑧𝑛 𝑛𝑑(𝑤,𝑧𝑛 ) 𝑛 ≥ 1∕𝐾,
𝑛
and 𝐷𝑢𝑛 not surjecitve. Suppose 𝐽𝑛 → 𝐽 in 𝒥 ∞ . By elliptic regularity estimates
([MS12, Theorem B.4.2]) and compactness, passing to a subsequence, we may
assume 𝑢𝑛 → 𝑢 in 𝐶 ∞ , 𝑧𝑛 → 𝑧. Then 𝑢 is 𝐽 -holomorphic with ‖𝑑𝑢‖𝐿∞ ≤ 𝐾,
inf𝑤≠𝑧 𝑑(𝑢(𝑤),𝑢(𝑧))
𝑑(𝑤,𝑧)
∞
≥ 1∕𝐾, and 𝐷𝑢 = lim 𝐷𝑢𝑛 is not surjective, so 𝐽 ∉ 𝒥reg,𝐾 . The
ℓ
same argument shows 𝒥reg,𝐾 is open in 𝒥 ℓ for all ℓ.
∞
That 𝒥reg,𝐾 is dense in 𝒥 ∞ is proved by a 1∕ℓ-argument as before: Let 𝐽 ∈
𝒥 ∞ . For each ℓ, take 𝐽ℓ ∈ 𝒥regℓ
with ‖𝐽ℓ − 𝐽 ‖𝐶 ℓ ≤ 1∕ℓ, take 𝛿ℓ > 0 such that
̃ ̃
‖𝐽 − 𝐽ℓ ‖𝐶 ℓ < 𝛿ℓ implies 𝐽 ∈ 𝒥reg,𝐾ℓ
, then take 𝐽̃ℓ ∈ 𝒥 ∞ with ‖𝐽̃ℓ − 𝐽ℓ ‖𝐶 ℓ <
min(1∕ℓ, 𝛿ℓ ). Then 𝐽̃ℓ ∈ 𝒥 ∞ and 𝐽̃ℓ → 𝐽 in 𝒥 ∞ .
reg,𝐾 ◻
Remark 4.6. The smooth structure on ℳ ∗ (𝐴, Σ; 𝐽 ) ⊂ 𝑊 𝑘,𝑝 (Σ, 𝑀) does not de-
pend on 𝑘, 𝑝, since from the proof we see that charts are given by 𝜉 ↦ exp𝑢 𝜉 for a
fixed 𝐽 -holomorphic curve 𝑢, which does not depend on 𝑘, 𝑝.
Background material
The purpose of this subsection is mainly to fix notation.
178 Transversality Theorems by Example
Anti-self-dual connections
Let 𝑀 be a compact smooth 4-manifold, 𝑃 → 𝑀 a principal SU(2)-connection.
We shall usually identify it with the associated vector bundle 𝑃 ×𝜌 ℂ2 , where 𝜌
is the standard representation of SU(2) on ℂ2 . A connection on 𝑃 is reducible if
𝑃 ×𝜌 ℂ2 splits into the direct sum of two line bundles and the induced connection
on it also splits. Otherwise it is irreducible.
Fix a Riemannian metric 𝑔 on 𝑀. This induces the Hodge star operator
∗ ∶ Λ2 → Λ2 with ∗2 = id. The subbundles Λ2± ∶= ker(∗ ∓ 1) are called the
bundle of (anti-)self-dual forms. This definition extends to forms with coefficients
in any vector bundle. In particular, a connection 𝐷 on 𝑃 is (anti-)self-dual if so
is 𝐹𝐷 . Such connections are automatically Yang–Mills in the sense that they are
critical points of the Yang–Mills functional. Since we shall make no use of this
functional, we omit its definition.
4. Application of Sard–Smale.
The theorem now follows immediately from the Sard–Smale theorem since
ℳ ∗ (𝑃 ; 𝜑∗ 𝑔) = 𝜋 −1 (𝜑).
See [FU91] for the complete proof.
References
[FU91] Daniel S. Freed and Karen K. Uhlenbeck (1991). Instantons and Four-Manifolds.
second. Vol. 1. Mathematical Sciences Research Institute Publications. Springer.
[Lan99] Serge Lang (1999). Fundamentals of Differential Geometry. Vol. 191. Graduate
Texts in Mathematics. Springer.
[MS12] Dusa McDuff and Dietmar Salamon (2012). 𝐽 -holomorphic Curves and Sym-
plectic Topology. second. Vol. 52. Colloquium Publications. American Mathe-
matical Society.
[Sch93] Matthias Schwarz (1993). Morse Homology. Vol. 111. Progress in Mathematics.
Birkhäuser.
180 Transversality Theorems by Example
[Sma65] Stephen Smale (1965). “An infinite dimensoinal version of Sard’s theorem”.
American Journal of Mathematics 87 (4), 861–866.
超实数
《荷思》编辑部
摘 要
超实数是在实数中添加无穷大、无穷小元素得到的数系. 我们首先介绍
超实数的构造过程, 然后介绍它在数学分析中的应用, 也就是非标准分析,
以及超实数与博弈论的联系.
目录
1 超实数 182
超实数的构造 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
超实数的运算 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
2 超实数与非标准分析 186
非标准实数 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
非标准微积分 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
3 超实数与博弈论 190
游戏局面与超实数的对应 . . . . . . . . . . . . . . . . . . . . . . . . 190
游戏策略 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
182 超实数
1 超实数
超实数1 (surreal number) 是在实数中添加无穷大、无穷小元素得到的数系.
在这一节中,我们介绍超实数的构造过程,并在超实数上定义一些基本运算.
超实数的构造
超实数的构造过程如下. 首先, 记
0=( ∶ ),
然后让
1 = (0 ∶ ), −1 = ( ∶ 0).
2 = (1 ∶ ), −2 = ( ∶ −1).
利用这样的归纳构造, 我们可以构造出所有的整数.
接下来, 我们让
1 3
= (0 ∶ 1), = (1 ∶ 2).
2 2
用这种方式可以得到所有半整数:
2𝑘 + 1
= (𝑘 ∶ 𝑘 + 1).
2
进一步地, 我们令
4𝑘 + 1 1 4𝑘 + 3 1
= (𝑘 ∶ 𝑘 + ) , = (𝑘 + ∶ 𝑘 + 1 ) .
4 2 4 2
这样就得到了所有 1∕4 的整数倍. 如此下去, 我们能得到所有二进制下的有限
小数. 这些数在实数中是稠密的, 因此, 其它的实数就可以通过类似 Dedekind
分割的方式来得到.
超实数除了包含一般的实数之外, 还有一些其它的数. 例如,
𝜔 = (1, 2, 3, ⋯ ∶ )
是一个比所有实数都要大的数, 而
1 1 1
𝜀 = (0 ∶ , , ,⋯
2 4 8 )
1
也常译作超现实数.
1 超实数 183
是一个比所有正实数都要小的正数. 这两个数也是合法的超实数.
以上的操作可以一直持续下去: 只要我们有两个超实数的集合 𝑋𝐿 和 𝑋𝑅 ,
使得对任意的 𝑥𝐿 ∈ 𝑋𝐿 和 𝑥𝑅 ∈ 𝑋𝑅 , 都有 𝑥𝐿 < 𝑥𝑅 , 那么 (𝑋𝐿 ∶ 𝑋𝑅 ) 就是一个
合法的超实数. 这样不断递归得到的全体对象就是所有的超实数.
超实数的大小关系按照如下方式递归定义: 对于两个超实数
𝑥 = (𝑋𝐿 ∶ 𝑋𝑅 ), 𝑦 = (𝑌𝐿 ∶ 𝑌𝑅 ),
我们说 𝑥 ≤ 𝑦, 如果对任意的 𝑥𝐿 ∈ 𝑋𝐿 和 𝑦𝑅 ∈ 𝑌𝑅 , 都有 𝑥𝐿 ≱ 𝑦 和 𝑥 ≱ 𝑦𝑅 成
立; 𝑥 和 𝑦 满足 𝑥 < 𝑦, 如果 𝑥 ≤ 𝑦 且 𝑦 ≰ 𝑥. 我们记 𝑥 = 𝑦, 如果 𝑥 ≤ 𝑦 且 𝑦 ≤ 𝑥,
并将 𝑥 和 𝑦 视作相等的超实数.
全体超实数构成的类,我们叫它 𝕊. 这是一个真类, 也就是说, 它的大小比
任何的集合都要大, 它因而不是集合.
根据超实数的构造过程, 我们可以通过超实数的 “复杂度” 来对超实数进
行分类. 我们记 𝕊0 为可以用空集 ∅ 构造的超实数的集合, 即
𝕊0 = {0}.
𝕊<𝛼 = 𝕊𝛽 .
⋃
𝛽<𝛼
𝕊1 = {−1, 0, 1};
这样对于超实数的分类方式可以帮助对超实数的大小进行比较: 𝕊𝛼 中超
实数和 𝕊𝛽 中超实数的大小比较, 依赖于 𝕊<𝛼 中超实数和 𝕊𝛽 中超实数的大小比
较, 以及 𝕊𝛼 中超实数和 𝕊<𝛽 中超实数的大小比较. 利用归纳法, 我们可以验证
以下的关于超实数的大小关系的性质:
• 任何两个超实数都可以比较大小;
• 超实数的大小关系具有自反性和传递性;
• 若 𝑥 = (𝑋𝐿 ∶ 𝑋𝑅 ) 是一个超实数, 那么 𝑋𝐿 < 𝑥 < 𝑋𝑅 .
2 = (1 ∶ ) = (0, 1 ∶ ).
184 超实数
超实数的运算
超实数不仅仅可以做大小比较, 它们还可以像实数一样, 做加减法, 乘除
法. 我们首先定义加减法, 为此只需要定义超实数的加法, 和超实数的相反数.
定义 1.1. 设有超实数
𝑥 = (𝑋𝐿 ∶ 𝑋𝑅 ), 𝑦 = (𝑌𝐿 ∶ 𝑌𝑅 ).
我们定义
−𝑥 = (−𝑋𝑅 ∶ −𝑋𝐿 ),
以及
𝑥 + 𝑦 = ((𝑥 + 𝑌𝐿 ) ∪ (𝑦 + 𝑋𝐿 ) ∶ (𝑥 + 𝑌𝑅 ) ∪ (𝑦 + 𝑋𝑅 )).
这里
−𝑆 = {−𝑠 ∣ 𝑠 ∈ 𝑆}, 𝑥 + 𝑆 = {𝑥 + 𝑠 ∣ 𝑠 ∈ 𝑆}.
例 1.2. 我们给出一些超实数间加减法的例子.
𝑥 + 0 = (𝑋𝐿 ∶ 𝑋𝑅 ) + ( ∶ ) = (𝑋𝐿 ∶ 𝑋𝑅 ) = 𝑥.
• 我们有
1 + 1 = (0 ∶ ) + (0 ∶ ) = (1 ∶ ) = 2.
• 我们有
𝜔 + 1 = (0, 1, ⋯ , 𝜔 ∶ ) = (𝜔 ∶ ),
又比如,
2𝜔 = 𝜔 + 𝜔 = (𝜔 + 1, 𝜔 + 2, ⋯ ∶ ).
例 1.3. 我们给出一些超实数乘法的例子.
• 2 和 𝜔 的乘积就是 2𝜔 = 𝜔 + 𝜔.
• 超实数 𝜀 和 𝜔 的乘积是 1, 也就是说这两个超实数互为倒数.
1 超实数 185
𝜔𝑥 = (𝑌𝐿 ∶ 𝑌𝑅 ),
其中
𝑌𝐿 = {0} ∪ {𝑠 ⋅ 𝜔𝑥𝑙 ∣ 𝑠 ∈ ℝ, 𝑠 > 0, 𝑥𝑙 ∈ 𝑋𝐿 },
𝑌𝑅 = {𝑠 ⋅ 𝜔𝑥𝑟 ∣ 𝑠 ∈ ℝ, 𝑠 > 0, 𝑥𝑟 ∈ 𝑋𝑅 }.
• 序列 (𝑦𝛽 ) 是单调递减的;
• 有
𝑥 = ∑ 𝑟 𝛽 ⋅ 𝜔 𝑦𝛽 .
𝛽<𝛼
这个表达式叫做 𝑥 的 𝝎 进制展开.
exp ∶ 𝕊 → 𝕊+
例 1.5. 我们有
186 超实数
• exp 𝜔 = 𝜔𝜔 .
• log 𝜔 = 𝜔𝜀 .
2 超实数与非标准分析
非标准实数
∗
ℝ = 𝕊<𝜔1 ,
2
也常译作超实数.
2 超实数与非标准分析 187
𝑥 = (𝑥, 𝑥, 𝑥, 𝑥, 𝑥, … ) (𝑥 ∈ ℝ),
𝜔 = (1, 2, 3, 4, 5, … ),
𝜔 + 1 = (2, 3, 4, 5, 6, … ),
𝜔2 = (1, 4, 9, 16, 25, … ),
𝜀 = 1∕𝜔 = (1, 1∕2, 1∕3, 1∕4, … ),
√𝜔 + 𝜀 = (1 + 1, √2 + 1∕2, √3 + 1∕3, … ),
等等. 严格的定义如下:
𝐼𝑘 = {{𝑎𝑛 }𝑛∈ℕ ∈ ℝℕ | 𝑎𝑘 = 0}
∗
ℝ = ℝℕ ∕𝐼.
在无歧义时, 也直接把 ∗ 𝑓 记为 𝑓 .
我们也可以得出一个奇怪但正确的等式:
sin(𝜔π) = (0, 0, 0, 0, … ) = 0.
然而,
cos(𝜔π) = (−1, 1, −1, 1, … ),
cos(𝜔π) = ±1.
例 2.7. 我们也有
这是因为, 我们知道
𝜀𝑛 = (1, 1∕2𝑛 , 1∕3𝑛 , … ),
非标准微积分
使用非标准分析的语言, 我们可以重新构建微积分的理论. 作为例子, 我们
使用非标准分析来定义函数的极限、导数等概念.
𝑥 = ∑ 𝑎 𝑛 𝜔𝑛
𝑛∈𝕊
st 𝑓 (𝑥0 + ℎ) = 𝑦,
就说 𝑦 是 𝑓 (𝑥) 在 𝑥 → 𝑥0 时的极限, 记为
lim 𝑓 (𝑥) = 𝑦.
𝑥→𝑥0
这样的定义也能够简化很多命题的证明, 这里我们举一个例子.
190 超实数
例 2.14. 我们来证明求导的链式法则
如果 𝑔(𝑥 + 𝜀) ≠ 𝑔(𝑥), 那么
3 超实数与博弈论
在这一小节, 我们给出超实数在博弈论中的简单应用.
游戏局面与超实数的对应
我们考虑满足如下条件的博弈问题: 一个双人游戏, 给定初始状态, 每人轮
流进行操作; 在游戏状态满足给定条件时游戏结束, 且有唯一的胜利者; 在有限
步操作后游戏必然结束 (例如, 严格禁止全局同形再现规则下的围棋即满足此
条件).
我们接下来将以一种广义超实数来表示一个游戏状态. 这里的广义超实数
与超实数定义类似, 即仍形如 (𝑋𝐿 ∶ 𝑋𝑅 ), 但不要求 𝑋𝐿 ≤ 𝑋𝑅 . 例如 (0 ∶ 0) 是
广义超实数, 却不是超实数.
称两人为 L,R. 如果形成轮到 L 操作的局面时触发结束条件, 且 L 失败,
则记 𝑋𝐿 = ∅, 反之亦然. 由此可以归纳地定义每个局面所对应的广义超实数
(𝑋𝐿 ∶ 𝑋𝑅 ). 例如轮到谁操作谁输的局面即为 (∅ ∶ ∅) = 0; 轮到 R 操作时 R 失
败, 但轮到 L 操作时 L 可以将此局面变为 0 的游戏局面即为 (0 ∶ ∅) = 1; 如果
此局面任意人操作以后均变为 0, 则为 (0 ∶ 0), 一个不是超实数的广义超实数.
我们用记号 ∗ 表示 (0 ∶ 0).
反之, 给定一个广义超实数 (𝑋𝐿 ∶ 𝑋𝑅 ), 可以构造如下的游戏: 轮到 L 操作
时, 其可以将此超实数变为 𝑋𝐿 中的一个元素, 如果轮到 L 操作时 𝑋𝐿 = ∅, 则
𝐿 失败, 反之亦然. 注意到集合只能嵌套有限次, 游戏必然在有限步操作后结
束. 由此我们构造了游戏局面和广义超实数的一一对应, 接下来我们使用广义
超实数来研究游戏.
3 超实数与博弈论 191
超实数的序关系则对应游戏的胜负关系. 设某游戏局面对应某广义超实数
𝑥, 则:
注 3.3. 超实数的乘法在游戏理论中没有好的对应.
广义超实数中序关系十分微妙. 这由下面的例子可以看出:
同样的方法可以定义其下界 inf(𝑥).
游戏策略
对于一般的游戏局面, 我们并无法使用广义超实数对其进行好的分析. 有
时一个游戏局面可以表示为若干较为简单的子局面之和 (和的定义见上述), 则
可以使用广义超实数给出一个较好的策略. 这类局面的一个有代表性的例子是
围棋的官子理论.
我们首先定义一步操作的价值.
𝛼𝐿 = sup{mean(𝑦) ∣ 𝑦 ∈ 𝑋𝐿 } − mean(𝑥),
即 L 操作后与操作前得到的局面的均值之差.
对一个局面 𝑥 = (𝑋𝐿 ∶ 𝑋𝑅 ), 使得
𝑥 = 𝑥1 + ⋯ + 𝑥 𝑛 , 𝑥𝑖 = (𝑋𝐿𝑖 ∶ 𝑋𝑅𝑖 ).
然而这样的策略未必是最佳的. 这从下面的例子可以看出.
这说明了使用广义超实数分析游戏的局限. 只有在一些简单的情况之下,
可以给出相同均值的广义超实数的序关系 (例如 (0 ∶ ∗) > 0), 从而给出精确的
游戏策略, 参见 [BW97], 一般情况只能进行近似估计.
参考文献 193
参考文献
[BW97] E. R. Berlekamp and D. Wolfe (1997). Mathematical Go: Chilling gets the last
point. A K Peters/CRC Press.
[Ehr12] P. Ehrlich (2012). “The Absolute Arithmetic Continuum and the Unification of
All Numbers Great and Small”. Bulletin of Symbolic Logic 18 (1), 1–45.
[Gon86] H. Gonshor (1986). An Introduction to the Theory of Surreal Numbers. Cam-
bridge University Press.
[Ros04] E. E. Rosinger. Short Introduction to Nonstandard Analysis. arXiv: math /
0407178.
折纸问题
黄虞来1
摘 要
本文将完整地给出折纸构造点的刻画, 并给出折纸的两个有趣实例: 三
等分角与折正 𝑁 边形.
1 折纸公理
折纸实际上是利用已知几何对象构造新的几何对象的过程. 我们称折纸中
构造出的点、线为可构造点、可构造线. 我们将折纸的纸面与复平面 ℂ 等同起
来, 于是可构造点、可构造线实际上是复平面的一些特殊子集.
实际操作中, 我们发现主要会使用以下六种构造方式 (称之为折纸的六公
理):
(1) 连接两个可构造点的直线是可构造的;
(2) 两条可构造线的交点是可构造点;
(3) 两个可构造点的中垂线是可构造线;
(4) 两条可构造线的角平分线是可构造的;
2 代数结构
设 𝒫 ⊂ ℂ 为复平面 ℂ 的子集, ℒ 为 ℂ 上一些直线构成的集合. 我们可以
把公理 (1) 到 (6) 理解为集合 𝒫 和 ℒ 之间的关系: 比如公理 (1) 说的就是过
𝒫 中两点的直线属于 ℒ .
我们先来刻画通过前五个折纸公理构造出的点与线:
𝑃 𝑄 = ⃖⃖⃖⃖⃗
(a) 若 𝑃 , 𝑄, 𝑅 ∈ 𝒫 , 则满足 ⃖⃖⃖⃖⃗ 𝑅𝑆 的点 𝑆 ∈ 𝒫 ;
(c) 若 𝑡 ∈ 𝒫𝑟 , 𝑡 ≠ 0, 则 1∕𝑡 ∈ 𝒫𝑟 ;
(d) 若 𝑎, 𝑏 ∈ 𝒫𝑟 , 则 𝑎𝑏 ∈ 𝒫𝑟 .
2 代数结构 197
R S’ S
P Q
命题 2.4. 𝒫 是 ℂ 的子域且对共轭封闭.
𝒫 = {𝑥 + 𝑖𝑦 ∣ 𝑥, 𝑦 ∈ 𝒫𝑟 }.
最后推论 (c)(d) 告诉我们 𝒫 是 ℂ 的子域. ◻
Q
P
Figure 2: 对开方封闭
𝑎𝑥 + 𝑏𝑦 = 0 (𝑎, 𝑏 ∈ ℳ𝑟 ).
由于 ℳ 对开平方封闭, 故 √𝑎2 + 𝑏2 ∈ 𝒫𝑟 . 而
𝑏 𝑎
cos 𝜃 = ± , sin(𝜃) = ± ,
√𝑎2 + 𝑏2 √𝑎2 + 𝑏2
因此 sin(𝜃), cos(𝜃) ∈ ℳ𝑟 . 同理 sin(𝜂), cos(𝜂) ∈ ℳ. 角平分线方程的系数
𝜃±𝜂 𝜃±𝜂
sin( ), cos( )
2 2
可由 𝜃, 𝜂 的三角函数经过四则运算及开方得到, 故角平分线可构造. 最后请读
者自行验证公理 (5), 从而 (ℳ, 𝒮 ) 是 Euclid 构造. ◻
3 第六公理
公理 (6) 是折纸公理中最不平凡的一条, 它使折纸操作超出了尺规作图的
界限.
设 ℳ ⊂ ℂ 是 ℂ 中包含 𝑇 且对共轭、开平方、开立方封闭的最小子域, 𝒮 为
系数在 ℳ𝑟 中的直线全体, 那么只需要证明 (𝒫 , ℒ ) = (ℳ, 𝒮 ). 由上一节可知
𝒫 为对共轭、开平方封闭的子域, 剩下只用说明:
• 𝒫 对开立方封闭:
𝑋 𝑡 𝐴𝑋 = 0
𝑋0𝑡 𝐴𝑋 = 0,
⎧ 𝑡
⎪𝑋 𝐴𝑋 = 0;
⎨ 𝑡
⎩𝑌 𝐵𝑌 = 0.
⎪
消去 𝑌 得到 (𝑥, 𝑦) 满足的两个二次方程:
⎧ 𝑡
⎪𝑋 𝐴𝑋 = 0; (∗)
⎨ 𝑡 −1
⎩𝑋 𝐴𝐵 𝐴𝑋 = 0.
⎪ (∗∗)
200 折纸问题
4 折纸三等分角
给定平面上两条射线夹出的角, 仅凭直尺和圆规不一定能将其三等分. 这
是因为三等分角实际上涉及了开立方运算:
Figure 3: 三等分角
5 折纸正 𝑁 边形
用纸条折出五角星的游戏能勾起许多人美好的童年回忆. 我们很自然地会
考虑折纸正 𝑁 边形的问题: 对于怎样的 𝑁, 可以折出一个正 𝑁 边形?
在之前的讨论中, 我们默认了这样一件事: 由 𝑇 ⊂ ℂ 生成的 Origami 构造
(𝒫 , ℒ ) 确为实际操作中可以折出的点和线. 事实上 “生成” 的概念是用集合的
交定义的, 这只是出于数学表述上的简洁, 但未必符合实际.
为了刻画折纸过程中可构造点集的变化, 引入根式塔的概念:
2
平面上五个点, 任三点不共线, 则唯一确定一条圆锥曲线.
参考文献 201
定理 5.4. 正 𝑁 边形可由折纸得到的充要条件是 𝑁 = 2𝛼 3𝛽 𝑝1 … 𝑝𝑟 , 其中
𝛼, 𝛽 ≥ 0, 𝑝𝑖 为互异素数, 且形如 2𝑎 3𝑏 + 1.
参考文献
[Hul96] Thomas Hull (1996). “A Note on ”Impossible” Paper Folding”. The American
Mathematical Monthly 103 (3), 240–241.
[Vid97] Carlos R. Videla (1997). “On Points Constructible from Conics”. Mathemat-
ical Intelligencer 19 (2), 53–57.
[毛天一 08] 毛天一,石权,林洁,王芝兰 (2008). “折纸的代数结构”. 荷思 (2), 4–14.
202 折纸问题
Figure 4: 折纸正 7 边形
A Game About Competing Area
尚鉴桥1
ABSTRACT
This article is about a problem in daily life: KFC and McDonald’s both
want to open 𝑛 restaurants in a city. McDonald’s opens its restaurant first, KFC
opens its restaurants based on locations of McDonald’s. If everyone goes to
the nearest restaurant, could McDonald’s make sure that it has at least half of
the customers?
Contents
1 Introduction 203
2 Lemmas 204
5 Summary 219
1 Introduction
Before the formal discussion of this question. Let us consider a question in daily
life: why the McDonald’s and KFC are always together on a street. To solve this
problem, we should establish a mathematical model. Assume that the population
density on this street is invariant and everyone would choose the closest restaurant
to have meal (If the two restaurants are at the same point, people will choose one
1
清华大学数学系数 80 班
204 A Game About Competing Area
randomly). If you are the owner of McDonald’s and you want to build a restaurant
first before KFC. You want to get as more customers as you can, which place will
you choose?
As the picture shown behind, if Mc and KFC open their restaurant at 𝐴 and
𝐵. Let 𝐶 be the middle point of the segment 𝐴𝐵. We can easily prove that the
customers in one side of C would go to Mc and others would go to KFC. So if you
want more customers than your competitor, you should choose the middle point of
the street or he could open his restaurant at the middle point of the street and get
more than half customers.
𝐴 𝐶 𝐵
If you want to open 𝑛 restaurants on the street, it can be proved that the best
choice is dividing the street into 𝑛 equal parts and choose middle point of each
part.
𝐴1 𝐵1 𝐴2 𝐵2 𝐵3 𝐴3 𝐴4𝐵4
So we naturally want to know that what will happen if we change the street
into the whole city. Could the first one always make sure he get at least half of the
customers?
We express this question in another way:
For a convex set 𝑇 in ℝ2 and a constant 𝑛, can we find a set of points 𝐴 =
{𝐴1 , 𝐴2 ...𝐴𝑛 } of 𝑛 points in 𝑇 which satisfy the following conditions? For any set
𝐵 = {𝐵1 , 𝐵2 ...𝐵𝑛 } of 𝑛 points on the plane , we color 𝑇 by two colors. For a point
𝑠 in 𝑇 , find the closet point to it in 𝐴 ∪ 𝐵. If it is a point in 𝐴, color it in red,
otherwise color it in green. The area of red parts in the diagram is always not less
than the green parts (Here the metric is always the Euclidean metric).
In this article we will try to find out all pairs of (𝑇 , 𝑛) which meets the require-
ment. At first we would give some lemmas in the second section. Then we would
consider a special case (𝑇 is a square) as an example in the third section. In the
forth section, we will give out two strong conclusion to solve the case when 𝑛 is
big enough.
We always use capital letters for points and domains, lower-case letters for lines,
|𝐴𝐵| ∶= 𝑑(𝐴, 𝐵). For a domain 𝑇 we always use 𝑆𝑇 to denote the area of 𝑇 .
2 Lemmas
In this chapter we will give some trivial but useful lemmas for this problem.
Definition 1. For any point 𝑃 ∈ 𝐴, we call the set 𝕋𝑃 ∶= {𝐾 ∶ 𝐾 ∈ 𝑇 , |𝐾𝑃 | =
min𝑛𝑖=1 |𝐾𝐴𝑖 |} the domain of 𝑝 and we call 𝑃 the origin point of 𝕋𝑃 .
2 Lemmas 205
Lemma 1. The intersection line of two adjacent domains is the perpendicular bi-
sector of the two origin points they correspond to.
𝐴1
𝑄 𝑅
𝐴2
Proof. If 𝕋𝐴1 is not convex, we may as well assume that ∠𝑃 𝑄𝑅 is more than 𝜋,
here 𝑃 , 𝑄, 𝑅 are adjacent vertexes of 𝕋𝐴1 . Because 𝑇 is convex, 𝑃 𝑄 can’t be the
side of 𝑇 . So the symmetric point of 𝐴1 about 𝑄𝑅 must be another origin point.
We call it 𝐴2 . Since ∠𝑃 𝑄𝑅 > 𝜋, the length of |𝑃 𝐴1 | is larger than |𝑃 𝐴2 |. So P
couldn’t be on the side of 𝕋𝐴1 . Then we get the contradiction. ◻
There are some lemmas about the area of each domain. For convenience, we
sometimes use 𝑆𝑖 to denote 𝑆𝕋𝑖 , 𝑆𝐴𝑖 to denote 𝑆𝕋𝐴𝑖
Lemma 3. For any domain 𝕋𝐴𝑖 and 𝜖 > 0, we can choose a point as one of points
in 𝐵 and let the point change almost half area of 𝐴 into green. (i.e. it can change
𝑆
the domain with area at least 𝕋2𝐴𝑖 − 𝜖 into green.). And we can also choose two
points 𝐵1 and 𝐵2 so that they change almost the whole 𝕋𝐴𝑖 into green.
206 A Game About Competing Area
𝑘
𝐵2
𝐴1 𝑙
𝐵1
Proof. For any line 𝑙 passing through 𝐴1 , it divides 𝐴 into two parts. We consider
the line 𝑘 perpendicular to 𝑙 and go through 𝐴1 .
𝑘 divides 𝐴1 into two parts 𝕋1 , 𝕋2 (We assume 𝑆1 ≥ 𝑆2 ). Let the diameter of
𝜖
𝐴1 be 𝑑, choose a point 𝐵1 on 𝑘 in 𝑆1 such that |𝐵1 𝐴1 | ≤ 2𝑑 . This point meets the
𝜖
first requirement. Then choose another point 𝐵2 on 𝑘 in 𝑆2 such that |𝐵2 𝐴1 | ≤ 2𝑑 .
The two points, 𝐵1 and 𝐵2 , meet the second requirement. ◻
Proof. Suppose that 𝑆𝐴1 ≥ 𝑆𝐴2 ≥ ... ≥ 𝑆𝐴𝑛 . By lemma 1.3, for every 𝜖 we can
choose two points in 𝕋𝐴1 one point in 𝕋𝐴2 and one point in 𝕋𝐴3 ... until one point in
𝑆𝐴2 +...𝑆𝐴𝑛−1
𝕋𝐴𝑛−1 so that these points change the domain of area at least 𝑆𝐴1 + −𝜖
2
𝑆𝐴2 +...𝑆𝐴𝑛−1 𝑆𝐴1 +...𝑆𝐴𝑛
into green. Then we get that 𝑆𝐴1 + 2
−𝜖 ≤ 2
is true for every 𝜖.
So 𝑆𝐴1 = 𝑆𝐴𝑛 , we get the corollary. ◻
Lemma 4. Every line goes through the origin point divides the domain into two
parts of equal area.
Proof. Using the same notation in lemma 3, for each line passing through 𝐴𝑖 ,
𝑆 −𝑆
choose a point 𝐵 which changes the area of area 1 2 2 − 𝜖 into green and we
choose other 𝑛 − 1 points which change at almost half area left into green. We can
get that 𝑆1 ≤ 𝑆2 , so 𝑆1 = 𝑆2 . ◻
𝐵1
𝐶1
𝐴1
𝐶2
𝐵2
Proof. For each line passing through 𝐴1 , it intersects the boundary of the domain
at two points 𝐵1 and 𝐵2 . If |𝐴1 𝐵1 | > |𝐴2 𝐵2 |, we rotate the line by a small angle
𝜃 and let the two new points of intersection 𝐶1 and 𝐶2 satisfies |𝐴1 𝐶1 | ≥ |𝐴2 𝐶2 |.
Then we know 𝑆𝐴1 𝐵1 𝐶1 = 12 |𝐴1 𝐵1 ||𝐴1 𝐶1 |𝑠𝑖𝑛𝜃 > 12 |𝐴1 𝐵2 ||𝐴1 𝐶2 |𝑠𝑖𝑛𝜃 = 𝑆𝐴1 𝐵2 𝐶2 .
From lemma 4 𝑆𝐴1 𝐵1 𝐶1 = 𝑆𝐴1 𝐵2 𝐶2 , then we get the contradiction. Hence |𝐴1 𝐵1 | =
|𝐴2 𝐵2 | for any line, which implies that 𝕋𝐴1 is centrally symmetric. ◻
Now we have got all tools we need. Then we will discuss an example to use
these tools.
In this section we will prove that if 𝑇 is a square, the only 𝑛 to make (𝑇 , 𝑛) a good
pair is 1. We can get an idea to consider general questions from this special case.
To achieve our goal, we need to consider what the shape 𝕋𝐴1 , 𝕋𝐴2 ... would be by
lemma 6, and then discuss some situations and details.
In the following discussion if we say lines and points without specified, we
mean the intersection line segments of domains and the endpoints of lines.
Theorem 3.1. The lines which intersect the boundary of the square are perpen-
dicular to the boundary.
𝐺 𝐻
𝑗 ℎ
𝐴 𝐶𝑘 𝐶𝑘+1 𝐵
By the theorem above we note that the domain on the sides must be rectangles.
If we draw it it can be naturally found out that whole graph must look like a table,
so there is naturally next theorem.
Theorem 3.2. If 𝐴𝑖 satisfies the conditions, then the domains are congruent rect-
angles.
3 The Situation of a Square 209
Proof. At first we show that the domains on one side are congruent rectangles
(Picture 1). For any two adjacent domains 𝕋𝐴1 and 𝕋𝐴2 , the intersection of them
must be perpendicular bisector of the two origin points 𝐴1 and 𝐴2 . Thus 𝕋𝐴1 and
𝕋𝐴2 are congruent rectangles since they have the same area. Then we prove that
all the domains are congruent rectangles (Picture 2). For a rectangle 𝕋𝐴3 in corner.
The symmetric point of 𝐴3 about the topside 𝐴4 is another origin point and we
can prove that 𝕋𝐴4 is also a rectangle by using theorem 3.1 again for the topside
of all the domains on the downside of the square. Then we can prove that 𝕋𝐴4
is congruent to 𝕋𝐴3 and we can show that there are a rows of rectangles on the
domain of the first row (Picture 3). We can repeat this operation until prove that all
the domains are congruent rectangles.
𝐴4
𝐴1 𝐴2 𝐴3
Now it only needs to discuss the situation theorem 3.2 gives. If 𝑛 = 1, it is clear
that one can choose 𝐴 in the middle of the square to turn at least half of the square
into green. Now we want to show that 𝑛 = 1 is the only good situation.
If 𝑛 > 1, there are only two cases. The one is that the table has two or more
rows and columns and the other is that it has only one row or one column.
𝐵4
𝐴1 𝐴2
𝐵1
𝐴3 𝐴4
𝐵2 𝐵3
For the first case, we consider 4 rectangles which aligned as the picture above.
Choose the midpoint of 𝐴1 𝐴3 as 𝐵1 and choose 𝐵2 , 𝐵3 , 𝐵4 behind and very close
to 𝐴2 , 𝐴3 , 𝐴4 (we can choose them well enough by lemma3). Then 𝐵1 changes at
210 A Game About Competing Area
𝑎𝑏 𝑎3
least 2
+ 16𝑏
into green, here 𝑎 and 𝑏 are the length and width of the rectangle.
3𝑎𝑏 𝑎3
And 𝐵2 , 𝐵3 , 𝐵4 change almost 2
into green. Because 16𝑏 is a positive number
𝑆𝑇 −𝑎𝑏
and other points can change 2
− 𝜖 into green, then we get contradiction.
O
𝐴1 𝐴2 𝐵2
𝐵1
𝐸
𝐷
𝑇 𝐷′
For the second case, select two adjacent rectangles 𝕋𝐴1 , 𝕋𝐴2 . Because the
length of the rectangles 𝑎 are twice more then there width 𝑏, we can choose a point
𝑂 on the perpendicular bisector of 𝐴1 𝐴2 such that 2|𝑂𝐴1 | < |𝑂𝑇 | by choosing
|𝑂𝑇 | an suitable number (for example 5𝑎 9
). Here 𝑇 is an endpoint of the intersec-
tion line of two domains. Let the intersection of |𝑂𝐴1 | and the left side of 𝕋𝐴1
be 𝐸, rotate 𝑂𝐸 counterclockwise a small angle 𝜃 which makes |𝑂𝐷| ≤ |𝑂𝑇 |.
Find a point 𝐷′ on the bottom side of 𝕋𝐴2 which makes ∠𝑇 𝑂𝐷′ be 𝜃. Then
𝑆𝐸𝑂𝐷 ≤ 𝑆𝑇 𝑂𝐷′ . Choose 𝐵 as the symmetry point of 𝐴1 about 𝑂𝑃 . Call the lower
left corner 𝐴1 be 𝐹 . Then we find that 𝐵 changes the area 𝐷𝑂𝐷′ 𝐹 into green.
The area of 𝐷𝑂𝐷′ 𝐹 is 𝑆𝑇 𝑂𝐷′ − 𝑆𝐸𝑂𝐷 , more than half area of a rectangle. Choose
𝑆 −𝑎𝑏
other points suitably to change 𝑇 2 − 𝜖 and we get the contradiction.
Now we have solved the case that 𝑇 is a square. It shows a basic idea to solve
the question, and we will see in the next section that this example is more important
than we think. The idea of the first main theorem comes from here and it is a case
needed to be discussed in the second main theorem.
Theorem 4.1. For a fixed 𝑛 there exists a 𝑇 which makes (𝑇 , 𝑛) a good pair.
Theorem 4.2. For a fixed 𝑇 and a large enough 𝑛, (𝑇 , 𝑛) is not a good pair.
4 Two Theorems when 𝑇 or 𝑛 is Fixed 211
The two theorems show that (𝑇 , 𝑛) may possibly be good or bad when 𝑇 or
𝑛 changes. Moreover we can see another fact by the proof then: the condition is
better when a domain would not “interfere” with another domain.
𝑇 𝐵
𝐴1 𝐸 𝐴2
𝐷
Label the points with letters as it shown in the picture above, we figure out that
𝐻𝐼 𝐿𝑀 are the perpendicular bisectors of 𝐴1 𝐵 and 𝐴2 𝐵. We need to prove that
𝑆𝐻𝐼𝑀𝐿 ≤ 4. Note that |𝐴1 𝐷| = |𝐵𝐷| ≥ |𝐷𝐸|. Then we get 𝑆𝐻𝐼𝐺𝐾 ≤ 𝑆𝐹 𝐻𝐼𝐽 .
Similarly we have that 𝑆𝐺𝐾𝑀𝐿 ≤ 𝑆𝑀𝐿𝑁𝑂 and then get the conclusion.
𝐽
𝐷
𝐴1 𝐴2
𝐾
𝐿
𝐵
𝐻 𝐸 𝐼 𝐹
We may as well let the coordinate of the two origin points be (4, 1), (12, 1) and
the coordinate of 𝐵 be (4 + 𝑎, 1 − 𝑏).
212 A Game About Competing Area
1 |𝐷𝐼| 𝑎 𝑏
𝑆𝐷𝐼𝐹 = |𝐷𝐼||𝐼𝐹 | = |𝐷𝐼| × ≤ 2|𝐷𝐼||𝑘𝐵𝐴2 | = 2|𝐷𝐸|
2 |𝑘𝐷𝐹 | √𝑎2 + 𝑏2 8 − 𝑎
𝑏
≤|𝐷𝐸| ≤ |𝐸𝐷||𝐾𝐴1 | = 2𝑆𝐸𝐷𝐴1 ≤ 𝑆𝐸𝐷𝐴1 + 𝑆𝐿𝐸𝐴1
2
≤𝑆𝐸𝐷𝐴1 + 𝑆𝐿𝐻𝐸𝐴1 = 𝑆𝐻𝐸𝐷𝐽
(4.0.1)
So 𝑆𝐷𝐸𝐹 ≤ 𝑆𝐼𝐽 𝐻 .
𝐴
𝐷
𝐵 𝐶
Let the coordinate of the three origin points be (4, 1), (12, 1), (20, 1) and coordinate
of 𝐵 be (12 − 𝑎, 1 − 𝑏). Here 𝑎 and 𝑏 are positive. We can calculate the area of
𝐴𝐵𝐶𝐷.
Remark 1. In fact we have not discussed all the cases, such as the case we discuss
in the last of section 3. However, we know that the other cases can be treated as a
part of these three cases we have discussed.
Remark 2. It is clear that the length of small rectangles is not the best constant
here. The best constant of length ∶ width is around 2.
4 Two Theorems when 𝑇 or 𝑛 is Fixed 213
Now it is time to go to our last main theorem. The main idea of proof is that
when 𝑛 is big enough there must be a lot of quadrilateral and hexagonal domains.
We need to find a contradiction in it.
At first we prove that there cannot be quadrilaterals in most cases.
Theorem 4.3. If 𝐴1 , 𝐴2 ...𝐴𝑛 satisfy the condition and there is a quadrilateral in
𝕋𝐴𝑖 , then the all the domains must be like the case in theorem 4.1 (a row of rectan-
gles).
Proof. By lemma 5 the quadrilateral must be a parallelogram. At first we prove
that it must be a rectangle. For the origin point of the quadrilateral 𝕋𝐴1 , we notice
that the symmetric point of 𝐴1 about the side of 𝕋𝐴1 which is not the bound of 𝑇
must be another origin point, called 𝐴2 . By lemma 4 and lemma 5 we know that
𝕋𝐴2 is also a quadrilateral. If 𝕋𝐴1 is not a rectangle, then two of its angle are obtuse
angles. We select the angle that one side is also the side of 𝕋𝐴2 and label it with
letter 𝑆. Then we label other points like the picture.
𝐵1
𝐴2
𝐻 𝐼
𝑈
𝑆 𝐵1
𝐾 𝐽
𝐴1
Hence only if 𝑇 is a rectangle then there could be quadrilateral in 𝕋𝐴1 , 𝕋𝐴2 ...𝕋𝐴𝑛 .
However, when 𝑇 is a rectangle and 𝑛 becomes big enough it becomes a case in
section 3 which turns out to be bad.
214 A Game About Competing Area
Then we prove that there would be a lot of hexagons, which will be done by
Euler’s Formula 𝑉 − 𝐸 + 𝐹 = 1.
Theorem 4.4. We divide a polygon 𝑇 into small polygons with even edges. If there
are not quadrilaterals, the numbers of “points on edges”, “the polygons which
have more than eight edges”, “the points in 𝑇 which is the endpoint of more than
four lines” can be controlled by a constant 𝑐 only depending on 𝑇 .
⎧𝑛 − 2 + 𝑝 + 2𝑟 ≥ 4ℎ + 6𝑡
⎪
⎨ℎ + 𝑡 + 𝑝 + 𝑟 − 𝑙 − 𝑝 = 1
⎪𝑙 ≥ 𝑝+3𝑟+𝑓
⎩ 2
The first inequality is got by calculating degrees, the second one follows from
Euler’s Formula and the third one is from every point inside 𝑇 has at least 3 lines
from it and every point on edges has at least one.
Solving the inequalities system we get that 𝑛 − 6 ≥ 𝑝 + 2𝑡 + 2𝑓 and the conclu-
sion. ◻
This theorem gives out a way leading to our goal. It tells us that there are enough
hexagons which are not on the edges. It is interesting that a hexagon will not cause
contradiction directly but when a hexagon is surrounded by other hexagons we can
get contradiction.
By theorem 4.4 we know that if 𝑛 is big enough we can select a lot of adjacent
hexagons, every of which is surrounded by other six hexagons and each vertex of
the original hexagon is just the vertex of three domains.
Consider such a hexagon:
𝐴 𝐵 𝐴𝐵 ∥ 𝐹 𝐶 ∥ 𝐸𝐷
𝐴𝐹 ∥ 𝐵𝐸 ∥ 𝐶𝐷
𝐹 𝐸 ∥ 𝐴𝐷 ∥ 𝐵𝐶
𝐹 𝐶
𝐸 𝐷
Now we give out the proof, the idea is based on translating 𝐵 a little distance
from 𝐴, because calculating of areas is a bit complicated. We use different colors
to sign different domain for convenience. We clarify our notation. Let 𝒯 be any
function such that lim𝜖→0 𝒯 𝜖(𝜖) = 1, and 𝑜(𝜖) be higher order infinitesimal of 𝜖.
𝑍
𝐴1 𝐻
𝐺 yellow
𝐼 𝑉 𝐽
𝑀𝑄𝑈 𝐵
black 𝑆
𝑂𝑅 orange
gray
𝐴2 purple 𝐴4
𝑇 𝑊 𝑌
𝑁 cyan 𝐿
𝑋
𝐾𝑃
𝐴3
𝑄
𝑀 𝑈
𝑆gray + 𝑆black
=|𝑈 𝑇 |(|𝑂𝑅| + 𝑜(|𝑂𝑅))
=|𝑈 𝑇 |(|𝑆𝑂|| cos ∠𝑆𝑂𝑅 + 𝑜(𝑆𝑂)| (4.0.3)
𝒯 (𝜖)
=|𝑈 𝑇 | cos 𝜃
2
and
𝑆purple − 𝑆black
1
=𝒯 ( |𝑁𝑂2 − 𝑀𝑂2 || sin ∠𝑃 𝑂𝑁|)
2
1 (4.0.4)
=𝒯 ( |𝑇 𝑆 2 − 𝑆𝑈 2 | sin ∠𝐴1 𝐴2 𝐵)
2
sin ∠𝐴2 𝐴1 𝐵
=𝒯 (𝑑𝑈 𝑇 |𝑈 𝑇 |𝜖 ).
𝐴1 𝐴2
4 Two Theorems when 𝑇 or 𝑛 is Fixed 217
It is the case that 𝐵 moves a little distance 𝜖 on the vertical line from 𝐴 to 𝑆𝑇 .
We let this 𝐵 be 𝐵1 . Similarly if 𝐵 moves 𝜖 on the perpendicular lines to other
edges and we calculate the average of them we get that 𝐵 changes the domain of
area
𝑆hexagon 1 1
+ 𝒯 (𝜖)( ∑(|𝑈 𝑇 | cos 𝜃 + |𝑅𝑆| cos 𝜃 ′ + |𝑇 𝑆| − |𝐺𝐻|)).
2 6 2 cyc
𝑃 𝑄 𝑃 𝑄
𝑈 𝑅
𝐺 𝐻 𝐺 𝐻
𝐴1 𝐴1
𝑅 𝑈
𝐿 𝑇 𝑆 𝐾 𝐿 𝑇 𝐾 𝑆
If the equation were not established, then we can always find a 𝐵 in the hexagon
and 𝐵 changes more than half area of the hexagon into its own domain. Then use
the tricks we have used many times to select other 𝐵𝑖 . Thus the equations are all
established.
There is one case left: when 𝑈 𝑇 ⟂ 𝑇 𝑆, the projection of 𝐺𝐻 is the same as
𝑈 𝑅. In that case, we can change the direction of 𝐵 moved from 𝐴 to get contra-
diction. ◻
When we have such a strong conclusion, we can select three adjacent hexagons
and calculate the angles to get further conclusion.
Theorem 4.6. If 3 hexagons are adjacent, they are all regular hexagons.
𝐴1
𝐶 𝐷
𝑇
𝐴2 𝐴3
Proof. We just need to calculate angles. In the equation we use the facts that
𝐴1 𝐵𝑇 𝐶, 𝐴3 𝐷𝑇 𝐵 and 𝐴2 𝐷𝑇 𝐶 are parallelograms, 𝐴1 and 𝐴2 and 𝐴3 are sym-
metric points about 𝑇 𝐶, 𝑇 𝐷, and 𝑇 𝐵.
Consider four adjacent regular hexagons. We can show that it’s not good by
date in this picture.
5 Summary 219
𝐴1 :(0.500,0.866)
𝐵:(0.404,2.121)
𝑆ℎ𝑒𝑥𝑎𝑔𝑜𝑛 =2.598
𝑆𝑂𝑃 𝑄𝑅 =1.309
𝑆𝑂𝑃 𝑄𝑅
𝑆
=0.504
ℎ𝑒𝑥𝑎𝑔𝑜𝑛
𝐴4
𝑅
𝐵 𝑄
𝐴2 𝐴3
𝑂 𝑃
𝐴1
We can choose a 𝐵 ∶ (0.404, 2.121) such that it changes more than half a
hexagon, so this situation is also not good.
We have finished all the theorems and given out some conclusions about this
question. However, it is a pity that we have not solved the problem completely, we
will give out some ideas about the question in next section.
5 Summary
Firstly we recall the conclusions we have gotten: for initial question, (square, 𝑛) is
good if and only if 𝑛 = 1. For any 𝑛 there exists 𝑇 such that (𝑇 , 𝑛) is good. For any
𝑇 , when 𝑛 is large enough (𝑇 , 𝑛) are always bad.
A fact is that it is hard to construct another good (𝑇 , 𝑛) when 𝑛 is more than
2 and 𝑇 is not a rectangle (except those in theorem 4.1). The difficulty is that we
cannot divide a convex polygon into centrally symmetric polygons which has more
than five edges easily. Moreover we need to use area-equivalent polygons: even
diagonal parallel (polygons with 4𝑛 + 2 edges) or every edge has two edges vertical
to it (polygons with 4𝑛 edges) when a polygon is surrounded by other polygons. It
is hard to achieve all such things. It can be conjectured there are no other good pairs
except the cases in theorem 4.1. However, on the other hand, as we do not have the
condition that 𝑛 is big enough, it is not easy to consider this problem locally. An
available idea is to prove that a polygon cannot be divided into such polygons. We
may prove this by considering the lines between origin points and using the length
and angles to calculating areas. It is also possible to calculating integral in some
certain area locally around an origin point to prove the existence of 𝐵𝑖 . However,
they all need nontrivial calculation.
220 A Game About Competing Area
徐凯1
摘 要
在这篇短文中, 我们从变号矩阵的计数问题出发, 将其表述为六顶点模
型中态和的计算, 并通过 Yang–Baxter 方程和可积性给出问题的显式解.
1 变号矩阵到六顶点模型
我们首先重新表述变号矩阵的计数问题, 将其化归为统计力学中的六顶点
模型.
在六顶点模型中, 我们关心平面上每边带定向, 并且每个顶点恰好有两条
边进入两条边离开的正方形图, 每一个这样的图称为一个态 (state). 给定一个
态, 在每个顶点附近有如下六种可能的定向:
𝑎 𝑏 𝑐 𝑑 𝑒 𝑓
1
原清华大学数学系数 41 班.
222 变号矩阵与可积系统
1 −1 0 0 0 0
2 YBE 的场论来源
3 六顶点模型的解
𝑞 𝑥∕2 −𝑞 −𝑥∕2
取一个未定元 𝑞, 记 [𝑥] = 𝑞 1∕2 −𝑞 −1∕2
, 对一个带 𝑥 标记的顶点
我们为六种定向赋权如下:
这可以视为指定了一个二维场论的弹性碰撞 (大小为 22 × 22 ) 的 𝑆 矩阵
(在可积格点模型中通常称为 𝑅 矩阵), 而 YBE 正是这个可积场论的相容性条
件:
𝑧 𝑧
𝑥 𝑦
= [𝑥] + [𝑥 − 1]
224 变号矩阵与可积系统
𝑥 =
𝑥−𝑦
𝑦
𝑥0
⋯
𝑥1
⋮ ⋱ ⋮
𝑥𝑛−1 ⋯
𝑦0 𝑦1 𝑦𝑛−1
(−1)𝑛 (∏𝑛−1
𝑖=0 𝑞
(𝑦𝑖 −𝑥𝑖 )∕2
) ∏0≤𝑖,𝑗<𝑛 [𝑥𝑖 − 𝑦𝑗 ][𝑥𝑖 − 𝑦𝑗 − 1]
𝑍(𝑛; 𝑋, 𝑌 ) = det 𝑀,
(∏0≤𝑗<𝑖<𝑛 [𝑥𝑖 − 𝑥𝑗 ]) (∏0≤𝑖<𝑗<𝑛 [𝑦𝑖 − 𝑦𝑗 ])
其中
1
𝑀𝑖,𝑗 = .
[𝑥𝑖 − 𝑦𝑗 ][𝑥𝑖 − 𝑦𝑗 − 1]
由第一节的论证我们知道 𝑛 × 𝑛 变号矩阵的数量恰好为方冰态数量, 而方
冰态中顶点 𝑎 比 𝑏 多 𝑛 个, 𝑐, 𝑑 数目相等, 𝑒, 𝑓 也相等, 所以可得变号矩阵数目
为
2
1 𝑛−𝑛
(2) (−1)𝑛 𝑞 𝑛∕4 𝑍1∕2 (𝑛)|𝑥=1 ,
其中 𝑍1∕2 (𝑛) = 𝑍(𝑛; 1∕2, ..., 1∕2, 0, ..., 0), 𝑥 = [ 21 ]−2 = 𝑞 1∕2 + 𝑞 −1∕2 + 2.
这是定理 2 中表达式的 (可去) 奇点, 使用 L’Hôpital 法则我们可以直接计
算得如下表达式 (计算细节请参见 [Kup96]):
参考文献 225
定理 3. 𝑛 × 𝑛 变号矩阵数目为
1! 4! 7! ⋯ (3𝑛 − 2)!
.
𝑛! (𝑛 + 1)! (𝑛 + 2)! ⋯ (2𝑛 − 1)!
由此我们就解决了变号矩阵的计数问题.
4 总结
可积系统对现代数学的深刻影响远不止于此: YBE 的解通常称为 𝑅 矩
阵, 是量子群与辫张量范畴的核心信息, 对其与场论的关系的研究启发了大量
重要的数学. 譬如著名的 Kazhdan–Lusztig 等价联系起量子群的 𝑅 矩阵与共
形场论中 Knizhnik–Zamolodchikov 方程解的单性 (monodromy). 又如 Maulik–
Okounkov 通过 𝑅 矩阵构造了瞬子模空间上同调上的量子群作用, 给出了量子
场论中 AGT 对应的数学验证. 希望本文对中学阶段的同学们有所启发, 借此机
会了解现代数学中众多深刻精微的洞见.
参考文献
[Kup96] Greg Kuperberg (1996). “Another proof of the alternative-sign matrix conjec-
ture”. International Mathematics Research Notices 1996 (3), 139–150.