![THE LOCAL LANGLANDS CORRESPONDENCE FOR …THE LOCAL LANGLANDS CORRESPONDENCE FOR GL n OVER p-ADIC FIELDS PETER SCHOLZE Abstract. We extend our methods from [24] to reprove the Local](https://static.fdocuments.net/doc/165x107/5e7dc0cf22850b473d13ccc4/the-local-langlands-correspondence-for-the-local-langlands-correspondence-for-gl.jpg)
Langlands correspondence and Bezrukavnikov’s ... - arXiv
274
Langlands correspondence and Bezrukavnikov’s equivalence Anna Romanov and Geordie Williamson March 4, 2021 This document contains notes (taken by the first author) from a course (taught by the sec- ond author) at the University of Sydney. The course took place over two extended semesters, and consisted of two hours of lectures per week between 2019-2020. The course is divided in two parts, reflecting the two terms. The first part (concluding with Lecture 11) is an attempt to give an informal introduction to what the Langlands program is about, from an arithmetical point of view. We assume the audience (like the lecturer) is a beginner in this subject, but had a first course in com- plex analysis, Galois theory, topology and representation theory. At times we also assume background in algebraic geometry. Not much is proved, but we try to give enough detail to convince the reader that there is a lot of marvellous mathematics here. The second longer part (Lectures 12 through 34) tries to give enough background in geometric representation theory to understand Bezrukavnikov’s equivalence. Taking the tamely ramified local Langlands correspondence as motivation, we pass through work of Deligne-Lusztig and Ginzburg giving a coherent construction of the affine Hecke algebra. Here we are roughly following the book of Chriss-Ginzburg, but our route is different at times. Bezrukavnikov’s equivalence is a categorification of this isomorphism, and is our aim for the rest of the notes. In order to even understand the statement of Bezrukavnikov’s equivalence, we need much of the toolbox of modern geometric representation theory: perverse sheaves; highest weight categories; Koszul duality; the geometric Satake equivalence, etc. A major role is also played by monoidal categories and their actions (“higher representation theory”). We try to spend enough time on each of these topics to make students feel somewhat comfortable with the ideas. It goes without saying that all of this took much longer to cover than anyone expected, and at the very end we arrive at a statement of Bezrukavnikov’s equivalence. Its proof will have to wait for the next lecture series! The spirit of these notes is very informal, and this is intentional. We hope that they are nonetheless useful for a casual reader trying to orient themselves in a fascinating but potentially intimidating landscape. We assume that the reader is willing to take some things on faith, and have tried to be honest. Audience members were encouraged to do exercises throughout, and this wouldn’t be bad advice for any potential reader either. 1 arXiv:2103.02329v1 [math.RT] 3 Mar 2021
Transcript of Langlands correspondence and Bezrukavnikov’s ... - arXiv
Langlands correspondence and Bezrukavnikov’sequivalence
Anna Romanov and Geordie Williamson
March 4, 2021
This document contains notes (taken by the first author) from a course (taught by the sec-ond author) at the University of Sydney. The course took place over two extended semesters,and consisted of two hours of lectures per week between 2019-2020. The course is divided intwo parts, reflecting the two terms.
The first part (concluding with Lecture 11) is an attempt to give an informal introductionto what the Langlands program is about, from an arithmetical point of view. We assumethe audience (like the lecturer) is a beginner in this subject, but had a first course in com-plex analysis, Galois theory, topology and representation theory. At times we also assumebackground in algebraic geometry. Not much is proved, but we try to give enough detail toconvince the reader that there is a lot of marvellous mathematics here.
The second longer part (Lectures 12 through 34) tries to give enough background ingeometric representation theory to understand Bezrukavnikov’s equivalence. Taking thetamely ramified local Langlands correspondence as motivation, we pass through work ofDeligne-Lusztig and Ginzburg giving a coherent construction of the affine Hecke algebra.Here we are roughly following the book of Chriss-Ginzburg, but our route is different at times.Bezrukavnikov’s equivalence is a categorification of this isomorphism, and is our aim for therest of the notes. In order to even understand the statement of Bezrukavnikov’s equivalence,we need much of the toolbox of modern geometric representation theory: perverse sheaves;highest weight categories; Koszul duality; the geometric Satake equivalence, etc. A major roleis also played by monoidal categories and their actions (“higher representation theory”). Wetry to spend enough time on each of these topics to make students feel somewhat comfortablewith the ideas. It goes without saying that all of this took much longer to cover than anyoneexpected, and at the very end we arrive at a statement of Bezrukavnikov’s equivalence. Itsproof will have to wait for the next lecture series!
The spirit of these notes is very informal, and this is intentional. We hope that theyare nonetheless useful for a casual reader trying to orient themselves in a fascinating butpotentially intimidating landscape. We assume that the reader is willing to take some thingson faith, and have tried to be honest. Audience members were encouraged to do exercisesthroughout, and this wouldn’t be bad advice for any potential reader either.
1
arX
iv:2
103.
0232
9v1
[m
ath.
RT
] 3
Mar
202
1
Contents
Lecture 1: Reciprocity Laws1.1 Reciprocity Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.2 Higher dimensional varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . 91.3 What is going on here? What does this have to do with representation theory? 91.4 Schematic picture of the Langlands correspondence . . . . . . . . . . . . . . 101.5 Chebotarev density theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Lecture 2: Review of some algebraic number theory2.1 Number fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132.2 An analogy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142.3 The fundamental exact sequence . . . . . . . . . . . . . . . . . . . . . . . . . 152.4 Ramification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162.5 The case of Galois extensions . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Lecture 3: L-functions3.1 The Riemann ζ-function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203.2 Why do we care? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213.3 Dirichlet L-functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223.4 Dedekind ζ-functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233.5 Walking across the bridge . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243.6 Artin L-functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Lecture 4: The Sato-Tate conjecture4.1 Equidistribution in representation theory . . . . . . . . . . . . . . . . . . . . 274.2 Elliptic curves and the Sato-Tate conjecture . . . . . . . . . . . . . . . . . . 304.3 Equidistribution and L-functions . . . . . . . . . . . . . . . . . . . . . . . . 34
Lecture 5: Infinite Galois theory and global class field theory5.1 Infinite Galois theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375.2 Global class field theory, first version . . . . . . . . . . . . . . . . . . . . . . 395.3 Global class field theory a la Artin . . . . . . . . . . . . . . . . . . . . . . . 40
Lecture 6: Local Galois groups and local class field theory6.1 A potted history of class field theory . . . . . . . . . . . . . . . . . . . . . . 436.2 A trip across the bridge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436.3 Local class field theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466.4 Structure of Galois groups of local fields . . . . . . . . . . . . . . . . . . . . 49
Lecture 7: Representation theory of p-adic groups7.1 Tying up some loose ends . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517.2 The no small subgroups argument . . . . . . . . . . . . . . . . . . . . . . . . 527.3 You could have guessed the LLC for GL2! . . . . . . . . . . . . . . . . . . . 537.4 Basic representation theory of p-adic groups . . . . . . . . . . . . . . . . . . 55
Lecture 8: Precise statement of local Langlands for GL2, p 6= 28.1 Basic representation theory of p-adic groups, continued . . . . . . . . . . . . 588.2 Canonical norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608.3 Smooth admissible representations of GL1(K) . . . . . . . . . . . . . . . . . 61
2
8.4 Smooth admissible representations of GL2(K) . . . . . . . . . . . . . . . . . 618.5 Weil-Deligne representations . . . . . . . . . . . . . . . . . . . . . . . . . . . 648.6 The local Langlands correspondence for GL2 . . . . . . . . . . . . . . . . . . 65
Lecture 9: The case of p = 2 and the Satake isomorphism9.1 Ramification filtration revisited . . . . . . . . . . . . . . . . . . . . . . . . . 669.2 More details on Weil–Deligne representations . . . . . . . . . . . . . . . . . . 689.3 Why is LLC for p = 2 special? . . . . . . . . . . . . . . . . . . . . . . . . . . 699.4 Unramified representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 709.5 Hecke algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Lecture 10: The big picture10.1 The very big picture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
10.1.1 Dimension 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7410.1.2 Dimension ≥ 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
10.2 Local Langlands correspondence for split groups . . . . . . . . . . . . . . . . 7810.3 The Deligne–Langlands conjecture . . . . . . . . . . . . . . . . . . . . . . . . 7810.4 Geometric Satake equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . 8010.5 Bezrukavnikov’s equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Lecture 11: Review of the first semester11.1 The local Langlands correspondence . . . . . . . . . . . . . . . . . . . . . . . 8111.2 The unramified story . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8211.3 The tamely ramified with unipotent monodromy (TRUM) story . . . . . . . 8311.4 Affine Weyl groups and affine Hecke algebras . . . . . . . . . . . . . . . . . . 84
Lecture 12: The Deligne-Langlands conjecture12.1 The Iwahori–Matsumoto Hecke algebra . . . . . . . . . . . . . . . . . . . . . 8812.2 The Deligne–Langlands conjecture . . . . . . . . . . . . . . . . . . . . . . . . 9012.3 Geometric setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9212.4 The Springer resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Lecture 13: Springer fibres and the Steinberg variety13.1 The Springer resolution, continued . . . . . . . . . . . . . . . . . . . . . . . 9613.2 Many examples of Springer fibres . . . . . . . . . . . . . . . . . . . . . . . . 9713.3 The conormal space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10113.4 The Steinberg variety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Lecture 14: Springer correspondence and Borel-Moore homology14.1 The Springer correspondence . . . . . . . . . . . . . . . . . . . . . . . . . . . 10514.2 Borel-Moore homology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10714.3 Geometric convolution algebras . . . . . . . . . . . . . . . . . . . . . . . . . 108
Lecture 15: The Kazhdan–Lusztig isomorphism15.1 More convolution algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11115.2 Equivariant K-theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11215.3 The Kazhdan–Lusztig isomorphism . . . . . . . . . . . . . . . . . . . . . . . 11415.4 Bezrukavnikov’s equivalence: rough outline . . . . . . . . . . . . . . . . . . . 116
Lecture 16: The constructible side of Bezrukavnikov’s equivalence16.1 Weil conjectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
3
16.2 The Hecke algebra, revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . 12116.3 Perverse sheaves on P1C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Lecture 17: Constructible and perverse sheaves on curves17.1 Constructible sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12517.2 Perverse sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12617.3 Nearby and vanishing cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Lecture 18: Beilinson gluing on curves18.1 A fact about unipotent monodromy . . . . . . . . . . . . . . . . . . . . . . . 13318.2 The unipotent vanishing cycles functor . . . . . . . . . . . . . . . . . . . . . 13418.3 Gluing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Lecture 19: The derived category of perverse sheaves19.1 Overview of Beilinson gluing . . . . . . . . . . . . . . . . . . . . . . . . . . . 14319.2 A theorem of Beilinson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14419.3 K(π, 1) spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14519.4 Yoneda extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14619.5 A key ingredient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14719.6 Proof of Beilinson’s theorem for curves . . . . . . . . . . . . . . . . . . . . . 149
Lecture 20: Perverse sheaves for affine stratifications20.1 Highest weight categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15420.2 Perverse sheaves for an affine stratification are a highest weight category . . 15720.3 Where are we going for the next few weeks? . . . . . . . . . . . . . . . . . . 160
Lecture 21: Equivariant K-theory of the Steinberg variety21.1 Equivariant K-theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16321.2 Equivariant K-theory of the Steinberg variety . . . . . . . . . . . . . . . . . 16621.3 The spherical and anti-spherical modules . . . . . . . . . . . . . . . . . . . . 168
Lecture 22: Proof of the Kazhdan–Lusztig isomorphism22.1 Notation and set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17022.2 The case of SL2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17122.3 The general case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
Lecture 23: Gaitsgory’s central sheaves23.1 A change of notation in the Hecke algebra . . . . . . . . . . . . . . . . . . . 17823.2 Gaitsgory’s central sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17823.3 Extended example: the natural representation of GL2 . . . . . . . . . . . . . 18123.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Lecture 24: A hitchhiker’s guide to the Hecke category24.1 The Hecke category: setting the scene . . . . . . . . . . . . . . . . . . . . . . 18924.2 First try . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19024.3 Second try . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19124.4 Third try . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19224.5 Fourth (and final) try . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Lecture 25: The categorical anti-spherical module and its symmetries25.1 The affine Hecke category . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19725.2 The categorical (anti-)spherical module . . . . . . . . . . . . . . . . . . . . . 198
4
25.3 Symmetries of categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20125.4 A notational interlude: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20225.5 The basic approach of Arkhipov-Bezrukavnikov . . . . . . . . . . . . . . . . 202
Lecture 26: Monoidal categories and their actions26.1 What is a monoidal category? . . . . . . . . . . . . . . . . . . . . . . . . . . 20326.2 Modules for monoidal categories . . . . . . . . . . . . . . . . . . . . . . . . . 20426.3 Examples of modules over monoidal categories . . . . . . . . . . . . . . . . . 20526.4 What is going on here? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20726.5 Example: Representations of the Verlinde category . . . . . . . . . . . . . . 208
Lecture 27: Abelian categories over stacks27.1 Modules over algebras in monoidal categories . . . . . . . . . . . . . . . . . . 21127.2 Linear algebra over a stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21227.3 A brief review of descent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21427.4 Back to sheaves over stacks . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Lecture 28: (De)equivariantisation28.1 A simple example of this phenomenon . . . . . . . . . . . . . . . . . . . . . 21828.2 Equivariantisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21928.3 Deequivariantisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21928.4 RepG enrichments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22028.5 Deequivariantisation principle . . . . . . . . . . . . . . . . . . . . . . . . . . 22128.6 Tannakian formalism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Lecture 29: Coherent sheaves on base affine space29.1 Serre’s description of coherent sheaves on Pn . . . . . . . . . . . . . . . . . . 22629.2 How can we describe coherent shaves on G/B? . . . . . . . . . . . . . . . . . 22729.3 Motivational interlude . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22829.4 Arrows between coherent sheaves on G/U . . . . . . . . . . . . . . . . . . . 229
Lecture 30: Constructible sheaves on finite flag varieties and braid groups30.1 A fun calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23130.2 The finite Hecke category . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23230.3 The Braid group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23430.4 Further properties of the braid group action . . . . . . . . . . . . . . . . . . 237
Lecture 31: Affine flags, affine braids, and Wakimoto sheaves31.1 Wakimoto sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24131.2 Motivational interlude . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24331.3 Beilinson–Gaitsgory–Kottwitz (BGK) central sheaves . . . . . . . . . . . . . 244
Lecture 32: Whittaker sheaves and Arkhipov-Bezrukavnikov theorem32.1 Averaging functors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24832.2 Another fun calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24832.3 Lightning introduction to the Whittaker world . . . . . . . . . . . . . . . . . 25132.4 Bird’s eye view of the rest of the proof . . . . . . . . . . . . . . . . . . . . . 253
Lecture 33: Soergel bimodules, Soergel calculus, and BGK central sheaves33.1 The Hecke category and Soergel bimodules . . . . . . . . . . . . . . . . . . . 25733.2 BGK central sheaves for G = SL2 . . . . . . . . . . . . . . . . . . . . . . . . 260
5
33.3 Braid group categorification . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
Lecture 34: Bezrukavnikov’s equivalence34.1 Potted history of Koszul duality . . . . . . . . . . . . . . . . . . . . . . . . . 26534.2 A toy example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26634.3 Torus monodromic sheaves a la Bezrukavnikov-Yun . . . . . . . . . . . . . . 26734.4 Bezrukavnikov’s equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
6
Lecture 1: Reciprocity Laws
If you do nothing else with this course this semester beyond attending the first lecture, youshould at least try to read [Lan90].
1.1 Reciprocity Laws
We start at the natural starting place: an equation. Consider the equation
x2 + 1 = 0.
If p is a prime, one might wonder: how many solutions does this equation have, modulo p?Some calculations will reveal the following table.
p 2 3 5 7 11 13 17 19 23# of sol’s mod p 1 0 2 0 0 2 2 0 0
p mod 4 2 3 1 3 3 1 1 3 3...
We see a pattern. The prime 2 is weird, so we ignore it. But for the rest, it seems that
# of solutions mod p 6= 2 =
2 if p = 1 mod 4,
0 if p = 3 mod 4.
This pattern is surprising. It appears to be saying that there is a global rule governing thenumber of solutions mod p; that is, that the different primes somehow “talk to one another”.
Here we can give a simple proof of why our claim above must be true. Assume p 6= 2.We have a short exact sequence
1→ (F×p )2 → F×p → ±1 → 1,
where the third arrow is given by x 7→ xp−1
2 . Therefore,
−1 is a square mod p ⇐⇒ (−1)p−1
2 = 1 ⇐⇒ p− 1
2is even ⇐⇒ p = 1 mod 4.
Let’s do another example. Consider the equation
x2 − 3 = 0.
We ask the same question and compute:
p 2 3 5 7 11 13 17 19 23 29 31# of sol’s mod p 1 1 0 0 2 2 0 0 2 0 0
p mod 12 2 3 5 7 11 1 5 7 11 5 7...
7
Again, we have some small weird primes (2 and 3), so we throw them out. For the rest, wemake a guess:
# of solutions mod p 6= 2, 3 =
2 if p = 1, 11 mod 12,
0 if p = 5, 7 mod 12.
To prove that this is indeed the case, we introduce a little more technology. Let p 6= 2 be aprime. Define
ε(p) =
0 if p = 1 mod 4,
1 if p = 3 mod 4
and the Legendre symbol(x
p
)= x
p−12 mod p =
1 if x is a square mod p,
−1 if x is not a square mod p.
(For example, we saw above that(−1p
)= (−1)ε(p).)
Theorem 1.1. (Gauss’s Law) Let p, q be distinct primes 6= 2. Then(p
q
)(q
p
)= (−1)ε(p)ε(q).
With this we can prove that our guess was correct. Assume p 6= 2, 3. Then
x2 − 3 has 2 solutions mod p ⇐⇒(
3
p
)= 1
⇐⇒(p
3
)(−1)ε(p)ε(3) = 1
⇐⇒(p
3
)(−1)ε(p) = 1
⇐⇒
p = 1 mod 3 and p = 1 mod 4
p = 2 mod 3 and p = 3 mod 4
⇐⇒ p = 1 or − 1 mod 12.
These are examples of reciprocity laws. All polynomials of degree 2 can be worked outanalogously to the ones above using quadratic reciprocity (Gauss’s law). There was muchactivity on this problem starting with Gauss’s work, which finally led to Artin’s reciprocitylaw. This implied all known reciprocity laws at the time, and in particular treats polynomialsof degree 3 and 4. However, we get stuck at 5. For example, consider
x5 + 20x+ 16 = 0.
We can construct a table
p 2 3 5 7 11 13 17 19 23 29 31 37 41 43 47 53# of sol’s mod p 1 0 1 2 2 0 2 0 1 2 0 1 2 2 0 0
...
but no obvious pattern emerges. (For a table that goes much further than this one, see thesheet on the course website.) It turns out that there is a pattern, but it is very well-hidden,and to find it, we need analysis.
8
1.2 Higher dimensional varieties
We could ask similar questions for polynomials in two variables. Consider the equation
y2 = x3 + 1.
How many solutions does this equation have modulo p? Let’s try to answer this for onespecific prime. Let p = 5, and we can compile our results in the following table:
y\x 0 1 2 3 40 × × × × X1 X × × × ×2 × × X × ×3 × × X × ×4 X × × × ×
So here we found that there are five solutions modulo 5. In general, how many solutions dowe expect? Well, the map x 7→ x3 + 1 in Fp is “roughly random,” about half the elementsof Fp are squares, and for every square we get two solutions, so we expect approximately psolutions. But how often is this actually the case? We can measure the accuracy of thisestimation by studying the Sato-Tate error term:
ST (p) = p−#(solutions modulo p).
Here is another table.
p 2 3 5 7 11 13 17 19 23 29 31 37 41 43 47 53# ST (p) 0 0 0 -4 0 2 0 8 0 0 -4 -10 0 8 0 0
...
Notice how frequently the Sato-Tate error term is zero! We can now study this tableand see if any patterns emerge. This is the content of the Sato-Tate conjecture, which isbasically known thanks to recent work of Harris, Taylor, Clozel, and many others.
1.3 What is going on here? What does this have to do with rep-resentation theory?
Let f(x) ∈ Z[x] be an irreducible polynomial with integral coefficients. We can consider thesplitting field of f
K = Q(α1, . . . , αn),
and the associated Galois group
Γ = Gal(K/Q),
which acts on the set of roots α1, . . . αn. As representation theorists, our natural instinctwhen we see a group action is to linearize. Doing this here results in the permutationrepresentation
Γ H =n⊕i=1
Cαi.
9
We can also consider the reduction of f modulo p, f(x) ∈ Fp[x], as we did in the previoussection. In general, f will be reducible. If p - ∆(f) (that is, p is not one of the “weird”primes we encountered earlier), then f(x) has n roots, α1, . . . , αn ∈ Fpn . Recall that theGalois group Gal(Fpn/Fp) = Z/nZ is generated by the Frobenius map
Frobp : x 7→ xp.
Then Fp = (Fpn)Frobp , and the number of solutions of f is the number of fixed points of Frobp
on α1, . . . , αn. For such a p (“unramified”) and after a choice (“prime in O above p”), weget a bijection
α1, . . . , αn −→ α1, . . . αn,
and an element Frobp ∈ Γ such that the action of Frobp on α1, . . . , αn aligns with theaction of Frobp on α1, . . . , αn under the bijection above.
Remark 1.2. Different choices of “prime in O over p” result in conjugate Frobp’s. Hence,it is best to think of Frobp as a conjugacy class instead of an individual element.
The upshot of the discussion above is that
# solutions modulo p = # fixed points of Frobp on α1, . . . , αn= # of fixed points of Frobp on α1 . . . , αn= Tr(Frobp, H),
where H is the permutation representation introduced at the beginning of this section. Thenumber Tr(Frobp, H) is completely canonical - it doesn’t depend on any of our choices!So we’ve reduced our question of finding solutions of polynomials modulo p to computingsomething that looks very much like the character of a representation.
The Punchline: If p - ∆(f),
# solutions of f mod p = Tr(Frobp, H).
1.4 Schematic picture of the Langlands correspondence
(Do not worry if this makes no sense!) A caricature of the Langlands correspondence iscaptured in the diagram below.
“geometric” reps V of Gal(Q/Q) “character” Tr(Frobp, V ) automorphic forms
L-functions (analytic)
From any “geometric” representation of Gal(Q/Q), we can take the trace of Frobenius, aswe did in the previous section for the permutation representation H. We should think ofthis procedure as taking the character of the representation. To Tr(Frobp, V ), we can attachthe associated “L-function,” which is an analytic object. (For example, when we start with
10
the trivial representation of Gal(Q/Q), the resulting L-function is the Riemann ζ-function.)On the other hand, there is also a procedure for constructing L-functions from automorphicforms. The Langlands correspondence is an attempt to align these two sources of L-functions.
This is very deep. For example, two-dimensional representations of Gal(Q/Q) result inHecke L-functions, and the corresponding automorphic forms are modular forms. It turnsout that working out the correspondence for 2-dimensional representations is enough to proveFermat’s last theorem.
1.5 Chebotarev density theorem
If we talk of Tr(Frobp, H) as a “character,” we would like to know at least that the setFrobp for all p unramified cover the set of all conjugacy classes of Γ. This is a deeptheorem.
Theorem 1.3. (Chebotarev density theorem) Fix a conjugacy class C ⊂ Γ. Then
p unramified |Frobp = C
has density |C|/|Γ|.
Here density refers to either the natural density or the analytic density of the set ofprimes.
Example 1.4. Let f(x) = x2 + 1 ∈ Z[x]. The set of roots of f(x) is i,−i. The splittingfield is K = Q(i) and Gal(Q(i)/Q) = Z/2Z = id, s. In this example, all p 6= 2 areunramified. Then for such an unramified p,
Frobp : i 7→ ip.
Hence,
Frobp =
id if p = 1 mod 4,
s if p = 3 mod 4.
Exercise 1.5. (Mandatory) Check that Frobp is indeed given as above!
Exercise 1.6. (Harder) By considering cyclotomic extensions (i.e. Q(e2πi/m)), show thatChebotarev’s density theorem implies Dirichlet’s theorem on primes in arithmetic progres-sion.
At the beginning of today’s lecture, we discussed patterns in the number of solutions ofa given polynomial modulo p. There is a sheet on the course webpage which shows tables ofthese patterns for the polynomials x2 + 1, x2 − 3, x2 + x + 1, x2 + 2x + 3, and x2 − x − 1.A somewhat mysterious feature of these tables was the modulus appearing in the patterns.(For example, we showed that x2 − 3 has two solutions modulo p 6= 2, 3 if and only if p = 1or 11 mod 12. Where did 12 come from?) We’ll complete today’s lecture with an exampleto demonstrate where this modulus comes from.
11
Example 1.7. Consider the polynomial f(x) = x2−x−1. We can see from the patterns onthe handout that f(x) has 2 solutions mod p if p = 1 or 9 mod 10 and f(x) has 0 solutionsmod p if p = 3 or 7 mod 10 (for p unramified). In this example, the splitting field is K =
Q(φ), where φ = 1+√
52
= 2 cos(π/5) is the golden ratio, and Gal(Q(φ)/Q) = Z/2Z = 〈s〉.Note that φ = ζ + ζ, where ζ = eπi/5 is a fifth root of unity. Hence we can embed
K → Q(e2πi/10) = Q(ζ)
via φ 7→ ζ + ζ. As in the last example, for unramified p,
Frobp : ζ 7→ ζp.
Hence,
Frobp :
φ 7→ φ if p = 1 or 9 mod 10,
φ 7→ s(φ) if p = 3 or 7 mod 10.
In general, the modulus for quadratic fields are determined by embeddings of the splittingfield into cyclotomic fields:
K → Q(e2πi/modulus).
Remark 1.8. Consider the degree 5 polynomial f(x) = x5 + 20x + 16 we discussed inthe first section. In all of the primes that occurred in our table, f(x) had either 0 or 2solutions. However, we would expect that for certain primes, f(x) should have 5 solutionsby Chebotarev’s theorem. Should we be worried that we haven’t seen any 5’s in our chart?Geordie reassures us that we shouldn’t be worried. In this example, Γ = A5 has order 60.Then, by our discussion earlier,
# solutions modulo p = 5 ⇐⇒ all solutions in Fpm are defined over Fp⇐⇒ all solutions are fixed by Frobp
⇐⇒ Frobp = id.
Since id is in its own conjugacy class, we expect f(x) to have 5 solutions modulo p about1/60th of the time by Chebotarev’s density theorem. We included fewer than 60 primes inour table, so we shouldn’t be surprised that we haven’t seen this happen yet.
It may seem like considering the number of solutions of a polynomial over a finite fieldis a cute, but not particularly important problem. However, it is actually of fundamentalimportance in number theory. A number field is a finite extension of Q. All number fields(which are Galois extensions) are splitting fields of polynomials f(x) ∈ Z[x]. One of the themost basic open questions in number theory is the following:
Question 1.9. How many number fields are there?
We can determine the field extension K corresponding to the polynomial f(x) by reducingmod p:
Theorem 1.10. The set p | p unramified and f(x) splits completely mod p completelydetermines K.
So our motivational problem may have been cute, but it certainly isn’t unimportant.
12
Lecture 2: Review of some algebraic number theory
Last time we discussed how by Chebotarev’s density theorem, the equation f(x) = x5 +20x+ 16 should have five solutions modulo p about 1/60th of the time. Joel (+ a computer)computed that in the set of all primes below 500, 000, there are 16, 613 where f(x) has nosolutions, 10, 367 where f(x) has one solution, 13, 885 with two solutions, and 673 with fivesolutions. In this case, we know that the Galois group is A5, so it is order 60, but if wedidn’t know the Galois group, we could use this data to predict its order.
Exercise 2.1. Check the consistency of the numbers above with Chebotarev’s density the-orem.
The goal of today’s lecture is to give the necessary background in algebraic number theoryto continue. It is roughly based on a lecture by Dick Gross [Gro11].
2.1 Number fields
A number field is a finite extension of Q. Given a number field K/Q of degree n (in thislecture, our field extensions will always be degree n), there is an associated ring of integersO ⊂ K consisting of all elements of K which satisfy a monic polynomial with coefficients inZ. The ring of integers O is a free Z-module of rank n, as well as a Dedekind domain (i.e.Noetherian, normal, Krull dimension 1).
Exercise 2.2. Show that the following field extensions have the following rings of integers:
1. K = Q(i), O = Z[i].
2. K = Z(√
2), O = Z[√
2].
3. K = Q(√
5), O = Z[φ], where φ = 1+√
52
.
Generally, for a complicated extension, it is not easy to find O.We can measure how complicated a number field is using something called the discrim-
inant. It is defined as follows. Let K/Q be a number field. Given x ∈ K, we get a Q-linearmap x· : K → K. Using this we define
Tr :K → QNm :K× → Q×
by Tr(x) := Tr(x·), Nm(x) := det(x·). This gives us a bilinear form called the trace form:
K ×K → Q(x, y) := Tr(xy).
The trace form is nondegenerate over Q because Tr(1) = n, hence Tr(xx−1) = n 6= 0.Since any element of O satisfies a monic polynomial with integer coefficients, the trace formrestricts to a map O×O→ Z. Choose a Z-basis α1, . . . , αn for O. Then the discriminantof the field K is
Disc(K) := det((αi, αj)).
This is a close relative of the discriminant of a polynomial.
13
Remark 2.3. We have no idea how many number fields there are, so it is useful to have ameasurement of how complicated a number field is. This is one of the reasons the discriminantis so useful.
Example 2.4. Let K = Q(i). Then O = Z[i] has basis α1, α2 = 1, i. We can compute
((αi, αj)) =
(2 00 −2
),
so Disc(K) = det((αi, αj)) = −4.
Exercise 2.5. 1. Let α ∈ Z be square-free. Let K = Q(√α). Then
O =
Z[√α] if α 6= 1 mod 4,
Z[
1+√α
2
]if α = 1 mod 4.
Hence, calculate
Disc(K) =
4α if α 6= 1 mod 4,
α if α = 1 mod 4.
2. Calculate the discriminant of Q(e2πi/3).
Let K/Q be an etale Q-algebra (i.e. a finite separable extension of Q). Then K ⊗Q Ris an etale R-algebra, so K ⊗Q R ' Rr1 × Cr2 , where n = r1 + 2r2. The field K is totallyreal if r2 = 0 (or, equivalently, if every embedding K → C lands in R). For example, thishappens if K is the splitting field of a polynomial with real roots.
Exercise 2.6. Show that the signature of the trace form on K totally real is (r1 + r2, r2).In particular,
K is totally real ⇐⇒ r2 = 0 ⇐⇒ (·, ·) is positive definite.
2.2 An analogy
Next we will explore a useful analogy which will be a theme of this course.finite extensions
K of C(x) = FracC[x](called function fields)
←
smooth complexprojective curvesC over P1C
↔
compact Riemannsurfaces with a
map to P1C
The first arrow going left is given by taking the function field of the curve. We can also goin the other direction. Given
K O
C(x) C[x]
14
we obtain a map SpecO → SpecC[x] = A1, so we have a unique compactification and anequivalence between the first two sets.
Similarly, there is a bijectionfinite extensionsK of Fp(x)
↔
smooth projectivecurves C over P1
Fp
.
Classically, people worked on problems in number theory in the algebraic world. Artin movedto the geometric world and proved deep results there. Many people in modern number theorywork on the geometric side and hope to prove something about number fields. Here is a (veryrough) schematic of difficulty:
function fields over C << function fields over Fp << number fields
For a very inspiring reference for all of this, see Andre Weil’s letter to his sister SimoneWeil on the role of analogy in mathematics [Kri05].
Exercise 2.7. (Use of Google allowed.) Show that Fermat’s last theorem is true in functionfields; i.e. if f, g, h ∈ k[x] are relatively prime and fn + gn = hn, then n = 2.
2.3 The fundamental exact sequence
Let K/Q be a number field and O the ring of integers of K. A fractional ideal is afinitely generated O-submodule of K. Given two fractional ideals I, J , we can constructtheir product:
IJ :=∑
αiβj|αi ∈ I, βj ∈ J.
(This is the “union” in the sense of algebraic geometry.) Since O is a Dedekind domain,
• every prime ideal p 6= 0 is maximal, and
• every fractional ideal has a unique factorization I =∏
peii , where pi are prime ideals.
Denote by J =⊕
p6=0 prime Zp the group of nonzero fractional ideals under this product. Wehave the following fundamental exact sequence:
1 → O× → K× → J → C`(K)→ 0.
Here C`(K) is the ideal class group of K, which measures the failure of O to be a PID.The ideal class group is difficult to calculate, and we know very little about it in general.The image of the second map in this exact sequence is P , the set of all principal ideals (thatis, ideals of the form xO for some x ∈ K×) of K.
Theorem 2.8. (Fundamental finiteness theorems)
1. The ideal class group C`(K) is finite.
2. The group O× is finitely generated of rank r1 + r2 − 1.
15
Exercise 2.9. Compute O× for Q(√
2) and Q(√
3). (Hint: Pell’s equation)
We can study an analogue of this exact sequence for a smooth projective curve. Let Cbe a compact Riemann surface. Then under the analogy,
0 6= p prime ideals = maximal ideals↔ points of C,
and we have the following exact sequence:
1 → C× → K× → P →⊕x∈C
Zx Pic(C)→ 0.
Here K is the function field of C, P is the set of divisors of meromorphic functions (“principaldivisors”), and Pic(C) is the Picard group of C (isomorphism classes of line bundles on C).
Remark 2.10. The group Pic(C) = Jac(C) × Z is very far from finite. Also, C× is notfinitely generated. So in this setting, neither of the fundamental finiteness theorems hold.
Exercise 2.11. (For those who know some algebraic geometry) Show that the analogues ofC`(K) and O× are finite if C is an affine curve defined over a finite field.
Exercise 2.12. (If you know what K0 is) Show that K0(O) = Z⊕ C`(K).
2.4 Ramification
First consider a smooth projective curve Cf−→ P1C (e.g. y2 = f(x) for some polynomial f(x)
without repeated roots.) After deleting a finite set of points from P1C (the “discriminant”of f), then f is etale, and hence gives us a finite covering of open sets U ⊂ P1C. So all fibresare “the same” away from finitely many points where f is “ramified”. A picture:
Over finite fields, almost the same thing happens. Again, if we have a smooth projective
curve Cf−→ P1
Fp , f is etale after deleting finitely many points. For example, consider C =
16
SpecFp[x, y]/(y2 = x3 + 1)→ SpecFp[x]. Fibres of this map over a point x = λ are singularif λ is a third root of unity, or else
Fp[x, y]/(y2 = x3 +1)⊗Fp[x]Fp = Fp[y]/(y2 = λ3 +1) =
Fp × Fp if λ3 + 1 is a square
Fp2 if λ3 + 1 is not a square.
A picture:
A similar picture holds for number fields:
We can make this picture rigorous. Let K/Q be a number field and O its ring of integers.For each prime p ∈ Z, the corresponding ideal in O decomposes into the product of primeideals in O:
(p) =
gp∏i=1
peii .
We say the primes pi appearing in this decomposition are the “primes above (p)”. Thenumber ei is the ramification index of pi. For each pi, the field O/pi is a finite extensionof Fp (so O/pi = Fpfi for some fi), and the degree fi of this extension is the inertia degree.
Exercise 2.13. (Important!) Show that n =
gp∑i=1
eifi.
17
Definition 2.14. (a) The ideal (p) is unramified if all ei = 1. Otherwise it is ramified.
(b) The ideal (p) splits completely if fi = ei = 1 for all i.
(c) The ideal (p) is inert if gp = 1 and e1 = 1.
Theorem 2.15. The ideal (p) is unramified in O if and only if p - Disc(K).
Exercise 2.16. Prove Theorem 2.15. (Hint: Show that a finite-dimensional commutativeFp-algebra is etale if and only if its trace form is nondegenerate.)
Example 2.17. Let K = Q(i), so O = Z[i] and Disc(K) = −4. (See Exercise 2.4.) Since(1 + i)2 = 1 + 2i− 1 = 2i, we see that the ideal (2) ⊂ O decomposes as (2) = (1 + i)2, and istherefore ramified. By Theorem 2.15, all other primes are unramified. Let p 6= 2 be prime.Then to determine if (p) splits, we notice that
Z[i]/(p) = Z[x]/(x2 + 1, p) = Fp[x]/(x2 + 1) =
Fp × Fp if(−1p
)= 1,
Fp2 if(−1p
)= −1.
Therefore (p) splits completely if and only if p = 1 mod 4. (For example, (5) = (2+i)(2−i)),and (p) is inert if and only if p = 3 mod 4.
We can adapt the strategy of Example 2.17 to determine splitting behavior in general.Let K/Q be a number field, and choose a primitive element (i.e. a generating element)θ of K. We can assume without loss of generality that θ ∈ O. Then θ satisfies a monicpolynomial f(x) ∈ Z[x] of degree n, so Z[θ] ⊂ O is of finite index m. Then for p such thep - m and p - Disc(K),
O/(p) = Z[θ]/(p) = Z[x]/(p, f(x)) = Fp[x]/(f(x)) = Fpf1 × · · · × Fpfk ,
where fi’s are the degrees of irreducible factors of f(x) modulo p. So in particular,
(p) splits completely ⇐⇒ f is reducible modulo p,
(p) is inert ⇐⇒ f is irreducible modulo p.
Hence we could have (and probably should have) phrased last week’s lecture in terms ofsplitting of primes in a ring of integers.
2.5 The case of Galois extensions
Assume K/Q is a Galois extension, with Galois group Gal(K/Q) = G. Then G acts on O,and Frobenius proved the following result.
Theorem 2.18. If (p) =∏
peii , then G acts transitively on the primes pi.
18
Therefore, all ei (resp. fi) are equal. Denote their common value e (resp. f). Exercise2.13 (n =
∑eifi) implies that n = efgp. Fix a prime pi =: p lying over (p). Denote by Gp
the stabilizer of p in G (the “decomposition group”). Then
|G/Gp| = # primes over p = gp,
so |Gp| = ef . Let Ip be the subgroup of Gp which acts trivially on Op (the “inertia group”).This group has order e. We have the following exact sequence
1→ Ip → Gp Aut(O/p) ' Z/fZ.
Since O/p ' Fpf , the group Aut(O/p) is generated by Frobp : x 7→ xp. If p is unramified(i.e. p - Disc(K)), then e = 1, so Ip is trivial and Gp ' Z/fZ. In this case, Frobp ∈ Gp isdefined to be the element that maps to Frobp ∈ Aut(O/p). A different choice of p lying overp results in conjugate Gp and conjugate Frobp. This explains rigorously the Frobp from thelast lecture.
19
Lecture 3: L-functions
Last class we reviewed some algebraic number theory. This class we will review some analyticnumber theory. The motivation for this lecture is the following. We start with an arithmeticproblem (for example, counting the number of x ∈ Fp such that x5 + 20x + 16 = 0), andassign to it an L-function (something like a “character” for the representation theorists inthe audience), which can be studied analytically.
3.1 The Riemann ζ-function
Defineζ(s) =
∑n≥1
n−s,
where s ∈ C is a complex variable. We can compare this sum to the integral∫ ∞1
x−sdx,
which converges to 1s−1
for real s > 1. When viewed as a holomorphic function, the integralconverges absolutely for s ∈ C such that <(s) > 1. Hence the sum ζ(s) converges for alls ∈ C such that <(s) > 1.
This function has a rich history. Euler computed special values (e.g. ζ(2) = π2
2), and
noticed that the ζ-function may also be given as the Euler product:∑n≥1
n−s = (1 + 2−s + (22)−s + · · · )(1 + 3−s + (32)−s + · · · ) · · · =∏
p prime
(1
1− ps
).
This product relates an analytic object, ζ(s), to the prime numbers. This relationship letsus study properties of primes using analysis! For example, the Euler product immediatelygives us two proofs of the infinitude of primes: (1) the divergence of
∑n≥1
1n
implies the
product on the right hand side must be infinite, and (2) since ζ(2) = π2
2, the irrationality of
π2 also implies that the right hand product must be infinite.Riemann showed that ζ admits a meromorphic continuation to all of C. This is the
Riemann zeta function. He also showed that ζ(s) has a simple pole at s = 1 (Exercise:Show that ζ(s)− 1
1−s converges for <(s) > 0), and “trivial zeros” at −2,−4, . . .. Furthermore,
he established the functional equation: Λ(s) := π−s/2Γ( s2)ζ(s) satisfies
Λ(s) = Λ(1− s).
Here Γ(s) =∫∞oxs−1e−xdx is the gamma function. And finally, he proposed the following
conjecture, which eventually became a millenium question.
Riemann hypothesis: If z is a non-trivial zero of ζ(z), then <(z) = 12.
The Riemann hypothesis is known to be true for the first 1012 zeros of ζ(s).
20
3.2 Why do we care?
Here’s the slogan of this story: “The zeros of the Riemann zeta function are the Fouriermodes of the primes”. We will spend the rest of the lecture trying to make this precise.
One of Riemann’s motivations was the following theorem, which was a conjecture duringhis lifetime.
Theorem 3.1. (Prime Number Theorem) Let π(x) be the number of primes less thanor equal to x ∈ R. Then
π(x) ∼ x
log x.
Here ∼ means that limx→∞π(x)x/ log x
= 1. Riemann was interested in two questions about thePrime Number Theorem:
• Why does it hold?
• What is the error term?
Instead of considering π(x) directly, we can examine the von Mangoldt function, which“makes noise at prime powers”:
Λvm(n) =
log p if n = pm,
0 otherwise.
Defineψ(n) =
∑m≤n
Λvm(m).
Exercise 3.2. Show that the Prime Number Theorem is equivalent to ψ(n) ∼ n.
Riemann discovered an explicit formula for ψ(x) at non-integers.
Theorem 3.3. (Riemann) For x ∈ R− Z, there is equality
ψ(x) = x−∑ρ
xρ
ρ− log(2π),
where the sum is taken over all zeros ρ of the Riemann ζ-function.
In other words,
x− ψ(x) = log(2π) +∑ρ
xρ
ρ,
so the zeros of the Riemann ζ-function measure the error term in the prime number theorem.We can examine what effect the different types of zeros have on the right hand side of theequality above.
• Trivial zeros: The function x−2n
−2n= − 1
2nxndecays very quickly, so for large x, trivial
zeros have almost no effect on the formula.
21
• A pair ρ, ρ of non-trivial conjugate zeros: Each such pair contribues
λ · x<(ρ) · cos(γ + log x),
where λ and γ depend in a simple way on ρ and ρ. So <(ρ) is crucial in the contributionof ρ and ρ to the error term, and if the Riemann hypothesis is true, the growth ofthis contribution looks roughly like the product of cos(x) and x1/2. Also, as =(ρ)gets bigger, λ gets smaller. Thus, if the Riemann hypothesis is true, small zeros willcontribute larger variations. A counterexample to the Riemann hypothesis would causeunexpectedly large fluctuations in x− ψ(x).
Exercise 3.4. (a) Show that the Prime Number Theorem is equivalent to ζ(s) having nozeros z with <(z) = 1.
(b) Show that the Riemann hypothesis is equivalent to x− ψ(x) ∈ O(x1/2).
(c) Find the error term in π(x)− xlog x
assuming the Riemann hypothesis.
Remark 3.5. It is unknown whether there exist non-trivial zeros ζ of ζ(s) with <(z) = 1−εfor any ε > 0.
3.3 Dirichlet L-functions
It is natural to ask questions about primes satisfying certain properties. For example, fixm ∈ Z≥0, and a ∈ (Z/mZ)×. Consider the set
p prime |p = a mod m.
Is this set infinite? Is there an analogue to the Prime Number Theorem in this setting? Anaive attempt to show that this set is infinite would be to consider the product∏
p
1
1− p−s
taken over all primes p such that p = a mod m and recreate one of the arguments for theinfinitude of primes given in the previous section. However, there is no Euler product in thissetting, so this approach fails. Another approach is representation theory.
What do we learn from representation theory?
Consider the set of all functions from
(Z/mZ)× → C.
This set has two natural bases:
1. Indicator functions: x 7→ δx,a for a ∈ (Z/mZ)×.
2. Irreducible characters: x 7→ e2πij/φ(m) for j = 0, . . . , φ(m)−1, where φ(m) = |(Z/mZ)| .
22
In many ways, the basis of characters is more natural. Dirichlet borrowed this idea ofcharacters to adapt our naive attempt above into something that works.
Definition 3.6. A Dirichlet character modulo m is a character χ : (Z/mZ)× → C×extended by zero to all of Z. In other words, it is a function χ : Z→ C× such that χ(n) = 0 if(n,m) > 1, χ(n) depends only on n mod m, and χ induces a character χ : (Z/mZ)× → C×.
Such a function χ has the property that χ(nn′) = χ(n)χ(n′). Using these characters,Dirichlet defined a Dirichlet L-function:
L(χ, s) =∑n≥1
χ(n)n−s =∏
p prime
1
1− χ(p)p−s.
With this adjustment by a character, the sum does admit an Euler product, as well as ameromorphic extension to all of C, and a (rather complicated) functional equation. If χ istrivial, we recover the Riemann ζ-function. If χ is not trivial, then L(χ, s) is entire.
Dirichlet’s theorem (that p|p = a mod m is infinite, with distribution 1φ(m)
πlog x
) is an
easy consequence of the fact that L(χ, 1) 6= 0 if χ is not trivial. Furthermore, there is ananalogue of the Riemann hypothesis in this setting, usually called the “generalized Riemannhypothesis,” and all non-trivial zeros of L(χ, s) lie on the critical line (i.e. satisfy <(z) = 1
2).
3.4 Dedekind ζ-functions
Another natural question of this flavor is how primes behave in Z[i], or other rings of integers.To answer this question, Dedekind introduced his version of a ζ-function.
Return to the setting of last week: Let K/Q be a number field with ring of integers O.For a nontrivial fractional ideal I ⊂ O, let NI = #O/I. (For example, if K = Q, N(p) = p.)The Dedekind ζ-function is
ζK(s) =∑
06=I⊂O
(NI)−s =∏
p⊂O prime
1
1− Np−s.
This sum admits an Euler product because of the uniqueness of factorization of ideals inO. Again, we have a meromorphic continuation and functional equation in this setting.(Historical note: The functional equation first appeared in Hecke’s thesis in the 1920’s, butthe proof was very complicated in Hecke’s work. A much simpler proof was given by Tatein his thesis in 1950.) The key example is the following.
23
Example 3.7. Let K = Q(i) ⊃ O = Z[i]. Then
ζK(s) =∏
p⊂O prime
1
1− Np−s
=
(1
1− 2−s
) ∏p s.t. (p) splits
(1
1− Np−s
)(1
1− Np−s
) ∏p s.t. (p) is inert
(1
1− Np−2s
)
=
(1
1− 2−s
) ∏p=1 mod 4
(1
1− p−s
)2 ∏p=3 mod 4
(1
1− p−s
)(1
1 + p−s
)=∏p
(1
1− p−s
)∏p 6=2
(1
1− χ(p)p−s
)= ζ(s)L(χ, s),
where
χ(p) =
1 if p = 1 mod 4,
−1 if p = 3 mod 4
is a Dirichlet character on Z/4Z. The moral of this example is that we can understand Z(i)in terms of Z! We are witnessing the beginnings of class field theory.
3.5 Walking across the bridge
Next we will cross our bridge of analogy and see what happens in the geometric world. Recallthat the objects analogous to a ring of integers O contained in a number field K are smoothprojective curves over Fp and their function fields over Fp.
Example 3.8. The simplest example of such a smooth projective curve is P1Fp . Instead we
will work with A1Fp, where the analogue of a ring of integers is O(A1Fp) = Fp[x]. Here,
ζA1Fp(s) =∑
o 6=I⊂Fp[x]
(NI)−s
=∑
f monic
(N(f))−s
=∑d≥1
( ∑f monic degree d
(pd)−s
)=∑d≥0
pd(pd)−s
=∑d≥0
(p−s+1)d
=1
1− p−s+1.
24
So ζA1Fp
(s) is a rational function with a unique pole at s = 1 and no zeros. Since ζA1Fp
(s)
measures the error term of the prime number theorem, this means that we can count primesexactly in this setting! In fact, we have (Gauss)
# irred. polys of degree d =1
d
∑m|d
µ
(d
m
)pm.
Here µ is the Mobius function (i.e. µ(n) is the sum of the primitive nth roots of unity). Soin this setting, we know exactly how many primes there are in a given interval.
Our example was a little too simple, so we should bump it up a notch. In Artin’s thesis(1923), he instead considered X = Fp[x, y]/(y2 − f(x)) for f(x) square free. (Under ouranalogy, this is the analogue of a quadratic field Q(
√a) for a square free.) Artin showed
that
ζX(s) =(1− αps)(1− αps)
1− p−s+1
is again a rational function, with α ∈ C of norm p1/2. Hence all zeros of ζX(s) have <z = 12,
and the Riemann hypothesis is true in this setting. However, unlike the zeros of ζ(s), thezeros of ζX(s) are distributed evenly along the critical line.
The analogue to this statement for all curves was proven by Weil, and the case of arbitraryvarieties was completed by Deligne (∼1970). These accomplishments were some of thecrowning glories of 20th century mathematics.
3.6 Artin L-functions
Now we enter the non-abelian world. The year is 1927, about 31 years after Frobeniusstarted developing the theory of group characters. Let K/Q be a Galois extension, withG = Gal(K/Q), and dK = Disc(K). Recall from previous lectures that:
• p is unramified if and only if p - dK , for unramified p,
• a choice of prime p over p results in an element Frobp ∈ G, and
• different choices of p lead to conjugate Frobp.
Fix a finite dimensional complex representation
ρ : G→ GL(V )
of the Galois group. Notice that the conjugacy class of ρ(Frobp) is completely determinedby its characteristic polynomial det(1 − tFrobp |V ). The (first approximation) of an ArtinL-function is
Lur(V, s) =∏p-dK
1
det(1− p−s Frobp |V ).
25
Example 3.9. (a) V is the trivial representation. Then
L(V, s) =∏p-dK
1
1− p−s= ζ(s) up to finitely many factors.
So the Artin L-function recovers the Riemann ζ-function as a special case.
(b) Let K = Z(√α)/Q be a quadratic field, G = ±1, and ρ : G → GL1(C) be the
identity map ±1 7→ ±1. Then for p - dK ,
ρ(Frobp) =
1 if p splits
(⇐⇒
(pα
)= 1),
−1 if p is inert(⇐⇒
(pα
)= −1
).
Therefore, by quadratic reciprocity,
Lur(ρ, s) =∏p
(1−
( pα
)p−s)−1
=∏p
(1− χ(p)p−s)−1
= L(χ, s) (up to finitely many factors)
for some Dirichlet character χ : Z/4αZ → C×. So Artin L-functions also recoverDirichlet L-functions.
Exercise 3.10. (a) Show that Lur(V1 ⊕ V2, s) = Lur(V1, s)Lur(V2, s).
(b) Show that L(V, s) converges absolutely for <s > 1. (Hint: Compare it to a product ofdimV copies of the ζ-function.)
(c) (Beautiful! Do it!) For a Galois extension K/Q, let Vreg be the regular representationof G. Show that Lur(Vreg, s) = ζK(s) up to finitely many factors.
A consequence of the exercise is the Artin decomposition:
ζK(s) =∏
V irrep of G
Lur(V, s)dimV (up to finitely many factors).
Hence ζQ(s) always divides ζK(s)! This is a generalization of the factorization ζQ(i)(s) =ζ(s)L(χ, s) that we saw in Example 3.7.
The Moral: All of the difficulty in K is contained in the difficulty of Q if we allow for“non-abelian” difficulty (i.e. general representations of the Galois group).
Remark 3.11. We can get rid of the “up to finitely many factors” caveat in all of these state-ments by modifying Lur(V, s) with “general local factors” at ramified primes. See Geordie’shand-written lecture notes or Gus’s lecture for a description of this procedure.
26
Lecture 4: The Sato-Tate conjecture
This will be our final lecture of global motivation before we zoom in on the local story. Thegoal of today’s lecture is to describe the Sato-Tate conjecture and its relationship to theglobal Langlands picture. We will see that seemingly innocuous statements in the Langlandscorrespondence can have very powerful repercussions.
4.1 Equidistribution in representation theory
We say that the real numbers α1, α2, . . . ∈ [0, 1] are equidistributed if
limn→∞
1
n#αi | αi ∈ (a, b) for i = 1, . . . , n = b− a =
∫ 1
0
1(a,b)dx
for any interval (a, b) ⊂ [0, 1]. Here 1(a,b) is the indicator function on (a, b). Because indicatorfunctions are dense in complex-valued Riemann integrable functions on [0, 1], this conditionis equivalent to saying that the discrete average of any function on this set and continuousaverage of the same function agree; that is,
limn→∞
1
n
n∑i=1
f(αi) =
∫ 1
0
f(x)dx
for all Riemann integrable f : [0, 1]→ C. Now we can approximate any Riemann integrablef : [0, 1]→ C with a Fourier series, so this is in turn equivalent to
limn→∞
1
n
n∑i=1
e(mαi) =
∫ 1
0
e(mx)dx =
1 if m = 0,
0 otherwise
for all m ∈ Z. Here e(x) = exp(2πix). The condition above always holds for m = 0, so tocheck if α1, α2, . . . are equidistributed, it suffices to check that for m 6= 0,
limn→∞
n∑i=1
e(mαi) = 0.
Here is an application of this observation. Let ξ ∈ R be irrational. Consider (ξ), (2ξ), (3ξ), . . .where (mξ) := mξ mod 1. Are these numbers equidistributed? Weyl used the observationabove to show that they are. Choose m 6= 0, and let η = mξ. Then∣∣∣∣∣ 1n
n∑j=1
e(mjξ)
∣∣∣∣∣ =
∣∣∣∣ 1n(e(η) + e(2η) + · · · e(nη))
∣∣∣∣ =
∣∣∣∣ 1n e((n+ 1)η)− e(η)
e(η)− 1
∣∣∣∣ ≤ 1
n
∣∣∣∣ 2
e(η)− 1
∣∣∣∣→ 0
as n→∞.
Remark 4.1. This leads to many questions of a similar flavor (e.g. what about (ξ), (4ξ), (9ξ), . . .?)Weyl’s paper [Wey16] gives an affirmative answer for polynomials f(x) ∈ nR[n] whose coef-ficients are not all rational!
27
We can reinterpret equidistribution via representation theory. The above reasoning shows(after identifying integers mod 1 with S1) that a sequence z1, z2, . . . ∈ S1 ⊂ C equidis-tributes if for every nontrivial rational character χ of S1 (e.g. χ : z 7→ zm for m 6= 0),
limn→∞
1
n
n∑i=1
χ(zi) = 0.
If this condition holds, we say that the sum is “little o of n,” and write
n∑i=1
χ(zi) = o(n).
Now suppose G is a compact group with Haar measure µ, and let X denote the spaceof conjugacy classes of G. (Recall that the Haar measure is the unique left (and thennecessarily also right) G-invariant measure on G with µ(G) = 1.) Let C(X) be the Banachspace of continuous complex-valued functions on X. Two properties of irreducible characterson compact groups are the following:
Theorem 4.2. (The Peter-Weyl Theorem) The irreducible characters span a densesubspace of C(X).
Theorem 4.3. (Orthogonality of characters) If χ, χ′ are irreducible characters, then∫G
χ(g)χ′(g)dµ =
1 if χ = χ′,
0 otherwise.
Hence, we have the following theorem about equidistribution of sequences in G.
Theorem 4.4. A sequence α1, α2, . . . ∈ X is equidistributed with respect to the (pushforward of the) Haar measure if and only if
n∑i=1
χ(αi) = o(n)
for all irreducible nontrivial characters χ.
Example 4.5. Let G = S3. The conjugacy classes are determined by cycle type: C1 = id,C2 = (12), (23), (13), C3 = (123), (132). The character table of S3 is the following.
Haar measure #Ci conj. class triv nat sgn1/6 1 C1 1 2 11/2 3 C2 1 0 -11/3 2 C3 1 -1 1
We can see from this table that for a sequence to be equidistributed, it should spend twicethe time in C3 as it does in C1 (as dictated by the column corresponding to the naturalrepresentation), and should spend as much time in C2 as in C1 ∪ C3 (as dictated by thecolumn corresponding to the sign representation).
28
Example 4.6. Let G = SU(2). Since all matrices in SU(2) are diagonalizable, conjugacyclasses are determined by eigenvalues, which come in conjugate pairs. SoX = (γ, γ)|γ ∈ S1can be identified with S1
+ = z ∈ S1|<(z) ≥ 0 ' [0, π]. By the Weyl character formula forSU(2), irreducible characters are of the form
χm(z) = zm + zm−2 + · · ·+ z−m,
where z = eiθ ∈ S1. (Here we are using the identification X ' S1+ when defining these
characters.) The Haar measure on the space of conjugacy classes (after identifying S1 with[0, 2π] via the exponential function) is then
2
πsin2 θdθ.
Exercise 4.7. Check that this is correct by showing that
2
π
∫ π
0
sin2 θdθ = 1
and ∫ π
0
χm(eiθ) sin2 θdθ = 0
for m 6= 0.
The figure below is a plot of 2π
sin2 θ. From this we see the distribution of eigenvaluesof matrices in SU(2). A first observation is that there are many matrices in SU(2) witheigenvalues i,−i (corresponding to θ = π
2), and very few matrices with eigenvalues 1, 1
and −1,−1 (corresponding to θ = 0 and θ = π, respectively). In fact, there is exactly onematrix in each case: I and −I.
Remark 4.8. This demonstrates that it is more likely for matrices in SU(2) to have eigen-values which are “far away” (meaning that the angle between them in the complex plane islarge), an important fact in random matrix theory.
Hence if you had a sequence of pairs of complex numbers which you suspect are eigenvaluesof random matrices in SU(2), you could tell pretty quickly whether or not it was plausible.
29
4.2 Elliptic curves and the Sato-Tate conjecture
Next we discuss elliptic curves. (This seems to be completely unrelated, but we should havefaith that it will come full circle.) Let k be a field, and E an elliptic curve. (That is, asmooth projective curve over k of genus 1 with a fixed rational point 0 ∈ E(k).) Assumethe characteristic of k is not 2 or 3. Then E can be made to be of the form (the projectiveclosure of)
y2 = x3 + ax+ b, with 0 = (0 : 0 : 1) its point at infinity.
Here smoothness translates into the fact that x3 + ax + b has no repeated roots, i.e. that4a3 − 27b2 6= 0.
First assume we are working over C. Then E is a compact Riemann surface of genusg = 1, so E = C/Λ for some lattice Λ ' Z2 → C. Hence
elliptic curvesover C
/ iso
'
latticesΛ ⊂ C
/ iso
“period ratio”−−−−−−−−→ SL2(Z)\H,
where H is the upper half-plane. The quotient SL2(Z)\H can be identified (up to someambiguity on the boundary) with its fundamental domain
so we can consider all complex elliptic curves as points in the region above.Certain elliptic curves have extra structure called complex multiplication. Let E be a
complex elliptic curve. As a real Lie group, E is isomorphic to S1 × S1, hence
EndLie gp(E) = Z2
(z, w) 7→ (zm, wn)↔ (m,n)
But E has additional structure - it is an elliptic curve (E ' C/Λ), and a complex algebraicgroup. By a miracle of abelian varieties, the elliptic curve endomorphisms of E fixing 0 arethe same as the complex algebraic group endomorphisms of E. Hence,
End07→0(E) = Endalg gp(E) = z | zΛ ⊂ Λ =
Z if Λ · Λ 6⊂ Λ,
Λ if Λ · Λ ⊂ Λ.
In the second case, we say E has complex multiplication. In the fundamental domain,the elliptic curves with complex multiplication are of the form i
√d (the “corner point” in
the picture above):
30
Thus one can think of curves with complex multiplication as being very special.
Exercise 4.9. Show that if Λ · Λ ⊂ Λ (i.e. E has complex multiplication), Λ ⊗ Q is animaginary quadratic field. In particular, if E has complex multiplication then Λ is what iscalled an order in an imaginary quadratic field.
Next we work over Q, and consider the elliptic curve E given by y2 = x3 + ax + b fora, b ∈ Z. To understand E, we reduce modulo p and study the number of points of E(Fp).If EFp = E × SpecFp is nonsingular, then p is a prime of good reduction. (This is theanalogue for algebraic varieties of a prime being unramified.) We can bound #E(Fp) by theHasse-Weil bound:
Theorem 4.10. (Hasse-Weil)
#E(Fp) = 1 + p− αp,
where |αp| ≤ 2√p.
The Hasse-Weil bound tells us that 1 + p is a “square root accurate” approximation for#E(Fp). By our discussions last week, hopefully you are convinced that this is an analogueof the Riemann hypothesis for elliptic curves.
Big question: How do the αp’s behave?
We can start to answer this question using the Grothendieck–Lefshetz trace formula.Let H∗(E) be the etale cohomology of E. For all p outside a finite set, we have an action ofFrobp on (etale, but don’t worry if you don’t know what this is) cohomology:
dim : 1 2 1H0(E) H1(E) H2(E)
Frobp = 1 Frobp ∼(γp 00 γp
)Frobp = p
31
The Grothendieck-Lefshetz trace formula is
#E(Fp) =∑
(−1)i Tr(Frobp : H i) = 1 + p− (γp + γp),
where γp, γp are the eigenvalues of Frobenius on H1(E). Hence to determine the number ofsolutions of E(Fp), it is enough to examine γp, and it is true (but not so easy to see) thatthe Riemann hypothesis for E is equivalent to |γp| = 1
2.
We can renormalize so that
θp :=1√pγp ∈ S1
+.
This leaves us with a sequence θ2, θ3, θ5, . . . of points on a semicircle S1+ controlling the
number of points of E modulo p.
Sato–Tate conjecture (1960): Suppose E does not have complex multiplication. If we
identify S1+'−→ [0, π], then
limn→∞
1
π(n)
∑p≤n
µθp =2
πsin2 θdθ.
Here µθp is the Dirac distribution.
In other words, the Sato–Tate conjecture is that the θp’s look like eigenvalues of randommatrices in SU(2)! Below is a very beautiful illustration of this phenomenon. For the ellipticcurve y2 + y = x3 + x + 3x + 5 (which has no complex multiplication), the following is theplot of θp for the first 5, 000 primes. (The further out the dot, the bigger the prime.)
32
In contrast to this, consider the elliptic curve y2 = x3 + 1. Here complex multiplication isgiven by the Eisenstein integers. The plot below shows θp for the first 5, 000 primes. Normis linear in the prime (so, again, the further the dot, the bigger the prime).
The difference between these two curves is striking!
Remark 4.11. The case of complex multiplication is well understood. The idea (Geordiethinks) is that the extra endomorphisms force the Frobp to lie in a subgroup of SU(2).(Geordie points out that he is “the exact opposite of an expert on this topic...”)
Remark 4.12. One can think of Sato-Tate (roughly) as a higher-dimensional analogue ofChebotarev density: Gal(K/Q)↔ SU(2).
33
4.3 Equidistribution and L-functions
In the final part of this lecture, we will describe how the Sato–Tate conjecture follows froma simple part of the Langlands correspondence (which is still conjectural). The Sato–Tateconjecture has been proven using other methods, but the proof is very involved and heavilyinfluenced by ideas from the Langlands program. This example is meant to showcase thepower of the Langlands correspondence.
Let G be a compact group, X its space of conjugacy classes, and xp ∈ X a family ofelements parameterized by primes p (perhaps outside some finite set of “bad” primes). Foran irreducible character χ, we can define an L-function analogously to how we defined ArtinL-functions for Galois groups:
L(χ, s) =∏p
1
det(1− ρ(xp)p−s),
where ρ is the representation afforded by χ. This converges for <(s) > 1. Assume addition-ally that L(χ, s) extends to a meromorphic function on <(s) ≥ 1 having neither zeros norpoles along <(s) = 1 except possibly at s = 1. Let −cχ be the order of L(χ, s) at s = 1 (socχ > 0 pole, cχ < 0 zero). With these assumptions, we have the following theorem.
Theorem 4.13. ∑p≤n
χ(xp) = cχ ·n
log(n)+ o
(n
log(n)
).
The proof of this theorem involves some complex analysis and tricks with sums, but isnot difficult. This theorem has an important corollary.
Corollary 4.14. If for all nontrivial χ, L(χ, s) is holomorphic and nonzero at s = 1, thenthe xp are equidistributed in X.
So we can use L-functions to test whether a sequence in a compact group is equidis-tributed!
Exercise 4.15. (a) Show that L(χ, 1) 6= 0 implies Dirichlet’s theorem.
(b) Let K be a number field. It’s known that ζK(s)/ζ(s) is holomorphic and nonvanishingat s = 1. Using this, deduce Chebotarev’s density theorem.
An important consequence of Corollary 4.14 is that it lets us reframe the Sato-Tateconjecture in terms of L-functions.
Example 4.16. (Serre) Assume that for all m ≥ 1, the symmetric L-power function
L(Smχ , s) :=∏p
1
(1− θmp p−s)(1− θm−2p p−s) · · · (1− θ−mp p−s)
satisfies the extra assumption above. (Here Smχ is the mth symmetric power of the repre-
sentation with character χ sending Frobp 7→(θp 00 θ−1
p
).) Then the Sato-Tate conjecture
holds!
34
Recall our cartoon of the Langlands correspondence:“geometric” n-dimensional
rep’ns of Gal(K) = Gal(K/K)
finite-to-one←−−−−−−→
“automorphic” rep’ns
of GLn(A)
L-functions
An extremely important part of this picture is functoriality. That is, the diagramshould be compatible with
1. composition of representations of the Galois group with algebraic representations of
GLn (so ρ : Gal(K) → GLnalg rep’n−−−−−→ GLm should correspond to some operation on
automorphic representations), and
2. changing the field K.
A key piece of the Langlands correspondence is that L-functions coming from automorphicrepresentations have many desirable properties, which are extremely difficult to establish forL-functions coming from geometric representations of Gal(K). For example, once we knowthat an L-function comes from an automorphic representation, we immediately know thatit admits a meromorphic continuation and has a functional equation.
The Punchline: In the simple example of the algebraic representation GL2 → GLm viasymmetric powers, the prediction of functoriality implies the Sato-Tate conjecture!
35
Lecture 5: Infinite Galois theory and global class field
theory
The topic of today’s lecture is class field theory. But before diving in, we will start witha motivating question and a review of infinite Galois theory.
Basic Question: What are all finite extensions of a number field K (e.g. Q)?
This is certainly a question of fundamental importance in number theory. One could alsoask more specific questions, such as “How many number fields have a given Galois group anddiscriminant?” For almost every question of this sort, we have no idea what the answer is.Class field theory develops techniques for answering these questions in the abelian setting.We will say more precisely what this means later in the lecture, but for now, let’s take alook at an example.
Example 5.1. Let K = Fp, and K its algebraic closure. For all n ≥ 1, there is a uniquesubfield Kn ⊂ K with pn elements, and K =
⋃n≥1Kn. We have the following picture of
field extensions and corresponding Galois groups:
Let’s calculate Gal(K/K). Because K =⋃n≥1Kn, we have an injection
Gal(K/K) →∏n≥1
Gal(Kn/K) =∏n≥1
Z/nZ.
For ϕ ∈ Gal(K/K), let the sequence (ϕn) be its image in∏
n≥1 Gal(Kn/K). A sequence(γn) ∈
∏n≥1 Gal(Kn/K) is in the image if and only if γn = γm mod m whenever m|n.
Hence,Gal(K/K) = lim
←−Z/nZ = Z
is the “profinite completion of Z”.
Exercise 5.2. (a) Show that Z =∏
p prime Zp.
(b) Show that Z has uncountably many subgroups. Hence a naive Galois correspondencecannot hold.
36
5.1 Infinite Galois theory
Let L/K be a Galois extension (algebraic, normal, separable, not necessarily finite). Thenwe have an injection
Gal(L/K) →∏
K⊂L′⊂L
Gal(L′/K),
where the product is taken over all towers of field extensionsK ⊂ L′ ⊂ L, where the extensionL′/K is finite Galois. For all towers of extensions K ⊂ L′ ⊂ L′′ ⊂ L, where the extensionsL′/K and L′′/K are finite Galois, there is a corresponding map Gal(L′′/K) → Gal(L′/K).This determines the image of this injection; that is,
Gal(L/K) = lim←−K⊂L′⊂L
Gal(L′/K).
This is a topological group. Indeed, if we give∏
K⊂L′⊂L Gal(L′/K) the product topology(which is compact, by Tychonov), then Gal(L/K) inherits the subspace topology.
Exercise 5.3. The group Gal(L/K) is closed (therefore compact) in∏
K⊂L′⊂L Gal(L′/K).
Example 5.4. We see from Exercise 5.3 and Example 5.1 that the group Z is compact,which might look strange.
Exercise 5.5. (Important, can be used as a definition) A basis of open neighborhoods of1 ∈ Gal(L/K) is given by kernels of the maps
Gal(L/K)→ Gal(L′/L)
for L′/K finite Galois.
For a general group G, define
G = limH⊆Gnormal
finite index
G/H.
The group G is profinite if G'−→ G, otherwise, say G is the profinite completion of G.
So Galois groups are profinite groups! The key example to keep in mind is the following.
Example 5.6. Consider the group Zp = lim←−
Z/pnZ. What does this group look like? Here’s
a picture for p = 3:
37
Exercise 5.7. Here are some fun thought experiments.
(a) Where is −1 in this picture?
(b) Think about Q3.
The motto: Galois groups are fractal-like objects!
Theorem 5.8. (Fundamental theorem of infinite Galois theory) Let L/K be a Galois ex-tension. Then there exists a canonical bijection
K ⊂ L′ ⊂ L ↔
closed subgroupsof Gal(L/K)
LH ←[ HL′ 7→ Gal(L/L′)
Moreover, under this bijection,
finite extensions ↔ closed and open subgroups
Galois extensions ↔ normal subgroups
Exercise 5.9. Show that the only closed subgroups of Z are nZ for n ∈ Z. If n ≥ 1, thenthe subgroup nZ corresponds to the extension Kn under the bijection above, and n = 0corresponds to K. (So 0Z is the only closed subgroup which isn’t open.)
Now we return to the problem posed at the start of this lecture:
describe all number fields over Q
or, equivalently,describe all closed subgroups of Gal(Q/Q).
However, after some thought, one sees that this isn’t really a well-defined question (froma philosophical point of view), because Q involves a choice (or many choices), so thereis no concrete canonical realisation of Q. Hence Gal(Q/Q) is only really a “group up toconjugacy,” in the sense that any meaningful statements one can make about this groupmust be invariant under conjugation. (One cannot talk about individual elements.)
The Punchline:
1. Isomorphism classes of representations of “a group up to conjugacy” are canonical, soit makes sense to talk about representations of Gal(Q/Q). This is one reason whyrepresentations are so important in the Langlands program!
2. The maximal abelian extension Qab of Q (which is the extension corresponding to
[Gal(Q/Q),Gal(Q/Q)]) is canonical, and we can hope to describe it by studying themaximal abelian quotient Gab := G/[G,G] of G = Gal(Q/Q). This is class field theory!
38
5.2 Global class field theory, first version
The key example to keep in mind is the maximal abelian extension of Q.
Example 5.10. Let Km := Q(ζm) for ζm = e2πi/m. Define Q(µ∞) :=⋃m≥1Km. The Galois
group of Km/Q is (Z/mZ)×, hence
Gal(Q(µ∞)/Q) = lim←−
(Z/mZ)× =∏p
Z×p .
Fact: (“Kronecker Jugendtraum”) Q(µ∞) is the maximal abelian extension of Q.
This fact is not easy! (It will follow from global class field theory.) The hope of Kroneckerwas to predict this starting just from Q, without calculating extensions.
Exercise 5.11. Use Kronecker Jugentraum to show that any continuous character χ :Gal(Q/Q) → C× “is” a Dirichlet character. (Part of the exercise is to work out what“is” means in this context.)
Given a number field K, it is useful to consider all norms at once. Let O ⊂ K be thering of integers. A place v is an equivalence class of nontrivial multiplicative norms
| · |v : K → R≥0
on K.
Theorem 5.12. All places of a number field K are of the following form.
• Finite places: |x|v := (#O/p)−valp(x) for p ⊂ O prime.
• Real places: |x|v := |i(x)| for some real embedding i : K → R,
• Complex places: |x|v := |i(x)|2 for some pair of conjugate i : K → C not landing inR.
These are all possible notions of distance on a number field.
Exercise 5.13. Show that there are no nontrivial norms on a finite field.
Note that we could have chosen any scalar λ > 1 in place of #O/p above. The reasonfor the the above normalization is the beautiful product formula: For x ∈ K×,∏
places v
|x|v = 1.
Note that this product makes sense because all but finitely many places are 1. Thefunction field case of this formula is the statement that the number of poles and number ofzeros (with multiplicity) agree.
39
5.3 Global class field theory a la Artin
Fix a finite abelian Galois extension L/K with abelian Galois group Gal(L/K). Let OL ⊂ Land OK ⊂ K be the rings of integers, and let Sf ⊂ OK be the set of ramified primes. Wehave seen that for a prime p ⊂ OK such that p 6∈ Sf , there is a corresponding conjugacyclass Frobp ∈ Gal(L/K). In general Frobp is only defined up to conjugacy, but since we areassuming that Gal(L/K) is abelian, Frobp is a single element. Hence we obtain a map (theArtin map):
J Sf =⊕p6∈Sf
Zp→ Gal(L/K)
∑mpp 7→
∏Frobp
mp
Here J Sf ⊂ J is a subgroup of the group of nonzero fractional ideals J =⊕
primes p Zpdiscussed in Lecture 2. By Chebotarev’s density theorem, the Artin map is surjective.
Question: What is the kernel of the Artin map?
The answer to this question is related to an observation we made in the very first lecture.Recall our motivating problem for the course of determining the number of solutions of apolynomial modulo p for various primes p. For quadratic polynomials, we used quadraticreciprocity to find some modulus m ∈ Z such that the number of solutions of the polynomialmodulo p was given by the residue of p modulo m. At first, the modulus m seemed to besomewhat mysterious, but eventually we observed that it was obtained from the ramifiedprimes (that is, the “weird primes” which we ignored). For example, for the polynomialx2 + 1, which has 2 solutions modulo p if p = 1 mod 4 and 0 solutions if p = 3 mod 4, themodulus 4 is the square of 2, our only ramified prime.
Returning to the setting of the Artin map, we define a modulus m supported in a setof places S to be a formal Z-linear combination of places m =
∑mivi such that mi ∈ 0, 1
for real places and mi = 0 for all complex places and places vi 6∈ S. Given a modulus msupported in a set of places S, we can define an associated group (the Ray class group) asfollows. Consider the following two subsets of K×:
KS = λ ∈ K× | valp(λ) = 0 for all p ∈ S, and
Km,1 = λ ∈ Ks | valp(λ− 1) ≥ mi for finite places, and i(λ) ∈ R×>0 for real places mi = 1.The set Km,1 is the set of λ which are “m close to 1”. We have the following maps:
K× J C`K
principal ideals
Ks J S
Km,1
val
val
val
40
The Ray class group associated to the modulus m is the quotient
C`mK := J S/val(Km,1).
Example 5.14. If m = 0, then C`0K = C`K is the class group.
Example 5.15. Let K = Q, and m = n(p) for a prime p ∈ Z. Then m is a modulussupported on S = (p). We have
KS =ab| a, b are coprime to p
= Z×(p), and
Km,1 =ab| a, b coprime to p and valp(
a
b− 1) ≥ n
=ab| ab
= 1 mod pn.
HenceKs/Km,1 = Z×(p)/K
m,1 = (Z/pnZ)× = C`mQ .Theorem 5.16. For any modulus m, the Ray class group is finite and surjects onto theclass group of K.
Theorem 5.17. (Artin) For any L/K as above, there exists a modulus m with supp(m) ∩finite places = Sf such that val(Km,1) is contained in the kernel of the Artin map. More-over, for any modulus m and any quotient q : C`mK Q, there exists an abelian extensionL/K ramified only at primes in supp(m) such that
C`mK Gal(L/K)
Q
Artin
commutes.
A weaker form of this theorem provides a more direct answer to our question from thebeginning of this section.
Theorem 5.18. For any abelian Galois extension L/K, there exists an ε > 0 such that ifλ ∈ KSf is ε close to 1 for all places v ∈ S, then val(λ) is in the kernel of the Artin map.
Example 5.19. Consider the extension Q(i)/Q, which is the splitting field of the polynomialx2 + 1. The ramified primes are 2 and ∞ (but we haven’t discussed what it means for ∞to be ramified, so we are sweeping this under the rug), so Sf = (2). For any p 6= 2,Frobp(i) = ip, hence in the Galois group, Frobp ↔ p mod 4 ∈ (Z/4Z)× = Gal(Q(i)/Q). Wehave
J Sf =ab| a, b > 0 coprime to 2
= Z×(2),>0.
Hence the Artin map is the natural map
Z×(2),>0 → (Z/4Z)× .
Its kernel is the set of all elements satisfying a “congruence condition at 2”:λ ∈ Z×(2),>0 | val2(λ− 1) ≥ 2
.
We see from this example that the finitely many “weird” (ramified) primes from the firstlecture are the primes determining our congruences.
41
Example 5.20. If m = 0, then C`0K = C`K . Hence the existence statement in Theorem 5.17
implies that there exists an unramified everywhere extension L/K with Gal(L/K) = C`K .This extension is the Hilbert class field.
42
Lecture 6: Local Galois groups and local class field the-
ory
6.1 A potted history of class field theory
Sometimes if we find something difficult, it can be comforting to know that other people alsofound it difficult. Accordingly, we’ll start today’s lecture with a potted history of class fieldtheory, following [Con01, Miy11].
• Kronecker, Weber (1850-1880): Kronecker’s Jugentraum (that Q(µ∞) is the max-imal abelian extension of Q), explicit class field theory (describing Kab explicitly, notjust its Galois group) for Q and Q(i), relevance of complex multiplication.
• Dedekind, Frobenius (1880): defined Frobp (then everyone promptly forgot for 40years).
• Hilbert (189-1900): first correct proof of Jugentraum for Q, emphasis on “places at∞,” introduction of Hilbert class field, 12th problem on Hilbert’s list was explicit classfield theory for any K. (Still open! Even for Q(
√d), d ≥ 0!)
• Hensel (1897): introduction of p-adic numbers, took a while to catch on.
• Takagi (1900): PhD student of Hilbert in Gottingen, thesis on Q(i)ab, proof of exis-tence theorem during WWI (when there was no contact between Germany and Japan),result announced at the ICM in 1920.
• Hasse (1922): local global principle, first time p-adics were taken seriously by thebroader mathematics community.
• Chebotarev (1927): density theorem.
• E. Artin (1927): introduction of Artin map (the return of Frobp!), reciprocity theo-rem.
• Schmidt (1930): deduced local class field theory from global class field theory (proofsstill analytic).
• Noether (1930s): local theory should be simpler and come first.
• Chevalley (1940): algebraic proof of local class field theory.
6.2 A trip across the bridge
Recall that the “bridge” in this course is the motivating analogy between number fields andfunction fields.
Remark 6.1. This bridge was what motivated Hensel’s advocation for the introduction ofthe p-adic numbers.
43
Let’s look at local class field theory through this analogy. Let L/K be a finite Galoisextension. Across the bridge, this should correspond to a surjective map (i.e. ramifiedcover) f : X → Y of algebraic curves/Riemann surfaces over C. Recall that our R-pictureand C-picture of such a map are the following:
For all y ∈ Y outside of a finite set, f is etale; that is, a smooth n : 1 cover in a neighborhoodof y:
For a finite set of points y ∈ Y , f is not smooth, and locally sends z 7→ zni , such that∑ni = n:
44
If f is Galois, then all ni are equal.
The Moral: The ramified cover f is determined by simple data (“local monodromies”) atfinitely many points (“ramified primes of Y ”).
Remark 6.2. We have
The algebraic closure of C((t)) is C((tQ)) :=⋃n≥1
C((t1/n)). Hence,
Gal(C((t))/C((t))
)= lim←
Z/nZ = Z.
This is “why” the local field information is so simple in function field/C case1. In languageto come, every extension of C((t)) is “tamely ramified”.
The upshot is that on the function field side of the bridge, local information is easy, andcan be patched together to form the global picture. On the number field side of the bridge,local information is much harder, but the philosophy we learn from our analogy is that itshould still be easier than global information, so we should focus on it first. Because of this,essentially the rest of this course will be local.
Now we return to number fields. Let L/K be a finite Galois extension. Fix a place v ofK. If v is finite (corresponding to some p ⊂ Ok), then we know what it means for a placev′ (corresponding to q ⊂ OL) of L to “lie over v”: v′ lies over v precisely when q is a primeabove p in the sense of lecture 2 (i.e. pOL ⊂ q).
If v is a real or complex place corresponding to i : K → K ∈ R,C, then a place v′ ofL lies over v if the corresponding injection i′ fits into a commutative diagram.
L K′
K K
i′
i
1Recall that we saw last lecture that Gal(Fp/Fp) = Z as well. This turns out to be a useful coincidence,but we won’t comment on it further here.
45
Hence if v is real, then v′ is either real or complex. A real place v is ramified if there existsa v′ lying above v that is complex in L, and unramified if all v′ above v are real. If v iscomplex, then all v′ above v are complex, and we say the place v is unramified.
Fix a place v of K and a place v′ of L over v. Let Lv′ (resp. Kv) be the completion of L(resp. K) with respect to the place v′ (resp. v). Then we have the diagram
L Lv′
K Kv.
Set
Gv = σ ∈ Gal(L/K) | σ acts continuously on Lv
= σ | σ preserves v′ =
Gq if v is finite,
1 if v is unramified infinite,
Z/2Z if v is ramified infinite.
Here Gq ⊂ Gal(L/K) is the decomposition group corresponding to the prime q ⊂ OL
determining v′.
Remark 6.3. In the last case (when v is ramified infinite), we get a canonical elementc ∈ Gal(L/K) corresponding to complex conjugation.
The point: Lv′/Kv is a finite Galois extension of local fields with Galois group Gv. We willfirst try to understand such extensions for all places v, then piece together this informationto understand L/K.
6.3 Local class field theory
Between the “easy” world of finite fields and the complicated world of global fields lies theworld of local fields. Let K be a field equipped with discrete valuation val : K → Z ∪ ∞.In K lies its ring of integers OK , with maximal idea m generated by a “uniformizer” π ∈ m:
K ⊃ OK = val−1(Z≥0 ∪ ∞) ⊃ m = val−1(Z>0 ∪ ∞) = (π).
The field kK = OK/m is the residue field. Note that in this setup, K, OK and m are allcanonical, but the uniformizer π ∈ m is not. For us, a local field will be a field K equippedwith a discrete valuation as above such that
1. K is complete with respect to val (i.e. K has the topology coming from OK = lim←
OK/mnK),
and
2. kK is finite.
Exercise 6.4. Show that 1. and 2. are equivalent to K being locally compact.
46
Example 6.5. The field Qp is locally compact because it is covered by dilates of Zp:
Qp =⋃n≥1
p−nZp.
Recall that Zp are compact open sets in Qp.
Remark 6.6. In some terminology, R and C are also referred to as local fields.
Let L/K be a finite Galois extension of local fields. Then any element σ ∈ Gal(L/K)preserves OL/OK and mL/mK , and thus acts on kL/kK . Hence we get maps
1→ IL/k → Gal(L/K) Gal(kL/kK)→ 1,
where IL/K is the inertia subgroup of lecture 2. The Galois group Gal(kL/kK) ' Z/dZ isgenerated by Frobq. For L = K, this short exact sequence becomes
1→ IK/K → Gal(K/K) Gal(Fq/Fq) ' Z→ 1.
Local class field theory: There exists a canonical map2 rK : K× → Gal(K/K)ab with
dense image such that rK induces an isomorphism K×∼−→ Gal(K/K)ab. Here K× is the
profinite completion of K× (that is, the completion with respect to subgroups of finiteindex). Moreover,
1. the diagram
1 IabK/K
Gal(K/K)ab Gal(kK/kK) ' Z
1 O×K K× Zval
rK
commutes, and
2. if L/K is finite Galois, then
L× Gal(L/L)ab
K× Gal(L/K)ab
rL
NormL/K res
rK
commutes.
Remark 6.7. The analogous statements for the fields R and C((t)) are the following:
2The canonical map rK is sometimes called the “reciprocity map”.
47
1. K = R: In the diagram
C× Gal(C/C) = 1
R× Gal(C/R)
rC
NormC/R
rR
the map rR : −1 7→ complex conjugation is continuous and surjective. The kernel ofrR is R×>0, the set of norms coming from C× (i.e. the image of NormC/R).
2. K = C((t)):3 The map
rC((t)) : C((t))× → Gal(C((t))/C((t))
)= Z
has dense image, so a reasonable choice is valuation val : C((t))× → Z.
It is useful to modify the Galois group Gal(K/K) slightly. Define the Weil group of
K to be the subgroup WK ⊂ Gal(K/K) of elements whose projection onto Z is an integralpower of Frobq; that is, WK fits into the short exact sequence
IK/K WK Z
IK/K Gal(K/K) Z.
The purpose for this modification of the following fact: the reciprocity map rK provides anisomorphism between K× and the abelianization of the Weil group:
rK : K×'−→ W ab
K .
With this, we can state the local Langlands correspondence for GLn(K).
Theorem 6.8. (Local Langlands correspondence for GLn(K)) (Harris-Taylor) Thereis a bijection
Homcts(WK ,GLn(C))/conj1:1←→
irreps of GLn(K)in C-vector spaces
.
The continuous group homomorphisms on the left hand side of this bijection are referredto as the Langlands parameters of the corresponding GLn(K)-representations on theright.
Remark 6.9. Actually, this is not quite correct. Instead we should consider Weil-Delignereps on the left, and smooth admissible reps on the right. These issues will be addressed incoming lectures.
3Note that K does not quite fit our assumptions so LCFT does not apply, but morally it fits into thesame picture.
48
Example 6.10. The n = 1 case of this theorem is true by local class field theory:
Homcts(WK ,GL1(C))/conj = Homcts(WK ,C×)
= Homcts(WabK ,C×)
= Homcts(K×,C×)
= irreps of GL1(K).
Example 6.11. We can see explicitly that local class field theory is true for Qp. By a localversion of the Jugentraum,
Qabp = Qp(µ∞) =
⋃Qp(ζn),
where ζn is an nth root of unity. Hence
Qabp = Qp(µp′) ·Qp(µp∞),
where Qp(µp′) :=⋃p-n
Qp(ζn) and Qp(µp∞) :=⋃n≥1
Qp(ζpn). As before, Gal(Qp(ζpn)/Qp) =
(Z/pnZ)× , soGal(Qp(µp∞)/Qp) = Z×p .
If p - n, note that Qp(ζn)/Qp is unramified, so Q(µp′) =: Qurp is the maximal unramified
extension of Qp. Because Fp =⋃p-n
Fp(ζn), we have that
Gal(Qp(µp′)/Qp) = Gal(Fp/Fp) = Z.
We conclude thatGal(Qab
p /Qp) = Z× Z×p ' Q×p ,
which is exactly what is predicted by local class field theory.
6.4 Structure of Galois groups of local fields
We will finish this lecture with some remarks on the structure of local Galois groups. Thissection is hands-on and explicit. Morally, it should come before local class field theory.
Let L/K be a finite Galois extension of local fields, and L ⊃ OL ⊃ mL = (πL), K ⊃OK ⊃ mK = (πK) the respective rings of integers, maximal ideals, and uniformizers. Asbefore, there is a corresponding extension of residue fields kL/kK . We have a short exactsequence
1→ IL/K → Gal(L/K) Gal(kL/kK) ' Z/nZ→ 1,
where IL/K = σ | σ acts trivially on kL, and Gal(kL/kK) is generated by the canonicalgenerator Frob. Note that Gal(L/K) preserves OK , hence mK , hence the valuation vK ,hence acts on OK/m
jK , hence acts continuously on L.
Lemma 6.12. (Key lemma) Any σ ∈ IK/L is determined by its action on πL.
49
Exercise 6.13. Prove Lemma 6.12. (Hint: use the fact that any σ ∈ Gal(L/K) is automat-ically continuous.)
Set I := IL/K , I0 := I, and Ij := σ ∈ I | σ(π)π−1 ∈ 1 + mjL for j ≥ 1.
Proposition 6.14. This defines a filtration
I = I0 ⊃ I1 ⊃ I2 ⊃ · · ·
of I by normal subgroups. Moreover,
1. This is a finite filtration; i.e. Im = 1 for large enough m.
2. We have natural injections
I0/I1
σ(π)π−1
−−−−→ k×L ,
Ij/Ij+1 −−−−→ (1 + mjL)/(1 + mj+1
L ) ' kL
for j ≥ 1. In particular, I is solvable, I0/I1 is of order prime to p, and I1 is the Sylowp subgroup of I.
This filtration is called the ramification filtration of I. The proof is an easy exercise.
Definition 6.15. We say that L/K is tamely ramified if I1 = 1, and L/K is unrami-fied4 if I0 = 1.
Remark 6.16. This agrees with our earlier notion of unramified from Lecture 2.
4Note that a quirk of this terminology is that unramified is tamely ramified. It is strange, but we willjust have to get used to it.
50
Lecture 7: Representation theory of p-adic groups
7.1 Tying up some loose ends
We start today’s lecture by tying up some loose ends from previous weeks. Recall that twolectures ago, we discussed how looking for a description of Gal(Q/Q) is misguided becauseGal(Q/Q) is only a “group up to conjugacy,” since its definition requires a choice of Q.There is an analogy for this idea that Geordie learned from Kevin Buzzard which might bemore familiar to us. Let X be a path-connected space. A choice of base point x ∈ X yieldsthe fundamental group π1(X, x). Another choice of base point y ∈ X yields the isomorphicgroup π1(X, y).
An isomorphism π1(X, x) ' π(X, y) requires a choice of path from x to y in X. Such achoice of path is not canonical. Grothendieck taught us an analogue of this for extensions offields.
Q↔ “etale site” SpecQ (something like a space)
choice of Q↔ choice of “base point” of SpecQ
Then the etale fundamental group πet1 (SpecQ,Q) = Gal(Q/Q).
The punchline: The fundamental group π1(X, x) depends on base point x, but Repπ1(X, x) 'local systems on X is canonical! Transporting this statement via Grothendieck’s analogywe see that, although the absolute Galois group is not defined canonically, its category of(continuous) representations is. It is this category that the Langlands correspondence triesto understand.
Last lecture we stated the local Langlands correspondence for GLn: Fix a local fieldK (i.e. K is a finite extension of Qp or K ' Fq((t))). There is a canonical bijection
cts reps of WK
in GLn(C)
/iso
1:1←→
irred blahreps of GLn(K)
/iso
.
Here WK is the Weil group of the field K. Last week we showed why this follows fromlocal class field theory for n = 1. However, this statement is not quite precise. On the lefthand side we need to consider Weil-Deligne representations of WK , and on the righthand side we need to establish exactly what conditions are captured by “blah”. We’ll keepstating versions of this theorem every lecture until we converge on something correct.
51
7.2 The no small subgroups argument
Our final piece of housekeeping is the no small subgroups argument. This is a veryuseful fact that has not fit in naturally to our story so far, so we will slot it in here.
Definition 7.1. A topological group G has no small subgroups if there exists a neigh-borhood U of the identity in G such that any subgroup contained in U is trivial.
Example 7.2. Here are some examples of groups with no small subgroups.
1. The circle group S1.
2. Discrete groups (e.g. finite groups). We can take U = id.
3. The real numbers R and the complex numbers C. (Powers of any non-identity elementmove far away from the identity.)
4. Any Lie group G. (Use that exp : LieG→ G is a local diffeomorphism and exp(mg) =exp(g)m.)
5. Any topological subgroup of a group with no small subgroups has no small subgroups(e.g. Q/Z → S1 has no small subgroups).
Remark 7.3. In contrast, profinite groups have “many small subgroups,” because a basisof neighbourhoods of the identity consists of subgroups of finite index.
Lemma 7.4. Let Γ be a profinite group and G a topological group with no small subgroups.Then any continuous group homomorphism ϕ : Γ→ G has finite image.
Proof. Let U ⊂ G be an open neighborhood of the identity containing no nontrivial sub-groups, and let ϕ : Γ → G be a continuous group homomorphism. Then ϕ−1(U) ⊂ Γ isopen. Since Γ is profinite, there exists a normal subgroup N ⊂ ϕ−1(U) such that G/N isfinite. The image ϕ(N) ⊂ U is a subgroup, so ϕ(N) = 1 since G has no small subgroups.Hence Γ factors through a finite group, ϕ : Γ→ Γ/N → G.
The Moral:Fractal-like objects
(p-adic groups, Galois groups)
∩
Euclidean-type objects(Lie groups)
= finite groups
A consequence of this is that we can not draw any good pictures of Zp in C (or for thatmatter in any Lie group) which respect the addition or multiplication structure.
This moral gives us a new perspective of the local Langlands correspondence. The “nosmall subgroups” lemma implies that the left hand side of the LLC consists (roughly) of acollection of finite subgroups of GLn(C), along with surjections from a Galois-group-typeobject to the subgroups. So very roughly, the LLC provides a classification of irreducibleadmissible representations of a Lie group over a local field by certain finite subgroups ofGLn(C).
52
7.3 You could have guessed the LLC for GL2!
The goal is this section is to give a heuristic explanation for the LLC. Warning: this isnot precise! Everything we say here will have to be tweaked later. Geordie learned thisperspecive from a series of lectures by Dipendra Prasad in Russia.
Starting place: Say we wanted to guess the representation theory of GLn(Qp). What wouldwe do?
Step 1: We might start by figuring out the representation theory of finite reductive groups.For example, let G = SL2(Fq). There are two maximal tori in G, up to conjugacy:
Ts =
(a 00 a−1
)| a ∈ F×q
' Z/(q − 1)Z, the “split torus,” and
Ta = λ ∈ F×q2 ⊂ GL2(Fq) | Norm(λ) = 1 ' Z/(q + 1)Z, the “anisotropic torus”.
In the definition of Ta above, we are using the fact that GL2(Fq) is the group of invertiblelinear transformations of the Fq-vector space F2
q, so
F×q2 ⊂ GLFq(F2q) = GL2(Fq).
Roughly,
irred reps of
SL2(Fq) over C
1:1←→ χ : Ts → C×/χ∼χ−1
“principal series”
⊔θ : Ta → C×/θ∼θ−1
“discrete series”
Let us check that we are not too far off by doing a count. Irreducible representations ofa finite group are in bijection with conjugacy classes, and conjucagy classes are roughly inbijection with characteristic polynomials, so the sizes of the sets above are roughly
# characteristic polynomialsof elements in SL2(Fq)
= |x2 + ax+ 1 | a ∈ Fq| = q =q − 1
2+q + 1
2.
For more details and a careful construction of the irreducible representation of SL2(Fq), seethe notes from Joe Baine’s talks on the Informal Friday Seminar webpage. The upshot isthat we obtain almost all irreducible representations of SL2(Fq) through some “induction”from characters of the two conjugacy classes of tori. (Note that the details are much morecomplicated as there is no actual induction functor. In the setting of finite reductive groupswe use Deligne-Lusztig induction.)
Step 2: Once we have a good idea of the representation theory of finite reductive groups, anext natural step might be to understand the representation theory of real reductive groups.For example, let G = SL2(R). Again, there are two conjugacy classes of maximal tori: the“split torus”
Ts =
(a 00 a−1
)| a ∈ R×
' R×,
53
and the “anisotropic torus”
Ta =
(cos θ sin θ− sin θ cos θ
)' SO2.
Something similar happens in this setting to what we saw with the finite reductive groups.Roughly,
irred admissiblereps of SL2(R)
1:1←→
cts characters
of Ts ' R× Z/2Z
“principal series”
⊔ cts characters
of Ta ' SO2 ' S1
“discrete series”
So again we see that, roughly, irreducible representations are all obtained by “inducing”characters of conjugacy classes of tori.
Step 3: We dream that something similar might be true for representations of p-adic groups.Let G = GL2(K) for a local field K. By descent, we have the following relationship:
conjugacy classes ofmax’l tori in GL2(K)
↔
semisimple K-algebrasL s.t. dimK L = 2
L× ↔ L
There are two cases:
1. Split torus: L ' K ×K, so maximal torus is of the form L× ' K× ×K×.
2. Anisotropic torus: L/K degree 2 extension, so maximal torus is of the form L×.
Applying our analogy from earlier, we might expectirred blah
reps of GL2(K)
roughly 1:1←−−−−−→
pairs of charactersχ1, χ2 : K× → C×
⊔characters θ : L× → C×where L/K is degree 2
.
Now, local class field theory tells us that W abK ' K× and W ab
L ' L×. Moreover, WL ⊂ WK
is an index 2 subgroup, soirred blah
reps of GL2(K)
roughly 1:1←−−−−−→ χ1 ⊗ χ2 : WK → GL2(C) t
IndWK
WL(θ) : WL → GL2(C)
.
It turns out that our dream is a reality:
Fact: If p 6= 2, all continuous representations of WK are either of the form χ1 ⊗ χ2 orIndWK
WL(θ) as above.
So we guessed LLC for GL2(K)! Though again, let us emphasize that this is not actuallythe correct version of the correspondence (it is for example not compatible with taking duals).However it will not take too much effort to make this into a correct statement next lecture.
Remark 7.5. For p = 2, the matching still works, but there are more objects on both sides.
54
7.4 Basic representation theory of p-adic groups
Let K be a local field. Then GLn(K) is a topological group, with a basis of open neighbor-hoods of id given by
Kj =g ∈ GLn(OK) | g = id mod mj
K
.
Note that GLn(OK)/Kj ' GLn(OK/mjK) is a finite group.
For example, if K = Qp, we have a natural surjective map
GLn(Qp) ⊃ GLn(Zp)ϕj−→ GLn(Z/pjZ)
for all j ∈ Z: ≥ 0, and Kj = ϕ−1j (id).
Remark 7.6. 1. Kj ⊂ K0 is normal.
2. K0 is a maximal compact subgroup.
Exercise 7.7. Let π ∈ OK ⊂ K be a uniformizer.
1. Establish the Bruhat decomposition:
GLn(K) =⊔
λ1 ≥ λ2 ≥ . . . ≥ λnλi ∈ Z
GLn(OK)
πλ1
πλ2
πλ3
· · ·πλn
GLn(OK)
(∗)(Hint: Gaussian elimination.)
2. Use (∗) to classify the subgroups GLn(OK) ⊂ H ⊂ GLn(K).
3. Hence or otherwise, show that GLn(OK) is a maximal compact subgroup of GLn(K).
Example 7.8. Consider GL1(Q3) = Q×3 = Z × Z×3 . Recall our picture of the 3-adics fromExample 5.6. A picture of the maximal compact subgroup K0 of this group is:
55
A space is totally disconnected if every point admits a family of compact open neigh-borhoods; e.g. GLn(K) is totally disconnected because each Kj is compact open. Let G bea totally disconnected topological group and V a vector space over a field K of characteristic0. We give V the discrete topology.
Definition 7.9. A representation ρ : G→ GL(V ) is
1. smooth if for all v ∈ V , stabG v is open, and
2. admissible if for all open K ⊂ G, V K is finite-dimensional.
Example 7.10. 1. The trivial representation K is smooth and admissible, K∞ is smoothbut not admissible.
2. The standard representation of GLn(K) on Kn is not smooth, as stabGLn(K) v is notopen for v 6= 0.
3. The groupG = (Zp,+) acts on the vector space F = ϕ : Zp → C | ϕ is locally constantin the natural way, forming the “smooth regular representation”. (This is the p-adicanalogue of L2(G).) We claim that this representation is smooth and admissible.
• Smooth: Let ϕ ∈ F . Then for all x ∈ Zp, there is a neighborhood Ux suchthat ϕ|Ux is constant. This forms a covering of Zp by open neighborhoods ofthe form Ux = x + pnxZp. Since Zp is compact, there exists a finite subcoveringUx1 , . . . , Uxm . Then ϕ is fixed by pnZp, where n = Maxni, so the the stablizerof ϕ is open, hence the representation is smooth.
• Admissible: A basis of open neighborhoods of 0 is given by pmZp, m ≥ 0. Then
FpmZp = ϕ | ϕ is constant on pmZp-orbits= ϕ : Z/pmZ→ C
is finite dimensional, so the representation is admissible.
4. (The most important example!) Recall that P1C is covered by the compact sets D≤1 =z | |z| ≤ 1 and D−1
≤1, so P1C is compact:
56
Similarly, P1K = K ∪∞ is covered by the compact sets D≤1 = z ∈ K | |z|p ≤ 1 =OK and D−1
≤1, so P1K is compact:
The vector spaceI = f : P1K → C | f is locally constant
admits a natural GL2(K)-action, and the same argument as in the previous example(using compactness) shows that I is a smooth, admissible representation of GL2(K).In fact,
constant functions → I St
Where St := I/constant functions is the Steinberg module. The module St isirreducible (exercise, might be hard with current technology!).
Exercise 7.11. Show that any smooth finite dimensional representation of GLn(K) factorsover det : GLn(K) → K×. (Hint: the kernel of a smooth finite dimensional representationis a finite intersection of stabilizers of a basis, so it must be open and normal, hence containSLn(K)).
Exercise 7.12. The representation F ′ = ϕ : Qp → C | ϕ is locally constant of Zp is notadmissible or smooth.
57
Lecture 8: Precise statement of local Langlands for GL2,
p 6= 2
8.1 Basic representation theory of p-adic groups, continued
We pick up where we left off in the previous lecture. Let
mK ⊂ OK ⊂ K
be the maximal ideal in the ring of integers of a local field. We are interested in the rep-resentation theory of the group GLn(K). (Or, more generally, the representation theory ofany totally disconnected group G, but for concreteness we will work with GLn.)
Recall that the sets
Kj = g ∈ GLn(OK) | g = id mod mjK
form a basis of open neighborhoods of id ∈ GLn(K). In addition to being open neighborhoodsof the identity, the Ki are subgroups of GLn(K). (Note the existence of such subgroups whichform a basis for open neighborhoods of the identity is only possible because GLn(K) is atotally disconnected group; a Lie group could not have such a family of subgroups becauseLie groups have no small subgroups.)
Last week we saw that K0 is a maximal compact subgroup of GLn(K). Let V be arepresentation of GLn(K). Because K0 ⊃ K1 ⊃ K2 ⊃ · · · , we have a chain
V K0 ⊂ V K1 ⊂ V K2 · · ·
If V is smooth, each vector lies in V U for some open U ⊂ G, hence lies in some V Ki . So thefiltration is exhaustive. If V is admissible, each V Ki is finite dimensional.
Because Ki ⊂ K0 is normal, the subspace V Ki is stable under action by K0. Thesubgroup Ki ⊂ K0 acts trivially on V Ki , so the K0-action factors through the finite groupK0/Ki ' GLn(OK/m
iK); e.g. for K = Qp, the K0 action on V Ki factors through GLn(Z/piZ).
(The key point here is that Ki ⊂ K0 is normal, so the quotient K0/Ki is a group.) Sincerepresentations of finite groups are completely reducible, we have a decomposition
V Ki =⊕
ρ∈K0/Ki
V Ki(ρ),
where V Ki(ρ) is the ρ-isotypic component of V Ki ; that is, V Ki(ρ) is the direct sum of allirreducible subrepresentations of V Ki which are isomorphic to ρ. Passing to the limit, weobtain a decomposition
V =⊕ρ∈K0
V (ρ).
Here K0 denotes all representations of K0 which factor over some quotient K0/Ki.
Lemma 8.1. The representation V is admissible if and only if each isotypic component V (ρ)in the decomposition above is finite-dimensional.
58
Proof. Assume that V (ρ) is infinite dimensional for some ρ ∈ K0. By definition, ρ factorsthrough K0/Ki for some i. Hence, V (ρ) ⊂ V Ki is an infinite dimensional subspace and V isnot admissible.
To prove the opposite implication, assume that each V (ρ) is finite-dimensional. For eachi, we have a decomposition
V Ki =⊕ρ∈K0
V (ρ)Ki .
But since V (ρ) is the direct sum of irreducible representations which are isomorphic to ρ,we have
V (ρ)Ki =
0 if ρ|Ki 6= triv,
V (ρ) otherwise.
HenceV Ki =
⊕ρ ∈ K0
ρ|Ki = triv
V (ρ).
Since K0/Ki is a finite group, there are only finitely many representations ρ ∈ K0 whichfactor through K0/Ki for any fixed i, so decomposition above is a finite direct sum of finite-dimensional representations, hence V is admissible.
Remark 8.2. This is like the theory of K-finite vectors in representation theory of realLie groups. A big difference is that the representation theory of, for example GLm(Z/pnZ)for large m and n is extremely complicated, whereas we know the representation theory ofcompact Lie groups rather well (highest weights, etc.).
Example 8.3. 1. Consider the Zp-representation F = ϕ : Zp → C | ϕ is locally constantfrom Example 7.10.3. For an open neighborhood pmZp of the identity, the invariantsare
FpmZp = ϕ | ϕ constant on pmZp orbits= regular representation of Zp/pmZp.
Hence,
F =⊕
continuousχ:Zp→C×
Cχ.
2. Consider the GL2(K)-representation
I = f : P1K → C | f is locally constant
from example 7.10.4. Here
IKn = ϕ : P1(OK/mnK)→ C,
soI = lim
→C[P1(OK/mn
K)].
59
Let V be a representation of GLn(K). A map ξ : V → C is smooth if stabGLn(K) ξ isopen. Define the smooth dual
V = smooth vectors ξ : V → C.
Lemma 8.4. Assume V is a smooth representation of GLn(K). If V =⊕
ρ∈K0V (ρ), then
V =⊕
ρ∈K0V (ρ)∗.
In particular, if V is smooth and admissible, then so is V , and V∼−→ V .
Proof. The map ξ : V → C is smooth if and only if ξ vanishes on all but finitely many V (ρ).The lemma follows.
The goal for the remainder of this lecture will be to give a bird’s eye view on the smoothadmissible representations of GL1(K) and GL2(K). But first, we need a digression on norms.
8.2 Canonical norms
Recall that to make the product formula of Section 5.2 hold, we define three types of equiv-alence classes of multiplicative norms (“places,” denoted by v) on a local field K:
• finite places: |x|v := (#OK/p)− valp(x) for some prime p ⊂ OK ,
• real places: |x|v := |i(x)| for some real embedding i : K → R, and
• complex places: |x|v := |i(x)|2 for some pair of conjugate embeddings i : K → Cnot landing in R.
Different normalizations would also yield multiplicative norms, but we chose the ones aboveto make the product formula ∏
places v
|x|v = 1
for x ∈ K× holds. For example, if K = Qp, |p| = ε gives a norm for any 0 < ε < 1, so whydo we choose |p| = 1/p? In some sense, this choice is justified by the product formula, butit is still a little mysterious.
Tate made the following observation which further justifies this choice. For a place v,the completion Kv is is locally compact. Let µ be the additive Haar measure on Kv. Themeasure µ is unique up to a scalar. Define
|x|v = factor by which x· scales the Haar measure;
i.e., |x|v = µ(x·A)µ(A)
for A ⊂ Kv measurable and 0 < µ(A) <∞.
Example 8.5. 1. Kv = R: For x ∈ R, |x|v = µ(x[0,1])µ([0,1])
= µ([0, x]) = |x|.
2. Kv = C: For z ∈ C, and
60
3. K = Qp: Recall that Zp =⊔
0≤m<p
m+ pZp, so pµ(pZp) = µ(Zp). Hence,
|x|v =µ(pZp)µ(Zp)
=1
p.
From now on, whenever we consider a norm on a locally compact field, we will alwaysconsider this canonical norm, denoted | · |.
8.3 Smooth admissible representations of GL1(K)
Let V be a smooth admissible representation of GL1(K) = K×. Since V is smooth admissi-ble,
V =⋃
V Ki
and each V Ki is finite-dimensional. Furthermore, since K× is abelian, each of the subgroupsKj := 1 + mj
K ⊂ O×K is normal in K×, and the group
K×/Kj ' Z× (OK/mjK)×
acts on V Kj . Hence if V is irreducible, V is one-dimensional and determined by a characterof the form | · |cχ : K× → C, where c ∈ C and χ : O×K → C is a continuous character.
Remark 8.6. The category of smooth admissible representations of GL1(K) is not semisim-ple. For example, the representation
x 7→(
1 log |x|0 1
)is a smooth, two-dimensional admissible representation which is not semisimple.
8.4 Smooth admissible representations of GL2(K)
Recall from our heuristic description of last lecture that we expect roughly two types ofrepresentations of GL2(K): “principal series” representations coming from a split torus, and“cuspidal” representations coming from an anisotropic torus.
Let B ⊂ GL2(K) be the subgroup of upper triangular matrices. Given continuous char-acters χ1, χ2 : K× → C, define
I(χ1, χ2) := ϕ : GL2(K)→ C |ϕ loc. const., and ϕ
((a b0 d
)· g)
= χ1(a)χ2(d)|ad|1/2ϕ(g)
for all
(a b0 d
)∈ B.
61
Example 8.7.
I(| · |−1/2, | · |1/2) = ϕ : GL2(K)→ C | ϕ loc. const. and ϕ
((a b0 d
)· g)
= ϕ(g)
= ϕ : P1K ' G/B → C | ϕ locally constant
We saw last time that I(| · |−1/2, | · |1/2) is smooth and admissible.
The representations I(χ1, χ2) formed in this way are called principal series represen-tations.
Theorem 8.8. 1. For all χ1, χ2, I(χ1, χ2) is smooth and admissible.
2. I(χ1, χ2) ' I(χ−11 , χ−1
2 ).
3. If χ1/χ2 = | · |−1, then we have an exact sequence of representations
0→ C(χ1, χ2)→ I(χ1, χ2)→ S(χ1, χ2)→ 0
with dimC(χ1, χ2) = 1 and S(χ1, χ2) irreducible.
4. If χ1/χ2 = | · |, then we have an exact sequence of representations
0→ S(χ1, χ2)→ I(χ2, χ2)→ C(χ1, χ2)→ 0
with dimC(χ1, χ2) = 1 and S(χ1, χ2) irreducible.
5. Otherwise, I(χ1, χ2) is irreducible.
6. If χ1/χ2 ' | · |−1, then S(χ1, χ2) ' S(χ2, χ1) and C(χ1, χ2) ' C(χ2, χ1), and ifχ1/χ2 6' | · |±1, then I(χ1, χ2) ' I(χ2, χ1).
Remark 8.9. Some remarks on the theorem:
(a) The representation I(χ1, χ2) is an example of an induced representation for a totallydisconnected group.
(b) Why the strange |ab|1/2 factor? It is necessary to make 2. hold! So why does 2. hold?
Consider
I(| · |1/2, | · |−1/2) = ϕ | ϕ((
a b0 d
)· g)
= |ad|ϕ(g).
We can define a function
Φ : I(| · |1/2, | · |−1/2)→ C
ϕ 7→∫K0
ϕdµ.
(One can think of I(| · |1/2, | · |−1/2) as being some “functions” on P1K and we areintegrating over P1K to get a number. More precisely, these are “densities”, but to
62
explain why would take us too far afield.) Here is a minor miracle: the function Φ isGL2(K)-invariant, hence
C(| · |1/2, | · |−1/2) = Cis the trivial representation. Now, given ϕ ∈ I(χ1, χ2) and ϕ′ ∈ I(χ−1
1 , χ−12 ), ϕϕ′ ∈
I(| · |1/2, | · |−1/2). By composing with Φ, we get a GL2(K)-invariant pairing:
I(χ1, χ2)× I(χ−11 , χ−1
2 )→ C,
which turns out to be non-degenerate, establishing 2.
(c) Part 6. is the most complicated to prove. It uses an intertwiner I(χ1, χ2)→ I(χ2, χ1)via analytic continuation (there are connections to the Jantzen filtration).
(d) Finally, note that Rep GL2(K) is not semisimple, so we cannot just compute homs asin the finite group case.
The other class of representations of GL2(K) are cuspidal representations.
Theorem 8.10. For every degree 2 extension L/K and continuous character θ : L× → Cwhich does not factor through the norm, there exists an irreducible representation BCL/K(θ).We have that BCL/K(θ) ' BCL/K(θ′) if and only if θσ ' θ′ for σ ∈ Gal(L/K).
The construction of BCL/K(θ) is complicated, via the Weil representation. What is goingon metaphorically?
• Consider SL2(Fq). We have seen in Joe’s Informal Friday Seminar talks that a characterθ : Ta → C× gives rise to a local system Lθ on P1
Fq \P1Fq. Taking the first cohomology
yields a cuspidal representation RGTa
(θ).
• Consider SL2(R). A character θ : SO(2) → C× (such characters are classified by Z)gives rise to a local system O(n) on the upper half plane, taking global sections yieldsa discrete series representation Γ(H,O(n)).
• Now take GL2(K). A character θ : L× → C gives rise to a local system Lθ on “Drinfeldspace” P1K/P1K. Taking first cohomology yields the representation BCL/K(θ). Notethat this is very technical, and is an active area of research.
This can also be viewed through the lens of Langlands functoriality. Let L/K be a degreen extension, so WL ⊂ WK is an index n subgroup. We have the following diagram:
1-dim’l repsof L×
irred. smooth ad.
reps of GL1(L)
n-dim’l reps
of K×
irred smooth
reps of GLn(K)
IndWKWL
BC=“base change”
Thus an innocuous induction functor on the left hand side predicts a highly non-trivialcorrespondence between irreducible representation on the right hand side!
63
8.5 Weil-Deligne representations
We are almost ready to make a precise statement of the local Langlands correspondence forGL2(K)! Recall local class field theory: there is a map
rK : WK → K×.
Composing with | · | gives us the norm character
| · | : WK → Q×.
An n-dimensional Weil-Deligne representation is a triple (ρ, V,N), where
• V is an n-dimensional complex vector space,
• ρ : WK → GL(V ) is a continuous representation, and
• N ∈ End(V ) is nilpotent such that
ρ(x)Nρ(x)−1 = |x|N (∗)
for all x ∈ WK . (In fact, (∗) forces N to be nilpotent.)
Example 8.11. 1. Any n-dimensional continuous representation of WK with N = 0 is aWeil-Deligne representation.
2. The representation ρ =
(| · |χ 0
0 χ
)for any character χ : WK → C× and N =
(0 10 0
)form a Weil-Deligne representation. Indeed, for x ∈ WK ,
ρ(x)Nρ(x)−1 =
(|x|χ(x) 0
0 χ(x)
)(0 10 0
)(|x|−1χ(x)−1 0
0 χ(x)−1
)=
(0 |x|0 0
)= |x|
(0 10 0
).
We denote this Weil-Deligne representation St(χ, | · |χ).
A Weil-Deligne representation is F -semisimple if V is semisimple as a representationof WK .
Exercise 8.12. Let Frob ∈ WK be any lift of Frobenius. Show that a Weil-Deligne repre-
sentation is F -semisimple if and only if Frob is semisimple.
64
8.6 The local Langlands correspondence for GL2
Theorem 8.13. (local Langlands correspondence for GL2, p 6= 2) Fix a local field K ofresidue characteristic p 6= 2. There is a canonical bijection.
F -semisimple2-dimensional
Weil-Deligne reps
/'
↔
irred. smooth admiss.reps of GL2(K)
/'
Moreover, this bijection is given as follows.
χ1/χ2 = | · |±1 :
((χ1 00 χ2
), N = 0
)↔ C(χ1, χ2)
χ1/χ2 = | · |±1 :
((χ| · | 0
0 χ
), N =
(0 10 0
))↔ S(χ, | · |χ)
χ1/χ2 6= | · |±1 :
((χ1 00 χ2
), N = 0
)↔ I(χ1, χ2)
L/K, θ : L× → C× :(IndWL
WK(θ), N = 0
)↔ BCL/K(θ)
Where the character θ : L× → C× does not factor through the norm.
65
Lecture 9: The case of p = 2 and the Satake isomorphism
Today’s lecture has two objectives: to explore what the LLC looks like when p = 2, and toexamine how a simple special case of the LLC for GLn leads to the Satake isomorphism.
9.1 Ramification filtration revisited
To begin, we revisit the ramification filtration of Section 6.4. Let L/K be a finite Galoisextension where K and L are both local fields. We have the following inclusions:
L OL mL (πL)
K OK mK (πL)
Here the uniformizers πK , πL are the only non-canonical objects in the diagram above. Wedenote by kK = OL/mL (resp. kL = OL/mL) the residue fields. There is a short exactsequence
IL/K → Gal(L/K) Gal(kL/kK) = 〈Frob〉 ' Z/fZ.
Lemma 9.1. An element σ ∈ IL/K in the inertia subgroup is determined by σ(πL).
This leads to the ramification filtration of the Galois group. Define
I0 := IL/K , Ij :=
σ ∈ IL/K |
σ(πL)
πL= 1 mod mj
L
.
ThenGal(L/K) ⊃ I0 ⊃ I1 ⊃ I2 ⊃ · · · ⊃ 1
is a filtration of Gal(L/K).
Key Facts:
1. The ramification filtration is a finite exhaustive filtration.
2. There is an injection I0/I1 → O×L/(1 + O×L) ' k×L : σ 7→ σ(πL)πL
. Hence I0/I1 is cyclic oforder prime to p.
3. For j ≥ 1, there is an injection Ij/Ij+1 → (1 + mjL)/(1 + mj+1
L ) ' (kL,+) : σ 7→ σ(πL)πL
.Hence Ij/Ij+1 is an abelian p-group, and I1 is a Sylow p-subgroup of IL/K .
This filtration leads to some nomenclature: The first subquotient Gal(L/K)/IL/K is canon-ically isomorphic to Z/fZ, and is referred to as the unramified part of the Galois group.The second subquotient IL/K/I1 is cyclic of order prime to p, and is referred to as the tamelyramified part of the Galois group. The remaining subquotients of the ramification filtrationare abelian p-groups (and hence I1 is a solvable p-group), and are referred to as the wildlyramified part of the Galois group.
66
The upshot is that there are significant constraints on which groups can appear as Galoisgroups of extensions of local fields. (For example, they must be solvable.) This is in sharpcontrast to the number field setting, where many types of groups can appear as Galois groupsof extensions of Q. (Though exactly which groups appear as Galois groups of number fieldsis still very much an open problem, the inverse Galois problem.)
Example 9.2. Consider the extension L = Qp( p−1√p) of K = Qp. Here πL = p−1
√p and
πK = p are uniformizers. (Exercise: Show that the set µp−1 of all (p− 1)st roots of unity iscontained in Qp.) The Galois group of this extension is
Gal(L/K) = σζ : p−1√p 7→ ζ p−1
√p for ζ ∈ µp−1 ' k×K → k×L .
One can check that Gal(L/K) is totally ramified; that is, IL/K = Gal(L/K). (This followsfrom the observation that we are adjoining roots of p, whose image is zero in the residue
field.) Furthermore, if σζ ∈ IL/K = Gal(L/K), thenσζ(πL)
πL= ζ ∈ k×L , hence I1 = 1 and
Gal(L/K) is tamely ramified.
Remark 9.3. Examples of wild ramification are almost always hard! We really should spenda lecture on such examples, but we are quickly running out of time, so sadly we will not.
Next we will examine the structure of the absolute Galois group of a local field. For alocal field K, we have an exact sequence
IK/K → Gal(K/K) Z.
We would like to pass to the limit to obtain a ramification filtration of the inertia subgroupIK/K from the ramification filtrations of the inertia subgroups of finite extensions. However,there is a problem: if L′/L/K is a tower of finite extensions, then the ramification filtrationof IL′/K is related to multiples of the ramification filtration of IL/K .
This can be fixed through an “upper numbering” procedure which replaces Ij with IλL/Kfor λ ∈ Q≥0 in a way that is compatible with extensions. (Exactly how one does this appearspretty crazy at first sight. It is explained in Serre’s Local Fields [Ser79].) This leads to aramification filtration Iλ
K/Kof the inertia subgroup of the absolute Galois group indexed by
rational numbers:
Gal(K/K) ⊃ IK/K ⊃ I>0K/K⊃ · · · ⊃ I
463/5
K/K⊃ · · · .
This filtration has the property that Gal(K/K)/IK/K = Z canonically, IK/K/I>0K/K'∏l6=p
prime
Zl
non-canonically, and other subquotients are pro-p groups.
Important points:
1. The first two steps depend only on the residue characteristic of the field.
2. Via class field theory, the image of this filtration in W abK ' K× corresponds to the
filtration by 1 + mjK ⊂ OK . The fact that the only jumps in this filtration are at
integers (as opposed to other elements of Q) is the Hasse–Arf Theorem.
67
9.2 More details on Weil–Deligne representations
Next we would like to show two things: (1) why any Weil–Deligne representation is “close” toa continuous representation of Gal(K/K) and (2) why p = 2 is special in the local Langlandscorrespondence.
Proposition 9.4. Any indecomposible F -semisimple Weil–Deligne representation is isomor-phic to Stn ⊗ ρ, where ρ is an irreducible representation of WK .
Here Stn is the Steinberg representation from the previous lecture; e.g.
St4 =
| · |3 0 0 0
0 | · |2 0 00 0 | · | 00 0 0 1
, N =
0 1 0 00 0 1 00 0 0 10 0 0 0
The proof of this proposition is left as a somewhat tricky exercise. It becomes easier ifyou know what the weight filtration associated to a nilpotent operator is.
Proposition 9.5. 1. Let ρ : WK → GL(v) be an irreducible representation. Then thereexists a continuous character χ : WK → C× such that ρ⊗χ has finite image and hencedefines a representation ρ⊗ χ : Gal(K/K)→ GL(V ).
2. Suppose that ρ : WK → GL(V ) is irreducible and not induced from any proper sub-group of WK . Then the restriction to wild intertia is irreducible. In particular, dimVis a power of p (since any irreducible module over a p-group has dimension divisible byp).
Remark 9.6. Proofs of these two statements can be found in [Tat79].
Propositions 9.4 and 9.5.1 show that every Weil–Deligne representation is “close” to arepresentation of the absolute Galois group, in the sense that every Weil-Deligne represen-tation can be obtained from the Steinberg representation and an irreducible representationof WK , and every irreducible representation of WK can be upgraded to a representation ofGal(K/K) by tensoring with a character.
The two statements of Proposition 9.5 are reasonably easy consequences of the followinglemma.
Lemma 9.7. Suppose a group G has the form
Γ → G Z
for some finite group Γ. Then any irreducible G-module is either irreducible over Γ or inducedfrom a subgroup of the form Γ omZ.
The proof of this lemma is a worthwhile exercise!
68
9.3 Why is LLC for p = 2 special?
Proposition 9.5 shows that for p 6= 2, all irreducible 2-dimensional representations of WK
are induced from a finite index subgroup. However, for p = 2, it’s possible that there areirreducible representations of WK which are not induced. So do such representations exist?Yes!
Consider a continuous two-dimensional representation ρ : Gal(K/K)→ GL2(C). By theno small subgroups lemma, the image of ρ must lie in a finite subgroup of GL2(C), so in thecomposition of ρ with the projection
GL2(C)→ PGL2(C),
the image must be conjugate to a subgroup of the maximal compact subgroup SO3 ⊂PGL2(C). The finite subgroups of SO3 were classified5! They are of the following types:
• cyclic (orientation preserving symmetries of the product of an m-gon and an interval,fixing one end)
• dihedral (orientation preserving symmetries of the product of an m-gon and an interval)
• A4 (orientation preserving symmetries of the tetrahedron)
• symmetries of the cube
• A5 (orientation preserving symmetries of the icosahedron)
Reducible representations have images in cyclic subgroups of SO3, and induced representa-tions have images which are dihedral groups. What about the other three? Are there anyrepresentations of Gal(K/K) whose image lies in any of the final three finite subgroups?Since Gal(K/K) is solvable, we can eliminate the non-solvable group A5 from our list. Let’sconsider the composition series of A4:
K4 = Z/2Z× Z/2Z → A4 Z/3Z
By the structure of the ramification filtration, this subgroup structure is only possible for alocal Galois group if p = 2. It turns out that it does indeed occur for some local fields!
The upshot is that for p = 2, there are more representations on each side of the LLC,and the extra representations on the Weil group side are this special class of irreduciblenon-induced representations whose image lies in A4. (Geordie isn’t sure if representationscorresponding to the symmetries of the cube exist. He is told that they do...)
A mystery to ponder: Let G be a compact Lie group (e.g. a finite group), and let R(G)Cbe its representation ring. What is a character
θ : R(G)C → C?
5by Klein in 1884, [Kle93]
69
9.4 Unramified representations
One way to convince yourself that the LLC is amazing is to see that simple special casesalready have deep consequences. The first example of this that we have seen is local classfield theory. The second example we will see now!
The local Langlands correspondence for GLn(K) says that there is a canonical bijection:F -semisimplen-dimensional
Weil-Deligne reps
/'
irred smoothadmissible
reps of GLn(K)
/'
1:1
On the left hand side of this bijection, we can consider a special class of unramified Weil-Deligne representations consisting of those representations of WK which are trivial onthe inertia subgroup. The corresponding representations on the right hand side are thespherical representations of GLn(K):
Weil-Deligne repss.t. N = 0 and ρ
factors through Z:Wk Z → GLn(C)
/'
reps of GLn(K)
admitting aGLn(OK)-fixed vector
/'
1:1
Semisimple representations ofWK which factor through Z are in bijection with semisimpleelements of GLn(C), and irreducible representations of GLn(K) admitting a GLn(OK)-fixedvector are in bijection with irreducible representations of the “spherical Hecke algebra”(which you are not expected to be familiar with and we will soon define). Thus, the restrictionof the local Langlands correspondence to this special case results in a bijection
semisimple elementsin GLn(C)
/conj
1:1←→
irreducible reps ofHsph := H(GLn(OK),GLn(K))
/'
This is the Satake isomorphism! We will spend the rest of the lecture explaining thisbijection (particularly the right hand side) in more detail.
Remark 9.8. The left-hand-side of the bijection above is independent of K, and even ofthe residue characteristic p!
9.5 Hecke algebras
Suppose G is a finite group.
Case 1: Consider N ⊂ G a normal subgroup. If V is a G-representation, then G acts onV N (because for n ∈ N, g ∈ G, and v ∈ V N , n · gv = g · g−1ng · v = gv), and the actionfactors over G/N . Moreover, one can check that End(IndGN C) ' C[G/N ]. Hence we have abijection
irred G-moduleswith an N -fixed vector
/'↔ irred G/N -modules/'.
70
Case 2: Consider H ⊂ G not necessarily normal. Given a G-representation V , what actson V H? The Hecke algebra! The operator
πH : V → V H
v 7→ 1
|H|∑h∈H
h · v
projects onto H-invariants. This can be used to define a “Hecke operator” [HgH] for everyg ∈ G which makes the following diagram commute:
V H V H
V V
[HgH]
πH
·g
Note that all g in the same double coset yield the same Hecke operator. Alternatively, thisoperator is the sum
[HgH] =1
|H|∑
g′∈HgH
g′.
The Hecke algebra H(H,G) of the pair (H,G) is the vector space HC[G]H with multipli-cation
(f ∗ f ′)(g) :=1
|H|∑g=hh′
f(h)f ′(h).
This is an associative unital algebra with unit
1H =1
|H|∑h∈H
h.
Example 9.9. 1. If N is normal, H(N,G) = C[G/N ].
2. If G = GLn(Fq) and B =
∗ · · · ∗0
. . ....
0 0 ∗
, then H(B,G) is the “Hecke algebra of Sn
at q = |Fq|”. This algebra is almost independent of q.
Exercise 9.10. (Do it!) Show that
End(IndGH C) ' H(H,G).
Hence〈IndGH C〉 ∼−→ H(H,G)-mod.
(Here the angle brackets mean the smallest abelian category generated by kernels, cokernels,extensions, and direct sums.) Deduce that
irred. G-modules
with H-fixed vector
1:1←→ irred H(H,G)-modules .
71
Remark 9.11. There is a tendency in the literature to consider one subgroup H at a time,but one can also consider all subgroups (or a particularly nice family of subgroups) at thesame time, resulting in a “Hecke algebroid”.
We can also define Hecke algebras of p-adic groups. Let G = GLn(K) for a local field K,and K0 = GLn(OK) the maximal compact subgroup. Then the “big” Hecke algebra of G is
Hbig =
ϕ : G→ C
∣∣∣∣ϕ locally constant
compact support
.
An alternate description is
Hbig =⋃i
ϕ : G→ C
∣∣∣∣ ϕ locally constant on Ki-double
cosets, non-zero on finitely many
.
Exercise 9.12. Prove that the two formulations of Hbig are equivalent.
The algebra structure on Hbig is given by
(f ∗ f ′)(g) =
∫h∈G
f(h)f ′(h−1g)dµ,
where µ is the Haar measure.
Example 9.13. Let 1Ki be the indicator function on Ki. Then
1Ki ∗ 1Ki(g) =
∫h∈G
1Ki(h)1Ki(h−1g)dµ =
0 if g 6∈ Ki,∫Ki
1dµ if g ∈ Ki.
In other words,1Ki ∗ 1Ki = µ(Ki)1Ki ,
so 1Ki is a quasi-idempotent.
Remark 9.14. Because any irreducible G-module has V Ki 6= 0 for some i, Hbig can be usedto understand all smooth admissible representations of G. However, it is very complicated.
Assume µ(K0) = 1 so 1K0 is idempotent. The spherical Hecke algebra is
Hsph = H(K0, G) := 1K0Hbig1K0 .
Exercise 9.15. (Do it!) Prove the Cartan decomposition of G:
G =⊔λ
λ1 ≥ λ2 ≥ . . . ≥ λnλi ∈ Z
K0
πλ1
πλ2
πλ3
. . .
πλn
K0
HenceHsph =
⊕λ
C1λ.
72
There are two miracles.
Theorem 9.16. 1. The spherical Hecke algebra Hsph is commutative.
2. (The Satake isomorphism) There exists a canonical bijection
Hsph ∼←→ R(LGLn(C)).
Remark 9.17. The Langlands dual group L GLn(C) ' GLn(C) so we could have replacedthe right hand side of the Satake isomorphism with the representation ring of GLn(C);however, the theorem also holds for general reductive groups and there the dual group isimportant.
Recall our mystery from earlier in the lecture: For a compact Lie group G, what is acharacter θ : R(G)C → C of its representation ring? By the Chevalley restriction theorem,R(G)C ' R(T )WC , where T ⊂ G is a maximal torus and W is the Weyl group of G. So acharacter of R(G)C is just a choice of a semisimple conjugacy class in G!
Theorem 9.16 can be used to establish unramified LLC:“spherical representations”;
i.e. smooth admissibleirred reps of G with
a K0-fixed vector
/'
1:1←−−−−−−−−−−→Hecke algebra magic
irreducible
H(K0, G) = Hsph−modules
/'
1:1←−−−−−−−−−−−→commutativity of Hsph
characters
χ : Hsph → C
1:1←−−−−−−−−−→
Satake isomorphism
characters
θ : R(LGLn(C))→ C
1:1←−−−−−−−−−−−−→
the mystery from earlier
conjugacy classes
of semisimpleelts in L GLn(C)
1:1←−−−−→
definition
unramified
n-dimensionalWeil–Deligne
representations
/'
73
Lecture 10: The big picture
Today is about the big picture. We start with the very big picture, and finish with themoderately big picture. This is also the final lecture of the first term of this course!
10.1 The very big picture
10.1.1 Dimension 0
Let us go back to the beginning. Let f(x) ∈ Z[x] be a polynomial; e.g. f(x) = x2 + 1.Back in March, we wondered: How many solutions does f(x) have modulo a prime p? Weconstructed tables:
p 2 3 5 7 11 13 17 19 23# of sol’s mod p 1 0 2 0 0 2 2 0 0
p mod 4 2 3 1 3 3 1 1 3 3...
Then we studied this via representation theory. The Galois group Gal(Q/Q) acts on theroots σ1, . . . , σn ⊂ Q of f(x), so we have a permutation representation
Gal(Q/Q)→ GL(H),
where H :=⊕n
i=1 = Cσi. Then for unramified primes,
# solutions of f(x) mod p = Tr(Frobp, H).
Even in this innocent (“dimension 0”) case, H is enormously complicated. To simplifythings, we instead considered a collection of local representations HQp , defined as follows.
For each p, consider roots σ′1, . . . , σ′n of f(x) in Qp. Then for each p we have “local Galois
representations”Gal(Qp/Qp)→ GL(HQp),
where HQp =⊕
Cσ′i. This gives us a “categorification” of the table above:
p 2 3 5 7 11 13 17 19 23HQ2 HQ3 HQ5 HQ7 HQ11 HQ13 HQ17 HQ19 HQ23
...
If p is unramified, then the inertia subgroup I ⊂ Gal(Qp/Qp) acts trivially on HQp , so thelocal Galois representation is unramified. By the Satake isomorphism (Theorem 9.16), thisimplies that HQp is determined by a semisimple conjugacy class [x] ∈ GLn(C), and
# solutions of f(x) mod p = Tr([x]).
For unramified primes, the representation HQp is rather simple. However, for ramified primes,the representation HQp can be quite complicated:
1. The study of HQp lets us define local factors in the Artin L-function.
2. We can hope to understand HQp through the local Langlands correspondence.
Remember our slogan: There is a lot of substance at ramified primes/points!
74
10.1.2 Dimension ≥ 1
The classic example is that of an elliptic curve E; e.g. the projective completion of the curvey2 + y = x3 +x2 + 3x+ 5 that we studied in the lecture on the Sato-Tate conjecture, Lecture4. What is the analogue of the Galois representation H in this setting?
Recall that E is a group, and the complex points of E are
E(C) = solutions over C = C/Λ,
where Λ ⊂ C is a lattice That is, we obtain E(C) by identifying opposite edges in thispicture, where the dots represent elements of Λ:
The Galois group Gal(Q/Q) does not act in any meaningful way on E(C). It does act onE(Q), but this is an enormously complicated set, a little too complicated for us! However,for any prime `, we can consider the “`m-torsion points”:
E[`m] := x ∈ E(Q) | `m · x = 0 ' (Z/`mZ)2 ;
e.g., for ` = 3,m = 1:
There is a natural action of Gal(Q/Q) on E[`m]. The Tate module is formed by taking thedirect limit of the E[`m]:
T`(E) := lim←E[`m] ' Z2
` .
The Tate module has a continuous action of Gal(Q/Q). Moreover, if EFp is smooth, then
#E(Fp) = 1 + p− Tr(Frobp, T`(E)).
So in this classic example of an elliptic curve, the Tate modules play the role of the repre-sentation H which appeared in the dimension 0 setting. Notice that in the previous section
75
we constructed a single representation H, but there is one Tate module for each prime `.This is an embarrassment of riches!
The Tate module T`(E) is an example of “`-adic cohomology”:
T`(E) = H1et(E,Z`)∗.
In general, given a variety X over a field k and a prime ` such that multiplication by ` isnon-zero in k, there is a continuous action of Gal(k/k) on the `-adic cohomology groupsH∗et(Xk,Q`). Again, it is useful to study these representations via their restriction to localGalois groups Gal(Qp/Qp). We can always calculate these after base change to SpecQp andat primes of good reduction after base change to SpecFp. (Exercise: Think about whatthis statement means for f(x) ∈ Z[x].)
In addition to etale cohomology, one has several other methods for associating cohomologygroups to the variety X:
1. singular cohomology: H i(X(C),Z), H i(X(C),Fp) (related via universal coefficienttheorem)
2. deRham cohomology: H idR(X), H i
dR(XFp) (cohomology of differential forms)
3. crystalline cohomology: H icrys(XFp/Zp) (a fancy theory that produces Zp-vector
spaces for Fp-schemes)
Grothendieck’s philosophy: All of these cohomology groups should be shadows of aunique object, the “motive” of X.
Scholze: Perhaps the “motive” is more like a sheaf/local system on SpecZ× SpecZ.
Scholze’s ICM picture:
76
Scholze also predicts an archimidean theory for varieties in characteristic p which has beenmissing since the beginning of this subject!
Recommendation/Exercise: Read section 10 of Scholze’s ICM paper.
How does one compare columns in Scholze’s picture? In other words, ifGQp = Gal(Qp/Qp),how can we compare
ρ : GQp → GLn(Q`) and ρ′ : GQp → GLn(Q`′)?
The problem is the topology. The solution is given by Weil–Deligne representations.A topological group Γ is pro-p if it is profinite and for all open normal subgroups N ⊂ Γ,
Γ/N is a p-group.
Example 10.1. Two examples of pro-p-groups are:
1. wild inertia ⊂ GQp , and
2. 1 + pMatnZp = K1 ⊂ GLn(Qp).
Lemma 10.2. Any continuous group homomorphism
ρ : Γ→ G
from a pro-p group Γ to a pro-` group G is trivial.
Corollary 10.3. For a pro-p group Γ, any continuous group homomorphism
ρ : Γ→ GLn(Qp)
has finite image.
We have seen (via the ramification filtration) the GQp has the following structure:
Grothendieck showed us that the pro-`′ group∏
`′ 6=p,` Z`′ ⊂ I must have finite image. More-over, ρ(1`) ∈ GLn(Q`) is almost unipotent. So what Grothendieck has shown us is that wecan “take logs to get Weil–Deligne representations”.
77
Theorem 10.4. (Grothendieck) After identifying Q` with C, one has a canonical injectioncts. reps.
ρ : GQp → GLn(Q`)
→
Weil–Deligne repsof GQp over C
.
Notice that the set on the right is independent of `!
Remark 10.5. We are being a bit lazy, but one can identify the image of this injection.
10.2 Local Langlands correspondence for split groups
Let G be a split reductive algebraic group over Z determined by the root datum (X∗ ⊃R,X∗ ⊃ R∨), and G its dual group, determined by the opposite root datum (X∗ ⊃ R∨, X∗ ⊃R). Fix a local field K and set q = |OK/mK |.
A Weil–Deligne representation in G is a pair (ρ, e), where
• ρ : WK → G(C) is a continuous group homomorphism, and
• e ∈ Lie G(C) is a nilpotent element
such that ρ(g)eρ(g)−1 = |g|e for all g ∈ WK . A Weil–Deligne representation in G is F -semisimple if ρ is semisimple.
Example 10.6. A Weil–Deligne representation in GLn is just an n-dimensional Weil–Delignerepresentation, as in Section 8.5.
Given a Weil–Deligne representation (ρ, e), consider
ZG(ρ, e) = g ∈ G | g · (ρ, e) = (ρ, e).
Theorem 10.7. (Local Langlands correspondence) There is a canonical correspondenceF -semisimple
WD reps in G
/G-conj
1:finite←−−→
irred smooth
admissible repsof G(K)
/'
Fibres of this map should be indexed by irreducible representations of ZG(ρ, e)/ZG(ρ, e) andare called “L-packets”.
10.3 The Deligne–Langlands conjecture
Last week we examined (for G = GLn) a simple special case of the local Langlands corre-spondence, the case of unramified WD representations, and found that it followed from theSatake isomorphism. Another slightly less simple special case of the LLC is given by tamelyramified WD representations with unipotent monodromy (TRUM). By restrictingthe correspondence in Theorem 10.7 to TRUM, we hope to obtain a correspondence:
TRUM; i.e. (ρ, e)s.t. ρ factors
WK Z, e arbitrary
1:finite←−−→
reps with anIwahori fixed
vector
78
By analogous arguments to the ones we made last week for GLn, the set of ρ which factorthrough WK Z is in bijection with the set of conjugacy classes of semisimple elements inG. Hence the left hand side of the correspondence above is in bijection with the set
(s, e) | s ∈ G semisimple , e ∈ Lie G nilpotent s.t. ses−1 = qe/G-conj.
The right hand side of the corrspondence above is in bijection with the set
irred reps of the “Iwahori-Hecke algebra” Haff := H(I,G(K)).
This motivates the following conjecture of Deligne–Langlands.
The Deligne–Langlands conjecture: As in the set-up above, let q = |OK/mK | be theresidue characteristic of the local field K. There is a bijection:(s, e, χ)
∣∣∣∣∣∣∣∣s ∈ G(C) semisimple,
e ∈ Lie G nilpotent, andχ irred rep of π0(ZG(ρ, e))
such that ses−1 = qe
/G-conj
'←−→ irred Haff-modules/'
Remark 10.8. The affine Hecke algebra H(I,G(K)) has a presentation in which q becomesa variable. The above conjecture can either be understood with fixed q = #|OK/mK | orwith q as a variable, in which case q is also a variable on the left hand side.
Recall that the unramified LLC followed from the Satake isomorphism:
Hsph = H(G(OK), G(K))
“constructible”
'←−→ R(G) = O(
semisimple conj.
classes in G
)'←−→K0(pt/G) = KG(pt)
“coherent”
Similarly, the TRUM case of the LLC (which reduces to the Deligne–Langlands conjecture)follows from the Kazhdan–Lusztig isomorphism:
Haff
“constructible”
'←−→ KG×C×(St)
“coherent”
Indeed, if π : N = T ∗B → N is the Springer resolution, Be = π−1(e) is the Springer fibre of anilpotent element e ∈ N , and St is the Steinberg variety, the Kazhdan–Lusztig isomorphism(which is not easy to establish!) implies that there is an action of the affine Hecke algebraHaff on KZ
G×C× (e)(Be). Here we can see that
ZG×C×(e) = (g, c) | c · geg−1 = e = (g, c) | geg−1 = c−1 · e
looks very close to the parameters in the Deligne–Langlands conjecture. This action showsus that the K-theory of Springer fibres provides all simple Haff-modules, thus proving theDeligne–Langlands conjecture.
79
10.4 Geometric Satake equivalence
There is a geometric upgrade of the Satake isomorphism which has proven to be a majortool in geometric representation theory. Set K = k((t)), so OK = k[[t]], where k = C or Fq.Then
H(G(OK), G(K)) =G(OK)-invariant functions on the
“affine Grassmanian” GrG := G(K)/G(OK).
The geometric Satake equivalence is the equivalence of categories:
(PervG(OK)(Gr,C), ∗)“constructible”
'−→ (Rep GC,⊗)
“coherent”
This equivalence was key in recent work by V. Laffourges giving an “automorphic to Galois”correspondence for global function fields.
10.5 Bezrukavnikov’s equivalence
There is also a geometric upgrade of the Kazhdan–Lusztig isomorphism. With K = k((t))as above, the affine Hecke algebra is
Haff = Iwahori-invariant functions on G(K)/I.
Here G(K)/I is the set of k-points of the “affine flag variety” F lG. Roughly, Bezrukavnikov’sequivalence is an equivalence of categories
(DbI×I(F lG), ∗)
“constructible”
'−→ (DbCohG×C×
(St), ∗)“coherent”
.
Remark 10.9. This is a bit of a lie! It would take several more lecture to precisely describethe categories on each side of this equivalence.
This equivalence has many applications in geometric representation theory. For example,a mod p version of this equivalence would imply everything that we know about modularrepresentations of algebraic groups!
80
Lecture 11: Review of the first semester
11.1 The local Langlands correspondence
Recall our setup from last semester. Let K be a local field (i.e. K is a finite extension of Qp
or K = Fq((t))) with ring of integers O and residue field k:
K ⊃ O k
Nothing is lost by just thinking in terms of the example
Qp ⊃ Zp Fp.
Last semester, we worked up to stating the local Langlands correspondence.
Local Langlands correspondence for GLn: There exists a canonical bijectionirred. smooth admiss.
reps of GLn(K) onC-vector spaces
/'
1:1←→
F -semisimplen-dimensional
Weil-Deligne reps
/'
It has been a while sine the last lecture, so let us remind ourselves what all of these wordsmean. The representations on the left hand side of this correspondence are irreducible repre-sentations which are usually infinite-dimensional. The adjective smooth means that everyvector has an open stabilizer, and the adjective admissible means that for any open sub-group U ⊂ GLn(K), V U is finite-dimensional. We consider the set of such representationsup to equivalence.
The objects on the right hand side are n-dimensional representations (where n is thesame n appearing in GLn on the left) of the Weil group attached to the field K, along withsome extra data. Recall that the Weil group is a subgroup of the absolute Galois groupGal(K/K) defined by the fact that it fits into the following diagram.
IK/K WK := ϕ−1(Z) Z = 〈Frob〉
IK/K Gal(K/K) Gal(k/k) = Zϕ
Here IK/K is the inertia subgroup. A Weil-Deligne representation is a triple (V, ρ,N),where
1. V is an n-dimensional C-vector space,
2. ρ : WK → GL(V ) is a continuous representation, and
3. N is a nilpotent endomorphism of V such that
ρ(x)Nρ(x)−1 = |x|N
for all x ∈ WK .
81
Here | · | is the canonical norm on Wk, which is uniquely determined by the property that|Frob| = |k|. (See Section 8.2 for a refresher on why this exists.) A Weil-Deligne representa-tion is F -semisimple if any lift of Frob acts semisimply.
Remark 11.1. For the experts: Weil-Deligne representations are distilled out of continuousrepresentations WK → GLn(Q`) via “log of monodromy”. Thus one can think of the righthand side as secretly being genuine representations of a group. It is phrased in this way toremove the auxilary choice of a prime number `.
In general, if we replace GLn(K) with any split reductive group G(K) (e.g. SLn, Sp2n,etc.), then the LLC changes as follows.
Local Langlands correspondence for split reductive groups: There exists a canonicalfinite-to-one map
irred. smooth admiss.reps of G(K) onC-vector spaces
/'
finite:1−−−−→
F -semisimple
Weil-Deligne repsof WK in G∨(C)
/'
In this setting, a Weil-Deligne representation of WK in G∨(C) is a pair (ρ,N), where
1. ρ : WK → G∨(C) is a continuous group homomorphism of the Weil group into thecomplex Langlands dual group of G, and
2. N ∈ LieG∨(C) such thatρ(x)Nρ(x)−1 = |x|N
for all x ∈ WK .
Note that condition 2 forces N ∈ G∨(C) to be a nilpotent element.
11.2 The unramified story
Last semester we ended the course by unpacking the simplest piece of the LLC, the case ofunramified representations. Let us remind ourselves how this story went. By restricting eachside of the correspondence above, we obtain a bijection
irred. smooth admissibleunramified reps of G(K)
(i.e. reps of G(K) that admita non-zero G(O)-fixed vector)
/'
unramified
Weil-Deligne reps;(i.e. reps which factorWK Z→ G∨(C)
with N = 0)
/'
1:1
By Hecke algebra theory, representations of G(K) with a G(O)-fixed vector are in bijectionwith irreducible modules for the spherical Hecke algebra,
Hsph = H(G(K), G(O)) := (FunG(O)×G(O)(G(K),C), ∗).
The functions in this definition are continuous with compact support. On the other hand,unramified Weil-Deligne representations are in bijection with the set of conjugacy classes ofsemisimple elements in G∨(C). Recall the Satake isomorphism.
82
Theorem 11.2. (Satake isomorphism) There exists a canonical isomorphism
Hsph∼−→ [RepG∨]⊗Z C.
This theorem will be a big feature of the course this semester. Let T∨ ⊂ G∨ be a maximaltorus, and W the Weyl group of G∨. By highest weight theory we have
[RepG∨] [RepT∨] = Z [X∗(T∨)]
Z [X∗(T∨)]W
∼
Hence[RepG∨]⊗Z C = C[X∗(T∨)]W = O(T∨/W ) = O(G∨ss/conj).
The spherical Hecke algebra is commutative. Therefore,
irred. Hsph-modules ↔ χ : [RepG∨]⊗ZC→ C = semisimple conjugacy classes in G∨(C)
The first bijection follows from the Satake isomorphism and the commutativity of Hsph, andthe second equality is the nullstellensatz. So the Satake isomorphism implies unramifiedLLC!
Remark 11.3. The goal of the next two Informal Friday Seminars (September 6 and 13) willbe to explain the geometric Satake equivalence, which is a categorification of Theorem11.2:
(PervG(O)×G(O)(G(K),C), ∗) ∼←→ (RepG∨C,⊗)
11.3 The tamely ramified with unipotent monodromy (TRUM)story
This semester we will focus on the next simplest piece of the LLC, the case of tamelyramified representations with unipotent monodromy (TRUM). In this setting, the LLC givesus a finite-to-one map:
TRUM reps of G(K)
(i.e. those which admita non-zero Iw-fixed vector)
/'
finite:1−−−−→
TRUM Weil-Deligne reps(i.e. reps which factorWK Z→ G∨(C)with N arbitrary)
/'
Here Iw ⊂ G(K) is the Iwahori subgroup, which sits in the group G(K) in the following way:
G(O) Iw
G(k) B (Borel)
83
By Hecke algebra theory, TRUM representations of G(K) are in bijection with irreduciblerepresentations of the Iwahori-Matsumoto Hecke algebra Haff = H(G(K), Iw). On the otherhand, TRUM Weil-Deligne representations are parameterized by the set
(s,N) ∈ G∨(C)×N | sNs−1 = qN/conj.
Here s ∈ G∨(C) is the semisimple image of Frobenius, N ⊂ LieG∨ is the nilpotent cone,and q = |k|. So the LLC predicts a parameterisation of irreducible Haff-modules. Thisprediction is the Deligne–Langland conjecture, and it served as an important early test caseof the Langland’s philosophy.
Goals for the next few weeks:
1. Discuss the Iwahori-Matsumoto Hecke algebra is some detail.
2. Discuss Kazhdan–Lusztig’s realisation of the affine Hecke algebra H via equivariantK-theory:
H∼−→ KG∨×C×(Steinberg)
3. Deduce the Deligne–Langlands conjecture6.
Then we will pass to categorifications!
11.4 Affine Weyl groups and affine Hecke algebras
Let (X ⊃ R,X∨ ⊃ R∨) be a root datum.
Example 11.4. 1. SL2: X = Z ⊃ R = ±2, X∨ = Z ⊃ R∨ = ±1
2. PGL2: X = Z ⊃ R = ±1, X∨ = Z ⊃ R∨ = ±2We see from this example that SL2 and PGL2 are interchanged by swapping roots and
coroots, so they are Langlands dual groups.
Let Wf be the finite Weyl group associated to this root datum. Then Wf acts on bothX and X∨. Assume that our root datum (X ⊃ R,X∨ ⊃ R∨) is adjoint; i.e., X = ZR. (Thisis the “most complicated case”.)
Definition 11.5. The extended affine Weyl group7 is
Wext = ZX∨ oWf .
6Actually, we will not end up deducing the Deligne-Langlands conjecture. We will simply promise thereader that the Deligne-Langlands conjecture is not a long walk from where we get to.
7This definition is the definition according to Iwahori-Matsumoto and Bourbaki, but we warn the readerthat it is not consistent across all sources! For example, this is not the convention employed in Chriss-Ginzburg.
84
The group Wext acts on X∨R := X∨ ⊗Z R by “affine transformations;” that is, w ∈ Wf
acts as usual, w(λ) = w(λ), and for γ ∈ X∨, tγ ∈ ZX∨ acts by
tγ(λ) = λ+ γ.
To understand Wext, we first consider the affine Weyl group W = ZR∨ oWf ⊂ Wext. Forα ∈ R, m ∈ Z, λ ∈ X∨R , define
sα,m(λ) := λ− 〈λ, α〉α∨ +mα∨.
Clearly,sα,m = tmα∨ sα,
so sα,m ∈ W , and tmα∨ ∈ 〈sα,m | α ∈ R,m ∈ Z〉. We conclude that
W = 〈sα,m | α ∈ R,m ∈ Z〉
is an affine reflection group generated by reflections sα,m through the hyperplanes
Hα,m = λ ∈ X∨R | 〈λ, α〉 = m.
We call the set Hα,m the set of reflecting hyperplanes. Denote by A the corresponding setof alcoves; that is, the closures of connected components of
X∨R\⋃α∈Rm∈Z
Hα,m.
Fix a set of positive roots R+ ⊂ R, and let
A0 = λ ∈ X∨R | 0 ≤ 〈λ, α〉 ≤ 1 for all α ∈ R+ ⊂ A
be the fundamental alcove. The general (very beautiful) theory of reflection groups gives:
1. W is a Coxeter group with Coxeter generators S := reflections in the walls of A0.
2. The length function may be described by
`(w) = #reflecting hyperplanes separating Aint0 and wAint0 .
3. A0 is a fundamental domain for the W -action on X∨R .
So we have an identification
W → Aw 7→ wA0
Example 11.6. C2 = B2
85
Now we move on to Wext. There is an action of Wext on A because
tγ ·Hα,m = Hα,m+〈γ,α〉.
Define
` : Wext → Z≥0
w 7→ `(w) := #hyperplanes separating Aint0 and wAint0
Define the length zero elements of Wext to be
Ω := `−1(0).
Lemma 11.7. Wext = Ω nW .
Proof. Step 1: W ⊂ Wext is normal.
Let γ ∈ X∨, λ ∈ X∨R . Then
tγsα,mt−1γ (λ) = tγ(sα,m(λ− γ)
= λ− γ − 〈λ− γ, α〉α∨ +mα∨ + γ
= λ− 〈λ, α〉α∨ + (〈γ, α〉+m)α∨
= sα,〈γ,α〉+m(λ).
Step 2: Wext = W · Ω.
Let w ∈ Wext. Then wA0 ∈ A, so by 3. above, there exists y ∈ W such that ywA0 = A0.Hence yw = ω for ω ∈ Ω, and w = y−1ω.
Step 3: W ∩ Ω = id.
Any w ∈ W ∩ Ω is length zero, so w = id by 2. above and the fact that W is a Coxetergroup.
86
Lemma 11.8. (Iwahori-Matsumoto) w ∈ Wf , λ ∈ X∨,
`(tλw) =∑α∈R+
w−1(α)>0
|〈λ, α〉|+∑α∈R+
w−1(α)<0
|〈λ, α〉 − 1|.
Proof. Let x = tλw. Then
`(x) = #hyperplanes separating Aint0 and xAint0
=∑α∈R+
#m | Hα,m separates Aint0 and xAint0 .
Now chose a point p ∈ A0, very close to zero. Then 〈p, α〉 is small and positive, and
`(x) =∑α∈R+
#integers between 〈p, α〉 and 〈xp = λ+ wp, α〉
=∑α∈R+
|〈λ, α〉| if 〈wp, α〉 > 0
|〈λ, α〉 − 1| if 〈wp, α〉 < 0
=∑α∈R+
w−1α>0
|〈λ, α〉|+∑α∈R+
w−1α<0
|〈λ, α〉 − 1|.
Example 11.9. For PGL2, X∨ = Z$1. Then
`(m$1) = |m|, and `(m$1s) = |m− 1|.
Hence $1s is length zero, and Ω = id, t$1s.
87
Lecture 12: The Deligne-Langlands conjecture
12.1 The Iwahori–Matsumoto Hecke algebra
Recall our setup from last week. From an adjoint root datum (X ⊃ R,X∨ ⊃ R∨), (i.e.;meaning X = ZR) we construct
• the finite Weyl group Wf ,
• the affine Weyl group W = ZR∨ oWf , and
• the extended affine Weyl group Wext = ZX∨ oWf .
We fix a set of positive roots R∨+ ⊂ R∨, then obtain the fundamental alcove A0, and thecorresponding set S ⊂ W of Coxeter generators. We define a length function ` by
` : Wext → Z
x 7→ #
reflecting hyperplanesbetween Aint0 and xAint0
.
We denoted the length zero elements by Ω = `−1(0), and showed that Wext is a “quasiCoxeter group,” meaning that
Wext = Ω nW,
W is a Coxeter group, and Ω acts on W via automorphisms of the Coxeter system.
Example 12.1. For PGL2,
The affine Weyl group and extended affine Weyl group are
W = 〈s, t | s2 = t2 = id〉 ⊂ Wext = 〈s, t, τ = $s | s2 = t2 = τ 2 = id, τs = tτ〉.
We can use the extended affine Weyl group to define a Hecke algebra.
Definition 12.2. The Iwahori–Matsumoto Hecke algebra Hext is the Z[v±1]-algebra withbasis Hx | x ∈ Wext and multiplication
HxHy = Hxy if `(xy) = `(x) + `(y)
H2s = Hid + (v−1 − v)Hs if s ∈ S.
Define H = 〈Hx | x ∈ W 〉 ⊂ Hext.
Remark 12.3. 1. H ⊂ Hext is a subalgebra.
88
2. For all x ∈ Wext, Hx is invertible.
3. If τ ∈ Ω, then `(xτ) = `(x) = `(τx). Hence
Hext = Ω nH.
Question: Where did the loop presentation Wext = ZX∨ oWf go?
For λ ∈ X∨, write λ = γ − γ′ with γ, γ′ ∈ X∨+. Define
Hλ := HtγH−1tγ′.
Note that Htλ 6= Hλ in general! For example, if λ ∈ X∨+, then Hλ = Htλ , but if λ ∈ −X∨+,then Hλ = H−1
t−λ.
Why is this well-defined? Assume that λ = γ−γ′ = µ−µ′ for γ, γ′, µ, µ′ ∈ X∨+. To showthat Hλ is well-defined, we need to show that
HtγH−1tγ′
= HtµH−1tµ′.
This is equivalent to showing that for ζ ∈ X∨,
HtγH−1tγ′Htζ = HtµH
−1tµ′Htζ .
If we choose ζ ∈ X∨+ very dominant, then ζ − µ′ is dominant and
Htζ−µ′Htµ′
= Htζ−µ′ tµ′= Htζ = Htµ′ tζ−µ′
= Htµ′Htζ−µ′
.
The first and fourth equalities follow from the fact that the lengths of tζ−µ′ and tµ′ add bythe Iwahori–Matsumoto lemma (Lemma 11.8). Hence,
HtγH−1tγ′Htζ = HtγHtζ−γ′
= Htγ+ζ−γ′= Htµ+ζ−µ′
= HtµHtζ−µ′= HtµH
−1tµ′Htζ ,
so Hλ is well-defined.
The upshot: When studying representations of an algebra, it is useful to have a largecommutative subalgebra. This is what we have just accomplished for the Iwahori–MatsumotoHecke algebra: the map λ 7→ Hλ determines an embedding
Z[v±1][X∨] → Hext.
We can use this commutative subalgebra to describe the center of Hext.
Theorem 12.4. (Bernstein) For any λ ∈ X∨+, define
zλ :=∑
µ∈Wfλ
Hµ.
Then the center Z(Hext) of Hext is a free Z[v±1]-module with basis zλ | λ ∈ X∨+, and
Z(Hext) =(Z[v±1][X∨]
)Wf .
89
Now we can state the Bernstein presentation of Hext.
Theorem 12.5. (Bernstein presentation) The Iwahori–Matsumoto Hecke algebra admitsthe following presentation.
1. 〈Hs | s ∈ Sf〉 generate a finite Hecke algebra (“finite part”).
2. HλHγ = Hλ+γ for all λ, γ ∈ X∨ (“lattice part”).
3. For λ ∈ X∨, sα ∈ Sf ,
HsαHsα(λ) −HλHsα = (v − v−1)
(Hλ −Hsα(λ)
1−H−α
)= (v − v−1)(Hλ +Hλ−α + · · ·+Hsα(λ)+α).
In other words, we haveHext ' Z[v±1][X∨]⊗Hf .
We will check the relations for PGL2. First note that if 3. holds for λ and γ, then itholds for λ+ γ:
HsαHsα(λ)+sα(γ) = (v − v−1)
(Hλ −Hsα(λ)
1−H−α
)Hsα(γ) +HλHsαHsα(γ)
= (v − v−1)
(Hλ+sα(γ) −Hsα(λ)+sα(γ) +Hλ+γ −Hλ−sα(γ)
1−H−α
)+Hλ+γHsα
= (v − v−1)
(Hλ+γ −Hsα(λ+γ)
1−H−α
)+Hλ+γHsα .
In particular, if 3. is true for λ, then it is true for −λ.
For PGL2: We will check 3. for λ = $. We have that 〈$,α〉 = 1, τ = t$sα, so t$ = τsα,and H$ = HτHsα . We compute
HsαH−$ −H$H−1sα = HsαH
−1sα Hτ −HτHsαHsα
= Hτ −Hτ (1 + (v−1 − v)Hsα)
= (v − v−1)HτHsα
= (v − v−1)H$.
12.2 The Deligne–Langlands conjecture
Now we return to the LLC. Let K be a local field with ring of integers O and residue fieldk. Define q := |k|. Let G/K be a split reductive group, and (X ⊃ R,X∨ ⊃ R∨) thecorresponding root datum.
90
Recall that the Deligne–Langlands conjecture tells us that we should expect the followingrelationships:
TRUM repsof G(K)
/'
TRUM Weil-Deligne reps
/'
irred H(G(K), Iw)-
modules
/'
(s, x) ∈ G∨C × g∨Cs.t. s is ss, x nilpand sxs−1 = qx
/conj
finite:1
∼ =
The following theorem relates the extended affine Hecke algebra of Iwahori–Matsumoto tothis story.
Theorem 12.6. (Iwahori–Matasumoto)
1. G(K) =⊔w∈Wext
Iw · w · Iw (“Bruhat decomposition’)
2. There is an isomorphism of algebras
Hext ⊗Z[v±1] C∼−→ H(G(K), Iw)
where C is a Z[v±1]-algebra via v 7→ (√q)−1 ∈ R+ ⊂ C.
Moreover, under 2., Hx is mapped to the indicator function on Iw · x · Iw, up to a scalar.
Example 12.7. Let G = GLn, and fix a uniformizer π ∈ O. Then
Wext =
⟨permutation
matrices
⟩“finite part”
n
⟨πλ1 · · · 0
0. . . 0
0 · · · πλn
⟩ .“lattice part”
By Theorem 12.6, we can understand TRUM representations of G(K) by studying ir-reducible Hext-modules. Denote by Z := Z(Hext) = (Z[v±1][X∨])Wf . By Quillen’s Lemma(which is an infinite-dimensional version of Schur’s Lemma), Z acts by scalars on any ir-reducible Hext-module. The Bernstein presentation tells us that Hext is finite over R :=Z[v±1][X∨]. Since R is also finite over RWf , we conclude that any irreducible Hext-moduleis finite-dimensional, and, in fact, is of dimension ≤ |Wf |2.
Hence, the Deligne–Langlands conjecture predicts the following relationships.irreps/Cof Hext
(s, x) ∈ G∨C × g∨Cs.t. s is ss, x nilpand sxs−1 = qx
/'
irreps of Z
pairs (s′, v) ∈ G∨C × C×/G∨C conj (s, x) ∈ G∨C × C×/G∨C conj
finite:1
central character
=
91
The dashed arrow should match v−1 ↔ √q.
Remark 12.8. 1. We are no longer forced to take q = |k|.
2. In the diagram above, we are repeatedly using the fact from last semester that
characters χ : Z[X∨]Wf → C ↔ semisimple elts of G up to conjugacy.
An easy and interesting case: Consider the case when s = id and q = 1. In this case,the right side of the dashed arrow is
x ∈ g∨C nilpotent/G∨C conjugacy = “nilpotent orbits”.
This provides a hint that we should not expect to have a good algebraic grip on the problemof understanding the irreducible representations of Hext, because nilpotent orbits are com-plicated and not combinatorial in general. For the next lecture and a half, we will dive intothis geometry.
12.3 Geometric setting
For notational convience, we will temporarily swap G ↔ G∨ in this section. Let G/C be acomplex reductive group, and N ⊂ LieG =: g the nilpotent cone.
Remark 12.9. For GLn, it is tempting to define N as the variety
x ∈ gln(C) | xn = 0.
However, this results in a non-reduced scheme, because the ideal corresponding to the equa-tion xn = 0 is not radical. A better definition is to consider
x ∈ gln(C) | coeffients of the characteristic polynomial vanish.
This still captures what we know as a nilpotent matrix, but results in a better geometricobject. (In particular, it is a reduced scheme.)
In general, we define the nilpotent cone as follows. Consider the adjoint quotient map
gq−→ g/G = h/W.
The equality g/G = h/W follows from Chevalley’s theorem. For gln, the map q is “takecoefficients of the characteristic polynomial,” so this captures what we wanted in Remark12.9. We define
N = q−1(0).
Fundamental facts about the nilpotent cone:
1. N is irreducible, reduced, and normal.
2. G has finitely many orbits on N , and all are even-dimensional (over C).
92
Example 12.10. If G = GLn, thennilpotentmatrices
/conj
↔
Jordannormalform
↔
partitionsλ ` n
Moreover, if we denote by Oλ the orbit corresponding to the partition λ, then
Oµ ⊂ Oλ ⇐⇒ µ ≤ λ in dominance order.
We can examine these orbit relations explicitly for small n.
n = 2:
N =
x =
(a bc d
)| Tr x = detx = 0
=
x =
(a bc −a
)| detx = −a2 − bc = 0
⊂ C2.
A R-picture of this is:
n = 3: For n = 3, the picture is more of a caricature.
Exercise 12.11. Try to do this for large n (perhaps 7 or 8) and see that this poset is ratherugly; in particular, it is not graded.
93
12.4 The Springer resolution
We can study these nilpotent orbits by using a resolution of singularities of the nilpotentcone. Let B be the variety of Borel subalgebras in g. For gln, this is the variety of completeflags in Cn.
Definition 12.12. The Springer resolution is the map
N = (x, b) ∈ N × B | x ∈ b ' T ∗B
N
which sends (x, b) 7→ x.
Next week we will study the Springer resolution more carefully, and show that it is proper,smooth, and a resolution of singularities. For GLn,
N = (x, 0 = V0 ⊂ V1 ⊂ · · ·Vn = Cn) | x preserves flag ⇐⇒ xV i ⊂ V i−1.
Example 12.13. We examine GLn for small n again.
n = 2: The variety of Borel subalgebras is B = P1C = lines in C2, and the Springerresolution looks like:
n = 3: The variety of Borel subalgebras is B = (`, P ) | ` ⊂ P ⊂ C2 and a caricature ofthe Springer resolution looks like:
94
We will explain in more detail how we arrived at this picture next week.
95
Lecture 13: Springer fibres and the Steinberg variety
13.1 The Springer resolution, continued
Today we are going to pick up where we left off last week and continue to discuss the geometryof nilpotent orbits. To begin, we will discuss in more detail our claim from last week thatN ' T ∗B.
Let G be an algebraic group and g its Lie algebra. Canonically, we can express thetangent bundle of G as
TG = G× g,
since the Lie algebra can be identified with Te(G). Hence if X is a homogeneous space forG (i.e. G acts on X transitively), then we have a surjection
TG = G× g→ TX,
and for x ∈ X, TxX = g/Lie (stabG x). Similarly, there is a canonical identification of thecotangent bundle of G
T ∗G = G× g∗
and for a homogeneous space X,T ∗X → T ∗G.
Moreover, for x ∈ X, TxX = (Lie (stabG x))⊥ ⊂ g∗.Now assume that G is semisimple and let B be the variety of Borel subalgeras of g. Once
we choose a Borel subalgebra B ⊂ G, we can identify B ' G/B since stabG b = B. Lastlecture we introduced the following space
N = (b, x) ∈ B ×N | x ∈ b.
Claim 13.1. There is a canonical isomorphism N ' T ∗B.
Proof. For a point b ∈ B, the tangent space to B at b is
TbB = g/(Lie (stabG b) = b).
Hence,T ∗B = (b, v) ∈ B × g∗ | v ∈ b⊥.
Recall the Killing form κ : g × g → C, κ(x, y) = tr(adx ad y). This is a symmetric,nondegenerate bilinear form. Fix a Cartan subalgebra h ⊂ b, and we obtain a triangulardecomposition g = n− ⊕ h ⊕ n+, with b = h ⊕ n+. The restriction κ|h×h is nondegenerate,and κ gives a nondegenerate pairing
κ : n− × n+ → C.
Hence under the identification g ' g∗ via the Killing form, b⊥ ⊂ g∗ corresponds to n+ ⊂ g.We conclude that
T ∗B = (b, x) ∈ B × g | x ∈ b is nilpotent= (b, x) ∈ B ×N | x ∈ b= N .
96
13.2 Many examples of Springer fibres
“The richness of Springer fibres cannot be underestimated.”
Recall the Springer resolution
N = (x, b) ∈ N × B | x ∈ b ' T ∗B
N
πs
Definition 13.2. Let x ∈ Oλ ⊂ N be a point in a G-orbit. The associated Springer fibreFλ is
Fλ := π−1s (x).
By equivariance, this is independent of the choice of x ∈ Oλ, up to isomorphism.
Let G = SLn. This section is devoted to studying Springer fibres for n = 2, 3, 4. Recallthat for G = SLn,
N = (F ·, x) ∈ complete
flags in Cn×N | x preserves F · = (F ·, x) | xF i ⊂ F i−1.
Example n = 2: The nilpotent cone is the quadric cone
N =
(a bc −a
)∈ sl2 | −a2 − bc = 0
⊂ C3.
A R-picture:
97
A table of orbits:
λ
dimOλ 2 0codimOλ 0 2
Springer fibre pt P1Cdim fibre 0 1
Example n = 3: A caricature:
We’d like to construct a table as we did in the previous example, but to do so, we need todetermine the dimension8 of the nilpotent orbits. This is easy for the regular nilpotent orbitand the zero orbit. But what about the other orbit Oλ? Well we know that the matrix
e =
0 1 00 0 00 0 0
is in the orbit, and we observe that
Lie (stabSL3 e) = stabLie SL3 e = ker ad e.
A computation shows that
ker ad e =
a b c
0 a 00 d −2a
| a, b, c, d ∈ C
8If you’d like to understand some variety, it is a good idea to know its dimension!
98
is 4-dimensional, hence the orbit Oλ is (8− 4 = 4)-dimensional. We also need to determinethe Springer fibre corresponding to this orbit. We have that
Fλ = (L ⊂ P ⊂ C3) | eC3 ⊂ P, eP ⊂ L, and eL = 0,
but what does this look like? Well we know that
0 ⊂ ime ⊂ ker e ⊂ C3,
and the condition that a flag in Fλ must satisfy eL = 0 implies that L ⊂ ker e. We see thatthere are two possibilities for flags in Fλ:
• L = ime, free choice of P , as long as L ⊂ P ' P1C
• P = ker e, free choice of L, as long as L ⊂ P ' P1C
These two cases intersect when L = ime and P = ker e, which is a single point, so ourSpringer fibre looks like two P1C’s joined at a point. With this, we can complete our tableof orbits:
λ
dimOλ 6 4 0codimOλ 0 2 6
Springer fibre pt Bdim fibre 0 1 3
Example n = 4: Drawing pictures is no longer so reasonable, but we can still count dimen-sions and make our table. The dimension of B is the number of positive roots (which forSL4 is 6), so we know that dimT ∗B = 12.
To find the dimensions of the Springer fibres, we can play a similar game to what we didin the previous example. Consider the orbit Oλ corresponding to
λ =
The matrix
e =
0 1 0 00 0 0 00 0 0 10 0 0 0
is in Oλ, and we have that
0 ⊂ ime = ker e ⊂ C4.
The Springer fibre is
Fλ = (L ⊂ P ⊂ H ⊂ C4) | eC4 ⊂ H, eH ⊂ P, eP ⊂ L, eL = 0.
Again, we have two possibilities:
99
• (“easy component”) P = ime = ker e, free choice of L,H as long as L ⊂ P ⊂ H 'P1C× P1C =: Σ1
• (“hard component”) H = e−1L =: Σ2. There is a natural map Σ2 → P1C sending(L ⊂ P ⊂ H) 7→ L, and the fibre over L′ is (L′ ⊂ P ⊂ e−1L′ = H) ' P1C, so Σ2 isa P1C-bundle over P1C.
The diagonal ∆ ⊂ Σ1 embeds into Σ2 as the zero section, so these two components are gluedtogether along a P1C.
We’ve established that the second component Σ2 is a P1C-bundle over P1C, but whichone?
Claim 13.3. Σ2 = P(O(1)⊕O(−1)) is the “second Hirzebruch surface”.
For justification of this claim, see Geordie’s hand-written notes. We finish this exampleby constructing our table.
λ
dimOλ 12 10 8 6 0codimOλ 0 2 4 6 12
Springer fibre pt Σ1 ∪P1C Σ2 BSL3 ∪ 2 other comp. BSL4
dim fibre 0 1 2 3 6
Now that we’ve constructed four tables, we can make some observations about patternswe see.
Observations:
• 2 dimFλ = codimOλ
• For any λ, one (“easy”) component of Fλ is isomorphic to a flag variety of a smallergroup.
• Fibres are equidimensional9 and components appear to be smooth10.
Remark 13.4. An audience member “Mr. Wiggins” also observed that in our examples,dimT ∗B is divisible by all of the codimOλ, a property we coined “Wiggins divisibility”.(Though we are not sure if this is a general phenomenon...)
Many of our observations hold in general. Here are some fundamental properties of Springerfibers.
9This is really remarkable!10Sadly this fails in bigger examples.
100
Theorem 13.5. Let G be an a semisimple algebraic group.
1. The Springer resolution πs : N → N is a G-equivariant, projective resolution ofsingularities, and is an isomorphism over Nreg = x ∈ N | dimZGX = rankg.
2. For a nilpotent orbit Oλ ⊂ N , the corresponding Springer fibre Fλ := π−1s (x), x ∈ Oλ
is equidimensional, and dimFλ = 12
codim(Oλ ⊂ N ).
3. Hodd(Fλ,Z) = 0.
Remark 13.6. 1. Heven(Fλ,Z) is well-studied (for example, its Betti numbers are known).
2. Part 2. of the theorem implies that πs is semismall and all strata are relevent.
3. In type A, Fλ have cell decompositions, C0 t (C1)? t · · · . This cell decompositionimplies Part 3. of the theorem for type A. With much more work cell decompositionshave been constructed in other classical types, but existence of cell decompositions isunknown in exceptional types.
4. In type A,
# components of Fλ = dim(irrep of Sn indexed by λ).
In fact, there exists a canonical bijectionstandard Young
tableaux of shape λ
↔
componentsof Fλ
.
13.3 The conormal space
LetX =
⊔λ∈Λ
Xλ
be a stratefied variety. For example, C = C× t 0. Consider the space⋃λ∈Λ
TXλ ⊂ TX.
This is a horrible space! In our example, TC = C× C, and⋃TXλ = (x, y) ∈ C2 | x 6= 0 or x = 0, y = 0.
Here’s a picture:
101
In contrast to this, the conormal space
T ∗ΛX :=⋃λ∈Λ
T ∗λX ⊂ T ∗X,
where T ∗λX = ξ ∈ T ∗xX for x ∈ Xλ | ξ vanishes on TXλ is very nice. In our example ofC = C× t 0, T ∗ΛC = (x, y) ∈ C2 | xy = 0:
Properties of T ∗ΛX:
1. T ∗ΛX is a closed subvariety of T ∗X.
2. dimT ∗λX = dimXλ + codim(Xλ ⊂ X) = dimX is independent of λ! (So we have a“democracy of strata”.)
3. The components of the conormal space are in bijection with the strata.
The conormal space is a fundamental object in microlocal geometry.
Warning: the intersection pattern of T ∗λX may be very complicated.
An important object in our story arises as a conormal variety, the Steinberg variety.
102
13.4 The Steinberg variety
Let H be a group that acts on a variety X on the right, and a variety Y on the left. We canform the balanced product
X ×H Y := X × Y/(xh, y) ∼ (x, hy).
This space may not exist as a variety in general, but in all examples we will encounter, itdoes.
Choose a Borel subgroup B ⊂ G, so B ' G/B. Then consider the variety
G×B G/B∼−→ G/B ×G/B
(g, g′B) 7→ (gg′B, g′B)
The set of G-orbits on G×B G/B is equal to the set of B-orbits on G/B, which is parame-terized by W by the Bruhat decomposition. So we have a stratification
B × B =⊔x∈W
Ox.
In type A, this stratification is given by “pairs of flags in relative position x”.
Example 13.7. There is a stratification of P1C× P1C given by
P1C× P1C = ∆ t (P1C× P1C)\∆.Pairs of flags in ∆ are in relative position id (i.e. they are equal), and flags in (P1C×P1C)\∆are in relative position s.
Remark 13.8. The variety of Borel subalgebras B does not depend on any choices, so theproduct B × B does not depend on any choices. Since we can define the Weyl group as theset of G-orbits on B×B, this gives us a canonical definition of the Weyl group that does notdepend on a choice!
Definition 13.9. The Steinberg variety
St = (b, b′, x) ∈ B × B ×N | x ∈ b, x ∈ b′is given by the fibre product
St
N N
Nwhere the maps are
(b, b′, x)
(b, x) (b′, x)
x
103
Exercise 13.10. (Solution can be found in Geordie’s hand-written notes.) Show that theSteinberg variety St is the conormal variety of
B × B =⊔x∈W
Ox.
Corollary 13.11. There is a canonical bijectionirred. comp.
of St
↔ W.
To finish this lecture, let us recall a remarkable geometric construction of the Robinson-Schensted correspondence, due to Steinberg. Geordie is given to understand that this is theorigin of the name “Steinberg variety”. Fix λ, and two components C,C ′ of Fλ. We havethe following diagram
π−1s (Oλ) Oλ × C
Oλ Oλ
The set (x, c, c′) | x ∈ Oλ, c ∈ C, c′ ∈ C ′ is a subvariety in St of dimension
dimOλ+dimC+dimC ′ = dimOλ+1
2codim(Oλ ⊂ N )+
1
2codim(Oλ ⊂ N ) = dimN = dimB × B.
The components of St all have dimension dimB × B, so we have a bijection
components of St ↔ (Oλ, C, C ′) as above.
In type A, this is a bijection between
W ↔ (λ, T, T ′) | T, T ′ are standard Young tableaux of shape λ.
This is the Robinson-Schensted correspondence!
104
Lecture 14: Springer correspondence and Borel-Moore
homology
14.1 The Springer correspondence
We pick up in the setting of the last lecture, the Springer resolution:
N ' T ∗B Fx := f−1(x)
N ⊂ g x
πs
Roughly, the Springer correspondence states that W acts on H∗(Fx;Q), and one obtainsall irreducible representations of W in top cohomology. Note that this does not come froman action of W on Fx! This is in contrast to Deligne-Lusztig theory and other settings wherewe obtain representations of a group in the cohomology of varieties on which the group acts.In the first part of today’s lecture, we’ll work towards a precise statement of the Springercorrespondence.
Remark 14.1. In type A, the Springer correspondence explains why irreducible represen-tations of the symmetric group and nilpotent orbits are both classified by the same combi-natorial data (Young diagrams).
Grothendieck-Springer alteration: Let
g = (x, b) ∈ g× B | x ∈ b.
In type A, this is equal to
(x, F ) ∈ g×F`ags | xF i ⊂ F i.
Remark 14.2. Note that g ⊂ B× g, and its dual is N = T ∗B. (This is because the dual ofx ∈ b is x ∈ b⊥ = n.)
The key diagram is the following:
N g gr.s.
N g gr.s.
πs
⊂πG
⊃
⊂ ⊃
where gr.s. = x ∈ g | x regular semisimple. Note that g 6= N ∪ gr.s..
Theorem 14.3. 1. The map πs is semismall.
2. The map πG is small (i.e., πG∗kg[dim g] = IC(gr.s.,L)).
105
3. Over gr.s., πG is a W -torsor. (Here W is the Weyl group of G.)
For example, in type A,
gr.s. = x ∈ g | x is semisimple with distinct eigenvalues.To give a flag F preserved by x is equivalent to an ordering of the eigenvalues of x. This isa Sn-torsor.
Now we are ready to state the Springer correspondence precisely. Given x ∈ N , let AG(x)be the component group of the centralizer: AG(x) := CG(x)/CG(x).
Theorem 14.4. (Springer correspondence)
1. πs∗QN [dim N ] =⊕x∈N/G,
ρ∈IrrAG(x)
Htop(Fx)ρ ⊗ IC(G · x,Lρ).
2. (**Most important**) End(πs∗QN [dim N ]) = End(πG∗Qg[dim g]) = QW .
3. Htop(Fx)ρ are all irreducible representations of W .
Remark 14.5. It is not difficult to see that in fact 1. and 3. are consequences of 2.
Remark 14.6. Part 2. of the theorem is true over Z, and indeed over any ring. But part1. fails over arbitrary rings because the decomposition theorem fails.
How might we approach part 2.? Three possible approaches:
(a) Borho - MacPherson:
• First note that because g → g is small, the fact that End(πG∗Qg[dim g]) = QWis obvious. Indeed,
End(πG∗Qg) = End(πG∗Qg|reg.ss.) = EndQW (QW ) = QW,with the first equality following from smallness.
• Then BM point out that we have a homomorphism
End(πG∗Qg)r−→ End(i∗πG∗Qg) = End(πs∗QN )
where i : N → g; and, miraculously, r is an isomorphism. (Proof sketch: Themap r is injective because the action of W on H∗(B) is faithful, then comparedimensions.)
(b) Fourier transform: (Springer’s original approach) Because N and g are dual in B×g,kN and kg are Fourier transforms of one another. This implies that the endomorphismrings are equal:
End(πs∗kN ) = End(πG∗kg) = kW.
(c) Convolution algebras: (in Chris-Ginzburg)
The next part of the lecture will explain this approach. For impatient readers, we’llgive away the ending:
The punchline: There is a notion of homology (Borel-Moore homology, HBM) suchthat End(πs∗kN ) is canonically equal to HBM(Steinberg variety). This gives a concreterealisation of this endomorphism algebra.
106
14.2 Borel-Moore homology
Let X be an algebraic variety, k a field, p : X → pt, and kX the constant sheaf on X. Hereare four notions of (co)homology:
• H∗(X; k) = H∗(p∗kX) - cohomology (cochains)
• H∗c (X; k) = H∗(p!kX) - cohomology with compact support
• H∗(X; k) = H−∗(p!ωX) - homology (chains)
• HBM∗ (X; k) = H−∗(p∗ωX) - Borel-Moore homology (locally finite chains)
Here ωX is the dualising sheaf, DkX . The canonical example is the following.
Example 14.7. Let X = z ∈ C | 0 < |z| < 1. Then
H0(X;Z) = Z , H1(X;Z) = Z , H2(X,Z) = 0.
If we compute Borel-Moore homology, we get
HBM0 (X;Z) = 0, HBM
1 (X,Z) = Z , HBM2 (X;Z) = Z “fundamental class”.
In the first computation (HBM0 ), the formal generator of H0 is now a boundary, so we get
0. In the second computation (HBM1 ), we can now have cycles from the edge to the center,
and these are not boundaries. In the third computation (HBM2 ), the “fundamental class” is
a triangulation of X.
From now on in this course, H∗ = HBM∗ .
Key properties of Borel-Moore homology:
1. If Xi−→M is a closed embedding of X into a smooth, C-dimension d variety M , then
we have local Poincare duality: ωM ' kM [2d]. Hence
H∗(X; k) = H−∗(X,ωX)
= H−∗(X, i!ωM)
= H−∗(M, i!i!kM [2d])
= H2d−∗(M, i!i!kM)
= H2d−∗(M,M\X; k).
2. H∗(−) is not functorial for arbitrary maps, but for proper maps p, we have p∗.
107
3. For an open inclusion U → X, we have restriction
H∗(X)→ H∗(U).
(Thought exercise: Why? In terms of chains?)
4. If X is equidimensional of dimension d with components X1, X2, . . . , Xm, then
H2d(X) =⊕
Z[Xi],
where [Xi] are the fundamental classes of the components.
14.3 Geometric convolution algebras
Let X1, X2, X3 be smooth varieties of dimensions d1, d2, d3, respectively, and
Z12 ⊂ X1 ×X2, Z23 ⊂ X2 ×X3
closed subvarieties (“correspondences”). (For example, we could take Z12 = graph(f) forsome f : X1 → X2.) Define
Z12 Z23 := (x1, x3) ∈ X1 ×X3 | there exists x2 ∈ X2 s.t. (x1, x2) ∈ Z12, (x2, x3) ∈ Z23.
(So in our example, graph(f) graph(g) = graph(g f).) We have projections
X1 ×X2 ×X3
X1 ×X2 X1 ×X3 X2 ×X3
p12p13
p23
We make the following properness assumption: From now on, assume that the map
p13 : p−112 (Z12) ∩ p−1
23 (Z23)→ X1 ×X3
is proper.
Definition 14.8. We define a convolution product on homology:
Hi(Z12)×Hj(Z23)→ Hi+j−2d2(Z12 Z23)
(c12, c23) 7→ p13∗(p∗12c12 ∩ p∗23c23)
Here ∩ is the intersection product in Borel-Moore homology.
Most important cases: Let Xf−→ X be a proper map of a smooth variety X to X, and
let Xi = X for i = 1, 2, 3. Here are two cases of the construction above:
1. Let Z12 = Z23 = Z12 Z23 = X ×X X ⊂ X × X. Then the convolution product givesan associative algebra structure on H∗(X ×X X) (with messy gradings).
108
2. Let Z12 = X ×X X, Z23 = f−1(X), Z12 Z23 = Z23. Then the convolution productgives us a map
H∗(X ×X X)×H∗(f−1(X))→ H∗(f−1(X)).
This gives H∗(f−1(X)) the structure of a H∗(X × X)-module.
So whenever we have a smooth proper map into our variety, we obtain from this formalisman associative algebra and a collection of modules (one for each fibre) over that algebra.Seems promising!
Conceptual explanation of convolution algebras: Let f be as above, with the fibreproduct diagram
X ×X X X
X X
g
g
f
f
Claim: End·(f∗kX) = H∗(X ×X X)
Proof. We will only check the statement on the level of vector spaces. Let dX := dimCX.We have
Hom·(f∗kX , f∗kX) = Hom·(f ∗f∗kX , kX)
= Hom·(f ∗f!kX , kX)
= Hom·(g!g∗kX , kX)
= Hom·(kX×XX , g!kX)
= Hom·(kX×XX , g!ωX [−2dX ])
= H∗−2dX (X ×X X, ωX×X)
= H2dX−∗(X ×X X).
Here we are using properness of f (second equality), adjunctions (first equality, fourth equal-ity), proper base change (third equality), and local Poincare duality (fifth equality).
The Upshot: Up to gradings,
H∗(X ×X X) = End·(f∗kX),
and with some work we can show that multiplication matches on both sides. Similarly,
(f∗kX)x = H∗(f−1(X)),
and the module structure comes from the action End·(f∗kX) (f∗kX)x.
Connection to the Spring correspondence:
109
Let N πs−→ N be the Springer resolution, and
St := N ×N N =conormal space tothe G-space B × B =
⋃x∈W
T ∗x ,
the Steinberg variety. (See lecture 13.) Here T ∗x is the conormal bundle to the G-orbitOx ⊂ B×B. Recall that St is equidimensional and all components have dimension equal toN := dimB × B = dimT ∗B.
The convolution product in Borel-Moore homology gives us a map
H2N(St)×H2N(St)∗−→ H4N−2 dimT ∗B(St) = H2N(St).
This gives H2N(St) the structure of an algebra!
Theorem 14.9. As algebras,H2N(St) = kW.
By the theory earlier, the action of kW on Hi(Fx) yields the Springer action. In this waywe end up with W -modules everywhere!
110
Lecture 15: The Kazhdan–Lusztig isomorphism
15.1 More convolution algebras
Last week we ended with a discussion on convolution algebras. We’ll start today by contin-uing this discussion. Consider the following two settings:
1. Let G be a group and H ⊂ G a subgroup. With the operation ∗ of convolutionof functions, the vector space Fun(G,C) of complex-valued functions on G has thestructure of an algebra. This is just the group algebra, Fun(G,C) ' C[G]. Thesubspace FunH×H(G,C) of H-biinvariant functions forms a subalgebra. This is a Heckealgebra. As many of us are aware, its representation theory is very complicated ingeneral!
2. Now let X be a finite set. The vector space Fun(X×X,C) of complex-valued functionson X×X can also be given the structure of an algebra. In this setting, the convolutionproduct is given as follows. Let
X ×X ×X
X ×X X ×X X ×X
p12p13
p23
be the natural projections. Then for f, g ∈ Fun(X×X,C), define f ∗g ∈ Fun(X×X,C)by
(f ∗ g)(x, z) :=∑y∈X
f(x, y)g(y, z)
for (x, z) ∈ X ×X. In other words,
f ∗ g = p13!(p∗12f p
∗23g).
With ∗, (Fun(X ×X,C), ∗) is an algebra. In fact, it’s a familiar11 algebra. Let ex,y bethe indicator function on (x, y) ∈ X ×X. Then ex,y ∗ ey′,z = δy,y′ex,z. Hence,
(Fun(X ×X,C), ∗) ' MatX×X(C).
Some Variants:
(a) Let X → Y be a map of sets, then we can construct the convolution algebra
Fun(X ×Y X,C) '∏y∈Y
Matf−1(y)×f−1(y)(C).
We can formulate Mashke’s theorem in this language. For a finite group G, letY = irreps of G, X =
⊔ρ∈Y basis of ρ, and X → Y the map which assigns to
a basis element the corresponding representation. Then Mashke’s Theorem thatthe group algebra is semisimple is the following statement:
11This is why we don’t meet this algebra as often as we meet the Hecke algebra - it’s “too easy,” in thesense that it is just a matrix ring. However, when we categorify, it becomes more interesting, so we are morelikely to encounter it (or regognize it!) in categorified settings.
111
Theorem 15.1. (Mashke’s Theorem)
(Fun(G,C), ∗) ' (Fun(X ×Y X,C), ∗).
This realizes a complicated algebra (the group algebra) in terms of much simplerpieces (matrix rings).
(b) If our sets X, Y come with the additional structure of an action by a group Γ,and X → Y is a Γ-equivariant map of sets, then we can construct the convolutionalgebra
FunΓ(X ×Y X,C).
The Upshot: There are two types of convolution algebras, one is hard (Hecke alge-bras), and the other is easy (Fun(X × X,C)). The hard one is of great significancein representation theory. A strategy that we use in representation theory is to try torealise the hard type as the easy type.
15.2 Equivariant K-theory
Let G be an algebraic group acting on a variety X. Then we can formulate the notion of anequivariant coherent sheaf,
F ∈ CohG(X),
as a coherent sheaf F on X, coupled with some extra data12 Roughly speaking, one canthink of an equivariant coherent sheaf on X as being an equivariant sheaf together with analgebraic action on its sections.
Remark 15.2. If F is locally free, then F corresponds to a vector bundle V → X on X.In this setting, G-equivariance of F corresponds to an algebraic G-action on V which iscompatible with projection. In particular,
CohG(pt) ' RepG,
where RepG is the category of algebraic representations of G.
Define two types of equivariant K-groups:
KG(X) := Grothendieck group of G-equivariant coherent sheaves on X;
KG(X) := Grothendieck group of G-equivariant vector bundles on X.
Remark 15.3. (a) Very loosely,
KG(X)↔ “Borel–Moore homology,”
KG(X)↔ “cohomology”.
12For an excellent description of this construction, see notes from Emily’s Sept 20, 2019 talk in the InformalFriday Seminar. Notes from Emily’s (and all other) IFS talks can be found at https://sites.google.com/view/ifssydney/home.
112
(b) There are higher K-groups, KGi (X), Ki
G(X), which we will ignore here. (AlreadyKG
0 (X) = KG(X) and K0G(X) = KG(X) are rich enough.)
(c) There is a natural mapKG(X)→ KG(X).
If X is smooth, then we can use resolutions by coherent sheaves to show that this mapis an isomorphism.
(d) We have to be careful with functors between K-groups. For example, if X → pt isprojection and X is affine, then f∗OX = Γ(X,OX) is usually infinite dimensional. Soin general, we need to work to justify that functors we wish to use preserve Coh, or atleast Db(Coh).
As we did for H∗, we have a convolution product in K-theory. We imitate the set-up ofthe previous lecture: Let X1, X2, X3 be smooth varieties of dimensions d1, d2, d3, respectively,and
Z12 ⊂ X1 ×X2, Z23 ⊂ X2 ×X3
closed subvarieties, with projections pij, i, j = 1, 2, 3 as before. Given F ∈ Coh(Z12),G ∈Coh(Z23), define
F ∗ G := p13∗(p∗12F
L⊗ p∗23G) ∈ Db(Coh(Z12 Z23)).
As we emphasized in Remark 15.3, we need to justify that this product is well-defined.But sure enough, the push-forward p13∗ preserves coherence because p13 is proper, and thepull-backs p∗12, p
∗23 are okay because p12 and p23 are flat. Hence ∗ induces a product
K0(Z12)×K0(Z23)→ K0(Z12 Z23)
which descends to a product
KG(Z12)×KG(Z23)→ KG(Z12 Z23).
Example 15.4. Let X be a finite set, and X → pt projection to a point. In this case,elements of Coh(X ×X) are “matrices of vector spaces,” and
∗ : Coh(X ×X)× Coh(X ×X)→ Coh(X ×X)
is “multiplication of matrices”. That is, for vector spaces Vij,(V11 V12
V21 V22
)∗(V ′11 V ′12
V ′21 V ′22
)=
(V11 ⊗ V ′11 ⊕ V12 ⊗ V ′21 V11 ⊗ V ′12 ⊕ V12 ⊗ V ′22
V21 ⊗ V ′11 ⊕ V22 ⊗ V ′21 V21 ⊗ V ′12 ⊕ V22 ⊗ V ′22
).
The K-group is K0(X ×X) = Fun(X ×X,Z).
Remark 15.5. Lusztig noticed that certain categories of the form CohΓ(X ×X), where Γis a group acting on a finite set X, are central to the classification of unipotent charactersheaves.
113
15.3 The Kazhdan–Lusztig isomorphism
With this we have enough machinery to state the Kazhdan–Lusztig isomorphism. Let (X ⊃R,X∨ ⊃ R∨) be a root datum, Wext = Wf nZX∨ the extended affine Weyl group, and Hext
the affine Hecke algebra. (See lecture 11 for a refresher on these objects.) Let G∨ be thecorresponding algebraic group, N ∨ ⊂ g∨ the nilpotent cone, and St∨ := T ∗B∨×N∨ T ∗B∨ theSteinberg variety. (See Lecture 13.)
Theorem 15.6. (Kazhdan–Lusztig) We have canonical isomorphisms.
Hext KG∨×C×(St∨)
ZWext KG∨(St∨)
∼
v=1forget
C×-action
∼
Here C× acts on T ∗B∨ by dilation along the fibres.
Here are two examples of how Theorem 15.6 gives us insight into the representationtheory of affine Hecke algebras.
1. (Bernstein) As we have discussed, there is an inclusion of algebras:
Z[v±1][X∨] ⊂ Hext
λ 7→ Hλ
Bernstein noticed that the center of Hext can be realized as Wf -invariants of thissubalgebra:
Z[v±1][X∨]Wf = Z(Hext).
We can see this in terms of the Kazhdan–Lusztig isomorphism. Recall that for amap X → Y between finite sets, we have
X X ×X X
X Y
diagonal
and Fun(diag,C) ⊂ Fun(X ×Y X,C) are “diagonal matrices”. Applying this to thediagonal T ∗B∨ ⊂ St∨, we can think of
KG×C×(T ∗B∨) ⊂ KG×C×(St∨)
as “diagonal matrices”. Now by homotopy,
KG∨×C×(T ∗B∨) = KG∨×C×(B∨),
and since KB∨(pt) = [RepB∨] = Z[X∨] and KC×(pt) = [RepC×] = Z[v±1],
KG∨×C×(B∨) = KG∨×C×(G∨/B∨) = KB∨×C×(pt) = Z[v±1][X∨].
114
So the subalgebra Z[v±1][X∨] sits inside Hext as “diagonal matrices” in the G∨ × C×-equivariant K-group of St∨. Moreover, KG∨×C×(pt) is clearly central in KG∨×C×(St∨),and this gives the Bernstein center:
KG∨×C×(pt) = (KB∨×C×(pt))Wf = (Z[v±1][X∨])Wf .
2. The Kazhdan–Lusztig basis of Hext leads to the notion of (left, right, two-sided) cells13.Lusztig noticed that the sets
poset of2-sided
cells
↔
nilpotentorbitsin N ∨
appear to match.
He also noticed other remarkable parallels, such as
2-sided cellis finite
⇐⇒reductive part ofthe centralizer
is finite.
These parallels convince one rather quickly that this correspondence is deep.
We can use Theorem 15.6 to understand this observation of Lustig. Consider theprojections
St∨
T ∗B∨ T ∗B∨
N ∨
p
Convolution makes it clear that any closed G∨-invariant subvariety Z ⊂ N ∨ gives riseto a two-sided ideal in KG∨×C×(St∨):
[M ] | suppM ⊂ p−1(Z).13Pictures of 2-sided cells in rank 2 can be found on Lusztig’s webpage: http://www-math.mit.edu/
~gyuri/picture.html.
115
Hence we get a filtration of Hext by nilpotent orbits in N ∨! Lusztig and Bezrukavnikovshowed that the filtration of Hext coming from 2-sided cells agrees with this filtrationcoming from geometry, explaining Lusztig’s observation that the 2-sided cell ordermatches the order on nilpotent orbits.
3. We won’t explain this in detail due to time constraints, but one can show relativelyeasily that Theorem 15.6 implies the Deligne–Langlands conjecture (Section 12.2) onrepresentations of Hext, as long as q is not a root of unity.
15.4 Bezrukavnikov’s equivalence: rough outline
Now we are finally in the position to approach the second half of the title of this course.Here is a rough outline of the flow of ideas that lead to Bezrukavnikov’s equivalence:
• Langlands correspondence: A correspondence between sets satisfying a whole hostof properties, and with many extraordinary consequences in number theory and be-yond.
• Weil: In the function field case, an automorphic form14 can be regarded as a (veryspecial) function
f : BunG(Fq) :=
iso classes ofG-bundles on
smooth curves/Fq
→ C.
• Grothendieck: Functions on a variety should be understood as shadows of sheaves(function-sheaf correspondence). Here are some examples.
– E.g. 1: Characters of GLn(Fq) are shadows (trace of Frob) of certain `-adicsheaves on G, “character sheaves”.
– E.g. 2: Interesting analytic functions should satisfy many differential equations(i.e. should be solutions of a holonomic D-module).
• Drinfeld: The Langlands correspondence should be approached using Grothendieck’sdictionary. At its most basic level, this philosophy asserts that automorphic formsin the function field case should arise as traces of Frobenius of certain sheaves onthe moduli space of G-bundles. At a more sophisticated level, we should expect anequivalence of categories
Db(Sh(BunG))???←→ Db(Coh(LocG)),
i.e. a “geometric Langlands correspondence”. This allows us to work over C, andprovides fertile connections to physics, higher category theory, etc.
14We haven’t gone into what an automorphic form is yet in this course. For a global field K it is a functionof a special form on a certain quotient of the adelic points of a reductive group. In the function field case,Weil realised that this quotient parametrises G-bundles on the corresponding curve. In other words, in thefunction field case an automorphic form can be regarded as a function of the form given above.
116
Remark 15.7. Geometric Langlands is for function fields. The classical Langlandscorrespondence isn’t. So they seem to be living in different worlds, but the hope isthat they are actually related. For many years this appeared to many as a pretty wildidea. However, recently the geometric Langlands program has been shown to haveconsequences that number theorists really care about. For example Fargues showed[Far18] that geometric Langlands for the Fargues-Fontaine curve implies (part of) theLLC for any p-adic group!
• Ginzburg: A small piece of geometric Langlands should be controlled by a categori-fication of the Kazhdan–Lusztig isomorphism
Hext ' KG∨×C×(St∨).
What is sought is a fundamental monoidal category that arises in two Langlands dualways. Recall that the geometric Satake equivalence categorifies Hsph = KG(pt) (The-orem 11.2). Thus this equivalence can be seen as one layer of difficulty beyond thegeometric Satake equivalence.
• Bezrukavnikov: realization of (several) such equivalences:
KG∨×C×(St∨) CohG∨×C×(St∨) “coherent side”
Hext DbIw(G((t))/Iw) “constructible side”
Theorem 15.8. (Bezrukavnikov’s equivalence, most basic version) Let Iw ⊂ G((t)) bean Iwahori subgroup, and Iw0 ⊂ Iw the pro-unipotent radical. There is an equivalenceof monoidal categories(
DbIw0
(G((t))/Iw) , ∗)'(Db CohG
∨×C×(St∨), ∗).
where St∨
= g∨ ×g∨ g∨ and the hat indicates the pro-unipotent completion.
Remark 15.9. There are several aspects of the above that need explanation, hopefullythis will occur over the coming weeks and months!
117
Lecture 16: The constructible side of Bezrukavnikov’s
equivalence
Last lecture we ended by stating (modulo several undefined pieces) Bezrukavnikov’s equiv-alence:
“constructible side” “coherent side”(DbIw0
(G((t))/Iw) , ∗)'(Db CohG
∨×C×(St∨), ∗).
Our goal for today is to motivate the constructible side of this equivalence. To do so,we will explain Grothendieck’s function-sheaf correspondence in slightly more detail than isusually done. The starting place is the Weil conjectures.
16.1 Weil conjectures
The Weil conjectures were a collection of statements (made by Weil while in jail during thesecond world war) about counting the number of points on an algebraic variety over a finitefield. Let X be a smooth projective algebraic variety over Fq. Then we can ask about thenumber of points #X(Fqn) for all n. Weil conjectured that a generating series built out of#X(Fqn) has remarkable properties. He also pointed out that his conjecture would followfrom an interpretation of #X(Fqn) as the trace of an operator on a vector space, by takingtraces of its powers. The Grothendieck–Lefschetz trace formula, which we briefly visited inthe context of elliptic curves in Lecture 4, provides such an interpretation.
Grothendieck–Lefschetz trace formula: Let ` be a prime not dividing q, and X as above(but not necessarily projective). Then
#X(Fqn) = sTr(Frobnq H ic(XFq ,Q`)) :=
∑i
(−1)i Tr(
Frobnq H ic(XFq ,Q`)
). (16.1)
Here H ic(XFq ,Q`) denotes the compactly supported etale cohomology of XFq , the base change
of X to SpecFq. If X/k, then H∗c (Xk,Q`) has a continuous action of Gal(k/k). The symbolsTr in (16.1) stands for “supertrace”. Let’s see what this formula means in examples.
Example 16.1. Let X = P1Fq , then the cohomology and action of Frobenius are given by
the following table:
i H ic(XFq ,Q`) Frob action
2 Q` q1 0
0 Q` 1
Hence,sTr(Frobnq H∗
)= 1 + qn = #X(Fqn).
118
Example 16.2. Let X = Gm. Then the action of Frobenius on cohomology is:
i H ic(XFq ,Q`) Frob action
2 Q` q
1 Q` 10 0
Hence,sTr(Frobnq H∗
)= qn − 1 = #Gm(Fqn) = #F×qn .
Example 16.3. Let X = SpecFq2/ SpecFq. Then
X(Fqn) = Hom(SpecFqn , SpecFq2)
= HomFq-alg(Fq2 ,Fqn)
=
0 n odd,
2 n even.
On the geometric side,
XFq = SpecFq ×SpecFq SpecFq2
= Spec(Fq ⊗Fq Fq2)
= SpecFq t SpecFq' Fq × Fq.
Remark 16.4. For analogy, an easier version of the computation above is the following. Wecan realize
C ' R[x]/(x2 + 1).
Then
C⊗R C = C[x]/(x2 + 1)
= C[x]/(x− i)(x+ i)
' C× C.
The Upshot: The variety XFq consists of 2 points which are interchanged by Gal(Fq/Fq).So our table of cohomology is
i H ic(XFq ,Q`) Frob action
> 0 0
0 Q` ⊕Q`
(0 11 0
)and
sTr(Frobnq H∗
)=
0 n is odd,
2 n is even.
119
The Grothendieck–Lefschetz formula and the examples above can be realized as shadowsof something happening on the level of sheaves. Let k be a field. Then, roughly,
Spec k = pt/Gal(k/k).
In other words, one can think of Spec k as something like the classifying space of its absoluteGalois group. There is a bijection
etale cohomologysheaves on Spec k
with Z/`mZ-coefficients
∼←→finitely generated
Z/`mZ-modules with
an action of Gal(k/k)
Q`-coefficients →finite-dimensional continuous
representations of Gal(k/k)
on Q`-vector spaces
Let F ∈ Dbc(X,Q`) be a object in the bounded derived category of constructible Q`-sheaves
on X. Then for any x ∈ X(Fqn), we get an inclusion
SpecFqni−→ X,
and hence an object i∗F ∈ Dbc(SpecFq,Q`). Applying the supertrace of Frobenius to this
object results in an element of Q`. In this way, we get a map
Dbc(X,Q`)
f−→∏n≥1
Fun(X(Fqn)→ Q`).
If F → G → F ′ +1−→ is a distinguished triangle in Dbc(X,Q`), then f(F) + f(F ′) = f(G). In
other words, f factors through the Grothendieck group:[Dbc(X,Q`)
] f−→∏n≥0
Fun(X(Fqn),Q`).
A consequence of the Chebotarev density theorem (Theorem 1.3) is that this map is injective;i.e. for a given qn, the collection of all of the functions f(F) completely determine the classof F in the Grothendieck group. Now, Grothendieck tells us how to view this relationship.
Grothendieck’s philosophy: Interesting functions X(Fqn)→ C,Q`, etc. should be shad-ows of interesting sheaves.
Lusztig provided us with an extraordinary example of a case where Grothendieck’s phi-losophy is exactly true.
Example 16.5. (Lusztig’s theory of character sheaves) There exists a setFχ ∈ Db
c(GLn,Q`)
such that Fχ yield all irreducible characters χ : GLn(Fqn) → C of all GLn(Fqn) “at once”.(More precisely, for a given Fqn , one should only consider those character sheaves which are“defined over Fq”, however we will not go into the details.)
120
16.2 The Hecke algebra, revisited
In the second half of this lecture, we will describe another example of Grothendieck’s phi-losophy: the Hecke category. To start, we will recall the origin of the Hecke algebra.
Let G be a split reductive group over a finite field Fq and B ⊂ G a Borel subgroup. Wecan define a convolution algebra
HFq =(FunB(Fq)×B(Fq)(G(Fq),C), ∗
)of complex-valued B(Fq)-biinvariant functions on G(Fq). A priori, the structure of thisalgebra seems to depend on q, but Iwahori found a presentation of HFq which dependsalmost only on the Weyl group. Let (W,S) be the Coxeter system associated to B ⊂ G.Iwahori’s presentation has generators Ts | s ∈ S, and relations
T 2s = (q − 1)Ts + q
TsTt · · · = TsTt · · · ,
where the products on the second line are of mst = order(st) generators. In Iwahori’s presen-tation, the element Ts corresponds to the indicator function 1BsB ∈ FunB(Fq)×B(Fq)(G(Fq),C).Iwahori’s presentation demonstrated that the Hecke algebra HFq is “defined over Z[q]”.
The quadratic relation may look a little mysterious at first, but it has a natural geometricorigin.
Origin of the quadratic relation: Take G = SL2(Fq) and B =
(∗ ∗0 ∗
). Then on one
hand,
1G ∗ 1G =pushforward of the constant function on G×B G
under the multiplication map G×B Gm−→ G
But on the other hand, there is a natural isomorphism
G×B G∼−→ G/B ×G
(g, h) 7→ (gB, gh)
and the multiplication map factors through this isomorphism:
G×B G G/B ×G
G
m
∼
proj to G
Hence,
1G ∗ 1G =pushforward of constant function onG/B ×G under the projection to G
= |G/B| · 1G= (1 + q)1G.
121
From this, we deduce the quadratic relation: Since 1G = 1BsB + 1B = Ts + 1, we have
T 2s + 2Ts + 1 = (Ts + 1)2
= (q + 1)(Ts + 1)
= (q + 1)Ts + q + 1
= (q − 1)Ts + q + 2Ts + 1.
Hence T 2s = (q − 1)Ts + q.
The geometric origin of the Hecke algebra suggests a categorification via Grothendieck’sphilosophy. Let G be a split reductive group over Fq, and B ⊂ G a Borel subgroup. Definethe (first incarnation of) the Hecke category to be
H := DbB×B(G,Q`),
the B×B-equivariant derived category of etale Q`-sheaves, in the sense of Bernstein-Lunts.We won’t describe the precise construction of this category, but in its first approximation15,we can consider objects in Db
B×B(G,Q`) to be etale sheaves which are constructible forB ×B-orbits.
The convolution of functions in HFq can be upgraded to a convolution product onsheaves. Let F ,G ∈ Db
B×B(G). We have maps
G×G G×B G G
G G
p1 p2
mult
Then if F ,G ∈ DbB×B(G,Q`), we define
F ∗ G := mult∗(FG) ∈ DbB×B(G,Q`),
where FG ∈ DbB×B(G×BG,Q`) corresponds to resB×B×BB×B×B×B(p∗1F⊗p∗2G) under the equivalence
DbB×B×B×B(G×G,Q`)
∼−→ DbB×B×B(G×B G,Q`).
Example 16.6. Let k = Q`, and G = SL2. Then
kSL2∗ kSL2
= mult∗(kSL2×B SL2) = H∗(P1)⊗ kSL2
.
This is the categorified version of the Hecke algebra equality 1G ∗ 1G = (1 + q)1G that wediscussed earlier!
15Be careful! This approximation is just an approximation and can get you in trouble if you take it tooliterally (see Lecture 24).
122
16.3 Perverse sheaves on P1CThe time has come for an interlude. For the rest of this lecture we will shift gears andconsider the variety P1C with the stratification Λ given by
P1C = 0 t C∞.
Denote by 0 i−→ P1C
j←− C∞ the natural inclusions. It is very important in what is coming
to understand perverse sheaves on P1C with the stratification Λ, so we will discuss this (andgeneralizations) in the upcoming lectures. So what are the perverse sheaves on P1C?
Here’s an algebraic answer (c.f. Emily’s talk on nearby cycles in the Informal FridaySeminar16):
PervΛ(P1C, k) 'V1 V0
c
v| v c = 0
,
where Vi are finite-dimensional k vector spaces representing the nearby cycles at 0 (the vectorspace V1) and the vanishing cycles at 0 (the vector space V0). From this perspective, we cansee that there are five indecomposible objects:
1. 0 k i∗k0, skyscraper at 0
2. k 0 constant sheaf
3. k k∼
0
4. k k0
∼
16All IFS talk notes can be found on the IFS website https://sites.google.com/view/ifssydney/home
123
5. k k ⊕ k “big projective/tilting sheaf”
In the next lecture we will go into more detail and describe perverse sheaves on generalcurves.
124
Lecture 17: Constructible and perverse sheaves on curves
Today we continue our interlude of the previous lecture and discuss perverse sheaves oncurves. We are working toward understanding Beilinson glueing on curves, which will be themain topic of next week’s lecture.
17.1 Constructible sheaves
Let k be a field, fixed throughout, and let X/C be a variety, viewed with the classicaltopology. Let V be a k-vector space. We have the notion of a constant sheaf VX withvalues in V :
VX(U) := f : U → V continuous,
where V is viewed with the discrete topology. This leads to the notion of a local system,which is a locally constant sheaf with finite-dimensional stalks.
Theorem 17.1. If X is connected, there is a bijectionlocal systems
on X
∼←→ Rep(π1(X, x)).
Remark 17.2. The difference between local systems and vector bundles is captured withthe following picture.
Here the pink lines are intended to denote the sections of our local systems/vector bundles.Vector bundles with flat connections are equivalent to local systems. However, local systemsform an abelian category, whereas vector bundles do not.
The notion of a local system leads us to the notion to a constructible sheaf: A sheafF on X is constructible if there exists a stratification X =
⊔λ∈ΛXλ of X by a finite number
of subvarieties Xλ such that F|Xλ is a local system for all λ ∈ Λ. Finally, the notion of aconstructible sheaf leads to the notion of the bounded derived category of constructiblesheaves:
Dbc(X, k) :=
F∣∣∣∣ Hi(F) is constructible for all i, and
Hi(F) = 0 for |i| >> 0
⊂ Db
(sheaves of k-vector
spaces on X
)
125
From now on, assume that X is a connected, smooth curve.
Choose a set of points x1, . . . , xm, and let U denote their complement in X. Fix thestratification Λ:
X = U t x1 t · · · t xm.
Theorem 17.3. There is an equivalence (a fun exercise if the reader is tempted!)
Λ-constructiblesheaves on X
∼−→
L local system on U ,
V1, . . . , Vm finite-dimensional vector spacesmaps φi : Vi → (Lni)
µi
.
Here ni ∈ X is a “nearby point” to xi, and Lni denotes the stalk, which carries a monodromyoperator µi given by a small loop around xi (see the diagram below). Note that the stalkdepends on the choice of nearby point ni, but the invariants in the stalk do not!
17.2 Perverse sheaves
One way to visualize an object F ∈ Dbc(X) is via its table of stalks:
i− 1 i i+ 1U · · · Hi(F|U) · · ·x1 · · · Hi(Fxi) · · ·x2 · · · Hi(Fx2) · · ·...
...
We will use these tables frequently in the remainder of this lecture.The category of perverse sheaves
PervΛ(X) ⊂ Dbc(X)
126
is defined as the heart of the t-structure
D≤0Λ :=
F∣∣∣∣ Hi(F|U) = 0 for i ≥ 0Hi(F|xj) = 0 for i > 0
, D≥0
Λ = D(D≤0Λ ).
Here D denotes Verdier duality. We have two possibilities for what the table of stalks canlook like for a perverse sheaf in PervΛ(X):
fullsupport
−2 −1 0 1U 0 L 0 0xi 0 V W 0
skyscraper−2 −1 0 1
U 0 0 0 0xi 0 0 W ′ 0
A general strategy when trying to understand an abelian category is to try to produceand study exact functors from that category to a well-understood category (like Vect orrepresentations of a group). For perverse sheaves, this strategy proves to be somewhatcomplicated, and brings nearby and vanishing cycles into our world.
Theorem 17.4. (to be explained)
PervΛ(X)∼←→
L local system on U ,
V1, . . . , Vm finite-dimensional vector spaces (“vanishing cycles”),
maps Vi Lniv
us.t. v u = id− µi
Remark 17.5. The category PervΛ(X) is Verdier self-dual. The category of Λ-constructiblesheaves is not:
D(Λ-constructible sheaves) 'φi : Lµini → Vi
Example 17.6. Here are some examples of perverse sheaves.
• The constant sheaf kxj on xj is perverse.
• For a local system L on X, L[1] is perverse. (This follows because DL[1] ∼= L∨[1],where L∨ denotes the dual local system.)
• Let j : U → X and L a local system on U . Then we claim that j!L[1] and j∗L[1] areboth perverse. Because j! is extension by zero, computing the stalks of j!L[1] is easy:
−2 −1 0 1U 0 L 0 0xi 0 0 0 0
∈ D≤0Λ
Computing the stalks of the direct image j∗ := Rj∗ is trickier. Le x ∈ X be a point.Then
H i((j∗L)x) = lim←−B(x,ε)
H i(B(x, ε) ∩ U,L)
=
Lx if x ∈ U,H i(B(x, ε)\x,L) if x 6∈ U.
127
How can we compute the cohomology H i(B(x, ε)\x,L)? The space B(x, ε)\x ishomotopic to S1. We can compute the cohomology of S1 with coefficients in L usingthe two-term complex
Vid−µ−−−→ V.
Hence H0(S1,L) = V µ, the invariants of µ V , and H1(S1,L) = Vµ, the coinvari-ants. (Alternatively, one can see that H0(S1,L) = V µ directly, and conclude thatH1(S1,L) = Vµ by Poincare duality.) Hence, the table of stalks of j∗L[1] is
−2 −1 0 1U 0 L 0 0xi 0 V µ Vµ 0
∈ D≤0Λ .
Because D(j∗L) = j!D(L), we can conclude from these computations that j∗L[1] andj!L[1] are both perverse. In other words, j∗ and j! are exact for the perverse t-structure.
Next we turn our attention to classifying the simple objects in PervΛ(X). Recall thatthere is a natural map
j! → j∗.
Given L a local system on U , define
IC(X,L) := j!∗(L) := Im(j!L[1]→ j∗L[1]).
This is not obvious, but by a construction of Deligne, we have
IC(X,L) = τ≤−1(j∗L[1]),
where τ≤−1 is the truncation functor in the standard (not perverse) t-structure. Hence thetable of stalks for IC(X,L) is
−2 −1 0 1U 0 L 0 0xi 0 V µ 0 0
.
Remark 17.7. We see from the remarks above that when X is a curve, all IC sheaves areshifts of actual constructible sheaves. This is not the case in general!
Example 17.8. Consider a local system L on U such that the monodromy µ does not have1 as an eigenvalue. Then id − µ is invertible. Hence Vµ = V µ = 0. Recall that the generalform of the tables of stalks of j!L[1] and j∗L[1] are
j!L[1]:−2 −1 0 1
U 0 L 0 0xi 0 0 0 0
j∗L[1] :−2 −1 0 1
U 0 L 0 0xi 0 V µ Vµ 0
.
So the fact that V µ = Vµ = 0 implies that the natural map j!L[1]→ j∗L[1] is an isomorphism.We conclude that in this case
IC(X,L) = j!L[1] = j∗L[1].
128
Theorem 17.9. The category PervΛ(X) is a finite-length abelian category with simpleobjects
IC(X,L) | L|U is irreducible ∪ i∗k | i : xi → X .
For any irreducible local system L on U , we have the distinguished triangle
τ≤−1(j∗L[1])→ j∗L[1]→ τ≥0(j∗L[1])+1−→
=
IC(X,L)→ j∗L[1]→ (Vµ)0+1−→
in Dbc(X, k). There is something distinctive about this triangle: all objects are perverse
sheaves! Hence, this is an exact sequence and we’ve found our first composition series. Wecan draw this composition series with an “egg diagram:”
j∗L[1] = (17.1)
Example 17.10. Here is a special case of the construction above. Let D be a disc and
D×j−→ D
i←− 0
the natural inclusions. Then the composition series of j∗kD× [1] is
j∗kD× [1] = .
Example 17.11. Let
Uj−→ X
i←− Z := X\U
be the natural inclusions. Then in Dbc(X, k) we have the distinguished triangle
j!j!kX → kX → i∗i
∗kX+1−→ . (17.2)
By turning triangles, we obtain
i∗kZ → j!kU [1]→ kX [1] = IC(X)+1−→ .
129
All objects in this triangle are perverse sheaves, so this is a composition series! Hence wehave the following egg:
j∗kD× [1] = .
We could also have obtained this by dualizing (17.1).
Remark 17.12. Constructible sheaves on X are not necessarily finite length (see (17.2)).However, perverse sheaves are. This is one of the reasons why we are so fond of them.
Before moving on to glueing, we will do one more fun calculation. We return to the localcase:
D×j−→ D
i←− 0
Let Ln be a local system on D× with monodromy given by a single Jordan block of size n:
Jn =
1 1
1 1
1. . .. . . 1
1
Our goal is to describe j!∗Ln[1] = Im(j!Ln[1] → j∗Ln[1]) in this setting. Recall that thefunctors j! and j∗ are exact because j is affine, and we have the building blocks
j∗kD× [1] = j!kD× [1] = .
130
In local systems on D×, we have
Remark 17.13. In this example, we see that the functor j!∗ preserves injections and sur-jections, but is NOT exact.
17.3 Nearby and vanishing cycles
Let D be the disc, D× the punctured disc, and D× its universal cover. Let
0 i−→ D
j←− D× p←− D×
be the natural maps.
Motivation: We want to define the “stalk” of a perverse sheaf at singular points.
Recall the nearby cycles functor ψf (c.f. Emily’s Oct 18, 2019 IFS talk or Laurentiu’sFeb 21/28, 2020 IFS series): For a local system F on D×,
ψf (F) = i∗j∗p∗p∗F
= i∗j∗p∗(p∗F ⊗ k
D×)
= i∗j∗(F ⊗ p∗kD×).
Here the third equality follows from the projection formula. The sheaf p∗kD× is the “uni-versal local system on D×”. It is an infinite-dimensional local system with stalk k[x±1] andmonodromy x.
Recall that the monodromy µ ψf (F), so we get a decomposition into generalizedeigenspaces:
ψf (F) =⊕
ψλf (F).
By far the most important of these eigenspaces corresponds to λ = 1, the “unipotent nearbycycles”. Indeed, from this, we can recover all others:
ψλf (F) = ψ1f (F ⊗ kλ−1).
131
From now on, setψ(F) := ψ1
f (F).
Let Lunip be the universal unipotent local system. This is the local system with stalkk[[u]] at 1 and monodromy given by multiplication by 1 + u. In other words,
Lunip = lim←−Ln,
where Ln is our local system from earlier with stalk k[u]/(un) and monodromy multiplicationby 1 + u. (Think: “completion of the augmentation ideal” should give the same answer.)
Idea:
1. We should think thatψ(F)“=” lim
←−(i∗(j∗(F ⊗ Ln))).
2. unipotent H0(D×,F) = V µ
H1(D×,F) = VµHow do we recover V ?
We will see in the next lecture that after tensoring with the local system given by a bigJordan block, we can recover V and its monodromy via taking global sections. This appearsto be a basic idea behind Beilinson’s construction.
Hence it makes sense to define
ψ(F) = H1(D×ε ,F ⊗ Ln)
for n large.
132
Lecture 18: Beilinson gluing on curves
Throughout this lecture, we will work in the following setting. Let D be the disc, D× thepunctured disc, and
D×j−→ D
i←− 0
the natural inclusions.Our goal for this lecture is to explain the proof of Beilinson gluing17. We’ll pick up where
we left off last week.
18.1 A fact about unipotent monodromy
Let
e =
(0 10 0
), h =
(1 00 −1
), f =
(0 01 0
)be the standard basis for sl(2,C) = Cf ⊕Ch⊕Ce. Clebsch–Gordan tells us about how thetensor product of two irreducible finite-dimensional sl(2,C)-modules decomposes:
Theorem 18.1. (Clebsch–Gordan) Let Ln, Lm be irreducible finite-dimensional sl(2,C)-modules of highest weight n,m, respectively. Then
Ln ⊗ Lm = L|n−m| ⊕ · · · ⊕ Ln+m−2 ⊕ Ln+m.
A consequence of this theorem is that for n large, there are m+ 1 summands, so
(Ln ⊗ Lm)e =highest weight vectors
in above sum= Cm+1 ' Lm.
(Here the superscript denotes Lie algebra invariants; i.e. the kernel of the action
e · v ⊗ w = ev ⊗ w + v ⊗ ew
of e on Ln ⊗ Lm.) Moreover, this is more than just an isomorphism of vector spaces: theaction of −e⊗1 on (Ln⊗Lm)e aligns with the action of e on Lm. Hence (be exponentiating)we have the following corollary to Clebsch–Gordan’s theorem.
Corollary 18.2. Let V be a finite-dimensional vector space and φ a unipotent endomor-phism of V . Let Jn be a (n+ 1)-dimensional vector space and
φ′ =
1 1
1. . .. . . 1
1
17In addition to Beilinson’s original paper [Bei87], the notes [Mor18] of Sophie Morel are an excellent
reference for this material.
133
a Jordan block of size n+ 1. For n large,
(V ⊗ Jn)φ⊗φ′ ' V,
and this isomorphism is equivariant with respect to the action of 1⊗ (φ′)−1 on the LHS andφ on the RHS.
Hence by tensoring with a Jordan block and taking invariants, we can recover the wholevector space and its monodromy action.
Remark 18.3. (Aside for the Lie theorists) Let g/C be a semisimple Lie algebra, and V afinite-dimensional g-module. For λ sufficiently dominant,
∆λ ⊗ V '⊕γi∈Γ
∆λ+γi ,
where ∆µ is the Verma module of highest weight µ and Γ is the multiset of weights of V .Moreover, there is an isomorphism of n+-modules:
(∆λ ⊗ V )b ' V.
18.2 The unipotent vanishing cycles functor
Now we return to the world of perverse sheaves. A perverse sheafM∈ Perv(D×) is the samething as a vector space V and monodromy endomorphism µ V . Let Ln be the local systemwith stalk k[x]/(xn) and monodromy φ = 1 + x, as in Lecture 17. In light of Corollary 18.2and the third bullet point in Example 17.6, for n large enough, we could define a functor ψby
ψ(M) = H0(j∗(M⊗Ln)) = (V ⊗ Jn−1)µ⊗φ = V.
This will end up being our definition which allows us to see certain features in the argumentsbelow. The key lemma of today’s lecture tells us that this definition stabilizes for large n.
Lemma 18.4. Let M∈ Perv(D×). For a fixed m, the kernel and cokernel of
j!(M⊗Ln)xm−→ j∗(M⊗Ln)
stabilize for large n, and are canonically isomorphic.
Proof. We will prove the lemma whenM has rank 1 (i.e. M corresponds to the vector spacek with monodromy given by multiplication by λ ∈ k×.)
Case 1: monodromy λ 6= 1. Then
j!(M⊗Ln) = j∗(M⊗Ln).
(See Example 17.8.) The theorem follows in this case because the kernel and cokernel of
Lnxm−→ Ln stabilize for m fixed, n large and j!, j∗ are exact.
134
Case 2: monodromy λ = 1; i.e. M = kD× [1] is the trivial local system. We will illustratewhat is happening in this case through an example. Let m = 2 and n = 3. We can see what
the maps L3x2
−→ L3, j!L3x2
−→ j!L3 and j!L3x2
−→ j∗L3 are doing on egg diagrams:
As n gets larger, the eggs in the illustration above get longer and the image of j!∗ gets bigger,but the kernel and cokernel stay the same size.
Remark 18.5. What Beilinson actually proves is that
lim↔j!(M⊗Ln) ' lim
↔j∗(M⊗Ln),
where ‘lim↔’ is appropriately defined. We’ll sum this up with the slogan “middles agree”.We already saw this happening above, and will see it happening again in our next example.
Example 18.6. Here is a 2-dimensional example. Let
C2 f−→xy
C,
and consider PervΛ(C2), where Λ is the stratification via coordinate hyperplanes and their
135
complement U .
Let M = kU [2], and denote by IC1 = ICy=0, IC2 = ICx=0, and IC0 = ICx=y=0. Thecomposition series of j!kU [2] is given by the following egg:
j!kU [2] =
Then the key Lemma can be seen on egg diagrams:
Definition 18.7. (Beilinson) Let Xf−→ D, U = f−1(D×), and Z = f−1(0). For a perverse
sheaf M∈ Perv(U) and n large, define functors ψf ,Ξf : Perv(U)→ Perv(X) by
• ψf (M) = ker(x0) ' coker(x0) “unipotent nearby cycles”
136
• Ξf (M) = ker(x1) ' coker(x1) “maximal extension”
The perverse sheaf ψf (M) is supported on Z and Ξf (M) is supported everywhere.
Example 18.8. Let M = kU [1], then we can see the images of these functors in eggs:
Example 18.9. In our 2-dimensional example, Example 18.6 we have:
Lemma 18.10. The functors ψf and Ξf are exact, and we have functorial short exactsequences
1. j! → Ξf ψf
2. ψf → Ξf j∗
Moreover, the canonical mapψf → Ξf → ψf
agrees with monodromy −1.
Proof. See Geordie’s handwritten notes on the course website.
137
Definition 18.11. Let M ∈ Perv(D). The unipotent vanishing cycles functor isdefined by
φf (M) = H0
Ξf (MU)
j!(MU) ⊕ j∗(MU)
M
−1
All maps in the diagram above are the canonical ones.
Remark 18.12. 1. The map j! → Ξf is injective, and the map Ξf → j∗ is surjective,so the cocomplex above only has cohomology in degree 0, and hence φf is an exactfunctor on Perv(D).
2. x induces a monodromy endomorphism of Ξf , ψf , φf .
Example 18.13. Let X = D and M = j!kD× [1]. To compute φf (M), we need to computethe zeroeth cohomology
.
The image of the first map is
138
The kernel of the second map is
Notice that because of the sign in the second map, the image will always be contained in thekernel. Hence φf (M) is equal to
.
Example 18.14. If M = kD[1], then using the same method as above, we compute
φf (M) = .
Here, the kernel and image are equal:
139
Notice that the extra piece of the kernel in the previous example is now a subsheaf of theimage:
Hence φf (M) = 0.
Exercise 18.15. Show thatφf (Ξf (kD× [1])) = IC⊕2
0 ,
and show that the monodromy endomorphism is not trivial.
Example 18.16. Let M = ICX , for X = C2, and C2 f−→xy
C as earlier. Then
φf (M) =
18.3 Gluing
With this machinery, Beilinson shows us how we can glue perverse sheaves. To state thetheorem we return to a slightly more general setting.
Z X U
0 D D×
i
f
j
Theorem 18.17. (Beilinson)
Perv(X) '
MU ∈ Perv(U),MZ ∈ Perv(Z)
+
ψf (MU) ψf (MU)
MZ
monodromy -1
“pieces” “glue”
140
Remark 18.18. Beilinson’s theorem glues perverse sheaves on an open set to perversesheaves on a closed set. Gluing perverse sheaves on two open sets is much easier becauseperverse sheaves form a stack.
Example 18.19. Returning to our setting of X = D = D× t 0, a perverse sheaf on D×
(resp. 0) is the same thing as a vector space with automorphism µ V (resp. a vectorspace W ). Hence
Pervconstructible
w.r.t. D×, 0
(D) =
V µ invertible,W vector space
+V V
W
f
µ−id
g
.
Let us explain why Beilinson gluing holds. We start with a general definition. Let A bean abelian category. A diad in A is
Q :=
A
C− C+
B
α+
β−
α−
β+
These form a category Diads(A) in an obvious way.Given a diad, we can associate a complex
Q· = C−(α−,β−)−−−−→ A⊕B (α+,−β+)−−−−−→ C+,
and another diad
r(Q) :=
A
kerα+ cokerα−
H0(Q·)
Lemma 18.20. r2 = id
Proof. We will take this lemma as a black box.
141
Given this,
Perv(X) ' diads of the form
Ξf (M)
j!M ⊕ j∗M
M
' diads of the form
Ξf (M)
ψf (M) ⊕ ψf (M)
φf (M)
' pairs(MU , ψf (M)→ φf (M)→ ψf (M) | ψf (M)
monodromy -1−−−−−−−−→ ψf (M))
The second equality is obtained via applying our equivalence r on diads.
142
Lecture 19: The derived category of perverse sheaves
19.1 Overview of Beilinson gluing
Recall the setting of the last lecture: Let X/C be a variety with the metric topology andPX := Perv(X, k). Beilinson described for us how to glue perverse sheaves on X. Morespecifically, we have the following set-up:
Z X U
0 A1 A1\0
i
f
j
We can move between the bounded derived categories of constructible sheaves on X, Z, andU by pushing and pulling:
Dbc(Z) Db
c(X) Dbc(U)
PZ PX PU ,
i∗=i!
i∗
i!
j∗=j!j∗
j!
i∗=i! j∗=j!
j∗ (j affine)
j! (j affine)
The functors i∗ = i!, j∗ = j! preserve perverse sheaves, and if j is affine, then j∗, j! are exact
and also preserve perverse sheaves. However, i∗, i! do not!Last lecture, we constructed two other exact functors
x ψf : PU → PZ “unipotent nearby cycles”
x Ξf : PU → PX “maximal extension”
which each have a “monodromy” x. These functors fit into short exact sequences
0→ j!M→ ΞfM→ ψfM→ 0
0→ ψfM→ ΞfM→ j∗M→ 0
for any M∈ PU . We then used ψf and Ξf to construct another exact functor
x φf : PX → PZ “unipotent vanishing cycles”
which we should think about as the “stalk of a perverse sheaf along Z”. It was defined as
φfM = H0
ΞfMU
j!MU ⊕ j∗MU
Madj −adj
143
for M ∈ PX .
Easy (and useful) fact: For M∈ PZ ,
φf (i∗M) 'M.
The main result of last lecture was Beilinson gluing, which gave an equivalence of cate-gories:
PX∼−→
M∈ PUN ∈ PZ
∣∣∣∣∣∣∣∣∣ψfM ψfM
N
1−µ
F 7→
FU ∈ PUφfF ∈ PZ
∣∣∣∣∣∣∣∣∣ψfFU ψfFU
φfF
19.2 A theorem of Beilinson
The goal of today’s lecture is to use the tools above to prove the following theorem.
Theorem 19.1. Db(PX) ' Dbc(X).
This result is a beautiful example of a bigger philosophy. Often we have a triangulatedcategory we wish to understand, and within it a collection of well-behaved objects which weunderstand better than the rest. Then we might hope to reconstruct the entire triangulatedcategory from our special collection of objects. In many situations we cannot, but this is anexample where we can. In the poetry of Beilinson, “the niche D where PX dwells may berecovered from PX”.
Recall: Let Λ be a stratification of X. Associated to this stratification is the categoryPervΛ(X) of Λ-constructible perverse sheaves on X. Our category PX is the limit of suchcategories:
PX = Perv(X) = lim→Λ
PervΛ(X).
Remark 19.2. If Λ is a fixed stratification, usually
Db(PervΛ(X)) 6' DbΛ(X).
Example 19.3. Let X = P1C with the trivial stratification Λ. Then
PervΛ(X) = local systems on X ' Vectf.d.k .
However,Db(Vectf.d.k ) 6' Db
Λ(X)
because Exti(kX , kX) = H i(X) 6= 0 for i = 2, but the category Vectf.d.k is semisimple, sothere are no higher exts.
144
19.3 K(π, 1) spaces
We see in this example that differences in Ext groups prevented us from obtaining our desiredderived equivalence. This illustrates a more general phenomenon which we will examine now.
Definition 19.4. A nice, path connected space X with base point x ∈ X is K(π,1) if
1. πi(X) = 1 for i > 1, and
2. π1(X) = π.
Equivalently, X is K(π, 1) if its universal cover is contractible.
Exercise 19.5. Show (easier) that
Repf.d. kπ∼−→ k-local systems on X
V 7→ LV
and (harder) thatExti(V, V ′) ' Exti(LV ,LV ′).
Lemma 19.6. Let F : D1 → D2 be a triangulated functor of triangulated categories. Assumethat D1,D2 have t-structures with hearts C1, C2, respectively, such that F : C1
∼−→ C2 is anequivalence, and Di = 〈Ci〉∆ for i = 1, 2 (i.e. the t-structures giving Ci are nondegenerate).Then the following are equivalent:
(a) F is an equivalence.
(b) For any objects M,N ∈ C1, F : HomiD1
(M,N)∼−→ Homi
D2(F (M), F (N)). In other
words, the “Exts agree”.
(c) (Won’t be used) Assume that D1 = Db(C1). For any x ∈ HomiD2
(F (M), F (N)), thereexists an injection N → N ′ such that x is zero in Homi
D2(F (M), F (N ′)); i.e. under
the natural map
HomiD2
(F (M), F (N))→ HomiD2
(F (M), F (N ′))
x 7→ 0.
This condition is called “effaceability”.
Proof. (a) =⇒ (b) is immediate.(b) =⇒ (a) can be shown by induction and the long exact sequence.
Example 19.7. Let X be K(π, 1). Then
Db(Repf.d. kπ)∼−→ Db
Λ(X).
Example 19.3 illustrated that this is not necessarily the case when X is not K(π, 1).
145
19.4 Yoneda extensions
To understand part (c) of Lemma 19.6, we need some facts about Yoneda extensions. Let Abe an abelian category. For objects M,N in A, define a category Ei(M,N) with
• objects: acyclic complexes
N → C1 → C2 → · · · → Ci →M
• morphisms: chain complex maps of the form
N C1 C2 · · · Ci M
N C1 C2 · · · Ci M
id id
Lemma 19.8. (/Definition)
Exti(M,N) = connected components of Ei(M,N)
0↔ split complexes in Ei(M,N)
Remark 19.9. In Lemma 19.8, two extensions are in the same “connected component” ifone can pass from one to the other going along arrows in our categories in either direction.More formally, we can build a space with 0 (resp. 1) simplices given by objects (resp. arrows)in Ei(M,N); then connected component takes on its topological meaning. One can showtwo objects
N → C1 → C2 → · · · → Ci →M and N → C1 → C2 → · · · → Ci →M
in Ei(M,N) are in the same connected component if there exists a commutative diagram
N C1 C2 · · · Ci M
N D1 D2 · · · Di M
N C1 C2 · · · Ci M
id
id
id
id
withN → D1 → D2 → · · · → Di →M
also in Ei(M,N).
Exercise 19.10. Show directly that with Exti defined as in Lemma 19.8,
Exti(M,N) = HomiD(A)(M,N).
Make no assumptions on A.
146
What about functoriality? In our usual definition of Exti in the derived category, it is obviousthat Exti is a functor. With this new definition, it is not so obvious. In other words, givenN → N ′, how to we obtain a morphism Exti(M,N)→ Exti(M,N ′)? Here is how:
• Given with N → N ′ and an element of Exti(M,N) represented by the complex
N → C1 → C2 → · · · → Ci →M,
we obtain a complex representing an element of Exti(M,N ′) by forming the push-out
N C1 C2 · · · Ci M
N ′ P C2 · · · Ci M
Here P = (C1 ⊕N ′)/N . This gives a morphism Exti(M,N)→ Exti(M,N ′).
• Similarly, for M → M ′, we can use the pull-back to form a morphism Exti(M,N) →Exti(M ′, N).
An important consequence of this is the following. If x ∈ Exti(M,N) is represented bythe complex
Nf−→ C1 → C2 → · · · → Ci →M,
then applying the construction above to the morphism f : N → C1, we obtain
N C1 C2 · · · Ci M
C1 (C1 ⊕ C1)/N C2 · · · Ci M
But since (C1⊕C1)/N = C1⊕ (C1/N), the lower sequence splits, and hence represents zeroin Exti(M,C1). In other words,
Exti(M,N)Exti(M,f)−−−−−−→ Exti(M,C1)
x 7−→ 0.
This is the origin of effaceability.
Exercise 19.11. Prove (ii) =⇒ (iii) and (iii) =⇒ (ii) in Lemma 19.6. (Hint: See Geordie’shand-written notes if you are stuck.)
19.5 A key ingredient
A key ingredient in Beilinson’s theorem is:
Theorem 19.12. Let X/C be a smooth variety. Then there exists a Zariski open set U ⊂ Xwhich is K(π, 1).
147
This theorem explains philosophically why Beilinson’s theorem is true: for perversesheaves, we can always refine our stratification to include this K(π, 1) open set. We willsee this more precisely when we discuss the proof of Beilinson’s theorem.
Example 19.13. Any curve becomes K(π, 1) after deleting a point!
(Here the arrows indicate homotopy equivalences and the x’s are missing points.)
We will roughly explain the proof of Theorem 19.12. There are two main ingredients:
1. “Noether normalization,” (NN): If Z ⊂ An is of dimension d, then a generic projectionAn → Ad is finite when restricted to Z. Here’s a caricature:
2. “Extensions of K(π, 1) are K(π, 1),” (EK(π,1)): If
F E
B
is a fibre bundle, then if B isK(π, 1), E is connected, and either F is discrete orK(π, 1),then E is K(π, 1). (Follows from the long exact sequence of homotopy groups.)
With these tools in our toolbox, the proof proceeds as follows.
148
Step 1: Reduction to X ⊂ A1, Zariski standard open18. Without loss of generality, we mayassume that X is affine. By (NN), we can find a finite map f : X → Ad. Over a standardopen U ′ ⊂ Ad, this map is etale. Hence we have a fibration
finite U
U ′
so by (EK(π,1)), if U ′ is K(π, 1), then U is K(π, 1).
Step 2: Induction on dimension. By (NN), there exists a projection f : An → An−1 suchthat on the hypersurface Z = X\U , f is finite with fibres consisting of m points:
U An Z m points
An−1 An−1
f f
Over an open V ⊂ An−1, this map will be etale, hence
U ′ U
V ′ V
is fibred in C\m points. By induction, we can shrink V to V ′ such that V ′ is K(π, 1).Hence U ′ is K(π, 1) too by (EK(π,1)).
19.6 Proof of Beilinson’s theorem for curves
The goal for the rest of the lecture is to prove Beilinson’s theorem for curves; that is, for acurve X, we wish to show that
Db(PX) ' Dbc(X).
To start, we have the following fact [DBB83, Bei87]. There exists a triangulated functor
real : Db(PX)→ Dbc(X)
called the “realization functor” which is the identity on PX . (This holds for general X, notjust curves.) We will show that real is an equivalence of categories.
By Lemma 19.6, it is enough to show that for M,N ∈ PX ,
HomiDb(PX)(M,N) ' Homi
Dbc(X)(M,N).
Moreover, because PX is finite length, we can use the long exact sequence in Ext to reduceto the case where M,N are irreducible objects.
18“Zariski standard open” means that X is of the form D(f) = z ∈ X | f(z) 6= 0 for some regularfunction f .
149
Recall that in lecture 17 we classified irreducible perverse sheaves on curves. There weretwo types: (1) skyscrapers, and (2) IC sheaves supported everywhere.
Step 1: Let M,N be skyscrapers: M = i∗kx, N = i′∗ky. Then in Dbc(X),
Homj(i∗kx, i′∗ky) = Homj(kx, i
!i′∗ky)
=
k if x = y, j = 0,
0 otherwise.
Showing thatHomDb(PX)(kx, ky) = 0
if x 6= y is easier, and left as an exercise. The tricker case is show that ExtiPX (kx, kx) = 0 fori > 0. To do this, we will use the quiver description of PX and Yoneda Exts. Recall that
PX 'V1 V0
c
v
∣∣∣ c v + id invertible.
Then an element of Ext1(kx, ky) has a representative of the form
kk → F → kx.
In quiver language, such an extension is a complex
k → V0 → k.
But any such complex fits into a diagram
0 V1 0
k V0 k
cv .
The top sequence is exact, so V1 = 0. But the bottom sequence is also exact, so V1 = 0 impliesthat the bottom sequence must split. Hence our original element is zero in Ext1(kx, ky), soExt1(kx, ky) = 0.
Showing that Ext2(kx, ky) = 0 is more challenging. An element of Ext2(kx, ky) has arepresentative of the form
k → V 10 → V 2
0 → k.
This fits into a diagram
0 V 11 V 2
1 0
k V 10 V 2
0 k
c1 c2v1 v2
We want to show that we can reduce this to an Ext supported at x. But the basic issueis that there is nothing telling us that the middle terms must be supported on a point. Insheaf language, this is equivalent to the fact that the existence of a sequence
i∗kx → F → G → i∗kx
150
whose first and last terms are supported on a point does not imply that the middle termsF ,G are.
There are two ways we can get around this. First, the “stupid way”. We can assumethat c v is nilpotent (exercise!), then we have a natural map
0 V 11 V 2
1 0
k V 10 V 2
0 k
c1 c2v1 v2 ←0 Im v1 Im v2 0
k V 10 V 2
0 k
c′1 c′2v1 v2 .
Continuing this process, we have a map
0 Im v1 Im v2 0
k V 10 V 2
0 k
c′1 c′2v1 v2 ←0 Im v1 Im v2 0
k Im c′1 Im c′2 k
c′1 c′2v′1 v′2 .
Continuing in this way, we eventually obtain a top sequence of zeros, which lets us concludethat our lower sequence splits. This procedure works for all higher Exts.
Alternatively, we could use Beilinson’s approach using vanishing cycles, which works ingreater generality (and in fact provides the skeleton of his argument in the general case).Choose a map f : X → C with zero set x ∈ X. Let C · be the complex
i∗kx → F1 → · · · → F j → i∗kx
representing an element in Extj(i∗kx, i∗kx). Then Beilinson showed that
C · ' φf (C·), (19.1)
where φf : PX → Px is the vanishing cycles functor defined in Lecture 18. The complexφf (C
·) is supported on x, so this proves that C · = 0 ∈ Extj(i∗kx, i∗kx). To prove (19.1), wecan use the exact sequences from Lecture 18:
C · C · ⊕ Ξf (C·) (C · ⊕f (C ·))/j!(C
·)
φf (C·)
Step 2: Let M be irreducible with full support and N a skyscraper supported at x. Let
j : U = X\x → X.
Recall that because the immersion j is affine, j! and j∗ are exact, preserve perverse sheaves,and fit into adjoint pairs (j∗, j∗), (j!, j
!). The adjunction maps give an exact sequence
K → j!MU M,
151
where K is a perverse sheaf supported on x. The corresponding long exact sequence givesthe diagram
· · · HomiPX
(M,N) HomiPX
(j!MU , N) HomiPX
(K,N) · · ·
· · · HomiDbc(X)(M,N) Homi
Dbc(X)(j!MU , N) HomiDbc(X)(K,N) · · ·
Using the adjunctions, we have
HomiPX
(j!MU , N) = HomiPU
(MU , NU) = 0
because NU = 0. Similarly, HomiDbc(X)(j!MU , N) = 0, so the middle vertical arrow is an
isomorphism 0 ' 0. The right vertical arrow is an isomorphism by Step 1. Hence weconclude by induction that all other vertical arrows must also be isomorphisms.
Step 3: Let M and N be irreducible objects in PX with full support. Then by Lemma 19.12,we can choose U ⊂ X such that MU , NU are (shifts of) local systems and U is K(π, 1). Notethat the inclusion j : U → X is still affine. Because N is irreducible, we have a short exactsequence
N → j∗j∗N = j∗NU k.
Then by the long exact sequence in Hom, we have the diagram
· · · HomiPX
(M,N) HomiPX
(M, j∗NU) HomiPX
(M,k) · · ·
· · · HomiDbc(X)(M,N) Homi
Dbc(X)(M, j∗NU) HomiDbc(X)(M,k) · · ·
By adjunction, we have
HomiPX
(M, j∗NU) = Homi(MU , NU).
Hence the middle vertical arrow is an isomorphism because U is K(π, 1). The right verticalarrow is an isomorphism by Step 2, and hence the left vertical arrow must also be anisomorphism.
The Moral: If we fix a stratification, there can be complicated structure going on in Db(PX)which doesn’t only come from fundamental groups. However, if we’re allowed to refine ourstratification as much as we want, then we can choose open strata which are K(π, 1), andall information in the category comes from local systems.
152
Lecture 20: Perverse sheaves for affine stratifications
Recall the setting of the last few lectures: Fix k to be our field of coefficients, and let X bean algebraic variety over C, with stratification
X =⊔λ∈Λ
Xλ.
The abelian category of Λ-constructible perverse sheaves, PervΛ(X), is a subcategory of thetriangulated category Db
Λ(X), the derived category of Λ-constructible sheaves. Moreover,there is a realization functor
real : Db(PervΛ(X))→ DbΛ(X).
Last week we saw that for a fixed stratification, real is rarely an equivalence; however, ifwe are allowed to refine stratifications, we have Beilinson’s theorem:
Theorem 20.1. (Beilinson) Db(Perv(X))∼−→ Db
c(X).
Today, we will show that if each strata Xλ is an affine space, then we have such anequivalence for a fixed stratification.
Theorem 20.2. (Beilinson–Ginzburg–Soergel) If Λ is a stratification of X such that Xλ isaffine, then Db(PervΛ(X))
∼−→ DbΛ(X).
Remark 20.3. Last week we saw that the failure of the strata to be K(π, 1) was theobstruction to real being an equivalence. Affine spaces are contractible, so this obstructiondisappears19.
Remark 20.4. Serious homological algebra goes into Beilinson’s construction of real. How-ever, if we allow ourselves to use the Riemann–Hilbert correspondence, we have
D −modhol,r.s.Λ-const PervΛ(X)
Db(D −modhol,r.sΛ-const
)Db
Λ(X)
∼
R.H.
.
Under the equivalence between perverse sheaves and D-modules given by the top arrow, thebottom arrow in this diagram is real. This gives a high concept (rather than high homologicalalgebra!) construction of the realization functor.
The BGS theorem and the techniques involved in proving it will be invaluable as weapproach Bezrukavnikov’s equivalence. The proof of the BGS theorem is algebraic, usingwhat are today called highest weight categories.
19Caution: take this explanation with a grain of salt. The proof of Beilinson’s theorem is geometric (viavanishing cycles), and the proof of the BGS theorem is algebraic (via highest weight categories), so a directcomparison doesn’t exactly work.
153
20.1 Highest weight categories
Definition 20.5. Let A be a k-linear category. We say that A is highest weight if thefollowing six conditions hold.
1. A is finite length.
2. The set simple objects in A/' is finite.
3. For any simple object L ∈ A, End(L) = k.
Let Λ be an indexing set for isomorphism classes of simple objects and denote byLλ ∈ A the simple object corresponding to λ ∈ Λ. Assume that Λ is a poset (this ispart of the data determining a highest weight category). For any closed subset T ⊂ Λ(that is, if λ′ ≤ λ and λ ∈ T , then λ′ ∈ T ), denote by
AT = 〈Lλ | λ ∈ T 〉Serre
the Serre subcategory generated by the simple objects Lλ with λ ∈ T . Assume more-over that for each λ ∈ Λ, there are objects ∆λ (“standard” object), ∇λ (“costandard”object) in A and maps ∆λ → Lλ, Lλ → ∇λ.
4. If T ⊆ Λ is closed, then ∆λ → Lλ (resp. Lλ → ∇λ) is a projective cover20 (resp.injective hull) in AT 21.
5. For λ ∈ Λ,
ker(∆λ → Lλ) ∈ A<λ, and
coker(Lλ → ∇λ) ∈ A<λ.
This implies that the composition series eggs of ∆λ and ∇λ must have the followingform
6. Ext2(∆λ,∇µ) = 0 for all λ, µ ∈ Λ.22
20A projective cover of an object M in a category C is a morphism Pf−→ M out of a projective object
P ∈ C which is a superfluous epimorphism, meaning that every morphism Ng−→ P with the property that
f g is an epimorphism is itself an epimorphism.21Note that we are not requiring ∆λ (resp. ∇λ) to be projective (resp. injective) objects in A, just in the
Serre subcategory AT .22This condition is the most mysterious, and often the hardest to show.
154
Theorem 20.6. (Beilinson–Ginzburg–Soergel) LetA be highest weight. ThenA has enoughprojective and injective objects. Moreover, the projective cover Pλ of Lλ has a standardfiltration of the form
Similarly, the injective hull Iλ has a costandard filtration of the form
Proof. For full details, see [BGS96, Theorem 3.2.1].
Sketch: Instead of proving the theorem as stated, we prove that for any closed subset T ⊆ Λ,AT has enough projectives (resp. injectives), and they admit standard (resp. costandard)filtrations. We construct the projective cover P T
λ of Lλ inductively on T :
• Let µ ∈ T be maximal, and λ 6= µ ∈ T . Assume that the projective cover PT\µλ of
Lλ in AT\µ is already constructed.
• Let E = Ext1(PT\µλ ,∇µ) be the ext group.
• Every element of E gives rise to an extension
∆µ → P ′ Pλ.
It turns out that there exists a “universal” extension
E∗ ⊗∆µ → P PT\µλ
from which all others are constructed via push-out. (Challenge: construct it!)
155
• Claim: P is the projective cover we are seeking.
Proof. Use long exact sequence in Hom:
Hom(PT\µλ ,∆µ) Hom(P ,∆µ) E ⊗ Hom(∆µ,∆λ)
Ext1(PT\µλ ,∆µ) Ext1(P ,∆µ) E ⊗ Ext1(∆µ,∆µ)
∼
The last term is zero by axoim (4) of a highest weight category. Hence Ext1(P ,∆µ) = 0.
A bit more work using axiom (6) shows that P = P Tλ .
Definition 20.7. In a highest weight category A, an object T is tilting if it has a standardand costandard filtration. The indecomposable tilting objects in A are also indexed by Λ.
Similar arguments give injective and tilting objects.
Important objects in a highest weight category:
If our category also has a notion of duality compatible with the highest weight structure,then the costandard/standard and projective/injective objects are dual, as indicated by thepink arrows above.
156
Exercise 20.8. Let B ⊂ A be a Serre subcategory of an abelian category A. By definition,for objects M,N ∈ B,
Ext1B(M,N)
∼−→ Ext1A(M,N).
Show thatExt2
B(M,N) → Ext2A(M,N).
Hint: Use effaceability and the long exact sequence of Ext.
20.2 Perverse sheaves for an affine stratification are a highestweight category
Theorem 20.9. If X =⊔λ∈ΛXλ is stratified by affine spaces, then PervΛ(X) is a highest
weight category.
Proof. First, note that jλ : Xλ → X is affine and hence
∆λ := jλ!kXλ [dimXλ] and
∇λ := jλ∗kXλ [dimXλ]
are perverse sheaves. Moreover, we have canonical maps
∆λ ICλ → ∇λ, (20.1)
where ICλ := jλ!∗kXλ is the IC sheaf corresponding to the trivial local system on Xλ.
1. PervΛ(X) is finite-length. X
2. Simple objects in PervΛ(X) are parameterized by pairs (L, λ), where L is a local systemon Xλ. Because Xλ is affine, it is contractible, and thus admits a single local system.Hence there are finitely many simple objects in PervΛ(X). X
3. By Schur’s Lemma, End(ICλ) is a division algebra over k.
Claim 20.10. If D is a division algebra over a field k, any non-zero algebra homomor-phism D → k is an isomorphism.
Proof. Let ϕ : D → k be a nonzero algebra homomorphism. The kernel of ϕ is an idealin D, but the division algebras have no non-trivial ideals, so kerϕ = 0. Because ϕis a k-algebra homomorphism, ϕ(1) = 1. Hence for all ` ∈ k, ϕ(` · 1) = `ϕ(1) = ` andimϕ = k.
Assume Xλ ⊂ X is open. Then ICλ|Xλ = kXλ [dimXλ]. Hence the restriction map
End(ICλ)restriction−−−−−→ End(ICλ|Xλ) ' k
is a nonzero morphism from a division algebra to k. The claim lets us conclude thatEnd(ICλ) ' k.
157
For any λ, the inclusion Xλ → Xλ is open because Xλ is locally closed. There is anequivalence of categories
perverse sheavessupported on a
closed subvarietyZ ⊂ X
∼←→ Perv(Z),
soEndPervΛ(X)(ICλ) ' EndPervΛ(Xλ)(ICλ).
Hence by applying the argument above to the open inclusion Xλ → Xλ, we concludethat End(ICλ) = k for all λ. X
4. The maps in (20.1) give a surjection ∆λ ICλ and an injection ICλ → ∇λ. Becausethe partial order on Λ is given by closure of strata, for a closed subset ≤ λ ⊂ Λ, wehave
PervΛ(X)≤λ = PervΛ(Xλ).
To establish axiom (4), we must show two things: (a) that ∆λ is a projective object inPervΛ(Xλ), and (b) that ∆λ ICλ is a projective cover.
To show that ∆λ is projective, we will show that HomPervΛ(Xλ)(∆λ, ·) is exact. Because
Xλ
j−→ Xλ is open, for any F ∈ PervΛ(Xλ),
Hom(j!kXλ [dimXλ],F) = Hom(kXλ [dimXλ], j!F)
= Hom(kXλ [dimXλ], j∗F)
= Hom(kXλ [dimXλ],F|Xλ)
= F|Xλ .
Restriction to an open subvariety is an exact functor, so we conclude that Hom(∆λ, ·)is exact.
Because PervΛ(Xλ) is a finite-length abelian category (in particular, it is Krull-Schmidt),to show that ∆ ICλ is a projective cover, it suffices to show that ∆λ is indecompos-able. Now, the endomorphism ring
End(∆λ) = Hom(j!kXλ , j!kXλ) = Hom(kXλ , j!j!kXλ) = End(kXλ) = k
is local, so ∆λ is indecomposable.
Showing that the injection ICλ → ∇λ is an injective hull follows from a similar argu-ment. X
5. The fact that ker(∆λ ICλ), coker(ICλ → ∇λ) ∈ Perv<λ(X) is clear from the com-
158
position series eggs and the definition of ICλ = Im(Jλ! → jλ∗):
X
6. To see why condition (6) holds, first observe that by Exercise 20.8,
Ext2PervΛ(X)(∆λ,∇µ) → Ext2
Perv(X)(∆λ,∇µ).
Then, by Beilinson’s theorem,
Ext2Perv(X)(∆λ,∇µ) = Hom2
Dbc(X)(∆λ,∇µ).
Once we are in the derived category, we have full access23 to the adjoint pairs (j!, j!), (j∗, j∗).
Hence we can use adjunctions to compute
Hom2Dbc(X)(∆λ,∇µ) = HomDbc(X)(jλ!kXλ [dimXλ], jµ∗kXµ [dimXµ + 2])
= HomDbc(X)(j∗µjλ!kXλ [dimXλ], kXµ [dimXµ + 2]).
Because the functor jλ! is extension by zero, j∗µjλ!kXλ = 0 for µ 6= λ. If µ = λ, thenj∗µjλ!kXλ is a 1-dimensional vector space. There are no higher Exts in the category ofvector spaces, so we conclude that
Ext2Pervλ(X)(∆λ,∇µ) = 0.X
Remark 20.11. It’s unusual for a category of perverse sheaves to have enough projectives.By proving that PervΛ(X) is a highest weight category, we have just algebraically produceda collection of projective perverse sheaves Pλλ∈Λ. As far as we are aware, it is not knownhow to construct these perverse sheaves geometrically.
Corollary 20.12. If Λ is a stratification of X by affine spaces,
real : Db(PervΛ(X))∼−→ Db
Λ(X)
is an equivalence of categories.
23What we mean by this is that the functor j∗λ is not exact if Xλ is not open, so it does not preserveperverse sheaves (see beginning of lecture 19).
159
Proof. Both of the sets ∆λλ∈Λ and ∇λλ∈Λ generate Db(PervΛ(X)) and DbΛ(X). By the
upper triangularity of Pλ, the set Pλλ∈Λ also generates each category. Hence it is enoughto show
real : HomiDb(PervΛ(X))(Pλ,∇µ)
∼−→ HomiDbλ(X)(Pλ,∇µ).
In Db(PervΛ(X)),
Homi(Pλ,∇µ) =
0 if i > 0,
HomPervΛ(X)(Pλ,∇µ) if i = 0,
by the projectivity of Pλ.In Db
Λ(X),
Homi(Pλ,∇µ) =
0 if i > 0,
HomDbΛ(X)(Pλ,∇µ) if i = 0
as well. This is because HomiDbΛ(X)(∆λ,∇µ) = 0 for i > 0 by adjunction, so the long exact
sequence in Ext and the standard filtration of Pλ imply that Homi(Pλ,∇µ) = 0 for i > 0.
Remark 20.13. In the proof above, the vanishing of Homi for i > 0 happens for quitedifferent reasons in each of the two categories. In Db(PervΛ(X), the reason is algebraic(projectivity of an object), whereas in Db(X), the reason is topological.
20.3 Where are we going for the next few weeks?
For the rest of this lecture, we will reconnect with the big picture of this course and describeour plan for the upcoming weeks.
Let (X ⊃ R,X∨ ⊃ R∨) be a root datum, and let G,G∨ be the corresponding dual groupsover C. Associated to this datum, we have a finite Weyl group Wf , and an affine Weyl groupW = Wf n ZX∨. Let H be the affine Hecke algebra, and Z its center. In October andNovember of last year, we constructed the following diagram.
Z ' (ZX∨)Wf RG∨ = [Coh pt/G∨ ]
H[KG∨×C×(St)
]∼
pull-back
∼KL
The isomorphism on the bottom line is the Kazhdan–Lusztig isomorphism, which was merelystated (not yet even sufficiently explained, let alone proved!), and the injection on the left isBernstein’s description of the center of of the affine Hecke algebra. Our goal for the rest ofthe course is to categorify this picture. This is done via Bezrukavnikov’s equivalence, whichvery roughly is an equivalence of the form constructible
affine Heckecategory
, ∗
coherentaffine Hecke
category, ∗
.∼B
160
The constructible affine Hecke category on the LHS should be (a variant of) a category ofconstrucible sheaves on the affine flag variety of G, and the coherent affine Hecke category onthe RHS should be (a variant of) a category of G×C×-equivariant coherent sheaves on theSteinberg variety. When we take Grothendieck groups, we should recover the first diagram.
Philosophy for now: Before we can understand this story on the level of categories, weneed to better understand the Kazhdan–Lusztig isomorphism.
General strategy for understanding KL:
1. On the Bernstein generators Ti, Oλ, define
Ti 7→ Qi (to be explained next week)
Oλ 7→ Oλ (pull-back of O(λ) on G∨/B∨ to T ∗G∨/B∨diagonal−−−−→ St)
Remark 20.14. It is very challenging to verify the relations directly! (See LeonardoMaltoni’s May 24, 2019 talk in the Informal Friday Seminar.)
2. The affine Hecke algebra H has two important modules, defined as follows. Let
triv : Hf → Z[q±1], Ti 7→ q
sgn : Hf → Z[q±1], Ti 7→ −1
be the (quantized versions of) the trivial and sign representation of the finite Heckealgebra Hf . Define two H-modules
M := H ⊗Hf triv, N := H ⊗Hf sgn
by left multiplication on the first tensor factor. These are, respectively, the spheri-cal module and antispherical module of the Hecke algebra corresponding to theparabolic subgroup Wf ⊂ W .
Remark 20.15. By the Bernstein presentation,
H = Z[q±1][X∨]⊗Hf .
Hence, as Z[q±1]-modules,M ∼= N ∼= Z[q±1][X∨].
Thus, as modules over Z[q±1][X∨], both the spherical and anti-spherical modules arefree of rank 1.
3. By convolution formalism, bothKG×C×(G∨/B∨) andKG×C×(T ∗G∨/B∨) areKG∨×C×(St)-modules.
Remark 20.16. Note that
Z[q±1][X∨] = KB∨×C×(pt) ∼= KG∨×C×(G∨/B∨)∼−−−−−→
pull-backKG∨×C×(T ∗G∨/B∨).
161
4. Compute actions of generators in M (resp. N) and match them with actions inKG∨×C×(G∨/B∨) (resp. KG∨×C×(T ∗G∨/B∨) under (a choice of) the above isomor-phisms. Because all of these modules are faithful, this implies the Kazhdan–Lusztigisomorphism.
Remark 20.17. We really only need to do this computation for either M or N to usethis argument to deduce the Kazhdan–Lusztig isomorphism.
Next week: We implement 1-4.
Following week: We take the first step toward Bezrukavnikov’s equivalence by explainingthe Arkhipov–Bezrukavnikov’s theorem that
AS ∼−→ Db(
CohG∨×C×(T ∗G∨/B∨)
),
where AS is the categorical antispherical module.
162
Lecture 21: Equivariant K-theory of the Steinberg va-
riety
Our goal for the next few lectures is to prove the Kazhdan–Lusztig isomorphism:
Haffine ' KG∨×C×(Steinberg).
This is an isomorphism of the affine Hecke algebra with the G∨ × C×-equivariant K-theoryof the Steinberg variety. Recall from the end of last lecture that our general strategy forproving this isomorphism is to find a vector space on which each algebra acts faithfully bythe same operators.
Remark 21.1. The reader might have the (correct) impression that this is a rather indirectway of seeing that two algebras are isomorphic. However, there are several instances of veryindirect techniques to obtain isomorphisms, equivalences, or correspondences in the Lang-lands program. Another example of this is Soergel’s functor, which proves an equivalence oftwo categories (one geometric and one representation theoretic) by matching their images ina third world of algebra (so-called Soergel bimodules).
References for this lecture are:
• Kazhdan–Lusztig, Proof of the Deligne–Langlands conjecture for Hecke algebras, [KL87],
• Chriss-Ginzburg, Representation theory and complex geometry, Chapters 6 & 7, [CG09],
• Henderson, Notes on affine Hecke algebras and K-theory.
21.1 Equivariant K-theory
Let X be a scheme with an action by a group G. Associated to X are the G-equivariantK-groups for i ∈ Z≥0:
G X KGi (X).
When i = 0, this is
KG0 (X) =
Grothendieck group of the exact categoryof G-equivariant perfect complexes on X
(i.e. bounded complexes of vector bundles).
If X is smooth, then every G-equivariant coherent sheaf has a resolution by G-equivariantvector bundles and thus
KG0 (X) =
Grothendieck group of the categoryof G-equivariant coherent sheaves on X
.
Remark 21.2. At some future point we should expand on higher K-groups, but that timeis not today. For those who are interested, Quillen’s work and the groundbreaking paper
163
[TT90] of Thomason-Trobaugh are highly recommended. In [TT90], they show that if U ⊂ Xis open, we have a sequence24
· · · → Ki(X on Y )→ Ki(X)→ Ki(U)→ Ki−1(X on Y )→ · · ·
which is almost exact, except that K0(X)→ K0(U) might not be onto for singular X. (Thisresult was already proved by Quillen for smooth X.)
In the arguments to come, we only care about KG0 , and at certain points we will simply
assert that certain boundary maps vanish; e.g., we have a short exact sequence
KG0 (X on Y ) → KG
0 (X) KG0 (U).
We set KG(X) := KG0 (X).
Examples
1. KG(pt) = RG =Grothendieck group of finite-dimensional
algebraic representations of G
2. We claim that K(P1) = Z[x±1]/(x − 1)2. How can we see that this is true? To start,every vector bundle on P1 is a sum of line bundles. Hence there is a surjection
Z[x±1] K(P1)
xm 7→ O(m).
The tautological exact sequence of vector bundles
O(−1) → O⊕2P1 O(1)
shows that under the surjection above,
x−1 − 2 + x 7→ 0.
Hence (x− 1)2 7→ 0. This shows us that we have a map φ : Z[x±1]/(x− 1)2 → K(P1).
Claim: K(P1) = Z[O]⊕ Z[O(1)].
The arguments above imply that [O] and [O(1)] span K(P1). It remains to show thatthey are linearly independent.
Proof 1: For X proper, K(X) carries an intersection form:
〈F ,G〉 := χ(F ⊗ G),
24We can think of this like the long exact sequence in cohomology, but unlike cohomology, where thelong exact sequence came almost simultaneously with its definition, this took twenty years after the initialdefinition of K-theory to prove.
164
where χ is the Euler characteristic. Computing the pairings with respect to this form,we get the following table.
O O(1)O 1 2O(1) 2 3
We can see that the determinant is −1, so O and O(1) are linearly independent.
Proof 2: We have seen in Tom’s course that
Db(CohP1) ' Db(Rep( • • )
)O 7→ L0 := k 0
O(1) 7→ L1 := 0 k
The K-group of any finite length abelian category has a basis given by the classes ofsimple objects. In the case of Rep( • • ), the two simple objects are L0 and L0.Hence we have
K(P1) = K(Db(Coh(P1))) = K(Db(Rep( • • )
))= Z[L0]⊕ Z[L1].
Remark 21.3. The second proof illustrates a common phenomenon: derived equiva-lences can have interesting consequences for K-theory.
Fact: Proper maps f : X → Y induce maps in K-theory:
p∗ : KG(X)→ KG(Y )
The map p∗ is given by p∗(F ) =∑
(−1)iRif∗(F ).
Example 21.4. 1. Projection to a point, p : P1 → pt, induces the following map on(non-equivariant) K-theory:
p∗ : K(P1) ∼= Z[x±1]/(x− 1)2 → K(pt) ∼= Zxm 7→ χ(O(m)) = m+ 1
2. Let B =
(∗ ∗0 ∗
)⊂ SL2 act on P1 in the standard way. Then the map that p
induces on B-equivariant K-theory is given by the Weyl character formula. Indeed, wecan identify
KG(P1) = KB(pt) = Z[X(B)] = Z[x±1].
Then under the composition
KG(P1)→ KG(pt)→ KB(pt) = Z[x±1].
the element xm ∈ KB(P1) ∼= Z[x±1] maps to xm−x−m−2
1−x−2 ∈ KB(pt) ∼= Z[x±1].
165
21.2 Equivariant K-theory of the Steinberg variety
LetG ⊃ B ⊃ T
be a complex reductive group containing a Borel subgroup containing a maximal torus. LetX = X(T ) be the character lattice, Wf the finite Weyl group, and W = Wf nZX the affineWeyl group of the dual25 root system. Let
N ⊂ g
be the nilpotent cone in the Lie algebra of G, and
N = (x, b) ∈ N × B | x ∈ b = T ∗B → N
be the Springer resolution. Recall that the Steinberg variety is defined to be
St := N ×N N = (x, b, b′) | x ∈ b, x ∈ b′.
We also explained in Lecture 13 how this could be realized as the conormal space to thespace B × B with the stratification by G-orbits (which are parameterized by W ).
Example 21.5. Let G = SL2. Then B = P1, and the G-orbit stratification on B×B is givenby
B × B = ∆ t Y,
where Y := (P1 × P1)\∆. An illustration:
Then the Steinberg variety is
T ∗∆(P1 × P1) t T ∗Y (P1 × P1) = T ∗P1 t Y.25This convention is introduced to avoid including many checks (to indicate Langlands dual groups) in the
rest of this lecture and the next.
166
An illustration:
In the “easy part,” projection onto each component of B × B is an isomorphism. In the“tricky part,” projection onto each component hits the zero section in T ∗P1.
The group G× C× acts on St via
(g, z) · (z, b, b′) = (z2gx, gb, gb′). (21.1)
Algebra structure on KG×C×(St):
Recall that KG×C×(St) has an algebra structure given by convolution in Db(Coh) (seeLectures 14 and 15 for a refresher):
F ∗ G := p13∗(p∗12F ⊗ p∗23G),
where pij are the canonical projections
N ×N N ×N N
St St St
p12 p13p23
Module structure on KG×C×(N ):
Similarly, convolution also gives KG×C×(N ) the structure of a module for KG×C×(St):
F ∗ G := p1∗(F × p∗2G),
where
N ×N N
N N
p1 p2
Recall that our goal is to prove that
KG×C×(St) ∼= H.
167
We will accomplish that through a subgoal, which is to show that
KG×C×(N ) ∼= anti-spherical module.
We can work toward the subgoal by examining the vector space structure of KG×C×(N )more closely.
First, note that we have a map
KG×C×(N )pull←−−back
KG×C×(B) = KB×C×(pt) = Z[v±1][X].
Because N is a vector bundle over B, pull-back is an isomorphism. Hence
KG×C×(N ) = Z[v±1][X].
Now we turn to the vector space structure on KG×C×(St). The space B×B has a filtrationvia closures of G-orbits:
∅ = Z−1 ⊂ Z0 ⊂ · · · ⊂ Zm = B × B,
with Zi\Zi+1 ' G · (xiB,B) ' G×BXi, where Xi = BxiB/B is a Bruhat cell. This inducesa filtration of the Steinberg variety:
∅ = Z−1 ⊂ Z0 ⊂ · · · ⊂ Zm = St,
with Zi\Zi+1 ' T ∗Xi(B × B) =: T ∗xi . We have
KG×C×(T ∗xi) = KG×C×(Xi) = KB×C×(Xi) = KB×C×(pt).
All boundary maps vanish when we apply KG×C× to the above filtration, and a little work(see [CG09]) yields:
KG×C×(St) =⊕x∈Wf
Z[v±1][X][OT ∗xi ].
Remark 21.6. Compare this to the decomposition
H =⊕x∈Wf
Z[v±1][X]Tx.
21.3 The spherical and anti-spherical modules
Recall Bernstein’s presentation of the affine Hecke algebra (Theorem 12.5): H is an algebraover Z[v±1] with generators
Hs | s ∈ Sf(finite part)
andθλ | λ ∈ X(lattice part)
.
The generators Hs | s ∈ Sf generate a copy of the finite Hecke algebra Hf and thegenerators θλ | λ ∈ X generate Z[v±1][X]. The relations are:
168
1. H2s = (v−1 − v)H1 + 1 for all s ∈ S + braid relations (finite part),
2. θλθγ = θλ+γ for all λ, γ ∈ X (lattice part), and
3. the most important relation: For a simple reflection s = sα ∈ S and λ ∈ X,
Hsθsλ − θλHs = (v − v−1)
(θλ − θsλ1− θ−α
).
The first relation can be rewritten as
(Hs + v)(Hs − v−1) = 0,
which implies that Hf has two natural rank 1 modules:
triv : Hs 7→ v−1,
sgn : Hs 7→ −v.
From these we construct two induced H-modules:
M = H ⊗Hf triv, “spherical module”
N = H ⊗Hf sgn, “anti-spherical module”
As modules over the lattice part, each is isomorphic to Z[v±1][X].
Remark 21.7. In what follows, we will abuse notation and write
θλ := θλ ⊗ 1 ∈ N.
Using the third relation in the affine Hecke algebra, we can compute the action of thegenerator θs ∈ H on θsλ ∈ N in the anti-spherical module:
Hs · θsλ = (v − v−1)
(θλ − θsλ1− θ−α
)+ θλHs
= v
(θλ − θsλ − θλ + θλ−α
1− θ−α
)− v−1
(θλ − θsλ1− θ−α
)= v
(θλ−α − θsλ
1− θ−α
)− v−1
(θλ − θsλ1− θ−α
).
This formula looks nicer if we instead compute the action in terms of the Kazhdan–Lusztiggenerator bs := Hs + v:
bs · θsλ = v
(θλ−α − θsλ
1− θ−α
)− v−1
(θλ − θsλ1− θ−α
)+ vθsλ
= v
(θλ−α − θsλ−α
1− θ−α
)− v−1
(θλ − θsλ1− θ−α
)= (vθ−α − v−1)
(θλ − θsλ1− θ−α
).
In other words,
bs · θλ = (v−1 − vθ−α)
(θλ − θsλ1− θ−α
). (21.2)
We’ll pick up here next week with a very similar looking computation in K-theory.
169
Lecture 22: Proof of the Kazhdan–Lusztig isomorphism
In today’s lecture we will complete the proof of the Kazhdan–Lusztig isomorphism. First,we establish and streamline notation.
22.1 Notation and set-up
LetG ⊃ B ⊃ T
be a complex reductive group, Borel subgroup, and maximal torus. Let X be the characterlattice, and Wf ⊂ W the finite and affine Weyl groups with simple reflections Sf ⊂ Wf andS ⊂ W . We denote by B the flag variety, which we realize as the variety of Borel subalgebrasof g = LieG. The group G acts on the product B × B, and the orbits give a stratificationwith strata parameterized by Wf :
G B × B =⊔x∈Wf
Ox.
The orbit Ox consists of pairs of flags/Borel subalgebras in relative position x. For example,Oid = ∆ (the diagonal in B × B), and Ow0 (where w0 ∈ Wf is the longest element) is openin B × B.
LetN = T ∗B → N ⊂ g
be the Springer resolution. The Steinberg variety is
St = N ×N N =⊔x∈Wf
T ∗Ox(B × B) =⋃
w∈Wf
Λx,
where T ∗Ox(B×B) is the conormal bundle of the G-orbit Ox, whose closure Λx := T ∗Ox(B × B)is an irreducible component of St.
Last week we discussed the following example.
Example 22.1. Let G = SL2, so B = P1. The G-orbit stratification of B × B is:
The corresponding stratification of the Steinberg variety is
St = T ∗Oid(B × B) t T ∗Os(B × B).
170
The closures of the strata give the irreducible components Λx, x ∈ Wf of St, which are gluedtogether as follows:
St =
Remark 22.2. In general, the components range from Λid ' N to Λw0 = B × B. They arenot smooth in general, and their intersection pattern is extremely complicated. For example,it is not known in general when two Λx and Λy intersect in codimension 1.
Last week we introduced the anti-spherical module
N = H ⊗Hf sgn ' Z[v±1][X]
θλ := θλ ⊗ 1←[ eλ
for the affine Hecke algebra H and established the Demazure–Lusztig formula
bs · θλ = (v−1 − vθ−α)
(θλ − θsλ1− θ−α
). (22.1)
Here Hf = 〈Hs〉s∈Sf is the finite Hecke algebra and bs := Hs + v is the Kazhdan–Lusztiggenerator corresponding to the simple reflection s = sα ∈ Sf .
Remark 22.3. It is not immediately obvious that the formula(θλ−θsλ1−θ−α
)gives an element in
the lattice part of H. However, by rewriting(θλ − θsλ1− θ−α
)= (θλ + θλ−α + · · ·+ θsλ+α),
we see that bs · θλ ∈ Z[v±1][X]. (Here we assume 〈λ, α∨〉 ≥ 0.)
We want to show that a similar formula holds for the KG×C×(St)-action on KG×C×(N ).We will first do so in the special case of G = SL2, then move on to the general formulation.
22.2 The case of SL2
We start with a baby version of the calculation. We identify KG(P1) with Z[x±1] via thechain of isomorphisms
KG(P1) ' KG(G/B) ' KB(pt) ' Z[x±1]
171
Under this identification, O(m) 7→ xm. We have a Cartesian square
P1 × P1 P1
P1 pt
p2
p1 p′1
p′2
Claim 1: The map p1∗p∗2 : KG(P1)→ KG(P1) is given by xm 7→ xm−x−m−2
1−x−2 .
Proof. By smooth base change, p1∗p∗2 = p′∗2 p
′1∗. Then the claim follows from Weyl’s character
formula.
Claim 2: p1∗(O(−2, 0)⊗ p∗2(−)) : KG(P1)→ KG(P1) is given by xm 7→ x−2(xm−x−m−2
1−x−2
).
Proof. The notationO(m,n) refers to p∗1O(m)⊗p∗2O(n). By the projection formula, p1∗(p∗1(−2)⊗
p∗2(−)) = O(−2)⊗ p1∗p∗2(−). The claim then follows from Claim 1.
Now we move up to N . Recall that the pull-back q∗ of the projection N = T ∗P1 q−→ P1
gives an isomorphism
KG×C×(N )∼←−q∗KG×C×(P1) = Z[v±1][X].
Recall that C× acts on N via scaling by z2 (equation (21.1)). For the rest of this section,set O := ON .
We have projections
SetQs := OP1×P1(−2, 0)
172
on Λs ' P1 × P1 ⊂ St. We will see that (up to some simple scalar factors), Qs acts viaconvolution on KG×C×(St) in the same way that bs acts on N . To do so, we want to calculate
p1∗(Qs ⊗ p∗2O(m)) ∈ KG×C×(N ).
Claim 3: Let P1 i−→ N . Then p1∗(Qs ⊗ p∗2O(m)) =
(xm−x−m−2
1−x−2
)[i∗OP1(−2)] .
Exercise 22.4. 1. Suppose π : E → Y is a vector bundle, F a quasi-coherent sheaf onY , G a locally free sheaf on Y , and i : Y → E the zero section. Show that
i∗F ⊗ π∗G = i∗(F ⊗ G).
2. Use 1. applied to N × N → P1 × P1 to deduce Claim 3.
It remains to express [i∗OP1(−2)] in terms of our basis of KG×C×(N ); that is, to under-stand i∗OP1(−2) in terms of vector bundles.
Useful basic technique: Given a vector bundle q : V → Y , there is a bijectionquasi-coherent sheaveson the total space of V
'
quasicoherent sheaves ofmodules over q∗OV ' Sym•(V ∗)
.
(More generally, this holds for any affine morphism, see Hartshorne.)
In our case, the projectionN = T ∗P1 → P1
is the line bundle associated to O(−2), and
q∗O ' Sym•(OP1(2)) = OP1 ⊕OP1(2)⊕OP1(4)⊕ · · · .
Additionally, C× acts on T ∗P1 → P1 as multiplication by z2 along the fibres, which corre-sponds to multiplication by z−2 in the second factor in the direct sum decomposition above.Hence we have an exact sequence
z−2O(2) → O i∗OP1 .
Remark 22.5. Locally, this is the short exact sequence
k[x]·x−→ k[x] k,
but globally, we have some twisting. This is a special example of the Koszul resolution ofthe zero section of a vector bundle.
Hence in KG×C×(N ), we have
[i∗OP1 ] = [O]− v−2[O(2)].
173
By tensoring with O(−2) we obtain
[i∗OP1(−2)] = [O(−2)]− v−2[O].
Putting it all together, we see that
xm[Qs]∗7−−−→ (x−2 − v−2)
(xm − x−m−2
1− x−2
).
Hence
xm−[vQs]∗7−−−−→ (v−1 − vx−2)
(xm − x−m−2
1− x−2
).
We only need two last pieces of information to establish the Kazhdan–Lusztig isomorphismfor SL2, which we leave as exercises.
Exercise 22.6. The map θm 7→ xm−1 intertwines the actions of bs and −[vQs]∗.
Exercise 22.7. Both the representation N of H and the representation KG×C×(N ) ofKG×C×(St) are faithful.
22.3 The general case
Roughly speaking, in the SL2 example, we have seen the meaning in K-theory of theDemazure–Lusztig formula
bs · θλ = (v−1 − vθ−α)
(θλ − θsλ1− θ−α
).
The first factor (v−1 − vθ−α) comes from the “Koszul resolution,” and the second factor(θλ−θsλ1−θ−α
)from the “push-pull for P1”. We will now explain that the same philosophy works
in general.
Relative cotangent bundle: Let
Zf−→ Y
be a smooth map. (Think submersion = fibration.)
The relative tangent bundle is
TZ/Y := ker(df : TZ → TY ).
174
We have a short exact sequence
TZ/Y → TZ f ∗TY .
Dually,f ∗T ∗Y → T ∗Z T ∗Z/Y .
A caricature:
In our setting, we have a P1 fibration
B πs−→ Bs,
where Bs is the variety of parabolic subalgebras of type s.
Fundamental Cartesian diagram:
Os B
B Bs
p1
p2 πs
πs
Here Os ⊂ B × B is the G-orbit consisting of pairs of flags in relative position s. All mapsin this diagram are P1-fibrations. We identify KG(B) ' Z[X] via OB(λ) 7→ eλ.
Claim 4: πs∗π∗s : KG(B)→ KG(B) maps eλ 7→ eλ−es(λ)−α
1−e−α .
175
Proof. Again, this is simply an instance of the Weyl character formula for P1.
We claim that the corresponding Cartesian diagram “upstairs” is
Λs Ns
Ns T ∗Bs
To see that this is the correct lift of the first Cartesian square, we start with the lower rightcorner
T ∗Bs = (p, x) | x ∈ nilrad(πs(b),
which implies that the upper right and lower left corners are
Ns = π∗sT∗B/Bs = (b, x) | x ∈ nilrad(πs(b)) ⊂ N .
Hence the fibre product completing the diagram in the upper left corner is given by(b, b′, x) | b, b′ in relative position s, and
x ∈ nilrad(πs(b)) = nilrad(πs(b′))
= Λs.
Together, these two Cartesian squares form a “Cartesian cube:”
Λs Ns
Ns T ∗Bs
Os B
B Bs
DefineQs := q∗ΩOs/B
to be the relative 1-forms with respect to the second projection. As earlier, Claim 4 gives
p1∗(Qs ⊗ p∗2O(λ)) =
(eλ − es(λ)−α
1− e−α
)[i∗ONs(−α)
],
where Ns → N is the inclusion.What remains is to express
[i∗ONs(−α)
]∈ KG×C×(N ) in the basis of line bundles. Again,
we can do so using a Koszul-type resolution. There is a short exact sequence of B-modules
ps/b → g/b g/ps.
176
The corresponding exact sequence of vector bundles on B = G/B is
Lα → TB π∗sTBs ,
where Lα is the line bundle on B associated to α ∈ X. Passing to symmetric algebras, weobtain the Koszul resolution
z−2 Sym•(TB)(α) → Sym•(TB) Sym•(π∗sTB).
Remark 22.8. This should be thought of as the vector bundle version of the short exactsequence
k[x1, . . . , xn]·xn−→ k[x1, . . . , xn] k[x1, . . . , xn−1].
The z−2 comes from the C×-action.
In other words,z−2ON (α) → ON i∗ONs .
The result: [i∗ONs(−α)
]=[ON (−α)
]− v−2
[ON].
From here, the proof follows from two exercises analogous to those in the previous section.
Exercise 22.9. Check that θλ 7→ eλ±ρ intertwines bs· and −[vQs]∗.
Exercise 22.10. Complete the proof by showing that N (resp. KG×C×(N )) are faithful
modules over H (resp. KG×C×(N )).
Remark 22.11. Geordie isn’t quite sure about whether we should have α or −α above, sothe reader should take the final arguments with a grain of salt.
177
Lecture 23: Gaitsgory’s central sheaves
Last week we finished the proof of the Kazhdan–Lusztig isomorphism:
KG∨×C×(St) ' H.
This is an isomorphism of algebras. We established this isomorphism by proving that twofaithful modules for these algebras are isomorphic:
KG∨×C×(N ) ' N = H ⊗Hf sgn. (23.1)
For the next several weeks (months?), we will work toward categorifying the Kazhdan–Lusztig isomorphism via Bezrukavnikov’s equivalence. The first step towards a categorifica-tion is the Arkhipov-Bezrukavnikov theorem, which categorifies (23.1).
Theorem 23.1. [AB09]
Db(CohG∨×C×(N )) Dmix
IW
Db(CohG∨(N )) DIW
∼
∼
Here DmixIW and DIW are the mixed and unmixed version of the antispherical category.
Remark 23.2. If one wants to convince oneself that this is a deep equivalence, one onlyneeds to note that the “easy” C×-action on the LHS corresponds to the “hard” Frobeniusaction on the RHS!
For the remainder of the lecture, we will build the machinery of Gaitsgory’s centralsheaves, then focus on the example of the natural representation of GL2. We will return tothe Arkhipov-Bezrukavnikov theorem next week.
23.1 A change of notation in the Hecke algebra
We will make a slight change of notation from previous lectures. In the affine Hecke algebra,replace
Hx δx
for all x ∈ Wf , so our standard basis of the finite Hecke algebra is now called δxx∈Wfand
the Kazhdan–Lusztig basis is bxx∈Wf. We make this change so that from now on, small
letters indicate objects in the Hecke algebra H and big letters indicate objects in the Heckecategory H.
23.2 Gaitsgory’s central sheaves
For more details on this construction, see notes from Emily’s IFS talks on the Beilinson-Drinfeld Grassmannian26.
26Talk notes available at https://sites.google.com/view/ifssydney/home
178
Recall that within the affine Hecke algebra H, we have a large commutative subalgebra
L =⊕λ∈X∨
Z[v±1]θλ ⊂ H.
In Lecture 12 we discussed Bernstein’s description of the center of H in terms of this subal-gebra.
Theorem 23.3.Z(H) ' LWf ' Z[v±1]⊗Z [RepG∨].
Another perspective from which we can view this description of the center of the affineHecke algebra was discussed in Emily’s IFS talk. Temporarily, let K = Fq((t)) ⊃ O = Fq[[t]],and build Hecke algebras
Haff = Func.s.I (F`,C), Hsph = Func.s.
G(O)(Gr,C).
Here “c.s.” stands for “compactly supported”, F` = G(K)/I is the affine flag variety andGr = G(K)/G(O) is the affine Grassmannian. These function spaces gain the structure ofalgebras via convolution. The spherical Hecke algebra Hsph is a commutative convolutionalgebra, Haff is not. There are natural maps
Haff
Func.s.G(O)(F`,C).
Hsph
∫G(O)/I
pull-back
Under these maps, the image of the center Z(Haff) and the image of Hsph agree:
Z(Haff)
images agree
Hsph
∼
∼
This provides an isomorphismZ(Haff) ' Hsph. (23.2)
The analogous statement for the affine Hecke algebra H is as follows. Let wf ∈ Wf bethe longest element, and
bwf =∑x∈Wf
v`(wf )−`(x)δx ∈ H
179
the corresponding Kazhdan–Lusztig basis element. Define
Hsph := (bwf ·H) ∩ (H · bwf ).
By the Satake isomorphism,Hsph ' LWf .
Then we have maps
H
H · bwf ,
Hsph
h7→h·bwf
identity
sending
Z(H)
images agree.
Hsph ' LWf
∼
∼
This gives us the isomorphism in Theorem 23.3:
Z∼−→ Hsph
h 7→ h · bwfGaitsgory lifted this statement to the level of categories. For the remainder of the lecture
set O = C[[t]],K = C((t)), Gr = G(K)/G(O), F` = G(K)/I. Recall that
π : F`→ Gr
is a G/B-fibration. We replace
Hsph PervG(O)(Gr), “Satake category,” and
Haff PervI(F`), “Hecke category ”.
Both of these categories obtain a monoidal structure, see Emily’s IFS talk. Gaitsgory up-graded the isomorphism 23.2 to a central functor.
Theorem 23.4. There exists a central functor27
Z : PervG(O)(Gr)→ PervI(F`).
Moreover, the diagram
DbI(F`)
DbI(Gr)
PervG(O)(Gr)
π∗
forget
equivariance
commutes.27This means that for all F ∈ PervG(O)(Gr),G ∈ PervI(F`), Z(F) ∗ G is perverse, so is G ∗ Z(F), and we
have a canonical isomorphism Z(F) ∗ G ' G ∗ Z(F).
180
Remark 23.5. 1. It is best to think of a “central functor to a monoidal category (M, ∗)”as a “functor to the Drinfeld centre of (M, ∗)”. Geordie is not 100% sure, but hethinks that Z actually lands in the symmetric center (i.e. set of objecs with symmetricbraiding) of the Drinfeld center.
2. Everything that Bezrukavnikov does is built on Z, so we should get used to thinkingabout it!
3. If G ′,G ∈ PervI(F`), then usually G ′ ∗ G is not perverse. It’s a miracle that Z(F) ∗ Galways is!
23.3 Extended example: the natural representation of GL2
For the remainder of this lecture, we will work out the details of this construction for the sim-plest non-trivial example: the natural representation of GL2. This example is very beautifuland instructive, so we will discuss it in detail.
Set G = GL2, G∨ = GL2, and nat = the natural representation of GL2. Let Fnat ∈PervG(O)(Gr) be the corresponding perverse sheaf under geometric Satake. Our goal is todescribe Z(Fnat). Here is an outline of what we’ll do:
Algebra: (See exercises at end of section.) To start, we have
HGL2 = HSL2 n 〈δ$〉 = 〈δs, δs0 , δ$〉 = Hf ⊗ Z[v±1][θ1, θ2].
Here δ$ is the generator of length zero elements, s ∈ Sf is the finite simple reflection, ands0 is the affine simple reflection. The elements θ1 = δ$δs and θ2 = δ$δ
−1s0
are the Bernsteingenerators. Hence
znat = θ1 + θ2 = δ$(δs + δ−1s0
) = δ$(δ−1s + δs0).
A natural question arises: How can we produce znat geometrically?
Geometry: Let k ∈ Q,R,C. Both F` and Gr have components which are indexed byZ (measuring the valuation of the determinant of the lattice/the polynomial degree of therepresentation under geometric Satake), and the projection
F`GL2 = F`SL2 × Z→ GrGL2 = GrSL2 × Z
is a P1 fibration. The natural representation corresponds to a sheaf kP1 [1] on GrSL2 × 1.Hence all of our arguments will take place on the components GrSL2 ×1 and F`SL2 ×1,so we can ignore the rest.
181
Gaitsgory’s family: Gaitsgory constructs a family of varieties over the disc:
The P1 × B/B on which the sheaf kP1 [1] is supported in the fibre Gr ×G/B degeneratesto P1 × P1 (the intersection of two Schubert curves) in F`:
We will see that znat is categorified by nearby cycles. Moreover, we will see that the “Waki-moto filtration”
categorifies the relationships
znat = δ$δs + δ$δ−1s0
= δ$δ−1s + δ$δs0
in H. Here it goes!
Beilinson gluing: We want to understand perverse sheaves on xy = 0 ⊂ P2; i.e. twoP1’s meeting transversally at a point. To do so, we can start by examining perverse sheaveson xy = 0 ⊂ C2, with stratification
Λ = 0 t C×x-axis t C×y-axis.
182
Beilinson gluing (Lectures 17 and 18) tells us how! Take f = x + y. The zero locus of frestricted to xy = 0 is 0. By Beilinson gluing,
PervΛ(xy = 0) =
F perverse on C×x t C×y ,V0 perverse on 0 +
ψf (F) ψf (V )
V0
µ−1
=
Vy Vx
V0
µy
cy
µx
cxvyvx
∣∣∣∣∣∣∣∣∣for all z ∈ x, y,
vz cz = µz − 1, andcy vx = 0 = cx vy
.
Hence if X = P1x ∪ P1
y = xy = 0 ⊂ P2 (these will be our intersection of Schubert varietieslater), and
Λ = 0 t Cx t Cy,
then we obtain the same description as above, except that the point at infinity forces µx =µy = 1. In other words,
PervΛ(X) =
Vy Vx
V0
∣∣∣∣∣∣∣∣∣vz cz = 0
for z ∈ x, y, andcy vx = 0 = cx vy
.
Exercise 23.6. Let
Y = ((x : y : z), λ) ∈ P2 × A1 | xy = λz2 ⊂ P2 × A1,
and Y0 = Y\f−1(0), where f : P2 × A1 → A1 is projection:
183
1. Show that ψfkY0 [2] is described under the equivalence above by the diagram
2. Deduce that ψfkY0 [2] has composition series
3. Check that the monodromy µ is given by
Remark 23.7. Let
Cx
j−→ X
j′
←− Cy
be the natural inclusions. Then
This is the Wakimoto filtration, which categorifies the relation
znat = δ$δs + δ$δ−1s0
= δ$δ−1s + δ$δs0 .
in H.
184
Connect this all to Gaitsgory’s picture: Let Etriv be the trivial rank 2 vector bundleon A1, and Ltriv = C[t]e1 ⊕ C[t]e2 the trivial lattice in C((t))2. In Emily’s IFS talks, sheintroduced the Beilison–Drinfeld Grassmannian, BD, which provides a Gr-fibration over A1:
BD :=
(x,E , β)
∣∣∣∣ E rank 2 vector bundle
β : E |X−x∼−→ Etriv|X−x
Gr−→ A1.
She also introduced Gaitsgory’s souped-up version of the the Beilinson-Drinfeld Grassman-nian, which in the example of GL2 is a F`-fibration over 0 and a Gr × P1-fibration overA1\0: (x,E , β,F)
∣∣∣∣∣∣E rank 2 vector bundle
β : E |X−x∼−→ Etriv|X−x
F flag in E0
0 A1\0
F` Gr×P1
Note that a flag in E0 is simply the choice of a line in the two-dimensional vector spaceE0. Within BD we have a finite-dimensional closed subvariety G ⊂ BD. Under the latticedescription of GrGL2 ,
G := L ⊂ Ltriv | dimLtriv/L = 1.
We can cover G with two charts: if λ is the coordinate on A1,
U0 =
⟨(1a
),
(0
t− λ
)⟩, U∞ =
⟨(b1
),
(t− λ
0
)⟩.
Each chart is isomorphic to A2, and we see that G is a trivial P1-bundle over A1.The analogous subvariety in Gaitsgory’s version is
Y =
(L, `)
∣∣∣∣∣∣L ⊂ Ltriv a lattice s.t.dimLtriv/L = 1, and
` ⊂ L/tL
P1-bundle−−−−−→ G.
Consider the following closed subvariety of Y :
YSp = (L, `) | ` ⊂ ker(L → Ltriv Ltriv/tLtriv = C⊕ C C⊕ C/(C⊕ 0)).
What are the fibres of YSp over A1? Well, for any (L, `) ∈ YSp, L fits into an exactsequence
L → Ltriv Ox,
where Ox is a skyscraper. There are two cases:
1. If x 6= 0, thenL0 ' (Ltriv)0.
Hence ` is uniquely determined, and the fibre is P1.
185
2. If x = 0, then
L0φ−→ (Ltriv)0
has rank 1. We have two possibilities: either (1) Imφ = C⊕ 0, in which case ` is free,but L is fixed, so the fibre is P1, or (2) ` = kerφ, in which case L is free and ` is fixed,so the fibre is again P1. We conclude that the fibre over 0 is P1
⋃0 P1, as we had
hoped.
In Geordie’s hand-written notes, there are charts which show that locally, this degener-ation is given by the following picture:
Furthermore, one can check that this degeneration is isomorphic to the degenerating quadricin P2 discussed earlier.
23.4 Exercises
These exercises will examine the structure of the affine Hecke algebra for GL2. We have
X = Ze1 ⊕ Ze2, X∨ = Ze∗1 ⊕ Ze∗2, Φ = ±(e1 − e2), Φ∨ = ±(e∗1 − e∗2).
Remark 23.8. The group GL2 is Langlands self-dual, so we don’t need to worry too muchabout which side of Langlands duality we are on.
The affine Weyl group isW = Wf n (Ze2 ⊕ Ze2) .
A picture:
186
Given γ ∈ Ze1 ⊕ Ze2, write tγ for the corresponding translation.
1. Show that the set of length zero elements
Ω = x ∈ W | x preserves strip
is a free abelian group generated by te1s ∈ W , where s = sα is the finite simplereflection.
2. Set $ = te1s, s0 = tαs. Check that $s = s0$, $s0 = a$, and $2 is central.
3. Show that the Bernstein generators are given by
θe1 = δ$δs, θ−1e2
= δ$−1δsθe2 = δ−1s δ$ = δ$δ
−1s0.
Hint:
4. Verify thatznat = θe1 + θe2 = δ$(δs + δ−1
s0) = δ$(δ−1
s + δs0)
is central.
5. Check that znat · bs = δ$bs0s.
6. (Challenge) Prove that Z(H) is generated by δ2$ and znat and show that
Z(H)∼−→ [Rep GL2]
znat 7→ C2
$2 7→ det
is an isomorphism of rings.
187
Lecture 24: A hitchhiker’s guide to the Hecke category
In Lecture 16, we motivated the constructible side of Bezrukavnikov’s equivalence via Grothendieck’sfunction-sheaf dictionary. The material from that lecture is useful background for today’slecture, and the reader may benefit by reviewing Lecture 16 before continuing.
Let X be a quasi-projective variety over Fq. There is a generating series built out of#X(Fqm) for m ≥ 1, the zeta function:
ζ(X, s) = exp
(∑m≥0
#X(Fqm)
mq−ms
).
Weil noticed that when X is a smooth projective n-dimensional variety, its ζ function hassome remarkable properties, which he packaged into the Weil conjectures:
1. Rationality: ζ(X, s) can be written as a rational function in q−s.
2. Functional equation: ζ(X, s) and ζ(X, 1−s) agree (up to a simple and explicit scalar).
3. Riemann hypothesis: roots of ζ have a specific form.
Example 24.1. If X = P1Fq , then
ζ(X, s) = exp
(∑m≥0
1 + qm
mq−ms
)
= exp
(∑m≥0
(q−s)m
m+∑m≥0
(q1−s)m
m
)
=1
1− q−s· 1
1− q1−s .
We see from this that: (1) ζ(X, s) is rational, (2) ζ(X, s) = −q1−sζ(X, 1 − s), and (3) wecan write
ζ(X, s) =1
P0(q−1)P2(q−s),
with P0 = (1 − T ) and P2 = 1 − qT . Here the Riemann hypothesis is the (elementary)statement that all roots of P0 (resp. P2) have roots of norm 1 (resp. q).
Example 24.2. If X is a smooth elliptic curve, then
ζ(X, s) =P1(q−s)
P0(q−s)P2(q−s)=
P1(q−s)
(1− q−s)(1− q1−s),
where P1 has two roots, which are conjugate and have norm q1/2 (“Weil numbers of weight1”).
Remark 24.3. The case of an elliptic curve (due to Hasse) predates the Weil conjecturesand was an important ingredient in their formulation. Another important ingredient wasArtin’s computation of the zeta function of a hyperelliptic curve, discussed in Lecture 4.
188
Remark 24.4. We saw the significance of the roots of P1 (which determine how many pointsE has) in Lecture 4 on the Sato-Tate conjecture.
In the Weil conjectures, rationality follows from the Grothendieck-Lefschetz trace for-mula:
#X(Fq) = supertrace(Frob H∗et(X,Q`)
).
The functional equation follows from Poincare duality in etale cohomology, and the Rie-mann hypothesis (the most difficult part) follows from purity. This is the statement thatthe eigenvalues of Frobenius in degree i are all “Weil numbers of weight i”.
Remark 24.5. Notice the strange appearance of the auxiliary prime ` 6= p in the Grothendieck-Lefschetz trace formula. Why do we need to make this choice? Such questions are usuallylabelled questions of “independence of `” in the literature. We will also see such questionsarise in the Hecke category below.
The proof of the Weil conjectures needed the full artillery of Dbc(X,Q`) and its six functor
formalism. It used a relative version of the Grothendieck-Lefschetz trace formula, whichassociated to a sheaf F a collection of functions given by “trace of Frobenius”:
F ∈ Dbc(X,Q`)
fmF : X(Fqm)→ Q`
.
Then, as we saw in Lecture 16,[Dbc(X,Q`)
]→∏m≥1
Fun(X(Fqm)→ Q`),
so the collection of functions associated to F completely determines its class in the Grothendieckgroup. This leads to Grothendieck’s philosophy of “dictionaire functions faisceaux”:
functions ↔ sheaves.
This philosophy can be summed up with the slogan
“interesting functions should arise from interesting sheaves”.
24.1 The Hecke category: setting the scene
Let G be split reductive over Fq, and B ⊂ G a Borel subgroup. Let K = Fq((t)), and denoteby GK the loop group and I ⊂ GK the corresponding Iwahori subgroup. Recall that theHecke algebra (either finite or affine) has its origins as a convolution algebra of bi-invariantfunctions on a group (see Lecture 16 for more on this perspective). Using Grothendieck’sfunction-sheaf dictionary, a natural categorification of these functions is the following:
Hf “ = ”(FunB(Fq)×B(Fq)(G(Fq),C), ∗
) (DbB×B(G,Q`), ∗
)v 7→ 1√
q
categorify
189
H “ = ”(FunI(Fq)×I(Fq)(GK(Fq),C), ∗
) (DbI×I(GK,Q`), ∗
)v 7→ 1√
q
categorify
Remark 24.6. The loop group GK is very infinite-dimensional, so we usually work with(DbI(F`,Q`), ∗
),
instead. (Here F` = Gk/I is the affine flag variety.)
Remark 24.7. One can see from the above that Hf is “independent of q”. We have oneabstract algebra which specialises to all Hecke algebras at once. This is one desirable featurethat we are hoping to categorify.
Our goal for this lecture is to find a good categorification of the Hecke algebra. It mightappear that we’ve already accomplished this. However, the categories
(DbB×B(G,Q`), ∗
)and(
DbI×I(GK,Q`), ∗
)are not quite right for several reasons. In the remainder of the lecture, we
will explain why these categories are wrong, then slowly fix them. We will concentrate onthe finite case (i.e. the Hecke algebra for the finite Weyl group). The affine case is similar.
What we want: A triangulated monoidal category H such that:
1. ([H], ∗) ' H (i.e. H categorifies the Hecke algebra),
2. H is “independent of q and `,” and
3. H admits a triangulated monoidal functor to DbB×B(G,Q`) such that the diagram
H DbB×B(G,Q`)
H(FunB(Fq)×B(Fq)(G(Fq),C), ∗
)v 7→ 1√q
commutes for all ` and q.
Remark 24.8. The existence of such an object is tacitly implied by Bezrukavnikov’s equiv-alence, as ` and q are nowhere to be seen on the coherent side.
Now we try and fail and try and fail and try and fail, and then, finally, succeed.
24.2 First try
We start with the most obvious choice:
H = DbB×B(G,Q`).
Objections:
190
1. Depends on q and `.
2. If G is the trivial group, then
H = Dbc(SpecFq,Q`) → Db
c
continuous representations
of πet1 (SpecFq) = Z on
Q`-vector spaces
.
This is almost an equivalence. In fact, one has
[H] = Z[Z×` ], (24.1)
which is way too big. (See [DBB83, Proposition 5.1.2] and the remark following it.)
Note that the fact that this Grothendieck group is way too big persists for any group:rather than being an algebra over Z[v±1], our putative definition is an algebra over(24.1).
24.3 Second try
In the first try, we failed to obtain requirement 1 of our desired categorification because ourcategory had a Grothendieck group which was too big. We can attempt to fix this by passingto the algebraic closure. Try
H = DbB×B(G,Q`) for G/Fq, or
H = DbB×B(G,Q) for G/C.
Objections:
1. If G is the trivial group, then
DbB×B(G,Q) = Db
(finite dimensional
vector spaces
).
Hence[H] = Z,
which is too small. (We expect Z[v±1] for the Grothendieck group in this case.) Butperhaps this is just a trivial-case phenomenon, we can test with a slightly bigger ex-ample.
2. Recall (c.f. Lecture 16) that in SL2, the quadratic relation in FunB×B(SL2(Fq),C) camefrom the fact that
1SL2(Fq) ∗ 1SL2(Fq) = (1 + q)1SL2(Fq).
Why was this again? Well, the multiplication map mult : SL2×B SL2 → SL2 factorsthrough SL2 /B × SL2 = P1 × SL2:
SL2×B SL2 P1 × SL2
SL2
mult
∼
projection
191
Here the horizontal arrow is the map (g, h) 7→ (gB, gh). So we obtain the formulaabove by pushing forward the constant sheaf on either side of the isomorphism, andthe (1 + q) comes from #P1(Fq) = 1 + q.
In DbB×B(SL2,Q), the same diagram shows
QSL2 ∗QSL2 = p∗QP1×SL2= H∗(P1)⊗QSL2 = QSL2 ⊕QSL2 [−2]
In the Grothendieck group, this gives
[QSL2 ]2 = 2[QSL2 ]. (24.2)
Exercise 24.9. The map
Z[W ]→[DbB×B(SL2,Q)
]s 7→ [QSL2 ]− [QB/B]
is an isomorphism of algebras.
We see from the exercise that the trivial group wasn’t just an anomaly - with this defi-nition of H, the Grothendieck group really is too small. Also notice that we want to replacethe 2 in (24.2) by (1 + q). With the Grothendieck-Lefschetz trace formula in mind, we wishto replace an Euler characteristic (2) by the trace of Frobenius (1 + q). Therefore we are ledto the following conclusion:
The moral: Somehow we need to remember weights!
24.4 Third try
Again, let G/Fq. This time, we set
H = 〈ICx | x ∈ Wf〉[Z],(Z),∆ ⊂ DbB×B(G,Q`).
In this incarnation ofH, we are keeping track of weight by introducing “Tate twists,” denotedabove by (Z).
What is a Tate twist? Define
Q`(−1) := H2c (A1) = Q` ∈ Shet(SpecFq).
On Q`(−1), Frob acts via multiplication by q. Let p : X → SpecFq. For F ∈ Dbc(X,Q`), set
F(m) := F ⊗(p∗(Q`(−1)⊗(−m)
).
(Here we are using that Q`(−1) is invertible, as an etale sheaf on a point. Thus it makessense to take any integral tensor power.)
Very nontrivial fact: This definition of H is closed under convolution. (For the experts:This is a consequence of the Decomposition Theorem and the fact that the objects are“already semi-simple over Fq”.)
192
Remark 24.10. (Technical point) In the Hecke algebra, it is very useful to introduce asquare root of q. To reflect this, we often choose a square root of q in Q` and use it to defineQ`(−1/2), the “square root of Tate twist”. This can also be done purely formally if oneprefers.
Now we are in good shape: If G is the trivial group,
H =⟨Q`
(m2
)∣∣∣m ∈ Z⟩
∆⊂ Db(SpecFq).
Hence the Grothendieck group is
[H]∼−→ Z[v±1], v 7→ Q`(−1/2).
So this H passes our first test. Great!Using Grothendieck’s theory of weights, there is an alternative version of this third try.
Recall (see “Scholze’s motivic plane” in Lecture 10):
motives
etale
Q`-sheavesSaito’s mixed
Hodge modules
Frobeniusactions on etale
cohomology
mixed Hodgestructures
Weilconjectures
X/Fq X/C Hodgetheory
Under the red arrow above, we have a second incarnation of our third try:
H = 〈ICx | x ∈ Wf〉[Z],( 12Z),∆ ⊂ Db
B×B
(mixed Hodge
modules
).
This version of H also has the correct Grothendieck group. It is closed under convolutionby the Decomposition Theorem.
A subtlety: These two incarnations Hetale and HmHm of our third attempt are reallydifferent. If G is the trivial group, then
• in Hetale,
Ext1(Q`,Q`
(m2
))=
Q` if m = 0;
0 otherwise.
193
A non-trivial self-extension of Q` is given by Q2
` , with Frobenius acting by Jordan
block
(1 10 1
). Hence, there are extensions between objects of the same weights, but
no extensions between objects of different weights.
• However, in HmHm, there are no extensions between objects of the same weight, butthere are extensions between objects of different weights28.
Remark 24.11. Under Bezrukavnikov’s equivalence, we expect
weight↔ weight of C×-action.
So there should be no Exts at all for the trivial group!
A historical incarnation of this issue: LetO0 be the principal block of categoryO. Thereis a finite-dimensional C-algebra A such that O0 ' A-modf.g.. In 1990, Soergel showed that
A admits a Z-grading A defined over Q, and A-gmodf.g. provides a grading on categoryO. This grading explains the q in the Kazhdan–Lusztig polynomials. Hence we have thefollowing diagram:
A-modf.g. A-gmodf.g.
O0certainD-modules
Perv(B)(G/B)
∼
∼ ∼
What is the geometric meaning of the grading on O0 coming from A? This was explainedin [BGS96]:
AQ`-gmod ' some geometriccategory
↔ DbB(G/B,Q`),
AC-gmod ' some geometriccategory
↔ DbB(MHMG/B).
In each case, the graded version is explained by some category related, but not equal, tomixed sheaves. In both cases, some cooking is involved to remove the problematic Exts fromearlier. The “cooking” is different in each case.
Remark 24.12. Recent motivic versions (Soergel-Wendt [SW18] and Soergel-Wendt-Virk,[SVW18]) explain how to remove the cooking.
28Stating exactly what these extension groups are would require us to be more precise about what versionof mixed Hodge structures we are using (integral, real, complex, etc.). However, the basic phenomenon thatdifferent weights extend persists in all of them.
194
24.5 Fourth (and final) try
We have finally converged on the correct formulation. Let
Hs.s = 〈ICx | x ∈ Wf〉⊕,[Z] ⊂ DbB×B(G,Q).
This is an additive category, but it is not triangulated. Thus it consists of all semi-simplecomplexes in Db
B×B(G,Q). It is closed under convolution by the Decomposition Theorem.We need to make make one necessary change of notation.
Redefine: (1) := [1] on Hs.s..
Definition 24.13. The geometric Hecke category is
H := Kb(Hs.s.).
This category has two shift functors: (n) denotes the shift in Hs.s. and [n] denotes the shiftin Kb.
Before declaring victory, we should verify that this satisfies all of our desired propertiesfor the trivial group. If G is the trivial group, then
Hs.s. = Dbc(pt) = Db
(finite-dimensional
vector spaces
)=
finite-dimensionalgraded vector spaces.
(Here the final equality is as additive categories.) Hence
H = Kb
(finite-dimensional
graded vector spaces
).
Note that this category has no extensions, so our woes of our third attempt have disappeared.Moreover, we have
[H] = Z[v±1],
with v corresponding to the shift of grading. Moreover, there are no q’s or `’s in sight.Hurrah!
Remark 24.14. 1. We have [H] = [Hs.s.] = Hf .
2. A similar definition over Fq with Q`-sheaves leads to an equivalent category (afterextensions of scalars to Q`). This proves independence of q and `. The proof is byshowing that Hs.s. ' SBim, the category of Soergel bimodules29.
3. With difficulty, one can construct monoidal realisation functors:
(Hetale, ∗)
(H, ∗)
(HmHm, ∗)29We hope to explain this sometime in the future!
195
This provides considerable evidence that H is the right object.
Remark 24.15. It is possible that although H is the right object, it may not yet have theright definition. Recent motivic versions [SVW18] probably provide the “correct” definition.
196
Lecture 25: The categorical anti-spherical module and
its symmetries
Recall that we are working towards proving the following theorem of Arkhipov–Bezrukavnikov:
Theorem 25.1. [AB09] There exists an equivalence of triangulated categories
DG∨×Gm(N ) Masph.∼
This equivalence categorifies the isomorphism
KG∨×Gm(N ) H ⊗Hf sgn.∼
Today we will define the category Masph and discuss the philosophy of higher represen-tation theory.
25.1 The affine Hecke category
For the remainder of this course, k ∈ Q,R,C. Last lecture in our quest to find the correctdefinition of the geometric Hecke category, we defined
Hssf := 〈ICx | x ∈ Wf〉⊕,[Z] ⊂ DB×B(G, k).
Here “ss” stands for semi-simple. This category is additive, but not triangulated.
Remark 25.2. (Technical point) It is convenient to normalize ICx so that it correspondsto the usual IC sheaf on G/B under the equivalence
DB×B(G, k) ' DB(G/B, k);
e.g.ICid = kB, ICs = kPs [1], . . . , ICx|BxB = kBxB[`(x)].
We used Hssf to define the geometric Hecke category:
Hf := Kb(Hssf ).
This is a triangulated category (in contrast to Hssf ). It has two natural shift functors:
[1] = triangulated shift functor, and
(1) = functor induced by the shift functor on Hssf .
A picture of these shifts on an object F ∈ Hf is
Here each F i ∈ Hssf .
197
Theorem 25.3. The map bs 7→ [ICs] induces an isomorphism of Z[v±1]-algebras
Hf
[Hssf
]⊕ ' [Hf ]∆.
∼
Here the subscript ⊕ denotes the split Grothendieck group of an additive category, andthe subscript ∆ denotes the triangulated Grothendieck group of a triangulated category.Under this isomorphism, the Kazhdan–Lusztig basis element bx corresponds to the class ofan intersection cohomology complex, in formulas
bx 7→ [ICx].
The same construction works in the affine case. Define an additive category
Hss := 〈ICx | x ∈ W 〉⊕,[Z] ⊂ DbI(F`, k)
for F` := G((t))/I, and a triangulated category
H := Kb(Hss).
The analogue of Theorem 25.3 holds:
Theorem 25.4.H ' [Hss]⊕ ' [H]∆, bx 7→ [ICx].
25.2 The categorical (anti-)spherical module
In the Hecke algebra, we have the quadratic relation for each s ∈ S:
δ2s = (v−1 − v)δs + 1, or reformulated, (δs + v)(δx − v−1) = 0.
This leads to the existence of two Hf -modules, sgn and triv, where δs acts by −v and v−1,respectively, and two corresponding induced modules for H:
M sph := triv ⊗Hf H the spherical module, and
Masph := sgn⊗Hf H the antispherical module.
Remark 25.5. We have made a notational switch from denoting the anti-spherical moduleby N (c.f. Lecture 21) to Masph. This is because we wish to denote categorifications by thecorresponding script letters and the symbol N has already been assigned to the nilpotentcone.
Remark 25.6. We have also made a switch from left to right modules. From now on,the spherical and anti-spherical module will be considered as right H-modules to align with[AB09].
198
In Lecture 20, we discussed how as right modules over the lattice part L =⊕
λ∈X∨ Z[v±1]θλof H, M sph is a free module of rank 1:
M sph = triv ⊗Hf H = triv ⊗Hf Hf ⊗ L ' L.
A similar statement holds for Masph. This description of the (anti-)spherical module, whichfollows from the Bernstein presentation of the affine Hecke algebra, is the “coherent per-spective” of M (a)sph. On the other hand, using the Coxeter structure of W , we can viewM (a)sph from a “constructible perspective” as follows. Any w ∈ W can be written as wfin
fwfor wfin ∈ Wf and fw ∈f W , where fW is the set of minimal coset representatives in Wf\W .Hence we have an isomorphism
Hf ⊗
(⊕x∈fW
Z[v±1]δx
)∼−→ H,
and M (a)sph has a “standard basis”
δ(a)sphx := 1⊗ δx | x ∈ fW.
Exercise 25.7. (Fun!) Compute the bijection
dominantalcoves
∼←→ fW∼←→ Wf\W = Wf\(Wf n ZΦ∨) = ZΦ∨
in a few examples (e.g. for SL2, SL3, . . .).
Using Kazhdan–Lusztig combinatorics, we can give descriptions of the spherical and anti-spherical module in terms of the Kazhdan–Lusztig basis of H. We dedicate the next part ofthe lecture to doing this.
Exercise 25.8. Prove the following:
1. If x ∈ W and s ∈ S are such that xs < x, then bxbs = (v + v−1)bx.
2. If t ∈ S and x ∈ W are such that tx < x and s ∈ S is arbitrary, then in thedecomposition
bxbs =∑
nyby,
ny 6= 0 implies ty < y. Conclude from this that bx | tx < x span a right ideal in H.
Lemma 25.9. (realisations of M sph and Masph via Kazhdan–Lusztig theory) Let wf be thelongest element in Wf .
1. The map 1⊗ 1 7→ bwf induces an isomorphism of right H-modules
M sph ∼−→ bwfH.
2. The Z[v±1]-span of bs | s 6∈ fW is a right ideal. Moreover,
Masph ' H/〈bx | x 6∈ fW 〉.
199
Proof. By Exercise 25.8.1, bwf bs = (v + v−1)bwf for all s ∈ Sf , so
bwf δs = bwf (bs − vδid) = v−1bwf
for all s ∈ Sf . By Frobenius reciprocity, we have a map
M sph φ−→ bwfH, δsphid 7→ bwf .
For x ∈ fW
δsphx
φ7−→ bwf δx = δwfx + lower terms.
By upper triangularity, we conclude that φ is an isomorphism. This proves 1.By Exercise 25.8.2, 〈bx | x 6∈ fW 〉 is a right ideal in H. Hence H/〈bx | x 6∈ fW 〉 is a right
H-module. In H/〈bx | x 6∈ fW 〉,0 = 1 · bs = δs + v
for s ∈ Sf , so1 · δs = −v · 1.
By Frobenius reciprocity, we have a map
Masph → H/〈Bx | x 6∈ fW 〉,
and an analogous argument to the one above shows that this map is an isomorphism. Thisproves 2.
Remark 25.10. One can use the other Kazhdan–Lusztig basis b′x to realise M sph as aquotient and Masph as a submodule. This is categorified by Koszul duality.
Now it makes sense to define the categorical versions of M sph and Masph:
Msph,ss := 〈ICwf ∗ Hss〉 ⊂ Hss,
Msph := 〈ICwf ∗ H〉∆ ⊂ H,Masph,ss := Hss/〈ICx | x 6∈ fW 〉⊕,(Z),
Masph := H/〈ICx | x 6∈ fW 〉∆,(Z).
Here the quotients are quotients of additive categories, and denotes the closure underdirect summands.
Lemma 25.11.Masph = Kb(Masph,ss).
Theorem 25.12. The categoryM(a)sph is a right H-module. The map δasphid 7→ ICid induces
an isomorphism of right H-modules
M (a)sph ' [M(a)sph,ss]⊕ = [M(a)sph]∆.
Remark 25.13. In future lectures, we will discuss in more detail what it means for acategory to be a module over a monoidal category like H.
200
25.3 Symmetries of categories
The remainder of this lecture will be a discussion of higher representation theory. We startat the beginning: Representation theory is the study of linear symmetry (of groups,algebras, etc.). This field stems from the underlying observation:
Group actions are difficult, but they become easier once we linearize.
Example 25.14. 1. Linearizing the action of S1 S1 (resp. SO(3) S2) leads to thetheory of Fourier series (resp. spherical harmonics).
2. If X is a variety over Q, studying X(Q) is hard. This can be expressed as a questionabout the action of the absolute Galois group of Q on all points of X(Q). A centraltechnique in modern number theory is instead to study the linear problem Gal(Q/Q) H i(X,Q`).
2-representation theory is the study of symmetries of linear categories (additive,abelian, triangulated, etc.). These symmetries take the form of monoidal categories.
Philosophy: (learned from Manin) In differential geometry, if we have a group acting ona manifold G M , we get a lot of leverage by linearizing and studying G L2(M,C). Inalgebraic geometry, if we want to do the same thing for a group G acting on an algebraicvariety X, a useful technique is to study is G Coh(X),QCoh(X), Db(Coh(X)). Thecategories QCoh(X),Coh(X), Db(Coh(X)) can be thought of as 2-linearizations of X.
Remark 25.15. (Side remark, c.f. [Man14]) Let F be a sheaf on X. Early approaches tostudying such objects (∼ 1950s) emphasized Cech coverings. Later, Grothendieck shiftedthe emphasis to injective resolutions. In sheaf cohomology
H i(X,F)
there is a non-linear variable (X), and a linear variable (F). We can interpret Grothendieck’sshift as a movement from the non-linear variable to the linear variable.
Long range hope: From the perspective of 2-representation theory, our current goal is toshow that the natural symmetries of
DG∨×Gm(N ) and Masph
agree, and yield the Hecke category.
A priori the symmetries on each side look rather different:
RHS HLHS RepG∨,RepGm,PicG
∨×Gm(N ), . . .
201
25.4 A notational interlude:
We will use the language of stacks.
• All stacks we study will be of the form Y/H, where Y is a quasi-projective variety andH is an affine algebraic group.
• See Emily’s IFS talk “Sheaves on Stacks” for an excellent introduction to coherentsheaves on stacks.
• The most important fact (which will be used repeatedly) is
QCoh(Y/H) ' QCohH(Y ),
Coh(Y/H) ' CohH(Y ).
• Basic observation: If Y is a stack as above, then
1. QCoh(Y ) and the category of vector bundles on Y are symmetric monoidal cate-gories.
2. If Xf−→ Y is a map of stacks, then QCoh is a module over QCoh(Y ) via
F ∗ G := f ∗(F)⊗OX Gfor F ∈ QCoh(Y ), G ∈ QCoh(X). (Note that here ∗ denotes the action map ofQCoh(X) and not convolution.)
25.5 The basic approach of Arkhipov-Bezrukavnikov
The basic approach of [AB09] is the following. We wish to show that
DG∨×Gm(N ) 'Masph.
(For the moment, we ignore Gm.) We have the following maps of stacks:
N /G∨ QCoh(N /G∨)
g∨/G∨ QCoh(g∨/G∨)
pt/G∨ QCoh(pt/G∨)
q
p
q∗
p∗
The maps of stacks in the column on the left induce the increasing chain of symmetricmonoidal categories in the column on the right. Step by step, we will construct an action ofthe symmetric monoidal categories in this chain onMasph. The action of QCoh(pt/G∨) willcome from Gaitsgory’s central functors, the action of QCoh(g∨/G∨) from the monodromy
endomorphism of nearby cycles, and the action of QCoh(N /G∨) from “Wakimoto arrows”(which are yet to be defined). This procedure yields a functor
QCoh(N /G∨)→Masph,
which we will argue is an equivalence.
202
Lecture 26: Monoidal categories and their actions
In representation theory, we study
A
algebra
M
vector space.
This process lets us draw conclusions about M (composition series, semi-simplicity, etc.)In 2-representation theory, we study
Amonoidal category
Mcategory
.
What does this process let us conclude about M? Today we’ll examine this question.
A theme of this lecture: There are two approaches to higher algebra:
(a) Generators and relations: carry “just enough” coherence data, or
(b) Holistic: carry “all” coherence data.
Approach (a) is rigid, but (sometimes) computable; whereas approach (b) is flexible but(often) incomputable. Historically, (a) is usually developed first, but (b) proves itself tobe more powerful in the long run. The following familiar example from algebraic topologyillustrates this.
Example 26.1. Given a reasonable topological space X, there are two approaches to com-puting homology H∗(X):
(a) triangulate your space as efficiently as possible and compute homology using a “small”complex (simplicial homology), or
(b) consider all possible singular n-simplices and build an enormous chain complex fromthe spaces Hom(∆n, X) (singular homology).
The first approach is by “generators and relations,” whereas the second is “holistic”.
Remark 26.2. (For those who know about simplicial sets) The two approaches in Example26.1 underly two approaches to studying the homotopy theory of X.
26.1 What is a monoidal category?
There are two approaches.
Generators and relations: A monoidal category is a category A equipped with a bifunctor
⊗ : A×A → A,
a unit1 ∈ A,
203
and natural transformations
αX,Y,Z : (X ⊗ Y )⊗ Z ∼−→ X ⊗ (Y ⊗ Z),
λX : 1⊗X ∼−→ X, and
ρX : X ⊗ 1∼−→ X
such that the diagrams
((X ⊗ Y )⊗ Z)⊗W
(X ⊗ (Y ⊗ Z))⊗W (X ⊗ Y )⊗ (Z ⊗W )
X ⊗ (Y ⊗ (Z ⊗W ))
X ⊗ ((Y ⊗ Z)⊗W )
αX,Y,Z⊗W αX⊗Y,Z,W
αX,Y⊗Z,W
αX,Y,Z⊗W
X⊗αY,Z,W
and
(X ⊗ 1)⊗ Y X ⊗ (1⊗ Y )
X ⊗ YρX⊗idY
ρX⊗λX
idX⊗λY
commute for all objects X, Y, Z,W . We refer to the first diagram as the pentagon and thesecond as the unit. With this set-up, we can also formulate the notion of a monoidal functorbetween monoidal categories, though we won’t state this definition precisely today.
Basic claim: (“MacLane’s coherence theorem”) Any two maps consisting of associatorsand units agree.
Holistic approach: It’s possible to give an alternative (and perhaps better?) definition ofa monoidal category using the holistic approach, it can be found in [Lur07, §1.1]. This defi-nition is complicated so we won’t state it precisely here, but the rough idea is the following.
Let A⊗n be the category of sequences of n-elements in A. A monoidal category is equiv-alence to a whole host of functors of the flavor
A⊗5 → A⊗3
(X1, X2, X3, X4, X5) 7→ (X1 ⊗X2, 1, X3 ⊗X4 ⊗X5),
together with many many compatibilities.
26.2 Modules for monoidal categories
Similarly, we can formulate the notion of a module for a monoidal category either in termsof generators and relations, or holistically.
204
Generators and relations: A left module M for a monoidal category A consists of afunctor
⊗ : A×M→M,
together with natural transformations
βX,Y,M : (X ⊗ Y )⊗M ∼−→ X ⊗ (Y ⊗M),
λM : 1⊗M ∼−→M, and
ρM : M ⊗ 1∼−→M
such that the analogous pentagon and unit diagrams commute for all objects X, Y ∈ A andM ∈M.
Holistic: A similar approach is possible, see [Lur07, §2].
In representation theory, we are familiar with the fact that representations of an al-gebra are equivalent to modules over that algebra. An analogous statement holds in 2-representation theory. A representation of A is a pair (M, φ) of a category M and a ⊗-functor φ : A → End(M).
Proposition 26.3.A-modules ' A-representations .
Exercise 26.4. (Which Geordie assumes is true but hasn’t done.) Check that this is anequivalence of 2-categories.
26.3 Examples of modules over monoidal categories
In representation theory, we started by studying groups, then realized that many of ourquestions about linear group symmetry could be formulated in terms of representations ofalgebras. In a similar vein, we can start 2-representation theory by building categories fromgroups.
Given a discrete group G, construct a category AG as follows:
• Objects: rg | g ∈ G satisfying rg ⊗ rh = rgh, 1 = rid, and α, λ, ρ are all the identity.
• Morphisms: Hom(rg, rh) = ∅ for g 6= h, and End(rg) = idrg.
What is a module over AG? Given a category M and a ⊗-functor F : AG → End(M),define Fg := F (rg). Then the ⊗-functor F gives us natural isomorphisms
µgh : FgFh∼−→ Fgh and ε : Fid
∼−→ idM
such that
FgFhFk FgFhk
FghFk Fghk
Fgµh,k
µg,hFk µg,hk
µgh,k
,
FidFg Fg
Fg
µid,g
εFgid
, and
FgFid Fg
Fg
µg,id
Fgεid
commute for all g, h, k ∈ G. This is called a strict action of G on M.
205
Exercise 26.5. Show that giving an action of Z on M is the same as giving an autoequiv-alence E of M.
Example 26.6. (An example to urge caution) What does it mean to give an action of Z/2Zon M?
• First guess: A category M with an autoequivalence E : M → M together with anatural isomorphism m : E2 ∼−→ idM.
This is wrong! Here’s why: It follows from the definition of a categorical action thatthe diagram
EEE E
E
mE
Em
id(26.1)
commutes. (This is a consequence of the previously displayed commutative square.)But we can cook up an example of data as above where it doesn’t. Let M be thecategory with two objects X and Y and morphisms
Hom(X, Y ) = Hom(Y,X) = 0, End(X) = End(Y ) = k,
where k is a field. Define E :M→M by E(X) = Y , E(Y ) = X, and idXE7−→ idY
E7−→idX . A picture of this category:
Note that E2 = idM, so any m : E2 ∼−→ idM is a natural transformation m : idM →idM; i.e. m ∈ Z(M). Hence
m =
a on End(X)
b on End(Y )
for some a, b ∈ k. Now in this example, the diagram (26.1) on X becomes
Y Y
Y
a
bidY
This diagram commutes if and only if a = b. However, any choice of a and b defines anatural transformation m : E2 → idM, so our first guess must be wrong!
• Revised guess (Exercise): Show that the data of our first guess along with therequirement that (26.1) commutes does determine an action of Z/2Z on M.
206
26.4 What is going on here?
We will briefly take a small diversion and discuss a beautiful dictionary related to theseconcepts. As we discussed in the beginning of this lecture, in the generators and relationsapproach to categorical actions, we want to include “just enough” coherence data. Example26.6 illustrates that sometimes it can be subtle to determine how much coherence data isenough. The reader might be wondering:
How do we determine the necessary coherences?
It turns out that they are determined by the “cells of BG”. We will illustrate what wemean by this through a series of examples.
Example 26.7. Associated to G is its classifying space BG30.
1. For G = Z, BG = S1. This is a CW complex consisting of one 0-cell and one 1-cell.Recall that in Exercise 26.5 we showed that an action of Z on a category M is givenby an autoequivalence E :M→M:
Hence we have one piece of “coherence data” for each cell of BG.
2. For G = Z/2Z, BG = RP∞ = S∞/±1. The sphere S∞ has a CW complex structurewith two n-cells for each n ∈ Z≥0:
Moreover, the action of Z/2Z on this cell complex is cellular. Hence RP∞ is a CWcomplex with one n-cell for each n ∈ Z≥0. We saw in Example 26.6 that an action ofZ/2Z on M is determined by (M, E,m : E2 ' idM, (26.1)):
30Recall that BG is only defined up to homotopy. We will ignore this point below.
207
Again, for each n-cell with 0 ≤ n ≤ 3, we have a corresponding piece of “coherencedata”.
3. For any group G, the Milnor construction of BG yields a CW complex strucure withone 0-cell, G 1-cells, G × G 2-cells, G × G × G 3-cells, etc. Comparing this to thedefinition of a G-action on M in Section 26.3, we see:
So each of piece of coherence data in the definition of a G-action corresponds to ann-cell of BG.
In each example above, the CW complex structure gives us the necessary coherence datato determine the action of G. In parts 1 and 2, we used a CW complex with a small numberof cells to obtain “just enough” coherence data, illustrating the “generators and relations”approach. In part 3, we used a CW complex with many cells to capture “all” coherence dataof the G-action, illustrating the “holistic” approach.
This pattern is very pretty, but why do we stop at 3-cells? This is because we areacting on 1-categories. In general,
26.5 Example: Representations of the Verlinde category
For some reason, an action of the category AG is not as powerful as a G-action on a vectorspace.
Reasons: (Speculative)
1. We need more structure to categorify linear algebra. For example, what is the eigen-value of a functor? If Z M, how do we take its logarithm? In examples of categoricalactions (e.g. of braid groups in highest weight representation theory), there are ofteninteresting morphisms Fg → Fh and these need to be studied.
208
2. It is possible that this isn’t the right generalization. In this course, a much moreimportant action will be
RepG M,
where G is a linear algebraic group. In other words, we will study “representations ofrepresentations of G”.
Next lecture we will discuss the general case; today we will give a simple and beautifulexample. To do so we need to introduce the Verlinde category.
Recall that for each m ≥ 0, sl2(C) has a unique simple module Vm of highest weight mand dimension m + 1. In the category Rep sl2(C), the module V0 is the monoidal unit, andother representations multiply according to the rule
V1 ⊗ Vm ' Vm ⊗ V1 ' Vm+1 ⊕ Vm−1. (26.2)
For any positive integer ` ≥ 1, there exists a monoidal category C` which can be seen asa “finite” version of Rep sl2(C). It has ` + 1 simple objects, V0, . . . , V`, V0 is the unit, andmultiplication31 is given by
V1 ⊗ Vm = Vm ⊗ V1 = Vm+1 ⊕ Vm−1 for 1 ≤ m < `, and V` ⊗ V1 = V1 ⊗ V` ' V`−1. (26.3)
Remark 26.8. There are two realizations of C` (see [Ost03], [Kac90]). The first is as the
category of level ` representations of the affine Lie algebra sl2(C) with its fusion product. Thesecond is as the semi-simplification of representations of the Lusztig form of the quantumgroup of sl2 at a root of unity. In both settings, C` is braided but (in contrast to Rep sl(2,C))not symmetric. The braiding will not play a role below.
Theorem 26.9. Let A = C`. Thensemi-simple abelian A-modules M such that1. M has finitely many simple objects, and2. M is generated by a single simple object
/'
∼←→
simply laced
Dynkin diagramswith Coxeter number `+ 2
.
Example 26.10. The category C10 admits 3 module categories on the right hand side,corresponding to the A11, D7, and E6 Dynkin diagrams.
Proof. (Sketch) Let us explain how to go from the LHS to RHS. Starting with M on theLHS, we can build a graph as follows (cf. McKay correspondence):
• Vertices: One vertex for every simple module in M.
• Edges: An edge between M and M ′ if M is a summand of V1 ⊗M ′.
Just as in Rep sl2(C), V1 is self-biadjoint in A, so
M is a summand of V1 ⊗M ′ ⇐⇒ M ′ is a summand of V1 ⊗M.
31One can remember (26.3) by noticing that it is the same as (26.2), except that all classes Vm for m > `are zero.
209
Hence the graph associated to any A-module M in this way is undirected.From this graph, one can use (26.3) to deduce how Vm acts on the Grothendieck group
[M] for all m. For example, if D denotes the adjacency matrix of the graph associated toM,then the action of V2 on M is described by D2 − id because [V2] = [V ⊗2
1 ]− [V0] in C`. Nowsome linear algebra allows one to deduce that the graph we construct from M must either
be a simply laced affine Dynkin diagram or a “tadpole” . Then we can work harder to
rule out the tadpole and show that the graph determines M up to equivalence.
Remark 26.11. In lectures, it was incorrectly stated that one can classify module cate-gories over Rep sl2(C) in a similar fashion. It turns out that the answer here is much more
complicated. For example, in this setting, graphs of the form with any n ≥ 2 loops
appear. For more details32, the reader is referred to [EO03].
32An explanation of why these flower-shaped graphs appear can also be found in the notes from Anna’stalk “Representations of representations of sl2(C)” at Macquarie University, available on her website www.
maths.usyd.edu.au/u/romanova.
210
Lecture 27: Abelian categories over stacks
27.1 Modules over algebras in monoidal categories
General goal: Get a feeling for module categories (i.e. categories acted on by monoidalcategories).
One of the first surprises in representation theory is that what appears to be externalsearch (the search for all representations of a group) turns out to be an internal search (allrepresentations occur in the regular representation). A similar phenomenon happens in therepresentation theory of monoidal categories.
Given a monoidal category (A,⊗), an algebra A in A is an object A together withmorphisms 1→ A and m : A⊗ A→ A satisfying the axioms that one might expect. Givenan algebra, let
ModA(A) = category of right A-modules in A,where a right A-module in A is an object M and a morphism a : M ⊗ A → M satisfyingthe usual right module axioms. For objects X ∈ A and M ∈ ModA(A), X ⊗M is a rightA-module via idX ⊗ a, so the category ModA(A) is a left A-module.
Meta Theorem: “A-modules are A-modules.”
The Meta Theorem captures the external-becomes-internal phenomenon we see in repre-sentation theory: the external search for A-modules becomes an internal search for algebrasin A. For a precise version of the Meta Theorem in the case of a rigid semisimple categoryA with finitely many simple objects, see [Ost03].
Example 27.1. Module categories for finite groups.Let G be a finite group and let k be an algebraically closed field, such that k[G] is
semi-simple. Let H ⊂ G be a subgroup. The restriction functor
RepG→ RepH
is a ⊗-functor, and it gives RepH the structure of a RepG-module. The following exerciseshows that this A-module is an A-module (for A = RepG, A = k[G/H]).
Exercise 27.2. Show that k[G/H] is an algebra in RepG, and RepH ' ModRepG−k[G/H].
Question: Are these all semisimple indecomposible RepG-modules?
Theorem 27.3. [Ost03]indecomposible
module categoriesover RepG
/'
∼←→
(H,ω)
∣∣∣∣ H ⊂ G subgroupω ∈ H2(H, k×)
G-conjugacy
.
Under this correspondence, the pair (H,ω) corresponds to the category Rep1(H) of repre-sentations of the central extension
k× → H → H
on which k× acts via the identity character.
211
We can sum up this example with the following motto:
“Module categories over RepG are the same thing as subgroups with sauce.”
Remark 27.4. 1. The proof of Theorem 27.3 is easy once one has the Meta Theorem.
2. Objects on the LHS are not always monoidal categories as they are when the centralextension is trivial.
3. Question: In stacks language, RepH as a RepG-module corresponds to pt/H →pt/G. Can one interpret H2(H, k×) as twistings of QCoh(pt/H)?
27.2 Linear algebra over a stack
For the remainder of this lecture, we will summarize a short note of Gaitsgory on cat-egories with actions of algebraic stacks [Gai05]. Recall our motivation: In lecture 25,we introduced the “constructable” anti-spherical category Masp. We wish to show thatMasp ' DG∨×Gm(N ), and our strategy for doing so is to viewMasp progressively as a mod-
ule over QCoh(pt/G∨), then QCoh(g∨/G∨), then QCoh(N /G∨). The following languagewill be useful.
The notion of a category over an algebraic stack:Let C be a k-linear category. Assume that C is closed under inductive limits. (E.g. Vect
and QCoh(X) are okay, Vectf.d. and Coh(X) are not.) Let Y be a stack. Our goal is toaddress the following question:
What does it mean for C to be linear over Y?
Step 1: Start by assuming Y is an affine scheme; i.e. Y = SpecA for some k-algebra A. Inthis case, C being linear over Y should mean that C is A-linear; i.e. we have a map
A→ Z(C) := End(idC).
Example 27.5. If Xf−→ SpecA, then QCoh(X) is A-linear.
Claim: If C is A-linear, then QCoh(Y) acts on C; i.e. we have
QCoh(Y)× C → C.
Example 4.5. (continued) QCoh(SpecA) acts on QCoh(X) via f ∗(−)⊗ (−).
Proof. We want to define M ⊗ X for M ∈ QCoh(Y), X ∈ C. If M = AI (recall thatY = SpecA), define M ⊗X := XI . In general, choose a presentation
AI → AJ M
of M , and define M ⊗X := coker(AI ⊗X → AJ ⊗X), where the ijth-matrix coefficient isgiven by the aij action on Hom(X,X).
212
Exercise 27.6. Show that this is independent of presentation.
We conclude that QCoh(Y) C.
On the other hand, given QCoh(SpecA) C, we can also go back; that is, we can recoverthe A-linear structure on C. Indeed, because 1QCoh(SpecA) = A, for any M ∈ C we have amap
End(A) = A→ End(1⊗M) = End(M).
Hence we obtain a map A→ Z(C); i.e. an A-linear structure on C.We would like to have some notion of base change in this setting. But first we need a
pull-back. For a map A′ ← A of k-algebras, we want functors f∗, f∗ fitting into a diagram
C ′ C
Spec(A′) =: S ′ Spec(A) =: S
f∗
f∗
We define the category C ′ = C ×S S ′ as follows:
• Objects: objects of C, equipped with an additional action of A′ such that
A
A′ End(M)
commutes.
• Morphisms: morphisms in C compatible with the A′-action.
Exercise 27.7. (If you are stuck, ask Emily!) Show that C ′ = C ×S S ′ is abelian.
Example 27.8. Let C = QCoh(S) = A-Mod. Then C ×S S ′ is the category of A-moduleswith A′-actions; that is,
C ×S S ′ = A′-Mod = QCoh(S ′).
Returning to the general setting, we see that our desired functors
C ×S S ′ C
S ′ S
f∗
f∗
are given by
f∗ = forget A′-structure,
f ∗ : X 7→ A′ ⊗X.
It is not difficult to see that they form an adjoint pair (f ∗, f∗).
213
27.3 A brief review of descent
The idea: Often in mathematics, we can construct an object by constructing it “locally”.
More specifically, the principle of descent is that for some mathematical object Z,
giving Z on X ←→ giving pieces of Z ona “cover” of X + glue.
Remark 27.9. One of Grothendieck’s great insights is that here “cover” can be interpretedvery generally.
Different types of objects require different glue:
• Function: Define a function f on X by giving functions fi on Ui such that
fi|Uij = fj|Uij .
In this setting the glue is a “truth value” (i.e. the functions agree or don’t). If youwant to be more fancy, you can interpret a function as a “sheaf of truth values”.
• Sheaf: Define a sheaf F by giving sheaves Fi on Ui and morphisms αij : F|Uij → F|Uijsuch that
Fi|Uijk Fj|Uijk
Fk|Uijk
αij
αik αjk
commutes. In this setting, the morphisms are the glue, which is subject to the restric-tion imposed by the diagram.
• Category: Define a category C by giving categories Ci on Ui, functors αij : Ci|Uij →Cj|Uij , and natural transformations βijk : αjk αij
∼−→ αik such that
αkl αjk αij
αjl αij αkl αik
αil
(27.1)
commutes. In this setting, both the functors and natural transformations make up theglue.
Topological approach: Again coherence conditions come from a topological space. Assumefor simplicity X =
⋃i∈I Ui is a cover (in the classical sense). Then there is a bijection
simplicial complexes with I vertices↔ the indexing set I of the cover
n-simplices↔ n-element subsets K of Isuch that ∩i∈KUi 6= ∅.
214
Example 27.10.
Coherence conditions come from cells:
• Function: 0 and 1 cells.
• Sheaf: 0, 1, and 2 cells.
• Category: 0, 1, 2, and 3 cells.
• Etc.
Example 27.11. Coherence condition (27.1) above corresponds to the 3-cell in the 3-simplex:
27.4 Back to sheaves over stacks
Return to the setting of the end of Section 27.2 and the diagram
C ×S S ′ C
S ′ S
f∗
f∗
Lemma 27.12. (Key Lemma) Sheaves of categories satisfy descent with respect to flatcovers of affine schemes; that is, if S ′ → S is a flat cover of affine schemes, then,
sheaf of categoriesover S
∼−→ sheaf of categories over C ′+ descent data
.
Now let Y be an algebraic stack in the faithfully-flat sense, with affine diagonal. Thismeans that any morphism
SpecA→ Y
215
with source an affine scheme is affine. Recall from Emily’s IFS talk Sheaves on stacks, thatwe define a quasi-coherent sheaf on Y by imagining that there is a quasi-coherent sheaf “F”on Y and axiomatising what this would mean for f ∗“F” for all f : SpecA→ Y .
Given an algebraic stack Y as above, let SchaffY denote the category of affine schemes over
Y . We define a sheaf of categories over Y to be an assignment Csh:
1. S ∈ SchaffY 7→ CS, linear over S;
2. for each S ′ → S in SchaffY , an S-linear functor f ∗ : CS → CS′ inducing an equivalence
CS ×S S ′∼−→ CS′ ;
3. for S ′′g−→ S ′
f−→ S in SchaffY , an isomorphism g∗ f ∗ ' (f g)∗ such that 3-way
compatibility holds.
Remark 27.13. This is holistic definition. We could also give a “generators and relations”definition via a fixed flat cover.
Example 27.14. If Y = pt, then the assignment S 7→ QCoh(S) is a sheaf of categories onY .
Exercise 27.15. Given Csh on Y , define Γ(Y , Csh) and show that it is a VectY-module. (HereVectY is the category of vector bundles on Y .)
Suppose that Y satisfies:
1. Y is locally Noetherian and every quasicoherent sheaf is a limit of coherent sheaves,and
2. every coherent sheaf is a quotient of a vector bundle (e.g. “enough ample line bundles”).
Then we have the following theorem. (See Gaitsgory’s note [Gai05] for a proof.)
Theorem 27.16.
sheaves of categoriesover Y ↔ VectY -modules
Csh 7→ Γ(Y , Csh).
Example 27.17. (Key example) Let G be a linear algebraic group. Then
sheaf of categorieson pt/G
↔ Vectpt/G ' Repf.d.G-module.
To see this, recall Emily’s approach to understanding sheaves on pt/G. She started witha prestack (pt/G)triv given by
(pt/G)triv(S) = trivial G-bundle on S G(S),
then applied stackification to obtain the stack pt/G. A coherent sheaf on (pt/G)triv is thedata
V +G(A) V ⊗ A
216
for every affine k-scheme S = SpecA. In other words, an algebraic representation of G.Similarly, a category over (pt/G)triv is the data
k-linear category C +G(A) CS
for all affine S = SpecA.Consider the case of a G-scheme X. Then the assignment
S/k affine 7→ QCoh(XS) G(S)
is a sheaf of categories over pt/G. On the other hand,
QCohG(X) = QCoh(X/G)
is a RepG-module. Under the above equivalence, these correspond to one another.
Remark 27.18. We will have more to say about this correspondence in a slightly differentlanguage in the next lecture.
217
Lecture 28: (De)equivariantisation
We explained last week that if Y is an algebraic stack satisfying certain assumptions, then
sheaf of abeliancategories Csh on Y ←→
VectY -module(i.e. category with an action
of vector bundles on Y)
The most basic example of this is the following. Let G be a linear algebraic group, andY = pt/G. Recall that pt/G is the stackification of the prestack (pt/G)triv, which is givenby the assignment
S 7→ G× Sof a scheme S to the trivial G-bundle on S. As a stack in groupoids, this assignment sends
Recall that a quasi-coherent sheaf on pt/G is the data of a vector space V (the value ofthis sheaf on a point), together with an action of the S-points G(S) on V × S for all testschemes S. This is known as an algebraic representation.
Similarly, a sheaf of categories on pt/G (an object on the left hand side of the correspon-dence above) is the data of a category C (the value of this sheaf on a point), together withan action of the S-points G(S) on CS for all test schemes S.
Hence for the stack pt/G, the correspondence above matches(S 7→ CS =
abelian categorywith G(S)-action
)←→ Repf.d.G-module.
28.1 A simple example of this phenomenon
Remark 28.1. This example should have come earlier, but Geordie didn’t discover the veryclear reference [DGNO10] providing this perspective until last week.
Let G be a finite group and C an additive monoidal category.
Theorem 28.2. There are functors “equivariantisation” and “deequivariantisation” whichallow us to move between categories with G-actions and categories with RepG-actions:
k-linear categorieswith G-action
k-linear categorieswith RepG-action
RepG-enriched
categories
equivariantisation
deequivariantisation
On Karoubian k-linear categories, the upper two arrows provide equivalences of categories.The lower arrows can be made into equivalences too, with additional minor assumptions.
How do these functors work?
218
28.2 Equivariantisation
Suppose G acts on a category M; i.e. we have morphisms Fg : M →M for each elementg ∈ G and natural isomorphisms µgh : Fg Fh
∼−→ Fgh for each pair g, h ∈ G. (See lecture 26for a review of this notion.)
An equivariant object in M is a tuple (X, ugg∈G) where X ∈ M and ug : Fg(X)∼−→ X
such that the diagram
Fg(Fh(X)) Fg(X)
Fgh(X) X
µgh(X)
Fg(uh)
ug
ugh
commutes.An equivariant morphism in M is a morphism X
f−→ Y in M which commutes with allug; that is, a morphism such that the diagram
Fg(X) X
Fg(Y ) Y
Fg(f)
ug
f
ug
commutes for all ug.The equivariantisation MG of M is the category whose objects are equivariant objects
in M and whose morphisms are equivariant morphisms in M.
Remark 28.3. The category MG is a RepG-module. Indeed, given V ∈ RepG and(X, ugg∈G) ∈MG, define V ⊗X := (V ⊗X, u∨g g∈G), where
u∨g : Fg(V ⊗X) ' V ⊗ Fg(X)ρg⊗ug−−−→ V ⊗X.
Here ρ : G→ Aut(V ) is the G-representation structure on V .
Exercise 28.4. If X is a G-scheme, show that
1. G acts on CohX, and
2. the equivariantisation of Coh(X) is equivalent to the category of G-equivariant sheaveson X:
(Coh(X))G ' CohG(X).
28.3 Deequivariantisation
Let A be the algebra of functions from G to k. The algebra A carries commuting left andright G-actions:
• Left action: (g · f)(x) = f(g−1x)
• Right action: (f · g)(x) = f(xg).
219
The left action of G gives A the structure of a G-representation, and it is an algebra objectin the category RepG. The right G-action means that G acts on A, preserving its RepG-algebra structure.
If N is a RepG-module, define a category NG, the deequivariantisation of N , by
• Objects: A-modules in N ; i.e. objects Y together with a morphism a : A ⊗ Y → Ysatisfying the appropriate module conditions
• Morphisms: A-linear morphisms in N .
Example 28.5. Let N be the category of G-equivariant coherent sheaves on X, and A =k[G] as above. Then
NG = A-modules in CohG(X)
= OX [G]-modules in CohG(X)
= CohG(G×X)
= Coh(X).
28.4 RepG enrichments
Both equivariantisation and deequivariantisation can be realised as a two-step process inwhich the first step passes through a RepG-enrichment (i.e. a category whose Hom spacesare objects in RepG) of our original category. We describe this perspective now.
Let M be a category with an action of G, and let M,M ′ be equivariant objects in M.
Claim 28.6. HomM(M,M ′) is a G-module, and
HomMG(M,M ′) = HomM(M,M ′)G.
Proof. The action is given by
(Mf−→M ′)
g·7−→
Fg(M) Fg(M
′)
M M ′
Fg(f)
∼ ∼
g·f
.
The second statement follows from the definitions.
FromM, we can construct a RepG-enriched category by considering just the equivariantobjects ofM, but keeping all morphisms between those objects. This is a category where Gfixes all objects but moves morphisms. Claim 28.6 shows that from this RepG-enrichmentof M, we can then pass to the equivariantisation of M by only keeping morphisms whichare fixed by G.
On the other hand, given a RepG-module N , we can also construct a RepG-enrichedcategory. For X, Y ∈ N , consider the functor
RepG→ k : V 7→ Hom(V ⊗X, Y ).
220
Because RepG is semisimple, any k-linear functor on RepG is representable. The functorabove is represented by
Hom(X, Y ) ∈ RepG;
i.e.HomRepG(V,Hom(X, Y )) ' HomN (V ⊗X, Y ).
Hence the category consisting of the objects of N with morphisms given by Hom(X, Y ) is aRepG-enriched category. Forgetting the RepG-enrichment and taking the Karoubi comple-tion of this category results in the deequivariantisation NG of N . Note that HomN (X, Y ) =Hom(X, Y )G.
Attempt at a big picture:
Remark 28.7. The fact that non-equivariant objects are consequences of equivariant ones isnot so surprising in the case of a finite group G: Any X is a summand of Ind(X) =
⊕Fg(X),
which has a canonical equivariant structure.
28.5 Deequivariantisation principle
Now we return to the setting of the beginning of the lecture. Let S = SpecA be an affinescheme with an action of a linear algebraic group H (e.g. S = g∨, H = G∨). A key tool in
221
[AB09, Bez16] is the deequivariantisation principle, which answers the question:
How can we make a category C linear over S/H?
The deequivariantisation principle states that it is enough to give:
1. a RepH-module structure on C; and
2. an H-equivariant A-linear structure on Cdeeq33 (which is equivalent to an H-equivariant
A-action on HomC(X,OH ⊗ Y )).
Big diagram:
Recall that since C is a RepH-module, Cdeeq consists of OH-modules in C, and OH is anind-object in Repf.d.H. Hence the deequivariantisation functor sends
C → Cdeeq : X 7→ OH ⊗X.
Exercise 28.8. The functor OH ⊗ (−) is left adjoint to the forgetful functor Cdeeq → C.
Moreover,HomCdeeq
(OH ⊗X,OH ⊗ Y ) = HomC(X,OH ⊗ Y ).
Example 28.9. What is a sheaf of categories on A1/Gm? By the deequivariantisationprinciple, we need to give:
1. A RepGm-action on C; i.e. an auto-equivalence M 7→ M(1) := nat⊗M of C. (Recallthat RepGm is freely generated34 by the natural representation nat.)
2. A Gm-equivariant OA1 = k[X]-linear structure on Cdeeq; i.e. a k[X]-graded modulestructure on
HomCdeeq(OGm ⊗M,OGm ⊗N) = HomC(M,
⊕m∈Z
N(m))
=⊕m∈Z
HomC(M,N(m))
33Here Cdeeq is the deequivariantisation of C, the category we denoted in the previous section by CG. Weare making this notational switch to align with Bezrukavnikov’s notation in what is coming.
34as a k-linear additive tensor category
222
with M in degree 1. In other words, this is the same data as a morphisms of functorsid→ (1).
We conclude that a sheaf of categories on A1/Gm consists of: (1) a category C, an auto-equivalence M 7→M(1) of C, and a natural transformation X : id→ (1).
The point of this example:
sheaf over A1/Gm
“geometric”
←→ auto-equivalence (1) + morphism id→ (1)
“combinatorial”
This pattern is a feature in [AB09] and [Bez16].
Exercise 28.10. (Worthwhile). For each Y in the diagram
A1
0 0/Gm A1/Gm
Gm/Gm
calculate CY in terms of the above data.
A bigger fish: What does it mean to give a sheaf of categories on g/G?
Based on the deequivariantisation principle, this is
RepG-module C +G-equivariant Og-linear
structure on Cdeeq
.
Key observation: (To be explained below.) Via Tannakian formalism, the Og-linear struc-ture on Cdeeq is simply an “endomorphism”.
28.6 Tannakian formalism
Let G/k be a group scheme, and For : RepG→ k the forgetful functor. We have a k-groupfunctor:
A 7→ Aut⊗(For⊗ A) = ⊗-automorphisms of A⊗ For.
(A ⊗-automorphism of A ⊗ For is a collection of functors Fv : V ⊗ A → V ⊗ A for allV ∈ Repf.d.G such that
(V ⊗ A)⊗A (V ′ ⊗ A) (V ⊗ A)⊗A (V ′ ⊗ A)
(V ⊗ V ′)⊗ A (V ⊗ V ′)⊗ A
∼
FV ⊗FV ′
∼FV⊗V ′
commutes.)We discover that “G can be recovered from its category of representations”:
223
Theorem 28.11.G(A) ' Aut⊗(A⊗ For).
How can we think about this theorem? The A-points of G clearly give ⊗-automorphisms,so we have a map
G(A)→ Aut⊗(For⊗ A).
The miracle is that this map is an isomorphism.
Question: How can we recover the Lie algebra g = LieG from Tannakian formalism?
Definition 28.12. A ⊗-derivation of For⊗ A is an endomorphism
NV : V ⊗ A→ V ⊗ A
for all V ∈ RepG such that NV⊗V ′ = NV ⊗ 1 + 1⊗NV ′ .
Theorem 28.13.g⊗ A ' End⊗(For⊗ A).
Exercise 28.14. Deduce Theorem 28.13 from Theorem 28.11 using
g = ker(G(k[ε]/(ε2)
))→ G(k).
A rather startling consequence of this is the following. If A is a k-algebra with a G-action,then
G-equivariant⊗-derivationV 7→ V ⊗ A
=element
x ∈ (g⊗ A)G⇐⇒ HomG
v.s.(g∗, A)
⇐⇒ G-equivariantHomalg.(Og, A)
⇐⇒ SpecA/G→ g/G.
The moral is that, for a G-equivariant algebra, equipping the functor V 7→ V ⊗A with aG-equivariant tensor derivation is the same as equipping its category of equivariant sheaveswith a g/G-linear structure.
224
Lecture 29: Coherent sheaves on base affine space
Today we will discuss coherent sheaves on base affine space. But before doing so, we’ll startwith a brief reminder on where we’re going.
Let CohG∨
free(N ) ⊂ CohG∨(N ) be the full subcategory consisting of V ⊗ ON (λ) for V ∈
RepG∨ and λ a character of T∨. As in Lectures 24 and 25, let H be the affine Heckecategory andMasph the categorical anti-spherical module. Our goal is to give an equivalenceof categories CohG
∨(N )
∼−→Masph. To do so, we will construct a monoidal functor F fittinginto the following diagram.
CohG∨(N ) CohG
∨
free(N ) H Masph
want
construct
F
Once we have constructed F : CohG∨
free(N ) → Masph, we can extend F to a functor from
CohG∨(N ) using the fact that any coherent sheaf admits a resolution via vector bundles.
(This is analogous to the extension from an A-linear structure to an action of QCoh(SpecA)which occurred at the start of Lecture 27.)
How might we go about constructing such a functor F? The key philosophical observationis that F can be built “softly” once one has a functor RepG∨ → H. More precisely, one ofthe main technical lemmas in [AB09] (in a somewhat diluted form, to aid comprehensibilityat this point) is the following.
Theorem 29.1. Let C be an additive monoidal category, and F : RepG∨ → C a ⊗-functor.
1. If N is a ⊗-derivation of F (cf. Lecture 28), then F extends to a ⊗-functor
F : Cohfree(g∨/G∨)→ C.
2. Suppose that we can upgrade F to a⊗-functor Rep (G∨ × T∨)→ C. Moreover, supposethat we have arrows
bλ : F (Vλ)→ F (kλ)
satisfying the “Plucker relations” (and an “acyclicity condition”). Then F extends toa functor
F : CohG∨
free(B∨)→ C.
(Here B∨ denotes the flag variety of G∨.) Combining 1 and 2 we obtain a functor
F : CohG∨
free(B∨ × g∨)→ C.
3. Suppose that for all λ, the arrow bλ NVλ : F (Vλ)→ F (Vλ)→ F (kλ) is equal to zero.Then F extends to
F : CohG∨
free(N )→ C.
For the remainder of the lecture we will work entirely on the Langland’s dual side, sofrom here on, we drop all checks from our notation.
225
29.1 Serre’s description of coherent sheaves on Pn
Serre’s paper [Ser55] was very influential, and is recommended reading for any readers whohave not done so. It was one of the first papers dealing in detail with coherent sheavesand their cohomology. The last part gives a beautiful description of coherent sheaves onprojective space, which we now recall.
The variety Pn has a standard affine cover
Pn =n⋃i=0
Ui, Ui ' An.
Using this cover, one way to view a coherent sheaf on Pn is as
collection of n+ 1 modulesover k[x1, . . . , xn]
+ glue.
But this is not very practical. Instead, we can make the observation that
CohPn = CohGm(An+1\0) ' CohGm(An+1)/CohGm(0).
If you haven’t thought about this before, the second equivalence shouldn’t be entirely obvi-ous. Now,
CohGm(An+1) =finitely generated gradedS = k[x0, . . . , xn]-modules
CohGm(pt) =finite-dimensional
modules
Hence
CohPn =
(finitely generatedgraded S-modules
)/⟨finite dimensional
S-modules
⟩.
Here 〈−〉 denotes a “Serre subcategory”, and the quotient is a “Serre quotient”. (Perhapsthis language shouldn’t be a surprise!) If this is new to you, the following exercise is recom-mended.
Exercise 29.2. 1. Convince yourself that it gives the right answer for Coh(P0).
2. The functor CohPn → S-grmod is given by
Γ := Hom(⊕n∈Z
O(n),−).
Hence describe Γ(O(n)) and Γ(skyscraper).
Remark 29.3. We can view these descriptions through the lens of descent. The “classical”cover Pn =
⋃Ui leads to our not-very-useful description. The cover
An+1\0 → Pn
leads to Serre’s description. Note that An+1\0 is “almost” affine.
226
29.2 How can we describe coherent shaves on G/B?
If G = SL2, G/B ' P1, and we can use Serre’s description of coherent sheaves on Pn todescribe coherent sheaves on G/B. For other groups, a similar philosophy can be employed.At first glance, three options present themselves for describing coherent sheaves on G/B:
1. via covers (not very useful);
2. using CohG(G/B) ' Coh(pt/B) ' RepB;
3. via base affine space (analogue of Serre’s description).
What is base affine35 space? Let U ⊂ B be the unipotent radical. Then G/U is aG× T -variety via the T -action
gU · t = gtU.
This is well-defined because T normalizes U . Hence we have a quotient map
G/U → G/B ' (G/U)/T.
Example 29.4. Let G = SL2. Then G C2 and
stab
(10
)=
(1 ∗0 1
)= U.
Hence G/U ' orbit of
(10
)' C2\0. The T -action is given by scaling: x · λ = λx, and
(C2\0
)/C× = P1C = G/B.
In general, G/U is quasi-affine, meaning that it is an open set inside an affine variety.
Why? Denote by Vλ a simple highest weight module of G and vλ ∈ Vλ a fixed highest weightvector. We can choose λ1, . . . , λn such that the stabiliser of
v := (vλ1 , . . . , vλn) ∈ V := Vλ1 ⊕ · · · ⊕ Vλn
in G is U . Hence
G/U → V
g 7→ gv
gives a map G/Uopen−−→ G/U ⊂ V and G/U is affine.
Ring of functions: Assume chark = 0. Then we have the Peter-Weyl theorem:
k[G] =⊕λ∈X+
Vλ ⊕ V ∗λ as G×G-representations.
35Note that it is almost never affine, so the nomenclature is a little strange!
227
Hence,
k[G/U ] ' k[G]U '⊕λ∈X+
Vλ ⊗ (V ∗λ )U '⊕λ∈X+
Vλ.
The first isomorphism follows from the fact that the right action of U on G is free, and thefinal isomorphism follows from (V ∗λ )U ' C. This gives us a description of k[G/U ] in termsof representations of G, but how do we describe multiplication from this perspective?
The torus T acts on (V ∗λ )U via the character λ, so T acts on Vλ ' Vλ ⊗ (V ∗λ )U withcharacter λ. The multiplication map
m :
⊕λ∈X+
Vλ
⊗⊕λ∈X+
Vλ
→ ⊕λ∈X+
Vλ
is T -equivariant, so it must mapVλ ⊗ Vµ → Vλ+µ.
The space HomG(Vλ ⊗ Vµ, Vλ+µ) is one-dimensional, hence there is a unique map
mλ,µ : Vλ ⊗ Vµ → Vλ+µ : vλ ⊗ vµ 7→ vλ+µ.
(Note that the multiplication map could also be zero on this component, but this can beruled out.) This gives the multiplication in k[G/U ].
Remark 29.5. This is sometimes referred to as Chevalley multiplication on k[G/U ].
Example 29.6. If G = SL2, k[G/U ] = k[x, y] =⊕
m∈Z k[x, y]m, and multiplication
k[x, y]m ⊗ k[x, y]m′ → k[x, y]m+m′
is the unique SL2-equivariant morphism which sends xm ⊗ xm′ to xm+m′ .
Important Notation:
G/U
base affine space
open dense−−−−−−→ G/U = Spec k[G/U ]
affine closure
We should think that these two spacs are “almost the same”, like An+1 and An+1\0.
Exercise 29.7. (Well worth thinking about!) Describe the ideal of the boundary of G/Uinside G/U in terms of the above.
29.3 Motivational interlude
The algebra k[G/U ] =⊕
λ∈X+Vλ is a T -algebra. The deequivariantisation principle of
Lecture 28 tells us that to make a category M linear over G/B = (G/U)/T is the same asgiving:
1. a RepT -module M ⇐⇒ an action of the character lattice X of T on a category M.We will denote this action via λ ·M = M(λ).
228
2. a T -equivariant k[G/U ]-linear structure on Mdeeq; i.e. for all M,M ′ ∈ M, a T -equivariant map
k[G/U ]φM,M′−−−→ Hom(M,OT ⊗M ′) = Hom(M,
⊕M ′(λ)).
By the description of k[G/U ] above, this is the same data as a collection of maps
Vλ → Hom(M,M ′(λ))
for each λ ∈ X+ (the λ-isotypic components of φM,M ′); or, equivalently (as we willexplain momentarily), a collection of “highest weight arrows”
bλ ∈ Hom(Vλ ⊗M,M ′(λ))
such that the associated map φM,M ′ is K[G/U ]-linear (“Drinfeld–Plucker relations”).
3. with the property that sheaves on the complement
G/U\G/U
act by zero.
Remark 29.8. This explains the arrows in Theorem 29.1.2. Note that C from Theorem 29.1is EndM here.
29.4 Arrows between coherent sheaves on G/U
Now we will examine the “highest weight arrows” in the previous section more carefully. Set
O := k[G/U ] =⊕λ∈X+
Vλ
with Chevalley multiplication and X-grading. Then
HomG×T -eqvtvect bndls
(Vλ ⊗O,O(λ)) = HomG×T -eqvt(Vλ,O(λ))
= HomG-eqvt(Vλ, Vλ)
= C
is spanned byBλ : Vλ ⊗O → O(λ),
with the property thatBλ|Vλ⊗Vµ = mλ,µ : Vλ ⊗ Vµ → Vλ+µ.
Lemma 29.9. These arrows satisfy Plucker relations:
Bλ ⊗Bµ = Bλ+µ (mλ,µ ⊗ idO).
229
That is, the diagram
(Vλ ⊗O)⊗O (Vµ ⊗O) O(λ)⊗O O(µ)
(Vλ ⊗ Vµ)⊗O O(λ+ µ)
Vλ+µ ⊗O
Bλ⊗Bµ
∼mult
mλ,µ Bλ+µ
commutes.
Proof. It is enough to check that the restriction to Vλ ⊗ Vµ commutes. This becomes thedefinition of Chevalley multiplication.
Remark 29.10. Recall that the classical Plucker relations describe the homogeneous equa-tions defining the Grassmannian of k-planes in Cn in its “Plucker” embedding inside P(ΛkCn).Analogously, the above theory can be used to describe the defining relations of the flag varietyin its embedding inside a product of Grassmannians. For more on this (and the connectionto moduli of vector bundles and Drinfeld’s compactification), see [FGV02, §4].
We will end today’s lecture by describing the analogue in this setting of the pt/G g/Gupgrade that we discussed last lecture. Here, we have an upgrade
pt/G G\(G/U)/T.
Proposition 29.11. Let A be a G × T -algebra. Suppose for all λ ∈ X+, we are given aG-equivariant morphism
bλ : Vλ ⊗ A→ A(λ)
satisfying the Plucker relations. Then there exists a unique G×T -equivariant homomorphism
φ : O → A
such that bλ = φ∗(Bλ) := idA ⊗Bλ.
Proof. What does the last condition mean? It means that
Vλ ⊗O Vλ ⊗ A
O(λ) A(λ)
Bλ bλ
φ(λ)
commutes; i.e.φ|Vλ = bλ|Vλ⊗1.
This determines φ uniquely: define φλ := bλ|Vλ⊗1 : Vλ → A and φ =⊕
φλ. This is clearlyG×T -equivariant. One can check that φ is a homomorphism, which amounts to the Pluckerrelations holding.
230
Lecture 30: Constructible sheaves on finite flag varieties
and braid groups
Today we return to the constructible world. In this lecture and the following one, we willsee how the formalism of the last two lectures is used to produce a ⊗-functor
CohG∨
free(N )→ PI ⊂ DbI(F`).
The key tool is provided by BGK central sheaves and certain easier sheaves called Wakimotosheaves. Wakimoto sheaves are special to the affine setting; however, to warm up (it’s beena few weeks since we’ve been on this side!), we’ll spend today’s lecture in the finite case.
30.1 A fun calculation
Choose points x 6= y on P1:
There are two ways to include these points into P1:
U := P1\x, y P1\x
P1\y P1
jy
jx jx
jy
Exercise 30.1. Show that
jx!jy∗QU ' jy∗jx!QU =: Qx!,y∗.
(Hint: construct a map and then show that it is an isomorphism.)
A picture:
231
Claim 30.2. H∗(P1,Qx!,y∗) = 0.
Proof. Let ix : x → P1 be the inclusion. By applying the functorial distinguished triangle
jx!j!x → id→ ix∗i
∗x
+1−→
to Qy∗ := jy∗QP1\y, we obtain a distinguished triangle
jx!j!xQy∗ → Qy∗ → ix∗i
∗xQy∗
+1−→ .
By base change, jx!j!xQy∗ ' Qx!,y∗, so the long exact sequence in cohomology
H∗(Qx!,y∗) H∗(Qy∗) = H∗(C) H∗(ix∗i∗xQy∗)
2 0 0 0
1 0 0 0
0 0 Q Q∼
implies the claim.
Exercise 30.3. • Prove the claim again as follows: use Artin vanishing (which we willstate later) and the previous exercise to show that H i(P1,Qx!,y∗[1]) = 0 for i 6= 0, thencompute that
χ(P1,Qx!,y∗) = 0
to deduce the claim.
• Find yet another proof of the claim by interpreting H∗(P1,Qx!,y∗) as certain locallyfinite chains.
30.2 The finite Hecke category
FixG ⊃ B ⊃ T
as usual, taken over C. In what follows we will take our coefficients in Q. To simplify thediscussion, we will work in
DbB×B(G,Q) ' Db
B(G/B,Q).
Remark 30.4. As we learned in Lecture 24, this is an approximation to the Hecke category.Everything we do below may be lifted to the Hecke category H with a little more effort.However, we choose to focus on Db
B×B(G,Q) in today’s lecture to illustrate the ideas moreclearly.
232
Recall that DbB(G/B) comes with a convolution structure
∗ : DbB(G/B,Q)×Db
B(G/B,Q)→ DbB(G/B,Q).
The B-orbits on G/B stratify G/B by the Bruhat decomposition
G/B =⊔x∈Wf
BxB/B.
Hence we have two related (but not equal!) derived categories, each containing a correspond-ing category of perverse sheaves: the construcible (with respect to the Bruhat stratification)derived category
Db(B)(G/B,Q) ⊃ P(B),
and the B-equivariant derived category
DbB(G/B,Q) ⊃ PB.
Convolution gives a right action
Db(B)(G/B,Q) Db
B(G/B,Q).
Notation 30.1. For F ∈ DB(G/B), denote by F := QB/B ∗ F = For(F) ∈ Db(B)(G/B) the
object we get by “forgetting equivariance”. This gives a map
PB → P(B).
Remark 30.5. The map F 7→ F is fully faithful, but the image is not closed under exten-sions.
Let jx : BxB/B → G/B. Define
∆x := jx!QBxB/B[`(x)], ∇x := jx∗QBxB/B[`(x)]
in PB.
Remark 30.6. Recall that in Lecture 20, we showed that ∆x and ∇x are standard andcostandard objects in a highest weight structure on P(B). In general, the category PB doesnot admit a highest weight structure.
Example 30.7. Let G = SL2. As we discussed in Lecture 16, we can use Beilinson’sdescription
Perv(B)(P1) ' Rep
(• •
e
f∣∣∣∣ fe = 0
)of the category P(B) to see that it has five indecomposible objects:
1. ˙ICid ←→(
0 k)
,
233
2. ˙ICs ←→(k 0
),
3. ∆s ←→(k k
∼
0)
,
4. ∇s ←→(k k
0
∼)
,
5. Ts ←→(k k ⊕ k
), where the arrow → embeds k into the first factor of k⊕ k
and the arrow ← projects onto the second factor.
The object Ts is called the “big projective” or “big tilting sheaf”.We can also use a modified version Beilinson’s description to describe the indecomposible
objects in the category PB. The B-equivariance implies that the micolocal monodromy iszero, so we have an extra condition on our quiver representations:
PervB(P1) ' Rep
(• •
e
f∣∣∣∣ fe = ef = 0
).
Hence there are four indecomposible objects in PB: ICid, ICs, ∆s, and ∇s, which correspondto the quiver representations 1− 4 above. The quiver representation 5 corresponding to thebig projective does not satisfy the condition that ef = 0, so we have no analogue to Ts inPB. With this description, it is easy to see that PB is not a highest weight category. Forinstance, minimal projective resolutions of ICid and ICs are
· · · → ∆s → ∇s → ∆s → ∇s → ∆s ICs,
· · · → ∇s → ∆s → ∇s → ∆s → ∇s ICid,
so PB does not have finite homological dimension, and hence cannot admit a highest weightstructure.
30.3 The Braid group
Recall that Wf has a presentation
Wf = 〈s ∈ Sf | s2 = id, st . . .︸ ︷︷ ︸mst
= ts . . .︸ ︷︷ ︸mst
〉.
The braid group associated to Wf is
BWf:= 〈σs, s ∈ Sf | σsσt . . .︸ ︷︷ ︸
mst
= σtσs . . .︸ ︷︷ ︸mst
〉.
Let BWfbe the corresponding monoidal category36. We have the following important and
beautiful theorem.
36See Lecture 26, where this was denoted ABWf.
234
Theorem 30.8. The assignment σs 7→ ∆s extends to a monoidal functor
BWf→ Db
B(G/B,Q).
Corollary 30.9. There is a (strict) action of BWfon Db
(B)(G/B).
Proof. (of Theorem 30.8) We will use the following “generators and relations” theorem ofDeligne [Del97].
Theorem 30.10. Let A be a monoidal category. Giving a monoidal functor
F : BWf→ A
is the same as giving
• a collection Fx ∈ Ax∈Wfof invertible objects, and
• a collection FxFy∼−→ Fxyx,y∈Wf with `(xy)=`(x)+`(y) of isomorphisms,
such thatFxFyFz FxyFz
FxFyz Fxyz
(30.1)
commutes whenever `(xyz) = `(x) + `(y) + `(z).
We will show that there exists a monoidal functor
F : BWf→ (Db
(B)(G/B), ∗)
sending σs 7→ ∆s by providing the data of Theorem 30.10. Most of this data can be obtainedfrom the following lemma:
Lemma 30.11. If `(xy) = `(x) + `(y), then ∆x ∗∆y ' ∆xy canonically.
Proof. Recall that ∆x := jx!(QBxB/B[`(x)]). Unpacking definitions, we have
∆x ∗∆y = m∗(j!QBxB×BByB/B[`(x) + `(y)]
).
If `(x) + `(y) = `(xy), then we have a commutative diagram:
BxB ×B ByB/B BxyB/B
G×B G/B G/B
m′
j jxy
m
.
Hence,∆x ∗∆y = jxy!
(QBxyB/B[`(xy)]
)= ∆xy.
235
Lemma 30.11 provides the canonical isomorphisms FxFy∼−→ Fxy of Deligne’s theorem, and
it is easy to check that (30.1) commutes if `(xyz) = `(x) + `(y) + `(z). Hence to completethe proof, it remains to show that ∆x are invertible. Because
∆x = ∆s1 ∗ · · · ∗∆sm
for x = s1 · · · sm, it is enough to show that ∆s is invertible for all s ∈ Sf .
Lemma 30.12. For s ∈ Sf ,
∆s ∗ ∇s ' ∇s ∗∆s ' ∆id = ∇id.
Proof. (Sketch #1, a “hygenic” proof) First, one can easily reduce to SL2. Here we have adistinguished triangle of perverse sheaves
QP1 [1] = ICs → ∇s → ICid = Qid+1−→ .
Now, using the fact that convolution with ∆s preserves distinguished triangles, we haveanother distinguished triangle
∆s ∗ ICs = QP1 → ∆s ∗ ∇s → ∆s+1−→ .
From this, we conclude that ∆s ∗∇s = ker(∆s → ICs) = Qid, using that the connecting mapis actully a map between perverse sheaves to make sense of the kernel.
Exercise 30.13. Fill in the details of this argument.
Proof. (Sketch # 2, a more intuitive proof) We have an isomorphism
SL2×B SL2 /B∼−→ P1 × P1
(g, g′B) 7→ (gB, gg′B)
Under this isomorphism, B ×B SL2 /B corresponds to Z := x-axis, and SL2×BB/B corre-
236
sponds to ∆ := diagonal in the following picture:
After forgetting equivariance,
∆s ∗ ∇s = p∗((j∆∗QP1×P1\∆)⊗ (jZ!QP1⊗P1\Z [2])
), (30.2)
where j∆ : P1×P1\∆ → P1×P1 and jZ : P1×P1\Z → P1×P1. By proper base change, wecan compute stalks:
(∆s ∗ ∇s)x =
H∗(P1,Q0!,x∗) if x 6= 0;
H∗+2(P1, j0!Q) if x = 0.
In the second case, we have H∗+2(P1, j0!Q) = Q, and by the fun computation from earlier,we have H∗(P1,Q0!,x∗) = 0. Hence,
∆s ∗ ∇s ' QB/B
which completes the proof.
Exercise 30.14. Fill in the details of this argument.
This completes the proof of Theorem 30.8.
30.4 Further properties of the braid group action
We will spend the rest of the lecture exploring the properties of the elements associated tobraid group elements. All of this is in preparation for the affine case next week.
237
The ∆y and ∇y form “exceptional collections”:
Hom•(∆x,∆y) = 0 unless x ≤ y;
Hom•(∇x,∇y) = 0 unless x ≥ y;
Hom•(∆x,∇y) = 0 if x 6= y.
Indeed, ignoring shifts,
Hom•(∇x,∇y) = Hom•(jx∗QXx , jy∗QXy) = Hom•(j∗yjx∗QXx ,QXy) = 0 unless y ≤ x,
and similarly for Hom•(∆x,∆y). (Here Xx = BxB/B is the Bruhat cell associated tox ∈ Wf .)
Remark 30.15. In Lecture 20, we showed that
Homi(∆x, ∇y) =
Q if i = 0 and x = y;
0 otherwise.
This was how we established the necessary Ext2 vanishing property of the highest weightstructure on P(B). In contrast,
Hom•(∆x,∇y) =
H•B(pt) if x = y;
0 if x 6= y.
Note that in this case, it is not true that the derived category of PB agrees with the equiv-ariant derived category, so we cannot use this computation to conclude that Ext2 vanishingfails in PB. However, it is an indication that PB might not be highest weight, and indeed itis not in general. (See Remark 30.6.)
Remark 30.16. In BWf, Hom spaces are very easy. By contrast, morphisms between ∆x’s
are complicated and quite mysterious. For example, Ext•(Vermas) is an unsolved problem.
Recall that convolution gives a right action
Db(B)(G/B) Db
B(G/B) ⊃ PB.
Within Db(B) we have the highest weight category P(B) ⊂ Db
(B)(G/B). The convolution oftwo perverse sheaves is not generally perverse, so the subcategory P(B) is not preserved underthe action of PB.
Definition 30.17. A perverse sheaf F ∈ P(B) is convolution exact if F ∗ G ∈ P(B) for anyG ∈ PB.
Mirkovic’s observation: A perverse sheaf F ∈ P(B) is convolution exact if and only if Fis tilting.
238
Proof. ( =⇒ ) By highest weight formalism,
F is tilting ⇐⇒ Hom>0(F , ∇y)(∗)= 0
(!)= Hom>0(∆x,F).
We will check that equality (!) holds for F convolution exact, (∗) is similar.Because F is convolution exact, F ∗∆w0 is perverse. This implies that
F ∗∆w0 ∈ pD≤0 = 〈∆y[d] | y ∈ Wf , d ∈ Z≥0〉ext,
where the subscript “ext” denotes the closure under extensions. Hence
F = F ∗∆w0 ∗ ∇w0 ∈ 〈∆y ∗ ∇w0 [d] = ∇yw0 | y ∈ Wf , d ∈ Z≥0〉ext.
Recall that w0 = y−1 · yw0 and `(w0) = `(y−1) + `(yw0), so ∇w0 = ∇y−1 ∗ ∇yw0 . Now any
object in 〈∆y ∗ ∇w0 [d] = ∇yw0 | y ∈ Wf , d ∈ Z≥0〉ext satisfies (!), so we are done.For the (⇐= ) direction, we need a very important property of affine morphisms.
Theorem 30.18. Let f : X → Y be affine. Then f∗ preserves pD≥0, and f! preserves pD≥0.
Remark 30.19. Recall Artin vanishing: if F/X is perverse and X is affine,
H>0(f∗F) = H>0(X,F) = 0 = H<0c (X,F) = H<0(f!F).
This is exactly Theorem 30.18 when Y = pt.
Exercise 30.20. Can you prove the theorem for D-modules?
Now back to our setting.
Proposition 30.21. The maps
BxB ×B G/Bλx−→ G/B and G×B BsB/B
ρx−→ G/B
are affine.
Proof. (Idea) We can write BxB ' Ux · x ·B, hence we have a commutative diagram
BxB ×B G G
Ux ×G G
mult
∼ ∼
mult
The space Ux × G is affine, so mult : BxB ×B G → G is affine, hence λx is affine (beingaffine is local in the flat topology).
Exercise 30.22. Show that
∆x ∗ (−) = λx!λ∗x(−)[`(x)],
∇x ∗ (−) = λx∗λ∗x(−)[`(x)],
(−) ∗∆x = ρx!ρ∗x(−)[`(x)],
(−) ∗ ∇x = ρx∗ρ∗x(−)[`(x)].
(Note: λ!x ' λ∗x[2`(x)] because λx is smooth of relative dimension `(x).)
239
Corollary 30.23. ∆x ∗ (−), (−) ∗∆x preserve pD≥0, and ∇x ∗ (−), (−) ∗∇x preserve pD≤0,whenever this makes sense.
Corollary 30.24. ∆x ∗ ∇y and ∇y ∗∆x are perverse for all x, y ∈ Wf .
With this we can complete the proof of Mirkovic’s observation. Recall that
pD≤0B = 〈∆x[d] | x ∈ Wf , d ∈ Z≥0〉.
By Corollary 30.24,〈∇y | y ∈ W 〉 ∗ pD≤0
B ⊂pD≤0.
Hencetilting ∗ pD≤0
B ⊂pD≤0.
Similary, pD≥0 = 〈∇x[−d] | x ∈ Wf , d ∈ Z≥0〉, so
tilting ∗ pD≥0B ⊂
pD≥0.
This lets us conclude thattilting ∗ PB ⊂ P(B).
This completes the proof of Mirkovic’s observation.
Exercise 30.25. 1. Use these ideas to show that if F ∈ PB is convolution exact, thenF = ∆id, (i.e. there are no interesting convolution exact sheaves in PB).
Remark 30.26. We will see next week that there are many convolution exact F ∈PI ⊂ DI(F`) in the affine case.
2. Let Tx be indecomposible tilting. Show that
Tx ∗QPs/B[1] = 0
unless x = id. (This is related to the fact that H∗(G/B, Tx) = 0 unless x = id.) Usethis to give another proof of the (⇐= ) direction of Mirkovic’s observation.
240
Lecture 31: Affine flags, affine braids, and Wakimoto
sheaves
Last week we established the existence of a monoidal functor
BWf→ (Db
B(G/B), ∗)σx 7→ ∆x
and described several of its nice properties. Today we’ll describe the situation in the affinecase.
As in previous lectures, let F` = G((t))/I, where I ⊂ G((t)) is an Iwahori subgroup.Denote by W = X∨ oWf the affine Weyl group. For simplicity, assume that G is simplyconnected, so W is a Coxeter group37.
The same construction as we described last week in the finite case yields a monoidalfunctor
BW → (DbB(G/B,Q), ∗)
σx 7→ ∆x.
31.1 Wakimoto sheaves
In Lecture 12, we described Bernstein’s presentation of the extended affine Hecke algebra.We briefly recall this construction now. In Lecture 11, we proved the Iwahori-MatsumotoLemma (Lemma 11.8): for w ∈ Wf , λ ∈ X∨,
`(tλwf ) =∑α∈R+wf (α)>0
|〈λ, α〉|+∑α∈R+wf (α)<0
|〈λ, α〉 − 1|.
A consequence of this is that if λ′, λ′′ ∈ X∨ are dominant, then `(tλ′tλ′′) = `(tλ′) + `(tλ′′).(This is also intuitively clear because tλ′ and tλ′′ translate in the same direction.) Hence inthe Hecke algebra H38,
δtλ′δtλ′′ = δtλ′+λ′′ = δtλ′′δt′λ .
We can write any λ ∈ X∨ as λ = λ′ − λ′′ with λ′, λ′′ ∈ X∨ dominant. Hence we have have awell-defined map
Z[v±1][X∨] → H : λ 7→ θλ := δtλ′δ−1tλ′′.
Bernstein used this map to describe the center of H.Wakimoto sheaves categorify this construction; i.e. given λ ∈ X∨, write λ = λ′ − λ′′ for
λ′, λ′′ ∈ X∨+ and defineJλ := ∆tλ′
∗ ∇t−λ′′= ∆tλ′
∗ (∆tλ′′)−1.
The Jλ are called Wakimoto sheaves.
37If we drop this assumption, W is a quasi-Coxeter group. This adds a few complications to proofs, butdoesn’t change the story in any major way.
38Here we are aligning notation with Lecture 23, which differs slightly from Lecture 12.
241
Example 31.1. If λ is dominant, then Jλ = ∆tλ . If λ is anti-dominant, then Jλ = ∇tλ .
Lemma 31.2. The map λ 7→ Jλ extends to a monoidal functor X∨ 7→ (DbI(F`), ∗), and
hence to a monoidal functor
(RepT∨,⊗) 7→ (DbI(F`), ∗).
We saw last week that ∆x ∗ ∇y is perverse for all x, y ∈ Wf . An easy adaptation of thisto the affine case implies the following corollary.
Corollary 31.3. For all λ ∈ X∨, Jλ is perverse; i.e. Jλ ∈ PI ⊂ DbI(F`).
Notation 31.1. Let
A = Wakimoto-filtered objects in PI ,
grA = image of RepT∨ in DbI(F`).
Note that grA is not a full subcategory in DbI(F`). This notation will make sense in a
moment.
Remark 31.4. It is immediate from the definitions that A and grA are monoidal subcat-egories in PI . Note that this is in contrast to the finite case, where PB has no interestingmonoidal subcategories.
Recall that in the finite case,
Hom•(∆x,∆y) = 0 unless x ≤ y,
Hom•(∇x,∇y) = 0 unless y ≤ x.
On X∨, we have the periodic order:
λ ≤ µ if and only if µ− λ ∈ ZX∨+.
Lemma 31.5. With respect to the periodic order,
Hom•(Jλ, Jµ) = 0 unless λ ≤ µ,
Hom•(Jλ, Jλ) = Q.
Proof. For γ sufficiently dominant,
Hom•(Jλ, Jµ) = Hom•(Jλ ∗ Jγ, Jµ ∗ Jγ)= Hom•(∆tλ+γ
,∆tµ+γ )
= 0
unless tλ+γ ≤ tµ+γ. It is a known fact about the Bruhat order in affine type that forsufficiently dominant weights we have
tλ+γ ≤ tµ+γ ⇐⇒ λ+ γ ≤ µ+ γ.
Hence Hom•(Jλ, Jµ) = 0 unless λ ≤ µ.
242
Example 31.6. Let G = SL2, so X∨ = Zα∨. The Homs between Wakimoto sheaves onlygo in one direction:
Lemma 31.5 implies:
Exercise 31.7. Every F ∈ A admits a functorial39 filtration (called a “Wakimoto filtration”)FI indexed by upper closet subsets I ⊂ X∨ such that if λ ∈ I is minimal, then
FI/FI\λ '⊕
mJ⊕mλλ .
Remark 31.8. This exercise is an analogue of a standard (but nonetheless beautiful) featureof highest weight categories: (co)standard filtrations are unique when they exist.
The exercise gives us a functor
gr : A 7→ grA
F 7→⊕F≥λ/F>λ,
which sends an object to its “associated graded under the Wakimoto filtration”.
Exercise 31.9. The Wakimoto filtration is compatible with ∗, and hence gr is a ⊗-functor.
31.2 Motivational interlude
We’ll take a brief motivational interlude to explain how Wakimoto sheaves fit into our biggerpicture.
Move back to the coherent side, and let G∨ ⊃ B∨, R∨+ be as usual. Any G∨-module Vhas a B∨-filtration indexed by upper closed subsets I ⊂ X∨, namely
VI =⊕λ∈I
Vλ ⊂ V,
and the associated graded of this filtration is isomorphic to a direct sum of one-dimensionalmodules kλ.
Example 31.10. If G = SL2, a picture of this filtration is:
39in particular unique!
243
In fact, Hom•(kλ, kµ) = 0 unless λ ≤ µ. Similarly on N , we have
HomCohG
∨(N )
(O(λ),O(µ)) = 0 unless λ ≤ µ.
Our equivalence will end up mapping
O(λ) 7→ Jλ.
31.3 Beilinson–Gaitsgory–Kottwitz (BGK) central sheaves
Let Gr = G((t))/G[[t]] be the affine Grassmannian and F` = G((t))/I the affine flag variety.
Theorem 31.11. There exists a central functor
Z : PG(O)(Gr)→ PI = PI(F`) ⊂ DI(F`).
Moreover, Z is equipped with a ⊗-derivation N .
Remark 31.12. The necessary compatibilities are checked in the appendix “braiding com-patabilities” by Gaitsgory to Bezrukavnikov’s [Bez04] and in more detail in a book in progressby Achar-Riche.
For the purposes of stating the next theorem, we (temporarily) extend the definition ofWakimoto sheaves: write w ∈ W as w = tλwf for tλ ∈ X∨, wf ∈ Wf , and let
Jw = Jλ ∗ ∇wf .
The affine analogue of Mirkovic’s observation from last lecture is the following.
Theorem 31.13. [AB09]
1. Any convolution exact F ∈ PI is Jw | w ∈ W-filtered.
2. If in addition F is “weakly central” (i.e. F ∗G ' G ∗F for all G), then F is Wakimoto-filtered (i.e. only Jλ for λ ∈ X∨ occur in the filtration above).
Corollary 31.14. Z(G) is Wakimoto-filtered! i.e.,
Z : PG(O)(Gr)→ A
We omit the proof of Theorem 31.13, but note that it is not much more difficult thanthe proof of Mirkovic’s observation that we described last week.
The following is beautiful and bears a striking resemblance to Mirkivic-Vilonen’s theoremon weight functors. (In fact, according to [AB09, Remark 6], they are “equivalent”.)
Theorem 31.15. 1. The following diagram commutes up to canonical isomorphism.
RepG∨ ' PG(O)(Gr) A
RepT∨ grA
Z
res gr
∼
244
2. We have
H ic(ι∗w(Z(V ))) =
Vν if w = ν ∈ X∨, i = `(ν);
0 otherwise.
Remark 31.16. Part 2. is included for experts. The notation involved is the following:The NC((t))-orbits on F` are indexed by W . Given w ∈ W , let Sw = NC((t)) · wI/I and letιw : Sw → F` denote its inclusion.
We now have all of the ingredients that we need to define our functor:
1. Z : RepG∨ → A;
2. F : RepT∨∼−→ grA ⊂ A (not full);
3. N = nilpotent ⊗-derivation of Z, coming from monodromy of vanishing cycles;
4. bλ : Jλ → Z(Vλ) highest weight arrow coming from Wakimoto filtration.
Now we need to check:
(a) Plucker relations: Thus we need to check that the following diagram commutes:
Jλ ⊗ Jµ Z(Vλ) ∗ Z(Vµ) = V (Vλ ⊗ Vµ)
Jλ+µ Z(Vλ+µ)
bλ⊗bµ
Z(mλ,µ)
bλ+µ
This is immediate from the multiplicativity of the Wakimoto functor.
(b) Compatibility between bλ and N : (relates g∨ and G∨/N∨)
We want a commutative diagram
Jλ Z(Vλ)
Z(Vλ)
bλ
0NVλ
Proof. Z(Vλ) has a Wakimoto filtration . Because Hom(Jλ, Jµ) = 0 for λ > µ,
NVλ bλ maps Jλ into Jλ ⊂ Z(Vλ), hence NVλ induces an endomorphism of Jλ, whichhas to be zero because End(Jλ) = Q and N is nilpotent.
245
Now we have the following set-up:
N aff g∨ ×G∨/U∨
N g∨ ×G∨/U∨
N g∨ × B
T∨-bundle
open open
T∨-bundle
(For the definition ofN , see [AB09].) By a variant of the Tannakian formalism discussed in
previous lectures, the above data gives a functor
CohG∨×T∨
free (N )→ A.
From this we obtain a functor
Kb
(CohG
∨×T∨free (
N )
)→ Kb(A).
The functor passes to derived categories, complexes on ∂N go to zero40, and we arrive at
our desired functor:F : DG∨(N )→ Db(PI).
All that remains is to see that F induces an equivalence with the anti-spherical module.
Example 31.17. We’ll end today’s lecture with an illustrative example. Recall that wedescribed Z(nat) explicitly for G = GL2 in Lecture 23. We arrived at the following picture.
40This fact must sadly remain a black box here, due to time constraints.
246
(Here the notation aligns with the exercises of Lecture 23.)
247
Lecture 32: Whittaker sheaves and Arkhipov-Bezrukavnikov
theorem
We’ll start today with a simple and beautiful example. But first we need to introduce somenotation.
32.1 Averaging functors
Let N X be the action of an algebraic group, and denote by m and p the correspondingaction and projection maps:
N ×X X
X
m
p
We define two functors
Dbc(X) Db
N(X)
AvN∗
AvN !
by
AvN∗ := m∗(QN (−)[dimN ]) ' m∗p∗[dimN ] ' m∗p
![− dimN ],
AvN ! := m!(QN (−)[dimN ]).
One can imagine these functors as “smearing out” (integrating) our sheaf over the N -orbitsto make it equivariant.
Remark 32.1. 1. We haveDAvN∗ ' AvN !D.
2. The functors AvN∗ and AvN ! fit into adjoint pairs
(For[− dimN ], AvN∗) and (AvN !,For[dimN ]).
3. The functor AvN∗ preserves pD≤0 and AvN ! preserves pD≥0.
Remark 32.2. These are a special case of the induction and restriction functors of Bernstein–Lunts.
32.2 Another fun calculation
Let X = P1 and N =
(1 ∗0 1
). Then one can show that N -equivariant perverse sheaves are
equivalent to perverse sheaves with respect to the B-orbit stratification: PN = P(B). (Thisfollows from the fact that the “forget equivariance” functor is fully faithful when the groupis unipotent.) As we’ve discussed in previous lectures, there are 5 indecomposable objects
248
in P(B): ∆s,∇s, ICs, ICid, and Ts = Pid. Their relationship is captured by the followingdiagram:
ICid
∆s Pid = Ts ∇s
ICs
We have described geometric constructions for ∆s,∇s, ICs, ICid, but we established the exis-tence of Ts via formal properties of highest weight categories. Is there a geometric construc-tion of Ts = Pid as well?
The answer to this question is yes, and it can be realised using averaging functors. Recallthe perverse sheaf Qx!,y∗[1] for x, y 6= 0 that we saw in Lecture 30:
Lemma 32.3.AvN∗(Qx!,y∗[1]) ' AvN !(Qx!,y∗[1]) ' Ts.
For x, y 6= 0, we have
Proof. Step 1: Let j : C → P1 be inclusion of the open Bruhat cell. Averaging functorscommute with equivariant inclusions, so we have
j∗AvN∗(Qx!,y∗[1]) ' AvN∗(j∗Qx!,y∗[1]).
This is an N -equivariant sheaf on C, so it is determined by its global sections.
Exercise 32.4. H i(C,Qx!,y∗[1]) =
Q for i = 0,
0 otherwise.
(Hint: See the “fun calculation” from Lecture 30.)
The shift by [1] in the definition of Av produces a shift by [2] above, hence by Exercise32.4, we have:
j∗AvN∗(Qx!,y∗[1]) ' AvN∗(j∗Qx!,y∗[1]) = QC[1].
Step 2: Let i : 0 → P1 be inclusion of the closed Bruhat cell. Then by adjunction,
Hom•DbN(i∗Q0, AvN∗Qx!,y∗[1]) ' Hom•(For(i∗Q0)[−1],Qx!,y∗[1])
' Hom•(Q0, i!Qx!,y∗[2])
=
Q if • = 0,
0 otherwise.
249
Step 3: By Remark 32.1, we know that AvN∗Qx!,y∗[1] ∈ pD≤0. Moreover, by steps 1 and 2,AvN∗Qx!,y∗ ∈ pD≥0, so we conclude that AvN∗Qx!,y∗ ∈ PN(P1). Hence AvN∗Qx!,y∗ decomposesinto a direct sum of indecomposable perverse sheaves:
AvN∗Qx!,y∗[1] ' IC⊕midid ⊕ · · · ⊕ T⊕ns
Step 4: Finally, by the fun calculation of last week, we have
H∗(AvN∗Qx!,y∗) = H∗(Qx!,y∗) = 0.
This implies that only Ts can occur in the decomposition of AvN∗Qx!,y∗[1] into indecompos-ables, and by steps 1 and 2, there must be only a single copy. We conclude that
AvN∗Qx!,y∗[1] = Ts,
as desired.
Another construction: (which Geordie learned from K. Vilonen)Rough idea: Perhaps, it is easy to calculate the composition series of Qx!,y∗. In this case,
one could dream of taking a limit as x, y → 0 to obtain the big tilting sheaf:
Question: How do we formalize “taking a limit”?
Answer: Nearby cycles!
250
LetF = (j!Q(P1×C)\∆)⊗ (j∗Q(P1×C)\Z)[2].
ThenF|P1×x ' (Qx!,0∗[1])[1]
is a “family of Qx!,0∗ for varying x”.
Exercise 32.5. (Beautiful!) Ts ' ψfF .
Remark 32.6. (Spoiler alert): It’s perverse and B-constructible . . . now what are its globalsections?!
32.3 Lightning introduction to the Whittaker world
Let N ⊂ B ⊂ G be (for a moment) finite groups of Lie type. Fix χ : N → C× a character.Given a representation V of G, consider
V (N,χ) = v ∈ V | n · v = χ(n)v.
This is the space of Whittaker vectors in the representation. By Frobenius reciprocity,
HomN(Cχ, V ) = HomG(IndGN Cχ, V ),
so Whittaker vectors detect whether a representation is “seen” by a 1-dimensional characterof N . Put another way, what irreducible representations can one get by inducing from acharacter of N? Precisely those that admit a non-zero Whittaker vector!
Suppose V = Fun(X,C) for some G-space X. In this setting, what do Whittaker vectorslook like?
Fix an N -orbit U ⊂ X. This is a homogeneous space for N , so U ' N/K for someK ⊂ N . Then
Fun(U) = IndNK C =⊕λ∈N
Vλ ⊗ (V ∗λ )K .
Hence
Fun(U)(N,χ) =
C if χ is trivial on K,
0 otherwise.
So up to scaling, we have “either 0 or 1 Whittaker vectors per N -orbit” and we can tellexactly which orbits admit non-zero Whittaker vectors.
Exercise 32.7. (Important) Let X = G/B and N− be the unipotent radical of the oppositeBorel. Then
N−/[N−, N−] '⊕simpleroots
Fαq .
Hence a character χ : N− → C determines a subset of simple reflections:
I = sα ∈ S | χ 6= 0 on Fαq .
Show that the orbit N−xB/B supports a Whittaker vector if and only if x ∈ IW (the set ofminimal coset representatives of WI\W ).
251
Example 32.8. If χ is nondegenerate (i.e. I = S), then Fun(G/B,C)(N−,χ) = C.
Categorifying Whittaker functions: Now let G,B,N,N−, etc. be defined over Fq. Justas
Fun(X)N−
is categorified by DN−(X),
we will see thatFun(X)(N−,χ) is categorified by D(N−,L)(X).
But what is this category D(N−,L)(X)? Our first step is to find an appropriate way tocategorify the character χ. We do this using the Artin-Schreier sheaf.
Before defining this sheaf, we give some motivation. Recall that
C/Z ∼−→ C×
via z 7→ exp(2πiz). The exponential map is the fundamental additive character of C. Anal-ogously,
A1/Fp∼−→ A1
via the Artin-Schreier map a : x 7→ xp−x. (The map a : A1 → A1 has kernel Fp, so providesthe isomorphism above.) Moreover, we have
a∗(Q`)A1 '⊕
χ:Fp→Q×`
Lχ.
The sheaves Lχ are examples of character sheaves on A1.Define a map p by the following diagram:
N−
N−/[N−, N−]∏
α simple Gαa Ga
p
∼ ∑Fix a non-trivial additive character χ of A1, and define
L := p∗Lχ.
Exercise 32.9. The sheaves Lχ and L are “multiplicative”; i.e. we have m∗L ' L L,where m : N− × N− → N− is multiplication. Multiplicative sheaves are “one-dimensionalcharacter sheaves” or “categorified characters”.
Now letN− X. A (N−,L)-equivariant complex on X is a pair (F , β), where F ∈ Dbc(X)
andβ : a∗F ∼−→ L F
is an isomorphism satisfying the usual cocycle condition. Let
D(N−,L)(X) = category of (N−,L)-equivariant complexes on X.
252
Morphisms in this category are morphisms in Dbc(X) which commute with β. We can define
averaging functorsAvL∗, AvL! : Db
c(V )→ Db(N−,L)(X)
as we did in the beginning of this lecture,
AvL∗ = m∗(L (−))[dimN−], etc.
Important fact: The forgetful functor
For : D(N−,L)(X)→ Dbc(X)
is fully faithful.
Exercise 32.10. LetA1 P1
A1 P1
a
• Show explicitly that the map a : x 7→ xp − x extends to P1. (Find a formula!)
• Show that a breaks all the rules you know about coverings of Riemann surfaces. (Themap a at ∞ is the simplest example of “wild ramification”.)
• Show that j!Lχ∼−→ j∗Lχ.
• Show that H∗(P1, j!Lχ) = 0. (We can interpret this as saying that Lχ is a bit like ourfriend Qx!,y∗ from earlier.)
• Show that j!Lχ is (N−,Lχ)-equivariant.
• Show that AvN∗(j!Lχ) ' AvN !(j∗Lχ) = Ts.
32.4 Bird’s eye view of the rest of the proof
We return to our usual setting: fix G, F` the corresponding affine flag variety, I ⊂ G((t))Iwahori subgroup, I− Iwahori for the opposite Borel, Iu ⊂ I, I−u ⊂ I− pro-unipotent radicals.Let
NK =
1 ∗ ∗0
. . . ∗0 0 1
⊂ G((t)), N−K =
1 0 0
∗ . . . 0∗ ∗ 1
.
We have seen that the antispherical module can be realized as
Masph ' H/〈bx | x 6∈ fW 〉.
Another realization is important in p-adic groups:
Masph = Fun(F`)(N−K ,ψ),
253
the “Whittaker vectors in the principal series”. We might hope that on the level of categories,we would also have
Masph ' D(N−K ,L)(F`).
Problem: N−K orbits on F` are “∞/2-dimensional” (i.e. they have neither finite dimensionnor finite co-dimension). It is difficult to work with sheaves on an ∞-dimensional space.
One solution: Use Drinfeld compactification. This approach is described in [FGV01]. Thisis not yet understood by Geordie.
Another solution: Use “Iwahori-Whittaker” or “baby Whittaker” techniques. The idea isto replace N−K by I−u .
Lemma 32.11. For nondegenerate characters ψ of N−K and χ of I−u ,
Fun(F`(Fq))(N−K ,ψ) = Fun(F`(Fq))(I−u ,χ).
The I−u -orbits on F` are just (opposite) Bruhat cells, so passing to I−u -orbits resolves ourproblem of infinte-dimensional orbits.
DefinePIW ⊂ DIW = D(I−u ,L)(F`), “Iwahori-Whittaker sheaves”.
The irreducible objects in PIW are ICLχχ for x ∈ fW . Last week we constructed a functor
F : Db(CohG∨(N ))→ DI
induced by
CohG∨
free(N )→ PI
V ⊗O 7→ Z(Sat(V )) central sheaf
O(λ) 7→ Jλ Wakimoto sheaf
LetfP I = PI/〈ICx | x 6∈ fW 〉.
The main theorem is the following.
Theorem 32.12. F induces an equivalence
Db(CohG∨(N))
∼−→ Db(fPI).
The averaging functorAvL : PI → DIW
factors over fP I . Now
Db(CohG∨(N )) Db(fP I) Db(PIW ) DIW .
FIW
F Avχ real
254
Theorem 32.13. 1. fP I ' PIW (reasonably easy consequence of part 3. below)
2. Db(PIW ) ' DIW (same argument as Corollary 20.12)
3. Db(CohG∨(N ))
∼−−→FIW
DIW (main difficulty)
As usual, fully-faithfulness of FIW is the main issue.
Step 1: FIW is faithful
A game with induction and restriction functors (using the G/B version of the fun calcu-lation at the start of this lecture) implies that Avψ is faithful. Hence it is enough to show
that F is faithful. The key idea is the following. We have
N → N ⊂ g∨,
and Nreg ⊂ N is open and dense. For V, V ′ ∈ CohG∨
free(N ),
Hom(V, V ′)→ Hom(V |Nreg , V′|Nreg)
is injective. We have a diagram
RepZG∨(N0) ' CohG∨(Nreg) (∗)
CohG∨
free(N ) PI Masph
res.
What corresponds to (∗)?DI ⊃ D 6=idI = 〈ICx | x 6= id〉∆.
Using this, we can define a ⊗-category
DidI := DI/D
6=idI .
Moreover, if P idI is the image of PI in Did
I , then P idI is a ⊗-category with one simple object.
Hence we haveRepG∨
Z−→ PI → P idI .
Using the central functor + derivation, Tannakian formalism (+ a bit of sauce) gives us thefollowing diagram
RepZG(N0) CohG∨(Nreg) P id
I
RepG∨ CohG∨
free(N ) fP I
∼ ∼
res
F
Because res is faithful, F is faithful.
Remark 32.14. The above is an instance of an important theme in Bezrukavnikov’s work:from a two-sided cell c, Lusztig constructed a semi-simple abelian tensor category Jc. Bezrukavnikovobserved that Z provides a central functor RepG∨ → Jc. Bezrukavnikov then uses Z to iden-tify Jc. The theorem is the case when c = id.
255
Step 2: FIW is full:This can be shown using Beilinson’s lemma and a generation argument. It reduces to
checking
ExtiCohG
∨(N )
(V ⊗O,O(λ))FIW−−→ Exti(AvL(Sat(V )), AvL(Jλ))
for V ∈ RepG∨ and λ dominant. The left side is
HomG∨(V,H i(N ,O(λ))) =
0 if i 6= 0 (Frobenius splitting of T ∗B),
Vλ if i = 0.
We can chek that the right side has the same dimension and we are done.
Potential moral of the proof: It is “obvious” that PIW is highest weight. Hence one cando calculations more more easily here than in fP I .
256
Lecture 33: Soergel bimodules, Soergel calculus, and
BGK central sheaves
We will begin today’s lecture by describing a more combinatorial approach to the Heckecategory using Soergel bimodules.
33.1 The Hecke category and Soergel bimodules
We start with some motivation. Recall that given G a (split) finite group of Lie type and Ba Borel subgroup, we can define a C-algebra
Hq := (FunB×B(G,C), ∗).
This is the first incarnation of the Hecke algebra. Iwahori showed that Hq admits a presen-tation which is “independent” of q, in that it only depends on the Coxeter system (W,S)determined by B ⊂ G. This leads us to define a Z[v±1]-algebra using this presentation,which allows us study “all q at once”. This construction now makes sense for any Coxetersystem (W,S). Can we do a similar thing for the Hecke category?
More specifically, recall from Lecture 24 that given an algebraic group G/C, we defineda monoidal41 category
Hs.s. :=
⟨additive category of
semi-simple complexes
⟩⊂ Db
B×B(G,Q).
From this category, we built our final incarnation of the Hecke category:
H := Kb(Hs.s.).
Questions: Can we present H or Hs.s. by generators and relations? Does H makes sensefor any Coxeter system? We can also define H using G/Fq and etale sheaves, do we getequivalent categories after extending scalars?
To answer these questions, we need to introduce the notion of equivariant cohomology.Let K be a group acting on a space X. Define
H∗K(X) := H∗(X ×K EK),
where EK is a classifying space for K (i.e. a path-connected, contractible space with a freeK-action). This construction is sometimes referred to as the “Borel construction”. There isa natural map
X ×K EK → pt×K EK,
which gives H∗K(X) the structure of a graded H∗K(pt)-module. Similarly, given a complexF ∈ Db
K(X), H∗K(X,F) is a graded module over H∗K(pt).
41by the Decomposition Theorem
257
Equivariant cohomology for tori: Let T = C×. Then a classifying space for T is
ET = C∞\∞ = lim→
Cn\0.
HenceH∗C×(pt) = H∗(pt×C× C∞\0) = H∗(P∞) = C[x],
where x is in degree 2. Similarly, if T ' (C×)n, then ET = (C∞\0)n, and
H∗T (pt) ' H∗((P∞)n) = C[x1, . . . , xn]. (33.1)
Note that the isomorphism (33.1) depends on the isomorphism T ' (C×)n. We cangive a more canonical description as follows. Let χ : T → C× be a character. We have anassociated C×-bundle
Lχ := C× ×T,χ ET → pt×T ET.
Borel showed that there is an isomorphism
X∼−→ H2
T (pt,Z)
x 7→ c1(Lχ),
where X is the character lattice of T and c1( Lχ) is the first Chern class of Lχ. This leads toa canonical isomorphism
S•(X) ' H∗T (pt,C),
where S•(X) is the symmetric algebra of the character lattice. This is the Borel isomor-phism.
Remark 33.1. Over C, this can be further simplified. Given χ : T → C×, we can differen-tiate to get a linear functional dχ : LieT → C. Then the Borel isomorphism becomes
O(LieT ) ' H∗T (pt,C).
This is the version which we will use today.
Useful trick: If K ⊂ G is a subgroup, then any model of EG is also a model of EK viarestriction. Examples:
1. T ⊂ B: The mapT\EB → B\EB
is a B/T ' Cn-bundle, so H∗B(pt) ' H∗T (pt).
2. T ⊂ G: The mapT\EG p−→ G\EG
induces a map H∗G(pt)p∗−→ H∗T (pt).
Theorem 33.2. p∗ is injective, and the image is in (O(LieT ))W .
258
By taking equivariant cohomology, we can give an algebraic description of the Hecke category.
Hs.s. ⊂ DbB×B(G)
H∗B×B−−−→H∗B×B(pt)-graded modules
= H∗T×T (pt)-graded modules
= R-gbim,
where R = O(LieT ). The category R-gbim is a monoidal category via − ⊗R −. In mostsituations, taking cohomology loses a lot of information, so the following theorem is especiallyremarkable.
Theorem 33.3. (Soergel) H∗B×B is fully faithful and monoidal on Hs.s..
Hence,Hs.s. → R-gbim.
How can we describe the image? To start, observe that Hs.s. is generated by ICs = CPsP [1]for s ∈ S under ∗,⊕,, [1]; i.e.
Hs.s. = 〈ICs | s ∈ S〉∗,[Z],⊕,
This gives us a way of describing the image:
Sbim := 〈Bs | s ∈ S〉∗,⊕,(1),,
where Bs = H∗B×B(ICs) = R ⊗Rs R(1). (Exercise: prove this!) This is the category ofSoergel bimodules. Unpacking definitions, we obtain equivalences of monoidal categories:
Hs.s. Sbim
H = Kb(Hs.s.) Kb(Sbim)
∼H∗B×B
∼HB×B
Remarkable consequence: To define Hs.s., we need G, B, hundreds of pages of sheaftheory, the decomposition theorem42, and more. To define Sbim, all we need is W LieTand a bit of algebra!
Remark 33.4. 1. This can be seen as a first step toward freeing H from its concreterealization as a category of sheaves. By replacing LieT with a reflection representationh of W , the definition makes sense for any Coxeter group. This led to the proofby Soergel and Elias-W. that the Kazhdan–Lusztig polynomials have non-negativecoefficients.
2. One of the main goals of this course is to approach Bezrukavnikov’s equivalence:
Haffine ' “coherent sheaves on Steinberg”.
Soergel’s theorem tells us that the Hecke category always has a coherent descriptionin terms of Soergel bimodules. One way of understanding Bezrukavnikov’s theorem isthat “Soergel bimodules have another name” in the affine setting.
42can be avoided via the theory of parity sheaves
259
3. One can see Soergel bimodules as “half way” towards a generators and relations de-scription. A generators and relations description was obtained by Elias-W., followingElias-Khovanov, Libedinsky, and Elias. See [EMTW20]. An interesting recent takehas been given by Abe in [Abe19].
33.2 BGK central sheaves for G = SL2
Goal: Describe Z : Rep SL2 → Hext by “generators and relations”.
We start with the left hand side, Rep SL2, which we can describe via the Temperly-Liebcategory TL:
• Objects: Z≥0
• Morphisms: crossingless matchings, up to isotopy,
Let V = C2 be the natural representation of SL2.
Theorem 33.5. 〈V ⊗n〉 ' TL, hence (TL)⊕, ' Rep SL2.
Remark 33.6. This description of Rep SL2 is very useful. For example, an immediateconsequence of the theorem is that giving a functor Rep SL2 → (C,⊗) with C additiveKaroubian is the same as giving a self-dual object X ∈ C of dimension 2.
Giving a generators and relations description of the right hand side Hext is more com-plicated. We’ll start with the unextended case and describe H, the Hecke category corre-sponding to the affine Weyl group W = 〈s, t〉. (Here s is the finite simple reflection.) Fix a“realisation” of W ; that is, fix the data of
• a complex C-vector space h,
• vectors αs, αt ∈ h∗, and
• vectors α∨s , α∨t ∈ h, such that the pairing between the α and the α∨ is given by the
Cartan matrix (2 −2−2 2
).
With such a realisation, W acts via automorphisms of h according to the usual formulas.
Most important choices of realisations:
1. h = C, αs = −αt, α∨s = −α∨t . (Realises canonical quotient W Wf .)
2. hloop = Cαs ⊕ Cδ, αt = −αs + δ, α∨t (δ) = αs(δ) = 0.
260
Remark 33.7. In geometry, h arises from I, whereas hloop arises from InC× (loop rotation).
Now we can describe a diagrammatic version of the Hecke category. Let
HBS = monoidal category generated by Bs and Bt.
Morphisms in HBS are isotopy classes of diagrams generated by
,
subject to a collection of relations. The most important of these relations are:
along with the same relations (and generators) obtained by swapping red↔ blue and s↔ t.
Example 33.8. The diagram
is a morphism BsBtBsBs → BtBsBt.
From here, we can obtain a version of Hs.s.:
HBS −→
formally addshifts, keeponly degreezero maps
−→
add formalsums and
take Karoubienvelope
=: Hdiags.s.
261
Theorem 33.9. (Elias-W.) Using hloop,
Hdiags.s. ' Sbim ' Hs.s..
As a corollary to this theorem, we obtain a rather explicit version of the Hecke categoryH where we can do calculations.
Corollary 33.10. Kb(Hdiags.s. ) ' H.
How do we get Hext? We can extend this constuction to Hext by adding
• a generator ω of Ω,
• morphisms , and
• relations
Now we can translate what we’ve done in previous lectures to the language of diagram-matic Soergel bimodules.
33.3 Braid group categorification
In Soergel bimodule world,
∆s 7−→ 0→ Bs −→ R(1)→ 0
∆w 7−→ “Rouquier complex” with many intriguing properties
∆ω 7−→ 0→ ω → 0
Here a double underline indicates degree zero. Recall that for the affine simple reflection t,we have
translation by $ = tω.
Hence the Wakimoto sheaves
J$ = Btω −→ ω(1), J−$ = ω(−1) −→ Bsω,
262
and
J2$ = J2$ = BtωBtω →
Btωω(1)⊕
ωBtω→ 1(2) ' BtBs −−−−→
Btsω(1)⊕
Bstω(1)−−−→ 1(2).
Exercise 33.11. Show that
J3$ = Btst →Btsω(1)⊕
Bstω(1)→
Bsω(2)⊕
Btω(2)→ Bidω
and
J−2$ = 1(−2)→Bs(−1)⊕
Bt(−1)→ BsBt.
Consider
F = 1(−1) −−−→Bs
⊕Bt
−−−−→ 1(1).
Note that α2 = =
δ in hloop,
0 in h.
We have
FBs ' Bs(−1)→BsBs
⊕BtBs
→ Bs(1) ' Bs(−1)→
Bs(−1)⊕
Bs(1)⊕
BtBs
→ Bs(1) ' BtBs.
Similarly,FBt ' BsBt, BsF ' BsBt, and BtF ' BtBs.
Exercise 33.12. Checek this.
Theorem 33.13. (Elias) z1 := Fω has a canonical central structure. Moreover, there existmaps
1→ z21 , z2
1 → 1
satisfying the Temperly-Lieb relations.
Hence we obtainZ : Rep SL2 → Hext.
263
Example 33.14.
264
Lecture 34: Bezrukavnikov’s equivalence
This is the final lecture of the course!
34.1 Potted history of Koszul duality
A more thorough description of the material in this section can be found in [Wil18, §3].Let O0 be the principal block of category O for g, and ∆x,∇x, Lx, Tx, Px, Ix the standard,costandard, simple, tilting, projective, and injective objects in O0 corresponding to x ∈ W .Let O∨0 be the principal block of category O for the Langlands dual Lie algebra g∨, and∆∨x ,∇∨x , etc. the corresponding objects in O∨0 .
Beilinson-Ginzburg (’86) conjectured that:
1. O0 admits a graded version O0 with a “shift of grading functor” 〈m〉, and all “canoni-
cal” objects above admit lifts ∆x, . . . , Ix.
2. There exists an equivalence of triangulated categories
κ : Db(O0)∼−→ Db(O∨0 )
such that κ 〈−1〉[1] ' 〈1〉 κ, and κ sends
∆x 7→ ∆∨x−1w0, Lx 7→ I∨x−1w0
, Px 7→ L∨x−1w0
The conjecture was proved by Beilinson-Ginzburg-Soergel in [BGS96].
Exercise 34.1. Decategorify κ and deduce something remarkable about Kazhdan–Lusztigpolynomials! (See the first chapter on Koszul duality chapter in [EMTW20].)
Beilinson-Ginzburg (’99): If we compose κ with the Radon transform (∗∆w0) + inversion(g 7→ g−1) and pass to a geometric setting, we obtain a functor
κ : Dmix(B)(G/B)
∼−→ Dmix(B∨)(B
∨\G∨)
is an equivalence with κ 〈−1〉[1] ' 〈1〉 κ, and
ICx 7→ T∨x , ∆x 7→ ∆∨x , ∇x 7→ ∇∨x , T∨x 7→ IC∨x .
After seeing this more symmetric version, one starts to start to dream a little.
Dream:
1. It should work for any Kac-Moody group. (We’ve removed w0 from the formulas!)
2. It should be “monoidal”; i.e. we should roughly have
(semisimple, ∗) ' (tiltings, ∗∨).
265
Unfortunately, this is very much in dreamland.
Basic problem: Tx ∈ D(B)(G/B) lifts to DB(G/B) if and only if x = id.
Remark 34.2. We have seen this a number of times for P1 (e.g. Example 30.7): Ts doesnot have an equivariant structure.
One way we could try to resolve this problem is to work in DU(G/B) instead, becauseTx does lift to this category. However, in DU(G/B) we no longer have convolution. To fixthis, we can go one step further and work in DU(G/U), which gives us access to Tx andconvolution. But unfortunately, this introduces new problems.
Second problem: ∗ is not exact on DU(G/U).
To illustrate this problem, we can consider a toy example
34.2 A toy example
Let G = S1 (if you like Lie groups) or Gm (if you prefer the algebraic setting).
Exercise 34.3. Given L1,L2 ∈ Loc(S1),
(L1 ∗ L2)1 ' H∗(L1 ⊗ L∗2).
The subscript 1 denotes the stalk at the identity. As an additional exercise, describe themonodromy of L1 ∗ L2 in terms of the monodromy on L1 and L2.
The basic issue is the following. We have seen (Example 17.6) that for a local systemL/S1 given by monodromy µ V ,
H0(S1,L) = V µ and H1(S1,L) = Vµ.
For V finite dimensional, V µ 6= 0 ⇐⇒ Vµ 6= 0. Hence for any finite-dimensional localsystem L on S1, either H∗(S1,L) = 0 or H∗(S1,L) 6= 0 in both degree 0 and degree 1. Thisimplies (by Exercise 34.3) that ∗ is either zero or not exact.
Solution: Infinite-dimensional local systems! Take L to be the local system correspondingto
C[[t]] µ = 1 + t.
Note that C[[t]]µ = 0, C[[t]]µ = C[[t]]/(1− µ) = C.
Remark 34.4. Roughly one can think of L as an infinite-dimensional Jordan block. . . 0
1 11 1
1 10 1
,
which has one-dimensional coinvariants and no invariants.
Exercise 34.5. m!(L L)[1] ' L, hence we have an exact ∗-product!
266
34.3 Torus monodromic sheaves a la Bezrukavnikov-Yun
The ideas we encountered in the toy example above were formalized by Bezrukavnikov-Yunin [BY13] (c.f. Beilinson’s 1983 ICM talk). Let
D(B999 G 99
9 B) := 〈q∗D(B\G/B)〉= 〈p∗D(U\G/B)〉,
whereU\G/U
U\G/B
B\G/B
p
q
Example 34.6. 1. If G = C×,
D(B999 G 99
9 B) = 〈p∗D(pt)〉= full subcategory of Db
c(C×) generated by QC
= F | Hi(F) local systems with unipotent monodromy.
2. If G = SL2, G/U ' C2\0, so D(U\G/B) ' 〈IC0, ICP1〉. Hence
D(B999 G 99
9 B) = 〈QC× ,QC2\0〉 ⊂ DU(G/U).
By replacing B with U , we get a convolution product
U∗ : D(B999 G 99
9 B)×D(B999 G 99
9 B)→ D(B999 G 99
9 B)
from the diagram
G×U G/U G/U
G/U G/U
using the same definition as previously, except that m : G×U G/U is not proper, so we needto make a choice of which pushforward m∗ or m! to use. We choose m!.
Very technical point: (cf. appendix by Yun to [BY13])
1. One can completeD(B999 G 99
9 B) to the “free monodromic completion” D(B999 G 99
9 B)
to allow pro-local systems (like C[[t]] from earlier) along the fibres of p.
267
2. The convolution productU∗ extends to D(B
999 G 999 B). Moreover, we can define a
categoryTilt ⊂ D(B
999 G 999 B)
of “free monodromic tilting sheaves,” andU∗ is exact on Tilt.
Theorem 34.7. [BY13]
(Tilt,U∗) ' (SBim,⊗
R
)
The category on the right side of this equivalence consists of Soergel bimodules for R =O(LieT∨) = O((LieT )∗) (note the dual!), completed along the grading.
Idea of proof: Suppose thatX
Y
p is a T -torsor. Define a category
Db(X 999 T ) = 〈p∗Db
c(Y )〉∆ ⊂ Db(X)
of “unipotently monodromic sheaves”. The fundamental observation of Verdier is that everyF ∈ Db(X 99
9 T ) has a canonical monodromy action of π1(T ). The action is unipotent. In
other words, Db(X 999 T ) is linear over Q[π1(T )]. Taking logs, we obtain that Db(X 99
9 T )
is R-linear.
Theorem 34.8. [BY13] The functor V : Tilt→ SBim is monoidal and fully-faithful.
Adding weights, one gets monoidal equivalences
(Tiltmixed
, ∗) ' (SBim,⊗R
) ' (HG∨
s.s., ∗),
and hence taking homotopy categories,
Dmixed(B999 G 99
9 B) ' H ' Dmixed(B∨\G∨/B∨).
This achieves the dream of a “monoidal Koszul duality”.
A beautiful feature: Let π : U\G/U → U\G/B be projection. Then
π∗(Tx)[rankT ] = Tx ∈ DU(G/B).
In other words, taking coinvariants of free monodromic tilting sheaves gives indecomposibletilting sheaves downstairs.
Remark 34.9. With enough43 homological algebra, one can do all of the above algebraically(and also mod p), c.f. [AMRW19]. Some of the constructions there are very adhoc. Recentfoundational work of Hogencamp and Makisumi explains much more satisfactorily what isgoing on.
43approximately 250 pages
268
34.4 Bezrukavnikov’s equivalence
The constructible side: Associated to our choice T ⊂ B ⊂ G, we have the familiar zooof characters:
• the pro-unipotent radical I0 ⊂ I ⊂ G((t)),
• the extended affine flag variety (aka “affine base affine space”) F` = G((t))/I0,
• the affine flag variety F` = G((t))/I,
• the affine Grassmannian Gr = G((t))/G[[t]],
• projections
F` T -torsor−−−−→ F`→ Gr,
• derived categories
DII := D(I\G((t))/I),
DI0I := D(I0\G((t))/I) DI0I , free monodromic completion
DI0I0 := D(I0 999 G((t)) 999 I0) DI0I0 , free monodromic completion
• and actions
DI0I0 DI0I DII
DI0I0 DI0I DII .
The coherent side: Associated to G∨, we have
N ∨ g∨
N ∨ g∨,
where the left vertical arrow is the Springer resolution and the right vertical arrow is theGrothendieck-Springer resolution sending (x, b) ∈ g∨ = (x, b) ⊂ g∨×B∨ | x ∈ b to x ∈ g∨.
Recall: Given Xf−→ Y proper, one can construct a convolution structure on (Db(X×Y X), ∗)
(e.g. if X is a finite set and Y is a point, then Db(X ×Y X) ' X × X matrices of chaincomplexes of vector spaces). Earlier we used this construction to give a coherent realizationof the affine Hecke algebra via the Kazhdan–Lusztig isomorphism:
Haff ' KG∨×C×(N ∨ ×N N ∨).
We need a derived version of this (which will mostly disappear in a moment).Recall that X ×Y X is constructed by gluing Spec(B⊗AB) for an affine cover SpecB →
SpecA of f . Similarly, we can construct XL×Y X by gluing the dg-schemes Spec(B
L⊗A B)
for an affine cover SpecB → SpecA of f .
Key points:
269
1. Db(XL×Y X) is still triangulated and monoidal.
2. If TorOY>0 (OX ,OX) = 0, then XL×Y X = X ×Y X.
From this we define our coherent characters:
• Stg∨ := g∨L×g∨ g
∨ = g∨ ×g∨ g∨,
• StN∨ := g∨L×g∨ N ∨ = g∨ ×g∨ N ∨, and
• StL := N ∨L×g∨ N ∨ (6= N ∨ ×g∨ N ∨ = N ×N∨ N ∨).
Theorem 34.10. (Bezrukavnikov) There exist vertical equivalences
Db CohG∨(Stg∨) Db CohG
∨(StN∨) Db CohG
∨(StL)
DI0I0 DI0I DII
(a)∼ (b)∼ (c)∼
making the diagram of module categories commute.
Remarks on the proof:
1. Equivalences (b) and (c) and compatibility are reasonably straightforward consequencesof (a), which is where Bezrukavnikov spends most of the paper.
2. It is technically convenient to instead prove
DI0I0 ' D(CohG∨(Stg∨)),
where St denotes the formal completion of Stg∨ along the preimage of N ∨.
3. The proof relies heavily on ideas in [AB09].
The theorem has several remarkable consequences, but exploring them will have to waituntil the next course!
270
References
[AB09] S. Arkhipov and R. Bezrukavnikov. Perverse sheaves on affine flags and Lang-lands dual group. Israel Journal of Mathematics, 170(1):135, 2009.
[Abe19] Noriyuki Abe. On Soergel bimodules. arXiv:1901.02336, 2019.
[AMRW19] Pramod Achar, Shotaro Makisumi, Simon Riche, and Geordie Williamson.Koszul duality for Kac–Moody groups and characters of tilting modules. Journalof the American Mathematical Society, 32(1):261–310, 2019.
[Art24] Emil Artin. Uber eine neue art von L-reihen. In Abhandlungen aus dem Mathe-matischen Seminar der Universitat Hamburg, volume 3, pages 89–108. Springer,1924.
[Bei87] Alexander A Beilinson. On the derived category of perverse sheaves. In K-theory,Arithmetic and Geometry, pages 27–41. Springer, 1987.
[Bez04] Roman and Bezrukavnikov. On tensor categories attached to cells in affine weylgroups. In Representation theory of algebraic groups and quantum groups, pages69–90. Mathematical Society of Japan, 2004.
[Bez16] Roman Bezrukavnikov. On two geometric realizations of an affine Hecke algebra.Publications mathematiques de l’IHES, 123(1):1–67, 2016.
[BGS96] Alexander Beilinson, Victor Ginzburg, and Wolfgang Soergel. Koszul dualitypatterns in representation theory. Journal of the American Mathematical Soci-ety, 9(2):473–527, 1996.
[Bum98] Daniel Bump. Automorphic Forms and Representations, volume 55. CambridgeUniversity Press, 1998.
[Buz17] Kevin Buzzard. Automorphic forms and the Langlands program, MSRIsummer school. Available at https: // www. youtube. com/ playlist? list=
PLhsb6tmzSpiysoRR0bZozub-MM0k3mdFR , 2017.
[BY13] Roman Bezrukavnikov and Zhiwei Yun. On koszul duality for Kac-Moodygroups. Representation Theory of the American Mathematical Society, 17(1):1–98, 2013.
[CF67] JWS Cassels and A Frohlich. Algebraic Number Theory. Thompson Book Com-pany, Inc., 1967.
[CG09] Neil Chriss and Victor Ginzburg. Representation theory and complex geometry.Springer Science & Business Media, 2009.
[Clo06] L Clozel. The Sato-Tate conjecture. Current Developments in Mathematics,2006(1):1–34, 2006.
271
[Con01] Keith Conrad. History of class field theory. This unpublished essay isavailable online as a PDF file at www. math. uconn. edu/ ~ kconrad/ blurbs/gradnumthy/ cfthistory. pdf , 2001.
[DBB83] Pierre Deligne, Alexander A Beilinson, and Joseph Bernstein. Faisceaux pervers.Asterisque, 100, 1983.
[Del97] Pierre Deligne. Action du groupe des tresses sur une categorie. Inventionesmathematicae, 128(1):159–175, 1997.
[DGNO10] Vladimir Drinfeld, Shlomo Gelaki, Dmitri Nikshych, and Victor Ostrik. Onbraided fusion categories i. Selecta Mathematica, 16(1):1–119, 2010.
[EMTW20] Ben Elias, Shotaro Makisumi, Ulrich Thiel, and Geordie Williamson. Introduc-tion to Soergel bimdules. Springer International Publishing, 2020.
[EO03] Pavel Etingof and Viktor Ostrik. Module categories over representations of q(2)and graphs. arXiv preprint math/0302130, 2003.
[Far18] Laurent Fargues. La courbe. In Proc. Int. Cong. of Math, volume 1, pages261–290. World Scientific, 2018.
[FGV01] Edward Frenkel, Dennis Gaitsgory, and Kari Vilonen. Whittaker patterns in thegeometry of moduli spaces of bundles on curves. Annals of Mathematics, pages699–748, 2001.
[FGV02] Edward Frenkel, Dennis Gaitsgory, and Kari Vilonen. On the geometric Lang-lands conjecture. Journal of the American Mathematical Society, 15(2):367–417,2002.
[Gai05] Dennis Gaitsgory. The notion of category over an algebraic stack.arXiv:math/0507192 [math.AG], 2005.
[Gro11] Benedict Gross. Representation Theory and Number Theory, Eilenberg lecturesat Columbia University. Available at https: // www. youtube. com/ playlist?list= PL5E0D6DC4BCD8309D , 2011.
[Kac90] Victor G Kac. Infinite-dimensional Lie algebras. Cambridge University Press,1990.
[KL87] David Kazhdan and George Lusztig. Proof of the Deligne-Langlands conjecturefor Hecke algebras. Inventiones mathematicae, 87(1):153–215, 1987.
[Kle93] Felix Klein. Vorlesungen uber das Ikosaeder und die Auflosung der Gleichungenvom funften Grade. Birkhauser Verlag, Basel; B. G. Teubner, Stuttgart, 1993.Reprint of the 1884 original, Edited, with an introduction and commentary byPeter Slodowy.
272
[Kri05] Martin H. Krieger. A 1940 letter of Andre Weil on analogy in mathemat-ics. Notices Amer. Math. Soc., 52(3):334–341, 2005. Excerpted from ıt Do-ing mathematics [World Scientific Publishing Co., Inc., River Edge, NJ, 2003;MR1961400].
[Lan90] R. P. Langlands. Representation theory: its rise and its role in number theory.In Proceedings of the Gibbs Symposium (New Haven, CT, 1989), pages 181–210.Amer. Math. Soc., Providence, RI, 1990.
[Lur07] Jacob Lurie. Derived algebraic geometry ii: Noncommutative algebra. arXivpreprint math/0702299, 2007.
[Man14] Yuri I. Manin. Forgotten Motives: the Varieties of Scientific Experience, 2014.
[Mil97] James S Milne. Class Field Theory. Lecture notes available at https: // www.jmilne. org/ math/ CourseNotes/ cft. html , 1997.
[Miy11] Katsuya Miyake. Takagi’s class field theory: From where? and to where? RIMSKokyuroku Bessatsu, B25:125–160, 2011.
[Mor18] Sophie Morel. Beilinson’s construction of nearby cycles and gluing. preprint,available at https: // web. math. princeton. edu/ ~ smorel/ gluing. pdf ,2018.
[MS16] Barry Mazur and William Stein. Prime Numbers and the Riemann Hypothesis.Cambridge University Press, 2016.
[Ost03] Victor Ostrik. Module categories, weak Hopf algebras and modular invariants.Transformation Groups, 8(2):177–206, 2003.
[Sch17] Peter Scholze. p-adic geometry. ICM Lecture, arXiv:1712.03708, 2017.
[Ser55] Jean-Pierre Serre. Faisceaux algebriques coherents. Annals of Mathematics,pages 197–278, 1955.
[Ser79] Jean-Pierre Serre. Local fields, volume 67 of Graduate Texts in Mathematics.Springer-Verlag, New York-Berlin, 1979. Translated from the French by MarvinJay Greenberg.
[SVW18] Wolfgang Soergel, Rahbar Virk, and Matthias Wendt. Equivariant motives andgeometric representation theory.(with an appendix by f. hormann and m. wendt).arXiv:1809.05480, 2018.
[SW18] Wolfgang Soergel and Matthias Wendt. Perverse motives and graded derivedcategory O. Journal of the Institute of Mathematics of Jussieu, 17(2):347–395,2018.
273
[Tat79] J. Tate. Number theoretic background. In Automorphic forms, representa-tions and L-functions (Proc. Sympos. Pure Math., Oregon State Univ., Corval-lis, Ore., 1977), Part 2, Proc. Sympos. Pure Math., XXXIII, pages 3–26. Amer.Math. Soc., Providence, R.I., 1979.
[TT90] Robert W Thomason and Thomas Trobaugh. Higher algebraic K-theory ofschemes and of derived categories. In The Grothendieck festschrift, pages 247–435. Springer, 1990.
[Wey16] Hermann Weyl. Uber die gleichverteilung von zahlen mod. eins. MathematischeAnnalen, 77(3):313–352, 1916.
[Wil18] Geordie Williamson. Parity sheaves and the Hecke category. Proc. Int. Cong.of Math., Rio de Janeiro, Vol. 1:979 – 1016, 2018.
[You16] Alexander Youcis. Weil-Deligne representations and p-adic Hodge theory: moti-vation. Available at: https: // ayoucis. files. wordpress. com/ 2016/ 11/
weil-deligne-representations-2. pdf , 2016.
274
![On the Tate and Langlands–Rapoport conjectures for special … · 2017. 12. 30. · arXiv:1210.6629v1 [math.NT] 24 Oct 2012 On the Tate and Langlands–Rapoport conjectures for](https://static.fdocuments.net/doc/165x107/5fd0a1a17473fd40f74160fa/on-the-tate-and-langlandsarapoport-conjectures-for-special-2017-12-30-arxiv12106629v1.jpg)
![arXiv:1004.3516v1 [math.NT] 20 Apr 2010 · 2010-04-21 · arXiv:1004.3516v1 [math.NT] 20 Apr 2010 Tel Aviv University School of Mathematical Sciences. The Langlands-Shahidi Method](https://static.fdocuments.net/doc/165x107/5f8fa5afee0c333b8777a567/arxiv10043516v1-mathnt-20-apr-2010-2010-04-21-arxiv10043516v1-mathnt.jpg)
![GAUGE THEORY AND LANGLANDS DUALITY …arXiv:0906.2747v1 [math.RT] 15 Jun 2009 S´eminaire BOURBAKI Juin 2009 61`eme ann´ee, 2008-2009, no 1010 GAUGE THEORY AND LANGLANDS DUALITY by](https://static.fdocuments.net/doc/165x107/5e322d15fa0bc263eb1b61c0/gauge-theory-and-langlands-duality-arxiv09062747v1-mathrt-15-jun-2009-seminaire.jpg)
