Geometric Langlands Correspondence

⚡ Key Takeaways

Geometric Langlands Correspondence — A vast conjectural (now substantially proven) equivalence between two categories attached to a smooth projective curve $X$ over an algebraically closed field: the category of (D-modules / perverse sheaves) on the moduli stack $\mathrm{Bun}_G$ of principal $G$ -bundles, and the category of quasi-coherent sheaves on the moduli stack $\mathrm{LocSys}_{\check G}$ of $\check G$ -local systems. Key result: "Hecke eigensheaves" on the automorphic side correspond to points (local systems) on the spectral side, geometrizing the classical Langlands dictionary "automorphic forms ↔ Galois representations." Where it applies: algebraic geometry, representation theory, conformal field theory, and the physics of $S$ -duality in 4d $\mathcal N=4$ gauge theory. Who/when: envisioned by Drinfeld and Laumon (1980s–90s), formulated categorically by Beilinson–Drinfeld, Frenkel, Gaitsgory; the de Rham conjecture for $\mathrm{GL}_n$ and reductive $G$ was announced as proven by Gaitsgory–Raskin and collaborators (2024).

💡 The Core Insight

Classical Langlands says: automorphic representations of $G(\mathbb A_F)$ are parametrized by $\check G$ -valued Galois representations, with the Hecke eigenvalues at each unramified place $v$ recovered from the Frobenius conjugacy class via the Satake isomorphism. Geometrize each ingredient. A function on $G(F)\backslash G(\mathbb A_F)/G(\mathcal O)$ becomes, over a function field $F = k(X)$ , a function on the $k$ -points of $\mathrm{Bun}_G$ ; by Grothendieck's sheaf-function dictionary, such functions arise as traces of Frobenius on $\ell$ -adic sheaves. So replace automorphic functions by sheaves on $\mathrm{Bun}_G$ . A Galois representation $\mathrm{Gal}(\bar F/F) \to \check G$ unramified everywhere becomes a $\check G$ -local system on $X$ , i.e. a point of $\mathrm{LocSys}_{\check G}$ . The Hecke eigenvalue condition becomes the condition that a sheaf is a Hecke eigensheaf with eigenvalue the chosen local system. Promoting the parametrization to a full categorical equivalence is the geometric Langlands conjecture.

DefinitionModuli of bundles and the Hecke stack

Let $X$ be a smooth projective geometrically connected curve over an algebraically closed field $k$ , and $G$ a connected reductive group over $k$ .

$\mathrm{Bun}_G = \mathrm{Bun}_G(X)$ is the algebraic stack classifying principal $G$ -bundles on $X$ . It is smooth, of dimension $(g-1)\dim G$ , not quasi-compact, with stratifications by Harder–Narasimhan type.
The Hecke stack $\mathrm{Hecke}_G$ classifies triples $(\mathcal P_1, \mathcal P_2, x, \beta)$ where $\mathcal P_1, \mathcal P_2 \in \mathrm{Bun}_G$ , $x \in X$ , and $\beta: \mathcal P_1|_{X\setminus x} \xrightarrow{\sim} \mathcal P_2|_{X\setminus x}$ is an isomorphism away from $x$ . It sits in a correspondence:

1

\quad \mathrm{Bun}_G \xleftarrow{\;\overleftarrow{h}\;} \mathrm{Hecke}_G \xrightarrow{\;\overrightarrow{h}\times \mathrm{supp}\;} \mathrm{Bun}_G \times X

Derivation hint: the two maps record $\mathcal P_1$ and $(\mathcal P_2, x)$ respectively; the fibers of $\overleftarrow h$ over a fixed $\mathcal P_1$ and $x$ are the affine Grassmannian $\mathrm{Gr}_{G,x}$ .

DefinitionHecke functors via geometric Satake

The geometric Satake equivalence furnishes a tensor functor from finite-dimensional $\check G$ -representations to $G(\mathcal O)$ -equivariant perverse sheaves on the affine Grassmannian:

2

\quad \mathrm{Sat}:\ \mathrm{Rep}(\check G)\ \xrightarrow{\ \sim\ }\ \mathrm{Perv}_{G(\mathcal O)}(\mathrm{Gr}_G),\qquad V \mapsto \mathcal A_V

Derivation hint: convolution on $\mathrm{Gr}_G$ defines a symmetric monoidal structure; Tannakian reconstruction recovers $\check G$ as the automorphism group of the fiber functor $H^*(\mathrm{Gr}_G,-)$ . For each $V \in \mathrm{Rep}(\check G)$ one obtains a Hecke functor acting on sheaves over $\mathrm{Bun}_G \times X$ :

3

\quad \mathrm{H}_V(\mathcal F) \;=\; \overrightarrow{h}_!\big(\overleftarrow{h}{}^*\mathcal F \otimes \mathcal A_V\big)

Derivation hint: pull back along $\overleftarrow h$ , twist by the Satake sheaf supported on Hecke modifications of type $V$ , push forward along $\overrightarrow h$ recording the new bundle and modification point.

DefinitionHecke eigensheaf

A nonzero object $\mathcal F \in \mathrm{D\text{-}mod}(\mathrm{Bun}_G)$ is a Hecke eigensheaf with eigenvalue $E \in \mathrm{LocSys}_{\check G}(X)$ if there are functorial isomorphisms, compatible with the tensor structure of $\mathrm{Rep}(\check G)$ ,

4

\quad \mathrm H_V(\mathcal F)\ \simeq\ \mathcal F\ \boxtimes\ V_E,\qquad V \in \mathrm{Rep}(\check G),

where $V_E$ is the local system on $X$ associated to $E$ and the representation $V$ . Derivation hint: the eigenvalue equation generalizes " $T_v f = a_v f$ "; the appearance of $V_E$ on $X$ encodes that the eigenvalue depends on the modification point $x$ through the local system.

TheoremGeometric Langlands Conjecture — de Rham form

Let $k = \mathbb C$ (or char $0$ ), $G$ reductive with Langlands dual $\check G$ . There is a canonical equivalence of DG categories

5

\quad \mathbb L_G:\ \ \mathrm{D\text{-}mod}\big(\mathrm{Bun}_G\big)\ \xrightarrow{\ \sim\ }\ \mathrm{IndCoh}_{\mathcal N}\big(\mathrm{LocSys}_{\check G}\big),

exchanging the Hecke functors $\mathrm H_V$ with the operations of tensoring by the tautological vector bundles $\mathrm{ev}_x^* V$ , and normalized so that the skyscraper-like automorphic sheaf attached to a $\check G$ -local system $E$ goes to (a twist of) the skyscraper $\mathcal O_E$ supported at the point $[E]$ . Derivation hint: $\mathcal O_E$ is a Hecke eigenobject with eigenvalue $E$ by construction of the tautological bundles on $\mathrm{LocSys}_{\check G}$ ; transporting it back via $\mathbb L_G^{-1}$ yields the eigensheaf with eigenvalue $E$ .

⚠️ The subscript

\mathcal N

is essential, not decorative

The naive statement with $\mathrm{QCoh}(\mathrm{LocSys}_{\check G})$ is false for nonabelian $G$ . The correct spectral category is $\mathrm{IndCoh}_{\mathcal N}$ : ind-coherent sheaves with nilpotent singular support, where $\mathcal N \subset \mathrm{Sing}(\mathrm{LocSys}_{\check G})$ is the global nilpotent cone (Arinkin–Gaitsgory). $\mathrm{LocSys}_{\check G}$ is a (derived) stack with self-dual cotangent complex, so its singular support theory is governed by the Hitchin nilpotent cone. Forgetting this condition breaks the equivalence — the automorphic side is "smaller" than $\mathrm{QCoh}$ and matches exactly the nilpotent-singular-support subcategory.

ProofStrategy of the proof — Gaitsgory–Raskin et al.

The 2024 proof for general reductive $G$ in the de Rham setting proceeds through a sequence of structural reductions. We sketch the architecture. 1. Construct the functor. Build $\mathbb L_G$ (or its inverse $\mathbb L_G^{-1}$ ) using the Poincaré / Whittaker sheaf on $\mathrm{Bun}_G$ and the action of Hecke functors. The vacuum Whittaker object provides a "kernel" via which one defines the comparison. 2. Compatibility with Hecke. Geometric Satake (eq. $[2]$ ) guarantees the Hecke functors form a $\mathrm{Rep}(\check G)$ -action; this matches the $\mathrm{QCoh}(\mathrm{LocSys}_{\check G})$ -module structure on the spectral side. Both categories become module categories over the same symmetric monoidal category $\mathrm{Rep}(\check G)^{\otimes X}$ , organized by the factorization structure along $X$ . 3. Localize on $\mathrm{LocSys}_{\check G}$ . Decompose both sides according to the support over the coarse space of local systems. The hardest pieces are the cuspidal/elliptic part and the contributions of reducible local systems, where the singular-support condition $\mathcal N$ enters. 4. Reduce to Eisenstein and constant-term functors. Parabolic induction (geometric Eisenstein series) builds objects from Levi subgroups $M \subsetneq G$ . By induction on the semisimple rank one controls the non-cuspidal part of $\mathrm{D\text{-}mod}(\mathrm{Bun}_G)$ in terms of $\mathrm{Bun}_M$ . 5. Establish the cuspidal equivalence. On the cuspidal piece (where the spectral side is supported on irreducible local systems, an honest scheme-like locus), one proves the equivalence using vanishing of certain functors and the Hecke eigensheaf existence (Frenkel–Gaitsgory–Vilonen for $\mathrm{GL}_n$ generalized). 6. Glue. Reassemble cuspidal and Eisenstein contributions; verify the gluing is compatible across the recollement, yielding the global equivalence $[5]$ . The compatibility of all charges and the matching of $t$ -structures requires the full derived/ $\infty$ -categorical machinery. $\blacksquare$ (architecture only — each step is itself a major theorem)

∎

EW

Edward Witten

b. 1951

Though geometric Langlands is a theorem of pure mathematics, its deepest conceptual reframing came from physics. In the 2006 paper of Anton Kapustin and Edward Witten, "Electric–Magnetic Duality and the Geometric Langlands Program," the correspondence is realized as mirror symmetry of Hitchin moduli spaces, itself descending from the $S$ -duality of 4-dimensional $\mathcal N=4$ super Yang–Mills theory compactified on a Riemann surface. The Langlands dual group $\check G$ is exactly the Goddard–Nuyts–Olive (Montonen–Olive) magnetic dual of $G$ . Witten's perspective turned the Hecke eigensheaf condition into the statement that certain 't Hooft / Wilson line operators act diagonally, and recast $A$ -branes and $B$ -branes on the two Hitchin moduli spaces as the automorphic and spectral categories. This is the prototype of the quantum geometric Langlands family interpolated by the level parameter $\Psi$ .

🧠 Intuition

Think of three rungs on a ladder of abstraction.

Rung 0 (numbers): Classical Langlands over $\mathbb Q$ : an eigenform $f$ (e.g. a modular form) has Hecke eigenvalues $a_p$ ; these are the Frobenius traces of a Galois representation $\rho$ . Dictionary: eigenform ↔ Galois representation.
Rung 1 (functions): Over a function field $k(X)$ , automorphic forms are functions on $\mathrm{Bun}_G(\mathbb F_q)$ . Hecke operators act by averaging over modifications at a point. The Frobenius eigenvalues now organize into a $\check G$ -local system on $X$ in the étale sense.
Rung 2 (sheaves): Replace the function by the sheaf whose trace-of-Frobenius is that function (Grothendieck's dictionary). Now $X$ can live over $\mathbb C$ , Frobenius disappears, and the eigenvalue is a flat connection (a de Rham local system). The Hecke eigenvalue equation $[4]$ says: applying a Hecke modification at $x$ tensors the sheaf by the fiber of the local system at $x$ .

The miracle is that on the spectral side these eigensheaves organize into a geometry: the space of all possible eigenvalues is the moduli $\mathrm{LocSys}_{\check G}$ , and the equivalence $[5]$ says the entire automorphic category is "functions/sheaves on that space of eigenvalues." A Hecke eigensheaf is then literally a skyscraper sheaf on $\mathrm{LocSys}_{\check G}$ — the categorical analogue of a delta function concentrated at one eigenvalue.

Figure

The geometric Langlands correspondence as a duality of moduli categories

```svg

```

ExampleAbelian case:

G = \mathrm{GL}_1 = \mathbb G_m

, the geometric class field theory

Given. Let $G = \mathbb G_m$ , so $\check G = \mathbb G_m$ as well. Then $\mathrm{Bun}_{\mathbb G_m} = \mathrm{Pic}(X)$ is the Picard stack of line bundles, and $\mathrm{LocSys}_{\mathbb G_m}$ is the moduli of rank-1 local systems (flat line bundles).

Steps. 1. A rank-1 local system $E$ on $X$ is a homomorphism $\pi_1(X) \to \mathbb G_m$ (or a flat line bundle), equivalently a character $\chi_E$ . 2. The Hecke modification at $x$ for $\mathrm{GL}_1$ adds the divisor $[x]$ : it is the Abel–Jacobi map $X \to \mathrm{Pic}^{d+1}/\mathrm{Pic}^d$ . 3. The eigensheaf condition becomes a character sheaf $\mathcal F_E$ on $\mathrm{Pic}(X)$ — a rank-1 local system on the Picard stack — whose pullback along Abel–Jacobi recovers $E$ . 4. Such $\mathcal F_E$ exists and is unique up to scalar, constructed from $E$ via the Fourier–Mukai / theta machinery (Deligne; Laumon's geometric class field theory).

Boxed answer.

\boxed{\ \mathbb L_{\mathbb G_m}:\ \mathrm{D\text{-}mod}(\mathrm{Pic}\,X)\ \xrightarrow{\sim}\ \mathrm{QCoh}(\mathrm{LocSys}_{\mathbb G_m}\,X)\ \text{ is the Fourier–Mukai transform.}\ }

Here $\mathcal N = 0$ (no correction needed): for tori, $\mathrm{LocSys}$ is smooth and the naive $\mathrm{QCoh}$ statement is correct. This is the historical seed of the whole subject.

ExampleGenus-0 baby case: rigidity on

X = \mathbb P^1

Given. $X = \mathbb P^1$ , $G = \mathrm{GL}_n$ . There are no nontrivial unramified local systems because $\pi_1(\mathbb P^1) = 1$ , so the only everywhere-unramified $\check G$ -local system is trivial.

Steps. 1. $\mathrm{LocSys}_{\check G}(\mathbb P^1)$ as a derived stack is the (derived) point $\mathrm{pt}/\check G$ — local systems are forced trivial, with automorphisms. 2. Spectral side: $\mathrm{IndCoh}_{\mathcal N}(\mathrm{pt}/\check G) = \mathrm{Rep}(\check G)$ (with the nilpotent condition automatically satisfied; $\mathrm{Sing} = \check{\mathfrak g}^*//\check G$ enters via the derived structure). 3. Automorphic side: $\mathrm{Bun}_G(\mathbb P^1)$ is, by Grothendieck's classification, a disjoint union indexed by dominant coweights (splitting types $\mathcal O(a_1)\oplus\cdots$ ). 4. The equivalence pairs the combinatorics of Grothendieck splitting types with $\mathrm{Rep}(\check G)$ via the geometric Satake equivalence localized at the single special point.

Boxed answer.

\boxed{\ \mathrm{D\text{-}mod}\big(\mathrm{Bun}_G(\mathbb P^1)\big)\ \simeq\ \mathrm{IndCoh}_{\mathcal N}\big(\mathrm{pt}/\check G\big)\ \cong\ \mathrm{Rep}(\check G)\text{-mod}\ }

Even in this degenerate case the derived/nilpotent structure is necessary to make both sides match — illustrating why naive sheaf theory fails.

⚠️ It is not a bijection of objects, and not "QCoh"

Three frequent errors: 1. "It's a bijection automorphic forms ↔ Galois reps." — No; it is an equivalence of categories, far richer than a parametrization of irreducibles. The classical bijection is a shadow obtained by passing to simple objects / eigenobjects. 2. "The spectral side is $\mathrm{QCoh}(\mathrm{LocSys}_{\check G})$ ." — Only for tori. In general you need $\mathrm{IndCoh}_{\mathcal N}$ (eq. $[5]$ ), with nilpotent singular support. 3. "D-modules can be replaced by ordinary sheaves with no consequence." — The choice of sheaf theory (D-modules vs. $\ell$ -adic vs. constructible $\mathbb C$ -sheaves) determines whether you are doing the de Rham, étale (arithmetic function-field), or Betti version. These are genuinely different theorems with different available proofs.

CounterexampleWhy "naive QCoh" fails: a Hecke eigensheaf has the wrong category around it

Take $\check G$ nonabelian and a reducible local system $E$ sitting on the locus where $\mathrm{LocSys}_{\check G}$ is singular (its tangent complex has $H^{-1} \neq 0$ ). The structure sheaf $\mathcal O_E$ as an object of $\mathrm{QCoh}$ "sees" infinite-dimensional Ext groups (from the stacky automorphisms and the derived self-intersection), which on the automorphic side would demand eigensheaves that do not exist as D-modules of finite type. Imposing nilpotent singular support cuts the spectral category down to exactly those objects whose self-Exts are controlled by the global nilpotent cone $\mathcal N$ — matching the automorphic side. Thus $\mathrm{QCoh}$ is literally too big, and the failure is detected at reducible/non-elliptic local systems.

connection · Neighbouring structures

Geometric Satake Equivalence supplies the Hecke operators $[2]$ – $[3]$ and defines the dual group $\check G$ geometrically; it is the "local, unramified" heart of the global story.
Hitchin System and Hitchin Fibration provide the classical limit: $T^*\mathrm{Bun}_G$ and $T^*\mathrm{Bun}_{\check G}$ are dual integrable systems whose fibers are dual abelian varieties (SYZ/Fourier–Mukai), realizing classical geometric Langlands as fiberwise duality.
Kac–Moody Algebras at Critical Level and the Feigin–Frenkel isomorphism identify the center of the critical-level enveloping algebra with functions on $\mathrm{Opers}_{\check G}$ , building eigensheaves via localization (Beilinson–Drinfeld).
<a class="wiki-link" href="/wiki/conformal-field-theory">Conformal Field Theory</a>: conformal blocks of the WZW model and chiral algebras provide the factorization structure underlying the gluing in the proof.
<a class="wiki-link" href="/wiki/topological-field-theory">Topological Field Theory</a> ( $S$ -duality of 4d $\mathcal N=4$ ): the Kapustin–Witten realization.
<a class="wiki-link" href="/wiki/quantum-groups">Quantum Groups</a> and the Kazhdan–Lusztig equivalence govern the quantum/ramified deformations.

🔬 Where the machinery is used

⏳ History

- 1967–70 — Langlands. Robert Langlands formulates the original conjectures relating automorphic representations and Galois representations; the dual group is introduced.

- 1970s — Grothendieck. The sheaf–function dictionary over finite fields makes "functions = traces of Frobenius on sheaves" precise, enabling geometrization.

- 1980s — Drinfeld. Vladimir Drinfeld proves geometric Langlands for

\mathrm{GL}_2

over function fields and, with the theory of shtukas, the arithmetic

\mathrm{GL}_2

case; introduces key moduli ideas.

- 1987 — Laumon. Gérard Laumon formulates the geometric correspondence for

\mathrm{GL}_n

via automorphic sheaves and the Whittaker/Laumon sheaf; geometric class field theory (

\mathrm{GL}_1

) is established.

- 1991–96 — Beilinson–Drinfeld. Quantization of Hitchin Hamiltonians and the opers/critical-level construction produce Hecke eigensheaves for the "tame oper" local systems; the theory of chiral/factorization algebras is created.

- 2002 — Frenkel–Gaitsgory–Vilonen. Proof of the geometric Langlands existence theorem for

\mathrm{GL}_n

(vanishing conjecture).

- 2006 — Kapustin–Witten. Physical derivation via

S

-duality and Hitchin mirror symmetry; the quantum geometric Langlands family is articulated.

- 2015 — Arinkin–Gaitsgory. Singular support of coherent sheaves and the geometric Langlands conjecture: the precise categorical statement with

\mathrm{IndCoh}_{\mathcal N}

and the global nilpotent cone.

- 2024 — Gaitsgory–Raskin and collaborators (Arinkin, Beraldo, Chen, Faergeman, Lin, Rozenblyum and others). Announcement and posting of a multi-paper proof of the de Rham geometric Langlands conjecture for general reductive $G$ .

open-problem · What remains and what is hardest

Quantum geometric Langlands. For a nonzero level $\Psi$ , the conjectural equivalence $\mathrm{D\text{-}mod}_\Psi(\mathrm{Bun}_G) \simeq \mathrm{D\text{-}mod}_{-1/\Psi}(\mathrm{Bun}_{\check G})$ (Gaitsgory–Lurie, Stoyanovsky) is far less complete; the classical case is a degeneration.
Betti geometric Langlands. The Ben-Zvi–Nadler topological version using constructible sheaves and the Betti stack of local systems has a precise conjecture but a different (and partly open) proof landscape.
Arithmetic/étale function-field categorical Langlands. Upgrading V. Lafforgue's spectral decomposition to a full categorical equivalence over $\mathbb F_q$ remains active (related to Fargues–Scholze local theory).
Ramification. Wild and tame ramification, parahoric structures, and the precise spectral side for irregular singular connections are only partially understood.
Functoriality. Geometrizing Langlands functoriality (transfer between groups) beyond endoscopy is wide open.
Compatibility with the trace formula and analytic refinements (the "relative" and "spectral" geometric Langlands) are emerging frontiers.

Formal Statement

Why It Holds / Derivation Sketch

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits