Geometric Langlands Duality can be formulated as follows: Let C be a Riemann surface (compact, without boundary), G be a compact reductive Lie group, G_{C} be its complexification, and M_{flat}(G, C) be the moduli space of stable flat G_{C}-connections on C. The Langlands dual of G is another compact reductive Lie group ^{L}G defined by the condition that its weight and coweight lattices are exchanged relative to G. Let Bun(^{L}G, C) be the moduli stack of holomorphic ^{L}G-bundles on C. One of the statements of Geometric Langlands Duality is that the derived category of coherent sheaves on M_{flat}(G, C) is equivalent to the derived category of D-modules over Bun(^{L}G, C).

M_{flat}(G, C) is mirror to another moduli space which, roughly speaking, can be described as the cotangent bundle to Bun(^{L}G, C). The category of A-branes on T ^{∗} Bun(^{L}G, C) (with the canonical symplectic form) is equivalent to the category of B-branes on a noncommutative deformation of T ^{∗} Bun(^{L}G, C). The latter is the same as the category of (analytic) D-modules on Bun(^{L}G, C).

So, what exactly is, the relationship between A-branes and noncommutative B-branes. This relationship arises whenever the target space X is the total space of the cotangent bundle to a complex manifold Y. It is understood that the symplectic form ω is proportional to the canonical symplectic form on T ^{∗} Y. With the B-field vanishing, and Y as a complex, we regard ω as the real part of a holomorphic symplectic form Ω. If q^{i} are holomorphic coordinates on Y, and p_{i} are dual coordinates on the fibers of T ^{∗} Y, Ω can be written as

Ω = 1/ħdp_{i} ∧ dq^{i} = dΘ

Since ω (as well as Ω) is exact, the closed A-model of X is rather trivial: there are no nontrivial instantons, and the quantum cohomology ring is isomorphic to the classical one.

We would like to understand the category of A-branes on X = T ^{∗} Y. The key observation is that ∃ a natural coisotropic A-brane on X well-defined up to tensoring with a flat line bundle on X. Its curvature 2-form is exact and given by

F = Im Ω

If we denote by I the natural almost complex structure on X coming from the complex structure on Y , we have F = ωI, and therefore the endomorphism ω^{−1}F = I squares to −1. Therefore any unitary connection on a trivial line bundle over X whose curvature is F defines a coisotropic A-brane.

Now, what about the endomorphisms of the canonical coisotropic A-brane, i.e., the algebra of BRST-closed open string vertex operators? This is easy if Y is an affine space. If one covers Y with charts each of which is an open subset of C^{n}, and then argues that the computation can be performed locally on each chart and the results “glued together”, one gets closer to the fact that the algebra in question is the cohomology of a certain sheaf of algebras, whose local structure is the same as for Y = C^{n}. In general, the path integral defining the correlators of vertex operators does not have any locality properties in the target space. Each term in perturbation theory depends only on the infinitesimal neighbourhood of a point. This shows that the algebra of open-string vertex operators, regarded as a formal power series in ħ, is the cohomology of a sheaf of algebras, which is locally isomorphic to a similar sheaf for X = C^{n} × C^{n}.

Let us apply these observations to the canonical coisotropic A-brane on X = T ^{∗} Y. Locally, we can identify Y with a region in C^{n} by means of holomorphic coordinate functions q_{1}, . . . , q_{n}. Up to * BRST-exact terms*, the action of the A-model on a disc Σ takes the form

S = 1/ħ ∫_{∂Σ} φ ^{∗} (p_{i}dq^{i})

where φ is a map from Σ to X. This action is identical to the action of a particle on Y with zero Hamiltonian, except that q^{i} are holomorphic coordinates on Y rather than ordinary coordinates. The BRST-invariant open-string vertex operators can be taken to be holomorphic functions of p, q. Therefore quantization is locally straightforward and gives a noncommutative deformation of the algebra of holomorphic functions on T ^{∗} Y corresponding to a holomorphic Poisson bivector

P = ħ∂/∂p_{i} ∧ ∂/∂q^{i}

One can write an explicit formula for the deformed product:

(f ⋆ g)(p, q) = exp(ħ/2(∂^{2}/∂p_{i}∂q̃^{i} − ∂^{2}/∂q^{i∂p̃i })) f(p, q) g (p̃, q̃)|_{p̃ = p, q̃ = q}

This product is known as the Moyal-Wigner product, which is a formal power series in ħ that may have zero radius of convergence. To rectify the situation, one can restrict to functions which are polynomial in the fiber coordinates p_{i}. Such locally-defined functions on T ^{∗} Y can be thought of as symbols of differential operators; the Moyal-Wigner product in this case reduces to the product of symbols and is a polynomial in ħ. Thus locally the sheaf of open-string vertex operators is modelled on the sheaf of holomorphic differential operators on Y (provided we restrict to operators polynomial in p_{i}).

Locally, there is no difference between the sheaf of holomorphic differential operators D(Y ) and the sheaf of holomorphic differential operatorsD(Y, L) on a holomorphic line bundle L over Y. Thus the sheaf of open-string vertex operators could be any of the sheaves D(Y, L). Moreover, the classical problem is symmetric under p_{i} → −p_{i} combined with the orientation reversal of Σ; if we require that quantization preserve this symmetry, then the algebra of open-string vertex operators must be isomorphic to its opposite algebra. It is known that the opposite of the sheaf D(Y, L) is the sheaf D(Y, L^{−1} ⊗ K_{Y}), so symmetry under p_{i} → −p_{i} requires L to be a square root of the canonical line bundle K_{Y}. It does not matter which square root one takes, since they all differ by flat line bundles on Y, and tensoring L by a flat line bundle does not affect the sheaf D(Y, L). The conclusion is that the sheaf of open-string vertex operators for the canonical coisotropic A-brane α on X = T ^{∗} Y is isomorphic to the sheaf of noncommutative algebras D(Y, K^{1/2}). One can use this fact to associate Y to any A-brane β on X a twisted D-module, i.e., a sheaf of modules over D(Y, K^{1/2}). Consider the A-model with target X on a strip Σ = I × R, where I is a unit interval, and impose boundary conditions corresponding to branes α and β on the two boundaries of Σ. Upon quantization of this model, one gets a sheaf on vector spaces on Y which is a module over the sheaf of open-string vertex operators inserted at the α boundary. A simple example is to take β to be the zero section of T ^{∗} Y with a trivial line bundle. Then the corresponding sheaf is simply the sheaf of sections of K_{Y}^{1/2}, with a tautological action of D(Y, K_{Y}^{1/2}).

One can argue that the map from A-branes to (complexes of) D-modules can be extended to an equivalence of categories of A-branes on X and the derived category of D-modules on Y. The argument relies on the conjectural existence of the category of generalized complex branes for any generalized Calabi-Yau. One can regard the Geometric Langlands Duality as a nonabelian generalization.