Geometric Representation Theory | Flag Varieties, Borel–Weil and Beilinson–Bernstein

Geometric representation theory constructs and studies representations using geometric objects on spaces with symmetry. A representation can arise from sections of a line bundle, from cohomology, or from a sheaf on which differential operators act. Geometry supplies the object; taking sections or cohomology returns a linear representation.

The easiest complete example is the projective line. A line bundle of degree n has n+1 global sections when n is nonnegative. Those sections form an irreducible SL₂ representation. The same example also exposes the limits: a nonzero line bundle can have no nonzero global sections, so taking sections does not always preserve all information. Beilinson–Bernstein localization succeeds under much more specific conditions than the slogan “turn geometry into algebra” suggests.

This guide follows one route in depth: flags, line bundles, polynomial sections, Lie algebra operators, higher cohomology and localization. It includes a precise untwisted localization statement and an explicit failed-twist example. The general theorems are sourced, while the projective-line calculations are developed step by step so that the geometry can be checked against ordinary algebra.

Scope: complex algebraic geometry and finite-dimensional complex semisimple Lie algebras. The initial examples require only linear algebra and polynomial differentiation. The later sections introduce sheaves and D-modules rather than assuming they are familiar. For algebraic prerequisites, start with Lie Representations and Highest Weights.

Flags · Projective-line sections · Borel–Weil · Cohomology · Localization · Practice · Worked answers · BTT Mathematics Hub

Why moving to geometry can reveal algebra

A list of matrices describes how a representation acts in one basis. It may conceal why the representation exists, why a certain dimension occurs or why two modules share structural features. A geometric construction can answer those questions by replacing arbitrary matrix choices with a space, a bundle and a natural group action.

For example, the n+1 in the SL₂ irreducible of highest weight n can be seen by counting homogeneous monomials xⁿ,xⁿ⁻¹y,…,yⁿ. Geometry explains why precisely those polynomials are the sections of a line bundle on a compact projective space. Differentiating the group action then produces the familiar raising and lowering operators.

The benefit is not merely a second notation. The geometric model adds notions such as support, local behaviour, gluing, cohomological degree and singularity. These can distinguish objects whose simple numerical summaries look the same.

A flag is a nested sequence of subspaces

A complete flag in Cⁿ is a chain 0⊂V₁⊂V₂⊂···⊂Vₙ=Cⁿ with dim Vᵢ=i. It records a line, a plane containing that line, and successively larger containing subspaces. It does not record a preferred basis inside each subspace.

An ordered basis determines a flag by taking the spans of its first one, first two and first three vectors, and so on. Many bases determine the same flag. Multiplying an adapted basis matrix by an invertible upper-triangular matrix changes those vectors without changing the sequence of spans.

The collection of flags is a flag variety. For G=SLₙ(C) and B its subgroup of upper-triangular matrices, it is G/B. The group acts transitively, and B is the stabiliser of the standard coordinate flag. Source [1] develops this homogeneous-space description.

The smallest flag variety is the projective line

In C², a complete flag consists only of a line L⊂C². Thus SL₂(C)/B is the projective line P¹. A point [x:y] represents the line through the nonzero vector (x,y), and [x:y]=[cx:cy] for every nonzero c.

Use two affine charts. On U₀, where x≠0, set z=y/x. On U₁, where y≠0, set w=x/y. Their overlap has w=1/z. A single z-chart misses the point [0:1], so a formula regular for all finite z may still misbehave at infinity.

This missing point explains why arbitrary polynomials in z do not all become global functions on P¹. A global function must also be regular in the w-chart. Replacing z by 1/w shows that every positive power creates a pole at w=0. Only constants survive.

Count the dimension of a flag variety

The complex dimension of the complete flag variety in Cⁿ is n(n−1)/2. One way to count is to subtract the dimension of the upper-triangular stabiliser from the dimension of the general linear group: n²−n(n+1)/2=n(n−1)/2. Passing to determinant-one matrices reduces both dimensions by one and leaves the difference unchanged.

For C³, a flag is a line inside a plane, and the flag variety has complex dimension 3. A Grassmannian, by contrast, remembers only a k-dimensional subspace and has dimension k(n−k). These spaces answer different geometric questions even though both are homogeneous quotients.

Dimension here belongs to the geometric base space. It is not the dimension of a representation obtained from a bundle over that space. P¹ has complex dimension 1 while its line bundles can produce representations of arbitrarily large finite dimension.

A line bundle is not the same as a line in one vector space

A line bundle assigns a one-dimensional vector space to every point of a base space, with compatible local trivialisations. A section chooses a vector in each fibre in a way satisfying the relevant regularity and gluing conditions.

The fibre dimension is always one, but the vector space of global sections can have dimension zero, one or much larger. A bundle with three global sections has not suddenly acquired three-dimensional fibres.

On P¹, write O(n) for the standard line bundle of degree n. Positive and negative n give genuinely different gluing. The degree is part of the object, not a label that can be changed without affecting its sections.

Derive the global sections of O(n) from gluing

A regular section of O(n) can be described by polynomials p₀(z) on U₀ and p₁(w) on U₁ satisfying p₁(w)=wⁿp₀(1/w) on the overlap. This fixes the transition convention used throughout the article.

Write p₀(z)=a₀+a₁z+···+aᵣzʳ. Substitution gives p₁(w)=a₀wⁿ+a₁wⁿ⁻¹+···+aᵣwⁿ⁻ʳ. Regularity at w=0 requires all nonzero terms to have nonnegative exponents.

If n≥0, this is exactly the condition r≤n. Thus H⁰(P¹,O(n)) has basis 1,z,…,zⁿ in this chart and dimension n+1. If n<0, every possible term has a negative w-exponent, so the only global section is zero.

The calculation is a concrete instance of the projective-space cohomology formulas in source [2]. No integration or differential equation was needed: global regularity across two charts imposed a finite-dimensional constraint.

Work O(2) and test a false extra section

For O(2), take p₀(z)=a+bz+cz². The other chart has p₁(w)=aw²+bw+c, which is regular. There are three freely chosen coefficients, agreeing with the expected three-dimensional representation.

Now try p₀(z)=z³. The transition gives p₁(w)=w²(1/w³)=1/w, which is not regular at w=0. The candidate is a valid local expression on U₀ but not a global section.

This is a useful diagnostic exercise: a local formula can be correct on its own chart and still fail to define the claimed global object. The problem is not bad algebra. It is a missing gluing or boundary condition.

Homogeneous polynomials give the same sections

Multiplying p₀(y/x) by xⁿ gives a homogeneous polynomial a₀xⁿ+a₁xⁿ⁻¹y+···+aₙyⁿ. Conversely, every degree-n homogeneous polynomial gives the chart expression p₀(z) by setting x=1,y=z.

The global section space is therefore Symⁿ((C²)*) under the natural action on polynomial functions. SL₂ has an invariant nondegenerate alternating form, which identifies its standard representation with its dual. Consequently this section space is also isomorphic to the usual highest-weight-n representation.

The dual must be recorded in higher rank, where a representation and its dual need not be isomorphic. Ignoring it because the first SL₂ example hides the difference is a common convention error.

Borel–Weil generalises the polynomial construction

Let G be simply connected and complex semisimple, with a chosen Borel subgroup B. For a dominant integral weight λ, choose the ample-sign line bundle ℒ_λ on G/B. In our convention, H⁰(G/B,ℒ_λ) is the dual of the irreducible highest-weight representation V_λ. For SL₂, ℒ_n=O(n). This is the Borel–Weil theorem. [1]

References may attach the opposite sign to the bundle label. The MIT lecture in source [1] writes its bundle so that the dominant representation is obtained from the negative of that label. Our notation places positive degree directly on O(n). Compare the actual line bundle and action before comparing the printed symbol λ.

The theorem constructs representations from geometry. It is stronger than noticing that a vector space happens to have the correct dimension, because the group action on sections is part of the identification.

Differentiate the symmetry into explicit operators

After making the standard SL₂ self-duality identification just described, a convenient model on polynomials of degree at most n uses

e = d/dz
h = n − 2z d/dz
f = nz − z² d/dz.

Thus ezʳ=rzʳ⁻¹, hzʳ=(n−2r)zʳ and fzʳ=(n−r)zʳ⁺¹. The constant polynomial is a highest-weight vector of weight n, and the lowering action stops at zⁿ because its coefficient n−r has become zero.

These formulas can also be obtained from homogeneous polynomials: use e=x∂/∂y, f=y∂/∂x and h=x∂/∂x−y∂/∂y, then write the homogeneous polynomial as xⁿp(y/x). The chart expressions are the result of an actual transformation, not an unrelated differential-operator guess.

Verify the commutator instead of trusting the notation

Apply [e,f] to zʳ. First efzʳ=(n−r)(r+1)zʳ. Next fezʳ=r(n−r+1)zʳ. Their difference is (n−2r)zʳ, exactly hzʳ. Similar calculations give [h,e]=2e and [h,f]=−2f.

For n=2, f1=2z, fz=z² and fz²=0. The raising action sends z² to 2z and z to 1. Every nonzero weight vector can reach the top and every weight can be generated from it, so this three-dimensional representation is irreducible.

The coefficients differ from the Verma basis vᵣ=fʳv₀ because the polynomial basis zʳ has a different normalisation. Comparing coefficients term by term without accounting for that basis change would produce a false contradiction.

Sheaves keep track of local sections and restrictions

A sheaf assigns data to each open set, with restriction maps to smaller open sets and a rule for gluing compatible local data. The sheaf of sections of O(n) is one example. The section space H⁰ is the data assigned to the whole base space.

A local section and a global section are therefore different requests to the same sheaf. The polynomial z³ in the O(2) example is available locally, but it cannot be glued across infinity. A sheaf does not erase that local information merely because the global section space is smaller.

This is why geometric representation theory works with sheaves rather than only with their global vectors. Local support and gluing constraints may be essential to the representation-theoretic answer.

Higher cohomology measures information missed by global sections

For line bundles on P¹, the cohomology dimensions are h⁰(O(n))=max(n+1,0) and h¹(O(n))=max(−n−1,0). There is no higher cohomology. Source [2] provides the general projective-space result; a two-chart calculation explains these particular numbers.

On the overlap, use Laurent polynomials in z. The first chart supplies ordinary powers z⁰,z¹,z²,… . In the same trivialisation, the second chart supplies zⁿ times polynomials in z⁻¹, hence powers with exponent at most n.

The first cohomology is represented by the quotient C[z,z⁻¹]/(C[z]+zⁿC[z⁻¹]). The only exponents not covered are n+1,n+2,…,−1 when n≤−2. There are −n−1 of them. For n≥−1, no gap remains.

Three negative-degree examples

O(−1) has no global sections, and its Laurent quotient has no missing exponents. Thus both H⁰ and H¹ vanish even though the line bundle itself is nonzero.

O(−2) has H⁰=0 and one cohomology class represented by z⁻¹. Therefore H¹ has dimension 1. For O(−4), the missing powers are z⁻³,z⁻²,z⁻¹, so H¹ has dimension 3.

The Euler characteristic h⁰−h¹ equals n+1 in every case. For O(−4), it is −3, not a negative vector-space dimension. Euler characteristic is an alternating sum of dimensions, so it can be negative without inconsistency.

Borel–Weil–Bott explains the shifted pattern

The Borel–Weil–Bott theorem extends the global-section construction to higher sheaf cohomology and organises the answer using the shifted Weyl action. For an integral weight with regular shifted position, one cohomological degree supplies the relevant irreducible representation; a singular shifted weight gives vanishing. A general proof is linked in source [5].

In the SL₂ coordinate, the shift is n↦n+1. For n≥0, the weight is already on the dominant side and H⁰ gives highest weight n. At n=−1 the shifted weight is zero, explaining complete vanishing. For n≤−2, reflect and shift back to get −n−2; the representation occurs in H¹ and has dimension −n−1.

Our Laurent calculation therefore reproduces the numerical rank-one pattern without requiring the full theorem as a black box. The theorem adds the general group action and higher-rank organisation.

An exact sequence shows why taking sections can fail

On P¹, the homogeneous coordinates x and y have no common zero. They produce an exact sequence 0→O(−2)→O(−1)⊕O(−1)→O→0. The last map is multiplication by x and y, and the first records their relation using the column (−y,x).

Take global sections. The first two sheaves have zero H⁰, while H⁰(O)=C. The resulting map 0→C cannot be surjective. Thus the global-sections functor is not generally exact on coherent sheaves on P¹.

The missing surjectivity is recorded by H¹(O(−2))=C in the long exact sequence. Cohomology is not an optional decoration here: it records the failure of a simpler operation to preserve the exact relationship among geometric objects.

Differential operators add another kind of local structure

On the affine line with coordinate z, multiplication by z and differentiation ∂ satisfy ∂z−z∂=1. Their algebra is the first Weyl algebra. It is noncommutative because differentiating a product adds the derivative of the coordinate.

A D-module is, roughly, a sheaf on which local differential operators act compatibly. On a smooth variety, the relevant sheaf of algebras is denoted D_X. In this article we mean left D-modules whose underlying O_X-modules are quasi-coherent, unless a restriction is stated. Source [3] introduces this setting.

The letter D refers to differential operators, not a numerical derivative assigned independently at every point. A D-module includes algebraic compatibility among multiplication, derivatives, restriction and coordinate change.

A local differential equation becomes a module

To encode the equation ∂u=au with constant a, form the left module D/D(∂−a) on the affine line. A D-linear map from this quotient to a suitable function sheaf is determined by the image of the class of 1, and that image must satisfy the stated differential equation.

This distinguishes the equation module from its solution space. The module stores the operator relation. Solutions depend on the sheaf into which we map, for example algebraic, holomorphic or another analytically specified class of functions.

Replacing a D-module by one list of numerical solutions would discard the local operator structure. Geometric representation theory uses the module itself because it behaves well under categorical operations and carries more information than one chosen solution realisation.

The Lie algebra acts by global differential operators

A group action on a smooth variety differentiates to vector fields. This gives a Lie algebra map 𝔤→Γ(X,D_X) and hence an algebra map from U(𝔤). On the flag variety, central elements impose relations on this action.

For P¹ in the untwisted case, take n=0 in the earlier formulas: e=∂, h=−2z∂ and f=−z²∂. Although these are written on one chart, they extend as global vector fields with the proper coordinate transformations.

The operator C=h²+2h+4fe acts as zero. This can be checked on an arbitrary polynomial zʳ using the displayed formulas. The central relation is therefore visible in differential operators before any finite-dimensional representation is selected.

A precise untwisted Beilinson–Bernstein statement

Let G be connected, simply connected and complex semisimple, X=G/B, and χ₀ the central character of the trivial 𝔤-representation. Put A=U(𝔤)/U(𝔤)ker χ₀. In the untwisted setting, Γ(X,D_X) is isomorphic to A, and global sections give an equivalence between quasi-coherent left D_X-modules and left A-modules. Localization is an inverse functor. These are the untwisted results stated in source [4].

This is an equivalence of categories, not merely a dimension match. Morphisms are transported as well as objects. A module can be recovered after moving to a sheaf, and a D-module can be recovered after taking its global sections.

The theorem does not say that all sheaves on every projective variety can be recovered from sections. It applies to a specified category of D-modules on a flag variety and a specified central reduction of the enveloping algebra. Removing those specifications changes the claim.

What localization actually does

Given an A-module M, localization forms D_X⊗_AM, interpreted as a sheaf with local differential-operator action. In the other direction, Γ takes a D-module to its global section space with the action of Γ(D_X)=A.

The two constructions are naturally adjoint even in broader settings. The theorem supplies the stronger conclusion that their unit and counit maps are isomorphisms in the permitted situation. An adjunction alone would not guarantee this.

This distinction connects to induction and restriction elsewhere in representation theory. Two operations can be paired by a mapping property without undoing one another. A categorical equivalence requires an additional theorem, not simply two arrows drawn in opposite directions.

Twisted differential operators require a convention ledger

Differential operators may act on local sections of a line bundle rather than on functions. On P¹, let D⁽ⁿ⁾ mean the sheaf of differential operators on O(n). The earlier e,h,f formulas then have the parameter n, and the Casimir acts by n(n+2).

For integers n≥0 in this O(n) convention, the twisted localization theorem gives the corresponding equivalence. General weights require the correctly shifted dominance and regularity conditions. Source [4] uses the opposite sign for its bundle index, so its displayed antidominant condition translates accordingly.

A safe comparison records four choices: positive roots, the character defining the associated line bundle, whether modules act on the left or right, and the shift used to label the central character. The words dominant and antidominant alone do not settle a disagreement between two references.

A nonzero object with zero sections disproves an overbroad claim

O(−1) is a nonzero left D⁽⁻¹⁾-module: its own local differential operators act on its sections. But Γ(P¹,O(−1))=0. A functor that sends this nonzero object to zero cannot be faithful, so it cannot be an equivalence by taking global sections in this twist.

The same observation works for negative-degree line bundles in their corresponding differential-operator categories. It does not contradict the untwisted theorem because O(−1) with this action belongs to a different twisted D-module category.

This is a particularly useful theorem-boundary test. Instead of accepting or rejecting a very broad slogan, identify one concrete object, compute what the proposed functor does to it, and compare the result with a necessary property of an equivalence.

Symbols connect differential operators to cotangent geometry

Differential operators are filtered by order. Their leading symbols commute and live on the cotangent bundle. On an affine coordinate chart, the derivative ∂ is replaced at the symbol level by a cotangent coordinate ξ, turning a noncommutative operator algebra into a commutative polynomial description.

The enveloping algebra has its own degree filtration, whose associated graded algebra is Sym(𝔤). Thus filtered maps from enveloping algebras to differential operators have a geometric shadow relating cotangent spaces and Lie algebra duals. Sources [3] and [4] explain this relationship in the flag-variety setting.

Passing to symbols intentionally forgets lower-order terms and commutators. The classical geometry helps organise the operator problem, but it is not the full operator algebra. Recovering the differential structure requires putting those terms back.

The SL₂ Springer picture can be written with two-by-two matrices

The cotangent bundle of P¹ can be described by pairs (N,L), where L is a line in C² and N satisfies im N⊆L⊆ker N. Such an N is nilpotent. The map (N,L)↦N forgets the line and lands in the nilpotent cone of sl₂.

A trace-zero matrix N=[[a,b],[c,−a]] is nilpotent precisely when a²+bc=0. For a nonzero nilpotent two-by-two matrix, image and kernel are the same line. That line uniquely determines L, so the map has one point over a nonzero nilpotent.

For N=0, every line satisfies the condition. The fibre is all of P¹. The additional flag therefore resolves the singular behaviour at zero rather than giving a one-to-one correspondence everywhere. This rank-one calculation illustrates the Springer-resolution geometry discussed in source [1].

Schubert cells organise the flag variety by permutation data

The flag variety has a decomposition into Schubert cells indexed by the Weyl group. In type A this is the symmetric group. A cell has complex dimension equal to the length of its indexing permutation, under the usual choice of cells. [1]

For flags in C³, the six permutations have lengths 0,1,1,2,2,3. The resulting ordinary cohomology has Poincaré polynomial 1+2t²+2t⁴+t⁶ and Euler characteristic 6. Each complex cell contributes in an even real degree.

This ordinary topological cohomology is not the same as the sheaf cohomology Hⁱ(O(n)) computed earlier. Both use the letter H, but their coefficient objects and meanings differ. Naming the cohomology theory is part of a correct calculation.

What the geometry adds to representation theory

Line bundles explain the existence of finite-dimensional highest-weight representations. D-modules encode differential equations and infinitesimal symmetry. Cohomology records failures of global sections. Cotangent geometry exposes the leading-order structure of filtered algebras, while Schubert strata supply finite combinatorial organisation.

These are distinct contributions. Borel–Weil is not Beilinson–Bernstein, a line bundle is not a D-module without a specified action, and a cell count is not automatically a representation multiplicity. A useful geometric argument identifies exactly which construction converts the geometric object into the claimed algebraic one.

The same discipline applies when moving toward category O, intersection cohomology or geometric Langlands. The relevant subcategory, equivariance and parameter conventions must be stated. One successful projective-line example motivates those theories; it does not prove all of them.

Practice questions

1. What does a complete flag in C³ specify? 2. Compute the complex dimension of the complete flag variety in C⁴. 3. Determine h⁰(O(4)). 4. Test whether z⁴ defines a global section of O(3).

5. For n=3, compute e(z²), h(z²) and f(z²). 6. Verify [e,f] on z² in that model. 7. Find h⁰ and h¹ of O(−3), and compute the Euler characteristic. 8. List the Laurent powers representing H¹(O(−5)).

9. Why does O(−1) show that global sections are not faithful for every twist? 10. For N=[[0,1],[0,0]], find the line in its Springer fibre. 11. What is the fibre over N=0? 12. Explain why an adjunction between localization and global sections is weaker than an equivalence.

Worked answers

1–4. Flags and global sections

1. A complete flag in C³ consists of a one-dimensional line inside a two-dimensional plane inside C³. The line and plane must be nested. Choosing a line and an unrelated plane does not produce a flag.

2. The dimension is 4·3/2=6. Equivalently, subtract the ten-dimensional upper-triangular stabiliser from the sixteen-dimensional GL₄. This is a complex geometric dimension, not the number of flags or the dimension of every section representation.

3. The global sections of O(4) are polynomials p₀(z) of degree at most 4, with basis 1,z,z²,z³,z⁴. Thus h⁰=5. In homogeneous coordinates the same basis is x⁴,x³y,x²y²,xy³,y⁴.

4. The transition gives p₁(w)=w³(1/w⁴)=w⁻¹. It has a pole at w=0, so z⁴ is not a global O(3) section. It is only a section on the affine chart U₀.

5–8. Operators and cohomology

5. With n=3, e(z²)=2z, h(z²)=(3−4)z²=−z², and f(z²)=(3−2)z³=z³. The output weights differ by the prescribed raising and lowering increments.

6. ef(z²)=e(z³)=3z², while fe(z²)=f(2z)=4z². Their difference is −z², equal to h(z²). This directly verifies the commutator on the selected vector.

7. O(−3) has h⁰=0 and h¹=2. The Euler characteristic is 0−2=−2, which agrees with n+1. Its negative value does not mean either cohomology space has negative dimension.

8. For n=−5, the first chart supplies exponents at least zero and the second supplies exponents at most −5. The missing exponents are −4,−3,−2,−1. Their Laurent monomials represent a four-dimensional H¹.

9–12. Theorem boundaries and reconstruction

9. The line bundle O(−1) is nonzero and is naturally a module for its own twisted differential operators, but its global section space is zero. Its identity morphism therefore maps to the zero map. A faithful functor cannot do this, so global sections cannot be an equivalence in that twist.

10. The displayed N sends (x,y) to (y,0). Both its image and its kernel are span{(1,0)}. Therefore the only permitted line L is the first coordinate line.

11. For N=0, its image is zero and its kernel is all of C². Every line lies between them, so the fibre is P¹. This exceptional fibre is why the map is a resolution rather than an everywhere one-to-one parametrisation.

12. An adjunction identifies appropriate Hom spaces. An equivalence additionally requires the unit and counit maps to be isomorphisms, so that composing the functors recovers each object. The Beilinson–Bernstein theorem supplies that stronger conclusion in a specified setting.

Teach geometry through a complete return calculation

Begin with the two projective charts, not with a list of theorem names. Ask the learner to turn p₀(z)=1+2z+3z² into its O(2) expression on the second chart. Then replace it with z³ and identify the precise failure at infinity.

Next differentiate the polynomial action and verify one commutator. The learner can now see a representation emerge from a geometric section condition. Compare n=2 with n=−2 to make clear that a nonzero bundle need not have global sections.

Finally present the localization theorem with its category and central character attached. Ask which earlier counterexample would invalidate an unrestricted version. This teaching sequence rewards both construction and boundary checking: the learner should be able to explain why the theorem is powerful and why its hypotheses cannot be omitted.

Sources and further study

[1] Pavel Etingof, MIT 18.757, Geometry of Complex Semisimple Lie Groups, for flag varieties, Borel–Weil and the Springer geometry. [2] The Stacks Project, Cohomology of Projective Space, for line-bundle cohomology. [3] Etingof, D-Modules, Part I. [4] Etingof, The Beilinson–Bernstein Localization Theorem, including the precise untwisted theorem and the effect of twisting. [5] Man Shun John Ma, A Proof of the Borel–Weil–Bott Theorem. Bundle and dual conventions must be translated when moving between references.

For prerequisites, use Lie Representations and Tensor Products, Duality and Symmetric Powers. Return to the BTT Mathematics Learning Hub for the wider route.