Category O | Verma Modules, Highest Weights and BGG Reciprocity

A highest weight can generate an infinite representation even when the irreducible representation eventually obtained from it is finite-dimensional. Category O provides a setting in which that larger representation, its submodules, its simple quotient and its projective cover can be studied together.

The decisive example is already present in sl₂. Start with a highest-weight vector of weight 2 and repeatedly apply a lowering operator. The resulting Verma module does not stop after three vectors. Instead, a new highest-weight vector appears farther down the same infinite chain. Dividing out the submodule that it generates leaves the familiar three-dimensional irreducible representation. Confusing that quotient with the original module hides the structure this subject is designed to reveal.

This extended guide develops the example completely before introducing the general organisation. Its central questions are practical: how do we recognise a singular vector, when is a Verma module simple, what does a formal character remember, and how does BGG reciprocity turn composition multiplicities into information about projectives? The general theorems are identified with sources; the rank-one calculations are worked explicitly.

Scope: finite-dimensional complex semisimple Lie algebras and their algebraic modules. These are advanced enrichment topics, not a claim about Singapore school examination content. Familiarity with vector spaces, kernels, quotients and the sl₂ commutation relations is helpful. The prerequisite route is Lie Group and Lie Algebra Representations.

Build a Verma module · Find singular vectors · Understand category O · Use BGG reciprocity · Practice questions · Worked answers · BTT Mathematics Hub

First separate four objects that often get merged

A highest-weight vector is one vector. A highest-weight module is the entire module generated by such a vector. A Verma module is the universal module obtained by imposing only the highest-weight relations. A simple highest-weight module is the irreducible quotient of that Verma module. These are related objects, but they are not synonyms.

The word universal means that any other module generated by a highest-weight vector with the same weight receives a surjective map from the Verma module. Further equations, including those forcing a finite-dimensional representation, are introduced by taking a quotient. The source therefore contains the possible relations rather than assuming them away at the beginning. This construction and its universal property are established in the highest-weight theory referenced in source [1].

For notation, write M(λ) for the Verma module and L(λ) for its simple quotient. Later P(λ) will denote a projective cover of L(λ) in the relevant category O block. The common weight label λ records their relationship; it does not make their dimensions, submodules or mapping properties identical.

Construct the sl₂ Verma module

Use generators e, f and h with [h,e]=2e, [h,f]=−2f and [e,f]=h. Here [a,b]=ab−ba. The operator e raises weights and f lowers them. Choose a complex number λ and a nonzero vector v₀ satisfying ev₀=0 and hv₀=λv₀.

Define vᵣ=fʳv₀ for every integer r≥0. In M(λ), these vectors form a basis and there is no final vᵣ. The actions are

f v_r = v_(r+1)
h v_r = (λ − 2r) v_r
e v_r = r(λ − r + 1) v_(r−1),  r ≥ 1
e v_0 = 0.

The basis statement follows from the Poincaré–Birkhoff–Witt theorem: after the highest-weight relations have been imposed, powers of the lowering generator remain independent. It is not justified by merely observing the first few vectors in a computation. Source [1] gives the general construction through the universal enveloping algebra.

Every vector is a finite linear combination of the vᵣ. The module is infinite-dimensional, but it is generated by the single vector v₀ under the algebra action. This distinction between finite generation and finite dimension will be important throughout the article.

Derive the lowering and raising coefficients

From hf=fh−2f, induction gives hfʳv₀=(λ−2r)fʳv₀. Thus each application of f changes the weight by −2. That establishes the h formula without guessing a pattern from a diagram.

For the e formula, start with efv₀=(fe+h)v₀=λv₀. Suppose efʳv₀=r(λ−r+1)fʳ⁻¹v₀. Then efʳ⁺¹v₀=(fe+h)fʳv₀. The first term contributes r(λ−r+1)fʳv₀ and the second contributes (λ−2r)fʳv₀. Their sum is (r+1)(λ−r)fʳv₀, which is the required formula at r+1.

The coefficient r(λ−r+1) therefore comes from the commutation relation. It determines exactly when the upward action vanishes and is the main diagnostic quantity in the rank-one theory. Keeping its two factors visible is more useful than repeatedly multiplying matrices of increasing size.

Check the representation law on a general basis vector

Write cᵣ=r(λ−r+1), with c₀=0. Then efvᵣ=cᵣ₊₁vᵣ and fevᵣ=cᵣvᵣ. Subtracting gives [e,f]vᵣ=(cᵣ₊₁−cᵣ)vᵣ. Expanding the difference produces λ−2r, exactly the eigenvalue of h on vᵣ.

Likewise, e moves from weight λ−2r to weight λ−2r+2, so [h,e]=2e. The operator f moves in the opposite direction, giving [h,f]=−2f. We have verified all three relations on an arbitrary basis vector, not just on v₀.

This illustrates a reliable calculation strategy. Derive the action from the defining relations, then substitute it back into those relations. A table of coefficients without this return check can conceal an index error that later corrupts every singular-vector calculation.

Singular vectors create new highest-weight submodules

A singular vector here is a nonzero weight vector killed by e. The vector v₀ is singular by construction. A further basis vector vᵣ with r≥1 is singular precisely when r(λ−r+1)=0. Over C, r is nonzero, so the condition is λ=r−1.

Consequently an additional singular vector exists exactly when λ is a nonnegative integer. Write λ=n. The new singular vector is vₙ₊₁, of weight n−2(n+1)=−n−2. Its descendants form a submodule isomorphic to M(−n−2).

For λ outside the nonnegative integers, every upward coefficient below v₀ is nonzero. Any nonzero submodule contains a weight vector; repeatedly applying e then reaches a nonzero multiple of v₀, so that submodule is the whole module. Thus M(λ) is simple in this case. The one-dimensional weight spaces make the rank-one argument particularly transparent.

Work the complete example λ = 2

The weights in M(2) are 2,0,−2,−4,−6 and so on. The first upward actions are ev₁=2v₀, ev₂=2v₁, ev₃=0 and ev₄=−4v₃. The zero at v₃ does not say that v₃ itself is zero. It says that v₃ is a new highest-weight vector inside the existing module.

The span of v₃,v₄,v₅,… is stable under e, f and h. Its highest weight is −4. Divide M(2) by this submodule. The quotient has basis classes v̄₀,v̄₁,v̄₂, and f v̄₂=0 because v₃ has become zero in the quotient.

The resulting L(2) has weights 2,0,−2 and dimension 3. All of its nonzero weight vectors connect to the highest-weight vector through e and to the other weights through f, so it is irreducible. The exact sequence is

0 → M(−4) → M(2) → L(2) → 0.

This one sequence contains an infinite-dimensional submodule, an infinite-dimensional middle module and a finite-dimensional quotient. Dimension language must therefore be attached to the correct object every time.

Why the finite quotient is not a finite submodule

The span of v₀,v₁,v₂ inside M(2) is not stable under f, because fv₂=v₃ lies outside it. It is not a copy of L(2) sitting inside the Verma module. The quotient exists because we deliberately identify the tail with zero.

The exact sequence above does not split. A splitting would place a highest-weight-2 copy of L(2) inside M(2). Its top vector would have to be a nonzero multiple of v₀, because that weight space is one-dimensional. Applying f three times would give a nonzero multiple of v₃, contradicting the relation f³=0 on the highest vector of L(2).

The argument distinguishes two operations that a truncated diagram may make look identical: cutting a picture off after three rows and forming a quotient by an invariant tail. Only the second is automatically a representation-theoretic construction here.

Three nearby weights give three different outcomes

For λ=0, the vector v₁ is singular, and 0→M(−2)→M(0)→L(0)→0 has a one-dimensional trivial quotient. For λ=−1, the upward coefficient is −r², never zero when r≥1, so M(−1) is simple and infinite-dimensional.

For λ=1/2, the coefficient vanishes only at r=3/2, which is not a permitted basis index. There is no additional singular vector and the Verma module is again simple. Solving the coefficient equation over C is not enough: the index r must be a positive integer.

The calculation also explains why “negative highest weight” is not a contradiction. Highest refers to the ordering induced by the raising direction, not to numerical positivity. A highest weight of −1 can sit above −3,−5,−7,… in an infinite module.

What category O requires

Choose a triangular decomposition 𝔤=𝔫₋⊕𝔥⊕𝔫₊ of a finite-dimensional complex semisimple Lie algebra. In the standard definition, category O consists of finitely generated U(𝔤)-modules that are semisimple over the Cartan subalgebra 𝔥 and locally finite under 𝔫₊. These conditions give finite-dimensional weight spaces and controlled upward behaviour. An equivalent bounded-weight formulation is used in source [2].

Being semisimple over 𝔥 means a module is a direct sum of simultaneous weight spaces. It does not mean that the entire 𝔤-module is semisimple. The raising directions can connect these weight spaces in nontrivial extensions.

Local finiteness means that applying the algebra generated by the raising operators to one vector spans a finite-dimensional space. In our sl₂ Verma example, every vector has finite support in the vᵣ basis, and sufficiently many applications of e kill it. The downward f-chain may nevertheless be infinite.

Why these restrictions make an infinite problem manageable

Finite generation prevents the category from admitting an arbitrary independent supply of new generators. Weight decomposition separates simultaneous eigenvalue data. Controlled raising makes highest-weight arguments possible. None of these restrictions alone says that the underlying vector space is finite.

A direct sum of infinitely many copies of the trivial representation has very simple individual summands, but it is not finitely generated as a U(𝔤)-module: finitely many chosen vectors span only a finite-dimensional trivial submodule. It therefore lies outside category O.

By contrast, a Verma module has infinitely many basis vectors but only one generator. This is why an intuitive test such as “there are infinitely many vectors, so it is excluded” is wrong. The category controls how vectors are generated and organised, not merely their total number.

General Verma modules are induced modules

Let 𝔟=𝔥⊕𝔫₊. A weight λ defines a one-dimensional 𝔟-module C_λ: the Cartan acts by λ and 𝔫₊ acts by zero. The associated Verma module is M(λ)=U(𝔤)⊗_U(𝔟)C_λ.

This is induction from a subalgebra. It is related to the subgroup induction discussed earlier in the series, but its carrier is an enveloping-algebra module and is generally infinite-dimensional. A subgroup index formula from finite-group induction is not the appropriate dimension formula.

The lowering part U(𝔫₋) supplies the remaining vector-space structure. In higher rank there are several lowering root directions, so a weight can be reached by several ordered monomials. Weight multiplicities can exceed one, and finding singular vectors can require solving linear equations rather than spotting one vanished scalar.

Formal characters count weight spaces

For a module with finite-dimensional weight spaces, its formal character is ch M=Σ_μ(dim M_μ)e^μ. The symbols e^μ are bookkeeping devices satisfying e^μe^ν=e^μ+ν. They are not numerical exponentials that must be evaluated at a chosen real argument.

For an integral sl₂ highest weight, write t instead. Then ch M(n)=tⁿ+tⁿ⁻²+tⁿ⁻⁴+···, abbreviated tⁿ/(1−t⁻²) with its downward formal expansion understood. At each weight there is one basis vector.

The general Verma character has one denominator factor for each positive root, as given in source [2]. That compact expression records weight multiplicities. It does not display every submodule or extension map. A character is a useful summary precisely because it forgets some of the full representation.

Character subtraction removes the infinite tail exactly

Apply additivity to the sequence for M(2). Its quotient character is ch L(2)=ch M(2)−ch M(−4). The first series begins t²+1+t⁻²+t⁻⁴+··· and the second begins t⁻⁴+t⁻⁶+···. Subtract coefficient by coefficient.

ch L(2)=t²+1+t⁻².

The answer agrees with the three basis classes already constructed. More generally ch L(n)=ch M(n)−ch M(−n−2) for n≥0. The cancellation leaves n+1 weights, from n down to −n in steps of two.

Do not set t=1 in the two infinite Verma series and attempt to subtract infinities. The valid operation is coefficientwise subtraction in the formal character setting, followed by evaluating the resulting finite Laurent polynomial when a dimension is wanted.

Finite length is not finite dimension

Objects of category O have finite composition length in this semisimple-Lie-algebra setting. Their successive simple quotients may themselves be infinite-dimensional. Thus finite length and finite dimension measure different things. The category properties used here are developed in source [2].

M(2) has composition length two: its submodule M(−4) is simple, and its quotient L(2) is simple. It still has infinitely many vector-space basis elements. M(−4) has composition length one and infinite vector-space dimension.

Whenever a table records a number beside a module, ask which number it is: highest weight, weight-space dimension, total dimension, composition length, or multiplicity in a filtration. Many seemingly contradictory statements become consistent once their measurement is named.

The Casimir explains the shifted reflection

For our sl₂ convention, C=h²+2h+4fe is central in U(sl₂). On the highest vector, fev₀=0, so C acts by λ(λ+2). The same scalar acts on all of M(λ), because the module is generated by v₀ and C commutes with the generators.

Direct substitution confirms that λ(λ+2)=(−λ−2)(−λ). Thus M(λ) and M(−λ−2) have the same central scalar. The reflection appropriate to highest weights is the dot action s·λ=−λ−2, not the unshifted reflection −λ.

The shift comes from the half-sum of positive roots: w·λ=w(λ+ρ)−ρ. For sl₂ in our weight coordinate, ρ=1. At λ=2 the paired weight is −4, exactly the singular-submodule weight found by the elementary coefficient calculation. At λ=−1 the shifted weight is fixed.

Central characters organise blocks

The centre of U(𝔤) acts by a character on a simple highest-weight module. Generalised central-character conditions organise category O into smaller pieces. In the regular integral sl₂ case relevant below, a block contains the two simple modules labelled by n and −n−2, for n≥0. Source [3] works out this rank-one block structure.

The word generalised matters for non-simple modules: a central element may act with a nilpotent part around the prescribed scalar. Restricting attention only to modules on which every central element is literally scalar can omit extensions needed by the full block.

Our principal block uses the pair 0 and −2. The two simples are the one-dimensional L(0) and the infinite-dimensional L(−2)=M(−2). They share the Casimir value zero. Sharing a central character is not the same as being isomorphic.

Projective covers answer a mapping question

A projective object P has the property that maps from P can be lifted through surjective morphisms. A projective cover P(λ)→L(λ) is a minimal projective source covering the chosen simple object, in the appropriate categorical sense.

Projective does not mean projective space, a projective representation up to phase, or a projection matrix. Here it is a property of how module maps lift. The surrounding category matters: a module projective in a category O block need not be projective among all U(𝔤)-modules.

Category O has enough projectives, and these projectives admit finite Verma filtrations. Rather than decomposing as direct sums of Verma modules, they are built through successive submodules whose quotients are Vermas. This is the structure to which BGG reciprocity applies. [3]

BGG reciprocity exchanges two multiplicity questions

Use [M(μ):L(λ)] for the composition multiplicity of L(λ) in M(μ). Use (P(λ):M(μ)) for the multiplicity of M(μ) as a quotient in a Verma filtration of P(λ). The BGG reciprocity theorem states

(P(λ):M(μ)) = [M(μ):L(λ)].

The two sides concern different filtrations of different modules. On one side, a simple module appears inside the composition series of a standard module. On the other, that standard module appears in a filtration of a projective. Source [3] proves the theorem using morphism spaces into contragredient Verma modules.

The theorem does not say P(λ) is a direct sum of the Vermas counted on the right. Multiplicity is a count of successive quotients, not a licence to erase their extensions. This warning is essential when turning a multiplicity table into a structural statement.

Compute the principal sl₂ block

We already know 0→M(−2)→M(0)→L(0)→0, and M(−2)=L(−2). Set the row order of standard modules to M(0),M(−2), and the column order of simples to L(0),L(−2). The composition matrix is

D = [1  1]
    [0  1].

Read the first column to describe the Verma filtration of P(0): one M(0), no M(−2). Thus P(0)=M(0). Read the second column for P(−2): one M(0) and one M(−2). Their extensions produce the larger projective in this block.

In the latter projective, the simple composition factors are one L(0) and two L(−2). Its composition length is three, not its vector-space dimension. The rank-one projective example in source [3] confirms that its top and socle are L(−2), with L(0) between them.

A second calculation: the Cartan matrix

Let C have row label λ for P(λ) and column label ν for L(ν), recording composition multiplicities [P(λ):L(ν)]. BGG reciprocity gives C=DᵀD with our row and column ordering.

D^T D = [1  0] [1  1] = [1  1]
        [1  1] [0  1]   [1  2].

The first row agrees with P(0)=M(0): one copy of each simple. The second agrees with the two-step Verma filtration of P(−2): one L(0) and two L(−2).

The entries also match dimensions of appropriate Hom spaces between projective covers. But the matrix is not, by itself, a complete presentation of the block algebra. Composition rules for maps and relations among them contain additional information that a table of dimensions does not display.

What characters cannot see: reverse an extension

Category O has a restricted contragredient duality preserving weights and formal characters. It reverses the direction of exact sequences and fixes the simple objects up to isomorphism. The duality is described in source [3].

Dualise 0→L(−2)→M(0)→L(0)→0. The resulting sequence has L(0) as submodule and L(−2) as quotient. The dual Verma has the same character as M(0), but its simple submodule is in the opposite place.

These modules are not isomorphic. M(0) has no nonzero invariant vector: a vector of weight zero is a multiple of v₀, and fv₀=v₁ is nonzero. Its dual, by contrast, contains a copy of the trivial module. Equal formal characters therefore do not classify all modules in this non-semisimple category.

Do not import finite-group character conclusions unchanged

Finite-dimensional complex representations of a finite group are completely reducible, and ordinary characters determine their equivalence classes. Category O contains non-split extensions and infinite-dimensional simples. The same conclusion cannot be copied merely because both theories use the word character.

Our formal characters record weight-space dimensions rather than traces of every element of a finite group. They are additive in exact sequences, which is precisely why a module and a reversed extension can have identical character data.

The correct comparison is useful rather than discouraging: character data tells us which simple contributions occur, while extension and morphism data tells us how they are attached. The two layers should be studied together, not forced into one invariant.

A safe route toward higher rank

For larger semisimple Lie algebras, the Cartan has several independent directions. Weights become linear functionals, several positive roots can raise a vector, and singularity requires annihilation by all positive simple-root generators. A vector killed by just one raising operator need not be highest weight.

The Verma construction and BGG reciprocity remain useful, but composition multiplicities are no longer all visible from one scalar coefficient. Kazhdan–Lusztig theory connects such multiplicities with Weyl-group combinatorics and, through geometric representation theory, with intersection cohomology. This article does not attempt a general computation of those polynomials.

The appropriate next step is to carry the rank-one habits forward: declare the positive-root convention, distinguish standard from simple and projective objects, record the indexing of every multiplicity matrix, and verify that a proposed subspace is stable under every generator.

A calculation checklist that actually catches errors

When studying a candidate highest-weight module, first verify the commutators. Next compute weights from h rather than infer them from basis position. Solve the singular-vector equation with the permitted integer index range. Test invariance of the resulting tail, and state explicitly whether the finite object is a quotient or a submodule.

After that, compare formal characters coefficient by coefficient. If a finite quotient is expected, its character should terminate for a mathematical reason. For projectives, identify the category and block before applying reciprocity. Label the rows and columns before transposing any matrix.

Finally separate what has been proved from what has merely been counted. A character identity verifies composition data. A dimension calculation verifies a numerical constraint. Neither automatically supplies a splitting, an isomorphism of modules or a presentation of an endomorphism algebra.

Practice questions: work before opening the answer route

1. In M(3), calculate the weight of v₄ and decide whether it is singular. 2. Write the first four nontrivial coefficients of e in M(1). 3. Decide whether M(−2) and M(1/2) are simple. 4. Construct the finite-dimensional quotient of M(3) and give its dimension.

5. Explain why span{v₀,v₁} is not a submodule of M(1). 6. Compute ch L(3) by subtracting Verma characters. 7. Find the shifted-reflection partner of λ=5 and compare their Casimir values. 8. Explain how an infinite-dimensional module can have composition length one.

9. Starting with D=[[1,1],[0,1]], read the standard multiplicities of P(−2). 10. Compute C=DᵀD and interpret its bottom-right entry. 11. Does equal formal character imply isomorphism in category O? Give a concrete reason. 12. Explain why infinitely many copies of L(0) do not form an object of category O.

Worked answers

1–3. Weights and vanishing coefficients

1. In M(3), hv₄=(3−8)v₄=−5v₄. Also ev₄=4(3−4+1)v₃=0. Thus v₄ is singular and generates a copy of M(−5). The vector exists in the Verma module; it becomes zero only after passing to the finite quotient.

2. For λ=1, the coefficient is r(2−r). At r=1,2,3,4 the values are 1,0,−3,−8. Thus ev₁=v₀, ev₂=0, ev₃=−3v₂ and ev₄=−8v₃. The zero at r=2 marks the highest vector of the proper tail.

3. Both M(−2) and M(1/2) are simple. Neither highest weight is a nonnegative integer, so there is no additional singular basis vector. In either module, any nonzero weight vector can be raised to v₀, which then generates the entire module.

4–6. Quotients and characters

4. Quotient M(3) by the span of v₄,v₅,… . The four surviving basis classes have weights 3,1,−1,−3. They form L(3), of dimension 4. The removed tail is M(−5), so the exact sequence is 0→M(−5)→M(3)→L(3)→0.

5. In M(1), fv₁=v₂ is nonzero and lies outside span{v₀,v₁}. Therefore the proposed span is not stable. The two-dimensional representation L(1) is obtained by quotienting out the tail beginning at v₂, not by keeping the first two basis vectors as a submodule.

6. Subtract the tail: ch M(3)−ch M(−5)=(t³+t+t⁻¹+t⁻³+t⁻⁵+···)−(t⁻⁵+t⁻⁷+···). The result is t³+t+t⁻¹+t⁻³. Its four terms agree with the quotient basis in answer 4.

7–10. Central data and reciprocity

7. The dot-action partner is −5−2=−7. The Casimir scalars are 5·7=35 and (−7)(−5)=35. Using the unshifted partner −5 would instead give 15 and fail the comparison.

8. Composition length counts simple factors, not basis vectors. M(−2) is simple, hence has length one, but contains basis vectors v₀,v₁,v₂,… . There is no contradiction because the simple factor itself is infinite-dimensional.

9. Read the L(−2) column of D. Both M(0) and M(−2) contain L(−2) once. Reciprocity therefore gives one copy of each Verma module in a standard filtration of P(−2). This is a filtration statement, not a direct-sum assertion.

10. The product is [[1,1],[1,2]]. Its bottom-right entry says that L(−2) occurs twice among the composition factors of P(−2). It does not say that P(−2) is two-dimensional or that its highest-weight space has dimension two.

11–12. Diagnose the missing hypothesis

11. No. M(0) and its restricted contragredient dual have the same formal character. The former has L(−2) as its simple submodule; the latter has L(0). In particular the dual has a nonzero invariant vector while M(0) does not, so the modules cannot be isomorphic.

12. In an infinite direct sum of trivial modules, the algebra action cannot move a chosen vector into new independent trivial directions. Finitely many generators span only a finite-dimensional submodule. The entire direct sum is therefore not finitely generated and fails a defining requirement of category O.

A teaching route from one chain to a category

Begin with M(2) and ask the learner to calculate ev₁ through ev₄ without being told the answer. Then ask whether ev₃=0 implies v₃=0. That single question identifies whether singularity has been confused with truncation.

Next compare the non-invariant finite span with the valid finite quotient. Require an explicit test under f. Only after that introduce the formal character and use it to reproduce the quotient weights. The learner then sees what the character summarises and what it omits.

Finish with the two-by-two multiplicity matrix. Read one column aloud as a BGG statement before multiplying matrices. The goal is not simply to obtain [[1,1],[1,2]], but to name what every row, column and entry measures. This turns a small calculation into a transferable method for reading much larger representation-theoretic tables.

Sources and further study

[1] Pavel Etingof, MIT 18.757, Highest Weight Modules and Verma Modules: the universal construction, highest weights and simple quotients. [2] Etingof, Category O of 𝔤-Modules, Part I: category O, finite weight spaces and formal characters. [3] Etingof, BGG Reciprocity and BGG Theorem: Verma filtrations, reciprocity, duality and the explicit sl₂ block. The cited lectures supply the general theory; the computations in this guide keep their conventions visible at each step.

For the earlier foundations, return to Lie Representations and Highest Weights or compare non-split structure with Quiver Representations. The BTT Mathematics Learning Hub connects the full representation-theory route.