Tensor-Product Multiplicities | Littlewood–Richardson Coefficients, Honeycombs and Saturation

Tensor-product multiplicities answer a basic but deep representation-theoretic question: when two irreducible representations are tensored, which irreducibles appear, and how many times does each one appear? For polynomial representations of GL_n, the answers are the Littlewood–Richardson coefficients.

The same nonnegative integers appear in several mathematical languages. They are coefficients in products of Schur functions, counts of Littlewood–Richardson tableaux, multiplicities of irreducible GL_n modules, structure constants in Grassmannian cohomology, counts of integral honeycombs or hives, and counts of highest-weight components in type-A crystal tensor products.

The existence of so many models is useful because each makes a different property visible. Tableaux make integrality and positivity combinatorial. Honeycombs turn positivity into a convex-geometric feasibility problem. The saturation theorem says that for GL_n, positivity after scaling already implies positivity before scaling.

This guide follows Tensor Products and Duality, Young Tableaux and Specht Modules and Crystal Bases.

GLn representations · Schur functions · LR tableaux · Multiplicity two · Honeycombs · Saturation · Crystals · Practice

Partitions index polynomial GLn representations

A partition λ=(λ₁≥λ₂≥···≥0) with at most n nonzero parts indexes an irreducible polynomial representation V_λ of GL_n(C).

The size |λ|=λ₁+λ₂+··· is the total number of boxes in the Young diagram. Tensoring V_λ and V_μ gives

V_λ⊗V_μ ≈ ⊕_ν c^ν_λμ V_ν.

The integer c^ν_λμ is the Littlewood–Richardson coefficient.

A necessary size condition is

|ν|=|λ|+|μ|.

If the sizes do not add, the multiplicity is zero before any tableau is drawn.

The first tensor products

Let V be the defining GL_n representation, indexed by partition (1). Then

V⊗V≈Sym²V⊕Λ²V.

In partition notation,

V₍₁₎⊗V₍₁₎≈V₍₂₎⊕V_(1,1).

Therefore c⁽²⁾_(1),(1)=1 and c^(1,1)_(1),(1)=1.

The decomposition is familiar from symmetric and alternating tensors. Littlewood–Richardson theory extends it to arbitrary partitions.

Schur functions encode the same multiplication

Associate a Schur symmetric function s_λ to each partition λ. Then

s_λs_μ=Σ_νc^ν_λμs_ν.

The same coefficients are therefore structure constants of the Schur basis.

This gives an immediate algebraic reason for associativity identities among LR coefficients: multiplication of symmetric functions is associative.

It also gives a character-theoretic interpretation because Schur functions are the polynomial characters of irreducible GL_n representations in sufficiently many variables.

The Pieri rule is the first efficient special case

Multiplying by s_(r) adds r boxes with no two in the same column. Multiplying by s_(1^r) adds r boxes with no two in the same row.

For r=1, one simply adds one box at any location that still produces a partition.

For example,

s_(2,1)s₍₁₎=s_(3,1)+s_(2,2)+s_(2,1,1).

Each coefficient is one because each resulting diagram is obtained by one permitted single-box addition.

The Littlewood–Richardson tableau rule

Assume λ⊆ν as Young diagrams. The skew shape ν/λ consists of the boxes of ν not already in λ.

An LR tableau of shape ν/λ and content μ is a semistandard filling of the skew boxes with μ₁ copies of 1, μ₂ copies of 2, and so on, subject to a lattice-word condition.

We use the reverse-row reading convention: read each row from right to left, starting at the top row and moving downward. In every initial segment of this word, the number of 1s must be at least the number of 2s, the number of 2s at least the number of 3s, and so on.

Then

c^ν_λμ = number of LR tableaux of shape ν/λ and content μ.

Other reading-word conventions can state an equivalent rule differently. One should not mix the word direction from one convention with the lattice condition from another.

Worked tableau: V(2) tensor V(1)

Take λ=(2) and μ=(1). We must add one box to λ. The only partition shapes of size 3 containing (2) are (3) and (2,1).

Each skew shape has one box, and the content has one symbol 1. Each therefore gives exactly one LR tableau.

Hence

V₍₂₎⊗V₍₁₎≈V₍₃₎⊕V_(2,1).

This is the representation version of the Pieri rule.

A genuine multiplicity-two example

Take λ=μ=(2,1) and ν=(3,2,1). The skew shape ν/λ has three boxes:

• row 1, column 3;
• row 2, column 2;
• row 3, column 1.

The content μ=(2,1) means two 1s and one 2.

Because the three skew boxes lie diagonally, the row and column semistandard constraints impose no extra comparison between them. The lattice-word condition leaves exactly two fillings.

Tableau A reading word: 1, 1, 2
Tableau B reading word: 1, 2, 1

Every prefix has at least as many 1s as 2s. Therefore both tableaux are valid and

c^(3,2,1)_(2,1),(2,1)=2.

This is an important conceptual step. Tensor multiplicity is not restricted to zero or one. The same irreducible representation can occur through more than one independent highest-weight channel.

The complete square of s(2,1)

In the ring of symmetric functions,

s_(2,1)² = s_(4,2) + s_(4,1,1) + s_(3,3) + 2s_(3,2,1) + s_(3,1,1,1) + s_(2,2,2) + s_(2,2,1,1).

For GL₃, partitions with more than three rows vanish from the polynomial representation list. The decomposition becomes

V_(2,1)⊗V_(2,1)≈V_(4,2)⊕V_(4,1,1)⊕V_(3,3)⊕2V_(3,2,1)⊕V_(2,2,2).

Dimension verification in GL3

The GL₃ representation V_(2,1) has dimension 8. Its tensor square therefore has dimension 64.

The dimensions of the summands are

dim V_(4,2)   = 27
dim V_(4,1,1) = 10
dim V_(3,3)   = 10
dim V_(3,2,1) = 8
dim V_(2,2,2) = 1

Therefore

27+10+10+2·8+1=64.

The dimension check independently confirms that the multiplicity two is necessary.

The Weyl dimension formula supplies the check

For a GL_n highest weight λ=(λ₁,…,λₙ),

dim V_λ=∏_1≤i<j≤n(λᵢ−λⱼ+j−i)/(j−i).

For λ=(2,1,0) in GL₃, the three factors are

2, 2, 2,

so the dimension is 8.

A tensor decomposition that fails the total-dimension check is definitely wrong, although passing the dimension check alone does not prove the decomposition.

Schubert calculus gives a geometric interpretation

Schur functions also describe Schubert classes in the cohomology ring of a Grassmannian. Multiplying two Schubert classes expands with Littlewood–Richardson coefficients.

Geometrically, these coefficients count suitable intersections of general-position Schubert varieties, with the indexing constrained by the ambient rectangle.

Thus the same integer can be read as a representation multiplicity or as an intersection number.

This is another example of geometric representation theory turning an algebraic coefficient into a geometric count.

Honeycombs turn multiplicity into convex geometry

Knutson and Tao introduced a honeycomb model for GL_n tensor products. Boundary data encodes the three dominant weights, while internal edges satisfy balancing conditions.

For integral boundary data, the Littlewood–Richardson coefficient can be realised as the number of integral honeycombs, equivalently through related hive or Berenstein–Zelevinsky polytope models after translating conventions.

The combinatorial tableau count has therefore become a lattice-point count in a convex polytope.

LR tableau count ↔ integer points in a hive/honeycomb polytope ↔ tensor multiplicity.

Why the polytope model changes the positivity question

To ask whether c^ν_λμ>0 is to ask whether the relevant polytope contains at least one integer point.

If one scales all three boundary weights by N, the polytope scales. Positivity after scaling therefore becomes a question about rational feasibility versus integral feasibility.

The saturation theorem says that in type A, this distinction disappears for positivity: if a scaled integral point exists at some positive integer scale, an integral point already exists at scale one.

Hive inequalities

An equivalent hive model labels the vertices of a triangular grid by numbers satisfying rhombus inequalities. The three boundary difference sequences encode λ, μ and ν.

The inequalities are linear. Therefore the set of real hives with fixed boundary data is a convex polytope.

Integer hives count Littlewood–Richardson coefficients. Linear inequalities make the feasibility problem accessible to convex optimization and polyhedral methods, while integrality retains the discrete multiplicity.

Horn’s eigenvalue problem

Suppose A and B are Hermitian matrices and C=A+B. Which triples of eigenvalue lists can occur?

Horn conjectured a finite family of linear inequalities describing the answer. Klyachko connected the problem to representation theory, and Knutson–Tao’s honeycomb theorem completed the saturation route to Horn’s conjecture.

For integral dominant spectra, positivity of the appropriate LR coefficient captures the same feasibility structure after the trace-sum condition and indexing conventions are aligned.

The tensor-product problem has therefore encoded a matrix eigenvalue-sum problem.

The saturation theorem

For partitions λ, μ and ν appropriate to GL_n, Knutson and Tao proved:

c^Nν_Nλ,Nμ>0 for some integer N≥1 if and only if c^ν_λμ>0.

The reverse direction is immediate by scaling a valid combinatorial/geometric solution. The deep direction says that positivity cannot first appear only after dilation.

The theorem concerns positivity, not equality of coefficients. In general

c^Nν_Nλ,Nμ

can be much larger than c^ν_λμ.

A simple saturation check

We know c⁽²⁾_(1),(1)=1. Scaling by N=2 gives the question whether V₍₄₎ appears in V₍₂₎⊗V₍₂₎. By the Pieri rule it does, so c⁽⁴⁾_(2),(2)>0.

This verifies the easy direction in one example.

Saturation becomes informative when one knows only that a high-scale coefficient is positive and wishes to conclude scale-one positivity.

Saturation is special and should not be overgeneralised

The factor-one saturation theorem is a type-A phenomenon for GL_n/SL_n tensor multiplicities.

Other reductive groups can require nontrivial saturation factors, and naive transplantation of the GL_n statement can fail.

Therefore a positivity statement for symplectic, orthogonal or exceptional groups should identify the relevant saturation theorem rather than citing type A automatically.

Littlewood–Richardson cones

The triples (λ,μ,ν) with positive LR coefficient form the integral points of a rational polyhedral cone after the correct linear equalities and inequalities are imposed.

Saturation says this semigroup of positive triples is saturated inside its lattice for GL_n: if a positive integer multiple lies in the semigroup, the primitive lattice point lies there too.

This is the semigroup-theoretic form of the theorem and explains the word saturation.

Crystal interpretation

Let B(λ) and B(μ) be type-A highest-weight crystals. Their tensor product decomposes into connected highest-weight components.

The number of highest-weight nodes of weight ν equals c^ν_λμ.

Thus the LR coefficient can be found by scanning a crystal graph for nodes annihilated by every raising operator ẽ_i.

The tableau LR rule and the crystal highest-weight rule are two combinatorial presentations of the same tensor-product decomposition.

Why multiplicities exceed one

If two distinct highest-weight nodes in B(λ)⊗B(μ) have the same weight ν, they generate two disconnected components isomorphic to B(ν).

The multiplicity-two coefficient c^(3,2,1)_(2,1),(2,1)=2 means exactly that there are two independent highest-weight channels of that weight.

This explains multiplicity structurally rather than treating the coefficient 2 as an unexplained integer in a polynomial expansion.

Associativity imposes convolution identities on LR coefficients

Because tensor products are associative up to canonical isomorphism, decomposing (V_λ⊗V_μ)⊗V_κ or V_λ⊗(V_μ⊗V_κ) gives the same final multiplicities.

Therefore

Σ_ρc^ρ_λμc^ν_ρκ=Σ_σc^σ_μκc^ν_λσ.

This identity is automatic in Schur-function multiplication and tensor categories but becomes a nontrivial combinatorial identity when written purely as tableau counts.

Duality and determinant shifts

For GL_n, adding the same integer r to every highest-weight coordinate corresponds to tensoring by det^r.

This can shift partitions inside a fixed-rank representation problem without changing the underlying semisimple SL_n content.

When comparing LR conventions across GL_n, SL_n and eigenvalue problems, record determinant shifts and duals explicitly. Otherwise identical multiplicity data can appear under different partitions.

A practical LR workflow

• 1. Check sizes. Require |ν|=|λ|+|μ|.
• 2. Check containment. For the tableau rule, λ must lie inside ν.
• 3. Fix the reading-word convention.
• 4. Fill the skew shape semistandardly with content μ.
• 5. Apply the lattice-word test to every prefix.
• 6. Count valid tableaux.
• 7. Verify with Schur/Pieri rules where available.
• 8. Perform a dimension check on the full tensor decomposition.
• 9. For positivity at large scale, use saturation only in an appropriate type-A setting.
• 10. When using honeycombs or hives, state boundary and orientation conventions.

Common mistakes

• Forgetting the size equation: a tableau cannot repair a degree mismatch.
• Mixing reading conventions: the lattice condition depends on how the word is read.
• Assuming coefficients are only 0 or 1: multiplicities such as 2 occur quickly.
• Treating a dimension check as proof: different decompositions can have the same total dimension.
• Confusing coefficient value with positivity: saturation preserves positivity, not the numerical coefficient.
• Applying GL_n saturation unchanged to every group.
• Counting real honeycombs instead of integral honeycombs when computing an LR coefficient.
• Ignoring rank restrictions: partitions with too many rows do not label polynomial GL_n irreducibles.

Practice questions

1. What size must ν have for c^ν_(3,1),(2) to be nonzero? 2. Expand s₍₁₎s₍₁₎. 3. Use Pieri to expand s_(2,1)s₍₁₎. 4. What does c^ν_λμ=3 mean representation-theoretically?

5. Why is c^(3,2,1)_(2,1),(2,1)=2? 6. In GL₃, verify the dimension of V_(2,1). 7. What does an integer honeycomb count? 8. State the saturation theorem.

9. Does saturation say the scaled coefficient equals the original coefficient? 10. What does a highest-weight crystal node count? 11. Why can the same LR coefficient be a Schubert intersection number? 12. What is the main warning when moving from GL_n to other Lie types?

Worked answers

1. |ν|=4+2=6.

2. s₍₂₎+s_(1,1).

3. s_(3,1)+s_(2,2)+s_(2,1,1).

4. The irreducible V_ν appears three times as a direct summand in the semisimple tensor product.

5. The skew shape (3,2,1)/(2,1) has exactly two LR fillings with content (2,1), whose reading words are 112 and 121.

6. Weyl’s formula gives factors 2,2,2, so the dimension is 8.

7. With the appropriate boundary data and convention, integral honeycombs count the LR multiplicity.

8. c^Nν_Nλ,Nμ>0 for some N≥1 if and only if c^ν_λμ>0 in type A.

9. No. It preserves positivity, not coefficient value.

10. Highest-weight nodes of weight ν count irreducible B(ν) components in the tensor-product crystal.

11. Schur functions also represent Schubert cohomology classes; their product structure constants are LR coefficients.

12. Type-A saturation with factor one need not hold unchanged; different groups can require different saturation factors and combinatorics.

Sources and further study

[1] Allen Knutson and Terence Tao, The Honeycomb Model of GL(n) Tensor Products I: Proof of the Saturation Conjecture, the foundational honeycomb proof of saturation and its Horn-conjecture consequences. [2] Alexander Postnikov’s MIT Combinatorial Theory course includes Schur functions, LR coefficients, tableaux, Berenstein–Zelevinsky triangles and honeycombs. [3] Richard Stanley’s survey on positivity reviews LR coefficients as Schur structure constants and representation multiplicities. [4] The previous BTT Crystal Bases guide gives the highest-weight component interpretation.

Representation Mathematics — Batch 08

Build the symmetry algebra in Kac–Moody Algebras, specialise to loop symmetry in Affine Lie Algebra Representations, and recover highest-weight multiplicities graphically in Crystal Bases. Return to the BTT Mathematics Learning Hub.