Fusion Categories | Fusion Rules, Quantum Dimensions and Modular Tensor Categories

A fusion category describes finitely many basic types of objects and the ways their tensor products break into those types. The decomposition multiplicities are integers, but the dimension associated with a basic type can be irrational. Understanding why both statements are compatible is the first useful step into the subject.

The Fibonacci example makes the distinction concrete. A basic object τ satisfies τ⊗τ ≅ 1⊕τ. Its positive dimension solves d²=1+d, giving the golden ratio. Nevertheless, six copies of τ with total output 1 have a five-dimensional fusion space, not a space with an irrational number of basis vectors. The irrational dimension measures multiplicative growth; the integer multiplicity counts actual independent channels.

This guide develops that calculation, compares it with Ising fusion rules, and then asks what the rules leave out. Associators specify how different bracketings are related. Braiding describes exchanges. A modular S-matrix carries enough information to recover fusion coefficients through the Verlinde formula. Each new layer adds requirements that a multiplication table alone cannot provide.

Scope: finite semisimple complex tensor categories. Quantum dimensions are discussed with explicit pivotal or unitary assumptions when required. The word modular here is not the modular representation theory of a field whose characteristic divides a group order. For prerequisites, use Tensor Products and Duality and Quantum Groups and Braided Representations.

Definition · Fibonacci calculations · Associators · Ising calculations · Modularity and Verlinde · Practice · Worked answers · BTT Mathematics Hub

What a fusion category contains

Over C, a fusion category is a C-linear semisimple rigid tensor category with finite-dimensional morphism spaces, finitely many isomorphism classes of simple objects, and a simple tensor unit. Every object is a finite direct sum of simples. Rigidity supplies dual objects. These are the standard hypotheses used in sources [1] and [2].

The tensor unit, written 1, satisfies 1⊗X ≅ X ≅ X⊗1. The zero object is different: tensoring with zero gives zero. A simple object has no nonzero proper subobject, but it need not have an ordinary one-dimensional vector-space realisation.

A fusion category is not automatically braided. A braiding is additional compatible structure relating X⊗Y to Y⊗X. Likewise, a spherical structure or a ribbon twist is not part of the bare definition. These distinctions matter before one starts interpreting diagrams as freely interchangeable strands.

Fusion coefficients count copies, not probabilities

Choose simple representatives X₀=1,X₁,…,Xᵣ. Write Xᵢ⊗Xⱼ ≅ ⊕ₖ NᵢⱼᵏXₖ. Each coefficient Nᵢⱼᵏ is a nonnegative integer. Over C it equals the dimension of Hom(Xₖ,Xᵢ⊗Xⱼ), because the endomorphism algebra of a simple object is C.

The expression Nᵢⱼᵏ=2 means there are two independent copies of the output type Xₖ. It does not mean a probability of two, an amplitude of two, or two differently named simple types. A physical interpretation requires a state and an inner-product structure in addition to the fusion rule.

Duality constrains these coefficients. The tensor unit occurs in Xᵢ⊗Xⱼ exactly when Xⱼ is the dual of Xᵢ, and then its multiplicity is one. The unit rules themselves give N₀ⱼᵏ=δⱼₖ. These are efficient consistency checks on a proposed table.

A familiar example: representations of S3

The finite-dimensional complex representations of a finite group form a fusion category. For S3, the simples are 1, the sign representation ε, and the two-dimensional standard representation V. Their nontrivial rules are ε²=1, εV=V and V²=1+ε+V, where multiplication abbreviates tensor product.

Ordinary dimensions respect every rule: 1·1=1, 1·2=2 and 2²=1+1+2. These dimensions are integers because the category already comes with the usual forgetful functor to vector spaces.

We know more than the multiplication table in this example. We know actual group actions, intertwiners, tensor associativity and the ordinary flip of vector factors. Fusion categories retain this style of organisation while allowing examples that cannot be realised by simply forgetting to finite-dimensional vector spaces.

Fusion matrices turn repeated products into linear algebra

For each simple Xᵢ, define a matrix Nᵢ with entry (Nᵢ)ₖⱼ=Nᵢⱼᵏ. The column index j specifies the input type; the row index k specifies the output type after tensoring on the left by Xᵢ. State this convention before multiplying matrices.

If an object has multiplicity column vector m, the object Xᵢ⊗X has multiplicity vector Nᵢm. Repeated tensoring uses powers of Nᵢ. Associativity of the tensor product implies NᵢNⱼ=ΣₖNᵢⱼᵏNₖ.

These matrices act on the free abelian group generated by simple isomorphism classes. They are not automatically matrices acting on the underlying physical states of one particle. The coordinates count simple components of objects, not amplitudes in a chosen state vector.

Fibonacci fusion: one rule, an entire sequence

The unitary Fibonacci example has two simple types, 1 and τ, with τ²=1+τ. In the ordered basis (1,τ), its multiplication matrix is

N_τ = [0  1]
      [1  1].

The first column records τ·1=τ. The second records τ·τ=1+τ. Squaring gives [[1,1],[1,2]], equal to I+N_τ, exactly as the fusion rule requires.

Suppose τⁿ=aₙ1+bₙτ. Multiplying by τ gives aₙ₊₁=bₙ and bₙ₊₁=aₙ+bₙ. Starting with a₁=0,b₁=1, the successive decompositions are τ²=1+τ, τ³=1+2τ, τ⁴=2+3τ, τ⁵=3+5τ and τ⁶=5+8τ.

This is the Fibonacci recurrence derived directly from the fusion multiplication, not inferred from a suggestive name. Source [3] develops the same model in terms of fusion channels and associativity moves.

The integer five is a genuine vector-space dimension

The coefficient of 1 in τ⁶ is the dimension of Hom(1,τ⁶), a complex vector space of independent ways to obtain total output 1. It is 5. The coefficient of τ is the dimension of Hom(τ,τ⁶), which is 8.

The total output must be specified. Saying “six Fibonacci objects give eight states” without naming the output sector omits part of the problem. Choosing total output 1 gives five channels, while choosing τ gives eight.

Changing the parenthesisation of the sixfold tensor product changes a natural channel basis, but not these dimensions. The associator provides the map between the different bracketed descriptions. That map is additional information beyond the integer count.

Where the golden ratio enters

A positive multiplicative dimension function has d(1)=1 and must satisfy d(τ)²=1+d(τ). The positive solution is φ=(1+√5)/2. The other root, (1−√5)/2, is negative and cannot be the Frobenius–Perron dimension.

The characteristic polynomial of N_τ is x²−x−1, so φ is also its largest positive eigenvalue. The vector (1,φ) is an eigenvector because N_τ(1,φ)ᵀ=(φ,1+φ)ᵀ=φ(1,φ)ᵀ.

There is no vector space with φ basis vectors. Instead, φ controls the growth of multiplicities under repeated fusion. The exact sixfold identity passes the dimension test φ⁶=5+8φ. One number tracks multiplicative size while the coefficients retain integer channel counts.

Frobenius–Perron dimension versus quantum dimension

The Frobenius–Perron dimension is obtained from the positive spectral data of fusion matrices. It is additive on direct sums and multiplicative on tensor products. Its existence in a general fusion category is developed in source [2]. With our matrix convention, the dimension row vector satisfies dᵀNᵢ=dᵢdᵀ.

A categorical or quantum dimension instead uses evaluation, coevaluation and a suitable pivotal structure. A spherical structure makes the left and right traces agree. In the unitary examples considered here, the canonical positive quantum dimensions agree with Frobenius–Perron dimensions. That agreement should not be asserted for every possible choice of pivotal structure in every fusion category.

For Fibonacci, the Frobenius–Perron dimension of the category is 1+φ²=(5+√5)/2. In the unitary spherical setting this is also the global dimension. The total quantum dimension is its positive square root. These two category-level quantities differ by a square and should not share an ambiguous symbol in the same calculation.

A dimension equation rules out an ordinary fibre functor

Suppose the Fibonacci category had a faithful exact tensor functor to finite-dimensional complex vector spaces. Let m be the ordinary dimension of the vector space assigned to τ. Tensor compatibility would force m²=1+m.

No nonnegative integer solves that equation. The positive root is irrational. Therefore the proposed ordinary vector-space fibre functor cannot exist.

This short argument explains a boundary of the Tannakian reconstruction route. A fusion category may have finite-dimensional morphism spaces and perfectly coherent tensor products without being the ordinary representation category of a group with a forgetful vector-space functor.

Associativity needs a map, not only matching numbers

The objects (X⊗Y)⊗Z and X⊗(Y⊗Z) are related by a specified associator isomorphism. In a chosen channel basis, that map is recorded by F-matrices. Merely checking that both sides contain the same simple multiplicities does not determine the isomorphism.

For four objects, there are several sequences of rebracketing moves between the same endpoints. The pentagon identity requires the resulting maps to agree. Without this coherence, an answer could depend on an arbitrary sequence of parenthesis changes.

This is the categorical upgrade from an associative multiplication table to an associative tensor system. The table says that the counts match. The associator says how the spaces are identified. The pentagon says that those identifications fit together.

An explicit Fibonacci F-matrix

For three τ objects with total output τ, there are two intermediate channels, 1 and τ. In a standard unitary choice of channel bases, the associator is represented by

F = [ φ^(-1)      φ^(-1/2) ]
    [ φ^(-1/2)   −φ^(-1)   ].

Source [3], equations 9.122–9.125, gives this choice after a phase convention is fixed. Let a=φ⁻¹ and b=φ⁻¹/². Because φ²=φ+1, a²+b²=φ⁻²+φ⁻¹=1. The off-diagonal terms in F² cancel, so F²=I.

The calculation verifies that this matrix is real orthogonal and its own inverse. It does not, by itself, verify every pentagon equation of a full fusion category. Other output sectors and compatibility conditions must also be included. A useful local check is not a replacement for the complete coherence theorem.

Changing bases changes F without changing the category

A fusion channel space may have several choices of basis. Rephasing or changing those bases alters the printed F-matrices. It does not automatically produce a different tensor category.

Conversely, identical fusion rules do not guarantee that two associator systems can be related by such a basis change. To decide equivalence, one must compare the coherent data, not just one matrix or one table.

The distinction resembles ordinary representation theory: matrices may differ after a change of basis while encoding the same action. Here the coordinate choices occur in an entire family of morphism spaces, and the basis changes must remain compatible with the tensor structure.

A small example with the same fusion rules but a different associator

Consider vector spaces graded by Z₂, with simple objects labelled 0 and 1 and tensor rule a⊗b=a+b modulo 2. The unit is label 0. The fusion table is simply the group law.

One may multiply the usual associator on homogeneous labels a,b,c by a scalar ω(a,b,c). Coherence requires a normalized group 3-cocycle. Such pointed categories, conventionally written Vec_G^ω, are standard examples in source [1].

Take ω(a,b,c)=(−1)ᵃᵇᶜ for bits a,b,c. This leaves every fusion coefficient unchanged. Nevertheless it changes the associator when all three labels are 1. The effect occurs at the level of a map, not at the level of the multiplication table.

Verify the 3-cocycle directly

For an additive group, the pentagon scalar condition reads ω(b,c,d)ω(a,b+c,d)ω(a,b,c)=ω(a+b,c,d)ω(a,b,c+d). All sums in this example are modulo 2.

The left exponent is bcd+a(b+c)d+abc. The right exponent is (a+b)cd+ab(c+d). Expanding gives ab d+ac d+bc d+ab c on both sides, so the identity holds for every choice of the four bits.

This twist cannot be removed by a normalized two-cochain β. At (1,1,1), its coboundary is β(1,1)β(1,0)/[β(0,1)β(1,1)]=1, whereas ω(1,1,1)=−1. Thus the twisted and untwisted tensor structures have the same labelled fusion ring but inequivalent associator classes.

The example is an explicit demonstration of lost information under passage to fusion rules. It also distinguishes this 3-cocycle problem from the 2-cocycle phase factors in Projective Representations.

Ising fusion gives a second complete calculation

The Ising fusion rules use three simple types 1,ψ,σ with ψ²=1, ψσ=σψ=σ and σ²=1+ψ. These rules and their modular realizations are discussed in sources [1] and [4]. The symbol σ here is a simple-object label, not a permutation or a standard deviation.

Positivity gives dψ=1 from dψ²=1. Then dσ²=1+dψ=2, so dσ=√2. The global dimension in the unitary spherical example is 1²+1²+(√2)²=4, and the total quantum dimension is D=2.

In the ordered basis (1,ψ,σ), the nontrivial fusion matrices are

N_ψ = [0  1  0]       N_σ = [0  0  1]
      [1  0  0]             [0  0  1]
      [0  0  1]             [1  1  0].

Direct multiplication gives Nψ²=I, NψNσ=Nσ and Nσ²=I+Nψ. Each identity is a matrix version of one fusion rule. This is a practical way to test that row and column labels have not been interchanged.

Count six Ising σ objects

Starting with σ²=1+ψ, we obtain σ³=2σ and σ⁴=2(1+ψ). Multiplying twice more gives σ⁵=4σ and σ⁶=4(1+ψ).

Thus the total-output-1 fusion space for six σ objects has dimension 4, and the total-output-ψ space also has dimension 4. No σ output occurs in an even power. The dimension check is (√2)⁶=8=4·1+4·1.

More generally, σ²ʳ=2ʳ⁻¹(1+ψ) for r≥1 and σ²ʳ⁺¹=2ʳσ for r≥0. These formulas are derived by repeatedly multiplying the displayed rules. They are exact channel counts, not approximations based on the asymptotic meaning of quantum dimension.

Braiding adds the exchange maps

A braiding is a natural isomorphism c_X,Y:X⊗Y→Y⊗X satisfying hexagon identities with the associator. On fusion channel bases, exchange information is recorded by R-matrices.

Associativity and exchange solve different problems. F changes a bracketing while preserving object order. R exchanges neighbouring objects. Combining them lets one express an exchange in a basis adapted to a different fusion order.

The coefficients N alone cannot tell us which exchange operator occurs. Two choices of braided or ribbon structure can have the same fusion multiplicities. Therefore a claimed braid-group representation needs F and R data with coherence, not just a matrix N counting how many channels exist.

What makes a tensor category modular?

In the setting used here, a modular tensor category is a ribbon fusion category with nondegenerate braiding, equivalently with an invertible S-matrix. A ribbon structure supplies a compatible twist and duality, and the S-matrix records traces of double braiding. Source [1], Section 8.14, develops this criterion and the Verlinde formula.

A transparent object is one whose double braiding with every object is the identity. In a nondegenerate braided fusion category, the transparent objects are just sums of the tensor unit. Symmetry in the ordinary vector-space sense makes every object transparent, which is too much transparency rather than too little structure.

For this reason Rep(S3) with its usual flip is braided and symmetric but not modular. Its double-braiding trace matrix is proportional to the outer product of the dimension vector with itself, so it has rank one rather than three.

A normalized Ising S-matrix

For the standard unitary Ising example, choose the normalized matrix

S = (1/2) [ 1    1    √2 ]
          [ 1    1   −√2 ]
          [ √2  −√2   0 ].

The order of labels is 1,ψ,σ. Direct multiplication gives S²=I. This agrees with charge conjugation because all three labels are self-dual. The first row gives dimensions through dᵢ=S₀ᵢ/S₀₀: the answers are 1,1,√2.

Our normalization has S₀₀=1/D=1/2 and makes S unitary. Some texts, including source [1] in its displayed modular formulas, use an unnormalized matrix with the unit entry equal to 1. Scaling by D changes the factors in the formulas. A normalization mismatch can turn correct integer fusion coefficients into incorrect fractions.

The Verlinde formula reconstructs fusion coefficients

For a normalized unitary modular S-matrix, the formula is

Nᵢⱼᵏ = Σₗ SᵢₗSⱼₗ overline(Sₖₗ) / S₀ₗ.

The sum ranges over simple labels. The bar means complex conjugation. All entries in our Ising example are real, so conjugation does not visibly change them, but it belongs in the general formula. Sources [1] and [4] explain the modular setting.

The formula is not a rule for an arbitrary invertible matrix. It applies to S arising from the required categorical structure. Conversely, checking that one proposed matrix produces nonnegative integers is useful evidence but does not construct its underlying category or verify all associators and twists.

Recover σ² = 1 + ψ entry by entry

For Nσσ¹, the conjugated unit row cancels the denominator. The remaining sum is (√2/2)²+(−√2/2)²+0=1. Thus the unit occurs once.

For Nσσψ, the first two terms are each 1/2 and the third vanishes because Sσσ=0. The result is again 1. For Nσσσ, the two nonzero terms are √2/2 and −√2/2, which cancel, giving 0.

We have reconstructed exactly the three coefficients of σ². Checking the other products similarly gives ψ²=1 and ψσ=σ. The exercise returns a spectral matrix calculation to the original integer multiplication table.

S also diagonalises the fusion operators

The eigenvalues of multiplication by σ are Sσₗ/S₀ₗ. For the three columns they are √2,−√2 and 0. These are exactly the roots of the characteristic polynomial of Nσ.

The positive eigenvalue √2 is its Frobenius–Perron dimension. The negative eigenvalue does not become a negative physical state count; it is part of the spectral decomposition of a matrix with nonnegative integer entries.

This connects the subject to Fourier-style representation analysis. A multiplication operator becomes simpler in coordinates adapted to its simultaneous spectral structure. The categorical hypotheses explain why this particular change of coordinates exists.

Quantum groups supply examples, but not by ignoring non-semisimplicity

At a root of unity, the full category of finite-dimensional quantum-group modules can be non-semisimple. It is not automatically a fusion category merely because the parameter has a special value.

Standard constructions of semisimple fusion categories instead use an appropriate tilting-module category and quotient by negligible morphisms, with conditions on the parameter and chosen quantum group. Source [1] treats the resulting Verlinde categories.

The quotient is part of the mathematical construction. Removing it from the explanation makes an incorrect claim about the entire original module category. This is another instance where a useful simplified object is obtained by a controlled operation, not by pretending the discarded structure never existed.

What the anyon interpretation does and does not establish

In mathematical models of topological phases, simple labels can represent charge types, fusion spaces describe permitted total charges, and braiding acts on those spaces. Sources [3] and [4] develop this interpretation.

The category supplies exact algebraic constraints. It does not by itself establish that a proposed laboratory system realises the model, that a device is error-free, or that a particular braid implementation is available. Those would require separate physical evidence and engineering analysis.

Likewise, nuclear fusion is unrelated to the word fusion used here. The present operation is a tensor-product decomposition rule in a mathematical category, with a particular interpretation in some quantum models.

A layered verification strategy

First test the unit, duality and associativity constraints on the integer fusion coefficients. Construct fusion matrices with labelled rows and columns. Calculate the positive dimension vector and verify dᵢdⱼ=ΣₖNᵢⱼᵏdₖ. Recompute a few tensor powers directly.

Next specify the associator and its basis convention. Test the pentagon, not just invertibility of individual F-matrices. For a braided theory add R and the hexagons. For a modular theory identify a ribbon structure, state the S normalization and test nondegeneracy.

Finally use Verlinde as a return check and keep its logical direction clear. Valid local matrix calculations do not replace construction of the whole category. A fusion ring, a fusion category, a braided category and a modular tensor category are four increasingly structured objects.

Practice questions

1. Compute τ⁷ from τ⁶=5+8τ. 2. Find the dimensions of Hom(1,τ⁷) and Hom(τ,τ⁷). 3. Solve the positive dimension equation for τ. 4. Explain why this does not give an irrational-dimensional vector space.

5. Calculate σ⁴ and σ⁵ in the Ising ring. 6. Find the global dimension and total quantum dimension of the unitary Ising category. 7. Verify the top-left entry of the displayed Fibonacci F². 8. Does that check establish the entire pentagon system?

9. Use the Ising S-matrix to calculate Nψψ¹. 10. Explain why Rep(S3) with its ordinary flip is not modular. 11. Can a faithful tensor functor send Fibonacci objects to ordinary finite-dimensional complex vector spaces? 12. Explain what changes, and what stays the same, when the Z₂ associator is twisted by (−1)ᵃᵇᶜ.

Worked answers

1–4. Fibonacci multiplicity and dimension

1. Multiply 5+8τ by τ. The result is 5τ+8(1+τ)=8+13τ. No associator matrix is needed to compute these multiplicities, because the fusion ring already records products up to isomorphism.

2. The unit multiplicity is 8 and the τ multiplicity is 13. Therefore the two Hom-space dimensions are 8 and 13 respectively. The output label selects which coefficient answers the question.

3. Solve d²−d−1=0 to get (1±√5)/2. Positivity selects φ=(1+√5)/2. The same value is the Perron eigenvalue of [[0,1],[1,1]], so the ring and matrix calculations agree.

4. The number φ is a multiplicative dimension attached to an abstract tensor object. Actual fusion channel spaces are ordinary complex vector spaces with integer dimensions such as 8 and 13. The recurrence makes their counts grow at a rate governed by φ.

5–8. Ising products and associator checks

5. σ⁴=(1+ψ)²=1+2ψ+ψ²=2+2ψ. Multiplying by σ gives σ⁵=2σ+2ψσ=4σ. The rules ψ²=1 and ψσ=σ must both be used.

6. The squared simple dimensions sum to 1+1+2=4. That is the global dimension in this unitary spherical example. Its positive square root, 2, is the total quantum dimension. Reporting both as 4 would mix conventions.

7. The top-left entry is φ⁻²+φ⁻¹. Divide φ²=φ+1 by φ² to obtain 1=φ⁻¹+φ⁻². Thus the entry is 1; the off-diagonal entries cancel because the diagonal entries have opposite signs.

8. No. F²=I checks one inverse relation for one chosen channel matrix. Pentagon coherence compares sequences of associators for four objects and includes all permitted intermediate and output labels. A full category requires those additional equations.

9–12. Spectral reconstruction and categorical limits

9. In Nψψ¹, the unit row cancels the denominator. The sum is (1/2)²+(1/2)²+(−√2/2)²=1/4+1/4+1/2=1. Thus the unit occurs once in ψ².

10. With the ordinary flip, swapping twice is the identity for every pair of representations. Every object is transparent. The double-braiding trace matrix is the outer product of dimensions (1,1,2) with itself, so its rank is one. It cannot be an invertible three-by-three modular S-matrix.

11. No. The image of τ would have a nonnegative integer dimension m. Preservation of τ²=1+τ would require m²=1+m, which has no integer solution. This supplies a direct obstruction without appealing to an abstract classification theorem.

12. The simple labels and tensor multiplication remain those of Z₂. The map relating threefold bracketings gains a sign when a=b=c=1. The sign satisfies the 3-cocycle pentagon but is not a normalized coboundary, so it changes the associator class without changing any fusion coefficient.

A teaching route from integer counts to coherent structure

Begin with τ²=1+τ and ask for the next four powers. Then ask two separate questions about τ⁶: how many unit-output channels occur, and what multiplicative dimension does the whole object have? Requiring both answers prevents the golden ratio from being confused with a basis count.

Next move to Ising and reconstruct its matrices from the rules rather than copying them. Use one Verlinde calculation to recover a known coefficient. The learner should identify which conventions made the cancellation work.

Finish with the Z₂ associator twist. It gives a small, explicit reason that a multiplication table is not the complete category. The main transferable habit is to ask which layer of data has been retained: object types, multiplicities, maps between bracketings, exchange maps, or modular traces.

Sources and further study

[1] Pavel Etingof, Shlomo Gelaki, Dmitri Nikshych and Viktor Ostrik, Tensor Categories, for fusion categories, pointed twists, dimensions and modular categories. [2] Etingof, Nikshych and Ostrik, On Fusion Categories, for Frobenius–Perron dimension and general structural results. [3] John Preskill, Lecture Notes, Chapter 9: Topological Quantum Computation, especially the Fibonacci fusion and F-matrix calculations. [4] Eric C. Rowell and Zhenghan Wang, Mathematics of Topological Quantum Computing, for the modular-category and anyon-model framework. The tensor-power recurrences and small matrix checks above are developed explicitly with their label order and normalization stated.

Continue the prerequisites through Hopf Algebras, Quantum Groups and Tannakian Categories. Return to the BTT Mathematics Learning Hub.