Annals of Physics 303 (2003) 2–30

www.elsevier.com/locate/aop

Fault-tolerant quantum computation by anyons

A.Yu. Kitaev*
L.D. Landau Institute for Theoretical Physics, 117940, Kosygina St. 2, Germany
Received 20 May 2002

Abstract

A two-dimensional quantum system with anyonic excitations can be considered as a quan-

tum computer. Unitary transformations can be performed by moving the excitations around
each other. Measurements can be performed by joining excitations in pairs and observing the
result of fusion. Such computation is fault-tolerant by its physical nature.

2002 Elsevier Science (USA). All rights reserved.

1. Introduction

A quantum computer can provide fast solution for certain computational prob-
lems (e.g., factoring and discrete logarithm [1]) which require exponential time on
an ordinary computer. Physical realization of a quantum computer is a big challenge
for scientists. One important problem is decoherence and systematic errors in unitary
transformations which occur in real quantum systems. From the purely theoretical
point of view, this problem has been solved due to ShorÕs discovery of fault-tolerant
quantum computation [2], with subsequent improvements [3–6]. An arbitrary quan-
tum circuit can be simulated using imperfect gates, provided these gates are close to
the ideal ones up to a constant precision d. Unfortunately, the threshold value of d is
rather small;1 it is very difficult to achieve this precision.
Needless to say, classical computation can be also performed fault-
tolerantly. However, it is rarely done in practice because classical gates are reliable
enough. Why is it possible? Let us try to understand the easiest thing—why classical

*
Present address: Caltech 107-81, Pasadena, CA 91125, USA.
E-mail addresses: kitaev@itp.ac.ru, kitaev@cs.caltech.edu.
1
Actually, the threshold is not known. Estimates vary from 1/300 [7] to 10
6 [4].

0003-4916/02/$ - see front matter
2002 Elsevier Science (USA). All rights reserved.
PII: S 0 0 0 3 - 4 9 1 6 ( 0 2 ) 0 0 0 1 8 - 0
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 3

information can be stored reliably on a magnetic media. Magnetism arise from spins
of individual atoms. Each spin is quite sensitive to thermal fluctuations. But the spins
interact with each other and tend to be oriented in the same direction. If some spin
flips to the opposite direction, the interaction forces it to flip back to the direction of
other spins. This process is quite similar to the standard error correction procedure
for the repetition code. We may say that errors are being corrected at the physical
level. Can we propose something similar in the quantum case? Yes, but it is not so
simple. First of all, we need a quantum code with local stabilizer operators.
I start with a class of stabilizer quantum codes associated with lattices on the torus
and other two-dimensional surfaces [6,8]. Qubits live on the edges of the lattice
whereas the stabilizer operators correspond to the vertices and the faces. These op-
erators can be put together to make up a Hamiltonian with local interaction. (This is
a kind of penalty function; violating each stabilizer condition costs energy.) The
ground state of this Hamiltonian coincides with the protected space of the code. It
is 4g -fold degenerate, where g is the genus of the surface. The degeneracy is persistent
to local perturbation. Under small enough perturbation, the splitting of the ground
state is estimated as expð
aLÞ, where L is the smallest dimension of the lattice. This
model may be considered as a quantum memory, where stability is attained at the
physical level rather than by in explicit error correction procedure.
Excitations in this model are anyons, meaning that the global wavefunction ac-
quires some phase factor when one excitation moves around the other. One can op-
erate on the ground state space by creating an excitation pair, moving one of the
excitations around the torus, and annihilating it with the other one. Unfortunately,
such operations do not form a complete basis. It seems this problem can be removed
in a more general model (or models) where the Hilbert space of a qubit have dimen-
sionality >2. This model is related to Hopf algebras.
In the new model, we do not need torus to have degeneracy. An n-particle excited
state on the plane is already degenerate, unless the particles (excitations) come close
to each other. These particles are non-abelian anyons, i.e., the degenerate state un-
dergoes a non-trivial unitary transformation when one particle moves around the
other. Such motion (‘‘braiding’’) can be considered as fault-tolerant quantum com-
putation. A measurement of the final state can be performed by joining the particles
in pairs and observing the result of fusion.
Anyons have been studied extensively in the field-theoretic context [9–13]. So, I
hardly discover any new about their algebraic properties. However, my approach
differs in several respects:
• The model Hamiltonians are different.
• We allow a generic (but weak enough) perturbation which removes any symmetry
of the Hamiltonian.2
• The language of ribbon and local operators (see Section 6.2) provides unified de-
scription of anyonic excitations and long-range entanglement in the ground state.
An attempt to use one-dimensional anyons for quantum computation was made by
Castagnoli and Rasetti [14], but the question of fault-tolerance was not considered.

2
Some local symmetry still can be established by adding unphysical degrees of freedom, see Section 4.
4 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

Fig. 1. Square lattice on the torus.

2. Toric codes and the corresponding Hamiltonians

Consider a k  k square lattice on the torus (see Fig. 1). Let us attach a spin, or
qubit, to each edge of the lattice. (Thus, there are n ¼ 2k 2 qubits). For each vertex s
and each face p, consider operators of the following form:
Y Y
As ¼ rxj ; Bp ¼ rzj : ð1Þ
j2starðsÞ j2boundaryðpÞ

These operators commute with each other because star (s) and boundary (p) have
either 0 or 2 common edges. The operators As and Bp are Hermitian and have ei-
genvalues 1 and )1.
Let N be the Hilbert space of all n ¼ 2k 2 qubits. Define a protected subspace
L  N as follows:3
 
L ¼ jni 2 N : As jni ¼ jni; Bp jni ¼ jni for all s; p : ð2Þ
This construction gives us a definition of a quantum code TOR(k), called a toric code
[6,8]. The operators As , Bp are the stabilizer operators of this code.
To find the dimensionality of the subspace Q L, we can observe
Q that there are two
relations between the stabilizer operators, s As ¼ 1 and p Bp ¼ 1. So, there are
m ¼ 2k 2
2 independent stabilizer operators. It follows from the general theory of
additive quantum codes [15,16] that dim L ¼ 2n
m ¼ 4.
However, there is a more instructive way of computing dim L. Let us find the al-
gebra LðLÞ of all linear operators on the space L—this will give us full information
about this space. Let F  LðNÞ be the algebra of operators generate by As , Bp .
Clearly, LðLÞ ffi G=I, where G  F is the algebra of all operators which commute
with As , Bp , and I  G is the ideal generated by As
1, Bp
1. The algebra G is gen-
erated by operators of the form
Y Y
Z¼ rzj ; X ¼ rxj ;
j2c j2c0

3
We will show that this subspace is really protected from certain errors. Vectors of this subspaces are
supposed to represent ‘‘quantum information,’’ like codewords of a classical code represent classical
information.
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 5

Fig. 2. Loops on the lattice and the dual lattice.

where c is a loop (closed path) on the lattice, whereas c0 is a cut, i.e., a loop on the
dual lattice (see Fig. 2). If a loop (or a cut) is contractible then the operator Z is a
product of Bp , hence Z  1 (mod I). Thus, only non-contractible loops or cuts are
interesting. It follows that the algebra LðLÞ is generated by 4 operators Z1 , Z2 , X1 , X2
corresponding to the loops cz1 , cz2 and the cuts cx1 , cx2 (see Fig. 2). The operators Z1 ,
Z2 , X1 , X2 have the same commutation relations as rz1 , rz2 , rx1 , rx2 . We see that each
quantum state jni 2 L corresponds to a state of 2 qubits. Hence, the protected
subspace L is four-dimensional.
In a more abstract language, the algebra F corresponds to 2-boundaries and 0-
coboundaries (with coefficients from Z2 ), G corresponds to 1-cycles and 1-cocycles,
and LðLÞ corresponds to 1-homologies and 1-cohomologies.
There is also an explicit description of the protected subspace which may be not so
useful but is easier to grasp. Let us choose basis vectors in the Hilbert space N by
assigning a label zj ¼ 0, 1 to each edge j.4 The constraints Bp jni ¼ jni say that the
sum of the labels at the boundary of a face should be zero (mod 2). More exactly,
only such basis vectors contribute to a vector from the protected subspace. Such a
basis vector is characterized by two topological numbers: the sums of zj along the
loops cz1 and cz2 . The constraints Ap jni ¼ jni say that all basis vectors with the same
topological numbers enter jni with equal coefficients. Thus, for each of the 4 possible
combinations of the topological numbers v1 , v2 , there is one vector from the pro-
tected subspace,
2
X X X
jnv1 ;v2 i ¼ 2
ðk
1Þ=2 jz1 ; . . . ; zn i : z j ¼ v1 ; z j ¼ v2 : ð3Þ
z1 ;...;zn j2cz1 j2cz2

Of course, one can also create linear combinations of these vectors.

4
0 means ‘‘spin up,’’ 1 means ‘‘spin down.’’ The Pauli operators rz , rx have the standard form in this
basis.
6 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

Now we are to show that the code TOR(k) detects k
1 errors5 (hence, it corrects
bk
1
2
c errors). Consider a multiple error
Y  aj Y  bj
E ¼ rða1 ; . . . ; an ; b1 ; . . . ; bn Þ ¼ rxj rzj ðaj ; bj ¼ 0; 1Þ:
j j

This error cannot be detected by syndrome measurement (i.e., by measuring the

eigenvalues of all As , Bj Þ if and only if E 2 G. However, if E 2 F then Ejni ¼ jni for
every jni 2 L. Such an error is not an error at all—it does not affect the protected
subspace. The bad case is when E 2 G but E 62 F. Hence, the support of E should
contain a non-contractible loop or cut. It is only possible if jSuppðEÞj P k. (Here
Supp(E) is the set of j for which aj 6¼ 0 or bj 6¼ 0.)
One may say that the toric codes have quite poor parameters. Well, they are not
‘‘good’’ codes in the sense of [17]. However, the code TOR(k) corrects almost any
multiple error of size Oðk 2 Þ. (The constant factor in Oð. . .Þ is related to the percola-
tion problem.) So, these codes work if the error rate is constant but smaller than
some threshold value. The nicest property of the codes TOR(k) is that they are local
check codes. Namely,
1. Each stabilizer operators involves bounded number of qubits (at most 4).
2. Each cubit is involved in a bounded number of stabilizer operators (at most 4).
3. There is no limit for the number of errors that can be corrected.
Also, at a constant error rate, the unrecoverable error probability goes to zero as
expð
akÞ.
It has been already mentioned that error detection involves syndrome measure-
ment. To correct the error, one needs to find its characteristic vector ða1 ; . . . ; an ;
b1 ; . . . ; bn ) out of the syndrome. This is the usual error correction scheme. A new
suggestion is to perform error correction at the physical level. Consider the Hamil-
tonian
X X
H0 ¼
As
Bp : ð4Þ
s p

Diagonalizing this Hamiltonian is an easy job because the operators As , Bp com-

mute. In particular, the ground state coincides with the protected subspace of the
code TOR(k); it is 4-fold degenerate. All excited states are separated by an energy
gap DE P 2, because the difference between the eigenvalues of As or Bp equals 2.
This Hamiltonian is more or less realistic because in involves only local interac-
tions. We can expect that ‘‘errors,’’ i.e., noise-induced excitations will be removed
automatically by some relaxation processes. Of course, this requires cooling,
i.e., some coupling to a thermal bath with low temperature (in addition to the
Hamiltonian (4)).
Now let us see whether this model is stable to perturbation. (If not, there is no
practical use of it.) For example, consider a perturbation of the form

5
In the theory of quantum codes, the word ‘‘error’’ is used in a somewhat confusing manner. Here it
means a single qubit error. In most other cases, like in the formula below, it means a multiple error, i.e., an
arbitrary operator E 2 LðNÞ.
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 7
X X
V ¼
~
h rj

~ Jjp ð~ rp Þ:
rj ;~
j j<p

It is important that the perturbation is local, i.e., each term of it contains a small
number of r (at most 2). Let us estimate the energy splitting between two orthogonal
ground states of the original Hamiltonian, jni 2 L and jgi 2 L. We can use the
usual perturbation theory because the energy spectrum has a gap. In the mth order of
the perturbation theory, the splitting is proportional to hnjV m jgi or
hnjV m jni
hgjV m jgi. However, both quantities are zero unless V m contains a product
of rzj or rxj along a non-contractible loop or cut. Hence, the splitting appears only in
the dk=2eth or higher orders. As far as all things (like the number of the relevant
terms in V m ) scale correctly to the thermodynamic limit, the splitting vanishes as
expð
akÞ. A simple physical interpretation of this result is given in the next section.
(Of course, the perturbation should be small enough, or else a phase transition may
occur).
Note that our construction is not restricted to square lattices. We can consider an
arbitrary irregular lattice, like in Fig. 6. Moreover, such a lattice can be drawn on an
arbitrary two-dimensional surface. On a compact orientable surface of genus g, the
ground state is 4g -fold degenerate. In this case, the splitting of the ground state is
estimated as expð
aLÞ, where L is the smallest dimension of the lattice. We see that
the ground state degeneracy depends on the surface topology, so we deal with topo-
logical quantum order. On the other hand, there is a finite energy gap between the
ground state and excited states, so all spatial correlation functions decay exponen-
tially. This looks like a paradox—how do parts of a macroscopic system know about
the topology if all correlations are already lost at small scales? The answer is that
there is long-range entanglement6 which cannot be expressed by simple correlation
functions like hraj rbl i. This entanglement reveals itself in the excitation properties
we are going to discuss.

3. Abelian anyons

Let us classify low-energy excitations of the Hamiltonian (4). An eigenvector of

this Hamiltonian is an eigenvector of all the operators As , Bp . An elementary excita-
tion, or particle occurs if only one Q of the constraints
Q As jni ¼ jni, Bp jni ¼ jni is vio-
lated. Because of the relations s As ¼ 1 and p Bp ¼ 1, it is impossible to create
a single particle. However, it is possible to create a two-particle state of the form
jwz ðtÞi ¼ S z ðtÞjni or jwx ðt0 Þi ¼ S x ðt0 Þjni, where jni is an arbitrary ground state, and
Y Y
S z ðtÞ ¼ rzj ; S x ðt0 Þ ¼ rxj ð5Þ
j2t j2t0

(see Fig. 3). In the first case, two particles are created at the endpoints of the ‘‘string’’
(non-closed path) t. Such particles live on the vertices of the lattice. We will call them

6
Entanglement is a special, purely quantum form of correlation.
8 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

Fig. 3. Strings and particles.

z-type particles, or ‘‘electric charges.’’ Correspondingly, x-type particles, or ‘‘mag-

netic vortices’’ live on the faces. The operators S z ðtÞ, S x ðt0 Þ are called string operators.
Their characteristic property is as follows: they commute with every As and Bp , ex-
cept for few ones (namely, 2) corresponding to the endpoints of the string. Note that
the state jwz ðtÞi ¼ S z ðtÞjni depends only on the homotopy class of the path t while the
operator S z ðtÞ depends on t itself.
Any configuration of an even number of z-type particles and an even number
of x-type particles is allowed. We can connect them by strings in an arbitrary
way. Each particle configuration defines a 4g -dimensional subspace in the global
Hilbert space N. This subspace is independent of the strings but a particular vec-
tor S a1 ðt1 Þ    S am ðtm Þjni depends on t1 ; . . . ; tm . If we draw these strings in a topo-
logically different way, we get another vector in the same 4g -dimensional
subspace. Thus, the strings are unphysical but we cannot get rid of them in
our formalism.
Let us see what happens if these particles move around the torus (or other sur-
face). Moving a z-type particle along the path cz1 or cz2 (see Fig. 2) is equivalent
to applying the operator Z1 or Z2 . Thus, we can operate on the ground state space
by creating a particle pair, moving one of the particles around the torus, and anni-
hilating it with the other one. Thus, we can realize some quantum gates. Unfortu-
nately, too simple ones—we can only apply the operators rz and rx to each of the
2 (or 2g) qubits encoded in the ground state.
Now we can give the promised physical interpretation of the ground state split-
ting. In the presence of perturbation, the two-particle state jwz ðtÞi is not an eigenstate
any more. More exactly, both particles will propagate rather than stay at the same
positions. The propagation process is described by the Schr€ odinger equation with
some effective mass mz . (x-type particles have another mass mx ). In the non-perturbed
model, mz ¼ mx ¼ 1. There are no real particles in the ground state, but they can be
created and annihilate virtually. A virtual particle can tunnel around the torus before
annihilating with the other one. Such processes contribute terms bz1 Z1 , bz2 Z2 , bx1 X1 ,
bx2 X2 to the ground state effective Hamiltonian.
pffiffiffiffiffiffiffiffiffiffiffiffi
ffi Here bai  expð
aa Li Þ is the tunnel-
ing amplitude whereas aa  2mDE is the imaginary wavevector of the tunneling
particle.
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 9

Next question: what happens if we move particles around each other? (For this,
we do not need a torus; we can work on the plane). For example, let us move an
x-type particle around a z-type particle (see Fig. 4). Then
jWinitial i ¼ S z ðtÞjwx ðqÞi; jWfinal i ¼ S x ðcÞS z ðtÞjwx ðqÞi ¼
jWinitial i;
because S x ðcÞ and S z ðtÞ anti-commute, and S x ðcÞjwx ðqÞi ¼ jwx ðqÞi. We see that the
global wavefunction ( ¼ the state of the entire system) acquires the phase factor )1. It
is quite unlike usual particles, bosons and fermions, which do not change their phase
in such a process. Particles with this unusual property are called abelian anyons.
More generally, abelian anyons are particles which realize non-trivial one-dimen-
sional representations of (colored) braid groups. In our case, the phase change can be
also interpreted as an Aharonov–Bohm effect. It does not occur if both particles are
of the same type.
Note that abelian anyons exist in real solid state systems, namely, they are intrin-
sicly related to the fractional quantum Hall effect [18]. However, these anyons have
different braiding properties. In the fractional quantum Hall system with filling fac-
tor p=q, there is only one basic type of anyonic particles with (real) electric charge
1=q. (Other particles are thought to be composed from these ones). When one par-
ticle moves around the other, the wavefunction acquires a phase factor expð2pi=qÞ.
Clearly, the existence of anyons and the ground state degeneracy have the same
nature. They both are manifestations of a topological quantum order, a hidden
long-range order that cannot be described by any local order parameter. (The exis-
tence of a local order parameter contradicts the nature of a quantum code—if the
ground state is accessible to local measurements then it is not protected from local
errors.) It seems that the anyons are more fundamental and can be used as a univer-
sal probe for this hidden order. Indeed, the ground state degeneracy on the torus fol-
lows from the existence of anyons [19]. Here is the original EinarssonÕs proof applied
to our two types of particles.
We derived the ground state degeneracy from the commutation relations between
the operators Z1 , Z2 , X1 , X2 . These operators can be realized by moving particles

Fig. 4. An x-type particle moving around a z-type particle.

10 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

Fig. 5. A fly-over crossing geometry for a two-dimensional electron layer.

along the loops cz1 , cz2 , cx1 , cx2 . These loops only exist on the torus, not on the plane.
Consider, however, the process in which an x-type particle and a z-type particle go
around the torus and then trace their paths backward. This corresponds to the op-
erator W ¼ X1
1 Z1
1 X1 Z1 which can be realized on the plane. Indeed, we can deform
particlesÕ trajectories so that one particle stays at rest while the other going around
it. Due to the anyonic nature of the particles, W ¼
1. We see that X1 and Z1 anti-
commute.
The above argument is also applicable to the fractional quantum Hall anyons [19].
The ground state on the torus is q-fold degenerate, up to the precision  expð
L=l0 Þ,
where l0 is the magnetic length. This result does not rely on the magnetic transla-
tional symmetry or any other symmetry. Rather, it relies on the existence of the en-
ergy gap in the spectrum (otherwise the degeneracy would be unstable to
perturbation). Note that holes ( ¼ punctures) in the torus do not remove the degen-
eracy unless they break the non-trivial loops cx1 , cx2 , cz1 , cz2 . The fly-over crossing
geometry (see Fig. 5) is topologically equivalent to a torus with 2 holes, but it is al-
most flat. In principle, such structure can be manufactured,7 cooled down and placed
into a perpendicular magnetic field. This will be a sort of quantum memory—it will
store a quantum state forever, provided all anyonic excitation are frozen out or lo-
calized. Unfortunately, I do not know any way this quantum information can get in
or out. Too few things can be done by moving abelian anyons. All other imaginable
ways of accessing the ground state are uncontrollable.

4. Materialized symmetry: is that a miracle?

Anyons have been studied extensively in the gauge field theory context [9–11,13].
However, we start with quite different assumptions about the Hamiltonian. A gauge
theory implies a gauge symmetry which cannot be removed by external perturbation.

7
It is not easy. How will the two layers (the two crossing ‘‘roads,’’ one above the other) join in a single
crystal layout?
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 11

To the contrary, our model is stable to arbitrary local perturbations. It is useful to

give a field-theoretic interpretation of this model. The edge labels zj (measurable
by rzj ) correspond to a Z2 vector potential, whereas rxj corresponds to the electric
field. The operators As are local gauge transformations whereas Bp is the magnetic
field on the face p. The constraints As jni ¼ jni mean that the state jni is gauge-invari-
ant. Violating the gauge invariance is energetically unfavorable but not forbidden.
The Hamiltonian (which includes H0 and some perturbation V) need not obey the
gauge symmetry. The constraints Bp jni ¼ jni mean that the gauge field corresponds
to a flat connection. These constraints are not strict either.
Despite the absence of symmetry in the Hamiltonian H ¼ H0 þ V , our system ex-
hibits two conservation laws: electric charge and magnetic charge (i.e., the number of
vortices) are both conserved modulo 2. In the usual electrodynamics, conservation of
electric charge is related to the local ( ¼ gauge) U(1) symmetry. In our case, it should
be a local Z2 symmetry for electric charges and another Z2 symmetry for magnetic
vortices. So, our system exhibits a dynamically created Z2  Z2 symmetry which ap-
pears only at large distances where individual excitations are well-defined.
Probably, the reader is not satisfied with this interpretation. Really, it creates a
new puzzle rather then solve an old one. What is this mysterious symmetry? How
do symmetry operators look like at the microscopic level? The answer sounds as non-
sense but it is true. This symmetry (as well as any other local symmetry) can be found
in any Hamiltonian if we introduce some unphysical degrees of freedom. So, the
symmetry is not actually being created. Rather, an artificial symmetry becomes a
real one.
The new degrees of freedom are spin variables vs , wp ¼ 0; 1 for each vertex s and
each face p. The vertex spins will stay in the state 2
1=2 ðj0i þ j1iÞ whereas the face
spins will stay in the state j0i. So, all the extra spins together stay in a unique quan-
tum state jfi. Obviously, rxs jfi ¼ jfi and rzp jfi ¼ jfi, for every vertex s and every face
p. From the mathematical point of view, we have simply defined an embedding of the
space N into a larger Hilbert space T of all the spins, jwi 7! jwi  jfi. So we may
write N  T. We will call N the physical space (or subspace), T the extended space.
Physical states (i.e., vectors jwi 2 N) are characterized by the equations
rxs jwi ¼ jwi; rzp jwi ¼ jwi
for every vertex s and face p.
Now let us apply a certain unitary transformation U on the extended space T.
This transformation is just a change of the spin variables, namely
X X
vs 7! vs ; zj 7! zj þ vs ; wp 7! wp þ zj ð6Þ
s¼endpointðjÞ j2boundaryðpÞ

(all sums are taken modulo 2). The physical subspace becomes N0 ¼ U N. Vectors
jwi 2 N0 are invariant under the following symmetry operators:
Ps ¼ U rxs U y ¼ rxs As ; Qp ¼ U rzp U y ¼ rzp Bp : ð7Þ
The transformed Hamiltonian H 0 ¼ UHU y commutes with these operators. It is
defined up to the equivalence Ps  1, Qp  1. In particular,
12 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30
X X
H00 ¼ UH0 U y ¼ H00 
rxs
rzp : ð8Þ
s p

In the field theory language, the vertex variables vs (or the operators rzs ) are a Higgs
field. The operators Ps are local gauge transformations. Thus, an arbitrary Hamil-
tonian can be written in a gauge-invariant form if we introduce additional Higgs
fields. Of course, it is a very simple observation. The real problem is to understand
how the artificial gauge symmetry ‘‘materialize,’’ i.e., give rise to a physical con-
servation law.
Electric charge at a vertex s is givenQby the operator rxs . The total electric charge
on a compact surface is zero8 because s rxs  1. This is not a physically meaningful
statement as it is. It is only meaningful if there are discrete charged particles. Then
the charge is also conserved locally, in every scattering or fusion process. It is difficult
to formulate this property in a mathematical language, but, hopefully, it is possible.
(The problem is that particles are generally smeared and can propagate. Physically,
particles are well-defined if they are stable and have finite energy gap.) Alternatively,
one can use various local and non-local order parameters to distinguish between
phases with an unbroken symmetry, broken symmetry, or confinement.
The artificial gauge symmetry materialize for the Hamiltonian (8) but this is not
the case for every Hamiltonian. Let us try to describe possible symmetry breaking
mechanisms in terms of local order parameters. If the gauge symmetry is broken then
there is a non-vanishing vacuum average of the Higgs field, /ðsÞ ¼ hrzs i 6¼ 0. Electric
charge is not conserved any more. In other words, there is a Bose condensate of
charged particles. Although the second Z2 symmetry is formally unbroken, free mag-
netic vortices do not exist. More exactly, magnetic vortices are confined. (The duality
between symmetry breaking and confinement is well known [25]). It is also possible
that the second symmetry is broken, then electric charges are confined. From the
physical point of view, these two possibilities are equivalent: there is no conservation
law in the system.9
An interesting question is whether magnetic vortices can be confined without the
gauge symmetry being broken. Apparently, the answer is ‘‘no.’’ The consequence is
significant: electric charges and magnetic vortices cannot exist without each other. It
seems that materialized symmetry needs better understanding; as presented here, it
looks more like a miracle.

5. The model based on a group algebra

From now on, we are constructing and studying non-abelian anyons which will
allow universal quantum computation.

8
Strictly speaking, the electric charge is not a numeric quantity; rather, it is an irreducible
representation of the group Z2 . ‘‘Zero’’ refers to the identity representation.
9
The two possibilities only differ if an already materialized symmetry breaks down at much large
distances (lower energies).
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 13

Let G be a finite (generally, non-abelian) group. Denote by H ¼ C½G the corre-

sponding group algebra, i.e., the space of formal linear combinations of group ele-
ments with complex coefficients. We can consider H as a Hilbert space with a
standard orthonormal basis fjgi : g 2 Gg. The dimensionality of this space is
N ¼ jGj. We will work with ‘‘spins’’ (or ‘‘qubits’’) taking values in this space.10 Re-
mark: This model can be generalized. One can take for H any finite-dimensional
Hopf algebra equipped with a Hermitian scalar product with certain properties.
However, I do not want to make things too complicated.
To describe the model, we need to define 4 types of linear operators, Lgþ , Lg
, Tþh , T
h
acting on the space H. Within each type, they are indexed by group elements, g 2 G
or h 2 G. They act as follows:
Lgþ jzi ¼ jgzi; Tþh jzi ¼ dh;z jzi;
ð9Þ
Lg
jzi ¼ jzg
1 i; T
h jzi ¼ dh
1 ;z jzi:
(In the Hopf algebra context, the operators Lgþ , Lg
, Tþh , T
h correspond to the left and
right multiplications and left and right comultiplication, respectively.) These oper-
ators satisfy the following commutation relations:

1
Lgþ Tþh ¼ Tþgh Lgþ ; Lgþ T
h ¼ T
hg Lgþ ;

1
ð10Þ
Lg
Tþh ¼ Tþhg Lg
; Lg
T
h ¼ T
gh Lg
:
Now consider an arbitrary lattice on an arbitrary orientable two-dimensional surface,
see Fig. 6. (We will mostly work with a plane or a sphere, not higher genus surfaces).
Corresponding to each edge is a spin which takes values in the space H. Arrows in Fig.
6 mean that we choose some orientation for each edge of the lattice. (Changing the
direction of a particular arrow will be equivalent to the basis change jzi 7! jz
1 i for the
corresponding qubit). Let j be an edge of the lattice, s one of its endpoints. Define an
operator Lg ðj; sÞ ¼ Lg" ðjÞ as follows. If s is the origin of the arrow j then Lg ðj; sÞ is Lg
ðjÞ
(i.e., Lg
acting on the jth spin), otherwise it is Lgþ ðjÞ. This rule is represented by the
diagram at the right side of Fig. 6. Similarly, if p is the left (the right) adjacent face of
the edge j then T h ðj; pÞ is T
h (resp. Tþh ) acting on the jth spin.
Using these notations, we can define local gauge transformations and magnetic
charge operators corresponding to a vertex s and an adjacent face p (see Fig. 6). Put
Y
Ag ðs; pÞ ¼ Ag ðsÞ ¼ Lg ðj; sÞ;
j2starðsÞ

X Y
k ð11Þ
Bh ðs; pÞ ¼ T hm ðjm ; pÞ;
h1 hk ¼h m¼1

where j1 ; . . . ; jk are the boundary edges of p listed in the counterclockwise order,

starting from, and ending at, the vertex s. (The sum is taken over all combinations of
h1 ; . . . ; hk 2 G, such that h1    hk ¼ h. Order is important here!.) Although Ag ðs; pÞ

10
In the field theory language, the value of a spin can be interpreted as a G gauge field. However, we
do not perform symmetrization over gauge transformations.
14 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

Fig. 6. Generic lattice and the orientation rules for the operators Lg" and T"h .

does not depend on p, we retain this parameter to emphasize the duality between
Ag ðs; pÞ and Bh ðs; pÞ.11 These operators generate an algebra D ¼ DðGÞ, DrinfieldÕs
quantum double [20] of the group algebra C½G. It will play a very important role
below. Now we only need two symmetric combinations of Ag ðs; pÞ and Bh ðs; pÞ, namely
X
AðsÞ ¼ N
1 Ag ðs; pÞ; BðpÞ ¼ B1 ðs; pÞ; ð12Þ
g2G

where N ¼ jGj. Both AðsÞ and BðpÞ are projection operators. (AðsÞ projects out the
states which are gauge invariant at s, whereas BðpÞ projects out the states with
vanishing magnetic charge at p). The operators AðsÞ and BðpÞ commute with each
other.12 Also AðsÞ commutes with Aðs0 Þ, and BðpÞ commutes with Bðp0 Þ for different
vertices and faces. In the case G ¼ Z2 , these operators are almost the same as the
operators (1), namely AðsÞ ¼ 12 ðAs þ 1Þ, BðpÞ ¼ 12 ðBp þ 1Þ.13
At this point, we have only defined the global Hilbert space N (the tensor product
of many copies of H) and some operators on it. Now let us define the Hamiltonian.
X X
H0 ¼ ð1
AðsÞÞ þ ð1
BðpÞÞ: ð13Þ
s p

It is quite similar to the Hamiltonian (4). As in that case, the space of ground states is
given by the formula
L ¼ fjni 2 N : AðsÞjni ¼ jni; BðpÞjni ¼ jni for all s; pg: ð14Þ
The corresponding energy is 0; all excited states have energies P 1.
It is easy to work out an explicit representation of ground states similar to Eq. (3).
The ground states correspond 1-to-1 to flat G-connections, defined up to conjuga-
tion, or super-positions of those. So, the ground state on a sphere is not degenerate.
However, particles (excitations) have quite interesting properties even on the sphere
or on the plane. (We treat the plane as an infinitely large sphere). The reader prob-
ably wants to know the answer first, and then follow formal calculations. So, I give a

11
In the Hopf algebra setting, Ag ðs; pÞ does depend on p.
12
This is not obvious. Use the commutation relations (10) to verify this statement.
13
Here As and Bp are the notations from Section 2; we will not use them any more.
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 15

brief abstract description of these particles. It is a mixture of general arguments and

details which require verification.
The particles live on vertices or faces, or both; in general, one particle occupies a
vertex and an adjacent face same time. A combination of a vertex and an adjacent
face will be called a site. Sites are represented by dotted lines in Fig. 7. (The dashed
lines are edges of the dual lattice.)
Consider n particles on the sphere pinned to particular sites x1 ; . . . ; xn at large dis-
tances from each other. The space L½n ¼ Lðx1 ; . . . ; xn Þ of n-particle states has dimen-
sionality N 2ðn
1Þ , including the ground state.14 Not all these states have the same
energy. Even more splitting occurs under perturbation, but some degeneracy still sur-
vive. Of course, we assume that the perturbation is local, i.e., it can be represented by a
sum of operators each of which acts only on few spins. To find the residual degeneracy,
we will study the action of such local operators on the space L½n. Local operators gen-
erate a subalgebra P½n  LðL½nÞ. Elements of its center, C½n, are conserved classical
quantities; they can be measured once and never change. (More exactly, they cannot
be changed by local operators.) As these classical variables are locally measurable, we
interpret them as particleÕs types. It turns out that the types correspond 1-to-1 to irre-
ducible representations of the algebra D, the quantum double. Thus, each particle can
belong to one of these types. The space L½n and the algebra P½n split accordingly:
L½n ¼ # Ld1 ;...;dn ; P½n ¼ # Pd1 ;...;dn ðPd1 ;...;dn  LðLd1 ;...;dn ÞÞ; ð15Þ
d1 ;...;dn d1 ;...;dn

where dm is the type of the mth particle. The ‘‘classical’’ subalgebra C½n is generated
by the projectors onto Ld1 ;...;dn .
But this is not the whole story. The subspace Ld1 ;...;dn splits under local perturba-
tions from Pd1 ;...;dn . By a general mathematical argument,15 this algebra can be char-
acterized as follows:
Ld1 ;...;dn ¼ Kd1 ;...;dn  Md1 ;...;dn ; Pd1 ;...;dn ¼ LðKd1 ;...;dn Þ: ð16Þ
The space Kd1 ;...;dn corresponds to local degrees of freedom. They can be defined
independently for each particle. So, Kd1 ;...;dn ¼ Kd1      Kdn , where Kdm is the
space of ‘‘subtypes’’ (internal states) of the mth particle. Like the type, the subtype of
a particle is accessible by local measurements. However, it can be changed, while the
type cannot.
The most interesting thing is the protected space Md1 ;...;dn . It is not accessible by
local measurements and is not sensitive to local perturbations, unless the particles
come close to each other. This is an ideal place to store quantum information and
operate with it. Unfortunately, the protected space does not have tensor product
structure. However, it can be described as follows. Associated with each particle type
a is an irreducible representation Ud of the quantum double D. Consider the product
representation Ud1      Udn and split it into components corresponding to differ-
ent irreducible representations. The protected space is the component corresponding
to the identity representation.
14
The absence of particle at a given site is regarded as a particle of special type.
15
Pd1 ;...;dn is a subalgebra of LðLd1 ;...;dn Þ with a trivial center, closed under Hermitian conjugation.
16 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

Fig. 7. A ribbon on the lattice.

If we swap two particles or move one around the other, the protected space under-
goes some unitary transformation. Thus, the particles realize some multi-dimensional
representation of the braid group. Such particles are called non-abelian anyons. Note
that braiding does not affect the local degrees of freedom. If two particles fuse, they
can annihilate or become another particle. The protected space becomes smaller but
some classical information comes out, namely, the type of the new particle. So, the
we can do measurements on the protected space. Finally, if we create a new pair of par-
ticles of definite types, it always comes in a particular quantum state. So, we have a
standard toolkit for quantum computation (new states, unitary transformations
and measurements), except that the Hilbert space does not have tensor product struc-
ture. Universality of this toolkit is a separate problem, see Section 8.
Our model gives rise to the same braiding and fusion rules as gauge field theory mod-
els [10,11]. The existence of local degrees of freedom (subtypes) is a new feature. These
degrees of freedom appear because there is no explicit gauge symmetry in our model.

6. Algebraic structure

6.1. Particles and local operators

This subsection is also rather abstract but the claims we do are concrete. They will
be proven in Section 6.4.
As mentioned above, the ground state of the Hamiltonian (13) is not degenerate
(on the sphere or on the plane regarded as an infinitely large sphere). Excited states
are characterized by their energies. The energy of an eigenstate jwi is equal to the
number of constraints ðAðsÞ
1Þjwi ¼ 0 or ðBðpÞ
1Þjwi ¼ 0 which are violated.
Complete classification of excited states is a difficult problem. Instead of that, we will
try to classify elementary excitations, or particles.
Let us formulate the problem more precisely. Consider a few excited ‘‘spots’’ sep-
arated by large distances. Each spot is a small region where some of the constrains are
violated. The energy of a spot can be decreased by local operators but, generally, the
spot cannot disappear. Rather, it shrinks to some minimal excitation (which need not
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 17

be unique). We will see (in Section 6.4) that any excited spot can be transformed into
an excitation which violates at most 2 constraints, AðsÞ
1  0 and BðpÞ
1  0,
where s is an arbitrary vertex, and p is an adjacent face. Such excitations are be called
elementary excitations, or particles. Note that definition of elementary excitations is a
matter of choice. We could decide that an elementary excitation violates 3 constraints.
Even with our definition, the ‘‘space of elementary excitations’’ is redundant.
By the way, the space of elementary excitations is not well-defined because such
an excitation does not exist alone. More exactly, the only one-particle state on the
sphere is the ground state. (This can be proven easily.) The right thing is the space
of two-particle excitations, Lða; bÞ. Here a ¼ ðs; pÞ and b ¼ ðs0 ; p0 Þ are the sites occu-
pied by the particles. (Recall that a site is a combination
Q of a vertex
Q and an adjacent
face.) The projector onto Lða; bÞ can be written r6¼s;s0 AðrÞ l6¼p;p0 BðlÞ. Note that
introducing a third particle (say, c) will not give more freedom for any of the two.
Indeed, b and c can fuse without any effect on a.
Let us see how local operators act on the space Lða; bÞ. In this context, a local
operator is an operator which acts only on spins near a (or near b). Besides that,
it should preserve the subspace Lða; bÞ  N and its orthogonal complement. (N
is the space of all quantum states). Example: the operators Ag ðaÞ and Bh ðaÞ, where
a ¼ ðs; pÞ, commute with AðrÞ, BðlÞ for all r 6¼ s and l 6¼ p. Hence, they commute
with the projector onto the subspace Lða; bÞ. These operators generate an algebra
DðaÞ  LðNÞ. It will be shown in Section 6.4 that DðaÞ includes all local operators
acting on the space Lða; bÞ, and the action of DðaÞ on Lða; bÞ is exact (i.e., different
operators act differently).
Actually, the algebra DðaÞ ¼ D does not depend on a, only the embedding
D ! LðNÞ does. This algebra is called the quantum double of the group G and de-
noted by DðGÞ. Its structure is determined by the following relations between the op-
erators Ag ¼ Ag ðaÞ and Bh ¼ Bh ðaÞ:
Af Ag ¼ Afg ; Bh Bi ¼ dh;i Bh ; Ag Bh ¼ Bghg
1 Ag : ð17Þ
The operators Dðh;gÞ ¼ Bh Ag form a linear basis of D. (In [10,11] these operators were
.

denoted by h g .) The following multiplication rules hold

Dðh1 ;g1 Þ Dðh2 ;g2 Þ ¼ dh1 ;g1 h2 ;g
1 Dðh1 ;g1 g2 Þ :
1

This identity can be also written in a symbolic tensor form, with h and g being
combined into one index:
ðh;gÞ
Dm Dn ¼ Xkm n Dk Xðh1 ;g1 Þðh2 ;g2 Þ ¼ dh1 ;g1 h2 ;g
1 dh;h1 dg;g1 g2 ð18Þ
1

(summation over k is implied). Actually, D is not only an algebra, it is a quasi-tri-

angular Hopf algebra, see Sections 6.2 and 6.3.
Note that D ¼ DðaÞ is closed under Hermitian conjugation (in LðNÞÞ which acts
as follows:
Ayg ¼ Ag
1 ; Byh ¼ Bh ; Dyðh;gÞ ¼ Dðg
1 hg;g
1 Þ : ð19Þ

Thus, D ¼ DðaÞ is a finite-dimensional C $ -algebra. Hence it has the following

structure:
18 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

D ¼ #LðKd Þ; ð20Þ
d

where d runs over all irreducible representations of D. We can interpret d as particleÕs

type.16 The absence of particle corresponds to a certain one-dimensional represen-
tation called the identity representation. More exactly, the operators Dðh;gÞ act on the
ground state jni as follows:
Dk jni ¼ ek jni; where eðh;gÞ ¼ dh;1 : ð21Þ

The ‘‘space of subtypes,’’ Kd actually characterize the redundancy of our definition

of elementary excitations. However, this redundancy is necessary to have a nice
theory of ribbon operators (see Section 6.2).
Irreducible representations of D can be described as follows [10]. Let u 2 G be an
arbitrary element, C ¼ fgug
1 : g 2 Gg its conjugacy class, E ¼ fg 2 G : gu ¼ ugg its
centralizer. There is one irreducible representation d ¼ ðC; vÞ for each conjugacy
class C and each irreducible representation v of the group E (see below). It does
not matter which element u 2 C is used to define E. The conjugacy class C can be
interpreted as magnetic charge whereas v corresponds to electric charge. For exam-
ple, consider the group S3 (the permutation group of order 3). It has 3 conjugacy
classes of order 1, 2, and 3, respectively. So, the algebra DðS3 Þ has irreducible repre-
sentations of dimensionalities 1, 1, 2; 2, 2, 2; 3, 3.
The simplest case is when v is the identity representation, i.e., the particle carries
only magnetic charge but no electric charge. Then the subtypes can be identified with
the elements of C, i.e., the corresponding space (denoted by BC ) has a basis
fjvi : v 2 Cg. The local operators act on this space as follows:
Dðh;gÞ jvi ¼ dh;gvg
1 jgvg
1 i: ð22Þ
ðvÞ
Now consider the general case. Denote by Wf ¼ the irreducible action of f 2 E on
Wf
an appropriate space Av . Choose arbitrary elements qv 2 G such that qv uq
1
v ¼ v for
each v 2 C. Then any element g 2 G can be uniquely represented in the form g ¼ qv f ,
where v 2 C and f 2 E. We can define a unique action of D on BC  Av , such that
Bh ðjvi  jgiÞ ¼ dh;v jvi  jgi;
ð23Þ
Aqvf ðjui  jgiÞ ¼ jvi  Wf jgi ðv 2 C; f 2 EÞ:
More generally, Dðh;gÞ ðjvi  jgiÞ ¼ dh;gvg
1 jgvg
1 i  Wf jgi, where f ¼ qv ðqgvg
1 Þ
1 g.
This action is irreducible.

6.2. Ribbon operators

The next task is to construct a set of operators which can create an arbitrary two-
particle state from the ground state. I do not know how to deduce an expression

16
Caution. The local operators should not be interpreted as symmetry transformations. The true
symmetry transformations, so-called topological operators, will be defined in Section 7. Mathematically,
they are described by the same algebra D, but their action on physical states is different.
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 19

for such operators; I will just give an answer and explain why it is correct. In
the abelian case (see Section 3) there were two types of such operators which
corresponded to paths on the lattice and the dual, lattice, respectively. In the
non-abelian case, we have to consider both types of paths together. Thus, the
operators creating a particle pair are associated with a ribbon (see Fig. 7). The
ribbon connects two sites at which the particles will appear (say, a ¼ ðs; pÞ
and b ¼ ðs0 ; p0 Þ). The corresponding operators act on the edges which constitute
one side of the ribbon (solid line), as well as the edges intersected by the other
side (dashed line).
For a given ribbon t, there are N 2 ribbon operators F ðh;gÞ ðtÞ indexed by g; h 2 G.
They act as follows:17

ð24Þ

These operators commute with every projector AðrÞ, BðlÞ, except for r ¼ s, s0 and
l ¼ p, p0 . This is the first important property of ribbon operators.
The operators F ðh;gÞ ðtÞ depend on the ribbon t. However, their action on the space
Lða; bÞ depends only on the topological class of the ribbon This is also true for a
multi-particle excitation space Lðx1 ; . . . ; xn Þ. More exactly, consider two ribbons, t
and q, connecting the sites x1 ¼ a and x2 ¼ b. The actions of F ðh;gÞ ðtÞ and F ðh;gÞ ðqÞ
on Lðx1 ; . . . ; xn Þ coincide provided none of the

17
Horizontal and vertical arrows are the two types of edges. Each of the two diagrams (6 arrows with
labels) stand for a particular basis vector.
20 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

sites, x3 ; . . . ; xn lie on or between the ribbons. This is the second important property
of ribbon operators. We will write F ðh;gÞ ðtÞ  F ðh;gÞ ðqÞ, or, more exactly,
M ðh;gÞ
F ðh;gÞ ðtÞ ðqÞ, where M ¼ fx1 ; . . . ; xn g.
Linear combination of the operators F ðh;gÞ ðtÞ are also called ribbon operators.
They form an algebra FðtÞ ffi F. The multiplication rules are as follows:
ðh ;g Þðh2 ;g2 Þ
F m ðtÞF n ðtÞ ¼ Kmk n F k ðtÞ; 1 1
Kðh;gÞ ¼ dh1 h2 ;h dg1 ;g dg2 ;g ð25Þ
(summation over m and n is implied).
Any ribbon operator on a long ribbon t ¼ t1 t2 (see figure below) can be repre-
sented in terms of ribbon operators corresponding to its parts, t1 and t2

ðh;gÞ
F k ðt1 t2 Þ ¼ Xkm n F m ðt1 ÞF n ðt2 Þ; Xðh1 ;g1 Þðh2 ;g2 Þ ¼ dg;g1 g2 dh1 ;h dh2 ;g
1 hg1 ð26Þ
1

(Note that F m ðt1 Þ and F n ðt2 Þ commute because the ribbons t1 and t2 do not overlap.)
By some miracle, the tensor X$$$ is the same as in e.g., (18). From the mathematical
point of view, (26) defines a linear mapping Dðt1 ; t2 Þ : Fðt1 ; t2 Þ ! Fðt1 Þ  Fðt2 Þ, or
just D : F ! F. Such a mapping is called a comultiplication.
The comultiplication rules (26) allow to give another definition of ribbon opera-
tors which is nicer than Eq. (24). Note that a ribbon consists of triangles of two types
(see Fig. 7). Each triangle corresponds to one edge. More exactly, a triangle with two
dotted sides and one dashed side corresponds to a combination of an edge and its
endpoint, say, i and r. Similarly, a triangle with a solid side corresponds to a com-
bination of an edge and one of the adjacent faces, say, j and l. Each triangle can
be considered as a short ribbon. The corresponding ribbon operators are

1
F ðh;gÞ ði; rÞ ¼ dg;1 Lh ði; rÞ; F ðh;gÞ ðj; lÞ ¼ T g ðj; lÞ:
The ribbon operators on a long ribbon can be constructed from these ones.
It has been already mentioned that the multiplication in D and the comultiplica-
tion in F are defined by the same tensor X$$$ . Actually, D and F are Hopf algebras
dual to each other.
(For general account on Hopf algebras, see [21–23].) The multiplication in F cor-
responds to a comultiplication in D defined as follows:
DðDk Þ ¼ Kmk n Dm  Dn : ð27Þ
P
(More explicitly, DðDðh;gÞ Þ ¼ h1 h2 ¼h Dðh1 ;gÞ  Dðh2 ;gÞ .) The unit element of F is
1F ¼ ek F k , where ek are given by (21); the tensor e$ also defines a counit of D (i.e., the
mapping e : D ! C : eðDk Þ ¼ ek ). The unit of D and the counit of F are given by
eðh;gÞ ¼ dg;1 : ð28Þ
The Hopf algebra structure also includes an antipode, i.e., a mapping S : D ! D :
$
SðDk Þ ¼ Skm Dm , or S : F ! F : SðF m Þ ¼ Skm F k . The tensor S$$ is given by the equation
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 21

ðh ;g Þ
Sðh21;g21Þ ¼ dg
1 h1 g1 ;h
1 dg1 ;g
1 : ð29Þ
1 2 2

Here is the complete list of Hopf algebra axioms.

lj m n mj m
Klm in
i Kk ¼ Kk Kj ; ei Kim
k ¼ Kk ej ¼ dk ; ð30Þ

Xilm Xkin ¼ Xklj Xjm n ; ei Xkim ¼ Xkm j ej ¼ dkm ; ð31Þ

q
Klm l m ir js
q Xkn ¼ Xij Xrs Kk Kn ; eq Xqkn ¼ ek en ; Klm q l m
q e ¼ e e ; ð32Þ

q n k lm q n
Slk Klm q
p Xkn dm ¼ dl Kp Xkn Sm ¼ ep e : ð33Þ
Most of these identities correspond to physically obvious properties of ribbon op-
erators. Eq. (30) is a statement of the usual multiplication axioms in the algebra F,
namely, ðF l F m ÞF n ¼ F l ðF m F n Þ and 1F m ¼ F m 1 ¼ F m . The first equation in (31)
(coassociativity of the comultiplication in F) can be proven by expanding F k ðt1 t2 t3 Þ
as Xkin F i ðt1 t2 ÞF n ðt3 Þ or Xklj F l ðt1 ÞF j ðt2 t3 Þ—the result must be the same.18 Equation (32)
mean that the multiplication and comultiplication are consistent with each other. To
prove the first equation in (32), expand F l ðt1 t2 ÞF m ðt1 t2 Þ in two different ways. The
second equation follows from the fact that eq F q ðt1 t2 Þ is the identity operator.
The antipode axiom (33) does not have explicit physical meaning. Mathemati-
cally, it is a definition of the tensor S$$ : the element c ¼ Skl F k  Dl 2 F  D is the in-
verse to the canonical element d ¼ F i  Di . The antipode have the following
properties which can be derived from (30)–(33):
Sli Smj Klm ji q
p ¼ K q Sp ; Sqp Xqij ¼ Xpm l Sjm Sil : ð34Þ
Finally, we can define a so-called skew antipode S~$$ as follows:
S~im Sni ¼ Sjm S~nj ¼ dmn : ð35Þ
In our case, S~im ¼ Sim , but this is not true for a generic Hopf algebra. The skew an-
tipode have the following properties similar to (33) and (34):
S~ln Klm q k n lm q ~k q
p Xkn dm ¼ dl Kp Xkn Sm ¼ ep e ; ð36Þ

S~li S~mj Klm ji ~q

p ¼ K l Sp ; S~qp Xqij ¼ Xpm l S~jm S~il ð37Þ
(Note the distinction between (33) and (36), however.)
The reader may be overwhelmed by a number of formal things, so let us summa-
rize what we know by now. We have denned two algebras, D and F, and their ac-
tions on the Hilbert space N. In this context, we denote them by DðaÞ and FðtÞ
because the actions depend on the site a or on the ribbon t, respectively. Operators
from DðaÞ affect one particle whereas operators from FðtÞ affect two particles. The

18
The coassociativity is necessary and sufficient for that. The sufficiency is rather obvious; the necessity
follows from the fact that the mapping F ! FðtÞ is injective, i.e., the operators F k ðtÞ with different k are
linearly independent.
22 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

action of FðtÞ on the space of n-particle states Lðx1 ; . . . ; xn Þ depends only on the to-
pological class of the ribbon t. This space have not been found yet, even for n ¼ 2. (It
will be found after we learn more about local and ribbon operators.) The algebra F
is a Hopf algebra. The comultiplication allows to make up a long ribbon from parts.
There is a formal duality between F and D. The comultiplication in F is dual to the
multiplication in D. The multiplication in F is dual to a comultiplication in D. (The
meaning of the latter is not clear yet.)

6.3. Further properties of local and ribbon operators

Let us study commutation relations between ribbon operators. Consider two rib-
bons attached to the same site, as shown in Fig. 8a or b. Then

1 hv;v
1 gÞ
F ðh;gÞ ðt1 ÞF ðv;uÞ ðq1 Þ ¼ F ðv;uÞ ðq1 ÞF ðv ðt1 Þ;
ðh;gu
1 vuÞ
F ðh;gÞ ðt2 ÞF ðv;uÞ ðq2 Þ ¼ F ðv;uÞ ðq2 ÞF ðt2 Þ:
In a tensor form, these equations read as follows:
F m ðt1 ÞF n ðq1 Þ ¼ Rik Xnij Xmkl F j ðq1 ÞF l ðt1 Þ;
ð38Þ
F m ðt2 ÞF n ðq2 Þ ¼ F i ðq2 ÞF k ðt2 ÞXn Xm Rjl ;
ij kl

where
Rðh;gÞðv;uÞ ¼ dh;u dg;1 ; Rðh;gÞðv;uÞ ¼ dh
1 ;u dg;1 : ð39Þ
Note that
Rik Xnij Xmkl Rjl ¼ Rik Xnij Xmkl Rjl ¼ en em : ð40Þ
19
To prove (and to sec the physical meaning of) this equation, consider the config-
uration shown in Fig. 9b. Clearly, F r ðt2 t1 Þ and F s ðq0 Þ commute. On the other hand.
F s ðq0 Þ  F s ðq2 q1 Þ, so F r ðt2 t1 Þ and F s ðq2 q1 Þ commute. It follows that Rik Xnij Xmkl ¼ en em .
This identity can be easily written in an invariant form, namely, RR ¼ lDD , where
R ¼ Rjl Dj  Dl and R ¼ Rik Di  Dk . It also implies that RR ¼ lDD because the
algebra D  D is finite dimensional. Thus. R ¼ R
1 .
The tensor R$$ (or the element R 2 D  D) is called the R-matrix. It satisfies the
following axioms:

Kijk Rkm ¼ Ril Rjn Xmln ; Rmk Kjik ¼ Xmln Rli Rnj ; ð41Þ

Kjik ¼ Xilm r Xjpn s Rlp Kmk n Rrs ; ð42Þ

where Xilm r ¼ Xilu Xum r ¼ Xulm Xiu r . Eq. (41) follow from (38). Conversely, these equa-
tions ensure that the commutation relation are consistent. To prove the first equation
in (41), commute F m ðt1 ÞF i ðq1 ÞF j ðq1 Þ in two different ways. You will get
Wabijm F a ðq1 ÞF b ðt1 Þ, with two different expressions for Wabijm . Then calculate Wabijm ea eb

19
This proof is not rigorous, but an interested reader can easily fix it. Anyway, you can just substitute
(39) into (40) and check it directly.
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 23

Fig. 8. Two ribbons attached to the same site.

Fig. 9. Checking consistency of the commutation relations.

using the axioms (30)–(32). The second equation in (41) can be proven in a similar
way.
To prove Eq. (42), consider the configuration shown in Fig. 9a. Clearly,
F i ðq1 q2 Þ  F i ðt1 t2 Þ, so F j ðt1 t2 ÞF i  Kjik F k ðt1 t2 Þ. On the other hand, we can first expand
F j ðt1 ; t2 Þ and F i ðq1 q2 Þ using the comultiplication rules, and then apply the commuta-
tion relations (38). The result must be the same.
Let t be a ribbon connecting sites a and b. The local and ribbon operators com-
mute as follows:

F m ðtÞDi ðaÞ ¼ Kjk m l

i Xkl Dj ðaÞF ðtÞ; Di ðbÞF m ðtÞ ¼ Xmlk Kkj l
i F ðtÞDj ðbÞ: ð43Þ
These commutation relations can be also written in the form
Dj ðaÞF l ðtÞ ¼ Kikj S~kn Xlnm F m ðtÞDi ðaÞ;
ð44Þ
F l ðtÞDj ðbÞ ¼ Xl S~n Kki Di ðbÞF m ðtÞ;
mn k j

where S~$$ is the skew antipode (see (35) and (36)).

Finally, we introduce some special elements C 2 D and s 2 F. The first one has a
clear physical meaning: the corresponding operator CðaÞ ¼ AðaÞBðaÞ projects out
states with no particle at the site a. The element C can be represented in the form
24 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

C ¼ ci Di ; where cðh;gÞ ¼ N
1 dh;1 : ð45Þ

It has the following properties:
CX ¼ XC ¼ eðX ÞC for any X 2 D eðCÞ ¼ 1;
or, in tensor notations,
Xkij ci ¼ Xkji ci ¼ ej ck ek ck ¼ 1: ð46Þ
The element s 2 F is dual to C; it is defined as follows:
s ¼ si F i ; where sðh;gÞ ¼ N
1 d1;g : ð47Þ
Its properties are as follows:
Kijk si ¼ Kjik si ¼ ej sk ; ek sk ¼ 1: ð48Þ
Note that sk ck ¼ N
2 . Using these properties, we can derive an important conse-
quence from the commutation relations (43)
ss Xsm p Sqp F m ðtÞCðaÞF q ðtÞ ¼ ss Xspm Spq F q ðtÞCðbÞF m ðtÞ ¼ N
2 : ð49Þ

6.4. The space Lða; bÞ

Now we are in a position to find the space Lða; bÞ and to prove the assertions from
Section 6.1. The first assertion was that any excited spot can be transformed into one
particle. It is simple if we can transform two particles into one by ribbon operators. Let
us choose an arbitrary site b the excited spot to be compressed to. Let some constraint,
AðsÞ
1  0 or BðpÞ
1  0, be violated. Choose any site a containing the vertex s or
the face p. Connect a and b by a ribbon. By the assumption, we can clean up the site a
while changing the state of b, but without violating any more constraint. We can re-
peat this procedure again and again to clean up the whole spot.
So, we only need to show that two particles can be transformed into one. What
does it mean exactly? Physically, any transformation must be unitary, but it can in-
volve also some external system. (Otherwise, it is impossible to ‘‘decrease entropy,’’
i.e., to convert many states into fewer.) On the other hand, it is clear that unitarity is
not relevant to this problem. However, we should exclude degenerate transforma-
tions, such as multiplication by the zero operator. So, it is better to reformulate
the assertion as follows: any two-particle state (plus some other excitations far away)
can be obtained from one-particle states (plus the same excitations far away). Let
jwi 2 Lða; b; . . .Þ be such a two-particle state. We are going to use the formula
(49). Let
Gq ¼ N 2 ss Xsm p Sqp F m ðtÞ; jgq i ¼ CðaÞF q ðtÞjwi: ð50Þ
Then jwi ¼ Gq jgq i. The states jgq i belong to Lðb; . . .Þ, i.e., do not contain excitation
at a. This is exactly what we need.
The other two assertions were about the action of local operators on the
space Lða; bÞ, so we need to find this space first. We can consider this space as a
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 25

representation of the algebra E ffi EðtÞ generated by the operators Dj ¼ Dj ðaÞ,

F l ¼ F l ðtÞ, and D0j ¼ Dj ðbÞ. As a linear space, E ¼ D  F  D. (Thus, E has dimen-
sionality N 6 .) Multiplication in E is defined by the commutation relations (43). We
will call E ffi EðtÞ the algebra of extended ribbon operators. It is just an algebra,
not a Hopf algebra. More exactly, it is a finite-dimensional C $ -algebra. The involu-
tion ( ¼ Hermitian conjugation) is given by the formulas (cf. (19))
y y
1 ;gÞ y
ðDðh;gÞ Þ ¼ Dðg
1 hg;g
1 Þ ; ðF ðh;gÞ Þ ¼ F ðh ; ðD0ðh;gÞ Þ ¼ D0ðg
1 hg;g
1 Þ : ð51Þ

Remark. Apparently, the algebra E will play the central role in a general theory of
topological quantum order. Indeed, we were lucky to define ribbon operators sep-
arately from local operators. In the general case, ribbon operators should be mixed
with local operators.

So, we are looking for a particular representation L of the algebra E. This rep-
resentation must contain a special vector jni (the ground state) such that
Dk jni ¼ ek jni; D0k jni ¼ ek jni: ð52Þ
We start with constructing a representation L  spanned by the vectors jw i ¼ F jni. k k

(It will be proven after that L  ¼ L.) We assume that the vectors jwk i are linearly
independent. This need not be the case in the representation L but we can postulate
jwk i being linearly independent in L . Thus, L contains a factor-representation of
L.
The representation L  is given by the formulas

Dj jwk i ¼ S~jn Xknm jwm i; F j jwk i ¼ Kjk m

m jw i; D0j jwk i ¼ Xkmj jwm i: ð53Þ
 , i.e.,
It is easy to show that this representation is irreducible. Hence, L contains L
the vectors jwk i are linearly independent in L. The scalar products between the
vectors jwk i can be found from (53) and (51),
hwðv;uÞ jwðh;gÞ i ¼ N
1 dv;h du;g : ð54Þ
To prove that L  ¼ L, we use the Eq. (49) again. For an arbitrary two-particle
state jwi 2 L, define the vectors jgq i and operators Gq as in Eq. (50). Then
jwi ¼ Gq jgq i. One could say that jgq i 2 LðbÞ but, actually, the space LðbÞ is
spanned by the sole vector jni. It follows that jwi 2 L  —the assertion has been
proven. Thus, the action of local and ribbon operators on the space L ¼ Lða; bÞ
is given by Eq. (53).
It is easy to see that the action of DðaÞ on Lða; bÞ is exact (though it is reducible).
Besides that, DðaÞ is the commutant of DðbÞ in LðLða; bÞÞ and vise versa. (That is,
DðaÞ consists of all operators X 2 LðLða; bÞÞ which commute with every Y 2 DðbÞ.)
Hence, DðaÞ includes all local operators acting on the space Lða; bÞ. Indeed, a local
operator, which involves only spins near the site a, must commute with any operator
acting on distant spins. Of course, there are many such operators, but their action on
the two-particle space Lða; bÞ coincides with the action of the operators from DðaÞ.
This is also true for a multi-particle excitation space Lðx1 ; . . . ; xn Þ.
26 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

The space Lðx1 ; . . . ; xn Þ can be described as follows. Let us connect the sites
x1 ; . . . ; xn by n
1 ribbons t1 ; . . . ; tn
1 in an arbitrary way so that the ribbons form
a tree. Then the vectors jwk1 ;:::;kn
1 i ¼ F k1 ðt1 Þ . . . F kn
1 ðtn
1 Þjni form a basis of
Lðx1 ; . . . ; xn Þ. Choosing different ribbons means choosing a different basis. In
the next section we will give another description of multi-particle excitation
spaces.

7. Topological operators, braiding, and fusion

Let us consider again the n-particle excitation space L ¼ Lðx1 ; . . . ; xn Þ. The alge-
bra LðLÞ includes the local operator algebras Dðx1 Þ; . . . ; Dðx1 Þ. An operator
Y 2 LðLÞ which commute with every X 2 Dðxj Þðj ¼ 1; . . . ; nÞ is called a topological
operator. Physically, topological operators correspond to non-local degrees of free-
dom. For n ¼ 2, the algebra of topological operators coincides with the center of
Dðx1 Þ or Dðx2 Þ. (The two centers coincide). Hence, the only non-local degree of free-
dom is the type of either particle. (The two particles correspond to dual representa-
tions of D; in other words, these are a particle and an anti-particle.) So, there is no
hidden (i.e., quantum non-local) degree of freedom in this case. Such hidden degrees
of freedom appear for n P 3.
To describe the space L and operators acting on it, let us choose an arbitrary site
x0 (distinct from x1 ; . . . ; xn ) and connect it with x1 ; . . . ; xn by non-intersecting ribbons
t1 ; . . . ; tn , see Fig. 10a. As stated above, the space Lðx0 ; x1 ; . . . ; xn Þ is spanned by the
vectors

jwk1 ;:::;kn i ¼ F k1 ðt1 Þ    F kn ðtn Þjni: ð55Þ

The space in question, L ¼ Lðx1 ; . . . ; xn Þ is contained in Lðx0 ; x1 ; . . . ; xn Þ. It consists

of all vectors jwi 2 Lðx0 ; x1 ; . . . ; xn Þ which are invariant under the action of Dðx0 Þ on
the latter space.
The advantage of this description is that we can easily find all operators on the
space Lðx0 ; x1 ; . . . ; xn Þ which commute with Dðx1 Þ      Dðxn Þ. These are simply
operators which act on the ends of the ribbons t1 ; . . . ; tn attached to the site x0 . More

Fig. 10. Braiding and fusion in terms of ribbon transformations.

 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 27

ðrÞ
exactly, an operator Dj ðr ¼ 1; . . . ; nÞ acts on the rth ribbon as D0j ¼ Dj ðx0 Þ (see Eq.
(53)), but does not affect the other ribbons,
 
ð1Þ ðnÞ
Dj1      Djn jwk1 ;:::;kn i ¼ Xkm11 j1    Xkmnn jn jwm1 ;:::;mn i: ð56Þ

Thus, we arrive to an interesting physical conclusion. Let us consider only one

particle attached to an end of a semi-infinite ribbon (an analog of DiracÕs string).
Then the topological operators act on the far end of the ribbon.

Example. Let us see how the topological operators act on magnetic vortices. As
shown in Section 6.1, a vortex type is characterized by a conjugacy class C of the
group G. Individual topological states of the particle are characterized by particular
elements v 2 C. In terms of the notation (55), such a state can be represented as
follows:
1=2
X
ju; vi ¼ jCj jwðu;xÞ i;
x:x
1 ux¼v

where u 2 C characterize the local state of the particle. One can easily check that
D0ðh;gÞ ju; vi ¼ dh;gvg
1 ju; hi. This is consistent with Eq. (22). Note that the local degree
of freedom, u, is not affected.
How can we physically apply topological operators to particles? We can just move
the particles around each other; this process is called braiding. Let us see what
happens if we interchange two particles, xs and xsþ1 , counterclockwise, as shown
in Fig. 10b. The state jw:::;k;l;::: i becomes a new state

jw:::;k;l;::: inew ¼ Rx jw:::;k;l;::: i ¼    F k ðts0 ÞF l ðtsþ1

0
Þ    jni:

To represent this state in the old basis, we should represent the operator
F k ðts0 ÞF l ðtsþ1
0
Þ in terms of F m ðts Þ and F n ðtsþ1 Þ. Obviously, F k ðts0 Þ ¼ F k ðtsþ1 Þ; also
l 0
F ðtsþ1 Þ  F ðts Þ as long as there is no particle at xsþ1 , i.e., the operator F k ðtsþ1 Þ is not
l

applied yet. Hence,

F k ðts0 ÞF l ðtsþ1
0
Þ  F k ðtsþ1 ÞF l ðts Þ:

Now we can apply the second commutation relation from (38). (Actually, we should
reverse it.) It follows that:

F k ðtsþ1 ÞF l ðts Þ ¼ Rji Xlm i Xkn j F m ðts ÞF n ðtsþ1 Þ;

 
ðsÞ ðsþ1Þ
Rx jw:::;k;l;::: i ¼ Rji Di  Dj jw:::;l;k;::: i:

(see Eq. (56)). Consequently, the counterclockwise interchange operator has the
form
   
Rx ¼ Rji D0i  D0j r ¼ rRij D0i  D0j ; ð57Þ
28 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

where r is the permutation operator, and D0i , D0j are understood as topological op-
erators. (Note that the operator r permutes both topological and local degrees of
freedom.)

Example. Consider two magnetic vortices characterized by topological parameters

v1 ; v2 2 G. The operator Rx acts on the state jv1 ; v2 i as follows:
Rx jv1 v2 i ¼ jv1 v2 v
1
1 ; v1 i: ð58Þ
(The local parameters, u1 and u2 , are suppressed in this formula.)
Finally, let us see what happens if two particles, xs and xsþ1 , fuse into one. The
resulting particle can be characterized by the action of topological operators on it.
From this point of view, we can just glue parts of the corresponding ribbons (see
Fig. 10c) instead of fusing the particles themselves. Then


0  p
F k ðts ÞF l ðtsþ1 Þ  Xkm i Xln j Kijp F m ts0 F n tsþ1 F ðqÞ;



jw...;k;l;... i  Xkm i Xln j Kijp F m ts0 F n tsþ1
0
jw...;p;... i;
ðsÞ ðsþ1Þ
Kuv
r Du  Dv  D0r ;

where D0r acts on the end of the ribbon q. Thus, fusion is described by the comul-
tiplication in the algebra D, see Eq. (27). (To avoid confusion, one should replace D$
with D0$ in that equation.) The topological operator DðD0k Þ acts on a particle pair as
the topological operator D0k on the particle resulting from fusion.

Example. Consider a pair of opposite magnetic vortices jv; v
1 i. The operators DðD0k Þ
act on this state as follows:
 
D D0ðh;gÞ jv; v
1 i ¼ dh;1 jgvg
1 ; gv
1 g
1 i: ð59Þ

It terms of the representation classification (see Section 6.1), this action corresponds
to the pair ðC; vÞ, where C ¼ f1g, and v is the adjoint representation of G. Thus,
when opposite magnetic vortices fuse, the resulting particle has no magnetic charge
but may have some electric charge.

8. Universal computation by anyons

(This section should be considered as an abstract of results to be presented else-

where.)
Universal quantum computation is possible in the model based on the permuta-
tion group S5 . (Unsolvability of the group seems to be important.) Vortex pairs
jv; v
1 i, where v is a transposition, are used as qubits. It is possible to perform the
following operations:
1. To produce pairs with zero charge. If a pair is created from the ground state, it
has no charge automatically.
 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30 29

2. To measure the electric charge of a vortex pair destructively. For this, we should
simply fuse the pair into one particle.
3. To perform the following unitary transformation on two pairs
ju; u
1 i  jv; v
1 i 7! jvuv
1 ; vu
1 v
1 i  jv; v
1 i: ð60Þ
For this, we pull the first pair (as a whole) between the particles of the second pair.
4. To measure the value of v and produce an unlimited number of pure states jv; v
1 i
for any given transposition v (say, ð1; 2Þ or ð2; 3Þ). [At first sight, it is impossible
because we can only measure the conjugacy class of a v. However, we can agree on
a given state to correspond to v ¼ ð1; 2Þ. Then we use it as a reference to produce
an unlimited number of copies.]
The operations 3 and 4 are sufficient to perform universal classical computation.
It is relatively simple to run quantum algorithms based on measurements [24]. Sim-
ulating a universal gate set is more subtle and requires composite qubits. That is, a
usual qubit (with two distinct states) is represented by several vortex pairs.

9. Concluding remarks

It has been shown that anyons can arise from a Hamiltonian with local interactions
but without any symmetry. These anyons can be used to perform universal quantum
computation. There are still many things to do and questions to answer. First of all, it
is desirable to find other models with anyons which allow universal quantum compu-
tation. (The group S5 is quite unrealistic for physical implementation). Such models
must be based on a more general algebraic structure rather than the quantum double
of a group algebra. A general theory of anyons and topological quantum order is lack-
ing. [In a sense, a general theory of anyons already exists [10]; it is based on quasi-tri-
angular quasi-Hopf algebras. However, this theory either merely postulates the
properties of anyons or connects them to certain field theories. This is quite unlike
the theory of local and ribbon operators which describes both the properties of exci-
tations and the underlying spin entanglement.] It is also desirable to formulate and
prove some theorem about existence and the number of local degrees of freedom.
(It seems that the local degrees of freedom are a sign that anyons arise from a system
with no symmetry in the Hamiltonian.) Finally, general understanding of dynamically
created, or ‘‘materialized’’ symmetry is lacking. There one may find some insights for
high-energy physics. If we adopt a conjecture that the fundamental Hamiltonian or
Lagrangian is not symmetric, we can probably infer some consequences about the par-
ticle spectrum.

Acknowledgments

I am grateful to J. Preskill, D.P. DiVincenzo and C.H. Bennett for interesting dis-
cussions and questions which helped me to clarify some points in my constructions.
This work was supported, in part, by the Russian Foundation for Fundamental
30 A.Yu. Kitaev / Annals of Physics 303 (2003) 2–30

Research, Grant No. 96-01-01113. Part of this work was completed during the 1997
Elsag-Bailey—I.S.I. Foundation research meeting on quantum computation.

References

[1] P. Shor, in: Proceedings of the 35th Annual Symposium on Fundamentals of Computer Science,
IEEE Press, Los Alamitos, CA, 1994, pp. 124–134.
[2] P. Shor, in: Proceedings of the Symposium on the Foundations of Computer Science, IEEE Press,
Los Alamitos, CA, 1996, e-print quant-ph/9605011.
[3] E. Knill, R. Laflamme, Concatenated quantum codes, 1996 (e-print quant-ph/9608012).
[4] E. Knill, R. Laflamme, W. Zurek, Accuracy threshold for quantum computation, 1996 (e-print
quant-ph/9610011).
[5] D. Aharonov, M. Ben-Or, Fault-tolerant quantum computation with constant error, 1996 (e-print
quant-ph/9611025).
[6] A.Yu. Kitaev, Quantum computation: algorithms and error correction, Russian Math. Surveys, 52
(6) (1997) 1191.
[7] C. Zalka, Threshold estimate for fault-tolerant quantum computing, 1996 (e-print quant-ph/
9612028).
[8] A.Yu. Kitaev, in: O. Hirota, A.S. Holevo, C.M. Caves (Eds.), Quantum Communication, Computing
and Measurement, Plenum, New York, 1997.
[9] F. Wilczek, Fractional Statistics and Anyon Superconductivity, World Scientific, Singapore, 1990.
[10] R. Dijkgraaf, V. Pasquier, P. Roche, Nucl. Phys. B (Proc. Suppl.) 18B (1990).
[11] F.A. Bais, P. van Driel, M. de Wild Propitius, Phys. Lett. B 280 (1992) 63.
[12] F.A. Bais, M. de Wild Propitius, Discrete gauge theories, 1995 (e-print hep-th/9511201).
[13] H.K. Lo, J. Preskill, Phys. Rev. D 48 (1993) 4821.
[14] G. Castagnoli, M. Rasetti, Int. J. Mod. Phys. 32 (1993) 2335.
[15] D. Gottesman, Phys. Rev. A 54 (1996) 1862.
[16] A.R. Calderbank, E.M. Rains, P.M. Shor, N.J.A. Sloane, Phys. Rev. Lett. 78 (1997) 405.
[17] A.R. Calderbank, P.W. Shor, Good quantum error-correcting codes exist, 1995 (e-print quant-ph/
9512032).
[18] D. Arovas, J.R. Schrieffer, F. Wilczek, Phys. Rev. Lett. 53 (1984) 722–723.
[19] T. Einarsson, Phys. Rev. Lett. 64 (1984) 1995–1998.
[20] V.G. Drinfeld, in: Proc. Int. Cong. Math. (Berkley, 1986), 1987, pp. 798–820.
[21] M. Sweedler, Hopf Algebras, W.A. Benjamin, New York, 1969.
[22] S. Majid, Int. J. Mod. Phys. A 5 (1990) 1–91.
[23] C. Kassel, Quantum Groups, Springer-Verlag, New York, 1995.
[24] A.Yu. Kitaev, Quantum measurements and Abelian stabilizer problem, 1995 (e-print quant-ph/
9511026).
[25] G. tÕHooft, Nucl. Phys. B 138 (1978) 1.