-
Topological quantum compiling
L. Hormozi, G. Zikos, and N. E. BonesteelDepartment of Physics
and National High Magnetic Field Laboratory, Florida State
University, Tallahassee, Florida 32310, USA
S. H. SimonBell Laboratories, Lucent Technologies, Murray Hill,
New Jersey 07974, USA
�Received 17 October 2006; published 11 April 2007�
A method for compiling quantum algorithms into specific braiding
patterns for non-Abelian quasiparticlesdescribed by the so-called
Fibonacci anyon model is developed. The method is based on the
observation thata universal set of quantum gates acting on qubits
encoded using triplets of these quasiparticles can be builtentirely
out of three-stranded braids �three-braids�. These three-braids can
then be efficiently compiled andimproved to any required accuracy
using the Solovay-Kitaev algorithm.
DOI: 10.1103/PhysRevB.75.165310 PACS number�s�: 73.43.�f,
03.67.Lx, 03.67.Pp
I. INTRODUCTION
The requirements for realizing a fully functioning quan-tum
computer are daunting. There must be a scalable systemof qubits
which can be initialized and individually measured.It must be
possible to enact a universal set of quantum gateson these qubits.
And all this must be done with sufficientaccuracy so that quantum
error correction can be used toprevent decoherence from spoiling
any computation.
The problems of error and decoherence are particularlydifficult
ones for any proposed quantum computer. While thestates of
classical computers are typically stored in macro-scopic degrees of
freedom which have a built-in redundancyand thus are resistant to
errors, building similar redundancyinto quantum states is less
natural. To protect quantum infor-mation it is necessary to encode
it using quantum error-correcting code states.1,2 These states are
highly entangled,and have the property that code states
corresponding to dif-ferent logical qubit states can be
distinguished from one an-other only by global �“topological”�
measurements. Unlikestates whose macroscopic degrees of freedom are
effectivelyclassical �think of the magnetic moment of a small part
of ahard drive�, such highly entangled “topologically degener-ate”
states do not typically emerge as the ground states ofphysical
Hamiltonians. One route to fault-tolerant quantumcomputation is
therefore to build the encoding and fault-tolerant gate protocols
into the software of the quantumcomputer.3
A remarkable recent development in the theory of quan-tum
computation which directly addresses these issues hasbeen the
realization that certain exotic states of matter in twospace
dimensions, so-called non-Abelian states, may providea natural
medium for storing and manipulating quantuminformation.4–7 In these
states, localized quasiparticle excita-tions have quantum numbers
that are in some ways similar toordinary spin quantum numbers.
However, unlike ordinaryspins, the quantum information associated
with these quan-tum numbers is stored globally, throughout the
entire system,and so is intrinsically protected against
decoherence. Further-more, these quasiparticles satisfy so-called
non-Abelian sta-tistics. This means that when two quasiparticles
are adiabati-cally moved around one another, while being
keptsufficiently far apart, the action on the Hilbert space is
rep-
resented by a unitary matrix which depends only on the to-pology
of the path used to carry out the exchange. Topologi-cal quantum
computation can then be carried out bymoving quasiparticles around
one another in two spacedimensions.4,5 The quasiparticle
world-lines form topologi-cally nontrivial braids in three �=2+1�
-dimensional space-time, and because these braids are topologically
robust �i.e.,they cannot be unbraided without cutting one of the
strands�the resulting computation is protected against error.
Non-Abelian states are expected to arise in a variety ofquantum
many-body systems, including spin systems,8–10 ro-tating Bose
gases,11 and Josephson junction arrays.12 Ofthose states which have
actually been experimentally ob-served, the most likely to possess
non-Abelian quasiparticleexcitations are certain fractional quantum
Hall states. Mooreand Read13 were the first to propose that
quasiparticle exci-tations which obey non-Abelian statistics might
exist in thefractional quantum Hall effect. Their proposal was
based onthe observation that the conformal blocks associated
withcorrelation functions in the conformal field theory
describingthe two-dimensional Ising model could be interpreted
asquantum Hall wave functions. These wave functions describeboth
the ground state of a half-filled Landau level of spin-polarized
electrons, as well as states with some number offractionally
charged quasihole excitations �charge e /4�. Theparticular ground
state this construction produces, the so-called Pfaffian or
Moore-Read state, is considered the mostlikely candidate for the
observed fractional quantum Hallstate at Landau level filling
fraction �=5/2 ��=1/2 in thesecond Landau level�.14,15
In this conformal field theory construction, states withfour or
more quasiholes present correspond to finite-dimensional conformal
blocks, and so the correspondingwave functions form a
finite-dimensional Hilbert space. Themonodromy—or braiding
properties—of these conformalblocks are then assumed to describe
the unitary transforma-tions acting on the Hilbert space produced
by adiabaticallybraiding quasiholes around one another.13 Explicit
wavefunctions for these states were worked out in Ref. 16, and
thenon-Abelian braiding properties have been verified numeri-cally
in Ref. 17. In an alternate approach, the Moore-Readstate can be
viewed as a composite fermion superconductorin a so-called weak
pairing px+ ipy phase.
18 In this descrip-
PHYSICAL REVIEW B 75, 165310 �2007�
1098-0121/2007/75�16�/165310�19� ©2007 The American Physical
Society165310-1
http://dx.doi.org/10.1103/PhysRevB.75.165310
-
tion, the finite-dimensional Hilbert space arises from
zero-energy solutions of the Bogoliubov–de Gennes equationsin the
presence of vortices,18 and the vortices themselves arenon-Abelian
quasiholes whose braiding properties have beenshown to agree with
the conformal field theory result.19,20
Recently, a number of experiments have been proposedto directly
probe the non-Abelian nature of theseexcitations.21–24
Unfortunately, the braiding properties of quasihole excita-tions
in the Moore-Read state are not sufficiently rich tocarry out
purely topological quantum computation, although“partially”
topological quantum computation using a mixtureof topological and
nontopological gates has been shown tobe possible.25,26 However,
Read and Rezayi27 have shownthat the Moore-Read state is just one
of a sequence of stateslabeled by an index k corresponding to
electrons at fillingfractions �=k / �2+k�, with k=1 corresponding
to the �=1/3 Laughlin state and k=2 to the Moore-Read state.
Thewave functions for these states can be written as
correlationfunctions in the Zk parafermion conformal field
theory,
27 andthe braiding properties of the quasihole excitations
wereworked out in detail in Ref. 28. There it was shown that
thequasiholes are described by the SU�2�k Chern-Simons-Witten �CSW�
theories, up to overall Abelian phase factorswhich are irrelevant
for quantum computation. More re-cently, explicit quasihole wave
functions have been workedout for the k=3 Read-Reazyi state,29 with
results consistentwith the predicted SU�2�3 braiding properties.
The elemen-tary braiding matrices for the SU�2�k CSW theory for
k=3and k�5 have been shown to be sufficiently rich to carry
outuniversal quantum computation, in the sense that any
desiredunitary operation on the Hilbert space of N
quasiparticles,with N�3 for k�3,k�4,8 and N�4 for k=8, can be
ap-proximated to any desired accuracy by a braid.5,6
The main purpose of this paper is to give an efficientmethod for
determining braids which can be used to carryout a universal set of
a quantum gates �i.e., single-qubit ro-tations and controlled-NOT
gates� on encoded qubits for thecase k=3, thought to be physically
relevant for the experi-mentally observed30 �=12/5 fractional
quantum Halleffect27,31 ��=12/5 corresponds to �=2/5 in the second
Lan-dau level, and this is the particle-hole conjugate of
�=3/5corresponding to k=3�. We refer to the process of findingsuch
braids as “topological quantum compiling” since thesebraids can
then be used to translate a given quantum algo-rithm into the
machine code of a topological quantum com-puter. This is analogous
to the action of an ordinary compilerwhich translates instructions
written in a high-level program-ming language into the machine code
of a classical com-puter.
It should be noted that the proof of universality forSU�2�3
quasiparticles is a constructive one,5,6 and therefore,as a matter
of principle, it provides a prescription for com-piling quantum
gates into braids. However, in practice, fortwo-qubit gates �such
as controlled-NOT gates� this prescrip-tion, if followed
straightforwardly, is prohibitively difficultto carry out,
primarily because it involves searching thespace of braids with six
or more strands. We address thisdifficulty by dividing our
two-qubit gate constructions into a
series of smaller constructions, each of which involvessearching
only the space of three-stranded braids �three-braids�. The
required three-braids then can be found effi-ciently and used to
construct the desired two-qubit gates.This divide and conquer
approach does not, in general, yieldthe most accurate braid of a
given length which approxi-mates a desired quantum gate. However,
we believe that itdoes yield the most accurate �or at least among
the mostaccurate� braids which can be obtained for a given
fixedamount of classical computing power.
This paper is organized as follows. In Sec. II we reviewthe
basic properties of the SU�2�k Hilbert space, and showthat the case
SU�2�3 is, for our purposes, equivalent to thecase SO�3�3—the
so-called Fibonacci anyon model. SectionIII then presents a quick
review of the mathematical machin-ery needed to compute with
Fibonacci anyons. In Sec. IV weoutline how, in principle, these
particles can be used to en-code qubits suitable for quantum
computation. Section Vthen describes how to find braiding patterns
for three Fi-bonacci anyons which can be used to carry out any
allowedoperation on the Hilbert space of these quasiparticles to
anydesired accuracy, thus effectively implementing the proce-dure
given in Ref. 5 for carrying out single-qubit rotations.In Sec. VI
we discuss the more difficult case of two-qubitgates, and give two
classes of explicit gate constructions—one, first discussed by the
authors in Ref. 32, in which a pairof quasiparticles from one qubit
is “woven” through the qua-siparticles in the second qubit, and
another, presented herefor the first time, in which only a single
quasiparticle is wo-ven. Finally, in Sec. VII we address the
question of to whatextent the constructions we find are special to
the k=3 case,and in Sec. VIII we summarize our results.
II. FUSION RULES AND HILBERT SPACE
Consider a system with quasiparticle excitations describedby the
SU�2�k CSW theory. It is convenient to describe theproperties of
this system using the so-called quantum grouplanguage.28 The
relevant quantum groups are “deformed”versions of the
representation theory of SU�2�, i.e., thetheory of ordinary spin,
and much of the intuition for think-ing about ordinary spin can be
carried over to the quantumgroup case.
In the quantum group description of an SU�2�k CSWtheory, each
quasiparticle has a half-integer q-deformed spin�q-spin� quantum
number. Just as for ordinary spin, there arerules for combining
q-spin known as fusion rules. The fusionrules for the SU�2�k theory
are similar to the usual trianglerule for adding ordinary spin,
except that they are truncatedso that there are no states with
total q-spin �k /2. Specifi-cally, the fusion rules for the level k
theory are33
s1 � s2 = �s1 − s2� � �s1 − s2� + 1 � ¯
� min�s1 + s2,k − s1 − s2� . �1�
Note that, in the quantum group description of
non-Abeliananyons, states are distinguished only by their total
q-spinquantum numbers. The q-deformed analogs of the Sz quan-tum
numbers are physically irrelevant—there is no degen-
HORMOZI et al. PHYSICAL REVIEW B 75, 165310 �2007�
165310-2
-
eracy associated with them, and they play no role in
anycomputation involving braiding.28 The situation is
somewhatanalogous to that of a collection of ordinary spin-1/2
par-ticles in which the only allowed operations, including
mea-surement, are rotationally invariant and hence independent
ofSz, as is the case in exchange-based quantum computation.
34
The fusion rules of the SU�2�k theory fix the structure ofthe
Hilbert space of the system. For a collection of quasipar-ticles
with q-spin-1/2, a useful way to visualize this Hilbertspace is in
terms of its so-called Bratteli diagram. This dia-gram shows the
different fusion paths for N q-spin-1/2 qua-siparticles in which
these quasiparticles are fused, one at atime, going from left to
right in the diagram. Bratteli dia-grams for the cases k=2 and 3
are shown in Fig. 1.
The dimensionality of the Hilbert space for N
q-spin-1/2quasiparticles with total q-spin S can be determined
bycounting the number of paths in the Bratteli diagram fromthe
origin to the point �N ,S�. The results of this path count-ing are
also shown in Fig. 1, where one can see the well-known 2N/2−1
Hilbert space degeneracy for the k=2 �Moore-Read� case,13,16 and
the Fibonacci degeneracy for the k=3case.27
In this paper we will focus on the k=3 case, which is thelowest
k value for which SU�2�k non-Abelian anyons areuniversal for
quantum computation.5,6 In fact, we will showthat two-qubit gates
are particularly simple for this case. Be-fore proceeding, it is
convenient to introduce an importantproperty of the SU�2�3 theory,
namely, that the braidingproperties of q-spin-1/2 quasiparticles
are the same as thosewith q-spin 1 �up to an overall Abelian phase
which is irrel-evant for topological quantum computation�. This is
a usefulobservation because the theory of q-spin-1 quasiparticles
inSU�2�3 is equivalent to SO�3�3, a theory also known as
theFibonacci anyon theory35,36—a particularly simple theorywith
only two possible values of q-spin, 0 and 1, for whichthe fusion
rules are
0 � 0 = 0, 0 � 1 = 1 � 0 = 1, 1 � 1 = 0 � 1. �2�
Here we give a rough proof of this equivalence. Thisproof is
based on the fact that for k=3 the fusion rules in-
volving q-spin-3/2 quasiparticles take the following
simpleform:
3
2� s =
3
2− s . �3�
The key observation is that, since for k=3 the highest pos-sible
q-spin is 3/2, when fusing a q-spin-3/2 object with anyother object
�here we use the term object to describe either asingle
quasiparticle or a group of quasiparticles viewed as asingle
composite entity�, the Hilbert space dimensionalitydoes not grow.
This implies that moving a q-spin-3/2 objectaround other objects
can, at most, produce an overall Abelianphase factor. While this
phase factor may be important physi-cally, particularly in
determining the outcome of interferenceexperiments involving
non-Abelian quasiparticles,21–24 it isirrelevant for quantum
computing, and thus does not matterwhen determining braids which
correspond to a given com-putation. Because �3� implies that a
q-spin 1/2 object can beviewed as the result of fusing a q-spin-1
object with aq-spin-3/2 object, it follows that the braid matrices
forq-spin-1/2 objects are the same as those for q-spin-1 objectsup
to an overall phase �as can be explicitly checked�.
In fact, based on this argument we can make a strongerstatement.
Imagine a collection of SU�2�3 objects which eachhave either q-spin
1 or q-spin 1/2. It is then possible to carryout topological
quantum computation, even if we do notknow which objects have
q-spin 1 and which have q-spin1/2. The proof is illustrated in Fig.
2. Figure 2�a� shows abraiding pattern for a collection of objects,
some of whichhave q-spin 1/2 and some of which have q-spin 1.
Figure2�b� then shows the same braiding pattern, but now all
ob-jects with q-spin 1/2 are represented by objects with q-spin
1fused to objects with q-spin 3/2. Because, as noted above,
theq-spin-3/2 objects have trivial �Abelian� braiding
properties,the unitary transformation produced by this braid is
thesame, up to an overall Abelian phase, as that produced
bybraiding nothing but q-spin-1 objects, as shown in Fig. 2�c�.It
follows that, provided one can measure whether the totalq-spin of
some object belongs to the class 1��1,1 /2� or theclass 0��0,3
/2�—something which should, in principle, be
FIG. 1. Bratteli diagrams for SU�2�k fork��a� 2 and �b� 3. Here
N is the number ofq-spin-1/2 quasiparticles and S is the total
q-spinof those quasiparticles. The number at a given�N ,S� vertex
of each diagram indicates the num-ber of paths to that vertex
starting from the �0,0�point. This number gives the dimensionality
ofthe Hilbert space of N q-spin-1/2 quasiparticleswith total q-spin
S.
TOPOLOGICAL QUANTUM COMPILING PHYSICAL REVIEW B 75, 165310
�2007�
165310-3
-
possible by performing interference experiments as describedin
Refs. 37 and 38—then quantum computation is possible,even if we do
not know which objects have q-spin 1/2 andwhich have q-spin 1.
III. FIBONACCI ANYON BASICS
Having reduced the problem of compiling braids forSU�2�3 to
compiling braids for SO�3�3, i.e., Fibonaccianyons, it is useful
for what follows to give more detailsabout the mathematical
structure associated with these qua-siparticles. For an excellent
review of this topic see Ref. 35,and for the mathematics of
non-Abelian particles in generalsee Ref. 39.
Note that for the rest of this paper, except for Sec. VII,
itshould be understood that each quasiparticle is a
q-spin-1Fibonacci anyon. It should also be understood that, from
thepoint of view of their non-Abelian properties quasihole
ex-citations are also q-spin-1 Fibonacci anyons, even thoughthey
have opposite electric charge and give opposite Abelianphase
factors when braided. Because it is the non-Abelianproperties that
are relevant for topological quantum compu-tation, for our purposes
quasiparticles and quasiholes can beviewed as identical non-Abelian
particles. Unless it is impor-tant to distinguish between the two
�as when we discuss cre-ating and fusing quasiparticles and
quasiholes in Sec. IV� wewill simply use the terms quasiparticle or
Fibonacci anyon torefer to either excitation.
Figure 3 establishes some of the notation for
representingFibonacci anyons which will be used in the rest of the
paper.This figure shows SU�2�3 Bratteli diagrams in which theq-spin
axis is labeled by both the SU�2�3 q-spin quantumnumbers and, in
boldface, the corresponding Fibonacciq-spin quantum numbers, i.e.,
0 for �0,3 /2� and 1 for�1/2 ,1�. In Fig. 3�a� Bratteli diagrams
showing fusion pathscorresponding to two basis states spanning the
two-dimensional Hilbert space of two Fibonacci anyons areshown.
Beneath each Bratteli diagram an alternate represen-tation of the
corresponding state is also shown. In this rep-resentation dots
correspond to Fibonacci anyons and ovalsenclose collections of
Fibonacci anyons which are in q-spineigenstates whenever the oval
is labeled by a total q-spin
quantum number. �Note: If the oval is not labeled, it shouldbe
understood that the enclosed quasiparticles may not be ina q-spin
eigenstate.�
In the text we will use the notation • to represent a Fi-bonacci
anyon, and the ovals will be represented by paren-theses. In this
notation, the two states shown in Fig. 3�a� aredenoted �• , • �0,
and �• , • �1.
Figure 3�b� shows a Bratteli diagram, again with bothSU�2�3 and
Fibonacci quantum numbers, with fusion pathsthat this time
correspond to three basis states of the three-dimensional Hilbert
space of three Fibonacci anyons. Be-neath these diagrams the oval
representations of these threestates are also shown, which in the
text will be represented(�• , • �0 , • )1, (�• , • �1 , • )1, and
(�• , • �1 , • )0.
In addition to fusion rules, all theories of non-Abeliananyons
possess additional mathematical structure which al-lows one to
calculate the result of any braiding operation.This structure is
characterized by the F �fusion� and R �rota-tion�
matrices.35,39,40
To define the F matrix, note that the Hilbert space of
threeFibonacci anyons is spanned by both the three states
labeled(�• , • �a , • )c, and the three states labeled (• , �• , •
�b)c. The Fmatrix is the unitary transformation which maps one of
thesebases to the other,
„•,�•, • �a…c = �b
Fabc„�•, • �b, • …c, �4�
and has the form
F = � ��
� − �1
, �5�
where �= �5−1� /2 is the inverse of the golden mean. In
thismatrix the upper left 2�2 block Fab
1 acts on the two-dimensional total q-spin-1 sector of the
three-quasiparticleHilbert space, and the lower right matrix
element F11
0 =1 actson the unique total q-spin-0 state. Note that this F
matrix canbe applied to any three objects which each have q-spin
1,where each object can consist of more than one Fibonaccianyon.
Furthermore, if one considers three objects for whichone or more of
the objects has q-spin 0, then the state of
FIG. 2. �Color online� Graphical proof of the equivalence of
braiding q-spin-1/2 and q-spin-1 objects for SU�2�3. �a� shows a
braidingpattern for a collection of objects, some having q-spin 1/2
and some having q-spin 1. �b� shows the same braiding pattern but
with theq-spin-1/2 objects represented by q-spin-1 objects fused
with q-spin-3/2 objects, which, for SU�2�3, has a unique fusion
channel. Finally, �c�shows the same braid with the q-spin-3/2
objects removed. Because these q-spin-3/2 objects are effectively
Abelian for SU�2�3, removingthem from the braid will only result in
an overall phase factor which will be irrelevant for quantum
computing.
HORMOZI et al. PHYSICAL REVIEW B 75, 165310 �2007�
165310-4
-
these objects is uniquely determined by the total q-spin of
allthree, and in this case the F matrix is trivially the
identity.Thus, for the case of Fibonacci anyons, the matrix �5� is
allthat is needed to make arbitrary basis changes for any num-ber
of Fibonacci anyons.
The R matrix gives the phase factor produced when twoFibonacci
anyons are moved around one another with a cer-tain sense. One can
think of these phase factors as theq-deformed versions of the −1 or
+1 phase factors one ob-tains when interchanging two ordinary
spin-1/2 quasiparti-cles when they are in a singlet or triplet
state, respectively.This phase factor depends on the overall q-spin
of the twoquasiparticles involved in the exchange, so for
Fibonaccianyons there are two such phase factors which are
summa-rized in the R matrix,
R = �e−i4�/5 00 ei3�/5
� . �6�Here the upper left and lower right matrix elements are,
re-spectively, the phase factor that two Fibonacci anyons ac-quire
if they are interchanged in a clockwise sense when theyhave total
q-spin 0 or q-spin 1. Again, this matrix also ap-plies if we
exchange two objects that both have total q-spin1, even if these
objects consist of more than one Fibonaccianyon. And if one or both
objects has q-spin 0 the result ofthis interchange is the identity.
Again we emphasize that inthe k=3 Read-Rezayi state, there will be
additional Abelianphases present, which may have physical
consequences for
some experiments, but which will be irrelevant for topologi-cal
quantum computation.
Typically the sequence of F and R matrices used to com-pute the
unitary operation produced by a given braid isnot unique. To
guarantee that the result of any such compu-tation is independent
of this sequence, the F and R matricesmust satisfy certain
consistency conditions. These consis-tency conditions, the
so-called pentagon and hexagonequations,35,39,40 are highly
restrictive, and, in fact, for thecase of Fibonacci anyons
essentially fix the F and R matricesto have the forms given above
�up to a choice of chirality,and Abelian phase factors which are
again irrelevant to ourpurposes here�.35
Finally, we point out an obvious, but important, conse-quence of
the structure of the F and R matrices. When inter-changing any two
quasiparticles which are part of a larger setof quasiparticles with
a well-defined total q-spin quantumnumber, this total q-spin
quantum number will not change.
IV. QUBIT ENCODING AND GENERAL COMPUTATIONSCHEME
Before proceeding, it will be useful to have a specificscheme in
mind for how one might actually carry out topo-logical quantum
computation with Fibonacci anyons. Herewe follow the scheme
outlined in Ref. 7, which, for com-pleteness, we briefly review
below.
The computer can be initialized by pulling
quasiparticle-quasihole pairs out of the “vacuum” �by vacuum we
meanthe ground state of the k=3 Read-Rezayi state or any other
FIG. 3. �Color online� Basis states for the Hilbert space of �a�
two and �b� three Fibonacci anyons. SU�2�3 Bratteli diagrams
showingfusion paths corresponding to the basis states for the
Hilbert space of two and three q-spin-1/2 quasiparticles are shown.
The q-spin axes onthese diagrams are labeled by both the SU�2�3
q-spin quantum numbers 0, 1/2, 1 and 3/2 and, to the left of these
in bold, the correspondingFibonacci q-spin quantum numbers 0��0,3
/2� and 1��1/2 ,1�. Beneath each Bratteli diagram the same state is
represented using anotation in which dots correspond to Fibonacci
anyons, and groups of Fibonacci anyons enclosed in ovals labeled by
q-spin quantumnumbers are in the corresponding q-spin
eigenstates.
TOPOLOGICAL QUANTUM COMPILING PHYSICAL REVIEW B 75, 165310
�2007�
165310-5
-
state that supports Fibonacci anyon excitations�. Each suchpair
will consist of two q-spin-1 excitations in a state withtotal
q-spin 0, i.e., the state �• , • �0. In principle, this pair
canalso exist in a state with total q-spin 1, provided there
areother quasiparticles present to ensure the total q-spin of
thesystem is 0, so one can imagine using this pair as a
qubit.However, it is impossible to carry out arbitrary
single-qubitoperations by braiding only the two quasiparticles
formingsuch a qubit—this braiding never changes the total q-spin
ofthe pair, and so only generates rotations about the z axis inthe
qubit space.
For this reason it is convenient to encode qubits usingmore than
two Fibonacci anyons. Thus, to create a qubit,
twoquasiparticle-quasihole pairs can be pulled out of thevacuum.
The resulting state is then (�• , • �0 , �• , • �0)0 whichagain has
total q-spin 0. The Hilbert space of four Fibonaccianyons with
total q-spin 0 is two dimensional, with basisstates, which we can
take as logical qubit states �0L= (�• , • �0 , �• , • �0)0 and �1L=
(�• , • �1 , �• , • �1)0 �see Fig. 4�a��.The state of such a
four-quasiparticle qubit is determined bythe total q-spin of either
the rightmost or leftmost pair ofquasiparticles. Note that the
fusion rules �2� imply that thetotal q-spin of these two pairs must
be the same because thetotal q-spin of all four quasiparticles is
0.
For this encoding, in addition to the
two-dimensionalcomputational qubit space of four quasiparticles
with totalq-spin 0, there is a three-dimensional noncomputational
Hil-bert space of states with total q-spin 1 spanned by the
states(�• , • �0 , �• , • �1)1, (�• , • �1 , �• , • �0)1, and (�• ,
• �1 , �• , • �1)1.When carrying out topological quantum
computation it iscrucial to avoid transitions into this
noncomputational space.
Fortunately, single-qubit rotations can be carried out
bybraiding quasiparticles within a given qubit and, as discussedin
Sec. III, such operations will not change the total q-spin ofthe
four quasiparticles involved. Single-qubit operations cantherefore
be carried out without any undesirable transitionsout of the
encoded computational qubit space.
Two-qubit gates, however, will require braiding quasipar-ticles
from different qubits around one another. This will ingeneral lead
to transitions out of the encoded qubit space.
Nevertheless, given the so-called “density” result of Ref. 6
itis known that, as a matter of principle, one can always
findtwo-qubit braiding patterns which will entangle the two
qu-bits, and also stay within the computational space to what-ever
accuracy is required for a given computation. The mainpurpose of
this paper is to show how such braiding patternscan be efficiently
found.
Note that the action of braiding the two leftmost
quasipar-ticles in a four-quasiparticle qubit �referring to Fig.
4�a�� isequivalent to that of braiding the two rightmost
quasiparti-cles with the same sense. This is because as long as we
are inthe computational qubit space both the leftmost and
right-most quasiparticle pairs must have the same total q-spin,
andso interchanging either pair will result in the same phasefactor
from the R matrix. It is therefore not necessary tobraid all four
quasiparticles to carry out single-qubitrotations—one need only
braid three.
In fact, one may consider qubits encoded using only
threequasiparticles with total q-spin 1, as originally proposed
inRef. 5. Such qubits can be initialized by first creating a
four-quasiparticle qubit in the state �0L, as outlined above,
andthen simply removing one of the quasiparticles. In this
three-quasiparticle encoding, shown in Fig. 4�b�, the logical
qubitstates can be taken to be �0L= (�• , • �0 , • )1 and �1L= (�•
, • �1 , • )1. For this encoding there is just a single
non-computational state �NC= (�• , • �1 , • )0, also shown in
Fig.4�b�. As for the four-quasiparticle qubit, when carrying
outsingle-qubit rotations by braiding within a
three-quasiparticlequbit the total q-spin of the qubit, in this
case 1, remainsunchanged and there are no transitions from the
computa-tional qubit space into the state �NC. However, just as
forfour-quasiparticle qubits, when carrying out two-qubit
gatesthese transitions will in general occur and we must workhard
to avoid them. Henceforth we will refer to these un-wanted
transitions as leakage errors.
Note that, because each three-quasiparticle qubit has
totalq-spin 1, when more than one of these qubits is present
thestate of the system is not entirely characterized by the
“in-ternal” q-spin quantum numbers which determine the
com-putational qubit states. It is also necessary to specify the
stateof what we will refer to as the “external fusion
space”—theHilbert space associated with fusing the total q-spin-1
quan-tum numbers of each qubit. When compiling braids for
three-quasiparticle qubits it is crucial that the operations on
thecomputational qubit space not depend on the state of
thisexternal fusion space—if they did, these two spaces wouldbecome
entangled with one another leading to errors. Fortu-nately, we will
see that it is indeed possible to find braidswhich do not lead to
such errors.
For the rest of this paper �except Sec. VII� we will usethis
three-quasiparticle qubit encoding. It should be notedthat any
braid which carries out a desired operation on thecomputational
space for three-quasiparticle qubits will carryout the same
operation on the computational space of four-quasiparticle qubits,
with one quasiparticle in each qubit act-ing as a spectator. The
braids we find here can therefore beused for either encoding.
We can now describe how topological quantum computa-tion might
actually proceed, again following Ref. 7. A quan-tum circuit
consisting of a sequence of one- and two-qubit
FIG. 4. �Color online� �a� Four-quasiparticle and �b�
three-quasiparticle qubit encodings for Fibonacci anyons. �a� shows
twostates that span the Hilbert space of four quasiparticles with
totalq-spin 0 which can be used as the logical �0L and �1L states
of aqubit. �b� shows two states spanning the Hilbert space of
threequasiparticles with total q-spin 1 which can also be used as
logicalqubit states �0L and �1L. This three-quasiparticle qubit can
be ob-tained by removing the rightmost quasiparticle from the two
statesshown in �a�. The third state shown in �b�, labeled �NC for
non-computational, is the unique state of three quasiparticles that
hastotal q-spin 0.
HORMOZI et al. PHYSICAL REVIEW B 75, 165310 �2007�
165310-6
-
gates which carries out a particular quantum algorithmwould
first be translated �or “compiled”� into a braid bycompiling each
individual gate to whatever accuracy is re-quired. Qubits would
then be initialized by pullingquasiparticle-quasihole pairs out of
the vacuum. These local-ized excitations would then be
adiabatically dragged aroundone another so that their world-lines
trace out a braid inthree-dimensional space-time which is
topologically equiva-lent to the braid compiled from the quantum
algorithm. Fi-nally, individual qubits would be measured by trying
to fuseeither the two rightmost or two leftmost excitations
withinthem �referring to Fig. 4�a�� for four-quasiparticle qubits,
orjust the two leftmost excitations �referring to Fig. 4�b��
forthree-quasiparticle qubits. If this pair of excitations
consistsof a quasiparticle and a quasihole �and it will always
bepossible to arrange this�, then, if the total q-spin of the pair
is0, it will be possible for them to fuse back into the
vacuum.However, if the total q-spin is 1 this will not be possible.
Theresulting difference in the charge distribution of the
finalstate would then be measured to determine if the qubit was
inthe state �0L or �1L. Alternatively, as already mentioned inSec.
II, interference experiments37,38 could be used to initial-ize and
read out encoded qubits.
As a simple illustration, Fig. 5 shows a computation inwhich a
four-quasiparticle qubit �which can also be viewedas a
three-quasiparticle qubit if the top quasiparticle is ig-nored� is
initialized by pulling quasiparticle-quasihole pairsout of the
vacuum, a single-qubit operation is carried out bybraiding within
the qubit, and the final state of the qubit ismeasured by fusing a
quasiparticle and quasihole togetherand observing the outcome.
V. COMPILING THREE-BRAIDS AND SINGLE-QUBITGATES
We now focus on the problem of finding braids for threeFibonacci
anyons �three-braids� which approximate any al-lowed unitary
transformation on the Hilbert space of these
quasiparticles. This is important not only because it allowsone
to find braids which carry out arbitrary single-qubitrotations,5
but also because, as will be shown in Sec. VI, it ispossible to
reduce the problem of constructing braids whichcarry out two-qubit
gates to that of finding a series of three-braids approximating
specific operations.
A. Elementary braid matrices
Using the F and R matrices, it is straightforward to deter-mine
the elementary braiding matrices that act on the three-dimensional
Hilbert space of three Fibonacci anyons. If, as inFig. 6, we take
the basis states for the three-quasiparticleHilbert space to be the
states labeled (�• , • �a , • )c then, in theac= �01,11,10� basis,
the matrix 1 corresponding to aclockwise interchange of the two
bottommost quasiparticlesin the figure �or leftmost in the (�• , •
�a , • )c representation� is
1 = �e−i4�/5 0
0 ei3�/5
ei3�/5
, �7�
where the upper left 2�2 block acts on the total q-spin-1sector
��0L and �1L� of the three quasiparticles, and thelower right
matrix element is a phase factor acquired by theq-spin 0 state
��NC�. This matrix is easily read off from theR matrix, since the
total q-spin of the two quasiparticlesbeing exchanged is well
defined in this basis.
To find the matrix 2 corresponding to a clockwise inter-change
of the two topmost �or rightmost in the (�• , • �a , •
)crepresentation� quasiparticles, we must first use the F matrixto
change bases to one in which the total q-spin of
thesequasiparticles is well defined. In this basis, the braiding
ma-
FIG. 5. �Color online� Space-time paths corresponding to
theinitialization, manipulation through braiding, and measurement
ofan encoded qubit. Two quasiparticle-quasihole pairs are pulled
outof the vacuum, with each pair having total q-spin 0. The
resultingstate corresponds to a four-quasiparticle qubit in the
state �0L �seeFig. 4�a��. After some braiding, the qubit is
measured by trying tofuse the bottommost pair �in this case a
quasiparticle-quasiholepair�. If they fuse back into the vacuum the
result of the measure-ment is �0L; otherwise it is �1L. Because
only the three lowerquasiparticles are braided, the encoded qubit
can also be viewed asa three-quasiparticle qubit �see Fig. 4�b��
which is initialized in thestate �0L.
FIG. 6. �Color online� Elementary three-braids and the
decom-position of a general three-braid into a series of elementary
braids.The unitary operation produced by this braid is computed by
mul-tiplying the corresponding sequence of elementary braid
matrices,1 and 2 �see text� and their inverses, as shown. Here the
�unla-beled� ovals represent a particular basis choice for the
three-quasiparticle Hilbert space, consistent with that used in the
text. Inthis and all subsequent figures which show braids,
quasiparticles arealigned vertically, and we adopt the convention
that reading frombottom to top in the figures corresponds to
reading from left to rightin expressions such as (�• , • �a , • )c
in the text. It should be noted thatthese figures are only meant to
represent the topology of a givenbraid. In any actual
implementation of topological quantum compu-tation, quasiparticles
will certainly not be arranged in a straight line,and they will
have to be kept sufficiently far apart while beingbraided to avoid
lifting the topological degeneracy.
TOPOLOGICAL QUANTUM COMPILING PHYSICAL REVIEW B 75, 165310
�2007�
165310-7
-
trix is simply 1, and so, after changing back to the
originalbasis, we find
2 = F−11F = �− �e
−i�/5 �e−i3�/5�e−i3�/5 − �
ei3�/5
. �8�
The unitary transformation corresponding to a giventhree-braid
can now be computed by representing it as asequence of elementary
braid operations and multiplying thecorresponding sequence of 1 and
2 matrices and their in-verses, as shown in Fig. 6.
If we are only concerned with single-qubit rotations, thenwe
only care about the action of these matrices on the en-coded qubit
space with total q-spin 1, and not the totalq-spin-0 sector
corresponding to the noncomputational state.However, in our
two-qubit gate constructions, various three-braids will be embedded
into the braiding patterns of sixquasiparticles, and in this case
the action on the full three-dimensional Hilbert space does
matter.
To understand this action note that 1 can be written
1 = �±e−i�/10�±e−i7�/10 0
0 ± ei7�/10�
ei3�/5
, �9�
where the upper 2�2 block acting on the total q-spin 1sector is
an SU�2� matrix, �i.e., a 2�2 unitary matrix withdeterminant 1�,
multiplied by a phase factor of either +e−i�/10or −e−i�/10, and the
lower right matrix element ei3�/5 is thephase acquired by the total
q-spin-0 state. The phase factorpulled out of the upper 2�2 block
is only defined up to ±1because any SU�2� matrix multiplied by −1
is also an SU�2�matrix.
From �8� it follows that 2 can be written in a similarfashion,
with the same phase factors. Each clockwise braid-ing operation
then corresponds to applying an SU�2� opera-tion multiplied by a
phase factor of ±e−i�/10 to the q-spin-1sector, while at the same
time multiplying the q-spin-0 sectorby a phase factor of ei3�/5.
Likewise, each counterclockwisebraiding operation corresponds to
applying an SU�2� opera-tion multiplied by a phase factor of
±e+i�/10 to the q-spin-1sector and a phase factor of e−i3�/5 to the
q-spin-0 sector.
We define the winding W�B� of a given three-braid B tobe the
total number of clockwise interchanges minus the totalnumber of
counterclockwise interchanges. It then followsthat the unitary
operation corresponding to an arbitrary braidB can always be
expressed
U�B� = �±e−iW�B��/10�SU�2��ei3W�B��/5
� , �10�where �SU�2�� indicates an SU�2� matrix. Thus, for a
giventhree-braid, the phase relation between the total q-spin-1
andtotal q-spin-0 sectors of the corresponding unitary operationis
determined by the winding of the braid. We will refer to�10� often
in what follows. It tells us precisely what unitaryoperations can
be approximated by three-braids, and placesuseful restrictions on
their winding.
B. Weaving and brute force search
At this point it is convenient to restrict ourselves to
asubclass of braids which we will refer to as weaves. A weaveis any
braid that is topologically equivalent to the space-timepaths of
some number of quasiparticles in which only asingle quasiparticle
moves. It was shown in Ref. 41 that thisrestricted class of braids
is universal for quantum computa-tion, provided the unitary
representation of the braid group isdense in the space of all
unitary transformations on the rel-evant Hilbert space, which is
the case for Fibonacci anyons.
Following Ref. 41 we will borrow some weaving termi-nology and
refer to the mobile quasiparticle �or collection ofquasiparticles�
as the “weft” quasiparticle�s� and the staticquasiparticles as the
“warp” quasiparticles.
One reason for focusing on weaves is that weaving willlikely be
easier to accomplish technologically than generalbraiding. This is
true even if the full computation involvesnot just weaving a single
quasiparticle, as was proposed inRef. 41, but possibly weaving
several quasiparticles at thesame time in different regions of the
computer—carrying outquantum gates on different qubits in
parallel.
Considering weaves has the added �and more immediate�benefit of
simplifying the problem of numerically searchingfor three-braids
which approximate desired gates. For the fullbraid group, even on
just three strands, there is a great dealof redundancy since braids
which are topologically equiva-lent will yield the same unitary
operation. Weaves, however,naturally provide a unique
representation in which the warpstrands are straight, and the weft
weaves around them. Thereis therefore no trivial double counting of
topologicallyequivalent weaves when one does a brute force
numericalsearch of weaves up to some given length.
The unitary operations performed by weaving three
qua-siparticles in which the weft quasiparticle starts and ends
inthe middle position will always have the form
Uweave��ni�� = 1nm2
nm−1¯ 1
n32n21
n1. �11�
Here the sequence of exponents n2 ,n3 , . . . ,nm−1 all take
theirvalues from �±2, ±4�, and n1 and nm can take the values�0, ±2,
±4�. Because these exponents are all even, each fac-tor in this
sequence takes the weft quasiparticle all the wayaround one of the
two warp quasiparticles either once ortwice with either a clockwise
or counterclockwise sense, re-turning it to the middle position. We
allow n1 and nm to be 0to account for the possibility that the
initial or final weavingoperations could each be either 1
n or 2n with n= ±2 or ±4.
Note that we need only consider exponents ni up to ±4
�i.e.,moving the weft quasiparticle at most two times around awarp
quasiparticle� because of the fact that i
10=1 for Fi-bonacci anyons, implying, e.g., i
6=i−4. We define the
length L of such weaves to be equal to the total number
ofelementary crossings; thus L=�i=1
m �ni�.We will also consider weaves in which the weft
quasipar-
ticle begins and/or ends at a position other than the
middle.These possibilities can easily be taken into account by
mul-tiplying Uweave��ni��, as defined in �11�, by the
appropriatefactors of 1 or 2 on the right and/or left. Thus, for
ex-ample, the unitary operation produced by a weave in which
HORMOZI et al. PHYSICAL REVIEW B 75, 165310 �2007�
165310-8
-
the weft quasiparticle starts in the top position and ends inthe
middle position can be written Uweave��ni��2, where, be-cause of
the extra factor of 2, the first braiding operationscarried out by
this weave will be 2
n where n is an odd power,n= ±1, ±3 or 5. This will weave the
weft quasiparticle fromthe top position to the middle position
after which Uweave willsimply continue weaving this quasiparticle
eventually endingwith it in the middle position. �Note that by
multiplyingUweave on the right by 2, and not 2
−1, we are not requiringthe initial elementary braid to be
clockwise, since Uweave mayhave n1=0 and n2=−2 or −4 so that the
initial 2 is imme-diately multiplied by 2 to a negative power.�
Similarly, theunitary operation produced by a weave in which the
weftparticle starts in the top position and ends in the
bottomposition can be written 1Uweave��ni��2, and so on.
To find a weave for which the corresponding unitary op-eration
Uweave��ni�� approximates a particular desired unitaryoperation,
the most straightforward approach is to simplyperform a brute force
search over all weaves, i.e., all se-quences �ni� as described
above, up to a certain length L, inorder to find the Uweave��ni��
which is closest to the targetoperation. Here we will take as a
measure of the distancebetween two operators U and V the operator
norm distance
�U ,V�= �U−V� where �O� is the operator norm, defined tobe the
square root of the highest eigenvalue of O†O. Again, ifwe are
interested in fixing the relative phase of the totalq-spin-1 and
total q-spin-0 sectors then we would restrict thewinding of the
weaves so that the phases in �10� match thoseof the desired target
gate.
For example, imagine our goal is to find a weave
whichapproximates the unitary operation
iX = �0 ii 01
. �12�
If the resulting weave were to be used only for a
single-qubitoperation, then we would only require that the weave
ap-proximate the upper left 2�2 block of iX up to an overallphase
and we would not care about the phase factor appear-ing in the
lower right matrix element. There would then beno constraint on the
winding of the braid. However, for thisexample we will assume that
this weave will be used in atwo-qubit gate construction, for which
the overall phaseand/or the phase difference between the total
q-spin-1 andtotal q-spin-0 sectors will matter.
In this case, by comparing iX to �10�, we see that thewinding W
of any weave approximating iX must satisfyei3�W/5=1 or W=0 �modulo
10�. Results of a brute forcesearch over weaves satisfying this
winding requirementwhich approximate iX are shown in Fig. 7. In
this figure,ln�1/� is plotted vs braid length L, where is the
minimumdistance between Uweave and iX for weaves of length L. It
isexpected that, for any such brute force search for
weavesapproximating a generic target operation, the length
shouldscale with distance according to L� log�1/�, because
thenumber of braids grows exponentially with L. The resultsshown in
Fig. 7 are consistent with such logarithmic scaling.
All the brute force searches used to find braids in thispaper
are straightforward sequential searches, meant mainlyto demonstrate
proof of principle. No doubt more sophisti-cated brute force search
methods �e.g., bidirectional search�could be used to perform deeper
searches resulting in longerand more accurate braids. Nevertheless,
the exponentialgrowth in the number of braids with L implies that
findingoptimal braids by any brute force search method will
rapidlybecome infeasible as L increases. Fortunately one can
stillsystematically improve a given braid to any desired accuracyby
applying the Solovay-Kitaev algorithm,42,43 which wenow briefly
review.
C. Implementation of the Solovay-Kitaev algorithm for braids
The general result of the Solovay-Kitaev theorem tells usthat we
can efficiently improve the accuracy of any givenbraid without the
need to perform exhaustive brute forcesearches of ever improving
accuracy.42,43 The essential ingre-dient in this procedure is an
-net—a discrete set of operatorswhich in the present case
correspond to finite braids up tosome given length, with the
property that for any desiredunitary operator there exists an
element of the -net that iswithin some given distance 0 of that
operator. Provided 0 issufficiently small, the Solovay-Kitaev
algorithm gives us aclever way to pick a finite number of braid
segments out ofthe -net and sew them together so that the resulting
gate willbe an approximation to the desired gate with improved
accu-racy.
The implementation of the Solovay-Kitaev algorithm weuse here
follows closely that described in detail in Refs. 44and 45. The
first step of this algorithm is to find a braidwhich approximates
the desired gate, U, by performing abrute force search over the
-net. Let U0 denote the result ofthis search. Since we know that
�U0 ,U��0 it follows thatC=UU0
−1 is an operator that is within a distance 0 of
theidentity.
The next step is to decompose C as a group commutator.This means
that we find two unitary operators A and B for
FIG. 7. ln�1/� vs braid length L for weaves approximating
thegate iX. Here is the distance �defined in terms of operator
norm�between iX and the unitary transformation produced by a weave
oflength L which best approximates it. The line is a guide to the
eye.
TOPOLOGICAL QUANTUM COMPILING PHYSICAL REVIEW B 75, 165310
�2007�
165310-9
-
which C=ABA−1B−1. The unitary operators A and B are cho-sen so
that their action on the computational qubit spacecorresponds to
small rotations through the same angle butabout perpendicular axes.
For this choice, if A and B are thenapproximated by operators A0
and B0 in the -net, it canreadily be shown that the operator
C0=A0B0A0
−1B0−1, will ap-
proximate C to a distance of order 03/2. It follows that the
operator U1=A0B0A0−1B0
−1U0 is an approximation to U withina distance 1�c0
3/2, where c is a constant which determinesthe size of the -net
needed to guarantee an improvement inaccuracy.
What we have just described corresponds to one iterationof the
Solovay-Kitaev algorithm. Subsequent iterations arecarried out
recursively. Thus, at the second level of approxi-mation each
search within the -net is replaced by the pro-cedure described
above, and so on, so that at the nth level allapproximations are
made at the �n−1�st level. The result ofthis recursive process is a
braid whose accuracy grows su-perexponentially in n, with the
distance to the desired gatebeing of order n��c20��3/2�
nat the nth level of recursion,
while the braid length grows only exponentially in n,
withL�5nL0, where L0 is a typical braid length in the initial
-net. Thus, as the distance of the approximate gate from
thedesired target gate, , goes to zero, the braid length growsonly
polylogarithmically, with L� log��1/� where �=ln 5/ ln�3/2��3.97.
While this scaling is, of course, worsethan the logarithmic scaling
for brute force searching, it isstill only a polylogarithmic
increase in braid length which issufficient for quantum
computation. Similararguments44,45can be used to show that the
classical com-puter time t required to carry out the Solovay-Kitaev
algo-rithm also only scales polylogarithmically in the desired
ac-curacy, with t� log�1/� where =ln 3/ ln�3/2��2.71.
It is worth noting that there is a particularly nice featureof
this implementation of the Solovay-Kitaev algorithmwhen applied to
compiling three-braids. Recall that whencarrying out two-qubit
gates it will be crucial to maintain thephase difference between
the total q-spin-1 and total q-spin-0sectors of the
three-quasiparticle Hilbert space associatedwith a given
three-braid, and, according to �10�, this can bedone by fixing the
winding of the braid �modulo 10�. Be-cause of the group commutator
structure of the Solovay-Kitaev algorithm, the winding of the
nth-level approximationUn will be the same as that of the initial
approximation U0.This is because all subsequent improvements
involve multi-plying this braid by group commutators of the
formAnBnAn
−1Bn−1 which automatically have zero winding. The
phase relationship between the total q-spin-1 and totalq-spin-0
sectors is therefore preserved at every level of
theconstruction.
Figure 8 shows the application of one iteration of
theSolovay-Kitaev algorithm applied to finding a braid
whichgenerates a unitary operation approximating iX. The
braidlabeled U0 is the result of a brute force search with
L=44corresponding to the best approximation shown in Fig. 7.�Note
that although this braid is drawn as a sequence ofelementary braid
operations, it is topologically equivalent toa weave. In fact
precisely this braid, drawn explicitly as aweave, is shown in Fig.
13.� The braids labeled A0 and B0
generate unitary operations which approximate operators Aand B
whose group commutator gives UU0
−1 where U= iX.Finally, the braid labeled U1 is the new, more
accurate, ap-proximate weave.
VI. TWO-QUBIT GATES
We have seen that single-qubit gates are “easy” in thesense that
as long as we braid within an encoded qubit therewill be no leakage
errors �the overall q-spin of the group ofthree quasiparticles will
remain 1�. Furthermore, the space ofunitary operators acting on the
three-quasiparticle Hilbertspace �essentially SU�2�� is small
enough to find excellentapproximate braids by performing brute
force searches andsubsequent improvement using the Solovay-Kitaev
algo-rithm. We now turn to the significantly harder problem
offinding braids which approximate entangling two-qubitgates.
A. Divide and conquer approach
Figure 9 depicts six quasiparticles encoding two qubitsand a
general braiding pattern. To entangle these qubits, qua-siparticles
from one qubit must be braided around quasipar-ticles from the
other qubit, and this will inevitably lead toleakage out of the
encoded qubit space, �i.e., the overallq-spin of the three
quasiparticles constituting a qubit may no
FIG. 8. �Color online� One iteration of the Solovay-Kitaev
al-gorithm applied to finding a braid which approximates the
operationU= iX. The braid U0 is the result of a brute force search
overweaves up to length 44 that best approximate the desired gate
U= iX, with an operator norm distance between U and U0 of
�8.5�10−4. The braids A0 and B0 are the results of similar brute
forcesearches to approximate unitary operations A and B whose
groupcommutator satisfies ABA−1B−1 � UU0
−1. The new braid U1=A0B0A0
−1B0−1U0 is then five times longer than U0, and the accuracy
has improved so that the distance to the target gate is now
1�4.2�10−5. Given the group commutator structure of theA0B0A0
−1B0−1 factor, the winding of the U1 braid is the same as
the
U0 braid. Note that, when joining braids to form U1, it is
possiblethat elementary braid operations from one braid will
multiply theirown inverses in another braid, allowing the total
braid to be short-ened. Here we have left these “redundant” braids
in U1, as thecareful reader should be able to find.
HORMOZI et al. PHYSICAL REVIEW B 75, 165310 �2007�
165310-10
-
longer be 1�. Furthermore, the space of all operators actingon
the Hilbert space of six quasiparticles is much bigger thanfor
three, making brute force searching extremely difficult.Here the
unitary operations acting on this space are inSU�5� � SU�8� �up to
winding-dependent phase factors as in�10��, which has 87 free
parameters as opposed to three forthe three quasiparticle case of
SU�2�.
Still, as a matter of principle, it is possible to perform
abrute force search of sufficient depth so that it corresponds toa
fine enough -net to carry out the Solovay-Kitaev algo-rithm in this
larger space.42 This is essentially the programoutlined in Ref. 5
as an “existence proof” that universalquantum computation is
possible; however, it is not at allclear that, even if one could do
this, it would be the mostefficient procedure for compiling braids.
For the sameamount of classical computing power required to
directlycompile braids in SU�5� � SU�8�, we believe one can
findmuch more efficient braids �in the sense of having a
moreaccurate computation with a shorter braid� by breaking
theproblem into smaller problems, each consisting of finding
aspecific three-braid embedded in the full six-braid space. Aswe
have shown above, these three-braids can then be veryefficiently
compiled.
Here we present two classes of two-qubit gate construc-tions
based on this divide and conquer approach. The first ofthese were
originally introduced by the authors in Ref. 32and are
characterized by the weaving of a pair of quasiparti-cles from one
qubit through the quasiparticles, forming thesecond qubit. The
second class, presented here for the firsttime, can be carried out
by weaving only a single quasipar-ticle from one qubit around one
other quasiparticle from thesame qubit, and two quasiparticles from
the second qubit.
B. Two-quasiparticle weave construction
We now review the two-qubit gate constructions first dis-cussed
in Ref. 32. The basic idea behind these constructionsis illustrated
in Fig. 10. This figure shows two qubits and abraiding pattern in
which a pair of quasiparticles from thetop qubit �the control
qubit� is woven through the quasipar-ticles forming the bottom
qubit �the target qubit�. Throughoutthis braiding the pair is
treated as a single immutable objectwhich, at the end of the braid,
is returned to its originalposition.
If, as in Fig. 10, we choose the pair of weft quasiparticlesto
be the two quasiparticles whose total q-spin determinesthe logical
state of the qubit, then we refer to this pair as thecontrol pair.
We can then immediately see why this construc-tion naturally
suggests itself. If the control qubit is in thestate �0L the
control pair will have total q-spin 0, and weav-ing this pair
through the target qubit will have no effect. Weare thus guaranteed
that if the control qubit is in the state �0Lthe identity operation
is performed on the target qubit.
The only nontrivial effect of this weaving pattern occurswhen
the control qubit is in the state �1L. In this case, thecontrol
pair has total q-spin 1 and so behaves as a singleFibonacci anyon.
The problem of constructing a two-qubitcontrolled gate then
corresponds to finding a weaving patternin which a single Fibonacci
anyon weaves through the threequasiparticles of the target qubit,
inducing a transition on thisqubit without inducing leakage error
out of the computa-tional qubit space, or at least keeping such
leakage as smallas required for a particular computation. This
reduces theproblem of finding a two-qubit gate to that of finding a
weav-ing pattern in which one Fibonacci anyon weaves aroundthree
others—a problem involving only four Fibonaccianyons. However,
following our divide and conquer philoso-phy, we will further
narrow our focus to weaving a singleFibonacci anyon through only
two others at a time.
We define an “effective braiding” weave to be a woventhree-braid
in which the weft quasiparticle starts at the topposition, and
returns to the top position at the end of theweave, with the
requirement that the unitary transformationit generates be
approximately equal to that produced by mclockwise interchanges of
the two warp quasiparticles. Tofind such weaves we perform a brute
force search, as out-lined in Sec. V, over sequences �ni� which
approximatelysatisfy
2Uweave��ni��2 � 1m. �13�
If both sides of this equation are expressed using �10�
itbecomes evident that the winding of any effective braiding
FIG. 9. �Color online� Two encoded qubits and a generic
braid.Because quasiparticles are braided outside of their starting
qubitsthese braids will generally lead to leakage out of the
computationalqubit space, i.e., the q-spin of each group of three
quasiparticlesforming these qubits will in general no longer be
1.
FIG. 10. �Color online� A two-qubit gate construction in which
apair of quasiparticles from the top �control� qubit is woven
throughthe bottom �target� qubit. The mobile pair of quasiparticles
is re-ferred to as the control pair and has a total q-spin of 0 if
the controlqubit is in the state �0L, and 1 if the control qubit is
in the state �1L.Since weaving an object with total q-spin 0 yields
the identity op-eration, this construction is guaranteed to result
in a transformationof the target qubit state only if the control
qubit is in the state �1L.Note that in this and subsequent figures
world-lines of mobile qua-siparticles will always be dark blue.
TOPOLOGICAL QUANTUM COMPILING PHYSICAL REVIEW B 75, 165310
�2007�
165310-11
-
weave must satisfy W=m �modulo 10�. Since the weft par-ticle
starts and ends in the top position, W must be even; thuseffective
braiding weaves exist only for even m.
An example of an m=2 effective braiding weave foundthrough a
brute force search is shown in Fig. 11. The corre-sponding unitary
operation approximates that of interchang-ing the two warp
quasiparticles twice to a distance �10−3.�This is a typical
distance for a woven three-braid of lengthL�46 which approximates a
desired operation—precise dis-tances of approximate weaves are
given in the figure cap-tions.� As for all approximate weaves
considered here, theSolovay-Kitaev algorithm outlined in Sec. V C
can be usedto improve the accuracy of this weave so that can be
madeas small as required with only a polylogarithmic increase
inlength.
The construction of a two-qubit gate using this
effectivebraiding weave is also shown in Fig. 11. In this
constructionthe control pair is woven through the top two
quasiparticlesof the target qubit using this weave. As described
above, ifthe control qubit is in the state �0L, the control pair
hasq-spin 0 and the target qubit is unchanged. But, if the
controlqubit is in the state �1L, the control pair has q-spin 1 and
theaction on the target qubit is approximately equivalent to thatof
interchanging the top two quasiparticles twice, with
theapproximation becoming more accurate as the length of
theeffective braiding weave is increased, either by deeper
bruteforce searching or by applying the Solovay-Kitaev algo-rithm.
Because this effective braiding all occurs within anencoded qubit,
leakage errors can be reduced to zero in thelimit →0. The resulting
two-qubit gate is then a controlled-2
2 gate which corresponds to controlled rotation of the
targetqubit through an angle of 6� /5.
Unfortunately, due to the even m constraint, it is impos-sible
to find an effective braiding gate which corresponds toa controlled
� rotation of the target qubit. Such a gate wouldbe equivalent to a
controlled-NOT gate up to single-qubitrotations.43 Nonetheless, it
is known that any entangling two-qubit gate, when combined with the
ability to carry out arbi-trary single-qubit rotations, forms a
universal set of quantumgates.46 Thus, the efficient compilation of
single-qubit opera-tions described in Sec. V and the effective
braiding construc-
tion just given provide direct procedures for compiling
anyquantum algorithm into a braid to any desired accuracy.
Although it can be used to form a universal set of gates,this
effective braiding construction is still rather restrictive. Itis
clearly desirable to be able to directly compile acontrolled-NOT
gate into a braid. We now give a constructionwhich can be used to
efficiently compile any arbitrary con-trolled rotation of the
target qubit—including a controlled-NOT gate. This construction is
based on a class of woventhree-braids which we call “injection
weaves.”
In an injection weave the weft quasiparticle again starts atthe
top position but in this case ends at a different position.At the
same time we require that the unitary operation gen-erated by this
weave approximate the identity. Thus the ef-fect of an injection
weave is to permute the quasiparticlesinvolved without changing any
of the underlying q-spinquantum numbers of the system.
Comparing the identity matrix to �10� we see that anythree-braid
approximating the identity must have windingW=0 �modulo 10�. The
fact that this winding must be evenimplies that the final position
of the weft particle must be atthe bottom of the weave. Thus
injection weaves correspondto sequences �ni� which approximately
satisfy the equation
1Uweave��ni��2 � �1 00 11
. �14�
An injection weave obtained through brute force search isshown
in Fig. 12. The unitary operation produced by thisweave
approximates the identity operation to a distance
�10−3.
Our two-qubit gate construction based on injection weav-ing is
carried out in three steps. In the first step, also shownin Fig.
12, the control pair is woven into the target qubitusing the
injection weave. If the control pair has total q-spin1 �the only
nontrivial case� the effect of this weave is merelyto replace the
middle quasiparticle of the target qubit with
FIG. 11. �Color online� An effective braiding weave, and a
two-qubit gate constructed using this weave. The effective
braidingweave is a woven three-braid which produces a unitary
operationwhich is a distance �2.3�10−3 from that produced by
simplyinterchanging the two target particles �1
2�. When the control pair iswoven through the target qubit using
this weave the resulting two-qubit gate approximates a
controlled-�2
2� gate to a distance
�1.9�10−3 or 1.6�10−3 when the total q-spin of the two qubits
is0 or 1, respectively.
FIG. 12. �Color online� An injection weave, and step 1 in
ourinjection-based gate construction. The box labeled I represents
anideal �infinite� injection weave which is approximated by the
weaveshown to a distance �1.5�10−3. In step 1 of our gate
construc-tion, this injection weave is used to weave the control
pair into thetarget qubit. If the control qubit is in the state �1L
then a=1 and theresult is to produce a target qubit with the same
quantum numbersas the original, but with its middle quasiparticle
replaced by thecontrol pair.
HORMOZI et al. PHYSICAL REVIEW B 75, 165310 �2007�
165310-12
-
the control pair. Because the unitary operation approximatedby
the injection weave is the identity, in the →0 limit thisinjection
is accomplished without changing any of the q-spinquantum numbers.
The injected target qubit is therefore �ap-proximately� in the same
quantum state as the original targetqubit.
In the second step of our construction, illustrated in Fig.13,
we carry out an operation on the injected target qubit bysimply
weaving the control pair within the target. Becausefor a=1 all of
this weaving takes place within the injectedtarget qubit, there
will be no leakage error �again, strictlyspeaking, only in the
limit of an exact injection weave�. Theonly constraint on this
weave is that the control pair mustboth start and end in the middle
position, and so it must haveeven winding.
If our goal is to produce a gate which is equivalent to
acontrolled-NOT gate up to single-qubit rotations then we mustapply
a � rotation to the target qubit. Unfortunately, thiscannot be
accomplished by any finite weave with even wind-ing, so we must
again consider approximate weaves. Figure
13 shows the control pair being woven through the injectedtarget
qubit using a weave found by a brute force searchwhich approximates
a particular � rotation—the operator iXdefined in �12�—to a
distance �10−3 �this is, in fact, thesame weave shown at the top of
Fig. 8�.
The third step in our construction is the extraction of
thecontrol pair from the target qubit. This is accomplished,
asshown in Fig. 14, by applying the inverse of the injectionweave
to the control pair. The effect of this extraction is torestore the
control qubit to its original state, and replace thecontrol pair
inside the target qubit with the quasiparticlewhich originally
occupied that position.
The full construction is summarized in Fig. 15, whichprovides a
recipe for compiling a controlled-NOT gate into atwo-quasiparticle
weave. A quantum circuit showing that acontrolled-NOT gate is
equivalent to a controlled-�iX� gateand a single-qubit operation is
shown in the top part of thefigure. The single-qubit operation can
be compiled to what-ever accuracy is required following Sec. V, and
thecontrolled-�iX� gate can be decomposed into injection, iX,and
inverse injection operations, as is also shown in the top
FIG. 13. �Color online� A weave that approximates iX �see
Eq.�12��, and step 2 in our injection-based construction. The box
la-beled iX represents an ideal �infinite� iX weave which is
approxi-mated by the weave shown to a distance =8.5�10−4 �this is
thesame weave that appears at the top of Fig. 8�. In step 2 of our
gateconstruction the control pair is woven within the injected
targetqubit, following this weave, in order to carry out an
approximate iXgate when a=1, as shown.
FIG. 14. �Color online� An inverse injection weave and step 3
inour injection-based construction. The box labeled I−1 represents
anideal �infinite� inverse injection weave which is approximated
bythe inverse of the injection weave shown in Fig. 12, again to
adistance �1.5�10−3. This weave is used to extract the controlpair
out of the injected target qubit and return it to the control
qubit,as shown.
FIG. 15. �Color online� Injection-weave based compilation of a
controlled-NOT gate into a braid. A controlled-NOT gate can be
expressedas a controlled-�iX� gate and a single-qubit operation
R�−� /2ẑ�=exp�i�z /4� acting on the control qubit. The
single-qubit rotation can becompiled following the procedure
outlined in Sec. V, and the controlled-�iX� gate can be decomposed
into ideal injection �I�, iX, and inverseinjection �I−1� operations
which can be similarly compiled. The full approximate
controlled-�iX� braid obtained by replacing I, iX, and I−1with the
weaves shown in the previous three figures is shown at bottom. The
resulting gate approximates a controlled-�iX� gate to a
distance
�1.8�10−3 and 1.2�10−3 when the total q-spin of the two qubits
is 0 or 1, respectively.
TOPOLOGICAL QUANTUM COMPILING PHYSICAL REVIEW B 75, 165310
�2007�
165310-13
-
part of the figure. These operations can then all be
similarlycompiled following Sec. V.
The full braid shown at the bottom of Fig. 15 correspondsto
using the approximate woven three-braids shown in Figs.12–14 to
carry out a controlled-�iX� gate. In this braid, if thecontrol
qubit is in the state �0L the control pair has totalq-spin 0 and
the resulting unitary transformation is exactlythe identity.
However, if the control qubit is in the state �1Lthe control pair
has total q-spin 1 and behaves like a singleFibonacci anyon. This
pair is then woven into the target qubitusing an injection weave,
woven within the target in order tocarry out the iX operation, and
finally woven out of the targetand back into the control qubit
using the inverse of the in-jection weave. The resulting gate is
therefore a controlled-�iX� gate.
By replacing the iX weave with an even winding weavewhich
carries out an arbitrary operation U this constructionwill give a
controlled-U gate. The only restriction on U isthat its overall
phase must be consistent with �10� with evenwinding W. However,
this phase can be easily set to anydesired value by applying the
appropriate single-qubit rota-tion to the control qubit, as in Fig.
15.
Finally, note that at no point in either the effective braid-ing
or injection weave constructions described above did wemake
reference to the total q-spin of the two qubits involved.It follows
that, in the limit of exact effective braiding orinjection weaves,
the action of the corresponding two-qubitgates on the computational
qubit space does not depend onthe state of the external fusion
space associated with theq-spin-1 quantum numbers of each qubit
�see Sec. IV�. Thesegates will therefore not entangle the
computational qubitspace with this external fusion space.
C. One-quasiparticle weave constructions
We now show that two-qubit gates can be carried out withonly a
single mobile quasiparticle. This possibility followsfrom the
general result of Ref. 41 that for any system ofnon-Abelian
quasiparticles in which general braids are uni-versal for quantum
computation �such as Fibonacci anyons�,single quasiparticle weaves
are universal as well. However,the “proof of principle” weaves
constructed in that workwere extremely inefficient—involving a huge
number of ex-cess operations. Here we show how to efficiently
construct asingle-quasiparticle weave corresponding to a
controlled-NOT gate �up to single-qubit rotations�.
Our construction is based on a class of weaves that aresimilar
to injection weaves in that they can be used to swaptwo q-spin-1
objects—where one object is a pair of Fi-bonacci anyons with total
q-spin 1 and the other object is asingle Fibonacci anyon—while
acting effectively as the iden-tity operation so that none of the
other q-spin quantum num-bers of the system are disturbed. However,
unlike injectionweaves, this new class of weaves accomplishes this
swapwithout moving the pair as a single object, and in fact can
becarried out by moving just one quasiparticle.
The class of weaves we seek are those that approximatethe
transformation
U„�•, • �a, • …c = ei�„•,�•, • �a…c, �15�
where � is an overall �irrelevant� phase that does not dependon
a or c. The relevant case for showing the similarity with
injection is when a=1, for which the initial and final states
in�15� consist of two q-spin-1 objects—a single Fibonaccianyon and
a pair of Fibonacci anyons with total q-spin 1. Ifboth these
objects are represented as single Fibonacci anyonsthen �15� can be
written U�• , • �c=ei��• , • �c. In this represen-tation U
therefore acts effectively as the identity operation�times an
irrelevant phase�, similar to injection.
Using the F matrix �5� to expand the right-hand side of�15� in
the (�• , • � , • ) basis yields
U„�•, • �a, • …c = ei��b
Fabc„�•, • �b, • …c. �16�
Comparing this with the action of a unitary operation U
withmatrix representation
U = �U001 U01
1
U101 U11
1
U110 , �17�
on the state (�• , • �a , • )c,
U„�•, • �a, • …c = �b
Uabc„�•, • �b, • …c, �18�
we see that the matrix representation of the U we seek
isprecisely the F matrix �up to a phase�: U=ei�F. While the Fmatrix
describes a “passive” operation, i.e., a change of ba-sis, the
operator U can be viewed as an “active” F operationwhich acts
directly on the states of the Hilbert space. Notethat, since F=F−1,
we also have
U„•,�•, • �a…c = ei�„�•, • �a, • …c. �19�
We will refer to weaves that approximate the operation�15� �and
thus also �19�� as F weaves. As we have seen, theunitary operation
U produced by an F weave need only ap-proximate the F matrix �5� up
to an overall irrelevant phase.To be consistent with �10� this
phase must be −1, as can beseen by writing the matrix −F as
FIG. 16. �Color online� An F weave, and step 1 of
ourF-weave-based two-qubit gate construction. The box labeled F
rep-resents an ideal �infinite� F weave which is approximated by
theweave shown to a distance �3.1�10−3. Applying the F weave tothe
initial two-qubit state, as shown, produces an intermediate
statewith q-spins labeled a and b� which depend simply on a and
b—theinitial states of the two qubits �see Table I�.
HORMOZI et al. PHYSICAL REVIEW B 75, 165310 �2007�
165310-14
-
− F = �±i� ±i� ± i�
±i� � i� �− 1
, �20�where a factor of ±i has been pulled out of the upper left
2�2 block, leaving an SU�2� matrix �det=�2+�=1�. Compar-ing �20�
with �10�, it is also evident that any F weave musthave winding W=5
�modulo 10�, which is necessarily odd.
The fact that F weaves must have an odd number of wind-ings
implies that if the weft quasiparticle starts at the topposition of
the weave it must end at the middle position. Forthis choice the F
weave must then approximately satisfy theequation
Uweave��ni��2 � − F . �21�
The result of a brute force search for an F weave
whichapproximates the operation −F to a distance �10−3 isshown in
Fig. 16.
The first step in our single-quasparticle weave construc-tion is
the application of an F weave to two qubits, alsoshown in Fig. 16.
Note that in this figure for convenience wehave made a change of
basis on the bottom qubit, so that thepair which determines its
state �the control pair� consists ofthe top two quasiparticles
within it rather than the bottomtwo. There is no loss of generality
in doing so since this justcorresponds to a single-qubit rotation
on the bottom qubit.
With this basis choice the initial state of the two qubits
isdetermined by the q-spins of their respective control pairswhich
are indicated in Fig. 16 as a �top qubit� and b �bottomqubit�.
After carrying out the F weave, taking the middlequasiparticle of
the top qubit as the weft quasiparticle, andweaving it around both
the bottom quasiparticle of the topqubit and the top quasiparticle
of the bottom qubit, the re-sulting state �again, strictly
speaking, only in the limit of anexact F weave� is shown at the end
of the two-qubit weave inFig. 16. From �19� it follows that the
newly positioned weftquasiparticle and the quasiparticle beneath
will have totalq-spin a. When the quasiparticle beneath these two
is alsoincluded, the three quasiparticles form what we will refer
toas the intermediate state (• , �• , • �a)b�, where the total
q-spinof all three quasiparticles, b�, has a well-defined value
pro-vided a and b are well defined, as we now show.
First consider the case a=1. As described above, the ef-fect of
the F weave is then similar to that of the injectionweave from the
previous construction—it replaces the top-
most quasiparticle in the bottom qubit with a pair of
quasi-particles with q-spin 1, and the bottom-most pair of
quasi-particles in the top qubit �which also has total q-spin 1�
witha single quasiparticle, without changing any of the otherq-spin
quantum numbers of the system. In the limit of anideal F weave,
this means that the b quantum number doesnot change after this swap
and so b�=b. The case a=0 issimpler, since in this case the
intermediate state is(• , �• , • �0)b� for which the fusion rules
�2� imply b�=1, re-gardless of the value of b. The resulting
dependence of b� ona and b is summarized in Table I.
Having used the F weave to create the intermediate state(• , �•
, • �a)b�, the next step in our construction is the applica-tion of
a weave which performs an operation on this statewhich does not
change a and b� but which does yield an a-and b�-dependent phase
factor. After carrying out such aweave, which we will refer to as a
phase weave, we can thenapply the inverse of the F weave to restore
the two qubits totheir initial states a and b.
For any phase weave we will require that the weft quasi-particle
both start and end in the top position so that whenwe join it to
the F weave and its inverse there will be a singleweft
quasiparticle throughout the entire gate construction.
TABLE I. Values of b� for different values of a and b
afterapplying the F weave as shown in Fig. 16, and the phase
applied tothe resulting state by a phase weave with zero winding.
The valueof b� is determined by the fact that b�=1 when a=0 and
b�=b whena=1, as shown in the text.
a b b� Phase factor
0 0 b�=1 1 ei�
0 1 1 ei�
1 0 b�=b 0 1
1 1 1 e−i�
FIG. 17. �Color online� A phase weave with �=� �see text�which
gives a � phase shift to the intermediate state when b�=1,and step
2 of our F-weave-based construction. The box labeled Prepresents an
ideal �infinite� �=� phase weave which is approxi-mated by the
weave shown to a distance �1.9�10−3. Applyingthis phase weave to
the intermediate state created by the F weave,as shown, results in
a b�-dependent � phase shift �see Table I with�=��.
FIG. 18. �Color online� An inverse F weave and step 3 in our
Fweave construction. The box labeled F−1 is an ideal �infinite�
in-verse F weave which is approximated by the inverse of the F
weaveshown in Fig. 16, again to a distance �3.1�10−3. By
applyingthe inverse F weave to the state obtained after applying
the phaseweave, as shown, the two qubits are returned to their
initial states,but now with an a- and b-dependent phase factor �see
Table I�.
TOPOLOGICAL QUANTUM COMPILING PHYSICAL REVIEW B 75, 165310
�2007�
165310-15
-
The phase weave must therefore have even winding, andwith no
loss of generality we can consider the case for whichthe winding
satisfies W=0 �modulo 10�. The unitary opera-tion produced by such
a phase weave must then approxi-mately satisfy the equation
2Uweave��ni��2 � F�ei� 0
0 e−i�
1
F−1, �22�
where the F matrices are needed to change the Hilbert spacebasis
from that in which the operation produced by the phasebraid must be
diagonal �the (• , �• , • �) basis�, to that in whichthe 1 and 2
matrices are defined �the (�• , • � , • ) basis�.
We will see that a phase weave with �=� produces atwo-qubit gate
which is equivalent to a controlled-NOT gateup to single-qubit
rotations. The result of a brute force searchfor such a phase weave
which approximates the desired op-eration to a distance �10−3 is
shown in Fig. 17. This figurealso shows the action of the phase
weave on the intermediatestate produced in Fig. 16. In this weave,
the weft quasiparti-cle is now woven through the two quasiparticles
beneath it,and returns to its original position. Because the phase
weaveproduces a diagonal operation in the basis shown for
theintermediate state, it does not change the values of a and
b�.Its only effect is to give a phase factor of ei� to the state
witha=0 �which necessarily has b�=1� and e−i� to the state witha=1
and b�=1. The state with a=1 and b�=0 is unchanged.These phase
factors are also shown in Table I.
The final step in this construction is to perform the inverseof
the F weave to return the two qubits to their originalstates. This
is shown in Fig. 18. In the limit of exact F andphase weaves, the
resulting operation on the computationalqubit space in the basis
ab= �00,01,10,11� is then
U =�ei� 0 0 0
0 ei� 0 0
0 0 1 0
0 0 0 e−i�
. �23�
If we take the top qubit to be the control qubit and the bot-tom
qubit to be the target qubit, then this gate corresponds,up to an
irrelevant overall phase, to a controlled-�e−i3�/2ei�z/2�
operation. For the case �=� this is acontrolled-�−Z� gate �where
Z=z�, i.e., a controlled-phase
FIG. 19. �Color online� F-weave-based compilation of a
controlled-NOT gate into a braid. A controlled-NOT gate is
equivalent to acontrolled-�−Z� gate with the single-qubit operation
R�� /2ŷ�=exp�−i�y /4� and its inverse applied to the target qubit
before and after thecontrolled-�−Z�. Again, the single-qubit
operations can be trivially compiled, and the controlled-�−Z� gate
decomposed into ideal F, phase�P�, and inverse F �F−1� weaves which
can be similarly compiled. The full approximate controlled-�−Z�
weave obtained by replacing F, P,and F−1 with the approximate
weaves shown in the previous three figures is shown at the bottom.
The resulting gate approximates acontrolled-�−Z� to a distance
�4.9�10−3 and 3.2�10−3 when the total q-spin of the two qubits is 0
or 1, respectively.
FIG. 20. �Color online� Two four-quasiparticle qubits and
abraiding pattern in which only two quasiparticles from each
qubitare braided. Here the quasiparticles are SU�2�k excitations
withq-spin 1/2. The state of the top qubit is determined by the
totalq-spin of the quasiparticle pairs labeled a and the state of
the bot-tom qubit is determined by the total q-spin of the
quasiparticle pairslabeled b. The overall q-spin of the four
braided quasiparticles is d�a dashed oval is used because when
a=b=1 these quasiparticleswill not be in a q-spin eigenstate�. For
this braid to produce noleakage errors, the unitary operation it
generates must be diagonalin a and b, though it can, of course,
result in an a- and b-dependentphase factor. For k�3, d can take
the values 0, 1, or 2, while fork=3 the only allowed values for d
are 0 and 1. The existence of thed=2 state for k�3 makes it
impossible to carry out an entanglingtwo-qubit gate by braiding
only four quasiparticles �see text�.
HORMOZI et al. PHYSICAL REVIEW B 75, 165310 �2007�
165310-16
-
gate, which, up to single-qubit rotations, is equivalent to
acontrolled-NOT gate.
The full F-weave-based gate construction is summarizedin Fig.
19. A quantum circuit showing a controlled-NOT gatein terms of a
controlled-�−Z� gate and two single-qubit op-erations is shown in
the top part of the figure. As in ourinjection based construction,
the single-qubit operations canbe compiled to whatever accuracy is
required following theprocedure outlined in Sec. V. The
controlled-�−Z� gate canthen be decomposed into ideal F, phase, and
inverse Fweaves as is also shown in the top part of the figure.
Woventhree-braids which approximate these operations can then
becompiled to whatever accuracy is required, again followingSec. V.
The full controlled-�−Z� weave corresponding to us-ing the
approximate F and phase weaves shown in Figs.16–18 is shown in the
bottom part of the figure.
Finally, in this construction, as for the constructions
de-scribed in Sec. VI B, we at no point made reference to thetotal
q-spin of the two qubits involved. Thus, in the limit ofexact F and
phase weaves, the action of the two-qubit gatesconstructed here
will not entangle the computational qubitspace with the external
fusion space associated with theq-spin 1 quantum numbers of each
qubit.
VII. WHAT IS SPECIAL ABOUT k=3?
All of the gate constructions discussed in this paper ex-ploit
the fact that the braiding and fusion properties of a pairof
Fibonacci anyons are either trivial if their total q-spin is 0,or
equivalent to those of a single Fibonacci anyon if theirtotal
q-spin is 1. The fact that these are the only two possi-bilities is
a special property of the Fibonacci anyon model,and hence also the
SU�2�3 model, given their eff