Encoding a qubit in an oscillator∗

Daniel Gottesman,(1,2)† Alexei Kitaev,(1)‡ and John Preskill(3)§

(1)
Microsoft Corporation, One Microsoft Way, Redmond, WA 98052, USA
(2)
Computer Science Division, EECS, University of California, Berkeley, CA 94720, USA
(3)
Institute for Quantum Information, California Institute of Technology, Pasadena, CA 91125, USA

analog classical information seems to pose insuperable

Quantum error-correcting codes are constructed that em- difficulties.
bed a finite-dimensional code space in the infinite-dimensional
In this paper we achieve a more modest goal: we
Hilbert space of a system described by continuous quantum
describe quantum error-correcting codes that protect a
variables. These codes exploit the noncommutative geometry
arXiv:quant-ph/0008040v3 13 May 2001

state of a finite-dimensional quantum system (or “qu-

of phase space to protect against errors that shift the values
dit”) that is encoded in an infinite-dimensional system.
of the canonical variables q and p. In the setting of quantum
optics, fault-tolerant universal quantum computation can be
These codes may be useful for implementing quantum
executed on the protected code subspace using linear opti- computation and quantum communication protocols that
cal operations, squeezing, homodyne detection, and photon use harmonic oscillators or rotors that are experimentally
counting; however, nonlinear mode coupling is required for accessible.
the preparation of the encoded states. Finite-dimensional We also explain how encoded quantum states can be
versions of these codes can be constructed that protect en- processed fault tolerantly. Once encoded states have
coded quantum information against shifts in the amplitude been prepared, a universal set of fault-tolerant quan-
or phase of a d-state system. Continuous-variable codes can tum gates can be implemented using, in the language
be invoked to establish lower bounds on the quantum capacity of quantum optics, linear optical operations, squeezing,
of Gaussian quantum channels. homodyne detection, and photon counting. However, for
preparation of the encoded states, nonlinear couplings
must be invoked.
I. INTRODUCTION Our continuous-variable quantum error-correcting
codes are effective in protecting against sufficiently weak
Classical information can be carried by either a dis- diffusive phenomena that cause the position and momen-
crete (digital) signal or a continuous (analog) signal. Al- tum of an oscillator to drift, or against losses that cause
though in principle an analog signal can be processed, the amplitude of an oscillator to decay. By concatenat-
digital computing is far more robust – a digital signal ing with conventional finite-dimensional quantum codes,
can be readily re-standardized and protected from dam- we can also provide protection against errors that heavily
age caused by the gradual accumulation of small errors. damage a (sufficiently small) subset of all the oscillators
Quantum information can also be carried by either in a code block.
a discrete (finite-dimensional) system like a two-level A different scheme for realizing robust and efficient
atom or an electron spin, or by a continuous (infinite- quantum computation based on linear optics has been
dimensional) system like a harmonic oscillator or a rotor. recently proposed by Knill, Laflamme, and Milburn [3,4].
Even in the finite-dimensional case, quantum information We begin in §II by describing codes that embed an
is in a certain sense continuous – a state is a vector in n-state quantum system in a larger d-state system, and
a Hilbert space that can point in any direction. Never- that protect the encoded quantum information against
theless, we have known for nearly five years that cleverly shifts in the amplitude or phase of the d-state system.
encoded quantum states can be re-standardized and pro- A realization of this coding scheme based on a charged
tected from the gradual accumulation of small errors, or particle in a magnetic field is discussed in §III. Our
from the destructive effects of decoherence due to uncon- continuous-variable codes are obtained in §IV by con-
trolled interactions with the environment [1,2]. sidering a d → ∞ limit. Formally, the code states of
One is tempted to wonder whether we can go still fur- the continuous-variable codes are nonnormalizable states,
ther and protect the quantum state of a system described infinitely squeezed in both position and momentum; in
by continuous quantum variables. Probably this is too §V we describe the consequences of using more realis-
much to hope for, since even the problem of protecting tic approximate code states that are finitely squeezed.
In §VI we outline the theory of more general continuous-
variable codes based on lattice sphere packings in higher-
dimensional phase space.
∗
CALT-68-2273 We discuss in §VII how continuous-variable codes pro-
†
[email protected] tect against quantum diffusion, amplitude damping, and
‡
[email protected] unitary errors. In §VIII, we establish a lower bound on
§
[email protected] the quantum capacity of the Gaussian quantum channel.

1
 We then proceed to develop schemes for fault-tolerant We will now imagine that errors with |a|, |b| small com-
manipulation of encoded quantum information, starting pared to d are common, but errors with large |a| and |b|
in §IX with a discussion of the symplectic operations that are rare. This type of error model could be expected
can be implemented with linear optics and squeezing. In to apply in the case of a continuous quantum variable,
§X we discuss the measurement of the error syndrome which is formally the d → ∞ limit of a qudit. For exam-
and error recovery, which can be achieved with symplec- ple, decoherence causes the position q and momentum p
tic operations and homodyne detection. Completion of of a particle to diffuse with some nonzero diffusion con-
the fault-tolerant universal gate set by means of photon stant. In any finite time interval q and p will drift by
counting is described in §XI, and the preparation of en- some amount that may be small, but is certainly not
coded states is explained in §XII. zero. How can we protect encoded quantum information
Finally, §XIII contains some further remarks about the under these conditions?
physical realization of our coding schemes, and §XIV con- Fortunately, the general “stabilizer” framework [9,10]
tains concluding comments. for constructing quantum codes can be adapted to this
setting. In this framework, one divides the elements of
a unitary operator basis into two disjoint and exhaustive
II. SHIFT-RESISTANT QUANTUM CODES classes: the set E of “likely errors” that we want to pro-
tect against, and the rest, the “unlikely errors.” A code
The main novelty of our new codes is that they are subspace is constructed as the simultaneous eigenspace of
designed to protect against a different type of error than a set of commuting “stabilizer generators,” that generate
has been considered in previous discussions of quantum an Abelian group, the “code stabilizer.” The code can
coding. This distinction is more easily explained if we reverse errors in the set E if, for each pair of errors Ea and
first consider not the case of a continuous quantum vari- Eb , either Ea† Eb lies in the stabilizer group, or Ea† Eb fails
able, but instead the (also interesting) case of a “qudit,” to commute with some element of the stabilizer. (In the
a d-dimensional quantum system. Quantum codes can latter case, the two errors alter the eigenvalues of the gen-
be constructed that encode k protected qudits in a block erators in distinguishable ways; in the former case they
of N qudits, so that the encoded qudits can be perfectly do not, but we can successfully recover from an error of
recovered if up to t qudits are damaged, irrespective of type a by applying either Ea† or Eb† .) In typical discus-
the nature of the damage [5–8]. Error recovery will be sions of quantum coding, E is assumed to be the set of
effective if errors that act on many qudits at once are all tensor products of Pauli operators with weight up to t
rare. More precisely, a general error superoperator act- (those that act trivially on all but at most t qudits). But
ing on N qudits can be expanded in terms of a basis of the same principles can be invoked to design codes that
operators, each of definite “weight” (the number of qu- protect against errors in a set E with other properties.
dits on which the operator acts nontrivially). Encoded Quantum codes for continuous variables have been de-
information is well protected if the error superoperator scribed previously by Braunstein [11] and by Lloyd and
has nearly all its support on operators of weight t or less. Slotine [12]. For example, one code they constructed can
But consider instead a different situation, in which the be regarded as the continuous limit of a qudit code of the
amplitude for an error to occur on each qudit is not small, type originally introduced by Shor in the binary (d = 2)
but the errors are of a restricted type. The possible errors case, an [[N = 9, k = 1, 2t + 1 = 3]] code that protects
acting on a single qudit can be expanded in terms of a single qudit encoded in a block of 9 from arbitrary
a unitary operator basis with d2 elements, the “Pauli damage inflicted on any one of the 9. The 8 stabilizer
operators”: generators of the code can be expressed as

X aZ b, a, b = 0, 1, 2, . . . , d − 1 . (1) Z1 Z2−1 , Z2 Z3−1 , Z4 Z5−1 , Z5 Z6−1 , Z7 Z8−1 , Z8 Z9−1 ,

(X1 X2 X3 ) · (X4 X5 X6 )−1 , (X4 X5 X6 ) · (X7 X8 X9 )−1 , (4)
Here X and Z are generalizations of the Pauli matri-
ces σx and σz , which act in a particular basis {|ji, j = and encoded operations that commute with the stabilizer
0, 1, 2, . . . , d − 1} according to and hence act on the encoded qudit can be chosen to be

X : |ji → |j + 1 (mod d)i , Z̄ = Z1 Z4 Z7 ,

Z : |ji → ω j |ji , (2) X̄ = X1 X2 X3 . (5)

where ω = exp(2πi/d). Note that it follows that In the d → ∞ limit, we obtain a code that is the simul-
taneous eigenspace of eight commuting operators acting
ZX = ωXZ. (3) on nine particles, which are

For N qudits, there is a unitary operator basis with d2N q1 − q2 , q2 − q3 , q4 − q5 , q5 − q6 , q7 − q8 , q8 − q9 ,

elements consisting of all tensor products of single-qudit (p1 + p2 + p3 ) − (p4 + p5 + p6 ) ,
Pauli operators. (p4 + p5 + p6 ) − (p7 + p8 + p9 ) . (6)

2
Logical operators that act in the code space are If an amplitude error occurs that shifts j by ±1, the
error can be diagnosed by measuring the stabilizer gen-
q̄ = q1 + q4 + q7 , erator Z 6 , which reveals the value of j modulo 3; the
p̄ = p1 + p2 + p3 . (7) error is corrected by adjusting j to the nearest multiple
of 3. Phase errors are shifts in the Fourier transformed
This code is designed to protect against errors in which conjugate basis, and can be corrected similarly.
one of the particles makes a large jump in q or p (or both) This code is actually perfect, meaning that each possi-
while the others hold still. But it provides little protec- ble pair of eigenvalues of the generators X 6 and Z 6 is a
tion against small diffusive motions of all the particles, valid syndrome for correcting a shift. There are nine pos-
which allow q̄ and p̄ to drift. sible errors {X a Z b , |a|, |b| ≤ 1}, and the Hilbert space
Entanglement purification protocols for continuous of the qudit contains nine copies of the two-dimensional
variable systems have also been proposed — good entan- code space, one corresponding to each possible error.
gled states can be distilled from noisy entangled states These “error spaces” just barely fit in the qudit space
via a protocol that requires two-way classical commu- for d = 18 = 9 · 2.
nication [13,14]. These purification protocols work well Similar perfect codes can be constructed that protect
against certain sorts of errors, but their reliance on two- against larger shifts. For d = r1 r2 n, consider the stabi-
way classical communication makes them inadequate for lizer generators
accurately preserving unknown states in an imperfect
quantum memory, or for robust quantum computation. X r1 n , Z r2 n . (12)
Returning to qudits, let us consider an example of a There is an encoded qunit, acted on by logical operators
quantum code that can protect against small shifts in
both amplitude and phase, but not against large shifts. X̄ = X r1 ,
It is already interesting to discuss the case of a system Z̄ = Z r2 , (13)
consisting of a single qudit, but where the dimension n
of the encoded system is (of course) less than d. For which evidently commute with the stabilizer and satisfy
example, a qubit (n = 2) can be encoded in a system Z̄ X̄ = ω r1 r2 X̄ Z̄ = e2πi/n X̄ Z̄ . (14)
with dimension d = 18, and protected against shifts by
one unit in the amplitude or phase of the qudit; that is, The commutation relations of the Pauli operators with
against errors of the form X a Z b where |a|, |b| ≤ 1. The the generators are
stabilizer of this code is generated by the two operators (X a Z b ) · X r1 n = ω r1 nb X r1 n · (X a Z b )
X 6, Z6 , (8) = e2πib/r2 X r1 n · (X a Z b ) ,
a b r2 n
(X Z ) · Z = ω̄ r2 na Z r2 n · (X a Z b )
and the commutation relations of the Pauli operators
with these generators are = e−2πia/r1 Z r2 n · (X a Z b ) . (15)

(X a Z b ) · X 6 = ω 6b X 6 · (X a Z b ) , The phases are trivial only if a is an integer multiple of

r1 and b an integer multiple of r2 . Therefore, this code
(X a Z b ) · Z 6 = ω̄ 6a Z 6 · (X a Z b ) . (9)
can correct all shifts with
Therefore, a Pauli operator commutes with the stabilizer r1
|a| < ,
only if a and b are both multiples of 3 = 18/6; this prop- 2
erty ensures that the code can correct single shifts in r2
|b| < . (16)
both amplitude and phase. Logical operators acting on 2
the encoded qudit are The number of possible error syndromes is r1 r2 = d/n,
so the code is perfect.
X̄ = X 3 , Z̄ = Z 3 , (10) Expressed in terms of Z eigenstates, the codewords
which evidently commute with the stabilizer and are not contain only values of j that are multiples of r1 (since
contained in it. Z r2 n = 1), and are invariant under a shift of j by r1 n
Since the codewords are eigenstates of Z 6 with eigen- (since X r1 n = 1). Hence a basis for the n-dimensional
value one, the only allowed values of j are multiples of code subspace is
three. And since there are also eigenstates of X 6 with 1
eigenvalue one, the codewords are invariant under a shift |0̄i = √ (|0i + |nr1 i + . . . + |(r2 − 1)nr1 i) ,
r2
in j by six units. A basis for the two-dimensional code 1
space is |1̄i = √ (|r1 i + . . . + |((r2 − 1)n + 1)r1 i) ,
r2
1 ·
|0̄i = √ (|0i + |6i + |12i) ,
3 ·
1 1
|1̄i = √ (|3i + |9i + |15i) . (11) |n̄ − 1̄i= √ (|(n − 1)r1 i + . . . + |(r2 n − 1)r1 i) . (17)
3 r2

3
If the states undergo an amplitude shift, the value of j ity of about 1/2r2 , and a Z̄ error with a probability of
modulo r1 is determined by measuring the stabilizer gen- about 1/2r1 . The code is more effective if the typical
erator Z r2 n , and the shift can be corrected by adjusting j errors gently deform the state in each cell, rather than
to the nearest multiple of r1 . The codewords have a sim- strongly deforming it in one cell.
ilar form in the Fourier transformed conjugate basis (the
basis of X eigenstates), but with r1 and r2 interchanged.
Therefore, amplitude shifts by less than r1 /2 and phase IV. CONTINUOUS VARIABLE CODES FOR A
shifts by less than r2 /2 can be corrected. SINGLE OSCILLATOR

Formally, we can construct quantum codes for systems

III. A QUDIT IN A LANDAU LEVEL described by continuous quantum variables by consider-
ing the large-d limit of the shift-resistant codes described
A single electron in a uniform magnetic field in two in §II. We might have hoped to increase d to infinity
dimensions provides an enlightening realization of our while holding r1 /d and r2 /d fixed, maintaining the abil-
codes. General translations in a magnetic field are ity to correct shifts in both amplitude and phase that are
noncommuting, since an electron transported around a fixed fraction of the ranges of the qudit. However, since
a closed path acquires an Aharonov-Bohm phase eieΦ , the perfect codes satisfy
where Φ is the magnetic flux enclosed by the path.
Two translations T and S commute only if the opera- r1 1 r2 1
= , = , (18)
tor T ST −1S −1 translates an electron around a path that d nr2 d nr1
encloses a flux Φ = kΦ0 , where Φ0 = 2π/e is the flux
this is not possible. Nonetheless, interesting codes can
quantum and k is an integer.
be obtained as the amplitude and phase of the qudit ap-
Translations commute with the Hamiltonian H, and
proach the position q and momentum p of a particle – we
two translations T1 and T2 form a maximally commuting
can hold fixed the size of the shifts ∆q and ∆p that can
set if they generate a lattice that has a unit cell enclosing
be corrected, as the ranges of q and p become unbounded.
one quantum of flux. Simultaneously diagonalizing H, T1 1
Another option is to take d → ∞ with r1 /d ≡ m fixed
and T2 , we obtain a Landau level of degenerate energy iθ
eigenstates, one state corresponding to each quantum of and r2 = m/n fixed, obtaining a rotor Z = e (or a
magnetic flux. Then T1 and T2n are the stabilizer gen- particle in a periodically identified finite box) that can
1/n be protected against finite shifts in both the orientation
erators of a code, where Z̄ = T1 and X̄ = T2 are the θ of the rotor and its (quantized) angular momentum L.
logical operators on a code space of dimension n. The stabilizer of this code is generated by
Suppose the system is in a periodically identified box
(a torus), so that T1r1 = (T2n )r2 = 1 are translations Z r2 n → eiθm ,
around the cycles of the torus. The number of flux quanta
X r1 n = X d/r2 → e−2πiL(n/m) (19)
through the torus, and hence the degeneracy of the Lan-
dau level, is nr1 r2 . The code, then, embeds an n dimen- and the logical operations are
sional system in a system of dimension d = r1 r2 n.
In this situation, the logical operations X̄ and Z̄ can Z̄ = eiθm/n
be visualized as translations of the torus in two different
X̄ = e−2πiL/m (20)
directions; the stabilizer generator X̄ n is a translation by
a fraction 1/r2 of the length of the torus in one direction, Since X̄ shifts the value of θ by 2π/m, and Z̄ shifts the
and the stabilizer generator Z̄ n is a translation by 1/r1 value of L by m/n = r2 , this code can correct shifts in θ
of the length in the other direction. Therefore, for any with ∆θ < π/m and shifts in L with |∆L| < m/2n.
state in the code space, the wave function of the electron Alternatively, we can consider a limit in which r1 and
in a cell containing n flux quanta is periodically repeated r2 both become large. We may write r1 = α/ε and r2 =
altogether r1 r2 times to fill the entire torus. Our code 1/nαε, where d = nr1 r2 = 1/ε2 , obtaining a code with
can be regarded as a novel kind of “quantum repetition stabilizer generators
code” – identical “copies” of the wave function are stored
in each of r1 r2 cells. But of course there is only one Z r2 n → e2πiqε

(1/αε)
= e2πiq/α ,
electron, so if we detect the electron in one cell its state
(nα/ε)
is destroyed in all the cells. X r1 n → e−ipε = e−inpα , (21)
This picture of the state encoded in a Landau level cau-
tions us about the restrictions on the type of error model and logical operations
that the code can fend off successfully. If the environ-
ment strongly probes one of the cells and detects noth- Z̄ = e2πiq/nα , X̄ = e−ipα , (22)
ing, the wave function is suppressed in that cell. This
causes a X̄ error in the encoded state with a probabil- where α is an arbitrary real number. Using the identity
eA eB = e[A,B] eB eA (which holds if A and B commute

4
with their commutator) and the canonical commutation The Wigner function associated with the codeword
relation [q, p] = i, we verify that wave function ψ (j) (q) ≡ hq|Z̄ = ω j i is a sum of delta
functions positioned at the sites of a lattice in phase
Z̄ X̄ = ω X̄ Z̄ , ω = e2πi/n . (23) space, where three quarters of the delta functions are
positive and one quarter are negative. Explicitly, we have
Since X̄ translates q by α and Z̄ translates p by 2π/nα,
Z ∞
the code protects against shifts with (j) 1
W (q, p) ≡ dx eipx ψ (j) (q + x/2)∗ ψ (j) (q − x/2)
α 2π −∞
|∆q| < , ∞
2 X  π 
π ∝ (−1)st δ p − ·s
|∆p| < . (24) nα
nα s,t=−∞
 nα 
Note that the shifts in momentum and position that the × δ q − αj − ·t ; (28)
code can correct obey the condition 2
π the delta functions are negative on the sublattice with
∆p∆q < h̄ . (25) s, t odd. If we integrate over p, the oscillating sign causes
2n
the terms with odd t to cancel in the sum over s, and the
In typical situations, errors in q and p arepof comparable surviving positive delta functions have support at q =
magnitude, and it is best to choose α = 2π/n so that (n · integer + j)α. If we integrate over q, the terms with
r ! r ! odd s cancel in the sum over t, and the surviving positive
2π 2π delta functions have support at p = (2π/nα) · integer.
Z̄ = exp iq , X̄ = exp −ip . (26)
n n Wigner functions for the X̄ eigenstates are similar, but
with the roles of q and p interchanged.
Formally, the codewords are coherent superpositions of It is also of interest to express the encoded states in
infinitely squeezed states, e.g. (up to normalization) terms of the basis of coherent states. Consider for exam-
∞ ple the encoded state with X̄ = 1, which is the unique
simultaneous eigenstate with eigenvalue one of the oper-
X
|Z̄ = ω j i = |q = α(j + ns)i ,
s=−∞
ators e2πiq/α and e−ipα . In fact starting with any state
∞ |ψi, we can construct the encoded state (up to normal-
X 2π ization) as
|X̄ = ω̄ j i = |p = (j + ns)i . (27)
s=−∞
nα P∞
P∞
e−ispα · 2πitq/α


t=−∞ e |ψi
Ps=−∞
(See Fig. 1.) Of course, realistic codewords will be nor- = s,t exp [i (−spα + 2πtq/α + πst)] |ψi . (29)
malizable finitely squeezed states, rather than nonnor-
malizable infinitely squeezed states. But squeezing in at In particular, if we choose |ψi to be the ground state
least one of p and q is required to comfortably fulfill the |0i P of the oscillator, then the opera-
condition eq. (25). tor s,t exp [i (−spα + 2πtq/α + πst)] displaces it to a
coherent state centered at the point (q, p) = (sα, 2πt/α)
in the quadrature plane. Thus the encoded state is an
equally weighted superposition of coherent states, with
2α centers chosen from the sites of a lattice in the quadra-
| 0〉 : q ture plane whose unit cell has area 2π. Since the coherent
states are overcomplete the expansion is not unique; in-
2α deed, if we choose |ψi to be a coherent state rather than
| 1〉 : q
the vacuum, then the lattice is rigidly translated, but the
2π / α encoded state remains invariant.
| 0〉 + | 1 〉 : p We can envision the stabilizer of the code as a lat-
2π / α
tice of translations in phase space that preserve the code
words, the lattice generated by the translations e2πiq/α
| 0〉 − | 1〉 : p
and e−inpα . In fact, this lattice need not be rectangular
– we can encode an n-dimensional system in the Hilbert
space of a single oscillator by choosing any two variables
Q and P that satisfy the canonical commutation relation
FIG. 1. Codewords of the n = 2 code. The states |0̄i, |1̄i [Q, P ] = i, and constructing the code space as the simul-
are superpositions of q eigenstates, periodically spaced with taneous eigenstate of e2πiQ and e−inP . The unit cell of
period 2α; the two basis states √ differ by a displacement in the lattice has area 2πh̄n, in keeping with the principle
q by α. The states (|0i ± |1i)/ 2 are superpositions of p that each quantum state “occupies” an area 2πh̄ in the
eigenstates, periodically spaced with period 2π/α; the two phase space of a system with one continuous degree of
basis states differ by a displacement in p by π/α. freedom.

5
 V. FINITE SQUEEZING again a superposition of Gaussians weighted by a Gaus-
sian envelope. (See Fig. 2.)
Strictly speaking, our codewords are nonnormalizable
states, infinitely squeezed in both q and p. In practice, Probability
we will have to work with approximate codewords that
will be finitely squeezed normalizable states. We need
to consider how using such approximate codewords will
affect the probability of error.
We will replace a position eigenstate δ(0) by a normal-
ized Gaussian of width ∆ centered at the origin,
Z ∞
dq 1 2 2
|ψ0 i = 2 )1/4
e− 2 q /∆ |qi
−∞ (π∆
Z ∞
dp 1 2 2 q
= 2 )1/4
e− 2 ∆ p |pi . (30) -6 -3 3 6
−∞ (π/∆
A codeword, formally a coherent superposition of an infi- FIG. 2. Probability distribution in position space
P (q) = 21 |hq|(|0̃i + |1̃i)|2 for an approximate codeword with
nite number of δ-functions, becomes a sum of Gaussians
∆ = κ = .25. The dashed line is the distribution’s Gaussian
weighted by a Gaussian envelope function of width κ−1 ;
envelope.
in the special case of a two-dimensional code space, the
approximate codewords become
The approximate codewords |0̃i, |1̃i have a small over-
∞
X
− 12 κ2 (2sα)2
lap if ∆ is small compared to α and κ is small compared
|0̃i = N0 e T (2sα)|ψ0 i , to π/α. For estimating the error probability caused by
s=−∞ the overlap, let’s consider the special case where q and p
∞ √
X 1 2
[(2s+1)α)]2
are treated symmetrically, α = π and κ = ∆; then
|1̃i = N1 e− 2 κ T [(2s + 1)α]|ψ0 i , (31)
s=−∞ ∞
2 X −4π∆2 s2 √
|hq|0̃i|2 ≈ √ e exp[−(q − 2s π)2 /∆2 ]
where T (a) translates q by a, N0,1 are normalization fac- π s=−∞
tors, and we use e.g. |0̃i rather than |0̄i to denote the
approximate codeword. We will assume that κα and (35)
∆/α are small compared to one, so that N0 ≈ N1 ≈
and
(4κ2 α2 /π)1/4 ; then in momentum space, the approximate
codeword becomes e.g., 1 2
hp|0̃i + hp|1̃i
√
 2 2 1/4 Z ∞ 2
κ α dp 1 2 2 ∞
(|0̃i + |1̃i)/ 2 ≈ e− 2 ∆ p 2 X −∆2 p2 √
π −∞ (π/∆ 2 )1/4 ≈ √ e exp[−(p − 2m π)2 /∆2 ] . (36)
∞
π m=−∞
1 2 2
X
× e− 2 κ (sα) eip(αs) |pi . (32)
s=−∞ To perform√error recovery, we measure the value of q and
p modulo π and then correct for the observed shift. In
By applying the Poisson summation formula, the state |0̃i, the probability of failure is the
√ probabil-
∞
X 2
∞
X 2
ity that q is closer to an odd multiple of√ π than an
e−πa(m−b) = (a)−1/2 e−πs /a 2πisb
e , (33) even multiple, and in the state (|0̃i + |1̃i)/ 2, the error
m=−∞ s=−∞ probability √is the probability that p is closer to an odd
multiple of π than an even multiple. For both the am-
this approximate codeword can be rewritten as
plitude and phase errors, the intrinsic error probability
√
 2 2 1/4 Z ∞
κ α dp 1 2 2
arising from the imperfections of the approximate code-
(|0̃i + |1̃i)/ 2 ≈ 2 1/4
e− 2 ∆ p words becomes exponentially small for small ∆. Using
π −∞ (π/∆ )
√ the asymptotic expansion of the error function,
∞
"  2 #
2π X 1 2π 2 Z ∞
× exp − p− m /κ |pi
 
−t2 1 2
κα m=−∞ 2 α e−x 1 − O(1/x2 ) ,


dt e = (37)
x 2x
Z ∞ 1/4 X ∞
4π∆2

dp 1 2 2
= 2 )1/4
· 2
e− 2 ∆ p we may estimate the error probability by summing the
−∞ (πκ α m=−∞ contributions from the tails of all the Gaussians, obtain-
ing
"  2 #
1 2π
× exp − p− m /κ2 |pi , (34)
2 α

6
 ∞
!
∞
2 space of a single particle. We can construct more sophis-
Z
−4π∆2 n2 2
/∆2
X
Error Prob < √ e ·2 √
dq e−q ticated codes from lattices in the (2N )-dimensional phase
π n=−∞ π/2
space of N particles. Then codes of higher quality can
2 1 ∆ 2
be constructed that take advantage of efficient packings
∼ √ · · 2∆ · √ e−π/4∆
π 2∆ π of spheres in higher dimensions.
2∆ −π/4∆2 For a system of N oscillators, a tensor product of Pauli
= e . (38) operators can be expressed in terms of the canonical vari-
π
ables qi and pi as
This error probability is about 1% for ∆ ∼ .5, and is
already less than 10−6 for a still modest value ∆ ∼ .25. N
" !#
√ X
Using finitely squeezed approximate codewords does not Uαβ = exp i 2π αi pi + βi qi , (44)
badly compromise the error-correcting power of the code, i=1
since a gentle spreading in p and q is just the kind of error
the code is intended to cope with. where the αi ’s and βi ’s are real numbers. (In this setting,
The mean photon number of a finitely squeezed ap- the Pauli operators are sometimes called “Weyl opera-
proximate codeword is tors.”) Two such operators commute up to a phase:
′ ′
1 Uαβ Uα′ β ′ = e2πi[ω(αβ,α β )] Uα′ β ′ Uαβ , (45)
†
ha ai + 1/2 = hp2 + q 2 i ≈ ∆−2 (39)
2
where
for small ∆. Therefore, an error probability of order 10−6
can be achieved with Gaussian approximate codewords ω(αβ, α′ β ′ ) ≡ α · β ′ − α′ · β (46)
that have mean photon number of about (.25)−2 ∼ 16.
More generally, a finitely squeezed codeword |ψi can is the symplectic form. Thus two Pauli operators com-
be regarded as a perfect codeword |ξi that has undergone mute if and only if their symplectic form is an integer.
an error; we may write Now a general continuous variable stabilizer code is the
Z simultaneous eigenspace of commuting Pauli operators,
|ψi = du dv η(u, v) ei(−up+vq) |ξi, (40) the code’s stabilizer generators. If the continuous vari-
able phase space is 2N -dimensional and the code space
is a finite-dimensional Hilbert space, then there must be
where η(u, v) is an error “wave function.” In the special
2N independent generators. The elements of the sta-
case of a Gaussian finitely squeezed codeword, we have
bilizer group are in one-to-one correspondence with the
1  1  points of a lattice L in phase space, via the relation
η(u, v) = √ exp − (u2 /∆2 + v 2 /κ2 ) , (41)
πκ∆ 2 " 2N
!#
√ X
where ∆ and κ are the squeezing parameters defined U (k1 , k2 , . . . k2N ) = exp i 2π ka va . (47)
above. a=1

If η(u, v) vanishes for |u| > α/2 or |v| > π/(nα), then Here {va , a = 1, 2, . . . , 2N } are the basis vectors of the
the error is correctable. In this case, the interpreta- lattice (each a linear combination of q’s and p’s), the
tion of η(u, v) as a wave function has a precise mean- ka ’s are arbitrary integers, and U (k1 , k2 , . . . k2N ) is the
ing, since there is an unambiguous decomposition of a corresponding element of the stabilizer. For the stabi-
state into codeword and error. Indeed, if |ξ1 i, |ξ2 i are lizer group to be Abelian, the symplectic inner product
perfect codewords and |ψ1 i, |ψ2 i are the corresponding of any pair of basis vectors must be an integer; that is,
finitely squeezed codewords with error wave functions η1 , the antisymmetric 2N × 2N matrix
η2 , then
Aab = ω(va , vb ) (48)
hψ1 |ψ2 i = hξ1 |ξ2 i hη1 |η2 i , (42)

where has integral entries. The lattice L has a 2N × 2N gener-

Z ator matrix M whose rows are the basis vectors,
hη1 |η2 i = du dv η1 (u, v)∗ η2 (u, v) . (43) 
v1

 v2 
M = ·  . (49)
 
 · 
VI. CONTINUOUS VARIABLE CODES FOR v2N
MANY OSCILLATORS
In terms of M , the matrix A can be expressed as
The continuous variable codes described in §IV are
based on simple lattices in the two-dimensional phase A = M ωM T , (50)

7
where ω denotes the 2N × 2N matrix or

T
T
M ⊥ω M ⊥ = A−1 . (61)
 
0 I
ω= , (51)
−I 0 If the lattice basis vectors are chosen so that A has the
standard form eq. (54), then
and I is the N × N identity matrix.  
The generator matrix of a lattice is not unique. The −1 T

0 D−1
A = . (62)
replacement −D−1 0

M → M ′ = RM (52) In the special case of a self-dual lattice, corresponding to

a code with a one-dimensional code space, both A and
leaves the lattice unmodified, where R is an invertible in- A−1 must be integral; hence D = D−1 and the standard
tegral matrix with determinant ±1 (whose inverse is also form of A is
integral). Under this replacement, the matrix A changes 
0 I

according to A= =ω. (63)
−I 0
A → A′ = RART . (53) Since the code subspace is invariant under the trans-
lations in L, we can think of the encoded information
By Gaussian elimination, an R can be constructed such as residing on a torus, the unit cell of L. The encoded
that the antisymmetric matrix A is transformed to Pauli operators {X̄ a Z̄ b } are a lattice of translations on
  this torus, corresponding to the coset space L⊥ /L. The
′ 0 D number of encoded Pauli operators is the ratio of the vol-
A = , (54)
−D 0 ume of the unit cell of L to the volume of the unit cell of
L⊥ , namely the determinant of A, which is therefore the
where D is a positive diagonal N × N matrix.
square of the dimension of the Hilbert space of the code.
There are also Pauli operators that provide a basis for
Thus the dimension of the code space is
the operations acting on the code subspace – these are
the Pauli operators that commute with the stabilizer but n = |Pf A| = det D , (64)
are not contained in the stabilizer. The operators that
commute with the stabilizer themselves form a lattice where Pf A denotes the Pfaffian, the square root of the
L⊥ that is dual (in the symplectic form) to the stabilizer determinant of the antisymmetric matrix A.
lattice. The basis vectors of this lattice can be chosen to The stabilizer lattice unit cell has volume |Pf A| in
be {ub , b = 1, 2, 3, . . . , 2N } such that units with h = 2πh̄ = 1, and the unit cell of the lattice of
encoded operations has volume |Pf A|−1 in these units.
ω(ua , vb ) = δab ; (55) So the code fits an n-dimensional code space into n units
of phase space volume, as expected.
then the generator matrix Codes of the CSS type (those analogous to the binary
quantum codes first constructed by Calderbank and Shor
u1
 
[15] and by Steane [16]) are constructed by choosing one
 u2  lattice Lq describing stabilizer generators that are linear
M⊥ = ·  (56) combinations of the q’s, and another lattice Lp ⊂ L⊥
 
 ·  q
describing stabilizer generators that are linear combina-
u2N tions of the p’s. (Here L⊥q denotes the Euclidean dual of
the lattice Lq .) The generator matrix of a CSS code has
of L⊥ has the property the form
 
M ⊥ ωM T = I . (57) M=
Mq 0
, (65)
0 Mp
It follows from eq. (48) and eq. (55) that the L basis
vectors can be expanded in terms of the L⊥ basis vectors where Mq and Mp are N × N matrices, and the integral
as matrix A has the form
Mq MpT
 
0
A= (66)
X
va = Aab ub , (58) −Mp MqT 0
b
For single-oscillator codes described in §IV, A is the
or 2 × 2 matrix
 
M = AM ⊥ , (59) A=
0 n
, (67)
−n 0
and hence that
where n is the code’s dimension. For a single-oscillator
−1 CSS code, the lattice is rectangular, as shown in Fig. 3.
ω(ua , ub ) = (A)ba , (60)

8
 [[7, 1, 3]] block code to protect one logical qubit against
p severe damage to any one of the oscillators.
For n ≥ 5, there is a [[5,1,3]] polynomial code [17],
also well suited to fault-tolerant processing, encoding one
qunit in a block of 5. (Actually, [[5, 1, 3]] quantum codes
exist for n < 5 as well [6,7], but these codes are less con-
ducive to fault-tolerant computing.) The larger value of
q n increases the vulnerability of each qunit to shift errors.
Hence, whether the [[7,1,3]] binary code or the [[5,1,3]]
should be preferred depends on the relationship of the
size of the typical shift errors to the rate of large errors.

VII. ERROR MODELS

FIG. 3. The stabilizer lattice and its dual for an n = 2 code What sort of errors can be corrected by these codes?
of a single oscillator. Solid lines indicate the stabilizer lattice; The codes are designed to protect against errors that shift
solid and dotted lines together comprise the dual lattice. In the values of the canonical variables p and q. In fact the
units of (2πh̄)2 , the unit cell of the stabilizer lattice (shaded) Pauli operators are a complete basis, so the action of
has area 2, and the unit cell of its dual has area 1/2. a general superoperator E acting on the input density
matrix ρ of a single oscillator can be expanded in terms
The closest packing of circles in two dimensions is of such shifts, as in
achieved by the hexagonal lattice. The generator ma- Z
trix for a hexagonally encoded qunit can be chosen to
E(ρ) = dαdβdα′ dβ ′ C(α, β; α′ β ′ )
be
′ ′

2
1/2   × ei(αp+βq) ρ e−i(α p+β q)
. (70)
1 √0
M= √ ·n , (68)
3 1/2 3/2 If the support of C(α, β; α′ , β ′ ) is concentrated on suffi-
ciently small values of its arguments, then the input ρ
and the dual lattice is generated by
can be recovered with high fidelity.
1 A useful model of decoherence is the special case of a
M⊥ = ·M . (69) “Pauli channel” in which C(α, β; α′ , β ′ ) is diagonal and
n
the superoperator can be expressed as
The shortest vector of the dual lattice has length
√
1/2 √
Z
2/n 3 , compared to length 1/ n for the square E(ρ) = dαdβ P (α, β) ei(αp+βq) ρ e−i(αp+βq) . (71)
lattice. Therefore the size of the smallest uncorrectable
shift is larger for the hexagonal code than for the square Since E is positive and trace preserving, we infer that
√
1/2
lattice code, by the factor 2/ 3 ≈ 1.07457. P (α, β) ≥ 0 and
An important special class of quantum codes for many Z
oscillators are the concatenated codes. In particular, we dαdβ P (α, β) = 1 . (72)
can encode a qubit in each of N oscillators using the
code of §IV. Then we can use a binary stabilizer code
that encodes k qubits in a block of N oscillators, and pro- Thus, we may interpret P (α, β) as a probability distri-
tects against arbitrary errors on any t oscillators, where bution: the phase space translation
2t + 1 is the binary code’s distance. The concatenated (q, p) → (q − α, p + β) (73)
codes have the important advantage that they can pro-
tect against a broader class of errors than small diffusive is applied with probability P (α, β).
shifts applied to each oscillator – if most of the oscillators Weak interactions between an oscillator and its envi-
undergo only small shifts in p and q, but a few oscillators ronment drive a diffusive process that can be well mod-
sustain more extensive damage, then concatenated codes eled by a Pauli channel. If the environment quickly “for-
still work effectively. gets” what it learns about the oscillator, the evolution
For example, there is a binary [[7, 1, 3]] quantum code, of the oscillator can be described by a master equation.
well suited to fault-tolerant processing, that encodes one Over a short time interval dt, the shifts applied to the
logical qubit in a block of seven qubits and can pro- oscillator may be assumed to be small, so that the Pauli
tect against heavy damage on any one of the seven [2]. operator can be expanded in powers of α and β. Suppose
Given seven oscillators, we can encode a qubit in each one that the shifts are symmetrically distributed in phase
that is resistant to quantum diffusion, and then use the space such that

9
 hαi = hβi = 0 , For dt small, the annihilation operator can be expanded
hα2 i = hβ 2 i , in terms of Pauli operators as
hαβi = 0 , (74) √ i  i√Γdt/2 q √ 
Γdt a ≈ − e − e−i Γdt/2 q
2
where h·i denotes the mean value determined by the prob- 1  i√Γdt/2 p √ 
ability distribution P (α, β). Suppose further that the + e − e−i Γdt/2 p (81)
shifts are diffusive, so that the mean square displacement 2
increases linearly with dt; we may write Thus, if the time interval ∆t between error correction
steps is small compared to the damping time Γ−1 , the
hα2 i = hβ 2 i = Ddt , (75) displacements applied to codewords are small, and error
where D is a diffusion constant. We then obtain correction will be effective.
Z Aside from decoherence, we also need to worry about
ρ(t + dt) = dαdβ P (α, β) ei(αp+βq) ρ e−i(αp+βq) “unitary errors.” For example, an inadvertent rotation
  of the phase of the oscillator induces the unitary trans-
1 1 formation
= ρ(t) + Ddt pρp − p2 ρ − ρp2
2 2
U (θ) ≡ exp iθa† a


  (82)
1 2 1 2
+ Ddt qρq − q ρ − ρq + O(dt3/2 ) , (76) Like any unitary transformation, this phase rotation can
2 2
be expanded in terms of Pauli operators. It is convenient
or to introduce the notation for the phase-space displace-
D D ment operator
ρ̇ = − [p, [p, ρ]] − [q, [q, ρ]] . (77)
2 2
D(γ) ≡ exp γa − γ ∗ a†


The interpretation of D as a diffusion constant can be √ 
confirmed by computing = exp i 2 [(Im γ)q − (Re γ)p] , (83)
d d where γ is a complex number. The displacements satisfy
tr p2 ρ = D = tr q 2 ρ ;



(78)
dt dt the identity
the mean square values of p and q increase with time as
tr D(γ)D(η)† = πδ 2 (γ − η) ,


(84)
Dt.
More generally, the master equation contains a diffu- so the operator U (θ) can be expanded in terms of dis-
sive term determined by the covariance of the distribu- placements as
tion P (α, β), and perhaps also a nondissipative drift term
1
Z
determined by the mean of P (α, β). Our quantum error- U (θ) = d2 γ uθ (γ)D(γ) , (85)
correcting codes can successfully suppress decoherence π
caused by quantum diffusion, if the recovery operation is where
applied often enough; roughly, the time interval ∆t be-
uθ (γ) = tr U (θ)D(γ)†


tween error correction steps should be small compared to . (86)
the characteristic diffusion time D−1 .
Interactions with the environment might also damp the Evaluating the trace in the coherent state basis, we find
amplitude of the oscillator, as described by the master that
equation i eiθ/2

i 2

uθ (γ) = exp − |γ| cot(θ/2) . (87)
2 sin(θ/2) 2
 
1 1
ρ̇ = Γ aρa† − a† aρ − ρa† a ; (79)
2 2 For small θ, the coefficient
√
here a = (q + ip)/ 2 is the annihilation operator and Γ
 
i i
is a decay rate. This master equation cannot be obtained uθ (γ) ≈ exp − |γ|2 (88)
θ θ
from a Pauli channel, but as for quantum diffusion, the
effects of amplitude damping over short time intervals has a rapidly oscillating phase, and can be regarded as
can be expressed in terms of small phase-space displace- a distribution with support concentrated on values of γ
ments. such that |γ|2 ∼ θ; indeed, formally
The master equation for amplitude damping can be
lim uθ (γ) = π δ 2 (γ) . (89)
obtained as the dt → 0 limit of the superoperator θ→0
√  √ 
ρ(t + dt) = E(ρ(t)) = Γdt a ρ(t) Γdt a† Thus a rotation by a small angle θ can be accurately
    expanded in terms of small displacements – error correc-
Γdt † Γdt † tion is effective if an oscillator is slightly overrotated or
+ I− a a ρ(t) I − a a . (80)
2 2 underrotated.

10
 VIII. THE GAUSSIAN QUANTUM CHANNEL asymptotically for large N by either of two methods.
In one method, the n = 2 code described in §IV is in-
At what rate can error-free digital information be con- voked for each oscillator, and concatenated with a bi-
veyed by a noisy continuous signal? In classical informa- nary quantum code. In the other method, which more
tion theory, an answer is provided by Shannon’s noisy closely follows Shannon’s construction, a code for N os-
channel coding theorem for the Gaussian channel [18]. cillators is constructed as in §VI, based on a close packing
This theorem establishes the capacity that can be at- of spheres in 2N -dimensional phase space. However (in
tained by a signal with specified average power, for a contrast to the classical case), neither method works if
channel with specified bandwidth and specified Gaussian σ 2 is too large. For large σ 2 , encodings can be chosen
noise power. The somewhat surprising conclusion is that that protect against q shifts or against p shifts, but not
a nonzero rate can be attained for any nonvanishing value against both.
of the average signal power. To establish an attainable rate using concatenated cod-
A natural generalization of the Gaussian classical chan- ing (the method that is easier to explain), we first recall a
nel is the Gaussian quantum channel. The Gaussian result concerning the quantum capacities of binary chan-
quantum channel is a Pauli channel: N oscillators are nels [15,23]. If X and Z errors are independent and each
transmitted, and the q and p displacements acting on occur with probability pe , then binary CSS codes exist
the oscillators are independent Gaussian random vari- that achieve a rate
ables with mean 0 and variance σ 2 . A code is an M -
dimensional subspace of the Hilbert space of the N oscil- R > 1 − 2H2 (pe )
lators, and the rate R of the code (in qubits) is defined ≡ 1 + 2pe log2 pe + 2(1 − pe ) log2 (1 − pe ) ; (93)
as
this rate is nonzero for pe < .1100.
1 Now, for the Gaussian quantum channel, if we use the
R= log2 M . (90)
N n = 2 continuous variable code, errors afflicting the en-
coded qubit are described by a binary channel with in-
The quantum-information capacity CQ of the channel is
dependent X and Z errors. Since the code √ can correct
the maximal rate at which quantum information can be
shifts in q or p that satisfy ∆q, ∆p < π/2, the error
transmitted with fidelity arbitrarily close to one.
probability is
The need for a constraint on the signal power to define
the capacity of the Gaussian classical channel can be un- 1
Z ∞
2 2
derstood on dimensional grounds. The classical capacity pe < 2 · √ √
dx e−x /2σ . (94)
2πσ 2 π/2
(in bits) is a dimensionless function of the variance σ 2 ,
but σ 2 has dimensions. Another quantity with the same
Since the expression bounding pe in eq. (94) has the value
dimensions as σ 2 is needed to construct a dimensionless .110 for σ ≈ .555, we conclude that the Gaussian quan-
variable, and the power fulfills this role. But no power tum channel has nonvanishing quantum capacity CQ pro-
constraint is needed to define the quantum capacity of the vided that
quantum channel. The capacity (in qubits) is a function
of the dimensionless variable h̄/σ 2 , where h̄ is Planck’s σ < .555 . (95)
constant.
An upper bound on the quantum capacity of the Gaus- One might expect to do better by concatenating the
sian quantum channel was derived by Holevo and Werner hexagonal n = 2 single-oscillator code with a binary sta-
[19]; they obtained (reverting now to units with h̄ = 1) bilizer code, since the hexagonal code can correct larger
shifts than the code derived from a square lattice. For the
CQ ≤ log2 (1/σ 2 ) , (91) Gaussian quantum channel, the symmetry of the hexag-
onal lattice ensures that X, Y , and Z errors afflicting the
for 0 < σ 2 < 1, and CQ = 0 for σ 2 ≥ 1. They also
encoded qubit are equally likely. A shift is correctable if
computed the coherent information IQ of the Gaussian
it lies within the “Voronoi cell” of the dual lattice, the
quantum channel, and maximized it over Gaussian signal
cell containing all the points that are closer to the origin
states, finding [19]
than to any other lattice site. By integrating the Gaus-
(IQ )max = log2 (1/eσ 2 ) , (92) sian distribution over the hexagonal Voronoi cell, we find
that the probability pe,total of an uncorrectable error sat-
for 0 < σ 2 < 1/e (where e = 2.71828 . . .). The coher- isfies
ent information is conjectured to be an attainable rate Z r Z x/√3
12 2 2 2
[20–22]; if this conjecture is true, then eq. (92) provides pe,total < 1 − 2
dx dy e−(x +y )/2σ , (96)
a lower bound on CQ . 2πσ 0 0
Using our continuous variable codes, rigorous lower √ 1/2
bounds on CQ can be established. For σ 2 sufficiently where r = (π/2 3) is the size of the smallest un-
small, a nonzero attainable rate can be established correctable shift. For a binary quantum channel with

11
equally likely X, Y , and Z errors, it is known [24] that group of unitary transformations that, acting by conju-
there are stabilizer codes achieving a nonvanishing rate gation, take tensor products of Pauli operators to tensor
for pe,total < .1905; our bound on pe,total reaches this products of Pauli operators (one says that they preserve
value for σ ≈ .547. the “Pauli group”). Since for N oscillators the tensor
Somewhat surprisingly, for very noisy Gaussian quan- products of Pauli operators have the form eq. (44), the
tum channels, square lattice codes concatenated with Clifford group transformations, acting by conjugation,
CSS codes seem to do better than hexagonal codes con- are linear transformations of the p’s and q’s that preserve
catenated with stabilizer codes. The reason this hap- the canonical commutation relations. Such transforma-
pens is that a CSS code can correct independent X tions are called symplectic transformations. The sym-
and Z errors that occur with total probability pe,total = plectic group has a subgroup that preserves the photon
pX + pZ − pX · pZ , which approaches .2079 > .1905 as number
pX = pZ → .1100. For a given value of σ, the qubit en-
N
coded in each oscillator will have a lower error probability
a†i ai .
X
total photon number = (97)
if the hexagonal code is used. But if the square lattice
i=1
is used, a higher qubit error rate is permissible, and this
effect dominates when the channel is very noisy. The transformations in this subgroup can be imple-
We remark that this analysis is readily extended to mented with linear optics [26]. The full symplectic group
more general Gaussian quantum channels. We may con- also contains “squeeze operators” that take an a to a lin-
sider Pauli channels acting on a single oscillator in which ear combination of a’s and a† ’s; equivalently, the squeeze
the probability distribution P (α, β) is a more general operators rescale canonical operators by a real number λ
Gaussian function, not necessarily symmetric in p and q. along one axis in the quadrature plane, and by λ−1 along
In that case, a symplectic transformation (one preserving the conjugate axis, as in (for example)
the commutator of p and q) can be chosen that transforms
the covariance matrix of the Gaussian to a multiple of q1 → λq1 , p1 → λ−1 p1 . (98)
the identity; therefore, this case reduces to that already
discussed above. We may also consider channels acting With squeezing and linear optics we can in principle im-
on N oscillators that apply shifts in the 2N -dimensional plement any symplectic transformation.
phase space, chosen from a Gaussian ensemble. Again Aside from the symplectic transformations, we will also
there is a symplectic transformation that diagonalizes the assume that it is easy to do displacements that shift q
covariance matrix; therefore, this case reduces to N in- and p by constants. A displacement of q1 by c is actually
dependent single oscillator channels, each with its own the limiting case of a symplectic transformation on two
value of σ 2 [25]. oscillators q1 and q2 :

q1 → q1 + εq2 , p1 → p1 + εp2
IX. SYMPLECTIC OPERATIONS q2 → q2 − εq1 , p2 → p2 − εp1 (99)

To use these codes for fault-tolerant quantum compu- where ε → 0 with εq2 = c held fixed.
tation, we will need to be able to prepare encoded states, Since for the code with stabilizer generators eq. (21)
perform error recovery, and execute quantum gates that the Pauli operators acting on our encoded qunits are
act on the encoded quantum information. The most dif- X̄ = eipα and Z̄ = e2πiq/nα , the Clifford group trans-
ficult task is encoding; we will postpone the discussion formations acting on N qunits constitute a subgroup of
of encoding until after we have discussed encoded opera- the symplectic transformations (including shifts) acting
tions and error recovery. on N oscillators, the subgroup that preserves a specified
Suppose, for example, that we have N oscillators, each lattice in phase space. Thus we can do any encoded Clif-
encoding a qunit. We wish to apply U (nN ) transforma- ford group gate we please by executing an appropriate
tions that preserve the code subspace of the N qunits. symplectic transformation (possibly including a shift).
As is typical of quantum codes, we will find that there is A similar comment applies to the case of a qunit en-
a discrete subgroup of U (nN ) that we can implement coded in a qudit. Since the logical Pauli operators are
“easily;” but to complete a set of universal gates we X̄ = X r1 and Z̄ = Z r2 , each Clifford group transforma-
must add further transformations that are “difficult.” In tion in the n-dimensional code space is also a Clifford
the case of our continuous variable codes, the easy gates group transformation on the underlying qudit.
will be accomplished using linear optical elements (phase But we must also be sure that our implementation of
shifters and beam splitters), along with elements that can the Clifford group is fault tolerant. In previous discus-
“squeeze” an oscillator. For the “difficult” gates we will sions of quantum fault tolerance for [[N, k, 2t + 1]] codes,
require the ability to count photons. the central theme has been that propagation of error from
The easy gates are the gates in the Clifford group. In one qudit to another in the same code block must be very
general, the Clifford group of a system of N qunits is the carefully controlled [27,28]. For shift-resistant codes the
main issue is rather different. Since each qudit typically

12
has a (small) error anyway, propagation of error from one of the oscillator. For example, suppose that we measure
qudit to another is not necessarily so serious. But what q for a state in the code subspace. If there are no errors
must be controlled is amplification of errors – gates that and the measurement has perfect resolution, the only al-
turn small errors into large errors should be avoided. lowed values of q will be integer multiples of α. If there
The Clifford group can be generated by gates that are are errors or the measurement is imperfect, classical error
fault-tolerant in this sense. The Clifford group for qunits correction can be applied to the outcome, by adjusting
can be generated by three elements. The SUM gate is it to the nearest α · k, where k is an integer. Then the
a two-qunit gate that acts by conjugation on the Pauli outcome of the measurement of Z̄ is ω k .
operators according to To diagnose errors in a coded data state, we must mea-
sure the stabilizer generators. This measurement can be
SUM : X1a X2b → X1a X2b−a , Z1a Z2b → Z1a+b Z2b . implemented by “feeding” the errors from the code block
to a coded ancilla, and then measuring the ancilla de-
(100)
structively, following the general procedure proposed by
Here qunit 1 is said to be the control of the SUM gate, Steane [29] (see Fig. 4). For example, to measure the
and qunit 2 is said to be its target; in the binary (n = 2) generator e2πiq/α (i.e., the value of q modulo
√ α), we pre-
case, SUM is known as controlled-NOT, or CNOT. The pare the ancilla in the state (|0̄i + |1̄i)/ 2, the equally
Fourier gate F acts by conjugation as weighted superposition of all |q = s · αi, s an integer.
Then a SUM gate is executed with the data as control
F : X →Z , Z → X −1 ; (101) and the ancilla as target – acting according to

for n = 2 the Fourier Transform is called the Hadamard q2 → q1 + q2 , (105)

gate. The phase gate P acts as
where q1 , q2 are the values of q for the data and ancilla
P : X → (η)XZ , Z→Z , (102) respectively, prior to the execution of the SUM gate. By
measuring q of the ancilla, the value of q1 +q2 is obtained,
where the n-dependent phase η is ω 1/2 if n is even and and this value modulo α determines the shift that should
1 if n is odd. Any element of the Clifford group can be applied to the data to recover from the error.
be expressed as a product of these three generators. (In Similarly, to measure the stabilizer generator einpα , we
Ref. [8] another gate S was included among the genera- prepare the ancilla in the state |0̄i, the equally weighted
tors of the Clifford group, but in fact the S gate can be superposition of all |p = s · 2π/nαi, s an integer. Then a
expressed as a product of SUM gates.) SUM gate is executed with the ancilla as control and the
For an n-dimensional system encoded in a continuous data as target. Finally, the p quadrature of the ancilla
variable system, these Clifford group generators can all is measured. The outcome reveals the value of p2 − p1
be realized as symplectic transformations. In the case prior to the SUM gate, where p1 is the momentum of
where the stabilizer generators are symmetric in q and p, the data, and p2 is the momentum of the ancilla. The
r ! r ! measured value modulo 2π/nα then determines the shift
2π 2π that should be applied to the data to recover from the
X̄ = exp −ip , Z̄ = exp iq , (103) error.
n n

the required symplectic transformations are data ✉

(a)
SUM : q1 → q1 , p1 → p1 − p2 ,
|X̄ = 1i SUM Measure q
q2 → q1 + q2 , p2 → p2 ,
F : q→p, p → −q ,
P : q→q , p→p−q+c , (104)
data SUM
(b)
p
where the n-dependent shift c is 0 for n even and π/2n
for n odd. Under these symplectic transformations, small |Z̄ = 1i ✉ Measure p
deviations of q and p from the stabilizer lattice remain FIG. 4. Measurement of the error syndrome. (a) To diag-
small; in this sense the transformations are fault tolerant. nose the q shift, an ancilla is prepared in the encoded X̄ = 1
state, a SUM gate is executed with the data as control and
the ancilla as target, and the position of the ancilla is mea-
X. ERROR RECOVERY sured. (b) To diagnose the p shift, the ancilla is prepared in
the Z̄ = 1 state, a SUM gate is executed with the ancilla
If we are willing to destroy the encoded state, then as control and the data as target, and the momentum of the
measuring the encoded X̄ or Z̄ is easy – we simply con- ancilla is measured.
duct a homodyne measurement of the q or p quadrature

13
 Of course, the ancilla used in the syndrome measure- (recalling that X̄ 2 = Z̄ 2 = I on the code subspace). This
ment can also be faulty, resulting in errors in the syn- transformation is just the Fourier transform
drome and imperfect recovery. Similarly, the measure-  π 
ment itself will not have perfect resolution, and the shift F : exp i a† a (108)
applied to recover will not be precisely correct. Fur- 2
thermore, as is discussed in §V, the ideal codewords are (where a† a is the photon number), which describes the
unphysical nonnormalizable states, so that the encoded natural evolution of the oscillator for one quarter cycle.
information will always be carried by approximate code- Thus the phase of the Hadamard operator is simply the
words. For all these reasons, deviations from the code photon number modulo four; we can measure the eigen-
subspace are unavoidable. But if a fresh supply of ancilla value of the encoded Hadamard transformation by count-
oscillators is continuously available, we can prevent these ing photons.
small errors from accumulating and eventually damaging In fact the photon number in the code space is even –
the encoded quantum information. all codewords are invariant under a 180◦ rotation in the
quadrature plane. Because of this feature, the prepara-
tion of the Hadamard eigenstate has some fault tolerance
XI. UNIVERSAL QUANTUM COMPUTATION built in; if the photon count is off by one, the number will
be odd and an error will be detected. In that case we re-
Symplectic transformations together with homodyne ject the state we have prepared and make a new attempt.
measurements are adequate for Clifford group computa- If the photon number is large, then obtaining a reliable
tion and for error recovery (assuming we have a supply determination of the photon number modulo four will re-
of encoded states). But to achieve universal computa- quire highly efficient photodetection. But on the other
tion in the code space, we need to introduce additional hand, the photon number need not be very large – the
operations. Fortunately, the quantum optics laboratory mean value of a† a is about ∆−2 where ∆ is the squeeze
offers us another tool that can be used to go beyond the factor, and we have seen that the intrinsic error rate due
symplectic computational model – the ability to count to imperfect squeezing is quite small for ∆ ∼ 1/4, or
photons. ha† ai ∼ 16.
There are a variety of ways in which photon count- An alternative to preparing an encoded EPR pair and
ing can be exploited to complete a universal set of fault- destructively measuring one member of the pair is to pre-
tolerant gates. We will describe two possible ways, just pare |0̄i and then perform a quantum nondemolition mea-
to illustrate how universal fault-tolerant quantum com- surement of the photon number modulo 4. This might
putation might be realized with plausible experimental be done by coupling the oscillator to a two-level atom
tools. For this discussion, we will consider the binary as proposed in Ref. [32]. Indeed, since only one bit of
case n = 2. information needs to be collected (the photon number
is either 0 or 2 modulo 4), the measurement could be
made in principle by reading out a single atom. Suppose
A. Preparing a Hadamard eigenstate that the coupling of oscillator to atom is described by the
perturbation
We can complete the universal gate set if we have the
H ′ = λ a† a σz , (109)
ability to prepare eigenstates of the Hadamard operator
H [30,31]. For this purpose it suffices to be able to de- where σz = −1 in the atomic ground state |gi and σz =
structively measure H of an encoded qubit. Assuming we 1 in the atomic excited state |ei. By turning on this
are able to prepare a supply of the encoded Z̄ eigenstate coupling for a time t = π/4λ, we execute the unitary
|0̄i, we can make an encoded EPR pair using symplectic transformation
gates. Then by destructively measuring H for one en-
coded qubit in the pair, we prepare the other qubit in an U = exp(−i(π/4) a† a σz ) . (110)
encoded eigenstate of H with known eigenvalue. √
But how can we destructively measure H? The Then the atomic state (|gi + |ei)/ 2 evolves as
Hadamard gate acts by conjugation on the encoded Pauli 1
operators according to U: √ (|gi + |ei)
2
H: X̄ → Z̄ , Z̄ → X̄ . (106) 1 ia† aπ/4 †
→√ e (|gi + e−ia aπ/2 |ei) . (111)
2
If we use the code that
√ treats q and p symmetrically
√ so √
that X̄ = exp(−ip π) and Z̄ = exp(iq π), then the By measuring the atomic state in the basis (|gi±|ei)/ 2,
Hadamard gate can be implemented by the symplectic we read out the value of the photon number modulo 4
transformation. (assumed to be either 2 or 4). Since this is a nondemoli-
tion measurement, it can be repeated to improve reliabil-
q→p, p → −q (107) ity. By measuring the photon number mod 4 many times

14
(perhaps with rounds of error correction in between the Completing the universal gate set by measuring the
measurements), we obtain a Hadamard eigenstate with Hadamard transformation has some drawbacks. For one
excellent fidelity. thing, while photon number modulo four corresponds to
How does the ability to construct the Hadamard eigen- the Hadamard eigenvalue in the ideal code space, this
state enable us to achieve universal quantum computa- correspondence will not apply to approximate codewords
tion? We can make contact with constructions that have unless they are of a special type.
been described previously in the literature by observing Recall that the imperfections of the codewords arising
that the Hadamard eigenstate can be transformed by ap- from finite squeezing can be described by an “embedded
plying symplectic gates to the “π/8 phase state.” First error” |ηi as in eq. (40); a Gaussian approximate code-
note that the two Hadamard eigenstates can be converted word has a Gaussian embedded error
to one another by applying the encoded gate X̄ Z̄, which
1  1 
can be implemented by shifting both p and q. Therefore η(u, v) = √ exp − (u2 /∆2 + v 2 /κ2 ) , (117)
it is sufficient to consider the eigenstate corresponding to π∆κ 2
the eigenvalue 1,
where ∆ is the width in q and κ is the width in p. Sym-
|ψH=1 i = cos(π/8) |0i + sin(π/8) |1i . (112) plectic gates act separately on the encoded qubit and the
“embedded error” |ηi; for example, the Fourier transform
By applying the symplectic single-qubit gate gate and the SUM gate act on the error according to
   
1 1 1 1 0 F : |u, vi → |v, −ui ,
H · P −1 ≡ √ ·
2 1 −1 0 −i
SUM : |u1 , v1 ; u2 , v2 i → |u1 , v1 + v2 ; u2 − u1 , v2 i . (118)
 
1 1 −i
=√ , (113)
2 1 i By measuring the photon number modulo 4, we actually
measure the product of the eigenvalue of the Hadamard
we obtain the π/8 state gate acting on the codeword and the eigenvalue of F act-
1   ing on the embedded error. The latter always equals 1 if
|ψπ/8 i = √ e−iπ/8 |0i + eiπ/8 |1i . (114) we use symmetrically squeezed codewords, with ∆ = κ.
2
Symmetric squeezing is not in itself sufficient to ensure
Now this π/8 state can be used to perform the non- that the measurement of the photon number modulo 4
symplectic phase gate will prepare the desired encoded Hadamard eigenstate.
 −iπ/8  We also need to consider how the embedded error is af-
e 0 fected by the preparation of the EPR pair that precedes
S= , (115)
0 eiπ/8 the measurement. To prepare the EPR pair, we use the
SUM gate. Suppose that we start with two symmet-
which completes the universal gate set [33,34]. The gate
rically squeezed states. Then the SUM gate yields the
is constructed by executing the circuit shown in Fig. 5.
error wave function
We perform a CNOT gate with the arbitrary single-qubit
state |ψi = a|0i + b|1i as the control, and the π/8 phase η ′ (u1 , v1 ; u2 , v2 )
state as the target; then the target qubit is measured
= exp − u21 + (v1 − v2 )2 + (u1 + u2 )2 + v22 /∆2 . (119)



in the basis {|0i, |1i}. If the measurement outcome is
|0i (which occurs with probability 1/2), then the control
Not only is it not symmetric, but the error is entangled
qubit has become aeiπ/8 |0i + be−iπ/8 |1i = S|ψi and we
between the two oscillators. The Fourier transform mea-
are done. If the measurement outcome is |1i, then the
control qubit has become ae−iπ/8 |0i + beiπ/8 |1i, and we surement will not give the desired result when applied to
either oscillator.
obtain S|ψi by applying the symplectic single-qubit gate
To ameliorate this problem, we could perform error
correction after the preparation of the EPR pair and be-
 −iπ/4 
e 0
e−iπ/4 P = . (116) fore the measurement, where the error correction proto-
0 eiπ/4
col has been designed to produce symmetrically squeezed
states. Or we could avoid preparing the EPR state by
|ψi ✉ S|ψi using the nondemolition measurement of photon number
P
modulo 4, as described above.

|ψπ/8 i ✐ Measure
B. Preparing a cubic phase state
FIG. 5. Implementation of the S gate. An ancilla is pre-
pared in the state |ψπ/8 i, and a CNOT gate is executed with
Now we will describe another way to use photon count-
the data as control and the ancilla as target; then the ancilla
is measured in the basis {|0i, |1i}. A P gate is applied to the
ing to implement non-symplectic gates, which is less sen-
data conditioned on the measurement outcome. sitive to the codeword quality. Again, we will complete

15
 Z
the universal gate set by constructing the π/8 phase gate |φi = dx eiφ(x) |xi . (126)
S.
For our binary (n = 2) code, the code subspace has
the basis If we perform the gate SUM−1 with position eigenstate
|qi as control and |φi as target, and then measure the
+∞
X position of the target obtaining the outcome |ai, the state
|0̄i = |q = 2sαi , of the control oscillator has become eiφ(q+a) |qi. We can
s=−∞ therefore complete the construction of Vφ by applying the
+∞
X transformation
|1̄i = |q = (2s + 1)αi . (120)
s=−∞ U (a) = ei[φ(q)−φ(q+a)] . (127)

(For now we ignore the embedded error due to imperfect If the function φ(q) is cubic, then the argument of the
squeezing; it will be taken into account later.) An S gate exponential is quadratic and hence U (a) is a symplectic
acting on the encoded qubit is implemented (up to an transformation.
irrelevant overall phase) by the unitary operator Now the problem of implementing universal quantum
computation in the code subspace has been reduced to
the problem of preparing the cubic phase state |γi. We
 
iπ 
2(q/α)3 + (q/α)2 − 2(q/α)

W = exp . (121) can accomplish this task by preparing an EPR pair, and
4
then performing a suitable photon counting measurement
Indeed, we can check that (a nonideal homodyne measurement) on one member of
 the pair.
0, if x = 2s , Of course, the EPR pair will not be perfect. To be
2x3 + x2 − 2x (mod 8) = (122)
1, if x = 2s + 1 . definite, let us suppose (although this assumption is not
really necessary) that it is a Gaussian state
The operator W is the product of a symplectic gate
and the cubic phase gate  1/2 Z "  2 #
σp 1 2 q1 + q2
|ψσp ,σq i = dq1 dq2 exp − σp
Vγ = exp(iγq 3 ) , (123) πσq 2 2
 
1
where γ = π/(2α3 ). But how do we implement the cubic 2
× exp − (q1 − q2 ) /σq2 |q1 , q2 i (128)
gate? In fact, if we are able to prepare a “cubic phase 2
state” with σp , σq ≪ 1.
Now suppose that the second oscillator is mixed with
Z
3
|γi = dx eiγx |xi , (124) a coherent light beam, resulting in a large shift in mo-
mentum,
then we can perform the gate Vγ by executing the circuit
shown in Fig. 6. |ψi → eiwq |ψi , w ≫ σq−1 , σp−1 ; (129)
then the photon number is measured and n photons are
✉ Vγ |ψi detected. Thus the state of the first oscillator becomes
|ψi U (a)
(up to normalization)
 1/2 Z
(n) σp 1 2 2
|ψ1 i ≈ dq1 |q1 i e− 2 σp q1
|γi SUM−1 Measure q πσq
Z
1 2 2

FIG. 6. Implementation of the cubic phase gate. An ancilla × dq2 ϕ∗n (q2 )eiwq2 e− 2 (q1 −q2 ) /σq , (130)
is prepared in the state |γi, and a SUM−1 gate is executed
with the data as control and the ancilla as target; then the where |ϕn i denotes the photon number eigenstate, the
position of the ancilla is measured. A symplectic gate U (a) eigenstate with eigenvalue n + 12 of the Hamiltonian H =
is then applied to the data, conditioned on the outcome a of 1 2 2
2 (p + q ).
the measurement. We can evaluate the q2 integral in eq. (130) by ap-
pealing to the semiclassical approximation. For q2 in the
To understand how the circuit works, consider the classically allowed region and far from the classical turn-
more general problem of implementing a phase gate that ing points, we may write
acts on the position eigenstates according to  Z q2 
1
ϕ∗n (q2 ) ∼ p exp −i dx p(x)
Vφ : |qi → eiφ(q) |qi (125) 2πp(q2 )
 Z q2 
(where φ(q) is a real-valued function), using the prepared 1
+p exp +i dx p(x) , (131)
phase state 2πp(q2 )

16

−1

where Vγ = Sγ/γ ′ Vγ ′ Sγ/γ ′ , (138)
p
p(x) = 2n + 1 − x2 . (132) where Sr is a squeeze operation that acts according to

For w ≫ σq−1 , the rapid phase oscillations strongly sup- Sr : q → (r)1/3 q ,

press the contribution to the integral arising from the p → (r)−1/3 p . (139)
left-moving part of ϕ(n) (q2 ). A contribution from the
right-moving part survives provided that Alternatively, we could squeeze the phase state |γ ′ i before
we use it to implement the cubic phase gate.
|p(q1 ) − w| < σq−1 . (133) Is this procedure fault tolerant? Before considering the
errors introduced during the implementation of the cubic
When this condition is satisfied, it is a reasonable approx-
1 2 2 phase gate, we should check that the gate does not catas-
imation to replaceqthe Gaussian factor e− 2 (q1 −q2 ) /σq in trophically amplify any preexisting errors. In general, a
the q2 integral by 2πσq2 δ(q1 − q2 ), so that we obtain phase gate can transform a small position shift error into
a potentially dangerous momentum shift error. Com-
muting V (φ) = eiφ(q) through the shift operator e−iup ,
Z
(n) 1/2 1 2 2
|ψ1 i ≈ (2σp σq ) dq1 |q1 i e− 2 σp q1 we find
 Z q1 
1 eiφ(q) e−iup = e−iup eifu (q) eiφ(q) , (140)
× p exp −i dx (p(x) − w) . (134)
2πp(q1 )
where fu (q) = φ(q + u)− φ(q); the operator eifu (q) can be
The probability that n photons are detected is given by expanded in terms of momentum shift operators of the
(n)
the norm of this |ψ1 i. The values of n that occur with form eivq by evaluating the Fourier transform
appreciable probability satisfy eq. (133) for some q1 with
dq i(fu (q)−vq)
Z
|q1 | < σp−1 ; thus typical measurement outcomes are in ˜
fu (v) = e . (141)
the range 2π

1 1
2 1 Assuming we use a code where the parameter α is of

n+ ∼ w ± σq−1 + σp−2 , (135) order one, uncorrectable errors will be likely if f˜u (v) has
2 2 2
significant support on values of v that are order one.
with a flat probability distribution Suppose that V (φ) acts on an approximate codeword
σq whose wave function is concentrated on values of q in
(n) (n)
Prob(n) = hψ1 |ψ1 i ∼ . (136) the domain |q| < L. Phase cancellations will strongly
w
suppress f˜u (v), unless the stationary phase condition
Heuristically, after the momentum shift is applied, the fu′ (q) = v is satisfied for some value of q in the domain of
oscillator that is measured has momentum of order w ± the approximate codeword. Therefore, V (φ) can propa-
σq−1 , and position of order σp−1 , so that the value of the gate a preexisting position shift u to a momentum shift

2 error of magnitude
energy is n + 21 = 21 (p2 + q 2 ) ∼ 12 w ± σq−1 + 21 σp−2 .
For a particular typical outcome of the photon-
(n) |v| ∼ max |fu′ (q)| . (142)
counting measurement, since |ψ1 i has its support on |q|≤L
−1
|q1 | < σp ≪ w, we can Taylor expand p(x) about x = q1
(n) The cubic phase gate needed to implement the encoded
to express |ψ1 i as S gate is W = eiφ(q) where φ(q) = πq 3 /2α3 , so that
 Z q1  fu (q) = 3πuq 2 /2α3 + · · · (ignoring small terms linear and
(n)
p
constant in q), and fu′ (q) = 3πuq/α3 ; the gate transforms


ψ1 (q1 ) ∝ exp −i (2n + 1) − x2 − w dx
the position shift u to a momentum shift
i √
q13 − i 2n + 1 − w q1


∝ exp √ v ∼ 3πLu/α3 . (143)
6 2n + 1
For α of order one, then, to ensure that v is small we
!
+O(q15 /w3 ) . (137) should use approximate codewords with the property
that the typical embedded position shift u satisfies
This is a cubic phase state to good precision if w is large |u| ≪ L−1 . (144)
enough.
The coefficient γ ′ of q13 in the phase of ψ1 is of order In particular, if the approximate codeword’s embedded
n−1/2 , while the phase γ of the operator Vγ that we wish errors are Gaussian, where κ is the typical size of a mo-
to execute is of order one. However, we can construct Vγ mentum shift and ∆ is the typical size of a position shift,
from Vγ ′ as we require

17
 ∆≪κ. (145) C. Purification

We assume that shift errors due to other causes are no Either of the above two methods could be used to im-
larger than the embedded error. plement a nonsymplectic phase transformation that com-
In the circuit Fig. 6 that implements the cubic phase pletes the universal gate set. Of course, experimental
gate, position shift errors in either the encoded state |ψi limitations might make it challenging to execute the gate
or the ancilla state |γi might cause trouble. A shift by with very high fidelity. One wonders whether it is pos-
u in |ψi is transformed to a phase error eifu (q) , and a sible to refine the method to implement fault-tolerant
shift by u in |γi infects |ψi with a phase error eifu (q+a) . universal gates of improved fidelity.
Therefore, we should require that position shift errors in In fact, such refinements are possible. We have seen
both |ψi and |γi satisfy the criterion eq. (144), where L that we can reach beyond the symplectic transformations
is the larger of the two wave packet widths. and achieve universal quantum computation if we have a
When a cubic phase state is prepared by measuring supply of appropriate “nonsymplectic states” that can’t
half of an EPR pair, the packet width is of order σp−1 be created with the symplectic gates. If the nonsymplec-
and typical position shift errors have u ∼ σq . However, tic states have the right properties, then we can carry
we must also take into account that either the encoded out a purification protocol to distill from our initial sup-
state or the ancilla must be squeezed as in eq. (139). ply of noisy nonsymplectic states a smaller number of
Suppose that the ancilla is squeezed, by a factor of order nonsymplectic states with much better fidelity [35,36].
n1/6 ∼ w1/3 ; the wave packet is rescaled so that, after An example of a nonsymplectic state that admits such
squeezing, the width L′ and the typical shifts u′ are given a purification protocol is a variant of the state originally
by introduced by Shor [27], the three-qubit state
L′ ∼ σp−1 w−1/3 , u′ ∼ σq w−1/3 . (146) X
2−3/2 (−1)abc |ai1 |bi2 |ci3 ; (150)
a,b,c∈{0,1}
Then the condition |u′ | ≪ L′−1 is satisfied provided that
σq ≪ σp w2/3 . We also require that the rescaled packet it can be characterized as the simultaneous eigenstate of
has width large compared to 1, or σp ≪ w−1/3 . three commuting symplectic operators: Λ(Z)1,2 X3 and
For the derivation of eq. (137), we used the approxima- its two cyclic permutations, where Λ(Z) is the two-qubit
tions wσq ≫ 1 and wσp ≫ 1. We also need to check that conditional phase gate
the remainder terms in the Taylor expansion give rise to
a phase error that is acceptably small. This error has the Λ(Z) : |a, bi → (−1)ab |a, bi (151)
form eif (q1 ) , where f (q1 ) = O(q15 /w3 ), corresponding to
a momentum shift As Shor explained, this nonsymplectic state can be em-
ployed to implement the Toffoli gate
v ∼ f ′ (q1 ) ∼ σp−4 w−3 . (147)
T : |a, b, ci → |a, b, c ⊕ abi , (152)
Squeezing amplifies this momentum shift error to v ′ ∼
vw1/3 ∼ σp−4 w−8/3 , which will be small compared to and so provides an alternative way to complete the uni-
1 provided that σp ≫ w−2/3 . To summarize, our im- versal gate set.
plementation of the cubic phase gate works well if the To purify our supply of nonsymplectic states, symplec-
approximate codewords have embedded errors satisfying tic gates are applied to a pair of nonsymplectic states and
∆ ≪ κ, and if widths σq and σp of the approximate EPR then one of the states is measured. Based on the outcome
state satisfy w ≫ σq−1 and of the measurement, the other state is either kept or dis-
carded. If the initial ensemble of states approximates the
w−1/3 ≫ σp ≫ w−2/3 . (148) nonsymplectic states with adequate fidelity, then as pu-
rification proceeds, the fidelity of the remaining ensemble
Finally, how accurately must we count the photons? converges rapidly toward one.
An error ∆n in the photon number results in a phase The details of the purification protocol will be de-
(n)
error eivq1 with |v| ∼ n−1/2 ∆n in ψ1 (q1 ), which will scribed elsewhere; here we will only remark that these
Shor states can be readily created using symplectic gates
be amplified by squeezing to |v | ∼ |v|w1/3 ∼ n−1/3 ∆n.
′
and π/8 phase gates. The Shor state is obtained if we
Therefore, the precision of the photon number measure-
apply the transformation
ment should satisfy
Λ2 (Z) : |a, b, ci → (−1)abc |a, b, ci (153)
∆n ≪ n1/3 (149)
to the state
to ensure that this error is acceptably small.
X
H1 H2 H3 |0, 0, 0i = 2−3/2 |a, b, ci . (154)
a,b,c∈{0,1}

18
As shown in Fig. 7, Λ2 (Z) can be applied by executing a (coherent or squeezed) states to Gaussian states. Hence
circuit containing 5 S gates, 4 S −1 gates, and 8 CNOT no symplectic transformation can take (say) the oscilla-
gates. tor’s ground state to a state in the code subspace.
So encoding requires nonsymplectic operations, and as
far as we know it cannot be accomplished by counting
✉ ✉ ✉ S photons either – we must resort to a nonlinear coupling
(a) = between oscillators, such as a χ(3) coupling. We will de-
P ✐ S −1 ✐ S scribe one possible encoding scheme: First, we prepare
a squeezed state, an eigenstate of the momentum with
✉ ✉ ✉ ✉ p = 0. This state is already an eigenstate with eigen-
value one of the stabilizer generator einpα , but not an
(b) ✉ = ✉ ✉ P
eigenstate of e2πiq/α ; rather its value of q is completely
indefinite. To obtain an encoded state, we must project
Z ✐ P −1 ✐ P out the component with a definite value of q modulo α.
This can be achieved by coupling the oscillator to an-
FIG. 7. Construction of the three-qubit gate Λ2 (Z). (a) A other oscillator that serves as a meter, via the perturba-
Λ(P ) gate can be constructed (up to an overall phase) from tion of the Hamiltonian
two S gates, an S −1 gate, and two CNOT’s. The circuit is exe-
H ′ = λ q b† b ,


cuted from left to right. (b) A Λ2 (Z) gates can be constructed (155)
from two Λ(P ) gates, a Λ(P −1 ) gate, and two CNOT’s.
where b is the annihilation operator of the meter.1 This
Therefore, if we can apply symplectic gates accurately, perturbation modifies the frequency of the meter,
and are also able to create a supply of π/8 states of rea-
∆ωmeter = λ q ; (156)
sonable fidelity (or can otherwise implement S gates of
reasonable fidelity), then we can use the purification pro- then if this coupling is turned on for a time t = 2π/λnα,
tocol to implement Toffoli gates with very good fidelity. the phase of the meter advances by

∆θmeter = 2πq/nα . (157)

XII. ENCODING
By reading out the phase, we can determine the value of
Now we have discussed how to execute universal quan- q modulo nα, and apply a shift if necessary to obtain the
tum computation fault tolerantly, and how to perform state with q ≡ 0 (mod nα), the known state |0̄i in the
error recovery. But the discussion has all been premised code subspace. (See Fig. 8.)
on the assumption that we can prepare encoded states.
It is finally time to consider how this can be done. In
fact, preparing simultaneous eigenstates of the stabilizer
generators exp(2πiq/α) and exp(−inpα) is a challenging
task.
For the [[N, k]] stabilizer codes that have been dis- (a)
cussed previously, encoding is not intrinsically difficult q

in that it can be accomplished with Clifford group gates.

Acting by conjugation, Clifford group transformations
take tensor products of Pauli matrices to tensor products
of Pauli operators. In particular, there is a Clifford group
transformation that takes the state |0i⊗N (the simultane- (b)
q
ous eigenstate with eigenvalue one of all N single-qubit nα
Z’s) to the encoded |0̄i⊗k (the simultaneous eigenstate
with eigenvalue one of (N − k) stabilizer generators and
k encoded Z̄’s).
Where our codes are different, in both their finite- FIG. 8. Preparation of an encoded state. (a) An eigenstate
dimensional and infinite-dimensional incarnations, is of p is prepared, which has an indefinite value of q. (b) The
that a single qudit or oscillator is required to obey two in- value of q modulo nα is measured, projecting out a state that
dependent stabilizer conditions – i.e., to be the simultane- differs from the encoded Z̄ eigenstate by a shift in q.
ous eigenstate of two independent Pauli operators. Hence
there is no Clifford group encoder. In the continuous
variable case, the problem can be stated in more familiar
1
language: the symplectic transformations take Gaussian There is an extensive literature on the experimental real-
ization and applications of this kind of coupling; see [37].

19
 Of course, in practice the state squeezed in p prepared S = M1−1 M2 (162)
in the first step will be only finitely squeezed, and the
measurement of q modulo nα will have imperfect resolu- that takes one lattice to the other is also symplectic.
tion. If the squeezed state is Gaussian and the measure- Thus, while the task of preparing the encoded states
ment has a Gaussian acceptance, then this procedure will of the single-oscillator codes can be accomplished only
produce an approximate codeword of the sort described by introducing a nonlinear coupling between oscillators,
in §V. proceeding from single-oscillator encoded states to many-
If we are able to prepare “good enough” encoded oscillator encoded states can be achieved with linear op-
states, we can distill better ones. The distillation proto- tical operations and squeezing.
col is similar to the error recovery procedure, but where
the ancilla used for syndrome measurement may be fairly
noisy. We might improve the convergence of the distil- XIII. PHYSICAL FAULT TOLERANCE?
lation procedure by discarding the data oscillator if the
measurement of the ancilla oscillator yields a value of q In a physical setting, making use of the continuous
or p that is too distant from the values allowed by the variable quantum error-correcting codes proposed here
code stabilizer. (or “digital” quantum codes that have been proposed
So far, we have described how to prepare encoded previously) is a daunting challenge. We must continually
states for the “single-oscillator” codes described in §IV. measure the stabilizer operators (the “error syndrome”)
To prepare an encoded state for one of the N -oscillator to diagnose the errors; to recover we must apply frequent
codes described in §VI, we proceed in two steps. First shifts of the canonical variables that are conditioned on
we prepare each of N oscillators in a single-oscillator en- the measurement outcomes. Cold ancilla oscillators must
coded state. Then we apply a symplectic transformation be provided that are steadily consumed by the syndrome
to obtain the encoded state of the N -oscillator code. measurements. The ancillas must be discarded (or re-
A particular known encoded state of a lattice stabilizer freshed) to rid the system of excess entropy that has been
code can itself be regarded as a code with an (n = 1)- introduced by the accumulated errors.
dimensional code space. Hence it can be characterized An alternative to this complex scheme was suggested
by a self-dual symplectic lattice. For example, the X̄ = 1 in Ref. [38]. Perhaps we can engineer a quantum system
state of a qunit encoded in a single oscillator is the simul- whose (degenerate) ground state is the code subspace.
taneous eigenstate with eigenvalue one of the operators Then the natural coupling of the system to its environ-
e−ipα and e2πiq/α – the state associated
√ with the √ self-dual ment will allow the system to relax to the code space, re-
lattice whose basis vectors are pα/ 2π and q 2π/α. moving errors introduced by quantum and thermal noise,
One encoded state can be transformed to another by or through the imperfect execution of quantum gates.
symplectic gates if there is a symplectic linear transfor- Such a system, if it could be built, would be a highly
mation that takes the self-dual lattice associated with the stable quantum memory.
first state to the self-dual lattice associated with the sec- Continuous variable coding suggests new approaches to
ond. In fact, such a symplectic transformation exists for implementing this type of physical fault tolerance. For
any pair of self-dual lattices. example, the Hamiltonian
A linear transformation acting on the p’s and q’s mod-
ifies the generator matrix M of a lattice according to H = 2 − [cos p + cos(2πnq)] (163)
M → MS ; (158) has an n-fold degenerate (but nonnormalizable) ground
state that is just the code space of a continuous vari-
this transformation is symplectic if
able code. (The operators cos p and cos 2πnq commute
SωS T = ω , (159) and can be simultaneously diagonalized.) The low-lying
states of a real system whose Hamiltonian is a reason-
where able approximation to H would resemble the approxi-
mate codewords described in §V.
 
0 I
ω= . (160) One possible way to realize physical fault tolerance is
−I 0
suggested by the codes for an electron in a Landau level,
We saw in §VI that we can always choose the generator described in §III. The wave functions in the code space
matrix M of a self-dual lattice so that the matrix A has are doubly periodic with a unit cell that encloses n flux
the form quanta, where n is the code’s dimension. If we turn on a
tunable periodic potential whose unit cell matches that
A ≡ M ωM T = ω ; (161) of the code, then the Landau level is split into n energy
that is, so that M is a symplectic matrix. Therefore, bands, and the codewords are the states with vanishing
the generator matrices M1 and M2 of two self-dual lat- Bloch momentum. Therefore, an encoded state could be
tices can each be chosen to be symplectic; then the linear prepared by turning on the potential, waiting for dissipa-
transformation tive effects to cause the electrons to relax to the bottom

20
of the lowest band, and then adiabatically turning off the tum codes can be invoked to investigate the efficacy of
potential. If dissipative effects cause electrons to relax to quantum cryptographic protocols, even in cases where
the bottom of a band on a time scale that is short com- the protocol makes no direct use of the encoded states
pared to spontaneous decay from one band to another, [39]. With continuous variable codes, we can demon-
then more general encoded states could be prepared by strate the security of key distribution protocols based on
a similar method. Furthermore, turning on the potential the transmission of continuous variable quantum infor-
from time to time would remove the accumulated Bloch mation. This application is discussed in a separate paper
momentum introduced by errors, allowing the electron to [40].
relax back to the code space.

ACKNOWLEDGMENTS
XIV. CONCLUDING COMMENTS
We gratefully acknowledge helpful discussions with
We have described codes that protect quantum states Isaac Chuang, Sumit Daftuar, David DiVincenzo, An-
encoded in a finite-dimensional subspace of the Hilbert drew Doherty, Steven van Enk, Jim Harrington, Jeff
space of a system described by continuous quantum vari- Kimble, Andrew Landahl, Hideo Mabuchi, Harsh
ables. With these codes, continuous variable systems can Mathur, Gerard Milburn, Michael Nielsen, and Peter
be used for robust storage and fault-tolerant processing Shor. This work has been supported in part by the
of quantum information. Department of Energy under Grant No. DE-FG03-92-
For example, the coded information could reside in ER40701, and by the Caltech MURI Center for Quan-
the Hilbert space of a single-particle system described tum Networks under ARO Grant No. DAAD19-00-1-
by canonical quantum variables q and p. In practice, 0374. Some of this work was done at the Aspen Center
these variables might describe the states of a mode of for Physics.
the electromagnetic field in a high-finesse microcavity, or
the state of the center of mass motion of an ion in a
trap. Or the continuous Hilbert space could be the state
space of a rotor described by an angular variable θ and
its conjugate angular momentum L; in practice, these
variables might be the phase and charge of a supercon-
ducting quantum dot. Our coding scheme can also be [1] P. W. Shor, “Scheme for reducing decoherence in quan-
applied to a charged particle in a magnetic field. tum computer memory,” Phys. Rev. A 52, R2493 (1995).
Our codes are designed to protect against small errors [2] A. Steane, “Error-correcting codes in quantum theory,”
that occur continually – diffusive drifts in the values of Phys. Rev. Lett. 77, 793 (1996).
the canonical variables. The codes are less effective in [3] E. Knill, R. Laflamme, and G. Milburn, “A scheme for
protecting against large errors that occur rarely. In some efficient quantum computation with linear optics,” Na-
ture 409, 46-52 (2001); “Efficient linear optics quantum
settings, we may desire protection against both kinds of
computation,” quant-ph/0006088.
errors. One way to achieve that would be to concatenate
[4] E. Knill, R. Laflamme, and G. Milburn, “Thresholds for
our continuous-variable codes with conventional finite-
linear optics quantum computation,” quant-ph/0006120.
dimensional quantum codes. [5] E. Knill, “Non-binary unitary error bases and quan-
When we consider how to manipulate continuous- tum codes,” quant-ph/9608048; E. Knill, “Group rep-
variable quantum information fault tolerantly, the issues resentations, error bases and quantum codes,” quant-
that arise are rather different than in previous discussions ph/9608049.
of quantum fault tolerance. With continuous variable [6] H. F. Chau, “Correcting quantum errors in higher
codes, propagation of error from one oscillator to another spin systems,” Phys. Rev. A 55, R839 (1997), quant-
is not necessarily a serious problem. More damaging are ph/9610023; H. F. Chau, “Five quantum register error
processes that amplify a small shift of the canonical vari- correction for higher spin systems,” Phys. Rev. A 56, R1
ables to a large shift. We have described how to im- (1997), quant-ph/9702033.
plement a universal set of fault-tolerant quantum gates; [7] E. M. Rains, “Nonbinary quantum codes,” quant-
with these, harmful error amplification can be avoided as ph/9703048.
the encoded state is processed. [8] D. Gottesman, “Fault-tolerant quantum computation
Apart from encouraging the intriguing possibility that with higher-dimensional systems,” Lect. Notes. Comp.
continuous quantum variables might prove useful for the Sci. 1509, 302 (1999), quant-ph/9802007.
construction of robust quantum memories and comput- [9] A. R. Calderbank, E. M. Rains, P. W. Shor, and
ers, these new quantum codes also have important the- N. J. A. Sloane, “Quantum error correction and orthog-
oretical applications. In this paper we have discussed onal geometry,” Phys. Rev. Lett. 78, 405 (1997), quant-
an application to the theory of the quantum capacity ph/9605005.
of the Gaussian quantum channel. Furthermore, quan- [10] D. Gottesman, “A class of quantum error-correcting

21
 codes saturating the quantum Hamming bound,” Phys. 56 (IEEE, Los Alamitos, CA, 1996), quant-ph/9605011.
Rev. A 54, 1862 (1996), quant-ph/9604038. [28] D. Gottesman, “A theory of fault-tolerant quantum
[11] S. Braunstein, “Error correction for continuous quan- computation,” Phys. Rev. A 57, 127 (1998), quant-
tum variables,” Phys. Rev. Lett. 80, 4084 (1998), quant- ph/9702029.
ph/9711049. [29] A. Steane, “Active stabilization, quantum computation,
[12] S. Lloyd and J. E. Slotine, “Analog quantum error and quantum state synthesis,” Phys. Rev. Lett. 78, 2252
correction,” Phys. Rev. Lett. 80, 4088 (1998), quant- (1997), quant-ph/9611027.
ph/9711021. [30] E Knill, R. Laflamme, W. H. Zurek, “Resilient quantum
[13] S. Parker, S. Bose, and M. B. Plenio, “Entangle- computation: error models and thresholds,” Proc. Roy.
ment quantification and purification in continuous vari- Soc. London, Ser. A 454, 365 (1998), quant-ph/9702058.
able systems,” Phys. Rev. A 61, 32305 (2000), quant- [31] D. Gottesman and I. Chuang, “Quantum teleportation
ph/9906098. is a universal computational primitive,” Nature 402, 390
[14] L. M. Duan, G. Giedke, J. I. Cirac, and P. Zoller, “En- (1999), quant-ph/9908010.
tanglement purification of Gaussian continuous variable [32] S. Schneider, H. M. Wiseman, W. J. Munro, and G. J.
quantum states,” Phys Rev. Lett. 84, 4002-4005 (2000), Milburn, “Measurement and state preparation via ion
quant-ph/9912017; L. M. Duan, G. Giedke, J. I. Cirac, trap quantum computing,” Fort. der Physik 46, 391
and P. Zoller, “Physical implementation for entangle- (1998), quant-ph/9709042.
ment purification of Gaussian continuous variable quan- [33] P. O. Boykin, T. Mor, M. Pulver, V. Roychowdhury, and
tum systems,” Phys. Rev. A 62, 032304 (2000), quant- F. Vatan, “A new universal and fault-tolerant quantum
ph/0003116. basis,” Inform. Process Lett. 75, 101-107 (2000), quant-
[15] A. R. Calderbank and P. W. Shor, “Good quantum error- ph/9906054.
correcting codes exist,” Phys. Rev. A 54, 1098 (1996), [34] X. Zhou, D. W. Leung, and I. L. Chuang, “Methodology
quant-ph/9512032. for quantum logic gate construction,” Phys. Rev. A 62,
[16] A. Steane, “Multiple particle interference and quantum 052316 (2000), quant-ph/0002039.
error correction,” Proc. Roy. Soc. London, Ser. A 452, [35] A. Yu. Kitaev, unpublished.
2551 (1996), quant-ph/9601029. [36] E. Dennis, Fault-tolerant computation without concate-
[17] D. Aharonov and M. Ben-Or, “Fault-tolerant quantum nation, quant-ph/9905027.
computation with constant error,” Proc. 29th Ann. ACM [37] V. Giovannetti, S. Mancini, and P. Tombesi, “Radi-
Symp. on Theory of Computing, p. 176 (ACM, New York, ation pressure induced Einstein-Podolsky-Rosen para-
1998), quant-ph/9611025; D. Aharonov and M. Ben-Or, dox,” quant-ph/0005066, and references therein.
“Fault-tolerant quantum computation with constant er- [38] A. Yu. Kitaev, “Fault-tolerant quantum computation by
ror rate,” quant-ph/9906129. anyons,” quant-ph/9707021.
[18] T. M. Cover and J. A. Thomas, Elements of Information [39] P. W. Shor and J. Preskill, “Simple proof of security
Theory, Wiley, New York (1991). of the BB84 quantum key distribution protocol,” Phys.
[19] A. S. Holevo and R. F. Werner, “Evaluating capacities of Rev. Lett. 85, 441-444 (2000), quant-ph/0003004.
bosonic Gaussian channels,” quant-ph/9912067. [40] D. Gottesman and J. Preskill, “Secure quantum key
[20] S. Lloyd, “The capacity of the noisy quantum channel,” distribution using squeezed states,” Phys. Rev. A 63,
Phys. Rev. A 56, 1613 (1997), quant-ph/9604015. 022309 (2001), quant-ph/0008046.
[21] B. W. Schumacher and M. A. Nielsen, “Quantum data
processing and error correction,” Phys. Rev. A 54, 2629
(1996), quant-ph/9604022.
[22] H. Barnum, M. A. Nielsen, and B. Schumacher, “Infor-
mation transmission through a noisy quantum channel,”
Phys. Rev. A 57, 4153 (1998), quant-ph/9702049.
[23] C. H. Bennett, D. P. DiVincenzo, J. A. Smolin, and W. K.
Wootters, “Mixed state entanglement and quantum er-
ror correction,” Phys. Rev. A 54, 3824 (1996), quant-
ph/9604024.
[24] P. W. Shor and J. A. Smolin, “Quantum error-correcting
codes need not completely reveal the error syndrome,”
quant-ph/9604006; D. P. DiVincenzo, P. W. Shor and
J. A. Smolin, “Quantum channel capacity of very
noisy channels,” Phys. Rev A 57, 830 (1998), quant-
ph/9706061.
[25] We thank the anonymous referee for this comment.
[26] M. Reck, A. Zeilinger, H. J. Bernstein, and P. Bertani,
“Experimental realization of any discrete unitary opera-
tor,” Phys. Rev. Lett. 73, 58 (1994).
[27] P. W. Shor, “Fault-tolerant quantum computation,”
Proc. 37th Annual Symp. on Found. of Comp. Sci., p.

22