PHYSICAL REVIEW A 76, 012325 共2007兲

Protocols for optimal readout of qubits using a continuous quantum nondemolition measurement

Jay Gambetta,1 W. A. Braff,1 A. Wallraff,1,2 S. M. Girvin,1 and R. J. Schoelkopf1

1
Department of Applied Physics and Physics, Yale University, New Haven, Connecticut 06520, USA
2
Department of Physics, ETH Zurich, CH-8093 Zurich, Switzerland
共Received 4 January 2007; published 24 July 2007兲
We study how the spontaneous relaxation of a qubit affects a continuous quantum nondemolition measure-
ment of the initial state of the qubit. Given some noisy measurement record ⌿, we seek an estimate of whether
the qubit was initially in the ground or excited state. We investigate four different measurement protocols, three
of which use a linear filter 共with different weighting factors兲 and a fourth which uses a full nonlinear filter that
gives the theoretically optimal estimate of the initial state of the qubit. We find that relaxation of the qubit at
rate 1 / T1 strongly influences the fidelity of any measurement protocol. To avoid errors due to this decay, the
measurement must be completed in a time that decrease linearly with the desired fidelity while maintaining an
adequate signal to noise ratio. We find that for the nonlinear filter the predicted fidelity, as expected, is always
better than the linear filters and that the fidelity is a monotone increasing function of the measurement time. For
example, to achieve a fidelity of 90%, the box car linear filter requires a signal to noise ratio of ⬃30 in a time
T1, whereas the nonlinear filter only requires a signal to noise ratio of ⬃18.

DOI: 10.1103/PhysRevA.76.012325 PACS number共s兲: 03.67.Lx, 42.50.Lc, 03.65.Yz

I. INTRODUCTION to produce N ⬃ 106 fluorescence photons. Given the photo-

multiplier quantum efficiency and typically small solid angle
In this paper we consider the following problem. Given a coverage, only a very small number n̄d will be detected on
qubit initially either in the ground or excited state with finite average. The probability of getting zero detections 共ignoring
lifetime T1, how can we best make use of a continuous-in- dark counts for simplicity兲 and hence misidentifying the hy-
time noisy quantum nondemolition 共QND兲 measurement to perfine state is P共0兲 = e−n̄d. Even for a very poor overall de-
optimally estimate the initial state 共ground or excited兲 of the tection efficiency of only 10−5, we still have n̄d = 10 and
qubit? A number of authors have previously considered the nearly perfect fidelity F = 1 − P共0兲 ⬃ 0.999 955. It is important
related problem of optimal estimation of the present state of to note that the total time available for measurement is not
the qubit based on the past and current measurement record limited by the phase coherence time 共T2兲 of the qubit or by
关1–6兴, but, to the best of our knowledge, the problem we the measurement-induced dephasing 关3,4,11,12兴, but rather
consider here has not been previously studied. only by the rate at which the qubit makes real transitions
QND measurements play a central role in the theory and between measurement 共␴ˆ z兲 eigenstates. In a perfect QND
practical implementation of quantum measurements 关7兴. In a measurement there is no measurement-induced state mixing
QND measurement, the interaction term in the Hamiltonian 关4兴 and the relaxation rate 1 / T1 is unaffected by the measure-
coupling the system to the measuring apparatus commutes ment process.
with the quantity being measured, so that this quantity is a The ability to read out a qubit with high fidelity is of
constant of the motion. This does not imply that the quantum central importance to the successful construction of a quan-
state of the system is totally unaffected, but it does imply that tum computer 关13兴. In order to successfully measure a qubit,
the measurement is repeatable. For example, a Stern-Gerlach its quantum state must be mapped into a piece of classical
measurement of ␴ˆ z for a spin-1/2 particle initially prepared information by measuring the relative occupation of its two
in an eigenstate of ␴ˆ x will randomly yield the results +1 and states with the highest possible fidelity. Possible qubit imple-
−1 with equal probability. However, all subsequent measure- mentations include superconducting circuits, silicon based
ments of ␴ˆ z will yield exactly the same result as the initial electron and nuclear spins, and trapped ions, among others
measurement. 关14–23兴. In order for qubits prepared in different states to be
The fact that QND measurements are repeatable is of fun- distinguishable, the measurement must be completed before
damental practical importance in overcoming detector ineffi- the excited qubits decay 关4,24兴. Many atomic qubits have
ciencies. A prototypical example is the electron-shelving sufficiently long lifetimes so that relaxation is not a major
technique 关8,9兴 used to measure trapped ions. A related tech- concern 关14,16,20兴, but most solid state qubits have lifetimes
nique is used in present implementations of ion-trap based on the order of microseconds or less, and spontaneous relax-
quantum computation. Here the 共extremely long-lived兲 hy- ation plays a significant role in the measurement. The qubit
perfine state of an ion is read out via state-dependent optical relaxation affects different measurement schemes differently,
fluorescence. With properly chosen circular polarization of but in all cases, it can limit the maximum fidelity.
the exciting laser, only one hyperfine state fluoresces and the Although the behavior of a qubit during continuous mea-
transition is cycling; that is, after fluorescence the ion almost surement has been studied using Monte Carlo simulations
always returns to the same state it was in prior to absorbing 关2–6兴, no attempt has been made to derive an analytical ex-
the exciting photon. Hence the measurement is QND. Typi- pression for the probability distribution of the initial state or
cal experimental parameters 关10兴 allow the cycling transition to study how spontaneous emission impacts the measurement

1050-2947/2007/76共1兲/012325共11兲 012325-1 ©2007 The American Physical Society

GAMBETTA et al. PHYSICAL REVIEW A 76, 012325 共2007兲

fidelity. This is at least partially because there exist very few qubit prepared in the excited state may have decayed during
high fidelity continuous QND measurements, and even fewer tarm and be misidentified. Thus if tarm is not infinitesimally
that operate in a regime where qubit relaxation is a limiting small, the qubit lifetime places a limit on fidelity even with a
factor. With the application of low temperature amplifiers, perfect detector.
high fidelity 共though not necessarily weak continuous兲 mea-
surements of superconducting qubits are now becoming fea-
sible 关19,23,25–32兴. These measurements have found asym- II. QUBIT WITH INFINITE LIFETIME
metries between the probability distributions for the
integrated signal corresponding to the ground and excited It is worthwhile to first consider the case when the qubit
states, but could not accurately predict them. cannot relax from the state it is initially prepared in, so it is
Here we find that when we use a measurement protocol “fixed” for all time. This allows us to formalize our intuitive
that only records the integrated signal, the qubit relaxation understanding of a general continuous measurement and to
induces asymmetry in the probability distributions, and that give us a result against which we can compare the finite
with a sufficiently precise detector, the distributions become lifetime case. We consider the measurement of a qubit with
distinctly non-Gaussian. Unlike measurement of a perfect two states 兩 + 典 共excited兲 and 兩−典 共ground兲 and assume that the
共i.e., nondecaying兲 qubit, where fidelity is always improved measurement result is given by the actual value of the qubit
by a longer measurement, we show that there is some opti- state plus Gaussian noise. This assumption is justified for
mal measurement fidelity that depends on the signal to noise example in the current circuit quantum electrodynamic ex-
ratio 共rsn兲 of the detector and the filter used. The first filter we periments 关11,12,22,23兴 in which a cavity is dispersively
consider is the linear box car filter and optimize over the coupled to the qubit 共no energy is exchanged between the
integrated time tf. Choosing a longer or shorter integration cavity and the qubit兲. A homodyne measurement on the cav-
time will lower the fidelity of the measurement. Next we ity output will reveal the cavity state 共which is proportional
show that by choosing a filter that gives exponentially less to ␴ˆ z兲 plus Gaussian noise 关12,21兴. This Gaussian noise will
importance to results at later times slightly increases the fi- be at least the photon shot noise but in present experiments it
delity. We then numerically find the optimal linear filter and is dominated by the following amplifier. In other words, the
compare these linear filters to a nonlinear filter that yields the measurement is faithful and given that the system is in state
theoretically optimal estimate of the initial state of the qubit i = ± 1 our detector for a time interval d␶ outputs ␺共␶兲 with
given some measurement record ⌿. We find that we can statistics

冑
reach the same fidelity as the linear filters at a substantially
lower rsn. Furthermore, due to the nature of the updating d␶ rsn
P共兩␺兩i兲 = exp关− 共␺ − i兲2d␶ rsn/2兴. 共2.1兲
protocol, the fidelity is a nondecreasing function of the mea- 2␲
surement time. In summary, in this paper we determine the
optimal measurement fidelity given four measurement proto- For convenience we have also introduced a dimensionless
cols for continuous measurement experiments currently be- time ␶ = t / T1 where T1 is an arbitrary but finite number, that
ing performed, and also provides a guideline for the neces- will become the relaxation lifetime when we treat the finite
sary detector signal to noise ratio in order to reach a lifetime case. Here rsn is the ratio of integrated signal power
particular desired fidelity in future experiments. to noise power. It is linear in the integration time and we will
There are two major ways of measuring qubits. The first adopt the convention of specifying the rsn as that achieved
method is a latching measurement, for example by having after integrating for time T1.
the qubit state modify the switching current 共or state兲 of an From this distribution we can write ␺共␶兲 in terms of the
adjacent Josephson junction 关28,29兴 or the bifurcation point Wiener increment dW共␶兲 关35兴 as
of the nonlinear Josephson plasma oscillation 关30–32兴. In
such latching measurements, the qubit is measured very ␺共␶兲d␶ = i±共␶兲d␶ + 冑rsn
−1
dW共␶兲. 共2.2兲
quickly with very high signal to noise, and after only a short Here we have introduced the subscript ± to indicate a pos-
waiting time 关15,17,28,32兴. Some versions of such strong sible realization of the dynamics of the qubit given the initial
measurements can in principle be QND 关32兴. The second condition ±1. For this case the qubit can be initialized in
method is to perform a sequence of repeated or weak con- either state, but because it has an infinite lifetime, it is fixed
tinuous quantum nondemolition 共QND兲 measurements, in whatever state it starts in for the duration of the measure-
which each leave the populations of the qubit unchanged. ment. That is, i±共␶兲 = ± 1.
Several recent experiments with solid-state qubits 关23,33,34兴, We define our measurement signal s as the output of the
have used continuous QND measurement schemes in which detector integrated over time ␶ f
the qubit lifetime imposes the main limitation on the mea-
surement fidelity, for which the analysis of this paper should
apply.
s= 冕
0
␶f
d␶ ␺共␶兲. 共2.3兲
It is not just continuous measurements that are affected by
qubit relaxation. For example, consider an idealized latching Formally we are restricting ourselves here to a simple box
measurement scheme where a qubit is prepared in an eigen- car linear filter which uniformly weights the measurement
state, but there is some finite arming time tarm before a per- record ␺共␶兲 in the interval 0 ⬍ ␶ ⬍ ␶ f . Using Eq. 共2.2兲 it is
fect measurement is made. In this case, a qubit prepared in simple to carry out the above integral and rewrite the mea-
the ground state will always be measured correctly, but a surement signal as s±共␶ f 兲 = ± ␶ f + XG关0 , ␴2兴 where XG关0 , ␴2兴 is

012325-2
PROTOCOLS FOR OPTIMAL READOUT OF QUBITS USING… PHYSICAL REVIEW A 76, 012325 共2007兲

TABLE I. The required minimum signal to noise ratio rsn after T1 and the optimal measurement time ␶opt
for a continuous measurement of a qubit with infinite lifetime 共rsn共fixed兲兲, a simple box car linear filter
共rsn共BC兲兲, an exponentially decaying linear filter 共rsn共exp兲兲, the optimal linear filter 共rsn共OL兲兲, and a nonlinear
Bayesian filter 共rsn共NL兲兲. The last column is the maximum allowed waiting time ␶arm for an idealized latching
共instantaneous兲 measurement with infinite signal to noise performed after time ␶arm in order to achieve the
desired fidelity F.

F rsn共fixed兲 rsn共BC兲共␶opt兲 rsn共exp兲共␶opt兲 rsn共OL兲 rsn共NL兲 ␶arm

50% 0.50 1.47共0.82兲 1.23共1.59兲 1.17 1.1 0.7

67% 1.0 4.05共0.55兲 3.58共0.75兲 3.10 2.9 0.40
90% 2.7 29.9共0.17兲 28.7共0.18兲 22.1 18. 0.11
95% 3.8 77.3共0.087兲 75.7共0.090兲 55.9 48. 0.05
99% 6.7 574.共0.018兲 572.共0.018兲 420 269. 0.01

a Gaussian random variable of mean 0 and standard devia- a fundamental asymmetry in how the excited and ground
tion ␴ = 冑␶ f / R共S/N兲. The signals then follow the familiar states behave. We assume 共in our model兲 that the excited
Gaussian distributions state decays at rate 1 / T1 but the transition rate upward out of
the ground state is zero. A qubit prepared in the ground state
1 2/共2␴2兲
P±fixed共s兲 = e−共s ⫿ ␶ f 兲 . 共2.4兲 will never experience any excitations, so it can be treated
␴ 冑2 ␲ much the same as the fixed qubit discussed above and hence
P−共s兲 = P−fixed共s兲. By contrast, an initially excited qubit will
Because these distributions are symmetric about s = 0, the
produce an ensemble averaged output which will decay ex-
most obvious analysis is to set a signal threshold ␯th = 0 and
ponentially with characteristic dimensionless time ␶1 ⬅ 1. Al-
call every measurement with s ⬎ ␯th a 共+1兲 state, and every
though most qubits dephase in some shorter dimensionless
measurement with s ⬍ ␯th a 共−1兲 state. Calculation of fidelity time ␶2, our only concern here is the relative population of
in this case is accomplished using the definition the two qubit states, so we are not limited by decoherence of

F⬅1− 冕 −⬁
␯th
ds P+共s兲 − 冕 ⬁

␯th
ds P−共s兲. 共2.5兲
the qubit. For a single qubit, this translates into a single,
abrupt relaxation of the qubit at some dimensionless time ␶d
that is exponentially distributed with mean dimensionless
For the case of infinite qubit lifetime we have the simple time 1, P共␶d兲 = exp共−␶d兲. That is, if the qubit was initially in
result the excited state then i±共␶兲 would obey

F = erf 冉冑 冊 ␶ f rsn
2
. 共2.6兲 i±共␶兲 = ␪共␶d − ␶兲 − ␪共␶ − ␶d兲. 共3.1兲

A fidelity of zero corresponds to a completely random Thus given a possible realization i±共␶兲, we can generate a
measurement that extracts no information, a fidelity of one typical record an experimentalist would measure by using
corresponds to a perfect faithful measurement, and in be- Eq. 共2.2兲. A typical trajectory for ␺共␶兲 is shown in Fig. 1 for
tween the measurement conveys varying degrees of certainty. a rsn of 570. Here we see, at this value of rsn, the jump in
As ␶ f becomes large, Eq. 共2.5兲 predicts that the fidelity rap- output is clearly visible and occurs at ␶d ⫽ 1 共td ⫽ T1兲.
idly approaches unity. Higher rsn serves to speed up the con- In the case where the qubit happens to decay early, ␶d
1,
vergence, but as long as rsn is non zero, any desired fidelity is the signal s from a qubit initially prepared in the excited state
attainable simply by measuring the qubit for long enough. In would be almost indistinguishable from that of a qubit ini-
Table I, the required rsn共fixed兲 is listed in order to achieve a tially prepared in the ground state, and even if the measure-
given fidelity within T1. Note the same results for the fidelity ment apparatus were nearly perfect, almost no information
would be obtained if we used the optimal nonlinear filter of could be extracted. The probability distribution P+共s兲 for the
Sec. VI. That is, for a qubit with infinite lifetime the simple measurement signal of an initially excited qubit can be de-
box car filter is equivalent to the optimal nonlinear filter, it is termined analytically with a simple derivation. This will al-
only when we include relaxation this is not the case. This low a more quantitative discussion of fidelity, and will even-
will be discussed in detail in the next four sections. tually allow for optimization of integration time, ␶ f . The
critical difference between the infinite lifetime system de-
III. BOX CAR LINEAR FILTER scribed in Sec. II and an actual QND measurement is that s is
now a function not only of the dimensionless integration
A. Probability distributions for QND measurement
time ␶ f , but also the exponentially distributed random relax-
Here we consider the same measurement protocol as in ation time ␶d.
Sec. II, however unlike a qubit with infinite lifetime, where If the qubit is initially in the excited state, then from Eqs.
both states behave similarly, a qubit with a finite lifetime has 共2.2兲, 共2.3兲, and 共3.1兲 a possible realization of s± will be

012325-3
GAMBETTA et al. PHYSICAL REVIEW A 76, 012325 共2007兲

Detector output, ψ(τ) 2 sary to include these extra terms, but for now the demolition
is taken to be small compared to the spontaneous relaxation
1
and can be ignored.
The Gaussian first term in Eq. 共3.4兲 is dominant for
␶ f
1, and P+共s兲 is nearly symmetrical to P−共s兲. For these
0 fast measurements, there is little chance that the qubit decays
during the measurement, and the distributions will be very
−1 similar to those of fixed state bits. At some point, the prob-
ability that the qubit has relaxed will be large enough that it
−2 τd significantly affects the distributions. At this point, predic-
tions based on the assumption of no relaxation are no longer
0 0.5 1.0 1.5 2.0 valid, and the non-Gaussian term in Eq. 共3.4兲 becomes very
Time, τ [T1] important. This effect can be seen in the non-Gaussian tail of
P+共s兲 in the second two time cuts in Fig. 2.
FIG. 1. 共Color online兲 Example detector output as a function A strongly non-Gaussian P+ distribution is not required
of time for a continuous QND measurement with rsn = 570 for the fidelity to exhibit a maximum at some finite measure-
共d␶ = 1 ⫻ 10−4兲, which allows for a measurement fidelity of 99%. ment time. Even though for experiments with rsn ⬇ 1, the
The dashed red line corresponds to a large ensemble average of non-Gaussian tail of P+ is not very prominent, the asymme-
qubit outputs, and decays exponentially with characteristic time T1. try between the distributions in both height and width is
The solid blue line corresponds to the measurement record for a quite clear. For long enough dimensionless integration times,
particular single shot. At this value of rsn, the jump in output is
a qubit initially in the excited state will have probably re-
clearly visible and occurs at ␶d.
laxed relatively early in the measurement, so its mean should
be very similar to that of a qubit initially in the ground state,
s± = ␶ f ␪共␶d − ␶ f 兲 + 共2␶d − ␶ f 兲␪共␶ f − ␶d兲 + XG关0, ␴2兴. the distributions will be almost identical, and all resolving
共3.2兲 power will be lost.
B. Optimal box car filter
That is, the probability distribution for s given a realization
with a decay at time ␶d is The behavior of fidelity as a function of integration time
for a QND qubit measurement is very different from a mea-
1 surement of a fixed state qubit that never relaxes. Recall that
P+共兩s兩␶d兲 =
冑2␲␴ exp关− 共s − ␶ f 兲 /2␴ 兴␪共␶d − ␶ f 兲
2 2
in the fixed state case, the fidelity eventually converges to
one, independent of rsn. We define fidelity as we did for the
1 fixed state case in Eq. 共2.5兲. The difference here is that be-
冑2␲␴ exp兵− 关s − 共2␶d − ␶ f 兲兴 /2␴ 其␪共␶ f − ␶d兲,
2 2
+ cause the distributions are not necessarily symmetrical, the
signal threshold ␯th is not always zero. Maximizing F with
共3.3兲 respect to ␯th yields the following implicit equation for ␯th

and from this one can easily obtain the probability distribu- P+共␯th兲 = P−共␯th兲. 共3.5兲
tions for s by averaging over all possible realizations 共decay As an aside we note that from this we see that we can write
times兲. Doing this gives the fidelity 共optimized with respect to ␯th兲 in the alternative
form
1 1
冕
2/共2␴2兲−␶
P+共s兲 = e−共s − ␶ f 兲 + e−共s+␶ f 兲/2 +⬁
冑2␲␴
f
4 1
ds兩P+共s兲 − P−共s兲兩. 共3.6兲

再冉 冊冎
F=

冊 冉
2 −⬁
2/8 ␴ − 2共s − ␶ f 兲
2
␴ − 2共s + ␶ f 兲
2
⫻ e␴ erf − erf .
2 冑2 ␴ 2 冑2 ␴ Despite the complications of finite lifetime, the integrations
in Eq. 共2.5兲 can still be carried out analytically to yield

再冉 冊
共3.4兲
1 2 ␴2 − 2共␯th − ␶兲
F = e␴ /8e−共␯th+␶兲/2 erf
2 冑2 ␴
Although it does not figure directly into this analysis, it is
2

冊冎
not difficult to expand this treatment to consider a measure-
ment with finite “demolition” that stimulates both excitation
and relaxation of the qubit. In this case, both P+ and P− will
be non-Gaussian, because the qubit may excite and relax
− erf冉 ␴2 − 2共␯th + ␶兲
2 冑2 ␴
. 共3.7兲

several times during the measurement interval. To calculate The fidelity is maximized by numerically solving for ␯th such
the distributions, all we need to do is extend the possible that P−共␯th兲 = P+共␯th兲. Using the correct value of ␯th共␶ f 兲, it is
realization to include multiple relaxations and excitations straightforward to compute F共␶f兲 and then vary ␶ f to obtain
keeping in mind that the relaxation and excitation times are the optimal value of the integration time.
not independent variables: ␶dn must occur before ␶dn+1. For Roughly speaking, fidelity is a measure of how separate
stronger or less ideal measurements, it may become neces- the probability distributions are, ranging between 0 for com-

012325-4
PROTOCOLS FOR OPTIMAL READOUT OF QUBITS USING… PHYSICAL REVIEW A 76, 012325 共2007兲

7
6
5
7 4
6 3
5 2
7 4 1
6 3 0
5 2 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6
4 1 τf = 0.3
P+/-(s)

3 0
2 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6
νth = -0.03 τf = τopt = 0.2
1
0
-0.6 -0.4 -0.2 0.0 0.2 0.4 0.6
τf = 0.1
Signal, s [T1]

FIG. 2. 共Color online兲 Evolution in time of probability distributions for QND measurement of qubits initialized in the ground 共blue兲 and
excited 共red兲 states for the box car linear filter and a rsn = 30. For short ␶ f , the distributions are almost symmetrical because an excited qubit
has probably not decayed yet. As ␶ f gets longer, the qubit is much more likely to have decayed, and the mean of P+共s兲 begins to drift back
towards that of P−共s兲. At some optimal point in between, ␶ f = ␶opt, and if the threshold ␯th is chosen so that P+共␯th兲 = P−共␯th兲, the misidentified
tails are minimized, and fidelity is maximized.

plete overlap and 1 for no overlap. Fidelity is limited by

detector noise for short times and spontaneous relaxation for
long times. For ␶ f
1, the probability of a relaxation during
the measurement is very low, and the fidelity behaves simi-
larly to the fixed state case: it increases with increasing ␶ f .
For longer ␶ f ⬇ 1, the qubit is more and more likely to have
relaxed during the measurement and the mean of P+共s兲 will
stop increasing linearly and in the long time limit will actu-
ally decrease in time as can be inferred from the decrease in
optimal threshold value plotted in Fig. 3共b兲 共solid line兲. This
implies the existence of some intermediate time ␶opt共rsn兲 that
maximizes the fidelity, this is clearly seen in the red solid
line of Fig. 3共a兲 where the fidelity for a rsn = 10 has a maxi-
mum of 0.79 at ␶opt = 0.34.
To find this time for a given value of rsn, we compute
F共␶ f 兲, then numerically solve dF共␶ f 兲 / d␶ f = 0 for ␶ f = ␶opt. The
solution, Fopt ⬅ F共␶opt兲 is the best possible fidelity for the
simple linear box car filter; no improvement can be made
from this value without improving the measurement appara-
tus or the qubit lifetime. This optimal fidelity is only achiev-
able by correctly setting ␶ f = ␶opt and ␯th. Any variation of
these parameters will reduce the fidelity. It should be pointed
out that while the condition P−共␯th兲 = P+共␯th兲 does maximize
the fidelity, it implies that the measurement protocol is biased FIG. 3. 共Color online兲 共a兲 The fidelity F as a function of time ␶ f
for all measurement protocols. The box car filtered integrated signal
towards the ground state. That is, if the qubit is prepared in
is the red 共solid兲 line and the exponential filtered integrated signal is
the ground state we are more likely to assign it correctly than
the blue 共dashed兲 line. Here we see that both these schemes have an
if it was prepared in the excited state. An unbiased measure-
optimal measurement time and that the latter case is less sensitive to
ment protocol is obtained by setting ␯th such that the measurement time. The nonlinear filter is shown by the green
␯th
兰−⬁ ds P−共s兲 = 兰␯⬁ ds P+共s兲. Doing this results in a slightly 共dashed-dotted兲 line. Here we see that it is clearly better than the
th
lower optimal fidelity. other cases and that there is no optimal measurement time. 共b兲 The
Since Fopt depends only on rsn, it is possible to first derive threshold ␯th as a function of time ␶ f for both the box car filtered
the required signal to noise ratio after one lifetime, and then integrated signal 共red solid line兲 and the exponential filtered inte-
choose the correct measurement time and signal threshold grated signal 共blue dashed line兲. In both subplots the rsn after time
required in order to attain any arbitrary fidelity as shown in ␶ f = 1 is 10 and time is measured in units of T1.

012325-5
GAMBETTA et al. PHYSICAL REVIEW A 76, 012325 共2007兲

F0 ⬃ 1 −
1
rsn
ln 冉 冊
1 + rsn
冑␲ . 共3.10兲

This can be rewritten as in terms of the optimal measurement

time as F0 ⬃ 1 − ␶opt / 2. Here we see that to achieve the opti-
mal, the measurement must be completed in a time that de-
creases linearly with the desired fidelity. The slight devia-
tions form this zeroth order result is shown in the inset of
Fig. 4共b兲. This result is consistent with the intuitive picture
that to achieve a fidelity F = 1 − ␩ with ␩
1, we must have
sufficient rsn to be fooled only by those extremely early de-
cays which occur in time ␶* ⬃ 2␩. This set of decays occurs
with probability ⬃2␩ and is the main source of the infidelity.

IV. EXPONENTIAL DECAYING LINEAR FILTER

In this section we consider the case when we include an

exponential weighting factor in our integrated signal s. This
is chosen as this would be the optimal linear filter if our
signal was simply a decaying exponential with a random
initial amplitude, this can be proven by minimizing our esti-
mate of the amplitude in a least square sense. However, even
though our signal on average is of this form, in one particular
FIG. 4. 共Color online兲 共a兲 Optimal fidelity Fopt and 共b兲 measure- run it is not. Thus this will not be the optimal linear filter for
ment time ␶opt as functions of rsn. The inset in 共b兲 shows the optimal estimating the initial state of the qubit. For this reason we
fidelity as a function of the measurement time. The red 共solid兲 line introduce an integration time ␶f and optimize the measure-
is for the box car filter, blue 共dashed兲 is for the exponential filter, ment fidelity over this. That is, the signal s for this filter is
black 共dotted兲 is for the optimal linear filter, and green 共dashed- given by
dotted兲 is for the nonlinear filter. Here we see that by using the

冕
optimal nonlinear filter we get an improved fidelity in comparison
␶f
to the simple linear filters. Particular values are shown in
s= d␶ ␺共␶兲e−␶ . 共4.1兲
Table I. 0

Fig. 4 关solid line in 共a兲兴 along with the optimal measurement As in the last section we first determine a possible s given
time 关solid line in 共b兲兴. For example, one standard initial goal that the qubit started in the excited state. Doing this gives
is a fidelity of 90%, sufficient to violate Bell’s inequalities
关36兴. As shown in Table I, this fidelity requires a minimum
rsn of 30 after time T1.
The following argument shows how rapidly the required
s± = 冕
0
␶f
d␶关␪共␶d − ␶兲 − ␪共␶ − ␶d兲兴e−␶ + 冑rsn
−1
冕
0
␶f
dW共␶兲e−␶
rsn diverges for very high values of fidelity. For large rsn, it is
a good approximation to set the threshold ␯th = 0 and then the 共4.2兲
fidelity is approximately
and by treating the noise integral as simply a linear combi-
F0 = exp共− ␶ f /2兲erf共冑rsn␶ f /2兲. 共3.8兲 nation of infinitesimal Gaussian variables gives

Optimizing this with respect to ␶ f and then using the s± = 关1 − 2e−␶d + e−␶ f 兴␪共␶ f − ␶d兲 + a␪共␶d − ␶ f 兲 + XG关0, ␴2兴,
asymptotic form for the error function of large argument 共4.3兲
yields the following expression for the optimal integration
where ␴ = 冑共1 − e−2␶ f 兲 / 2rsn and a = 1 − e−␶ f . Here we see that
time

␶opt ⬇
2
再
rsn
1
x0 − ln共x0兲 ,
2
冎 共3.9兲
unlike before as the measurement time becomes large the
variance saturates at 共2rsn兲−1 rather than continuing to in-
crease linearly with time. That is, we have designed our filter
such that when all the information about the qubit state has
where x0 ⬅ ln[共1 + rsn兲 / 冑␲]. been lost into the T1 environment the noise in the integrated
Approximating the error function in Eq. 共3.8兲 by unity signal will remain constant.
and neglecting ln共x0兲 relative to x0 leads to the following Following the same procedure as before the excited state
simple asymptotic form for large rsn: distribution is

012325-6
PROTOCOLS FOR OPTIMAL READOUT OF QUBITS USING… PHYSICAL REVIEW A 76, 012325 共2007兲

P+共s兲 =
1
4
冋 冉 冊 冉 冊册
erf
a+s
冑2␴ + erf
a−s
冑2␴ a共l兲 = 冕 l
k共␶⬘兲d␶⬘ , 共5.3兲

冋 册
0

共s − a兲2

冑冕
1
+ exp − − ␶f . 共4.4兲
冑2␲␴ 2␴2
␴=
⬁
k2共␶⬘兲d␶⬘/rsn . 共5.4兲
0
Repeating the above for the ground state initial condition
gives s f = −a + XG关0 , ␴2兴 and a ground state distribution of the If the qubit was initially in the ground state then the signal
form would be

P−共s兲 =
冑2␲␴
1
exp −冋 共s + a兲2
2␴2
册. 共4.5兲
s− = − a共⬁兲 + XG关0, ␴2兴.
From the above equations the excited state and ground state
共5.5兲

With the above two distributions the fidelity, as before, can probability distribution for the signal s are

冕
be solved analytically and in terms of the ␯th is ⬁
exp兵− ␶d − 关s − 2a共␶d兲 + a共⬁兲兴2/2␴2其

冉 冋 册 冋 册冊 P+共s兲 = d␶d ,
␴ 共a − ␯th兲 2
共a + ␯th兲 2
0 冑2␲␴2
F=
冑8␲ exp − 2␴2 − exp −
2␴2 共5.6兲

+
共e−␶ f + 1 − ␯th兲
4
冋 冉 冑 冊 冉 冑 冊册
erf
a − ␯th
2␴
+ erf
a + ␯th
2␴
.
P−共s兲 =
exp兵− 关s + a共⬁兲兴2/2␴2其
共5.7兲
冑2␲␴2
Using the same procedure as before we can numerically
determine the ␯th which maximizes the fidelity. For a rsn of and by using Eq. 共2.5兲 the fidelity is
10 共in time T1兲 the fidelity as a function of measurement time
is shown in Fig. 3共a兲 as a blue dashed line. Here we see as
before there is an optimal measurement time. To measure F=
1
2
冕 ⬁

0
e−␶derf冉 2a共␶d兲 − a共⬁兲 − ␯th
冑2 ␴ 冊 d␶d

冉 冊
any longer than this time results in a lower fidelity. The
optimal fidelity and measurement time are shown in Fig. 4 1 − a共⬁兲 − ␯th
− erf . 共5.8兲
共blue dashed line兲 as a function of the rsn. Here we see that 2 冑2 ␴
by using this filter, the fidelity is slightly better than the
simplest case 共see Table I for some values兲, but more impor- Maximizing this gives the following set of coupled differen-
tantly the curvature of the fidelity at ␶opt is less. This means tial equations
that this filter is less sensitive to errors in the measurement
time. That is, this protocol would be more practical to imple- d␶a共␶兲 = k共␶兲, 共5.9兲
ment than the simple box car integrated signal of Sec. III.
d␶k共␶兲 = − exp兵− ␶ − 2a共␶兲关a共␶兲 − a共⬁兲 − ␯th兴/␴2其,
V. OPTIMAL LINEAR FILTER 共5.10兲
In this section we calculate the optimal linear filter for with the initial conditions k共0兲 = 1, a共0兲 = 0, and the boundary
estimating the initial state of the qubit. We define the linear condition k共⬁兲 = 0. It is this latter condition which determines
signal s by the relation ␯th. This system of equations can be solved numerically us-

s= 冕
0
⬁
k共␶兲␺共␶兲d␶ , 共5.1兲
ing a shooting method 关37兴. The results are shown in Fig. 5
for a rsn of 1.0 and 172. Here we see that for the small rsn,
the kernel k共␶兲 can be approximated well by
exp共−␤␶兲 where ␤ is a fit parameter that is approximately
where the kernel k共t兲 is unknown and is determined by maxi-
equal to 1 + rsn / 2. In the large rsn limit the k共␶兲 cannot be fit
mizing the measurement fidelity. This, as before, is defined
by an exponential. For illustrative purposes the optimal lin-
as the difference between the probability of us making a
ear filter is compared with the optimal box car linear filter in
correct assignment and an incorrect assignment 关Eq. 共2.5兲兴.
Fig. 5共b兲 for a rsn of 172. Here we see that the time when the
The assignment criteria we use is again if s is above ␯th then
box car linear filter turns off, is comparable with the time
we say the qubit was initially up and if it is below ␯th then it
scale of the optimal linear filter.
was down. ␯th like k共t兲 is determined by the maximization
Numerically solving for k共␶兲 for a given ␴ we can use
procedure which we will describe now.
Eqs. 共5.4兲 and 共5.8兲 to plot the fidelity of the optimal linear
For an unknown kernel the signal conditioned on the qu-
filter as a function of the rsn. This is shown in Fig. 4 as a
bit being initially up will be given by
dotted black line and in Table I as column 4. Here we see
s+ = 2a共␶d兲 − a共⬁兲 + XG关0, ␴2兴, 共5.2兲 that the optimal linear filter out performs the other linear
filters and is almost as good as the optimal filter 共which is
where nonlinear兲 and is described in the next section.

012325-7
GAMBETTA et al. PHYSICAL REVIEW A 76, 012325 共2007兲

0 is calculate P共⌿ 兩 i0兲. This is the nontrivial step, but can be

determined by
−2 P共兩⌿兩i0兲 = 兺 P共兩⌿兩i p兲P共兩i p兩i0兲, 共6.2兲
Log[k(τ)]

−4 where p labels a possible realization of the qubit trajectory.

That is, P共⌿ 兩 i0兲 can be determined by taking the ensemble
−6 average of P共⌿ 兩 i p兲 for all possible realizations, i p 关i p is
given by Eq. 共3.1兲 if the qubit is initially in the excited state
or i p = −1 if initially in the ground state兴. From our simple
−8 model for the noise, Eq. 共2.1兲, P共⌿ 兩 i p兲 is simply a multipli-
(a) 1 0 0.5 1 1.5 2
cation of many Gaussians each centered on the instantaneous
value of i p.
0.8 For an excited state initial condition, the probability of
getting record ⌿ will be given by

冕
0.6 ␶f
k(τ)

P共兩⌿兩i0 = 1兲 = A d ␶ de −␶d
0.4 0

0.2 ⫻exp − 冋冕 0
␶d
d␶关␺共␶兲 − 1兴2rsn/2 册
(b)
0
0 0.02 0.04
Time, τ [Τ1]
0.06 0.08 0.1
⫻exp − 冋冕 ␶f

␶d
d␶关␺共␶兲 + 1兴2rsn/2 册
FIG. 5. 共Color online兲 The optimal linear filter for a rsn of 1.0
blue 共dashed兲 and rsn of 172 red 共solid兲 as a function of time. The
green 共dashed-dotted兲 line in 共b兲 is the optimal box car linear filter
冋冕
+ e−␶ f exp −
0
␶f
册
d␶关␺共␶兲 − 1兴2rsn/2 .

for a rsn of 172 共␶opt = 0.046兲. 共6.3兲

Here the first term represents all possible trajectories that
VI. OPTIMAL NONLINEAR FILTER have had a decay in time ␶ f 共the integration time兲, the second
term represents all the trajectories that did not decay in this
In the previous sections we considered only the case time, and A is the normalization constant. If the qubit were
where one looks at or only has access to the integrated sig- initially in the ground state then the probability of getting
nal, s共␶兲. Here we assume that we have access to the full record ⌿ will be given by

冋冕 册
record ⌿ = 兵␺共␶兲其 and ask how much better can we do with a
␶f
nonlinear filter. In particular, given this record what is our
P共兩⌿兩i0 = − 1兲 = A exp − d␶关␺共␶兲 + 1兴2rsn/2 .
best guess at the initial state of the qubit. Mathematically, our 0
best guess can be represented by the probability distribution
P共i0 兩 ⌿兲. This is the probability that the initial condition i0 is 共6.4兲
±1 given the record, ⌿. As with all probability distributions, Now that we have expressions for P共⌿ 兩 i0兲 all that we
this will range from 0 to 1, and the closer it is to one the need to do to get our best estimate of the initial state is to use
more we are certain that the qubit was in the initial state i0 Eq. 共6.1兲. Doing this gives

冋冕
= ± 1. This by definition is the optimal protocol as it is the
␶f
best estimate of i0 given the complete set of information 1
P共兩i0 = 1兩⌿兲 = d␶de−␶dexp兵关s共␶d,0兲 − s共␶, ␶d兲兴rsn其
available. To find this distribution we use Bayes theorem, N 0

P共兩i0兩⌿兲 =
P共兩⌿兩i0兲P共i0兲
兺i 0
P共兩⌿兩i0兲P共i0兲
, 共6.1兲 + e−␶ f exp关s共␶ f ,0兲rsn兴 , 册 共6.5兲

where P共i0兲 is the initial probability distribution and P共⌿ 兩 i0兲 1

P共兩i0 = − 1兩⌿兲 = exp关− s共␶ f ,0兲rsn兴, 共6.6兲
is the probability that we would measure record ⌿ given that N
the experiment was initially prepared in state i0. We will
where the two time integrated signal s共␶ , ␶⬘兲 is
assume that our experiment can prepare unbiased initial
states and as such we take P共i0 ± 1兲 = 1 / 2. The conditional
distribution P共⌿ 兩 i0兲 is the probability that we would mea- s共␶, ␶⬘兲 = 冕␶⬘
␶
d ␶ ⬙␺ 共 ␶ ⬙兲 共6.7兲
sure record, ⌿, given that the experiment was initially pre-
pared in state i0. Thus to calculate P共i0 兩 ⌿兲 all we need to do and the norm is simply

012325-8
PROTOCOLS FOR OPTIMAL READOUT OF QUBITS USING… PHYSICAL REVIEW A 76, 012325 共2007兲

Estimate state, z(τf) 1.0 predicting the real initial condition and we are almost certain
for the excited case. However this is only a typical trajectory
and to get a better understanding of the predictability of this
0.5 method we use the same fidelity measure as before, that is
we subtract our wrong guesses from our correct guesses. To
be more specific, we define the following assignment proce-
0.0 dure: if z̃ ⬎ 0 then we assign the qubit as up and if z̃ ⬍ 0 we
assign it as down. If z̃ is equal to zero we ignore the result 共or
flip an unbiased coin to make our decision兲. Given this as-
-0.5
signment criteria we can define the fidelity as

(a) 0 0.5 1.0

Time, τf [Τ1]
1.5 2.0
F = lim
1
M→⬁ 2M 冋兺
z̃+1⬎0
− 兺
z̃+1⬍0
+ 兺
z̃−1⬍0
− 兺
z̃−1⬎0
册 , 共6.9兲
Number of estimates

6000 where M is the number of randomly generated ⌿ for both +1

5000 and −1 initial conditions. This is shown in Figs. 3共a兲 and 4共a兲
as the dashed-dotted line for M = 104 共specific optimal fideli-
4000 ties are listed in Table I兲. Here we see that the fidelity is
3000 always better than the other cases and that there is no optimal
measurement time. That is, unlike the box car filter and the
2000 exponential linear filter, the fidelity is a nondecreasing func-
tion of the integration time.
1000
It should be noted that as in the previous protocols, this is
0 a biased measurement which favors ground state prepara-
-1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0 tions. This can be seen by looking closely at Fig. 6共b兲. This
(b) Estimate state, z(τf → ∞) figure shows a histogram of 104 estimates when the system is
prepared in the excited state 共dark bars, column 1兲 and the
ground state 共light bars, column 2兲 for a rsn = 10. If we use
FIG. 6. 共Color online兲 共a兲 A typical trajectory for the estimated
initial condition, z̃共␶ f 兲, given a record ⌿ which is randomly gener-
our assignment procedure and subtract the wrong guesses
ated for an excited state initial condition 共blue dashed line兲 and a from the correct guesses for each prepared initial state we get
ground state initial condition 共red solid line兲. 共b兲 A histogram of a fidelity of 0.92 for the ground state and 0.76 for the excited
z̃共␶ f → ⬁ 兲 for 104 trajectories for both an excited state initial con- state 共an average fidelity of 0.84兲. This is because when the
dition 共dark blue bars column 1兲 and a ground state initial condition qubit is prepared in the excited state, rare, early decays will
共light red bars column 2兲. The rsn after ␶ f = 1 is 10 and time is fool the detector.
measured in units of T1.
VII. COMPARISON WITH LATCHING MEASUREMENT

N= 冕 0
␶f
d␶de−␶dexp兵关s共␶d,0兲 − s共␶, ␶d兲兴rsn其 The challenges from qubit relaxation to attaining high fi-
delity in a continuous measurement also exist in other mea-
+ e−␶ f exp关s共␶ f ,0兲rsn兴 + exp关− s共␶ f ,0兲rsn兴. 共6.8兲 surement schemes. For comparison, consider a latching mea-
surement where the qubit state triggers a classical switching
Here we see that to solve this equation we need to evalu- event in the measurement apparatus 关28–32兴. Such a mea-
ate a double integral over a stochastic process. This is im- surement has the advantage that the detector state stays
practical to solve numerically, however, as shown in the Ap- latched for a very long time, so that noise from subsequent
pendix we can easily recast these integrals in terms of two amplification stages is completely negligible and the rsn is
sets of two coupled stochastic differential equations which effectively infinite. We can roughly model this process as an
require similar computational resources to that used with the instantaneous measurement with no errors but a finite arming
linear filters. time tarm needed to set up the prelatched state of the detector.
To show a typical trajectory for this estimated initial con- We assume that the arming stage occurs after the qubit is
dition, we randomly generated records for both an excited prepared but before it is measured. The measurement of a
and ground state initial condition. Rather then plotting both qubit in the excited state will be wrong if the qubit relaxes
P共i0 = 1 兩 ⌿兲 and P共i0 = −1 兩 ⌿兲 we define z̃ = P共i0 = 1 兩 ⌿兲 before the measurement is made. The probability that the
− P共i0 = −1 兩 ⌿兲 共this is the estimator that replaces s used in qubit relaxes during tarm rises exponentially towards 1, so the
the linear filters兲, this will range from −1 to 1 and the closer fidelity falls exponentially, F共␶arm兲 = e−␶arm. As shown in Table
it is to one of these limits, the more certain we are that the I, ␶arm must be slightly smaller than ␶opt. Although these
initial condition which generated ⌿ is this value. The results times are different in that ␶arm is a maximum value whereas
of this simulation are shown in Fig. 6共a兲. The solid line cor- ␶opt is an optimal value for a given rsn, they provide a valu-
responds to the case when the initial state was the ground able comparison between the measurement schemes.
state and the dashed line is for the excited state. Here we see Of course, no latching measurement is truly instantaneous
that for this typical trajectory our estimate is fairly good at with perfect accuracy in translating the qubit state into latch-

012325-9
GAMBETTA et al. PHYSICAL REVIEW A 76, 012325 共2007兲

ing events. Moreover, a latching measurement usually in- P̄−1,␺

duces some mixing of states, so the actual fidelity may be P共i0 = − 兩1兩⌿兲 = , 共A2兲
lower than for an equivalent continuous QND measurement. P̄−1,␺ + P̄+1,␺
Both measurement schemes have their own strengths and
weaknesses, but in either case, qubit decay can be a signifi- where
cant limiting factor on the fidelity.

VIII. CONCLUSION
P̄+1,␺ = e−␶ f exp关s共␶ f ,0兲rsn兴 + 冕
0
␶f
d ␶ de −␶d

⫻exp兵关s共␶d,0兲 − s共␶ f , ␶d兲兴rsn其, 共A3兲

We have examined the effect of qubit relaxation on the
estimation of the qubit initial state 共excited or ground兲 using
a continuous-in-time noisy quantum nondemolition measure- P̄−1,␺ = exp关− s共␶ f ,0兲rsn兴, 共A4兲
ment for four different measurement protocols. In these pro-
tocols the measurement results are integrated with a box car and the signal s共␶ , ␶⬘兲 is defined in Eq. 共6.7兲. We first differ-
linear filter, an exponentially decaying filter, an optimal lin- entiate P̄+1,␺ with respect to ␶ f and by defining

冕
ear filter and a nonlinear Bayesian filter which by definition ␶f
is the optimal theoretical filter. We found that in all these
␭+1,␺ = e−␶ f exp关s共␶ f ,0兲rsn兴 − d ␶ de −␶d
protocols there exists a theoretical limit on the measurement 0
fidelity. The determining factor of this limit is the signal to
noise ratio of the measurement. Our results are summarized ⫻exp兵关s共␶d,0兲 − s共␶ f , ␶d兲兴rsn其, 共A5兲
in Table I where we see that the non-linear filter reaches the we get the following set of coupled stochastic differential
same fidelity as the linear filters even for substantially lower equations
required signal to noise ratio. Lastly we compare the con-
tinuous quantum nondemolition results with latching mea- d
surements and found that there is a quantitatively different P̄+1,␺ = rsn␺共␶兲␭+1,␺ , 共A6兲
d␶ f
but qualitatively similar limit on the fidelity of latching mea-
surements also due to relaxation. The signal to noise ratio
required to do successful qubit single shot quantum non- d
␭+1,␺ = rsn␺共␶兲P̄+1,␺ − 共␭+1,␺ + P̄+1,␺兲, 共A7兲
demolition measurements should be attainable in the near d␶ f
future, and there is no fundamental reason why significantly
higher fidelity measurements cannot be performed. and the initial conditions ␭+1,␺ = 1 and P̄+1,␺ = 1. If we use Eq.
共A4兲 the two additional coupled equations are
ACKNOWLEDGMENTS
d
P̄−1,␺ = rsn␺共␶兲␭−1,␺ , 共A8兲
We thank Michel Devoret, Isaac Chuang, Alexandre d␶ f
Blais, Jens Koch, Andrew Houck, and David Schuster for
discussions. This work was supported in part by NSA under
d
ARO Contract No. W911NF-05-1-0365, and the NSF under ␭−1,␺ = rsn␺共␶兲P̄−1,␺ , 共A9兲
Grants ITR-0325580, DMR-0342157, and DMR-0603369, d␶ f
and the W. M. Keck Foundation.
with initial conditions ␭−1,␺ = −1 and P̄−1,␺ = 1. Note that
since there is no relaxation, we do not need to have two
APPENDIX: NUMERICAL PROCEDURE USED
BY THE OPTIMAL NONLINEAR FILTER equations for the ground state, we can just use d␶ f P̄−1,␺
= −rsn␺共␶兲P̄−1,␺, but to keep the problem symmetrical we
In this appendix we present the method used to simulate
have decided to leave both equations in. This makes it easier
the optimal nonlinear filter. This filter requires simulating
to extend the theory to cases where upward jumps are pos-
Eqs. 共6.5兲 and 共6.6兲 which contains a double integral over a
sible.
stochastic process ␺共␶兲. This is not practical numerically and
Thus to simulate Eqs. 共6.5兲 and 共6.6兲 we simply solve the
a much better method can be implemented by deriving a set
above two sets of two coupled differential equations and then
of coupled stochastic differential equations. To do this we
combine them using Eqs. 共A1兲 and 共A2兲. Note an equivalent
start by rewriting Eqs. 共6.5兲 and 共6.6兲 as
derivation of these equations can be made by using the
Kushner-Stratonovich equation 关38兴 and then simply using
P̄+1,␺
P共兩i0 = 1兩⌿兲 = , 共A1兲 Bayes theorem to invert these equations for estimating un-
P̄−1,␺ + P̄+1,␺ known parameters 关39兴.

012325-10
PROTOCOLS FOR OPTIMAL READOUT OF QUBITS USING… PHYSICAL REVIEW A 76, 012325 共2007兲

关1兴 H. J. Carmichael, An Open Systems Approach to Quantum Op- 关22兴 A. Wallraff, D. I. Schuster, A. Blais, L. Frunzio, R. S. Huang,
tics 共Springer-Verlag, Berlin, 1993兲. J. Majer, S. Kumar, S. M. Girvin, and R. J. Schoelkopf, Nature
关2兴 Y. Makhlin, G. Schon, and A. Shnirman, Phys. Rev. Lett. 85, 共London兲 431, 162 共2004兲.
4578 共2000兲. 关23兴 A. Wallraff, D. I. Schuster, A. Blais, L. Frunzio, J. Majer, M.
关3兴 A. N. Korotkov, Phys. Rev. B 63, 115403 共2001兲. H. Devoret, S. M. Girvin, and R. J. Schoelkopf, Phys. Rev.
关4兴 Y. Makhlin, G. Schoen, and A. Shnirman, Rev. Mod. Phys. 73, Lett. 95, 060501 共2005兲.
357 共2001兲. 关24兴 D. DiVincenzo, G. Burkard, D. Loss, and E. Sukhorukov, in
关5兴 A. N. Korotkov and D. V. Averin, Phys. Rev. B 64, 165310 Quantum Mesoscopic Phenomena and Mesoscopic Devices in
共2001兲. Microelectronics, edited by I. Kulik and R. Ellialtioglu 共NATO
关6兴 H.-S. Goan, G. J. Milburn, H. M. Wiseman, and H. B. Sun, Advanced Study Institute, Turkey, 1999兲, e-print arXiv:cond-
Phys. Rev. B 63, 125326 共2001兲. mat/9911245.
关7兴 Asher Peres, Quantum Theory: Concepts and Methods 共Klu- 关25兴 H. A. Engel, V. N. Golovach, D. Loss, L. M. K. Vandersypen,
wer Academic Publishers, Dordrecht, Netherlands, 1993兲. J. M. Elzerman, R. Hanson, and L. P. Kouwenhoven, Phys.
关8兴 W. Nagourney, J. Sandberg, and H. Dehmelt, Phys. Rev. Lett. Rev. Lett. 93, 106804 共2004兲.
56, 2797 共1986兲. 关26兴 K. B. Cooper, M. Steffen, R. McDermott, R. W. Simmonds, S.
关9兴 T. Sauter, W. Neuhauser, R. Blatt, and P. E. Toschek, Phys. Oh, D. A. Hite, D. P. Pappas, and J. M. Martinis, Phys. Rev.
Rev. Lett. 57, 1696 共1986兲. Lett. 93, 180401 共2004兲.
关10兴 D. J. Wineland, C. Monroe, W. M. Itano, D. Leibfried, B. 关27兴 P. Bertet, I. Chiorescu, K. Semba, C. J. P. M. Harmans, and J.
King, and D. M. Meekhof, J. Res. Natl. Inst. Stand. Technol. E. Mooij, Phys. Rev. B 70, 100501共R兲 共2004兲.
103, 259 共1998兲. 关28兴 J. M. Martinis, S. Nam, J. Aumentado, and C. Urbina, Phys.
关11兴 D. I. Schuster, A. Wallraff, A. Blais, L. Frunzio, R. S. Huang, Rev. Lett. 89, 117901 共2002兲.
J. Majer, S. M. Girvin, and R. J. Schoelkopf, Phys. Rev. Lett. 关29兴 N. Katz, M. Ansmann, R. C. Bialczak, E. Lucero, R. McDer-
94, 123602 共2005兲. mott, M. Neeley, M. Steffen, E. M. Weig, A. N. Cleland, J. M.
关12兴 J. Gambetta, A. Blais, D. I. Schuster, A. Wallraff, L. Frunzio, Martinis, and A. N. Korotkov, Science 312, 1498 共2006兲.
J. Majer, M. H. Devoret, S. M. Girvin, and R. J. Schoelkopf, 关30兴 I. Siddiqi, R. Vijay, F. Pierre, C. M. Wilson, M. Metcalfe, C.
Phys. Rev. A 74, 042318 共2006兲. Rigetti, L. Frunzio, and M. H. Devoret, Phys. Rev. Lett. 93,
关13兴 M. Nielsen and I. Chuang, Quantum Computation and Quan- 207002 共2004兲.
tum Information 共Cambridge University Press, Cambridge, 关31兴 I. Siddiqi, R. Vijay, F. Pierre, C. M. Wilson, L. Frunzio, M.
UK, 2000兲. Metcalfe, C. Rigetti, R. J. Schoelkopf, M. H. Devoret, D. Vion,
关14兴 B. Kane, Nature 共London兲 393, 133 共1998兲. and D. Esteve, Phys. Rev. Lett. 94, 027005 共2005兲.
关15兴 C. van der Wal, A. ter Haar, F. Wilhelm, R. Schouten, C. 关32兴 I. Siddiqi, R. Vijay, M. Metcalfe, E. Boaknin, L. Frunzio, R. J.
Harmans, T. Orlando, S. Lloyd, and J. Mooij, Science, 290, Schoelkopf, and M. H. Devoret, Phys. Rev. B 73, 054510
773 共2000兲. 共2006兲.
关16兴 R. Vrijen, E. Yablonovitch, K. Wang, H. W. Jiang, A. Balandin, 关33兴 T. Duty, D. Gunnarsson, K. Bladh, and P. Delsing, Phys. Rev.
V. Roychowdhury, T. Mor, and D. DiVincenzo, Phys. Rev. A B 69, 140503共R兲 共2004兲.
62, 012306 共2000兲. 关34兴 K. W. Lehnert, B. A. Turek, K. Bladh, L. F. Spietz, D. Gun-
关17兴 D. Vion, A. Aassime, A. Cottet, P. Joyez, H. Pothier, C. Ur- narsson, P. Delsing, and R. J. Schoelkopf, Phys. Rev. Lett. 91,
bina, D. Esteve, and M. Devoret, Science 296, 886 共2002兲. 106801 共2003兲.
关18兴 I. Chiorescu, Y. Nakamura, C. Harmans, and J. Mooij, Science 关35兴 C. Gardiner, Handbook of Stochastic Methods: for Physics,
299, 1869 共2003兲. Chemistry and the Natural Science 共Springer, Berlin, 1985兲.
关19兴 O. Astafiev, Y. A. Pashkin, T. Yamamoto, Y. Nakamura, and J. 关36兴 T. Jennewein, G. Weihs, J. W. Pan, and A. Zeilinger, Phys.
S. Tsai, Phys. Rev. B 69, 180507共R兲 共2004兲. Rev. Lett. 88, 017903 共2001兲.
关20兴 C. Langer, R. Ozeri, J. D. Jost, J. Chiaverini, B. DeMarco, A. 关37兴 W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flan-
Ben-Kish, R. B. Blakestad, J. Britton, D. B. Hume, W. M. nery, Numerical Recipes in C⫹⫹ 共Cambridge University
Itano, D. Leibfried, R. Reichle, T. Rosenband, T. Schaetz, P. O. Press, Cambridge, UK, 2002兲.
Schmidt, and D. J. Wineland, Phys. Rev. Lett. 95, 060502 关38兴 T. P. McGarty, Stochastic Systems and State Estimation 共John
共2005兲. Wiley & Sons, New York, 1974兲.
关21兴 A. Blais, R.-S. Huang, A. Wallraff, S. M. Girvin, and R. J. 关39兴 J. Gambetta and H. M. Wiseman, Phys. Rev. A 64, 042105
Schoelkopf, Phys. Rev. A 69, 062320 共2004兲. 共2001兲.

012325-11