Phase Sensitivity of a Mach-Zehnder Interferometer
Luca Pezzé and Augusto Smerzi
arXiv:quant-ph/0511059v1 7 Nov 2005
Theoretical Division, Los Alamos National Laboratory,
Los Alamos, New Mexico 87545, USA
and
Istituto Nazionale per la Fisica della Materia BEC-CRS & Dipartimento di Fisica,
Università di Trento, I-38050 Povo, Italy
(Dated: February 1, 2008)
The best performance of a Mach-Zehnder interferometer is achieved with the input state |NT /2 +
1i|NT /2 − 1i + |NT /2 − 1i|NT /2 + 1i, being NT the total number of atoms/photons. This gives:
i) a phase-shift error confidence C68% = 2.67/NT with ii) a single interferometric measurement.
Different input quantum states can achieve the Heisenberg scaling ∼ 1/NT but with higher prefactors
and at the price of a statistical analysis of two or more independent measurements.
Introduction. The central goal of interferometry is to
estimate phase shifts with the highest sensitivity given
a finite resource of atoms/photons. There are several
possible interferometric configurations, which are generally reducible to the “drosophila” Mach-Zehnder (MZ),
see Fig.(1). The MZ transforms a two-mode input state
ˆ
as |ψiout = e−iJy θ |ψiinp [1], where the generator of the
phase translation θ is the y angular momentum component of the three-dimensional rotation group Jˆx =
(↠b̂ + âb̂† )/2, Jˆy = (↠b̂ − âb̂† )/2i, Jˆz = (↠â − b̂† b̂)/2
(â, b̂ are bosonic annihilation operators for the two input quantum modes). How precisely can the unknown
phase shift θ be estimated from the output |ψiout ? It
is well known that, if one of the two input arms of the
interferometer is fed with vacuum, the phase measurement uncertainty is bounded by √the standard quantum
limit (or “shot noise”) ∆θ ≥ 1/ NT , where NT is the
total/average number of atoms in the input state. The
first breakthrough came in the early ’80, when Caves
showed that it is possible to reach sub shot-noise sensitivities after squeezing the vacuum of the unused port
of the interferometer [2]. Sub shot-noise was experimentally demonstrated few years later [3]. This inaugurated
a large body of literature proposing new states and optimal performances [1, 4, 5, 6, 7, 8] to push the sensitivity
up to the Heisenberg limit ∼ 1/NT [9]. Paradoxically,
however, fewer efforts have been devoted to the correct
inference of phase measurement uncertainties. This fact
went (and it is still going) almost unnoticed, despite of
the seminal works of Helstrom and Holevo [10] and Lane,
Braunstein and Caves [11]. Apart from a few noticeable
exceptions, in the current literature the phase measurement uncertainty is still retrieved from a simplified error propagation theory, ∆θ = (∆Â)/ ∂h∂θÂi , where hÂi
is the average of a generic phase dependent observable
Â(θ) and ∆ = hÂ2 i − hÂi2 is the mean square fluctuation. Unfortunately, this simple approach can give correct estimates only when the probability distributions are
Gaussian [12], which is seldom the case in quantum interferometry. Moreover, optimized phase estimations often
require the distribution of the available finite resource of
NT particles among p independent interferometric experiments, each with N = NT /p particles. This optimization
cannot be analyzed with the linear error propagation formula, which only considers the dependence on the num-
ber of particles in each experiment N and not on the
total number of particles NT . As a matter of fact, even
if the Mach-Zehnder has been extensively studied, there
is not published literature about the exact dependence
of phase uncertainties on the total (or mean) number of
particles of any input state.
Matter waves Mach-Zehnder interferometers have been
realized with cold atoms [13] and electrons [14]. Evidences of non-classical light properties after a beam splitter have been experimentally showed in [15]. Entangled
“NOON” states have been created with few photons [16]
and ions [17] with applications in quantum lithography
[18]. Current efforts are devoted to the creation of the MZ
with dilute Bose-Einstein condensates (BEC). BEC offers
almost unique possibilities to create the large intensity,
highly squeezed waves needed for Heisenberg interferometry. Recently, double-slit [19] as well as Michelson interferometers [20] have been experimentally demonstrated.
By trapping BEC in two wells and in periodic potentials,
the beam splitters and mirrors are replaced by the dynamical tailoring of interwell barriers and magnetic wells.
Fock and number squeezed condensates have been realized in [21].
In the following, we study a MZ interferometer fed by
1
|ψiinp = √ |j + mia |j − mib + |j − mia |j + mib , (1)
2
where N = 2j and 2m are the total and the relative number of particles at the a and b input ports, respectively,
cfr. Fig. (1). We rigorously calculate error confidences
within a Bayes framework [10]. We first consider the
twin-Fock (m = 0) introduced by Holland and Burnett
[5]. We then study a “NOON” state (m = j), showing
that it cannot overcome shot-noise sensitivity. We will
finally demonstrate that the best Mach-Zehnder performance is obtained when m = 1. This state can be created
by constructing a twin-Fock state followed by the measurement of one particle [22].
Bayes analysis. Quantum Mechanics provides the conditional probability P (µ|j, θ) to measure, at the output
ports, the relative number of atoms µ = (N1 − N2 )/2,
given 2j = N1 + N2 particles and the unknown phase
shift θ. With the general state Eq.(1) we have
P (µ|j, θ) = |hj, µ|ψiout |2 =
1 j
|d (θ) + djµ,−m (θ)|2 , (2)
2 µ,m
2
phase
shift
beam
splitter 1
111111
000000
000000
111111
000000
111111
beam
splitter 2
b
D2
mirror
FIG. 1: Schematic representation of the Mach-Zehnder interferometer. Atoms/photons enter the a and b input ports, mix
and recombine in the beam splitters and are finally detected
in D1 and D2. The phase shift is inferred from the number
of atoms/photons measured in each output port.
where |j, µi ≡ |j + µiD1 |j − µiD2 , being D1 and D2 the
detectors at the output ports (see Fig. (1)), and djµ,ν (θ)
being the angular momentum matrices [23]. The goal is
to estimate θ after detecting a certain value µ. According to the Bayes theorem, P (θ|j, µ)P (µ) = P (µ|j, θ)P (θ),
where P (θ) and P (µ) take into account any further a priori information about the real value of the phase shift and
the output measurement µ. We assume complete ignorance, so P (θ) = H π/2 − |θ| , being H(x) the Heaviside
step function, while P (µ) is fixed by the normalization.
We notice that, with the state Eq.(1) as input of the
MZ, it is possible to estimate only the absolute value of
phase shifts in the interval [0, π/2]. In most cases, the
best interferometric performance requires the statistical
analysis of several independent measurements. Once obtained the results µ1 ...µp , theQphase probability distribup
tion reads Pp (φ|j, µ1 ...µp ) ∝ k=1 P (φ|j, µk ). Averaging
over all the possible p-uple µ1 ...µp we find the phase distribution for fixed θ and j:
Pp (φ|j, θ) =
X
{µ1 ...µp }
Pp (φ|j, µ1 ...µp )Pp (µ1 ...µp |j, θ), (3)
Qp
where Pp (µ1 ...µp |j, θ) = k=1 P (µk |j, θ) is provided by
quantum mechanics. As phase estimator, we choose the
value of the phase φ̄p (j, θ) corresponding to the maximum of the probability distribution Eq.(3). The phase
uncertainty is estimated as confidence, namely the γprobability that the real value of the phase shift is within
R φ̄+c
a given interval cγ around in φ̄: γ = φ̄−cγγ dφ Pp (φ|j, θ).
In the following we consider the case of a null phase
shift θ = 0. The probability distribution Eq.(2) reduces
to P (µ|j, 0) = |δµ,m + δµ,−m |2 /2, where δα,β is the Kronecker delta. Therefore, we have two possible results
for the measurement of the relative number of particles:
µ = +m, and µ = −m. Because of the symmetry properties of the MZ interferometer and of the state Eq.(1), we
have that P (φ|j, m) = P (φ|j, −m), and the sum Eq.(3)
reduces to Pp (φ|j, 0) ≈ |djm,m (φ) + djm,−m (φ)|2p . When
θ 6= 0 the probability of obtaining values of µ 6= ±m increases. In this case we might need several measurements
in order to explore all the relevant probability distribu-
A
amplitude
11111
00000
00000
11111
4
2
0
6
probability
D1
probability
6
a
C
0.4
0
6
B
4
2
0
−π/2
0.8
−0.4
probability
mirror
−π/4
0
φ
π/4
π/2
4
D
2
0
−π/2
−π/4
0
φ
π/4
π/2
FIG. 2: A,B: Probability distribution of the m = 0 twin-Fock
state with A) N = 40 particles and a p = 1 measurement, (A);
N = 20 particles and p = 2 measurements, (B). By combining the independent measurements it is possible to strongly
reduce the weight of the tail of the distribution with respect
to the central peak. C,D: Amplitude dj1,1 (φ) (red line), and
dj1,−1 (φ) (blue line), for N = 40 particles (j = 20) [23], (C).
The tails of the amplitudes oscillate out of phase and interfere
destructively, so to enhance the central peak of the probability
distribution P (φ|j, 0) ∼ |dj1,1 (φ) + dj1,−1 (φ)|2 , (D).
tion, so the sensitivity of the interferometer for a given
number of particles is, in general, reduced [24].
Mach-Zehnder interferometer with twin-Fock States.
We first consider the input state |ψiinp = |jia |jib (m = 0
in Eq.(1)).
The probability of measuring a relative particle number µ at the output port, given a phase shift θ, is [25]:
µ
2
µ
P (µ|j, θ) = (j−µ)!
(j+µ)! Pj (cos θ) , where Pj (x) are the Associated Legendre Polynomials. When θ = 0 the probability P (µ|j, θ) reduces to a single peak centered in µ = 0.
Therefore, repeating the experiment p times we have
Pp (φ|j, 0) ≈ [Pj0 (cos φ)]2p . With a single measurement
1/3
(p = 1), the 68%-confidence scales as ∼ 1/NT while,
as discussed in [26], the mean square phase fluctuation
scales as σ ∼ 1/ ln NT . In both cases the phase uncertainty is worse than shot noise [27]. This problem is
generally overlooked when the sensitivity is calculated
from the simplified error propagation formula, or looking
at the width of the central peak of the Legendre polynomials, which indeed shrink as 1/NT . The large tails
of the distribution contain most of the probability and
cannot be ignored, see Fig.(2,A). This clarifies the need
of considering several independent measurements, which
increases the relative weight of the central peak with respect to the tails, see Fig. (2,B). As discussed in [26],
the mean square phase fluctuation is minimized combining p = 4 measurements.
In Fig. (3,A) we study different values of confidence
for the twin-Fock state distribution as a function of the
number p of independent measurements and a fixed total number of atoms NT . We see that the bigger the
confidence, the larger is the number of experiments, p,
necessary to reach the minimum. It is important to no-
3
1
1
A
−1
−1
10
−2
−2
10
10
−3
−3
10
∆θ
confidence interval
10
−4
10
1
B
−1
10
−4
10
1
10
−2
−2
10
−3
−3
10
10
−4
1
2
3
4
5 6 7 8 9 10
tice that the number of measurements needed to reach
the minimum for a given confidence does not depend on
the total number of atoms NT . The main results of our
analysis can be summarized as follows: i) with a single
measurement the MZ interferometer gives a sensitivity
worse than the shot noise, ii) the highest sensitivity is
reached with two independent measurements, for the 68%
confidence, and iii) with three measurements for the 95%
confidence:
p = 2;
∆θ95% =
6.654
,
NT
100
1000
10000
NT
FIG. 3: Confidence interval as a function of the number of
independent measurements p for A) the twin-Fock m = 0 and
B) the twin m = 1 states. Here NT = 2000 particles. In both
plots, from the bottom line to the top one, the black line is
the confidence at 38.29%, the blue at 68.27%; the green at
95.45%, the red at 99.73%, the yellow at 99.994%, and the
brown at 99.99994%. These correspond to σ/2, σ, 2σ, 3σ, 4σ
and 5σ when the distribution is Gaussian.
2.915
,
NT
B
−4
10
20
number of measurements
∆θ68% =
A
−1
10
10
10
p = 3.
(4)
The number of measurements needed to reach higher confidences is shown in Fig.(3,A). In Fig.(4,A) we summarize the scaling of the phase uncertainties reached with
the twin-Fock state, as a function of NT .
How does the sensitivity scale with the number of measurements, p, given a fixed number of atoms, N , for each
experiment? The Cramer-Rao theorem does not allow a
√
better scaling than 1/ p, and we have verified numerically that the limit is reached when p > 4, obtaining the
√
mean square fluctuation ∆θ ≈ 2 p/NT . The 1/p scaling
claimed in [25] was probably an artifact of low statistics.
Mach-Zehnder interferometer with NOON state. We
now consider the state (1) with m = j. If θ = 0
the only possible outcomes are µ = ±j. According
to the Bayes theorem, we obtain the probability distribution Pp (φ|j, 0) ≈ [cos2j (φ/2) + sin2j (φ/2)]2p . For
large j we have that cos2j (φ/2) ∼ exp[−(j/4)φ2 ], while
sin2j (φ/2) ∼ exp[−(j/4)(φ − π)2 ]. Therefore, the phase
probability distribution
is a Gaussian with width scal√ √
ing as σ = 2/ NT (see Fig.(4,B)). The NOON state
with a Mach-Zehnder interferometer, assuming a complete “a priori” ignorance of the real phase shift, achieves
a shot noise sensitivity. A previous analysis [30], di-
FIG. 4: Phase uncertainty ∆θ as a function of NT . A) TwinFock state: (from the top line to the bottom one) black line, σ
for p = 1 (σ ≈ 1/ ln NT ); brown line, 68% confidence for p = 1
1/3
1/2
(∆θ = 0.8/NT ); yellow line, σ for p = 2 (σ ≈ 1/NT ); violet
line, σ for p = 4 (σ = 4/NT ); green line: 68% confidence for
p = 2 (∆θ = 2.91/NT p
). B) Phase sensitivity of the NOON
state (blue line), σ = 2/NT , independent of p, and of the
twin m = 1 state (red line), ∆θ = 2.67/NT at p = 1. Points
are numerical data.
rectly addressing current experiments [16, 17], showed
that NOON states and balanced homodyne detection can
3/4
provide phase sensitivity scaling as ∼ 1/NT .
Mach-Zehnder interferometer with a twin m = 1 state.
We have previously shown how the large tails of the
twin-Fock probability prevent the sensitivity to scale as
∼ 1/NT . To kill these tails, and enhance the central
peak of the distribution, we have to statistically analyze several independent measurements. Here we follow a different strategy: is it possible to cancel out the
tails by destructive interference? The answer is positive. The tails of the probability distribution P (φ|j, 0) ∼
|djm,m (φ) + djm,−m (φ)|2 obtained with the input state (1)
cancel out when m is odd. This special feature, not
present in the twin-Fock state, is a consequence of the
interference between the two matrices in P (φ|j, 0), and
it can be easily recognized by expanding the Jacobi Polynomials [31]. Since the cancellation has a precision of the
order of O[(j − m)−3 ] we can conclude that, among this
class of states, the case m = 1 offers the highest performance. States with larger odd values of m do not
benefit of an equally efficient destructive interference.
With semi-integer values of m, as for the Yurke state
(m = 1/2) [4], there is no cancellation at all, so we expect that a 1/NT scaling can be reached only with p > 1
as in the twin-Fock case. In Fig.(2,C) we plot the amplitude dj1,1 (φ) (red line), and dj1,−1 (φ) (blue line), for
N = 40 particles (j = 20). Outside the central region the
two functions oscillate out of phase and, when summed
to calculate P (φ|j, 0), interfere destructively. The phase
probability distribution is shown in Fig.(2,D). The comparison with the Fock state distribution (Fig. (2,A)), for
the same total number of particles, highlights the strong
4
reduction of the tails. As a consequence, for a given confidence, we reach the minimum with a smaller number
of measurements with respect to the Fock state case, see
Fig. (3,B). The sensitivity of the twin-m states is:
∆θ68% =
2.67
,
NT
p = 1;
∆θ95% =
5.376
,
NT
p = 1.
(5)
The scaling of ∆θ68% as a function of NT is shown in
Fig.(4,B). The main advantage is, apart from the smaller
prefactors, that the 1/NT scaling is reached with a single (p = 1) measurement for both the 68% and the 95%
confidence, as compared with the p = 2 and p = 3 measurements requested with Fock states, respectively. For
higher confidences, the required number of measurements
is shown in Fig.(3,B) [32, 33].
[1] B. Yurke, S.L. McCall, and J.R. Klauder, Phys. Rev. A,
33, 4033 (1986).
[2] C.M. Caves, Phys. Rev. D 23, 1693 (1981).
[3] M. Xiao, L.A. Wu, and H. J. Kimble, Phys. Rev. Lett.
59, 278 (1987); P. Grangier, R.E. Slusher, B. Yurke, and
A. LaPorta, Phys. Rev. Lett. 59, 2153 (1987).
[4] B. Yurke, Phys. Rev. Lett. 56, 1515 (1986).
[5] M. Holland and K. Burnett, Phys. Rev. Lett. 71, 1355
(1993).
[6] M. Hillery, L. Mlodinow, Phys. Rev. A,48, 1548 (1993).
[7] J.J. Bollinger, W.M. Itano, D.J. Wineland, and D.J.
Heinzen, Phys. Rev. A 54, R4649 (1996).
[8] G.M. D’Ariano, M.G.A. Paris and M.F. Sacchi, Phys.
Rev. A62, 023815 (2000).
[9] For a review see, for instance, A. Luis and L.L. SanchezSoto, in Progress in Optics, Elsevier, New York, Vol. 41,
421 (2000).
[10] C.W. Helstrom, Quantum Detection and Estimation
Theory Academic Press, New York, 1976; A.S.
Holevo, Probabilistic Aspects of Quantum Theory, NorthHolland, Amsterdam, 1982.
[11] A.S. Lane, S.L. Braunstein, C.M. Caves, Phys. Rev. A
47, 1667 (1992).
[12] With non-Gaussian probability distributions, ∆hÂi cannot be approximated by ∆ in a error propagation approach. Indeed, according to the central limit theorem,
the variance of the mean, (∆hÂi), becomes equal to the
√
average of the variance, (∆Â)/ p only asymptotically
with the number of independent measurements p.
[13] P. R. Berman, Atom Interferometry, New York, 1997.
[14] Yang Ji, et al., Nature 422, 415 (2003).
[15] S. Feng and O. Pfister Phys. Rev. Lett. 92, 203601 (2004).
[16] M.W. Mitchell, et. al., Nature 429, 161 (2004); P.
Walther, et. al., Nature 429, 158 (2004).
[17] D. Leibfried, et. al., Science 304, 1476 (2004); V. Meyer,
M.A. Rowe, D. Kielpinski, C.A. Sackett, W.M. Itano, C.
Monroe, and D.J. Wineland, Phys. Rev. Lett. 86, 5870
(2000).
[18] A. N. Boto, et al., Phys. Rev. Lett. 85, 2733 (2000).
[19] Y. Shin, et al., Phys. Rev. Lett. 95,170402 (2004);
M. Albiez, et al., M. Albiez, R. Gati, J. Fölling, S.
Hunsmann, M. Cristiani and Markus K. Oberthaler,
Conclusions. There are not known protocols for the
direct measurement of phase shifts. Phases can only be
inferred after the measurement of a different (phase dependent) observable. What is the highest achievable sensitivity? Within
√ a rigorous Bayesian analysis we find that
|ψiinp = 1/ 2(|NT /2+1i|NT /2−1i+|NT /2−1i|NT /2+
1i), gives the highest sensitivity C68% = 2.67/NT , with
a single interferometric measurement. Different input
quantum states can achieve the Heisenberg scaling ∼
1/NT but with higher prefactors and at the price of a
statistical analysis of two or more independent MachZehnder measurements.
Acknowledgment. We thank O. Pfister and B.C.
Sanders for useful discussions.
Phys.Rev.Lett, 95, 010402 (2005).
[20] Y. Wang, et al., Phys. Rev. Lett. 94, 090405 (2005).
[21] A.K. Tuchman et al., cond-mat/0504762; M.Greiner, et
al., Nature 415, 39 (2002); C. Orzel et al. Science 291,
2386 (2001).
[22] Y. Castin and J. Dalibard, Phys. Rev. A 55, 4330 (1997).
[23] L.C. Biedenharn, J.D. Louck, Angular Momentum
in Quantum Physics:
Theory and Applications,
Cambridge University Press (1984). The angular
ˆ
momentum matrices are djµ,ν (θ) ≡ hj, µ|e−iθJy |j, νi =
q
ν−µ
ν+µ ν−µ,ν+µ
(j+ν)!(j−ν)!
) sin 2θ
cos θ2
Pj−ν
cos θ ,
(j+µ)!(j−µ)!
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
where Pnα,β (x) are Jacobi Polynomials.
A θ-indipendent sensitivity can be achieved with positive
operator valued measurements, B.C. Sanders and G.J.
Milburn, Phys. Rev. Lett. 75, 2944 (1995).
T. Kim, O. Pfister, M.J. Holland, J. Noh, and J.L. Hall,
Phys. Rev. A, 57, 4004 (1998).
Z. Hradil, J. Rehaecek Physics Letters A 334, 267 (2005).
We emphasize that when θ = 0, the Bayesian analysis of
the measurement of the parity operator [28] or of the Jˆz2
operator [29] give the same results as the measurement
of the relative number of particles.
R.A. Campos, C.C. Gerry, and A. Benmoussa, Phys. Rev.
A 68, 023810 (2003)
P. Bouyer, M.A. Kasevich, Phys. Rev. A, 56, R1083
(1997).
L. Pezzé and A. Smerzi, quant-ph-0508158 (2005).
For integer values of m 6= j, and φ > 0, P (φ|j, m) ≈
2
cos2 [(j+ 1
)φ− π
]
2
4
1 + (−1)m + O[(j − m)−3 ].
(j−m) sin φ
The possibility to estimate a phase with a single measurement has some important advantages. Indeed, when
the external perturbation has a short life-time, the need
to perform several independent measurements might require the building of an equal number of interferometers.
We have extended our analysis to consider classes of
states different from Eq.(1). We have always found that
the twin m = 1 state provides the best performance,
and, on the ground of some general arguments, we believe that this remains true with any possible two-mode
input quantum state.