Maximum-Likelihood Multiple Quantum Measurements: Analysis

VOLUME
12 OCTOBER 1992
69
Maximum-Likelihood
NUMBER
Analysis of Multiple Quantum
Samuel L. Braunstein,
Center for Laser Studies, University
15
Phase Measurements
Alistair S. Lane, and Carlton M. Cavestbi

of Southern California, Los Angeles, California 90089 1112(Received 1 June 1992)
&~i
Shapiro, Shepard, and Wong [Phys. Rev. Lett. 62, 2377 (1989)] suggested that a scheme of
multiple phase measurements,
using quantum states with minimum "reciprocal peak likelihood, "
could achieve a phase sensitivity scaling as 1/Nwhere N&oq is the mean number of photons
available for all measurements. We have simulated their scheme for as many as 240 measurements
and havefoundoptimumphasesensitivities
for 3 & Nt t & 120, A power-law fit to thesimulated data
yields a phase sensitivity that scales as 1/N, ', + '
We conclude that reciprocal peak likelihood is
not a good measure of sensitivity.
PACS numbers:
03.65.Bz, 42. 50.Dv, 89.70.+c
Precise determination of small changes in the phase of

light has been an important tool for physics and measurement science for over a hundred years [1] and continues in
that role today [2]. Even with this long history, we have
begun only recently to understand the ultimate bounds
on how efficiently we can measure phase.
For interferometers using coherent states of light, the
Poisson counting statistics produced by the independent
arrivals of photons at the detector yields an optimum
phase sensitivity AC' = 1/2/N&, where Nt is the
mean total number of photons available during the time
the phase to be determined is stable. The 1/2v'N&, & behavior, called the "shot-noise" limit, is not fundamental.
It has been shown experimentally [3] that interferometry with squeezed states can beat the shot-noise limit, as
was predicted theoretically [4]. The optimum sensitivity
for squeezed-state interferometry is believed [5) to scale
as AC
1/N&, with a multiplicative constant of order
unity.
The theoretical basis for handling phase measurements quantum mechanically has generated controversy.
Although there have been many approaches to the
phase [610], a common feature
quantum-mechanical
has been the introduction of a probability distribution
P($~4) for the phase P, which is derived from the quantum state [@c,). Here C, the phase shift to be determined,
can be thought of as a parameter that gives the "location" of ~@g, ) on a phase-space diagram or, more generally, along a curve in Hilbert space. The goal of phase
'.
measurements is to determine C as precisely as possible;

the AC quoted above for interferometric measurements
are precise statements of the statistical confidence in an
experimentally determined value of C.
In order to address fundamental limits on phase measurements, Shapiro, Shepard, and Wong (SSW) [11] (see
also Ref. [12]) assumed and analyzed ideal phase measurements, whose statistics are given by P(Q~C'). The
question of how well we can determine C then reduces to
finding the optimal shape of P(Q~O), given a constraint
on the number of photons. Using their familiarity with
ordinary parameter estimation, SSW sought the quantum state whose phase distribution has the highest peak.
To ensure normalizability, they introduced a cutoff in the
number-state basis and so found a family of states
M
IVc)ssw
= ~)
~inC
. n+1 ~n),
with normalization constant A = ~6/vr + O(M

phase distribution for the SSW state (1) is
M
l)=2 )
(y @)
). The
(2)
These states have a mean photon number

(n)
= 1
determined
6+[p
l+n(M
1+)] /vr
+ O(M ),
solely by the cutoff M, where p
1992 The American Physical Society
(3)
0.57721 is
2153
VOLUME
69, NUMBER 15
PH
YSICAL
R EV I E%'
Euler's constant.
The parameter C gives the location of the peak of the
SSW phase distribution (2). Figure 1 shows examples of
the SSW distribution for small values of (n). For (n) & 2
the distribution takes on a universal shape in the domain
IP
CI 1/M:
A2
Pssw(QI@)
27r
ln(1
)I
(4)
For IP 4I
1/M the photon-number cutoff softens the
logarithmic singularity into a Gaussian peak with height
Pssw(CIC)
= ((n) +1)
/2vrA
For (n) & 2 the picture is that of a universal shape,

consisting of broad tails and a runup to a central peak,
whose height
(n)2 and width
1/M are the only features that depend on (n).
By analogy with the eKciency of maxirnurn-likelihood
(ML) estimation to locate the edge of a rectangular distribution, SSW introduced the "reciprocal peak likelihood, "
LETTERS
12 OC'TOBER 1992
To perform
determine the peak's location accurately.
N measurements,
however, requires N copies of the
SSW state and thus a mean total number of photons
Nt
(n)N. We interpret SSW's proposal as a suggestion that the phase sensitivity of multiple SSW measurements scales as AC
1/Nt2 t.
We investigate this suggestion by performing Monte
Carlo simulations of multiple SSW phase measurements.
"
We assume that N field modes, which we call "pulses,
are prepared in the same SSW state and are shifted in
phase by the same amount C. An ideal phase measurement is performed on each. The measurements yield X
results Pi, . . . , P~ (our N "samples" ), from which we
must estimate the actual phase shift C. Because SSW's
work is based on ML estimation and because ML estimation is optimal asymptotically in N, we use the ML
estimator for C,
CMLE
= arg
. . . , Q~IC),
max
~(C &~ l:(Pi,
where
):
N
bP
sup
7l (C (Vr
2~A2
P(PI 4 )
ssw
(6)
as their measure of phase sensitivity and concluded that

the sensitivity of ideal phase measurements using their
states scales as 1/(mean photon number) 2. A single sampling of Pssw($]4), however, yields almost no information about 4&, because of the broad tails. In contrast,
for mean photon number (n)
1, the phase probability distribution of a coherent state or a squeezed state
is very nearly a Gaussian, with central peak at P = 4
and standard deviation 1/2(n)iiz for a coherent state or
1/(n) for an optimized squeezed state. A Gaussian has
no broad tails, almost all of the probability being concentrated within a few standard deviations of the central
peak; a single sampling provides an estimate of C, which
is good to within a few standard deviations.
SSW argued that many samplings of Pssw(QIO) could
~(41
NivIC'):
Pssw(&i IC')
(8)
This entire process preparing and

phase shifting the pulses, collecting the phase data, and
an idealization
estimating C is a single "experiment"
of a multiple-measurement
laboratory experiment. The
ultimate result of an experiment is a single number, the
ML estimate C MLE.
An experimenter wants to know how much confidence
should be placed in the estimate @Mr,E. Thus we seek to
determine the probability distribution P(4MLE]4) of the
ML estimator, given an actual phase shift C. This "estimator distribution" contains all statistical information
about the estimator. To summarize important information in a single number, we give the one-sided 68.26'%%uo
confidence interval (AC)ss% defined by saying that the
estimator has a 68.26%%uo probability to lie within an interval of width 2(AC)as% which brackets symmetrically
the actual phase shift C. Such a confidence interval is
unquestionably a good measure of phase sensitivity. We
use the one-sided 68.26'%%uo confidence interval because it
corresponds to 1 standard deviation of a Gaussian.
The asymptotic behavior of ML estimation is known
oo the estimator
from Fisher's theorem [13]: As N
distribution approaches a Gaussian with standard deviation
is the log-likelihood.
AC
z/2
Tt/2
FIG. 1. SSW distribution Pssw($]4), with peak at 4 = 0

and with mean photon numbers (n) = 0.56 (dotted line), 1.02
(dashed line), and 2.00 (solid line).
2154
= 1/QNF
Here F, the Fisher information [13], is determined by

the phase distribution. For the SSW distribution, F is a
function only of (n) (or M) and increases roughly linearly
with M. ML estimation is also asymptotically eKcient:
The Cramer-Rao lower bound [13] states that for a very
broad class of estimation techniques, AO & 1/QNF for
VOl UME
PH YSICAL REVIEW
69, NUMBER 15
10
20
LETTERS
12 OCTOBER 1992
50 100 200
0.5
10
20
50
100
0.5
0.2
0.2
0. 1
Q
0. 1
Q5
CI
o.os
0.02
0.01
0.02
FIG. 2. Log-log plot of 68%%uo confidence intervals (DC)es%

of pulses N for SSW states with
(n) = 2.00 (error bars represent 95'%%uo confidence in the plotted confidence intervals). The dashed line gives the (NF)
Cramer-Rao lower bound on the standard deviation DC of
any estimator, with the Fisher information F calculated for
(n) = 2.00.
(in radians)
vs number
all sample sizes
Monte Carlo simulation proceeds by doing a sufFicient

number of experiments for particular values of (n) and
N to map out the estimator distribution P(CMi, E~O) for
those values. Since the simulations are demanding of
computer time, we have only been able to do a sufficient
number of numerical experiments for representative values in the range (n) & 5, N & 240, and Nt, t & 480.
Figure 2 shows the convergence of the 68% confidence
interval with increasing numbers of pulses N for the SSW
state with (n) = 2.00. For N & 3 the ML estimator
distribution P(@Mi,E~0) is so wide that it is sensitive to
the finite 2vr range of phases. At the other extreme, for
N 50 the asymptotic I/y'NF behavior of Fisher's theorem takes over as the confidence intervals become small
enough to probe the 1/M width of the central peak. Between these two regions lies a region of "universal convergence, " where ML estimation is unable to "see" the
narrow central peak. In this region the confidence intervals are determined by the universal shape of the SSW
distribution, and hence the convergence of the confidence
intervals must follow some universal form independent of
(n). Unfortunately, we have been unable to discover the
scale-invariant shape that characterizes this preasymptotic universal convergence. A power-law Gt to our data
for (n) = 3, 4, and 5, with 3 & N & 40 all of which data
are comfortably within the preasymptotic region yields
a 1/Ni is scaling for the preasymptotic 68% confidence
intervals.
The universal convergence in the preasymptotic region
is easier to see in Fig. 3, where we plot 68% confidence intervals for selected values of Nt- t. Points with the same
Nt t, are joined by straight line segments.
By following the confidence intervals with Nt t fixed, we can see
the effect of changing the split between the number of
FIG. 3. Log-log plot of 68'%%uo confidence intervals (AO)6s%

(in radians) vs number of pulses N for selected values of Nial, .
Error bars are suppressed, and points with the same Nt t are
joined by line segments. For each N&, t, there is an optimum
number of pulses which gives the best phase sensitivity.
pulses, N, and the mean number of photons per pulse,

(n). As we split up the photon-number resources into
pulses, the preasymptotic universal convergence gives initially an improvement in the 68% confidence intervals,
until they approach the Cramer-Rao lower bound. Then
the curves turn up into the asymptotic I/QNF behavior of Fisher's theorem, which here has an upward slope
because Nt, t is fixed. The optimal split occurs at the
"knee" between the preasymptotic universal convergence
and Fisher's asymptotic behavior.
Figure 4 plots the optimized 68'%%uo confidence intervals, obtained by using the optimal split of resources
from Fig. 3, against the mean total number of photons
available, i.e. , against the constrained quantity Nt t. A
straight-line fit yields
I/NQ. s5+0.01
(gC, )
for the optimized 68% confidence intervals over the range
0.5
10
20
50
100
0.2
0. 1
o.os
0.02-
FIG. 4. Log-log plot of optimized 68% confidence intervals

(EC)6s% (in radians; optimum number of pulses from Fig. 3)
vs mean total number of photons Ni, i (error bars represent
95% confidence in the plotted confidence intervals). The bestfit line corresponds to (AC)es% = 0.97N, ,
2155
VOLUME
69, NUMBER 15
PH
YSICAL
R EV I EW
Nt- t
120 that is accessible to our simulations.
3
Over this range the convergence rate of confidence intervals, obtained from ML estimation, is significantly slower
than the 1/N2 ~ scaling which SSW extracted from the
We conclude that reciproreciprocal peak likelihood.
cal peak likelihood is not a good measure of sensitivity.
Others [1416] have argued that reciprocal peak likelihood is not a good measure of phase sensitivity, but our
work, by finding the sensitivity of multiple SSW measurements, demonstrates the inadequacy of reciprocal peak
likelihood. Furthermore, over the range of our simulations, we find that multiple SSW measurements do not
have a phase sensitivity as good as the I/Nt, scaling of
squeezed-state interferometry.
The relevant scale in the SSW distribution is the
2
("lis. This width is excentral-peak width 1/M
e
ponentially small in (n) [14], but many pulses are required to locate the central peak with an accuracy approaching the width, because of the small probability
(n)z/M
(lnM) /M under the peak. Indeed, the
central-peak width is, crudely speaking, the phase sensitivity at the knee transition from preasymptotic universal
convergence to asymptotic Fisher behavior.
Optimal SSW measurements for a fixed Nt require
many pulses X with only a few mean photons per pulse,
(n). This behavior should be contrasted with that for
coherent states and squeezed states, for which the nearGaussian phase distribution means that the sensitivity
of multiple measurements using ML estimation improves
as I/~N. For coherent states, with phase standard desensitivity
viation I/2(n) i2, the multiple-measurement
~
=
of
how the
is
independent
I/2y'N&
I/2((n)N)
photon-number
resources are split between the number
of pulses and the number of photons per pulse. For optimized squeezed states, with standard deviation
I/(n),
the optimum sensitivity is attained by using a single pulse
that carries all the photons.
convergence rate
The large-Nq, t confidence-interval
cannot be obtained directly from our simulations. If the
form of the preasymptotic universal convergence were
known, then its approach to the I/QNF Cramer-Rao
lower bound for a particular (n) would give us the optimum number of pulses for that (n); extrapolations based
on various assumptions for the universal convergence will
be discussed elsewhere [17]. Alternatively, working backwards from a knowledge of where the asymptotic Gaussian approximation to the ML estimator breaks down [18]
behavior
can give a direct way to attack the large-Nt
[17, 19].
We do not know at present how to formulate the general question of quantum limits on phase sensitivity, because it becomes entangled in dificult issues of estima&
t-,
2156
LETTERS
12 OcToBFR 1992
tion theory for multiple measurements. Nonetheless, our

work shows that there are surprises lurking in multiplemeasurement schemes
we did not expect SSW states to
achieve even the shot-noise limit and it indicates that
any investigation of the ultimate phase sensitivity must
allow for a "divide-and-conquer" strategy that divides up
the available photons among many measurements of the
same phase datum.
This work was supported in part by the Once of Naval
Research (Contract No. N00014-88-K-0042 and Grant
No. N00014-91- J-1167). S.L.B. appreciated the support
of the Lady Davis Fellowship Trust in writing up this
work, and he thanks A. Mann for a critical reading of
the manuscript.
Present address: Department of Physics, Technion Israel

Institute of Technology, 32 000 Haifa, Israel.
Present address: Department of Physics and Astronomy, University of New Mexico, Albuquerque, NM 87131-
1156.
[1] A. A. Michelson and
E. W.
Morley, Am.
J. Sci. 84, 333
(1887) .
[2] See, for example, A. Abramovici
et al. , Science
256, 325
(1992).
L.-A. Wu, and H. J. Kimble, Phys. Rev. Lett.

59, 278 (1987).
C. M. Caves, Phys. Rev. D 23, 1693 (1981).
B. Yurke, S. L. McCall, and J. R. Klauder, Phys. Rev. A
88, 4033 (1986).
L. Susskind and J. Glogower, Physics (Long Island City,
N. Y.) 1, 49 (1964).
P. Carruthers and M. M. Nieto, Rev. Mod. Phys. 40, 411
(1968).
J.-M. Levy-Leblond, Ann. Phys. (N. Y.) 101, 319 (1976).
D. T. Pegg and S. M. Barnett, Phys. Rev. A 39, 1665
[3] M. Xiao,
[4]
[5
[6]
[7]
[8]
[9]
(1989).
[10 S. M. Barnett and D. T. Pegg, J. Mod. Opt. 86, 7 (1989).

[11 J. H. Shapiro, S. R. Shepard, and N. W. Wong, Phys.
Rev. Lett. 62, 2377 (1989) (cited in text as SSW).
[12] J. H. Shapiro and S. R. Shepard, Phys. Rev. A 43, 3795
(1991).
[13] H. Cramer, Mathematical Methods of Statistics (Princeton Univ. Press, Princeton, NJ, 1946), esp. pp. 500504.
[14] W. P. Schleich, J. P. Dowling, and R. J. Horowicz, Phys.
Rev. A 44, 3365 (1991).
[15 Z. Hradil and J. H. Shapiro, Quantum Opt. 4, 31 (1992).
[16 M. J. W. Hall (to be published).
[17 A. S. Lane, S. L. Braunstein, and C. M. Caves (to be
published) .
[18] S. L. Braunstein, J. Phys. A 25, 1 (1992).
[19] S. L. Braunstein (to be published).

Maximum-Likelihood Multiple Quantum Measurements: Analysis

Uploaded by

Copyright:

Available Formats

Maximum-Likelihood Multiple Quantum Measurements: Analysis

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Maximum-Likelihood Multiple Quantum Measurements: Analysis

Uploaded by

Copyright:

Available Formats

VOLUME

Analysis of Multiple Quantum

Alistair S. Lane, and Carlton M. Cavestbi

03.65.Bz, 42. 50.Dv, 89.70.+c

Precise determination of small changes in the phase of

measurements is to determine C as precisely as possible;

with normalization constant A = ~6/vr + O(M

These states have a mean photon number

solely by the cutoff M, where p

1992 The American Physical Society

For (n) & 2 the picture is that of a universal shape,

as their measure of phase sensitivity and concluded that

This entire process preparing and

FIG. 1. SSW distribution Pssw($]4), with peak at 4 = 0

Here F, the Fisher information [13], is determined by

FIG. 2. Log-log plot of 68%%uo confidence intervals (DC)es%

all sample sizes

Monte Carlo simulation proceeds by doing a sufFicient

FIG. 3. Log-log plot of 68'%%uo confidence intervals (AO)6s%

pulses, N, and the mean number of photons per pulse,

for the optimized 68% confidence intervals over the range

FIG. 4. Log-log plot of optimized 68% confidence intervals

tion theory for multiple measurements. Nonetheless, our

Present address: Department of Physics, Technion Israel

J. Sci. 84, 333

L.-A. Wu, and H. J. Kimble, Phys. Rev. Lett.

[10 S. M. Barnett and D. T. Pegg, J. Mod. Opt. 86, 7 (1989).

You might also like