Phase detection at the quantum limit with multi-photon Mach-Zehnder interferometry L. Pezz´e, A. Smerzi BEC-CNR-INFM and Dipartimento di Fisica, Universit`a di Trento, I-38050 Povo, Italy G. Khoury, J. F. Hodelin and D. Bouwmeester Department of Physics, University of California, Santa Barbara, California 93106, USA We study a Mach-Zehnder interferometer fed by a coherent state in one input port and vacuum in the other. We explore a Bayesian phase estimation strategy to demonstrate that it is possible to achieve the standard quantum limit independently from the true value of the phase shift and specificassumptionsonthenoiseoftheinterferometer. Wehavebeenabletoimplementtheprotocol 7 usingparalleloperationoftwophoton-number-resolvingdetectorsandmultiphotoncoincidencelogic 0 electronicsattheoutputportsofaweakly-illuminatedMach-Zehnderinterferometer. Thisprotocol 0 2 is unbiased and saturates the Cramer-Rao phase uncertainty bound and, therefore, is an optimal phaseestimation strategy. n a PACSnumbers: 42.50St,42.50-p J 6 2 TheMach-Zehnder(MZ)interferometer[1,2]isatruly ubiquitous device that has been implemented using pho- 2 tons, electrons [3], and atoms [4, 5]. Its applications v 8 range from micro- to macro-scales, including models of 5 aerodynamics structures, near-field scanning microscopy 1 [6]andthemeasurementofgravityaccelerations[7]. The 1 central goal of interferometry is to estimate phases with 0 thehighestpossibleconfidence[8,9,10]whiletakinginto 7 0 account sources of noise. Recent technological advances / make it possible to reduce or compensate the classical h FIG. 1: (a) Schematic of a Mach-Zehnder interferometer. A noisetothelevelwhereadifferentandirreduciblesource p phase sensitive measurement is provided by the detection of - of uncertainty becomes dominant: the quantum noise. t the number of particles Nc and Nd at the two output ports. n Givenafiniteenergyresource,quantumuncertaintyprin- (b)Pulseheightdistributionforavisiblelightphotoncounter a ciples and back reactions limit the ultimate precision of (VLPC) used in the experiment. The power incident on the u a phase measurement. In the standard configuration of detector is 144 fW at a wavelength of 780 nm. The vertical q the MZ interferometer, a coherent optical state with an lines show the decision thresholds [23]. : v average number of photons n¯ = α2 enters input port i | | X a and the vacuum enters input port b, as illustrated in Figure(1). Thegoalistoestimatethevalueofthephase r eventuallyvanishes. Asaconsequence,thisinterferomet- a shift θ after measuring a certain number of photons N c ricprotocoldoesnotallowthe measurementofarbitrary and N at output ports c and d, which, in the experi- d phaseshifts. Thiscanbe a seriousdrawbackforapplica- ment discussed in this Letter, is made possible by two tions likelasergyroscopes,the synchronizationofclocks, number-resolving photodetectors. orthealignmentofreferenceframes. Furthermore,toes- The conventional phase inference protocol estimates timatesmallphaseshiftswiththe highestresolution,the the true value of the phase shift θ as [11, 12, 13]: interferometer has to be actively stabilized around π/2. M This generally requires the addition of a feedback loop, p Θest =arccos(cid:16) n¯ (cid:17), (1) which can be quite costly in terms of time and energy resources. whereM = p (N(k) N(k))/pisthephotonnumber p Pk=1 c − d It was first noticed by Yurke, McCall and Klauder difference detected at the output ports, averaged over p (YMK) in [14] that the estimator (1) does not take into independent measurements. The phase uncertainty of account all the available information, and, in particular, estimator (1) is the fluctuations in the total number of photons at the output ports. The possibility to improve Eq.(2) is con- 1 ∆Θ= , (2) firmed by the analysis of the Cramer-Rao lower bound √pn¯sinθ (CRLB) [15, 16], which provides, given an input state which follows from a linear error propagation theory. andchoiceofobservables,the lowestuncertaintyallowed Eq.(2)predictsanoptimalworkingpointataphaseshift byQuantumMechanics. Forageneric,unbiased,estima- θ = π/2, where the average photon number difference tor, ∆Θ =1/ pF(θ), where F(θ) is the Fisher in- CRLB p varies most quickly with phase. As θ approaches 0 or π, formation[17],which,ingeneral,candependonthetrue theconfidenceofthemeasurementbecomesverylowand valueofthephaseshiftθ andthenumberofindependent 2 0.2 0.4 1 probability distribution 000... 1320 AD 0000000 ....0...3245132 CB (i)(i)φ P ( | {N , N } )pci=1...pd 0000 ....045232468 AC 010 1 . ..088242 BD 0.1 0.1 1 4 0 0 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 φ / π φ / π φ / π φ / π FIG. 2: (color online). Phase distribution P(φ|Nc,Nd), for: FIG.3: (color online). Phaseprobabilitydistribution Eq.(6), A) Nc =0, Nd =0; B) Nc =0, Nd =1; C) Nc =1, Nd =1; obtained after p independent experimental measurements Dco)lleNccte=din0,thNedca=lib2r.atTiohnepcairrctleosftahreeetxhpeereixmpeenritm,tehnetadladshaetda {Nc(i),Nd(i)}i=1..p: A) p = 1, B) p = 10, C) p = 100, D) p = 1000. The true value of the phase shift is θ/π = 0.24, line is the ideal phase distribution Eq.(4). The solid line is a shown bythe vertical dashed line. fit of thedata according toEq.(5). measurements p. A direct calculation of the Fisher in- vides both anestimate onthe phase andthe uncertainty formation for the coherent vacuum input state, gives in this estimate. ⊗ F(θ)=n¯. Therefore, the Cramer-Rao lower bound is There are several advantages to using a Bayesianpro- tocol. Notably, it can be applied to any number p of 1 independent measurements, it does not require statisti- ∆Θ = , (3) CRLB √pn¯ calconvergenceoraveraging,anditprovidesuncertainty estimates tailored to the specific measurement results. which, in contrast with the result of Eq.(2), is indepen- For instance, with a single measurement, p = 1, it pre- dent of the true value of the phase shift. The only as- dicts anuncertaintythatscalesas ∆Θ 1 . Since sumptionhereisthattheobservablemeasuredattheout- ≈ √Nc+Nd Eq.(4)doesnotdependonn¯,theestimationisinsensitive putportsisthenumberofparticles. Itiswellknown(see to fluctuations of the input laser intensity. Most impor- for instance [18]) that the Maximum Likelihood (ML) tantly,itsuncertainty,inthelimitp>>1,is∆Θ= 1 estimator, defined as the maximum, Θ , of the Like- √pn¯ ML which coincides with the CRLB. lihood function P(N ,N φ) (see below), saturates the c d CRLB, but only asympto|tically in the number of mea- To implement the proposed protocol we have real- surements p. In the current literature there have been ized a polarization Mach-Zehnder interferometer with alternative suggestions to obtain an unbiased estima- photon-number-resolving coincidence detection. In a re- tor and a phase independent sensitivity with a Mach- cent paper we reported on the analysis of a coherent Zehnder interferometer [10, 14, 19, 20, 21]. They will be state using a single photon-number-resolving detector specifically addressed at the end of this Letter. [23]. We have extended this experimental capability to Here we develop a protocol based on a Bayesian two simultaneously operating visible light photon coun- analysis of the measurement results [8, 9, 18, ters(VLPCs)[24],cryogenicphotodetectorsthatprovide 19]. The goal is to determine P(φ N ,N ), acurrentpulseofapproximately40,000electronsperde- c d the probability that the phase equals |φ given tected photon. The VLPCs were maintained at 8 K in a the measured N and N . Bayes’ theorem pro- helium flow cryostat, and their photocurrent was ampli- c d videsthis: P(φN ,N )=P(N ,N φ) P(φ)/P(N ,N ), fiedbylow-noise,roomtemperatureamplifiers. Wemea- c d c d c d where P(N ,N| φ) is the probabili|ty to detect N and suredadetectionefficiencyof35%andadarkcountrate c d c N when the ph|ase is φ [22], P(φ) quantifies our prior of 3 105 for each detector under our operating condi- d × knowledge about the true value of the phase shift, and tions. Customelectronicsprocessedthe amplifiedVLPC P(N ,N ) is fixed by normalization. Assuming no prior current pulses to perform gated, fast coincidence detec- c d knowledge of the phase shift, P(φ) = 1/π. In the ideal tion. Wewerethusabletodetermine,foreachpulse,how case, the Bayesian phase probability distribution can be many photons were detected at both ports c and d. A calculated analytically for any value of N and N , Ti:sapphire pulsed laser, attenuated such that n¯ = 1.08 c d photons,providedtheinputstate. Sinceacoherentstate φ 2Nc φ 2Nd maintains its form under linear loss [11], the presence of P(φN ,N )= cos sin , (4) | c d C(cid:16) 2(cid:17) (cid:16) 2(cid:17) loss after the interferometer is completely equivalent to a losslessinterferometerfed by a weakerinput state. We where = Γ(1/2+Nc)Γ(1/2+Nd) is a normalization con- use n¯ to signify the average number of photons in the C Γ(1+Nc+Nd) stant. In practice, one must measure P(N ,N φ) and, detected state per pulse, after all losses. We were lim- c d | fromthis,determineP(φN ,N ). Thisdistributionpro- ited by the amplifiers to measuring up to four photons c d | 3 0.03 10 0.02 8 Θ ) / πest 0.0 01 ∆Θ 6 − p θ −0.01 4 ( −0.02 2 −0.03 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 θ / π θ / π FIG. 4: (color online). Difference between the mean value FIG. 5: (color online). Phase sensitivity as a function of of the phase estimator, Θ , obtained after 150 replica of est the true value of the phase shift. Circles are obtained from h i p=1000independentmeasurementsandthetruevalueofthe Bayesian distributions Eq.(6), with p=1000. Theerror bars phase shift θ. The vertical bars are the mean squarefluctua- give the fluctuations of √p∆Θ obtained with 150 indepen- tions σe2st =h(Θest−hΘesti)2i. The result σest ≫(θ−Θest) dent replica of the experiment. The solid black line is the proves that our protocol provides an unbiased experimental theoretical prediction Eq.(3). The dashed blue line is the phase estimation. CRLB calculated with the experimental distributions. Tri- angles are the uncertainty obtained with a generalization of theestimatorEq.(1)takingintoaccounttheexperimentalim- perfections, while the dotted red line is the phase sensitivity per pulse Fig. 1(b), but at n¯ = 1.08 the probability of predicted by Eq.(2). detecting five or more photons is negligible. The phase shiftθwaschangedbytiltingabirefringentcrystalinside the interferometer. distributions associated with each experimental result: The first part of the experiment consists of the cali- brationoftheinterferometer. Atdifferent,known,values p P φ N(i),N(i) P φN(i),N(i) . (6) of the phase shift, we measured Nc and Nd for each of (cid:0) |{ c d }i=1...p(cid:1)∝Y fit(cid:0) | c d (cid:1) 200,000 laser pulses. This procedure allows us to deter- i=1 mineexperimentallybothP(N ,N φ)andP(φN ,N ). c d c d The phase estimator Θ is given by the mean value of | | est In Fig.(2) we compare the ideal and the experimental the distribution, Θ = πdφφP φ N(i),N(i) , phase distributions. The agreement is quite good, and est R0 (cid:0) |{ 1 2 }i=1...p(cid:1) and the phase uncertainty ∆Θ is the 68.27% con- the discrepanciescanbe attributed to imperfect photon- fidence interval around Θ . An example of est number discrimination by the detectors. To fit the data (i) (i) P φ N ,N isgiveninFig.(3),forθ/π =0.24 inFig.(2),weintroducetheprobabilityP(Nc,Nd|Nc′,Nd′) an(cid:0)d|f{ordcifferednt}vi=a1lu..e.ps(cid:1)ofp. Sincetheaveragenumberof tomeasureN andN whenN andN photonswerere- c d c′ d′ photonsofthecoherentinputstateissmall,forp 1the ally present. Taking this into account, the experimental ∼ phase uncertainty is of the order of the prior knowledge, phase probability distribution is ∆Θ π. As p increases,the probability distribution be- ≃ comesGaussianandthesensitivityscalesas∆Θ 1/√p, ∝ Pfit(φ|Nc,Nd)= X P(φ|Nc′,Nd′)P(Nc′,Nd′|Nc,Nd), in agreement with the central limit theorem. N′,N′ In Fig.(4), we show the difference between the mean c d (5) value of the phase estimator Θ , obtained from 150 est h i whereP(φ|Nc′,Nd′)aretheidealprobabilitiesEq.(4). The phaseestimations,eachwithp=1000,andthetruevalue weights P(Nc′,Nd′|Nc,Nd) can be retrieved from a fit of ofthephaseshift. Thebarsarethemeansquarefluctua- the experimental calibration distributions P(φNc,Nd), tion. The important result is that our protocol provides | seeFig.(2). ThequantityP(Nc,Nd Nc,Nd),equaltoone anexperimentallyunbiasedphaseestimationovertheen- | in the ideal case, is 0.54 in Fig.(2A) (corresponding to tire phase interval. the worstcaseamongalldistributions), 0.67in (2B) and The main result of this Letter is presented in Fig.(5). (2C), and 0.87 in (2D). We show the phase sensitivity for different values of the After the calibration, we can proceed with the phase shift θ, calculated from the distribution Eq.(6) Bayesian phase estimation experiment. For a certain with p = 1000 photon-number measurements. The value of the phase shift, we input one laser pulse and circles are the mean value of ∆Θ, and the bars give detect the number of photons N and N . We repeat the corresponding mean square fluctuation, obtained c d this procedure p times obtaining a sequence of indepen- from 150 independent phase measurements. The dashed dent results N(i),N(i) . The p photon-number blue line is the CRLB calculated with the experimen- { c d }i=1...p tal probability distributions, ∆θ =1/ pF (θ), where measurements comprise a single phase estimation. The fit p fit overall phase probability is given by the product of the Ffit(θ) = PN1,N2 Pfit(N11,N2θ)(∂Pfit(N∂θ1,N2|θ))2 [26]. For | 4 0.1 . θ/π . 0.9, it follows the theoretical prediction put ports of a weakly-illuminated interferometer. Yet, (solid black line), Eq.(3), where n¯ = 1.08 has been in- the method can be generalizedto the case of high inten- dependently calculated from the collected data. Around sity laser interferometry and photodiode detectors. In θ = 0,π, where the photons have higher probability to this case, the limit Eq.(3) becomes harder to achieve be- exit through the same output port, ∆θ increases as a causeoflargerelectronicnoiseandlowerphotonnumber fit consequence of the decreased sensitivity of our detectors resolution, however it should still be possible to demon- to higher photon number states. Even though the phase strateaphaseindependentsensitivity. Ourresultsareof sensitivityofourapparatusbecomesworsenearθ =0,π, importance to quantum inference theory and show that it never diverges. Triangles show the phase uncertainty the MZ interferometer does not require phase-locking in obtainedwithEq.(1), but takinginto accountthe exper- order to reach an optimal sensitivity. imental noise. The estimator is obtained by inverting Acknowledgment. This work has been partially sup- the equation M = n¯cos(a+θ)+b [25]. This strategy ported by the US DoE and NSF Grant No. 0304678. p provides an unbiased estimation with a sensitivity close to the one predicted by Eq.(2) (dotted red line). The reasonfor the superiorperformance ofthe Bayesianpro- tocol can be understood by noticing that, in Eq.(2), the [1] L. Mach, Zeitschr. f. Instrkde. 12, 89 (1892). phase estimate is retrieved only from the measurement [2] L. Zehnder, Zeitschr. f. Instrkde. 11, 275 (1891). of the photon number difference, which, simply does not [3] Y. Ji, et al.,Nature 422, 415 (2003). exploit all of the available information. [4] AtomInterferometry.P.R.Bermanedt.,AcademicPress, In [14] YMK first proposed a generalization of the New York,1997. estimator Eq.(1) to take into account the whole infor- [5] Y. Wang et al.,Phys. Rev. Lett. 94, 090405 (2005). mation in the output measurements. Their estimator, [6] F. Zenhausern, Y.Martin & H.K.Wickramasinghe, Sci- ence 269, 1083 (1995). Θ =arccos[(N N )/(N +N )], gives a phase in- YMK c− d c d [7] A.Peters,K.Y.Chung&S.Chu,Nature400,849(1999). dependent sensitivity, ∆Θ = 1/√Nc+Nd. Notice that [8] M.J. Holland & K. Burnett. Phys. Rev. Lett. 71, 1355 Θ coincides with the Maximum Likelihood estima- YMK (1993). tor in the ideal, noiseless, MZ interferometer. However, [9] L. Pezz´e & A. Smerzi, Phys. Rev. A 73, 011801(R) it is not obvious how to generalize Θ for real inter- (2006). YMK ferometry, where classical noise is present and the YMK [10] B.C. Sanders& G.J. Milburn, Phys. Rev. Lett. 75, 2944 estimator is different from Θ . In general, because of (1995). ML [11] M.O. Scully & M.S. Zubairy, Quantum Optics, Cam- correlations between N N and N +N , this estima- tor becomes strongly bcia−seddin the pcresendce of noise as [12] Jb.rPid.gDeoUwnliinvge,rsPithyysP.rResesv1.9A975.7, 4736 (1998). we haveverifiedusing our experimentaldata [27]. More- [13] C.M. Caves, Phys. Rev. D 23, 1693 (1981). over,the YMK estimatorcannotbe extendedwhenboth [14] B.Yurke,S.L.McCall &J.R.Klauder,Phys. Rev. A33, input ports of the interferometer are illuminated. Con- 4033 (1986). versely, the Bayesian analysis holds for general inputs [15] H. Cramer. Mathematical methods of statistics, Prince- and, in particular, it predicts a phase-independent sen- ton, Princeton university press, 1946. [16] C.R. Rao, Bull. Calcutta Math. Soc. 37, 81 (1945). sitivity when squeezed vacuum is injected in the unused [17] R.A. Fisher, Proc. Camb. Phi. Soc. 22, 700 (1925). portoftheMZ[27],whichreachesasubshot-noisesensi- [18] C.W. Helstrom, Quantum Detection and Estimation tivity[13]. Itshouldbenotedthatdetectionlossesagain Theory Academic Press, New York,1976. become important when attempting to use nonclassical [19] Z. Hradil, et al.,Phys. Rev. Lett. 76, 4295-4298 (1996). light to overcome the shot-noise limit Eq.(3). [20] D.W.Berry&H.M.Wiseman, Phys. Rev. Lett.85,5098 (2000). In [19], Hradil et al. used a Bayesian approach for [21] J.W. Noh, A. Fougeres & L. Mandel, Phys. Rev. Lett. a Michelson-Morley neutron interferometer (single out- 67,1426(1991);J.Rehacek,etal.,Phys. Rev.A60,473 putdetection)anddiscussedtheoreticallythe MZ.Their (1991). analysis wasbased onspecific assumptions about the in- [22] The Quantum Mechanical probability to detect Nc and terferometricclassicalnoisewhicharenotsatisfiedinthe N particles at the output port c and d, respectively, d case discussed in this Letter. Different approaches with is given by P(Nc,Nd φ) = ( Nc Nd )e−iφJˆy ψinp 2, | | h |h | | i| adaptive measurements [20] and positive operator value whereJˆy =(aˆ†ˆb ˆb†aˆ)/2i, withaˆ,ˆbbosonicannihilation − measurements[10]havebeenalsosuggested. Whilethese operators. For the configuration studied in this Letter sHteraisteengbieesrgmliigmhitt,bteheiymparoertnaonttnfoecreisnstaerryfeirnomouertrcyasaet. the [23] GP.(NKch,oNudr|yφ,)e=t a|lα.(,N|2PcN)h!e(y−Ns|dα.)|!R2`evsi.nLφ2et´t2.N9c6`,co2s03φ26´021Nd(.2006). In conclusion, we have presented a Bayesian phase es- [24] G.B.Turner,etal.,ProceedingsoftheWorkshoponScin- timationprotocolforaMZinterferometerfedbyasingle tillating Fiber Detectors, Notre Dame University, edited by R. Ruchti(World Scientific, Singapore, 1994). coherent state. The protocol is unbiased and provides a phase sensitivity that saturates the ultimate Cramer- [25] AfitofthecalibrationdatagiveshNc−Ndi(θ)=n¯cos(θ+ a)+b, with a=0.024 and b=0.008. Rao uncertainty bound imposed by quantum fluctua- [26] ThedistributionsP (N ,N θ)areobtainedfromEq.(5) fit 1 2 tions. We have been able to implement the protocol with the Bayes theorem. | with two photon-number-resolving detectors at the out- [27] Manuscript in preparation.