On-LineEstimationofLONoiseandOptimisationofServoParametersinAtomicClocks ∗ IanD.Leroux, NilsScharnhorst,StephanHannig,JohannesKramer,LennartPelzer,MariiaStepanova,andPietO.Schmidt QUESTInstituteforExperimentalQuantumMetrology,Physikalisch-TechnischeBundesanstalt,38116Braunschweig,Germany (Dated:25thJanuary2017) Foratomicfrequencystandardsinwhichfluctuationsofthelocaloscillatorfrequencyarethedominant noisesource,weexaminetheroleofthetheservoalgorithmthatpredictsandcorrectsthesefrequencyfluc- tuations. Wederivetheoptimallinearpredictionalgorithm,aswellaspracticalproceduresforoptimising theparametersofaconventionalintegratingservoandcharacterisingtheflickerandrandom-walknoise ofthelocaloscillatorwithoutinterruptingthenormaloperationofthestandard. Inparticular, weshow thatnear-optimalservoparameterscanbecalculatedwithoutindependentknowledgeofthelocal-oscillator noise,basedonlyonmonitoringoftheerrorsignal.Usingsimpleanalyticalmodelsandnumericalsimula- tions,weestablishoptimumprobetimesasafunctionofclockatomnumberandofthedominantnoisetype inthelocaloscillator. Thisleadstoanargumentforthesuperiorityofmethodsthatincreasethemeasure- mentsignal-to-noiseratio(increasingatomnumber,spinsqueezing)overthosethatincreasetheeffective transitionfrequency(maximally-correlatedstates)whentryingtoimprovethestabilityoflocal-oscillator- 7 limitedfrequencystandards. 1 0 2 The instability of frequency standards limits the total termfractionalinstability[11] n uncertainty achievable in a measurement of finite dura- (cid:115) a tion[1,2]. Thislimitcanbepracticallyrelevantevenwhen 1 T J performingmeasurementsofstaticfrequencyratios,since σC(τ)=ωT(cid:112)N τc (2) 4 many-month-long measurement campaigns place strin- 2 gent demands on the reliability of all components in an where T is the duration of a single Ramsey interrogation experiment. Instability becomes a fundamental concern and T is the length of the frequency standard’s operating ] c h whenattemptingtomeasuretime-varyingfrequencyratios. cycle, such that τ/T measurements can be performed in c p For instance, in the emerging field of chronometric level- an averaging time τ. This quantum projection noise limit - ing[3–5],directobservationoftidalfluctuationsexpectedin (QPN)1forclocksusinguncorrelatedatomsdependsonthe m thegravitationalredshift[6]requiresfrequencyratiomeas- experimenter’schoiceofprobetimeT,becomingarbitrarily o urementswithafractionaluncertaintyatthelevelof10−18 smallforsufficientlylongprobetimes. Thus,Eqn.2setsno t a tobecompletedinamatterofhours. Physicsbeyondthe limitonachievableclockinstabilityatlongaveragingtimes s. StandardModelmightbedetectableinclockfrequencyra- unless some additional scale in the problem restricts the c tiomeasurementsaspostulatedtransientshiftsassociated choiceofT. si withdark-matterdomainwalls[7]orultralightscalardark- One such restriction is set by excited-state decay in the y mattercandidates[8,9].Searchesforsuchsignalswouldre- atoms, which sets a fundamental limit to interrogation h quireevenfastermeasurements,completedinseconds. At times. Theperformanceofopticalfrequencystandardsop- p suchtimescales,thestatisticaluncertaintyduetoinstabil- eratingatthislimithasbeenanalysedinRefs.[12–15].How- [ ityplaysafargreaterrolethanlong-termsystematicuncer- ever, formanyoftheopticalfrequencystandardsnowbe- 1 taintyincurrentstandards. ing investigated, frequency fluctuations of the local oscil- v Of the noise processes contributing to the instability of latorrestrictT tolessthanasecondevenwhentheatoms’ 7 atomicfrequencystandards, themostfundamentaloneis excited-statelifetimeismeasuredinminutes(87Sr)oreven 9 quantum projection noise [10], which arises from the dis- years (171Yb+) [1]. Because the local oscillator’s noise is 6 cretenessinthemeasurementresultsobtainablefromafi- common to all the atoms in the standard, and because 6 0 nitenumberofatoms. ForanensembleofN uncorrelated it typically exhibits significant power-law temporal correl- . two-levelatoms, thisnoiseimposesaminimumstatistical ations, its effects are qualitatively different from those of 1 uncertainty excited-state decay. In fact, it might at first glance seem 0 7 oddthatlocal-oscillatornoiselimitsclockstabilityatall:the 1 1 ∆φ =(cid:112) (1) localoscillatorfrequencyisinsomesensethemeasurandin QPN : N anatomicfrequencystandardanditsfluctuationsarecon- v stantlymonitoredandcorrected. Local-oscillatornoiseaf- i X on any measurement of the phase accumulated in an fectsthestabilityofthestandardonlytotheextentthatit r atomic superposition state. For a standard operating at a cannotbecorrectedbyfeedbackfromtheatoms. Thiscan a frequencyωandintheidealcaseofdead-time-freeRamsey happen,forinstance,ifthecyclicatomicinterrogationpro- interrogationwithouttechnicalnoise, thisleadstoalong- tocol allows undetected aliased frequency components of ∗[email protected] 1sometimesreferredtoastheStandardQuantumLimit(SQL) 2 thelocal-oscillatornoisespectrumtocontaminatetheout- x x−h LO Output put signal of the standard, a phenomenon known as the Dick effect [16]. Even in the absence of the Dick effect, h however,thequantisedmeasurementsignalfromtheatoms hasafundamentallylimiteddynamicrange:onecannotex- ω tract more than log2(N+1) bits of frequency information Servo e=y−h≈x−h ? from a single measurement of N atoms [17]. The useful Reference domain of the measurement, i.e. the frequency band in whichitcanbeunambiguouslyinterpreted,mustbebroad enoughto cover the frequencies whichthe local oscillator Figure1. Generalmodelofaperiodically-stabilisedatomicclock. islikelytoemitintheinterrogation. Frequencyexcursions Alocaloscillator(LO)emitsasignalwithafractionaldeviationx fromthenominalclockfrequency. Theservocontrollerattempts beyond the domain for which the reference provides use- topredict x andcorrectsthefrequencybyitspredictionh. The fulinformationleadtolessinformativemeasurementres- correctedsignal(withfractionalfrequencydeviationx−h)isthen ults,andhencetodegradedinstability.Intheworstcasethe usedtointerrogateanatomicreference,whichproducesanestim- servo,workingfromambiguousoruninformativemeasure- atee ofthepredictionerroror,equivalently,anestimatey ofthe mentresults,maybeunabletokeeptheoutputfrequency LO’sunknownfrequencydeviationx.Theseestimatescanbeused lockedtotheatomicreference. Theoutputfrequencythen bytheservoinfuturepredictions. eitherhopsbetweendifferentzero-crossingsofafrequency- periodicRamseyerrorsignalordriftsaimlesslyfarfromthe resonanceofaRabierrorsignal. Thiscaseiscatastrophic tionVIconsidersthemeritsofusingentangledatomicstates and the operating parameters must be chosen to make it to modify the phase resolution of Eqn. 1, giving a simple vanishinglyunlikely. Thus,evenintheabsenceoftheDick dimensional argument for the disappointing performance effect(e.g. withdead-time-freeRamseyinterrogation[18]), ofmaximally-correlatedstatesinatomicclocksandarguing theachievablemeasurementresolutionultimatelydepends forthesuperiorityofstatesthatenhancethedynamicrange onthescaleoflocal-oscillatorfrequencyfluctuationsseen ofatomicmeasurements[17,24–26]. Thisresultiscomple- bytheatoms,andhenceontheperformanceoftheclock’s mentarytothatofRef.[12],whichconsideredindependent feedbackloopwhichcorrectsthesefluctuations. dephasingoftheatomsratherthanthecollectivedephas- ingassociatedwiththeLO,andtakesintoaccounttemporal In this work, we study the limits to the stability of fre- correlations intheLOnoise rather thanassumingawhite quencystandardsdominatedbylocal-oscillatornoisewith spectrum as in Ref. [20, 27]. Section VII considers the in- realistic temporal correlations. We focus on clocks using stabilityoftheclockatshorttimes,whichmaybelimitedby asingleensembleofatomsperiodicallyinterrogatedusing finitefeedbackgainratherthanmeasurementnoise,show- thesameprotocolforeveryinterrogationcycle, whosein- ingthatthesecondintegratorrecommendedinRef.[15]to stabilitywequantifyusingtheAllanvarianceatlongtimes. correctforlineardriftoftheLOisalsonecessarytosatur- Our work is thus less general, but more directly relevant atetheQPNlimitinthepresenceofrandom-walknoise.We to current experiments, than analyses of multi-ensemble concludewithsomeremarksonproposedfrequencystand- clocks or of interrogation protocols which are modified ardswhosedesigndoesnotfollowtheconventionalpattern on-the-fly [19–22], and our approach is a more concrete weconsiderinthiswork. complement to the derivation of universal performance bounds in mathematically idealised settings [22, 23]. Us- ingsimpleanalyticalargumentsandnumericalsimulations of clocks with different local-oscillator noise spectra, we I. SETUPANDNOTATION study the performance of the servo controller which pre- dicts and corrects local-oscillator noise and then analyse Figure1sketchesthestructureofthefrequencystandards its implications. After establishing notation and conven- we consider and summarises the notation we will use for tionsinSec.I,webeginbyderivingtheoptimallinearpre- thevarioussignalswemustconsider.Thegoalofthestand- dictionalgorithmandevaluatingitsperformanceinSec.II. ard is to produce a continuous classical oscillatory signal In Sec. III we show that a feedback controller with near- whosefrequencycorrespondstoareferencetransitionfre- optimalperformancecanbedesignedwithoutpriorknow- quencyωinaparticularatomicspecies.Astheclassicalsig- ledgeofthenoisespectrum,bymonitoringtheerrorsignals nalisgeneratedbyamacroscopiclocaloscillator(LO)sub- innormalclockoperation.Wealsoshowthatthesametech- ject to environmental perturbations, its frequency ω will L niques provide useful diagnostic information on the local differfromthetargetfrequencyωbyafluctuatingfractional oscillator’s noise, allowing on-line monitoring of its per- discrepancyx=(ω /ω)−1. Thescaleofthesefluctuations L formance. We turn to the effects of the noise in Sec. IV, issummarisedintheAllandeviationσ (τ)oftheLO.Inor- L in which we derive a modification to the QPN that takes dertosuppressthesefluctuations,aservocontrollergener- intoaccounttheperformanceoftheservocontroller. This atesapredictionhoftheLOfrequencyerror,whichisused modified QPN formula predicts an overall limit to achiev- tofrequency-shifttheLOoutputsignalbacktoatomicres- ableclockinstability,whichisattainedforanoptimalchoice onance. Theresultingsignal,withnetfractionalfrequency ofatomicinterrogationtimethatwediscussinSec.V.Sec- errorx−h,isprovidedtotheusersofthestandardandhas 3 anAllandeviationσ (τ). Thiscorrectedsignalissupplied power-lawnoise,suchthatσ2(τ)∝τµ,withµ=−1,0,1for C L toareference,whereitinteractswithN atomsaccordingto whitefrequencynoise,flickerfrequencynoiseandrandom somefixedinterrogationprotocol,suchasRamseyorRabi walkoffrequencynoise, respectively. Asarguedinthein- interrogation. Measurementoftheatoms’stateattheend troduction,theLOnoisegivestheproblemacharacteristic oftheinterrogationprotocolconveyssomeinformationon time scale which ultimately limits the useful resolution of theresidualfrequencyerror x−h,whichweexpressasan measurementsontheatoms.WedefinethistimeZ,without error estimate e. The error estimate might, for instance, assumingaparticularformofLOnoisespectrum,bytheim- correspondtotheimbalanceofatomicstatepopulationsat plicitequation theendofaRamseysequencedividedbytheaccumulated phase ωT. For consistency, we express the error estimate σL(Zc)ωZ=1rad, (3) inthesameunitsasx andh,sothat y =h+e isanestim- ateoftheLO’suncorrectedfrequencyerrorx,onethatuses whereZcisthecycletimeoftheclockwhenoperatedwith onlythemostrecentatomicdataandtakesnoaccountof a probe time Z. In other words, Z is the choice of probe previousmeasurements. Notethaty andedifferfrom,and timeforwhichtheLOAllandeviationatoneclockcycleis fluctuatemorethan,x andx−hrespectively,becausethey as large as the quantum projection noise of a single atom areaffectedbythenoiseoftheatomicreference. (Eqn. 2 with N = 1). This definition lets us combine the We consider periodically-stabilised frequency standards LOnoiseandthechoiceofprobetimeintoasingledimen- with an operating cycle of period T , where the reference sionless parameter T/Z which can be compared between c providesaseriesoferrorestimates{e }. Atanygivenpoint clocksofdifferenttypesusingLOswithdifferentperform- i intime,welabelthemasfollows:e isthemostrecentavail- ance.NotethatZ willbeontheorderofafewsecondsfora 1 ableerrorestimate,e theprecedingerrorestimate,andso typicalcurrentopticalfrequencystandardwithafractional 2 forth. e isthentheerrorestimatewhichwillbeproduced LOinstabilityaround10−16. 0 attheendofthecurrentoperatingcycle.Welabeltheother signalssimilarly:h istheservo’spredictionofthe(average) 0 LO frequency error x0 during the current operating cycle, II. SERVOCONTROLLERDESIGN while h and x correspond to the jth most recent com- j j pletedcycle. Causalityrequiresthattheservocomputethe Wenowfocusourattentionontheservo.Givenahistory predictionh withoutknowledgeofe ,usingonly{e ,e ,...} 0 0 1 2 ...,h ,h ,h of its own past predictions and of the corres- 3 2 1 or,equivalently,{y ,y ,...}. 1 2 pondingerrorsignals...,e ,e ,e obtainedfromtheatomic 3 2 1 The Allan deviation σC(τ) is that of the physical signal reference,itmustmakeapredictionh0oftheLOfrequency producedbythefrequencystandardasitisoperating.With inthenextoperatingcycle.Thepredictionshouldtakeinto the exception of Sec. VII, most of the analysis presented account the temporal correlations of the LO noise, which in this work also applies to “paper clocks”, i.e. virtual sig- dictatethetimescaleoverwhichpastmeasurementresults nalsgeneratedbypost-processingmeasurementdata. Al- remainrelevanttopredictingfutureLObehaviour. thoughthepost-processingneednotrespectcausalityand Webeginbyconsideringthesimpleintegrator,thebasic canuselatermeasurementstocorrectestimatesofthefre- buildingblockoftheservoalgorithmusedinmostcontem- quency at earlier times, the quality of the measurements poraryopticalfrequencystandards[15].Inournotation,the themselves still depends on the ability of the (causality- simpleintegratormakestheprediction respecting)servotokeepthecorrectedLOfrequencynear the atomic resonance frequency ω while the clock is run- h =h +ge (4) 0 1 1 ning,andconstraintsonthisabilityaffecttheperformance of the reference no matter how the resulting data is sub- where g is a dimensionless gain specifying the fraction of sequentlyused. thefrequencyerrormeasuredinthelastcycletoapplyasa Whereitisnecessarytoassumeadefiniteinterrogation correctiontothelastprediction.Thepredictioncanalsobe protocol in the atomic reference, we will focus on dead- expressed in terms of past estimates of the LO’s fractional time-freeRamseyinterrogation,wherethemeasuredsignal frequencydeviation{y }asfollows: k depends on the average of the corrected signal frequency ∞ duringsomeinterrogationtimeT. Whileweassumeinour h = (cid:88)g(1−g)k−1y . (5) examplesthatT,whichsetsthefrequencyresolutionofthe 0 k k=1 interrogation,isequaltoT ,whichsetstherepetitionrateof c theinterrogationcycle,thetwotimesareconceptuallydis- While Eqn. 4 is easier to implement, Eqn. 5 is easier to tinct and we will use separate symbols for them through- reasonaboutbecausethestatisticalpropertiesoftheestim- out. Referenceswhoseoperatingcycleincludesdeadtime atedLOfrequencyyaremostlydeterminedbytheLOnoise (T < Tc) or which use a different interrogation protocol andbythemeasurementnoiseofthereference,depending (suchasRabiorhyper-Ramsey[28,29])willsufferfromthe onlyweaklyonthedesignoftheservocontrolleritself.Toa Dickeffect,whichcanbemodelledasadditionalmeasure- goodfirstapproximation,then,wecantakethefluctuations mentnoiseintheatomicreference. andcorrelationsofthe{y }asgiven,andtrytochooseg so k InnumericalexampleswewillconsiderLOswithsimple astominimisetheerrorofthepredictionh . 0 4 It is instructive to study the broader class of linear pre- Althoughlinearpredictorswitharbitrarycoefficientsare dictors,whosepredictionsareweightedaveragesofpastLO not difficult to implement following Eqn. 6, one can also frequencyestimatesoftheform use the preceding formalism to optimise the gain of con- ventionalintegrators. AppendixBderivesanexplicit,albeit h0=(cid:88)wkyk, (6) cumbersome,formulafortheoptimalintegratorgaingiven k knownnoisemodelparameters. Alternatively,onecanuse wheretheweightsw arerequiredtosatisfythenormalisa- Eqn.11tochooseavectorofweightsforahypotheticallin- k tioncondition earpredictorandthensimplysettheintegratorgaintothe leadingentryofthisvectorg =w . Simulationsshowthat (cid:88)wk=1. (7) for common power-law noise pro1cesses, the resulting in- k tegratingservoperformsalmostaswellastheoptimallin- earpredictor,withapenaltyoflessthan10%inthepredic- ThesimpleintegratorofEqn.5isaspecialcaseofalinear predictor,withw =g(1−g)k−1. Theoptimisationofsuch tionvariance. Theformalismwehavedevelopedcanthus k beusedtooptimisetheparametersofaconventionalinteg- linearpredictorshasbeenstudiedextensivelysincethepi- ratingservoalgorithm,withoutrequiringanymodifications oneeringworkofWiener[30]andKolmogorov[31](seee.g. toanalready-runningclockexperiment.Aswewillsee,this Ref. [32]). Here we derive the minimum-mean-squared- optimisation can be performed even without prior know- error predictor in a form similar to that used for ordinary ledgeoftheexperiment’snoisecharacteristics. krigingingeostatistics(seee.g. [33]). Webeginbycomput- We quantify the scale of theservo’s predictionerrors by ingthemeansquareddifferencebetweenthepredictionh 0 thedimensionlessvariance andthenextfrequencyestimatey : 0 (cid:42) (cid:43) v=(cid:173)((x−h)ωT)2(cid:174) (12) 〈(h −y )2〉= (cid:88)(cid:88)w w (y −y )(y −y ) (8) 0 0 j k j 0 k 0 j k of the phase accumulated in the Ramsey interrogation of =w(cid:214)Cw, (9) durationT. Thevariancev playsanimportantroleinde- terminingboththerobustnessofthelocktoatomicreson- where we have collected the weights {w } into a vector w anceandtheachievablelong-termstability(seeSec.IV),so k and introduced the two-sample covariance matrix for the that it is worth studying its behaviour. The solid lines in estimatedLOfrequency,whoseentriesaredefinedas Fig.2illustratetheperformanceoflinearpredictors,based onsimulationsofclockoperationwithbetween1and104 Cjk=〈(yj−y0)(yk−y0)〉. (10) atomsinthereference.Asafunctionofthechoiceofprobe timeT/Z,andthusoftheratiobetweenLOnoiseandmeas- Notethat〈(h −y )2〉isnotthesameasthemeansquared 0 0 urementnoise, onecandistinguishthreequalitativelydif- prediction error 〈(h −x )2〉, since it also includes the 0 0 ferentregimes.Inthelimitoflargeatomnumbersandlong noise of the atomic reference which estimates that error. probetimes, thesimulationsapproachan N-independent Provided, however, that the atomic reference is unbiased, limit thesamechoiceofweightswillminimiseeithermeasureof noise, soweproceedtominimiseEqn.9andfindthatthe (cid:181)T(cid:182)2+µ v=ξ . (13) optimumweightssatisfy Z 1 This is the scaling one expects for the case where the LO 1 noisecompletelydominatesthemeasurementnoiseofthe Cw=λ1 (11) reference,giventhepostulatedpower-lawscalingoftheLO 1 noisewithexponentµ(c.f. Sec.I)andthedefinitionof Z .. in Eqn. 3. The proportionality constant ξ can be estim- . ated from the simulations or derived using the formulae withλaLagrangemultiplierthatmustbechosentosatisfy in Appendix A, and varies between 1 (for a white-noise- dominatedLO)and2(forarandom-walk-noise-dominated the normalisation constraint of Eqn. 7. Thus the optimal weightscanbefoundbysolvingEqn.11forw/λandnor- LO). In the opposite limit of low atom number or short probe time, the (white)quantumprojectionnoisedomin- malisingtheresult,providedthatoneknowsthecovariance atesandtheservoperformancedependsonlyonthenum- matrixC. Ifthenoisepropertiesofthecomponentsinthe berofmeasurementsnwhichitaveragesinmakingitspre- frequency standard are known, then C can be computed diction,andthus simplyasthesumofmatricesforeachindependentnoise process. Explicit expressions for the C matrix associated 1 withaknownnoisespectrumareprovidedinAppendixA. v≈ (14) Nn As discussed in Sec. III,C can also be estimated, and the > servocontrolleroptimised,without priorknowledgeofthe for sufficiently short probe times. For n ∼ 20 this limit systemnoiseproperties, usingonlydatageneratedduring hasnoimpactoncorrectlyoptimisedclockoperation(see normalclockoperation. Sec.V).Betweenthetwolimitsconsideredabovethereisa 5 So far, we have discussed the simple integrator and its (a)White(µ=−1) generalisations. Practical frequency standards, however, 1100−−11 mustuseadouble-integratortocorrectforsteadydriftsin theLOfrequency[15]. Withtheadditionofthesecondin- 1100−−22 tegrator,theservopredictionsbecome 1100−−33 h0=h1+ge1+g2(cid:88)ek (16) k 1100−−44 or (b)Flicker(µ=0) =(cid:88)wkyk+g2(cid:88)ek. (17) 1100−−11 k k Asidefromitsroleinsuppressingsteady-statefrequencyer- 1100−−22 rors with drifting LOs, the second integrator is necessary for the servo to have enough low-frequency gain to attain v 1100−−33 theprojectionnoiselimitinmany-atomclocks, apointto which we will return in Sec. VII. However, as long as its 1100−−44 gain g2 is chosen low enough to avoid servo oscillations, the additional integrator has only a negligible impact on the variance of the prediction errors2. The controller can (c)Random-Walk(µ=+1) thusbedesignedby optimisingasimple integrator or lin- 10−1 earpredictorasdiscussedabove,andthenaddingthedrift- correctionintegratorwithagaing (cid:191)w =g. 2 1 Besides minimising prediction variance, another desir- 10−2 ablefeatureinpracticalservocontrollersisrobustness,the quality of remaining locked to the (correct) atomic reson- 10−3 anceforlongperiods.Inprinciplethetwoqualitiesaredis- tinct,butwefindempiricallythatforwell-optimisedservos 10−4 theyaretightlycoupled.Insimulationsofclockswithawide range of atom numbers (i.e. reference signal-to-noise ra- tios),wefindthattherateatwhichaclockhopstodifferent 10−2 10−1 Ramseyfringesdepends,foragivenLOandafullyoptim- T/Z isedservo, onlyonthepredictionvariance v. Suboptimal servos (such as integrators with incorrectly chosen gain) Figure2. Varianceofservopredictionerrors, expressedasthe havebothgreaterpredictionvariancevandahigherrateof variancevofthephaseaccumulatedinaRamseyinterrogation,as fringehopsforagivenv,sothattheyarelessrobustaswell afunctionofprobetimeforawhite-noise(top)flicker-(middle)or asnoisier. Weconjecturethatthebestservosaresimultan- random-walk-limited(bottom)LO.ThesimulatedclocksuseRam- eouslythemostrobustandtheleastnoisy,sothatthereis seyinterrogationof1(solidblackcircles),10(openbluecircles), 102 (solidredtriangles), 103 (opengreentriangles)or104 (solid noneedtochoosebetweenthetwoqualitiesprovidedthat greysquares)uncorrelatedatoms.Symbolsshowtheperformance onecan,infact,findthisoptimalservodesign. ofintegratingcontrollersoptimisedwithoutknowledgeoftheLO noise as discussed in Sec. III. Solid lines show the performance oflinearpredictorsusingthelastn=50frequencyestimates,de- III. ON-LINESERVOOPTIMISATIONANDNOISE signedwithknowledgeoftheLOnoisepropertiesasdiscussedin CHARACTERISATION Sec.II. In practice, the noise spectrum of the LO may not be trade-offbetweenaveragingmanymeasurementstoreduce known accurately. A significant benefit of the formalism theimpactofmeasurementnoiseandconsideringonlythe presentedintheprevioussectionisthatitallowsonetoop- recentmeasurementsmostrelevanttotheLO’scurrentfre- timisetheservocontrollerwithoutpriorknowledgeofthe quency. Weknowofnosimple,accurateexpressionforthe achievableservoperformanceinthisintermediateregime, buttheroughscalingthatonewouldexpectfromtheafore- mentionedtrade-off 2 Thisisbestunderstoodbyconsideringtheactionoftheservointhe frequency domain: the variance of the prediction errors depends on 1 (cid:181)T(cid:182)1+µ/2 thefeedbackgainatafrequencycorrespondingtotheclockcyclerate, v∝(cid:112) (15) N Z whereasthesecondintegratoronlycontributesfeedbackgainatmuch lowerfrequencies.Theeffectofthesecondservoisvisibleinthecorrela- doesholdinsimulations. tionsbetweenpredictionerrors,notintheirvariance. 6 LOnoiseproperties. Thisispossiblebecausethedefinition 100 ofC inEqn.10involvesonlytheestimatedLOfrequencyer- T rorineachclockcycle,whichisroutinelyrecordedinnor- ω malclockoperation3.Asademonstrationofsuchoptimisa- T)c 10−1 tion,wehaverunclocksimulationswithintegratingservos (L σ whosegainswerechosen,withoutknowledgeofthetrueLO d e noise,bythefollowingempiricalprocedure: t uc 10−2 r 1. Start by setting the gain to g =0.2, an arbitrary but st n reasonable initial value chosen to allow reliable, if o c suboptimal, clock operation under a wide range of Re 10−3 conditions. 10−3 10−2 10−1 100 2. Simulatetheclockfor104 cycles,correspondingtoa TrueσL(Tc)ωT fewhoursofoperationforatypicalcontemporaryfre- quencystandard. Figure3. NormalisedAllandeviationofflicker(bluecircles)or random-walk(redtriangles)LOnoisereconstructedfromcorrela- 3. ComputeC accordingtoEqn.10andthencethevec- tionsintheerrorsignalofasingle-atomclockasafunctionofthe torw ofoptimalweights.Setg=w 4. actualnoiselevel. Blacklinemarkscorrectreconstruction. Light 1 and dark grey regions mark the noise levels at which the clock 4. Simulate the clock with the newly optimised servo servohopsbetweenRamseyfringesforrandomwalkandflicker andreoptimise,repeatingasnecessary. noiserespectively. Evenwhentheservogainisinitiallychosenblindly,wefind that five rounds of optimisation suffice for the gain g to generallyoptimistic. AsdiscussedinSec.IV,eveninacor- convergetoavaluethatyieldsperformanceindistinguish- rectly optimised clock there will be unavoidable ambigu- able from that of an integrator designed with full know- ities in interpreting the error signal (e.g. 2π phase slips ledgeoftheLOnoisespectrum. Undermorerealisticcon- in Ramsey interrogation) and the resulting measurement ditions,wheretheinitialchoiceofservoparametersreflects errors contribute to the servo prediction variance without somepriorknowledgeoftheLOperformance,theoptimisa- beingobservableintheexperimentallyrecordedmeasure- tioncouldbeperformedmuchmorequickly. Thesymbols mentdata. inFig.2showthepredictionvarianceofsuchempirically- One can, however, use the correlation matrix C estim- optimisedintegrators, whichcanbecomparedtotheper- ated during clock operation to partially characterise the formanceofoptimallinearpredictorsshownassolidlines. LO. Although white noise of the LO is indistinguishable For clocks operated near their optimal probe times (to be frommeasurementnoiseintheatomicreference,flickeror discussedinSec.IV),thedifferenceinpredictionvarianceis random-walk noise can produce detectable temporal cor- lessthan10%.Thus,itispossibletodevelopcontrollersthat relationsevenwhentheircontributionstothetotalmeas- takefulladvantageofthetimecorrelationsintheLOnoise urementvariancearesmall. ByfittingtheestimatedC toa evenwithoutindependentknowledgeofthosecorrelations. linearcombination Unfortunately,itisnotalwayspossibletoverifytheservo performancedirectly,becausetheobservedvarianceofthe C=σ2(T )C +σ2C +σ2(T )C (18) w c w f f r c r error signal e contains contributions both from the servo predictionerrorandfrommeasurementnoiseoftherefer- of the correlation matrices expected for white, flicker and ence. Forasingle-atomclockthisproblemisinsurmount- random-walknoise, onecanobtainestimatesoftheAllan able:theobservedfluctuationsofabinaryerrorsignalmust variancesassociatedwitheachclassofnoiseprocess,which correspond to quantum projection noise, independent of aresimplythecoefficientsinthelinearcombination.Expli- servoperformance. Foramany-atomclockwherethede- citexpressionsforthecorrelationmatricesassociatedwith tectionnoiseiswell-characteriseditispossibletomeasure arbitrarynoisespectraareprovidedinAppendix.A.Fig.3, the servo prediction variance as an increase in the fluctu- forinstance,showstheflickerandrandom-walkAllandevi- ations of the error signal, but the resulting estimates are ationsreconstructedinthisfashionfromacorrelationmat- rixC estimatedfrom2×106 cyclesofoperationofasimu- latedsingle-atomclock,normalisedtoquantumprojection noise,asafunctionofthetrueleveloftherespectivenoise 3 Insomeimplementations,theestimatedfrequencymightnotberecor- processes. Random-walk noise, whose correlations differ dedassuch,butitcanbeobtainedbyaddingtheservopredictionstothe morestronglyfromthoseofthewhitemeasurementnoise, frequencyerrorsignalreportedbytheatomicreference. iseasiertodetect,butasseeninFig.3bothflickerandran- 4 Incaseswherethenoiseisnearlywhite,thiscanleadtoatheoretically dom walk noise can be reliably estimated from levels too optimalbutpracticallyuselesscontrollerwithnear-zerogain. Toavoid thisproblem,weimposeaminimumgainof0.04toensureafiniteservo lowtoaffecttheclock’sinstability(variancelessthan1%of attacktime.Thisboundislowenoughthatitdoesnotaffecttheachiev- projection noise) up to levels that would be unacceptably ableclockstability. highinnormaloperation,whentheclockservoisjumping 7 1 chain-dotted distribution, corresponding to the distribu- tion of LO prediction errors. Our best knowledge after a y measurementonasingleatom,intheeventthatitisfound t bili intheexcitedstate,isshownbythedashedprobabilitydis- a tribution. In this case we can be certain that the phase b o was not near φ = −π/2 (since the excitation probability r P would then have been 0), and we can be reasonably con- n o fidentthatitliesintheregionbetween0andπ,butwecan- ti ta notruleoutthatitliesnear−π,whereboththeexcitation xci probabilityandthepriorprobabilitydistributionarenon- E negligible.Varyingtheprobetime,andhencethespacingof 0 −3π −π −π 0 +π +π +3π theRamseyfringes,involvesatrade-off: increasingT nar- 2 2 2 2 rowsthemainlobeoftheposteriordistribution,thanksto PhaseErrorφ thesteeperslopeoftheexcitationsignal,butalsoincreases the weight of secondary lobes due to other fringes of the Figure4. Schematicillustrationofinformationgaininasingle Ramseyspectrum. cycleofclockoperation.Bluechain-dottedGaussian:Priorknow- To quantify this trade-off more formally, we consider ledgeoftheLOfrequency. Blacksolidsinusoid: Ramseyexcita- Ramseyinterrogationof N uncorrelatedatoms, takingthe tionspectrum. Reddashedline: PosteriorknowledgeoftheLO frequencygiventhatanatomwasdetectedintheexcitedstate. If prior distribution (chain-dotted curve in Fig. 4) to be a noexcitationisdetected,thentheposteriordistributionisthemir- Gaussianofvariancev rorimageaboutφ=0oftheoneshownhere. −φ2 P(φ)=(cid:112)e 2v . (20) 2πv between Ramsey fringes. Although this method provides muchlessdetailedinformationontheLOnoisespectrum Thisdistributionencodes,formally,allthatisknownabout than does the optical spectrum analyser of Ref. [34], it re- theLOfrequencybeforetheatomicmeasurementresultbe- quires no measurements beyond those performed as part comesavailable.Inapracticalsense,itisthedistributionof oftheclock’snormaloperation. Itcanthereforebeusedto servopredictionerrors: iftheservopredictionwereperfect monitortheLOwhiletheclockisrunning,eveninasingle- (h=x)thenthephaseaccumulatedintheRamseyinterrog- ionfrequencystandard,providinganearlywarningofper- ationwouldbezero. Theassumptionofnormallydistrib- formancedegradationsaswellasinformationusefulforthe utedservopredictionerrorsisborneoutbyoursimulations, optimisationofinterrogationparametersintheatomicref- which show that even with a single-atom reference and a erence(seeSec.V). binaryerrorsignaltheskewnessandexcesskurtosisofthe errordistributionarenomorethanafewpercentaslongas theclockremainslockedtothe(correct)Ramseyfringe. IV. IMPACTOFSERVOPERFORMANCEONLONG-TERM Thereferencedoesnot,unfortunately,supplyuswiththe STABILITY expectationvalueP .Rather,themeasurementyieldsaran- e domfractionF ofatomsdetectedintheexcitedstatethat Any measurement performed on a finite number of fluctuatesabouttheexpectationvaluePe duetomeasure- atomscanyieldonlyafinitenumberofpossibleresults.The mentnoiseoftheatomicreference. Intheabsenceoftech- optimisationofthemeasurementprotocolthusinvolvesa nicalnoiseonthereferencesignal,thevarianceofF is compromisebetweenfineresolutionoveranarrowusable (cid:191) P (1−P )(cid:192) domainofLOfrequenciesandcoarsefrequencyresolution var(F)= (Pe−〈Pe〉)2+ e e (21) N overabroaderdomain. Thebestcompromisedependson e−vsinhv 1−e−vsinhv therangeoffrequencieswhichmightplausiblyhavetobe = + (22) measuredbythereference,i.e.onthevarianceofservopre- 4 4N diction errors. In this section we study this compromise, where the averages 〈·〉 are taken over the prior distribu- showingthatitleadstoafiniteoptimalprobetimeandan tion P(φ). The first term expresses the fluctuations in the overalllimitonthelong-termstabilityofclockswithnoisy measured excitation fraction due to actual changes in the LOs. φ-dependent excitation probability, while the second cor- Bywayofillustration, considerRamseyinterrogationof responds to quantum projection noise of the binomially- asingleatom,wheretheprobabilitytofindtheatominthe distributed excitation signal. The usefulness of the excit- excitedstatedependsonthephaseerrorφ=(x−h)ωT as ationfractioninestimatingthefrequencydependsonthe 1+sinφ covarianceofthetwoquantities P = . (19) e 2 cov(φ,F)=(cid:173)φF(cid:174)−〈φ〉〈F〉= ve−v2 (23) 2 ThisexcitationspectrumisshownasasolidlineofFig.4.Let us assume that, before any measurement, our best know- whichdetermines how muchweight should be given to F ledgeofthe(corrected)LOfrequencyisrepresentedbythe inconstructingtheposteriorestimateoftheLOfrequency. 8 (cid:48) Choosingtheweighttominimisethevariancev oftheerror 101 (a)White(µ=−1) inthisposteriorestimate,wefind(c.f.AppendixC): v(cid:48)=v−cov(φ,F)2 (24) 100 var(F) (cid:181) (cid:182) Nv =v 1− . (25) (N−1)sinhv+ev 10−1 This posterior variance combines information from the measurement with information that was known before- hand. Inordertoisolatethecontributionoftheformer,we defineaneffectivemeasurementvariancevmby 101 (b)Flicker(µ=0) 1 1 1 = + . (26) v(cid:48) v vm τZ 100 (cid:112) Thisistheusualrelationforthevarianceofa(posterior)es- ω ) τ timateobtainedbyanoptimallinearcombinationoftwoin- ( C dependentpiecesofinformation. vmisthusthevarianceof σ 10−1 ahypotheticalmeasurement,onethatcouldbeinterpreted withoutanypriorknowledge,andwhichwouldreduceour uncertaintyontheLOfrequencyasmuchasdidtheactual measurement.Forthecaseweconsider, 101 (c)Random-Walk(µ=+1) ev (cid:181) 1(cid:182) v = + 1− sinhv−v (27) m N N 100 1 v2 v3 ≈ + + +··· (28) N 2N 6 where, in the second line, we have expanded the effective 10−1 measurementvarianceinpowersofthepriorvariance.The firsttermistheconventionalQPNonthemeasurementof thephase, validwhen v issmallandthecorrectedLOfre- 10−2 10−1 100 quency is known a priori to be well centred on the Ram- T/Z seyfringe.Thethirdtermreflectstheadditionaluncertainty arisingwhenthecorrectedLOfrequencycanlieoutsidethe Figure5. Long-terminstabilityasafunctionofprobetimefor range where the reference produces a meaningful result. clocks using Ramsey interrogation of 1 (solid black circles), 10 Thistermisindependentofatomnumber,anddominates (openbluecircles),102 (solidredtriangles),103 (opengreentri- theeffectivemeasurementvarianceasvapproaches1. angles)or104(solidgreysquares)uncorrelatedatoms.Solidlines Replacing the standard phase variance of Eqn. 1 by the showthepredictionofEqn.29,withoutfreeparameters. Vertical effectivemeasurementvarianceinEqn.2yieldsanewpre- barsmarktherecommendedinterrogationtimegiveninTableII. dictionforclockstabilityinthelimitoflargeaveragingtime Dashedlinemarksthelimitforperfectphaseestimationwithno τ,onethataccountsfortheeffectsoflimitedpriorinforma- projectionnoise. Thethreegraphsare,fromtoptobottom,fora white-,flicker-,orrandom-walk-dominatedLO. tionineachinterrogation: (cid:115) (cid:115) 1 T ev (cid:181) 1(cid:182) σC(τ)→ωT τc N + 1−N sinhv−v. (29) achievablelong-terminstabilityasafunctionofthechoice ofprobetime. In Fig. 5 we compare this prediction with simulations for Whentheprobetimeisshort,theRamsey(cid:112)fringeisbroad clockswithwhite-,flicker-,andrandom-walk-noise-limited andv issmall,theinstabilityimprovesas1/ T asconven- LOs, using Ramsey interrogation of uncorrelated atoms tionallyexpected. Theimprovementwithincreasingprobe withnodeadtime(T =T).Eachpointinthesegraphsisob- time stops either when the additional v-dependent terms c tainedfromasimulationof2×106cyclesofclockoperation. inEqn.29growimportantorwhentheservocannolonger TheAllandeviationiscomputedforatimeτlon(cid:112)genough reliablylocktheLOtothereferencetransition: thecurves thattheinstabilityhasreachedtheasymptotic1/ τregime inFig.5endwhenthefringe-hopratereaches1per2mil- (corresponding to 2×104 cycles of clock operation), then lioncycles.AsN increases,quantumprojectionnoiseisre- rescaled to a fixed averaging time Z and normalised to a ducedrelativetoLOnoiseanditbecomesadvantageousto fixednoiselevelσ (Z)toobtainadimensionlessresultthat reducetheprobetimesoastobelesssensitivetothelatter. L is comparable across systems. The graphs thus show the Thus, the optimal probe time gets shorter with increasing 9 LONoiseType Asymptoti(cid:112)cScaling LONoise Safe AsymptoticOptimum ofσC(τ)ω Zτ T/Z T/Z White ∝N−1/3 White – min(cid:179)N−1/5,1.4N−1/3(cid:180) Flicker ∝N−5/12 Flicker 0.4−0.15N−1/3 0.76N−1/6 Rnd.Walk ∝N−4/9 Rnd.Walk 0.4−0.25N−1/3 0.79N−1/9 Table I. Asymptotic scaling of LO-limited clock instability with TableII.RecommendationsforthechoiceofRamseyinterrogation atom number. The scaling differs from the conventional N−1/2 time. Thelastcolumngivestheoptimalprobetimeinthelimitof QPN limit because the optimum probe time decreases with in- manyatoms.Thereisnothingtobegainedbyprobinglongerthan creasingatomnumber. thistime. Itmaybenecessarytouseshorterprobetimestoavoid fringehops; asuggestedsafeupperboundontheprobetimeis giveninthesecondcolumn. atomnumber,andthefullyoptimisedclockinstabilitydoes notscaleasN−1/2. TheasymptoticscalingwithN isgiven inTableI.Forwhitenoise,themostextremecase,theop- obtaintheasymptoticallyoptimalprobetime(lastcolumn timalprobetimescalesasN−1/3 inthelarge-N limit,lead- ofTableII).Increasingtheprobetimebeyondthisoptimum ingtoaN−1/3scalingofthelong-termAllandeviation. For alwaysleadstoanincreaseineffectivemeasurementvari- flickerorrandom-walknoise,v fallsoffmoresteeplyasthe ance and long-term instability, and is of no practical in- probe time is shortened (see Eqn. 13), so that the the op- terest. At small atom numbers, shorter probe times are timalprobetimeislesssensitivetoatomnumberandascal- requiredtokeeptheservocontrollerrobustagainstfringe ingclosertotheconventionalQPNlimitisobtained. Inthe hops. Thepurelyphenomenologicalboundinthesecond absence of projection noise, i.e. in the limit N →∞, the columnofTableIIischosentobeslightlyshorterthanthe servoperformancelimitofEqn.13combineswithEqn.29 time for which we observe fringe-hops at a rate of 1 per toyieldageneralmeasurement-noise-independentlimiton million simulated clock cycles, with a 20% safety margin. clockinstability,whichweplotasdashedlinesinFig.5.This Ourchoiceofmaximumacceptablefringe-hoprate,corres- limitarisessolelyfromtheunpredictabilityoftheLOnoise ponding to a requirement that the clock remain locked to andfromthefinitedomainoverwhichtheRamseyerrorsig- thecorrectfringeforafewdays,isarbitrary,butastheonset nalcanbeunambiguouslyinterpreted. offringe-hoppingisextremelysteep(thefringe-hopratein Strictlyspeaking,Eqn.24holdsonlyiftheestimatedfre- simulationsincreasesbytwotothreeordersofmagnitude quency error is a linear function of the measured excita- whentheprobetimeisdoubled),themaximumsafeprobe tionfraction.Sincetheexcitationprobabilityisanon-linear timeisonlyweaklydependentonthischoiceofthreshold.A (e.g. sinusoidal) function of the LO frequency error, one fulloptimisationofallcommonprobeprotocolsinthepres- might hope to do better than the estimated performance enceofrealisticexperimentalimperfectionsisbeyondthe of Eqn. 29 by using a non-linear function to convert the scopeofthiswork,butweexpectqualitativelysimilarbeha- excitation fraction to a frequency error estimate. Simula- viourfromRabiorhyper-Ramseyprobing,withsomewhat tions show, however, that correcting for the curvature of longeroptimalprobetimesandslightlydegradedinstability theRamseyfringebyestimatingtheaccumulatedphaseas due to the increased width of the observed atomic reson- arcsin(2F−1)ratherthansimply2F−1hasnosignificantef- anceinthesesschemes. Conversely,weexpectthatclocks fectonvorontheachievablelong-termclockstability.One withsignificantdeadtimeintheiroperatingcyclewillneed canunderstandthisfindingbynotingthat,whenv islarge tousesomewhatshorterprobetimestocompensateforthe enoughthatthecurvaturewithinasingleRamseyfringeis servo’sinabilitytocorrectunobservedLOfrequencyfluctu- significant, the effect of unavoidable ambiguities such as ations,whichwillleadtoav higherthaninourdead-time- thesecondarylobeinFig.4ismuchlargeranddominates freesimulations. theposteriorvariance. VI. CONSTRAINTSONTHEBENEFITSOFENTANGLED V. GUIDELINESFORINTERROGATIONPARAMETERS ATOMICREFERENCES To choose the operating parameters for a clock, one Theargumentsdevelopedintheprecedingsectionsalso can in general use the formalism of Sec. II to predict the apply to certain Ramsey-like protocols using entangled servo error variance v as a function of those operating states in the atomic reference, provided that LO noise is parametersandEqn.29topredicttheresultinglong-term thelimitingformofdecoherence.Forinstance,thescheme stability, which can then be optimised. Table II, for ex- proposed in Ref. [35] and demonstrated in Refs. [36–38], ample,providesrecommendedRamseyinterrogationtimes employing N ato(cid:112)msinamaximallycorrelatedstate[|ψ〉= for clocks dominated by different types of power-law LO (cid:161)|g〉⊗N+|e〉⊗N(cid:162)/ 2, where |g〉 and |e〉 are the atomic ei- noise, expressed as multiples of Z. In the many-atom genstates], isfullyequivalenttoRamseyinterrogationofa limit, the servo prediction errors become independent of singleatomwithanN-foldenhancedtransitionfrequency thequantumprojectionnoiseandwecansolveEqn.29to by the corresponding harmonic of the LO radiation. Now 10 thelong-terminstabilityofasingle-atomfrequencystand- 102 ardcanbeexpressedas (a)White(µ=−1) s 101 σ (τ)= (cid:112) , (30) C ω Zτ 100 τ1 withsadimensionlessconstantoforderunityencodingthe choiceofprobetimeT/Z andtheadditionalcontribution 10−1 τ 2 ofLOnoiseatthisprobetime. Thelong-terminstabilityof theclockusingmaximally-correlatedatomsthusbecomes 10−2 σ (τ)= (cid:112)s , (31) 10−3 C Nω ZNτ 102 (b)Flicker(µ=0) where Z isthenoisetimescaleforthe N-foldfrequency- N multipliedLO: 101 τ1 σL(ZNc)NωZN=1rad. (32) ωZ 100 τ τ) 2 In the dead-time-free limit ZNc = ZN, one can compare (C 10−1 σ Eqn.32withEqn.3andfind 10−2 ZN =N2−+2µ, (33) Z 10−3 withµagaindescribingthetime-dependenceoftheLOAl- 102 lanvariance(seeSec.I).Theentangledclock’slong-termin- (c)Random-Walk(µ=+1) stabilitythusscalesas 101 τ 1 s 1 σ (τ)= (cid:112) . (34) 100 τ C ω ZτN12++µµ 2 10−1 Thus,iftheclockstabilityislimitedbywhiteLOnoise(µ= −1), a reference using a maximally-correlated state of N 10−2 atoms performs no better than a reference using a single atom,andisinfactworsethanareferenceusinguncorrel- 10−3 atedinterrogationoftheNatoms.Thisscalinghasbeenob- 10−4 10−2 100 102 104 servedexperimentallyforcorrelatedmagneticfieldnoisein τ/Z a14ionGHZstate[38].Forflicker-floorLOnoise,theAllan deviation improves as N−1/2 with N maximally-entangled Figure6. SimulatedAllandeviationasafunctionofaveragingtime atoms,veryslightlybetterthantheasymptoticN−5/12scal- for optimised clocks using either 100 uncorrelated atoms (solid ing achievable without entanglement, but worse than the bluecircles)oramaximally-correlatedstateof100atoms(open scalingachievedwithunentangledatomsforN <102. Itis red circles). The latter must operate with shorter interrogation only for random-walk LO noise that maximally entangled times,andthusthecorrespondingcurvesstartearlier. Theblack states offer measurable benefits, with an N−2/3 scaling of linemarkstheAllandeviationofthefree-runningLO.Thethree graphsare,fromtoptobottom,forawhite-,flicker-,orrandom- thelong-termAllandeviation. Weillustratethesescalings walk-dominated LO. Arrows mark the two integrator time con- inFig.6,whichplotstheAllandeviationspectrumrecorded stantsforthecaseofuncorrelatedatoms,asdiscussedinSec.VII. insimulationsoffullyoptimisedclocksusingeither100un- correlatedatomsora100-atommaximally-correlatedstate forallthreeLOnoisetypes. Maximallyentangledstatesreducethesignal-to-noisera- tioofameasurementonN atomstothatofasinglequbit, providing less new information per measurement but ac- as in Ramsey interrogation of uncorrelated atoms, and so celeratingtheclockcyclesothatmoremeasurementscan theyaresubjecttothesameprojection-noise-independent be averaged. That is why their use is advantageous with sinhv−v≈v3/6limitontheireffectivemeasurementvari- random-walkLOnoise, whenfastmeasurementscantake ance(c.f. Eqn.28). Theadditionalnoiseintroducedwhen advantage of the reduced LO noise at short time scales. servopredictionerrorsallowtheanti-squeezedquadrature Otherapproachestotheuseofentanglementinatomicref- tocontaminatethemeasurementresult[42]caninprinciple erences, such as spin squeezing [24–26, 39–41], focus in- beeliminatedbyasuitablereadoutprocedure[27]inwhich stead on improving the signal-to-noise ratio of the meas- caseweexpectsuchinterrogationprotocolstoofferbenefits urement, and thus increasing the amount of new inform- comparabletosuppressingtheprojectionnoisebyincreas- ationobtainedineachinterrogation. Theerrorsignalpro- ingatomnumber,evenwithwhite-orflicker-floorlimited ducedinsuchschemeshasthesameperiodicambiguities LOs.