ebook img

No-go theorem and optimization of dynamical decoupling against noise with soft cutoff PDF

0.27 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview No-go theorem and optimization of dynamical decoupling against noise with soft cutoff

No-gotheoremandoptimizationofdynamicaldecouplingagainstnoisewithsoftcutoff Zhen-Yu Wang and Ren-Bao Liu∗ Department of Physics and Center for Quantum Coherence, The Chinese University of Hong Kong, Shatin, N. T., Hong Kong, China Westudytheperformanceofdynamicaldecouplinginsuppressingdecoherencecausedbysoft-cutoffGaus- sian noise, using short-time expansion of the noise correlations and numerical optimization. For the noise correlationswithoddpowertermsintheshort-timeexpansion,thereexistsnodynamicaldecouplingschemeto eliminatethedecoherencetoarbitraryordersoftheshorttime,regardlessofthetimingorpulseshapingofthe controlunderthepopulationconservingcondition. Weformulatetheequationsforoptimizingpulsesequences fordynamicaldecouplingthatminimizesdecoherenceuptothehighestpossibleorderoftheshort-timeexpan- sion. Inparticular,weshowthattheCarr-Purcell-Meiboom-Gillsequenceisoptimalinshort-timelimitforthe noisecorrelationswithalinearorderterminthetimeexpansion. 2 PACSnumbers:03.67.Pp,03.65.Yz,03.67.Lx,82.56.Jn 1 0 2 I. INTRODUCTION periments, finding efficient DD schemes with fewer control pulses is desirable. A remarkable advance is the Uhrig DD t c (UDD)[26–29]. UDDisoptimalintheshort-timelimitinthe O Quantum information processing [1] relies on the coher- sensethatitsuppressesthepuredephasingofaqubitcoupled enceofquantumsystems. Unavoidableinteractionsbetween toafinitebathtothe Mthorderusingonly Mqubitflips. The 4 aquantumsystemanditsenvironment(bath)introducenoise performance bounds for UDD against pure dephasing were ] on the system and lead to error evolutions (decoherence) of established [30]. Shaped pulses [31, 32] of finite amplitude h the quantum system. Various methods have been proposed canbeincorporatedintoUDD[31]. Manyrecentexperimen- p to combat the decoherence, including decoherence-free sub- tal studies [20, 23, 33–37] demonstrated the performance of - t spaces[2–4],error-correctioncodes[5,6],anddynamicalde- UDD. n coupling(DD)[7–9]. Inparticular,theDDschemeusesrapid Itisimportanttofindefficientschemestosuppressgeneral a unitarycontrolpulsesactingonlyonthesystemstosuppress u decoherence(includingpuredephasingandpopulationrelax- theeffectsofthenoisefromtheenvironments.DDhasthead- q ation). Yang and Liu extended UDD to the suppression of [ vantages of suppressing decoherence without measurement, populationrelaxation[29].Thisinspiredefficientwaystosup- feedback,orredundantencoding[8]. DDoriginatedfromthe pressthegeneraldecoherenceofsinglequbits,includingcon- 1 seminal spin echo experiment [10], in which the effect of a v catenationofUDDsequences(CUDD)[38]andamuchmore 8 static random magnetic field (inhomogeneous broadening) is efficientonecalledquadraticDD(QDD)[25,39–42]discov- canceled.AndmorecomplexDDpulsesequences,suchasthe 3 eredbyWestetal[39]. Basedontheproofin[29],Mukhtar 3 Carr-Purcell-Meiboom-Gill(CPMG)sequence[11,12],were et al generalized UDD to protect arbitrary multilevel sys- 1 designedtoprolongthespincoherencetime[13]. temswithfullpriorknowledgeoftheinitialstates[43]. One . 0 The early DD schemes only eliminate low-order errors, can actually preserve the coherence of arbitrary multi-qubit 1 i.e., the errors of quantum evolutions up to some low order systems by protecting a mutually orthogonal operation set 2 in the Magnus expansion. By unitary symmetrization pro- (MOOS)[25]. BynestingUDDsequencesforprotectingthe 1 cedure [7, 8], DD cancels the first order (i.e., leading order) elementsintheMOOS,thenestedUDD(NUDD)[25,42,43] : v errors. To eliminate errors to the second order in short time, requires only a polynomially increasing number of pulses in Xi mirror-symmetric arrangement of two DD sequences can be thedecouplingorder. TheseuniversalDDschemesalsowork used [8]. The first explicit arbitrary Mth order DD scheme, foranalyticallytime-dependentbaths[44]. r a which suppresses errors to O(TM+1) for short evolution time Theabove-mentionedvariationsofUDD,however,relyon T,istheconcatenatedDD(CDD)[14,15]proposedbyKhod- the finiteness of the baths, i.e., the existence of hard high- jastehandLidar.CDDsequencesagainstpuredephasingwere frequency cutoff in the noise spectra. Legitimate questions investigatedforelectronspinqubitsinrealisticsolid-statesys- are:Forquantumsystemscoupledtoaninfinitequantumbath tems with nuclear spins as baths [16–18]. Experiments [19– or affected by soft-cutoff noise, can any DD be designed to 23] have tested the performance of CDD. CDD works for eliminate the decoherence to arbitrary orders of precision in generic quantum systems coupled to a finite bath [24, 25]. the short-time limit? And if yes, how can such DD be de- However, since CDD uses recursively constructed pulse se- signed? Such questions have been previously addressed in quences to suppress decoherence, the number of pulses in- somespecificnoisemodels. Comparingtheefficiencyofvar- creases exponentially with the decoupling order. As pulse ious DD sequences in suppressing pure dephasing of a qubit errors are inevitably introduced in each control pulse in ex- duetoclassicalnoise,Cywin´skietalobservedthatifthenoise spectrumcutoffisnotreached,CPMGsequences[11,12]ac- tually performs better than CDD and UDD sequences [45]. Also, Pasini and Uhrig derived the equations for minimizing ∗Electronicaddress:[email protected] decoherenceforpower-lawspectra,andfoundthatthenumer- 2 ically optimized sequences resemble CPMG [46]. Chen and destroys the quantum coherence. We can suppress the deco- LiuprovedthatfortelegraphlikenoisetheCPMGsequences herencebyDDcontrolonthequbit. Thereisonlyonenoise are the most efficient scheme in protecting the qubit coher- sourceβ(t)inthemodelEq.(1),andtosuppressthedecoher- ence in the short-time limit and the decoherence can be sup- ence we need to protect a MOOS which consists of a Pauli pressed at most to the third order of short evolution time by operatorσ (moregenerallyσ cosφ+σ sinφwithφbeing x x y DD[47]. Theseresultssuggestthatfornoisewithsoftcutoff real) [25]. We will prove later that DD can suppress the de- inthespectrum,therearecertainconstraintsontheefficacyof coherence (i.e., the protection of the MOOS {σ }) only to a x DD. However, no conclusion has been drawn on the perfor- certainorderofshortevolutiontimefornoisecorrelationsthat manceofDDwitharbitrarytimingandpulseshapingforthe haveodd-powerexpansiontermsintime. Weexpectthatthe generalcasesofsoft-cutoffnoise. proofalsoappliestootherquantumsystems(e.g.,multi-qubit In this paper, we address the general question of the per- systems) when the noise correlations have odd-power terms, formanceofDDagainstsoft-cutoffnoise,basedontheshort- sinceinthosesystemstherearemorenoisesourcesandmore timeexpansionofthenoisecorrelations. Theessenceoftime- system operators (e.g., a MOOS consisting of L > 1 Pauli expansionmethodistoavoiddivergenceintheintegrationof operators)shouldbeprotected. soft-cutoffnoiseinthefrequencydomain. Wefindthatwhen Whenweapplyasequenceofinstantaneousunitaryopera- the expansion of the noise correlation has odd-power terms tionsσ atthemomentsT ,T ,...,T ,thecontrolledevolu- x 1 2 N intime,thereexistsnoDDschemetoeliminatethedecoher- tionoperatorreads encetoanarbitraryorderoftheshorttime, regardlessofthe timing and shaping of the control pulses under the popula- U(T) = (σx)NU(TN+1,TN)σxU(TN,TN−1)··· tionconservingcondition. Theodd-powertermsinthetime- ×σ U(T ,T )σ U(T ,T ), (2) x 2 1 x 1 0 expansionindicatethatthenoisehasnohardhigh-frequency cutoff. Sinceforsoft-cutoffnoisethedecoherencecannotbe whereT0 =0,TN+1 =T,andthefreeevolutionoperator eliminatedatacertainorderofshorttime,wederiveasetof eexqpuaantisoinosntowmhiilneimeliizmeitnhaetilnegadtihneg-loorwdeerrtoerrdmerisn.thTehsehsoerte-tqiumae- U(Tj+1,Tj)=e−iσ2z(cid:82)TTjj+1[ωa+β(t)]dt. (3) tionsarenumericallysolvedforoptimalsolutions. Inparticu- NotethatwhenNisodd,wemayapplyanadditionalσ pulse lar,fornoisecorrelationswithalinearordertermintime,we x attheendofthesequencefortheidentityevolution. Using provethattheCPMGsequencesareoptimal. Forothernoise correlationswithodd-orderterms,theminimumpulseinterval oftheoptimizedsequencesislargerthaninUDDsequences. σxU(Tj+1,Tj)σx =e−iσ2z(cid:82)TTjj+1[−ωa−β(t)]dt, (4) This feature is important in realistic experiments when there isaminimumpulseswitchingtime[48]. wewritetheevolutionoperatoras This paper is organized as follows: In Sec. II, we analyze theperformanceofDDviatheshort-timeexpansionofcorre- U(T)=e−iσ2z(cid:82)0TωaFπ(t)dte−iσ2z(cid:82)0Tβ(t)Fπ(t)dt, (5) lationfunctions,andwegivetheconditionunderwhichdeco- herencesuppressiontoarbitraryorderisimpossible.Therela- where we have defined the modulation function for instanta- tionbetweenthehigh-frequencycutoffandtheshort-timeex- neousπ-pulsesequences[26,45] pansionofcorrelationsisalsodiscussed.InSec.III,wederive the equations for sequence optimization and obtain optimal  DD for noise correlations with odd-power expansion terms. F (t)=(−1)j fort∈(Tj,Tj+1] . (6) Finally,theconclusionsaredrawninSec.IV. π 0 fort>T, ort≤0 Under DD control, the qubit is flipped at different moments, II. NO-GOTHEOREMONDYNAMICALDECOUPLING and the random field β(t) is modulated by the modulation AGAINSTNOISEWITHSOFTCUTOFF function F (t). Formultilevelsystems, themodulationfunc- π tions resulting from instantaneous π-pulse sequences may We consider the pure-dephasing Hamiltonian for a single havevaluesnotrestrictedto{±1}fort∈(0,T]. InRef.[49],it spin(qubit) isshownthatforDDcomposedofspeciallyengineeredfinite- duration pulses, the effective modulation functions can take 1 valuesfrom{+1,−1,0}alternatively. Wemayalsoencounter H = σ [ω +β(t)], (1) 2 z a effectivemodulationfunctionswhicharetrianglewavefunc- tions during the time of system evolution [50]. For a more whereσ =|+(cid:105)(cid:104)+|−|−(cid:105)(cid:104)−|isthePaulioperatorofthequbit,ω z a general analysis, we assume that the control conserves the istheenergysplittingofthequbit,andβ(t)describesrandom populations and the phase modulation function F (t) has a noisewithaverageβ(t) = 0. Heretheoverbardenotesaver- π generalformas agingoverthenoiserealizations.Weassumethatthestatistics ofthenoisefluctuationsareGaussian.  AfteradurationoffreeevolutiontimeT,thenoiseinduces F(t)=aboundedfunction fort∈(0,T] . (7) (cid:82)T 0 fort>T ort≤0 between the two states |±(cid:105) a random phase shift β(t) that 0 3 AtthemomentT, theoff-diagonaldensitymatrixelement This kind of noise can be caused by Markovian (or instanta- ofanensembleis neous) collisions in the bath [53]. We can see from Eq. (13) thatwhenthenoisespectrumhashardhigh-frequencycutoff, ρ↑↓(T)=ρ↑↓(0)e−i(cid:82)0TωaF(t)dte−i(cid:82)0Tβ(t)F(t)dt, (8) the noise correlation function β(t)β(0) is analytic as a func- tionoftimeandtheexpansioncontainsonlyeven-orderterms, where F(t) = Fπ(t) for ideal instantaneous σx pulses on β(t)β(0) =(cid:80)∞ C t2k. the qubit. The coherence is characterized by the ensemble- cutoff k=0 2k Expanding the noise correlation in time has the advantage averagedphasefactor of avoiding the divergence of integration over frequency for noisewithasofthigh-frequencycutoffinthepowerspectrum. W(T)≡e−i(cid:82)0Tβ(t)F(t)dt. (9) Thereforewecanderivecertaingeneralresultsontheefficacy of general dynamical decoupling control against soft-cutoff ForGaussiannoise,theensemble-averagedphasefactorW(T) noise. Also,theperformanceofDDintheshort-timelimitis isdeterminedbythetwo-pointcorrelationfunctionβ(t1)β(t2) directlyderivedfromthetime-domainexpansion. andW(T)becomes[45,51,52] UsingtheexpansionEq.(14),wewrite W(T)=e−χ(T), (10) (cid:88)∞ χ(T)= C φ , (15) k k wherethephasecorrelation k=0 1(cid:90) T (cid:90) T wherethedecoherencefunctions χ(T)= dt dt β(t )β(t )F(t )F(t ), (11) 2 0 1 0 2 1 2 1 2 φ ≡(cid:90) Tdt (cid:90) t1dt (t −t )kF(t )F(t ). (16) k 1 2 1 2 1 2 can be written as the overlap between the noise power spec- 0 0 trumandafilterfunctiondeterminedbytheFouriertransform ThereforeifthemodulationfunctionF(t)isdesignedtomake ofthemodulationfunction[45]. TheaimofDDdesignisto minimizeχ(T). {φ =0}M ≡{φ =0,··· ,φ =0}, (17) k k=0 0 M We assume the noise is stationary, i.e., of time translation symmetry, β(t )β(t ) = β(t −t )β(0). The noise correlation the decoherence in the short-time limit is eliminated to the functionhasth1esy2mmetry1β(t)β2(0) = β(0)β(t),sinceβ(t)and Mthorder, i.e., χ(T) = O(TM+3). Notethate−i(cid:82)0TωaF(t)dt = 1 inEq.(8)whenφ =0.Theeven-orderdecoherencefunctions β(0)commute[Forquantumnoise,thismaynotbetrue.Butin 0 Eq.(11),t andt canbeexchangedwithoutchangingtheinte- φ2k describethedecoherenceduetothelow-frequencynoise, 1 2 gration. Sothenoisecorrelationcanalwaysbesymmetrized]. while the odd-order functions φ2k+1 arise from the soft high- frequencycutoffinthenoisespectrum. Thesesymmetriesindicatethatthenoisecorrelationisaneven functionoftime,i.e., β(t)β(0)≡C (t)=C (|t|). (12) A. Analysisofthedecoherencefunctions corr corr The noise correlation can be transformed from the noise Belowweshowthatwhenthenoisehassofthigh-frequency powerspectrumS(ω)as cutoff,itisnotpossibletosuppressthedecoherencetoanar- (cid:90) ∞ dω bitraryorder,nomatterhowtheDDmodulationfunctionF(t) β(t)β(0)= S(ω)e−iωt. (13) isdesigned. 2π −∞ The general filter function is defined as the Fourier trans- Toanalyzethedecoherencesuppressioninshort-timelimit, formofthegeneralmodulationfunction, weexpandthenoisecorrelationfunctioninpowerseries (cid:90) ∞ F˜(ω)≡ F(t)eiωtdt= F˜∗(−ω). (18) (cid:88)∞ −∞ β(t)β(0)≡ C |t|k. (14) k Weexpanditinpowerseries k=0 The existence of odd-order terms means that the correla- (cid:88)∞ (iω)m−1 F˜(ω)= Λ , (19a) tionfunctionisnon-analytical,whichindicatesthatthenoise m! m sourcecannotbeafinitequantumbath. Thenoisewithnon- m=1 (cid:90) T analytical correlation functions must come from the fluctua- Λ ≡m F(t)tm−1dt. (19b) tions of an infinite bath, since otherwise the unitary evolu- m 0 tion of a finite quantum system will always lead to analyti- NotethatΛ = O(Tm). Forasequenceof N instantaneousπ cal correlation functions. For example, the noise correlation m β(t)β(0) = e−|t|/tc has odd-order terms in the time expansion, pulses,F(t)isgivenbyEq.(6)andΛmreads and the noise has a Lorentz spectrum, which has a power- (cid:90) T (cid:88)N lawdecreaseathighfrequencies. Wecallthispower-lawde- Λ(π) =m F (t)tm−1dt= (−1)j(Tm −Tm). (20) creaseasofthigh-frequencycutoffinthenoisespectrum[46]. m 0 π j=0 j+1 j 4 AsF(t)=0fort(cid:60)(0,T],weextendtheboundsofintegra- whichisdecomposedas(withthechangesofsummationorder tionforttoinfinityandtransformEq.(16)to andindices) φk = ∂(∂iηk)k (cid:90)−∞∞ d2ωπ1 (cid:90)−∞∞ d2ωπ2 (cid:90)−∞∞dt1(cid:90)0t1dt2 φ2m = (2m)!(cid:90)−∞∞ d2ωπ 2(cid:88)rm=+112m(cid:88)n−=r1+22(iω−)(2−m1−)r+r3−nΛr!r Λn!n whereweset×ηF˜→(ω01)aF˜f(teωr2d)eiff−ieωr1et1net−iaiωt2iot2ne.iη(Itn1−tet2g),rationsove(r2t12), +2(cid:88)m+12m(cid:88)−r+22(iω−)(2−m1−)rn+3−nΛr!r Λn!n t andω give r=1 n=1 1 1 ∂k (cid:90) ∞ dω F˜(ω) −(2m)!2(cid:88)m+1(−1)2m−rΛr Λ2m−r+2 . (29) φ = [F˜(η)−F˜(−ω)]. (22) 2 r! (2m−r+2)! k ∂(iη)k 2π i(ω+η) r=1 −∞ Theintegralsofthetermswithodd-powerofωvanish. And Using for even functions of ω, the sum n+r is an odd number, so ∂(∂iηr)rF˜(η)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = (r+r!1)!Λr+1, forr≥0, (23) (th−e1c)rh+an(−ge1)onf=su0m.mThautisonthienidnetxegorfalthineElaqs.t(l2in9e)ivnanEiqsh.e(2s.9)W,withe η=0 write andchangingthesummationindex,weobtain φk =k!(cid:90)−∞∞ d2ωπ (|−F˜i(ωω))k|+21 −(cid:88)kr=+11 (−iFω˜()ωk−)r+2Λr!r. (24) φ2m =(2m)!(−21)m (cid:32)(mΛ+m+11)!(cid:33)2−(cid:88)r=m1(−1)rΛr!r (2mΛ−2mr−r++2(23)0!). FromEq.(30),wefindthatthefollowingtwosetsofequations Note that the summation over r and the integration over fre- areequivalent quency cannot be exchanged when the integration does not convergeforeachindividualterm. UsingEq.(19a),wehave {φ2m =0}mM=−01 ⇔{Λm =0}mM=1. (31)  φk = k!(cid:90)−∞∞ d2ωπ (|−F˜i(ωω))k|+21 −(cid:88)kr=+11k(cid:88)−n=r+12(iω(−)k1−)rk+−3r−nΛr!r Λn!n ofFthoereinqsutaatniotannseeotu{sΛπm-p=ulΛse(mπ)se=qu0e}mnM=c1esi,sthe optimal solution − (cid:88)k+1 (cid:88)∞ (iω(−)k1−)rk+−3r−nΛr!r Λn!n. (25) TUjDD =Tsin2(cid:20)2Nπ+j 2(cid:21), (j=1,2,...,N), (32) r=1n=k−r+3 which is the timing of UDD sequences [26]. Note that WesimplifythelastlinebyusingEq.(19b)andtheequality the decoherence functions φ describe the noise with high- 2m frequency cutoff in spectrum. Our result is consistent with (cid:90) ∞ dω(cid:88)∞ (±iω)n−rtn = 1 tr−1 , forr≥1,t≥0, (26) Refs.[28,29]. TheconditionsEq.(31)aremoregeneralthan 2π n! 2(r−1)! the one that leads to UDD, and Eq. (31) may lead to more −∞ n=r generaldesignsofoptimalDD. whichisprovedinAppendixA.Weobtain   φk =k!(cid:90) ∞ d2ωπ (|−F˜i(ωω))k|+21 −(cid:88)k+1k(cid:88)−r+2(iω(−)k1−)rk+−3r−nΛr!r Λn!n 2. Leadingodd-orderdecoherencefunctionφ2M−1 −∞ r=1 n=1 −k!(cid:88)k+1(−1)k−rΛr Λk−r+2 . (27) relFatoironthefunnocitsieonwβit(ht)sβo(f0t)hiisghn-oftreaqnuaelnyctiyc.cuIttosffe,xtphaennsoioisneccoonr-- 2 r=1 r! (k−r+2)! tains odd-order terms C2m+1φ2m+1 (cid:44) 0. Suppose the DD scheme has eliminated the leading order errors by making {φ = 0}M−1, i.e., Eq. (31). We want to study the leading 2m m=0 1. Even-orderdecoherencefunctionsφ2m odd-order decoherence functions φ2M−1 due to the soft high- frequencycutoffinthenoisespectrum. Wefirstconsidertheeven-orderdecoherencefunctionsφ , The odd-order decoherence function φ of Eq. (27) 2m 2M−1 whichiscausedbythenoisewithhardhigh-frequencycutoff. reads Fgroarlevvaennisnhuems.bEeqr.2(m27,)(−g|Fi˜iω(vω)e2)m|s2+1 isanoddfunctionanditsinte- φ = (2M−1)!(cid:88)2M(−1)rΛr Λ2M+1−r 2M−1 2 r! (2M+1−r)!   r=1 φ2m =(2m)!(cid:90)−∞∞ d2ωπ 2(cid:88)rm=+112m(cid:88)n−=r1+2(iω−)2(m−−1r)+r3−nΛr!r Λn!n +(2M2−π 1)!(cid:90)−∞∞dω(cid:34)(|−F˜i(ωω))2|M2  −(2m2)!2(cid:88)m+1(−1)2m−rΛr!r (2mΛ−2mr−r++22)!, (28) +(cid:88)2M 2M(cid:88)−r+1(iω)(2−M1−)rr+2−nΛr!r Λn!n. (33) r=1 r=1 n=1 5 TheconditionEq.(31)gives C = 0 andC = −1. We assume that the noise correlation 1 3 decreasessmoothlyatlongcorrelationtimes,thatis, (cid:90) ∞ dω(cid:34) |F˜(ω)|2 φ2M−1 =(2M−1)! −∞ 2π (−iω)2M (34) lim dk C (t)=0, fork=0,1,..., (39) + (cid:88)2M 2M(cid:88)+1−r (iω)(2−M1−)rr+2−nΛr!r Λn!n. and t→∞dtk corr r=M+1 n=M+1 (cid:90) ∞ dL Noticeinthesummation2M+1−r< M+1. Weobtain IL(ω)≡ eiωtdtLCcorr(t)dt, (40) 0 φ =(−1)M(2M−1)!(cid:90) ∞ dω|F˜(ω)|2. (35) vanishes at large ω for L = 0,1,... [54]. The noise correla- 2M−1 −∞ 2π ω2M tionsthatdecayinthecorrelationtimesmoothlywithoutfast oscillationsatisfyEq.(40).Forexample,thenoisecorrelation InEq.(35),theintegrand|F˜(ω)|2/ω2M ≥0anditcannotvan- e−|t|hasI (ω)=i(−1)L/(ω+i)→0whenω→∞. ish for all ω from −∞ to ∞. Thus we have the following L We consider the high frequency behavior of the Fourier theorem. transformofC (t), corr Theorem1. Forthenoisecorrelationβ(t)β(0) = (cid:80)∞k=0Ck|t|k (cid:90) ∞ with{C2m (cid:44)0}mM=−01andC2M−1 (cid:44)0,thereisnoDDschemethat S2K(ω)= Ccorr(t)eiωtdt. (41) caninduceafilterfunction F(t)toeliminatetheerrorsupto −∞ O(T2M+1)forshortevolutiontimeT. Thatis,thereisnoDD IntegrationbypartsL≥(2K+1)timesgives schemesatisfying{φ =0}M−1andφ =0. 2m m=0 2M−1   tioTnhFis(tt)h,eaonrdemithsohlodwssfothraatrbointrearcyannnoont-zseurpopmreossduthlaetidoencfouhnecr-- S2K(ω)=2(cid:60)(cid:88)r=L1 ((r−−iω1))r!Cr−1+ (I−Li(ωω))L, (42) encetoarbitraryorderwhenthenoisecorrelationisnotana- wherewehaveusedEqs.(37)and(39). lytic. Forexample,forthenoisewiththecorrelationfunction e−γ|t|,onecannotsimultaneouslyeliminatethetwoleadingde- UsingEqs.(38)and(40),weobtainforlargeω, coherencetermsC0φ0 andC1φ1,andtheerrorinducedbythe (2K−1)! (cid:32) 1 (cid:33) noiseisatleastoftheorderO(T3). S2K(ω)≈2 (iω)2K C2K−1+O ω2K+1 , (43) TheresultinRef.[47]thatDDcansuppressdecoherenceat mosttothethirdorderofshortevolutiontimefortelegraphlike whichisapower-lawdecreaseathighfrequencies. noiseisgeneralfornoisewitharbitrarystatistics. Inderiving When the noise correlation expansion contains only even- Theorem1,wehavemadetheassumptionthatthestatisticsof order terms, from Eq. (42) we have the noise spec- thenoiseareGaussianandthedecoherenceisdeterminedby trum S (ω) = 2(cid:60)I (ω)/(−iω)L for an arbitrarily large even L thetwo-pointnoisecorrelation.Fornon-Gaussiannoise,there L. From the assumption lim I (ω) = 0, we have ω→∞ L existsthedecoherenceeffectfromhigher-ordernoisecorrela- lim S (ω)ωL = 0 for an arbitrarily large L and there- ω→∞ even tions. forethenoisespectrumhasahardhigh-frequencycutoff. One example is the correlation function e−t2, which has the noise spectrumofexponentialform∼ exp(−ω2),andobviouslythe B. Noisecorrelationexpansionandhigh-frequencycutoff UDD sequence can suppress the noise4effect order by order. Thelargeωbehaviorofothercorrelationfunctionsoftheform Here we discuss further on the relation between the noise exp(−(cid:80)p α t2j)canbecalculatedbythesaddlepointinte- correlationexpansionandhigh-frequencycutoff. Letuscon- j=1 2j gration method, which gives a result of an exponential de- siderthecorrelationfunctionsoftheexpansionform crease at high frequencies (i.e., hard cutoff). For example, (cid:88)∞ when ω is very large, (cid:82)∞ e−t4eiωtdt (cid:39) 1(cid:61)(cid:113) 2π eg(ω), where β(t)β(0)=C (t)= C |t|k, (36) −∞ 2 a(ω) corr k=0 k g(ω)=3(ω4)43e−i2π/3anda(ω)=12(ω4)2/3ei2π/3. Wenowconsiderthenoisespectrum wheretheexpansioncoefficients 1 C ≡ 1 dkCcorr(t)(cid:12)(cid:12)(cid:12)(cid:12) (37) SP(ω)≈ ωP, forω≥Ω, (44) k k! dtk (cid:12)(cid:12) t→0+ approximated by a power law (P > 0) decrease beyond the frequencyΩ. Thenoisecorrelationfunctionis arefiniterealnumberswith C2k−1 =0, fork< K, (38a) β(t)β(0)=CΩ,cutoff(t)+CΩ,P(t), (45) C2K−1 (cid:44)0. (38b) wherethecorrelationduetohigh-frequencynoise The leading odd-order term in the short-time expansion of (cid:90) ∞ dω 1 Ccorr(t) is C2K−1|t|2K−1. An example is Ccorr(t) = e−|t|3 with CΩ,P(t)=2(cid:60) Ω 2π ωPe−iωt, (46) 6 for K ∈ {1,2,...}, which can be transformed to a more gen- eral form of noise c by using some scaling of the a+(ω/ωd)2K amplitude and ω. For example, the measured ambient noise for ions in a Penning trap has an approximate 1/ω4 spec- trum at high frequencies and a flat dependence at low fre- quencies[33],whichapproximatelycorrespondstothenoise spectrumEq.(52)withK =2. Thecorrespondingcorrelation functionofEq.(52)isobtainedbytheinversetransform (cid:90) ∞ dω e−iωt FIG.1: Thepathsfortheintegralof1/zP. β(t)β(0) = . (53) 2K 2π 1+ω2K −∞ Usingtheresiduetheorem,wehave andthenoisecorrelationatlowfrequencies CΩ,cutoff(t)=2(cid:60)(cid:90) Ω d2ωπSP(ω)e−iωt, (47) β(t)β(0)2K = 2iK (cid:88)K−1exp[eixπp[(−2inei+2πK1(2)n(+21K)|t|−] 1)]. (54) 0 n=0 2K causesadditionaldecoherence. TheeffectfromCΩ,cutoff(t)can Expandingtheright-handsideinpowersoft,weget beeliminatedefficientlybyDDsequences, suchasUDDse- quences,asthenoisespectrumofCΩ,cutoff(t)hasahardcutoff i (cid:88)∞ (−i|t|)m (cid:88)K−1 Ω. β(t)β(0)2K = 2K m! ei2πK(2n+1)(m−2K+1), (55) AsCΩ,P(t)isanevenfunction,wejustcalculatetheintegral m=0 n=0 forthecaseoft>0. Wehave(fort>0) andsummingthetermsgives (cid:34)(cid:90) (cid:90) (cid:90) (cid:35) (cid:60) 1 CΩ,P(t)= π cΩ+ ci+ c∞ zPeiztdz, β(t)β(0)2K = 2−Ki (cid:88)∞ (−mi|t!|)mei(m−e22iKK(+m1+)π1[)(π−/K1)−m1+1]. (56) m=0 wherethepathscΩ,ci,andc∞areshowninFig.(1). Since the maximum of 1/zP → 0 as |z| → ∞ in the upper InEq.(56),theleadingodd-ordertermofthetimeexpansion (cid:82) half-plane,thecontribution c∞ z1Peiztdz=0. Thecontribution is 2((2−K1−)K1)!|t|2K−1,aspredictedbyEq.(51). ByTheorem1,we (cid:60)(cid:90) 1 (cid:60)(cid:90) ∞ havethefollowingcorollary. eiztdz= i1−Py−Pe−ytdy (48) π ci zP π Ω Cdeocroohlelarernyc1e.upDytonaamnicearlrodrecOo(uTp2lKin+g1)ciannthoenltyimseu-pepxrpeasnssitohne vanishesforevenP. And ofχ(T)forthenoisespectrumwithhigh-frequencyasymptote (cid:60)(cid:90) 1 eiztdz= (cid:60)(cid:90) π2 1 eiΩteiθ(iΩ)eiθdθ ∼ω−2K. π zP π ΩPeiPθ For a more general noise spectrum with high-frequency cΩ 0 = I(1)+I(2), (49) asymptote ω−α, one can not suppress the error higher than cΩ cΩ the order of O(Tmα+1) with mα being the smallest even num- where ber larger than α. This is because comparing with ω−mα, the asymptoticnoisespectrum 1 withα<m decreasesslower I(1) = (cid:60) (cid:88)∞ 1 Ω1−P(iΩt)r (cid:16)ir+1−P−1(cid:17), (50a) andtheoverlapbetweenS(1ω+)ωαand|F˜(ω)|2iαslarger. cΩ π r!(r+1−p) The 1/f noise is a special case (the case of P = 1) to Ic(2Ω) = 21(cid:60)r=(0Pi,Pr(cid:44)−tPP−−111)!(cid:88)r∞=0δr,P−1. (50b) fbπ1ue(cid:82)nΩcd∞ttiiscoconusszχs/ePzd=d1.z(TdT)ivh[eeErgqne.os(is1ae1t)tc]o→irnrde0lua;cteihodonwbefyuvnethrc,teitohhneigdChe-Ωcf,roPe=hq1eu(rtee)nncc=ye noiseisfinite. WhenthereisnoDDcontrol,i.e.,forthecase ForevenP,I(1)isanexpansionofevenpowersoft,and offreeinductiondecay, cΩ Ic(2Ω) = 2((−P1−)P1/2)!|t|P−1, forevenP, (51) χP=1(T)= 2π1Ω2 [1−cos(ΩT)+ΩTsin(ΩT)] T2 (cid:90) ∞ cosz + dz, (57) istheleadingoddorderexpansiontermofthenoisecorrela- 2π ΩT z tionfunction. Asanexample,weconsidertheflowingnoisespectrum, whichhastheshort-timeexpansion 1 1 (cid:16) (cid:17) S2(cid:48)K(ω)≡ 1+ω2K, (52) χP=1(T)= 4π(−2ln(ΩT)+3−2γe)T2+O Ω2T4 , (58) 7 with the Euler constant γ ≈ 0.577. Notice the interesting e 18 resultthatχP=1(T)doesnothavethesimplepowerlawscaling atshorttimeT. Aftereliminatingtheeffectoflow-frequency 16 noisebyhigh-orderDDsequenceswhichsatisfy{Λ =0}M 14 m m=1 (Eq. (31)) w(cid:16)ith 2(cid:17)M > P−1, χP(T) has a power la(cid:16)w s(cid:17)caling 12 χP(T) = O TP+1 [46],andwehaveχP=1(T) = O T2 when 10 Λ =0. 1 8 6 ODD III. SEQUENCEOPTIMIZATIONINSHORT-TIMELIMIT 4 UDD CPMG 2 TheaimofoptimalDDistominimizeχ(T). Asindicated 0 0.2 0.4 0.6 0.8 1 in Eq. (35), a smaller F˜(ω) at low frequencies will give a smallerφ . HerewefocusonDDwithidealinstantaneous 2M−1 FIG.2: (Coloronline). ComparisonoftheODD,UDDandCPMG π pulses. We use more pulses to construct a more efficient sequences for different pulse number N. Squares (blue), triangles modulation function F(t) = Fπ(t) to minimize φ2M−1, with (green),andcircles(red)correspondtoUDD,CPMG,andODD.The the conditions {φ2m = 0}mM=−01 [Eq. (31)]. Using the method ODDsequencesareoptimizedtominimizeφ3 undertheconstraints of Lagrange multipliers, we need to solve a set of nonlinear φ =φ =0. 0 2 equationsas ∇{y,T}GM =0, (59a) anN-pulseCPMGsequencereads (cid:88)M GM ≡ yjΛ(jπ)+φ2M−1, (59b) TCPMG = 2j−1T, for j=1,...,N. (62) j=1 j 2N ∂ ∂ ∂ ∂ ∇ ≡( ,..., , ,..., ). (59c) TheCPMGsequencesobviouslysatisfytheconstraintΛ(π) = {y,T} ∂T ∂T ∂y ∂y 1 1 N 1 M 0 [see Eq. (20)], which is the so-called echo condition that The introduced variables {y } are the Lagrange multipliers. eliminates the effect of static inhomogeneous broadening. j NotethatthesequenceoptimizationinRef.[46]alsousedthe Eqs.(61)and(20)give constraints {Λ(mπ) = 0}, but the constraints were used there to ∂φ (cid:12)(cid:12) T2 (cid:104) (cid:105) guarantee the convergence of the calculation of χ(T). Here, 1(cid:12)(cid:12) =(−1)k+1 1+(−1)N , (63a) the constraints eliminate the lowest orders of errors ({φ2m = ∂Tk(cid:12)CPMG (cid:12) 4N2 d0}emMc=o−0h1)er[esneceeEfqro.m(31in)]hoinmoshgoenrte-otiumseblriomaidte.niInngpiasrteicliumlairn,attehde y1∂∂ΛT(1π)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =2(−1)k+1y1. (63b) whenφ =0. k (cid:12)CPMG 0 WecalculateEq.(16)byseparatingthedomainofintegra- Thus the CPMG sequences also satisfy Eq. (59) with y = tion according to the value of F (t )F (t ). For k ≥ 0, we (cid:104) (cid:105) 1 π 1 π 2 − T2 1+(−1)N ,sotheyareatleastthelocallyoptimalpulse obtain 8N2 sequences. It has been proved that CPMG sequences are the  φk = (k+1−)(1k+2)4(cid:88)N (cid:88)j−1(Tj−Ti)k+2(−1)i+j menocsetaegffiaicniestnttepleuglsreapshe-qliukeencneosisienipnrtohteecsthinogrt-thtiemqeulbimiticto[h4e7r]-. j=2 i=1 Withnumericalevidence,weconjecturethatitisalsotruethat (cid:88)N theCPMGsequencesareoptimalintheshort-timelimitwhen +(TN+1−T0)k+2(−1)N+1+2 (Tj−T0)k+2(−1)j the time expansion of the noise correlation function has the j=1 twoleadingtermsC0andC1|t|. +2(cid:88)N (TN+1−Ti)k+2(−1)N+1+i. (60) {ΛF(mπo)r=ot0h}emMr=1c,aosense coafnmsieneimthiaztinthgeφs2hMo−r1t-twimitehotphteimcioznedditDioDn i=1 (ODD) coincides with UDD for pulse number N ≤ M. And as N increases, the ODD sequences gradually approach the Thenwehave CPMG sequences. For example, ODD for the noise correla-  ∂φ∂2TM−1 = (−M1)k 2 (cid:88)N (Tj−Tk)2M(−1)j−(Tk−T0)2M tion k j=k+1 β(t)β(0)=C0+C2t2+C3|t|3+O(t4), (64)  +(TN+1−Tk)2M(−1)N+1−2(cid:88)k−1(Tk−Tj)2M(−1)j. (61) isshowninFig.2incomparisonwithUDDandCPMG.The j=1 ODD sequences resemble the CPMG sequences when N is large. ForthespecialcaseofM =1,G =y Λ(π)+φ ,wefindthat WeshowinFig.3(a)theperformanceofthreeDDschemes 1 1 1 1 theCPMGsequencesaresolutionstoEq.(59). Thetimingof againstthenoisedescribedbyEq.(64). Acomparisonisalso 8 model,weexpectthattheresultoftheexistenceofthelargest decoupling order in short-time limit can be extended to the generaldecoherencemodel(includingbothdephasingandre- laxation)ofquantumsystems. ItisdesirabletostudytheDD in suppressing the general decoherence of quantum systems inthesoft-cutoffnoiseinthefuture. Acknowledgments WethankYi-FanLuoandBoboWeifordiscussions. This workwassupportedbytheHongKongGRFCUHK402209, the CUHK Focused Investments Scheme, and the National FIG.3:(Coloronline).Thedecoherencefunctionχ(T)asafunction ofpulsenumberN (a)withouthardcutoffand(b)withahardhigh- NaturalScienceFoundationofChinaProjectNo. 11028510. frequency cutoff ω = 40. Here the noise spectrum S(ω) = κ , c 1+ω4 with κ = 105 and T = 0.5. Squares (blue), triangles (green), and circles(red)correspondtoUDD,CPMG,andODD. TheODDse- AppendixA:ProofofEquation(26) quencesarethesameasthoseshowninFig.2. ToproveEq.(26),wejustneedtoprove showninFig.3(b)byconsideringahardhigh-frequencycut- (cid:90) R (cid:88)∞ (±ix)n−r π offω =40.InFig.3(a),wecanseethatODDsequencesgive lim dx = , forr≥1, (A1) bettercperformance than UDD and CPMG sequences. These R→∞ −R n=r n! (r−1)! ODDsequencesareoptimalforawiderangeofnoisewhich wheretheboundsintheintegralguaranteethatthemodulation has the noise correlation given by Eq. (64). When we intro- duceahigh-frequencycutoffinthenoisespectrum,asthecase functionF(t)isarealfunction. Using inFig.3(b),theODDisthebestinitiallywhenthenumberof pulses N (cid:46) ωcT/2 ≈ 10, and the UDD sequences become (cid:90) Rdx(cid:88)∞ (±ix)n−r =(cid:88)∞ Rn (in−1+c.c.), (A2a) better and suppress the decoherence order by order when N n! (n+r−1)!n islargeandthehardcutoffisreached. InFig.3(b), forlarge −R1 n=(cid:90)r R sinx n(cid:88)=∞1 Rn 1 N UDD is better than ODD, since the ODD sequences are dx= (in−1+c.c.), (A2b) optimized for soft-cutoff noise rather than hard-cutoff noise. (r−1)! −R x n=1 n!n(r−1)! InFig.3(a)thedecreasingofχ(T)isalineardecreaseinthe double-logarithmplot,butinFig.3(b)itismuchfaster. This andlim (cid:82)R sinxdx=π,wejustneedtoprove confirmsthatDDisnotsoefficientagainstsoft-cutoffnoise. R→∞ −R x (cid:88)∞ (cid:34) Rn Rn (cid:35) lim − (in−1+c.c.)=0. (A3) R→∞ (n+r−1)!n n!n(r−1)! IV. SUMMARYANDCONCLUSIONS n=1 For r = 1, it obviously holds. For r ≥ 2, we can show the Wehavestudiedthedynamicaldecouplingcontrolofdeco- difference herence caused by Gaussian noise with soft cutoff. We have tpperaronmvse,iodDnTDohfecnoaorneismoenc1loywrrsheulipacpthirosenhssohwadssectthohehate(,2rweKnhc−een1t)tohtheOos(dhTdo2rKet-x+t1pi)ma.nesTieohxne- ∆≡(cid:88)n∞=1 R(nn(+in−r1−+1c).!cn.)(r−1)!−(cid:89)rj−=11(n+ j)=O(cid:32)R1(cid:33), existenceofodd-ordertermsintheshort-timeexpansioncor- (A4) respondstoasofthigh-frequencycutoffinthenoisespectrum solimR→∞∆=0. Byexpandingthetermsinthesquarebrack- (i.e., a power-law asymptote at high frequencies). For these etsofEq.(A4),wehave noise spectra, we have derived the equations for pulse se-   qdueceonhceeroenpctiemfiuznactitoionn,wanhdicehlimmiinniamteizseesvtehne-olerdaedrindgecoodhde-roerndceer ∆=(cid:88)∞ R(nn(i+n−1r+−c1.)c!.)(cid:88)r−2aknk, (A5) functions of lower orders. The ODD sequences obtained by n=1 k=0 thismethodcoincidewiththeUDDsequenceswhenthepulse wherea isanumberindependentofn. Wearrangetheterms k numberNissmall,andtheyresembleCPMGsequenceswhen inthesquarebracketsandget N islarge. Forthespecialcasethattheshort-timecorrelation function expansion has a linear term in time, the ODD se-   quencesareexactlytheCPMGsequences. ∆=(cid:88)∞ R(nn(i+n−1r+−c1.)c!.)(cid:88)r−2bk(cid:89)k [(n+r− j)]+b0, (A6) Although we derived Theorem 1 from a pure dephasing n=1 k=1 j=1 9 with b independent of n. After some simplification it be- Hence∆=O(1),andEq.(26)isproved. j R comesforr≥2 (cid:88)r−1 ibkr+−1kR−1k eiR−(cid:88)k−1 Rn!nin+c.c.=O(cid:32)R1(cid:33). (A7) k=1 n=0 [1] M. A. Nielsen and I. L. Chuang, Quantum Computation 129901(2011). andQuantumInformation(CambridgeUniversity,Cambridge, [27] B.Lee,W.M.Witzel,andS.DasSarma,Phys.Rev.Lett.100, 2000). 160505(2008). [2] L.-M.DuanandG.-C.Guo,Phys.Rev.Lett.79,1953(1997). [28] G.S.Uhrig,NewJ.Phys.10,083024(2008). [3] P.Zanardi,Phys.Rev.A57,3276(1998). [29] W.YangandR.-B.Liu,Phys.Rev.Lett.101,180403(2008). [4] D.A.Lidar,I.L.Chuang,andK.B.Whaley,Phys.Rev.Lett. [30] G.S.UhrigandD.A.Lidar,Phys.Rev.A82,012301(2010). 81,2594(1998). [31] S.Pasini,T.Fischer,P.Karbach,andG.S.Uhrig,Phys.Rev.A [5] P.W.Shor,Phys.Rev.A52,2493(1995). 77,032315(2008). [6] A.M.Steane,Proc.R.Soc.Lond.A452,2551(1996). [32] B. Fauseweh, S. Pasini, and G. S. Uhrig, Phys. Rev. A 85, [7] P.Zanardi,Phys.Lett.A258,77(1999). 022310(2012). [8] L. Viola, E. Knill, and S. Lloyd, Phys. Rev. Lett. 82, 2417 [33] M. J. Biercuk, H. Uys, A. P. VanDevender, N. Shiga, W. M. (1999). Itano,andJ.J.Bollinger,Nature458,996(2009). [9] W.Yang,Z.-Y.Wang,andR.-B.Liu,Front.Phys.6,2(2011). [34] J. Du, X. Rong, N. Zhao, Y. Wang, J. Yang, and R.-B. Liu, [10] E.L.Hahn,Phys.Rev.80,580(1950). Nature461,1265(2009). [11] H.Y.CarrandE.M.Purcell,Phys.Rev.94,630(1954). [35] M. J. Biercuk, H. Uys, A. P. VanDevender, N. Shiga, W. M. [12] S.MeiboomandD.Gill,Rev.Sci.Instrum.29,688(1958). Itano,andJ.J.Bollinger,Phys.Rev.A79,062324(2009). [13] M. Mehring, Principles of High Resolution NMR in Solids [36] H.Uys,M.J.Biercuk,andJ.J.Bollinger,Phys.Rev.Lett.103, (Spinger-Verleg,Berlin,1983),2nded. 040501(2009). [14] K. Khodjasteh and D. A. Lidar, Phys. Rev. Lett. 95, 180501 [37] E.R.Jenista,A.M.Stokes,R.T.Branca,andW.S.Warren,J. (2005). Chem.Phys.131,204510(2009). [15] K. Khodjasteh and D. A. Lidar, Phys. Rev. A 75, 062310 [38] G.S.Uhrig,Phys.Rev.Lett.102,120502(2009). (2007). [39] J.R.West,B.H.Fong,andD.A.Lidar,Phys.Rev.Lett.104, [16] W.Yao,R.-B.Liu,andL.J.Sham,Phys.Rev.Lett.98,077602 130501(2010). (2007). [40] G.QuirozandD.A.Lidar,Phys.Rev.A84,042328(2011). [17] W.M.WitzelandS.DasSarma, Phys.Rev.B76, 241303(R) [41] W.-J.KuoandD.A.Lidar,Phys.Rev.A84,042329(2011). (2007). [42] L.JiangandA.Imambekov,Phys.Rev.A84,060302(2011). [18] W. X. Zhang, V. V. Dobrovitski, L. F. Santos, L. Viola, and [43] M.Mukhtar,T.B.Saw,W.T.Soh,andJ.Gong,Phys.Rev.A B.N.Harmon,Phys.Rev.B75,201302(R)(2007). 81,012331(2010). [19] X.Peng, D.Suter, andD.A.Lidar, J.Phys.B:At.Mol.Opt. [44] Z.-Y.WangandR.-B.Liu,Phys.Rev.A83,062313(2011). Phys.44,154003(2011). [45] Ł. Cywin´ski, R. M. Lutchyn, C. P. Nave, and S. Das Sarma, [20] G.A.A´lvarez,A.Ajoy,X.Peng,andD.Suter,Phys.Rev.A82, Phys.Rev.B77,174509(2008). 042306(2010). [46] S.PasiniandG.S.Uhrig,Phys.Rev.A81,012309(2010). [21] A. M. Tyryshkin, Z.-H. Wang, W. Zhang, E. E. Haller, J. W. [47] K.ChenandR.-B.Liu,Phys.Rev.A82,052324(2010). Ager,V.V.Dobrovitski,andS.A.Lyon,arXiv:1011.1903v2. [48] K. Khodjasteh, T. Erde´lyi, and L. Viola, Phys. Rev. A 83, [22] Z.-H. Wang, W. Zhang, A. M. Tyryshkin, S. A. Lyon, J. W. 020305(2011). Ager, E. E. Haller, and V. V. Dobrovitski, Phys. Rev. B 85, [49] G.S.UhrigandS.Pasini,NewJ.Phys.12,045001(2010). 085206(2012). [50] Z.-Y.WangandR.-B.Liu,unpublished. [23] C.Barthel,J.Medford,C.M.Marcus,M.P.Hanson,andA.C. [51] P.W.Anderson,J.Phys.Soc.Jpn.9,316(1954). Gossard,Phys.Rev.Lett.105,266808(2010). [52] R.Kubo,J.Phys.Soc.Jpn.9,935(1954). [24] L.F.SantosandL.Viola,NewJ.Phys.10,083009(2008). [53] P.R.BermanandR.G.Brewer,Phys.Rev.A32,2784(1985). [25] Z.-Y.WangandR.-B.Liu,Phys.Rev.A83,022306(2011). [54] C.P.SunandX.J.Liu,ActaPhys.Sin.5,343(1996). [26] G. S. Uhrig, Phys. Rev. Lett. 98, 100504 (2007), ibid, 106,

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.