quantum statistics
Kahn, J.
Citation
Kahn, J. (2008, June 17). Quantum local asymptotic normality and other questions of quantum statistics. Retrieved from
https://hdl.handle.net/1887/12956
Version: Corrected Publisher’s Version
License: Licence agreement concerning inclusion of doctoral thesis in the Institutional Repository of the University of Leiden
Downloaded from: https://hdl.handle.net/1887/12956
Note: To cite this publication please use the final published version (if
applicable).
Normality
and
other questions
of Quantum Statisti s
Proefs hrift
ter verkrijging van
de graad van Do tor aan de UniversiteitLeiden,
opgezag vanRe tor Magni us prof.mr. P.F. van der Heijden,
volgens besluit van het Collegevoor Promoties
teverdedigen opdinsdag 17juni 2008
klokke 15uur
door
Jonas Kahn
geboren te Maisons-Alfort,Fran e
in1982
promotor prof. dr. R. Gill
opromotor prof. dr. P. Massart (Orsay)
referent prof. dr. M. Hayashi (Tokyo)
overige leden
prof. dr. W. Th. F. den Hollander
dr. F. Redig
prof. dr. P. Stevenhagen
and
other questions
of Quantum Statisti s
Chapter 2(Dis rimination)
G.M. D'Ariano,M.F. Sa hi,andJ.Kahn. Minimax quantum statedis rimi-
nation. Phys. Rev. A,72:032310,2005a. arXiv:quant-ph/0504048.
G.M. D'Ariano, M.F. Sa hi,and J.Kahn. Phys. Rev. A, 72:052302,2005b.
arXiv:quant-ph/0507081.
Chapter 3(Fast estimationof unitary operations)
Fast rateestimation ofunitary operationsin
SU (d)
. Phys. Rev. A, 75:022326,2007b. arXiv:quant-ph/0603115.
Chapter 4(Clean positive operator valued measures)
Clean positiveoperator valued measures for qubits and similar ases. J. Phys.
A,Math. Theor., 40:48174832,2007a. arXiv:quant-ph/0603117.
Chapter 5(Complementarysubalgebras)
J.KahnandD Petz. Complementary redu tionsfortwoqubits. J. Math. Phy.,
48:012107,2007. arXiv:quant-ph/0608227.
Chapter 6(QLAN forqubits)
M.Guµ andJ.Kahn. Lo alasymptoti normalityforqubit states. Phys. Rev.
A,73:052108,2006. arXiv:quant-ph/0512075.
Chapter 7 (Optimal estimation of qubit states with ontinuous time
measurements)
M. Guµ , B. Janssens, and J. Kahn. Optimal estimation of qubit states with
ontinuous time measurements. Comm. Math. Phy., 277(1):127 160, 2008.
arXiv:quant-ph/0608074.
Chapter 8(QLAN nitedimension)
M.Guµ andJ.Kahn. Lo alasymptoti normalityfornite-dimensionalsystems.
SoumisàComm. Math. Phy.,2008. arXiv:0804.3876
ISBN978-90-9023254-6
A knowledgments
Ringraziamenti
Köszönetnyilvánítás
This thesis has been written under the s ienti dire tion of Ri hard Gill. I
thank him for introdu ing me to the world of quantum statisti s, and to lo al
asymptoti normality. The periods I spent with himwere alwaysenlightening,
be it onquantum statisti s, oron lassi al statisti s, orthe way astatisti ian
should behave. I also admire his will of introdu ing morepoeple to this eld,
andhisalwayssuggestingproblemsofinterest.
Pas al Massart a été mon deuxième dire teur de thèse. Son aide dans la ba-
taille administrative a été des plus pré ieuses. Il a également su me fournirla
bibliographieidoinequandl'o asions'yprêtait.
Inmymind, M d linGuµ was athird supervisoraswellasmymain ollabo-
rator. Wehaveworkedoutstrongquantumlo alasymptoti normalitytogether.
I havedeeply appre iatedmystaysin Nottingham, andhis wayof ndingnew
quantumproblems outof lassi alones. Thankyouforthepi turethat adorns
thisthesis' over,too.
Nottingham wasalso the pla e to speak with Belavkin, and takeadvantageof
hishugehistori alknowledge.
SpeakingofQLAN,IalsothankBasJanssensforourworkonquantumsto has-
ti dierentialequations. Dis ussions with Anna Jen£ováfurther deepened my
understandingofQLAN,byyieldingtheotherside,namelyweakequivalen e.
Nelle due settimane he ho passato in Pavia, all'inizio della tesi, Professore
milianoSa hi.Ringraziotuttalasquadralo ale, hehoan oravistoin ongressi
duranteglianniseguenti.
KöszönömPetzDénesfelhívásotBudapestre. Nem sakkészültünkela íkkeim-
mel,desokattanítottis aCCRalgebrák-ról.
I regretnotbeingableto swit hto JapanesetothankHayashiandMatsumoto
formystayinJapaninMar h2007,andthedis ussionswehavehadonquantum
statisti s,espe iallyquantum Cramér-Raobounds.
Cristina Butu ea a été mon interlo uteur en Fran e, la seule ave laquelle je
puissedialoguersurlesujetdesstatistiquesquantiques.
Jeremer ieaussitous euxquiontétéautourdemoi,élèves ommeprofesseurs,
quand j'étais àl'ENS. Les dis ussions s ientiques permanentes, les problèmes
quenousnousposions,restentlaplusbelleformationquej'aiepuavoir.
Mer iYanPautrat,pourlesparolesé hangéessurlaphysiquestatistiquequan-
tique du point de vue mathématique, et pour avoir invité Vladimir Jak²i¢ à
donnerun oursinspiré.
Mer i enn à eux qui ont relu mathèse, Cristina,Patri iaReynaud, Borg,et
surtout mon voisin de bureau Sylvain Arlot, dont je me suis outrageusement
inspirédansl'espoir himérique d'appro hersa larté.
J'aiessayéd'être aussineutrequepossibledansleslignespré édentes, bienque
plusieursdespersonnes évoquéesméritentletitre d'ami.Mais j'espèrequetous
euxqui omptentpourmoi,autrementplusquelesmagniquesstatuesdegla e
quej'étudie,n'ontpasbesoinquejeles ite pourlesavoir.
Remer iementsA knowledgments Ringraziamenti
Köszönetnyilvánítás i
1 Introdu tion 1
1.1 Statisti s. . . 3
Classi al Statisti s,3.
•
QuantumObje tsandOperations, 10.•
Quantumstatisti s,19.
1.2 Dis rimination. . . 26
Motivation,26.
•
Formerresults,27.•
Contributionsofthethesis,31.1.3 FastEstimationofUnitaryOperations . . . 31
Motivation,31.
•
Formerresults,32.•
Contributionsofthethesis,34.1.4 Clean PositiveOperatorValuedMeasures. . . 34
Motivation,34.
•
Formerresults,35.•
Contributionsofthethesis,37.1.5 Complementarysubalgebras. . . 38
Motivation,38.
•
Formerresults,39.•
Contributionsofthethesis,39.1.6 Quantumlo alasymptoti normality. . . 40
Classi al lo al asymptoti normality, 40.
•
Motivation, 43.•
Former andrelated results,43.
•
Contributionsofthethesis,45.•
Outlook,46.I Mis ellaneous Problems in Quantum Statisti s 47
2 Dis rimination 49
2.1 Introdu tion . . . 49
2.2 Optimalminimaxdis riminationoftwoquantumstates. . . 52
2.3 Optimalminimaxdis riminationof
N ≥ 2
quantumstates. . . . . 552.4 Optimalminimaxunambiguousdis rimination . . . 58
2.5 Bayesiandis riminationoftwoPauli hannels . . . 59
2.6 Minimaxdis riminationofPauli hannels . . . 61
3 Fastestimationof unitary operations 71 3.1 Introdu tion . . . 71
3.2 Des riptionoftheproblem . . . 74
3.3 Whywe annotexpe tbetterratethan
1/N 2
. . . . . . . . . 783.4 Formulasfortherisk . . . 80
3.5 Choi eofthe oe ients
c(~λ)
andproofoftheire ien y . . . . . 813.6 Evaluation ofthe onstantinthespeedof onvergen eandnalresult 85 3.7 Con lusion . . . 87
4 Clean positive operatorvalued measures 89 4.1 Introdu tion . . . 89
4.2 Denitions andnotations. . . 91
4.3 AlgorithmandIdeas . . . 92
Algorithm,92.
•
Heuristi s: whatthealgorithmreallytests,93. 4.4 Su ient ondition . . . 954.5 Ne essary onditionforquasi-qubitPOVMs . . . 99
4.6 Summaryforquasi-qubitPOVMsandaspe ial ase . . . 110
5 Complementarysubalgebras 113
5.1 Introdu tion . . . 113
5.2 Preliminaries . . . 114
5.3 Complementarysubalgebras. . . 115
II Quantum Lo al Asymptoti Normality 121
6 Quantum lo alasymptoti normalityfor qubits 123
6.1 Introdu tion . . . 124
6.2 Lo alasymptoti normalityin statisti sand itsextension to quantum
me hani s . . . 128
6.3 Thebigballpi tureof oherentspinstates . . . 130
6.4 Lo alasymptoti normalityformixedqubit states. . . 133
Blo kde omposition,134.
•
Irredu iblerepresentationsofSU (2)
,136.6.5 Constru tionofthe hannels
T n
. . . . . . . . . . . . . . 1376.6 Constru tionoftheinverse hannel
S n
. . . . . . . . . . . . 1426.7 Appli ations . . . 143
Lo alasymptoti equivalen eoftheoptimalBayesianmeasurementand the
heterodyne measurement, 143.
•
The optimal Bayes measurement is also lo ally asymptoti minimax, 146.•
Dis rimination of states, 152.•
Spinsqueezedstatesand ontinuoustimemeasurements,154.
7 Optimalestimationof qubitstates with ontinuoustime
measurements 155
7.1 Introdu tion . . . 156
7.2 Stateestimation . . . 160
7.3 Lo alasymptoti normality . . . 164
Introdu tiontoLANandsomedenitions,165.
•
Convergen etotheGaus- sianmodel,166.7.4 Timeevolutionoftheintera tingsystem . . . 170
Quantumsto hasti dierentialequations,170.
•
SolvingtheQSDEfor theos illator,172.
•
QSDEforlargespin,173.7.5 These ondstage measurement. . . 175
Theheterodynemeasurement,176.
•
Energymeasurement,177.7.6 Asymptoti optimalityoftheestimator. . . 178
7.7 Con lusions. . . 183
7.A Appendix: Proof ofTheorem7.3.1 . . . 185
ProofofTheorem7.3.1;themap
T n
,185.•
ProofofTheorem7.3.1;themapS n
,189.7.B Appendix: Proof ofTheorem7.4.1 . . . 190
8 Quantum lo alasymptoti normalityfor
d
-dimensionalstates 197 8.1 Introdu tion . . . 1978.2 Lo alasymptoti normalityforqubits . . . 199
8.3 Classi alandquantumstatisti alexperiments . . . 201
Classi alandquantumrandomizations,202.
•
TheLeCamdistan eanditsstatisti almeaning,205.
8.4 Lo alasymptoti normalityinstatisti s . . . 207
8.5 Lo alasymptoti normalityinquantumstatisti s . . . 209
QuantumGaussianshift experiment,210.
•
Symmetri Fo kspa es, 210.•
Fo kspa es,211.
•
Gaussianstates,212.•
Maintheorem,215.8.6 Grouptheoryprimer . . . 216
Irredu ible unitary representations, 216.
•
Irredu ible representations ofSU (d)
,219.•
Tensorprodu trepresentation,224.8.7 Parametrisationofthedensitymatri esand onstru tionofthe
hannels
T n
. . . . . . . . . . . . . . . . . . . . . . 230Thenite-dimensionalexperiment,230.
•
Des riptionofT n
,231.8.8 Mainstepsoftheproof . . . 233
Why
T n
doesthework,233.•
DenitionofS n
andproofofitse ien y,239.8.9 (Evenmore)te hni alproofs . . . 241
A few more tools, 241.
•
Proof of Lemma8.7.1, 255.•
Proof of Lemmas8.6.9and8.7.2andworkaroundsfornon-orthogonalityissues,256.
•
ProofofLemma8.8.4,259.
•
ProofofLemma8.8.2,262.•
ProofofLemma8.8.1andLemma8.8.8,263.
•
ProofofLemma8.8.3,269.•
ProofofLemma8.8.5,271.Bibliography 287
Samenvatting 289
Résumé 293
Curri ulumVitae 297
Introdu tion
Statisti sisthes ien e ofpullinginformationoutofdata. Thoughthey anbe
wildlypolymorphi ,anystatisti alproblemmaybesplitintothree omponents:
theobje twestudy,theoperationsweareallowedtouse,andtheexa tmathe-
mati alquestion. Inotherwords,what wehave,what we ando, andwhat we
wantto know.
Quantum statisti sdiverge from lassi al statisti son the rst point, what we
have. Hen e theydieralsoonwhatisallowed,sin ethetwoarelinked.
In lassi al statisti s, we often immediately start from the result of measure-
ments, whi h are modeled by randomvariables with probability laws. Indeed,
ifwe anmeasure quantity AorquantityB,we antheoreti allymeasure both
simultaneously. Experiments often measure every useful and easily a essible
quantity. In theory, what we ando is applyinganymathemati al treatment
onthedata to transformit. Mathemati ally,this meansapplying anyfun tion
onthedata,possiblywitharandomout ome. Inpra ti e, omputationalpower
mightbound su h latitude.
In some ases, however, wemust already onsider the obje tunder study, and
hoose what measurement we arry out. A typi alexamplewould be tryingto
understandwhatabla kboxdoes. Wemustprobeitwithinputs,andea htime
wemust hoosetheinput. Thisthemati is alleddesign ofexperiments. What
we andomaydependhugelyontheproblemathand. Inthebla kbox ase,we
an hoosethe input. The mathemati aldes riptionof this hoi e mightdier
from one bla k box to another, though. Yet, on e the measurement is arried
out,weagainhaveprobabilitylawsandweareba ktothepreviousparagraph.
In quantum statisti s, the design of experiments annot be avoided. Indeed,
measuring A and B, in general. We must then hoose the measurement that
yields the information we need most. Nevertheless, quantum physi s gives a
framework paralleling that of lassi alprobability,whi h tellsus exa tly what
we ando. Initially,whatwearegiven isaquantumobje t,whi hismodeled
byaquantumstate. Whatwe ando ismeasuringthestate,gettinga lassi al
randomvariableas aresult,ormoregenerallytransformingthequantum state.
The sets of both measurements and transformations have pre ise and general
mathemati aldenitions,allowingtotreatmanyquestionsin auniedway.
Whatwewanttoknowseldomdiersinquantumand lassi alstatisti s. Most
often,wewanteithertosummarizetheinformationinthedata(statisti alinfer-
en e),to disproveahypothesis ortosee whathypothesisinaniteset best ts
thedata (testing),orto guess with pre isionwhat theunderlying phenomenon
wasthatgeneratedthedata(estimation). Allthese anusuallybedes ribedbya
lassi alparameter. Theex eptionwouldbewhenourben hmarkisintrinsi ally
quantum,forexamplewhentryingtoapproximately loneaquantumstate.
Thisthesis,inPartI,studiesanumberofparti ularsystems. Namelywe onsider
in Chapter 2 how to best de ide in whi h state among a nite set a quantum
obje t anbe;in Chapter3,wegiveafast(
1/n
)pro edureto estimateabla kboxunitarytransformation. Chapter4andChapter5dwellmoreonthegeneral
stru ture of quantum experiments: the former deals with an order relation on
measurements, and the latter on ndingmaximally dierent subsystemsof a
quantumsystem,inthesimplest ase.
Now,wemayhaveverydierentquestionsonagivensystem. Forsu hasystem,
orexperiment,whatwehave and whatwe ando will remainthesame. We
maythenwonderaboutwhatwe ansaydire tlyonthesystem,withoutreferen e
to aparti ular question. Thetheory of onvergen e of experimentsin lassi al
statisti sworksouthowwellwe anapproximateanexperimentbyanother. We
an then translate all the pro edures we use in one experiment to the other.
Hen e we get answers to what we want to know in both experiments when
solvingthequestionin one.
PartII,themain ontributionofthisthesis,generalizestothequantumworldthe
mostbasi aseof onvergen eofexperiments,namelylo al asymptoti normal-
ity. We provethat asu iently smoothexperimentwith identi alindependent
(i.i.d.) quantum states onverge to aquantum Gaussian shift experiment. The
pointisthatthisexperimentisverywell-known,andeverythingweknowabout
it anbetranslatedtothelarge lassofsmoothi.i.d. experiments.
Theremainder ofthe introdu tion rst makespre ise the rulesof lassi aland
quantumstatisti s,andthenintrodu eea hofthe haptersofthethesis,andthe
1.1 Statisti s
1.1.1 Classi al Statisti s
LeCam[1986℄andvanderVaart[1998℄maybe onsultedforfurtherreferen es,
amongmanyotherbooksonstatisti s. Wesummarizein Table1.1,onpage24,
the mostbasi ingredientsof lassi alstatisti s. The sister Table 1.2 givesthe
orrespondingquantumnotions.
What we have
In lassi alstatisti s,wearegivendata,that anbemodeledasarandomvariable
X
with probability lawp
. What we knowbeforehandis thatp
isa probability lawinasetE = {p θ , θ ∈ Θ} ,
(1.1)with no onstraint in general on the parameter set
Θ
. Thep θ
are all denedon the sameprobability spa e
(Ω, A)
. ThisE
is alled the experiment or thestatisti al model.
Remarks:
•
Thedataareoftenmadeofmanymeasurements,yieldingasmanyrandom variablesX 1 , . . . , X n
,withprobabilitylawsp 1 , . . . , p n
onpotentiallydier- entprobabilityspa es. However,wemaystill onsiderallthedataasasin-glerandomvariable
X = (X 1 , . . . , X n )
withprobabilitylawp = p 1 ⊗· · ·⊗p n
,andwestayinthe urrentframework.
•
Although there is no onstraint onΘ
at this pointof the theory, this setis often either nite orareasonablesubset of
R d
. Therst aseleads to dis retestatisti s,andsomefamilies oftestsinparti ular,these ond aseto parametri statisti s. When theset
Θ
is innite-dimensional, weenter the omplexrealmofnon-parametri statisti s,themain fo usofresear hin re entyears.
Examples: Bernoulli experiment,Gaussian shiftexperiment
The mostbasi probabilityspa e wemay nd is the two-elementspa e
{0, 1}
.Anexperiment orrespondingtoa ointosswouldbe
E Ber = {p θ = (θ, 1 − θ), θ ∈ [0, 1]} .
(1.2)Alternatively, we mighttoss the oin
n
times. DenotingX = (X 1 , . . . , X n )
theresults,wewouldgetthisexperimenton
{0, 1} ⊗n
:E Bin = n
p θ : {X} 7→ θ P X i (1 − θ) n−P X i , θ ∈ [0, 1] o
.
(1.3)Whendealingwith ontinuousfun tions, themostpervading ofthem allis the
Gaussian. Weareespe iallyinterestedinGaussian shiftexperiments,wherethe
varian eoftheGaussianisxed andtheparameteristhemean:
E gs =
N (θ, I −1 ), θ ∈ R d
,
(1.4)where
N
meansnormallaw,andI
isanyxedpositivematrix1.What we an do
On ewehaveourdata
X
,how anwepro essthem?Themostgeneralpro edure onsistsin drawinganew randomvariable
Y
withprobabilitylaw
p X
dependingonlyonX
,andmeasurableasafun tion ofX
.We anviewthisproto olintwoways. Therstis onsideringthat
Y
isananswerto what wewantto know. Then
Y
is a(randomized) estimator, typi allyan estimatorofθ
,inwhi h asewealsodenoteitbyθ ˆ
.Alternatively,we an onsiderthat
Y
isanewrandomvariable,andthatwehavetransformedourexperiment. Ournewexperiment onsistsof
Y
withprobability lawq
in theset{q θ , θ ∈ Θ}
onaspa e(Ω 1 , B)
,withdensity2q θ (y) = T (p θ )(y) ˆ = Z
Ω
p X (y)dp θ (X).
(1.5)Thetransformation
T
isaMarkov kernel.Inthe lassi al ase,thetwonotionsarethesame. However,Iinsistonseparating
themsin etheywillbedierentinthequantum ase.
1
Weusethisstrangenotationbe ause thismatrixistheinverseoftheFisherinformation
matrix(1.13).
2
We ouldequivalentlyworkwithnon-dominatedsetsofprobabilitylaws,butthatwould
onlymakenotationsheavier. Wethenassumethatallprobabilitylawshaveadensity,anduse
Examples
Letusgoba ktoour
n
-sampleBernoulliexperimentE Bin
(1.3). Ourprobability spa eis{0, 1} ⊗n
. WemayuseaMarkovkernelfromthatspa eto[0, n] ∩ N
thatsimply send
X = (X 1 , . . . , X n )
toY = P
X i
. Here, thep X
are merely deltafun tions. Wethenobtainabinomialprobabilitylawfor
Y
,thatisq θ = B(n, θ)
.The orrespondingexperimentis
E = {q θ , θ ∈ Θ}
.Alternatively, we might want to build an estimator
θ ˆ
. The most obvious onewould be
X 7→ P
X i /n = Y
. The law of our estimatoris the abovebinomialdividedby
n
.Wemightalsolookforanestimatorin
E gs
(1.4). Therstthoughtisyetsimpler:wejustkeep
X
. The orrespondingMarkovkernelwouldbetheidentity.What we wantto know
We usually want to haveinformation on the unknown underlying pro ess that
gaverisetoourdata. Inotherwords,wewanttoguesstheparameter 3
θ
.We angiveananswereither with a onden einterval,orwithaguess ofour
quantity, maybe with estimates on the varian e of the estimate. This guess
orrespondstogivinganestimator
θ ˆ
ofθ
.Wewanttobuildagoodestimator. Wethereforeneedawaytorateestimators.
Inde isiontheory,we onsidera ostfun tion
c(θ, ˆ θ)
. Thatisthe ostwehavetopayifourestimatoryields
θ ˆ
whenthetrueparameterisθ
. Hen e, ostfun tionsare usually zero on the diagonal, and grow when
θ
andθ ˆ
get farther apart insomesense.
Atypi al ostfun tionwhen
Θ
isdis reteand ountablewouldbec(θ, ˆ θ) = δ θ, ˆ θ
.When
Θ
isanopensubsetofR d
,themostmathemati allytra table ostfun tion is the square of the Eu lidean distan ec(θ, ˆ θ) = kθ − ˆθk 2 2
, or more generallyany quadrati ostfun tion
(θ − ˆθ) ⊤ G(θ − ˆθ)
for apositivematrixG
, possiblydepending on
θ
.Sin e
θ ˆ
isarandom variable, we wantto minimizethe expe tation of the ost, alledtheriskatpointθ
:r θ (ˆ θ) = Z
Ω 1
c(θ, ˆ θ)dq θ (ˆ θ).
(1.6)3
Moregenerally,wemaybeinterestedmerelyinafun tion
f
ofθ
.However,we analwaysuse
(θ, f (θ))
asparameter. Wethen hoosethe ostfun tions introdu ed belowsothat theydependonlyon
f(θ)
.However,we annot dire tly minimizethis expression, sin ethe best guess de-
pendson
θ
, whi h is unknown. Wemust thennd away to hoosean e ientestimatorforany
θ
wearelikelytoen ounter. Therearemainlytwoapproa hes.A favourite of physi ists is theBayesianparadigm, where we assumethe exis-
ten eof ana priori probabilitylaw ontheparameter
θ
. Mathemati ians often preferminimax riteria,whereastrategyisratedbytheworst ase.Bayesian riteria
Wehave onsideredourdatatobe
X
withprobabilitylawp
. Weassumedthattheonlyinformationwehad wastheexperiment,thesetweknow
p
belongsto.Suppose nowthat we havemore information. Namely, weare told beforehand
that
θ
is hosenatrandomwithaprobabilitylawπ
. Then,onaverage,thebestestimatorwouldbetheonethat minimizestheaverageoftherisk(1.6),that is:
R π (ˆ θ) = Z
Θ
π(dθ)r θ (ˆ θ)
= Z
Θ
Z
Ω 1
c(θ, ˆ θ)dq θ (ˆ θ)π(dθ).
(1.7)Fromthe Bayesriskofaspe i estimator
θ ˆ
,we anwrite theBayesriskasso-iatedtotheprior
π
astheinmumoftherisksforallθ ˆ
:R π = inf
θ ˆ
R π (θ).
(1.8)Theweaknessofthisapproa histhatthereisnoreasonwhythereshouldbean
a priori probability lawon
Θ
,ex ept adelta fun tion on therealθ
... whi h isexa tlywhatwewanttoknow. Wehaveto hooseapriorand onsideritasthe
realone. Theriskofthenal estimatorwillbeunderestimated,however.
Themain strengthofaBayesianestimatoristheoptimaluseoftheinformation
we get from measurements, given the prior. The prior orresponds to a priori
information, whi h is generally wrong. The best priors try then to minimize
theinformation in theprior 4
. Foranite
Θ
,weusually hooseequiprobability a priori forea h possibleθ
. For anopen pre ompa tsubset ofR d
, we hoose Jereys [1946℄prior, proportionalto thesquare root of theFisher information(1.13)denedbelow. Apointwiseanalysisshowsthattheseestimatorsareoften
verygoodestimators.
4
Subje tiveBayesians onsidertheprobabilitylawsasdegreesofbelief. Hen ethey anuse
Bayesian estimators an be omputed through the al ulations of a posteriori
distributions. Insomesimple ases, these anbe arriedoutexpli itly andthe
estimatoristhebary enterofthe
θ
withweightsthelikelihoods. Inmore omplex situations,we anresorttoMonte-CarloMarkov hains.Minimax riteria
Themathemati ianiseitherpessimisti ormegalomania ,andassumesheplays
againsttheDevil. Therefore,hewantstodesignastrategythatwillbee ient
whateverthereal
θ
is. Hen e theben hmarkofanestimatorθ ˆ
isitsvalueintheworst ase:
R M (ˆ θ) = sup
θ
r θ (ˆ θ).
(1.9)Theminimaxriskistheriskofthebest possibleestimator:
R M = inf
θ ˆ
R M (ˆ θ) = inf
θ ˆ
sup
θ
r θ (ˆ θ).
(1.10)The weakness of this method is that we might have to worsen mu h an esti-
mator on intuitively many
θ
for it to bee ienton somespe ial ases. Theworkaroundistorequireadaptiveness,thatis,minimaxe ien yonawhole lass
of subsets of
{p θ }
. The latter te hnique is essentially used for non-parametri statisti s.Theinterestofthesemethodsisthat theyrequirenoassumption. Theygivean
e ien yweknowweattaininreality,aslongastheexperiment(ormodel)itself
wasright.
Linksbetween Bayesian and minimax riteria
The main link between the two riteria omes from the following remark. Ifa
strategy
θ ˆ
is Bayesoptimal, and su h that the riskofθ ˆ
doesnotdepend onθ
,then
θ ˆ
isalsominimaxoptimal.Indeed,forany
π
,theBayesriskofθ
ismorethantheminimaxrisk:R π (ˆ θ) ≤ sup
θ
r θ (ˆ θ) = R M (θ),
(1.11)withequalityifandonlyiftheriskat
θ
isthesameπ
-almost everywhere.Under some onditions, a onverse statement is true: a minimax estimator is
optimalfor somepre iseprior, theoneforwhi h theBayesianriskis maximal.
Example
We ompute the risk of the aforementioned estimator for the Gaussian shift
family(1.4). Thelawof
θ ˆ
isthelawoftheoriginaldata,thatis thenormallawN (θ, I −1 )
. Sothatr θ (ˆ θ) = E θ
h (θ − ˆθ) ⊤ G(θ − ˆθ) i
= Tr(G I −1 ).
(1.12)Thisriskatpoint
θ
doesnotdependonθ
,sothatthesamevalueistheminimaxriskandtheBayesianriskforanyprioroftheestimator. Weshallseebelowthat
theestimatorisminimaxforthemodel.
Theremainderofthese tiongivesaqui ksummaryofwhatriskswe anexpe t
inregularenough ases,forquadrati ostfun tions.
Fisherinformation
Theriskswegiveabovedepend onthequestion(the ostfun tion) and onthe
experiment
{p θ , θ ∈ Θ}
,butnotonanyparti ularestimator. Wemaythenreadinformationaboutthemdire tlyontheexperiment.
ThemostimportantnotiontothatendistheFisher informationmatrix. Itisa
lo alnotion,that anbeinterpretedasameasureofhowfastwe andistinguish
p θ
from the surroundingp θ+dθ
. The Cramér-Rao bound des ribed in thenextse tionmakesthatexpli it. Noti ethatinthefollowing,weneedsomeregularity
inthemodel. Twi edierentiableismorethanenough.
TheFisherinformationatpoint
θ = (θ α ) α=1...d
isgivenbyI α,β (θ) = Z
Ω
∂ ln(p θ (X))
∂θ α
∂ ln(p θ (X))
∂θ β
dp θ (X).
(1.13)The Fisher information matrix is positive denite, and denes a metri on
Θ
,whi hisinvariantbyanysmooth hangeofvariables. Thisfa t anbeviewedas
themostbasi onne tionbetweenstatisti sanddierentialgeometry. Dieren-
tialgeometry anbeused to study higher-orderasymptoti s,asexemplied by
Amari[1985℄.
Developingthelogarithmsofprodu ts,itiseasilyseenthat having
n
samplesofthedatamultipliestheFisherinformationby
n
,thatisI (n) (θ) = n I (1) (θ)
whereI (n)
istheFisherinformationmatrixoftheexperimentE (n) = {p ⊗n θ , θ ∈ Θ}
.Cramér-Raobound
We anusetheFisherinformationmatrixtoderivealowerboundonthevarian e
matrixoflo allyestimators:
Z
Ω 1
(θ − ˆθ)(θ − ˆθ) ⊤ dq θ (ˆ θ) ≥ I −1 (θ).
(1.14)The bound holds 5
for all lo ally unbiased estimators
θ ˆ
, that is as long asR ˆ θdq θ (ˆ θ) = θ
and∂/∂θ i R ˆ θ j dq θ (ˆ θ) = δ i,j
.Animmediate onsequen eisthat,forlo allyunbiasedestimators,andaquadrati
ostfun tion
(θ − ˆθ) ⊤ G(θ − ˆθ)
, wegetthislowerbound ontheriskatpointθ
:r θ (ˆ θ) ≥ Tr(GI −1 ).
(1.15)Thisboundisknowntobeasymptoti allysharp. Indeed,a
n
-sampleexperimentin reasingly resembles aGaussian shift experiment, forwhi h it is sharp. The
pre iseexplanation omesfrom thetheoryof onvergen e ofexperimentsbyLe
Cam,thatwefurthersket hinSe tion1.6.1.
Examples
We ompute the Fisher information for the Bernoulli experiment, at point
θ
dierent from
0
and1
. Theexpression isslightlyeasier sin ewehaveonlyoneparameter.
I(θ) = θ
d ln(θ) dθ
2
+ (1 − θ)
d ln(1 − θ) dθ
2
= 1 θ + 1
1 − θ
= 1
θ(1 − θ) .
Fromthatandourpreviousremarkfor
n
samples,weseethatI(θ) = n/(θ(1−θ))
inthebinomialexperiment
E bin
.Aslightlymoretedious al ulationwouldshowthat theFisherinformationma-
trixofaGaussianshiftexperimentistheinverseofthevarian eoftheGaussians.
5
Supere ient estimators su h as Stein estimator provethat we annot simplydrop the
unbiasedness ondition. However, addingsomete hni ality(essentially onsideringe ien y
on a whole neighborhood of
θ
, through either a Bayesianor a minimaxapproa h), we anHen e our hoi e ofnotationin equation (1.4). Moreover,after omparison be-
tweenthebound(1.15)andtherisk(1.12)oftheestimator onsistingin
X
itself,weobtainoptimalityofthelatterestimatoramongthe lassoflo ally unbiased
estimators.
Wenowtryto givetheequivalentsofthosenotionsin thequantum world.
1.1.2 Quantum Obje ts and Operations
The books by Helstrom [1976℄ and Holevo [1982℄ are the usual referen es for
quantum statisti s. We also add the more re ent review arti le by Barndor-
Nielsenet al.[2003℄. Asalreadymentionned,wehavesummarized inTable1.2,
onpage25, themostbasi ingredientsof quantum statisti s,withTable1.1for
lassi al orrespondan eonthepagebefore.
States,density operators
Thebasi obje tinquantumprobabilityisthestate. Thestateistheequivalent
ofaprobabilitylaw.
Wedene it overaHilbert spa e
H
. Its mathemati alexpression isgivenbya densityoperator.Denition 1.1.1. A densityoperator
ρ
overa Hilbert spa eH
isa tra e- lass operator withthe followingproperties:•
Self-adjointness:ρ
isself-adjoint.•
Positivity:ρ
isnon-negative.•
Normalization:Tr(ρ) = 1
.Thosearetheequivalentof onditionsforprobabilitymeasures: probabilitymea-
sures are real (
=
self-adjointness), non-negative (=
positivity) and normalized to1
(=
normalization).For nite-dimensional Hilbert spa es, the operators are matri es, and density
matri esalsosatisfytheabove onditions. Therealdimensionofthemanifoldof
statesis
d 2 − 1
ifthe omplexdimensionofH
isd
.Example: Qubits
Themostelementary situationarises when
dim( H) = 2
. Physi ally,thesystem ouldbeanele tronspin. Those statesare alled qubit statesand heavilyusedinquantuminformation.
WedenePaulimatri es as
σ x =
0 1 1 0
, σ y =
0 i
−i 0
, σ z =
1 0 0 −1
.
(1.16)Self-adjointness implies that adensity matrix must be a linear ombination of
those matri esand theidentity
1
. Positivity andnormalization further impose that:ρ = 1 2
1 + ~ θ · ~σ
, k~θk ≤ 1,
(1.17)with
~σ = (σ x , σ y , σ z )
ave torofmatri es.We see that we already need three real parameters to des ribe a qubit state,
onfer the one parameterwe need to des ribe aprobability law on a lassi al
two-out omespa e.
Pure states
Thesetof lassi alprobabilitymeasures anbeseenasthe onvexhullof delta
fun tions. Similarly,theset ofstatesisthe onvexhullofpurestates.
Purestatesare hara terizedby beingrank-one operators,with eigenvalueone.
We anwritethem
|ψi hψ|
,where|ψi
isanorm-oneve torofH
. Purestates anthusberepresentedaspointsoftheproje tivespa easso iatedto
H
.Theyareveryimportant: manytreatmentsofquantum me hani sfeature only
purestates. Generalstates anbeseenasa lassi almixingofpurestates.
Unlike for delta fun tions, where we merely draw a random variable with the
unknown law, there is no measurement that an identify unambiguously any
purestate,evenifweknowbeforehandthatthestateispure. Thisfundamental
dieren e with the lassi alworldis a hallmarkof non- ommutativitybetween
dierentstates. Thestudyofpurestatesinthemselvesisalready hallenging.
For qubits with the above parameterization, the pure states orrespond
to
k~θk = 1
. Thisparameterizationby asphere, alledthe Blo h sphere,givesa graphi alintuitionforproblemsonqubits.Therealdimensionofthepurestatesis
2(d − 1)
ifdim H = d
.Example: Coherentstates
Qubits aretheparadigmfor nite-dimensionalquantum states. Theother fun-
damentalfamilyofstatesisthat of oherentstates 6
.
ThosestatesliveontheFo kspa e
7
F(C)
,thatistheinnite-dimensionalHilbert spa eℓ 2 (N)
. We denote{|ki} k∈N
the anoni albasis onℓ 2 (N)
. Physi ists all|ki
thek
-thFo kstate.StatesonFo kspa esarestatesoftheharmoni os illator,anexampleofwhi h
is the state of mono hromati light (laser). We are thus on the playground of
quantumopti s. Amongthosestates, oherentstatesarein somewaythemost
lassi al: theysaturateHeisenbergun ertaintyrelations.
Theyaregivenbyone omplex,hen etworeal, oe ient
θ
. Sin etheyarepurestates,we andes ribethemwithave torin
F(C)
,ratherthananoperator8:|θ) = exp(−|θ| 2 /2) X n k=0
θ k
√ k! |ki .
(1.18)Multipartite states,entangled states
Letus onsidertwoquantumobje ts
ρ 1
andρ 2
onH 1
andH 2
. They anbeseenasasinglequantumobje ton
H = H 1 ⊗ H 2
,withstateρ = ρ 1 ⊗ ρ 2
.Any state on su h omposite Hilbert spa e is alled a multipartite state. Now
somemultipartite states annot be writtenas
P c i ρ i 1 ⊗ ρ i 2
withpositivec i
. Wemightneedsomenegative
c i
. Inotherwords,thosestatesarenota lassi alran-domizationofa hoi eofapairofstates. They ontainanintrinsi allyquantum
oupling. Theyare alledentangledstates.
Let us provethey do exist. We write
dim H 1 = d 1
anddim H 2 = d 2
. Hen edim H = d 1 d 2
. PuremultipartitestatesarepurestatesonH
,sothey onstitutea
2(d 1 d 2 − 1)
manifold. Ontheotherhand,apurestateoftheformP
c i ρ i 1 ⊗ ρ i 2
withpositive
c i
onlyallowoneterminthesum,withbothρ 1
andρ 2
purestates.The orrespondingdimensionis
2(d 1 + d 2 − 2) < 2(d 1 d 2 − 1)
. Hen e there aremanyentangledpurestates.
6
Moregenerally,allpossiblysqueezedGaussianstatesplayanimportantroleinquantum
opti sand,asweshallsee,inquantumstatisti s. Westi kto oherentstatesforsimpli ityof
theexample.
7
Multidimensional oherentstatesaretensorprodu tsof oherentstatesonthetensorized
Fo kspa e
F (C d ) = F (C) ⊗d
.8
Weusethe notation
|θ)
instead of the usualket|θi
soas to avoid onfusionwithFo kstates,inparti ularwhen
θ
happenstobeapositiveinteger.A typi al example are maximally entangled states, that is states of the form
|Ψi hΨ|
, with|Ψi = √ 1 d P ψ i
⊗ ψ i
, where
H 1 = H 2
and{ ψ i
}
is an or-thonormalbasisof
H 1
. Astheir nameimply,they arry asmu h entanglement aspossible.Entanglementmaybethesinglemostbasi and pervasiveresour ein quantum
information. Itliesat theheartof quantum teleportation,mostquantum ryp-
tographyproto olsandthein reasedpro essingpowerofaquantum omputer.
Literatureonthesubje tistoodauntingtobeevens rat hedupon. Inquantum
statisti s, apart from the problems linked to estimating entangled states, they
anbeusedtospeedupestimationofquantumtransformations.
A tions on states
In the lassi al ase, we noti ed that givingan estimator of a parameter
θ
ormoregenerallyofanyfun tionof
θ
wasthesameastransformingourinitialdata togetanewrandomvariableY
withlawT (p θ )
.Inthequantum ase,thetwonotionsaredistin t. Indeed,transformingthedata
meansgettinganewquantumstate,thatisanoperatoronaHilbertspa e. States
undergoatransformationwhentheyaresentthrougha hannel. Anestimatorof
a lassi alparameter,ontheotherhand,isa lassi alquantity. Wethenendup
with a lassi alrandomvariable. We retrievethis lassi aldata from thestate
throughameasurement.
Ifwemerelywantto onsiderestimators,whyarewealsointerestedin hannels?
Indeed,applyingmany hannelsandthenameasurement anbesummedupto
usingonlyamore omplexmeasurement.
Therstreasonisthat wemighttransformourstatestoanewfamilyforwhi h
weknowwhatmeasurementtouse. Infa t,thewholeaimofstronglo alasymp-
toti normality, whosestudy onstitutesmostof this thesis, is to transforman
experimenttoaquasi-equivalentandeasierone.
Se ondly, hannels des ribephysi al transformations. We mightwant to study
thetransformationitselfrather thanthestate. Typi ally,thephysi al transfor-
mation ouldbe generated bya for e we wantto measure. Wedwell on these
mattersinChapter3.
We allinstrument afun tionyielding lassi alandquantumdataoutofaquan-
tum input. Real measurementapparatuses are essentiallyinstruments, even if
we may forget about the out ome state. In parti ular, ontinuous-time mea-
surementsare ommon in pra ti e. Typi ally, we measure the ele tromagneti
seen asa sequen e of innitesimal instruments, and writing the orresponding
evolutionequationsisthepurposeofquantumltering,pioneeredbyDaviesand
Belavkin[Boutenetal.,2006,foranintrodu tion℄.
Measurements, POVMs
Ifwewanttomake lassi alstatisti alinferen eontheunknownparameters,we
havetotranslateourquantuminformationto lassi alinformation. Tothatend,
we apply ameasurement. Sin e mixed states are lassi almixing of states, we
requirelinearityofthetransformation. Theout omeshouldalwaysbea lassi al
probabilitylaw. Wededu e from that thefollowingform of physi ally allowed
measurements:
Denition1.1.2. Apositiveoperatorvaluedmeasure,or POVM,overa mea-
suredspa e
(Ω, A)
isaset{M(A)} A∈A
of boundedoperatorsonH
su hthat:• M(Ω) = 1 H
.• M(A)
ispositive.•
For any ountable olle tion(A i ) i∈N
of disjointA i
, we haveM ( S A i ) = P M (A i )
.We noti e that those are exa tly the usual axioms for a probability measure,
ex eptthatweworkwithoperatorsinsteadofrealnumbers. We allea h
M (A)
aPOVMelement.
Applyingameasurement
M
onastateρ
yields aprobabilitylawP ρ
on(Ω, A)
,givenbyBorn'srule:
P ρ (A) = Tr(ρM (A)).
(1.19)InChapter4,wes rutinize aspe i orderrelationonPOVMs.
Afewremarksareinorder.Firstofall,we anin ludeany lassi alpro essingof
thedatainthePOVM.Indeed,applyingameasurement
M
andthenaMarkovkernel
T
(denedby(1.5))ontheoutputrandomvariableisthesameasapplyingthemeasurement
N
on(Ω 1 , B)
withN (B) = R
Ω p ω (B)M (dω)
. SothatworkingonPOVMsisequivalenttoworking onestimators.
Se ondly, we annotingeneralmeasuresimultaneously
M 1
andM 2
on(Ω 1 , A 1 )
and
(Ω 2 , A 2 )
. In ontrastto the lassi al ase,where we ouldhavesimultane-ously the results of applying
T 1
andT 2
. Indeed, measuring bothM 1
andM 2
meansmeasuring
N
on(Ω 1 × Ω 2 )
withN (A 1 × Ω 2 ) = M 1 (A 1 )
andN (Ω 1 × A 2 ) =
M 2 (A 2 )
. An easy ounterexample illustratingthe role of non- ommutativity is givenbyM 1
andM 2
bothdened on{0, 1}
,withM 1 (0) =
1 0 0 0
, M 1 (1) =
0 0 0 1
, M 2 (0) = 1
2
1 1 1 1
, M 2 (1) = 1 2
1 −1
−1 1
.
Allthosematri esarerank-one. Wewouldnowneed
N (0, 0) + N (0, 1) = M 1 (0)
.Sin eallPOVMelementsarepositive,wehave
M 1 (0) ≥ N(0, 0)
. Sin emoreoverM 1 (0)
is rank-one, we haveN (0, 0) = c 1 M 1 (0)
for some0 ≤ c 1 ≤ 1
. We alsoknow
N (0, 0) + N (1, 0) = M 2 (0)
. SothatN (0, 0) = c 2 M 2 (0)
. Theonlysolutionis
c 1 = c 2 = 0
andN (0, 0) = 0
. ThesameholdsforN (0, 1)
,N (1, 0)
andN (1, 1)
.Ontheotherhandweneed
N ( {0, 1} 2 ) = 1 C 2
. Contradi tion.Finally,all those measurementsare believedto bephysi ally feasible. However
theymightbeveryhardtoimplementin pra ti e. Inparti ular,ifthestateisa
multipartite state,it anmakesense to restri tourattention tosmaller lasses
ofmeasurements. Notably,ifdierentpeoplehold dierentparti lesindierent
pla es, they annot implement ageneral measurement, even if they ooperate.
The best they an do is: one of them measures his parti le (possibly with a
non-trivialoutput quantum state), tells the resultto the other, who hoosesa
measurement onhis parti le, keeps theoutput state and tellsthe resultto the
rstone,andtheyiterateontheoutputstates. Su h measurements,usingonly
lo alquantumoperationsand lassi al ommuni ation,aredubbedLOCC:Lo al
Operations,Classi alCommuni ation.
Inquantum information whenthe(usuallyentangled)quantum stateisdivided
betweenseveralpeople,wenaturallyrestri tto LOCCmeasurements. Inquan-
tum estimation of a state with
n
opies of theinitial state, weare at least in-terestedin what anbea hievedthroughLOCCmeasurements,mu h easierto
implementthangeneral( olle tive)measurements. We aningeneralreallygain
pre isionwith olle tivemeasurements. Thismightbesurprisingfromthepoint
ofviewofphysi ists,sin ethe
n
opiesaretotallyindependent. Insome ases,no- tablywhenweknowthattheunknownstateispure[Matsumoto,2002℄, olle tivemeasurementsdonotyieldmu h improvementoverLOCCmeasurements. This
mightbesurprisingfromthepointofviewofmathemati ians,sin ethespa eof
olle tivemeasurementsismu hbiggerthanthat ofLOCCmeasurements.
Example: Spin
z
Considerthebinaryout omemeasurementonqubits givenby
M ( ↑) =
1 0 0 0
= 1
2 (1 + σ z ), M ( ↓) =
0 0 0 1
= 1
2 (1 − σ z ).
Thismeasurementappliedto thestate
ρ = 1+~ 2 θ·~σ
yields↑
with probabilityTr(ρM ( ↑)) = 1 2
Tr(1M ( ↑)) + X
α=x,y,z
θ α Tr(σ α M ( ↑))
= 1
2 (1 + θ z ).
Inparti ular,if
θ z = 1
,thentheout omeisalways↑
. Conversely,ifθ z = −1
,theout omeisalways
↓
. Ontheotherhand,ifθ x = 1
,sothatθ z = 0
,theout omeiseither
↑
or↓
withprobabilityonehalf,eventhoughthestateρ
ispure.This kind of measurements, where all the POVM elements are proje tors, are
also alledobservables. Theyonlyyieldinformationonthebasisinwhi hallthe
POVM elementsarediagonal. Noti e that usualaxiomsofquantum me hani s
restri tmeasurementsto observables. However,weget ba kallthe POVMsby
applyinganobservableonamultipartite stateofwhi hourstateis onlyapart
(Naimarktheorem).
Heterodyne measurement
Theheterodynemeasurementgetsitsnamefromthete hniqueusedtoimplement
itin laboratory, withlasersthat are o-phase. This POVM without omein
C
hasamathemati alexpressiongivenby:
M (A) = 1 π
Z
A |z)(z|dz,
(1.20)where
|z)
isa oherentstate(1.18).Theprobabilitylawoftheout omewhenmeasuring
ρ
hasthus adensity(z |ρ|z)
with respe t to Lebesgue at point
z
. Inparti ular, the lawof the result when measuringa oherentstateisaGaussian:q θ (dz) = 1
π (z |θ)(θ|z) = 1
π exp( −|θ − z| 2 ).
(1.21)Ifwe onsiderallthe omplex
θ
,were ognizea lassi alGaussianshiftexperiment(1.4)in
R 2
.Moregenerally,theprobabilitydensityfun tionof theout omeof themeasure-
mentonastate
ρ
is alled theHusimi fun tion ofthestate:H ρ (dz) = 1
π (z |ρ|z).
(1.22)Channels
We now des ribe how to make anew quantum state out of the original state.
Noti ethattherststateisdestroyedin thepro ess.
A physi al transformationof aquantum obje ttakesastate andyield another
state,possiblyonadierentspa e. It isdes ribedbya hannel, theequivalent
ofaMarkovkernel.
We re all that a positive superoperator
E
is amap su h that for any positiveoperator
A
, theoutputE(A)
isalsopositive.Denition 1.1.3. A hannel
E
is amap from the setT (H 1 )
of tra e- lassop- eratorstoT (H 2 )
,withthe following properties:•
Linearity:E
islinear.•
Completepositiveness: foranyauxiliaryspa eH 3
,thesuperoperatorE⊗Id : T (H 1 ⊗ H 3 ) → T (H 2 ⊗ H 3 )
givenby( E ⊗ Id)(ρ ⊗ σ) = E(ρ) ⊗ σ
ispositive.•
Tra e-preserving:Tr( E(A)) = Tr(A)
.Noti ethatMarkovkernelssatisfyallthese riteria,whenrepla ingoperatorsby
measures 9
.
Thene essity of linearity an be provedfrom the axiomof unitaryevolution 10
andin ludingtheobserverin thesystem.
Wewanttheimageofastatetobeastate, soapositiveoperatormustbesent
toapositiveoperator. Tounderstandwhyweneed ompletepositivity,wemust
onsider apossiblyentangledstate on
H 1 ⊗ H 3
. If we transformstatesonH 1
,wealsotransformstateson
H 1 ⊗ H 3
,withE ⊗ Id
asthe hannel. Thereforethelattertransformationmustbepositive. Hen eweneed ompletepositivity.
Finally,the output isastate ifthe inputis astate,and bothare tra e-one,so
tra emustbepreserved.
Weoften onsider the hannelsin the(pre)dualpi ture,thatisasa tingonthe
elementsof
B(H)
. SothatTr( E(ρ)A) = Tr(ρE ∗ (A))
forallstateρ
andallboundedoperator
A
. Inthis aseE ∗
isalsoa ompletelypositivelinearmap,butwemust9
Inthe moregeneral settingof
C ∗
-algebras,the spa esoffun tions are ommutativeC ∗
-algebrasandallpositivesuperoperatoronthosespa esis ompletelypositive.
10
Quantumme hani sstatethattheevolutionofasystemisgivenby
ρ(t) = U (t)ρ(0)U ∗ (t)
,where
U(t)
isaunitaryoperatorthat anbe omputedfromtheself-adjointoperatorH
alledtheHamiltonian.IftheHamiltoniandoesnotdependontime,then
U(t) = e itH
.repla ethetra e-preserving onditionbytheidentity-preserving ondition,that
is
E ∗ (1) = 1
.Notations: Weusuallywrite
E
orF
for hannels. Abusingnotations,weusuallydrop the star for the pre-dual and also write
E
in that ase. However, thosestandardnotationsare alsothe standardnotationsfor experiments. Sothat in
the hapterswhere weuse that notion,weusefor hannels thesamenotations
asforMarkovkernels,thatis
T, T n , S, S n
.Kraus representation,Stinespring theorem
Theabovedenitiondoesnotmakeitobvioustodealwith hannels. Fortunately,
tworepresentationtheoremsdes ribe ompletelypositivemapsinamoreusable
way. ThebookbyPaulsen[1987℄isagoodreferen eonthosematters.
Kraus[1983℄representationisthemain toolwhentheHilbert spa esare nite-
dimensional.
Theorem1.1.4. A ompletely positive map
E
fromM (C d 1 )
toM (C d 2 )
an bewrittenas
E(A) = X
α
R α AR α ∗ ,
(1.23)with
α
runningfrom1
toatmostd 1 d 2
,andR α ∈ M d 2 ,d 1 (C)
. Staristheadjoint.Moreover, the hannel istra e-preservingifandonly if
P R ∗ α R α = 1 C d1
.Thede ompositionisnotunique. Thedual hannelisgivenby
A 7→ P
R ∗ α AR α
.Ininnitedimension,weratherusethemorepowerfulStinespring[1955℄dilation
theorem 11
.
Theorem 1.1.5. Let
E : B(H 1 ) → B(H 2 )
be a ompletely positive map. ThenthereisaHilbertspa e
K
anda*-homomorphism(orrepresentation)π : B(H 1 ) → B(H 2 )
su hthatE(A) = V π(A)V ∗ ,
(1.24)where
V : K → H
isaboundedoperator.Moreover, if
E
isidentity-preserving, thenV
isanisometry, thatisV V ∗ = 1 H
.If we further impose that
K
is the losed linear span ofπ(A)V ∗ H
, then thedilationisuniqueuptounitarytransformations.
11
Infa t,Stinespringtheoremwasprovedforanyunital
C ∗
algebraasinitialspa e.It anbeshowntoimplyKrausrepresentation,butalsotheGNSrepresentation,astapleof
C ∗
-algebras.Instruments
We give the representation of instruments for nite dimensions 12
. To further
simplify notations, werestri tourselvesto the asewhenthe measurementhas
anitenumberofout omes.
Denition1.1.6. Aninstrument isgiven byaset
{N ω,k }
ofmatri esfromH 1
to
H 2
,su hthatX
ω
X
k
N ω,k ∗ N ω,k = 1 H 1 .
The orresponding measurementisgiven by
M (ω) = X
k
N ω,k ∗ N ω,k ,
andthe outputstatewhenthe resultof themeasurement is
ω
isgiven byN (ρ, ω) =
P
k N ω,k ρN ω,k ∗ Tr(ρM (ω)) .
The outputstatelives on
H 2
.Wenowhave anotherwayto understand whywe annot measure twoPOVMs
simultaneously: aftermeasuring
M
,thequantumobje t,thatisourdata,hasingeneralbeenperturbed. Infa t, ifthemeasurementisri henough, theoutput
statedependsonlyontheout ome
ω
,andnotanymoreontheinputstate.Wenowhaveallthetoolsto opythesetupfrom lassi alstatisti stoquantum
statisti s.
1.1.3 Quantum statisti s
Usually,weworkonquantumstates;o asionallywemaywanttogainknowledge
ona hannel. Wetreatthetwo asesseparately.
States: What we have,what we an do,what we wantto know
Inanalogywiththe lassi al ase,weareusually givenaquantum state
ρ
,thatweknowtobein aset
E = {ρ θ , θ ∈ Θ} .
(1.25)12
Ininnitedimension,wehavetousethe
C ∗
-algebrasettingandaninstrumentismerelyahannelbetween
C ∗
-algebras.Weagain all thissetanexperiment,oramodel.
With theexamplesof thequbits, the usualmodelswouldbethe3D full mixed
model
E m = {ρ θ , kθk < 1}
and the 2D pure state modelE p = {ρ θ , kθk = 1}
,where wehaveused ourformerparameterizationfor thestate
ρ θ
(1.17). Whenhaving
n
opiesofthestate,werepla eρ θ
byρ ⊗n θ
.Another typi al experiment would be
E t = {ρ θ , θ ∈ {θ 1 , θ 2 }}
, where the usualquestion is to dis riminate between the twopossible
θ
. We study this kind ofprobleminSe tion 1.2andChapter2.
We anapriori useanysequen eofinstrumentsonthestate. Ifwemerelywant
lassi alinformationon
θ
,wemayrestri ttomeasurementsM
,thatisPOVMs.We then asso iate to
M
an estimator, sayθ ˆ
, with law depending on the trueparameter
θ
throughq θ (B) ˆ = P θ
h θ ˆ ∈ B i
= Tr(ρ θ M (B)).
Dependingonthe ir umstan es,wemightallowanyphysi al measurement, or
asmaller lass,su hasseparateorLOCCmeasurement.
Finally, what we want to know is the same as in the lassi al ase. We want
to knowsomefun tion oftheparameter
θ
. Sothatwewanttoestimateθ
, andwe rate our estimator
θ ˆ
through a ost fun tionc(θ, ˆ θ)
. As before, the mostommon ostfun tionsare
(1 − δ θ, ˆ θ )
,iftheparametersetisnite,andquadratiostfun tions
(ˆ θ − θ) ⊤ G(ˆ θ − θ)
forapositivematrixG
,iftheparameterlivesonanopensubsetof
R d
. TheweightmatrixG
mightdependonθ
.We anagain write the risk (1.6)of an estimator at point
θ
. Sin e we donotknow
θ
, wethen either use the Bayesianrisk(1.7) for anappropriate prior, or the minimax risk (1.9), and optimize (1.8, 1.10)over the available estimators.Noti ethatthelaststagedependonthesetofallowedestimators.
QuantumFisher informationand Cramér-Raobounds
We antrytomimi thedenitionof lassi alFisherinformationandgetsimilar
bounds onvarian e of estimators. Infa t, we anbuild su h anequivalent for
any hoi eofalogarithmi derivative. We hoosetherightlogarithmi derivative
(RLD),dened forea h
θ
andea h oordinateθ α
asamatrixλ α,θ
su hthat:∂ρ θ
∂θ α = ρ θ λ α,θ
(1.26)onthesupport of