PeterHyer 1;?
,Mi heleMos a 2;??
, andRonalddeWolf 3;???
1
Dept.ofComputerS ien e,Univ.ofCalgary,Alberta,Canada.
hoyer ps .u algary. a
2
Dept.ofCombinatori s&Optimization,Univ.ofWaterloo,andPerimeter
InstituteforTheoreti alPhysi s,Ontario,Canada.mmos auwaterloo. a
3
CWI.Kruislaan413,1098SJ,Amsterdam,theNetherlands.rdewolf wi.nl
Abstra t. Suppose we have n algorithms, quantumor lassi al, ea h
omputing some bit-value with bounded error probability. We des ribe
a quantum algorithm that uses O(
p
n ) repetitions of the base algo-
rithmsandwithhighprobabilityndstheindexofa1-bitamongthese
n bits (if there is su han index). This shows that it is not ne essary
torstsigni antlyredu e theerrorprobabilityinthebase algorithms
toO(1=poly(n))(whi hwouldrequireO(
p
nlogn)repetitions intotal).
Ourte hniqueisare ursiveinterleavingofamplitudeampli ationand
error-redu tion,andmaybeofmoregeneralinterest.Essentially,itshows
thatquantumamplitudeampli ation anbemadetoworkalsowitha
bounded-error verier.Asa orollaryweobtainoptimalquantumupper
bounds of O(
p
N) queries for all onstant-depth AND-ORtrees on N
variables,improvinguponearlierupperboundsofO(
p
Npolylog(N)).
1 Introdu tion
Oneofthemainsu essesofquantum omputingisGrover'salgorithm[10,7℄.It
ansear hann-elementspa einO(
p
n)steps,whi hisquadrati allyfasterthan
any lassi alalgorithm.Thealgorithmassumesora lea ess totheelementsin
the spa e, meaning that in unit time it an de ide whether the ith element
is a solution to its sear h problem or not. In some more realisti settings we
aneÆ ientlymakesu hanora leourselves.Forinstan e,ifwewanttode ide
satisability of an m-variable Boolean formula, the sear h spa e is the set of
all n = 2 m
truth assignments, and we an eÆ iently de ide whether a given
assignmentsatises the formula. However, in these asesthe de ision is made
withoutanyerrorprobability.Inthispaperwestudythe omplexityofquantum
sear hifweonlyhavebounded-error a esstotheelementsinthespa e.
Morepre isely, supposethatamong nBooleanvaluesf
1
;:::;f
n
wewantto
ndasolution(ifoneexists),i.e.,anindexjsu hthatf
j
=1.Forea hiwehave
?
SupportedinpartbytheAlbertaIngenuityFundandthe Pa i Institutefor the
Mathemati alS ien es.
??
Supported by St. Jerome's University, the Canada Resear h Chair programme,
NSERC(CROandDis overyGrant),CFI,OIT,PREA,ORDCFandMITACS.
???
Thisresear hwas(partially)fundedbyproje tsQAIP(IST{1999{11234)andRESQ
i i i
is1thenthealgorithmoutputs1withprobability,say,atleast9=10,andiff
i
=0
thenitoutputs0withprobabilityatleast9=10.Grover'salgorithmisnolonger
appli ableinthisbounded-errorsetting,atleastnotdire tly,be ausetheerrors
inea hstepwillqui klyadduptosomethingun ontrollablylarge.A ordingly,
weneedtodosomethingdierenttogetaquantumsear halgorithmthatworks
here.Wewillmeasurethe omplexityofourquantum sear halgorithmsbythe
numberoftimesthey alltheunderlyingalgorithmsF
i
.Clearly,the( p
n)lower
bound for the standard error-lesssear h problem, due to Bennett, Bernstein,
Brassard,andVazirani[4℄,also appliestoourmoregeneralsetting. Ouraimis
togiveamat hingupperbound.
Anobviousbut sub-optimal quantum sear halgorithm isthe following.By
repeating F
i
k = O(logn) times and outputting the majority value of the k
out omes,we an omputef
i
witherrorprobabilityatmost1=100n.Ifwethen
opytheanswertoasafepla eandreversethe omputationto leanup(mostof)
theworkspa e,thenwegetsomethingthatissuÆ iently\ lose"toperfe tora le
a esstothef
i
bitstojusttreatitassu h.Nowwe anapplyGrover'salgorithm
ontopofthis, andbe ausequantum omputationalerrorsaddlinearly[5℄,the
overalldieren ewithperfe tora lea esswillbenegligiblysmall.Thissolves
thebounded-errorquantumsear hproblemusingO(
p
nlogn)repetitionsofthe
F
i
's,whi hisanO(logn)-fa torworsethanthelowerbound.Belowwewillrefer
tothisalgorithm as\thesimplesear halgorithm".
Arelativelystraightforwardimprovementoverthesimplesear halgorithmis
thefollowing.Partitionthesear hspa einton=log 2
nblo ksofsizelog 2
nea h.
Pi konesu hblo katrandom.We anndapotentialsolution(anindexjinthe
hosenblo ksu hthatf
j
=1,ifthereissu haj)in omplexityO(lognloglogn)
using the simplesear h algorithm, and then verifythat itis indeed 1 wither-
rorprobabilityatmost1=nusinganotherO(logn)invo ationsofF
j
.Applying
Groversear h onthespa eof alln=log 2
nblo ks,weobtainanalgorithmwith
omplexityO(
q
n=log 2
n)O(lognloglogn+logn)=O(
p
nloglogn).
A further improvement omes from doing the splitting re ursively:we an
use the improved upper bound to do the omputation of the \inner" blo ks,
insteadofthesimplesear halgorithm.UsingT(n)todenotethe omplexityon
sear hspa eof sizen,thisgivesusthere ursion
T(n)d
T(log 2
n) r
n
log 2
n
+logn
for some onstantd >0. This re ursion resolvesto omplexity O(
p
n log
n
)
for some onstant > 0.It is similar to (and inspired by) the ommuni ation
omplexityproto olforthedisjointnessproblemofHyeranddeWolf[11℄.
Apartfrombeingrathermessy,thisimprovedalgorithmisstillnotoptimal.
Themainresultofthispaperistogivearelatively leanalgorithmthatusesthe
optimalnumberO(
p
n)ofrepetitionstosolvethebounded-errorsear hproblem.
Ouralgorithmuses akindof\ arrot-and-sti k"approa hthat maybeofmore
i
bran hesin ludesolutions,buttheyalsoin lude\falsepositives":bran hes or-
responding to the 1=10 errorprobability of F
i
's where f
i
= 0.We then \push
these ba k"bytesting whethera1-bran hisarealpositiveorafalseone(i.e.,
whether f
i
=1ornot) andremoving mostof thefalseones.Interleavingthese
amplifyandpush-ba kstepsproperly,we anamplifytheweightofthesolutions
toa onstantusingO(
p
n)repetitions.Atthispointwejustdoameasurement,
seeapotentialsolutionj, andverifyit lassi allybyrunningF
j
afewtimes.
As anappli ation of ourbounded-error quantum sear h algorithm,in Se -
tion4wegiveoptimalquantumalgorithmsfor onstant-depthAND-ORtreesin
thequery omplexitysetting.Forany onstantd,weneedonlyO(
p
N)queries
forthed-levelAND-ORtree,improvingupontheearlierO(
p
N(logN) d 1
)algo-
rithmsofBuhrman,Cleve,andWidgerson[9℄.Mat hinglowerboundsof( p
N)
werealreadyshownforsu hAND-ORtrees,usingAmbainis'quantumadversary
method[1,2℄.Finally,inSe tion5weindi atehowtheideaspresentedhere an
be astmoregenerallyin termsofamplitudeampli ation.
2 Preliminaries
Herewebrie ysket hthebasi sandnotationofquantum omputation,referring
to thebook byNielsen andChuang[12℄ for moredetail.An m-qubit stateisa
linear ombinationofall lassi alm-bitstates
ji= X
i2f0;1g m
i jii;
wherejiidenotesthebasisstatei(a lassi alm-bitstring),theamplitude
i is
a omplexnumber,and P
i j
i j
2
=1.Weviewjiasa2 m
-dimensional olumn
ve tor.Ameasurementof statejiwill givejii withprobabilityj
i j
2
, andthe
statewillthen ollapsetotheobservedjii.Anon-measuringquantumoperation
orrespondstoapplyingaunitary(=linearandnorm-preserving)transformation
U to theve torof amplitudes.Ifjiand j i arequantumstateson mand m 0
qubits, respe tively,then the two-registerstatejij i=jij i orresponds
to the2 m+m
0
-dimensionalve torthatisthetensorprodu tofjiandj i.
Thesettingofquery omplexityisasfollows.Forinputx2f0;1g n
,aquery
orresponds tothe unitarytransformationO that maps ji;b;zi!ji;bx
i
;zi.
Here i 2 [n℄ and b 2 f0;1g; the z-part orresponds to the workspa e, whi h
is notae ted by thequery. A T-queryquantum algorithmhas theform A=
U
T OU
T 1
OU
1 OU
0
, wherethe U
k
areunitarytransformations,independent
of x. This A depends on x only via the T appli ations of O. The algorithm
startsin initialall-zerostatej0ianditsoutput(whi hisarandomvariable)is
In this se tion we des ribe our quantum algorithm for bounded-error sear h.
Thefollowingtwofa tsgeneralize,respe tively,theGroversear handtheerror-
redu tionusedinthealgorithmswesket hedin theintrodu tion.
Fa t1(Amplitude ampli ation[8℄) LetS
0
betheunitarythat putsa`-' in
front of the all-zero state j0i, and S
1
be the unitary that puts a`-' in front of
all basis states whose last qubit isj1i. Let Aj0i =sin()j
1
ij1i+ os()j
0 ij0i
whereangle is su hthat0=2 andsin 2
()equals theprobability thata
measurementofthelastregisterofstateAj0iyieldsa'1'.SetG= AS
0 A
1
S
1 .
Then GAj0 i=sin(3)j
1
ij1i+ os(3)j
0 ij0i.
Amplitude ampli ation is a pro ess that is used in many quantum algo-
rithms to in rease the su ess probability. Amplitude ampli ation ee tively
implementsarotationbyanangle2inatwo-dimensionalspa e(aspa edier-
entfromtheHilbertspa ea tedupon)spannedbyj
1
ij1iandj
0
ij0i.Notethat
we analwaysapplyamplitudeampli ationregardlessofwhether theangle
isknownto usornot.
Fa t2(Error-redu tion) Suppose Aj0i = p
pj
b ijbi+
p
1 pj
1 b
ij1 bi,
where b 2 f0;1g and p 9=10. Then using O(log(1=")) appli ations of A
and majority-voting, we an build a unitary E su h that Ej0i = p
qj
b ijbi+
p
1 qj
1 b
ij1 bi with q 1 ", and j
b=1 b
i possibly of larger dimension
thanj
b=1 b
i(be auseofextraworkspa e).
Wewillre ursivelyinterleavethese twofa ts to getaquantum sear halgo-
rithmthatsear hesthespa ef
1
;:::;f
n
2f0;1g.Weassumeea hf
i
is omputed
byunitaryF
i
withsu essprobabilityatleast9=10.Let =fj:f
j
=1gbethe
set ofsolutions,and t=j j itssize (whi h isunknown toour algorithm).The
goalistondanelementin ift1,andtooutput `nosolutions'ift=0.
Wewill buildanalgorithmthat hasasuperpositionof allj 2[n℄in itsrst
register,agrowingse ondregisterthat ontainsworkspa eandotherjunk,and
a 1-qubit third registerindi ating whether somethingis deemed a solution or
not.Thealgorithmwillsu essivelyin reasetheweightofthebasisstatesthat
simultaneouslyhaveasolutionintherstregisteranda1in thethird.
Consideranalgorithm Athat runs all F
i
on e in superposition, produ ing
thestateAj0i,whi hwerewriteas
1
p
n n
X
i=1 jii
p
p
i j
i;1 ij1i+
p
1 p
i j
i;0 ij0i
=sin()j
1
ij1i+ os()j
0 ij0i;
where p
i
is the probability that F
i
outputs 1, the states j
i;b
i des ribe the
workspa eoftheF
i
,andsin() 2
= P
n
i=1 p
i
9t=10n.
Theideaistoapplyaround ofamplitudeampli ationtoAtoamplifythe
j1i-partfromsin()tosin(3).Thiswillamplifyboththegoodstatesjjij1ifor
toredu etheamplitudeofthefalsepositives,setting\most"ofitsthirdregister
to 0.These twostepstogetherform anew algorithmthat puts almost 3times
asmu hamplitudeonthesolutionsasAdoes,andthat putslessamplitudeon
thefalsepositivesthanA.Wethenrepeattheamplify-redu estepsonthisnew
algorithmto getanevenbetteralgorithm,andsoon.
Letusbemorepre ise.Ouralgorithmwill onsistofanumberofrounds.In
roundk wewillhaveaunitaryA
k
thatprodu es
A
k j0i=
k j
k
ij1i+
k j
k ij1i+
q
1
2
k
2
k jH
k ij0i;
where
k
;
k
arenon-negativereals,j
k
iisaunitve torwhoserstregisteronly
ontains j 2 , j
k
i is aunit ve torwhose rst registeronly ontainsj 62 ,
andjH
k
iisaunit ve tor.Ifwemeasure therstregisteroftheabovestate,we
willseeasolution(i.e.somej2 )withprobabilityatleast 2
k .A
1
istheabove
algorithm A, whi h runs the F
i
in superposition. Initially, 2
1
9t=10nsin e
ea hsolution ontributesatleast9=10n.Wewanttomakethegoodamplitude
k
growbyafa torofalmost3inea hround.
Amplitude ampli ation step.Forea h roundk, dene
k
2[0;=2℄by
sin(
k )
2
= 2
k +
2
k
. Applying amplitude ampli ation (G
k
= A
k S
0 A
1
k S
1 )
givesusthestateG
k A
k
j0i,whi hwemaywriteas
sin(3
k )
sin(
k )
k j
k ij1i+
sin(3
k )
sin(
k )
k j
k ij1i+
s
1
sin(3
k )
sin(
k )
2
( 2
k +
2
k )jH
k ij0i:
WeappliedA
k
twi eandA 1
k
on e,sothe omplexitygoesupbyafa torof3.
Error-redu tionstep.Conditionalonthequbitin thethirdregisterbeing
1,theerror-redu tionstepE
k
nowdoesmajorityvotingonO(k)runs oftheF
j
(forallj insuperposition)tode idewitherroratmost1=2 k +5
whether f
j
=1.
Itaddsone0-qubitasthenewthirdregisterandmaps (ignoringitsworkspa e,
whi hisaddedtothese ondregister)
E
k
jjij1ij0i=a
jk
jjij1ij1i+ q
1 a 2
jk
jjij1ij0i
E
k
jjij0ij0i=jjij0ij0i
where a 2
jk
1 1=2 k +5
if f
j
=1 and a 2
jk
1=2 k +5
if f
j
= 0. This way, E
k
removesmostofthefalsepositives.
Putting A
k +1
=E
k G
k A
k
anddening
k +1 ,
k +1 ,j
k +1 i,j
k +1
i, andjH
k +1 i
appropriately,wenowhave
A
k +1 j0i=
k +1 j
k +1 ij1i+
k +1 j
k +1 ij1i+
q
1
2
2
jH
k +1 ij0i:
stepE
k
,aswellasbythequbitthatpreviouslywasthethirdregister.Thegood
amplitudehasgrowninthepro ess:
k +1
k sin(3
k )
sin(
k )
q
1 1=2 k +5
:
Sin ex x 3
=6sin(x)x,wehave
sin(3
k )
sin(
k )
3 9
2
k
=2:
A ordingly,aslongas
k
issmall,
k
willgrowbyafa torofalmost3inea h
round.Ontheotherhand,theweightofthefalsepositivesgoesdownrapidly:
k +1
k sin(3
k )
sin(
k )
1
p
2 k +5
:
We nowanalyze thenumberm of roundsthat weneed to makethe good am-
plitudelarge.Ingeneral,wehavesin(
k )
2
= 2
k +
2
k
,hen e 2
k
2(
2
k +
2
k )for
thedomainweareinterestedin.Here 2
k
9 k 1
2
1 and
2
k
1
10 (9=2
6
) k 1
.Note
m 1
X
k =1
2
k
2 m 1
X
k =1
2
k +
2
k
2 m 1
X
k =1 9
k 1
2
1 +2
m 1
X
k =1 1
10 (9=2
6
) k 1
29 m 1
2
1 +1=4:
Therefore,mroundsoftheabovepro essamplies thegoodamplitude
k to
m
1 m 1
Y
k =1 sin(3
k )
sin(
k )
q
1 1=2 k +5
1 m 1
Y
k =1
3 9
2
k
=2
1 1=2 k +5
=
1 3
m 1 m 1
Y
k =1
1 3
2
k
=2
1 1=2 k +5
1 3
m 1
1 3
2 m 1
X
k =1
2
k m 1
X
k =1 1
2 k +5
!
1 3
m 1
1 3
2 (29
m 1
2
1
+1=4) 1=16
1 3
m 1
1=2 39 m 1
2
:
[n=9 m+1
;n=9 m
℄,equivalently9 m
2[n=9t;n=t℄,thenwehave
1
3 m
p
10
r
9t
10n
1
r
t
n
1
3 m
:
Thisimplies
m
0:04;
so the probability of seeing a solution after m rounds is at least 0:0016. By
repeating this lassi ally a onstantnumber of times, say 1000 times, we an
bring the su ess probability lose to 1 (note to avoid onfusion: these 1000
repetitions arenotpartofthedenition ofA
m itself).
The omplexityC
k
oftheoperationA
k
,intermsofnumberofrepetitionsof
theF
i
algorithms,isgivenbythere ursion
C
1
=1andC
k +1
=3C
k
+O(k);
where the 3C
k
is the ost of amplitude ampli ation and O(k) is the ost of
error-redu tion.This impliesC
m
=O(
P
m 1
k =1 k3
m k 1
)=O(3 m
):
Wenowgivethefullalgorithmwhenthenumberofsolutionsisunknown:
Algorithm:Quantumsear h on bounded-errorinputs
1. form=0todlog
9
(n)e 1do:
(a) runA
m
1000times
(b) verifythe1000measurementresults,ea hbyO(logn)runsofthe orre-
sponding F
j
( ) ifasolutionhasbeenfound,thenoutputasolutionandstop
2. Output`nosolutions'
Thisndsasolutionwithhighprobabilityifoneexists.The omplexityis
dlog
9 (n)e 1
X
m=0
1000O(3 m
)+1000O(logn)=O(3 log
9 (n)
)=O(
p
n):
Ifweknowthatthereisatleastonesolutionbutwedon'tknowhowmanythere
are,then,usingamodi ationofouralgorithmasin[7℄,we anndasolution
usinganexpe tednumberofrepetitionsinO(
p
N=t),wheretisthe(unknown)
numberofsolutions.Thisisquadrati allyfasterthan lassi ally,andoptimalfor
anyquantumalgorithm.
4 Optimal Upper Bounds for AND-OR Trees
A d-levelAND-OR tree on N Boolean variables is a Boolean fun tion that is
inputvariable,andiff
1
;:::;f
n
allared-levelAND-ORtreesonmvariables,ea h
withanAND(resp.OR) asroot,thenOR(f
1
;:::;f
n
)(resp.AND) isa(d+1)-
levelAND-ORtreeonN =nmvariables.AND-ORtrees anbe onvertedeasily
intoOR-ANDtreesandvi eversausingDeMorgan'slaws,ifweallownegations
to beaddedtothetree.
Considerthetwo-leveltree onN =n 2
variableswith anORasroot,ANDs
asits hildren,andfanoutninbothlevels.Ea hAND-subtree anbequantum
omputedbyGrover'salgorithm withone-sidederrorusingO(
p
n)queries(we
letGroversear hfora`0',andoutput1ifwedon'tnd any),andthevalueof
theOR-ANDtreeisjusttheORofthosenvalues.A ordingly,the onstru tion
ofthepreviousse tiongivesanO(
p
n p
n)=O(
p
N)algorithmwithtwo-sided
error.Thisisoptimalupto a onstantfa tor[1℄.
Moregenerally,ford-levelAND-ORtreeswe anapplytheabovealgorithm
re ursivelyto obtainanalgorithm withO(
d 1 p
N)queries.Here isthe on-
stant hidden in the O() of the result of the previous se tion. For ea h xed
d, this omplexity is O(
p
N), whi h is optimal up to a onstant fa tor [2℄. It
improvesupontheO(
p
N(logN) d 1
)algorithmgivenin [9℄.
Our query omplexity upper bound also implies that the minimal degree
among N-variatepolynomialsapproximatingAND-OR isO(
p
N)[3℄.Whether
thisupperbound onthedegreeisoptimalremainsopen.Thebest knownlower
boundforthe2-level aseis(N 1=4
p
logN)[13℄.
5 Amplitude Ampli ation with Imperfe t Verier
Inthisse tionweviewour onstru tioninamoregenerallight.
Suppose weare givensome lassi alrandomizedalgorithmA that su eeds
insolvingsomeproblemwithprobabilityp.Inaddition,wearegivenaBoolean
fun tionthattakesasinputanoutputfromalgorithmA,andoutputswhether
itisasolutionornot.Then,wemayndasolutiontoourproblembyrepetition.
Werst apply algorithmA, obtainingsome andidate solution,whi hwethen
giveasinputtotheverier.Ifoutputsthatthe andidateindeedisasolution,
weoutput it and stop, and otherwise we repeat the pro ess by reapplying A.
Theprobabilitythatthispro essterminatesbyoutputtingasolutionwithinthe
rst(
1
p
)iterationsoftheloop,islowerboundedbya onstant.
A quantum analogueof boosting the probability of su ess is to boost the
amplitude ofbeingina ertainsubspa eofaHilbertspa e.Thusfar,amplitude
ampli ation[6℄hasassumedthatwearegivenaperfe tverier: whenevera
andidatesolutionisfound,we andeterminewith ertaintywhetheritisasolu-
tionornot.Formally,wemodelthisbylettingbe omputedbyadeterministi
lassi alsubroutineoranexa tquantumsubroutine.
Themain resultofthispapermaybeviewedasanadaptationofamplitude
ampli ation to the situation where the verier is not perfe t, but sometimes
tum subroutine,wearegivenabounded-errorquantumsubroutine. Previously,
the onlyknown te hniqueforhandlingsu h aseshasbeenby straightforward
simulationofaperfe t verier: onstru tasubroutinefor omputing wither-
ror 1
2 k
byrepeatingagivenbounded-errorsubroutine oforder (k)times and
thenusemajorityvoting.Usingsu hdire tsimulations,wemay onstru tgood
but sub-optimal quantum algorithms, liketheO(
p
nlogn)queryalgorithm for
quantumsear hoftheintrodu tion.Here,wehaveintrodu edamodi ationof
the amplitude ampli ationpro ess that allows us to eÆ iently dealwith im-
perfe tveriers.Essentially,ourresultsaysthatimperfe t veriersareasgood
asperfe tveriers(up toa onstantmultipli ativefa torinthe omplexity).
A knowledgments
We thankRi hard Clevefor usefuldis ussions, aswellasfor hostingMM and
RdWattheUniversityofCalgary,wheremostofthisworkwasdone.
Referen es
1. A. Ambainis. Quantumlowerbounds byquantumarguments. InPro eedings of
32ndACMSTOC,pages636{643,2000.
2. H. Barnum and M. Saks. A lower boundon the quantumquery omplexity of
read-on efun tions. quant-ph/0201007,3Jan2002.
3. R. Beals, H. Buhrman, R. Cleve, M. Mos a, and R. de Wolf. Quantum lower
boundsbypolynomials. InPro eedingsof39thIEEEFOCS,pages352{361,1998.
4. C.Bennett,E.Bernstein,G.Brassard,andU.Vazirani. Strengthsandweaknesses
ofquantum omputing. SIAMJournalonComputing,26(5):1510{1523, 1997.
5. E. Bernstein and U. Vazirani. Quantum omplexity theory. SIAM Journal on
Computing,26(5):1411{1473, 1997.
6. G.Brassardand P.Hyer. Anexa tquantumpolynomial-time algorithmfor Si-
mon'sproblem.InPro eedingsofFifthIsraeliSymposiumonTheoryofComputing
andSystems(ISTCS'97),pages12{23,1997.
7. M.Boyer,G.Brassard,P.Hyer,andA.Tapp. Tightboundsonquantumsear h-
ing. Forts hrittederPhysik,46(4{5):493{505, 1998.
8. G.Brassard,P.Hyer,M.Mos a,andA.Tapp.Quantumamplitudeampli ation
andestimation. InLomona o, S.J.,Jr.andBrandt,H.E.(eds.):QuantumCom-
putation andQuantumInformation:A MillenniumVolume.AMSContemporary
Mathemati sSeries,305:53{74,2002.
9. H.Buhrman,R.Cleve,andA.Wigderson. Quantumvs. lassi al ommuni ation
and omputation. InPro eedingsof30thACMSTOC,pages63{68,1998.
10. L. K. Grover. A fast quantum me hani al algorithm for database sear h. In
Pro eedingsof28thACMSTOC,pages212{219,1996.
11. P.HyerandR.deWolf.Improvedquantum ommuni ation omplexityboundsfor
disjointnessandequality.InPro eedingsof19thAnnualSymposiumonTheoreti al
Aspe ts ofComputer S ien e(STACS'2002), Le ture NotesinComputer S ien e,
Vol.2285,pages299{310.Springer-Verlag,2002.
12. M.A.NielsenandI.L.Chuang.QuantumComputationandQuantumInformation.
manus ript,2002.