You are on page 1of 61

Strengths and Weaknesses

Charles H.Bennett

IBM Research

7
9

9
1

Gilles Brassard

Universit

e de Montr

eal

To appear inSIAM Journal


(special issue

on Computing

on quantum

computing)

of Quantum Computing
Ethan Bernstein

Microsoft Corporation

Umesh Vazirani

UC Berkeley

12 December 1996

Abstract

Recently

a sequence

a great

deal of attention has focused

on quantum

computation following

of results [4,16, 15] suggesting that quantum computers

are more

powerful

than classical probabilistic computers. Following Shors result that factoring and the

are both solvable

inquantum polynomial time, it is

extraction of discrete logarithms

natural to ask whether allof NP can be effi ciently solved inquantum polynomial time.

u
q

Inthis

paper, we address

this question by proving that relative to an oracle chosen

uniformly at random, with probability 1,the class NP cannot be solved

Turing machine intime o(2

n/2

chosen uniformly

with probability

ona quantum

).We also show that relative to a permutation oracle

at random,

1,the class
NP
coNP
chosen
uniformly

at

r
random,
cannot be
with probability
a

1,
solved
the class
onaNP
quantum

Turing machine intime o(2

n/3

).The former

bound is tight since

recent work of Grover [13]


shows how to accept the class NP relative to any

oracle

on
a quantum computer

O(2

n/2

intime

).

IBM T.J. Watson Research Laboratory,

Yorktown Heights, New York, NY 10598, USA. email:


bennetc@watson.ibm.com.

1
Microsoft Way, Redmond, WA 980526399,

USA. email: ethanb@microsoft.com.

Supported inpart by Canadas

nserc

and

Qu ebecs fcar.

D epartement IRO, Universit

6128, succursale centre-ville,

e de Montr

eal, C.P.

Montr eal (Qu ebec),

Canada H3C 3J7. email: brassard@iro.umontreal.ca.

Supported by NSF Grant No. CCR-9310214.


Computer Science Division, University

of

California, Berkeley, CA 94720, USA. email:


vazirani@cs.berkeley.edu.

Introduction

Quantum

exciting

computational

new area

foun-dations

of

that

both

complexity
touches

theoretical

is

upon

an

the

computer

science

and quantum

eighties,

Feynman

on a

require

In the early

time;

he

quantum

defined

general

computation:

appear to

overhead

in the size of the


asked

that

the

computer.

quantum

is

system and the


whether

model

that

quantum

of

this

is

to design

inherent, and whether it is possible


universal

out

pointed

clas-sical computer

simulation

exponential
simulated

[12]

simulations

straightforward

mechanics

physics

Deutsch

of

[9]

quantum

Turing

machine.

Bernstein and Vazirani [4] proved that there is

an effi cient

universal quantum Turing machine.

Yao [17] extended

this by proving

that

quan-

tum circuits (introduced by Deutsch [10])


polynomially

equivalent

to quantum

are

Turing

machines.
The

computational

Turing machines
by

several

Deutsch
exploit

researchers.

and

some

power

(QTMs)

Jozsa

Early

[11]

inherently

of

quantum

has been explored


work

showed

quantum

how

by

to

mechanical

features

conj

of

unction

Berthiaume

Their

QTMs.

subse-quent

with

and Brassard

the existence

computational problems

we

require

classical

to produce

machine

results

by

which there

can

that QTMs

time with

in

[5, 6], established

of oracles under

in polynomial

results,

cer-tainty,

solve

whereas

probabilistic

if

Turing

answer

the correct

are

with

certainty, then it must take exponential time

on some

inputs. On the other

computational

problems

to the

relative

same

are

oracle,

hand, these

in
and

analogue

class BQP

there-fore

sense.

The

of the class BPP is

the

effi ciently solvable in the classical

quantum

[5]. Bernstein

and Vazirani

proved that BPP BQP PSPACE,


establishing

that

conclusively

that

BQP

6=

[4]

thus

it will not be possible

prove

BPP

to

BPP

le
to conclus
ively the
provmajor
e tha t
BQP
6= BPP
without
resolving
open
problem
P
?

= PSPACE.
that

BQP

They also

6=

gave

BPP

quantum Turing machines

the first evidence


(polynomial-time

are more

powerful

than
quantum
polynomial-time
Tu ring ma chin
probabilistic
es are mo Turing
re po

werful
machines),
than

by

proving

the

existence

are

oracle rela-tive to which there


BQP that cannot

by

restricted

to running

probabilistic

as

BPP is regarded

problems)

this

powerful

than

of

an

even

cannot

are

simulated

run

machines

allowed

to

Simons

paper

important

new

that

Turing
1/3

Simon

[16]

by

proving

the

to which BQP
by

for

also

probabilistic
n/2

steps.

introduced

In

an

technique

BPP is the class of decision problems

can

in

oracle relative

be

that

more

computers

way.
evidence

addition,

evidence
inherently

classical

model-independent

existence

machines

steps. Since

(computational

provided

computers

this

in

error

the class of all effi ciently

quantum

strengthened

o(log n)

in

languages

computable

problems

be solved with small

probability

an

of

(languages)

be solved in polynomial time by probabilistic

machines

(for

all

with

error

inputs).

probability

Using

bounded

standard

by

boosting

techniques,

error

can

then be made

small in k by iterating

the algorithm

the

exponentially

probability

answer.

k times and returning the majority


2

is

BQP

(languages)

that

the

can

quantum

Turing

bounded

by

1/3

as

is the

with

prove

problems

error

probability

see

[4] for

with BPP, the

error

paper

in Section 4 of this

can be made

of BQP machines

decision

all inputs)

(for

We

case

of

be solved in polynomial time by

machines

formal definition.

that,

class

probability

exponentially

small.

2
which

was

remarkable
Shor

one
result

logarithm

problems

have

presumed

BQP;

i.e.

for the factoring


These

been well-studied,
forms

cryptography.

these results, it is natural

by

polynomial-time

problems

intractability

much of modern

in

subsequently

gave

algorithms

discrete

ingredients

proved

[15]. Shor

quantum

the

of

the

and

two

and their
basis

of

In view

of

to ask whether NP

can quantum
to ask
computers
whether

solve
NP

NPcomplete
In this

problems in polynomial time?

paper, we

address

by proving that relative to

this question

an oracle

chosen

at random [3], with probability

uniformly

the class NP cannot be solved


Turing machine in time o(2

that relative

at

uniformly

to

on a quantum

n/2).

permutation

random,

with

1,

We also show

oracle chosen

probability

the
uniforml
class NP
y at
ran
coNP
dom,cannot
wit h proba
be solved
bilityon
1
n/3

the
a quantum
class NP
Turing
co N
machine
Pca nnot
in time
be solv
o(2
ed on
)
The
a
former bound is tight since recent work
of

Grover

NP

[13] shows

relative

computer

in

to

any

time

how to accept

O(2

the class

on a quantum

oracle

n/2).

See

[7]

for

detailed analysis of Grovers algorithm.


What
results?

is the

relevance

We should emphasize

not rule out the possibility


What
is

of

no

black-box

uniquely

approach

problems

oracle

that they do

that NP BQP.

these results do establish

NPcomplete

these

by

quantum- mechanical

is that there

to

solving

using

features

some

of

QTMs. That

was a

this

real possibility

is

clear from Grovers

[13] result, which gives

black-box

to solving

approach

as

in square-root

problems

NPcomplete

much time

as

is

required classically.
One

way to

think

special subroutine

costs

unit

no

has
that

any

bits

residual

can

calls

pose a

leaves

as a
only

of QTMs,

special problem that

must not leave around

beyond its computed


computational

information
achieved

any

the

open

whether

do not interfere. This is

deterministic

subroutine

to show

answer

the

BQP

more

computation

so

remain.

that only the

However,

general

machine

can

subroutines

any

BQP

this

question

be used

Our final result in this

how

because

with different

for deterministic

classical

and

answer,

paths

be carried out reversibly

input

is

classical counterpart. The problem is

otherwise

since

oracle

call whose invocation

the subroutine

easily

an

In the context

time

subroutine

of

machine

of

as a

paper

can

is

be

into

modified

tidy

final superposition consists

tape configuration
and the single

subroutines,
BQP

answer.

can

this

these

The result

tidy

used

safely

us to

show

as

that

also justifies

the definition of oracle quantum machines

we now

just the input

Since

be

allows

= BQP.

BQP

whose

almost entirely of

containing

bit

machines

BQP

machine

BQP

that

give.

it is not

Actually

even

clear whether

BQP

NP

BPPAct ;
ually
i.e. it
it is
is not
unclear
ev en
whether
clear wh
nondeterminism
eth er BQP

together
BPP

with

quantum
Vazi-ranis

They

BQP, but cannot

games

[1] with

evidence

does not
Fourier

even

a time

unrelativized

stated

an

can

and

above.

oracle, the

be solved in

be solved by Arthur-Merlin

bound of

cannot

o(logn)

on top

They conjecture

sampling

to

problem

that nondeterminism

help).

than

that relative

sampling

to simulate

In fact, Bernstein

is stronger

proved

Fourier

is suffi cient

machines.

[4] result

actually

recursive

randomness

Turing

even

(thus giving

of probabilism

that the recursive

be

polynomial-time hierarchy.

solved

in the

Oracle Quantum Turing

Machines
In this

assume
Turing

section

without
machine

and

the

loss of generality
alphabet

(for

is {0, 1, #}, where

tape)

we

next,

shall

that

each

track

denotes

the

or
the

track
blank or
symbol.
tap e)is
Initially
{0,1,#},
all
where
tapes
# are
d enotes
blank

the
except
blank
that
symbol.
the input tape contains the actual
input surrounded

by blanks. We shall

use

to

denote {0,1}.
In the classical setting,

described informally

some

Boolean

function

arguments,

arbitrary

an

as a device
A

at

oracle

may

be

for evaluating

unit

cost

on

per

,
evaluation.
on ar bitrThis
ar y argu
allows
ments,
to formulate
at unit cost
questions
pe re
such

as

if A

Turing

languages)

Turing

were

machine,
could

ma-chines?

effi ciently computable

which other functions

be effi ciently

computed

In the quantum

by

(or

by

setting,

an equivalent
we

quantum

oracle

define

can be

question

sectionand

provided

An oracle

for

single

may

rules

QTM has

block

of

oracle

allow

are

blank except

non-blank

this

region

qq

and

executed

Oracle

a post-query state qa

whenever

no-op
directly

change.

to

can be written

grow

string

in the form

b , and denotes

a pre-query
A

query

enters

machine

post-query

query

into

have

state

passes

no

with

is nonempty,

xb where

is

the

string is empty,

the

and

QTMs

machine

query

and

the

If the

the

If the

occurs,

In

machine

it from fragmenting

blocks.

pre-query state.

cells.

to

two distinguished internal states:

state

we

query tape

special

QTM, the Turing

shrink, but prevent


non-contiguous

this

quantum

paper.

all of whose cells

well-formed

in

be composedwhich

show inSection 4 of this

(or track)

do

machines

bounded-error

can

Turing machines

Turing

we

appropriatelywhich

asked, provided

it

concatenation.
xb w he re
In that
x

case,

the result

of

call

on

oracle

A is that

that
internal
c ase,
control
the resul
passes
t o fto
a call
the on
post-query
while the contents

of the

query

state

tape changes

from |x bi to |x (b A(x))i, tape


where
ch ang

denotes
from
|x
the
exclusive-or
bi t o |x

Except
for the
the
de notes

(addition
(b
A(x)
modulo
)i,where
2)

tape
exquery
clusive -or
(adand
ditioninternal

control, other parts of the oracle QTM do not


change

during the

is supplied

state

will

query.

in initial
be

|A(x)i,

If the target

bit
do
|bi
n

state |0i, then its

as

just

in

final

classical

final
oracle
sta
machine.
te will bConversely,
e |A(x) i,just
if the
as in
target
a clasbit
sica
is

already
orac le in
mastate
chi ne.
|A(x)i,
C onverse
calling
ly,
the
if the
oracle
target
will
b

reset
it tody
|0i.
ability
tocalling
uncompute
it
is
alrea
in sThis
tate |A(x
)i,
the ora

reset
willwill
often

it to |0
i.
This ability
to un
co
prove
essential
to allow
proper

mput
interference
e will

among

computation

take place. Using this fact, it is also


that

the

above

machines yields

ourselves

that

in particular

they enter the

the post-query

of

to

easy to see

oracle

Turing

unitary evolutions if we restrict

to machines

other respects,

as

definition

paths

are

pre-query

state.

well-formed

in

evolving unitarily

state and leave

power

The

their

from

superpo-sition
oracle

of quantum

machines

ability

follow

of computation

quantum

from the

to

ability

computers

comes

coherent

paths. Similarly

derive

great

power

to perform superpositions

of

queries. For example, oracle A might be called


Px

when the

query tape

x |x0i,

where
query
x tape
are is
complex
in state coeffi
|0i
cients,

is in state |0i

corresponding
|x 0i,w heto
re an
x arbitrary
queries

with

a constant

an

to thisarbitrary
In
case,
4

This

generality

allowing

superposition of

restriction
and

superposition

|0i in the target

it

can

can

only machines

be made

be

verified

that make

break the rule before writing

of

bit

quer
without

loss

syntactically

sure

of

by

they do not

on the query tape.

4
after the

query,

the

query

string will be left

Px

inthe entangled state

x |x A(x)i.

It is

Px

also useful to be able to put the target bit b


into

superposition.

ditional

phase

For example,

inversion

used

the

con-

in Grovers

can

algorithm

be

achieved

by performing
2.

with the target bit b in the nonclassical


2.

queries

superposition
with
th e tar
(|0i
get
bit
|1i)/
b inth enon
It clas
can

readily be

an

verified that

query tape

in state

oracle call with the

leaves

the

machine
with the state,
qu ery ta
including
pe in sta the
te x
query
lea
unchanged

state

= 0, and

if A(x)

unchanged

while

It is often convenient

on

as

(x, i)

This

as

defining

interpreting

is

th

the i

to think of

easily

the oracle

any

interpreted

as a
n

an

by

the pair

as a binary

value

string

function.

oracle which, when

length-preserving

function,
n

length-pr
eservi ngon
fun
ctio
0 as
a permutation

n,ac
Henceforth,
ts fo r
when
each no
n
mutation
shall
use
onA(x)

length-preserving
oracle

function

accomplished

answer on

pairing

oracle is

acts for each

phase

a Boolean

bit of the function

standard

permutation

length-preserving

The pair (x,i) is encoded


using

leaves the entire

introducing

= 1.

factor 1 if A(x)

oracle

entire

ves
tape,
th

A rather

confusion
0 as
amay
per arise,

to

denote

function
than

we

the

associated

the Boolean

with

function

that gives rise to it.

us

Let

define

sets of languages
at least 2/3 by
running

time

on

bound

by

whose

T(n).

This

time applies

to each

on

average.

just

the

or not

upon

might depend
be

the

is

the oracle

a BQP-machine

while

might not be one.

Note:
oracle

The
for

above

an

suffi ce for the

but

not

as

with probability

bounded

might

oracle QTM M

whether

BQP-machine

is

input,

that

Athus

accepted

some

the running

individual
Notice

BQTime(T(n))

the

perform

suggests

that

broader

perform

hiding

a quantum
function

of the present

unitary

will

paper,

computers

to

transformations

definition,

which

may

be

contexts. For example, oracles

more

operations

computational

of

Boolean

of quantum

general

useful in other

unitary

purposes

ability

definition

arbitrary

learning

information

general,

have

been
theory

against

non-Boolean
considered
[8]

classical

and

in

for

queries

[14].
Most broadly,
defined

as a device
unitary

fixed

current

a quantum

oracle

may

be

that, when called, applies

transformation

to

the

contents
a fixed u
|zi
nitar
of
y trthe
ansfor
query
ma tion
tape,
U

o
replacing
the c urr
it en
by U|zi.
t contents
Such an
|zi
oracle
of the
U
query
must tap
be
defined
repla cing
on ita
Hilbert

countably

space,

as

such

...,

infinite-dimensional

that

spanned

by

the

binary basis vectors |i, |0i,|1i,|00i, |01i, |10i,

he
|11i,
bina
|000i,
ry basis

vecwhere
tors |
i,|0i,
denotes
|1i,|00i
the,|01i,|1
empty

string. Clearly, the


oracles

still

yields

of such general unitary

unitary

oracle

well-formed
Naturally,

use

these oracles

superpositions

of

can map

outputs,

and they need not

evolution

Turing

even

and

for

machines.

inputs onto
vice

versa,

be length-preserving.

However, in order to obey the dictum that

single machine cycle ought not to make infinite


changes

in the

tape,

one

might

require
make
that
in

inite
U|zi have
chang amplitude
es in the tape,
zeroone
onmall
ight
but
requir
finitely
e th

at
many
U|zi
basis
ha ve
vectors.
am plitude
(One
zero
could
on even
all but
insist

uniform

and effective

version of the

on

above

restriction.)

may

involution,
oracle call
the

Another

same

can

= I,so

proper

that

the

that the effect

a further

be undone by

interference

may

on

be crucial to

in

to evaluating

corresponds

an

unitary

of

considered

transformation

of

call

an

to take place. Note

case

special

one

restriction

U is that it be

oracle. Again this

allow

which

natural

upon

wish to impose

this

paper,
classical

Boolean function, is an involution.

Diffi culty of Simulating

Nondeterminism

The
their

power

computational
ability

expo-nentially
tempting
parallelism

However,

to maintain
large

to try

to
there

of

QTMs

to

use

lies

and compute

superpositions.

simulate

are

on QTMs

this

in

with
It

is

exponential

non-determinism.

inherent

constraints

on

scope

the

imposed

by
5

mechanics.

this

of

parallelism,

the

formalism

In this section,

quantum

of

we

are

which

some

explore

of these constraints.
To

not

see
up

speed

exponentially,

distinguishing

from

an

unknown string

oracle,

y of

never

and

probability

to

query

one query

quantum

respecting
be

single

yes

distinct

random

n(i.e. A(y)

on an empty
its

success

yes on a nonempty

computer

a success

can do no

nbit

probability

better

do

at

strings
1/2

after

after k queries. How

computer

of

= 0)

better,

can

while

the rule that its overall evolution

unitary,

nonempty

to maximize

and k/2

problem

We require that the

of answering

random, giving

the

oracle (x A(x)

known length

answer
x 6=y

seek

oracle. A classical
than

but

containing

=1,but x6=y A(x)=0).

computer

can

consider

the empty

oracle

interference
quadratically

quantum

why

NP problems

and,
oracle,

in

all

computation
computation

with

paths

empty

querying
they

would

quantum
would

for

of 2

collapse

the

the

of

an

in
n

classical

This yields

same as

the

parallelism.

the state
with
and th
an

vector

query

computer.

whether

location
1/2

the

However,

way to exploit quantum


should be to maximize

between

after k interactions

by asking

probability

classical

Our goal

the separation

paths,

each path, and finally

a success

this is not the best

algorithm

the nonempty

had found

A direct

equally-weighted

the superposition

query

as

exactly

oracle?

computation

on

string

different

evolve

an empty

analog

start

superposition

locations

with

the state

vector

an empty

oracle, and

|k i

|k (y)i after
ankem
interactions
pty oracl

e
oracle
statenonempty
vect
or | atk (an
y)i
unknown
after

location

y.

with

Starting

uniform

superposition

|0 i

it is easily

seen that

1
2

Xx

|xi,
n

the separation after

one

query

a unitary

is maximized by

|1 (y)i

evolution to

Xx

(1)

x,y

|xi

= |0

|yi.
n

This is a phase inversion of the term


corresponding to the nonempty location. By
testing

whether the post-query

with |0 i
we obtain

a success

1
|h0 |1 (y)i|
5

There

is

exponential
and

the fact

probability
domains.

case,

superficial

parallelism
that

The difference

of

quantum-mechanical

therefore

paths

it is

superposition

at each

between

quantum

computations

exponentially

is that

this

computation

yield

large

in the probabilistic

path is chosen by making

case,

it

is possible

step to accurately

for

several

destructively,

to keep track

system.

for each step. In the

to interfere

necessary

4/21

similarity

random choicesone

computational

over

agrees

probability

probabilistic

distributions

the computational

sequence

in

state

and

of the entire

simulate

the

approximately

times

four

classical value. Thus, if

one query, quantum


improvement

interaction

the

is

same

oracle.

as

The

difference

in the

the state

way

only

one

the
only

a modest

overwhelmingly

vector

after

oracle is almost

after interaction

after

concentrate

still

a nonempty

with

than

allowed

parallelism gives

but

to fail because

likely

better

we are

an empty

with

of producing

query

would

large

be

to

much of the initial superposition

term before the

query,

which cannot

be done because that location is unknown.


Having

after

achieved

one query,

be increased
strategies
(called
Grover

[13])

is

change

the

same

about

to

the phase

difference,
sign

as

but

the

perform

unitary

separation

that separation

queries? Various

be imagined,

inversion

amplitude

can

by subsequent

can

oracle-independent

as to

the maximum

how best

good

average

one

by

an

transformation

so

into

an

difference

leaving the

all the other

term with

terms

but

magnitude

approximately

threefold

larger

Subsequent

phase-inverting

interactions

with

the oracle, alternating with oracle-independent


phase-to-amplitude

cause

conversions,

the

distance
pha se-tobetween
am plitude
|0 i
and
conver
|k (y)i
sio ns,caus
to grow
N /2.
2
n

linearly
the di sta
with
nce k,
be tween
approximately
|0 i
a nd as
2k/ to
| k (y)i
N/2.

row
when
link
ear

ly

growth
as 2k/

the
w
he

of

proof

success
n
k
4k

/2

probability,

for

small

k. The

of Theorem 3.5 shows that this approach

is essentially

can

with
Thisk,
results
approximate
in a quadratic

as

approximately

more

gain

success

NP-type

algorithm

than this quadratic

probability

algorithms,

no quantum

optimal:

when

questions

to

attempting

formulated

in

factor

to

compared

classical

answer

relative

to

random oracle.

3.1

Lower Bounds

on Quantum

Search

We

will

measure
calculating

sometimes

the

find

accuracy

of

the Euclidean

it

convenient

to

simulation

by

distance

between

target

the

and

The following

theorem

the simulation

worse

are

samples

If two

two

Definition

unit-length

distance

distance

superpositions

from distributions

total variation

which

are

gives

within

3.2 Let |i i
be the superposition

on input x at

time i.
We denote by

of squared

of configurations

of M which

magnitudes

are

of
oracle
configurations
on string y.
of Mwhich
We refer to

magnitude

of

y in |i

qy

in |i i

querying

the

(|i i) as the

i.

Theorem 3.3 Let |i i


be the superposition
A

at most 4.

qy
of(|i
M i) the sum

query

that

is at most 4 times

within Euclidean

observ-ing the

from [4] shows

accuracy

3.1

superpositions
then

superpositions

than this Euclidean distance.

Theorem

of M

simulation

Theorem
Let
be the
of
M
on3.3
input
x|i
ati
time
i.

super
Let >0.

Let F [0,T 1]

be a set of
P(i,y)F

qy

time-strings pairs such that

LetT F [0,T 1]
6

and |i
7

P(i,y)F

The Euclidean distance between |i

Px

as (

Px

1/2

x |)

|x

x |xi
P

Px

The total variation distance between two


Px

distributions

Now

Px

|xi is defined
Px

(|i i)

Dand D

suppose

the

F is modified

(these

answers

oracle).
Let |
(th ese answe

is

|D(x) D

answer to
to

(x)|.

some

each

query

arbitrary

need not be consistent


i

(i,y)

fixed ai,y
with

i
be thed
time
superposition
r
snee
not bi
ec
onsistent

M
input Let
x |with
Ai
modified
anon
oracle).
i
beoracle
the time
su
stated above. Then ||T i
|

operator of M

F then
Ai (y)
=then
A(y).
Let
that
if (i,y)
F
A i
time
Fevolution
then Ai

operator

as

time evolution

Let Ai denote

that if (i,y) F then Ai (y)

of
wi

i| .

Proof. Let U be the unitary

A.

an

an oracle

= ai,y

such

and if (i,y)

Ui be the
unitary
an d
if (i,y

of M

Ai

Let |i i
be

the superposition of M

We define
caused

by

replacing

on input x
Le
at
ttime
| ii
b
i.

|Ei i
to be the
the

error

th

in the i

oracle

step

with

Ai

Then

= Ui |i i U|i

|Ei i

i.

So we have
T1

= U|T1

|T i

= UT

= =

|T1 i
|ET1 i

Xi=0

UT

U1

Since

|0 i

Xi=0

Ui

UT1

are

all of the Ui

|Ei i.

unitary,

|UT1

= ||Ei i|.

|Ei i|

The

sum

of squared

magnitudes

of all of

P(i,y)F

the

Ei

therefore
UT1

is

however,
U

at most

at most

Ui |Ei is

qy

to

equal

(|i i)

and

P(i,y )F

In the worst

could interfere

case,

the

constructively;

the
squared
magnitude
their
sum is
i
|E i
is co uld
inter fer of
econ
structive
T times

magnitudes,

i.e.

the

2.

sum

Therefore

of their

squared

||T i |

i|

3.4

Corollary

alphabet

. For

such that

time

Let

y Ay

6=

(x)

isuperposition

an

be

let Ay

= A(x).

of M

over

oracle

be

any

oracle

Let |i i
be the

on

input

and

(y)

|i i
be the time i

superposition of M

cardinality at

most

on input
???|T x. Then

for

2T

every

Ay

such that

>

0, there is

???

(y)

i
|T i

a set

S of

PT1

Proof. Since each |t i


has unit length,
Py

qy
PT1

such

that

Therefore

PT1

Py

card(S)

If

i=0

(|i i) T. Let S be the set of strings

PT1
2
2T

i=0

S then

by

qy

(|i i)

2T

Clearly

PT1

qy

i=0

Theorem

3.3

(|i i) <

2T

???|i

???

(y)

i
|i i

Theorem 3.5 For

any

T(n) which is o(2

n/2),

relative to a random oracle, with probability 1,


BQTime(T(n)) does not contain NP.

Proof.

Recall

can

oracle

from

be

length-preserving

mean

Section

thought

of

function:

this

below by A(x). Let LA

is

we

what

= {y :x A(x)
is
contained
in
= {y
:x A (x)

=
y}.belo
Clearly,
language
mean
w by this
A(x).Let
LA
A.
n/2).
Let T(n) = o(2
We show

NP

any

an

that

as

that for

bounded-error

8
A

oracle

QTM M

T(n),

with

accept

the language

taken

over

LA

in time at most

.
.

1, M

countable

oracle A
number

does

not

The probability

thethe
choice
of LA
a
la nguage

length-preserving

are a

running

probability

is

random

Then, since there

of QTMs

and the

intersection

conclude

that

error

bounded

countable

1
events

probability

we

of

enough
the

so

1,

no

with

oracle

way

not

over

equal

function

over

us

fix

pick

to

We will show that

M gives

the

wrong
every

answers on inputs

n.

in

n large

is at least 1/8 for

The

probability

of

is

n.

an

arbitrary

from strings

alphabet

oracles

can

LA

the random choices of the oracle for

inputs of length
Let

n/2

20

fixing the oracle

length

taken

that

input 1

accepts

= o(2 n/2), we

probability

of

probability

QTM

that T(n)

answer on

of

still has probability

time bounded by T(n).


Since T(n)

number

length-preserving

of lengths

Let

other than

C denote

the

set

n
of

consistent
thisC darbitrary
alp habwith
et . Let
enot e the

function.
Letes
A consiste
be the set
in
C such
set of o racl
nt of
witoracles
hth is
arbit
rar
n

that
1 on.
hasLet
noAinverse
fun cti

(does

not belong

to

ch
LA t).
hat
If the
1
has
oracle
noanswers
inve rs e
to
(does
length
n ot
n
bel
strings
ong

are

chosen

probability

uniformly

that the

at random,

oracle

then the

is in A is at least

1/4. This

has
1/ 4.
noThis
inverse
is beca
is (
(for

the
probability
that
1
oracle
is in A is
a t lea

is because

n suffi ciently

use
the
)
which
prob abis
ilit
at
y th
least
at 11/4
n
2

large). Let B be the set of


n

oracles in C such that 1

a unique

has

inverse.

As
above,
the
probability
that
randomly
oracle
s in C
su ch
th at
has
a ua
nique
inv ers

chosen
As abo
oracle is in B is (

1
n

which is at

least 1/e.
Given

an oracle

answer on any

A in A, we

single input,

can

modify its

say y, to

and

therefore
its a nsweget
ron
an any
oraclesingle
Ay
in
in pu
B.t,
Wesay
will

show

that

acceptance

for

most

probability

choices
of M

on input

almost
equal to the acceptance
ac cepta nce
Ay

of M

must

on

reject

Therefore

input 1

easy to

and

with probability

is

uniformly

length

n, and

Ay

must

other strings.

accept

is

that M fails

on

and LAy

more

carefully,

input

at least 1/8 when the oracle

random function
is

probability

cannot accept both LA

show

On the other hand, M

By working through the details


it is

y,
the
inB

of

an

arbitrary

on

strings of

function

on

all

Let Ay

be the oracle such that Ay (y)

and z 6=

Ay (z)

= A(z). By

that

the

2(n)

superpo-sition
input

Theorem

Ay

of M

3.1

we

can

most 1/13

that M

each

arec
single
cor
tly deanswer
cide s

mapped

to at least (2

different

oracles

correctly

Af

the

is at

should

we can

either LA

for

which

at least 2/3,

on

Using

should is
accept
at m

fails to accept

oracle

th

probabilities

decides
So, ewhether
ach ora 1
cle
A LA

correctly

changing
hich
M

Ay

that

at least 2/3 and M

conclude

So,

1/13.

on input

4 < 1/3. Since M

with probability

strings

the

and M

conclude

1 and M

reject 1

LAy

norm at most

1 with probability
n

on input

has

on input

Ay

between the acceptance

difference
of M

between

difference

3.4

Corollary

there
is a
S of
at most 338T
a nd
set
z 6=
y Ay

such

= 1n

or

can,
by
A fo r
w
n

of A to
be
whether
1 1 LA

card(S))

n1

B for which M fails to


n

decide whether 1

LAf

Moreover,

rectly
any particular
de cid e wAf
hethe
B
ris
1 the
image
LAf
under this
mapping

of at most

1
oracles

A A,

since where

it

given
one of
2
since wher
ethe
it

Therefore,

which
answers

the

now answers
n

it must have

answers.

1
other possible

of 1
oracles
Bible
for

other in
poss

number

M
fails must
at least
1/2 in
the
T herefor
e,t he be
number
of or acles

number
for which
of M
oracles
fai ls
in
must
A for
be
which
at leas
M
tsucceeds.
1/2 the
So,
number
calling
of oracles
a the number
in A

of oracles

in A for

which M fails, M must

9
fail for at least

Therefore
n

1 LA

a + 1/2(card(A)

M fails to correctly

with probability

a) oracles.

decide whether

at least (1/2)P[A]

1/8.
It is

easy to conclude

in
withto
probability
ItLA
is easy
c

membership
uniformly

that M decides

0 for a

chosen oracle A.

Note:
isolate

Theorem
the

parallelism

rest

of

3.3

and

constraints

its

Corollary

on

3.4

quantum

imposed by unitary evolution. The

the

proof

of

the

above

theorem

is

similar

to standard

in spirit

to separate
random

oracle

be

For

to

example,

to

used

random

oracle

probabilistic
n

[3]

can

to

used

BPP from NP relative

techniques
relative

techniques

can

machine

these

show

no

A,

that,

classical

recognize
LAssica
in
no cla

time
probabili
o(2 stic
)
However,
machine quantum
c an recognize
ma-chines
LA
2

recognize

this language
2

in time

O(

[13].

This

quadratically

), using

explains

Grovers

why

of the standard

modification

can

faster,

algorithm

substantial

technique

was

required to prove the above theorem.

next

The

to

relative

result

about

random

NP

coNP

permutation

oracle

NP
requires
re lativ
a more
e t oa
subtle
ra argument;
ndom per mut
ideally
ation
we
would

like

to

asserting that

which

1(1 n

apply

the total

Theorem

query

) is probed

this is precisely

what

3.3

after

magnitude

with

is small. However,

we are

trying to prove in

the first place.

Theorem 3.6 For

any

T(n) which is o(2

n/3

),

relative to a random permutation oracle, with

1, BQTime(T(n)) does not

probability

contain NP coNP.

Proof.

For

any

permutation

oracle

A, let

1(y)
=of.
{y For
: first
anbit
yof
permuta
A
tio
is 1}nora
Clearly,
cl
this
language

A,let
LA
= {yis: contained in (NP
is 1}.
A.
n/3).
coNP)
Let language
T(n) = o(2is cont We
show in
that
Cle
arly,
this
ained
(N
LA
Pro

for co
any

N
bounded-error
P)

running

in

LA

at

time

1,M

probability

The probability

the intersection

we

oracle A

of

1
events

conclude

bounded

error

enough

so

with

that

countable

of

1,

no

probability

accepts

QTM

= o(2 n/3), we

that T(n)

number

still has probability


with

oracle

Then, since

number of QTMs and

time bounded by T(n).


Since T(n)

T(n),

is taken
not ac over
cep tthe
the choice
lan guage
of

are a countable

probability

most

QTM

does not accept the language

LA
a random permutation
there

oracle

n/3

100

can

pick

LA

in

n large

We will show that

arge
the en
probability
oug hso t hat
that
T (n)
Mgives

the

wrong

answer on

way

is at least 1/8 for

input 1

fixing the oracle

of

length

not

equal

taken

over

the

to

answers on inputs

n.

The

random

the

random

defining

of

probability

choices

of

method

following

on

permutations

is

the

n.

permutation oracle for inputs of length


Consider

every

of

{0, 1}

n:

let
definin
x0 ,x1
g
,...xT+1
random be
p ermutatio
a sequence
ns
ofon
strings
{0,
chosen

random
be a
in

at

uniformly

sequenc
{0, 1}

.,

n.

Pick
of strings
0
uniformly
c hosen unif at
ormly
random
a t ra ndo
among
m in

= 1n

{0,1
permutations
}

i1

i.e.

such that (x0 )

, where

i (xi )

i1 (xi ).

{0,1}

Ai

and

is

xi),

i (xi1 )

random

n.

sequence

of

such that Ai (y)

and Ai (y)

Denote
{0,1}

each

on {0,1}

Consider

oracles

is the transposition (xi1


i1 (xi1 )

Clearly

permutation

Let i

= i (y)

permutation

= Aj (y)
if

if

n.
y
(y)
{0,1}
if y

by |i i
the time i
superposition

of

AT(n)

10

on

input

n,

and

by

i
the

time

superposition

of M

con-struction,

with

the string 1

is

AT(n)1

probability

a member

the two languages LAT(n)

.
.

on input

exactly

of exactly

and LAT(n)1

???|T(n)

By

1/2,

one

of

We
?? ]

T(n)

will

show

1/50.

Here

that

i
???|T
(n) |

E[

the expectation

the
random
1/50.

choice ???|T(n)
of the

???

is i
taken
|

over
i

oracles.

By

???

T(n)

3/4.

bound,

i
|

i 2/25]
???

Applying Theorem 3.1


we conclude

???|T(n)

Markovs

???|T(n)
P[

T(n)

that

i |
i
???| T(n )

if

acceptance
M

AT(n)1

probability

differ

2/25, then the


???

of

AT(n)

and

< 1/3,

by at most 8/25

and
n

hence either both machines accept input 1


both reject that
M

AT(n)1

input. Therefore

give the

with probability
the probability

same answers on

AT(n)

or

and

input

at least 3/4. By construction,


n

that the string 1

belongs

exactly

one

LAT(n)1

is equal to P[first bit of xT(n)1

of the

two languages

=1/2. Therefore,

bit of xT(n) ]

bit ofwith
xT( n) probability
that

LAT(n)

to

and

6= first

we can conclude

at least

1/4, either

AT(n)

input

or M
n

AT(n)1

are chosen

same

from the

conclude that M
n

on input

gives the

wrong answer on

each of AT(n)

Since

AT(n)

and AT(n)1

distribution,

gives the

we can

wrong answer

with probability at least 1/8.


???|T(n)
???],
T(n)

To bound E[

???|T (n

T(n)

we
are

show

that
???],

each close to

To define

|T(n) i and

a certain

superposition |
T(n) i

this superposition,

run

on

input

d efin
t his suporacle
erpo sition,
run
M on
input
1T o
with
ae
different
on each
step:
on
step
i,use Ai

to

answer

the oracle queries. Denote

by |i i,to
the
superposition
thatDresults.
antime
sw eri
t he
oracle
qu eries.
en ote
the set of time-string

Consider

xj )

:
j

pairs S

= {(i,

i,0 i T}. It is easily checked

that
{(i,xjthe
):
oracle
j i queries
0 i

inTthe
}. computation

described
M

AT(n)+1

above

and

differ only

that the expected

those

on

query

think of xj

AT(n)

and

the set S. We claim

magnitude

in the set is at most 1/2

we may

of

n,

of

any

since for j

as having

been

pair

i,

randomly

chosen during step j,after the superposition


of oracle queries to be performed

has already

been written

sum

of the

on the

query

oracle tape. Let be the

magnitudes

time-string

for

pairs in S.Then

E[]

card(S)/2

!/2
T(n)

+1

T(n)

for T(n) 4. Let be a random variable


2/2T(n).
such that

Then by Theorem
???|i
???|i

???

???

T(n)

3.3,

|T(n) i and

i .

We showed above that

E[

2/T(n)]

= E[]

T(n)

q2T(n)]
2

But E[/

E[

q2T(n)E[/
E[]

/2T(n)]. Therefore

q2T(n)]

s2T(n)
q2T(n)E[
2/2T(n)]

T(n)

2
100

< 1/100.
???|i

???]

Therefore E[

|T(n) i

???|i
T(n)

and E[

???]

< 1/100.

???|T(n)

that E[

It

???]
T(n)

follows

< 1/100

E[]

???]
???|i
i E[]

i
|

<1/50.

11
Finally, it is

easy to conclude

decides membership

in LA

that M

with probability

for a uniformly random permutation oracle A.


2

Note: In view

we

know

that

improved

constant

Theorem

of

algorithm
1/2

3.5

[13],

in the

cannot

be

On the other hand, there is

evidence

that

statement

may

the

state-ment

of Grovers

the

constant

1/3

no

in the

of Theorem 3.6 is fundamental

It

well be that Theorem 3.6 would still hold

(albeit

not

its

current

proof)

with

1/2

substituted for 1/3.

to

Corollary

3.7

permutation

oracle, with probability

Relative

exists

Given

the oracle,

one-way

quantum

this

computed

effi ciently

deterministic

exponential

with

yet

machine,

time to invert

there

permutation

permutation

even

,.

random

it

can

be

classical
requires

even on a quantum

machine.

Proof.

an

Given

oracle A for which A

time

o(2

is just

n/3)

can

on a quantum

permutation

be computed

of Theorem

proof

that

when A is

Turing machine, it

this

3.6. It follows

happens

uniformly

from that

with probability
random

Using

permutation

in

as easy on
toadecide
qua ntu
LA m
as
Turi
defined
ng ma
in
chin
the

proof

oracle

arbitrary
1

a Bounded-Error

QTM

as a Subroutine

The notion of
invocation

subroutine

provides

abstrac-tion

in

computation.

context

the

making

Before

in the context of

call

classical

quantum computation, there

thought

through.

For

computes

would like to think

on

useful

of

subtle considerations

subroutine

oracle

and

this abstraction

are some

subroutine

or an

simple

the

example,

the
of

that must be

an

the

if

function

f,

we

invo-cation of the

x as

string

magically

writing f(x) in some designated spot (actually


xoring it to

ensure

unitarity). In the context

of quantum algorithms, this abstrac-tion is only


valid if the subroutine
its intermediate
the final

cleans

calculations,

answer on the tape.

if the subroutine is invoked

of xs, then different


in different
would

these

on

leaves

just

This is because

on a

different

all traces of

and

values of

scratch-work

prevent

up

superposition

x would
the

result

tape,

and

computational

paths from interfering.

cannot, in

unitary operation, the scratch-work

general,

case

be erased

where

deterministically,

so

subroutine

post-facto.

can

be

into

f(x)

safe

to get

design

erases

compute

BQP

storage,

rid

However, in the

machine,

and

the

of

case

scratch

the correct

copy

f(x)

uncompute

paths

answer

f (x)

work

situation

some

storage

computational

by

of the

lead to

f(x), and therefore


safe

[2]

more

is

of the machine

into

the

uncompute

that f is computed

the

the

copy

f(x),

then

complicated. This is be-cause only


computational

special

computed

easy to

is

scratch-worksimply
f(x)

In the

effi ciently

it reversibly

it

that

is not

Since erasing

if

we

and

then

paths

with

different values of f(x) will no longer interfere

with each other, and

we will not

12

reverse

the first

phase

We show, nonetheless,

of the computation.

that

if

we

boost

the

suc-cess

probability

copying

of the BQP machine before

into

f(x)

the final superposition


only the input

has

and the

such tidy BQP machines


subroutines,

BQP

BQP

careful

such

as

f(x). Since

be safely used

us to

4.13

the

same
that

and

outline

we

machinery

show

that

procedure

4.14.

as

in

The
the

have to be much

insimple programming constructs

looping, etc. We therefore


developed

and present
lemmas and
section.

can

of the boosting

case, except

more

answer

The result also justifies

in Theorems

follows

classical

clean tape with

of oracle quantum machines.

The correctness

proof

this allows

= BQP.

our definition

is proved

and

f(x), then most of the weight of

uncomputing

as

storage

safe

The

theorems
main

of the relevant

in the first part of this

new

contribution

proofs of Theorems

4.14. The reader

purpose,

in [4] for this

the statements

section is in the

borrow the

may

therefore

directly ahead to these proofs.

in this

4.13 and

wish to skip

4.1

Some Programming

Primitives for QTMs

we present

Inthis subsection,

several

definitions, lemmas and theorems from [4].


Recall that

a QTM Mis defined

by

triplet (,Q,) where: is a finite alphabet

with

an identified

blank symbol #,Q is a finite

set of states with an identified initial state q0


and final state qf

6= q0

and , the quantum

transition function, is a function

where

to within 2

Definition

{L,R}

is

can be approximated

intime polynomial inn.

4.1 A final configuration

any

configuration

when QTM M is
the

C is the set of complex numbers whose

real and imaginary parts

QTM

run

superposition

in state

with input

contains

x,at
only

qf

of

If

time T
final

any

configurations

and at

superposition

contains

then M halts

with running time T

x.

The superposition

called

the final

input

x.

time less than T the

no

final configuration,

superposition

of M

polynomial-time

well-formed QTM which

4.2

on every

well-behaved if it halts

a final

QTM

input

is

is

x halts

x.

is

on all input

called

strings in

superposition where each configuration

has the tape head in the

cell

run on

QTM

intime polynomial inthe length of

Definition

on input

of M at time T is

always

same

the

start

that

a QTM

cell,

cell. If this

we

call

the

QTM stationary.

We will

say

form if all transitions

state qf

M is in normal

symbol in the scanned


and the head

moves

from the distinguished

lead to the distinguished

state q0

the

cell is left unchanged,

right, say. Formally:

13
Definition 4.3 A QTM M

=(,Q,)

is in

normal form if

Theorem

strings

(qf ,)

4.4

If f

to strings

deter-ministic

= |i|q0

a function

is

can

which

polynomial

x,

then

stationary,
input

there

normal

x, produces

is

form

be computed

a one-to-one

on the

polynomial-time,
given

and whose

on the length

of

in deterministic

can be

polynomial

time,

and such that the length of f(x) depends

on

the

length

of

polynomial-time,
QTM which

given

f(x), and whose

x,

x.

function from strings

to strings that such that both f and f


computed

in

length

Q TM which

x; f(x),

output

running time depends only


If f is

mapping

time and such that

the length of f(x) depends only

of

i|Ri

then

stationary,
input
running

x,

there

is

normal
produces

only

form

output

time depends

only


on the length

of

4.5

Definition

machine

x.

multi-track

with k tracks is

whose alphabet is of the form 1

with

a special

the input

by

blank symbol # ineach


2
i so

that the blank in is (#,...,#).

We specify

the string

specifying

(separated by

track

Turing

Turing machine

;),

on

each

and optionally by

of the contents of the

specifying the alignment

tracks.

Lemma 4.6

(,Q,)

QTM

behaves

any

Given

any set

and

= (

,Q,

exactly

as

QTM

there

such

M while

is

that

leaving

its

second track unchanged.

,,

Lemma 4.7 Given


k

any

QTM M

Q,) and permutation

= (1

:[1,k]

[1,k],

)
k
thereQ,)
is a QTM
an d pe
Mrmutatio
((1)
n
(k)
:
[1,k ] ,Q,
[1,k],

such that the M

behaves exactly

as

M except

that its tracks

Lemma

are permuted

according to .

4.8 If M1 and M2

normal form QTMs


then there is

are

with the

well-behaved,

same

alphabet,

normal form QTM M which

carries out the computation of M1 followed by


the computation of M2

4.9

Lemma

well-behaved,

Suppose

normal

that

form QTM M

x;k with

> 0, the

such that

machine M

on its first

k iterations

is

form QTM. Then there

is a normal
k

on input

runs

M for

track.

14
Definition
the

same

reverses

4.10 If QTMs M1 and M2


alphabet,

then

we

say

have

that

M2

the computation of M1 if the following

x on which

holds: for

any

cx

be the initial configuration and final

and

input

superposition of M1
input

M1 halts, let

on input x.Then

the superposition

halts

M2

on

with final

superposition

configuration

consisting

cx

of

reverse

to

Note that for M2

the final state of M2

M1

entirely

must be equal to the

initial state of M1 and vice

versa.

Lemma 4.11If Mis a normal form QTM


which halts
normal

on all inputs,

form QTM M

then there is a

that

reverses

the

computation of Mwith slowdown by

a factor

of 5.

Finally, recall the definition of the class BQP.

4.12

Definition

Let

M be

normal form, multi-track


track

has alphabet

accepts
of the

rejects

#)

say

that M

1
in the last track

cell. Otherwise

we say

that

x.

A QTM accepts

stationary,

QTM M whose last

{#, 0,1}. We

if it halts with

start

the language

L (

with
A QT
probability
Macc ep ts
p the
if M
langua
accepts
ge L
with

prob-ability

at least

p every

string

and rejects
with
probability
at strin
least
a bility
at lea
st p every
gp
x every
L

string
and r

x
( with
#)prob

ejects

L.
We define
the p
class
ability
at
least
eve

BQP
(bounded-error
str ing

time)

as

accepted

set

the
with

the

2/3 by

any

QTM.

class

of languages

are

which

2/3

More

accepted

QTM whose

input of length

4.2

languages

are

some

by

we

generally,

as

BQTime(T(n))

which

some

polynomial

probability

polynomial-time

define

of

quantum

set

the

with probability

running

nis bounded

on

time

by T(n).

Boosting and Subroutine

Calls

Theorem

4.13 If QTM M accepts

language

L with
The
orem
probability
4.13 If2/3
QTM
in time
M a T(n)
ccep ts
>lan,
ngwith
uag

T(n)
time-constructible,
for T(n)
any
e
L with
pro ba bility 2/3then
in time

0, there
with
with

is

probability

QTM

which

accepts

in time
which
cT(n)
accepts
where

polynomial in log1/ but independent

Proof.

Let

M be

>
>
n

stationary

of

is

n.

QTM which

accepts the language L in time T(n).


We will build

machine

that

runs

independent

copies

the majority
input

x,

of M and then takes

vote of the k answers.

will

some

have
Pi

superposition

of strings

final

i |xi i
Pi

If

PiA

on separate

we

call

has
correct
|x the
ii.
If we

A the set of i
for which xi

call
answer
A t he
M(x)
set then
of i
for which
|i |xi
running

On any

copies

2/3. Now

of its input

P iA

times

will

produce

Pi1

nseparate
ik |xi1 i
c opi
|xik
es of
i. it

ning
M
i1o
,...,ik
Then

|xi1 i|xik

correct

Pi1

the

probability

i
such that

answer

of

seeing

the majority have the

M(x) is the

,...,

sum
of
|i1
th
majo r
it y
ha|
ve

|ik
c orrec
| such
t anthat
sw er
the
M(x)
majority
is theof
sum
i1 of |i1
ik lie
|
in

|
A.
ik But
|
this is just like taking the majority
k independent

coin flips each

at least

of

2/3

some constant
the probability

heads.

with probability
there

Therefore

b such that when

of seeing

of

is

= blog 1/,

the correct

answer

will be at least 1
.
So,

we

will build

the following steps.

machine

to

carry out

15

n = T(|x|).

1. Compute

2.Write out k copies of the input

x spaced

out with 2n blank cells inbetween, and write


down k and

non other

3.Loop k times
and then steps

tracks.

ona machine

that

runs

ntimes to the right.

4.Calculate the majority of the k answers


and write it back inthe start cell.

construct

We

building

the

desired

QTM

by

QTM for each of these four steps

and then dovetailing them together.


Since

Steps

computable
depend

functions

only

can carry
normal
Theorem

1, 2, and

on

4.4,

depend only

whose

output

easily
length

k and the length of

them

form

4 require

out

QTMs,

whose

on k and the

using

x, we

well-behaved,

constructed
running

length of

times

x.

using
also

we

So,

complete

the

to

constructing

QTM

by

proof

run

the

given

machine k times. First, using Theorem 4.4

can construct a stationary,


which drags the

to the right

none square

integers k and

on its

we

normal form QTM

work track

If

we

add

single step right

to the end of this QTM and

apply

4.9,

Lemma

we

can

build

well-behaved, normal form QTM

n squares

to the right, dragging

along with it. Dovetailing


M, and

then

applying

normal form QTM that

this

a machine to return

to

the

a QTM

k and

after

4.9 gives

M on each of the

we can dovetail

with k and

start cell by using Lemma 4.9 two


around

which

machine

Lemma

runs

k copies of the input. Finally,


with

moves

n to the

more

which carries k and

times

none step

left.

The extra information

of

a QTM can be erased

on

the output

tape

by copying the desired

output to another track, and then running the

reverse
in

of the QTM. If the output is the

every

recover

the

then this reversal

final

will exactly

the input. Unfortunately, if the output

differs in different

the output
from

in

configuration

superposition,

same

then saving

configurations,

will prevent

these

when

interfering

configurations

the

machine

is

reversed, and the input will not be recovered.


We show

is the

superposition,

same

in most

of the final

the reversal must lead

then

us

close to the input.

Theorem
contained

4.14

the

If

language

is

in
The
the
orem
class 4.14
BQTime(T(n)),
If the langu age
with
L

T(n)
is c onta
> in
n and
ed in
T(n)
the ctime-constructible,
las s BQTi me( T(n)),
then

for
withany
T (n)
>
>n 0, there is

accepts L with

1
and haswthe
hi

following
h
a ccepts
of length

probability

property.
L
with

n,

cT(n), where

and produces

When

runs
is

final

QTM M

run on

for time

which

input

bounded

polynomial

superposition

by

in log1/,

in which

|xi|L(x)i,

with L(x)

otherwise, has squared

1
if

magnitude

L and 0

at least 1

Proof. Let M

= (,Q,)

be

a stationary,

normal form QTM which accepts language L


in time bounded by T(n).

16

to

According

expense
can

of

polynomial

slowdown

assume

running

that

1
/2

accepts

on every

we can construct
M, copying

simple
back
clean

copy

is

two-step
right
track.

construct

easily

while

normal

is

n,we

with

the desired M

the

by

another

reverse

of M
with

that steps left and

writing

Using

of

input.

accomplished

machine

the

which

answer to

the

track, and then running


The

factor

in log1/ but independent

probability
Then

by

at

4.13,

Theorem

the

Lemma
form

answer
4.11,

QTM

on a

we
R

can

which

reverses

M. Finally,

desired

we can construct

4.6 and 4.7,

Lemmas

stationary

machines

QTM

M and

use

with appropriate

by

around

of

the

dovetailing

the

copying

the

desired

machine.

see

To

properties,
length

n.

that

this

consider

has

running M

run

will first

on

M on

input

of

producing
Py

some

final superposition

of configurations
Py

y |yi of M

on input x.Then

or
|yi
1
inthe
of M
extra
configura- tion
and

|i
If

run

=
we

it will write

a0

track of the start cell of each

on

this

superposition

Py

y |yi|by i.

were

to

Py

run

instead
i

superposition
|yi|by i.If|we

on

the

Py

wer
= e to inste
y |yi|M(x)i
ad run M

we

Py

the
would
superp
afterosition
T(n) steps
|
have the superposition
|yi|M(x)i w
consisting
wou ld afteentirely
rT (n)of
st

eps
the have
final the
configuration
su pe rpos

i
Clearly, h|
is real,
1
n
and
with
since
outpu
Mt
has
x ;M(x).
success Clearly,
probability
h|
at least
is real
1
1
i
and
/2,
h| M
.
sinceatthe
since
has suc
ces Therefore,
s proba bility
lea

tion
with output

x; M(x).

time
1
evolution
/2
h

of
| M

is unitary
and refor
hence
. The

preserves
since the

ti
the
me inner

superposition
product

of

product,

must

the

final

an

have

inner

which must
is realhave
and an
at

with |xi|M(x)i

least
inner prod
1
uc t
/2.
with|x
Therei|M(x)i
fore, which
the
squared
is rea l
a

nd
magnitude
at lea st
in
1
the
/2.
final
T superposition
hereof M
the final configuration

must

be

at

least

output

with

(1

/2)

of

x; M(x)
1

BQP

Corollary 4.15 BQP

=BQP.

Acknowledgement

We wish to thank Bob Solovay for several


useful discussions.

References

[1] Babai, L.and Moran, S., ArthurMerlin

games:

A randomized proof system, and

hierarchy

of complexity

classes

Journal

of Computer and System Sciences, vol. 36,

1988,

pp. 254276.

[2] Bennett,

reversibility

C.H., Logical

computation,

IBM Journal of Research and

Development, vol. 17, 1973,

pp. 525532.

[3] Bennett, C.H. and Gill, J., Relative


A

random oracle A,P


with

6= NP
1

probability

Computing, vol. 10, 1981,

[4]

Bernstein,

E.

complexity

6= coNP

6=
SIAM
NP

to a

Journal
6=

on

pp. 96113.

17

Quantum

of

and

Vazirani,

theory

U.,

Proceedings

of

on

of

the 25th
Annual

ACM Symposium

Computing, 1993,

Theory

pp. 1120.

[5] Berthiaume, A. and Brassard, G., The

quantum challenge to structural complexity


theory
Conference

Proceedings

on

Structure

of

7th

IEEE

in

Complexity

Theory, 1992,

pp. 132137.
[6]

Berthiaume,

A.

and

Brassard,

G.,

quantum

Oracle

Journal

computing

of

Modern
Optics,

vol. 41,

no.

12, December

1994,

pp. 25212535.
[7] Boyer,
Tapp,

A.,

M., Brassard,
Tight

G., Hyer, P. and

bounds

on

quantum

searching,

Proceedings

of the Fourth Workshop

on

Physics and Computation, Boston, November 1996, New England Complex


Institute,

pp. 3643.

InterJournal

Systems

Available online inthe

at http

:
//interj

ournal

org.
[8] Bshouty, N. and Jackson,

DNF

over

tum example oracle

Deutsch,

Church-Turing

quantum
computer

D.,

Proceedings

of 8th

pp. 118127.
theory,

Quantum

principle

a quan-

on Computational

Learning Theory, 1995,


[9]

J., Learning

uniform distribution using

Annual ACM Conference

and

Proceedings

the

of

the

universal

the

Royal

Society, London, vol. A400, 1985,

[10] Deutsch,

networks,

D., Quantum

Deutsch,

solution

computational

Proceedings of the Royal Society,

London, vol. A425, 1989,


[11]

pp. 97117.

of

pp. 7390.

D. and

Jozsa,

problems

by

R.,

Rapid

quantum

computation,
Proceedings

of the Royal Society, London,

vol. A439, 1992,


[12]

Feynman,

computers

pp. 553558.
R., Simulating

International

physics

with

Journal

of

Theoretical
Physics,

vol.

21,

nos.

6/7,

1982,

pp.

467488.
[13] Grover, L., A fast quantum

mechanical

algorithm for database search, Proceedings of


the

28th

Annual

ACM

Theory of Computing, 1996,


[14]

Machta,

quantum

J.,

oracle

Symposium

on

pp. 212219.

Phase

information

computing

in

Physics

Department,
University

of Massachusetts

at Amherst

manuscript, May 1996.


[15]

Shor,

P. W., Algorithms

computation:

Discrete

quantum

for

logarithms

and

factoring,

Proceedings

on

Symposium

of

the

35th

Foundations

Annual

of

IEEE

Computer

Science,

1994,

pp. 124134.

[16] Simon, D., On

the

power

Proceedings

computation

of quantum

the

of

35th

Annual

IEEE

on

Symposium

Computer Science, 1994,

Foundations

of

pp. 116123.

[17] Yao, A., Quantum

circuit

complexity,

Proceedings of the 34th Annual IEEE Sympo-

sium

on

Science, 1993,

Foundations

pp. 352361.

18

of

Computer

You might also like