Академический Документы
Профессиональный Документы
Культура Документы
r
X
i
v
:
1
3
0
5
.
5
6
4
9
v
2
[
q
u
a
n
t
-
p
h
]
1
4
N
o
v
2
0
1
3
Optimal strategies for estimating the average delity of quantum gates
Daniel M. Reich, Giulia Gualdi,
act
] with
id/act
the ideal/actual state. Monte Carlo sampling estimates
the quantity F from a small random sample of measure-
ments. F thus needs to be expressed in terms of mea-
surement results. To this end,
id
and
act
are expanded
in an orthonormal basis of Hermitian operators, yielding
F
k
Tr[
id
W
k
] Tr[
act
W
k
]. The measurement results
are treated as a random variable X taking values X
T
=1
Pr() = 1
2
with T the size of the event space), i.e.,
F =
T
=1
Pr()X
. (1)
Introducing
() = Tr[W
], X
is given by
X
act ()
id()
=
Tr[
act
W
]
Tr[
id
W
]
and Pr() =
id()
2
A
with A ensuring proper normalization. Pr() is also
called relevance distribution. Two levels of stochastic
sampling are involved in the Monte Carlo estimation of a
delity: According to Eq. (1), F is the expectation value
of the random variable X taking values X
with known
probability Pr(). However, the X
cannot be accessed
directly, since they depend on another random variable,
the expectation value of W
for
act
. Due to the statisti-
cal nature of quantum measurements as well as random
errors in the experiment, it will be necessary to repeat-
edly measure W
in order to determine X
. Assuming
that the X
l=1
X
l
(2)
and
L
taking values between 1 and T. L is chosen such
that the probability for F
L
to dier from F by more than
is less than . The key point of the Monte Carlo ap-
proach is that L depends only on the desired accuracy
and condence level and is independent of the sys-
tem size. Note that is lower bounded by measurement
and state preparation errors. Since, as quantum me-
chanical expectation values, the X
l
are known only ap-
proximately, also F can be obtained only approximately:
F
L
=
1
L
l
=1
l
(with
X
l
denoting the approximate
values of X
l
). Therefore, in addition to ensuring that
F
L
approximates F with a statistical error of at most
, one also has to guarantee that
F
L
approximates F
L
with the desired accuracy. This implies repeated mea-
surements for a given element l (l = 1, . . . , L) of the
Monte Carlo sample. Denoting the number of respective
measurements by N
l
, the total number of experiments is
given by N
exp
=
L
l=1
N
l
. It can be shown that a proper
choice of N
l
and N
exp
guarantees the approximations
of F
L
by
F
L
and of F by F
L
to hold with the desired
condence level. While L is indepenent of system size,
the choices of N
l
and N
exp
in general depend on it.
For a quantum process, the average delity can be
obtained by Monte Carlo estimation when combining it
with the channel-state isomorphism [2, 3]. F
av
is then
expressed in terms of the entanglement delity F
e
via
F
av
= (dF
e
+ 1)/(d + 1) [19]. Since F
e
is a state delity
in Liouville space and Liouville space vectors correspond
to Hilbert space operators, this implies evaluation of F
av
with respect to an operator basis, comparing input to
output operators. Since a complete operator basis con-
sists of 2
2n
elements and the size of the event space is
given by all possible combinations of input and output
operators, T = 2
4n
. The fact that only states, not oper-
ators can be prepared as input is remedied by randomly
selecting eigenstates of the input operators. There are
6 eigenstates for the 3 Pauli operators for each qubit.
Therefore the number of experimental settings, i.e., pairs
of input state/output measurement operator, is given by
N
setting
= N
input
N
meas
= 6
n
2
2n
. The random selec-
tion of experimental settings requires classical computa-
tional resources (
class
that scale as n
2
2
4n
[2, 3]. Although
only some of the settings will be selected, the ability to
implement all of them in the experiment is implied. Due
to the statistical nature of measurements, all in all N
exp
i=1
Tr[
j,id
i
j,act
i
]
=
1
d
d
i=1
Tr[U[
j
i
j
i
[U
+
T([
j
i
j
i
[)] (3)
with [
j
i
the states of two mutually unbiased bases in d-
dimensional Hilbert space (i = 1, . . . , d, j = 1, 2, d = 2
n
),
T the dynamical map describing the actual evolution,
and U the desired unitary. Expanding the states
j,id
i
,
j,act
i
in terms of Pauli operators, Eq. (3) becomes
F
j
=
d
i=1
d
2
k=1
Pr
j
(i, k)
j
D
(i, k)
j
U
(i, k)
(4)
with characteristic function
j
U
(i, k) =
Tr[W
k
U[
j
i
j
i
[U
+
] and relevance distribution
Pr
j
(i, k) =
1
d
2
[
j
U
(i, k)]
2
. Note that Tr[W
k
W
k
] = d
k,k
.
We show in the supplementary material that Pr
j
(i, k)
is properly normalized such that we can estimate the
two classical delities F
j
, and thus an upper and a lower
bound to F
av
, by Monte Carlo sampling.
3
approach C
class
Ninput Nsetting Nexp
A O(n
2
2
4n
) 6
n
O(6
n
2
2n
) O(2
2n
)
B O(n
2
2
4n
) 2
n
(2
n
+ 1) O(2
4n
) O(2
n
)
C O(n
2
2
3n
) 2 2
n
O(2
3n
) O(2
n
)
TABLE I: Resources required for determining the average
gate error of a general unitary operation in terms of classi-
cal computational eort C
class
required for the random se-
lection, number Ninput of input states that need to be pre-
pared, the number of experimental settings Nsetting from
which the actual experiments will be randomly chosen, and
the average number Nexp of experiments to be performed.
Nsetting = Ninput Nmeas with the number of measurement
operators Nmeas = 2
2n
for all cases (A: Monte Carlo sam-
pling based on the channel-state isomorphism [2, 3]; B: Monte
Carlo sampling for two-designs; C: Monte Carlo sampling for
classical delities).
The expression for the average delity when using
two-designs, given in terms of d + 1 mutually unbiased
bases [11],
F
2des
av
=
1
d(d + 1)
d(d+1)
i=1
Tr[U[
i
i
[U
+
T([
i
i
[)],
(5)
is formally similar to Eq. (3), i.e., it can be interpreted
as the sum over d + 1 classical delities. Equation (5)
can thus be rewritten
F
2des
av
=
d(d+1)
i=1
d
2
k=1
Pr
2des
(i, k)
2des
D
(i, k)
2des
U
(i, k)
(6)
with characteristic function
2des
U
analogous to
j
U
, and
the relevance distribution diering only in normalization,
Pr
2des
(i, k) =
1
d
2
(d+1)
_
2des
U
(i, k)
2
. We show in the sup-
plementary material that also Pr
2des
(i, k) is properly nor-
malized such that F
2des
av
can be estimated by Monte Carlo
sampling.
Resources for estimating the gate error of general uni-
taries Evaluating Eqs. (4) or (6) by Monte Carlo esti-
mation involves randomly selecting L times a pair (i
l
, k
l
)
of input state/measurement operator. Compared to
Refs. [2, 3], the number of input states is signicantly
reduced for the two approaches based on state delities
in Hilbert space. This yields a correspondingly smaller
number of settings that an experimentalist needs to be
able to implement, cf. Table I. Moreover, the smaller
number of input states reduces the classical computa-
tional resources required for the random selection by
a factor 2
n
for the classical delities. This is due to
(
class
= N
input
(
single
with (
single
the classical compu-
tational cost for sampling a single state delity in Hilbert
space ((
single
n
2
2
2n
[3]). It reects the fact that the
relevance distribution for the classical delities depends
on O(d
3
) parameters whereas the relevance distribution
of Refs. [2, 3] depends on O(d
4
) parameters. The reduced
number of parameters is sucient to determine whether
the actual evolution matches the desired unitary [18].
Analogously to Refs. [2, 3], we determine the sample
size L by Chebychevs inequality. It provides an upper
bound for the probability of a random variable Z with
variance
Z
to deviate from its mean,
Pr
_
[Z Z[
Z
/
_
(7)
with > 0. In our case, Z = F with F = F
2des
av
or F
j
, Z = F
L
= 1/L
L
l=1
X
l
, and X
l
X
l
=
D
(i
l
, k
l
)/
U
(i
l
, k
l
), cf. Eqs. (4), (6). We show in the
supplementary material that the variance of X
l
is smaller
than one, and thus var(F
L
) 1/L. Then the choice
L = 1/(
2
) guarantees that the probability for the esti-
mate F
L
to dier from F by more than is smaller than
. Specifying the experimental inaccuracy and choosing
the condence level thus determines the sample size.
In order to estimate the number of required experi-
ments, we rst determine the number of experiments for
one setting, N
l
. For each l, the observable W
k
l
has to be
measured N
l
times to account for the statistical nature
of the measurement. The corresponding approximation
to X
l
is given by
X
l
=
1
U
(i
l
, k
l
)
1
N
l
N
l
j=1
w
lj
(8)
with w
lj
the measurement result for the jth repetition of
experimental setting l, equal to either +1 or -1 for Pauli
operators. Since
X
l
is given as the sum of independent
random variables w
lj
, N
l
can be determined using Hoe-
dings inequality. It provides an upper bound for the
probability of a sum S =
n
i=1
Y
i
of independent vari-
ables Y
i
with a
i
Y
i
b
i
to deviate from its expected
value by more than ,
Pr ([S S[ ) 2 exp
_
2
2
n
i=1
(b
i
a
i
)
2
_
(9)
> 0. In our case, S =
F
L
=
1
L
L
l=1
X
l
and, using
Eq. (8),
n
i=1
(b
i
a
i
)
2
=
L
l=1
4N
l
[LN
l
U
(i
l
, k
l
)]
2
.
Inserting this into Eq. (9), it is obvious that the choice
N
l
=
2
L[
U
(i
l
, k
l
)]
2
log
_
2
_
= N
l
(i
l
, k
l
) (10)
ensures the right-hand side of Eq. (9) to be . The
setting l is chosen with probability Pr
j/2des
(i
l
, k
l
). The
average number of times that this specic experiment
(with input state i
l
and measurement operator W
k
l
) is
4
approach C
class
Ninput Nsetting Nexp
A O(1) 6
n
O(6
n
2
n
) O(1)
B O(1) 2
n
(2
n
+ 1) O(2
3n
) O(1)
C O(1) 2 2
n
O(2
2n
) O(1)
TABLE II: Resources required for determining the average
gate error of a Cliord gate (Nsetting = Ninput 2
n
). Symbols
as in Table I.
carried out is therefore given by
N
l
=
d
i
l
=1
d
2
k
l
=1
Pr
j
(i
l
, k
l
) N
l
(i
l
, k
l
) (11)
=
1
d
2
d
i
l
=1
d
2
k
l
=1
[
j
U
(i
l
, k
l
)]
2
4
[
j
U
(i
l
, k
l
)]
2
L
2
log
_
2
_
1 +
2d
L
2
log
_
2
_
for the two classical delities (j = 1, 2). The same N
l
l=1
N
l
L
_
1 +
2d
L
2
log
_
2
__
1 +
1
+
2d
2
log
_
2
_
. (12)
This number is sucient to account for both the sampling
error due to nite L and statistical experimental errors in
the measurement results. Notably, N
exp
2
n
only, i.e.,
the average number of experiments to estimate F
av
scales
like that required for characterizing a general pure quan-
tum state [3]. This represents a reduction by a factor 2
n
compared to Refs. [2, 3], cf. Table I. These savings come
at the expense of (i) obtaining only bounds on the aver-
age delity when using two classical delities F
j
or (ii)
the necessity to prepare entangled input states when us-
ing two-designs. The latter scales quadratically in n [11].
Even factoring this additional cost in, Monte Carlo esti-
mation of the average delity for a general unitary opera-
tion using two-designs is signicantly more ecient than
that based on the channel-state isomorphism [2, 3].
Resources for Monte Carlo estimation of Cliord gates
The scaling of N
exp
with the number of qubits changes
dramatically for Cliord gates [2, 3]. This is due to
the property of Cliord gates to map eigenstates of a d-
dimensional set of commuting Pauli operators into eigen-
states from the same set. The mutually unbiased bases
in Eqs. (3), (5) can be chosen to be such eigenstates [20].
Given a generic eigenstate [
i
of a commuting set J of
Pauli operators, the characteristic function of a Cliord
gate, U
Cl
, becomes
U
Cl
(i, k) = Tr[W
k
U
Cl
[
i
i
[U
Cl
] (13)
= Tr[W
k
[
j
j
[] =
_
1 if W
k
J
0 otherwise
.
The relevance distribution for Cliord gates, Pr(i, k)
[
U
Cl
(i, k)]
2
, is thus zero for many settings and uniform
otherwise. Since settings with Pr(i, k) = 0 will never be
selected, the sampling complexity becomes independent
of system size. Calculating N
l
according to Eq. (11) for
a uniform relevance distribution, and accounting for the
correct normalizations of Pr(i, k), N
l
is found to be in-
dependent of d, N
l
1 +2 log(2/)/(L
2
), for all three
approaches. Consequently, also N
exp
does not scale
with system size, N
exp
1+1/(
2
)+2 log(2/)/
2
, cf.
Table II. For Cliord gates, the three approaches require
therefore a similar, size-independent number of measure-
ments. A dierence is found, however, for the number
of possible experimental settings. For each input state
i, there are only d (instead of d
2
) measurement opera-
tors W
k
with non-zero expectation value. This leads to
N
setting
= N
input
2
n
for Cliord gates, cf. Table II.
The larger N
setting
required for approaches A and B in
Table II comes with a potentially higher accuracy of the
estimate which is, however, limited by the experimental
error of state preparation and measurement.
Conclusions We nd the number of measurements re-
quired to estimate the gate error for a general unitary to
scale as 2
n
for n qubits. Our reduction by a factor of 2
n
compared to the best currently available approach [2, 3]
comes at the expense of either determining bounds from
two classical delities instead of the average delity itself
or allowing for entangled input states. For the classi-
cal delities, the number of experimental settings that
one needs to be able to prepare and the classical com-
putational resources required for the sampling are also
reduced by a factor 2
n
. All three approaches are signi-
cantly more ecient than traditional process tomography
requiring of the order 2
4n
measurements and a computa-
tional cost scaling as 4
6n
. For the special case of Cliord
gates, we nd the number of experiments to be indepen-
dent of system size, just as in Monte Carlo estimation
based on the channel-state isomorphism [2, 3] and ran-
domized benchmarking [6, 7].
We have shown earlier [18] that the minimum number
of pure input states for device characterization is of the
order 2
n
. This corresponds to the number of states re-
quired by the classical delities. Monte Carlo sampling of
two classical delities therefore realizes a strategy of min-
imal resources. Our comprehensive classication should
allow an experimentalist to choose the most suitable pro-
cedure to determine the average delity, dened in terms
of the number of experimental settings, from which a
Monte Carlo procedure randomly draws realizations, and
the actual number of experiments to be carried out.
5
We would like to thank Daniel Burgarth, Tommaso
Calarco and David Licht for helpful comments. Finan-
cial support from the EC through the FP7-People IEF
Marie Curie action Grant No. PIEF-GA-2009-254174 is
gratefully acknowledged.
k=1
i
[ W
k
[
j
n
[ W
k
[
m
= d
im
jn
(1)
with [
j
the canonical basis. Note that for each
pair [
i
, [
j
, there are exactly d operators W
k
with
i
[ W
k
[
j
, = 0. This is seen easily in the bit represen-
tation (W
k
=
1
k
. . .
N
k
). For the mth qubit, the
scalar product vanishes if there is a bit ip between the
two states at the mth qubit and
m
k
= 11
2
or
z
. Analo-
gously, the scalar product vanishes if the mth qubit has
the same value between the two states and
m
k
=
x
or
y
. There are thus only two choices of W
k
for each qubit
that lead to a non-zero scalar product. Repeating the
argument over all n qubits gives exactly 2
n
= d possible
operators W
k
for which
i
[ W
k
[
j
, = 0.
Consider now
i
[ W
k
[
j
n
[ W
k
[
m
for a certain
W
k
with [
i
,= [
m
and [
j
, [
n
xed. Since [
i
,=
[
m
there exists a qubit, l, where the two states dier.
We dierentiate two cases for this qubit:
1. [
j
and [
i
take the same value on the
lth qubit. Then
l
k
must be 11
2
or
z
for
i
[ W
k
[
j
n
[ W
k
[
m
not to vanish. How-
ever, there exists an operator W
k
such that the
contribution of the two operators to the sum,
Eq. (1),
i
[ W
k
[
j
n
[ W
k
[
m
+
i
[ W
k
[
j
n
[ W
k
[
m
,
vanishes. This operator is identical to W
k
except
that
l
k
= 11 if
l
k
=
z
and vice versa. Then
i
[ W
k
=
i
[ W
k
and W
k
[
m
= W
k
[
m
i
[ W
k
[
j
n
[ W
k
[
m
not to vanish. Again,
there exists an operator W
k
such that the contribu-
tion of the two operators to the sum, Eq. (1), van-
ishes. This operator is identical to W
k
except that
l
k
=
x
if
l
k
=
y
and vice versa. If
l
i
_
= [1
(and thus
l
m
_
= [0),
i
[ W
k
= i
i
[ W
k
and W
k
[
m
= iW
k
[
m
.
Otherwise, if
l
i
_
= [0 (and thus
l
m
_
= [1),
i
[ W
k
= i
i
[ W
k
and W
k
[
m
= iW
k
[
m
.
In both cases, the terms in the sum cancel.
Consequently, for each W
k
and [
i
, [
j
, [
m
, [
n
with [
i
, = [
j
there exists a pair operator W
k
d
2
k=1
i
[ W
k
[
j
n
[ W
k
[
m
= 0 if [
i
,= [
m
. Re-
peating the argument for [
j
,= [
n
leads to
d
2
k=1
i
[ W
k
[
j
n
[ W
k
[
m
=
im
jn
d
2
k=1
i
[ W
k
[
j
n
[ W
k
[
m
=
im
jn
d
2
k=1
[
i
[ W
k
[
j
[
2
using Hermiticity of W
k
in the last step. Now validity
of Eq. (1) follows simply from the fact that there exist,
for each pair [
i
, [
j
, exactly d operators W
k
with non-
vanishing
i
[ W
k
[
j
, and, in the canonical basis, these
matrix elements are equal to one.
Expanding a general vector [ in the canonical basis
and using Eq. (1), we nd our second intermediate result,
7
d
2
k=1
[[ W
k
[ [
2
=
d
2
k=1
i,j=1
c
i
c
j
i
[ W
k
[
j
2
=
d
2
k=1
d
i,j,n,m=1
c
i
c
j
c
n
c
m
i
[ W
k
[
j
n
[ W
k
[
m
=
d
i,j,n,m=1
c
i
c
j
c
n
c
m
d
im
jn
= d
_
d
i=1
[c
i
[
2
_
= d . (2)
It is now straightfoward to prove normalization of
Pr
j
(i, k), starting from the denition
Pr
j
(i, k) =
1
d
2
j
U
(i, k)
2
=
1
d
2
Tr[
j,id
i
W
k
]
2
and
j,id
i
= U
j
i
U
+
= U[
j
i
j
i
[U
+
. Then
d
i=1
d
2
k=1
Pr
j
(i, k) =
1
d
2
d
i=1
d
2
k=1
Tr
_
W
k
U
+
[
j
i
j
i
[U
_
2
=
1
d
2
d
i=1
d
2
k=1
d
n,m=1
n
[UW
k
U
+
[
j
i
j
i
[
n
m
[
j
i
j
i
[U
+
W
k
U[
m
=
1
d
2
d
i=1
d
2
k=1
j
i
[UW
k
U
+
[
j
i
j
i
[U
+
W
k
U[
j
i
=
1
d
2
d
i=1
d = 1 , (3)
where we have used Eq. (2) in the last line with [ =
U[
j
i
.
The relevance distribution for two-designs,
Pr
2des
(i, k) is normalized.
The normalisation of Pr
2des
(i, k) is veried analo-
gously to the previous subsection,
d(d+1)
i=1
d
2
k=1
1
d
2
(d + 1)
Pr
2des
(i, k) =
1
d
2
(d + 1)
d(d+1)
i=1
d
2
k=1
Tr
_
W
k
U [
i
i
[ U
+
2
(4)
=
1
d
2
(d + 1)
d(d+1)
i=1
d
2
k=1
i
[U
+
W
k
U[
i
i
[UW
k
U
+
[
i
=
1
d
2
(d + 1)
d(d+1)
i=1
d = 1 ,
using again Eq. (2) with [ = U[
i
. The variance of X
l
is smaller than one.
X
l
, the random variable of the top level of sampling
in the Monte Carlo estimation of the average delity,
8
has been dened in the main paper as the ratio of the
measurement outcomes for the actual state and the ideal
state,
X
l
=
D
(i
l
, k
l
)
U
(i
l
, k
l
)
.
We rst show that the variance of each X
l
in the estima-
tion of the classical delities is not too large,
Var (X
l
) = E
_
X
2
l
_
E(X
l
)
2
=
d
i=1
d
2
k=1
Pr
j
(i, k)
_
j
D
(i
l
, k
l
)
j
U
(i
l
, k
l
)
_
2
_
_
d
i=1
d
2
k=1
Pr
j
(i, k)
j
D
(i
l
, k
l
)
j
U
(i
l
, k
l
)
_
_
2
=
1
d
2
d
i=1
d
2
k=1
[
j
D
(i
l
, k
l
)]
2
1
d
2
_
d
i=1
Tr
_
U[
j
i
j
i
[U
+
T
_
[
j
i
j
i
[
__
_
2
=
1
d
2
d
i=1
d
2
k=1
Tr
_
W
k
T
_
[
j
i
j
i
[
__
2
F
2
j
(5)
with F
j
the classical delities (j = 1, 2) dened in Eq. (5)
of the main paper. Since 0 F
j
1, 0 F
2
j
1.
The same is true for the rst term. This can be seen
as follows. Each term T([
i
i
[) can be written as a
density matrix
i
,
T([
i
i
[) =
i
=
d
n=1
(i)
n
[
(i)
n
(i)
n
[ ,
with eigenvectors [
(i)
n
and eigenvalues
(i)
n
. Evaluating
the trace for each i in the corresponding eigenbasis yields
1
d
2
d
i=1
d
2
k=1
Tr [W
k
T([
i
i
[)]
2
=
1
d
2
d
i=1
d
2
k=1
n,m=1
(i)
m
[W
k
(i)
n
[
(i)
n
(i)
n
[
(i)
m
2
=
1
d
2
d
i=1
d
2
k=1
n=1
(i)
n
[W
k
(i)
n
[
(i)
n
1
d
2
d
i=1
d
n=1
_
(i)
n
_
2
d
2
k=1
(i)
n
[W
k
[
(i)
n
2
.
Using Eq. (2) with [ = [
(i)
n
and Tr[
2
i
] =
d
n
_
(i)
n
_
2
1, we obtain
1
d
2
d
i=1
d
2
k=1
[ Tr [W
k
T([
i
i
[)][
2
1
d
d
i=1
d
n=1
_
(i)
n
_
2
1
d
d
i=1
1 = 1 . (6)
Hence Var (X
l
) is the dierence between two numbers in
the interval [0, 1] and therefore smaller than one. Just as
Eq. (4) was derived by replacing in Eq. (3) the summation
limit of d by d(d+1) in the sum over states i and utilizing
the normalization of Pr
2des
, the proof of Var (X
l
) 1 for
the two-designs proceeds analogously to Eqs. (5) and (6).