Programmable projective measurement with linear optics

Ulysse Chabaud1 ,∗ Eleni Diamanti1 , Damian Markham1 , Elham Kashefi1,2 , and Antoine Joux3†
1
Laboratoire d’Informatique de Paris 6, CNRS,
Sorbonne Université, 4 place Jussieu, 75005 Paris
2
School of Informatics, University of Edinburgh, 10 Crichton Street, Edinburgh, EH8 9AB and
3
Chaire de Cryptologie de la Fondation SU, Sorbonne Université,
Institut de Mathématiques de Jussieu – Paris Rive Gauche, CNRS, INRIA,
Université Paris Diderot, Campus Pierre et Marie Curie, 4 place Jussieu, 75005 Paris
(Dated: May 8, 2018)
The swap test is a central primitive in quantum information, providing an efficient way of com-
paring two unknown quantum states. In this work, we generalise this test for the case where one
has access to multiple copies of a reference state, but to only one copy of the tested state. We
present a circuit implementing this test and propose a simple linear optical implementation using
arXiv:1805.02546v1 [quant-ph] 7 May 2018

the Hadamard interferometer. We show that both the circuit and the interferometer can be used as
programmable projective measurement devices, with a precision scaling with their size. This also
leads to a natural interpretation of the notion of projective measurement in quantum mechanics.

I. INTRODUCTION
Quantum information features various non-classical
properties, including for instance the non-cloning the-
orem [1, 2]. As a result, quantum information has to
be handled differently than classical information. For
example, the simple task of discriminating between two
classical strings of n bits can be done by a straightfor-
ward comparison in at most n steps, while comparing
two n qubits states is non trivial. Indeed, even assuming
that the complete classical descriptions of such two quan-
tum states were available, comparing these descriptions
would take up to 2n steps. Moreover, as soon as one of
the states is unknown, any deterministic discrimination
is hopeless due to the probabilistic nature of quantum
information [3].
The swap test, introduced in [4], provides an effi-
cient probabilistic tool to compare two unknown quan-
tum states. It takes as input two quantum states |φi and
|ψi that are not entangled and outputs 1 with probabil-
ity 12 + 12 | hφ|ψi |2 and 0 with probability 21 − 12 | hφ|ψi |2 ,
where hφ|ψi is the overlap between the states |φi and |ψi.
When the measurement outcome is 0 (resp. 1), we con-
clude that the states were identical, up to a global phase
(resp. different). A circuit implementing the swap test is
represented in Fig. 1, where the matrix
 
1 1 1
H=√ (1)
2 1 −1
is a Hadamard matrix. The swap test is a natural tool
when it comes to quantum state comparison, as it per-
forms a projection onto the symmetric subspace of the
input states Hilbert space [5–7]. It finds application in
various quantum communication protocols [4, 8, 9]. The
product test [10–12] uses the swap test as a subroutine for
∗ Electronic address: ulysse.chabaud@gmail.com
† Electronic address: antoine.joux@m4x.org
H H
SWAP
FIG. 1: Circuit representation of a swap test. The ancilla
qubit is measured in the computational basis.
entanglement detection. Moreover, the swap test allows
estimating the overlap of two states [13]. Hence, given
two vectors of classical data encoded into two quantum
states, the distance between these vectors can be esti-
mated using the swap test, which is useful in particular
for quantum machine learning [14].
From a practical point of view, it has been shown that
linear optics can be used to implement the swap test,
using the Hong-Ou-Mandel effect [15]. The linear op-
tics platform is particularly attractive in quantum infor-
mation. Indeed, it is possible to implement a universal
quantum computer solely with linear optical tools [16].
Furthermore, the Boson Sampling model – a linear opti-
cal model for non-universal quantum computing – is an
outstanding candidate for demonstrating the power of
quantum computers in the near-term [17], and a recent
proposal for demonstrating quantum superiority also re-
lies on linear optics [18]. In the optical implementation
proposed in [15], two single photons in different spatial
modes are sent at the same time through a beamsplitter.
Measuring the intensity of the output modes gives statis-
tics that can be post-processed in order to reproduce the
statistics of a swap test, where the states to be compared
are the states of the single photons. Two indistinguish-
able photons will always pass the test, while two photons
whose states have an overlap x < 1 will pass the test with
probability 12 + 21 x.
Despite the wide range of applications of the swap
test and the potential for its implementation, a difficulty

comes from the fact that for many applications it is nec-
essary to use multiple copies of both input states. Indeed,
the swap test satisfies the one-sided error requirement,
i.e. if the input states are identical, the test will always
declare them as identical. On the other hand, if the input
states are different, the test can obtain a wrong conclu-
sion and declare the states identical. The probability that
this happens is strictly less than 1, hence by repeating
the test various times, the probability that the sequence
of tests never answers 0 can be brought down arbitrarily
close to 0, exponentially fast. However, the swap test
is destructive, in the sense that the tested states cannot
be reused for a new test because they become maximally
entangled during the test [15]. This means that in or-
der to boost the correctness of the test when the input
states are different, multiple copies of both states must
be available.
In this work, we study a generalised swap test that
we call swap test of order M , where one has access to
M − 1 copies of a reference state, but to only a single
copy of the other tested state. In Sec. II we describe a
circuit implementing the swap test of order M when M
is a power of 2, and we show the optimality of this test
under the one-sided error requirement. In Sec. III, we
further show that the corresponding circuit can be used
as a programmable projective measurement device, with
a precision scaling with the inverse of its size. Generalis-
ing the equivalence between the swap test and the Hong-
Ou-Mandel effect [15], we study in Sec. IV the Hadamard
interferometer, and show that its output statistics can be
efficiently post-processed to reproduce those of the swap
test of order M , thus providing an easily implementable
programmable projective measurement device, with the
aforementioned precision. For completeness, we intro-
duce in Sec. V a general family of interferometers which
reproduce the appropriate statistics when M is not nec-
essarily a power of 2. We conclude with an interpretation
of our results and discuss various applications in Sec. VI.
II. SWAP CIRCUIT OF ORDER M
Let M ≥ 2. We introduce the following generalisation
of the swap test:
Definition 1. The swap test of order M is a binary test
that takes as input a state |φi and M − 1 copies of a state
|ψi, and outputs 1 with probability M1
+ MM−1 | hφ|ψi |2 and
M −1
0 with probability ( M )(1 − | hφ|ψi |2 ). If the outcome
1 (resp. 0) is obtained, the test concludes that the states
|φi and |ψi were identical (resp. different).
Such a test clearly satisfies the one-sided error require-
ment, i.e. if |φi = |ψi then the test will always say so.
In the following, we restrict to the swap test of or-
der M when M is a power of 2, writing n = log M .
We introduce the swap circuit of order M (Fig. 2), that
acts on M input qubits by applying n consecutive lay-
ers of products of swap gates controlled by n ancilla
2
H H
H H
H H
S0
S1
Sn-1
FIG. 2: Swap circuit of order M . The unitaries Sk are tensor
products of swap gates described in the main text (2). The
ancilla qubits are measured in the computational basis at the
end of the computation. The probability of obtaining 0 for
all measurement outcomes is M 1
+ MM−1 | hφ|ψi |2 .
qubits. These ancilla qubits are first initialised in the
|+i state. Then, they are used as control qubits for the
gates S0 , . . . , Sn−1 , which can be applied in any order,
where for all k ∈ {0, . . . , n − 1}
O
SWAP j2k+1 + i, j2k+1 + 2k + i ,
 
Sk =
i∈[0,2k −1],
j∈[0,2n−k−1 −1]
(2)
with SWAP[i, j] being the unitary operation that swaps
the ith and j th qubits for i, j ∈ {0, . . . , M − 1}.
These controlled gates are applied to the input states
|φi , |ψi , . . . , |ψi (one copy of a state |φi and M −1 copies
of a state |ψi). Finally, a Hadamard gate is applied to
each ancilla, which is then measured in the computational
basis. By a simple induction, we obtain that the proba-
bility of obtaining the outcome 0 for all ancilla qubits is
the squared norm of the following state:
1
(|φψ . . . ψi + |ψφ . . . ψi + · · · + |ψ . . . ψφi), (3)
M
which only depends on the overlap between the states |φi
and |ψi. More precisely,
1 M −1
Pr(0, . . . , 0) = + | hφ|ψi |2 . (4)
M M
The swap circuit of order M thus implements the swap
test of order M . Indeed, if the outcome (0, . . . , 0) is ob-
tained, the test outputs 1 and we conclude that the states
were identical, while for any other outcome the test out-
puts 0 and we conclude that the states were different.
Note that in the case where M = 2, the scheme reduces
to the original swap test.

Because the M − 1 last input qubits are in the same
state, swapping them acts as the identity. This can be
used to simplify the swap circuit of order M by replacing
the n layers of swap gates in Eq. (2) by the following
n layers S00 , . . . , Sn−1
0
, which have to be applied in this
order:
k
2O −1
Sk0 SWAP l2n−k , (l + 1)2n−k − 1 .
 
= (5)
l=0
This reduces the total number of swap gates from M log 2
M
to M − 1 without changing the number of ancilla qubits.
We now prove the optimality of the swap test of order
M under the one-sided error requirement, i.e. we show
that it achieves the lowest error probability for the iden-
tity testing of an input |φi , |ψi , . . . , |ψi. This identity
testing is directly linked to the state comparison of |φi
and |ψi. For this purpose, we first derive a more general
result. In Ref. [19], the authors consider the problem
of testing the identity of M quantum states with the
promise that all the states are pairwise identical or or-
thogonal. In particular, they show that the optimal value
for the error probability of any identity test satisfying the
one-sided error requirement is M 1
. We extend their result
to the case where the states to be compared are no longer
supposed pairwise identical or orthogonal:
Theorem 1. Under the one-sided error requirement,
any identity test of M unknown quantum states
|ψ0 i , . . . , |ψM −1 i has an error probability at least
M −1
1 X Y
hψk |ψσ(k) i, (6)
M!
σ∈SM k=0
where SM is the symmetric group over {0, . . . , M − 1}.
Proof. An identity test satisfying the one-sided error
requirement can only be wrong when declaring identical
(outputting 1) states that were not identical. Hence, to
prove Theorem 1, it suffices to lower bound the probabil-
ity of outputting 1 for any identity test. This is done by
showing that the optimal identity test consists in a pro-
jection onto the symmetric subspace of the input states
Hilbert space. We detail the proof in the Appendix A.
Applying Theorem 1 with |ψ0 . . . ψM −1 i = |φψ . . . ψi
implies that the value M 1
+ MM−1 | hφ|ψi |2 is a lower bound
for the error probability of any identity test of M states
|φi , |ψi , . . . , |ψi (one copy of a state |φi and M −1 copies
of a state |ψi). With Definition 1 we directly obtain the
following result:
Corollary 1. The swap test of order M has optimal er-
1
ror probability M + MM−1 | hφ|ψi |2 under the one-sided er-
ror requirement.
The swap circuit of order M is thus optimal for quan-
tum state comparison with an input |φi , |ψi , . . . , |ψi, un-
der the one-sided error requirement, since it implements
3
the swap test of order M . In the next section, we give
an alternative interpretation of the swap circuit of order
M in terms of a programmable projective measurement
device.
III. PROGRAMMABLE PROJECTIVE
MEASUREMENT
A projective measurement with respect to a state |ψi
is a process that takes as input a state |φi and out-
puts 1 with probability | hφ|ψi |2 and 0 with probability
1 − | hφ|ψi |2 . We introduce the notion of precision of a
projective measurement.
Definition 2. Given a quantum state |ψi and  > 0, a
projective measurement with precision  is a process that
takes as input a quantum state |φi and outputs 1 with
probability P (1) and 0 with probability P (0), such that
|(| hφ|ψi |2 ) − P (1)| ≤  and |(1 − | hφ|ψi |2 ) − P (0)| ≤ .
Note that the two conditions in the previous definition
are equivalent, since P (0) + P (1) = 1. It will thus suffice
to consider e.g. the first condition.
Theorem 2. A swap circuit of order M can be used to
1
perform a projective measurement with precision M .
Proof. Since Pr(0, . . . , 0) = M 1
+ MM−1 | hφ|ψi |2 , we can
consider all the circuit except |φi as a black box in Fig. 2,
and post-process the measurement outcomes D as fol-
lows: if D = (0, . . . , 0), output 1, and output 0 otherwise
(Fig. 3). The setup now takes a single state |φi in input
and outputs 1 with probability P (1) = M 1
+ MM−1 | hφ|ψi |2 ,
and 0 with probability P (0) = 1 − P (1). We have
1
|(| hφ|ψi |2 ) − P (1)| ≤ M , hence this device performs a
1
projective measurement with precision M .
Theorem 2 implies that given a large enough swap cir-
cuit and the ability to produce many copies of a state
|ψi, one can projectively measure any state with respect
to the state |ψi up to arbitrary precision. This preci-
sion scales as the inverse of the number of copies. The
circuit can thus be used as a programmable projective
measurement device, where the programmable resource
is the reference state |ψi whose number of copies can
be adjusted to control the precision of the measurement
(Fig. 3).
The implementation of the swap circuit of order M
is however challenging, due to the presence of many
controlled-swap gates. In order to lower the implemen-
tation requirements, we study in the next section the
Hadamard interferometer and show that its statistics can
be efficiently post-processed to reproduce those of a swap
circuit of order M . This comes at the cost that the device
no longer has a quantum output, which however does not
matter for most applications. In particular we show that

Post-processing
d1
dn
Swap circuit of
order M=2n
FIG. 3: The swap circuit of order M used as a pro-
grammable projective measurement device. It takes as in-
put a state |φi and the internal measurement outcomes are
post-processed such that the device outputs 1 with probability
M
1
+ MM−1 | hφ|ψi |2 and 0 with probability MM−1 (1 − | hφ|ψi |2 ).
The programmable resource is the state |ψi and the process
uses M − 1 copies of this state as well as n = log M ancillas.
the Hadamard interferometer provides a simple linear op-
tical platform for the programmable projective measure-
ment that we have described.
IV. HADAMARD INTERFEROMETER
In what follows, we consider optical unitary interfer-
ometers of size M which take as input one single photon
in a quantum state |φi and M − 1 indistinguishable sin-
gle photons in a state |ψi, one in each spatial mode (the
spatial modes of the interferometers will be indexed from
0 to M − 1). These states should be thought of as en-
coded in additional degrees of freedom of the photons
(e.g. polarisation, time-bins, phase). The output modes
are measured using photon number detection.
There exist complex amplitudes α and β and a state
|ψ ⊥ i with hψ|ψ ⊥ i = 0 such that
|φi = α |ψi + β |ψ ⊥ i ,
where α = hψ|φi and |α|2 + |β|2 = 1. We have the fol-
lowing homomorphism property for single photon states:
|1φ i = |1αψ+βψ⊥ i = α |1ψ i + β |1ψ⊥ i ,
where for any state |χi, |1χ i is the state of a single pho-
ton encoding the state |χi. It thus suffices to compute
the output statistics separately when |φi = |ψi (indis-
tinguishable case) and when |φi = |ψ ⊥ i (distinguishable
case) to obtain the output statistics in the general case by
linearity. The probability of detecting the photon num-
ber pattern D = (d0 , . . . , dM −1 ), or equivalently that the
4
k th detector detects dk photons for all k ∈ {0, . . . , M −1},
is then
1
Pr(D) = |α|2 Pri (D) + |β|2 Prd (D)
(9)
= Prd (D) + | hφ|ψi |2 [Pri (D) − Prd (D)] ,
0 where Pri (D) is the probability in the indistinguishable
case and Prd (D) is the probability in the distinguishable
case. The single photon encoding maps quantum state
comparison to distinguishability of single photons. Note
that for any measurement outcome D = (d0 , . . . , dM −1 ),
we have d0 + · · · + dM −1 = M since an interferometer is
a passive device that does not change the total number
of photons. For any interferometer of size M , we prove
in Appendix B the following inequality:
Pri (D)
Prd (D) ≥ , (10)
M
for any detection pattern D. Combining this inequality
with Eq. (9) yields
 
1 M −1
Pr(D) ≥ + | hφ|ψi |2 Pri (D). (11)
M M
This last expression is valid for any interferometer and
can be used it to retrieve, in the context of linear op-
tics, the error probability bound for state comparison
under the one-sided error requirement obtained in Corol-
lary 1. Indeed, assume that E is a detection event, which
could be a disjoint union of detection events, used for
an identity test: if E is obtained we conclude that the
states were identical (or equivalently that the photons
were indistinguishable), otherwise we assume that the
states were different (or equivalently that the first pho-
ton was distinguishable from the others). The one-sided
error requirement can thus be written as Pri (E) = 1:
indistinguishable photons always pass the test. For dif-
ferent input states |φi and |ψi, the error probability of
the corresponding test is then given by Pr(E), which by
Eq. (11) is lower bounded by M 1
+ MM−1 | hφ|ψi |2 .
We now study a particular unitary interferometer,
when the size M is a power of 2, namely the Hadamard
interferometer [20, 21] and show that it provides a prac-
tical and simple implementation of the swap test of order
(7) M . For M = 4 spatial modes (Fig. 4), this interferometer
is described by the Hadamard-Walsh transform of order
2:
 
1 H H
(8) √ (12)
2 H −H
where H is a Hadamard matrix, see Eq. (1). In the
general case, the Hadamard interferometer of order M
is described by the Hadamard-Walsh transform of order
n = log M , which is defined by induction:
 
1 Hk Hk
Hk+1 = √ , (13)
2 Hk −Hk

FIG. 4: Hadamard interferometer with 4 input modes. The
dashed red lines represent balanced beamsplitters. The in-
put states are one single photon in state |φi and three single
photons in state |ψi, one in each mode.
with H0 = 1 and H1 = H. We can now state our main
result linking the Hadamard interferometer and the swap
test of order M .
Theorem 3. The output statistics of the Hadamard in-
terferometer of order M can be classically post-processed
in time O(M log M ) to reproduce those of the swap test
of order M .
Proof. We give hereafter an overview of the proof and
refer to the Appendix C for further details.
Due to the structure of the Hadamard-Walsh trans-
form, we are able to show that there exists a collection of
detection patterns which saturates the bound in Eq. (11)
and to characterise this collection. We introduce the
M × M matrix
√ the measurement outcomes. The post-processing consists
S = (sij )0≤i,j≤M −1 = M Hn , (14)
thus omitting the normalisation factor. The matrix S
only has +1 and −1 entries. We show that its lines, to-
gether with the element-wise multiplication, form a group
n
isomorphic to (Z/2Z) . We define for all measurement
outcomes D = (d0 , . . . , dM −1 ) the function
M −1 M −1
X Y d Using the argument developed in the proof of Theo-
π(D) = (sij ) j , (15)
i=0 j=0
and exploit the aforementioned group structure to obtain
the following equivalences:
π(D) 6= 0 ⇔ π(D) = M
⇔ Pri (D) 6= 0
(16)
Pri (D)
⇔ Prd (D) = .
M
With the first two lines, the condition π(D) = 0 is di-
rectly equivalent to having a detection event D that can
only be witnessed in the distinguishable case. In other
5
words, the detection patterns D such that π(D) = 0 can
only occur if hφ|ψi = 6 0. On the other hand, with the
third equivalence, the detection patterns D such that
π(D) 6= 0 are those that saturate the bound obtained
in Eq. (11). The Hadamard interferometer can thus be
used to compare the states |φi and |ψi: if the outcome D
obtained satisfies π(D) = M , we conclude that the states
were identical, otherwise π(D) = 0 and we conclude that
the states were different. We show in particular that the
interferometer described by the unitary matrix Hn satis-
fies
1 M −1 2
Pr[π(D) = M ] = + |hφ|ψi| , (17)
M M
and
Pr[π(D) = 0] = 1 − Pr[π(D) = M ], (18)
for any detection pattern D. Hence the identity test using
the Hadamard interferometer of order M is a swap test of
order M . The measurement outcomes D have to be post-
processed by computing π(D). Using the group structure
of the matrix S, we show that this can be done in time
O(M log M ).
Note that the group structure invoked in the proof is
preserved under permutations, so Theorem 3 also applies
to the unitary interferometers described by permutations
of the Hadamard-Walsh transform.
The conclusion to be drawn from Theorem 3 is that
as long as a state |ψi can be encoded using single pho-
tons, then one can perform a swap test of order M with
respect to the state |ψi using the Hadamard interferome-
ter of order M and an efficient classical post-processing of
in the following parity test: given the measurement out-
come D = (d0 , . . . , dM −1 ), where d0 + · · · + dM −1√= M ,
construct the matrix SD from the matrix S = M Hn
by keeping the k th column only if dk is odd. If the lines
(1, 2, 4, . . . , 2n−1 ) of SD all have an even number of −1,
output 1. Output 0 otherwise. This means that the post-
processing only requires the parity of the photon number
in each output mode.
rem 2, by considering the M − 1 identical photons and
the interferometer as a black box whose outcomes are
post-processed as described above, we also deduce the
following result from Theorem 3:
Corollary 2. The Hadamard interferometer of order M
can be used to perform a projective measurement with
1
precision M , using a classical post-processing of its mea-
surement outcomes that takes time O(M log M ) (Fig. 5).
Interestingly, the unitary interferometers described by
the Hadamard-Walsh transform and its permutations are
not the only unitary interferometers which can reproduce

d0
1 time at most M · N . The lines of FG = M UG to-
d1
Post-processing
Hn
dM-2
0
dM-1
FIG. 5: The Hadamard interferometer of order M used as a
programmable projective measurement device. A single pho-
ton in the state |φi goes through a linear interferometer along
with M − 1 indistinguishable single photons in the state |ψi.
The parity of the number of photons in each output mode
is measured and efficiently post-processed, such that the de-
vice outputs 1 with probability M1
+ MM−1 | hφ|ψi |2 and 0 with
M −1 2
probability M (1 − | hφ|ψi | ).
the statistics of a swap test with efficient post-processing,
and indeed we present a generalisation in Sec. V. How-
ever, it is the simplicity of the Hadamard interferometer
in terms of experimental implementation that motivates
our interest towards this interferometer. In particular,
this interferometer can be simply implemented with a
few balanced beamsplitters. A result by Reck et al. [22]
states that any M × M unitary interferometer can be
implemented using phase shifters and at most M (M2 −1)
beamsplitters, possibly unbalanced. For the Hadamard
interferometer, only M log M
balanced beamsplitters are
2
needed and no phase shifters. The proof of this statement
is based on a simple induction detailed in Appendix D.
V. GROUP GENERALISATION FOR ANY
VALUE OF M
The Hadamard interferometer requires the size param-
eter M to be a power of 2. This requirement can be
relaxed, possibly raising the experimental requirements
at the same time. Indeed, for any value of M , one can
associate to any abelian group of order M an interferom-
eter of size M which has the desired statistics. This is
the object of the following result that uses the invariant
factor decomposition of an abelian group:
Theorem 4. Let G be an abelian group of order M .
Then there exists N ∈ N∗ and a1 , . . . , aN ∈ N∗ , where
ai |ai+1 for i ∈ {1, . . . , N − 1} and a1 . . . aN = M , such
that the interferometer described by the M × M unitary
matrix
1 processing of the measurement outcomes, while using
UG = √ Fa1 ⊗ · · · ⊗ FaN , (19)
M
2iπ
where Fa = (e a kl )0≤k,l≤a−1 is the Quantum Fourier
Transform (QFT) of order a for all a ∈ N∗ , can per-
6
1
form a M -approximate projective measurement with a
post-processing of its measurement outcomes√that takes
gether with the element-wise multiplication form a group
isomorphic to G.
Proof. We use the notations of the Theorem. The
invariant factor decomposition of G gives
G ' (Z/a1 Z) ⊗ · · · ⊗ (Z/aN Z) , (20)
where N ∈ N∗ and a1 , . . . , aN ∈ N∗ are unique, satis-
fying ai |ai+1 for i ∈ {1, . . . , N − 1} and a1 . . . aN = M .
Given that the lines of Fa together with the element-wise
multiplication form a group isomorphic to (Z/aZ)
√ for all
a ∈ N∗ , the lines of FG = (fij )0≤i,j≤M −1 = M UG to-
gether with the element-wise multiplication form a group
isomorphic to G.
Since the group structure was the only argument in-
voked in the proof of Theorem 3, the same conclusion
can be drawn here, by following the same argument:
1 M −1 2
Pr[π(D) = M ] = + |hφ|ψi| , (21)
M M
where
M
X −1 M
Y −1
d
π(D) = (fij ) j . (22)
i=0 j=0
The group G is finitely generated by N elements, so
N lines of FG are sufficient to generate all its lines by
element-wise multiplication. The condition π(D) = M
can thus be checked in time at most M · N .
In particular, for G ' (Z/M Z), the corresponding in-
terferometer is described by the (normalised) QFT of or-
der M , while for G ' (Z/2Z)n , we retrieve Theorem 3
and the Hadamard interferometer.
VI. CONCLUSION AND DISCUSSION
We have studied a generalisation of the swap test in
the case where many copies of a reference state can be
produced, but only a single copy of the tested state is
available. We have shown that this test has optimal error
probability under the one-sided error requirement. When
the number of input states M is a power of 2, we have
introduced a simple circuit implementing the swap test
of order M .
We have proven that the Hadamard interferometer
has the same statistics up to an efficient classical post-
only number resolving detectors and balanced beamsplit-
ters, which lowers the cost for an experimental imple-
mentation of the swap test of order M . The Hadamard
and the QFT interferometers have been studied e.g.
 7

in the context of benchmarking Boson Sampling ma- Appendix A: Proof of optimality

chines [20, 21, 23], and we have generalised the corre-
sponding class of interferometers by noticing a group
structure in the lines of the unitary matrices describing
these interferometers. We have interpreted our results
in an alternative way, which opens the way for a fruitful
application: being able to produce many copies of a state
allows to perform a projective measurement with respect
to this state, with a precision scaling as the number of
copies. We provided a programmable projective measure-
ment device which can be implemented using only linear
optics.
This result also gives rise to a natural interpretation
of the notion of projective measurement in quantum me-
chanics, as a comparison between one state and several
copies of another state using an interferometer: in the
macroscopic limit, when many copies of a reference eigen-
state are available, we retrieve a macroscopic classical
detector.
The Hadamard interferometer is easily implementable,
but this comes at the cost that we are detecting all modes,
i.e. that there is no more quantum output compared to
the swap circuit of order M . However, for most appli-
cations, it is only the classical output statistics of the
circuit that matters. Our scheme provides a lower error
probability than the swap test in the context where one
has access to many copies of only one of the states, e.g.
when one of the states is easier to produce. It thus finds
application for various quantum information tasks, such
as the aforementioned photon indistinguishability bench-
marking, but also for enhancing quantum communication
protocols, quantum state comparison, and entanglement
detection.
For completeness, it could be interesting to charac-
terise the full class of interferometers that are optimal
for state comparison under the one-sided error require-
ment, as we only gave a broad class of such interferome-
ters using a group construction. We conjecture that the
Hadamard interferometer will remain the simplest to im-
plement among this class of optimal schemes. It would
be also interesting to consider the influence of real ex-
perimental conditions, as our scheme assumes that the
input states are pure. The one-sided error requirement
is also a challenge experimentally, as any interferometer
or detector would suffer from the effects of imperfection
and noise. We leave these analyses open for future work.
An identity test on a Hilbert space H is a binary test
which can be written as a positive-operator valued mea-
sure {Π0 , Π1 }, with Π0 + Π1 = I. Such a test takes as
input a pure tensor product state |ψ0 . . . ψM −1 i ∈ H⊗M
and outputs 0 with probability
P (0) = Tr[Π0 |ψ0 . . . ψM −1 i hψ0 . . . ψM −1 |], (A1)
and 1 with probability
P (1) = 1 − P (0) = Tr[Π1 |ψ0 . . . ψM −1 i hψ0 . . . ψM −1 |].
(A2)
If the output 1 is obtained we conclude that we had
|ψ0 i = · · · = |ψM −1 i, whereas if the output 0 is obtained
we conclude that the states were not all identical. The
one-sided error requirement can thus be written as
⊗M
∀ |ψi , Tr[Π0 |ψi hψ| ] = 0. (A3)
Following [24], the symmetric subspace of H⊗M can be
characterised as
S = span{|ψi
⊗M
: |ψi ∈ H}, (A4)
and the orthogonal projector onto this space can be writ-
ten as
1 X
PS = Pσ , (A5)
M!
σ∈SM
where for all σ ∈ SM and all |ψ0 . . . ψM −1 i ∈ H⊗M we
have Pσ |ψ0 . . . ψM −1 i = |ψσ(0) . . . ψσ(M −1) i. Given the
characterisation of the symmetric subspace, the one-sided
error requirement in Eq. (A3) implies that the supports
of PS and Π0 are disjoint. The support of PS is thus
included in the support of Π1 , given that Π0 + Π1 = I
and this implies in turn that Π1 ≥ PS by positivity of
Π1 .
The error probability of the identity test under the
one-sided error requirement is given by the probability
of outputting the result 1 while the states were not all
identical:
P (1) = Tr[Π1 |ψ0 . . . ψM −1 i hψ0 . . . ψM −1 |]
≥ Tr[PS |ψ0 . . . ψM −1 i hψ0 . . . ψM −1 |]
1 X
≥ Tr[Pσ |ψ0 . . . ψM −1 i hψ0 . . . ψM −1 |]
M!
σ∈SM

VII. ACKNOWLEDGEMENTS 1 X
≥ Tr[|ψσ(0) . . . ψσ(M −1) i hψ0 . . . ψM −1 |]
M!
σ∈SM
We kindly acknowledge F. Grosshans and A. Olivo M −1
1 X Y
for interesting and fruitful discussions. This work has ≥ hψk |ψσ(k) i,
been supported in part by the European Union’s H2020 M!
σ∈SM k=0
Programme under grant agreement number ERC-669891 (A6)
and by the European Research Council Starting Grant where in the third line we used the expression of the
QUSCO. orthogonal projector PS onto the symmetric subspace.
 8

Appendix B: Statistics of an interferometer In order to obtain more readable expressions, we define

for all k ∈ {0, . . . , M − 1} and for any detection pattern
Recall that we consider optical unitary interferometers D,
of size M which take as input one single photon in a quan- ( u Per(U
0k 0,D−1k )
tum state |φi and M − 1 indistinguishable single photons √ if dk 6= 0,
pk (D) = D! (B5)
in a state |ψi, one in each spatial mode, indexed from 0 0 otherwise.
to M − 1. The output modes are measured using photon
number detection. A measurement outcome thus has the Using the Laplace expansion of the permanent, the pre-
form D = (d0 , . . . , dM −1 ), with d0 + · · · + dM −1 = M . vious equations (B2, B4) rewrite
The permanent of an M × M matrix T = M −1 2
(tij )0≤i,j≤M −1 is defined by X 
Pri (D) =  dk pk (D) , (B6)
 
 
X M
Y −1 k=0
Per(T ) = tkσ(k) , (B1)
and
σ∈SM k=0
M
X −1
where SM is the symmetric group over {0, . . . , M − 1}. Prd (D) = dk |pk (D)|2 . (B7)
We now compute Pri (D) and Prd (D) for all detection k=0
patterns D. PM −1
In the indistinguishable case, M indistinguishable pho- Since k=0 dk = M , we obtain, using √ Cauchy-Schwarz

tons, one in each mode, are sent through a linear op- inequality with the complex vectors dk 0≤k≤M −1 and
tical network described by an M × M unitary matrix
√
dk pk (D) 0≤k≤M −1 ,
U = (uij )0≤i,j≤M −1 . The probability of a detection event
D can be computed (see, e.g, [17]) as Pri (D)
Prd (D) ≥ , (B8)
M
|Per(UD )|2
Pri (D) = , (B2) for any detection pattern D.
D!
where D! = d0 ! . . . dM −1 ! and where UD is the matrix
obtained from U by repeating dk times the k th column Appendix C: Proof of Theorem 3
for k ∈ {0, . . . , M − 1}.
In the distinguishable case, M − 1 indistinguishable Let us define
photons are sent in modes 1, . . . , M − 1 through a linear √
optical network described by an M × M unitary matrix S = (sij )0≤i,j≤M −1 = M Hn , (C1)
U = (uij )0≤i,j≤M −1 , along with one additional photon
in the 0th mode in an orthogonal state. Since it is fully thus omitting the normalisation factor. We have
distinguishable from the others, the additional photon √ √
behaves independently, hence the probability of detect- S = 2H ⊗ · · · ⊗ 2H , (C2)
| {z }
ing the photon number pattern D for one distinguishable n times
photon and M − 1 indistinguishable photons in input is √
where H is a Hadamard matrix. The lines of 2H, to-
M
X −1 gether with the element-wise multiplication, form a group
Prd (D) = Pri (D − 1k ) · Pri (1k ). (B3) isomorphic to Z/2Z, thus the lines of S together with the
k=0 element-wise multiplication form a group isomorphic to
dk 6=0 n
(Z/2Z) . As a consequence, multiplying element-wise all
the lines of S by its ith line for a given i amounts to
This last expression formalises the fact that the M −1 in-
permuting the lines of S. Let D = (d0 , . . . , dM −1 ) and
distinguishable photons give a detection pattern D − 1k
k ∈ {0, . . . , M − 1} such that dk 6= 0. Let also SD−1k
which, completed by the additional distinguishable pho-
be the matrix obtained from S by repeating dl times the
ton in the k th output mode, forms the pattern D. Devel-
lth column for l 6= k and dk − 1 the k th column. For all
oping this expression with Eq. (B2) yields
i ∈ {0, . . . , M − 1}, one can obtain the matrix S0,D−1k
M −1 (with the 0th line removed) from the matrix Si,D−1k
1 X (with the ith line removed) by multiplying element-wise
Prd (D) = dk |u0k Per(U0,D−1k )|2 (B4)
D!
k=0
all lines by the ith line and permuting the lines. Since
dk 6=0 the permanent is invariant by line permutation we ob-
tain, for all i ∈ {0, . . . , M − 1} and all k ∈ {0, . . . , M − 1}
where U0,D−1k is the matrix obtained from U by remov- such that dk 6= 0,
ing the 0th line, then by repeating dl times the lth column
for l 6= k and by repeating dk − 1 times the k th column. Per(Si,D−1k ) = ik (D)Per(S0,D−1k ), (C3)
 9
QM −1 d
where ik (D) = sik j=0 (sij ) j . Finally, we use We finally conclude by combining Eqs. (C10,C11) and
the Laplace row expansion formula for the permanent Eq. (9):
of SD to obtain, for all D = (d0 , . . . , dM −1 ) and all X
k ∈ {0, . . . , M − 1} such that dk 6= 0, Pr[π(D) = M ] = Pr(D)
π(D)=M
M −1 (C12)
M −1
X
Per(SD ) = sik Per(Si,D−1k ) 1 2
= + |hφ|ψi| .
i=0 M M
−1
M
!
X The post-processing mentioned in the main test,
= sik ik (D) Per(S0,D−1k ) i.e. computing π(D), can be done efficiently in
i=0 (C4)
  time O(M log M ) for any detection pattern D =
M
X −1 M
Y −1 (d0 , . . . , dM −1 ). Indeed, let SD be the M × M matrix
d
= (sij ) j  Per(S0,D−1k ) obtained from S by repeating dk times the k th column
i=0 j=0 for k ∈ {0, . . . , M − 1}. The expression π(D) in Eq. (15)
= π(D)Per(S0,D−1k ), is the sum of the product of the elements of each line of
SD . Since the entries of the matrix S are only +1 and
where we used Eq. (C3) in the second line. With the −1, π(D) = M if and only if the number of −1 on the
general expressions of Pri (D) (B2) and Prd (D) (B4), this lines of SD is even for all lines. The condition π(D) = M
equation implies can thus be written as a system of M linear equations
n
modulo 2. Since (Z/2Z) is finitely generated by n ele-
M Pri (D) = π(D)2 Prd (D). (C5)
ments, the M lines of SD can be generated with at most
With the Laplace column expansion formula for the per- n lines using element-wise multiplication, for any mea-
manent of SD and the last line of Eq. (C4), we also obtain surement outcome D. Hence, computing the parity of
the number of −1 on each line of SD , which is equivalent
M 2 Pri (D) = π(D)2 Pri (D). (C6) to testing π(D) = M , can be done by computing at most
In particular, combining Eqs. (C5,C6), n = log M parity equations, with a number of terms in
each equation which is at most M .
M 2 π(D)2 Prd (D) = π(D)4 Prd (D). (C7) A simple induction shows that a possible choice for
Now Prd (D) is non-zero for all D, since by Eq. (B4) it is a the lines whose parity has to be tested is the lines with
sum of moduli squared of permanents of (2n −1)×(2n −1) index 2k for k ∈ {0, . . . , n − 1} (the lines of the matrix
matrices, which in turn cannot vanish by a result of [25]. being indexed from 0 to M − 1).
Hence the previous equation rewrites
M π(D) = π(D)2 . (C8)
As a consequence, π(D) = M or π(D) = 0 for all D. Appendix D: The Hadamard interferometer can be
Combining Eqs. (C5,C8) we obtain implemented with a few balanced beamsplitters
π(D) 6= 0 ⇔ π(D) = M
⇔ Pri (D) 6= 0 Let Ik be the k×k identity matrix for all k. The size M
(C9) is a power of 2, with n = log M . We prove by induction
Pri (D) over n that, there exist P0 (n), . . . , Pn−1 (n) permutation
⇔ Prd (D) = ,
M matrices of order M/2, such that
and thus
X n−1
Y
Pk (n) IM/2 ⊗ H Pk (n)T .


Pri [π(D) = M ] = Pri (D) Hn = (D1)
π(D)=M k=0
X
= Pri (D) (C10) Since multiplying matrices is equivalent to setting up ex-
Pri (D)6=0 perimental devices in sequence, and given that H is the
= 1. matrix describing a balanced beamsplitter, Eq. (D1) im-
plies the result we want to prove.
We also obtain For n = 1, we have M = 2 and Eq. (D1) is true with
X P0 (1) = I1 . For brevity, we define for all k
Prd [π(D) = M ] = Prd (D)
π(D)=M H (k) = Ik ⊗H. (D2)
1 X
= Pri (D) (C11) Assuming that Eq. (D1) is true for n, we use the recursive
M
π(D)=M definition of the Hadamard-Walsh transform
1
= . Hn+1 = H ⊗ Hn , (D3)
M
 10

along with properties of the tensor product of matrices where Q is a permutation matrix of order M and where in
in order to obtain the third line we have used Eq. (D1). Setting Pk (n+1) =
Q (I2 ⊗ Pk (n)) for k ∈ {0, . . . , n − 1} and Pn (n + 1) =
Hn+1 = (Hn ⊗ I2 ) H (M ) = Q (I2 ⊗ Hn ) QT H (M ) IM proves Eq. (D1) for n + 1, since these matrices are
permutation matrices of order M . This completes the
" n−1
#
Y
(M/2)
= Q I2 ⊗ Pk (n)H Pk (n) QT H (M )
T
induction and the proof of the result.
k=0
"n−1 #
Y
(M ) T
QT H (M )


=Q (I2 ⊗ Pk (n)) H I2 ⊗ Pk (n)
k=0
n−1
Y T
= [Q (I2 ⊗ Pk (n))] H (M ) [Q (I2 ⊗ Pk (n))] H (M ) ,
k=0
(D4)

[1] W. K. Wootters and W. H. Zurek, Nature 299, 802
(1982).
[2] D. Dieks, Physics Letters A 92, 271 (1982).
[3] A. Chefles, Contemporary Physics 41, 401 (2000).
[4] H. Buhrman, R. Cleve, J. Watrous, and R. De Wolf,
Physical Review Letters 87, 167902 (2001).
[5] S. M. Barnett, A. Chefles, and I. Jex, Physics Letters A
307, 189 (2003).
[6] M. Kleinmann, H. Kampermann, and D. Bruss, Physical
Review A 72, 032308 (2005).
[7] A. Montanaro and R. de Wolf, arXiv preprint
arXiv:1310.2035 (2013).
[8] J. N. de Beaudrap, Physical Review A 69, 022307 (2004).
[9] N. Kumar, E. Diamanti, and I. Kerenidis, Physical Re-
view A 95, 032337 (2017).
[10] F. Mintert, M. Kuś, and A. Buchleitner, Physical Review
Letters 95, 260502 (2005).
[11] S. Walborn, P. S. Ribeiro, L. Davidovich, F. Mintert, and
A. Buchleitner, Nature 440, 1022 (2006).
[12] A. W. Harrow and A. Montanaro, Journal of the ACM
(JACM) 60, 3 (2013).
[13] A. K. Ekert, C. M. Alves, D. K. Oi, M. Horodecki,
P. Horodecki, and L. C. Kwek, Physical review letters
88, 217901 (2002).
[14] S. Lloyd, M. Mohseni, and P. Rebentrost, arXiv preprint
arXiv:1307.0411 (2013).
[15] J. C. Garcia-Escartin and P. Chamorro-Posada, Physical
Review A 87, 052330 (2013).
[16] E. Knill, R. Laflamme, and G. J. Milburn, Nature 409,
46 (2001).
[17] S. Aaronson and A. Arkhipov, Theory of Computing 9,
143 (2013).
[18] J. M. Arrazola, E. Diamanti, and I. Kerenidis, arXiv
preprint arXiv:1711.02200 (2017).
[19] M. Kada, H. Nishimura, and T. Yamakami, Journal of
Physics A: Mathematical and Theoretical 41, 395309
(2008).
[20] A. Crespi, Physical Review A 91, 013811 (2015).
[21] A. Crespi, R. Osellame, R. Ramponi, M. Bentivegna,
F. Flamini, N. Spagnolo, N. Viggianiello, L. Innocenti,
P. Mataloni, and F. Sciarrino, Nature communications
7, 10469 (2016).
[22] M. Reck, A. Zeilinger, H. J. Bernstein, and P. Bertani,
Physical review letters 73, 58 (1994).
[23] N. Viggianiello, F. Flamini, M. Bentivegna, N. Spagnolo,
A. Crespi, D. J. Brod, R. Osellame, and F. Sciarrino,
arXiv preprint arXiv:1710.03578 (2017).
[24] A. W. Harrow, arXiv preprint arXiv:1308.6595 (2013).
[25] R. Simion and F. W. Schmidt, Discrete Mathematics 46,
107 (1983).