Generating entanglement with linear optics

Stasja Stanisic,1, 2 Noah Linden,3 Ashley Montanaro,3 and Peter S. Turner1

1
Quantum Engineering Technology Labs, H. H. Wills Physics Laboratory and
Department of Electrical & Electronic Engineering, University of Bristol, UK
2
Quantum Engineering Centre for Doctoral Training, University of Bristol, UK
3
School of Mathematics, University of Bristol, UK
(Dated: February 20, 2017)
Entanglement is the basic building block of linear optical quantum computation, and as such
understanding how to generate it in detail is of great importance for optical architectures. We
prove that Bell states cannot be generated using only 3 photons in the dual-rail encoding, and give
strong numerical evidence for the optimality of the existing 4 photon schemes. In a setup with
a single photon in each input mode, we find a fundamental limit on the possible entanglement
between a single mode Alice and arbitrary Bob. We investigate and compare other setups aimed
at characterizing entanglement in settings more general than dual-rail encoding. The results draw
attention to the trade-off between the entanglement a state has and the probability of postselecting
that state, which can give surprising constant bounds on entanglement even with increasing numbers
of photons.
arXiv:1702.05209v1 [quant-ph] 17 Feb 2017

I. INTRODUCTION deterministic, and the probability of success is often low;

for example, the best known Bell state generation scheme
Research into quantum technologies has gained signif- has success probability of 1/4 [12] and if the postprocess-
icant momentum in the last several years, with appli- ing technique known as procrustean distillation is not
cations ranging across metrology, communications, secu- allowed, then the probability drops to 0.1875 [13]. When
rity, simulation and computation [1–4]. One of the im- we consider the number of Bell states needed to construct
portant resources lying behind many of these advances two-dimensional cluster states [11], the requirements can
is quantum entanglement [5, 6]. Long before it was a be quite daunting, though promising proposals exist [14].
potential technological resource, entanglement was stud- This helps to motivate the study of entanglement gener-
ied as one of the phenomena lying at the foundations of ation in linear optics more generally; in particular, it is
quantum mechanics [7–9]. That there exist non-classical natural to consider entanglement between two subsets of
correlations between physical systems is now well estab- modes, foregoing encoding altogether. While this is cur-
lished, while how best to generate, verify and quantify rently not the preferred way of generating entanglement,
such entangled states in practice is an ongoing field of any bounds that can be found present fundamental limits
activity. What is practical in any given situation de- on linear optical architectures, as well as for other quan-
pends on the physical platform under consideration; here tum information processing tasks such as boson sampling
we will be interested in the generation of entanglement [15].
using linear optics and postselection. In this paper we will consider two main themes regard-
In linear optics we study collections of optical modes, ing bipartite entanglement in linear optics; that where
modelled as harmonic oscillators whose excitations cor- the parts are encoded qubits, and that where they are
respond to photons. Interactions are restricted to Hamil- collections of modes. Section II introduces the back-
tonians that leave the total number of photons fixed, ground and notation used throughout. Section III exam-
giving rise to unitary transformations on modes (inter- ines qubit entanglement within the standard linear opti-
ferometers), as well as possible measurement and posts- cal dual-rail encoding. When we speak of dual-rail encod-
election of quantum states (heralding). This realization ing, we mean qubit states that are post-selected such that
introduces an interesting set of constraints on the en- there is exactly one photon in each pair of modes. First
tanglement problem. Most work to date focuses on ei- we prove that one cannot generate a Bell state using only
ther single- or dual-rail encoding of photons into two- 3 photons, and then we give strong numerical evidence
dimensional qubits, and then applying the usual ap- for the known 4 photon Bell state generator (with a suc-
proaches to quantum computation such as the circuit cess probability of 0.1875) being optimal. In Section IV
model or measurement-based schemes. Gates are car- we compare qubit and mode entanglement, including an
ried out via ancilla modes and photon detection measure- investigation of the expected average entanglement over
ments [4]. The dual-rail encoding, where qubits are real- uniformly (Haar) distributed interferometers. In Section
ized as single photons in pairs of spatial or polarization V we shift our focus to mode entanglement, considering
modes, is the commonly accepted standard for quantum bipartite systems made from two sets of optical modes,
computation with linear optics, and allows us to discuss Alice and Bob, with a fixed total number of photons. We
entanglement in terms of standard concepts such as Bell see two types of behaviour. In the case of bunched photon
and GHZ states [4, 10, 11]. However, the requirement input and single mode Alice, we find the entanglement
of postselection means generation of such states is non- can grow as log n where n is the number of photons. On
 2

Bound (ebits) Parameters Input state Section

O(log n) MA = 1, MB = 1, MH = 0 Bunched VA1
2 MA = 1, MB ≥ 1, MH = 0 Unbunched VA2
log 
3 MA = MB = 1, MH ≥ 1 Unbunched VA3
+ n−1

log 2 MAM 2 MA = MB , n odd Any VB1
 A 
M + n −1

log 2 n+M
n
A A
M
2 MA = MB , n even Any VB1
A

n MA = MB Fock state V B 2

TABLE I. Entanglement bounds proven in this paper. The notation is as defined in Section II (see Figure 1).

the other hand, looking at the case of at most a single U(M ). The top MI input modes contain n input pho-
photon per input mode (as in, for example, boson sam- tons, while the bottom MV modes are ancilla vacua. The
pling [15]), a single mode Alice and no measurement, the representation of U carried by the n photon, M mode
entanglement is bounded above by 2 ebits regardless of Hilbert space in the number state (Fock) basis is de-
how many photons are present. If we also restrict Bob noted U (n) . The top MA output modes belong to one
to a single mode and furnish the remaining modes with party, Alice, the middle MB modes belong to Bob, and
number resolving detectors, the expected entanglement the bottom MH modes – Harold – get measured using
is bounded by log 3 ebits. We then find provable uni- photon counting detectors. Harold’s detection pattern is
versal bounds on the mode entanglement stemming from labelled h = (nMS +1 , · · · , nM ) where ni gives the pho-
the dimensionality of the bipartite Fock states involved, ton number of output mode i, and MS = MA + MB is
and from the linearity of the optical transformation. Fi- the number of modes in the “system”, i.e. modes that
nally, we conjecture a third bound due to unitarity which do not belong to Harold and are therefore unmeasured.
PM
extends the previously mentioned constant bound in the If nH = k=MS+1 nk = ||h||1 total photons have been
case of Alice having a single mode to multi-mode Alice, detected, the number of photons left in the system is
and we provide numerical evidence for this conjecture. nS = n − nH = nA + nB . The Hilbert space of subsys-
The maximum mode entanglement is summarized in Ta- tem X (a subset of modes), given that it contains exactly
ble I. nX photons, is denoted HX nX
.
Let the input to the interferometer be a Fock state
II. BACKGROUND |ψin i = |n1 , n2 , · · · , nMI , 0, · · · , 0i (1)
| {z }
MV
MI
n1 Y (â†k )nk
= √ |vaci , (2)
... nk !
k=1

MA which transforms according to

...

MI MI
Y 1  (n) † (n) † nk (n)
...

U (n) |ψin i = √ U âk U U |vaci (3)

k=1
nk !
MB nk
...

U

MI M
Y 1 X †
= √ âj Ujk  |vaci , (4)
n !
...

k=1 k j=1
MV
MH where Ujk are the matrix elements of the mode trans-
...

formation U , U (n) is the representation of U , and we

have used the fact that the vacuum is invariant under all
such transformations. We will usually be interested in
FIG. 1. The generic setup used throughout this paper; see the case of single photon Fock inputs, where nk = 1 or
text for an explanation of the notation. vacuum for all input modes k, a situation we will refer to
as unbunched. If all the photons are found in one mode
Figure 1 introduces the generic linear optical setup and and the rest contain vacuum, we will refer to the state as
notation used throughout the paper. The interferometer completely bunched.
has M input modes and M output modes. The mode When MH > 0 the ideal number resolving detectors
transformation describing this (photon number preserv- will register a detection pattern h = (nMS +1 , · · · , nM ) of
ing) interferometer is an M × M unitary matrix U ∈ nH photons. The output will be the post-measurement
 3

state consisting of nS = n − nH photons remaining in the relevant coefficients in particle notation by γ, which are
system modes 1, · · · , MS , given by related to the above mentioned permanent as
Cea,b (h, U )
hh|U (n) |ψin i γ1···1···M ···M (h, U ) = p . (11)
|ψS (h, U )i = . (5) n1 !...nMS !
k hh|U (n) |ψin i k
These are the coefficients of the output states as ex-
Note that this is a pure state on the system S = AB, pressed in terms of the creation operators assuming un-
because |hi only has support on subsystem H. We bunched input to the interferometer, see Eq.(12).
will denote the unnormalized output by |ψeS (h, U )i = Equation (9) provides a Schmidt decomposition we can
hh|U (n) |ψin i. The Hilbert space of the system is use to compute the entanglement. However, the fact
nS
that the total number of photons in the system, nS , is
HSnS =
M
nA
HA nB
⊗ HB , (6) preserved implies that not all conceivable bipartite basis
states |aiA ⊗ |biB are available, so the system should not
nA =0
simply be viewed as the tensor product of two qudits i.e.
where nB = nS − nA is the number of photons in Bob’s Eq.(8) is not simply the product of dimHA and dimHB .
subsystem. We are interested in entanglement with re- In particular, this means that states that are maximally
spect to this tensor product structure. The dimension of entangled in the usual sense do not exist. For example,
the Hilbert space of n photons in M modes is M +n−1 Alice can have many states with nS photons, but there


n ,
and so is only one possible Bob state to which they can be cor-
related, namely the vacuum (see Section V B 1).
nS 
The entanglement measure that will be used is the von
 
X MA + n A − 1 MB + n B − 1
dim HSnS = (7) Neumann entropy; given a pure state |ψS (h, U )i, its den-
nA =0
nA nB
  sity matrix is defined ρAB (h, U ) = |ψS (h, U )i hψS (h, U )|,
MS + n S − 1 and its reduced density matrices on subsystems
= (8)
nS are the marginals ρA (h, U ) = TrB [ρAB (h, U )] and
ρB (h, U ) = TrA [ρAB (h, U )]. The von Neumann entropy
as MS = MA + MB and nS = nA + nB . The totality of then S(ρA (h, U )) = − Tr[ρA (h, U ) · log ρA (h, U )] =
is P
states L
available to Alice can be thought of as the Hilbert − a λa · log λa where {λa }a are the non-zero eigenval-
nS nA
space nA =0 HA , and we may index its Fock basis as ues of the reduced state. Unless stated otherwise, loga-
{|aiA : a = (n1 , n2 , · · · , nMA ), ||a||1 = nA }. Similarly rithms will be assumed to be base 2. Finally, we will use
for Bob. Expanding the output in this basis, we have ebits as the unit of bipartite entanglement where 1 ebit
X corresponds to the von Neumann entropy of a Bell state.
|ψeS (h, U )i = Cea,b (h, U ) |ai ⊗ |bi .
A B (9)
a,b
III. QUBIT ENTANGLEMENT
The coefficients C e are related to permanents of the ma-
trix U [15, 16]. More specifically, consider an input In this section we will be considering the dual-rail en-
Fock state |ψi = |n1 · · · nM i and an output Fock state coding of two qubits. This means that MA = MB = 2
|φi = |n01 · · · n0M i both with a total number of n photons. and states are postselected so that subsystems A and B
Construct a new matrix Uψφ from U in two steps. First, have exactly one photon each, nA = nB = 1; all the
define the matrix Uψ consisting of nj copies of the j-th other states are discarded. (In general, the k-th qubit
column of U for all j ∈ {1, · · · , M }. Next, construct the consists of the modes 2k − 1 and 2k via the mapping
matrix Uψφ by using n0j copies of the j-th row of Uψ for |10i2k−1,2k → |0ik , |01i2k−1,2k → |1ik .) Despite the full
all j ∈ {1, · · · , M }. Then Hilbert space of the system being of dimension 10 (see
Eq. (13) ), these constraints limit the space of permissi-
perm(Uψφ ) ble states to dim HA = dim HB = 2, encoding two qubits.
hφ| U (n) |ψi = p . (10)
n1 ! · · · nM !n01 ! · · · n0M ! To entangle photons in this encoding using only passive
linear optics, the use of ancillas and postselection is nec-
In our notation, |ψi = |ψin i and |φi = |abhi, we therefore essary [4], so MH > 0.
have C ea,b (h, U ) = habh| U (n) |ψin i. The probability of
detecting pattern h is P (h, U ) = a,b |C
P ea,b (h, U )|2 , and
p A. Generating Bell states with three photons is
defining Ca,b = C ea,b / P (h, U ), the normalized state can impossible
P
be written as |ψS (h, U )i = a,b Ca,b (h, U ) |aiA |biB .
For future convenience we define coefficients of the It is known that generating a Bell state in dual-rail en-
output states in particle notation, where the Fock state coding with just two photons is impossible [4, 17]. Here
|n1 · · · nM i is written as |1| ·{z
· · 1} · · · M
| ·{z
· · M}i. We denote we prove that not only is it impossible with three pho-
n1 nM tons, it is only possible to create product states.
 4

where p = P ((1), U ), the probability of one photon being

MA = 2 detected in the last mode. We will now try to find a
MI = 3 unitary U that satisfies these constraints. Define Kk :=
U MB = 2
Uk2 U53 + Uk3 U52 , ∀k ∈ {1, . . . , 4}.
First, let us consider the case where at least one of U51 ,
MV = 2 U52 and U53 is 0. Without loss of generality (wlog) we
MH = 1
can label modes so that U51 = 0, because we can swap A
for B and mode 1 for 2 without affecting entanglement.
FIG. 2. The setup used in Section III A, with MI = n = 3, Then the equations in (12) can be rewritten as γkk =
MV = 2, MA = MB = 2, and MH = 1. We show that no
Uk1 Kk and γkj = Uk1 Kj + Uj1 Kk for k 6= j. Since γ11 =
such setup can create an entangled state in dual-rail qubit
encoding with any non-zero probability. On the other hand, U11 K1 = 0 and γ13 = U11 K3 + U31 K1 6= 0, then one and
with 4 input photons it is possible to create a Bell state with only one of U11 or K1 can be equal to 0. First, assume
probability of 1/4 [13]. that U11 = 0. Since K1 6= 0, from the constraints γ12 =
U21 K1 = γ14 = U41 K1 = 0 and γ24 = U21 K4 + U41 K2 6=
0, we see that there is no solution. Similarly, if K1 = 0,
Proposition III.1. In a passive linear optical setup us- then U11 6= 0 and the constraints γ12 = U11 K2 = γ14 =
ing dual-rail encoding, ancillas and postselection, it is not U11 K4 = 0 and γ24 = U21 K4 + U41 K2 6= 0 again results
possible to create an entangled state using 3 photon input. in no solution. Therefore there is no solution for which
Proof. First, let us consider the case where there are at least one of U51 , U52 , U53 is zero.
five modes (M = 5); four system modes (MA + MB = 4) Next we assume Kk 6= 0 ∀k ∈ {1, . . . , 4}, with
and one ancilla (MH = 1), as illustrated in Figure 2. Let U51 U52 U53 6= 0. Then solving for Uk1 from γkk = 0 we
the input be three unbunched photons (n = MI = 3). get Uk1 = −Uk2 Uk3 U51 /Kk , ∀k ∈ {1, . . . , 4}. Substitut-
Dual-rail encoding has a total of two photons in a valid ing this into the expression for γkj we get
qubit state output (nS = 2), implying here that one pho-
ton is detected (nH = 1). As there is only one measure- U51 U52 U53 (Uk2 Uj3 − Uj2 Uk3 )2
γkj = , (17)
ment ancilla, the only possible measurement pattern is Kk Kj
h = (1) (one photon in the fifth mode).
As discussed in Sec. II, the amplitudes are related to for all k, j ∈ {1, . . . , 4}, k 6= j. The only way γ12 =
the permanents of the matrix U : γ23 = 0, is if U12 U23 = U22 U13 and U22 U33 = U32 U23 . If
( P U22 U23 6= 0 then U12 U33 = U13 U32 , which means γ13 = 0
1
2 σ∈S3 Uk,σ(1) Uk,σ(2) U5,σ(3) , k = j also, thus cannot be a solution. If only one of U22 or
γkj ((1), U ) = P
k 6= j U33 is zero, assume U2j = 0 where j is 2 or 3. But
σ∈S3 Uk,σ(1) Uj,σ(2) U5,σ(3)
(12) then U1j = U3j = 0 and again γ13 = 0. If both are
defined ∀k, j ∈ {1, 2, 3, 4}. The unnormalized state fol- zero, then γ24 = 0. Therefore, there is no solution with
lowing detection is Kk 6= 0 ∀k ∈ {1, . . . , 4}.
√ √ Lastly, assume that at least one of the Kk = 0 and that
|ψ((1),
e U )i = 2γ11 |2000i + 2γ22 |0200i U51 U52 U53 6= 0; wlog, K1 = 0. Then U12 = −U13 U52 /U53
√ √ combined with the constraint γ11 = U12 U13 U51 = 0
+ 2γ33 |0020i + 2γ44 |0002i
means U12 = U13 = 0. This gives γ1j = U11 Kj , ∀j ∈
+ γ13 |1010i + γ24 |0101i {1, . . . , 4}. Since γ12 = γ14 = 0 and γ13 6= 0, then
+ γ12 |1100i + γ34 |0011i U11 6= 0, while K2 = K4 = 0. However, this implies
+ γ14 |1001i + γ23 |0110i , (13) U22 = U23 = 0 by a similar argument, further implying
P4 that γ24 = 0 and hence there is no solution.
occurring with probability P ((1), U ) = 2 k=1 |γkk |2 + We see that under no conditions is there a solution to
P4 2
k,j=1 |γkj | . the given equations where α 6= 0 and β 6= 0.
k6=j This proves the claim for 5 modes. To see that it is true
In dual-rail encoding it is possible to do any local uni- for any number of vacuum ancillas, notice that as long
tary deterministically by adding beamsplitters and phase as there are no photons added, Eqs. (12) do not change
shifters to each of the qubits [4]. Thus it suffices to show other than the mode number 5 being replaced with the
that it is not possible to create any state of the form new detection ancilla. Each new case therefore gives rise
α |0iA |0iB + β |1iA |1iB where |α|2 + |β|2 = 1 and α 6= 0, to the same constraints implied by Eqs. (16), with a lack
β 6= 0, because any entangled pure state can be trans- of solution in the same way. Thus, vacuum ancillas can
formed into one of this form by local unitary operations. only increase the probability of creating a state if that
The coefficients must therefore satisfy probability was nonzero in the first place.
γ11 = γ22 = γ33 = γ44 = 0, (14) Finally, if we allow inputs other than completely un-
bunched, Eqs. (12) become even more restrictive. For
γ12 = γ14 = γ23 = γ34 = 0 and (15)
√ √ example, if there were two photons in input mode 1 and
|γ13 | = α p, |γ24 | = β p, (16) one photon in input mode 2, then the matrix elements
 5

Ui3 , U3i would not appear in Eqs. (12), serving only to found that raising the overlap to the exponent 10 op-
make the constraints harder to satisfy. timized the numerical efficacy, penalizing states far from
 a Bell state heavily. Multiplying by the probability of
detection gives
P the targetPcost function to be minimized,
6
Corollary III.2. In a passive linear optical setup using f (h, U ) = − h P (h, U ) k=1 | hBk |ψ(h, U )i |10 .
dual-rail encoding, ancillas and postselection, it is not
possible to create a Bell state using 3 photon input∗ .

B. Optimal Bell state generation

MA = 2
MI
MB = 2
U
MV MH
FIG. 4. Results of optimization looking for interferometers
that generate Bell states with highest probability. The mini-
mum found of ≈ −0.1875 is exactly bounded by the values of
cost function for the known UBell interferometer as described
FIG. 3. The setup used in Section III B with four photons
in the text. Out of 50, 000 test runs, 1.2 % of minima found
in eight modes; MI = n = 4, MV = 4, MA = MB = 2,
were within 0.0001 of the minimum corresponding to UBell .
MH = 4. We give extensive numerical evidence for optimal
Besides the trials depicted in this graph, the cost function
Bell state generation using this setup when looking for specific
was also optimized with other parameters given to the op-
Bell states as output.
timizing algorithm as well as over the space of orthogonal
matrices. Thus the number of test runs for which a better
The previous section showed that Bell state generation solution could not be found is close to 100, 000.
with non-zero success probability requires at least four
photons. Two schemes which accomplish this task using Figure 4 shows the results of this minimization. The
four photons use six [10] and eight [12, 13] modes, with optimal known scheme, when evaluated for this cost
success probabilities of 2/27 and 1/4 respectively. function, gives a value of approximately −0.1875. It
We performed a numerical search for a linear optical produces one of these 6 Bell states with probability 1/32
Bell state generator that gives a higher success probabil- for 6 out of the 10 possible measurement patterns [13].
ity. We used a gradient descent based optimization al- We can see from the figure that the minimum achieved
gorithm over M = 8 unitaries with n = 4 photon input. by the numerical optimization over 50, 000 trials is also
Numerical optimization was carried out in Python, using approximately −0.1875, thus giving solutions which
the BFGS algorithm from the SciPy library [18]. This are equivalent to the known scheme in terms of this
algorithm finds local minima so it needs to be run many parameter. While not a proof, this numerical evidence
times with different seed unitaries, which were randomly strongly suggests that the known scheme is optimal
selected according to the Haar measure. for generating the above set of Bell states. Other
The cost function we consider is based on the overlap cost functions were also attempted, as well as other
with the desired Bell states. We allow for six different optimization libraries, but all gave the same results as
Bell states, which in the Fock basis after √ measurement the technique above.
correspond to |B1,2
√ i = (|1010i ± |0101i)/ 2, |B3,4 i√=
(|1001i ± |0110i)/ 2 and |B5,6 i = (|1100i ± |0011i)/ 2, We also investigated the case of non-orthogonal Bell
where the latter can be corrected to the usual dual-rail states; for example, allowing |00i + |11i as well as
qubit encoding using a switch [13]. After detecting mea- |00i + i |11i as target states. The possibility of both
surement pattern h, the overlap between each of these of these states being generated from the same U for
states with the post-selected state is calculated. We different measurement patterns was explored by running
similar numerical optimizations rewarding such situa-
tions. We found no such unitary, which is an interesting
result in itself.
∗ Kieling observed this using an algebraic geometry approach to
the problem [17]; here we offer an explicit proof applicable to any Though the complexity of the problem grows quickly,
entangled state. we also looked at how the situation changes with higher
 6

numbers of input photons and modes. We numerically maximum we can achieve in dual-rail qubit encoding.
optimized over n = 5, M = 10 using a similar algorithm Adding input photons for the same M increases the
and no improved solutions were found over 5000 runs. average entanglement, while adding vacuum ancillas
Similarly, we checked n = 6, M = 12 over 1000 runs and decreases it. We see that the average entanglement of
here as well there was no improvement over the −0.1875 n + 1 photons in M + 1 modes can be lower than that
result for n = 4, M = 8. for M and n (see n = M = 5 and 6). That is, we
do not expect more average entanglement by adding a
photon at the cost of adding another mode. Further, we
IV. RANDOM UNITARIES note that even with 2 photons, there is more average
entanglement generated than in the optimal Bell state
In this section, we move from the dual-rail qubit en- generator with 4 photons. We explore this in more detail
coding of Section III to mode entanglement in Section V. for a better comparison.
First, we look at how much mode entanglement can be
generated with random elements of the unitary group,
which we can then use to compare with the dual-rail en-
coding. We do so by setting Alice and Bob’s number
of modes to 2, and numerically computing the average
amount of entanglement over measurement patterns. No-
tice that this is different from the setting in Section III,
where we aimed to get a maximally entangled Bell state
with the highest possible probability. Here and in the
rest of this work we will study this average entanglement,
namely
X
hS(U )iH = P (h, U )S(ρA (h, U )). (18)
h

The expectation overRthe unitary group (for fixed M and

n) is then hSiH,U = U(M ) dU hS(U )iH , where dU is the FIG. 6. Numerical evaluation of hS(U )iH for 100, 000 uni-
normalized Haar measure. taries U chosen using the Haar measure in the case MA =
MB = 2, MH = 4, and n = 4 unbunched input photons.
Average entanglement for a given U was calculated according
to Equation (18) and then binned in one of 100 bins with a
minimum of 0 and maximum obtained in the samples. The
red dot marks the value of average entanglement that the
Bell generating unitary from Section III B can give, denoted
as UBell , if all of its output states were used.

In the usual Bell state generation scenario discussed in

Section III B, if the measurement outcome indicates that
the output state is outside of the qubit subspace, the
output is discarded. Here we include the entanglement
of the discarded states in accordance with Eq. (18). We
compare the optimal Bell state generator to random uni-
taries with the same parameters; MA = MB = 2, n = 4
and M = 8.
FIG. 5. The expectation, over the unitary group, of the av-
In Fig. 6 we see the results of the comparison. Firstly,
erage, over measurement patterns, mode entanglement versus in Section III B we saw that the probability of getting
the number of modes M , for various numbers of unbunched a Bell state for a state correctable with a single switch
input photons. MA = MB = 2, and if the number of photons is 3/16 [13]. A Bell state gives a single ebit, and if all
n is smaller than M , vacuum input modes are added. The the other states are discarded, the average entanglement
number of heralding detectors is MH = M − MA − MB . The would be 0.1875 ebits. If all the outputs from this unitary
entanglement for a single unitary U is averaged over all mea- were counted towards average entanglement as discussed
surement patterns, and subsequently averaged over 100,000 in the previous paragraph (where Equation 18 is utilized),
randomly Haar-sampled unitaries U . Colours represent dif- the entanglement obtained is marked on the Figure 6
ferent number of input photons, with 2 ≤ n ≤ 7. as UBell . As we can see from the graphs, UBell gives a
markedly lower amount of entanglement than what could
Figure 5 shows the numerical results. We notice that be generated on average with a random unitary on the
often the average is higher than 1 ebit, which is the same number of modes.
 7

V. MODE ENTANGLEMENT A. Entanglement when Alice has a single mode

1. Entanglement for bunched input can be unbounded

The previous section shows that, on average, random
unitaries give significantly more mode entanglement than
dual-rail encoding. We therefore turn our attention to First, we show that mode entanglement is unbounded
the investigation of mode rather than qubit entanglement if we are not restricted to unbunched input.
as defined in Section II. n
Equation (6) states that the total system Hilbert ...
MI = 1 MA = 1
space is a direct sum of Hilbert subspaces such that
the sum of Alice and Bob’s photon numbers is nS , the MV = 1 U MB = 1

number of photons left after heralding. Let ρAB =

|ψS (h, U )i hψS (h, U )| as in Eq. (5). Alice’s reduced den- FIG. 7. The setup used in Section V A 1, where we consider
sity matrix is only M = 2 modes. The input consists of all n photons
bunched in the top mode; MI = MV = MA = MB = 1,
MH = 0. We prove that in this setup maximal entanglement
ρA (h, U ) = TrB [ρAB (h, U )] (19) grows as log n.
 
X X
= hb00 |  Cab C a0 b0 |abi ha0 b0 | |b00 i Proposition V.1. Let the input into a M = 2 inter-
b00 a,b,a0 ,b0 ferometer consist of n photons bunched in a single mode
  (see Figure 7). Then the entanglement across the two
X X output modes is at most O(log n) ebits, which is achieved
=  Cab C a0 b  |ai ha0 | , (20) when U is a balanced beamsplitter.
a,a0 b
Proof. Parameterize the M = 2 unitary matrix U act-
ing on Alice and Bob’s single mode Hilbert spaces as
where only the terms with kak1 = ka0 k1 are non-zero, " #
because kbk1 = kb0 k1 = kb00 k1 and nS = kak1 + kbk1 = U=
c d
, (22)
ka0 k1 + kb0 k1 . Therefore, there exists a Fock basis or- −d∗ c∗
dering in which Alice’s reduced state is block diagonal,
which allows us to derive a bound on the entanglement where |c|2 + |d|2 = 1. The output state is
(see Section V B 1). In the case that Alice has a single  n √
mode, this implies her state is diagonal in Fock basis. |n0i = â†1 / n! |0i
The total number of orthogonal states available to Alice  n √
is 7→ câ†1 − d∗ â†2 / n! |0i
n  
1 X n
nS     =√ (câ†1 )k (−d∗ â†2 )n−k |0i
nS
X MA + n A − 1 MA + n S n! k=0 k
dim(HA )= = .
nA nS n   √ p
n =0 A 1 X n k
(21) = √ c (−d∗ )n−k k! (n − k)! |ki |n − ki
n! k=0 k
In Section V A, we find entanglement bounds when Al- n
s 
ice only has one mode. The bound depends on the input X n k
= c (−d∗ )n−k |ki |n − ki . (23)
state; if the input photons are bunched in a single mode, k
k=0
entanglement is unbounded as the number of photons in-
creases. Surprisingly, if the input is unbunched, we find When Alice has only one mode, her reduced density ma-
a constant bound independent of the number of Bob’s trix is diagonal in the Fock basis, so we can find the
modes and independent of the number of photons. More spectrum of her state directly from the above equations:
general bounds can be found, though they are also more    
n n
loose. In Section V B 1 we give the bound on entangle- λk = (|c|2 )k (|d|2 )n−k = (|c|2 )k (1 − |c|2 )n−k .
ment due to the block diagonal structure of Alice’s re- k k
duced density matrix in Fock basis. In Section V B 2 we (24)
give a bound which is a consequence of the linearity of This is a binomial distribution with a ‘success’ proba-
the mode transformations. Unlike in Sec. V A, neither of bility of p = |c|2 . The entropy of the binomial distribu-
these bounds depend on the unitarity of the mode trans- tion for a fixed p is 1/2 log2 (2πen · p · (1 − p))+O(1/n)† .
formations, which we expect should affect the amount of
entanglement that can be achieved. In Section V C we
conjecture a general unitarity bound based on numerical
† From, e.g., the de Moivre-Laplace Theorem
evidence.
 8

Thus we see that the entanglement bound is O(log n), In the following calculations we shall assume that n →
where n is the number of photons. The constant prefac- ∞ as any bound on the entropy found for this infinite
tor is maximized for p = |c|2 = |d|2 = 1/2, whence the en- case would also hold for a finite one with the same set of
tropy of Alice’s state is 1/2 log2 (2πen · 1/2 · (1 − 1/2)) + constraints.
O(1/n) = 1/2 log2 (πen/2)+O(1/n). Finally, notice that
solutions to Equation (22) where |c|2 = |d|2 = 1/2 are a Lemma V.3. Let {pj }∞ P be a probability distribution
j=0
family of balanced beamsplitters.  subject to the constraint j jpj ≤ N . Then the entropy


This is in stark contrast to the situation where the of this distribution is at most log (1 + N )1+N /N N .
input is unbunched, where we will see in the next section
Proof. The entropy P∞ of the probability distribution
that the entanglement is bounded by a constant.
{pj }∞
j=0 is S = − log pj . We maximize this
j=0 pj P
∞
subject to the constraints j=0 jpj = n ≤ N and
P∞
2. Entanglement for unbunched input is bounded j=0 p j = 1 using the method of Lagrange multipliers.
Let the Lagrangian be
We now consider situations where Alice only has one    
∞ ∞
mode, Bob can have many, and we do not use any mea- X X
surement. The following Lemma will be of use. L = S + (λ0 + log e)  pj − 1 + λ1  jpj − n .
j=0 j=0
(27)
MA = 1 Then ∂L/∂pj = − log pj + λ0 + λ1 j. Setting ∂L/∂pj = 0
... ...

MI gives pj = 2λ0 +λ1 j . Substituting the value of pj into the

constraints, we get
U
...

MB
∞
MV X
jpj = 2λ0 2λ1 /(1 − 2λ1 )2 = n (28)
j=0
∞
FIG. 8. The setup used in Section V A 2. The input is an
X
pj = 2λ0 /(1 − 2λ1 ) = 1 (29)
unbunched state with MI = n, with MV ≥ 0, MA = 1, MB ≥
j=0
1 and MH = 0. We prove that entanglement for this setup is
bounded by a constant. This allows us to solve for λ0 and λ1 , giving

λ0 = log [1/(1 + n)], λ1 = log [n/(1 + n)]. (30)

Lemma V.2. Consider inputting a Fock state |ni =
|n1 . . . nM i into an arbitrary interferometer that has M Notice that
modes. Let N = max {n1 , . . . , nM }. Then the mean pho- X X
ton number in each output port is bounded by N [19]. S=− pj log pj = − pj (λ0 + λ1 j)
j j

= −λ0 − λ1 n
Proof. Let |ni be an arbitrary Fock state.  
1+n
= log (1 + n) /nn (31)
†
hn̂j i = hn| U (n) n̂j U (n) |ni
† † The function above increases monotonically for n ≥ 0
= hn| U (n) â†j U (n) U (n) âj U (n) |ni and since n ≤ N we get
  
S ≤ log (1 + N )1+N /N N .


(32)
X † X
= hn|  âj 0 Ujj 0   âj 00 Ujj 00  |ni
j0 j 00

Ujj 0 Ujj 00 hn| â†j 0 âj 00 |ni
XX
=
j0 j 00
Corollary V.4. Let {pj }∞ j=0 bePsome probability distri-
X bution subject to the constraint j jpj ≤ N , N ∈ [0, 1].
= Ujj 0 Ujj 0 nj 0 (25) Then the entropy of this distribution is at most 2 ebits.
j0
Theorem V.5. Let Alice have one output mode, MA =
If, as hypothesized, nj ≤ N for all modes j, then 1, and Bob have MB = k. Let the input be a single
photon in each of the k+1 modes. Then the entanglement
X X between Alice and Bob is bounded by 2 ebits for all k.
hn̂j i = |Uj 0 j |2 nj 0 ≤ |Uj 0 j |2 N = N. (26)
j0 j0 Proof. Alice’s reduced density matrix is diagonal in
the Fock basis, where each entry hj| ρA |ji corresponds
 to the probability that Alice’s mode contains j photons.
 9

By Lemma V.2, this distribution satisfies the conditions

of Corollary V.4. Thus the von Neumann entropy of this 1 MA = 1
W
state is bounded by 2 for any k, as the bound which holds 2 MB = 1
for k → ∞ also holds for any finite k as well.  3
Notice that extra vacuum modes will not increase this
limit on the entanglement as the limit is due to the ex- 4
pected number of photons in Alice’s mode being at most MH

...
1. We see that despite the fact that the dimension of M −1
Alice’s Hilbert space grows with the number of photons
as n + 1, and Bob’s can be even larger, the maximum M
entanglement is severely constrained to be less than 2
ebits. FIG. 9. The setup used in Section V A 3, where M = n,
Because we are interested in the average entanglement, MA = MB = 1, and MH ≥ 1. An arbitrary M mode in-
the result will hold for heralding as well: terferometer can be decomposed into M (M + 1)/2 two-mode
interferometers [21, 22]. Note that this also applies to an ar-
Corollary V.6. Let Alice have one output mode, MA = bitrary M − 1 mode sub-interferometer (blue). By focusing
1, while Bob and Harold have MB + MH = k. Let the on the only component that entangles Alice and Bob (red),
input be a single photon in each of the k + 1 modes. Then we show that the maximum entanglement is the M = n = 2
value of log 3 ebits.
the entanglement between Alice and Bob is bounded by 2
ebits for all k.
Pn
Proof. No LOCC operation can increase the amount of j=0 pj log (j + 2) is bounded by log 3 which can be
entanglement Pin the system on average [20]. Therefore, achieved by p1 = 1 and pk = 0 for k ∈ {0, 2, 3, . . . , n}.
hS(U )iH = h P (h, U )S(ρA (h, U )) ≤ S(ρA (U )), where
ρA (U ) is Alice’s reduced density matrix before any mea-
Proof. Since f (x) = log (x + 2) is a concave
P function,
surement, and by Theorem V.5, S(ρA (U )) ≤ 2 ebits.  Pn n

We can also examine inputs that have different num- by Jensen’s inequality j=0 pj f (j) ≤ f j=0 p j j =
bers of bunched photons. If the highest number of pho- f (1) = log 3, which is achieved by substituting p1 = 1
tons in a single input mode is N , as per Lemma V.2, and pk = 0 for k ∈ {0, 2, 3, . . . , n}. 
the expected number of photons in Alice’s mode will
then be bounded by N . Because the function which
Theorem V.8. Consider an interferometer with M ≥ 3
bounds the entropy, Eq. (32), is monotonically increas-
modes, where both Alice and Bob have one mode and the
ing, the entropy of Alice’s (diagonal) state (p0 , . . . , pn ) is

other output modes are measured using photon counting
at most log (1 + N )1+N /N N by Lemma V.3. In the ex-
detectors. Let the input be the n = M unbunched Fock
treme case where all the photons are bunched in a single
state. Then the maximal average entanglement that can
mode, S scales as O(log(N + 1)), consistent with Propo-
be created between Alice and Bob is log 3 ebits.
sition V.1.

Proof. First, notice that the average entanglement

3. Entanglement when Bob also has a single mode achievable by an M = 2 interferometer can be achieved
for M ≥ 2 by having modes 3 to M transform trivially,
since photons in these modes will be detected with unit
In this section we consider a similar setup to the pre-
probability. Thus max hS(UM )iH ≥ max hS(UM =2 )iH =
vious section, except now we fix the number of Bob’s
log 3 ebits, ∀M ≥ 3. The interferometer given in Section
modes to 1 and assign the rest to Harold. Recall that
V D, Eq. (43) below achieves this.
we are interested in generating the highest amount of en-
tanglement between Alice and Bob on average, thus the Any U ∈ U(M ) can be decomposed as in Fig. 9. Then
probability of detection patterns must be taken into ac- the bottom left triangle (colored blue in the figure) is a
count. More precisely, we are looking for the maximum unitary V ∈ U(M − 1). Since the input is unbunched,
Lemma V.2 implies that each output from V has a mean
P
of hS(U )iH = h P (h, U )S(ρA (h, U )). Some patterns
might yield a state with high entanglement, but be very photon number of 1. In particular, Bob’s mode P before
unlikely to occur. In a practical setting we might prefer beamsplitter W (red in the figure) will satisfy k kqk = 1
states that are less entangled but we can generate more where k is the number of photons occuring with proba-
consistently. bility qk . Since the remaining beamsplitters (white in the
We first prove a technical lemma that will be useful figure) act only on Bob and Harold’s systems, they have
later. no effect on Alice’s reduced state and can therefore be
ignored.
Lemma V.7. Given Pa probability distribution Let the probability of detecting pattern h be ph , and
n
(p0 , . . . , pn ) such that j=0 jpj = 1, the sum the probability of detecting a total of nH = khk1 photons
 10
P
be pnH = h:khk1 =nH ph . The average entanglement is 1. Dimensionality

X By looking solely at the dimensions of Alice and Bob’s

hS(U )iH = ph S(ρA (h)) Hilbert spaces, we can derive an entanglement bound as
h
follows.
n
X X
= pnH ph /pnH S(ρA (h)) Proposition V.9. Let Alice’s and Bob’s joint postse-
nH =0 h:khk1 =nH
lected state have a total of nS photons. Let Alice and
n Bob have MA = MB modes. The Schmidt rank, ω, is at
most
X X
≤ pnH ph /pnH log (n − nH + 1)
MA + nS2−1
 
nH =0 h:khk1 =nH
ω=2 nS −1 nS odd, (34)
n−1
X 2
= pnH log (n − nH + 1), (33) nS + MA MA + n2S − 1
 
nH =0 ω=2 nS nS even. (35)
nS 2 −1

where we’ve used the fact that the entanglement of

S(ρA (h)) is upper bounded by the Schmidt rank log(n − Proof. Let
P Alice’s and Bob’s joint state be
nH + 1). |ψS (h, U )i = k,j Ckj (h, U ) |kiA ⊗ |jiB , where we in-
clude the possibility of no measurement (MH = 0). The
As the photon number found in modes 1 and 2 is
Schmidt decomposition is achieved by a state dependent
set before the beamsplitter W , if nH photons have
change of basis such that
been detected, then there were already nH photons in
modes 3 through M . Alice contributes one photon min(dim HA ,dim HB )
X
through her mode to their joint system, which implies |ψS (h, U )i = λq |qiA ⊗ |qiB , (36)
that Bob must contribute n − nH − 1 photons through q=1
mode 2, occurring with probability qn−nH −1 . Therefore
pnH = qn−nH −1 and recall that Bob’s probability distri- where {|qiA,B } are orthonormal bases for A and B, re-
Pn−1 spectively.
bution is constrained by k=0 kqk = 1. By Lemma V.7
Pn−1 Pn−1 Writing this state in terms of Alice and
nH =0 qn−nH −1 log (n − nH + 1) = j=0 qj log (j + 2) Bob’s photon numbers we have |ψS (h, U )i =
is maximized for j = n−nH −1 = 1, that is q1 = 1, yield- PnS nA ,nB
nA =0 |ψS (h, U )i with nB = nS − nA . The
ing hS(U )iH ≤ log 3. This also implies that nH = n − 2 n0 ,n0
photons are detected in the optimal situation.  overlap hψSnA ,nB (h, U )|ψS A B (h, U )i = 0 for nA 6= n0A ,
nB 6= n0B as these states belong to different Hilbert
Note that this agrees with the bound in Theo- subspaces in the direct sum. The reduced density matrix
rem V.5 found in the previous section, which follows is block diagonal – each block corresponds to a different
from
P the entanglement measure property hS(U )iH = (nA , nB ) combination. We may therefore consider each
h P (h, U )S(ρA (h, U )) ≤ S(ρA (U )), where ρA (U ) is Al- subspace individually, where the maximal Schmidt
ice’s reduced density matrix before any measurement. rank is min(dim HA nA
, dim HB nB
). The total number of
Here the maximum entanglement is log 3 < 2 ebits. Schmidt coefficients is therefore at most
Moreover, adding more vacuum input modes will not af- nS
fect this bound, as this would only change Bob’s expected
X
nA nB
ω= min{dim HA , dim HB } (37)
number of photons before the beamsplitter W to be at nA =0
most 1 instead of exactly 1 as per Lemma V.2. nS    
X MA + n A − 1 MB + n B − 1
= min , .
nA =0
nA nB
(38)
B. Entanglement when Alice has many modes
For MA = MB this gives the result. 
Since the entanglement is given by the number of
In this section we give two bounds on entanglement nonzero Schmidt coefficients, this gives a bound on the
for more general situations when Alice has more than entanglement S ≤ log(ω).
one mode, based on the Schmidt rank of Alice’s reduced
state. They are independent of the input state or any in-
terferometer transformation, depending only on the given 2. Linearity bound
number of photons and modes; we assume the latter is
the same for both Alice and Bob. This generality comes Here we consider a bound due to the linearity of the
at a price, however, in that the bounds loosen; we will interferometer transformations. In the following we do
discuss a conjectured tighter bound in the following sec- not assume anything about the form of the input Fock
tion. state, nor whether measurement occurs or not.
 11

Proposition V.10. Given an n photon Fock state as C. Hints of another bound

input to a M -mode linear optical device, with Alice and
Bob having MA and MB output modes respectively, the In this section we explore a potential bound that is
maximal entanglement achievable between Alice and the motivated by numerical evidence (see Figure 10). While
rest of the modes for any state is bounded by n ebits. adding more photons to the interferometer increases the
size of Alice’s and Bob’s Hilbert spaces, and according
Proof. Starting with the arbitrary linear optical mode to the results from the previous section should allow for
transformation in Eq.(4), we can group the sum into Al- higher amounts of entanglement, we see that this is not
ice’s modes and the ‘rest’: what happens in general (assuming the number of modes
that Alice and Bob have are fixed). Based on the analyt-
M MA M ical results from Section V A and the numerical evidence
â†k 7→ â†j Ujk = â†j Ujk + â†j Ujk
X X X
for all cases up to n = 7 photons and MA = MB = 3
j=1 j=1 j=MA +1 modes, we make a conjecture that there is another bound
=: Âk (U ) + R̂k (U ). (39) which seems to arise from the unitarity of the mode trans-
formation.
The degree one polynomials Âk (U ), R̂k (U ) in the cre- Conjecture V.1. For n unbunched photons input into
ation operators are not canonical raising operators, be- an interferometer with MA Alice and MB Bob output
cause e.g. [Âk (U ), Âk0 (U )] 6= δkk0 . This means that dif- modes with n > MA + MB , the average amount of
ferent monomials in {Âk (U )}k do not necessarily give rise entanglement, obtained over Harold’s measurements, is
to orthogonal states; however, this can only reduce the bounded above by the maximal average amount of en-
Schmidt rank of the resulting state. tanglement achieved when n = MA + MB .
An arbitrary input Fock state is of the form We provide numerical evidence supporting this “uni-
QM † nk √
(â
k=1 k ) / n k ! |vaci, so that the output state is of tarity bound” for various numbers of input photons and
the form modes. We assume that the input states are unbunched,
ancillas and measurement are allowed, and Alice and Bob
M
Y 1 have the same number of modes; MI = n, MV ≥ 0,
√ (Âk (U ) + R̂k (U ))nk |vaci (40) MA = MB ≥ 1 and MH ≥ 0.
nk !
k=1 Propositions V.9 and V.10 provide tight entanglement
bounds when all input photons are kept in the system,
i.e. it is a product of n terms, not all of which are neces- i.e. when there is no detection. We know that it is pos-
sarily different. We can rewrite it as sible to postselect states that exceed these bounds, but
because we are interested in average entanglement these
n
Y cases must be weighted with their heralding probabilities.
N (Âjk (U ) + R̂jk (U )) |vaci , (41) Our findings are consistent with a generic trade off be-
k=1 tween these two quantities, leading to a bounded average
entanglement.
where jk ∈ {1, . . . , M } and N is the necessary normal- Figure 10 shows the results of numerical optimization
ization. The highest number of monomial terms in this of the average entanglement given by Eq. (18) for vari-
product is bounded by 2n and after tracing out Bob and ous numbers of input photons and modes. We can see
Harold this also bounds the number of monomial terms how the linearity and dimensionality bounds of Sec. V B
that can be in Alice’s reduced state.  are indeed limiting the entanglement. We also see the
Consider a balanced 50:50 beam splitter coupling one appearance of what looks like a third bound, seemingly
of Alice’s modes (say k) to one of Bob’s modes (say when the number of photons is larger than the total num-
k + MA ). If Alice’s mode contained one input photon ber of modes in the system (MA + MB ). This new be-
and Bob’s none, we get 1 ebit of entanglement. Propo- haviour is not captured by the bounds we have obtained
sition V.10 tells us we can only get up to n ebits using and we conjecture that it is due to the unitarity of the in-
n photons, so as long as n ≤ MA = MB , a beamsplitter terferometric transformation. This leads to the hypoth-
coupling mode k with mode k + MA for k = 1 through n esis that the maximum possible average entanglement,
in this way would give us a state that achieves the bound. in situations with unbunched input and Alice and Bob
The dimensionality (Section V B 1) and linearity have the same number of modes, can be reached using a
bounds above hold for all M and all n. We can find (MA +MB )-mode interferometer with MA +MB photons.
numerically the photon number nL (MA ) ∈ N, which de-
pends on the number of Alice’s modes. For a given MA
it represents the number of photons up to which the lin- D. Optimal interferometers
earity bound is smaller than the dimensionality bound.
For n > nL (M ), the dimensionality bound is a tighter Finally, in this section we report some of the explicit
limit on the entanglement (see Figure 10). interferometers (unitaries) that produce the optimal en-
 12

photons, n = 2, with one in each mode, the unitary

" #
cos θ sin θ
BS2 = (43)
− sin θ cos θ

produces a state with log 3 ebits of entanglement. Con-

jecture V.1 says that for all higher photon numbers, log 3
will still be the maximum, achieved by using BS2 be-
tween any pair of Alice and Bob’s modes and identity
on all the others (they are just routed straight to the
detectors).
For MA = MB = 2, we have that all the optimal inter-
ferometers are actually combinations of BS1 and BS2 .
An example for M = n = 4 is:
 
FIG. 10. Plot of the maximum average entanglement found cos θ 0 0 sin θ
through numerical optimization, along with the dimension-
 0 cos θ sin θ 0 
 
ality and linearity bounds for MA = MB . The input are  , (44)
unbunched states. If n > MA + MB , the remaining MH =  0 − sin θ cos θ 0 
n − MA − MB modes contain detectors. The green dashed − sin θ 0 0 cos θ
line is the linearity bound. The dotted lines are dimension-
√
ality bounds for the value of MA whose dots have the same where as before θ = 21 arccos (1/ 3). This interferometer
colour. We can see that the values of nL for specific MA s are: corresponds to a BS2 beamsplitter between modes 1 and
nL (1) = 1, nL (2) = 2, nL (3) = 3, nL (4) = 4. The dots are
4 and another BS2 beamsplitter between modes 2 and 3,
solutions returned by the optimization.
giving log 9 ≈ 3.17 ebits of entanglement. When n = 3,
the optimal value of log 6 ≈ 2.58 ebits is achieved by
using BS2 on modes 1 and 4 and BS1 on modes 2 and
3. For n = 2, the maximum of 2 ebits is achieved by two
tanglement found for small number of modes. BS1 beamsplitters, similar to n = 4 case. Finally, for
n = 1 we just use a single BS1 to achieve 1 ebit.
In the case of MA = MB = 1 and a single photon
n = 1, the well known balanced 50:50 beamsplitter is
optimal, ACKNOWLEDGMENTS

" # SS was supported by the Bristol Quantum Engi-

1 1 1 neering Centre for Doctoral Training, EPSRC grant
BS1 = √ . (42) EP/L015730/1 and would like to thank P Birchall for
2 −1 1
bringing to our attention a more elegant proof of Lemma
V.2. AM was supported by EPSRC Early Career Fellow-
ship EP/L021005/1. PST was supported in part by the
This is familiar, as in single-rail
√ encoding it creates a U.S. Army Research Office under contract W911NF-14-
Bell state. Let θ = 12 arccos (1/ 3). When we input two 1-0133 and EPSRC First Grant EP/N014812/1.

[1] J. L. O’Brien, A. Furusawa, and J. Vučković, Nature [7] J. S. Bell, Physics 1, 195 (1964).
Photonics 3, 687 (2010). [8] A. Aspect, Physics Letters A 54, 117 (1975).
[2] M. Saffman, T. G. Walker, and K. Mølmer, Reviews of [9] S. Popescu and D. Rohrlich, Foundations of Physics 24,
Modern Physics 82, 2313 (2010). 379 (1994).
[3] Z.-l. Xiang, S. Ashhab, J. Q. You, and F. Nori, Reviews [10] J. Carolan, C. Harrold, C. Sparrow, E. Martin-Lopez,
of Modern Physics 85, 623 (2013). N. J. Russell, J. W. Silverstone, P. J. Shadbolt, N. Mat-
[4] P. Kok, W. J. Munro, K. Nemoto, T. C. Ralph, J. P. suda, M. Oguma, M. Itoh, G. D. Marshall, M. G. Thomp-
Dowling, and G. J. Milburn, Reviews of Modern Physics son, J. C. F. Matthews, T. Hashimoto, J. L. O’Brien, and
79, 135 (2007). A. Laing, Science 349, 711 (2015).
[5] R. Jozsa and N. Linden, Proceedings of the Royal Soci- [11] M. Gimeno-Segovia, P. Shadbolt, D. E. Browne, and
ety A: Mathematical, Physical and Engineering Sciences T. Rudolph, Physical Review Letters 115, 020502 (2015).
459, 2011 (2003). [12] J. Joo, P. L. Knight, J. L. O’Brien, and T. Rudolph,
[6] R. Horodecki, P. Horodecki, M. Horodecki, and Physical Review A 76, 052326 (2007).
K. Horodecki, Reviews of Modern Physics 81, 865 (2009).
 13

[13] Q. Zhang, X.-H. Bao, C.-Y. Lu, X.-Q. Zhou, T. Yang,

T. Rudolph, and J.-W. Pan, Physical Review A 77,
062316 (2008).
[14] T. Rudolph, (2016), arXiv:1610.07128.
[15] S. Aaronson and A. Arkhipov, Theory of Computing 9,
143 (2013).
[16] S. Scheel, (2004), arXiv:quant-ph/0406127.
[17] K. Kieling, Linear optics quantum computing–
construction of small networks and asymptotic scaling,
Ph.D. thesis, Imperial College London (2008).
[18] E. Jones, T. Oliphant, P. Peterson, et al., “SciPy: Open
source scientific tools for Python,” (2001–2017).
[19] J. P. Olson, K. R. Motes, P. M. Birchall, N. M. Studer,
M. LaBorde, T. Moulder, P. P. Rohde, and J. P. Dowling,
(2016), arXiv:1610.07128.
[20] M. Horodecki, Quantum Information & Computation 1,
3 (2001).
[21] A. Hurwitz, Nachrichten von der Gesellschaft der Wis-
senschaften zu Göttingen, Mathematisch-Physikalische
Klasse 1897, 71 (1897).
[22] M. Reck, A. Zeilinger, H. J. Bernstein, and P. Bertani,
Physical Review Letters 73, 58 (1994).