Arxiv:2004.12788V2 [Quant-Ph] 16 Nov 2020

Fourier Transform of the Orbital Angular Momentum of a Single Photon

Jaroslav Kysela,1, 2 Xiaoqin Gao,3, 1, 2, ∗ and Borivoje Dakić1, 2, † 1Institute for Quantum Optics and Quantum Information (IQOQI), Austrian Academy of Sciences, Boltzmanngasse 3, 1090 Vienna, Austria. 2Faculty of Physics, University of Vienna, Boltzmanngasse 5, 1090 Vienna, Austria. 3National Mobile Communications Research Laboratory, Quantum Information Research Center, Southeast University, Sipailou 2, 210096 Nanjing, China. (Dated: November 17, 2020) Optical networks implementing single-qudit quantum gates may exhibit superior properties to those for qubits as each optical element in the network can work in parallel on many optical modes simultaneously. We present an important class of such networks that implements in a deterministic and efficient way the quantum Fourier transform (QFT) in an arbitrarily high dimension. These networks redistribute the initial quantum state into the orbital angular momentum (OAM) and path degrees of freedom and offer two modes of operation. Either the OAM-only QFT can be implemented, which uses the path as an internal auxiliary degree of freedom, or the path-only QFT is implemented, which uses the OAM as the auxiliary degree of freedom. The resources for both schemes scale linearly O(d) with the dimension d of the system, beating the best known bounds for the path-encoded QFT. While the QFT of the orbital angular momentum states of single photons has been applied in a multitude of experiments, these schemes require specially designed elements with non-trivial phase profiles. In contrast, we propose a different approach that utilizes only conventional optical elements.

I. INTRODUCTION general approach, which works with in principle 100 % ef- ficiency for arbitrarily high dimension of the OAM space The field of quantum computation has gained ever- of a single photon. Our scheme decomposes the Fourier increasing attention thanks to the invention of the quan- transform into a series of elementary operations that tum factoring algorithm due to Shor [1–3], which utilizes can be directly implemented with basic commercially- as its key part a quantum Fourier transform. The quan- available optical elements such as beam-splitters, mirrors tum Fourier transform (QFT), or quantum Hadamard and Dove prisms. Such an explicit decomposition reveals gate, has been since then used in many areas of quantum how individual components participate in the evolution computation and communication [4] for systems of qubits of different OAM eigenstates and allows for modifica- as well as high-dimensional qudits. The application areas tions, such as miniaturization of the setup to a micro- of the high-dimensional QFT acting on a single photon’s chip level. The scheme’s implementation is recursive and state include, but are not limited to, generation of mu- makes use of an interplay between the OAM and path tually unbiased bases in the quantum state tomography degrees of freedom. The number of required optical ele- [5–8] and quantum key distribution [9, 10]; generation ments O(d) scales linearly in the dimension d, as opposed of angular states [11–14]; sorting of spatial modes of a to O(d log d) scaling of the setup using path-encoded qu- photon [15]; and representation of multiport devices em- dits [30–32]. This is made possible by the fact that a sin- ployed in Bell test experiments [16]. Single photon’s high- gle passive optical element can act on many OAM eigen- dimensional QFTs can be also used as building blocks of states at the same time, leading to a heavily parallelized programmable universal multi-port arrays [17–19]. operation of the network of optical elements. Moreover, The orbital angular momentum (OAM) of single pho- the setup for the OAM Fourier transform can be modi- tons is a quantized property with infinite-dimensional fied to act as the path-only Fourier transform. In such a Hilbert space, which allows for construction of qudits scheme, the OAM is present only in the inner workings of in arbitrarily high dimension [20–22]. By manipulation the transform. This OAM-enhanced setup preserves the of a single photon’s OAM the universal quantum com- linear scaling of the number of optical elements, which putation is possible [23, 24]. The Fourier transform of shows a clear advantage of our scheme over the setup the OAM eigenstates of single photons has been demon- that uses only the path degree of freedom. arXiv:2004.12788v2 [quant-ph] 16 Nov 2020 strated in a number of experiments [10, 14, 25–28] using free-space propagation and specially designed optical One of our scheme’s main components is the OAM- elements imparting non-trivial phase profiles. Alterna- Path swap operator, which interchanges the OAM and tive experimental schemes have been presented in special path degrees of freedom of a photon’s state. To the best cases [29]. of our knowledge, we demonstrate for the first time the In this paper we demonstrate a completely different implementation of such an operator in terms of conventional optical elements. The OAM-Path swap represents a multiport generalization of the OAM sorter and its implementation features efficient deployment of the OAM ∗ [email protected] parity sorter. Each instance of the parity sorter functions † [email protected] simultaneously as a series of many conventional beam- 2

a) Henceforth we denote a quantum state of a photon with ℋ = ℋ ⊗ ℋ ℋ 푂 푃 m quanta of OAM by |mi and call it an OAM eigen- 퐹(푑) O 푂퐴푀 푑 = 푑푂 × 푑푃 state. OAM eigenstates represent a basis of the subspace HO. Likewise, a basis vector |liP of the subspace HP b) OAM−Path swap is called a propagation mode or simply a path. For each ℋ푂 (푑푂) division of dimension d into a pair of smaller dimensions 퐹푂퐴푀 CZ dO and dP there is a one-to-one correspondence between ℋ푃 an index j ∈ {0, . . . , d − 1} and a pair of numbers (m, l) (푑푃) 퐹푃푎푡ℎ with 0 ≤ m < dO and 0 ≤ l < dP such that

|miO |liP ⇔ |ji = |m × dP + li . (2) FIG. 1. Circuit representation of the recursive scheme for the OAM Fourier transform. a) The OAM Fourier transform In cases when d is a prime number the decomposition into acting on the d-dimensional OAM space H, which can be dO and dP is trivial and the recursive scheme presented decomposed into a tensor product of two factor subspaces HO below is not applicable. For these dimensions a different and HP . b) A circuit equivalent to a), where a dO-dimensional approach has to be chosen and we discuss one possible OAM Fourier transform is applied first, followed by a phase alternative in the following section. gate CZ and a dP -dimensional path-only Fourier transform. The Fourier transform in dimension d is obtained in The swap operator then exchanges states between the two four steps as demonstrated in Fig. 1 b). In the first step, subspaces. For details refer to the main text. a dO-dimensional Fourier transform is applied only on the OAM subspace HO. Then a controlled phase gate acts on both subspaces, followed by a dP -dimensional Fourier splitters for different OAM eigenstates. transform applied only on the path subspace. As the last The manuscript is organized as follows. In section II we step, a swap gate is used, which effectively exchanges introduce the theoretical background for the construction the OAM and path subspaces. This procedure can be of the Fourier transform. Then we present in section III mathematically summarized by the formula the setup that implements the Fourier transform in the OAM of a single photon. We discuss the properties of F (d) = SWAP · F (dP ) · CZ · F (dO ) , (3) the OAM sorter, a key part of the setup, in section IV. OAM P ath OAM In section V, we demonstrate how to generalize the OAM where the phase gate CZ acts as a high-dimensional sorter into the OAM-Path swap operator. In section VI controlled-Z gate on the OAM degree of freedom the scaling of our scheme is presented. In section VII, we compare our scheme with the recursive scheme for the l m×l CZ(|miO |liP ) = (Z |miO) |liP = ω |miO |liP (4) path-only Fourier transform and summarize our results in the last section VIII. with ω = exp (2πi/d). The action of the swap on an input mode reads SWAP(|riO |qiP ) = |qiO |riP . Note that when dO 6= dP the swap operator effectively changes di- II. FOURIER TRANSFORM mensions of the two subspaces HO and HP . This imposes nevertheless no restrictions on our implementation. The In this section, we present a recursive scheme for the resulting mode |qiO |riP can be now identified with a construction of the Fourier transform that acts on the single index |ki in a way analogous to Eq. 2 as OAM of a single photon. The initial state of a pho- Pd−1 |qiO |riP ⇔ |ki = |q × dO + ri . (5) ton is a superposition of OAM eigenstates j=0 αj |ji. The Fourier image of such a state is then a superposition At the end, one obtains the Fourier image of the OAM Pd−1 k=0 βk |ki, where coefficients βk satisfy eigenstates of a single photon leaving the device along a single output path. For details see Appendix A. d−1 1 X i 2π jk We have thus demonstrated how to construct a high- βk ≡ √ e d αj. (1) d dimensional Fourier transform using lower-dimensional j=0 Fourier transforms that act on the OAM and path de- Our implementation of the Fourier transform is inspired grees of freedom. In the following section we present the by the classical fast Fourier transform algorithms [33], an corresponding implementation scheme. idea already applied in the field of quantum information for the path degree of freedom [31, 32, 34, 35]. To implement the d-dimensional Fourier transform in III. IMPLEMENTATION the OAM, we first decompose the total d-dimensional OAM space of a photon into a tensor product H = HO ⊗ The general recursive scheme based on the decompo- HP , as shown in Fig. 1 a). There, a dO-dimensional sub- sition in Eq. (3) works in principle for any dimension space HO is spanned by a subset of original OAM eigen- d, provided that d is not a prime number. Nevertheless, states and HP is a dP -dimensional subspace represented from now on we restrict ourselves only to dimensions that M by the path degree of freedom, such that d = dO × dP . are powers of two, i.e. d = 2 for some M ∈ N. For these 3

a) Fourier transform 푝0 (푑 ) 푑 = 푑 × 푑 퐹 푂 0 for 푂 푃 푂퐴푀 푝0 푝1 (푑푂) 1 훼 푝 퐹푂퐴푀 1 푝2 퐹(푑푂) 2 훼

푂퐴푀 OAM−

...... (푑) ... = 푆 (푑푃) −Path 푆−1 퐹푂퐴푀 (푑푃) 퐹푃푎푡ℎ (푑푂) 푝푑푃−3 (푑푂) swap 퐹 푑푃 − 3 훼 푂퐴푀 푝푑푂−2 푝푑푃−2 푑 = 푑 × 푑 (푑푂) 푂 푃 퐹 푑푃 − 2 훼 푂퐴푀 푝푑푂−1 푑 ≈ 푑 푝 푂 푑푃−1 (푑 ) 푂 푑 − 1 훼 퐹푂퐴푀 푃 푑푃 ≈ 푑

Sorter Fourier transform Phase gate Fourier transform Swap gate Combiner in OAM in Path b) Fourier transform for 푑 = 2 푘 휋 푘 훼 Dove prism rotated through: 푘 훼 = 푑 × 푑푃 (2) −1 푞-dimensional = 푆 푆 50:50 Beam-splitter 퐹푂퐴푀 (2) BS (2) 푆(푞) OAM sorter Mirror

FIG. 2. General scheme of the Fourier transform in the orbital angular momentum (OAM) of single photons. a) The Fourier transform in dimension d = dO × dP > 2 is constructed recursively making use of Fourier transforms in smaller dimensions dO and dP . The OAM eigenstates in the initial quantum superposition are at first redistributed by an OAM sorter S(dP ) such that along each of the dP different paths denoted by p0 through pdP −1 there propagate OAM eigenstates of the form |0i, |dP i, ..., |dP (dO − 1)i. The Fourier transform itself is then performed in four steps. The first step consists in the application of the (dO ) dO-dimensional Fourier transform FOAM on each path. Then a phase gate is applied on all paths, which is implemented by a (dP ) series of Dove prisms. In the third step a single dP -dimensional path-only Fourier transform FP ath is applied on the path degree of freedom for all OAM eigenstates. All beam-splitters in the path-only Fourier෍ transform are supplemented by two mirrors as demonstrated in b). The fourth step is represented by the OAM-Path swap gate. Finally, all states are recombined into a single output path by the second OAM sorter S−1 , which is operated in reverse. b) The elementary building block of the recursive (dO ) scheme – Fourier transform for d = 2 – consists of two OAM sorters and a single beam-splitter, which is complemented by two mirrors such that the OAM value of the incoming eigenstates is not affected by the reflection off the beam-splitter’s interface. dimensions an efficient implementation can be found for elementary building block – the Fourier transform each component of the general scheme. We discuss the for d = 2 – is depicted in Fig. 2 b). case with d 6= 2M at the end of this section. The efficient recursive scheme that implements the 3. The third module, which represents the phase gate Fourier transform is depicted in Fig. 2 a) and comprises CZ in Eq. 4, is implemented by a series of prop- six separate modules. Each module is described in detail erly rotated Dove prisms. They impart additional below. phases to OAM eigenstates that pass through them. 4. The fourth module comprises a single path-only 1. In the first module, the initial superposition of d dP -dimensional Fourier transform, which performs OAM eigenstates is split such that there is a smaller (dP ) transformation FP ath in Eq. 3. This transform number dO of OAM eigenstates propagating along affects only the path degree of freedom and acts each of dP different paths. This splitting, corre- identically onto each OAM eigenstate. The im- sponding to relabeling |ji → |miO |liP in Eq. 2, is plementation of the path-only Fourier transform in performed by the dP -dimensional OAM sorter. The terms of beam-splitters and phase-shifters is given structure and operation of the OAM sorter are de- in Refs. [31, 32]. To compensate for the OAM inver- scribed in detail in the following section. sion |miO → |−miO due to reflection off the beam- splitter’s interface, each beam-splitter in the path- 2. The second module consists of a collection of identi- only Fourier transform has to be complemented by cal d -dimensional OAM Fourier transforms, each O two extra mirrors as demonstrated in Fig. 2 b). of which is applied onto a different path to imple- (dO ) ment FOAM in Eq. 3. The dO-dimensional Fourier 5. The fifth module is the SWAP gate which reorders transform itself is constructed recursively following the coefficients of the joint OAM-Path state such the same pattern as the one presented in this sec- that the coefficient for mode |miO |liP becomes tion when one replaces d with the value of dO. The the coefficient for mode |liO |miP . The swap gate, 4

whose structure and working principle are one of a) OAM Sorter

the main results of this paper, is described in de- 푝0 tail in a separate section. Without the swap gate, ȁ0ۧ ȁ푑푃ۧ ȁ2 푑푃ۧ ⋯ ȁ(푑푂 − 1) 푑푃ۧ 푝1 the OAM sorter in the sixth module would also ȁ0ۧ ȁ푑푃ۧ ȁ2 푑푃ۧ ⋯ ȁ(푑푂 − 1) 푑푃ۧ

ퟐퟎ

… … … … perform undesirable additional permutation of the … output OAM eigenstates. 푆 푝푑푃−2 ȁ0ۧ ȁ푑푃ۧ ȁ2 푑푃ۧ ⋯ ȁ(푑푂 − 1) 푑푃ۧ 6. In the sixth module, all OAM eigenstates are 푝푑푃−1 rerouted into a single output path by the dO- ȁ0ۧ ȁ푑푃ۧ ȁ2 푑푃ۧ ⋯ ȁ(푑푂 − 1) 푑푃ۧ dimensional OAM sorter, which is operated in reverse. This action corresponds to the relabeling in b) Structure of OAM Sorter c) OAM Exchanger 푝0 푝0 Eq. 5. 0 1 2 ȁ2푚푘ۧ ȁ2푚푘ۧ 2 2 2 푘 푝4 It is important to note that OAM eigenstates enter- 푝2 ing the lower-dimensional OAM Fourier transforms are 22 ȁ 2푚 + 1 푘ۧ of the form |0iO, |dP iO,..., |dP (dO − 1)iO (in the first 푝6 푘 recursion), so the difference between two successive OAM ȁ2푚푘ۧ 푝1 eigenstates, or multiplicity, is dP . This fact has to 21 22 be reflected in the order of OAM exchangers (see the 푝5 ȁ푚ۧ 푐ȁ푚ۧ 푘 next section) and rotation of Dove prisms in the lower- 푝 푠ȁ푚 − 푘ۧ 3 dimensional Fourier transforms. Specifically, the order of 22 푝 all exchangers has to be multiplied by dP and the angle 7 푚 휋 푚 휋 푐 = cos , 푠 = sin of rotation for all Dove prisms has to be divided by dP 2 푘 2 푘 (compare also the form of angle α in Fig. 2). With each recursion the multiplicity of input eigenstates increases FIG. 3. OAM Sorter. a) The OAM sorter for a general power- correspondingly. of-two dimension d . All OAM eigenstates from the sub- The recursive scheme in Fig. 2 requires a number of P space {|0i ,..., |dP − 1i } get transformed into the eigen- optical elements that scales linearly with the dimension, O O state |0iO propagating in different paths. Eigenstates from as is shown in section VI. This scaling is made possible the subspace {|dP iO ,..., |2 dP − 1iO} are sorted analogously, by the efficient implementation of OAM sorters, swap op- but the resulting OAM eigenstate is |dP iO. This modulo erators and the path-only Fourier transforms. Such an property holds for arbitrarily large OAM values of input eigen- implementation is nevertheless available only for dimen- states. b) The binary-tree-like structure of OAM exchangers sions of the form d = 2M . For dimensions that are not a with an increasing order works as a sorter of OAM eigen- power of two alternative implementations exist, but these states. Here, a specific example for dP = 8 is shown. c) may require asymptotically more resources. The extreme The OAM exchanger EXk is a composition of two holograms and the elementary OAM parity sorter [36]. OAM eigenstates case is when d is a prime number, where the known recur- that are even multiples of the order k of the exchanger and sive schemes cannot be used. The brute-force alternative enter its upper port leave the output upper port unaffected. to our recursive scheme, which works for any dimension, The odd multiples entering the upper port are rerouted to the consists of three steps. In the first step, a d-dimensional lower output port and loose k quanta of OAM. Other OAM OAM sorter transforms the initial OAM eigenstates into eigenstates that are not multiples of k leave the exchanger propagation modes. In the second step, a d-dimensional in a superposition of eigenstates and paths. A single OAM path-only Fourier transform is applied, which can be im- exchanger can thus work as an identity, a switch and a beam- plemented using Reck et al. design [30]. In the third splitter for different input eigenstates. For eigenstates enter- step, the resulting propagation modes are transformed ing the lower input port the exchanger works analogously. back to the OAM with the help of another OAM sorter, which is operated in reverse. The number of elements in such a brute-force approach scales quadratically with the dimension. It is still an open question how to implement output paths p0 through pdP −1. All such input eigen- efficiently the scheme in Fig. 1 for a general dimension states leave the sorter in OAM eigenstate |0iO. Various d 6= 2M . designs of the sorter have been realized using e.g. the multi-plane light conversion [37, 38], light scattering in random media [39] or light propagation through two spe- IV. OAM SORTER cially designed phase plates [40]. Throughout this paper we consider the design due to Leach et al. [36, 41–44] The first stage of the setup for the Fourier transform and discuss alternative designs in section VI. This de- consists of an OAM sorter. The OAM sorter is a device sign has the following modulo property. When an OAM that transforms the OAM eigenstates of incoming pho- eigenstate |miO with m ≥ dP is injected into the sorter, tons into different propagation modes. For a fixed dimen- its propagation is analogous to that of OAM eigenstates sion dP the sorter sorts input eigenstates |0iO through with m < dP except that the output eigenstate is no |dP − 1iO, which enter the first path p0, into separate longer |0iO. It turns out that the input OAM eigenstate 5

miO |0iP gets transformed according to relations a) ȁ3ۧ ȁ3ۧ| ȁ3ۧ 4 2 1 1 2 ȁ−1ۧ m S (|mi |0i ) = d m mod d , (6) ȁ1ۧ dP ) O P P P 1 4 ȁ1ۧ) dP O P ȁ−3ۧ 1 ȁ5ۧ ȁ5ۧ where bxc is the integral part of x ∈ R and S(dP ) denotes ȁ2ۧ ȁ5ۧ 2 4 ȁ5ۧ a dP -dimensional OAM sorter. This modulo property is 2 ȁ1ۧ ȁ3ۧ illustrated in Fig. 3 a) and corresponds exactly to the 4 ȁ3ۧ relabeling introduced in Eq. 2. ȁ−1ۧ The sorter is constructed as a binary-tree network of b) 훼 휋 OAM-manipulating elements in a way shown in Fig. 3 ȁ푚ۧ 푐ȁ푚ۧ 푐 = cos α, 푘 2 푘 b). Each of these elements, henceforth referred to as 훼 휋 푠ȁ푚 − 푘ۧ 푠 = sin OAM exchangers, is an interferometric device composed 2 푘 of an OAM parity sorter [36, 43] and two holograms, which shift the OAM value of the input eigenstate [42]. c) Note that additional permutation of paths is necessary in ȁ2푚푘 + 훥푘ۧ ȁ2푚푘 + 훥푘ۧ −∆푘, 푘 푘 Fig. 3 b) to comply with the order of paths depicted in ȁ(2푚 + 1)푘 + 훥푘ۧ Fig. 3 a). The OAM exchanger EXk exhibits three modes ȁ2푚푘 + 훥푘ۧ of operation based on its order k and the state of the incoming photon, see Fig. 3 c). A photon in OAM eigenstate |m ki that enters the upper port of the exchanger d) 1 2 4 EXk leaves its upper or lower output port depending on 1 2 the parity of m ∈ Z. All other input OAM eigenstates, 2,4 1 4 H2 which are not multiples of k, leave the exchanger in a 1 E8 superposition of both output ports. A single exchanger 1,4 1,2 2 4 therefore behaves either as an identity, or as a switch, or 1,2 2 H4 as a beam-splitter with varying splitting ratio for differ- 3,4 4 ent OAM eigenstates. This beam-splitter-like property was first utilized in Ref. [45] for the special case when the order of the exchanger is k = 2. In this paper we show that a single exchanger of order k works effectively FIG. 4. OAM-Path swap operator. a) The network of OAM as 4k different beam-splitters simultaneously. For a more exchangers with increasing orders of the form 2j represents detailed description of the exchanger refer to Appendix a naive generalization of the OAM sorter. By propagation B. through this network, superpositions of OAM eigenstates and paths are introduced into the output state as exemplified for eigenstates |2i and |5i entering the fifth input port. b) Such V. OAM-PATH SWAP an undesirable behavior can be counteracted by adding a collection of holo-beam-splitters into the network. The holo- beam-splitter HBS(α,k) of order (α, k) is a device consisting In the final part of the setup of the Fourier transform, of a conventional beam-splitter with a splitting ratio απ/(2k) the swap operator is utilized. Its operation on individual and two holograms of opposite values k and −k. For fur- modes can be summarized as ther details refer to Appendix B. c) The crucial feature of the holo-beam-splitter is that it effectively shifts the split- SWAP(|miO |liP ) = |liO |miP , (7) ting properties of the OAM exchangers, such that eigenstates where |mi |li denotes an OAM eigenstate 0 ≤ m < d |mk + ∆ki, that would otherwise leave the exchanger of or- O P O der k in a superposition, leave in only one of the two output in path p . Analogously to the OAM sorter discussed in l ports of the exchanger. d) As demonstrated in Appendix the previous section, also the OAM-Path swap operator C, the resulting network implementing the OAM-Path swap exhibits the modulo property for m ≥ dO. comprises the network of exchangers EdP together with the The swap operator can be understood as a generaliza- series of gradually larger networks Hj of holo-beam-splitters. tion of the OAM sorter [46]. One could naively expect In the figure a special case of a swap operator for dP = dO = 8 that a complete network of OAM exchangers of increas- is shown. ing orders 20 to 2dP −1 works as a swap, see the example for dP = 8 in Fig. 4 a). Unfortunately, such a network not only permutes the input eigenstates as it is supposed with 0 < ∆k < k. The OAM value of this eigenstate to, but also gives rise to superpositions of multiple eigen- is not a multiple of the order of the OAM exchangers. states and paths. When an OAM eigenstate is injected From that point onward, all exchangers work on such an to any of the input ports pl of such a network, where eigenstate as beam-splitters with varying splitting ratios l ≥ 3, it leaves the network in a superposition of as and the output state is thus a superposition of different many as dP /2 different modes. This is caused by the eigenstates in different paths. Specifically, for eigenstate fact that at some point of its evolution through the net- |2 m k + ∆kiO entering the upper port pa of the OAM work the initial eigenstate attains the form |m k + ∆kiO exchanger EXk of order k one obtains (omitting a global 6 phase) Hilbert space is decomposed into a product of two subspaces of dimensions dO and dP , such that d = dO × dP . EXk(|2mk + ∆kiO |aiP ) = The values of dO and dP are nevertheless not fixed. One can search for such a combination of d and d that P O π∆k minimizes the number of beam-splitters in the setup of cos |2mk + ∆kiO |aiP 2k the Fourier transform, while still satisfying d = dO × dP . π∆k Our simulations show that the optimal number of beam- + sin |(2m − 1)k + ∆ki |bi , (8) 2k O P splitters scales approximately linearly as 6.037 × d. This optimal scenario tends to prefer choices with dP ≈ dO. In and analogously for eigenstates |(2m + 1) k + ∆kiO and Appendix D the linear scaling is analytically confirmed 2M the lower input port pb. The exchanger EXk in this spe- for a subset of dimensions of the form d = 2 with cific case therefore works effectively as a beam-splitter M ∈ N, for which dP = dO is used. Note that the log- with the splitting ratio equal to π∆k/(2k). arithmic scaling of the Fourier transform setup reported The undesirable emergence of superpositions is avoided in Ref. [24] relates to the number of elementary gates, when we augment the network with a collection of holo- not actual optical elements. When implementing the pro- beam-splitters. A holo-beam-splitter HBS(α,k) of order posal in Ref. [24] with beam-splitters, their number scales 5 (α, k) is a passive optical device consisting of a conven- as O(d(log2(d)) ). tional beam-splitter with a splitting ratio απ/(2k) and Apart from the stability, the performance of the setup two holograms of opposite values k and −k. Its oper- is also negatively affected by the losses introduced by ation on eigenstates entering its upper port is summa- beam-splitters and other optical elements. It turns out rized in Fig. 4 b) and its detailed structure is presented that the number of these additional optical elements also in Appendix B. A crucial observation is that one can scales linearly in the dimension d. For the optimal sce- force the OAM exchanger of order k to work as a mere nario described above, one needs approximately 4.860×d identity or switch even for eigenstates |m k + ∆kiO that Dove prisms, 3.037 × d holograms and 2.976 × d phase- are not multiples of k. One can do so by prepending a shifters. For details see Appendix D. The linear scaling holo-beam-splitter of order (−∆k, k) to the exchanger, as of the whole scheme is made possible by an efficient im- demonstrated in Fig. 4 c) [47]. We obtain transformation plementation of the swap gate, which requires approxi- rules (again omitting a global phase) mately 3d log2(d)/2 beam-splitters. There are alternative brute-force implementations of the swap gate, but these

(EXk · HBS(−∆k,k))(|2mk + ∆kiO |aiP ) = require asymptotically larger number of beam-splitters. |2mk + ∆ki |ai , (9) The OAM sorter design we adopted is the one due to O P Leach et al. [36], which is based on a network of interfer- where analogous relations hold also for eigenstates ometers. Our scheme makes use of the modulo property of this specific design in each recursion. Nevertheless, |(2m + 1) k + ∆kiO and the lower input port pb. In this formula the holo-beam-splitter and the exchanger act on after all the recursions are performed, the initial part of the same states as in Eq. 8, but this time no superposition the final setup represents a single d-dimensional sorter, emerges. where the modulo property plays no role any more. One One can stack multiple setups in Fig. 4 c) to create a can therefore replace this initial part with a sorter built larger network. This way we arrive at the setup that im- using alternative designs. For example, one could use plements an OAM-Path swap operator for general power- the design of Ref. [40] that uses only two plates with special phase-profiles. This way one can save around 2d of-two dimensions dO and dP . The swap gate consists of a network of exchangers and a series of networks of holo- beam-splitters and equal number of other elements, such beam-splitters of increasing size as shown in Fig. 4 d) for that the scheme then requires approximately 4d beam- splitters. the case of dO = dP = 8. For the detailed explanation of the construction and structure of the swap operator refer to Appendix C. The resulting network works as a proper sorter for all input ports, where the OAM value of VII. OAM VS. PATH output eigenstates contains information about the path into which the original eigenstate was injected. The proposed scheme makes use of an interplay between the OAM and path degrees of freedom to efficiently perform the Fourier transform in the OAM de- VI. SCALING PROPERTIES gree of freedom. No other properties of incoming photons, such as the polarization, are affected. Nevertheless, In real-world applications it is necessary to assess the the scheme can be after slight modification used also as effect of imperfections and noise on the stability and a path-only Fourier transform, where the OAM degree of overall feasibility of the setup. The stability of the freedom plays the role of an intermediary that does not scheme is the more of concern the more interferometers appear either in the input or the output state. Specif- are employed. Their number depends on the number of ically, the first part of the recursive scheme, see Fig. 2 beam-splitters. In our scheme the original d-dimensional a), represented by a series of OAM sorters of decreasing 7 dimensions, can be removed completely. We are then cussions. XQG thanks Bin Sheng and Zaichen Zhang left with d input ports. The last part of the scheme for support. This work was supported by the Austrian has to be adjusted by removing the very last reverted Academy of Sciences (OeAW), the European Research sorter and adding a series of additional sorters to obtain Council (SIQS Grant No. 600645 EU-FP7-ICT), and d output ports. This OAM-enhanced scheme of the path- the Austrian Science Fund (FWF): F40 (SFB FoQuS) only Fourier transform shows better scaling properties in and W 1210-N25 (CoQuS). XQG acknowledges support terms of the number of beam-splitters than the scheme from the National Natural Science Foundation of China presented in Refs. [31, 32]. The OAM-enhanced scheme (No. 61501109). BD acknowledges support from an ESQ requires O(d) beam-splitters as opposed to O(d log2(d)) Discovery Grant of the Austrian Academy of Sciences beam-splitters of the original scheme with the crossing (OAW) and the Austrian Science Fund (FWF) through point from which OAM-enhanced scheme prevails occur- BeyondC (F71). ring at d = 8192 = 213. This improvement is made JK and XQG contributed equally to this work. possible by optimal redistribution of coefficients of the quantum state between the OAM and path degrees of freedom. Appendix A: Decomposition of the Fourier Transform

VIII. CONCLUSION In this section we prove the validity of the decomposition formula in Eq. 3 by calculating explicitly the action We proposed a scheme for an efficient implementation of this decomposition on input OAM eigenstates. At first of the Fourier transform that acts on the orbital angular we relabel the input OAM eigenstate |ji in accord with momentum of single photons. Only commercially acces- the main text (cf. Eq. 2) as |ji = |m dP + li = |miO |liP . sible optical elements are used in our scheme. An in- Then the application of the OAM Fourier transform on tegral component of the scheme is the OAM-Path swap the first ket yields operator, which is a generalization of the OAM sorter dO −1 for multiple input ports. In its implementation a heavy 1 X i 2π mr |ji = |mi |li → √ e dO |ri |li . (A1) use is made of non-trivial parallel operation of OAM ex- O P d O P O r=0 changers, which are elementary building blocks of the

OAM sorter. A single exchanger, a passive element com- The phase gate multiplies each mode |riO |liP by posed among others of two conventional beam-splitters, exp (2πirl/d) and the swap gate exchanges the two kets. can work as many beam-splitters with varying splitting After the sequential application of the phase gate, the ratio at the same time. This property may be used in path-only Fourier transform, and the swap gate the state a more general framework, where each OAM eigenstate on the right-hand side of Eq. A1 transforms into undergoes a different complex, yet precisely tailored, evolution by propagating through the identical network of dO −1 dP −1 1 X X i 2π mr+i 2π rl+i 2π lq standard optical elements. Even though algorithms for √ e dO d dP |qi |ri . (A2) d d O P the decomposition of a general unitary into separate gates O P r=0 q=0 for path and OAM degrees of freedom exist [48–50], our scheme is to our knowledge the first explicit example of It is easy to see that the exponent simplifies into such a network. 2π 2π 2π 2π The number of optical elements required in our scheme mr + rl + lq ≡ (m dP + l)(q dO + r) dO d dP d scales as O(d), which is an improvement over the scal- 2π ing O(d log (d)) of the Fourier transform setup acting on ≡ jk (mod 2π), (A3) 2 d path-encoded qudits [31, 32]. Our scheme can be after a slight modification used also to implement the path-only where we defined |ki = |q dO + ri = |qiO |riP in accor- Fourier transform while preserving the scaling properties dance with Eq. 5 and we used relation d = dO × dP . As O(d). The modulo property of the scheme allows one a result we obtain the transformation rule to use different d-dimensional OAM subspaces, such as d−1 {|−d/2 + 1i ,..., |d/2i}, which is naturally produced in 1 X i 2π jk |ji → √ e d |ki , (A4) the process of parametric down-conversion and which im- d poses less stringent requirements on the precision of the j=0 OAM-manipulating elements. which is equivalent to formula in Eq. 1. This completes the proof.

ACKNOWLEDGMENTS Appendix B: Optical elements The authors thank Anton Zeilinger, Mario Krenn, Manuel Erhard, Armin Hochrainer, Marcus Huber, The OAM sorter as well as OAM-Path swap consist of Robert Fickler and Elizabeth Agudelo for valuable dis- two kinds of passive optical elements—OAM exchangers 8

FIG. 5. OAM exchangers and holo-beam-splitters. a) The OAM exchanger EXk of order k is an interferometric device made out of two holograms with opposite values and the OAM FIG. 6. Identities for OAM exchangers and holo-beam- parity sorter [36]. Its operation is captured by the transfor- splitters. a) A composition of a holo-beam-splitter of or- mation formulas depicted at the bottom. One exchanger of der (α, k) with α = −∆k and an OAM exchanger of order order k effectively works as 4k different beam-splitters with k effectively works as a single exchanger of order k. This varying splitting ratio for different OAM eigenstates. Note effective exchanger either passes on or reroutes OAM eigen- the reversed order of sine and cosine functions in the for- states of the form |m k + ∆ki that enter the upper port, where mula for |m, p2i. b) The holo-beam-splitter HBS(α,k) of order 0 < ∆k < k is fixed (and similarly for eigenstates entering (α, k) comprises a beam-splitter with splitting ratio α π/(2k) the lower port). b) Exchangers of order 2k and holo-beam- accompanied by two holograms with opposite values k and splitters of order (α, 2n) commute in a sense shown in the −k. In the diagram, the implementation of the variable split- figure for n ≥ m + 1. The minus sign is present only for the ting ratio beam-splitter is demonstrated with the help of two special case of n = m+1. c) Two pairs of holo-beam-splitters 50:50 beam-splitters. The operation of the holo-beam-splitter of orders (α, k) and (β, n), respectively, commute in a sense is captured by the transformation formulas depicted at the shown in the figure. d) Two holo-beam-splitters of orders bottom. (α, n) and (β, n), respectively, can be combined into a single holo-beam-splitter of order (α + β, n). and holo-beam-splitters. Their structure as well as operation are depicted in Fig. 5. Whereas the holo-beam- splitter has a fixed splitting ratio for all OAM eigenstates, the exchanger splits the incoming OAM eigenstates into the two output ports according to a splitting ratio that depends on the OAM eigenstate. As a result, a single to such an approach we note that the composition of an exchanger of order k works as 4k different conventional exchanger of order k and a holo-beam-splitter of order beam-splitters with splitting ratios 0, π/(2k), 2π/(2k), (α, k) with α = −∆k performs the identical rerouting ..., (4k − 1)π/(2k). operation, see Fig. 6 a). The advantage of the latter ap- The trivial example of an OAM-Path swap is a single proach is that additional exchangers can be inserted be- exchanger of order k. OAM eigenstates |m ki entering tween the two elements. These additional exchangers can its input port, where m k is a multiple of k, are either reroute unwanted terms away from the setup and inject not affected by the exchanger, or rerouted to the other wanted terms in instead. By using this idea iteratively, output port. All remaining OAM eigenstates |m k + ∆ki the OAM-Path swap operator can be constructed. with 0 < ∆k < k leave the exchanger in a superposition of the two output ports. The same exchanger can be used to reroute such non-multiple OAM eigenstates as It is a matter of several simple goniometric transfor- well if input eigenstates are first shifted by −∆k with a mations to prove the identities demonstrated in Fig. 6 hologram as shown on the left-hand side in Fig. 6 a). In b), c) and d). These identities will nonetheless greatly such a case though, the superpositions are introduced to simplify the following discussion of the working principle the multiple OAM eigenstates instead. As an alternative of the OAM-Path swap operator. 9

Appendix C: General structure of the OAM-Path there due to the d/¯ 2 dimensional swap. At the end, we swap obtain the OAM-Path swap in dimension d¯that correctly sorts OAM eigenstates injected to any of its d¯input ports To illustrate the operating principle of the OAM-Path as shown in Fig. 7 i). The network of exchangers forms swap for general dimensions d and d , let us focus first the routing part of the swap operator, which we refer P O to as an E block, and the presence of superpositions in on the simplest case with dP = dO = 4, which is depicted in Fig. 7 a)–d). When the network of OAM exchangers is output modes is corrected for by a series of networks of used to reroute OAM eigenstates entering different input increasing size made out of holo-beam-splitters. We refer ports, undesirable superpositions of eigenstates are cre- to these smaller networks as H blocks. ated by higher-order exchangers. All eigenstates entering In order for the setup to implement the swap trans- the first or the second path are rerouted correctly as in formation in the sense of Eq. 7, an additional per- the case of the OAM sorter. Nevertheless, for all other mutation of paths has to be appended to the setup paths, the incoming eigenstate leaves the network in a in each recursion (cf. Fig. 7 e)). This permutation superposition. Utilizing the observation in Fig. 6 a) we reroutes eigenstates from paths (p0, p1, p2, . . . , pd¯−1) to can insert holo-beam-splitters as demonstrated in Fig. 7 paths (p0, p2, p4, . . . , pd¯−2, p1, p3, . . . , pd¯−1). As the last c) to fix the undesirable splitting of eigenstates by the stage, a Dove prism rotated through the same angle has last column of exchangers (see also the previous section). to be inserted in each path, compare again with Fig. 7 e). Applying identity from Fig. 6 b) we can swap the first These Dove prisms correct for alternating phases of eigen- column of exchangers with the holo-beam-splitters. This states leaving the network. Let us note that H blocks way we fixed behavior of exchangers for eigenstates enter- can be simplified even more as identity similar to that in ing the third and forth input ports. Before the addition Fig. 6 b) exists for beam-splitters and holograms. This of holo-beam-splitters, eigenstates injected into the first way, all holograms present in holo-beam-splitters can be and second port were rerouted correctly. This property put to the sides of the block, whose middle part is then would be destroyed had we left the upper holo-beam- formed merely by conventional beam-splitters. splitter in place. When we remove it and keep only the In cases when dO 6= dP one constructs the network ¯ lower holo-beam-splitter, we arrive at the setup depicted for dimension d = max (dO, dP ) and then uses only first in Fig. 7 d), which sorts correctly all OAM eigenstates dP input ports and first dO output ports of the network. injected to any of the input ports. Obviously, some elements then do not enter the evolution of injected OAM eigenstates and can be removed from the The construction of the OAM-Path swap for general network with no effect on the swap functionality. dimensions is recursive and relies heavily on the property illustrated in Fig. 6 a). Let us assume for the moment ¯ ¯ M that dP = dO ≡ d. The swap in dimension d = 2 is constructed from two swaps in dimension d/¯ 2 = 2M−1 Appendix D: Scaling of the number of elements for high dimensions that are connected by a layer of additional exchangers. For a specific example in d¯ = 8 refer to Fig. 7 e), where the d/¯ 2-dimensional swap is presented in Fig. 7 d). All The number of optical elements in the setup reflects OAM eigenstates entering first d/¯ 2 ports are sorted cor- the setup’s complexity. In this section we estimate how rectly due to properties of d/¯ 2-dimensional swaps. Nev- this number scales with the dimension d of the OAM ertheless, eigenstates that are injected into the other d/¯ 2 Hilbert space. We focus first on the beam-splitters. In ports leave the network in a superposition of d/¯ 2 output our scheme the incoming state is transformed into a state ports. The reason is that the layer of additional exchang- with dO OAM eigenstates propagating in dP different ers adds one quantum of OAM to such eigenstates and paths, such that d = dO × dP . The two numbers dO and their value is thus no longer a multiple of the order of ex- dP are not fixed and their optimal values can be found, changers in the remaining layers. These exchangers then for which the number of beam-splitters in the setup is act not as switches, but rather as genuine beam-splitters. minimal. It turns out that the optimal scenario tends to prefer the choice with d ≈ d . Here we assume from As follows from Fig. 6 a), the switch-like behaviour O P√ the beginning that d = d = d in each recursion and can be restored if specifically chosen holo-beam-splitters O P 2k are added to the remaining layers of exchangers, see also study dimensions of the form d = 2 , where k ∈ N. Fig. 7 f). Making use of identities in Fig. 6 b) and c) all At first, we have to determine the number of beam- such holo-beam-splitters can be aggregated at the begin- splitters required by individual components of the scheme ning of the network as shown in Fig. 7 g). At this point, schown in Fig. 2. For the OAM sorter and the one has to recall that the holo-beam-splitters were intro- path-only Fourier transform in dimension dP one ob- ¯ (sort) (pFT) duced to correctly reroute eigenstates entering last d/2 tains NBS (dP ) = 2(dP − 1) [44] and NBS (dP ) = ¯ dP ports. Eigenstates entering first d/2 ports should not be 2 log2(dP ) [31, 32], respectively. The OAM-Path swap affected by the additional holo-beam-splitters. All such comprises one E block and a series of H blocks of increas- holo-beam-splitters are therefore removed from the first ing size. The E block is built out of log2(dP ) groups of ¯ d/2 paths. This step is illustrated in Fig. 7 h). To save re- k-order exchangers EXk, where each group contains dP /2 sources, one can merge the additional holo-beam-splitters exchangers, cf. Fig. 4 d). The number of conventional in last d/¯ 2 paths with those holo-beam-splitters that are beam-splitters implementing the E block is thus equal to 10

FIG. 7. The structure of OAM-Path swap operator. a), b) A specific case of a network of OAM exchangers for d¯ = 4 and propagation of OAM eigenstates |2i and |3i entering the third and forth paths. The OAM eigenstates leave the network in a superposition. c) When holo-beam-splitters are inserted as suggested in the figure the undesirable splitting of eigenstates is avoided. d) The first column of exchangers can be swapped with the holo-beam-splitters. Futhermore, the upper holo-beam- splitter can be removed such that it does not affect eigenstates injected into the first and second port. As a result, we obtain a setup that sorts correctly all OAM eigenstates injected to any of the input ports. e) The idea of inserting holo-beam-splitters can be generalized to higher dimensions. We demonstrate the general idea for d¯= 23. The setup then consists of two setups for d¯= 4 from d) that are connected by four exchangers of order 1. Notice also that orders of exchangers and holo-beam-splitters in the two setups have to be multiplied by two. To comply with the formula in Eq. 7 an additional permutation of paths and a series of properly rotated Dove prisms are necessary in the final part of the setup. As the last part does not alter the evolution of eigenstates through the network we omit it in the following. f) The holo-beam-splitters from the d/¯ 2-dimensional swaps can be moved to the left thanks to identity in Fig. 6 b). At this point, all OAM eigenstates that enter any of the first d/¯ 2 = 4 ports are sorted correctly. For the other four input ports the OAM eigenstates undergo more complex evolution and leave the setup in a superposition. In analogy to c), appropriately chosen holo-beam-splitters are inserted into the setup to preclude creation of such superpositions. g) Using identities from Fig. 6 b) and c) all holo-beam-splitters can be aggregated in front of the network of exchangers. h) Holo-beam-splitters aggregated this way in the first four paths nevertheless negatively affect OAM eigenstates propagating through these paths. We can remove these holo-beam-splitters to restore the sorting properties of the network for these input paths. In the case of the last four paths, we can use identity from Fig. 6 d) to merge all holo-beam-splitters originating in the setup for d¯ = 4 with the aggregated holo-beam-splitters. For d¯ = 4 there is only one holo-beam-splitter. We can merge it with its neighbour as made clear in the figure. i) This way we obtained the setup for the OAM-Path swap in d¯ = 8. All OAM eigenstates that enter any of the eight input ports are sorted correctly. We can identify two conceptually different parts of the swap operator. The network of exchangers, henceforth referred to as an E block and a series of networks of increasing size made out of holo-beam-splitters, which we refer to as H blocks.

(E) k NBS (dP ) = dP log2(dP ). An H block of size 2 has the beam-splitters for H blocks is thus 2k k same structure as the E block, requiring 2 log(2 ) holo- beam-splitters. (In Fig. 5 b) the asymetric beam-splitter log2(dP )−1 k (H) X 2 dP is constructed as an interfereometer, here we neverthe- N (d ) = log (2k) = log (d )−d +1. BS P 2 2 2 2 P P less take it as a single element.) The total number of k=1 (D1) To construct an OAM-Path swap in dimension dP one (swap) (E) (H) therefore needs NBS (dP ) = NBS (dP ) + NBS (dP ) = 3 2 dP log2(dP ) − dP + 1 beam-splitters. 11

semi-brute-force approach our approach

M d dO dP BS Dove holo phas total BS Dove holo phas total 1 2 1 2 5 4 2 2 13 5 4 2 2 13 2 4 1 4 16 12 6 7 41 16 12 6 7 41 3 8 2 4 40 28 14 19 101 39 35 20 19 113 4 16 4 4 92 60 30 47 229 89 75 40 43 247 5 32 4 8 204 124 62 111 501 181 143 84 87 495 6 64 4 16 444 252 126 255 1077 373 279 180 179 1011 7 128 8 16 956 508 254 575 2293 757 667 416 383 2223 8 256 16 16 2044 1020 510 1279 4853 1597 1355 768 799 4519 9 512 8 64 4348 2044 1022 2815 10229 3093 2587 1744 1551 8975 10 1024 32 32 9212 4092 2046 6143 21493 6205 4923 3008 3055 17191 11 2048 32 64 19452 8188 4094 13311 45045 12317 9659 5952 6015 33943 12 4096 32 128 40956 16380 8190 28671 94197 24605 19131 11904 11967 67607 13 8192 32 256 86012 32764 16382 61439 196597 49309 38075 23936 23935 135255

(d) TABLE I. Exact numbers of optical elements necessary to implement the OAM Fourier transform FOAM for several lowest M dimensions of the form d = 2 with corresponding dO and dP as determined by the optimization procedure. The numbers of beam-splitters (BS), Dove prisms (Dove), holograms (holo) and phase-shifters (phas) are presented together with their total sum (total) for each dimension. For comparison, the brute-force approach that consists in using the path-only Fourier transform (d) (d) FP ath supplemented with two OAM sorters is also presented. As we use the eﬃcient design of FP ath [31, 32], this approach is referred to as the semi-brute-force approach. One sees that from d = 8 onward our approach needs less beam-splitters than the semi-brute-force approach. When the total number of all the optical elements is considered, our approach is more resource-eﬃcient from d = 32 onward.

(FT) The number of conventional beam-splitters used in our and also directly calculate NBS (d1) = 16 so that in the implementation of the Fourier transform for dimensions end we arrive at 2k of the form dk := 2 is thus

(FT) (sort) (pFT) (FT) 27 N (dk) = 2N (dk−1) + N (dk−1) + N (d ) ≤ d = 6.75 d (D5) BS BS BS BS k 4 k k (swap) (FT) + NBS (dk−1) + dk−1 NBS (dk−1) = 2dk−1 log2(dk−1) + 3dk−1 − 3 + The number of beam-splitters required in our scheme of (FT) the Fourier transform therefore scales linearly with the + dk−1 NBS (dk−1), (D2) √ dimension where dO = dP = d = dk−1. Let us define ck := 2 dk log2(dk) + 3dk − 3 such that the last formula turns (FT) into NBS (d) ∼ O(d), (D6) (FT) (FT) N (dk) = ck−1 + dk−1 N (dk−1). (D3) BS BS k when we consider dimensions of the form d = 22 . The When we apply this relation recursively for k, k−1,..., 1, exact total numbers of beam-splitters required by our the final formula attains the form scheme are shown in Tab. I for several lowest dimensions M k k−1  k−1 ! of the general form d = 2 . Optimal choices of dO and (FT) X Y Y (FT) dP for these dimensions were found by an optimization NBS (dk) =  dl cj−1 + dl NBS (d1) j=2 l=j l=1 algorithm that searched for a setup with the minimum number of beam-splitters when a given dimension d was k X k j k (FT) fixed. Even in these cases with a general power-of-two = 22 −2 c + 22 −2N (d ) j−1 BS 1 dimension the number of beam-splitters still preserves j=2 the linear scaling. When fitting the data in Tab. I one k−1 j X 2 dk dk (FT) observes that this number relates to the dimension as = 2d + 3 − 1 + N (d ). k 22j 4 4 BS 1 j=1 N (FT)(d) ≈ 6.037 d. (D7) We can bound the sum in the previous expression as BS

k−1 k−1 ∞ X 2j X 2j X 1 ≤ ≤ = 1, (D4) In complete analogy to Eq. (D2) one can also find a re- 22j 22j 2j j=1 j=1 j=1 cursive formula for the number of other optical elements 12 present in the setup. Their explicit forms read for several lowest dimensions of the form d = 2M . By fitting the data from this table, one obtains the following (FT) NDove(dk) = dk−1 log2(dk−1) + 6dk−1 − 5 + behaviour (FT) + dk−1 N (dk−1), (D8) (FT) Dove NDove(d) ≈ 4.860 d, (D15) (FT) (FT) N (dk) = 2dk−1 log2(dk−1) + holo Nholo (d) ≈ 3.037 d, (D16) (FT) + dk−1 N (dk−1), (D9) (FT) holo Nphas (d) ≈ 2.976 d. (D17) 3 N (FT)(d ) = d log (d ) + d − 1 + (D18) phas k 2 k−1 2 k−1 k−1 (FT) Let us mention in this context that the Fourier trans- + dk−1 N (dk−1), (D10) phas form in the orbital angular momentum can be also im- (FT) (FT) (FT) plemented in a different way. In the brute-force approach where NDove, Nholo and Nphas stand for the number of Dove prisms, holograms and phase-shifters, respectively. the OAM Fourier transform is implemented with the help After performing the corresponding calculations it turns of a d-dimensional path-only Fourier transform and two out that the total numbers of Dove prisms, holograms d-dimensional OAM sorters, as is explained at the end and phase-shifters scale linearly with the dimension of section III in the main text. Since here we deal only with dimensions that are powers of two, we can use the (FT) 67 efficient design of Ref. [31] to implement the path-only N (d ) ≤ d < 5.6 d , (D11) Dove k 12 k k Fourier transform. For comparison, the number of opti- 7 cal elements required in this semi-brute-force approach is N (FT)(d ) ≤ d = 3.5 d , (D12) holo k 2 k k also shown in Tab. I. This approach needs more beam- 7 splitters than our scheme, especially in high dimensions. N (FT)(d ) ≤ d = 3.5 d . (D13) phas k 2 k k When the total number of all the elements in the setup (D14) are of concern, both our scheme and the semi-brute-force scheme are initially comparable, but already from d = 32 The exact numbers of these elements are listed in Tab. I onward our scheme is more resource-efficient.

[1] Peter W. Shor, “Algorithms for quantum computation: quantum cryptography with qutrits,” New Journal of discrete logarithms and factoring,” in Proceedings 35th Physics 8, 75–75 (2006). Annual Symposium on Foundations of Computer Science [10] Mehul Malik, Malcolm O’Sullivan, Brandon Rodenburg, (IEEE Comput. Soc. Press, 1994) pp. 124–134. Mohammad Mirhosseini, Jonathan Leach, Martin P. J. [2] Artur Ekert and Richard Jozsa, “Quantum computa- Lavery, Miles J. Padgett, and Robert W. Boyd, “In- tion and Shor’s factoring algorithm,” Reviews of Modern fluence of atmospheric turbulence on optical communi- Physics 68, 733–753 (1996). cations using orbital angular momentum for encoding,” [3] Peter W. Shor, “Polynomial-Time Algorithms for Prime Optics Express 20, 13195 (2012). Factorization and Discrete Logarithms on a Quantum [11] Stephen M. Barnett and D. T. Pegg, “Quantum theory Computer,” SIAM Journal on Computing 26, 1484–1509 of rotation angles,” Physical Review A 41, 3427–3435 (1997). (1990). [4] Andrew M. Childs and Wim van Dam, “Quantum al- [12] Sonja Franke-Arnold, Stephen M. Barnett, Eric Yao, gorithms for algebraic problems,” Reviews of Modern Jonathan Leach, Johannes Courtial, and Miles Padgett, Physics 82, 1–52 (2010). “Uncertainty principle for angular position and angular [5] William K. Wootters and Brian D. Fields, “Opti- momentum,” New Journal of Physics 6, 103–103 (2004). mal state-determination by mutually unbiased measure- [13] Eric Yao, Sonja Franke-Arnold, Johannes Courtial, ments,” Annals of Physics 191, 363–381 (1989). Stephen M. Barnett, and Miles J. Padgett, “Fourier re- [6] Stephen Brierley, Stefan Weigert, and Ingemar Bengts- lationship between angular position and optical orbital son, “All mutually unbiased bases in dimensions two to angular momentum,” Optics Express 14, 9071 (2006). five,” Quantum Info. Comput. 10, 803–820 (2010). [14] Yu Wang, Václav Potoˇcek,Stephen M. Barnett, and Xue [7] Thomas Durt, Berthold-Georg Englert, Ingemar Bengts- Feng, “Programmable holographic technique for imple- son, and Karol Zyczkowski,˙ “On mutually unbiased menting unitary and nonunitary transformations,” Phys- bases,” International Journal of Quantum Information ical Review A 95, 033827 (2017). 08, 535–640 (2010). [15] Radu Ionicioiu, “Sorting quantum systems efficiently,” [8] D. Giovannini, J. Romero, J. Leach, A. Dudley, Scientific Reports 6, 25356 (2016). A. Forbes, and M. J. Padgett, “Characterization of High- [16] Marek Zukowski,˙ Anton Zeilinger, and Michael A. Dimensional Entangled Systems via Mutually Unbiased Horne, “Realizable higher-dimensional two-particle en- Measurements,” Physical Review Letters 110, 143601 tanglements via multiport beam splitters,” Physical Re- (2013). view A 55, 2564–2579 (1997). [9] Simon Gröblacher, Thomas Jennewein, Alipasha Vaziri, [17] V´ıctor J. López-Pastor, Jeff S. Lundeen, and Flo- Gregor Weihs, and Anton Zeilinger, “Experimental rian Marquardt, “Arbitrary optical wave evolution 13

with fourier transforms and phase masks,” (2019), Mathematics of Computation 19, 297–297 (1965). arXiv:1912.04721. [34] S. Zhang, C. Lei, A. Vourdas, and J. A. Dunningham, [18] M. Yu. Saygin, I. V. Kondratyev, I. V. Dyakonov, S. A. “Applications and implementation of Fourier multiport Mironov, S. S. Straupe, and S. P. Kulik, “Robust Archi- devices,” Journal of Physics B: Atomic, Molecular and tecture for Programmable Universal Unitaries,” Physical Optical Physics 39, 1625–1637 (2006). Review Letters 124, 010501 (2020). [35] Gelo Noel M. Tabia, “Recursive multiport schemes for [19] Luciano Pereira, Alejandro Rojas, Gustavo Cañas,Gus- implementing quantum algorithms with photonic inte- tavo Lima, Aldo Delgado, and AdánCabello, “Universal grated circuits,” Physical Review A 93, 012323 (2016). multi-port interferometers with minimal optical depth,” [36] Jonathan Leach, Miles J. Padgett, Stephen M. Barnett, (2020), arXiv:2002.01371. Sonja Franke-Arnold, and Johannes Courtial, “Measur- [20] L. Allen, M. W. Beijersbergen, R. J. C. Spreeuw, ing the Orbital Angular Momentum of a Single Photon,” and J. P. Woerdman, “Orbital angular momentum of Physical Review Letters 88, 257901 (2002). light and the transformation of Laguerre-Gaussian laser [37] Nicolas K. Fontaine, Roland Ryf, Haoshuo Chen, modes,” Physical Review A 45, 8185–8189 (1992). David T. Neilson, Kwangwoong Kim, and Joel Carpen- [21] Mario Krenn, Marcus Huber, Robert Fickler, Radek Lap- ter, “Laguerre-Gaussian mode sorter,” Nature Commu- kiewicz, Sven Ramelow, and Anton Zeilinger, “Genera- nications 10, 1865 (2019). tion and confirmation of a (100 x 100)-dimensional en- [38] Guillaume Labroille, Bertrand Denolle, Pu Jian, Philippe tangled quantum system,” Proceedings of the National Genevaux, Nicolas Treps, and Jean-Fran¸coisMorizur, Academy of Sciences 111, 6243–6247 (2014). “Efficient and mode selective spatial mode multiplexer [22] Manuel Erhard, Robert Fickler, Mario Krenn, and An- based on multi-plane light conversion,” Optics Express ton Zeilinger, “Twisted photons: new quantum perspec- 22, 15599 (2014). tives in high dimensions,” Light: Science & Applications [39] Robert Fickler, Manit Ginoya, and Robert W. Boyd, 7, 17146–17146 (2018). “Custom-tailored spatial mode sorting by controlled [23] Juan Carlos Garc´ıa-Escart´ın and Pedro Chamorro- random scattering,” Physical Review B 95, 161108(R) Posada, “Universal quantum computation with the or- (2017). bital angular momentum of a single photon,” Journal of [40] Gregorius C. G. Berkhout, Martin P. J. Lavery, Johannes Optics 13, 064022 (2011). Courtial, Marco W. Beijersbergen, and Miles J. Padgett, [24] Xiaoqin Gao and Zhengwei Liu, “Universal Quan- “Efficient Sorting of Orbital Angular Momentum States tum Computation by a Single Photon,” (2019), of Light,” Physical Review Letters 105, 153601 (2010). arXiv:1909.09535. [41] Jonathan Leach, Johannes Courtial, Kenneth Skeldon, [25] Martin P. J. Lavery, David J. Robertson, Gregorius C. G. Stephen M. Barnett, Sonja Franke-Arnold, and Miles J. Berkhout, Gordon D. Love, Miles J. Padgett, and Jo- Padgett, “Interferometric Methods to Measure Orbital hannes Courtial, “Refractive elements for the measure- and Spin, or the Total Angular Momentum of a Single ment of the orbital angular momentum of a single pho- Photon,” Physical Review Letters 92, 013601 (2004). ton,” Optics Express 20, 2110 (2012). [42] Juan Carlos Garc´ıa-Escart´ın and Pedro Chamorro- [26] Malcolm N. O’Sullivan, Mohammad Mirhosseini, Mehul Posada, “Quantum multiplexing with the orbital angu- Malik, and Robert W. Boyd, “Near-perfect sorting of lar momentum of light,” Physical Review A 78, 062320 orbital angular momentum and angular position states (2008). of light,” Optics Express 20, 24444 (2012). [43] Manuel Erhard, Mehul Malik, and Anton Zeilinger, [27] Mohammad Mirhosseini, Mehul Malik, Zhimin Shi, and “A quantum router for high-dimensional entanglement,” Robert W. Boyd, “Efficient separation of the orbital an- Quantum Science and Technology 2, 014001 (2017). gular momentum eigenstates of light,” Nature Commu- [44] Xiaoqin Gao, Mario Krenn, Jaroslav Kysela, and Anton nications 4, 2781 (2013). Zeilinger, “Arbitrary d-dimensional Pauli X gates of a [28] Florian Brandt, Markus Hiekkamäki,FrédéricBouchard, flying qudit,” Physical Review A 99, 023825 (2019). Marcus Huber, and Robert Fickler, “High-dimensional [45] Xiaoqin Gao, Manuel Erhard, Anton Zeilinger, and quantum gates using full-field spatial modes of photons,” Mario Krenn, “Computer-inspired concept for high- Optica 7, 98 (2020). dimensional multipartite quantum gates,” (2019), [29] Xinbing Song, Yifan Sun, Pengyun Li, Hongwei Qin, arXiv:1910.05677. and Xiangdong Zhang, “Bell’s measure and implement- [46] Iulia Ghiu, “Simultaneous sorting many qudits using dif- ing quantum Fourier transform with orbital angular mo- ferent input ports,” Quantum Information Processing 18, mentum of classical light,” Scientific Reports 5, 14113 285 (2019). (2015). [47] This trick is somewhat similar to the use of an extra [30] Michael Reck, Anton Zeilinger, Herbert J. Bernstein, phase plate in the OAM parity sorter as reported in: and Philip Bertani, “Experimental realization of any dis- Haiqing Wei, Xin Xue, Jonathan Leach, Miles J. Pad- crete unitary operator,” Physical Review Letters 73, 58– gett, Stephen M. Barnett, Sonja Franke-Arnold, Eric 61 (1994). Yao, and Johannes Courtial, “Simplified measurement of [31] P. Törmä,I. Jex, and S. Stenholm, “Beam splitter real- the orbital angular momentum of single photons,” Optics izations of totally symmetric mode couplers,” Journal of Communications 223, 117–122 (2003). Modern Optics 43, 245–251 (1996). [48] Ish Dhand and Sandeep K. Goyal, “Realization of arbi- [32] Ronen Barak and Yacob Ben-Aryeh, “Quantum fast trary discrete unitary transformations using spatial and Fourier transform and quantum computation by linear internal modes of light,” Physical Review A 92, 043813 optics,” Journal of the Optical Society of America B 24, (2015). 231 (2007). [49] Hubert de Guise, Olivia Di Matteo, and Luis L. Sánchez- [33] James W. Cooley and John W. Tukey, “An algorithm Soto, “Simple factorization of unitary transformations,” for the machine calculation of complex Fourier series,” Physical Review A 97, 022328 (2018). 14

[50] Shreya P. Kumar and Ish Dhand, “Optimal modu- lar architectures for universal linear optics,” (2020), arXiv:2001.02012.