Simulation of integrated photonic gates

Andrei-Emanuel Dragomir, Cristian George Ivan, and Radu Ionicioiu

Horia Hulubei National Institute of Physics and Nuclear Engineering, 077125 Bucharest–Măgurele, Romania

Quantum technologies, such as quantum communication, sensing and imaging, need a platform which is flex-
ible, miniaturizable and works at room temperature. Integrated photonics is a promising and fast-developing
platform. This requires to develop the right tools to design and fabricate arbitrary photonic quantum devices.
Here we present an algorithm which, starting from a n-mode transformation U , designs a photonic device im-
plementing U . Using this method we design integrated photonic devices which implement quantum gates with
high fidelity. Apart from quantum computation, future applications include the design of photonic subroutines
or embedded quantum devices. These custom-designed photonic devices will implement in a single step a given
algorithm and will be small, robust and fast compared to a fully-programmable processor.

I. INTRODUCTION mode unitary transformation has been proposed recently [9–

11]. The new decomposition uses alternating layers of phase-
Quantum information brings a paradigm shift of how we shifts Pk (the variable elements) and mixing transformations
represent, store, process and read information, with huge im- Vk (the fixed elements)
arXiv:2006.02946v2 [quant-ph] 7 Jun 2020

pact on future technology. Successful applications of quantum

U = P1 V1 · · · PD VD PD+1 (1)
technologies include quantum communication/cryptography,
quantum sensing, quantum simulation and quantum imaging.
The mixing transformations Vk are acting globally an all n
However, the ultimate goal is to design and build a quan-
modes, in contrast to a local beamsplitter which and acts only
tum computer, i.e., a device which can implement any unitary
on two neighbouring modes. The new decomposition (1) is
transformation U on an arbitrary quantum state.
more robust to implementations errors, as the authors showed
One of the approaches towards this objective is quantum in Ref. [9].
optics, where photons are the main carriers of quantum in- In this context one problem emerges, namely how to de-
formation. Quantum algorithms are a set of transformations sign a photonic circuit which implements a given multi-mode
(gates) applied to qubits, the primary units of quantum infor- unitary, like the mixing gate Vk in (1). Here we propose an al-
mation. In quantum optics, these gates are optical elements, gorithm which addresses this problem. Given a n-port unitary
such as beam-splitters, phase shifts, polarising beam-splitters transformation U , our algorithm designs an integrated pho-
etc. Most of the research in this field is done with bulk optics, tonic circuit which implements the transformation U .
i.e., macroscopic elements on an optical table (lenses, beam- The structure of the article is the following. In Section II
splitters, wave-plates etc). However, the large size of these we describe the simulation algorithm, we discuss the main
components prevents miniaturisation and scalability towards ansatz and the optimisation strategy. In section III we present
more complex quantum algorithms. simulations for different quantum gates: Hadamard H, 4-
A solution to this problem is integrated quantum photonics dimensional Fourier transform F4 and 2D random unitaries,
[1]. The goal is to implement on-chip every part of a quan- together with error analysis. Finally, we conclude in Section
tum circuit: qubit generation [2], transformation and detec- IV.
tion. Bulk optics quantum gates are replaced by on-chip quan-
tum gates. So far basic integrated quantum photonic gates
have been successfully designed and fabricated [3, 4]. Nev- II. FROM QUANTUM GATES TO PHOTONIC DEVICES
ertheless, as quantum devices increase in size, the fabrication
errors become a problem and have to be kept under control.
The problem we address here is the following. Given a uni-
A unitary transformation U is usually decomposed in terms tary transformation U ∈ U(n) acting on n spatial modes
of simpler gates. These elementary gates are acting either on a
single optical mode (phase-shifts Pϕ ) or on two optical modes |outi = U |ini (2)
(beamsplitters) [5, 6]. To date, almost all optical experiments
use this decomposition in terms of beamsplitters and variable our goal is to design a n-mode photonic device implementing
phase-shifts [7]. U , see Fig. 1. In our case the n spatial modes are waveg-
This decomposition is convenient since beamsplitters and uides attached to the input (output) of the device. In terms
phase-shifts can be straightforwardly implemented both in of quantum information, the device implements a transforma-
bulk optics and in integrated photonics. For example, chip- tion U over a n-dimensional qudit space. We use the path
integrated photonics implement beamsplitters as multi-mode (spatial mode) representation for qudits, i.e., the basis state |ii
interference devices and phase-shifts as heating metallic pads is represented by a (single-photon) wave-function in the i-th
[8]. waveguide of the device, 0 ≤ i ≤ n − 1.
However, this decomposition is not robust against perturba- In this section we discuss several necessary ingredients: (i)
tions, i.e., fabrication errors in beamsplitters and phase-shifts. representation of the initial state; (ii) optimisation algorithm;
To address this problem, a different decomposition of a n- (iii) fidelity measure.
 2

|ini U |outi

FIG. 2: The structure of the simulation algorithm. The algorithm

FIG. 1: A photonic device implementing the unitary transformation applies iteratively to the silicon chip the following two steps: (1)
U on n optical modes (black input/output lines): |outi = U |ini. optimisation, and (2) fine graining; black (grey) pixels are silicon
(air).

Initial state. Since we use FDTD, which is a classical algo-

rithm, to solve a quantum problem, we need to know how to vice implementing the identity matrix I. We generate the tar-
make the transition between classical and quantum descrip- get output fields and record the discrete electric and magnetic
tions. We consider the following ansatz for the electric field fields in a matrix (the ’target’ output). To compare the target
E of the photon wave-function travelling in the waveguides: and test output we use the electromagnetic description of the
complex wave function, see Ref. [16]:
1 1 t−µ 2 "→
−#
E(t) = A √ e− 2 ( σ ) sin(ωt + φ) (3) 1 E
σ 2π |Ψi = √ →
− (4)
2 iB
i.e., a Gaussian-modulated sine-function and we consider this
as the classical description of a quantum wave-function of We define the fidelity (overlap) between the target |Ψ0 i and
phase φ. This is the transition from quantum to classical. test output |Ψi as:
Since the final classical states are mixed, we use an indi-
rect approach to make the transition back to the quantum de- 1
F= |hΨ0 |Ψi| (5)
scription. We calculate the overlap between the target and the N
simulated fields and we interpret this figure-of-merit as the p
quantum fidelity, as explained in the ’Fidelity’ subsection be- and 0 ≤ F ≤ 1 with the norm N = hΨ0 |Ψ0 ihΨ|Ψi; the
low. fidelity is clearly symmetric in Ψ0 , Ψ. From eq. (4) we have:
Optimisation. The algorithm has several parts. First, we need
1 −→ → − − → → −
to generate a photonic structure representing the device. Our hΨ0 |Ψi = (E0 · E + B0 · B ) (6)
device consists of blocks (called pixels) which are either solid 2
(silicon) or empty (air), see Fig. 2. Second, given a device We use transverse electric (TE) FDTD simulation, meaning
structure, we propagate the input state through the device to one electric component on the z direction and two magnetic
find out the output state. Finally, we need to optimise the components on x and y. Taking in account that the fields are
structure such that the device approximates as closely as pos- discrete, the fidelity F becomes:
sible the transformation U .  
Our algorithm is based on iterative designs of classical op-  
1  X z z x x z z

tical devices [12] and direct binary search (DBS) [13, 14]. We F= E0  E + B0  B + B0  B  (7)
start with a solid block of silicon connected by n input and N 
grid pts 
n output waveguides. We divide the active area of the device
into smaller blocks (pixels) which can be either on (filled with where  is the Hadamard product and the sum is performed
silicon) or off (filled with air). We alternate between optimis- over all FDTD grid points. We interpret this as the quantum
ing the structure and fine-graining until we reach the lowest fidelity between the two states.
pixel size which is still technologically feasible (Fig.2). The
optimisation algorithm is presented in Annex A.
Fine-graining is an essential step of our algorithm. Ini- III. RESULTS
tially we had developed only the optimisation part without
fine-graining, by starting with the maximum number of pix- In this section we present simulations for a single-qubit ap-
els (e.g., 64). However, we soon discovered that the algorithm plication, the Hadamard gate H, and a single-qudit applica-
was prone to getting stuck in a local minimum, far from an tion, the 4-dimensional Fourier gate F4 . For simulations we
optimal value. use λ = 650 nm and r = 11.7 (silicon).
We use a 2D finite-difference time-domain (FDTD) algo- Importantly, FDTD simulations are scale invariant, due to
rithm [15] and propagate the input state through the device the scale invariance of Maxwell equations. This implies that
in order to obtain the output state. We record the values of we can arbitrarily choose the scale a of the system. By taking
the FDTD discrete electric and magnetic fields over the entire the speed of light c = 1, then a (or a/c) defines the unit of
simulation in a matrix (the ’test’ matrix). time. Consequently, the frequency is expressed in units of
Fidelity. Next we need to evaluate how close is the resulting 1/a. We have chosen the scale a = 1 µm corresponding to the
quantum state compared to the desired (target) output state. smallest possible pixel being 125×125 nm, with λ = 650 nm.
We perform a separate simulation only with waveguides con- Clearly, every configuration (device) can be easily scaled for
necting the input and the output; this is equivalent to a de- different λ.
 3

FIG. 3: Propagation of a wave-function through the Hadamard gate.

Each figure is a superposition of three snapshots: the initial state, the
state during propagation inside the gate and the output state. Left:
simulation of the |0i state. Right: simulation of the |1i state. Notice
the π-phase difference in the output of the |1i state.

FIG. 5: Simulation of the 4-dimensional Fourier transform F4 . As

before, each figure is a superposition of three snapshots: the initial
state, the state during propagation inside the gate and the output state.
The four panels correspond to the four basis states |0i, |1i, |2i and
|3i.

FIG. 4: Fidelity F as a function of iteration steps for Hadamard H

gate; Fmax = 0.887.

A. Qubit

The Hadamard gate is defined as: FIG. 6: Fidelity F as a function of iteration steps for the Fourier gate
F4 ; Fmax = 0.774.
 
1 1 1
H=√ (8)
2 1 −1 B. Qudit

For our qudit application we chose a 4-dimensional Fourier

The two spatial modes correspond to the qubit basis states: gate F4 . The d-dimensional Fourier transform Fd is [17]:
|0i (top waveguide) and |1i (bottom waveguide).
d−1
For the Hadamard gate, the electric field E in both output 1 X 2πi kj
Fd |ki = √ e d |ji (9)
waveguides should have the same amplitude, but with differ- d j=0
ent phases depending on the initial state: ϕ = 0 for the input
state |0i and ϕ = π for the input state |1i. We clearly see the The Fourier Fd generalises the Hadamard gate H for d > 2.
phase difference in the output electric field for the input state As before, each waveguide corresponds to a basis state (top
|1i, Fig.3. to bottom): |0i, |1i, |2i and |3i. Our design for the Fourier
The fidelity increases with the number of iteration steps. gate (Fig. 5) is compact, compared to previous implementa-
Not surprisingly, the largest increase in fidelity occurs imme- tions which involved a large number of optical elements in a
diately after fine-graining, see Fig. 4. complex design [18].
 4

FIG. 7: Fidelity as a function of input wavelength. The vertical line FIG. 8: Fidelity as a function of pixel displacement.
corresponds to 650 nm, the wavelength used during optimisation.

IV. DISCUSSION
C. Random unitaries
Integrated photonics is one of the most promising platforms
for future quantum technologies. In order to fully take ad-
To show that our algorithm works well for arbitrary gates,
vantage of this platform, we need flexible tools to design and
we generate structures for random unitaries and compute their
simulate chip-integrated quantum gates.
fidelity. We simulated 2×2 unitaries and optimised up to
32×32 pixels, due to our limited computational power. The Here we presented an algorithm for designing integrated
unitary matrices are drawn from a uniform U(2) distribution silicon devices performing arbitrary quantum gates U on n
[19]: spatial modes. Starting from a uniform block of silicon, the al-
gorithm alternates optimisation and fine-graining steps in or-
  der to reach a photonic structure implementing U . We have
eiψ cos φ eiχ sin φ achieved fidelities up to 0.887 and we expect to surpass 0.9 by
U (α, φ, ψ, χ) := eiα −iχ (10)
−e sin φ e−iψ cos φ quadrupling the number of final pixels to 128×128, either by
enlarging the chip or by using smaller pixels.
with the sampling φ ∈ [0, π2 ] and α, ψ, χ ∈ [0, 2π]. So far the algorithm is intrinsically 2D due to our limited
computing power. In the future we plan to develop a fully
We obtain an average fidelity F = 0.881 ± 0.025. This is
3D implementation. This will allow us to design and simulate
similar to the fidelity for the Hadamard gate, showing that our
devices controlling other photonic degrees of freedom, like
algorithm generates consistent results for arbitrary gates.
the orbital angular momentum (OAM). Consequently, it will
be feasible to develop integrated spiral phase-plates and mode
converters.
Our algorithm can be used to design photonic subroutines,
D. Error analysis i.e., sets of quantum gates which are repeatedly used dur-
ing the execution of a program. An example is the Fourier
transform F2n on 2n modes. Usually F2n is decomposed in
Our goal is to design quantum gates which will be experi-
n(log2 n+1) beamsplitters and n(log2 n−1)+1 phase-shifts
mentally implemented. Thus it is important to know how fi-
and has optical depth d = n log2 n + 1 [20]. Thus it is more
delity varies in practice with different sources of errors.
efficient to have a specialised photonic circuit which performs
First, we are interested in analysing the effects of variable F2n in one step. This corresponds to optical depth 1, com-
photon wavelength. Not surprisingly, fidelity is robust for pared to the optical depth d for the standard decomposition in
wavelengths λ > λ0 larger than the optimised value λ0 , but terms of beamsplitters and phase-shifts.
decreases rapidly for shorter ones, see Fig. 7. Another future application are dedicated quantum devices,
The second source of errors is manufacturing imprecision. similar to classical embedded systems. Examples are quan-
To study this, we randomly shift each pixel relative to its orig- tum communication, quantum sensing and quantum imaging
inal position. Significantly, fidelity is almost constant for dis- devices, where full programmability is not required. In this
placement errors below 5 nm, then decreases almost linearly scenario embedded quantum systems need to execute a partic-
for larger values, Fig. 8. Thus if the fabrication errors are be- ular task fast and reliable without being fully programmable.
low 5 nm, the device will have a fidelity close to the simulated Thus, a custom-designed photonic device which implements
one. in a single step a given unitary will be small, robust and fast
 5

compared to a fully programmable processor. The optimisation algorithm starts by making a randomly or-
dered list of pixels. Then it goes through each pixel and flips
its state. If the new fidelity F is higher, it keeps the pixel
Acknowledgments flipped. After testing every pixel, it compares the improve-
ment of F across all steps. If this improvement is non-zero, it
The authors acknowledge support from a grant of the runs the DBS algorithm again. After the configuration cannot
Romanian Ministry of Research and Innovation, PCCDI- be improved anymore, each pixel is subdivided into 4 squares
UEFISCDI, project number PN-III-P1-1.2-PCCDI-2017- of the same type, which will be the new pixels, such that if
0338/79PCCDI/2018, within PNCDI III. R.I. acknowledges the initial pixel was on (off) the new pixels are on (off). We
support from PN 19060101/2019-2022. keep optimising and fine graining until we reach a certain size
threshold, given by fabrication constraints.

Appendix A: Optimisation algorithm

A detailed flowchart of the optimisation algorithm is shown

in Fig. 9.

[1] F. Flamini, N. Spagnolo, F. Sciarrino, Photonic quantum in- taries, Phys. Rev. Lett. 124, 010501 (2020).
formation processing: a review, Rep.Prog.Phys. 82, 016001 [10] V.J. Lopez-Pastor, J.S. Lundeen, and F. Marquardt, Arbitrary
(2019). optical wave evolution with Fourier transforms and phase
[2] J.W. Silverstone, D. Bonneau, M.J. Strain, M. Sorel, masks, arXiv:1912.04721 (2019).
J.L. O’Brien, and M.G. Thompson, Silicon quantum photonic [11] L. Pereira, A. Rojas, G. Canas, G. Lima, A. Delgado, and
circuits for on-chip qubit generation, manipulation and logic A. Cabello, Universal multi-port interferometers with minimal
operations, in Frontiers in Optics 2013, I. Kang, D. Reitze, N. optical depth, arXiv:2002.01371 (2020).
Alic, and D. Hagan, eds., OSA Technical Digest (online) (Op- [12] A.D. Verhoeven, F. Wyronski, J. Turunen, Iterative design of
tical Society of America, 2013), paper FW4C.5. diffractive elements made of lossy materials, JOSA 35, 45
[3] E. Heemskerk and B.I. Akca, On-chip polarization beam split- (2018).
ter design for optical coherence tomography, Opt. Express 26, [13] Z. Manna and R. Waldinger, The origin of a binary-search al-
33349 (2018). gorithm, Science of Computer Programming 9, 37-83 (1987).
[4] M. Burla, L.R. Cortes, M. Li, X. Wang, L. Chrostowski, and [14] B. Shen, P. Wang, R. Polson and R. Menon, An integrated-
J. Azana, On-chip ultra-wideband microwave photonic phase nanophotonics polarization beamsplitter with 2.4 × 2.4µm2
shifter and true time delay line based on a single phase- footprint, Nature Photonics 9, 378 (2015).
shifted waveguide Bragg grating, in Proc. IEEE Microw. Pho- [15] K. Yee, Numerical solution of initial boundary value problems
ton. Conf., 2013, pp. 92–95. involving Maxwell’s equations in isotropic media, IEEE Trans-
[5] M. Reck, A. Zeilinger, H.J. Bernstein, and P. Bertani, Experi- actions on Antennas and Propagation 14, 302 (1966).
mental realization of any discrete unitary operator, Phys. Rev. [16] M.G. Raymer and B.J. Smith, The Maxwell wave function of
Lett. 73, 58 (1994). the photon, SPIE conference Optics and Photonics (San Diego,
[6] W.R. Clements, P.C. Humphreys, B.J. Metcalf, W.S. Koltham- Aug. 2005)
mer, and I.A. Walsmley, Optimal design for universal multiport [17] M.A. Nielsen and I.L. Chuang, Quantum Computation and
interferometers, Optica 3, 1460 (2016). Quantum Information (Cambridge University Press, 2010).
[7] P. Kok, W.J. Munro, K. Nemoto, T.C. Ralph, J.P. Dowling and [18] R. Barak and Y. Ben-Aryeh, Quantum fast Fourier transform
G.J. Milburn, Linear optical quantum computing with photonic and quantum computation by linear optics, J. Opt. Soc. Am.
qubits, Rev. Mod. Phys. 79, 135 (2007). B24, 231 (2007).
[8] N.C. Harris, Y. Ma, J. Mower, T. Baehr-Jones, D. Englund, [19] K. Zyczkowski, M. Kus, Random unitary matrices, J. Phys. A:
M. Hochberg, and C. Galland, Efficient, compact and low loss Math. Gen. 27, 4235 (1994).
thermo-optic phase shifter in silicon, Opt. Express 22, 10487 [20] G.N.M. Tabia, Recursive multiport schemes for implementing
(2014). quantum algorithms with photonic integrated circuits, Phys.
[9] M. Saygin, I. Kondratyev, I. Dyakonov, S. Mironov, S. Straupe, Rev. A 93, 012323 (2016).
S. Kulik, Robust architecture for programmable universal uni-
 6

FIG. 9: Optimisation flowchart.