betaFeedback

Reading v1View latest version

WorksPhysics and Astronomy

Multiport photonic routing with a single cavity-based electro-optic phase shifter per port

Abstract

Programmable photonic meshes can perform general linear transformations, but universal transformations on NN modes require O(N2)O(N^2) programmable phase shifters, which cannot simultaneously achieve low drive voltage, gigahertz speed, zero static power, lossless modulation, and a compact footprint. Signal routing does not require universal transformations, which allows considerable circuit simplifications. We present a compact reflective circuit architecture that connects all input-output pairs across NN modes in NN steps using only O(N)O(N) programmable phase shifters. The circuit comprises a single passive N×NN\times N interferometer and a single column of programmable reflective phase shifters. We demonstrate the architecture by implementing a 3×33\times 3 router in x-cut thin-film lithium niobate with a passive multimode interferometer and phase shifters based on single-sided photonic-crystal nanobeam cavities. Our programmable devices exhibit a 1.921.92 GHz/V resonance tuning efficiency via the Pockels effect, a footprint of 100 μm2100\,μ\mathrm{m}^2, sub-5-V drive voltage, and a 6.2 GHz electro-optic bandwidth. Through voltage-space characterization and physical modeling, we identify routing states with extinction ratios exceeding 11 dB at each output and demonstrate sequential routing at an equipment-limited speed of 10 MHz. Our results establish cavity-enhanced Pockels switching as a routing layer for classical and quantum computing interconnects across visible and near-infrared wavelengths.

The growing demand for computation is driving a shift toward increasingly distributed architectures, from AI accelerator clusters to modular quantum processors. In these architectures, communication between nodes, whereas distributed quantum processors require coherent photonic links between remote quantum registers, with platform-specific transduction at the network endpoints [4, 5, 6]. Reconfigurable optical switches address both requirements by dynamically allocating paths within the photonic network without optical-to-electronic conversion (Figure 1a).

Image not extracted

Figure 1. Multiport photonic routing architecture and electro-optic cavity phase-shifter principle. a, Network-level role of photonic switches in connecting high-performance computing systems, quantum networks, and distributed quantum processors. b, Crossbar switch with an active element at each input–output crossing, requiring O(N2)O(N^{2}) active elements. c, Universal Mach–Zehnder-interferometer mesh implementing arbitrary linear transformations with N(N−1)/2N(N-1)/2 tunable interferometers, scaling as O(N2)O(N^{2}). d, Pairwise-routing architecture demonstrated in this work: a passive N×NN \times N multimode interferometer (MMI) terminated by NN independently controlled reflective phase shifters. e, Illustrative reflective phase-shifter design comprising a tapered photonic-crystal nanobeam cavity and an integrated metal electrode for voltage control of the reflected optical phase. The design shown differs from the cavities integrated in the 3×33 \times3 switch. f, Cross-section of the electro-optic cavity region showing the LiNbO3_3 waveguide and adjacent gold electrode on an SiO2_2 substrate. g, Calculated reflection amplitude (black) and phase (gray) versus wavelength detuning, with phase referenced to its resonant value. The linewidth uses the spectroscopic QL=5.07×104Q_{\mathrm{L}} = 5.07 \times10^{4}; κi/κe=0.737\kappa_{\mathrm{i}/\kappa_{\mathrm{e}}} = 0.737 is obtained from the reciprocal transfer-matrix fit described in Supplementary Note 4. h, Reachable output-power region of the ideal three-port architecture under phase-only control, shown on the (P1,P2,P3)(P_{1}, P_{2}, P_{3}) simplex; color indicates the relative power in each output port.

AI systems require high-throughput optical links between accelerator

Integrated photonic switches offer a chip-scale, low-latency, energy-efficient path to reconfigurable optical routing [7]. High speed, low optical loss, and wide wavelength coverage across the visible and near-infrared spectrum make thin-film lithium niobate (TFLN) an ideal material platform for quantum technology, optical interconnects, and biomedical devices [8, 9]. However, the scale and control of integrated switches are fundamentally limited by the size and number of programmable components required [3]. Crossbar arrays require N2N^2 active crossings, while universal Mach–Zehnder interferometer (MZI) meshes implementing arbitrary N×NN \times N linear transformations require N(N−1)/2N(N - 1)/2 tunable interferometers, corresponding, together with the output-phase layer, to O(N2)O(N^2) programmable phase controls, all scaling quadratically with port count (Figure 1b,Figure 1c) [10, 11]. Architectures for universal transformations based on layers of discrete Fourier transforms (DFTs) and programmable phase masks similarly require O(N2)O(N^2) tunable elements [12]. Restricting the target operation from an arbitrary linear transformation to routing reduces the required programmability [13, 14]. Networks that realize arbitrary input–output permutations, such as Beneš networks, require O(Nlog⁡N)O(N \log N) [15, 16] tunable elements.

In this work, we present a reflective circuit architecture capable of pairwise routing among NN ports using only NN programmable phase shifters. For N=2kN = 2^k (kk an integer), each phase shifter switches between two states separated by π\pi, and all N(N−1)/2N(N - 1)/2 distinct input–output pairs can be addressed in the minimum possible N−1N - 1 configurations. The architecture consists of a single passive N×NN \times N interferometer and one reflective phase shifter per port (Figure 1d) [17]. We demonstrate the architecture with a 3×33 \times3 router using compact photonic-crystal (PhC) cavity-based phase shifters operating at telecom wavelengths. Conventional non-resonant TFLN phase shifters require millimeter-scale interaction lengths [18, 19], whereas resonant phase shifters convert a small electro-optic resonance shift into a large reflected-phase change over a capacitive interaction length of only tens of micrometers. This enables low drive voltage, low switching energy, and static power consumption limited by leakage currents [20, 21]. Our implementation uses a multimode interferometer (MMI) as the passive mixing element, with a multimode section measuring (450×15)(450 \times15) µm. The single-sided PhC cavity-based phase shifters measure (100×1)(100 \times1) µm and simultaneously combine a 6.2-GHz electro-optic bandwidth with sub-5-V drive and negligible static power dissipation. We demonstrate sequential routing between three input–output pairs using 20-ns state slots, with extinction ratios above 11 dB across all three output ports. Automated characterization of 60 complete devices, comprising 180 one-sided switch cavities and 60 two-sided reference cavities, shows loaded quality factors reaching 7.1×1047.1 \times10^4, close resonance matching, and tuning efficiencies up to 1.92 GHz V−1^{-1}.

Results

Architecture and operating principle

Light entering any input port of the MMI (Figure 1d) is split equally among the reflector arms; each branch acquires a voltage-controlled phase shift at its reflector, and the reflected fields are mixed again on the return pass through the MMI.

The reflective phase shifters consist of single-sided PhC nanobeam cavities (Figure 1e), flanked by gold electrodes (Figure 1f). The reflection coefficient of each cavity follows from coupled-mode theory [22],

r(Δ)=iΔ−(κe−κi)/2iΔ+(κe+κi)/2(1)r(\Delta) = \frac{i\Delta- (\kappa_{\mathrm{e}} - \kappa_{\mathrm{i}})/2}{i\Delta+ (\kappa_{\mathrm{e}} + \kappa_{\mathrm{i}})/2} \tag*{(1)}

where Δ=ωres(V)−ω\Delta= \omega_{\mathrm{res}}(V) - \omega is the detuning between the cavity resonance and the probe field, κe\kappa_{\mathrm{e}} is the external coupling rate to the waveguide, κi\kappa_{\mathrm{i}} the intrinsic loss rate, and ωres(V)=ω0+2πηV\omega_{\mathrm{res}}(V) = \omega_0 + 2\pi\eta V shifts linearly with applied voltage VV through the Pockels effect with a signed tuning coefficient η\eta. Quoted tuning efficiencies denote its magnitude. In the ideal overcoupled limit (κi/κe→0\kappa_{\mathrm{i}}/\kappa_{\mathrm{e}} \to0), ∣r∣→1|r| \to1 while arg⁡(r)\arg(r) winds through a full 2π2\pi across resonance, controlled by the voltage VV. Figure 1g illustrates the resonant amplitude and phase response with κi/κe=0.737\kappa_{\mathrm{i}}/\kappa_{\mathrm{e}} = 0.737, showcasing a large phase swing across resonance. We thus realize a low-loss, voltage-tunable phase shifter with static power consumption limited only by leakage currents.

To describe the complete switch, we write the reflection coefficient of each cavity as rj=∣rj∣exp⁡(iϕj)r_j = |r_j| \exp(i\phi_j). Writing M\mathbf{M} for the unitary forward transfer matrix of the MMI, reciprocity makes the return pass MT\mathbf{M}^{\mathrm{T}}, giving the round-trip transfer matrix

T(ω,V)=MTdiag⁡[r1,r2,r3]M,(2)\mathbf{T}(\omega,V) = \mathbf{M}^{\mathrm{T}} \operatorname{diag}[r_1,r_2,r_3]\mathbf{M}, \tag*{(2)}

where diag⁡[x,y,z]\operatorname{diag}[x,y,z] represents a diagonal matrix with the arguments as the ordered non-zero entries. When only port mm is illuminated, the power routed to output port kk is Pk=∣Tkm∣2Pin,mP_k = |T_{km}|^2P_{\mathrm{in},m}. Independently tuning (ϕ1,ϕ2,ϕ3)(\phi_1,\phi_2,\phi_3) through the control voltages (V1,V2,V3)(V_1,V_2,V_3) changes the interference conditions at the output ports. The resulting output-power distributions define the reachable region within the (P1,P2,P3)(P_1,P_2,P_3) simplex (Figure 1h). For four ports, equal-split MMIs can have different reachable power regions even though each still permits perfect pairwise routing (Supplementary Note 7).

Routing configurations

In Supplementary Notes 1 and 2, we provide several proofs that explain the possibilities and limitations of the circuit architecture in Figure 1d. First, we show that any single pairwise connection among NN input–output ports is reachable if the MMI implements a DFT and the reflections are of the form rj=exp⁡(i2πjl/N)r_j = \exp(i2\pi jl/N), where the integer ll controls which ports are connected. Next, we generalize the result and show that any lossless, equal-split MMI with ∣Mnm∣=1/N|M_{nm}| = 1/\sqrt{N} suffices for single-pair routing. Finally, for Walsh matrices of size N=2kN = 2^k, all N(N−1)/2N(N-1)/2 distinct port pairs can be covered in N−1N-1 configurations using only binary reflection phases rj=±1r_j = \pm1. This means that no port is ever idle, and the phase shifters can be controlled with digital electronics. In each step, N/2N/2 pairings are active, but for N≥8N \ge8 there is insufficient freedom to choose among the (N−1)!!=(N−1)(N−3)⋯3⋅1(N-1)!! = (N-1)(N-3)\cdots3\cdot1 pairing configurations (where !!!! denotes the double factorial). This would require O(Nlog⁡N)O(N\log N) programmable elements, as in Beneš networks [15, 16]. Our reflective circuit connects all pairs in the minimum possible number of steps (N−1)(N-1), making it ideally suited for networks with a fixed, traffic-independent switching schedule [23].

Device and experimental setup

The switch is fabricated on x-cut TFLN and clad with SiO2\mathrm{SiO}_2 (Methods). The crystal orientation aligns the guided TE-like mode with the in-plane DC and radio-frequency control fields, maximizing the electro-optic response through the dominant Pockels coefficient r33r_{33} (Figure 1f).

Figure 2a shows a single-sided PhC nanobeam cavity divided into taper, front-mirror, cavity, and back-mirror sections. Light traverses a ∼15 μm\sim15\,\mu\mathrm{m} taper and the front mirror, which sets the external coupling rate κe\kappa_e, enters the central defect, and is retroreflected by the back mirror. The four sections contain NtapN_{\mathrm{tap}}, Nm,LN_{\mathrm{m,L}}, 2Ncav2N_{\mathrm{cav}}, and Nm,RN_{\mathrm{m,R}} periods with pitch a=418 nma = 418\,\mathrm{nm}. The cavity and mirror sections are 550 and 730 nm wide, respectively. Full design details are provided in Supplementary Note 3.

Image not extracted

Figure 2. Fabricated 3×33 \times3 photonic switch and experimental setup. a, Scanning electron micrographs of a dedicated cavity test structure after local removal of the SiO2_2 cladding, identifying the taper, cavity, and mirror sections and representative dimensions. The scale bar shown in the taper section applies to all three sections. b, Optical micrograph of the fabricated switch, showing the input/output grating couplers, central MMI, waveguide phase shifters, three tunable photonic-crystal cavities, and electrical contact region. The driven cavity electrodes V1V_1, V2V_2, and V3V_3, together with the intervening ground contacts, are identified. c, Experimental setup used to select the driven port, control the three cavity electrodes through bias tees, and simultaneously record the three reflected output powers. Black and light-gray lines denote optical and electrical connections, respectively.

Figure 2b shows the fabricated 3×33 \times3 switch. Uniform TE grating couplers (I/O1_1–I/O3_3) couple light to and from the three bidirectional ports of a single 3×33 \times3 MMI. Each MMI arm also carries a waveguide phase shifter for optional static phase control. These phase shifters were left unbiased in this work. Each arm then terminates in a single-sided PhC nanobeam cavity. Coplanar gold electrodes straddle the three cavities and are addressed by voltages (V1,V2,V3)(V_1,V_2,V_3) referenced to a shared ground in a G1V1G2V2G3GG_1V_1G_2V_2G_3G pad configuration. The electrodes are separated by a 6.8 μm\mu\mathrm{m} gap and cover 64 μm\mu\mathrm{m} of the cavity’s length so that the applied electric field overlaps the guided mode. Finite-element electrostatic simulations give a terminal capacitance C=18.6±0.3C = 18.6 \pm0.3 fF between each cavity drive electrode and ground (Supplementary Note 3). For a conventional non-energy-recovering driver, the resulting energy dissipated per 5-V transition is Etr=C(ΔV)2/2=232.5±3.8E_{\mathrm{tr}} = C(\Delta V)^2/2 = 232.5 \pm3.8 fJ.

Figure 2c shows the fiber-coupled routing setup. A tunable laser, circulator, and pair of MEMS fiber switches select the driven port and direct the three reflected outputs to detectors. Polarization paddles align the fiber array with the grating couplers’ TE mode. The outputs are recorded simultaneously on an oscilloscope. Each cavity electrode is driven through a bias tee combining a DC operating-point voltage with an arbitrary-waveform generator (AWG) signal bounded between −5-5 and +5+5 V at the high-impedance device. The AWG synchronization signal triggers the oscilloscope and aligns the drive waveforms with the optical responses. Instrument details are provided in Methods.

Cavity electro-optic characterization

Figure 3a shows the broadband reflection and transmission spectra of a two-sided reference cavity. The single-sided switch cavities use a partially transmitting front mirror with Nm,L=40N_{\mathrm{m,L}} = 40 periods and a highly reflecting back mirror with Nm,R=100N_{\mathrm{m,R}} = 100 periods. The two-sided reference cavity instead uses two symmetric 40-period mirrors, Nm,L=Nm,R=40N_{\mathrm{m,L}} = N_{\mathrm{m,R}} = 40. This symmetric geometry clearly reveals the PhC bandgap between approximately 1500 and 1560 nm and a high-QQ cavity resonance at λ0=1541.48\lambda_0 = 1541.48 nm. A joint fit to the reflection and transmission resonance gives a loaded quality factor Q=(3.30±0.02)×104Q = (3.30 \pm0.02) \times10^{4}, corresponding to an optical linewidth of Δνcav=ν0/Q=5.9\Delta\nu_{\mathrm{cav}} = \nu_0/Q = 5.9 GHz. The reference-cavity linewidth characterizes the fabricated PhC optical quality. The full-device model described below uses the one-sided switch-cavity parameters to determine the phase response relevant to routing.

Image not extracted

Figure 3. Cavity spectroscopy, electro-optic tuning and modulation bandwidth. a, Measured broadband transmission (black) and reflection (gray) spectra of a representative two-sided reference cavity, independently scaled for display; inset: the measured cavity resonance versus wavelength detuning. b, Measured resonance-frequency shift of a single-sided device cavity versus applied DC voltage at a −14-dBm laser-output setting (black points). The dashed gray line is a linear fit with an effective DC tuning rate of ηDC=1.92 GHz V−1\eta_{\mathrm{DC}} = 1.92\ \mathrm{GHz}\ \mathrm{V}^{-1}; a kink around zero voltage indicates a deviation from linearity. Inset: schematic of the electric field between the drive and ground electrodes. c, Measured normalized electro-optic response of a single-sided device cavity versus modulation frequency (black), with a third-order low-pass fit (dashed gray).

Figure 3b shows the voltage-dependent resonance shift of a single-sided device cavity. At a −14-dBm laser-output setting, a least-squares fit over the full voltage range gives an effective quasistatic tuning rate of ηDC=1.92±0.06\eta_{\mathrm{DC}} = 1.92 \pm0.06 GHz V−1^{-1}. A deviation from linearity is visible around zero voltage. At higher optical power, photo-induced space charge more strongly screens the applied field, making the DC-tuning kink more pronounced and reducing the effective ηDC\eta_{\mathrm{DC}}.

While ηDC\eta_{\mathrm{DC}} describes tuning under a slowly varying bias, we measure the normalized high-speed electro-optic response as a function of modulation frequency by applying an RF tone to a cavity electrode and recording the response with a vector network analyzer (VNA). The response stays approximately flat through the low-gigahertz range. A third-order empirical low-pass fit gives a 3-dB bandwidth of 6.2 GHz (Figure 3c; Supplementary Note 5).

We performed automated measurements of 60 complete devices, comprising 180 one-sided switch cavities and 60 two-sided reference cavities, spanning a range of designs. Independent fits to 44 reference-cavity spectra acquired without detector saturation and with R2≥0.9R^2 \ge0.9 give loaded quality factors reaching 7.1×1047.1 \times10^4. Across 50 complete devices for which a visual review identified two or three resolved switch-cavity resonances, we define the within-device resonance span Δλ\Delta\lambda as the wavelength separation between the outermost resolved resonances among the three nominally identical, closely spaced switch cavities. The median Δλ\Delta\lambda is 0.63 nm, with an interquartile range of 0.44–0.91 nm, and 82% have Δλ<1\Delta\lambda< 1 nm (Supplementary Note 3).

Experimental 3 × 3 routing

To map the switch across its accessible voltage space, we first align the three cavity resonances using static DC biases. We then explore the three-dimensional voltage space by driving the cavity electrodes (V1,V2,V3)(V_1, V_2, V_3) with orthogonal multisine waveforms, allowing the voltage cube to be sampled densely in a single acquisition. We map the output-power distributions accessible to the switch by recording P1P_1, P2P_2, and P3P_3 as the drive traverses the cube (Figure 4a).

Image not extracted

Figure 4. Voltage-space characterization and model-guided time-domain routing. a, Measured output-power distributions across the (V1,V2,V3)(V_1, V_2, V_3) voltage cube sampled by the multisine drive. The RGB value of each point encodes the relative output powers, with red, green, and blue components corresponding to P1P_1, P2P_2, and P3P_3, respectively. b, Output-power distributions predicted by the fitted transfer-matrix model on a regular voltage grid, using the same display scheme and RGB encoding as in a. c, Extinction ratio computed from the measured, channel-normalized output signals across the voltage cube; color identifies the target output port, and brighter regions indicate higher routing contrast. d, Model-predicted extinction ratio on a slice through the voltage cube containing the three optimized routing points, with orthonormal in-plane voltage coordinates (e1,e2)(e_1, e_2), expressed in volts. Color identifies the target output port, and brighter regions indicate higher routing contrast. The gray overlay marks extrapolation beyond the measured voltage range; S1S_1, S2S_2, and S3S_3 mark routing from input port 1 to output ports 1, 2, and 3, respectively. The colors show the calibrated model prediction, whereas the extinction ratios annotated next to the three states are the corresponding experimental values from the dynamic measurement in e. e, Time-domain sequential routing. Top: programmed AWG voltage waveforms V1V_1 (solid), V2V_2 (dashed), and V3V_3 (dotted). Bottom: simultaneously recorded normalized output powers P1P_1 (red), P2P_2 (green) and P3P_3 (blue); shaded windows mark the stable routing intervals for each target output port.

We fit the measured point cloud using the transfer-matrix model of Eq. (2), with cavity parameters initialized from reference-cavity measurements. The fitted model reproduces the measured output distribution (Figure 4b) and provides a continuous mapping from electrode voltages to output powers, allowing optimal routing points to be identified numerically.

To identify the routing state for each target output kk, we define the worst-case extinction ratio

ERk(V)=10log⁡10[P~kmax⁡j≠k(P~j)],(3)\mathrm{ER}_k(\mathbf{V}) = 10\log_{10}\left[\frac{\widetilde{P}_k}{\max_{j\ne k}(\widetilde{P}_j)}\right], \tag*{(3)}

which compares the target-port signal with the strongest off-target signal. Here, P~k\widetilde{P}_k denotes the independently normalized output power (Methods). Figure 4c plots the measured extinction ratio across the voltage cube. Maximizing ERk\mathrm{ER}_k in the fitted model identifies the optimal voltage tuple for each routing state; Figure 4d shows a model-predicted slice through the three optimized points. The corresponding model-reconstructed cavity spectra are discussed in Supplementary Note 4.

Applying the optimized voltage tuples sequentially demonstrates time-domain routing (Figure 4e). We use a 100-ns waveform divided into five 20-ns slots, corresponding to a 10-MHz repetition rate and a 50-MHz slot rate. Three slots apply the voltage tuples for routing states S1S_1, S2S_2, and S3S_3, which route light from input port 1 to output ports 1, 2, and 3, respectively. Because the bias tees block DC and frequencies below 10 MHz, each AWG waveform must have zero average over a complete cycle. We therefore add two compensation slots whose voltages cancel the average of the three routing slots. Within each slot, the voltage is held constant for 6 ns and changes during a 14-ns

The demonstrated timing was set by the experimental drive and acquisition setup, including the 10-MHz bias-tee lower cutoff and 200-MHz oscilloscope bandwidth, rather than by the intrinsic device response. The 14-ns transitions were chosen conservatively to preserve waveform fidelity and suppress electrical ringing. During the three routing slots, the switch directs light to output ports 1, 2, and 3 with maximum instantaneous, normalized-signal extinction ratios of 13.7, 12.3, and 11.3 dB, respectively, relative to the strongest off-target port, using AWG excursions below 5 V about the static cavity biases. The variation among output ports is consistent with differences in cavity coupling across the three arms.

An independent power-budget measurement at the 1541.66-nm switching wavelength gives a calibration-corrected summed-output excess loss of 3.65 dB relative to a same-session grating-reference path. For routing state S1S_1, the calibrated model attributes approximately 2.33 dB to cavity dissipation. The estimated remaining excess of approximately 1.3–1.6 dB is comparable to that measured over the off-resonant scan interval (Supplementary Note 4). This can include two MMI passages, routing-waveguide loss, and residual path mismatch, so it cannot be assigned to the MMI alone.

Discussion

The key performance metrics of our compact router architecture are insertion loss, extinction ratio, speed, control voltage, and dynamic energy consumption. Below, we discuss ways to improve these metrics in future demonstrations.

For the circuit architecture requiring only two phase settings, we can choose an operating point so the two voltages correspond to detunings ±Δπ=±12κe2−κi2\pm\Delta_{\pi} = \pm\frac{1}{2}\sqrt{\kappa_{\mathrm{e}}^{2} - \kappa_{\mathrm{i}}^{2}}. From Eq. (1), the corresponding phase difference is arg⁡{r(Δπ)}−arg⁡{r(−Δπ)}=π\arg\{r(\Delta_{\pi})\} - \arg\{r(-\Delta_{\pi})\} = \pi, and the reflected powers are ∣r(Δπ)∣2=∣r(−Δπ)∣2=κe−κiκe+κi|r(\Delta_{\pi})|^{2} = |r(-\Delta_{\pi})|^{2} = \frac{\kappa_{\mathrm{e}}-\kappa_{\mathrm{i}}}{\kappa_{\mathrm{e}}+\kappa_{\mathrm{i}}}. Such an operating point requires overcoupled cavities, κe>κi\kappa_{\mathrm{e}} > \kappa_{\mathrm{i}}. If all cavities have identical coupling rates, perfect extinction remains possible. An insertion loss of 10% requires κe/κi=19\kappa_{\mathrm{e}}/\kappa_{\mathrm{i}} = 19, and 1% corresponds to κe/κi=199\kappa_{\mathrm{e}}/\kappa_{\mathrm{i}} = 199.

For lossy and non-identical cavities, the extinction ratio cannot be perfect in general. Unequal arm lengths between the MMI and the cavities change the required operating points because the cavity phase and amplitude responses are coupled. Adding waveguide-based phase shifters in front of the cavities decouples these responses and enables operating-point optimization. In Supplementary Note 4, we investigate how device imperfections affect the achievable extinction. In an optical-only sensitivity model retaining the fitted cavity and MMI parameters but excluding empirical detector backgrounds and calibration, optimizing the two static relative operating-point phases increases the worst of the three separately optimized route extinctions from 16.0 to 24.0 dB and the minimum route-specific fraction of the voltage cube above 10 dB from 1.49% to 3.93%. Optimizing instead for local 10-dB voltage-error tolerance increases the worst-route half-width from approximately ±0.70\pm0.70 to ±1.13\pm1.13 V per electrode. Designs with matched arm optical path lengths reduce the required operating-point phase-shifter length. Alternatively, millimeter-scale waveguide-based phase shifters could implement both DC offsets and dynamic switching, enabling broadband operation with an interaction length comparable to that of a push–pull MZI.

Larger extinctions than the 11.3–13.7 dB shown in Figure 4e are possible for single input–output pairs. We achieved 20.8–22.6 dB with repeated returns to the same waveform phase, as described in Supplementary Note 6. The contrast available at a fixed operating point therefore exceeds that maintained throughout the full three-state sequence. At the highest-contrast point, the off-target signals approach the measurement repeatability floor, whereas noise and slow drift displace the device from its optimized operating points during extended sequences. Faster feedback and improved thermal stabilization should reduce this difference.

Photo-induced space charge screens the applied electric field, making the DC response dependent on optical power and reducing the effective tuning efficiency ηDC\eta_{\mathrm{DC}} at higher power. Related electro-optic relaxation and photorefractive resonance shifts are widely observed in TFLN modulators and resonators [24, 25]. This instability complicates long-term bias control but may be reduced through surface treatment, improved charge management, and air-clad device structures.

The speed, control voltage, and energy consumption are coupled in cavity-based phase shifters because their modulation bandwidth is limited by the linewidth, consistent with the gigahertz response in Figure 3c. The required switching voltage satisfies 2πηRFΔV=2Δπ=κe2−κi22\pi\eta_{\mathrm{RF}}\Delta V = 2\Delta_{\pi} = \sqrt{\kappa_{\mathrm{e}}^{2}-\kappa_{\mathrm{i}}^{2}}, where ηRF\eta_{\mathrm{RF}} is the resonance-frequency tuning efficiency at the modulation frequency. The switching energy, C(ΔV)2/2C(\Delta V)^{2}/2, is also coupled to the linewidth and speed. Improving the speed–energy trade-off requires increasing ηRF\eta_{\mathrm{RF}} and reducing the device capacitance CC. Future cavity designs should therefore be optimized for ηRF\eta_{\mathrm{RF}}, κi\kappa_{\mathrm{i}}, and size. In Supplementary Note 3, we show that our nanobeam PhC cavity design enables control of κe\kappa_{\mathrm{e}} without affecting κi\kappa_{\mathrm{i}}, allowing the linewidth to be adjusted to meet specified control-voltage and speed requirements.

Conclusion

We have theoretically shown that pairwise routing requires only O(N)O(N) programmable phase shifters in a compact reflective circuit architecture. Our experimental demonstration of a 3×33 \times3 router in TFLN illustrates the feasibility of the approach, providing a path toward dense, low-energy, electronically addressable routing layers for classical photonic interconnects. While our demonstration used TFLN, the reflective architecture could also be implemented on other platforms, including silicon, which benefits from mature foundry processes [26]. There, a ring-resonator phase shifter [27, 28] or a waveguide phase shifter could be placed before a passive Sagnac-loop mirror to provide the required controllable reflection phase.

Circuits described by permutation matrices also enable the routing of individual photons distributed among the input ports. The high speed, cryogenic compatibility [29, 30], and broad transparency [9, 31] make these circuits attractive for quantum photonics. They could enable cryogenic routing of photons between spin–photon interfaces in distributed matter-qubit modules [6, 32], control beams for trapped ions, neutral atoms, and solid-state emitters [33, 34, 35, 36], and provide essential building blocks in recently proposed recirculating quantum photonic networks [37].

Cavity-enhanced Pockels switching therefore provides a versatile path toward compact routing layers for both classical and quantum photonic interconnects.

Methods

Device fabrication

Devices were fabricated on CSEM’s x-cut TFLN platform, comprising a 600 nm LiNbO3_3 device layer on buried SiO2_2 and silicon. The 1550-nm waveguides and PhC cavities were patterned in the LiNbO3_3 device layer and etched to a depth of 400 nm, leaving a 200 nm slab. The routing waveguides had a nominal lithographic top width of 800 nm. The nominal exterior sidewall angle was 55° from the substrate plane (35° from vertical). For the 400-nm etch, this sidewall angle made the fabricated waveguide base 560 nm wider than the lithographic top width, giving a nominal base width of 1.36 µm for an 800-nm lithographic top width. Gold electrodes were formed using the 600-nm-thick platform metallization layer. The devices were clad with silicon dioxide. To expose the cavity for the SEM in Figure 2a, the silicon-dioxide cladding was locally removed from a dedicated test structure while the switch cavities remained oxide-clad.

Optical and electrical measurement setup

For optical measurements, the chip was mounted on a Thorlabs M30XY translation stage and coupled through the grating couplers using a fiber array mounted at an 8° angle on a Thorlabs NanoMax stage. The optical source was a Santec TSL-570 tunable laser with a tuning range of 1480–1640 nm. For the switching measurements in Figure 4, the laser output was set to 0 dBm. The laser light passed through a fiber-optic circulator and a fiber-switching network comprising two Agiltron 2 × 2 MEMS switches that selected the driven and detected ports. Polarization paddles aligned the launch to the TE mode before coupling to the chip. Three Thorlabs APD431C InGaAs avalanche photodetectors (40-MHz nominal bandwidth) recorded P1P_1, P2P_2, and P3P_3 simultaneously. The three cavity electrodes were contacted with a FormFactor Cascade Multi ∣Z∣|Z| RF probe (K20-MZN-9B01-100) and driven through Mini-Circuits ZX85-12G-S+ bias tees (10-MHz–12-GHz passband) that combined static DC biases from three Keithley 2400 source-measure units (100 µA current compliance) with AC waveforms from a three-channel arbitrary waveform generator (Zurich Instruments HDAWG8, 2.4 GSa/s, ±5\pm5 V high-impedance output span, or 10 V peak-to-peak). For the broadband insertion-loss budget, a plain grating-reference path and the device were measured on the same wavelength grid using mutually calibrated detector channels. The traces were median-filtered before comparison. Because the reference includes both grating couplers, external fiber components, polarization mismatch, and reference-waveguide propagation, the device values are reported as excess loss relative to this path rather than as independently calibrated component losses (Supplementary Note 4).

Cavity spectroscopy, DC electro-optic tuning, and device statistics

For cavity spectroscopy, the laser wavelength was swept with a nominal step of 1 pm over 1480–1640 nm, and the reflected power was normalized to the laser output. The spectrum of the two-sided reference cavity was fitted over a 0.30-nm window centered on the detected resonance. The reflection dip and transmission peak were fitted simultaneously with a shared Lorentzian center and linewidth, independent amplitudes, and independent linear backgrounds. This simultaneous Lorentzian fit extracted the resonance wavelength, linewidth, and loaded quality factor QtotalQ_{\mathrm{total}}. The reported uncertainty in QtotalQ_{\mathrm{total}} is the one-standard-error value propagated from the least-squares covariance. For the DC tuning measurement, the voltage was stepped sequentially, and a wavelength scan was acquired at each voltage using a −14-14-dBm laser output setting. At each voltage, the three detector spectra were fitted independently of the other voltage steps with a shared resonance center and linewidth and independent linear backgrounds and Lorentzian/dispersive amplitudes. The extracted resonance wavelengths were converted to frequency shifts and fitted by ordinary least-squares over all 15 points from −20-20 to +20+20 V, giving an effective quasi-static tuning rate ηDC=1.92±0.06\eta_{\mathrm{DC}} = 1.92 \pm0.06 GHz V−1^{-1}. The formal slope standard error is 0.044 GHz V−1^{-1}; grouping two to four neighboring voltage settings to allow correlated residuals gives descriptive standard errors up to 0.060 GHz V−1^{-1}, which sets the reported uncertainty. A deviation from linearity remains near zero voltage.

Device statistics were acquired with an automated measurement routine that measured all input–output port configurations for 60 complete devices, comprising 180 one-sided switch cavities and 60 two-sided reference cavities. The Thorlabs M30XY translation stage was used to optimize the fiber coupling before each wavelength sweep. After short translations, automatic alignment used three iterations of a climb-ladder search along xx followed by yy. After larger translations, a coarse raster scan first located the grating couplers, followed by the same climb-ladder procedure for fine alignment. For each complete device, the recorded sequence comprised a reference-cavity sweep, three MMI sweeps acquired under different programmed permutations of the MEMS fiber-switch matrix, and a final calibration configuration. In a representative uninterrupted block covering 27 device positions, successive device measurements began at a median interval of 6.5 min. The five recorded sweeps within each complete sequence spanned a median of 3.9 min; the remaining cycle time included translation to the next device and optimization of the fiber coupling.

Electro-optic bandwidth measurement

The electro-optic bandwidth of a single cavity was measured from its S21S_{21} response over 50 MHz–10 GHz using a four-port Rohde & Schwarz ZNA67 vector network analyzer connected to the cavity electrodes through one GSG pair of a GSGSGSG probe. VNA port 1 drove the contacted electrodes directly through the probe, without a bias tee or applied DC bias. The laser was tuned to a point of maximum absolute slope on the cavity reflection lineshape. The modulated optical output was detected by a Discovery Semiconductors DSC30S InGaAs photodetector (18-GHz nominal bandwidth), whose electrical output returned directly to VNA port 2. The response shown in Figure 3c was acquired with a nominal electrical VNA source power of 12 dBm and an optical power setting of 13 dBm at the laser output. Accounting for the measured 1.39 dB input-path loss, an estimated 5.4–6.1 dB input-grating loss, and ideal three-way MMI splitting gives an estimated optical power of 0.7–1.4 dBm in the addressed cavity arm, before uncalibrated MMI excess loss. A third-order low-pass model with a free vertical offset was fitted to the measured S21S_{21} magnitude, yielding a 3-dB bandwidth of 6.2 GHz. Additional cavity-channel measurements showed similar bandwidth responses. Supplementary Note 5 provides the fitting procedure and model comparison.

Multisine voltage-space acquisition and delay correction

To map the router response over its three-dimensional control space, we simultaneously varied the voltages applied to the three cavity electrodes while recording the optical power at all three output ports. The three voltage waveforms were designed to sample this space broadly while minimizing correlations between the drive channels. Each drive channel carried a 25 μ\mus-period waveform synthesized at 2.4 GSa/s from 1001 equally spaced tones between 10 and 50 MHz, with equal initial amplitudes and pseudorandom phases. The three waveforms were orthogonalized by QR decomposition so that their discrete pairwise inner products over one period satisfied ∑nVi[n]Vj[n]=0\sum_n V_i[n]V_j[n] = 0 for i≠ji \ne j. They were then scaled to equal peak range and rank-remapped to give each channel a uniform amplitude distribution over [−1,1][-1,1]. The nonlinear remapping and delay reconstruction introduced small residual correlations, and for the trajectories in Figure 4, the largest pairwise correlation magnitude was 0.018. The 10–50 MHz range refers to the initial multisine synthesis band before rank remapping. The resulting waveforms spanned ±5\pm5 V and were added to static cavity biases of (−17,+19,−17)(-17,+19,-17) V. With the laser set to 1541.66 nm, four waveform periods were acquired over 100 μ\mus at 1 GSa/s with 16 waveform averages.

Signal-propagation delays between each AWG drive channel and oscilloscope acquisition channel were calibrated using a common linearly swept chirp spanning 10–150 MHz with a raised-cosine taper. The chirp occupied 15 μ\mus of each 25 μ\mus period, and four periods were recorded together with a synchronous reference tap. For each drive–detector pair, the traces were filtered over the chirp band and cross-correlated, and the Hilbert-transform envelope was used to estimate the delay. The nine delays were jointly fitted as τij=din,i+dout,j\tau_{ij}=d_{\mathrm{in},i}+d_{\mathrm{out},j}, with one input delay fixed as the reference, and used to align the recorded traces. Synthetic tests recovered injected delays

within 0.2 ns, while repeated measurements gave a relative-delay standard deviation of 1.4 ns. For the color representation in Figure 4a, each detector trace was independently min–max normalized over the acquisition. For visualization, samples were grouped into a 15315^3 voltage grid. Each occupied voxel was assigned its samples’ mean color, and display points were distributed uniformly within occupied voxels. The extinction-ratio color maps in Figure 4c,Figure 4d evaluate the definition above using the min–max-normalized output power P~k\tilde{P}_k and a numerical offset ϵ=10−6\epsilon=10^{-6} added to each normalized power. For visualization, negative extinction ratios were set to zero.

Transfer-matrix model and parameter estimation

At the fixed probe frequency, the transfer-matrix model predicts the three output powers from the applied voltages, using the voltage-dependent cavity reflection coefficients rj(Vj)r_j(V_j) defined in Eq. (1). For each cavity, the numerical fit was parameterized by the zero-voltage resonance detuning Δν0=(ω0−ω)/(2π)\Delta\nu_0=(\omega_0-\omega)/(2\pi), external-to-internal coupling ratio, κe/κi\kappa_e/\kappa_i, electro-optic tuning efficiency, η\eta, and loaded quality factor, QL=ω0/(κe+κi)Q_L=\omega_0/(\kappa_e+\kappa_i). Here, ω\omega is the fixed probe angular frequency and ν0=ω0/(2π)\nu_0=\omega_0/(2\pi) is the zero-voltage resonance frequency. At the fixed probe frequency, however, the cavity response constrains the products QLΔν0/ν0Q_L\Delta\nu_0/\nu_0 and QLη/ν0Q_L\eta/\nu_0, rather than QLQ_L and η\eta independently. Equivalently, it determines the voltage required to tune through one loaded linewidth, ν0/(QLη)\nu_0/(Q_L\eta). The nonlinear cavity parameters were constrained to the following bounds: −60 to 60 pm for the resonance wavelength offset, 0.2 to 20 for the external-to-internal coupling ratio, 0.95 to 3.0 GHz/V for the tuning efficiency, and 2.5×1042.5\times10^4 to 6.0×1046.0\times10^4 for the loaded quality factor. This loaded-quality-factor range was used as a prior informed by full-device spectroscopy, including the reference-cavity measurements, rather than as an upper bound on quality factors observed across the device array. The MMI was modeled using a fixed phase convention based on the discrete Fourier transform (DFT) [14, 38]. Only two relative amplitude factors were fitted, each constrained to 0.85–1.15; thus, the fit did not use an unrestricted MMI matrix M. Reciprocal return propagation was evaluated with the transpose of the forward MMI matrix, giving the round-trip transformation MTRM\mathbf{M}^{\mathsf{T}}\mathbf{R}\mathbf{M}. The fit also included two relative arm phases constrained to [−π,π][-\pi,\pi], a common voltage-dependent phase correction constrained to ±0.15\pm0.15 rad/V, and a coherent background field at each detector with relative amplitude constrained to 0–0.35. A gain and offset for each detector were determined by linear least squares at every model evaluation.

For parameter estimation, the delay-aligned samples were binned into a 20×20×2020 \times20 \times20 voltage grid, and each detector signal was averaged within each occupied voxel. The residual from each voxel was weighted by the square root of its sample count. The nonlinear parameters were fitted by minimizing the weighted root-mean-square residual using bounded multistart optimization with a differential-evolution fallback and L-BFGS-B refinement. A weak dimensionless regularization term (overall scale 0.015) favored similar parameters among the three cavities and broad reference values of 1.7 GHz/V for tuning efficiency, 3.6×1043.6 \times10^{4} for loaded quality factor, and 5 for the mean external-to-internal coupling ratio. It also weakly favored MMI amplitude factors near unity, a small voltage-dependent phase correction, and small coherent background fields. The reciprocal fit achieved R2=0.85R^{2} = 0.85–0.96 across the three detector signals.

Routing-point optimization and time-domain measurement

Routing operating points were selected by maximizing ERk(V)ER_k(V) on the voltage grid used for model fitting. For the representative device used in Figure 3 and Figure 4 (device A46), the absolute electrode-voltage tuples (V1,V2,V3)(V_1,V_2,V_3) used at 1541.66 nm were

V(S1)=(−15.01,15.61,−20.07) V,\mathbf{V}(S_1) = (-15.01,15.61,-20.07)\ \mathrm{V},
V(S2)=(−20.08,21.44,−21.04) V,\mathbf{V}(S_2) = (-20.08,21.44,-21.04)\ \mathrm{V},
V(S3)=(−19.48,14.38,−13.75) V,\mathbf{V}(S_3) = (-19.48,14.38,-13.75)\ \mathrm{V},
Vbias=(−17,19,−17) V.\mathbf{V}_{\mathrm{bias}} = (-17,19,-17)\ \mathrm{V}.

The listed absolute voltages include the static bias shown above. The corresponding AWG excursions were

ΔV(S1)=(1.99,−3.39,−3.07) V,\Delta\mathbf{V}(S_1) = (1.99,-3.39,-3.07)\ \mathrm{V},
ΔV(S2)=(−3.08,2.44,−4.04) V,\Delta\mathbf{V}(S_2) = (-3.08,2.44,-4.04)\ \mathrm{V},
ΔV(S3)=(−2.48,−4.62,3.25) V.\Delta\mathbf{V}(S_3) = (-2.48,-4.62,3.25)\ \mathrm{V}.

The tuples are device- and wavelength-specific calibrated operating points rather than universal switching voltages. For time-domain routing, the AWG generated a 100-ns periodic waveform containing three routing slots and two compensation slots. Each 20-ns slot comprised a 6-ns constant-voltage interval followed by a 14-ns raised-cosine transition. Because the bias tees block DC and frequencies below 10 MHz, each channel’s waveform was made zero-mean over the complete cycle. Both compensation slots were therefore set to minus half the sum of the three routing-slot voltages in that channel.

The three detector signals and a synchronous reference were acquired on a Tektronix MSO2024B oscilloscope (200-MHz nominal bandwidth, 1 GSa/s) using the acquisition settings and delay correction described above. The demonstrated timing was set by these drive and acquisition constraints rather than by the intrinsic electro-optic response. For each target output, the reported extinction ratio was the maximum instantaneous value over a 25 μs analysis interval.

Theoretical analysis

Under the ideal assumptions of a lossless, equal-split MMI and unit-magnitude reflector coefficients with phases independently tunable over the full 2π2\pi range, every three-port coupler admits the factorization M=DoutPoutFPinDin\mathbf{M}=\mathbf{D}_{\mathrm{out}}\mathbf{P}_{\mathrm{out}}\mathbf{F}\mathbf{P}_{\mathrm{in}}\mathbf{D}_{\mathrm{in}}, where F\mathbf{F} is the three-point Fourier-transform matrix, the D\mathbf{D} matrices are diagonal phases, and the P\mathbf{P} matrices are permutations. Substituting this factorization into the reciprocal round-trip transformation T=MTRM\mathbf{T}=\mathbf{M}^{\mathsf{T}}\mathbf{R}\mathbf{M} shows that the internal phases can be absorbed into the controllable reflector phases, while external phases leave output powers unchanged, and permutations only relabel ports. Every such three-port coupler therefore has the same reachable power region as the Fourier-transform model, up to port relabeling.

For any NN, these ideal assumptions allow unit incident power at any input to be routed entirely to any chosen output, yielding Pk=1P_k=1 and Pj≠k=0P_{j\ne k}=0. Full proofs and the analytic three-port boundary are provided in Supplementary Notes 1 and 2. The routing theorems, the three-port reachable-region equivalence theorem, and supporting algebraic identities were machine-checked in Lean 4 using Mathlib; the formal source is supplied with the manuscript.

Data availability

The source data for the main-text numerical panels have been deposited in Zenodo under the reserved DOI https://doi.org/10.5281/zenodo.23194529 and will become publicly accessible when the record is published.

Code availability

Code supporting waveform generation, the transfer-matrix model, routing optimization, figure generation, and formal verification is maintained at https://github.com/sandrobuzzi/Multiport-Photonic-Routing. The repository is private during manuscript preparation; public release is planned with publication.

Acknowledgements

This work was supported by the DEVCOM Army Research Laboratory under Cooperative Agreement No. W911NF-22-2-0127. A.B. acknowledges support from the Air Force Research Laboratory under Award No. FA8750-25-2-0500 and from the Defense Advanced Research Projects Agency (DARPA) through the NanoWatt Platforms for Sensing, Analysis and Computation (NaPSAC) program as part of the TOPCHIP project under subcontract No. SCON-00005835. M. H. acknowledges support from The Novo Nordisk Foundation (REPEAT, grant no. NNF24OC0096351) and Innovation Fund Denmark (EQUAL, grant no. 4356-00007B). The authors thank Steven Nagle and the T. J. Rodgers RLE Laboratory for their continued support in developing the experimental setup and characterizing the electro-optic bandwidth. The authors thank CSEM, in particular Homa Zarebidaki, Amir H. Ghadimi, and Hamed Sattari, for fabricating the chips.

Author contributions

A.B. built the characterization setup, designed the control system, and measured the devices with support from B.A.F. and D.L.P.V. A.B., D.L.P.V., and M.H. analyzed the data. M.H. conceived the circuit concept, designed the devices, performed the optical simulations, and provided the DFT-based proof for pairwise routing. A.B. performed the electrostatic simulations and developed the Lean formalization, which D.R.E. reviewed. A.B. wrote the manuscript with support from M.H. and D.L.P.V. All authors reviewed the manuscript. D.L.P.V., M.H., and D.R.E. supervised the work.

Competing interests

The authors declare no competing interests.

Supplementary Note 1: Pairwise routing with a DFT MMI

Solution to the pairwise routing problem

We want to show that any input mode nn can be routed to any output mode mm of a reflection-based N×NN \times N interferometer using NN cavity-based phase shifters, of which only N−1N-1 relative phases are independent, as shown in Figure S1. We represent incident light in mode nn by an N×1N \times1 vector, x\mathbf{x}, which is zero everywhere except xn=1x_n = 1. The forward transfer matrix of the passive interferometer is denoted by M\mathbf{M}. Reciprocity makes the reverse transfer matrix MT\mathbf{M}^{\mathsf{T}}, so the round-trip input-output transformation is

Image not extracted

Figure S1. Sketch of an NN-port pairwise router with a passive MMI described by a transfer matrix M\mathbf{M} and NN single-sided cavity phase shifters with complex field-reflection coefficients rj(ω)r_j(\omega). Only N−1N-1 relative reflector phases are independent because a common phase does not affect the output powers.

z=T(R)x=MTRMx,(4)\mathbf{z} = \mathbf{T}(\mathbf{R})\mathbf{x} = \mathbf{M}^{\mathsf{T}}\mathbf{R}\mathbf{M}\mathbf{x}, \tag*{(4)}

where the transmission matrix, T(R)\mathbf{T}(\mathbf{R}), is a function of the diagonal matrix containing the reflection coefficients of the cavity phase shifters, R=diag⁡[rk(ω)]\mathbf{R} = \operatorname{diag}[r_k(\omega)]. Below, we show that the pairwise routing problem may be solved by choosing M\mathbf{M} to be the discrete Fourier transform (DFT) F=FT\mathbf{F} = \mathbf{F}^{\mathsf{T}},

y=Fx,yk=1N∑j=0N−1e−i2πNkjxj.(5)\mathbf{y} = \mathbf{F}\mathbf{x}, \qquad y_k = \frac{1}{\sqrt{N}}\sum_{j=0}^{N-1} e^{-i\frac{2\pi}{N}kj}x_j. \tag*{(5)}

First, we show that if all the phase shifts are zero, corresponding to R=I\mathbf{R} = \mathbf{I}, the transformation T(I)\mathbf{T}(\mathbf{I}) is a particular permutation matrix. Then, we use the shifting property of the DFT to provide solutions for R\mathbf{R} that result in NN cyclic permutations of the rows of T(I)\mathbf{T}(\mathbf{I}). Taken together, these NN configurations cover every ordered input–output pair: for each selected input nn and output mm, one configuration gives Tmn=1T_{mn} = 1, with all other elements in column nn equal to zero. Thus, any selected pair can be routed perfectly, although different pairs generally require different reflector settings.

For an input with light in only the nnth mode, only the nnth term in the sum in Eq. (5) is non-zero

yk=1Ne−i2πNkn.(6)y_k = \frac{1}{\sqrt{N}}e^{-i\frac{2\pi}{N}kn}. \tag*{(6)}

If R=I\mathbf{R} = \mathbf{I}, the output is

zm=1N∑k=0N−1e−i2πNkmyk=1N∑k=0N−1e−i2πNk(m+n).(7)z_m = \frac{1}{\sqrt{N}}\sum_{k=0}^{N-1}e^{-i\frac{2\pi}{N}km}y_k = \frac{1}{N}\sum_{k=0}^{N-1}e^{-i\frac{2\pi}{N}k(m+n)}. \tag*{(7)}

If we define the phase θmn=2π(m+n)/N\theta_{mn} = 2\pi(m+n)/N, we can write Eq. (7) as

zm=1N∑k=0N−1e−ikθmn=1N(1−e−iNθmn1−e−iθmn).(8)z_m = \frac{1}{N}\sum_{k=0}^{N-1}e^{-ik\theta_{mn}} = \frac{1}{N}\left(\frac{1-e^{-iN\theta_{mn}}}{1-e^{-i\theta_{mn}}}\right). \tag*{(8)}

Since mm and nn are integers, we see that Nθmn=2πpN\theta_{mn} = 2\pi p with p∈N0p \in\mathbb{N}_0. This means that zm=0z_m = 0 unless θmn=2πp\theta_{mn} = 2\pi p, which results in the requirement that m+n=0m+n=0 or m+n=Nm+n=N. When this requirement is fulfilled, we have from Eq. (7) that zm=1z_m=1. This means that an input in mode nn is routed to an output in mode mm, or, equivalently, the transmission matrix is a permutation matrix

Tmn(I)=δm,(−n) mod Nwith{m=0if n=0,m=N−nif n≠0.(9)T_{mn}(\mathbf{I}) = \delta_{m,(-n)\bmod N} \quad\text{with} \quad \begin{cases} m=0 & \text{if } n=0,\\ m=N-n & \text{if } n\ne0. \end{cases} \tag*{(9)}

The Kronecker delta introduced in Eq. (9) is 1 when its two indices agree and zero otherwise, with indices understood modulo NN. The relation in Eq. (9) shows that for R=I\mathbf{R} = \mathbf{I}, each input mode nn is routed to a particular output mode mm. To show an input mode nn may be routed to any output mode, we consider the shifting property of the Fourier transform

zm−l=1N∑k=0N−1e−i2πNkm(ykei2πNkl).(10)z_{m-l} = \frac{1}{\sqrt{N}} \sum_{k=0}^{N-1} e^{-i\frac{2\pi}{N}km}\left(y_k e^{i\frac{2\pi}{N}kl}\right). \tag*{(10)}

In this context, Eq. (10) means that the output light will shift from mode mm to mode m+lm+l (modulo NN) if we choose the reflection coefficients as

rk(ω)=ei2πNkl.(11)r_k(\omega) = e^{i\frac{2\pi}{N}kl}. \tag*{(11)}

This also becomes clear by inserting Eq. (6) into Eq. (10) and defining a new phase θmln=(m−l+n)2π/N\theta_{mln}=(m-l+n)2\pi/N. Following the arguments above, this results in

Tmn(R)=δm,(l−n) mod Nwith{m=l−nif n≤l,m=N−(n−l)if n>l.(12)T_{mn}(\mathbf{R}) = \delta_{m,(l-n)\ \mathrm{mod}\ N} \quad\text{with} \quad \begin{cases} m=l-n & \text{if } n\leq l,\\ m=N-(n-l) & \text{if } n>l. \end{cases} \tag*{(12)}

With the NN different choices of the phase in Eq. (11), corresponding to l∈{0,1,2,…,N−1}l\in\{0,1,2,\ldots,N-1\}, we get NN cyclic permutations of the rows of T(I)\mathbf{T}(\mathbf{I}). This means that for any choice of n∈{0,1,2,…,N−1}n\in\{0,1,2,\ldots,N-1\}, we can choose an ll, such that we get any m∈{0,1,2,…,N−1}m\in\{0,1,2,\ldots,N-1\}, which solves the routing problem. Note that r0(ω)=1r_0(\omega)=1 independently of our choice of ll, so we only need N−1N-1 tunable cavities.

At this point, we also note that only NN out of the N!N! possible N×NN\times N permutation matrices may be realized by the solution provided here. We have solved the pairwise routing problem, but not the realization of arbitrary permutation matrices.

Supplementary Note 2: General routing theorems and three-port reachable-region equivalence

This note gives three structural results and one quantitative consequence. First, fixed reciprocal phase delays between an MMI and its reflectors do not change its reachable set of output-power distributions. Second, every lossless equal-split N×NN\times N MMI supports unit-efficiency routing between any selected input–output pair, without requiring the MMI to be a DFT. Third, for 3×33\times3 devices, every lossless equal-split MMI has the same reachable output-power region as the DFT, up to a fixed relabeling of ports. Finally, we derive this common three-port region analytically, including the sharp (4/9,4/9,1/9)(4/9,4/9,1/9) balanced two-output limit.

All theorems in this note, together with their supporting algebraic lemmas, have been formalized and machine-checked in Lean 4 against Mathlib. The formal development is supplied with the manuscript and includes checks of the explicit three-port field identities and the sharp balanced-splitting result derived below.

Throughout this note, external ports are indexed by nn, mm, qq and the internal reflector arms by jj. The forward MMI matrix is M\mathbf{M}, reciprocity makes the reverse matrix MT\mathbf{M}^{\mathsf{T}}, and

D=diag⁡(d0,…,dN−1),dj=eiδj,\mathbf{D}=\operatorname{diag}(d_0,\ldots,d_{N-1}), \qquad d_j=e^{i\delta_j},
R=diag⁡(r0,…,rN−1),∣rj∣=1(13)\mathbf{R}=\operatorname{diag}(r_0,\ldots,r_{N-1}), \qquad|r_j|=1 \tag*{(13)}

denote the fixed arm phases and the controllable reflector phases, respectively. With ideal components, the round-trip matrix is

TM,D(R)=MTDTRDM.(14)\mathbf{T}_{\mathbf{M},\mathbf{D}}(\mathbf{R})=\mathbf{M}^{\mathsf{T}}\mathbf{D}^{\mathsf{T}}\mathbf{R}\mathbf{D}\mathbf{M}. \tag*{(14)}

For a fixed input port nn, the basis vector en\mathbf{e}_n represents unit excitation at that port. For a chosen reflector configuration R\mathbf{R}, the product TM,D(R)en\mathbf{T}_{\mathrm{M,D}}(\mathbf{R})\mathbf{e}_n is the complex output-field vector. Its qqth component is the field amplitude at output port qq, so the corresponding output power is

pq(R;n):=∣[TM,D(R)en]q∣2,q=0,…,N−1.(15)p_q(\mathbf{R};n) := \left|[\mathbf{T}_{\mathrm{M,D}}(\mathbf{R})\mathbf{e}_n]_q\right|^2,\qquad q=0,\ldots,N-1. \tag*{(15)}

Collecting these components gives the output-power vector p(R;n)=(p0(R;n),…,pN−1(R;n))\mathbf{p}(\mathbf{R};n)=(p_0(\mathbf{R};n),\ldots,p_{N-1}(\mathbf{R};n)). The reachable power set is

PM,D(n):={p(R;n):R diagonal and unitary}.(16)\mathcal{P}_{\mathrm{M,D}}(n):=\{\mathbf{p}(\mathbf{R};n):\mathbf{R}\ \text{diagonal and unitary}\}. \tag*{(16)}

Thus, PM,D(n)\mathcal{P}_{\mathrm{M,D}}(n) contains every output-power distribution obtainable by varying the independently controllable reflector phases while holding the input port nn fixed.

Fixed diagonal phases preserve the reachable set

Theorem 1 (Fixed-phase invariance). Let M\mathbf{M} be any N×NN\times N matrix and let D\mathbf{D} be any fixed diagonal unitary matrix. Then, for every input port nn,

PM,D(n)=PM,I(n).(17)\mathcal{P}_{\mathrm{M,D}}(n)=\mathcal{P}_{\mathrm{M,I}}(n). \tag*{(17)}

In fact, the sets of reachable round-trip matrices are identical. No equal-splitting or DFT assumption is required for this invariance.

Proof. Because D\mathbf{D} is diagonal, DT=D\mathbf{D}^{\mathsf{T}}=\mathbf{D}, and because both D\mathbf{D} and R\mathbf{R} are diagonal they commute. Consequently

DTRD=DRD=D2R.(18)\mathbf{D}^{\mathsf{T}}\mathbf{R}\mathbf{D}=\mathbf{D}\mathbf{R}\mathbf{D}=\mathbf{D}^{2}\mathbf{R}. \tag*{(18)}

Introduce the effective reflector matrix

R′=D2R.(19)\mathbf{R}'=\mathbf{D}^{2}\mathbf{R}. \tag*{(19)}

It is diagonal and unitary. Moreover, the map R↦R′\mathbf{R}\mapsto\mathbf{R}' is a bijection on the set of diagonal unitary matrices: for any desired R′\mathbf{R}', its unique preimage is R=D−2R′\mathbf{R}=\mathbf{D}^{-2}\mathbf{R}'. Hence

{MTDTRDM:R diagonal unitary}={MTR′M:R′ diagonal unitary}.(20)\{\mathbf{M}^{\mathsf{T}}\mathbf{D}^{\mathsf{T}}\mathbf{R}\mathbf{D}\mathbf{M}:\mathbf{R}\ \text{diagonal unitary}\} = \{\mathbf{M}^{\mathsf{T}}\mathbf{R}'\mathbf{M}:\mathbf{R}'\ \text{diagonal unitary}\}. \tag*{(20)}

Acting on en\mathbf{e}_n and taking componentwise squared magnitudes gives Eq. (S14). □\square

The theorem assumes that all NN diagonal phases can be varied. Only N−1N-1 independent controls are needed for powers. Indeed, multiplying every reflector coefficient by the same eiϑe^{i\vartheta} multiplies the complete output field by eiϑe^{i\vartheta} and leaves every output power unchanged. Every diagonal unitary can therefore be represented, up to this irrelevant common phase, by one whose reference entry is unity.

Pairwise routing for any equal-split unitary and any NN

Supplementary Note 1 establishes pairwise routing for a DFT MMI. We now show that the DFT structure is unnecessary: equal splitting and unitarity alone suffice.

By the fixed-phase-invariance theorem, the transformation R0=DTRD\mathbf{R}_0=\mathbf{D}^{\mathsf{T}}\mathbf{R}\mathbf{D} is a bijection on the set of diagonal unitary reflector matrices. We may therefore set D=I\mathbf{D}=\mathbf{I} while deriving the routing solution. For a device with fixed arm phases D\mathbf{D}, the corresponding physical reflector matrix is R=(DT)−1R0D−1=D−2R0\mathbf{R}=(\mathbf{D}^{\mathsf{T}})^{-1}\mathbf{R}_0\mathbf{D}^{-1}=\mathbf{D}^{-2}\mathbf{R}_0.

Theorem 2 (General pairwise routing). Let M∈U(N)\mathbf{M} \in U(N) be lossless and equal-split,

∣Mjk∣=1N,j,k=0,…,N−1.(21)|M_{jk}| = \frac{1}{\sqrt{N}}, \qquad j,k=0,\ldots,N-1. \tag*{(21)}

For every input port nn and target output port mm, there is a diagonal phase-only reflector matrix R0\mathbf{R}_0 such that

MTR0Men=eiθem,(22)\mathbf{M}^{\mathsf{T}}\mathbf{R}_0\mathbf{M}\mathbf{e}_n = e^{i\theta}\mathbf{e}_m, \tag*{(22)}

where θ\theta is an arbitrary common phase. Thus, all N(N+1)/2N(N+1)/2 distinct reciprocal routing connections are individually reachable with unit efficiency in the ideal case: N(N−1)/2N(N-1)/2 connect different pairs of ports, and the remaining NN are self-reflections.

Proof. Follow the input field through the device:

en→MMen→R0R0Men→MTMTR0Men.(23)\mathbf{e}_n \xrightarrow{\mathbf{M}} \mathbf{M}\mathbf{e}_n \xrightarrow{\mathbf{R}_0} \mathbf{R}_0\mathbf{M}\mathbf{e}_n \xrightarrow{\mathbf{M}^{\mathsf{T}}} \mathbf{M}^{\mathsf{T}}\mathbf{R}_0\mathbf{M}\mathbf{e}_n. \tag*{(23)}

The amplitude at an arbitrary output port qq is

zq:=[MTR0Men]q=∑j=0N−1Mjqr0,jMjn.(24)z_q := [\mathbf{M}^{\mathsf{T}}\mathbf{R}_0\mathbf{M}\mathbf{e}_n]_q = \sum_{j=0}^{N-1} M_{jq}r_{0,j}M_{jn}. \tag*{(24)}

Choose the reflector coefficients arm by arm as

r0,j=eiθMjm‾Mjn,(25)r_{0,j} = e^{i\theta}\frac{\overline{M_{jm}}}{M_{jn}}, \tag*{(25)}

where the overline denotes complex conjugation. The division is well-defined because equal splitting makes every MjnM_{jn} non-zero. More importantly, Eq. (25) is phase-only:

∣r0,j∣=∣Mjm∣∣Mjn∣=1/N1/N=1.(26)|r_{0,j}| = \frac{|M_{jm}|}{|M_{jn}|} = \frac{1/\sqrt{N}}{1/\sqrt{N}} = 1. \tag*{(26)}

Substitution into Eq. (24) gives

zq=eiθ∑j=0N−1MjqMjm‾=eiθδqm.(27)\begin{aligned} z_q &= e^{i\theta}\sum_{j=0}^{N-1} M_{jq}\overline{M_{jm}}\\ &= e^{i\theta}\delta_{qm}. \tag*{(27)} \end{aligned}

The final identity is the orthogonality of columns qq and mm of the unitary matrix M\mathbf{M}: ∑jMjqMjm‾=δqm\sum_j M_{jq}\overline{M_{jm}}=\delta_{qm}. Therefore, the desired output has unit amplitude and every other output has zero amplitude, proving Eq. (22).

For exact phase-only routing between every input–output pair, the relevant column magnitudes must satisfy

∣Mjn∣=∣Mjm∣for all j,n,m.(28)|M_{jn}| = |M_{jm}| \qquad\text{for all } j,n,m. \tag*{(28)}

Thus, for each row jj, all NN matrix elements have the same magnitude. Because M\mathbf{M} is unitary, each row has unit norm, so

∑k=0N−1∣Mjk∣2=1(29)\sum_{k=0}^{N-1} \lvert M_{jk}\rvert^2 = 1 \tag*{(29)}

implies

∣Mjk∣=1Nfor all j,k.(30)\lvert M_{jk}\rvert= \frac{1}{\sqrt{N}} \quad\text{for all } j,k. \tag*{(30)}

Equal splitting is therefore necessary for exact phase-only routing between every pair of ports, and the construction above shows that it is also sufficient. The arbitrary common phase θ\theta may be chosen so that one selected r0,jr_{0,j} is unity, leaving N−1N-1 independently controlled reflector phases. For the DFT convention of Supplementary Note 1, Mjn=N−1/2e−i2πjn/NM_{jn}=N^{-1/2}e^{-i2\pi jn/N}, the general solution reduces to

r0,j=eiθei2πj(m+n)/N,(31)r_{0,j}=e^{i\theta}e^{i2\pi j(m+n)/N}, \tag*{(31)}

which is the Fourier phase ramp after the fixed arm phases have been absorbed.

Simultaneous port swaps, self-reflections, and routing schedules

Pairwise routing completeness means that any selected input–output pair can be connected. It does not require that only one input be illuminated: a fixed phase setting acts on all inputs simultaneously, although the resulting connections cannot generally be chosen independently. Absorbing fixed arm phases into R\mathbf{R}, the lossless round-trip matrix T=MTRM\mathbf{T}=\mathbf{M}^{\mathsf{T}}\mathbf{R}\mathbf{M} is unitary and symmetric. We call a simultaneous routing configuration exact when every input nn is routed with unit efficiency to a single output π(n)\pi(n), so that

∣Tmn∣2=δm,π(n).(32)\lvert T_{mn}\rvert^2=\delta_{m,\pi(n)}. \tag*{(32)}

Equivalently, the power-transfer matrix P=∣T∣2\mathbf{P}=\lvert\mathbf{T}\rvert^2 is a permutation matrix. Reciprocity makes P\mathbf{P} symmetric. A symmetric permutation matrix satisfies P2=I\mathbf{P}^2=\mathbf{I}, so its permutation is an involution. An involution is a permutation that returns every port to itself when applied twice; its cycles can therefore only have length one or two. Thus, every exact simultaneous routing configuration consists of reciprocal two-port swaps and single-port self-reflections, and cycles involving three or more ports are excluded. For three ports, an exact swap necessarily leaves the third port fully self-reflected.

Discrete Fourier Transform MMI.

With the fixed arm phases absorbed into R\mathbf{R}, define the Fourier coefficients of the reflector settings as

cs=1N∑j=0N−1rje−i2πjs/N.(33)c_s=\frac{1}{N}\sum_{j=0}^{N-1}r_j e^{-i2\pi js/N}. \tag*{(33)}

For the DFT MMI, the round-trip matrix elements are then

Tmn=c(m+n) mod N.(34)T_{mn}=c_{(m+n)\bmod N}. \tag*{(34)}

From Eq. (S28), after absorbing the fixed arm phases, the reflector setting for routing between ports nn and mm depends only on

l=(m+n)mod⁡N.(35)l = (m + n) \mathbin{\operatorname{mod}} N. \tag*{(35)}

The same setting routes every input nn to the output

m=(l−n)mod⁡N.(36)m = (l - n) \mathbin{\operatorname{mod}} N. \tag*{(36)}

For a fixed setting ll, if m=nm = n, that port is self-reflected; otherwise, nn and mm form a reciprocal pair. For N=3N = 3, the settings l=0,1,2l = 0, 1, 2 respectively produce 0→00 \to0 with 1↔21 \leftrightarrow2, 2→22 \to2 with 0↔10 \leftrightarrow1, and 1→11 \to1 with 0↔20 \leftrightarrow2. In the physical one-based port labels, these are the three possible reciprocal swaps, each with the remaining port self-reflected. Thus these NN permutations exhaust the exact routing configurations, up to a common phase. For odd NN, each has one self-reflection and (N−1)/2(N - 1)/2 swaps. For even NN, odd ll gives N/2N/2 swaps; even ll gives two self-reflections and (N−2)/2(N - 2)/2 swaps, since fixed ports satisfy 2n=lmod⁡N2n = l \mathbin{\operatorname{mod}} N. Because every exact reciprocal configuration is an involution, at most ⌊N/2⌋\lfloor N/2 \rfloor swaps are possible and at least Nmod⁡2N \mathbin{\operatorname{mod}} 2 ports must self-reflect. The Fourier construction attains this bound for every pair when NN is odd and, when NN is even, for opposite-parity pairs; a same-parity pair instead gives two self-reflections, a restriction of the Fourier family rather than a fundamental requirement. By the DFT equivalence proved below, the same three-port result holds, up to port phases and relabeling, for every equal-split unitary 3×33 \times3 MMI. For N=5N = 5, the Fourier family supports five configurations, one for each self-reflected port, out of the 15 possible configurations comprising two swaps and one self-reflection. Once the self-reflected port is chosen, the Fourier construction fixes the other two pairs; without that restriction, the remaining four ports admit three pairings, giving 5×3=155 \times3 = 15 possibilities in total.

An all-to-all schedule for every $N.

As shown in Supplementary Note 1, the NN DFT reflector settings generate permutation matrices satisfying m=l−nmod⁡Nm = l - n \mathbin{\operatorname{mod}} N. Hence, for any input–output pair (n,m)(n, m), the unique required setting is l=n+mmod⁡Nl = n + m \mathbin{\operatorname{mod}} N. The NN settings therefore route every input to every output exactly once. Thus, as ll runs from 0 to N−1N - 1, each of the (N2)\binom{N}{2} possible reciprocal pairings of distinct ports occurs exactly once.

We consider the scheduling problem of realizing all (N2)\binom{N}{2} reciprocal pairings between distinct ports using the fewest reflector configurations. The combinatorial minimum is NN configurations for odd NN and N−1N - 1 for even NN. The DFT schedule uses NN configurations and is therefore optimal for odd NN, but uses one more configuration than the combinatorial minimum for even NN. For N=2kN = 2^k, the Walsh construction discussed below attains the N−1N - 1 minimum. Each port self-reflects in exactly one of the NN DFT settings, namely l=2nl = 2n. The five settings and their combined pair coverage are illustrated in Supplementary Fig. S2.

Walsh interferometer.

A Walsh interferometer implements the normalized Walsh–Hadamard transform, whose entries have equal magnitude and phases restricted to 0 or π\pi. It is defined for N=2kN = 2^k ports. We label each port index nn and arm index jj by the kk-bit binary representation of its integer index, so n,j∈F2kn, j \in\mathbb{F}_2^k. We denote the Walsh matrix by WW, with entries

Wjn=N−1/2(−1)j⋅n,W_{jn} = N^{-1/2}(-1)^{j \cdot n},

where j⋅n=∑t=0k−1jtntmod⁡2j \cdot n = \sum_{t=0}^{k-1} j_t n_t \mathbin{\operatorname{mod}} 2 is the parity of the positions in which both binary labels contain a one. Thus WjnW_{jn} is +1/N+1/\sqrt{N} for even parity and −1/N-1/\sqrt{N} for odd parity. The symbol ⊕\oplus denotes the exclusive-OR (XOR) of the binary labels, performed independently on every bit: 0⊕0=1⊕1=00 \oplus0 = 1 \oplus1 = 0 and 0⊕1=1⊕0=10 \oplus1 = 1 \oplus0 = 1. For example, in an eight-port device,

Image not extracted

Figure S2. Five-setting all-to-all schedule of the five-port Fourier router. The first five panels are the switching configurations; the sixth panel combines their color-coded reciprocal pairings and is not itself one switch setting. Setting ll implements n↦l−nn \mapsto l-n (mod 5). Each pentagon contains two simultaneous reciprocal pairings and one self-reflected port (colored ring), which is unavoidable for odd NN. Across the five panels, every side and every diagonal of the pentagon occurs exactly once, so all ten port pairs are connected in the minimum possible five settings.

3⊕5=0112⊕1012=1102=63 \oplus5 = 011_{2} \oplus101_{2} = 110_{2} = 6. Following the field from input nn through arm jj and back to output mm gives

Tmn=∑jWjmrjWjn=1N∑jrj(−1)j⋅m(−1)j⋅n=1N∑jrj(−1)j⋅(m⊕n).\begin{aligned} T_{mn} &= \sum_{j} W_{jm}r_jW_{jn}\\ &= \frac{1}{N}\sum_{j}r_j(-1)^{j\cdot m}(-1)^{j\cdot n}\\ &= \frac{1}{N}\sum_{j}r_j(-1)^{j\cdot(m\oplus n)}. \end{aligned}

The last equality follows because the exponent matters only modulo two:

j⋅m+j⋅n≡∑tjt(mt+nt)≡∑tjt(mt⊕nt)=j⋅(m⊕n)(mod2).j\cdot m+j\cdot n \equiv\sum_{t}j_t(m_t+n_t) \equiv\sum_{t}j_t(m_t\oplus n_t)=j\cdot(m\oplus n)\pmod{2}.

We define the Walsh-transform coefficients of the reflector settings by

Tmn=bm⊕n,ba=1N∑jrj(−1)j⋅a.(37)T_{mn}=b_{m\oplus n},\qquad b_a=\frac{1}{N}\sum_{j}r_j(-1)^{j\cdot a}. \tag*{(37)}

Like csc_s in the Fourier construction, bab_a is a coherently summed field coefficient, not a physical component. Selecting a binary mask aa and setting the reflector coefficients to

rj=eiθ(−1)j⋅ar_j=e^{i\theta}(-1)^{j\cdot a}

gives, for an arbitrary binary label uu,

bu=1N∑jeiθ(−1)j⋅a(−1)j⋅u=eiθN∑j(−1)j⋅(a⊕u)b_u = \frac{1}{N}\sum_j e^{i\theta}(-1)^{j\cdot a}(-1)^{j\cdot u} = \frac{e^{i\theta}}{N}\sum_j(-1)^{j\cdot(a\oplus u)}

In the second line we used j⋅a+j⋅u=j⋅(a⊕u)j\cdot a+j\cdot u=j\cdot(a\oplus u) (mod 2). If u=au=a, then a⊕u=0a\oplus u=0, so all NN terms in the sum equal +1+1. If u≠au\ne a, choose a bit position tt for which (a⊕u)t=1(a\oplus u)_t=1. Each term indexed by jj then cancels the term indexed by j⊕etj\oplus e_t, where ete_t has a one only in position tt. Hence

bu=eiθδu,a,Tmn=bm⊕n=eiθδm⊕n,a=eiθδm,n⊕a.(38)b_u=e^{i\theta}\delta_{u,a}, \qquad T_{mn}=b_{m\oplus n}=e^{i\theta}\delta_{m\oplus n,a}=e^{i\theta}\delta_{m,n\oplus a}. \tag*{(38)}

The last equality follows by applying XOR with nn to m⊕n=am\oplus n=a, using n⊕n=0n\oplus n=0. Consequently, every input nn exits from the unique output m=n⊕am=n\oplus a. For example, for N=4N=4 the labels are 00, 01, 10, 11; choosing a=01a=01 simultaneously pairs 00↔0100\leftrightarrow01 and 10↔1110\leftrightarrow11.

There are exactly NN configurations: identity for a=0a=0, and N−1N-1 complete pairings without self-reflections for a≠0a\ne0. For even NN, reciprocity permits (N−1)!!=(N−1)(N−3)⋯3⋅1(N-1)!!=(N-1)(N-3)\cdots3\cdot1 reflection-free permutation matrices: the lowest-index unpaired port has N−1N-1 possible partners, the lowest-index remaining port then has N−3N-3, and so on. The N−1N-1 Walsh matchings form an optimal all-to-all schedule: no setting can contain more than N/2N/2 disjoint swaps, and these matchings together cover each of the (N2)=N(N−1)/2\binom{N}{2}=N(N-1)/2 unordered port pairs exactly once. Including the NN self-reflections realized by the identity setting a=0a=0, the full schedule covers all N(N+1)/2N(N+1)/2 unordered input–output connections. Indeed, the swap n↔mn\leftrightarrow m is obtained with a=n⊕m≠0a=n\oplus m\ne0, while a fixed point would require n⊕a=nn\oplus a=n and hence a=0a=0. A Walsh 4×44\times4 therefore supports all three complete pairings, but cannot perform just one swap while self-reflecting the other two ports. For larger devices it still selects only a subset of complete matchings: Walsh 8×88\times8 supports seven of the 7×5×3=1057\times5\times3=105 possibilities, although each pair occurs in one of those seven reflection-free matchings.

The N=8N=8 schedule and its binary drive.

Table 1 writes the complete schedule explicitly. Ports and arms are labeled by the integers 0,…,70,\ldots,7, identified with their three-bit binary representations. In row aa, port nn is paired with n⊕an\oplus a, while arm jj receives a relative π\pi phase precisely when the binary inner product j⋅aj\cdot a is odd. Arm 0 is used as the phase reference and therefore remains at relative phase zero in every row.

routing mask aasimultaneous port pairs {n,n⊕a}\{n,n\oplus a\}1234567
binary arm label jj001010011100101110111
0010–1, 2–3, 4–5, 6–7π\pi0π\pi0π\pi0π\pi
0100–2, 1–3, 4–6, 5–70π\piπ\pi00π\piπ\pi
0110–3, 1–2, 4–7, 5–6π\piπ\pi00π\piπ\pi0
1000–4, 1–5, 2–6, 3–7000π\piπ\piπ\piπ\pi
1010–5, 1–4, 2–7, 3–6π\pi0π\piπ\pi0π\pi0
1100–6, 1–7, 2–4, 3–50π\piπ\piπ\piπ\pi00
1110–7, 1–6, 2–5, 3–4π\piπ\pi0π\pi00π\pi

Table 1. Seven-setting all-to-all schedule and binary phase drive for an eight-port Walsh router. Every row is a reflection-free complete matching, and every one of the 28 reciprocal pairings occurs in exactly one row. The entries give the reflector phase on each programmable arm relative to arm 0. Equivalently, arm jj is set to π\pi when j⋅aj\cdot a is odd and to zero otherwise. The omitted setting a=000a=000 is the identity and sets every arm to the same phase.

In the phase-reference convention of Table 1, moving between any two distinct switch configurations, including a transition to or from the identity, changes exactly four of the eight arm phases, that is, half of them. Because the zero-labeled arm is fixed as the phase reference, all four changes occur among the seven programmable arms shown in the table. Indeed, configurations aa and a′a' differ on arm jj precisely when j⋅(a⊕a′)=1j\cdot(a\oplus a')=1. For a≠a′a\ne a', this nonzero binary linear functional is balanced over the eight three-bit arm labels, taking the value one for exactly four labels. The seven nonidentity matchings and their combined pair coverage are illustrated in Supplementary Fig. S3.

Because the Walsh transform is invertible, the reflector pattern in Eq. (38) is unique up to its common phase. After removal of that common phase, every reflector therefore uses only the two values +1+1 and −1-1, or relative phases 0 and π\pi. The zero-labeled arm is fixed, and the remaining N−1N-1 arms are parity functions of the k=log⁡2Nk=\log_2N address bits.

Image not extracted

Figure S3. The seven nonidentity settings of the eight-port Walsh router. The first seven panels are the switching configurations; the eighth panel combines their color-coded reciprocal pairings and is not itself one switch setting. Each octagon shows one reflection-free complete matching n↔n⊕an\leftrightarrow n\oplus a; taken together, the panels contain every side and diagonal of the eight-vertex complete graph exactly once. The line below each panel lists the arms whose reflectors receive the relative π\pi phase (−1)j⋅a=−1(-1)^{j\cdot a}=-1. Port and arm labels use the same binary encoding but do not denote a direct physical port–arm connection.

Why concentrating all self-reflections in the identity requires N=2kN=2^k.

Suppose a single-layer reciprocal router realizes one identity setting and N−1N-1 reflection-free complete matchings that together cover every port pair. These NN settings send any chosen input once to each output. After taking the identity setting as a reference, their diagonal reflector patterns multiply arm by arm, so the corresponding routing permutations commute.

The product of two scheduled permutations also commutes with every schedule member. Choose the scheduled permutation that sends a reference input to the same output as this product. Since every other port is reached from the reference input by a scheduled permutation, commutativity forces the two permutations to agree on every port. The schedule is therefore closed under composition.

Reciprocity makes every reflection-free routing permutation a product of disjoint swaps, so applying any such operation twice gives the identity. The independent routing operations are therefore binary choices: each is either used once or not used, and adding one independent choice doubles the number of configurations. With kk independent choices the family contains 2k2^k configurations. Since the assumed schedule contains exactly NN,

N=2k.(39)N = 2^k. \tag*{(39)}

For six ports, for example, the identity and five matchings would require six configurations, whereas two binary choices give four and three give eight; six is impossible. This is an architectural, not a combinatorial, restriction: five complete matchings do exist for six ports, but realizing them requires a more general mixing/control architecture. The Walsh interferometer construction attains the binary schedule when N=2kN = 2^k: a=0a = 0 is the single all-self-reflecting identity, while every nonzero XOR mask is reflection-free. Other port counts can instead use the NN-setting Fourier schedule above.

Selected-pair routing versus simultaneous port permutations.

Equal splitting alone guarantees phase-only routing of one selected input to one selected output, as proved in Eq. (25). For N>3N > 3 it does not by itself require the remaining N−2N - 2 inputs to route one-to-one to distinct outputs without power splitting; they may instead undergo a smaller unitary mixing. The Fourier and Walsh matrices above are explicit constructions for which the entire matrix is a permutation in every listed setting. More generally, if a single-layer router does possess exact one-to-one routing settings (permutations up to phases) covering every port pair, the binary-choice argument above constrains their simultaneous routing and self-reflection patterns. This distinction separates complete point-to-point accessibility from the stronger ability to prescribe a complete simultaneous matching.

Planar realization with balanced 2×22 \times2 unitaries.

For N=2kN = 2^k, the Walsh transform can be implemented as a radix-two butterfly containing k=log⁡2Nk = \log_2 N sequential layers. Each layer contains N/2N/2 balanced 2×22 \times2 unitaries that act in parallel on disjoint pairs of modes, so every mode passes through one 2×22 \times2 unitary in each layer. Writing H=2−1/2(111−1)H = 2^{-1/2}\begin{pmatrix}1 & 1 \\ 1 & -1\end{pmatrix} and denoting the d×dd \times d identity by IdI_d, the factorization is

WN=H⊗k=∏r=0k−1(I2r⊗H⊗I2k−r−1).(40)W_N = H^{\otimes k} = \prod_{r=0}^{k-1}\left(I_{2^r} \otimes H \otimes I_{2^{k-r-1}}\right). \tag*{(40)}

Each factor mixes the N/2N/2 disjoint pairs of modes whose binary labels differ in one bit. The complete butterfly therefore contains (N/2)log⁡2N(N/2)\log_2 N passive 2×22 \times2 unitaries. Conversely, after LL layers of disjoint two-mode couplers, one input can reach at most 2L2^L outputs. Equal splitting among all NN outputs therefore requires L≥log⁡2NL \ge\log_2 N, so this depth is minimal. Fixed crossings or mode permutations provide the required butterfly connectivity. The DFT uses the same network with phase shifts between layers. All unitaries remain at a 50:50 power split; only relative phases are changed. Thus, this passive realization does not alter the O(N)O(N) number of actively controlled cavity reflectors in the routing architecture.

For the four-port case, let a symmetric balanced 2×22 \times2 unitary have the scattering matrix QQ. Two such unitaries acting in parallel on disjoint pairs of modes give the block-diagonal stage QsQ_s:

Q=12(1ii1),Qs=diag⁡(Q,Q)=12(1i00i100001i00i1).(41)Q = \frac{1}{\sqrt{2}}\begin{pmatrix}1 & i \\ i & 1\end{pmatrix}, \qquad Q_s = \operatorname{diag}(Q,Q) = \frac{1}{\sqrt{2}}\begin{pmatrix}1 & i & 0 & 0 \\ i & 1 & 0 & 0 \\ 0 & 0 & 1 & i \\ 0 & 0 & i & 1\end{pmatrix}. \tag*{(41)}

Let a waveguide crossing exchange the two middle rails; its permutation matrix is

P=(1000001001000001).(42)\mathbf{P}=\begin{pmatrix}1&0&0&0\\0&0&1&0\\0&1&0&0\\0&0&0&1\end{pmatrix}. \tag*{(42)}

The required phase-only interstage section is

Φ(α)=diag⁡(1,−i,−i,−eiα),(43)\mathbf{\Phi}(\alpha)=\operatorname{diag}(1,-i,-i,-e^{i\alpha}), \tag*{(43)}

corresponding, modulo 2π2\pi, to phase settings (0,−π/2,−π/2,α+π)(0,-\pi/2,-\pi/2,\alpha+\pi). With L=diag⁡(1,i,1,i)\mathbf{L}=\operatorname{diag}(1,i,1,i), direct multiplication gives

QsΦ(α)Qs=LPM4(α)PL,(44)\mathbf{Q}_{\mathrm{s}}\mathbf{\Phi}(\alpha)\mathbf{Q}_{\mathrm{s}}=\mathbf{L}\mathbf{P}\mathbf{M}_4(\alpha)\mathbf{P}\mathbf{L}, \tag*{(44)}

where multiplying from the left by PTL−1\mathbf{P}^{\mathsf{T}}\mathbf{L}^{-1} and from the right by L−1PT\mathbf{L}^{-1}\mathbf{P}^{\mathsf{T}} recovers

M4(α)=12(11111eiα−1−eiα1−11−11−eiα−1eiα).(45)\mathbf{M}_4(\alpha)=\frac{1}{2}\begin{pmatrix}1&1&1&1\\1&e^{i\alpha}&-1&-e^{i\alpha}\\1&-1&1&-1\\1&-e^{i\alpha}&-1&e^{i\alpha}\end{pmatrix}. \tag*{(45)}

Hence α=0\alpha=0 gives the Walsh matrix, α=π/2\alpha=\pi/2 gives the four-point DFT, and intermediate values continuously connect the two while preserving equal splitting. The external diagonal phases do not change routed powers and can either be absorbed into the port reference planes or implemented as fixed planar phase sections. A compact single-layer layout is shown in Supplementary Fig. S4.

Image not extracted

Figure S4. Planar realization of a four-port MMI with balanced 2×22\times2 unitaries. The external ports use the bit-reversed order (0,2,1,3)(0,2,1,3), so only the two middle internal rails are exchanged by a waveguide crossing. The four unitaries remain balanced; the setting of the interstage phase, α\alpha, selects the Walsh matrix (α=0\alpha=0), the DFT matrix (α=π/2\alpha=\pi/2), or an intermediate equal-split matrix.

Reachable-region equivalence for 3 × 3 equal-split MMIs

Pairwise routing uses only the vertices of the output power simplex. For 3×33\times3 equal-split MMIs, a stronger result holds: the complete interior reachable region is also independent of the MMI phases, up to a fixed port permutation.

Theorem 3 (3×33\times3 reachable-region equivalence). Let M∈U(3)\mathbf{M}\in U(3) satisfy ∣Mjk∣=1/3|M_{jk}|=1/\sqrt{3}, and let D\mathbf{D} be any fixed diagonal unitary. For every physical input port nn, there are a relabeled input index n′∈{0,1,2}n'\in\{0,1,2\} and a fixed permutation Π\mathbf{\Pi} of the three output power coordinates such that

PM,D(n)=Π(PF,I(n′)),(46)\mathcal{P}_{\mathbf{M},\mathbf{D}}(n)=\mathbf{\Pi}\left(\mathcal{P}_{\mathbf{F},\mathbf{I}}(n')\right), \tag*{(46)}

:::`

where F\mathbf{F} is the normalized 3×33 \times3 DFT matrix. Here n′n' is the input index in the canonical DFT representation corresponding to physical port nn; it is not an additional physical port. Thus fixed arm phases and the internal phases of an arbitrary equal-split 3×33 \times3 MMI do not change the shape of the reachable power region.

Proof. We first show that every equal-split 3×33 \times3 unitary is equivalent to the DFT through diagonal phases and permutations. Define H=3M\mathbf{H} = \sqrt{3}\mathbf{M}. Every entry of H\mathbf{H} has unit modulus, and its rows and columns are mutually orthogonal. Multiplying rows and columns by unit-modulus constants preserves these properties. We may therefore multiply on the right by a diagonal phase matrix to make the first row equal to (1,1,1)(1,1,1), and then multiply on the left by another diagonal phase matrix to make the first column equal to (1,1,1)T(1,1,1)^{\mathsf{T}}. The resulting dephased matrix has the form

H′=(1111ab1cd),∣a∣=∣b∣=∣c∣=∣d∣=1.(47)\mathbf{H}' = \begin{pmatrix} 1 & 1 & 1 \\ 1 & a & b \\ 1 & c & d \end{pmatrix}, \qquad|a| = |b| = |c| = |d| = 1. \tag*{(47)}

Orthogonality of the first two rows gives 1+a+b=01+a+b=0. Since b=−(1+a)b=-(1+a) and ∣a∣=∣b∣=1|a|=|b|=1,

1=∣b∣2=∣1+a∣2=2+2Re⁡(a),(48)1=|b|^2=|1+a|^2=2+2\operatorname{Re}(a), \tag*{(48)}

so Re⁡(a)=−1/2\operatorname{Re}(a)=-1/2. Together with ∣a∣=1|a|=1, this gives a∈{ω,ω2}a\in\{\omega,\omega^2\}, where ω=ei2π/3\omega=e^{i2\pi/3}; then bb is the other root. After possibly exchanging the last two columns, the second row is (1,ω,ω2)(1,\omega,\omega^2). Applying the same argument to the third row shows that (c,d)(c,d) is also a permutation of (ω,ω2)(\omega,\omega^2). Orthogonality with the second row forces the opposite ordering, so

H′=(1111ωω21ω2ω)=3F,(49)\mathbf{H}' = \begin{pmatrix} 1 & 1 & 1 \\ 1 & \omega& \omega^2 \\ 1 & \omega^2 & \omega \end{pmatrix} = \sqrt{3}\mathbf{F}, \tag*{(49)}

up to row or column permutations associated with the DFT sign convention. Undoing the dephasing proves the factorization

M=DoutPoutFPinDin,(50)\mathbf{M}=\mathbf{D}_{\mathrm{out}}\mathbf{P}_{\mathrm{out}}\mathbf{F}\mathbf{P}_{\mathrm{in}}\mathbf{D}_{\mathrm{in}}, \tag*{(50)}

where the D\mathbf{D} matrices are diagonal unitary and the P\mathbf{P} matrices are permutations.

By Theorem 1, the separate fixed phase matrix D\mathbf{D} in Eq. (14) does not change the reachable set, so it is sufficient to consider TM,I(R)=MTRM\mathbf{T}_{\mathbf{M},\mathbf{I}}(\mathbf{R})=\mathbf{M}^{\mathsf{T}}\mathbf{R}\mathbf{M}. Substituting Eq. (50) gives

MTRM=DinTPinTFTR′FPinDin,(51)\mathbf{M}^{\mathsf{T}}\mathbf{R}\mathbf{M} = \mathbf{D}_{\mathrm{in}}^{\mathsf{T}}\mathbf{P}_{\mathrm{in}}^{\mathsf{T}}\mathbf{F}^{\mathsf{T}}\mathbf{R}'\mathbf{F}\mathbf{P}_{\mathrm{in}}\mathbf{D}_{\mathrm{in}}, \tag*{(51)}

where

R′=PoutTDoutTRDoutPout.(52)\mathbf{R}'=\mathbf{P}_{\mathrm{out}}^{\mathsf{T}}\mathbf{D}_{\mathrm{out}}^{\mathsf{T}}\mathbf{R}\mathbf{D}_{\mathrm{out}}\mathbf{P}_{\mathrm{out}}. \tag*{(52)}

The map R↦R′\mathbf{R}\mapsto\mathbf{R}' is again a bijection on diagonal unitaries: diagonal phases are shifted, and the diagonal entries are reordered.

For a fixed physical input nn, there exist a relabeled input index n′n' and a phase eiγne^{i\gamma_n} such that

PinDinen=eiγnen′.(53)\mathbf{P}_{\mathrm{in}}\mathbf{D}_{\mathrm{in}}\mathbf{e}_n=e^{i\gamma_n}\mathbf{e}_{n'}. \tag*{(53)}

The factor eiγne^{i\gamma_n} is a common field phase. On the output side, DinT\mathbf{D}_{\mathrm{in}}^{\mathsf{T}} changes only the individual output phases, and PinT\mathbf{P}_{\mathrm{in}}^{\mathsf{T}} applies a fixed permutation to the output coordinates. Neither operation changes the set of output powers, apart from that permutation. Finally, as R\mathbf{R} ranges over every diagonal unitary, so does R′\mathbf{R}'. Thus Eq. (51) establishes Eq. (46).

Analytic three-port region and the balanced-splitting limit

The common reachable region of Theorem 3 can be written explicitly. It is sufficient to use the DFT, the zero-based matrix index n=0n=0 (physical input port 1), and no fixed phases. The preceding theorems then extend the result to every input, every fixed diagonal phase matrix, and every equal-split 3×33 \times3 unitary MMI. A common reflector phase is physically irrelevant, so write the three effective arm phases as

(u0,u1,u2)=(1,eiα,eiβ),α,β∈[0,2π).(54)(u_0,u_1,u_2) = (1,e^{i\alpha},e^{i\beta}), \qquad\alpha,\beta\in[0,2\pi). \tag*{(54)}

With ω=ei2π/3\omega=e^{i2\pi/3}, the three output amplitudes are

A1(α,β)=1+eiα+eiβ3,A2(α,β)=1+ωeiα+ω2eiβ3,A3(α,β)=1+ω2eiα+ωeiβ3.(55)\begin{aligned} A_1(\alpha,\beta) &= \frac{1+e^{i\alpha}+e^{i\beta}}{3},\\ A_2(\alpha,\beta) &= \frac{1+\omega e^{i\alpha}+\omega^2 e^{i\beta}}{3},\\ A_3(\alpha,\beta) &= \frac{1+\omega^2 e^{i\alpha}+\omega e^{i\beta}}{3}. \tag*{(55)} \end{aligned}

Thus, Pk=∣Ak∣2P_k=|A_k|^2 and the explicit two-phase power map is

P1=19[3+2cos⁡α+2cos⁡β+2cos⁡(α−β)],P2=19[3+2cos⁡(α+2π3)+2cos⁡(β+4π3)+2cos⁡(α−β−2π3)],P3=19[3+2cos⁡(α+4π3)+2cos⁡(β+2π3)+2cos⁡(α−β+2π3)].(56)\begin{aligned} P_1 &= \frac{1}{9}\left[3+2\cos\alpha+2\cos\beta+2\cos(\alpha-\beta)\right],\\ P_2 &= \frac{1}{9}\left[3+2\cos\left(\alpha+\frac{2\pi}{3}\right)+2\cos\left(\beta+\frac{4\pi}{3}\right)+2\cos\left(\alpha-\beta-\frac{2\pi}{3}\right)\right],\\ P_3 &= \frac{1}{9}\left[3+2\cos\left(\alpha+\frac{4\pi}{3}\right)+2\cos\left(\beta+\frac{2\pi}{3}\right)+2\cos\left(\alpha-\beta+\frac{2\pi}{3}\right)\right]. \tag*{(56)} \end{aligned}

Unitarity gives P1+P2+P3=1P_1+P_2+P_3=1, so the image lies in the two-dimensional probability simplex shown in Supplementary Fig. S5.

We next determine the boundary of this image. Since the phase torus has no boundary, the inverse function theorem implies that a boundary point of the image must be the image of a critical point of the map (α,β)↦(P1,P2)(\alpha,\beta)\mapsto(P_1,P_2). Direct differentiation of Eq. (56) gives

det⁡[∂(P1,P2)∂(α,β)]=−2381[sin⁡(α−2β)+sin⁡(α+β)−sin⁡(2α−β)].(57)\det\left[\frac{\partial(P_1,P_2)}{\partial(\alpha,\beta)}\right] = -\frac{2\sqrt{3}}{81}\left[\sin(\alpha-2\beta)+\sin(\alpha+\beta)-\sin(2\alpha-\beta)\right]. \tag*{(57)}

The trigonometric factor has the useful factorization

sin⁡(α−2β)+sin⁡(α+β)−sin⁡(2α−β)=2sin⁡(α−β2)[cos⁡(3β2)−cos⁡(α−β2)].(58)\begin{aligned} \sin(\alpha-2\beta)+\sin(\alpha+\beta)-\sin(2\alpha-\beta) &=2\sin\left(\alpha-\frac{\beta}{2}\right)\left[\cos\left(\frac{3\beta}{2}\right)-\cos\left(\alpha-\frac{\beta}{2}\right)\right]. \tag*{(58)} \end{aligned}

Modulo 2π2\pi, the critical set is therefore the union of the three line families

α=β2+kπ,α=2β+2kπ,α=−β+2kπ.(59)\alpha=\frac{\beta}{2}+k\pi,\qquad\alpha=2\beta+2k\pi,\qquad\alpha=-\beta+2k\pi. \tag*{(59)}
Image not extracted

Figure S5. Reachable output powers of the ideal three-port router. a, The lossless output powers lie on the conservation plane P1+P2+P3=1P_1 + P_2 + P_3 = 1; positivity restricts them to the triangular probability simplex, while phase-only control reaches the blue subset. b, Barycentric view of the simplex. The blue curve is the exact analytic boundary in Eq. (60). The orange diamonds mark the three cyclic balanced two-output optima, including (4/9,4/9,1/9)(4/9, 4/9, 1/9).

Substitution into Eq. (56) shows that the three families have the same image, differing only by cyclic reparametrization of the power coordinates. A single closed parametrization is

P∂(t)=19([1+2cos⁡t]2[1+2cos⁡(t−2π/3)]2[1+2cos⁡(t+2π/3)]2),0≤t<2π.(60)\mathbf{P}_{\partial}(t) = \frac{1}{9} \begin{pmatrix} [1 + 2\cos t]^2 \\ [1 + 2\cos(t - 2\pi/3)]^2 \\ [1 + 2\cos(t + 2\pi/3)]^2 \end{pmatrix}, \qquad0 \le t < 2\pi. \tag*{(60)}

Away from these critical lines, the Jacobian is non-zero, so each regular phase-domain component maps locally to an open subset of the simplex. The curve Eq. (60) is simple between its three routing cusps, and direct evaluation of any regular point (for example, a point near the equal-power center) selects its bounded side. Hence, the reachable set is exactly the closed three-arc region enclosed by Eq. (60), rather than the full simplex.

Theorem 4 (Sharp balanced two-output limit). For any reachable three-port power vector satisfying P1=P2P_1=P_2, one has

P3≥19,P1=P2≤49.(61)P_3 \ge\frac{1}{9}, \qquad P_1=P_2\le\frac{4}{9}. \tag*{(61)}

Both inequalities are attained simultaneously, so the optimum is (P1,P2,P3)=(4/9,4/9,1/9)(P_1,P_2,P_3)=(4/9,4/9,1/9). The two other choices of balanced output pair obey the corresponding cyclic statements. These bounds hold for every equal-split unitary 3×33\times3 MMI, every input, and arbitrary fixed reciprocal post-MMI phases.

Proof. Let a=A1a=A_1, b=A2b=A_2, and c=A3c=A_3. Direct substitution of Eq. (55), using 1+ω+ω2=01+\omega+\omega^2=0 and ∣uj∣=1|u_j|=1, gives the two field identities

∣a∣2+∣b∣2+∣c∣2=1,ab‾+bc‾+ca‾=0.(62)|a|^2+|b|^2+|c|^2=1, \qquad a\overline{b}+b\overline{c}+c\overline{a}=0. \tag*{(62)}

Suppose ∣a∣=∣b∣=s|a|=|b|=s. Rearranging the second identity and applying the triangle inequality gives

s2=∣ab‾∣=∣bc‾+ca‾∣≤∣b∣ ∣c∣+∣c∣ ∣a∣=2s∣c∣.(63)s^2=|a\overline{b}|=|b\overline{c}+c\overline{a}|\le|b|\,|c|+|c|\,|a|=2s|c|. \tag*{(63)}

If s=0s=0, the result is immediate. Otherwise, division by ss and squaring give s2≤4∣c∣2s^2\le4|c|^2. Combining this with 2s2+∣c∣2=12s^2+|c|^2=1 yields ∣c∣2≥1/9|c|^2\ge1/9 and s2≤4/9s^2\le4/9. The effective phase vector

(u0,u1,u2)=(1,ω,−ω2)(64)(u_0,u_1,u_2)=(1,\omega,-\omega^2) \tag*{(64)}

produces (∣a∣2,∣b∣2,∣c∣2)=(4/9,4/9,1/9)(|a|^2,|b|^2,|c|^2)=(4/9,4/9,1/9), proving sharpness for the DFT. Fixed-phase invariance and Theorem 3 transfer both the bound and the attaining state to arbitrary fixed phases and every equal-split 3×33\times3 unitary MMI.

The residual 1/9≃11.1%1/9\simeq11.1\% is unavoidable power in the undesired output when the other two outputs are required to be exactly balanced. It is therefore routing leakage (or crosstalk), not dissipative optical loss: the three powers still sum to unity.

Theorem 3 is stronger than Theorem 2, but it is special to dimensions in which all equal-split unitaries are DFT-equivalent. This reachable-region equivalence already fails at N=4N=4, as the following explicit example shows. Consider the two normalized equal-split unitary matrices

W4=12(111111−1−11−11−11−1−11),F4=12(11111i−1−i1−11−11−i−1i).(65)W_4=\frac{1}{2} \begin{pmatrix} 1&1&1&1\\ 1&1&-1&-1\\ 1&-1&1&-1\\ 1&-1&-1&1 \end{pmatrix}, \qquad F_4=\frac{1}{2} \begin{pmatrix} 1&1&1&1\\ 1&i&-1&-i\\ 1&-1&1&-1\\ 1&-i&-1&i \end{pmatrix}. \tag*{(65)}

For the zero-based matrix index n=0n=0 (physical input port 1), no fixed phases, and reflector vector r=(1,1,i,i)\mathbf{r}=(1,1,i,i), direct multiplication gives

W4Tdiag⁡(r)W4e0=(1+i2,1−i2,0,0)T,(66)W_4^{\mathsf{T}}\operatorname{diag}(\mathbf{r})W_4\mathbf{e}_0= \left(\frac{1+i}{2},\frac{1-i}{2},0,0\right)^{\mathsf{T}}, \tag*{(66)}

so the power vector (1/2,1/2,0,0)(1/2,1/2,0,0) is reachable with W4W_4.

It is not reachable with F4F_4. Indeed, suppose a unit-modulus reflector vector produced a field (a,b,0,0)T(a,b,0,0)^{\mathsf{T}} through F4F_4. Inverting the four-point DFT relation forces its four entries to be

a+b,a−ib,a−b,a+ib.(67)a+b,\qquad a-ib,\qquad a-b,\qquad a+ib. \tag*{(67)}

Their unit moduli imply, by comparing the first with the third and the second with the fourth,

abˉ+bˉa=0,abˉ−bˉa=0.(68)a\bar{b}+\bar{b}a=0,\qquad a\bar{b}-\bar{b}a=0. \tag*{(68)}

Hence abˉ=0a\bar{b}=0. This contradicts the required powers ∣a∣2=∣b∣2=1/2|a|^2=|b|^2=1/2. Thus even equal-split unitary MMIs can have different complete reachable power regions for N>3N>3. The fixed-phase invariance theorem shows that adding arbitrary fixed reciprocal post-MMI phases does not remove this difference. Theorem 2 nevertheless guarantees that every such region contains all NN routing vertices for every input port. Uniform loss across all arms reduces the total output power but leaves the normalized reachable region unchanged. Arm-to-arm loss imbalance deforms the region and generally removes the exact routing vertices, thereby limiting the achievable extinction ratio.

Supplementary Note 3: Device design, electrostatics and cavity-array characterization

The MMI is designed to split any input evenly among the output channels. Supplementary Figure S6 shows its port-numbering convention and geometrical parameters. We used the finite-difference eigenmode

Image not extracted

Figure S6. Sketch of the MMI Design. The coordinate system shown in the lower left corner corresponds to the crystal axes of the LiNbO3_3 membrane. The parameters of the device used in the main text are: h1=800h_1=800 nm, h2=2000h_2=2000 nm, h3=3150h_3=3150 nm, h4=15h_4=15 µm, w1=40w_1=40 µm, w2=435w_2=435 µm.

expansion solver in Ansys Lumerical to compute the scattering matrix between the input and output fundamental TE-like waveguide modes. The structure was optimized to reach ∣Mmn∣2=1/3|M_{mn}|^2=1/3 for every input mode nn and output mode mm. Supplementary Fig. S7 plots the calculated matrix elements corresponding to the optimized structure. The calculated transmission loss

Tloss,n=1−∑m=13∣Mmn∣2,(69)T_{\mathrm{loss},n}=1-\sum_{m=1}^{3}|M_{mn}|^2, \tag*{(69)}

is less than 2.2% for all input ports. The deviation from the ideal splitting ratio of 1/3 is less than 1% for all matrix elements.

The PhC cavity design follows Ref. [39] and uses band diagrams of infinite periodic waveguides with unit cells parametrized as shown in Supplementary Fig. S8(a). The structure is periodic in the yy-direction with lattice constant aa, and its band diagram gives the dispersion relation ω(ky)\omega(k_y) for the first few modes. We calculate these band diagrams using the open-source MIT Photonic Bands package [40]. The PhC cavity comprises four regions: 1) a taper of NtapN_{\mathrm{tap}} unit cells that converts the rib waveguide smoothly into a periodic structure; 2) a front mirror of Nm,LN_{m,L} identical unit cells that provides partial reflection at the target resonance frequency ω0\omega_0; 3) a cavity of 2Ncav2N_{\mathrm{cav}} unit cells with smoothly varied geometric parameters; and 4) a back mirror of Nm,RN_{\mathrm{m,R}} identical unit cells that ideally provides 100% reflection at ω0\omega_0. Supplementary Figure S8(b) shows how the parameters are varied in the taper section (unit cell 0 to

Image not extracted

Figure S7. Simulated Scattering Matrix of the MMI. a, Calculated intensity transmission, ∣Mmn∣2|M_{mn}|^2. b, Calculated transmission phase, arg⁡(Mmn)\arg(M_{mn}). The simulation-specific optimized parameters are h1=800h_1 = 800 nm, h2=2000h_2 = 2000 nm, h3=3200h_3 = 3200 nm, h4=15h_4 = 15 μ\mum, w1=40w_1 = 40 μ\mum and w2=435.4w_2 = 435.4 μ\mum; Supplementary Fig. S6 reports how the parameters are related to the geometrical structure of the MMI.

Ntap−1=31N_{\mathrm{tap}} - 1 = 31, left mirror section (unit cell 32 to Ntap+Nm,L−1=71N_{\mathrm{tap}} + N_{\mathrm{m,L}} - 1 = 71), and the left half of the cavity section (unit cell 72 to N_{\mathrm{tap}} + N_{\mathrm{m,L}} + N_{\mathrm{cav}} - 1 = 107). The full cavity section is symmetric around its center at unit cell 107, and the right mirror contains N_{\mathrm{m,R}} = 100$ unit cells.

Note that for the fabricated devices, we add an offset parameter, δoff\delta_{\mathrm{off}}, to the design parameters to compensate for over- or under-etching. This offset is added to the geometrical parameters in Supplementary Figure S8 as

Image not extracted

Figure S8. Design Parameters of the PhC Cavities. a, Sketch of a single unit cell of the PhC cavity indicating its geometrical parameters. b, Plot of the parameters as a function of unit cell number for the left part of the full cavity structure. The parameters plotted correspond to the device in the main text. The fixed parameters are: a=418a = 418 nm, w0=aw_0 = a, and w1=148w_1 = 148 nm. The lengths of the regions are: Ntap=32N_{\mathrm{tap}} = 32, Nm,L=40N_{\mathrm{m,L}} = 40, and Ncav=36N_{\mathrm{cav}} = 36. Note that the parameters plotted here differ from the design parameters by the offset parameter δoff=−16\delta_{\mathrm{off}} = -16 nm, as defined in Eq. (70).

hn→hn+δoffandwn→wn+2δoff.(70)h_n \to h_n + \delta_{\mathrm{off}} \qquad\text{and} \qquad w_n \to w_n + 2\delta_{\mathrm{off}}. \tag*{(70)}

Supplementary Figure S8 shows values for the device used in the main text for which δoff=−16\delta_{\mathrm{off}} = -16 nm. The nominal thickness of the LiNbO3_3 membrane is hmem=600h_{\mathrm{mem}} = 600 nm, which is etched by hwg=400h_{\mathrm{wg}} = 400 nm to form the photonic structures. The nominal exterior sidewall angle of the PhC cavity (Figure 2a) is 55∘55^\circ from the substrate plane (35∘35^\circ from vertical). The sidewalls inside the narrow PhC Bragg-mirror gaps are steeper. Following Supplementary Section V of Ref. [21], which reports an inner-gap angle of approximately 17∘17^\circ from vertical (73∘73^\circ from the substrate plane) based on SEM imaging and transmission modeling, our PhC design model uses 73∘73^\circ from the substrate plane for these internal sidewalls. The design assumes this angle on all sidewalls, since the inner-gap angle is the most significant to the PhC design.

Inside the cavity section, the parameters h1h_1, h2h_2, and w2w_2 are varied to create a linearly decreasing mirror strength, as shown in Supplementary Figure S9. The mirror strength is defined in terms of frequencies at the band edge of the band diagram, ky=π/ak_y = \pi/a, as [39]

Image not extracted

Figure S9. Calculated Mirror Strength and Quality Factors of the Single-Sided PhC Cavity. a, Calculated mirror strength, MSMS, from Eq. (71) as a function of unit cell number in the left half of the cavity section. b, Calculated quality factors as a function of the length of the left mirror section, Nm,LN_{m,L}. The parameters used for the simulations are the same as in Supplementary Figure S8, except for δoff\delta_{\mathrm{off}}, which is zero here. The length of the right mirror section is Nm,R=100N_{m,R} = 100.

MS=(ωair−ωdiel.)2(ωair+ωdiel.)2−(ω0−ωcen)2ωcen2,(71)MS = \sqrt{\frac{(\omega_{\mathrm{air}}-\omega_{\mathrm{diel.}})^2}{(\omega_{\mathrm{air}}+\omega_{\mathrm{diel.}})^2} - \frac{(\omega_0-\omega_{\mathrm{cen}})^2}{\omega_{\mathrm{cen}}^2}}, \tag*{(71)}

where ωair\omega_{\mathrm{air}}, ωdiel.\omega_{\mathrm{diel.}}, and ω0\omega_0 are the frequencies of the air band, dielectric band, and desired resonance, while ωcen=(ωair+ωdiel.)/2\omega_{\mathrm{cen}} = (\omega_{\mathrm{air}} + \omega_{\mathrm{diel.}})/2. For designs that are divided into sections, it is straightforward to vary the extrinsic coupling rate, κe\kappa_e, with a minimal effect on the intrinsic loss rate, κi\kappa_i. Supplementary Figure S9(b) shows the calculated quality factors (defined as Qe=ω0/κeQ_e = \omega_0/\kappa_e and Qi=ω0/κiQ_i = \omega_0/\kappa_i, respectively) as a function of the number of unit cells in the left mirror section, Nm,LN_{m,L}. The extrinsic coupling QQ increases exponentially with Nm,LN_{m,L}, while the intrinsic loss rate is roughly independent of Nm,LN_{m,L}. The simulations in

Supplementary Figure S9 were performed using Ansys Lumerical’s Finite-Difference Time-Domain solver. A cross-section of the field profile of the resonance corresponding to the geometry with Nm,L=40N_{m,L} = 40 in Supplementary Figure S9 is plotted in Supplementary Figure S10 with the structure overlaid. The colored contour corresponds to the real part of the yy-component, normalized as

Image not extracted

Figure S10. Simulated Mode Profile of the Single-Sided PhC Cavity. The plot shows the simulated real part of the electric field normalized as in Eq. (72). The black curve shows the cavity structure at the same cross-section, x=(hmem−hwg)+hwg/2x=(h_{\mathrm{mem}}-h_{\mathrm{wg}})+h_{\mathrm{wg}}/2. The four regions of the cavity are indicated with Ntap=32N_{\mathrm{tap}}=32, Nm,L=40N_{\mathrm{m,L}}=40, Ncav=36N_{\mathrm{cav}}=36, and Nm,R=100N_{\mathrm{m,R}}=100. The geometrical parameters are the same as in Supplementary Fig. (S9b).

Ey(y,z)=Ey(y,z)λ0∫∣E(y,z)∣2dydz(72)\mathcal{E}_y(y,z)=E_y(y,z)\frac{\lambda_0}{\sqrt{\int|E(y,z)|^2\mathrm{d}y\mathrm{d}z}} \tag*{(72)}

where λ0=1549.2\lambda_0=1549.2 nm is the corresponding resonance wavelength. The field is observed to be concentrated

in the cavity region, with some leakage to the left but none to the right, as required for a single-sided cavity. Assuming a uniform field V/gV/g and unit optical overlap, the Pockels effect gives ∣dν/dV∣≃ν0nr33/(2g)|\mathrm{d}\nu/\mathrm{d}V|\simeq\nu_0 n r_{33}/(2g). At a wavelength of 1542 nm, using ne=2.14n_e=2.14, r33=30.9r_{33}=30.9 pm/V and an electrode gap g=6.8g=6.8 μ\mum gives approximately 2.0 GHz/V. An approximate overlap calculation using local periodic cells of the patterned cavity and the nonuniform electrostatic field gives approximately 0.94 GHz/V. These estimates support the order of magnitude of the measured quasi-static tuning rate.

Electrode capacitance and switching energy

We estimated the terminal capacitance of the phase-shifter electrode pair with an electro-quasistatic finite-volume solution of ∇⋅(ϵ∇ϕ)=0\nabla\cdot(\epsilon\nabla\phi)=0. The three-dimensional model uses the exact 24-vertex electrode footprint extracted from the GDS, including the taper and end fringing, together with a continuous LiNbO3_3 rib and slab, the SiO2_2 claddings, the dielectric and electrically uncontacted silicon handle, air, and 600 nm thick gold electrodes. We use ϵr=3.9\epsilon_r=3.9, 11.7, and 1 for SiO2_2, Si, and air, respectively, and the anisotropic LiNbO3_3 values ϵr,∥=28.7\epsilon_{r,\parallel}=28.7 and ϵr,⊥=43.0\epsilon_{r,\perp}=43.0. Here, xx, yy and zz denote the longitudinal, transverse and vertical directions, respectively, corresponding to the crystal axes yy, zz and xx. The far-field boundaries are fixed at the balanced pair’s 0.5-V common-mode potential and placed at x=±1x=\pm1 mm, y=1y=1 mm from the symmetry plane, and z=±0.5z=\pm0.5 mm; doubling all of these distances changes the raw capacitance by only 0.096%. Calibrating the three-dimensional mesh against a converged two-dimensional straight-section calculation gives C=18.6±0.3C=18.6\pm0.3 fF for the complete electrode pair. The two-dimensional cross-section gives 0.1414 fF/μ\mum. Applying this value over the full 116.70-μ\mum GDS electrode span gives 16.5 fF; the remaining approximately 2.1 fF in the full three-dimensional result arises from the tapered and widened regions and end fringing. We therefore use the three-dimensional terminal capacitance for the switching-energy estimates. The electrode geometry and transverse field are shown in Supplementary Fig. S11, and the full three-dimensional field in Supplementary Fig. S12. For a conventional non-energy-recovering capacitive driver, the energy dissipated in one transition of one phase shifter is Etr=C(ΔV)2/2E_{\mathrm{tr}}=C(\Delta V)^2/2, corresponding to 232.5 ±\pm 3.8 fJ for ΔV=5\Delta V = 5 V. For a closed cycle among the three demonstrated routing states, the energy is

Ecyc=C2∑s∑j=13(Vj,s+1−Vj,s)2,(73)E_{\mathrm{cyc}} = \frac{C}{2} \sum_s \sum_{j=1}^{3} \left(V_{j,s+1} - V_{j,s}\right)^2, \tag*{(73)}

where the cyclic sum includes the last-to-first transition. We evaluate this expression using the measured AWG excursions listed in Methods and the closed routing-state sequence P1→P3→P2→P1P_1 \to P_3 \to P_2 \to P_1; the DC-balancing idle states are excluded. The common static biases cancel from every voltage difference. The result is 2.096 ±\pm 0.034 pJ per closed routing cycle across all three phase shifters. The contributions from the V1V_1, V2V_2, and V3V_3 electrodes are 0.429 ±\pm 0.007, 0.795 ±\pm 0.013, and 0.872 ±\pm 0.014 pJ per cycle, respectively. The average over the three routing transitions is 0.699 ±\pm 0.011 pJ per transition. Repeating this closed three-state cycle at 10 MHz corresponds to 21.0 ±\pm 0.3 μ\muW. These values describe capacitive driver dissipation and exclude driver quiescent power. Their quoted uncertainties propagate the numerical capacitance uncertainty; they do not include an independent voltage-calibration uncertainty.

The Pockels phase shifter is intrinsically capacitive and therefore consumes no static power in the ideal dielectric limit. No device leakage was resolved above the approximately 10-GΩ\Omega probe background, giving setup-limited upper bounds of ∣I∣<0.5\lvert I\rvert< 0.5 nA and Pstatic<2.5P_{\mathrm{static}} < 2.5 nW per phase shifter at 5 V. Establishing or changing a DC bias requires the one-time capacitive charging energy C(ΔV)2/2C(\Delta V)^2/2; maintaining that bias does not produce continuous intrinsic dissipation. Any continuous power is instead bounded by leakage and by driver quiescent power, which is not included here. This 10-GHz electro-quasistatic reference uses real permittivities and does not include dielectric loss or finite silicon conductivity.

Image not extracted

Figure S11. Electrostatic field of the phase-shifter electrode pair. a, Top view of the exact GDS electrode geometry, rotated so that the two electrodes appear on the left and right. The dashed line marks the transverse section used in panel b. b, Transverse field distribution through the active electrode gap. Color gives the electric field magnitude for a 1-V differential bias on the shared logarithmic scale, and arrows indicate the field direction. Black outlines mark the LiNbO3_3 boundary. The modeled waveguide comprises a 200-nm slab and a 400-nm rib with a 550-nm top width and 55° exterior sidewalls; the electrode gap is 6.8 μ\mum.

Image not extracted

Figure S12. Three-dimensional electrostatic field of the electrode pair. Field lines are integrated through the full three-dimensional potential solution and arrows point from the +1-V electrode toward the grounded electrode. The exact GDS electrode footprints are shown in muted red and blue. The pale-gold region shows the continuous 200-nm LiNbO3_3 slab and 400-nm trapezoidal rib, the pale-blue transparent volumes show the upper and buried SiO2_2 claddings, and the gray volume shows the dielectric silicon handle. The upper cladding terminates at the SiO2_2–air interface. The vertical coordinate is displayed with a fivefold exaggeration for clarity.

Cavity matching across complete devices

The cavity-matching analysis used the two-sided cavity reference and two complementary MMI transmission spectra for 60 complete three-cavity devices spanning the full 4×5×34 \times5 \times3 design set (four PhC types, five cavity offsets, and three left-mirror counts). The external MEMS fiber-switch matrix selected the measured optical input–output configurations, while the on-chip cavity electrodes were left unbiased. A reference-cavity resonance was assignable in 50 of the 60 spectra and, together with the corresponding design-family trend, guided identification of the three switch-cavity resonances in the MMI transmission spectra. After visual review, 50 devices contained an assignable cavity group; we refer to these as usable devices below: 34 showed three resolved resonances, and 16 showed two resolved resonances, consistent with two cavities overlapping within the spectral resolution. For each usable device, we define the observed resonance span as

Δλ=max⁡j(λj)−min⁡j(λj),(74)\Delta\lambda= \max_j(\lambda_j) - \min_j(\lambda_j), \tag*{(74)}

where λj\lambda_j are the resolved resonances in the two MMI transmission states. For devices with only two resolved resonances, Δλ\Delta\lambda assumes that the unresolved third resonance overlaps one of the observed features; the three-resonance subset is therefore shown separately in Supplementary Fig. S13. Across all 50 usable devices, the median observed span is 0.63 nm, the interquartile range is 0.44–0.91 nm, and 82% have Δλ<1\Delta\lambda< 1 nm.

For symmetric tuning of the extreme resonances toward their frequency midpoint, the required maximum voltage magnitude is estimated as

∣V∣max⁡=Δν2ηDC,Δν=c∣1λmin⁡−1λmax⁡∣(75)|V|_{\max} = \frac{\Delta\nu}{2\eta_{\mathrm{DC}}}, \qquad\Delta\nu= c\left|\frac{1}{\lambda_{\min}}-\frac{1}{\lambda_{\max}}\right| \tag*{(75)}

using the measured tuning efficiency ηDC=1.92\eta_{\mathrm{DC}} = 1.92 GHz V−1^{-1}. The resulting fractions alignable within ±10\pm10, ±20\pm20, ±30\pm30, and ±40\pm40 V are 10%, 48%, 76%, and 98%, respectively. The estimated ∣V∣max⁡|V|_{\max} has a median of 20.4 V and an interquartile range of 14.3–29.3 V. Thus, the approximately 30-V value corresponds to the upper quartile of the observed distribution, rather than a voltage sufficient for every device. In particular, ±30\pm30 V aligns 38 of 50 usable devices (76%; 95% Wilson interval, 63–86%). For the stricter subset with three resolved resonances, the median is 22.5 V, the interquartile range is 16.5–32.4 V, and 23 of 34 devices are alignable within ±30\pm30 V (68%; 95% Wilson interval, 51–81%). The ±0.06\pm0.06 GHz V−1^{-1} uncertainty of the measured tuning efficiency gives a 3.1% relative voltage-scale uncertainty (0.64 V at the all-device median) and changes either ±30\pm30-V device count by at most one. These estimates do not include device-to-device variation in tuning efficiency or uncertainty in identifying overlapping resonances; the two-resonance cases assume that the unresolved third resonance overlaps one of the observed features.

Image not extracted

Figure S13. Measured cavity matching and estimated alignment-voltage yield. a, Empirical cumulative distribution of the observed wavelength span Δλ\Delta\lambda for all 50 usable complete devices (solid) and the 34 devices with three resolved resonances (dashed). The vertical line marks the all-device median of 0.63 nm. b, Fraction of devices whose extreme resolved resonances can be tuned to their frequency midpoint within the indicated symmetric voltage magnitude, using ηDC=1.92\eta_{\mathrm{DC}} = 1.92 GHz V−1^{-1}.

The reviewed spectra also allow the measured resonance positions to be compared with the cavity design parameters. Within each PhC type, the mean resolved resonance wavelength follows the lithographic cavity offset (Supplementary Fig. S14a). The single-sided cavities used for the switching demonstration in device A46 belong to type 3, with (a,Ncav)=(418 nm,36)(a,N_{\mathrm{cav}}) = (418\ \mathrm{nm},36). For the 34 devices with three resolved resonances, type 0 has the smallest median span, although the unequal sample counts, particularly for type 1, make this comparison descriptive rather than conclusive. The usable-spectrum yield also varies between the four design types (Supplementary Fig. S14c).

Independent fits to the reference-cavity spectra show that the loaded quality factor increases mainly with the number of left-mirror cells, while no clear dependence on the lithographic cavity offset is resolved (Supplementary Fig. S15). This separation is consistent with the mirror count controlling the external coupling and the offset primarily selecting the resonance wavelength.

Image not extracted

Figure S14. Measured cavity design trends across complete devices. a, Mean wavelength of the resolved MMI-cavity features versus the designed cavity offset for all 50 usable devices. Lines are linear fits within each PhC type. The four types use (a,Ncav)=(424 nm,30),(420 nm,30),(422 nm,36),(a, N_{\mathrm{cav}}) = (424\ \mathrm{nm}, 30), (420\ \mathrm{nm}, 30), (422\ \mathrm{nm}, 36), and (418 nm,36)(418\ \mathrm{nm}, 36), respectively. b, Max–min wavelength span for the 34 devices with three resolved resonances; horizontal bars indicate medians. The accepted devices with two resolved resonances contribute to panel a but are excluded from panel b. c, Numbers of usable and rejected spectra for each PhC type. A spectrum was rejected when the cavity group could not be assigned confidently.

Image not extracted

Figure S15. Two-sided reference-cavity loaded quality-factor trends. Loaded quality factors extracted from 44 two-sided reference-cavity spectra acquired without detector saturation and whose line-shape fits have R2≥0.9R^2 \ge0.9. a, Loaded QQ versus the number of cells in the left mirror. b, Loaded QQ versus the designed cavity offset. Colors identify the four PhC types; horizontal bars show the median at each design value.

For a symmetric two-sided cavity driven from one side, the field reflection coefficient is

rtwo−sided(Δ)=iΔ+κi/2iΔ+κe+κi/2,(76)r_{\mathrm{two-sided}}(\Delta) = \frac{i\Delta+ \kappa_i/2}{i\Delta+ \kappa_e + \kappa_i/2}, \tag*{(76)}

where Δ=ω0−ω\Delta= \omega_0 - \omega, κe\kappa_e is the external coupling rate through each mirror, and κi\kappa_i is the total intrinsic loss rate. The loaded quality factor is therefore Q=ω0/(2κe+κi)Q = \omega_0/(2\kappa_e + \kappa_i).

The full-device model gives κe/κi=1.09\kappa_e/\kappa_i = 1.09–1.361.36 for the one-sided switch cavities. Taking this range as an estimate of the coupling through each 40-period mirror and assuming comparable intrinsic loss for the two-sided reference cavity gives Qi=Q(1+2κe/κi)=(1.05Q_i = Q(1 + 2\kappa_e/\kappa_i) = (1.05–1.23)×1051.23) \times10^5. This range is a model-based estimate rather than an independent fit of the reference-cavity intrinsic loss.

Image not extracted

Figure S16. Full broadband spectrum of the two-sided reference cavity. a, Reflection and transmission channels measured over the complete 1480–1640 nm laser sweep for device A46. Because the two detector channels were not absolutely calibrated relative to one another, the transmission trace is slightly offset for visual alignment with Figure 3a; this plotting offset has no physical significance. The shaded interval is the wavelength range displayed there. b, Sum of the simultaneously measured reflection and transmission powers and its rolling 90th-percentile upper envelope. The dashed trace is the calibrated transmission of a plain-grating reference measured in a separate 1480–1640 nm sweep. It is shown in relative decibels and discreetly offset to match the envelope on its long-wavelength shoulder, providing a qualitative comparison of spectral shape.

The complete broadband measurement is shown in Supplementary Fig. S16. Summing the simultaneously measured reflection and transmission channels suppresses much of the resonant redistribution between the two ports. Its smooth upper envelope therefore tracks the wavelength-dependent response shared by the optical path, including the grating couplers, but it is not an independent grating-coupler calibration and is not used to infer an absolute coupling loss. For qualitative comparison, a separately measured plain-grating test spectrum is vertically matched to the long-wavelength shoulder of this envelope in the figure.

The fit quality for the representative two-sided reference cavity in device A46 used in Figure 3a is shown in Supplementary Fig. S17. The transmission peak and reflection dip are fitted simultaneously with a shared Lorentzian center and linewidth, independent amplitudes, and independent linear backgrounds.

Finally, after subtracting a robust type-specific linear dependence on cavity offset, the residual mean wavelength has no resolved gradient across the chip (Supplementary Fig. S18). A planar fit to the non-outlier devices gives gradients of +0.07+0.07 and +0.05+0.05 nm mm−1^{-1} along xx and yy, respectively, with R2=0.04R^2 = 0.04. Device A48 is the only residual beyond three robust standard deviations.

Image not extracted

Figure S17. a, Normalized transmission (black) and reflection (light gray) measured around the resonance shown in Figure 3a. Dashed curves are the simultaneous Lorentzian fits. The shared fit gives λ0=1541.478\lambda_0 = 1541.478 nm, a full width at half maximum of 46.7 pm, and a loaded quality factor Q=3.30×104Q = 3.30 \times10^{4}; the coefficients of determination are RT2=0.9945R_T^2 = 0.9945 and RR2=0.9913R_R^2 = 0.9913. b, Data-minus-fit residuals for the two traces.

Image not extracted

Figure S18. Residual mean wavelength at each accepted device position after removing a robust, type-specific linear dependence on the designed cavity offset. The color scale is limited to ±2 nm to show the spatial variation among the non-outlier devices; the outlined and labeled point A48 is the sole residual beyond three robust standard deviations. Excluding A48, a planar fit gives gradients of +0.07 and +0.05 nm mm−1^{-1} along xx and yy, respectively, with R2=0.04R^2 = 0.04.

Supplementary Note 4: Cavity response, insertion loss and routing-performance analysis

The calibrated transfer-matrix model determines the voltage-dependent reflection coefficient rj(Vj)r_j(V_j) of each cavity from Eq. (1). Supplementary Fig. S19 shows the corresponding spectra at the three operating points obtained by maximizing the extinction ratio for routing to P1P_1, P2P_2, and P3P_3. The three resonances move relative to the fixed laser wavelength and thereby set the phases that interfere constructively at the selected output after the return pass through the MMI. To convert the normalized detunings constrained by the voltage cube to wavelength, we set the loaded quality factor to the independently extracted measured value QL=5.07×104Q_\mathrm{L}=5.07\times10^4 for all three nominally identical switch cavities. This value was independently extracted from a Fano fit to the reflection spectrum of one cavity measured in isolation on the same fabricated router device. The resonance offsets and tuning coefficients are rescaled together so that the modeled response at the fixed laser wavelength is unchanged. The figure also shows the continuous cavity reflection phases and the output powers obtained by coherently recombining the reflected fields using the calibrated reciprocal transformation MTRM\mathbf{M}^{\mathsf{T}}\mathbf{R}\mathbf{M}, including the fitted arm phases and coherent background. The wavelength positions are therefore conditional on applying the measured quality factor to all three cavities.

Using the fitted voltage linewidths (4.77–4.86 V) and external-to-internal coupling ratios (1.09–1.36), Eq. (1) predicts cavity phase excursions of 1.67π1.67\pi, 1.69π1.69\pi and 1.70π1.70\pi for cavities 1, 2 and 3, respectively, for a ±5\pm5 V sweep (10 V peak-to-peak) centered on the bias that brings each resonance to the laser wavelength. These correspond to 83.5%, 84.4% and 85.2% of 2π2\pi. These are model estimates of the cavity reflection phase alone, assuming constant tuning efficiency over the sweep, rather than direct phase measurements at the routing biases. A full 2π2\pi winding is approached asymptotically across resonance; it is not established by the finite drive range.

At one probe wavelength, the transformation QL→sQLQ_\mathrm{L}\to sQ_\mathrm{L}, η→η/s\eta\to\eta/s and Δν0→Δν0/s\Delta\nu_0\to\Delta\nu_0/s leaves the modeled voltage-cube response unchanged. The reciprocal fit therefore determines voltage linewidths of 4.78, 4.86 and 4.77 V for C1C_1, C2C_2 and C3C_3, respectively, but it does not independently determine QLQ_\mathrm{L}, η\eta or the absolute resonance offsets in wavelength. The separate QLQ_\mathrm{L} and tuning coefficients used in the numerical parameterization are consequently not interpreted as independent measurements. Experimental tuning claims use the direct spectroscopic value in Figure 3b.

Supplementary Table 2 Single-sided switch-cavity parameters constrained by the reciprocal three-cavity voltage-cube fit and the common Fano-QQ wavelength reconstruction. Here κe/κi\kappa_\mathrm{e}/\kappa_\mathrm{i} and Vlinewidth=ν0/(QLη)V_\mathrm{linewidth}=\nu_0/(Q_\mathrm{L}\eta). The last three columns give the reconstructed resonance wavelengths at the states routing to P1P_1, P2P_2 and P3P_3. The coupling ratios are model-conditioned point estimates because they can trade against coherent background and detector calibration. Statistical uncertainties are not quoted because the saved regularized fit does not contain a covariance matrix or a bootstrap ensemble.

Cavityκe/κi\kappa_\mathrm{e}/\kappa_\mathrm{i}VlinewidthV_\mathrm{linewidth} (V)λ(P1)\lambda(P_1) (nm)λ(P2)\lambda(P_2) (nm)λ(P3)\lambda(P_3) (nm)
C1C_11.3574.781541.64181541.67401541.6702
C2C_21.1684.861541.68381541.64731541.6916
C3C_31.0894.771541.67181541.67801541.6316

Table 2.

The A46 device uses uniform TE focusing grating couplers comprising 28 concentric circular teeth. The gratings have a constant radial period of 978 nm and a drawn tooth width of 166 nm, corresponding to a 17.0% duty cycle and an 812-nm gap. The innermost and outermost tooth centerlines have radii of 55.9 and 82.3 μm, respectively. Each grating is connected through a 55-μm-long taper that expands to a 17-μm aperture, corresponding to a 17.6° full in-plane opening angle.

Measured insertion-loss budget

We compared a broadband scan of device A46 with a plain grating-reference path acquired in the same measurement session (Supplementary Figure S20). The reference and physical output 1 were recorded on the same calibrated detector channel. The other two outputs were placed on the output-1 scale using the independently established detector calibration. All quoted values below are obtained after a 101-point median filter, corresponding to 0.101 nm on the 1-μm wavelength grid.

Image not extracted

Figure S20. Calibration-corrected insertion-loss budget for device A46. a, Broadband grating-reference power, the three device outputs, and their corrected sum. Red, green and blue denote P1P_1, P2P_2 and P3P_3, respectively. b, Detail around the 1541.66-nm switching wavelength. Output 2 is placed on the output-1 scale using an independently determined detector calibration; output 1 and the reference use the same detector channel. The 3.65-dB summed-output excess is relative to the complete grating-reference path and therefore includes cavity dissipation, two MMI passages, routing waveguides, and any residual path mismatch.

Image not extracted

Figure S19. Fano-Q-constrained cavity spectra, phases and output powers at the three switching states. a–c, Cavity reflectance ∣rj∣2|r_j|^2 versus wavelength at the calibrated voltage tuples that route light to P1P_1, P2P_2, and P3P_3, respectively. The dashed vertical line marks the fixed laser wavelength, 1541.660 nm. Triangles mark the reconstructed resonance positions, and horizontal bars show their loaded full widths at half maximum. A common QL=5.07×104Q_\mathrm{L} = 5.07 \times10^4, independently extracted from a Fano fit to one cavity measured in isolation on the same fabricated router device, gives linewidths of 30.4 pm. The same value is applied to all three nominally identical cavities; the absolute wavelength positions are conditional on this assumption. d–f, Corresponding continuous reflection phases ϕj/π\phi_j/\pi; only relative phases affect the recombined powers. g–i, Output powers after coherent recombination in the calibrated model, including the fitted arm phases, voltage-dependent phase correction and coherent background, before detector gain and offset calibration. Red, green and blue denote P1P_1, P2P_2 and P3P_3, respectively; the dashed black curve is their sum. Powers are in model units and do not represent absolute insertion-loss measurements. Columns correspond to routing to P1P_1, P2P_2 and P3P_3, respectively.

At the fixed switching wavelength of 1541.66 nm, the grating-reference power is −13.63 dBm, the physical output-1 power is −20.28 dBm, and the sum of the three calibration-corrected device outputs is −17.27 dBm. Thus, output 1 and the summed output exhibit 6.65 dB and 3.65 dB of excess loss, respectively, relative to the reference path. The corrected output fractions in this scan are 50.1%, 11.4%, and 38.6%. An optimized route-to-P1P_1 state would instead concentrate most of the surviving power in P1P_1; the fitted model predicts 94.7%. This scan therefore does not represent the optimized route-to-P1P_1 state.

Its 6.65 dB reference-to-output-1 difference describes this particular output distribution, rather than the insertion loss of an optimized routing state. The summed-output excess is used below to estimate the device-loss budget.

The reference value is itself a path measurement rather than a pure grating-coupler calibration: it includes two grating interactions, the fiber network, switches and connectors, polarization mismatch, and reference waveguide propagation. If the independently verified 0-dBm launch setting also applied to this acquisition, the measured powers correspond to total launch-to-detector losses of 13.63 dB for the reference, 17.27 dB for the summed device outputs, and 20.28 dB for output 1.

The fiber path from the source to the fiber array was independently measured to have 1.39 dB loss in the output-1 configuration used here. Subtracting this known single-sided external loss from the 13.63-dB reference-path loss leaves 12.24 dB for the two grating couplers, reference waveguide, and collection path, corresponding to a conservative upper estimate of 6.12 dB per grating if the remaining losses are neglected. If the collection-side external loss is comparable to the measured launch-path loss, the inferred value is approximately 5.43 dB per grating, before correcting for the reference-waveguide loss. We therefore estimate a grating-coupler loss of approximately 5.4–6.1 dB per coupler.

For comparison, the fitted route-to-P1P_1 model gives cavity power survivals of 0.600, 0.713, and 0.378 at 1541.66 nm. Weighting these by the fitted MMI arm amplitudes gives approximately 2.33 dB of cavity dissipation, and the fitted output distribution places 94.7% of the surviving power in P1P_1. Subtracting this model-conditioned cavity contribution from the measured 3.65-dB summed-output excess gives 1.32 dB for two MMI passages, routing waveguides, and residual path mismatch. Independently, the best 1-nm-averaged off-resonant summed-output ratio gives 1.57 dB near 1526.13 nm. Together, these measurements indicate a passive two-pass residual of approximately 1.3–1.6 dB. If assigned entirely and symmetrically to the MMI, this would bound each MMI passage by approximately 0.7–0.8 dB; the actual MMI loss is lower to the extent that waveguide loss or path mismatch contributes.

Ideal sensitivity to arm-loss imbalance

To isolate the effect of differential loss, consider an ideal 3×33 \times3 Fourier MMI with return coefficients rjr_j and returned arm-power fractions Rj=∣rj∣2R_j = |r_j|^2. For input power PinP_{\mathrm{in}}, the MMI sends Pin/3P_{\mathrm{in}}/3 into each arm, so arm jj returns PinRj/3P_{\mathrm{in}}R_j/3. The lossless MMI preserves their total power on the return pass, giving ∑kPk=(Pin/3)∑jRj\sum_k P_k = (P_{\mathrm{in}}/3)\sum_j R_j. Let aj=Rja_j = \sqrt{R_j} and ω=e2πi/3\omega= e^{2\pi i/3}. After choosing the reflector phases so that the returned fields add at the target, and relabeling it as output 0, the three output fields (up to a common phase, in units of Pin\sqrt{P_{\mathrm{in}}}) are

E0=a1+a2+a33,E1=a1+ωa2+ω2a33,E2=a1+ω2a2+ωa33=E1∗.\begin{aligned} E_0 &= \frac{a_1+a_2+a_3}{3},\\ E_1 &= \frac{a_1+\omega a_2+\omega^2 a_3}{3},\\ E_2 &= \frac{a_1+\omega^2 a_2+\omega a_3}{3}=E_1^*. \end{aligned}

Because the aja_j are real, the two unwanted fields are complex conjugates and therefore have equal power. The power at each output is then

Ptarget=Pin9(∑jRj)2,Poff,1=Poff,2=12(Pin3∑jRj−Ptarget).P_{\mathrm{target}}=\frac{P_{\mathrm{in}}}{9}\left(\sum_j\sqrt{R_j}\right)^2,\qquad P_{\mathrm{off},1}=P_{\mathrm{off},2}=\frac{1}{2}\left(\frac{P_{\mathrm{in}}}{3}\sum_jR_j-P_{\mathrm{target}}\right).

The maximum fraction of returned output power routed to the target and the corresponding target-to-strongest-unwanted extinction ratio are therefore

ηmax⁡=Ptarget∑kPk=(∑jRj)23∑jRj,\eta_{\max}=\frac{P_{\mathrm{target}}}{\sum_kP_k}=\frac{\left(\sum_j\sqrt{R_j}\right)^2}{3\sum_jR_j},
ERmax⁡=10log⁡10[Ptargetmax⁡(Poff,1,Poff,2)]=10log⁡10(2ηmax⁡1−ηmax⁡).(77)\mathrm{ER}_{\max}=10\log_{10}\left[\frac{P_{\mathrm{target}}}{\max(P_{\mathrm{off},1},P_{\mathrm{off},2})}\right]=10\log_{10}\left(\frac{2\eta_{\max}}{1-\eta_{\max}}\right). \tag*{(77)}

Here ηmax⁡=Ptarget/∑kPk\eta_{\max}=P_{\mathrm{target}}/\sum_kP_k measures routing of the power that returns, rather than transmission relative to the input. The factor of two in ERmax⁡\mathrm{ER}_{\max} follows because each unwanted port receives half of the remaining power. The extinction-ratio definition has the same target-to-strongest-unwanted form as Eq. (3), but uses ideal optical powers here rather than the independently normalized outputs of the main-text measurement. Uniform loss therefore increases insertion loss but does not change the maximum extinction ratio, whereas unequal arm losses prevent exact cancellation at the unwanted outputs. Supplementary Fig. S21 quantifies this sensitivity for independent Gaussian arm-loss variations: at σ=0.5\sigma= 0.5 and 11 dB, the median routed fractions are 99.85% and 99.39%, and the median extinction ratios are 31.2 and 25.2 dB, respectively.

Arm-phase and cavity-parameter effects on routing performance

In the fitted optical model, adjusting the static arm phases makes routing less sensitive to voltage errors. Optimizing those phases together with the routing voltages places all three states within a 2.46-V range on each electrode. Higher cavity quality factor reduces the required drive but narrows the bandwidth, while stronger overcoupling reduces cavity loss.

The fabricated device includes phase shifters in the three paths between the MMI arms and the one-sided cavities. To investigate how their operating points affect routing, we retained the fitted cavity and MMI parameters while varying the two independent relative round-trip arm phases. This optical-only sensitivity model excludes the empirical coherent detector backgrounds and detector gain and offset calibration used in the Fig. 4 fit. For each phase pair, we swept the three cavity voltages independently over the ±5\pm5 V cube and optimized each route separately. We recorded both the largest extinction ratio available for each route and the fraction of sampled voltage-cube points above a specified threshold. The reported region fraction is the smallest of the three route-specific fractions and therefore requires useful regions for every output port. It counts all qualifying sampled points regardless of connectivity and is not the voltage tolerance around a selected routing bias.

Image not extracted

Figure S21. Supplementary Figure S21 Ideal routing sensitivity to independent arm-loss imbalance. Monte Carlo calculation for an ideal 3×33 \times3 Fourier MMI with independently Gaussian-distributed arm losses. The common loss is removed from each realization, and the reflector phases are optimized for the target output. a, Maximum fraction of returned power routed to the target. b, Optimized worst-case extinction ratio. c, Fraction of realizations exceeding extinction-ratio thresholds of 20, 30 and 40 dB. Lines show medians and shaded bands show the 5th–95th percentiles over 3×1053 \times10^{5} realizations at each σ\sigma. This is a sensitivity analysis, not a measured loss distribution or fabrication-yield estimate.

Supplementary Fig. S22a shows 240 uniformly sampled operating-point phase pairs. The 5th, median, and 95th percentiles of the optimized worst-route extinction ratio are 9.9, 14.6, and 23.3 dB, respectively. Independently, the 5th, median, and 95th percentiles of the voltage-cube fraction above 10 dB are 0.00%, 0.21%, and 2.07%. At the fitted operating-point phases, the optical-only model gives an optimized worst-route extinction of 16.0 dB and a 10-dB region fraction of 1.49%. This 16.0-dB value is the smallest of three separately optimized route maxima, not the calibrated Figure 4 prediction at its displayed routing states. Optimizing solely for the highest extinction point produces a model value of 39.9 dB, but this solution relies on a narrow coherent cancellation. A more robust optimization that maximizes the volume of the 10-dB regions gives a 24.0-dB worst-route optimum and a 3.93% worst-route region fraction. These optical-only sensitivity values are distinct from the fully calibrated model and measured routing states in Figure 4.

To quantify local voltage-error tolerance, we find for each route the largest axis-aligned cube centered on a route-specific voltage bias and contained in the ±5\pm5 V domain for which the modeled extinction remains at least 10 dB throughout. We use one shared pair of operating-point phases across the three routes and report the smallest of their cube half-widths. This half-width is the allowed independent error on each electrode, including simultaneous errors on all three. Multistart numerical optimization gives approximately ±0.70\pm0.70 V per electrode at the fitted phases and ±1.13\pm1.13 V when the shared phases are optimized directly for this tolerance. At the latter setting, the optimized worst-route peak is 21.3 dB and the worst-route 10-dB region fraction is 3.77%. The cube interiors were checked numerically; these values are model estimates, not certified global optima or measurements of voltage-source stability. The region fraction and local tolerance thus describe different aspects of robustness.

For Supplementary Fig. S22b, we select the actual sampled phase pairs whose optimized worst-route extinction ratios are closest to the 5th, 50th, and 95th percentiles. These labels therefore identify three representative phase settings; they are not uncertainty bounds on the curves. The region-optimized setting is selected using a different objective: the voltage-cube fraction above 10 dB. It gives the largest fraction at that design threshold, marked by the vertical dotted line, but need not remain above the 95th-percentile representative at higher thresholds.

Image not extracted

(a)

Image not extracted

(b)

Figure S22. Optical-only sensitivity model: empirical coherent backgrounds and detector calibration excluded. Optical-only sensitivity to operating-point phases. Fitted cavity and MMI parameters are retained, while empirical coherent detector backgrounds and detector calibration are omitted. The extinctions shown are not Figure 4 model predictions. a, Optimized worst-route extinction ratio and fraction of the ±5\pm5 V voltage cube above 10 dB for 240 uniformly sampled pairs of relative round-trip arm phases (gray). The fitted phase setting, the phase setting maximizing the point extinction ratio, the setting maximizing the 10-dB region volume, and the setting optimizing local 10-dB voltage-error tolerance are highlighted. b, Worst-route voltage-cube fraction as a function of extinction-ratio threshold for representative 5th-, 50th-, and 95th-percentile phase settings selected by optimized worst-route extinction, together with the region-optimized setting, which is optimized specifically for the fraction above 10 dB. The percentile labels identify representative sampled settings rather than uncertainty bounds. These phase sweeps are a sensitivity analysis, not a measured phase distribution or a fabrication-yield estimate.

We jointly optimized the three routing voltage settings and one shared pair of operating-point phases, requiring at least 15 dB extinction for each route. The objective was the side length of the smallest axis-aligned cube containing all three voltage settings, equivalently the largest inter-state voltage span on any electrode. The smallest feasible value found by the numerical search was 2.46 V. Taking arm 3 as the reference, the additional round-trip operating-point phase offsets are 0.605 rad for arm 2 and 0.925 rad for arm 1. We then reevaluated the reciprocal model over the full fitted-baseline ±5\pm5 V cube at this phase setting to visualize the solution (Supplementary Fig. S23). These are static relative arm phases, not the global phase of the illuminated input. This inter-state span excludes the common DC voltages used to align the cavity resonances.

The intensity cube in Supplementary Fig. S23a encodes the three model output powers as red, green and blue after separately normalizing each output over the full cube. Consequently, its brightness and hue do not represent absolute throughput and should not be compared quantitatively with the measured cube in Figure 4. Panel b instead evaluates the same worst-case extinction ratio ERk\mathrm{ER}_k defined in the main text, here using the model powers PkP_k, and displays only its positive part, max⁡{ERk,0}\max\{\mathrm{ER}_k,0\}. At most one route can have a positive extinction ratio at a given voltage tuple, so this representation is intentionally less color-mixed than the intensity representation. The optimized phases relocate the three reference lobes and bring the three selected high-extinction regions close together, but the brighter or purer colors must not be interpreted as a phase-induced increase in absolute optical power.

The three selected voltage settings do not lie on a single straight line, so they define a plane in the three-voltage space. Its intersection with the full voltage cube is shown in Supplementary Fig. S23c, and panel d enlarges the optimized triangle and the cross-section of its 2.46-V bounding cube. The optimization did not impose planarity; the plane follows from the selected settings. The construction therefore provides a convenient two-coordinate description of transitions among these three states, rather than evidence that the full routing landscape is intrinsically two-dimensional or that only two physical electrodes are required.

Image not extracted

Figure S23. Voltage-space geometry of the jointly optimized routing states. The fitted reciprocal model is evaluated over the complete fitted-baseline ±5\pm5 V cube using one shared optimized pair of relative round-trip arm phases. Red, green and blue correspond to routing toward P1P_1, P2P_2 and P3P_3, respectively. a, Model output powers, independently normalized per output over the cube and RGB encoded; color brightness is not an absolute-power calibration. b, The same worst-case extinction ratio ERk\mathrm{ER}_k defined in the main text, evaluated from the model powers and clipped at zero for display, then normalized over the cube. This winner-only encoding is primary-colored by construction. c, Extinction-ratio map in the plane through the three optimized states. The black hexagon is the intersection of this plane with the full ±5\pm5 V cube; the gray exterior is outside the evaluated control range. The plotted (e1,e2)(e_1,e_2) coordinates describe positions within this plane, not individual electrode voltages; their ranges can therefore exceed ±5\pm5 V while each electrode remains within its ±5\pm5-V range. d, Enlargement of the optimized states. The white/black polygon is the plane section of their smallest axis-aligned bounding cube, whose side is ΔVmax⁡=2.46\Delta V_{\max}=2.46 V. Lines connecting the states are guides to the eye. The phase offsets and voltage tuples are numerical model optima and are not independent measurements.

Next, we varied the loaded quality factor and overcoupling in controlled sweeps (Supplementary Fig. S24). Scaling the loaded quality factors while holding the tuning efficiency (in GHz V−1^{-1}) and κe/κi\kappa_e/\kappa_i fixed isolates the linewidth effect. The voltage required to tune through one linewidth scales as

Vlinewidth=ν0QLη,(78)V_{\mathrm{linewidth}}=\frac{\nu_0}{Q_L\eta}, \tag*{(78)}

where η\eta is the tuning efficiency. Consistent with this scaling, we numerically optimized one shared pair of static operating-point phases and the three routing voltage tuples while requiring every target/off-target ratio to exceed 15 dB. For three routing states, we define the largest inter-state electrode-voltage excursion as

ΔVmax⁡=max⁡j[max⁡sVj(s)−min⁡sVj(s)],(79)\Delta V_{\max}=\max_j\left[\max_s V_j^{(s)}-\min_s V_j^{(s)}\right], \tag*{(79)}
Image not extracted

Figure S24. Model-derived cavity-design trade-offs. Here x=κe/κix = \kappa_{\mathrm{e}}/\kappa_{\mathrm{i}} is the external-to-internal coupling ratio. a, Best-found maximum electrode span across the three routing states, obtained by joint numerical optimization subject to ER ≥15\ge15 dB for every route. One shared pair of static arm phases and all three voltage tuples are optimized for the fitted cavities, doubled-loaded QQ, common overcoupling ratio x=5x = 5, their combination, and the lossless-cavity limit. b, Median optimized worst-route extinction ratio (green) and median fraction of the ±5\pm5 V cube above 10 dB (blue dashed) versus the common overcoupling ratio, evaluated over random relative arm phases at fixed loaded QQ. The vertical dotted line marks the geometric mean of the three fitted coupling ratios. c, Worst cavity-only loss after maximizing cavity survival subject to ER ≥15\ge15 dB for all routes and optimizing one shared pair of static arm phases. The first bar uses the fitted heterogeneous cavities; the following bars use the indicated common κe/κi\kappa_{\mathrm{e}}/\kappa_{\mathrm{i}} values and the lossless single-sided limit.

where ss indexes the three routing states. The best solution found with the fitted cavity parameters gives ΔVmax⁡=2.46\Delta V_{\max} = 2.46 V. For comparison, the same constrained span is 0.42, 0.48, and 0.54 V at thresholds of 10, 11, and 12 dB, respectively. Doubling all loaded quality factors reduces the 15-dB value to 0.35 V, while setting all three coupling ratios to κe/κi=5\kappa_{\mathrm{e}}/\kappa_{\mathrm{i}} = 5 at the fitted loaded QQ gives 1.65 V. Combining both changes gives 0.82 V. Higher QLQ_{\mathrm{L}} therefore reduces the required drive, but it also reduces the cavity-lifetime-limited bandwidth as 1/QL1/Q_{\mathrm{L}} and narrows the useful voltage regions, increasing sensitivity to voltage noise and drift.

The overcoupling sweep holds QLQ_{\mathrm{L}} fixed and assigns a common ratio x=κe/κix = \kappa_{\mathrm{e}}/\kappa_{\mathrm{i}} to the three cavities. At resonance, an overcoupled single-sided cavity has power reflectance

∣r(0)∣2=(x−1)2(x+1)2.(80)|r(0)|^2 = \frac{(x - 1)^2}{(x + 1)^2}. \tag*{(80)}

Increasing xx therefore suppresses the dissipative reflection dip and makes the cavity more nearly an all-pass phase shifter. In the random-phase ensemble, increasing xx between the sampled sweep points 1.58 and 2.90 raises the median optimized worst-route extinction ratio from 16.0 to 20.1 dB and expands the median 10-dB region fraction from 0.76% to 2.35%. The improvement largely saturates near x∼5x \sim5 in this fitted model because the remaining MMI amplitude imbalance and finite control range become limiting. The strictly lossless-cavity limit gives a best-found span of 2.07 V in the same constrained optimization. The nonmonotonic ordering across the five cases reflects both linewidth narrowing and changes in the cavity amplitude responses. In particular, the heterogeneous fitted responses can partially compensate the fitted MMI imbalance, so removing absorption does not necessarily minimize the voltage span for this specific nonideal model. This does not imply that absorption is generally beneficial: for a balanced MMI, the lossless all-pass limit remains preferable.

Maximizing extinction alone can place a cavity close to its absorption dip, so we separately maximized the worst-route cavity survival subject to a minimum extinction ratio of 15 dB. After optimizing one shared pair of static arm phases, the resulting worst cavity-only loss is 2.29 dB for the fitted heterogeneous cavities, 0.90 dB for x=3x = 3, 0.58 dB for x=5x = 5, and 0.32 dB for x=10x = 10, approaching zero in the lossless single-sided limit (Supplementary Fig. S24c). These are model-derived cavity losses evaluated before the return through the MMI; they are not total switch or fiber-to-fiber insertion losses. The QLQ_\mathrm{L} and overcoupling sweeps are varied independently to expose the trade-offs. In a physical redesign, changing intrinsic or external coupling can change both quantities simultaneously.

Taken together, the model identifies separate design levers: static arm-phase tuning can improve voltage tolerance, higher loaded quality factor reduces the required voltage span at the cost of bandwidth and narrower useful voltage regions, and stronger overcoupling reduces cavity-only loss while improving extinction until the remaining MMI imbalance and finite control range become limiting. We optimized voltage span and cavity loss separately, so their reported best values should not be attributed to the same operating settings.

Supplementary Note 5: Electro-optic bandwidth extraction

The electro-optic transfer-function measurement contains 201 frequency points from 50 MHz to 10 GHz. The laser output was set to 13 dBm. We correct the VNA transmission trace with a separately measured through-reference trace by subtracting the reference magnitude in decibels at each device frequency, interpolating the reference where needed. This calibration removes the cable response only; the probe, contact pads and photodetector remain part of the measured response. Denoting the corrected values by did_i, we fit all 201 points by nonlinear least squares to the Butterworth-like low-pass family [41]

dnfit(f)=bn−10log⁡10[1+(ff3dB)2n].(81)d_n^{\mathrm{fit}}(f) = b_n - 10\log_{10}\left[1+\left(\frac{f}{f_{3\mathrm{dB}}}\right)^{2n}\right]. \tag*{(81)}

For each fixed order nn, the two fitted parameters are the low-frequency level bnb_n (the model value at f=0f = 0) and the frequency f3dBf_{3\mathrm{dB}} at which the model falls 3.01 dB below that level. We compare n=1,2,3,4n = 1, 2, 3, 4 on the same samples; the model comparison below selects n=3n = 3. For the normalized plots, we subtract this third-order fit’s low-frequency level from both the measured trace and the model curves, setting that level to 0 dB. The fit gives a 3-dB bandwidth of 6.20 GHz. The measured trace itself crosses −3 dB at 6.07 GHz, found by interpolating between the two surrounding measurement points. Because every fixed-order candidate has the same two fitted parameters, we compare their root-mean-square residual errors on the same 201 points. We also allow nn to vary as a third parameter.

Supplementary Table 3 Comparison of fits to the through-corrected electro-optic magnitude response. The fixed-order model with the lowest RMSE is highlighted in bold.

ModelOrderf3dBf_{3\mathrm{dB}} (GHz)RMSE (dB)R2R^2
One-pole13.6282.0820.734
Low-pass25.3541.0220.936
Low-pass36.200.5430.982
Low-pass46.7650.8180.959
Free-order low-pass3.0286.2180.5420.982

Table S3.

The third-order model has the lowest fixed-order RMSE, with f3dB=6.20f_{3\mathrm{dB}} = 6.20 GHz, a root-mean-square residual of 0.543 dB, and R2=0.982R^2 = 0.982 (Supplementary Fig. S25). The free-order fit gives n=3.028n = 3.028, close to three, with nearly the same RMSE (0.542 dB). This supports the third-order empirical fit, from which we report the bandwidth as 6.20 GHz. The covariance matrix gives a formal standard error of 0.025 GHz for the fixed third-order bandwidth, but this does not include uncertainty from the through-reference measurement, model selection, or correlated ripple and is not quoted as a total measurement uncertainty.

Image not extracted

Figure S25. Electro-optic bandwidth-model comparison. The black curve is the response measured at a 13-dBm laser-output setting after through correction and subtraction of the fitted third-order low-frequency offset. The dotted gray curve is an order-one comparison constrained to the same low-frequency offset and 3-dB bandwidth as the best fixed-order model; it is not independently fitted. The dashed black curve is the best fixed-order model (n=3n = 3). This construction emphasizes the models’ different roll-offs after their common corner. The same fitted offset is subtracted from the measured trace and both model curves for display. The model curves cross −3.01-3.01 dB at f3dB=6.20f_{3\mathrm{dB}} = 6.20 GHz, marked by the horizontal and vertical guide lines. The directly interpolated measurement crossing at −3-3 dB is 6.07 GHz.

Finally, the magnitude response for n=1n = 1 in Eq. (81) is both the magnitude of a single-pole RC transfer function and the square root of a Lorentzian power response. Consequently, magnitude-only S21S_{21} data cannot distinguish those two physical labels. The empirical preference for a third-order roll-off does not by itself identify the physical processes that limit the measured response. The analysis code is included in the repository listed under Code availability.

Single-channel high-extinction routing

To distinguish the attainable contrast of an individually optimized routing state from the stability required to hold all three states in a switching sequence, we performed a targeted voltage-space acquisition around the high-contrast operating regions. The three detector traces were delay aligned using the same calibration as Figure 4 and independently min–max normalized over the full acquisition, consistent with the extinction ratio definition used for Figure 4c.

For routing from input port 1 to output port 1, the maximum instantaneous extinction ratio was ER1=22.63ER_1 = 22.63 dB (Supplementary Fig. S26). The periodic drive revisited the same waveform phase once in each 25 μ\mus period. The four visits gave 20.76, 20.98, 22.63, and 20.78 dB, corresponding to 21.29±0.9021.29 \pm0.90 dB (mean ±\pm sample standard deviation). Averaging the three detector signals across the four visits before evaluating their ratio gives 21.79 dB.

At the 22.63-dB sample, the normalized detector signals were (P~1,P~2,P~3)=(0.9542,0.00520,0.00237)(\tilde{P}_1, \tilde{P}_2, \tilde{P}_3) = (0.9542, 0.00520, 0.00237). The same-phase one-standard-deviation repeatabilities of the two off-target channels were 0.00197 and 0.00319, respectively. Thus, the P~3\tilde{P}_3 residual was below its measured repeatability and the P~2\tilde{P}_2 residual was only 2.6 times its repeatability. Using the larger off-target repeatability as the denominator gives an approximate one-standard-deviation resolution scale of 24.8 dB. The measured maximum is therefore close to the resolution of the present setup. Because the detector channels are normalized independently and the capture contains no separate dark/background calibration, this measurement does not determine absolute insertion loss or absolute optical crosstalk.

Image not extracted

Figure S26. Targeted single-channel high-extinction routing. a, Normalized measured detector signals in a 20-ns window centered on the highest-extinction sample. Red, green and blue denote P1P_1, P2P_2 and P3P_3, respectively. Solid curves show the mean over four repeated 25-µs periods and shaded bands show one sample standard deviation; the dashed vertical line marks the optimum waveform phase. b, Extinction ratio for routing to P1P_1. Pale red curves show the four individual periods, the black curve is evaluated from the phase-averaged detector signals, and red markers give the four individual-period values at the optimum. The maximum instantaneous extinction is 22.6 dB.

Supplementary Note 7: Four-port reachable-region extension

Supplementary Note 2 shows that every ideal equal-split unitary supports perfect routing between any selected input and output, but that the complete reachable power region need not be universal for N>3N > 3. Here, we describe this distinction geometrically for three 4×44 \times4 examples: the real Walsh matrix, the four-point DFT, and an intermediate equal-split unitary. We assume lossless propagation and independently controllable reflector phases with ∣rj∣=1|r_j| = 1. The plots describe power distributions for one excited input, not simultaneous arbitrary routing of several inputs.

A continuous family of equal-split four-port matrices

Consider the normalized family

M(α)=12(11111eiα−1−eiα1−11−11−eiα−1eiα).(82)\mathbf{M}(\alpha) = \frac{1}{2} \begin{pmatrix} 1 & 1 & 1 & 1 \\ 1 & e^{i\alpha} & -1 & -e^{i\alpha} \\ 1 & -1 & 1 & -1 \\ 1 & -e^{i\alpha} & -1 & e^{i\alpha} \end{pmatrix}. \tag*{(82)}

Every member is unitary and all entries have modulus 1/21/2. At α=0\alpha= 0 it equals W4\mathbf{W}_4, and at α=π/2\alpha= \pi/2 it equals F4\mathbf{F}_4 in Eq. (65). Thus the intermediate example α=π/4\alpha= \pi/4 is not a different splitting-ratio design: its single-pass power splitting remains exactly equal, while its internal phases differ.

The classification of order-four complex Hadamard matrices shows that every equal-split 4×44 \times4 unitary belongs to this one-parameter family up to row and column phases and permutations [42]. These operations change the reachable power set only through a fixed output-port relabeling. Complex conjugation also preserves the power set, since the controls can be conjugated together with the matrix. In particular, α↦α+π\alpha\mapsto\alpha+ \pi exchanges rows two and four, while α↦−α\alpha\mapsto-\alpha conjugates the matrix. For comparing power-region shapes, it therefore suffices to use 0≤α≤π/20 \le\alpha\le\pi/2. There is a continuous range of possibilities, rather than a finite list of three matrices. The three examples below illustrate that range.

Theorem 1 allows any fixed reciprocal arm phases D\mathbf{D} to be absorbed into the controllable reflector phases. For the zero-based matrix index n=0n = 0 (physical input port 1), M(α)e0=(1,1,1,1)T/2\mathbf{M}(\alpha)\mathbf{e}_0 = (1,1,1,1)^{\mathsf{T}}/2, so the output field can be written

z=12M(α)Tu,∣uj∣=1,Pk=∣zk−1∣2,k=1,…,4.(83)\mathbf{z} = \frac{1}{2}\mathbf{M}(\alpha)^{\mathsf{T}}\mathbf{u}, \qquad|u_j| = 1, \qquad P_k = |z_{k-1}|^2, \qquad k = 1,\ldots,4. \tag*{(83)}

Here u\mathbf{u} contains the effective reflector phases after absorbing D\mathbf{D}. The same power set is obtained for every other input: the entries of any column of an equal-split unitary have modulus 1/21/2, and their phases can likewise be absorbed into u\mathbf{u}. Therefore the figures apply to every input and to arbitrary fixed reciprocal post-MMI phases.

Coordinates and cross-section construction

Unitarity gives P1+P2+P3+P4=1P_1 + P_2 + P_3 + P_4 = 1. Consequently, the four nonnegative powers occupy a three-dimensional tetrahedron, the probability simplex. Its four vertices are perfect single-output routing states. To inspect the interior without overlap between front and back surfaces, define

s=P1+P3,x=P1−P3,y=P2−P4.(84)s = P_1 + P_3, \qquad x = P_1 - P_3, \qquad y = P_2 - P_4. \tag*{(84)}

The height ss sets the total power in outputs 1 and 3; the other pair carries 1−s1-s. Within a constant-ss plane, xx and yy specify the imbalances inside the two pairs. These coordinates retain all four powers, since

(P1,P2,P3,P4)=12(s+x,1−s+y,s−x,1−s−y).(85)(P_1,P_2,P_3,P_4) = \frac{1}{2}(s+x,1-s+y,s-x,1-s-y). \tag*{(85)}

Positivity restricts a full simplex section to the rectangle

0≤s≤1,∣x∣≤s,∣y∣≤1−s.(86)0 \le s \le1,\qquad|x| \le s,\qquad|y| \le1-s. \tag*{(86)}

In these linear coordinates the tetrahedron has vertices (1,0,1)(1,0,1), (0,1,0)(0,1,0), (−1,0,1)(-1,0,1) and (0,−1,0)(0,-1,0), with coordinates ordered as (x,y,s)(x,y,s). Its nonregular appearance in the sliced three-dimensional plot is only a consequence of the coordinate choice.

Removing one common, physically irrelevant phase, every control vector can be parametrized as

u=(eiϕ,ei(χ+ψ),e−iϕ,ei(χ−ψ))T.(87)\mathbf{u}=\left(e^{i\phi},e^{i(\chi+\psi)},e^{-i\phi},e^{i(\chi-\psi)}\right)^{\mathsf{T}}. \tag*{(87)}

Substitution into Eq. (83) gives the power map

s=cos⁡2ϕ+cos⁡2ψ2,x=cos⁡ϕcos⁡ψcos⁡χ,y=sin⁡ϕsin⁡ψcos⁡(χ+α).(88)\begin{aligned} s&=\frac{\cos^{2}\phi+\cos^{2}\psi}{2},\\ x&=\cos\phi\cos\psi\cos\chi,\\ y&=\sin\phi\sin\psi\cos(\chi+\alpha). \tag*{(88)} \end{aligned}

To construct a section at fixed ss, set k=cos⁡2ϕcos⁡2ψk=\cos^{2}\phi\cos^{2}\psi and ℓ=1−2s+k\ell=1-2s+k. The allowed interval follows from fixing the sum of cos⁡2ϕ\cos^{2}\phi and cos⁡2ψ\cos^{2}\psi:

max⁡(0,2s−1)≤k≤s2,ℓ=sin⁡2ϕsin⁡2ψ.(89)\max(0,2s-1)\le k\le s^{2},\qquad\ell=\sin^{2}\phi\sin^{2}\psi. \tag*{(89)}

After absorbing signs into χ\chi, the section is the union of the curves

x=kcos⁡χ,y=±ℓcos⁡(χ+α).(90)x=\sqrt{k}\cos\chi,\qquad y=\pm\sqrt{\ell}\cos(\chi+\alpha). \tag*{(90)}

For 0<α≤π/20<\alpha\le\pi/2 and kℓ>0k\ell>0, each curve is an ellipse. The complete section equals the union of their filled elliptical disks. To see why filling does not introduce extra states, decrease kk continuously from any disk containing the point to its lower endpoint. At that endpoint the ellipse degenerates to an already reachable line segment. A point either lies on that segment or crosses an ellipse boundary at an intermediate kk; the boundary is attained by Eq. (90). This construction preserves nonconvex sections and is not a convex-hull approximation. The intermediate sections in Supplementary Figs. S27 and S28 are evaluated by a numerically converged envelope over the interval Eq. (89); the Walsh and DFT sections use the analytic expressions below.

Walsh, intermediate and DFT regions

The DFT corresponds to α=π/2\alpha=\pi/2. The elliptical disks grow with kk, so the largest disk, k=s2k=s^{2} and ℓ=(1−s)2\ell=(1-s)^{2}, gives the complete section:

x2s2+y2(1−s)2≤1,0<s<1.(91)\frac{x^{2}}{s^{2}}+\frac{y^{2}}{(1-s)^{2}}\le1,\qquad0<s<1. \tag*{(91)}

Thus every nondegenerate constant-ss section is a filled ellipse. At s=1/2s=1/2 it is the circle x2+y2≤1/4x^{2}+y^{2}\le1/4 inscribed in the square simplex section. At s=0s=0 and s=1s=1 the sections are the complete opposite simplex edges joining outputs (2,4)(2,4) and (1,3)(1,3), respectively.

For Walsh, α=0\alpha=0, the section curves instead satisfy x=kcos⁡χx=\sqrt{k}\cos\chi and y=±ℓcos⁡χy=\pm\sqrt{\ell}\cos\chi. For 0<s<1/20<s<1/2, varying kk gives the exact section

y2≥(1−s)2s2x2,y2−x2≤1−2s.(92)y^{2}\ge\frac{(1-s)^{2}}{s^{2}}x^{2},\qquad y^{2}-x^{2}\le1-2s. \tag*{(92)}
Image not extracted

Figure S27. Sliced three-dimensional views of the ideal four-port regions. a, Walsh. b, Intermediate matrix, α=π/4\alpha= \pi/4. c, DFT. All panels use the same coordinates x=P1−P3x = P_1 - P_3, y=P2−P4y = P_2 - P_4 and s=P1+P3s = P_1 + P_3, and the same view. Gray edges outline the full power tetrahedron. Blue curves are reachable section boundaries at s=1/8,1/4,3/8,1/2,5/8,3/4,7/8s = 1/8, 1/4, 3/8, 1/2, 5/8, 3/4, 7/8; sections at s=1/4,1/2,3/4s = 1/4, 1/2, 3/4 are shaded where they have nonzero area. Their gray rectangular outlines show the corresponding full simplex sections. The sampled heights aid visualization only: all intermediate heights are included in the actual reachable region. Each tetrahedron vertex is a perfect routing state, while the different section shapes expose the nonequivalence of the interior reachable regions.

together with Eq. (86). Sections at 1/2<s<11/2 < s < 1 follow by exchanging xx and yy and replacing ss by 1−s1 - s. The central section degenerates to the two segments y=xy = x and y=−xy = -x within [−1/2,1/2]2[-1/2, 1/2]^2. This zero-area slice is a geometric restriction, not an absence of reachable states.

For the whole family, the central section has the particularly simple form

x2+y2−2cos⁡α ∣xy∣≤sin⁡2α4,∣x∣,∣y∣≤12.(93)x^2 + y^2 - 2\cos\alpha\,|xy| \le\frac{\sin^2\alpha}{4}, \qquad|x|, |y| \le\frac{1}{2}. \tag*{(93)}

This follows from Eq. (90) with ℓ=k\ell= k: the disks scale with k\sqrt{k} and the largest one has k=1/4k = 1/4. At α=π/4\alpha= \pi/4 the section is the union of two crossed elliptical disks, whose filled union is shown in Supplementary Fig. S28(e). At the endpoints, the same expression yields the Walsh diagonals and DFT circle. The maximum ∣x∣|x| on the line y=0y = 0 is 12sin⁡α\frac{1}{2}\sin\alpha, illustrating that the section changes continuously with the MMI phase parameter.

Image not extracted

Figure S28. Cross-sections of three ideal four-port reachable power regions. Rows show the Walsh matrix (α=0\alpha= 0; a–c), the intermediate matrix (α=π/4\alpha= \pi/4; d–f), and the DFT (α=π/2\alpha= \pi/2; g–i). Columns fix s=P1+P3s = P_1 + P_3 at 1/41/4, 1/21/2, and 3/43/4, respectively. Blue marks reachable states; the gray rectangle is the full allowed simplex section, ∣x∣≤s|x| \le s, ∣y∣≤1−s|y| \le1 - s. In the central column, the reachable section is two diagonal segments for Walsh, a union of two elliptical disks for the intermediate matrix, and a circular disk for DFT. The axes give normalized power differences, x=P1−P3x = P_1 - P_3 and y=P2−P4y = P_2 - P_4. Walsh and DFT curves are analytic; the intermediate noncentral boundaries use the converged ellipse envelope described in the text.

Finally, two different meanings of a “pair” must be distinguished. Input–output pairwise routing means sending all power from a selected input nn to a selected output mm. Every matrix considered here can do this, by Theorem 2; all four routing vertices remain reachable. Two-output splitting means distributing one input between two chosen outputs with zero power in the other two. Walsh reaches every point of all six simplex edges: choosing the two nonzero output amplitudes in quadrature makes all inverse-control entries have unit modulus. For 0<α≤π/20 < \alpha\le\pi/2, only the opposite edges (1, 3) and (2, 4) have reachable interiors; the other four edges contain only the routing vertices. The obstruction is the same phase-matching condition used in the four-port counterexample of Supplementary Note 2: for the other output pairs, the relative column phases include both ±1\pm1 and ±eiα\pm e^{i\alpha}, so two nonzero output amplitudes cannot give unit-modulus controls when sin⁡α≠0\sin\alpha\ne0. These edge restrictions do not contradict any-to-any single-output routing, and fixed post-MMI phases do not change them.

The figure-generation code is included in the repository listed under Code availability.

References

  1. [3]Bogaerts, W. et al. Programmable photonic circuits. Nature 586, 207–216 (2020).
  2. [4]Kimble, H. J. The quantum internet. Nature 453, 1023–1030 (2008).
  3. [5]Wehner, S., Elkouss, D. & Hanson, R. Quantum internet: A vision for the road ahead. Science 362, eaam9288 (2018).
  4. [6]Pompili, M. et al. Realization of a multinode quantum network of remote solid-state qubits. Science 372, 259–264 (2021).DOI
  5. [7]Chen, X., Lin, J. & Wang, K. A review of silicon-based integrated optical switches. Laser Photonics Rev. 17, 2200571 (2023).DOI
  6. [8]Zhu, D. et al. Integrated photonics on thin-film lithium niobate. Adv. Opt. Photon. 13, 242–352 (2021).DOI
  7. [9]Boes, A. et al. Lithium niobate photonics: Unlocking the electromagnetic spectrum. Science 379, eabj4396 (2023).DOI
  8. [10]Reck, M., Zeilinger, A., Bernstein, H. J. & Bertani, P. Experimental realization of any discrete unitary operator. Phys. Rev. Lett. 73, 58–61 (1994).pubmed.ncbi.nlm.nih.gov/10056719
  9. [11]Clements, W. R., Humphreys, P. C., Metcalf, B. J., Kolthammer, W. S. & Walmsley, I. A. Optimal design for universal multiport interferometers. Optica 3, 1460–1465 (2016).arxiv.org/abs/1603.08788
  10. [12]López Pastor, V., Lundeen, J. & Marquardt, F. Arbitrary optical wave evolution with Fourier transforms and phase masks. Opt. Express 29, 38441–38450 (2021).DOI
  11. [13]Jenkins, R. M. et al. Novel 1 × N and N × N integrated optical switches using self-imaging multimode GaAs/AlGaAs waveguides. Appl. Phys. Lett. 64, 684–686 (1994).DOI
  12. [14]Soldano, L. B. & Pennings, E. C. M. Optical multi-mode interference devices based on self-imaging: principles and applications. J. Lightwave Technol. 13, 615–627 (1995).DOI
  13. [15]Benes, V. E. Optimal rearrangeable multistage connecting networks. Bell Syst. Tech. J. 43, 1641–1656 (1964).DOI
  14. [16]Spanke, R. A. Architectures for guided-wave optical space switching systems. IEEE Commun. Mag. 25, 42–48 (1987).DOI
  15. [17]Heuck, M. Proposal for multiport photon routing using photonic crystal cavity phase shifters. In 2023 Conference on Lasers and Electro-Optics Europe and European Quantum Electronics Conference (CLEO/Europe-EQEC), 1 (IEEE, 2023).DOI
  16. [18]Wang, C. et al. Integrated lithium niobate electro-optic modulators operating at CMOS-compatible voltages. Nature 562, 101–104 (2018).DOI
  17. [19]Xu, M. et al. High-performance coherent optical modulators based on thin-film lithium niobate platform. Nat. Commun. 11, 3911 (2020).
  18. [20]Li, M. et al. Lithium niobate photonic-crystal electro-optic modulator. Nat. Commun. 11, 4123 (2020).
  19. [21]Larocque, H. et al. Photonic crystal cavity IQ modulators in thin-film lithium niobate. ACS Photonics 11, 3860–3869 (2024).DOI
  20. [22]Suh, W., Wang, Z. & Fan, S. Temporal coupled-mode theory and the presence of non-orthogonal modes in lossless multimode cavities. IEEE J. Quantum Electron. 40, 1511–1518 (2004).DOI
  21. [23]Mellette, W. M. et al. RotorNet: A scalable, low-complexity, optical datacenter network. In Proceedings of ACM SIGCOMM 2017, 267–280 (ACM, 2017).
  22. [24]Holzgrafe, J. et al. Relaxation of the electro-optic response in thin-film lithium niobate modulators. Opt. Express 32, 3619–3631 (2024).DOI
  23. [25]Xu, Y. et al. Mitigating photorefractive effect in thin-film lithium niobate microring resonators. Opt. Express 29, 5497–5504 (2021).DOI
  24. [26]Rahim, A., Spuesens, T., Baets, R. & Bogaerts, W. Open-access silicon photonics: Current status and emerging initiatives. Proceedings of the IEEE 106, 2313–2330 (2018).DOI
  25. [27]Li, Q. et al. A 10-Gb/s silicon microring resonator-based BPSK link. IEEE Photonics Technology Letters 26, 1805–1808 (2014).DOI
  26. [28]Huang, Y., Cheng, Q., Rizzo, A. & Bergman, K. Push–pull microring-assisted space-and-wavelength selective switch. Optics Letters 45, 2696–2699 (2020).DOI
  27. [29]Lomonte, E. et al. Single-photon detection and cryogenic reconfigurability in lithium niobate nanophotonic circuits. Nat. Commun. 12, 6847 (2021).DOI
  28. [30]Thiele, F. et al. Cryogenic electro-optic polarisation conversion in titanium in-diffused lithium niobate waveguides. Opt. Express 28, 28961–28968 (2020).DOI
  29. [31]Desiatov, B., Shams-Ansari, A., Zhang, M., Wang, C. & Lončar, M. Ultra-low-loss integrated visible photonics using thin-film lithium niobate. Optica 6, 380–384 (2019).DOI
  30. [32]Choi, H., Pant, M., Guha, S. & Englund, D. Percolation-based architecture for cluster state creation using photon-mediated entanglement between atomic memories. npj Quantum Inf. 5, 104 (2019).arxiv.org/abs/1704.07292
  31. [33]Mehta, K. K. et al. Integrated optical multi-ion quantum logic. Nature 586, 533–537 (2020).
  32. [34]Christen, I. et al. An integrated photonic engine for programmable atomic control. Nat. Commun. 16, 82 (2025).
  33. [35]Wan, N. H. et al. Large-scale integration of artificial atoms in hybrid photonic circuits. Nature 583, 226–231 (2020).
  34. [36]Buzzi, A. et al. Spectral tuning and nanoscale localization of single color centers in silicon via controllable strain. Nat. Commun. 16, 8829 (2025).arxiv.org/abs/2501.17290
  35. [37]Grovn, E., Bundgaard-Nielsen, M., Mørk, J., Englund, D. & Heuck, M. Recirculating quantum information networks for fast deterministic quantum information processing. Preprint at https://arxiv.org/abs/2602.11033 (2026).
  36. [38]Bachmann, M., Besse, P. A. & Melchior, H. General self-imaging properties in N×N multimode interference couplers including phase relations. Appl. Opt. 33, 3905–3911 (1994).DOI
  37. [39]Quan, Q., Deotare, P. B. & Lončar, M. Photonic crystal nanobeam cavity strongly coupled to the feeding waveguide. Appl. Phys. Lett. 96, 203102 (2010).DOI
  38. [40]Johnson, S. G. & Joannopoulos, J. D. Block-iterative frequency-domain methods for Maxwell’s equations in a planewave basis. Opt. Express 8, 173–190 (2001).pubmed.ncbi.nlm.nih.gov/19417802
  39. [41]Butterworth, S. On the theory of filter amplifiers. Experimental Wireless and the Wireless Engineer 7, 536–541 (1930).
  40. [42]Tadej, W. & Życzkowski, K. A concise guide to complex Hadamard matrices. Open Syst. Inf. Dyn. 13, 133–177 (2006).

Paper details

vlnc:W_fxdtyewg
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents