Optimized protocols for duplex quantum transduction

Introduction.—Quantum transducers convert quantum signals between physically distinct carriers, enabling quantum information exchange across multiple platforms in quantum networks Kimble (2008). For example, a microwave-to-optical quantum transducer Lambert et al. (2020); Han et al. (2021); Andrews et al. (2014); Vainsencher et al. (2016); Fan et al. (2018); Rueda et al. (2019); Jiang et al. (2020); Mirhosseini et al. (2020); McKenna et al. (2020); Xu et al. (2021); Tu et al. (2022); Shen et al. (2022) can distribute processed quantum states stored in superconducting qubits over optical fibers. Various designs of quantum transducers have been developed, utilizing hybrid interfaces like electro-optics Tsang (2010); Fan et al. (2018); Rueda et al. (2019); McKenna et al. (2020); Xu et al. (2021), optomechanics Aspelmeyer et al. (2014); Stannigel et al. (2010); Safavi-Naeini and Painter (2011); Hill et al. (2012); Palomaki et al. (2013); Andrews et al. (2014); Lecocq et al. (2016); Vainsencher et al. (2016); Jiang et al. (2020); Mirhosseini et al. (2020); Shen et al. (2022), and electro-magnonics Zhang et al. (2014); Tabuchi et al. (2015); Shen et al. (2022).

As devices for quantum state transfer, quantum transducers can be abstracted as quantum channels. Most bosonic quantum transducers use red-detuned pumps Han et al. (2021); Andrews et al. (2014); Fan et al. (2018); Rueda et al. (2019); Jiang et al. (2020); Mirhosseini et al. (2020); McKenna et al. (2020) to engineer a two-mode scattering process that is equivalent to a beam splitter. The input signal of one mode $\hat{a}_{1,\text{in}}(\hat{a}_{2,\text{in}})$ gets converted to the output signal of the other mode $\hat{a}_{2,\text{out}}(\hat{a}_{1,\text{out}})$, yielding two unidirectional transduction channels $\mathcal{E}_{1}$ and $\mathcal{E}_{2}$ (Fig. 1(a)). Oftentimes, only a single channel is utilized to transduce quantum signal from one mode to the other, which we refer to as unidirectional quantum transduction. The performance of unidirectional quantum transduction is characterized by the quantum capacity of either $\mathcal{E}_{1}$ or $\mathcal{E}_{2}$ Zhang et al. (2018); Wang et al. (2022).

By leveraging both unidirectional transduction channels, quantum signals can be converted bidirectionally which we refer to as duplex quantum transduction ¹¹1Previous works on “bidirectional conversion” Andrews et al. (2014); Vainsencher et al. (2016); Jiang et al. (2020); Xu et al. (2021) measures the two unidirectional transduction channels separately on a single device, while the duplex quantum transduction we consider here is more general.. Duplex quantum transduction can be modeled as a quantum interference channel Fawzi et al. (2012); Sen (2012); Das et al. (2021), with senders $A(\hat{a}_{1,\text{in}})$ and $B(\hat{a}_{2,\text{in}})$ and receivers $C(\hat{a}_{2,\text{out}})$ and $D(\hat{a}_{1,\text{out}})$. The senders and receivers can be distinct users by separating the input and output signals of each mode with circulators (Fig. 1(b)). For example, a lossless beam splitter with efficiency $T$ (Fig. 1(c)) implements a quantum interference channel:

One strategy for duplex quantum transduction is to alternate between using $\mathcal{E}_{1}$ and $\mathcal{E}_{2}$, while simultaneous transduction of uncorrelated input signals $\hat{a}_{1,\text{in}}$ and $\hat{a}_{2,\text{in}}$ may be more efficient. However, $\mathcal{E}_{1}$ and $\mathcal{E}_{2}$ can interfere with each other when put in use simultaneously, e.g., the input signal $\hat{a}_{1,\text{in}}$ for $\mathcal{E}_{1}$ acts as added noise for $\mathcal{E}_{2}$ in a beam splitter (Eq. (1)). As a result, characterizing the performance of duplex quantum transduction requires a new metric beyond the quantum capacities of the individual unidirectional transduction channels.

We propose to use achievable information rate region as the performance metric. The achievable rates of a quantum device depend on the quantum channels it implements as well as the input signal encodings, with channel parameters (such as $T$ in Eq. (1)) determined by the physical device parameters. In duplex quantum transduction, both transduction channels $\mathcal{E}_{1}$ and $\mathcal{E}_{2}$ transmit quantum information at rates $I_{1}$ and $I_{2}$ respectively. For simultaneous duplex transduction, the pair of achievable rates $(I_{1},I_{2})$ depends on how we encode quantum information into the quantum signals. By varying the encodings for $\hat{a}_{1,\text{in}}$ and $\hat{a}_{2,\text{in}}$, we obtain a two-dimensional region of achievable information rates $\{(I_{1},I_{2})\}$ (Fig. 1(d)). The rate region characterizes the performance of simultaneous duplex transduction and its boundary indicates the optimized coding strategies. Past studies have also employed rate regions or capacity regions to study the trade-off among multiple quantum channels, albeit limited to sending classical information Bennett et al. (2003); Childs et al. (2006); Fawzi et al. (2012); Sen (2012); Shi et al. (2021) or distributing entanglement in qubit-based quantum networks Pant et al. (2019); Shi and Qian (2020); Vardoyan et al. (2021); Dai et al. (2022). So far, there is no analysis investigating the achievable quantum information rate region at the hardware level, such as quantum transducers.

Furthermore, we can combine simultaneous duplex transduction with the time-sharing strategy, where we alternate between different signal encodings and even device parameters. This leads to a new region of achievable rates that is the convex hull of the original region, and we refer to the new region as the time-sharing achievable rate region. For example, we can perform transduction in one direction with $\mathcal{E}_{1}$ for 40% of the time and in the opposite direction with $\mathcal{E}_{2}$ for the rest 60% of the time (black dot Fig. 1(d)). Notably, when the original region is not convex, time-sharing can offer additional performance boost for duplex quantum transduction.

In this work, we define the (time-sharing) achievable rate region and apply the tool to characterize the performance of two-mode quantum transducers. We demonstrate that a sizable portion of quantum transducers can benefit from simultaneously transducing quantum signals in both directions. We also discuss how reflectionless scattering leads to the optimal duplex quantum transduction, as well as the effect of finite bandwidth.

Rate regions of duplex quantum transduction.—We consider quantum transducers with a linear input-output relation

where the scattering matrix $S$ is unitary and depends on the device parameters of the transducer Han et al. (2021). We choose ports 1 and 2 as the signal ports, and $\hat{a}_{n>2,\text{in}}$ are the injected vacuum noise from the internal loss channels. The two transduction channels are

Intuitively, the transmission coefficients $S_{12}$ and $S_{21}$ determine the transduction efficiency, while the reflection coefficients $S_{11}$ and $S_{22}$ lead to the interference between $\mathcal{E}_{1}$ and $\mathcal{E}_{2}$.

Here we define the achievable information rates for simultaneous duplex transduction. For a quantum channel $\mathcal{E}:\mathcal{L}(\mathcal{H})\rightarrow\mathcal{L}(\mathcal{H})$, the achievable rate of quantum information with an input state $\hat{\rho}$ is measured by the coherent information $I(\mathcal{E},\hat{\rho})$ Wilde (2013). Let $\left|\psi\right\rangle\in\mathcal{H}\otimes\mathcal{H}^{\prime}$ be a purification of $\hat{\rho}$, we have

where $H(\hat{\rho})$ is the von Neumann entropy of $\hat{\rho}$ and $\mathcal{I}^{\prime}$ is the identity map on $\mathcal{H}^{\prime}$. For $I(\mathcal{E},\hat{\rho})<0$, the achievable rate is 0. Generalizing to a quantum interference channel $\mathcal{E}_{(A,B)\rightarrow(C,D)}:\mathcal{L}(\mathcal{H}_{1}\otimes\mathcal{H}_{2})\rightarrow\mathcal{L}(\mathcal{H}_{1}\otimes\mathcal{H}_{2})$, the simultaneously achievable information rates $(I_{1},I_{2})$ with uncorrelated input state $\hat{\rho}_{1}\otimes\hat{\rho}_{2}$ are

Given the challenges in determining the quantum capacity for lossy channels with added noise Schumacher and Nielsen (1996); Lloyd (1997); Devetak (2005); Weedbrook et al. (2012), we focus on the rate region achievable with thermal input states as a lower bound. When the input signals of $\mathcal{E}_{1}$ and $\mathcal{E}_{2}$ are thermal states with average photon number $N_{1}$ and $N_{2}$, the outputs are also thermal states with photon number $N_{1}^{\prime}=TN_{1}+R_{2}N_{2}$ and $N_{2}^{\prime}=TN_{2}+R_{1}N_{1}$ (Fig. 2(a)). Here $R_{i}=|S_{ii}|^{2}$ is the power reflection coefficient from port $i$, $T_{ij}=|S_{ij}|^{2}$ is the power transmission coefficient from port $j$ to port $i$, and we assume $T_{12}=T_{21}\equiv T$. For finite $R_{1}$ and $R_{2}$, the reflected signal from one channel adds thermal noise to the other channel, which leads to the trade-off between $I_{1}$ and $I_{2}$ for simultaneous duplex transduction.

The achievable rates $(I_{1},I_{2})$ for Eq. (3) with thermal input states $(N_{1},N_{2})$ are (see Appendix A)

where $k=1,2$,

and

The rate region $\mathsf{R}\equiv\{(I_{1},I_{2})|\forall(N_{1},N_{2})\}$ only depends on channel parameters $(T,R_{1},R_{2})$. We could combine simultaneous duplex transduction with the time-sharing protocol, and the resulting time-sharing rate region is the convex hull $\tilde{\mathsf{R}}=\text{Conv}(\mathsf{R})$. Additionally, numerical evidences suggest that thermal encodings are likely optimal among general Gaussian encodings (see Appendix A.1). The rate regions can be calculated similarly when the environment injects thermal noise rather than vacuum noise via the internal loss channels $\hat{a}_{n>2,\text{in}}$ (see Appendix A.2).

The rate region $\mathsf{R}$ can be determined from its boundary $\partial\mathsf{R}$. For the special cases of unidirectional quantum transduction with $(N_{1},N_{2})=(\infty,0)$ and $(0,\infty)$, we achieve information rates $(I_{\text{max}},0)$ and $(0,I_{\text{max}})$ on $\partial\mathsf{R}$ (Fig. 2(b), blue dots). Here $I_{\text{max}}=\max\{\log_{2}(T/(1-T)),0\}$ is the quantum capacity of the pure-loss channel Holevo and Werner (2001). For $I_{1}>0$ and $I_{2}>0$, $\mathsf{R}$ corresponds to a continuous mapping $(N_{1},N_{2})\rightarrow(I_{1},I_{2})$ and $\partial\mathsf{R}$ can be solved numerically with the low-rank Jacobian condition $\det(J)=0$, where $J$ is the $2\times 2$ Jacobian matrix. In Fig. 2(b), we plot the rate regions $\mathsf{R}$ (blue lines and dots) and $\tilde{\mathsf{R}}$ (grey lines) for different reflection coefficients $(R_{1},R_{2})$. We choose $T=0.9$ with $I_{\text{max}}\approx 3.17$.

Finite reflection $R_{k}$ results in a noticeable discontinuity of the boundary $\partial\mathsf{R}$ at the $I_{k}$ axis (Fig. 2(b)i-iii). This can be explained from the upper bound on the thermal-loss capacity Rosati et al. (2018); Sharma et al. (2018); Noh et al. (2019). Assuming $R_{1}>0$, the channel $\mathcal{E}_{2,B\rightarrow C}$ is a thermal loss channel with noise photon $\bar{N}=R_{1}N_{1}/(1-T)$. From the upper bound Rosati et al. (2018); Sharma et al. (2018); Noh et al. (2019)

we must have $N_{1}<(2T-1)/2R_{1}$ to achieve a positive information rate $I_{2}>0$. On the other hand, when $I_{2}=0$ the quantum capacity $I_{1}=I_{\text{max}}$ is achieved at $N_{1}\rightarrow\infty$, which leads to the discontinuity at $I_{2}=0$.

If one side is reflectionless with $R_{k}=0$, the discontinuity of $\partial\mathsf{R}$ vanishes at the $I_{k}$ axis (Fig. 2(b)ii $I_{2}$ axis). If both sides are reflectionless, there is no interference between the two transduction channels and the maximal square region can be achieved (Fig. 2(b)iv). Therefore it is possible to outperform the time-shared unidirectional transduction (Fig. 1(d) grey dashed line) with the simultaneous duplex transduction, as long as the reflection coefficients are small (Fig. 2(b)ii-iv).

So far we have only considered direct transduction without adaptive control Zhang et al. (2018) or shared entanglement Zhong et al. (2020a, b); Wu et al. (2021). In Appendix B, we briefly discuss duplex quantum transduction assisted with local operations and classical communication (LOCC), along with the scenario where the senders are the same as the receivers which allows the interference-based techniques Lau and Clerk (2019); Zhang et al. (2022).

Optimized transduction protocols.—Here we apply the tool of rate regions to analyze a physical transducer model. The channel parameters $(T,R_{1},R_{2})$, and thus the achievable rates, depend on the device parameters of the transducer. We therefore generalize the achievable rate regions to include not only different signal encodings but also different device parameters. The boundary of the resulting rate region leads to optimized signal encodings and device parameters for the transducer.

We consider a transducer model for frequency conversion between two bosonic modes $\hat{a}_{1}$ and $\hat{a}_{2}$ (Fig. 3 (a)). The lab frame Hamiltonian is

where $\omega_{k}$ are the mode frequencies, $\omega_{p}$ is the pump frequency and $g$ is the interaction rate. An input signal at frequency $\omega$ in mode 1 gets converted to an output signal at frequency $\omega+\omega_{p}$ in mode 2, and vice versa. The Hamiltonian in the rotating frame of the signal is

where $\Delta_{1}=\omega_{1}-\omega$ and $\Delta_{2}=\omega_{2}-\omega_{p}-\omega$.

Assuming mode $k$ has external (internal) loss rate $\kappa_{k,e(i)},k=1,2$, the scattering matrix only depends on the ratios $\kappa_{k,e}/\kappa_{k,i}$ (see Appendix D). In practice, the loss rates of two modes may differ by orders of magnitude, but the ratios $\kappa_{1,e}/\kappa_{1,i}$ and $\kappa_{2,e}/\kappa_{2,i}$ are often close Fan et al. (2018); McKenna et al. (2020); Xu et al. (2021). Therefore, we assume symmetric loss rates $\kappa_{1,e(i)}=\kappa_{2,e(i)}\equiv\kappa_{e(i)}$ for simplicity and the more general case is discussed in Appendix D. The input-output relation is given by

where $\tilde{a}_{k,\text{in(out)}}$ are the internal loss channels, and

Here $I$ is the $2\times 2$ identity matrix, and

We focus on optimizing the detunings $(\Delta_{1},\Delta_{2})$ of the transducer, while keeping other relevant parameters $(g,\kappa_{e},\kappa_{i})$ fixed. Besides the signal encodings $(N_{1},N_{2})$, the achievable rates $(I_{1},I_{2})$ also depends on the device parameters $(\Delta_{1},\Delta_{2},g,\kappa_{e},\kappa_{i})$. We therefore define the rate region as $\mathsf{R}\equiv\{(I_{1},I_{2})|\forall(N_{1},N_{2},\Delta_{1},\Delta_{2})\}$. The optimized signal encodings $(N_{1},N_{2})$ and detunings $(\Delta_{1},\Delta_{2})$ can be obtained from the boundary $\partial\tilde{\mathsf{R}}$ of the time-sharing rate region $\tilde{\mathsf{R}}$.

The boundary of the rate region can be determined by exploring several possible solutions. On the $I_{1}$ and $I_{2}$ axes, the quantum capacity of unidirectional quantum transduction increases with the transmission coefficient. Therefore we choose $(\Delta_{1},\Delta_{2})$ that leads to the highest transmission rate $T$ to achieve the information rates $(I_{\text{max}},0)$ and $(0,I_{\text{max}})$ on $\partial\mathsf{R}$. For $I_{1}>0$ and $I_{2}>0$, $\mathsf{R}$ corresponds to a continuous mapping $(N_{1},N_{2},\Delta_{1},\Delta_{2})\rightarrow(I_{1},I_{2})$. The boundary $\partial\mathsf{R}$ as extreme values of the mapping can be obtained by comparing two possible solutions. One solution is from the low-rank Jacobian condition $\text{rank}(J)<2$, where $J$ is the $2\times 4$ Jacobian matrix. The other solution is from the reflectionless condition with $R_{k}=0$ and $N_{k}\rightarrow\infty$ where $k=1$ or $2$. We consider this solution separately since the Jacobian matrix may be undefined under the limit of $N_{k}\rightarrow\infty$. The reflectionless solution can be calculated analytically. For example, $R_{1}=0$ requires

which leads to the achievable rates at $N_{1}\rightarrow\infty$

In practice, we expect an approximate reflectionless solution with $R_{k}\approx 0$ and finite $N_{k}$, due to input power constraints and uncertainties in controlling the reflection coefficients.

We calculate the time-sharing rate region $\tilde{\mathsf{R}}(\kappa_{e})$ for several choices of $\kappa_{e}$ at $g=5$ and $\kappa_{i}=1$ (Fig. 3(b)). The boundary $\partial\tilde{\mathsf{R}}$ may be composed of one or more types of the protocols: reflectionless (red), low-rank Jacobian (green) and time-sharing (grey). For example, for $9.63<\kappa_{e}<9.68$ the boundary contains all three types of protocols (Fig. 3(b)ii) while for $\kappa_{e}>11.28$ or $\kappa_{e}<6.2$ the optimized protocols is time-shared unidirectional transduction. It is also worth mentioning that for $6.2<\kappa_{e}<11.28$ the transducers benefit from the simultaneous duplex transduction.

If both $\kappa_{e}$ and detunings are tunable, it can be proved that the highest transmission rate $T$ occurs when $R_{1}=R_{2}=0$ (see Appendix C). Therefore the optimal duplex quantum transduction is achieved with the two-side reflectionless condition at $\kappa_{e}=\sqrt{4g^{2}+\kappa_{i}^{2}}$ and $\Delta_{1}=\Delta_{2}=0$ (Fig. 3(b)iv). In other words, $\tilde{\mathsf{R}}(\sqrt{4g^{2}+\kappa_{i}^{2}})$ is the largest possible region in the sense that $\tilde{\mathsf{R}}(\kappa_{e})\subseteq\tilde{\mathsf{R}}(\sqrt{4g^{2}+\kappa_{i}^{2}})$ for any $\kappa_{e}$.

Frequency-integrated rate region.—Quantum transducer usually has a finite conversion bandwidth, which determines the range of signal frequencies that can be converted efficiently Zeuthen et al. (2020). Larger bandwidth enables higher operation speed of the transducer, and is preferable in presence of decoherence. Within the bandwidth, quantum signals at multiple frequencies can be transduced independently with frequency-dependent conversion efficiencies.

We can perform duplex quantum transduction in parallel for various signal frequencies, and the frequency-dependent scattering matrices result in distinct achievable rate regions that vary with frequency. Let $\omega_{2}=\omega_{p}+\omega_{1}$ for the transducer model Eq. (11), the signal detuning in the rotating frame becomes $\Delta_{1}=\Delta_{2}\equiv\Delta$, and the frequency-dependent rate region is $\mathsf{R}(\Delta)\equiv\{(I_{1}(\Delta),I_{2}(\Delta))|\forall(N_{1},N_{2})\}$. We plot the time-sharing rate regions $\tilde{\mathsf{R}}(\Delta)$ for multiple $\Delta$ (Fig. 4(a)), and compare $\max[I_{1}(\Delta)+I_{2}(\Delta)]$ with the quantum capacity $I_{\text{max}}(\Delta)$ (Fig. 4(b)). For $\Delta$ within the grey shaded region, simultaneous duplex transduction is advantageous, while outside this region time-shared unidirectional transduction is the optimal protocol.

To obtain the total achievable rate region, we sum the contributions from the individual rate regions at each signal frequency (Fig. 4(a)). The frequency-integrated rate region for time-shared duplex transduction is defined as

where $\{\omega_{j}\}$ is a set of evenly spaced frequencies with a frequency spacing $\delta$, and $\mathsf{A}\oplus\mathsf{B}\equiv\{\bm{a}+\bm{b}|\bm{a}\in\mathsf{A},\bm{b}\in\mathsf{B}\}$ is the Minkowski sum De Berg et al. (2008) of two sets $\mathsf{A}$ and $\mathsf{B}$. For general sets the complexity of Minkowski sum is $O(|A||B|)$, while for convex sets $\mathsf{A}$ and $\mathsf{B}$ in $\mathbb{R}^{2}$ the complexity is $O(|\partial A|+|\partial B|)$ De Berg et al. (2008). Therefore numerical evaluation of $\tilde{\mathsf{R}}_{\text{tot}}$ is efficient since $\tilde{\mathsf{R}}$ is convex in $\mathbb{R}^{2}$.

We calculate the Minkowski sum of all rate regions $\tilde{\mathsf{R}}(\Delta)$ in Fig. 4(a), and the resulting frequency-integrated rate region $\tilde{\mathsf{R}}_{\text{tot}}$ is shown in Fig. 4(c). The boundary $\partial\tilde{\mathsf{R}}_{\text{tot}}$ can be achieved with frequency dependent protocols. For example, to realize the orange part of $\partial\tilde{\mathsf{R}}_{\text{tot}}$ with a slope of -1, we choose the simultaneous duplex transduction protocol that maximizes $I_{1}(\Delta)+I_{2}(\Delta)$ (Fig. 4(a) orange dots) for signal detuning $\Delta$ within the grey shaded region in Fig. 4(b). For other $\Delta$, we perform the time-shared unidirectional transduction. Benefiting from simultaneous duplex transduction, the frequency-integrated rate region outperforms the time-shared unidirectional transduction (Fig. 4(c) grey dashed line).

Discussion.—We proposed the (time-sharing) rate region to quantify the performance of duplex quantum transduction, and studied optimized protocols for a two-mode quantum transducer. Unlike unidirectional quantum transduction, duplex quantum transduction is influenced by the reflection coefficients, and we explored how the reflectionless condition can be related to the optimal duplex quantum transduction. Furthermore, we incorporated the finite bandwidth of the transducer and introduced the frequency-integrated rate region. In future works, it would be interesting to consider non-Gaussian encodings (see Appendix E), as well as other approaches to quantum transduction such as adaptive control Zhang et al. (2018), shared entanglement Zhong et al. (2020a, b); Wu et al. (2021) and interference-based methods Lau and Clerk (2019); Zhang et al. (2022). Our method can also be extended to analyze the performance of multiplex quantum hardware with more than two quantum channels, such as characterizing the performance of a 3-port quantum circulator with a 3-dimensional rate region. Exploring alternative performance metrics Siddiqui and Wilde (2022) for duplex quantum transduction may provide further insights.