Continuous-Variable Quantum Reservoir Computers

• Continuous-variable quantum reservoir computers are architectures that use bosonic modes—such as harmonic oscillators and photonic fields—to create high-dimensional Hilbert spaces with inherent fading memory.
• They employ diverse dynamical evolutions including Hamiltonian/Lindblad dynamics, symplectic Gaussian updates, and measurement-based teleportation to encode inputs via displacements, squeezing, or cavity drives.
• These systems have demonstrated promising performance for tasks like chaotic forecasting and classification while highlighting challenges in surpassing classical Gaussian baselines and managing measurement overhead.

Searching arXiv for papers on continuous-variable quantum reservoir computing and closely related CV-QRC architectures. Continuous-variable quantum reservoir computers (CV-QRCs) are reservoir-computing architectures in which the reservoir is realized by bosonic degrees of freedom—harmonic modes, optical fields, multimode Gaussian states, interacting bosons on lattices, or hybrid qubit-boson systems—and only a classical readout layer is trained. Across the literature, the reservoir state is evolved either by Hamiltonian/Lindblad dynamics, by symplectic Gaussian updates, or by measurement-based teleportation circuits, while inputs are injected through displacements, squeezing parameters, mode frequencies, homodyne bases, pump phases, or cavity drives. The defining technical premise is that high-dimensional Hilbert spaces, fading memory, and nonlinear feature generation can be obtained without training the internal quantum dynamics, with linear or logistic regression used at readout (Nokkala et al., 2020, Das et al., 30 Sep 2025).

1. Foundational formulations

A central formalization of CV-QRC uses a network of $N$ interacting quantum harmonic oscillators, with quadrature vector $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$, a symplectic timestep propagator $S(\Delta t)$, and a partition into reservoir and ancilla blocks. In discrete time, the reservoir update is written as

$$\mathbf x^R_k = A\,\mathbf x^R_{k-1}+B\,\mathbf x^A_k, \qquad \langle\mathbf x^R_k\rangle =A\,\langle\mathbf x^R_{k-1}\rangle+B\,\langle\mathbf x^A_k\rangle,$$

with an analogous covariance recursion

$$\sigma_k^R=A\,\sigma_{k-1}^R A^T+B\,\sigma_k^A B^T.$$

Within this framework, the spectral-radius condition $\rho(A)<1$ is equivalent to the echo-state property and the fading-memory property on bounded input sets, and universality for fading-memory maps is obtained from separation plus Stone–Weierstrass arguments. The same work shows that Gaussian reservoirs can be universal for reservoir computing, and that the degree of linear versus nonlinear memory can be tuned by the encoding itself rather than by introducing non-Gaussian dynamics (Nokkala et al., 2020).

This Gaussian-state viewpoint is complemented by measurement-based CV-QRC built from cluster states. There, an $N$-mode CV cluster state

$$|\Phi_{\rm cl}\rangle = \prod_{(i,j)\in G}\hat C_Z^{(i,j)}(\xi_{ij}) \bigotimes_{k=1}^N |0\rangle_{p_k}$$

acts as a reconfigurable resource, while local teleportation circuits implement a multimode unitary $\hat U(\mathbf s)$ determined by the measurement bases. The reservoir covariance and first moments obey explicit state-update equations after teleportation, beam splitting, and partial trace,

$$\Gamma_R^{(k+1)} =T\,U_{\mathbf s_k}\Gamma_R^{(k)}U_{\mathbf s_k}^T+(1-T)I_{2N}, \qquad \mathbf R^{(k+1)}=\sqrt T\,U_{\mathbf s_k}\mathbf R^{(k)},$$

so losses regulate forgetting time while the graph couplings and encoding weights inject processing structure (García-Beni et al., 2024).

A distinct foundational line emphasizes that nonlinearity can arise from quantum measurement even when the dynamical system is minimal. In the single Kerr-oscillator model,

$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$0

with Lindblad damping $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$1, the output is the quadrature expectation $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$2. Numerical simulations in this setting show quantum-classical performance improvement and identify the likely source as the nonlinearity of quantum measurement (Govia et al., 2020).

2. Reservoir substrates and physical realizations

The CV-QRC literature spans Gaussian and non-Gaussian photonics, superconducting circuit QED, interacting bosonic lattices, and hybrid spin-boson systems.

Architecture	Representative formulation	Noted capability
Harmonic-network Gaussian reservoir	Symplectic update of moments and covariances	Universal fading-memory reservoir computing
Superconducting oscillator + qubit	Dispersive cavity-transmon Hamiltonian	Analog microwave signal processing
Multimode squeezed-state photonic reservoir	PDC-generated supermodes with feedback	Controlled fading memory
Cluster-state measurement-based reservoir	CZ-coupled squeezed states + teleportation	Static and temporal tasks without hardware modification
Bose-Hubbard lattice reservoir	1D interacting bosons with homogeneous couplings	Task-dependent optimality in chaotic or weak-interaction regimes
Hybrid qubit-boson JC/DJC reservoir	Jaynes–Cummings and dispersive Jaynes–Cummings models	Superior nonlinear over linear memory capacity

In superconducting hardware, a high-$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$3 microwave cavity dispersively coupled to a transmon realizes an analog reservoir with effective Hamiltonian

$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$4

where the reservoir directly accepts continuous-time microwave signals rather than discretized gate inputs. The system was operated without artificially discretizing the input data, and classification tasks were performed by repeated QND measurements of qubit state and oscillator parity (Senanian et al., 2023).

In optical CV-QRC, one experimental platform uses parametric down-conversion in ppKTP pumped by a $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$5, $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$6 pulsed laser, with interaction Hamiltonian

$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$7

and Schmidt decomposition $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$8. In the degenerate case, the device generates a product of single-mode squeezed vacua (“supermodes”), with up to $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$9 modes $S(\Delta t)$0 accessed. Programmable phase shaping and mode-selective homodyne detection then define the reservoir coordinates (Paparelle et al., 8 Jun 2025).

Integrated photonic circuits provide another substrate: a one-dimensional array of $S(\Delta t)$1 identical single-mode silicon waveguides coupled by directional couplers with Hamiltonian density

$S(\Delta t)$2

In the reported configuration, only one input mode is nonclassical, prepared as a small-amplitude even-cat “kitten” state, while the other modes carry coherent-state encodings (Świerczewski et al., 17 Mar 2026).

Bosonic many-body reservoirs also appear in atomic-lattice realizations. A 1D Bose-Hubbard chain with open boundary conditions uses

$S(\Delta t)$3

and transitions among Mott-insulator-like, chaotic, and superfluid-like regimes are controlled by $S(\Delta t)$4 (Llodrà et al., 2024).

Hybrid qubit-boson systems extend the CV-QRC notion by combining a single bosonic mode with a two-level system. The resonant Jaynes–Cummings model

$S(\Delta t)$5

and its dispersive limit

$S(\Delta t)$6

supply high-dimensional Hilbert spaces and intrinsic nonlinear dynamics in minimal spin-boson configurations (Das et al., 30 Sep 2025).

3. Input encoding, observables, and readout

Input injection in CV-QRC is highly architecture-dependent, but most schemes encode classical time series into physically native control parameters.

In Gaussian photonic loop reservoirs, a classical sequence $S(\Delta t)$7 is encoded into $S(\Delta t)$8 ancillary modes by preparing each ancilla in a zero-mean Gaussian state whose squeezing parameter $S(\Delta t)$9 is an affine function of $$\mathbf x^R_k = A\,\mathbf x^R_{k-1}+B\,\mathbf x^A_k, \qquad \langle\mathbf x^R_k\rangle =A\,\langle\mathbf x^R_{k-1}\rangle+B\,\langle\mathbf x^A_k\rangle,$$0, for example $$\mathbf x^R_k = A\,\mathbf x^R_{k-1}+B\,\mathbf x^A_k, \qquad \langle\mathbf x^R_k\rangle =A\,\langle\mathbf x^R_{k-1}\rangle+B\,\langle\mathbf x^A_k\rangle,$$1. The ancilla and reservoir modes interfere on a beam splitter of reflectivity $$\mathbf x^R_k = A\,\mathbf x^R_{k-1}+B\,\mathbf x^A_k, \qquad \langle\mathbf x^R_k\rangle =A\,\langle\mathbf x^R_{k-1}\rangle+B\,\langle\mathbf x^A_k\rangle,$$2 and pass through $$\mathbf x^R_k = A\,\mathbf x^R_{k-1}+B\,\mathbf x^A_k, \qquad \langle\mathbf x^R_k\rangle =A\,\langle\mathbf x^R_{k-1}\rangle+B\,\langle\mathbf x^A_k\rangle,$$3 nonlinear crystals, and the $$\mathbf x^R_k = A\,\mathbf x^R_{k-1}+B\,\mathbf x^A_k, \qquad \langle\mathbf x^R_k\rangle =A\,\langle\mathbf x^R_{k-1}\rangle+B\,\langle\mathbf x^A_k\rangle,$$4 cycle is repeated $$\mathbf x^R_k = A\,\mathbf x^R_{k-1}+B\,\mathbf x^A_k, \qquad \langle\mathbf x^R_k\rangle =A\,\langle\mathbf x^R_{k-1}\rangle+B\,\langle\mathbf x^A_k\rangle,$$5 times to build measurement statistics (Hahto et al., 4 Jul 2025).

In the hybrid JC/DJC reservoir, the discrete time series $$\mathbf x^R_k = A\,\mathbf x^R_{k-1}+B\,\mathbf x^A_k, \qquad \langle\mathbf x^R_k\rangle =A\,\langle\mathbf x^R_{k-1}\rangle+B\,\langle\mathbf x^A_k\rangle,$$6 is injected by setting the cavity-drive amplitude $$\mathbf x^R_k = A\,\mathbf x^R_{k-1}+B\,\mathbf x^A_k, \qquad \langle\mathbf x^R_k\rangle =A\,\langle\mathbf x^R_{k-1}\rangle+B\,\langle\mathbf x^A_k\rangle,$$7 for a duration $$\mathbf x^R_k = A\,\mathbf x^R_{k-1}+B\,\mathbf x^A_k, \qquad \langle\mathbf x^R_k\rangle =A\,\langle\mathbf x^R_{k-1}\rangle+B\,\langle\mathbf x^A_k\rangle,$$8. No “erase-and-write” reset is performed; instead, continuous drive plus dissipation implements a natural fading memory. Time multiplexing is introduced by sampling reservoir observables at $$\mathbf x^R_k = A\,\mathbf x^R_{k-1}+B\,\mathbf x^A_k, \qquad \langle\mathbf x^R_k\rangle =A\,\langle\mathbf x^R_{k-1}\rangle+B\,\langle\mathbf x^A_k\rangle,$$9 equally spaced times within each interval $$\sigma_k^R=A\,\sigma_{k-1}^R A^T+B\,\sigma_k^A B^T.$$0, multiplying the number of features from $$\sigma_k^R=A\,\sigma_{k-1}^R A^T+B\,\sigma_k^A B^T.$$1 to $$\sigma_k^R=A\,\sigma_{k-1}^R A^T+B\,\sigma_k^A B^T.$$2 (Das et al., 30 Sep 2025).

Multivariate CV-QRC with oscillator networks encodes a $$\sigma_k^R=A\,\sigma_{k-1}^R A^T+B\,\sigma_k^A B^T.$$3-dimensional real-valued input $$\sigma_k^R=A\,\sigma_{k-1}^R A^T+B\,\sigma_k^A B^T.$$4 into the mode frequencies $$\sigma_k^R=A\,\sigma_{k-1}^R A^T+B\,\sigma_k^A B^T.$$5 through three schemes: local encoding, clustered encoding, and global encoding. These include mappings such as

$$\sigma_k^R=A\,\sigma_{k-1}^R A^T+B\,\sigma_k^A B^T.$$6

for local injection, and

$$\sigma_k^R=A\,\sigma_{k-1}^R A^T+B\,\sigma_k^A B^T.$$7

for global injection with row-normalized $$\sigma_k^R=A\,\sigma_{k-1}^R A^T+B\,\sigma_k^A B^T.$$8 (Fellner et al., 9 Apr 2026).

Measurement-based cluster-state reservoirs encode a scalar input $$\sigma_k^R=A\,\sigma_{k-1}^R A^T+B\,\sigma_k^A B^T.$$9 through the local homodyne basis

$\rho(A)<1$0

while vector inputs can be distributed across modes by assigning different components to different measurement bases (García-Beni et al., 2024). In the optical memory-control platform, input and memory are merged in the feedback law

$\rho(A)<1$1

or, for global-phase encoding,

$\rho(A)<1$2

so past observables modulate the current pump (Paparelle et al., 8 Jun 2025).

Readout observables are equally diverse. JC/DJC reservoirs employ bosonic-mode observables such as

$\rho(A)<1$3

with $\rho(A)<1$4 or $\rho(A)<1$5, yielding $\rho(A)<1$6 features per time point. The readout is trained by ridge regression with regularization strength $\rho(A)<1$7, solving $\rho(A)<1$8 (Das et al., 30 Sep 2025).

Gaussian photonic reservoirs conventionally use homodyne statistics, estimating means $\rho(A)<1$9 and covariances $N$0, but improved protocols instead sample the measurement CDF. For thresholds $N$1, univariate nodes are $N$2, and bivariate nodes are $N$3. These replace moment-only readout by nonlinear features of the full measurement distribution (Hahto et al., 4 Jul 2025).

Other platforms read out quadrature covariances $N$4 (Fellner et al., 9 Apr 2026), second moments $N$5 from cluster-state modes (García-Beni et al., 2024), low-order quadrature moments $N$6 extracted from transformed covariance matrices (Paparelle et al., 8 Jun 2025), parity and qubit-state bits aggregated into trajectory moments (Senanian et al., 2023), or a feature vector

$N$7

in integrated optics (Świerczewski et al., 17 Mar 2026).

Training remains classically simple. Reported readouts include ridge regression (Nokkala et al., 2020), ordinary least squares (Fellner et al., 9 Apr 2026), logistic regression with $N$8 (Świerczewski et al., 17 Mar 2026), and softmax classifiers for multiclass tasks in cluster-state QRC (García-Beni et al., 2024). The reservoir parameters are typically fixed; only the linear or logistic output map is adapted.

4. Memory, nonlinearity, and capacity measures

CV-QRC performance is usually analyzed through memory-capacity decompositions, nonlinear benchmark tasks, and information-processing measures.

For JC/DJC reservoirs, linear short-term memory (STM) uses target $N$9, while nonlinear parity-check memory uses

$$|\Phi_{\rm cl}\rangle = \prod_{(i,j)\in G}\hat C_Z^{(i,j)}(\xi_{ij}) \bigotimes_{k=1}^N |0\rangle_{p_k}$$0

The delay-dependent capacity is the squared Pearson correlation

$$|\Phi_{\rm cl}\rangle = \prod_{(i,j)\in G}\hat C_Z^{(i,j)}(\xi_{ij}) \bigotimes_{k=1}^N |0\rangle_{p_k}$$1

with totals $$|\Phi_{\rm cl}\rangle = \prod_{(i,j)\in G}\hat C_Z^{(i,j)}(\xi_{ij}) \bigotimes_{k=1}^N |0\rangle_{p_k}$$2 and $$|\Phi_{\rm cl}\rangle = \prod_{(i,j)\in G}\hat C_Z^{(i,j)}(\xi_{ij}) \bigotimes_{k=1}^N |0\rangle_{p_k}$$3. A striking reported result is that in both JC and DJC reservoirs,

$$|\Phi_{\rm cl}\rangle = \prod_{(i,j)\in G}\hat C_Z^{(i,j)}(\xi_{ij}) \bigotimes_{k=1}^N |0\rangle_{p_k}$$4

with sample JC values $$|\Phi_{\rm cl}\rangle = \prod_{(i,j)\in G}\hat C_Z^{(i,j)}(\xi_{ij}) \bigotimes_{k=1}^N |0\rangle_{p_k}$$5, $$|\Phi_{\rm cl}\rangle = \prod_{(i,j)\in G}\hat C_Z^{(i,j)}(\xi_{ij}) \bigotimes_{k=1}^N |0\rangle_{p_k}$$6, and $$|\Phi_{\rm cl}\rangle = \prod_{(i,j)\in G}\hat C_Z^{(i,j)}(\xi_{ij}) \bigotimes_{k=1}^N |0\rangle_{p_k}$$7, $$|\Phi_{\rm cl}\rangle = \prod_{(i,j)\in G}\hat C_Z^{(i,j)}(\xi_{ij}) \bigotimes_{k=1}^N |0\rangle_{p_k}$$8 for $$|\Phi_{\rm cl}\rangle = \prod_{(i,j)\in G}\hat C_Z^{(i,j)}(\xi_{ij}) \bigotimes_{k=1}^N |0\rangle_{p_k}$$9, $\hat U(\mathbf s)$0, $\hat U(\mathbf s)$1 (Das et al., 30 Sep 2025).

In Gaussian photonic CV-QRC, capacity is often expressed through NMSE-based scores $\hat U(\mathbf s)$2 and the total information processing capacity

$\hat U(\mathbf s)$3

Replacing covariances by CDF readout increases IPC substantially. In the ideal $\hat U(\mathbf s)$4 setting, the normalized comparisons are: covariances-only $\hat U(\mathbf s)$5, univariate CDF with $\hat U(\mathbf s)$6 gives $\hat U(\mathbf s)$7, bivariate CDF with $\hat U(\mathbf s)$8 gives $\hat U(\mathbf s)$9, and bivariate CDF plus $$\Gamma_R^{(k+1)} =T\,U_{\mathbf s_k}\Gamma_R^{(k)}U_{\mathbf s_k}^T+(1-T)I_{2N}, \qquad \mathbf R^{(k+1)}=\sqrt T\,U_{\mathbf s_k}\mathbf R^{(k)},$$0 classical-memory steps yields $$\Gamma_R^{(k+1)} =T\,U_{\mathbf s_k}\Gamma_R^{(k)}U_{\mathbf s_k}^T+(1-T)I_{2N}, \qquad \mathbf R^{(k+1)}=\sqrt T\,U_{\mathbf s_k}\mathbf R^{(k)},$$1, averaged over $$\Gamma_R^{(k+1)} =T\,U_{\mathbf s_k}\Gamma_R^{(k)}U_{\mathbf s_k}^T+(1-T)I_{2N}, \qquad \mathbf R^{(k+1)}=\sqrt T\,U_{\mathbf s_k}\mathbf R^{(k)},$$2 random realizations (Hahto et al., 4 Jul 2025).

A different capacity concept is required for multivariate inputs. The multivariate CV framework defines the mixing capacity using targets

$$\Gamma_R^{(k+1)} =T\,U_{\mathbf s_k}\Gamma_R^{(k)}U_{\mathbf s_k}^T+(1-T)I_{2N}, \qquad \mathbf R^{(k+1)}=\sqrt T\,U_{\mathbf s_k}\mathbf R^{(k)},$$3

with coefficient of determination

$$\Gamma_R^{(k+1)} =T\,U_{\mathbf s_k}\Gamma_R^{(k)}U_{\mathbf s_k}^T+(1-T)I_{2N}, \qquad \mathbf R^{(k+1)}=\sqrt T\,U_{\mathbf s_k}\mathbf R^{(k)},$$4

and total

$$\Gamma_R^{(k+1)} =T\,U_{\mathbf s_k}\Gamma_R^{(k)}U_{\mathbf s_k}^T+(1-T)I_{2N}, \qquad \mathbf R^{(k+1)}=\sqrt T\,U_{\mathbf s_k}\mathbf R^{(k)},$$5

Because each term probes products of independent streams at arbitrary delays, $$\Gamma_R^{(k+1)} =T\,U_{\mathbf s_k}\Gamma_R^{(k)}U_{\mathbf s_k}^T+(1-T)I_{2N}, \qquad \mathbf R^{(k+1)}=\sqrt T\,U_{\mathbf s_k}\mathbf R^{(k)},$$6 measures nonlinear cross-stream mixing directly (Fellner et al., 9 Apr 2026).

The optical memory-control platform defines linear memory capacity as

$$\Gamma_R^{(k+1)} =T\,U_{\mathbf s_k}\Gamma_R^{(k)}U_{\mathbf s_k}^T+(1-T)I_{2N}, \qquad \mathbf R^{(k+1)}=\sqrt T\,U_{\mathbf s_k}\mathbf R^{(k)},$$7

and explicitly distinguishes real-time feedback memory from long-term dependencies created by spatial multiplexing. Experimentally, capacities up to $$\Gamma_R^{(k+1)} =T\,U_{\mathbf s_k}\Gamma_R^{(k)}U_{\mathbf s_k}^T+(1-T)I_{2N}, \qquad \mathbf R^{(k+1)}=\sqrt T\,U_{\mathbf s_k}\mathbf R^{(k)},$$8 were reported for $$\Gamma_R^{(k+1)} =T\,U_{\mathbf s_k}\Gamma_R^{(k)}U_{\mathbf s_k}^T+(1-T)I_{2N}, \qquad \mathbf R^{(k+1)}=\sqrt T\,U_{\mathbf s_k}\mathbf R^{(k)},$$9, $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$00, with decay as $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$01 increases (Paparelle et al., 8 Jun 2025).

Interacting bosonic lattices add a regime-based perspective. In the Mott-like regime $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$02, information fails to propagate and temporal memory is essentially absent; in the chaotic regime $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$03, level statistics cross to Wigner–Dyson behavior and both nontrivial memory and nonlinearity emerge; in the superfluid regime $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$04, memory persists over many $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$05 but nonlinear transformations are weaker. Benchmark capacities reflect this structure: for STM, $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$06 in the chaotic regime and $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$07 in the superfluid regime, while for parity-check tasks the superfluid regime can be superior (Llodrà et al., 2024).

Across platforms, expressivity is frequently enhanced by feature expansion rather than by parameter training: time multiplexing creates “virtual nodes” (Das et al., 30 Sep 2025), storing past outputs augments rank and mitigates Wishart noise (Hahto et al., 4 Jul 2025), and spatial multiplexing combines delayed components across multiple copies (Paparelle et al., 8 Jun 2025).

5. Benchmarks and empirical results

The benchmark suite for CV-QRC is broad, ranging from synthetic memory tasks to analog signal classification, chaotic forecasting, and image recognition.

For chaotic time-series prediction, JC/DJC reservoirs were tested on the Mackey–Glass system

$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$08

using both autonomous generation and one-step-ahead forecasting. Performance was measured by the normalized root-mean-square error

$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$09

For $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$10-step forecasts on a single MG segment, the JC reservoir achieved autonomous NRMSE $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$11 without multiplexing and $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$12 at $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$13, and one-step NRMSE $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$14 at $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$15 and $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$16 at $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$17. The DJC reservoir gave autonomous NRMSE $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$18 and one-step NRMSE $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$19 as multiplexing increased (Das et al., 30 Sep 2025).

In analog superconducting QRC, the reservoir classified multiple microwave-signal families without digitizing the input. On a two-arm spiral task in the $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$20 plane, the system reached accuracy $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$21 at $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$22 shots, while a linear classifier on $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$23 saturated at $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$24. On RF-modulation classification with ten digital modulation formats encoded at $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$25, accuracy was $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$26 in fewer than $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$27 shots, compared to $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$28 for a linear baseline. On filtered-noise classification, the reservoir reached $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$29 accuracy at $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$30 shots (Senanian et al., 2023).

The optical feedback platform demonstrates controlled memory and nonlinear temporal processing in a fully CV setting. Reported results include XOR accuracy $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$31 test accuracy with only $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$32 training points, a best linear memory capacity of $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$33 at $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$34, simulated parity-check accuracy $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$35 for $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$36 with general encoding, and double-scroll one-step prediction test capacities $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$37, $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$38, $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$39 using $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$40 reservoirs and $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$41 training points (Paparelle et al., 8 Jun 2025).

Measurement-based cluster-state QRC has been evaluated on both temporal and static tasks. Temporal XOR reached $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$42 test accuracy with a two-mode vacuum start and logistic readout. For NARMA$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$43, NMSE fell below $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$44 for $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$45 and $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$46. For MNIST, images compressed to $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$47 and injected column-by-column over $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$48 time steps into an $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$49 reservoir yielded a best test accuracy of $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$50 with a ring cluster and softmax readout (García-Beni et al., 2024).

Multivariate CV-QRC was benchmarked on Lorenz-63 prediction with data sampled at $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$51, using $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$52 steps partitioned into $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$53 washout, $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$54 training, and $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$55 test points. For $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$56, the CV reservoir outperformed the DV baseline: clustered encoding yielded NRMSE $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$57, and global encoding $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$58–$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$59 when both $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$60 inputs were used. Feeding multiple correlated components systematically lowered error compared with univariate input (Fellner et al., 9 Apr 2026).

Integrated optical CV-QRC with a single kitten state showed that higher-order correlations can dominate static classification performance. On a two-spirals dataset with $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$61 points, repeated $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$62 times with random splits of $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$63 training and $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$64 testing points, occupation-only baselines gave $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$65 accuracy both with all-classical inputs and with a single kitten if only $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$66 were used. Adding six $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$67 features raised accuracy to $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$68, corresponding to an error reduction factor $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$69 (Świerczewski et al., 17 Mar 2026).

The single nonlinear oscillator model also reported a systematic quantum-classical separation on sine-phase estimation. For moderate training-set sizes $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$70–$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$71, the QRC’s mean RMS error could be an order of magnitude smaller than the classical reservoir’s, and the error spread across parameter draws was markedly reduced (Govia et al., 2020).

6. Limits, quantum resources, and research directions

A recurring issue in CV-QRC is the distinction between what Gaussian reservoirs can already do and what additional resources are required for stronger nonlinear temporal processing.

Gaussian CV-QRC is classically efficiently simulable, so it defines a baseline rather than an automatic quantum advantage claim. Improved use of measurement statistics significantly raises that baseline: Gaussian covariance-only IPC is normalized to $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$72, whereas bivariate CDF sampling plus $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$73 classical memory reaches $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$74. The corresponding conclusion is explicit: any non-Gaussian CV-QRC claiming quantum advantage must surpass the combined capacity of Gaussian covariance+CDF+memory schemes, approximately $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$75 the conventional covariance IPC in that setting (Hahto et al., 4 Jul 2025).

Several works identify nonclassical resources that correlate with improved performance. In JC/DJC reservoirs, the bosonic mode introduces non-Gaussian dynamics, Wigner-negativity, and tunable nonlinearity relative to comparable two-qubit reservoirs (Das et al., 30 Sep 2025). In multivariate oscillator networks, peak mixing capacity and lowest Lorenz NRMSE coincide with the regime of moderate squeezing, suggesting that squeezing enhances the reservoir’s ability to mix and predict multivariate dynamics (Fellner et al., 9 Apr 2026). In integrated optics, a single kitten-state input combined with $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$76 features suffices to obtain the reported $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$77 reduction of classification error over the classical counterpart (Świerczewski et al., 17 Mar 2026).

A sharper limitation appears in recent work on cross-time nonlinear processing. For Gaussian reservoirs at minimal readout degree $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$78, connected Volterra kernels of order $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$79 are constrained by a rank ceiling set by the mean-response dimension: an $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$80-mode Gaussian reservoir reaches cross-time nonlinear rank at most $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$81. By contrast, a single Kerr mode in a delayed feedback loop reaches rank $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$82, where $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$83 is the feedback depth. The reported theorem and corollary imply an unbounded resource separation: for any finite $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$84, choosing $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$85 yields a connected kernel that no $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$86-mode Gaussian reservoir can match (Soh, 4 Jun 2026).

The physical mechanism in that model is the combination of Kerr nonlinearity and non-Markovian feedback. With Kerr Hamiltonian

$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$87

and per-trip power survival $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$88, the $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$89-th echo acquires phase

$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$90

Loss is not merely detrimental: because $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$91 becomes strictly monotonic in $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$92 for $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$93, each round-trip is fingerprinted by a different phase. The achievable feedback depth is estimated as

$\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$94

and integrated-photonics-typical $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$95 gives $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$96, so one Kerr mode can replace up to $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$97 Gaussian modes, at the price of measurement time (Soh, 4 Jun 2026).

Scalability claims in the literature are correspondingly nuanced. Optical multimode squeezing already accesses up to $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$98 modes (Paparelle et al., 8 Jun 2025); Gaussian harmonic networks offer observable counts scaling as $\mathbf x=\{q_1,p_1,\dots,q_N,p_N\}^T$99 for first moments and $S(\Delta t)$00 for covariances (Nokkala et al., 2020); positive-$S(\Delta t)$01 simulation of integrated optical reservoirs scales linearly with mode number $S(\Delta t)$02 and avoids Hilbert-space cutoff (Świerczewski et al., 17 Mar 2026). At the same time, finite sampling, shot noise, and measurement overhead remain central constraints. This suggests that future comparisons among CV-QRC platforms will depend as much on the measurement model and feature extraction protocol as on the internal Hamiltonian alone.