betaFeedback

WorksPhysics and Astronomy

To predict or to persist: how decoherence shapes learning in a spin-qubit quantum reservoir

Abstract

In this work we investigate how dephasing and the choice of measured observables affect learning in a simulated six-qubit spin chain with Markovian dissipation and exact expectation-value readout. For externally driven nonlinear autoregressive moving-average signals, combining single- and two-qubit observables gives better predictions than a population-only readout at comparable polynomial feature counts. Most of the configurations we tested become more accurate than a reference with relaxation and dephasing disabled, most strongly at T2=1 μT_2=1\,μs, where total memory capacity drops, and still at T2=5T_2=5-10 μ10\,μs, where memory is only redistributed across delays. For autonomous generation, we train the readout with scheduled sampling and iterative dataset aggregation to expose it to states reached under its own feedback. We do not find a clear systematic dependence of limit-cycle persistence on coherence across the reservoir ensemble. For chaotic forecasting, in contrast, prediction horizons at T2=1 μT_2=1\,μs shorten in all 10 reservoir realizations relative to that reference. For the noise-driven strange non-chaotic attractor, the model produces a smooth, approximately closed orbit, and neither captures the target's scatter nor follows its trajectory pointwise. We conclude that, within this simulated reservoir, the impact of coherence depends strongly on the task.

Introduction

Predicting a changing signal requires a model to retain useful information about the past and combine it with new observations. Reservoir computing (RC) approaches this problem by using the response of a dynamical system to an input sequence. It exploits the fact that if the system used is complex enough, it performs most of the computation that a recurrent neural network is otherwise required to learn [1, 2, 3]. After the system is driven with a temporal input, one proceeds by reading out a set of internal state variables and fitting a linear or low-degree polynomial map from those variables to a target sequence. Since only the readout is trained, the training reduces to a convex regression problem.

If we use a physical dynamical system as a reservoir, it reduces the need to train internal parameters. This has motivated implementation in photonics, spintronics and nanoelectronic materials [1, 4, 5, 6]. Quantum reservoir computing (QRC) is based on the same philosophy, only using a quantum system as a reservoir [7, 8, 9, 10, 11]. This ranges from superconducting qubits [12, 13], photonic [14, 15, 16], nuclear magnetic resonance (NMR) devices [17], single nonlinear oscillators [18], analog Rydberg-atom machines [19, 20], to interacting spins [21, 22] and applications to state tomography and quantum circuit compression [23, 24, 25]. Recent reviews collect the state of the field [26, 27]. Since the quantum state space is exponentially large in the number of qubits, we can in principle use a small number of physical degrees of freedom to get a large number of nonlinear functions of the input history. The role of a classical neural network is therefore substituted by the reservoir’s Hilbert space [28, 29, 30].

An open debate in the field is whether decoherence helps or hinders learning, and it is important for any potential use of real hardware as a quantum (reservoir) computer. So far in the literature there are arguments on both sides. On the one hand, the echo state property requires the decay of the influence of the past (i.e., finite memory). This favors (quantum) dissipation as a natural way for a quantum system to forget about its previous states and make the map well defined [31, 32]. A recent simulation of a six-qubit Heisenberg-coupled spin-qubit array found that larger decay rates improved performance on NARMA tasks up to order n=10n = 10, because they stopped the reservoir retaining too much of its input history [33]. On that note, ref. [34] proposes controllable damping as a design parameter of a system. On the other hand, coherence generates the interference between the many nonlinear functions that the Hilbert space supplies in the first place. A many-body study reports a monotonic relationship between coherence and information processing capacity in a transverse-field Ising reservoir, and finds that finite measurement resources and noise limit how far quantum effects can be exploited [35].

We believe that the disagreement is partly a matter of the specific task performed. Specifically, if a reservoir is required to reproduce a nonlinear filter of its input then it needs memory. If it is required to sustain a limit cycle autonomously it needs to have learned an attracting orbit; such orbits usually absorb small perturbations by construction. Conversely, if it is required to track a chaotic trajectory, Lyapunov divergence amplifies the perturbations. We find no reason for three such different demands on the same hardware to respond to decoherence in a similar fashion. In order to test whether coherence dependence changes with the task within a single model, we fix the physical parameters to a realistic operating point and sweep the coherence times across a range.

The reservoir we simulate here is a chain of six electron-spin qubits of the kind realized in Si/SiGe heterostructures, coupled by Heisenberg exchange and addressed through the field gradient of a micromagnet. This is a platform that is able to run multi-qubit circuits in practice [36, 37, 38], and experiments have directly characterized six-dot arrays of this kind. Their relaxation times are of the order T1∼100T_1 \sim100 ms and free-evolution coherence times of order T2∼10T_2 \sim10 µs at mK temperatures [39, 40, 41]. Here we hold T1=100T_1 = 100 ms and sweep T2T_2 from 1–5000 µs, together with a unitary evolution reference. The parameters we use for the configuration of the reservoir are representative of the platform and not a specific chip. We drive the reservoir stochastically with a non-zero reset probability [42]. Moreover, alongside the coherence sweep we test different sets of readout quantum observables used as learning features. These range from the six single-qubit populations ⟨Zi⟩\langle Z_i\rangle up to a set of 57 single-qubit and two-qubit expectation values spanning local X/ZX/Z measurements arranged in five settings. We benchmark on three temporal tasks with three different requirements sketched above: i) the nn-order nonlinear auto-regressive moving average (NARMA-nn) family, driven open-loop; ii) the Mackey–Glass (MG) system [43] generated autonomously, in both its non-chaotic and chaotic regimes, and iii) a noise-induced strange non-chaotic attractor (SNA) [44, 45], which is also autonomous. Figure 1 summarizes the scheme.

Figure 1. Schematic of the QRC protocol. The input uku_k controls the mixture of two completely positive trace-preserving (CPTP) maps T0\mathcal{T}_0 and T1\mathcal{T}_1, which differ in the phase ϕ\phi and amplitudes hih_i of the transverse drive, in exchange strength and in duration (2τ2\tau and 3τ3\tau). They act on a six-qubit chain with Heisenberg exchange JijJ_{ij} and a micromagnet field gradient that sets the detunings δi\delta_i. Every qubit relaxes with time T1=100T_1 = 100 ms and undergoes pure dephasing with time TφT_{\varphi}, where 1/Tφ=1/T2−1/(2T1)1/T_{\varphi} = 1/T_2 - 1/(2T_1) and we sweep the coherence time T2T_2. Five local X/ZX/Z measurement settings give 57 expectation values, and the ridge readout forms the prediction as a weighted sum of MM polynomial features Φj\Phi_j of these values. Only the weights WjW_j are trained. The input switch defines the driving mode: for NARMA the reservoir is driven by an external signal (open loop), while for MG and the SNA it is driven by its own prediction (closed loop), and the readout is trained in the loop with DAgger/scheduled sampling.

Autonomous generation requires a training procedure that is not necessarily included in standard QRC. Fitting the readout on teacher-forced states minimizes one-step error but places no constraint on the stability of the resulting closed-loop map. When the loop is closed, the trajectory can collapse onto a fixed point or drift away from the attractor. This is the same exposure-bias failure that is found in autoregressive sequence models [46, 47]. In order to mitigate it, we use scheduled sampling [48] combined with dataset aggregation in the manner of DAgger [49], so that we fit the readout on the off-attractor states reached by its own predictions, labeled with the true continuation (closed loop in Figure 1). We will show that on the MG tasks the in-loop fit is almost always selected over the teacher-forced one. Under the assumptions of our variance analysis, the reservoir realization is estimated to contribute more to performance variability than training randomness.

Here we investigate a device-grounded but still idealized reservoir, since we simulate the Lindblad dynamics and read out exact expectation values. This means that finite measurement statistics, readout infidelity, gate calibration error, and non-Markovian noise are not taken into account. The first one interacts with coherence [35] and we return to it in § 3. Readout infidelity depends on the charge sensor that detects the spin-to-charge conversion, such as a single-electron box [50]. We also treat the exchange couplings as a quenched disorder rather than as a controllable resource. This allows us to isolate the effect of coherence, and establish a baseline which could be employed by future work with a realistic measurement overhead.

The structure of the paper is as follows. In § 2 we present the results for all three tasks, covering what the readout observables buy and then the coherence dependence of each task. Also, we separate the contribution of the reservoir configuration from that of the training stochasticity. In § 3 we compare the findings across tasks and close with our conclusions and remarks. Finally, in § 4 we present the reservoir Hamiltonian and its Lindbladian used, the observable sets and the cost of the measurement settings for each of them, and the training and evaluation methodology, including the in-loop algorithm that makes autonomous generation possible.

Results

We test our quantum reservoir on three separate classes of timeseries, each one with different instantiations. The first one is the nonlinear autoregressive moving-average time series of order nn (NARMA-nn). The second one is the MG timeseries, with lags τMG=16,17\tau_{\mathrm{MG}} = 16, 17. The two different lags give rise to non-chaotic and chaotic dynamics, respectively. Finally, we attempt to predict a timeseries generated by the noise-driven map of Eq. (8), whose trajectory in the (yk,yk−1)(y_k, y_{k-1}) plane traces an SNA.

We sweep two axes throughout our calculations. One is coherence time T2T_2 from 1 μs1\ \mu\mathrm{s} to 5 ms5\ \mathrm{ms} at fixed T1=100 msT_1 = 100\ \mathrm{ms}, with a unitary reference in which the relaxation and dephasing are switched off in Eq. (2). The other is the quantum observable set O\mathcal{O} measured from the qubits, from 6 up to the full 57 of Eq. (9). Moreover, all open-loop results are reported for two independent reservoir instantiations, which we label R1\mathcal{R}_1 and R2\mathcal{R}_2. We show in § 2.5 that the choice of the reservoir is important and strongly affects the results.

Readout richness at matched feature count

We investigate whether the extra quantum observables offer any essential improvement in learning by themselves, or they only increase the number of regressors. We compare the full observable set at quadratic order generating 1710 features, against a ZZ-only readout of degree deg⁡=7\operatorname{deg} = 7 generating 1715 features. The two generate approximately the same number of features, to ensure a fair comparison in the regressor count and to distill the information content of them. We show the comparison in Table 1. We can see that the full set wins in all 27 cells of the grid on both reservoirs. We show the readout prediction in Figure 2 at T2=10 μsT_2 = 10\,\mu\mathrm{s}, the lower end of free-evolution coherence in Si/SiGe devices.

reservoirreadoutmedian NMSE, n=10n = 10n=15n = 15n=20n = 20# featureswins
R1\mathcal{R}_1OZ\mathrm{O}_{Z}, d=7d = 70.6320.6790.7401715—
R1\mathcal{R}_1Oall\mathrm{O}_{\mathrm{all}}, d=2d = 20.3050.4500.558171027/27
R2\mathcal{R}_2OZ\mathrm{O}_{Z}, d=7d = 70.6530.6900.7341715—
R2\mathcal{R}_2Oall\mathrm{O}_{\mathrm{all}}, d=2d = 20.2630.4810.576171027/27

Table 1. Observable readout at equal number of regressor terms. NMSE on NARMA-nn, median over the nine coherence columns of the sweep, including the unitary reference. Oall\mathrm{O}_{\mathrm{all}} at degree d=2d = 2 gives 1710 features and OZ\mathrm{O}_{Z} at d=7d = 7 gives 1715. The last column counts the cells in which the full set wins, out of 3 orders × 9 coherence values.

Figure 2. NARMA-nn timeseries prediction for orders n={10,15,20}n = \{10, 15, 20\}. We plot the full observable set Oall\mathcal{O}_{\mathrm{all}} for deg=2\mathrm{deg} = 2 on R1\mathcal{R}_1, at T2=10 μsT_2 = 10\ \mu\mathrm{s} on the three NARMA orders (columns). Top row: 250 out of the 750-step test window, target vs prediction, with the NMSE annotated. Bottom row: target vs prediction over the whole test window, on equal axes. We count the number of points that fall off the target axis limits. The prediction accuracy degrades with increasing orders as the memory demand grows. We can see (bottom row, from left to right) that the point cloud progressively scatters away from the diagonal, over-predicting smaller values while under-predicting larger ones. At n=20n = 20 in (f) the reservoir cannot predict the points beyond the center accurately. We highlight that T2=10 μsT_2 = 10\ \mu\mathrm{s} is the realistic device column instead of the best cell. Table 1 shows the medians over the whole coherence axis. R1\mathcal{R}_1 was also chosen for consistency, since from the same Table we see that R2\mathcal{R}_2 performs better on one of the three orders nn.

Given that it is not efficient to have 1715 regressors on six variables, we also ran the ZZ-only readout at deg⁡=6\operatorname{deg} = 6 which yields 923 features. Across the 54 matched cells, the median absolute NMSE difference between the deg⁡=6\operatorname{deg} = 6 and deg⁡=7\operatorname{deg} = 7 readouts is 8×10−58 \times10^{-5}, with a maximum of approximately 1.05×10−21.05 \times10^{-2}. This supports the idea that the improvement comes from the additional observables themselves and not from the number of monomials that participate in a regressor. Splitting the additional observables apart, the two-body correlators alone (O2q\mathrm{O}_{2q}, 1080 features) beat the ZZ-only readout at deg⁡=7\operatorname{deg} = 7 in all 27 cells on both reservoirs despite having fewer features. The single-qubit X/ZX/Z readout (O1q\mathrm{O}_{1q}, 1819 features) improves on the deg⁡=7\operatorname{deg} = 7 ZZ-only readout in 25/27 cells on R2\mathcal{R}_2. On R1\mathcal{R}_1, its median NMSE across the nine coherence settings differs from the deg⁡=7\operatorname{deg} = 7 ZZ-only baseline by less than 0.002 for each NARMA order. Since O2q\mathrm{O}_{2q} also contains the 15 ⟨ZaZb⟩\langle Z_a Z_b\rangle, which commute with the populations, these sets do not separate the role of non-commutativity from that of two-qubit correlations. The one-setting readout of § 4.3, containing the single-qubit ZZ observables and the 15 ⟨ZaZb⟩\langle Z_a Z_b\rangle correlators, would test how much improvement can be obtained from commuting observables alone. This would provide a useful control, although it would not by itself establish non-commutativity as the cause of any additional benefit from the full observable set. As we show in § 4.3, these extra quantum features also come at a measurement cost.

NARMA-nn: the effect of dephasing on memory

To study the effect of quantum coherence, we pool two reservoirs and five observable configurations over the three NARMA orders. This gives us a total of 30 paired comparisons against the unitary reference at each T2T_2. Since two of the five configurations are the same measurement at two polynomial degrees, excluding one leaves 24 paired comparisons. We can see in Figure 3 and Table 2, in the lower T2T_2 regime, that dephasing improves NARMA performance, and there is a gradual improvement. At T2=1 μsT_2 = 1\ \mu\mathrm{s} the reservoir outperforms its coherent version in 27/30 pairs with a median NMSE ratio of 0.786, at 5 μs\mu\mathrm{s} in 24/30, at 10 μs\mu\mathrm{s} in 23/30, and for ≥50 μs\ge50\ \mu\mathrm{s} the median NMSE ratio approaches unity.

Figure 3. Open-loop NARMA prediction accuracy against coherence time. The rows correspond to the two reservoir instantiations R1\mathcal{R}_1 and R2\mathcal{R}_2 and columns to the orders nn. We show three observables, OZ\mathcal{O}_Z-only at deg=7\mathrm{deg}=7, two-body correlators at deg=2\mathrm{deg}=2 and the full set Oall\mathcal{O}_{\mathrm{all}} at deg=2\mathrm{deg}=2. Each one has 1715, 1080 and 1710 regressors, respectively. The unitary value of each configuration is shown as a projected line with a marker. We can see that the OZ\mathcal{O}_Z readout is worse compared to the other two over the whole range and that the correlated readouts generally improve as T2T_2 decreases. On R2\mathcal{R}_2 their NMSE rises above the unitary value near T2=50T_2=50–100 μs100\ \mu\mathrm{s} and returns to it by 1000 μs1000\ \mu\mathrm{s}.

T2T_2 (μ\mus)better than unitarymedian NMSE ratio
127/300.786
524/300.879
1023/300.900
5012/301.016
10011/301.015
50012/301.003
100011/301.003
500010/301.000

Table 2. Coherence dependence of the NARMA timeseries. The statistics are calculated over two reservoirs, five observable configurations (i.e OZ\mathcal{O}_Z at d=6d = 6 and d=7d = 7, O1q\mathcal{O}_{1q} at d=4d = 4, and O2q\mathcal{O}_{2q} and Oall\mathcal{O}_{\mathrm{all}} at d=2d = 2) and three orders. In total, we have 30 paired comparisons against the unitary reference per row. A ratio below unity means the dephased reservoir is more accurate. Two of the five configurations, OZ\mathcal{O}_Z at degrees six and seven, are the same measurement and are not independent; dropping one of them leaves 24 pairs.

At first sight, our results appear inconsistent with the view that nonlinear filtering of an external input requires memory depth. If task performance were controlled primarily by memory, then shortening that memory should reduce accuracy. We investigate this apparent discrepancy by measuring the memory capacity (MC). Using standard construction, we fit each delay uk−du_{k-d} from the feature vector and record MCd=1−NMSEd\mathrm{MC}_d=1-\mathrm{NMSE}_d. The total capacity is MC=∑dMCd\mathrm{MC}=\sum_d\mathrm{MC}_d. As we can see in Figure 4, at T2=5T_2=5 and 10 μs10\ \mu\mathrm{s} the total capacity is between 0.95 and 0.99 of the unitary value. This shows that essentially no memory has been lost and what changes is its profile. The deepest delay still recovered above 0.90 extends from d=4d=4 to d=5d=5 steps in the unitary reservoir, to d=6d=6 and d=7d=7 under dephasing. This is true on both reservoir realizations and correlator configurations. The long weak tail is traded for a sharper cutoff: recall stays above 0.90 to longer delays and then falls faster. This does not

Figure 4. Linear MC of the Oall\mathcal{O}_{\mathrm{all}} readout set. (a,b) Per-delay profile MCd\mathrm{MC}_d for R1\mathcal{R}_1 and R2\mathcal{R}_2 at three coherence values. MC0=0\mathrm{MC}_0 = 0 by construction, since ρk\rho_k has not yet received uku_k (Eq. (4)). Each legend shows its curve’s total MC capacity, where the dotted lines show the 0.90 level used in the text. At T2=10 μsT_2 = 10\ \mu\mathrm{s} the curves lie above the unitary one at every delay up to d=9d = 9 on R1\mathcal{R}_1 and d=8d = 8 on R2\mathcal{R}_2, and then drop below it. (c) Total capacity relative to the unitary, showing that 5–10 μs\mu\mathrm{s} columns lose a small amount of memory, while the 1 μs\mu\mathrm{s} column loses approximately 26%.

improve linear recall of uk−nu_{k-n}, which is the far tap of the product uk−nuk−1u_{k-n}u_{k-1} through which the input enters the recurrence for yky_k. MC10_{10} falls from 0.49 and 0.46 in the unitary reservoir to between 0.26 and 0.33 at T2=5T_2 = 5–10 μ\mus, and MC15_{15} stays below 0.1 throughout, so the accuracy gain is not explained by better recall of uk−nu_{k-n}. We return to this below.

In order to calibrate the requirements of the tasks, we fit Eq. (6) at increasing tap depth. The error on NARMA-nn falls quickly once κ\kappa exceeds nn, at κ=11\kappa= 11, 18 and 21 for the three orders and is controlled by the tap count. The polynomial degree does not seem to affect it since on NARMA-10 a cubic in 10 taps does not show any improvement over a quadratic in 10, while a quadratic of 11 improves by a factor of 4.3. Moreover, on higher orders the same step from 10 to 11 taps does not change anything, since 11 is still short for their requirements. Therefore, NARMA-nn demands close to n+1n+1 steps of input memory. The reservoir does not have that, since the deepest delay it recovers above 0.90 is 7, and then only for the Oall\mathcal{O}_{\mathrm{all}} set on R1\mathcal{R}_1 at T2=5T_2 = 5 and 10 μ\mus. Moreover, it is 6 for the same set on R2\mathcal{R}_2, and 6 for the correlator-only readout on both. For the OZ\mathcal{O}_Z set it is at most 2, and on R1\mathcal{R}_1 this set does not reach 0.90 at any delay in six of its nine columns. We see that none of the three orders is being solved with the memory it requires, and the higher two are being solved with less. This is why the coherence dependence tells us more about how a short memory is shaped and less about how far back it reaches. On the other hand, the selected ridge parameter αr\alpha_r tells the same story from the other side, since the median log⁡10αr\log_{10}\alpha_r chosen by cross-validation rises monotonically from −6.0-6.0 at T2=1T_2 = 1 μ\mus to −2.75-2.75 at unitary. This shows that the dephased feature sets need less regularization, which is what we would expect if dephasing had removed redundant directions. We note that −6.0-6.0 is the floor of our αr\alpha_r grid rather than an interior optimum. Specifically, at T2=1T_2 = 1 μ\mus the selection remains on that minimum in 15/15 cells on R1\mathcal{R}_1 and 13/15 on R2\mathcal{R}_2, leaving open the possibility that smaller values would further improve validation performance.

At T2=1T_2 = 1 μ\mus the picture is different and we do not have a sufficient explanation for it. The total capacity falls to 0.74 of unitary on both reservoirs and recall at intermediate delays is reduced. MCd\mathrm{MC}_d is 0.89, 0.45, 0.24, 0.15 on R1\mathcal{R}_1 for d={5,…,8}d = \{5,\ldots,8\} against the unitary 0.93, 0.86, 0.80, 0.69, and 0.83, 0.48, 0.33, 0.18 against 0.92, 0.87, 0.76, 0.69 on R2\mathcal{R}_2. Nevertheless, this is the best performing column in the whole sweep. Moreover the NMSE there stops depending on the NARMA order, having 0.33, 0.30, 0.30 values for n=10n = 10, 15, 20 on R2\mathcal{R}_2, while the unitary reservoir degrades from 0.26 to 0.57. The similar errors across the three orders raise the question of which temporal information the predictor uses in this regime. To test this, we checked whether these cells had collapsed into a static map of the recent input, fitting Eq. (6) at tap depths that are too short to succeed in the task. We found that they did not, since a quadratic in 10 taps reached only 0.62 to 0.71 against the 0.30 these cells achieved. We therefore report the accuracy gain at T2=1 μsT_2 = 1\ \mu\mathrm{s} without proposing a mechanism for it, and note that this column lies an order of magnitude below the typical lower end of coherence on this platform.

Mackey–Glass: the non-chaotic limit cycle

We now close the loop and the output is shown in Figure 5. For τMG=16\tau_{\mathrm{MG}} = 16 we checked whether the trained readout has the target periodic orbit as an attractor. In our sweep, we ran 70 runs in total which included 10 reservoir realizations and seven coherence columns. We found that the reservoir sustains (i.e., V=Keval\mathcal{V} = K_{\mathrm{eval}}) all 1699 evaluated autonomous steps in 50 cases, collapsing in approximately 30% of the time. The sustain rate does not seem to depend on coherence since it is essentially flat throughout the T2T_2 sweep, remaining between 6/10 and 8/10 in every column, including the unitary reference.

Figure 5. Autonomous phase of the reservoir at T2=10 μsT_2 = 10\,\mu\mathrm{s}. The MG-16 panels show the best of the 10 realizations at this coherence; the MG-17 and SNA panels show R1\mathcal{R}_1, which is also the best realization for MG-17. Top row: autonomous trajectory vs truth, with the teacher-forced warmup removed. The valid prediction time V\mathcal{V} of Eq. (16) is marked where it falls inside the window. The MG panels show 1400 steps from the first autonomous one. The SNA panel shows 100 for clarity centered on V\mathcal{V}. Bottom row: the whole autonomous window phase space in delay coordinates, with target and prediction overlaid. Here we test whether the attractor topology was reproduced instead of just tracking. The MG panels use a lag of 15 steps and the SNA of one, since the latter's structure lives in successive steps. In (f) the target occupies a scattered band, while the prediction is a smooth, approximately closed curve in its middle. This non-scattered skeleton behavior is discussed in § 2.6.

On the other hand, coherence seems to show up instead in the accuracy of the runs that survive. Fixing the reservoir at R1\mathcal{R}_1 and re-rolling only the training stochasticity, the T2=10T_2 = 10 and 50 μs50\,\mu\mathrm{s} columns beat the unitary reference in all 10 training seeds, with median NMSE ratios of 0.447 and 0.059. Unfortunately, this does not survive changing the chip, since across the 10 reservoir realizations the same two columns win only 5/10 and 7/10, respectively. We flag this explicitly, to avoid overclaiming the beneficial effect of decoherence and to highlight that on one particular structure dephasing helps the non-chaotic task, while on another it does not. This makes the use of dephasing as a resource for this particular task device-dependent.

Mackey–Glass: the chaotic regime

At τMG=17\tau_{MG} = 17 the attractor becomes chaotic and the more relevant metric becomes the valid prediction time V\mathcal{V}, defined in Eq. (16). In Figure 6a, we can see that the effect at T2=1 μsT_2 = 1\,\mu\mathrm{s} is unambiguous and has the opposite sign to NARMA. The unitary reservoir sustains the trajectory for 597.7±116.0597.7 \pm116.0 steps averaged over 10 realizations, and at T2=1 μsT_2 = 1\,\mu\mathrm{s} this falls to 298.2±153.8298.2 \pm153.8. We see the degrading effect in all 10 realizations with a median loss of 313 steps. Among the remaining tested coherence times, T2≥5 μsT_2 \ge5\ \mu\mathrm{s}, none gives a shorter prediction horizon than the unitary reference in more than six of the ten reservoir realizations. Across these values, we observe no consistent increase in prediction horizon with increasing coherence time. The sweep therefore identifies a reduction at T2=1 μsT_2 = 1\ \mu\mathrm{s} but does not establish the precise location of a threshold.

Figure 6. Coherence and hardware effects on autonomous phase MG-τ\tau generation. (a) Valid prediction time on the chaotic target against coherence. Here we compare i) 10 values of a single seed driving both the couplings and the training RNG, and ii) one reservoir while sweeping 10 training seeds. The bands are 1σ1\sigma. The prediction horizon at T2=1 μsT_2 = 1\ \mu\mathrm{s} is shorter than the unitary reference in all 10 reservoir realizations. (b) Fraction of non-chaotic runs that sustain the orbit over all 1699 evaluated autonomous steps. The rate stays between 60% and 80% in every column across reservoirs, while on the fixed chain R1\mathcal{R}_1 it ranges from 0% to 100%.

Furthermore, given that the NARMA task was sensitive to the choice of the ridge parameter αr\alpha_r, we checked the same for this task. Specifically, we ran a comparison at log⁡10αr∈{−6,−4,−2,0}\log_{10}\alpha_r \in\{-6,-4,-2,0\} across three training seeds. The unitary reservoir was better in both V\mathcal{V} and NMSE in all 12 combinations with a median NMSE ratio of 2.25, which indicates that the effect is not an artifact of regularization. Moreover, the selected DAgger iteration was always different between the two coherences. The unitary runs peaked, in all combinations, at iterations 13–15 with feedback probability pq=0.69p_q = 0.69–0.800.80, while the T2=1 μsT_2 = 1\ \mu\mathrm{s} at 9–13 with pq=0.48p_q = 0.48–0.690.69. Hence, the dephased reservoir exits the scheduled-sampling ramp earlier, which may indicate that its closed-loop map loses stability at a lower feedback probability.

Separating reservoir disorder from training stochasticity

We swept two randomness axes to quantify the sensitivity to both the reservoir configuration (e.g. exchange JijJ_{ij} disorder), and to stochasticity in the training procedure (e.g. DAgger/ridge training). We compared a mixed sweep with a training-only control, in which the chain was fixed at R1\mathcal{R}_1 and only the training seed was re-sampled over 10 runs. We show the results in Figure 6.

Pooled over the 14 cells, re-rolling the seed that drives both the couplings and the training RNG gives a root-mean-square (RMS) spread in log⁡10NMSE\log_{10}\mathrm{NMSE} 2.12 times that of re-rolling only the training seed, and the spread is wider in every cell. Since the mixed sweep contains both sources, the reservoir contribution can be isolated by subtracting the training variance in quadrature. Here we assume that the training variance is chain-independent and the two sources add in variance, σboth2≈σchain2+σtraining2\sigma_{\mathrm{both}}^2 \approx\sigma_{\mathrm{chain}}^2 + \sigma_{\mathrm{training}}^2, with no interaction term. The reservoir contribution is then 1.87 times the training contribution, and it is larger in 10/14 cells, with all four exceptions lying in the chaotic regime.

The distinction between the two is clearer for τMG=17\tau_{\mathrm{MG}} = 17 at T2=5 μsT_2 = 5\ \mu\mathrm{s}. Across 10 reservoir configurations, the prediction horizon is 517.0±203.2517.0 \pm203.2 steps, while 10 independent training runs on the fixed reservoir R1\mathcal{R}_1 give 683.4±0.5683.4 \pm0.5 steps. In this cell, the training outcome is reproducible, while the modification of the disorder configuration can shift the prediction horizon substantially.

In addition, the failure statistics point the same way. For the chaotic τMG=17\tau_{\mathrm{MG}} = 17 case at T2=1 μsT_2 = 1\ \mu\mathrm{s}, we define collapse as a horizon below 70% of the corresponding unitary mean and find that it occurs for 8/10 reservoir configurations. On the contrary, we do not see a collapse when we fix the chain in any of the ten training seeds. The loss of predictive performance is therefore primarily associated with the reservoir realization.

For the non-chaotic τMG=16\tau_{\mathrm{MG}} = 16 case, fixing the chain leaves the pooled collapse rate almost unchanged, at 29% across the mixed sweep and 31% on R1\mathcal{R}_1. Within a column, however, the training seed rarely changes the outcome on R1\mathcal{R}_1. No seed sustains the orbit at T2=1 μsT_2 = 1\ \mu\mathrm{s}, whereas all 10 do at 10, 50 and 5000 μs\mu\mathrm{s}, and nine do at the unitary reference. The fixed-chain rate therefore records which coherence columns fail on R1\mathcal{R}_1, consistent with the chain term exceeding the training term in all seven non-chaotic cells. On a single chain the limit cycle depends strongly on coherence, and the dependence averages out to between 6/10 and 8/10 per column across reservoirs.

The strange non-chaotic attractor

The reservoir failed on the SNA by the pointwise measure. We ran the same protocol on 10 reservoir configurations, with T1=100 msT_1 = 100\ \mathrm{ms} and the eight T2T_2 values of the NARMA sweep plus the unitary reference, giving 90 cells in total. NMSE never falls below 1.471, with a median of 1.834, so the readout does worse than predicting the target mean in all 90 cells. Therefore, the failure is a property of the task on this substrate and does not depend on the realized chain.

In Figure 5f, the target occupies a broad scattered band, while the prediction follows a smooth, approximately closed curve through its middle. The prediction captures the broad orbital shape but misses the target’s scatter. The noise κk\kappa_k in Eq. (8) is redrawn independently at each step and is not determined by the state. An autonomous predictor without access to these future noise draws cannot reproduce the target’s specific random fluctuations.

However, this is not why NMSE >1> 1. Iterating Eq. (8) with D=0D = 0 gives a period T=13\mathcal{T} = 13 orbit and its variance is 1.01 times that of the noisy series. The noise-free and noisy series therefore have similar total variances. Applying the exact deterministic map one step ahead to the noisy series gives NMSE =0.0027= 0.0027. This means that the failure is most probably due to phase, not scatter. A copy of the target displaced along its own orbit by less than a period scores between 0.31 and 3.83 with a mean of 2.17. Our measured range, 1.47 to 1.86, sits inside that band and is expected to be produced by a model reproducing the orbit while drifting in phase. This is also clear from Figure 5f, where we can see that the geometry survives even though pointwise agreement is not achieved. Hence, the result tests what autonomous generation can achieve on a noise-driven attractor; this perspective is also supported by the separation horizon. It takes two values across the 90 cells, namely V=137\mathcal{V} = 137 and V=202\mathcal{V} = 202 in 67/90 and the remaining 23, respectively. Both are in agreement to 7 mod TT. More explicitly, 137 = 10 × 13 + 7 and 202 = 15 × 13 + 7, so the two differ by five loops of the attractor. We expect that a model that captures only the orbit without the phase would produce a horizon that moves in integer periods. In contrast, a model that had lost the attractor at some moment would have no reason to lose it on a lattice. NMSE is bimodal in the same way with 28 cells in [1.471, 1.650] and 62 in [1.804, 1.862]. Moreover, the two structures agree that 19/23 at V=202\mathcal{V} = 202 are in the lower cluster against 9/67 at V=137\mathcal{V} = 137. The teacher-forced bootstrap (q=0q = 0) is selected in 18/23 cells with the longer horizon and 15/67 with the shorter horizon. These associations suggest that phase alignment and readout selection influence the measured scores, without establishing phase as the only source of error.

Discussion

In this work, we have swept the coherence time T2T_2 of a six-qubit spin chain over a parameter range and analyzed how it affects our three temporal tasks. We observed four different behavioral regimes, since the MG task exhibits different dynamics on each side of its chaotic transition. For the open-loop NARMA family, dephasing helps since the reservoir outperforms its own coherent version in 27/30 paired comparisons at T2=1 μsT_2 = 1\ \mu\mathrm{s} and in 23/30 at 10 μs\mu\mathrm{s}, while the effect seems to diminish by 50 μs\mu\mathrm{s}. For autonomous generation of a non-chaotic limit cycle, coherence shows no systematic effect across reservoirs. The failure rate stays between 20 and 40% per column, including the unitary reference. In contrast, strong dephasing at T2=1 μsT_2 = 1\ \mu\mathrm{s} shortens the chaotic prediction horizon in all ten reservoir realizations. The mean horizon decreases from 598±116598 \pm116 steps in the unitary reference to 298±154298 \pm154 steps at this coherence time. At the other tested coherence times, T2≥5 μsT_2 \ge5\ \mu\mathrm{s}, we observe no consistent increase in prediction horizon with increasing coherence time. Furthermore, for autonomous generation of a noise-driven SNA no configuration at any column of any realization is able to beat the mean prediction. In this task, the variation across cells is consistent with differences in the phase to which the readout locks. In summary, whether decoherence helps or hinders learning in our simulated reservoir depends on the target task.

Our NARMA result agrees in direction with ref. [33], which found better performance at larger dissipation in a six-qubit Heisenberg-coupled spin-qubit array for NARMA up to order n=10n = 10; at order n=20n = 20, where our gain is largest, their larger decay rate is slightly worse. Our memory capacity measurements show that dephasing redistributes the memory profile. At T2=5T_2 = 5–10 μs\mu\mathrm{s}, the total capacity remains close to the unitary value and high-accuracy recall extends to longer delays. However, there is no improvement on the recall of the distant input uk−nu_{k-n}. These measurements therefore characterize a change accompanying the accuracy gain, without establishing its mechanism. For chaotic forecasting, the horizon is shorter at T2=1 μsT_2 = 1\,\mu\mathrm{s}, but the sampled coherence values do not establish the precise location of a threshold. Ref. [35] reports a monotonic relationship between coherence and information processing capacity in a transverse-field Ising reservoir, so the two results are not directly comparable. Task dependence is one possible explanation for the differing findings, but since we did not reproduce either setting, we cannot establish that explanation.

We highlight two caveats which limit how far this generalizes. First, we read out exact expectation values, a point which ref. [35] identifies as an important factor. With a finite-shot budget, the estimator variance on each observable competes with the signal. Therefore, an argument that dephasing improves the conditioning of the feature set has to be tested when the features are noisy. Our accounting of the measurement settings in § 4.3 is a first step in the resource direction. Second, our Lindblad evolution is Markovian, whereas the free-evolution dephasing of a spin qubit is dominated by non-Markovian quasi-static charge and nuclear noise. We expect it to behave differently, since a slowly drifting detuning is closer to a fluctuating reservoir than the decay channel we employed.

Under the assumptions of chain-independent training variance and additive variance contributions stated in § 2.5, we estimate that the reservoir component of the RMS spread of log⁡10NMSE\log_{10}\mathrm{NMSE} is 1.87 times the training component. For example, at τMG=17\tau_{\mathrm{MG}} = 17 and T2=5 μsT_2 = 5\,\mu\mathrm{s}, 10 realizations spread the horizon over 517.0±203.2517.0 \pm203.2 steps, while 10 training seeds on one reservoir give 683.4±0.5683.4 \pm0.5. The training protocol can be highly reproducible on a given chain, while the reservoir configuration remains a key factor in learning performance. This makes the coupling pattern itself a design parameter. The disorder strength alone moves a spin reservoir between localized and ergodic dynamics, with the best MG prediction found near the transition [51]. Beyond disorder strength, structured patterns such as dimerized chains of Su–Schrieffer–Heeger type [52] or couplings engineered for perfect state transfer [53] are possible alternatives to the random draws used here.

In conclusion, we have shown that readout quantum observables contribute information non-trivially and that this is not due to the fact that they add more regressors. This contribution comes at the cost of five measurement settings, which is minimal for six qubits, while a ZZ-basis readout needs one. Also, the aggregated fit was selected over the teacher-forced bootstrap in all 70 non-chaotic runs of the coherence sweep, although on the SNA the bootstrap was selected in 33/90 cells.

We studied NARMA-nn timeseries, which are solved efficiently by classical delay-embedding and next-generation reservoir methods [54, 55]. Here, we used them as a reference of how the reservoir’s behavior is changed by its coherence and readout. For the autonomous tasks we did not run the equivalent delay-line comparison, although next-generation reservoir computing also forecasts autonomously from a delay embedding [54], which makes it the natural baseline for future work.

For future research, we aim to put a shot budget on the readout and repeat the coherence sweep with sampled expectation values instead of the exact ones that we used here. We believe that this is the assumption most likely to change the open-loop conclusion. A second extension worth investigating is time multiplexing [7]. In that case, the reservoir is sampled multiple times within each input interval and its feature vector is built from those sub-steps. With it, a seven-qubit reservoir matched the NARMA performance of a 500-node echo state network. In our implementation, the feature vector was functional instead of temporal, where we expanded 57 observables measured once per step into a polynomial basis. Sub-step sampling requires no additional measurement bases, but on hardware each sub-step needs its own shots, so which of the two approaches is cheaper depends on the platform. It would also be worth replacing the Markovian dephasing with a quasi-static detuning drift, to make the noise model more realistic for spin-qubit devices. The reset channel of Eq. (4), which gives the reservoir its fading memory, could also be realized physically with metallic leads that initialize the chain by injecting electrons. In our previous work, we have modeled such leads with Anderson impurity models, and in a quantum dot array they changed quantum information scrambling dynamics [56], so they may also change what the reservoir retains. Lastly, we expect that a model with access to its own driving noise instead of a purely deterministic feedback map could potentially recover the SNA’s scatter, in addition to its deterministic orbit. That way we could determine whether the limit we report here is the attractor’s or the reservoir’s.

Methods

The spin-chain reservoir

Hamiltonian

We use a linear chain of N=6N = 6 electron spin qubits confined in (gate-defined) quantum dots, in a way similar to Si/SiGe devices. Neighboring spins interact through the exchange interaction, which in the single-band Hubbard description reduces to an isotropic Heisenberg coupling in the strong Coulomb interaction regime. Each qubit has its own Zeeman frequency because of the stray field gradient of an on-chip micromagnet for individual addressability. Writing the reservoir Hamiltonian in the rotating frame and ordinary frequency units, we have:

H=∑i=1NδiZi+∑⟨i,j⟩Jij(XiXj+YiYj+ZiZj)+∑i=1Nhi(cos⁡ϕXi−sin⁡ϕYi)(1)H = \sum_{i=1}^{N} \delta_i Z_i + \sum_{\langle i,j\rangle} J_{ij}\left(X_iX_j + Y_iY_j + Z_iZ_j\right) + \sum_{i=1}^{N} h_i\left(\cos\phi X_i - \sin\phi Y_i\right) \tag*{(1)}

where Xi,Yi,ZiX_i,Y_i,Z_i are the single qubit Pauli operators acting on qubit ii, δi\delta_i is the detuning of qubit ii from the rotating-frame reference, JijJ_{ij} the exchange coupling on the bond ⟨i,j⟩\langle i,j\rangle where ⟨⋯ ⟩\langle\cdots\rangle denotes nearest-neighbor indexing, hih_i the amplitude of the transverse drive on qubit ii and ϕ\phi its phase in the XY plane. Eq. (1) is written in frequency units so the time evolution operator U(t)=exp⁡(−2πiHt)U(t) = \exp(-2\pi iHt). In this normalization, qubit ii precesses at 2δi2\delta_i in the rotating frame, the drive produces Rabi oscillations at 2hi2h_i, and the exchange splitting of bond ⟨i,j⟩\langle i,j\rangle is 4Jij4J_{ij}.

The detunings created by the micromagnet gradient are taken as δ=(28,14,0,−14,−28,−42)\delta= (28, 14, 0, -14, -28, -42) MHz. This results in a nearest-neighbor spacing of 14 MHz. and a total spread of 70 MHz across the N=6N = 6 dots. The drive amplitudes are hi≈1h_i \approx1 MHz, i.e., Rabi frequencies of about 2 MHz. The exchange couplings with induced disorder are drawn from base values of (8, 5, 9, 6, 7) MHz on the five bonds of the chain, where each one is multiplied by an independent factor sampled from U(1,2)\mathcal{U}(1, 2). The final couplings are in the range between 5–18 MHz. The disorder is quenched and every result below at a given instantiation of the quantum dot array uses the same chain. The two realizations carried through the open-loop results as R1\mathcal{R}_1 and R2\mathcal{R}_2 have couplings (14.19, 7.19, 16.73, 10.18, 7.66) and (13.22, 5.22, 9.18, 11.04, 11.11) MHz, respectively. We show in § 2.5 that the reservoir configuration matters significantly.

Dissipative dynamics

We treat the reservoir as an open quantum system. Each qubit is subject to amplitude damping at a rate of 1/T11/T_1 and to pure dephasing at a rate 1/Tφ=1/T2−1/(2T1)1/T_\varphi= 1/T_2 - 1/(2T_1), with T2≤2T1T_2 \le2T_1. The Lindbladian superoperator is then:

L[ρ]=−2πi[H,ρ]+D[ρ](2)\mathcal{L}[\rho] = -2\pi i[H,\rho] + \mathcal{D}[\rho] \tag*{(2)}

where D[ρ]\mathcal{D}[\rho] is the dissipation superoperator acting as:

D[ρ]=∑i=1N1T1(σi+ρσi−−12{σi−σi+,ρ})+∑i=1N12Tφ(σizρσiz−ρ)(3)\mathcal{D}[\rho] = \sum_{i=1}^{N} \frac{1}{T_1}\left(\sigma_i^{+}\rho\sigma_i^{-} - \frac{1}{2}\{\sigma_i^{-}\sigma_i^{+},\rho\}\right) + \sum_{i=1}^{N} \frac{1}{2T_\varphi}(\sigma_i^z\rho\sigma_i^z-\rho) \tag*{(3)}

where σi±=(Xi±iYi)/2\sigma_i^\pm= (X_i \pm iY_i)/2 are the raising and lowering ladder operators on qubit ii and {⋅,⋅}\{\cdot,\cdot\} the anti-commutator. We integrate Eq. (2) by vectorizing the density matrix in column-major order ρ↦vec⁡(ρ)\rho\mapsto\operatorname{vec}(\rho), under which Hρ↦(1⊗H)vec⁡(ρ)H\rho\mapsto(1 \otimes H)\operatorname{vec}(\rho) and ρH↦(HT⊗1)vec⁡(ρ)\rho H \mapsto(H^T \otimes1)\operatorname{vec}(\rho) and exponentiating the resulting 4N×4N4^N \times4^N matrix once per drive setting. Since we simulate N=6N = 6, we can exponentiate without the need to use truncation or Trotterization and introduce Hamiltonian simulation errors.

Throughout this work we hold T1T_1 at 100 ms, comparable to the relaxation times measured in 28Si{}^{28}\mathrm{Si}/SiGe spin qubits [39]. Spin relaxation is far slower than dephasing, and it could affect the reservoir only within the few input steps that the reservoir remembers. One step lasts 2τ2\tau or 3τ3\tau, i.e., 1 or 1.5 μ\mus, so a qubit relaxes with probability below 2×10−52 \times10^{-5} per step. This is more than two orders of magnitude below the reset probability ϵ=0.02\epsilon= 0.02 of Eq. (4), which already drives the chain toward ∣0⟩\lvert0\rangle, the same state that relaxation favors. We sweep T2T_2 over {1,5,10,50,100,500,1000,5000} μ\{1, 5, 10, 50, 100, 500, 1000, 5000\}\ \mus together with the unitary reference (effectively T1,T2→∞T_1,T_2 \to\infty). The MG sweeps omit 500 and 1000 μ\mus. Over this range TφT_\varphi and T2T_2 differ by less than 3%, so the sweep is in effect a sweep of the pure-dephasing time. The 10 μ\mus column is the lower-end coherence that this type of platform typically reaches in free evolution, and it is the column we refer to whenever we mention a realistic operating point. The noise model and TφT_\varphi depend on how the device is operated and differ from device to device. For example, TφT_\varphi differs between an isolated qubit and qubits coupled by exchange, and between free and driven evolution. Our reservoir is driven with the exchange switched on, so 10 μ\mus is a free-evolution reference rather than a measured dephasing time for this operating mode.

Input-driven protocol

The input sequence u1:Lu_{1:L} is normalized to [0,1][0,1] and drives the reservoir by interpolating between two distinct evolutions. We define two drive settings, J0\mathcal{J}_0 and J1\mathcal{J}_1, that differ in exchange strength, drive phase, transverse amplitude and duration. J0\mathcal{J}_0 evolves under Eq. (1) with J0=2JijJ_0 = 2J_{ij}, phase ϕ=0\phi= 0 and h(0)=(1.0,0.95,1.1,0.9,1.05,1.0) MHz\mathbf{h}^{(0)} = (1.0, 0.95, 1.1, 0.9, 1.05, 1.0)\ \mathrm{MHz} for a time 2τ2\tau with τ=500 ns\tau= 500\ \mathrm{ns}. J1\mathcal{J}_1 uses couplings JijJ_{ij}, phase ϕ=π/4\phi= \pi/4 and h(1)=(1.1,1.0,0.95,1.05,1.0,0.9) MHz\mathbf{h}^{(1)} = (1.1, 1.0, 0.95, 1.05, 1.0, 0.9)\ \mathrm{MHz} for a time 3τ3\tau. These amplitudes are arbitrary values within ±10%\pm10\% of 1 MHz1\ \mathrm{MHz} rather than values from a specific device. Both vectors are kept fixed across reservoir realizations; only the exchange couplings change from one realization to the next. Both drive settings correspond to CPTP maps. That is, the two settings correspond to different Hamiltonians and runtimes. We should mention that although we drive the system in an analog fashion, one can always use combinations of quantum gates to drive it digitally [42]. In that case, the two maps would most probably differ also in their noise profile, since each gate has its own characteristic noise channel. The state then evolves as:

ρk+1=(1−ϵ)[ukJ0[ρk]+(1−uk)J1[ρk]]+ϵξ(4)\rho_{k+1} = (1-\epsilon)\left[u_k\mathcal{J}_0[\rho_k] + (1-u_k)\mathcal{J}_1[\rho_k]\right] + \epsilon\xi \tag*{(4)}

where ϵ\epsilon is the reset rate to a ξ\xi arbitrary fixed state, both tunable hyperparameters. Here, we take ϵ=0.02\epsilon= 0.02 (i.e., 2%2\% chance to force-restart) and ξ=∣0⟩⟨0∣\xi= \lvert0\rangle\langle0\rvert, the spin-down state ∣↓z⟩\lvert\downarrow_z\rangle into which spin qubits are initialized in experiment [39]; in principle it can be any state ∣ψ0⟩∈H⊗N\lvert\psi_0\rangle\in\mathcal{H}^{\otimes N} that is easily prepared, where H\mathcal{H} denotes the single qubit Hilbert space. Since the above evolution is a convex combination of CPTP maps, it is itself a CPTP map for every uk∈[0,1]u_k \in[0,1] and the ϵ\epsilon term makes it a contraction. This makes the reservoir have a fading memory and the state at step kk depends on the whole input history with exponentially decaying weight. All of the above claims were proven in ref. [42].

Benchmark tasks

NARMA-nn

The driving signal is an i.i.d. sequence drawn from a uniform distribution uk∼U(0,0.5)u_k \sim U(0,0.5), and the target obeys a NARMA-nn timeseries:

yk+1=tanh⁡[αyk+βyk∑j=0n−1yk−j+γuk−n+1uk+δ](5)y_{k+1} = \tanh\left[\alpha y_k + \beta y_k \sum_{j=0}^{n-1} y_{k-j} + \gamma u_{k-n+1}u_k + \delta\right] \tag*{(5)}

with (α,β,γ,δ)=(0.3,0.05,1.5,0.1)(\alpha,\beta,\gamma,\delta) = (0.3, 0.05, 1.5, 0.1), y0=0.16y_0 = 0.16, and n=10,15,20n = 10, 15, 20. The readout at step kk is fitted to yky_k, and the input enters the recurrence for yky_k only through the single product uk−nuk−1u_{k-n}u_{k-1}. This is why a quadratic polynomial in the last n+1n+1 taps quickly reduces the error. Moreover, the autoregressive term couples yky_k to a moving sum over the previous nn values, so the memory task demands grows with nn, with its nonlinearity remaining unchanged.

Although the tanh function is not usually part of the original NARMA definition, we apply it to avoid divergent feedback. Specifically, our drive is uniform on [0,0.5][0, 0.5], which is wider than the [0,0.2][0, 0.2] used with those coefficients in [7]. At the narrower range the recurrence is bounded at every order. In contrast, at ours the βyk∑jyk−j\beta y_k \sum_j y_{k-j} term diverges within a few hundred steps for n=15n = 15 and n=20n = 20, and the saturation is what keeps the higher orders finite. Although it is not required for stability, we also employ it at n=10n = 10 in order to keep a coherent definition across the three. In this way, the difference between n=10n = 10 and n=20n = 20 is a difference in memory demand rather than in recurrence. Moreover, we highlight that since we have a wider drive range than previous literature and to compensate we need a tanh function, the claims we make for NARMA are purely internal. Specifically, we compare one readout against another at matched feature count, or one coherence value against another on the same series with the same target. We flag it because a reader scanning for a familiar figure would be misled. We also note that ref. [7] drives its main NARMA demonstration with a product of three sinusoids. The same work uses random input, which carries no temporal structure of its own, for its memory and parity-check tasks, and repeats NARMA with white noise in an appendix. We use random input throughout, which we think is appropriate for a study on memory.

Since we need a way of checking how much input memory NARMA-nn actually demands independently of any reservoir, we use the delay-embedding of next-generation reservoir computing [54, 55]. We fit the same ridge readout of Eq. (12) to yky_k under the same protocol, but on polynomial features built directly from the last κ\kappa inputs:

Ψκ,d(k)=Φd(uk,uk−1,…,uk−κ+1),(6)\Psi_{\kappa,d}(k) = \Phi_d(u_k, u_{k-1}, \ldots, u_{k-\kappa+1}), \tag*{(6)}

which is independent of the reservoir employed. Since the input enters the recurrence for yky_k as the product uk−nuk−1u_{k-n}u_{k-1}, the feature set Ψκ,2\Psi_{\kappa,2} contains that monomial when κ≥n+1\kappa\ge n + 1. The prediction error as a function of κ\kappa therefore provides an empirical measure of the input history length needed by this baseline. We use it in § 2.2 as a memory demand indicator.

Mackey–Glass

The MG differential equation is defined as follows:

dy(t)dt=0.2y(t−τMG)1+y(t−τMG)10−0.1y(t)(7)\frac{\mathrm{d}y(t)}{\mathrm{d}t} = \frac{0.2y(t-\tau_{MG})}{1+y(t-\tau_{MG})^{10}} - 0.1y(t) \tag*{(7)}

and is integrated by explicit Euler at a step of 1/101/10 and subsampled by 10, so that one output step equals one unit of continuous time and the delay in output units is τMG\tau_{MG}. We take τMG=16\tau_{MG} = 16, where the attractor is a periodic orbit, and τMG=17\tau_{MG} = 17, past the bifurcation, where it is chaotic. The first 10310^3 output steps are discarded as transient before any normalization. The delay embedding used in Figure 5 takes a lag of 15 steps, close to the delay τMG\tau_{MG}.

Strange non-chaotic attractor

The SNA is generated by the noise-driven map:

yk+1=yk+σt(1−yk−μyk−12+g+Dζk),(8)y_{k+1}=y_k+\frac{\sigma}{t}\left(1-y_k-\mu y_{k-1}^{2}+g+D\zeta_k\right), \tag*{(8)}

with (σ,μ,g,t)=(1,4,−0.6,4/3)(\sigma,\mu,g,t)=(1,4,-0.6,4/3) and ζk\zeta_k standard Gaussian noise of intensity D=0.01D=0.01, discarding 5×1035\times10^3 steps of transient. Written as a two dimensional system in (yk,yk−1)(y_k,y_{k-1}) the deterministic part is close to marginal with the largest Lyapunov exponent λLE≈−0.017\lambda_{\mathrm{LE}}\approx-0.017. At this noise intensity the exponent remains negative λLE≈−0.005\lambda_{\mathrm{LE}}\approx-0.005, which makes the attractor non-chaotic. We should mention that the stochastic noise ζk∼N(0,1)\zeta_k\sim\mathcal{N}(0,1) is redrawn at every step and is not a function of the state. This makes any closed-loop model driven only by its output unable to have any access to it, which is the substance of our result in § 2.6.

Readout and the cost of a quantum observable

Quantum observable sets

After each step we measure a set O\mathcal{O} of expectation values on ρk\rho_k. Our maximal set consists of all six single-qubit populations, transverse magnetizations and the three families of two-qubit correlators:

Oall≡{⟨Zi⟩,⟨Xi⟩,⟨ZaZb⟩,⟨XaXb⟩,⟨ZaXb⟩}(9)\mathcal{O}_{\mathrm{all}}\equiv\{\langle Z_i\rangle,\langle X_i\rangle,\langle Z_aZ_b\rangle,\langle X_aX_b\rangle,\langle Z_aX_b\rangle\} \tag*{(9)}

with i∈{0,…,N−1}i\in\{0,\ldots,N-1\} and a<ba<b, giving a cardinality of ∣Oall∣=6+6+15+15+15=57|\mathcal{O}_{\mathrm{all}}|=6+6+15+15+15=57 observables in total. Our resulting vector is xk∈R∣Oall∣\mathbf{x}_k\in\mathbb{R}^{|\mathcal{O}_{\mathrm{all}}|} at step kk and x∈R∣Oall∣×nsteps\mathbf{x}\in\mathbb{R}^{|\mathcal{O}_{\mathrm{all}}|\times n_{\mathrm{steps}}} for the whole series. Moreover, we consider restricted sets in order to ask what the extra observables actually contribute. In particular, we consider OZ≡{⟨Zi⟩}\mathcal{O}_{Z}\equiv\{\langle Z_i\rangle\} with ∣OZ∣=6|\mathcal{O}_{Z}|=6, O2q≡{⟨ZaZb⟩,⟨XaXb⟩,⟨ZaXb⟩}\mathcal{O}_{2q}\equiv\{\langle Z_aZ_b\rangle,\langle X_aX_b\rangle,\langle Z_aX_b\rangle\} with a<ba<b and ∣O2q∣=45|\mathcal{O}_{2q}|=45, O1q≡{⟨Zi⟩,⟨Xi⟩}\mathcal{O}_{1q}\equiv\{\langle Z_i\rangle,\langle X_i\rangle\} with ∣O1q∣=12|\mathcal{O}_{1q}|=12 and Oall\mathcal{O}_{\mathrm{all}}. We can see that OZ⊂O1q⊂Oall\mathcal{O}_{Z}\subset\mathcal{O}_{1q}\subset\mathcal{O}_{\mathrm{all}} and O2q⊂Oall\mathcal{O}_{2q}\subset\mathcal{O}_{\mathrm{all}}, but O1q∩O2q=∅\mathcal{O}_{1q}\cap\mathcal{O}_{2q}=\varnothing.

This is where Hilbert space as a feature generator is tested. A classical recurrent network with six state variables offers six functions of the input history. In contrast, the same six spins measured locally in the X/ZX/Z bases give 57 one- and two-body expectation values, which are different nonlinear functionals of u1:ku_{1:k}. These additional observables probe transverse coherences and two-qubit correlations that are not captured by the single-qubit ZZ readout [28]. The question we ask here is whether feature richness is contributing, or whether it stands as a proxy for having more regressors. A ZZ-only readout can always be given more regressors by raising the polynomial degree. Therefore, we compare observable sets at matched feature count rather than at matched polynomial degree, which we set up next.

Measurement settings

Accessing the 57 quantum observables requires a measurement overhead. Observables compatible with the same local measurement bases can be estimated simultaneously from the same shots. This makes the number of distinct settings important for optimal resource count. Each setting requires an appropriate basis configuration and shot budget. Assigning either a ZZ or an XX measurement to each qubit defines a setting S:{0,…,N−1}→{Z,X}S : \{0,\ldots,N-1\} \to\{Z,X\}. Brute-force search over all (644)\binom{64}{4} sets of four settings confirms that none covers Oall\mathcal{O}_{\mathrm{all}}. Hence for N=6N=6, we find that five is minimal:

ZZZZZZ,XXXXXX,ZZZZXX,ZZXXZZ,ZXZXZX.(10)ZZZZZZ,\quad XXXXXX,\quad ZZZZXX,\quad ZZXXZZ,\quad ZXZXZX. \tag*{(10)}

The first two provide the six ⟨Zi⟩\langle Z_i\rangle and all 15 ⟨ZaZb⟩\langle Z_aZ_b\rangle, and the six ⟨Xi⟩\langle X_i\rangle and all 15 ⟨XaXb⟩\langle X_aX_b\rangle, respectively. The remaining three cover all 15 mixed correlators ⟨ZaXb⟩\langle Z_aX_b\rangle with a<ba<b.

The mixed settings can be generated directly from the binary representations of the qubit labels, without checking each pair individually. To see this, first write each label i∈{0,…,N−1}i \in\{0,\ldots,N-1\} using ⌈log⁡2N⌉\lceil\log_2 N\rceil binary digits, including leading zeros. Each digit position defines one setting: qubit ii is measured in ZZ if that digit is 0 and in XX if it is 1. For six qubits, the labels are 000, 001, 010, 011, 100, 101, whose three digit positions produce ZZZZXXZZZZXX, ZZXXZZZZXXZZ, and ZXZXZXZXZXZX.

For any a<ba<b, the most significant digit at which their binary representations differ is 0 for aa and 1 for bb. The corresponding setting therefore measures ZZ on qubit aa and XX on qubit bb, covering ⟨ZaXb⟩\langle Z_aX_b\rangle. Therefore, this construction requires at most 2+⌈log⁡2N⌉2+\lceil\log_2 N\rceil overall settings.

A simpler construction uses N+1N+1 settings. These are the all-ZZ and all-XX settings together with the N−1N-1 threshold settings

Sm(i)={Z,i<m,X,i≥m,m=1,…,N−1.(11)S_m(i)= \begin{cases} Z, & i<m,\\ X, & i\ge m, \end{cases} \qquad m=1,\ldots,N-1. \tag*{(11)}

These cover every required mixed correlator because, for any a<ba<b, the setting SbS_b places aa in the ZZ group and bb in the XX group. For six qubits, this gives seven settings, compared with five for the binary construction. For N≥4N\ge4, the binary construction reduces the setting count from N+1N+1 to 2+⌈log⁡2N⌉2+\lceil\log_2 N\rceil.

The full observable set Oall\mathcal{O}_{\mathrm{all}} can consequently be accessed using five settings, compared to one for OZ\mathcal{O}_{Z}. The all-ZZ setting also yields the 15 ⟨ZaZb⟩\langle Z_aZ_b\rangle, so a one-setting readout could use 21 observables; we did not test that set. The benefit of accessing additional observables should therefore be weighed against the platform-dependent cost of basis changes and the shots required to achieve the desired statistical precision. Polynomial features calculated classically from these expectation values require no additional measurement settings.

Polynomial features and ridge regression

The readout is a polynomial in the measured observables. We construct all monomials of total degree 1≤deg⁡≤d1\le\deg\le d in the components of x\boldsymbol{x} and then min–max normalize each resulting feature. Denoting this normalized feature vector by Φd(x)\boldsymbol{\Phi}_d(\boldsymbol{x}), the prediction at step kk is:

yˉk=W~T(Φd(xk)1)(12)\bar{y}_k = \widetilde{W}^{T}\begin{pmatrix}\Phi_d(x_k) \\ 1\end{pmatrix} \tag*{(12)}

with W~=(W,Wc)\widetilde{W} = (W,W_c) the readout weights and bias vector. Throughout this work we take deg⁡=2\deg= 2 for the full observable set Oall\mathcal{O}_{\mathrm{all}} which yields 57+(582)=171057 + \binom{58}{2} = 1710 features. The correlators enter as deg⁡=1\deg= 1 variables, so a quadratic readout on Oall\mathcal{O}_{\mathrm{all}} contains products of two-body expectation values. For the ZZ-only comparison we use deg⁡=7\deg= 7 on the six ⟨Zi⟩\langle Z_i\rangle, which gives 1715 features and therefore approximately matches the full set. This is the comparison that separates the effect of the observable set from that of the regressor count. For the two-qubit set we use deg⁡=2\deg= 2 with 45+(462)=108045 + \binom{46}{2} = 1080 features. For O1q\mathcal{O}_{1\mathrm{q}} we use deg⁡=4\deg= 4, which gives (164)−1=1819\binom{16}{4} - 1 = 1819 features.

The optimal weights W~⋆\widetilde{W}^{\star} are obtained in closed form by ridge regression, minimizing:

L(W~)=∑k∈T∣yk−yˉk∣2+αr(∥W∥22+Wc2)(13)\mathcal{L}(\widetilde{W}) = \sum_{k\in\mathcal{T}} \lvert y_k-\bar{y}_k\rvert^2 + \alpha_r\left(\lVert W\rVert_2^2 + W_c^2\right) \tag*{(13)}

over the training index set T\mathcal{T}, and αr>0\alpha_r > 0 is a tunable ridge penalty hyperparameter which is applied to all coefficients, including the bias. For MG and SNA, the yky_k in this objective is replaced by yk+1y_{k+1}, since yˉk\bar{y}_k predicts the next sample after the input yky_k. The NARMA fit is deterministic. For the autonomous tasks, training has two stochastic elements: the Gaussian perturbation added to the training features and the scheduled-sampling draws. For NARMA, αr\alpha_r is selected separately in every cell by blocked five-fold cross-validation over the training window, from 19 logarithmically spaced values in [10−6,103][10^{-6},10^3]. The readout is then refit on the whole training window. For MG and the SNA, we use αr=10−4\alpha_r = 10^{-4} throughout, apart from the regularization check on the chaotic task.

Training and evaluation methodology

Two ways of driving the reservoir

The three tasks differ not only in the target timeseries, but also in the way the reservoir is driven and trained.

For NARMA the drive is an external signal. We inject the i.i.d. sequence u1:Lu_{1:L}, collect the observables, and fit the readout to the target y1:Ly_{1:L} that is a nonlinear filter of that same sequence. The input and target are two different objects, the task is function approximation on a driven system, and the fit is a single ridge solve.

For the MG and SNA the drive is the signal itself, and the readout output yˉk\bar{y}_k predicts the next value yk+1y_{k+1}. The training involves teacher forcing, which drives the reservoir with the true yky_k. At evaluation, the loop is closed and the reservoir is driven with its own (clamped) previous output, uk+1=max⁡(0,min⁡(yˉk,1))u_{k+1} = \max(0,\min(\bar{y}_k,1)), with no external signal. The raw prediction used to calculate the error remains unclamped. This is a strictly harder problem than the NARMA, since the model must have the target trajectory as an attractor of the closed-loop map it defines instead of being only a function approximator. We call the NARMA setting open-loop and this one autonomous. It is worth highlighting that the teacher-forced phase is open-loop in the same sense, although drive is the target instead of an independent input. Nevertheless, their training method is different from the one we show here.

Training and evaluation windows are specified separately for each task. Using zero-based indices and half-open intervals, the NARMA experiments discard the first 100 samples as washout, fit the readout over [100,4250)[100, 4250), and evaluate over [4250,5000)[4250, 5000). For MG, the first 200 samples are used for washout, the training window is [200,10000)[200, 10000), and the evaluation window is [10000,12000)[10000, 12000). The first 300 evaluation steps use teacher forcing, followed by 1700 autonomous updates. Of the resulting predictions, 1699 have available next-step targets and enter the evaluation. For both autonomous tasks the reservoir is reinitialized to ξ\xi at the start of the evaluation window, so the teacher-forced steps that open it also serve as washout.

For SNA, the first 200 samples are used for reservoir washout, and the training window is the same as in MG. The evaluation window is [12000,22000)[12000, 22000). The first 1000 evaluation steps use teacher forcing, followed by 9000 autonomous updates. All resulting autonomous predictions have available next-step targets, corresponding to target indices [13001,22001)[13001, 22001), so that Keval=9000K_{\mathrm{eval}} = 9000. The retained readout is selected by the lowest autonomous NMSE on this window. For the autonomous tasks, the evaluation window is also used for model selection. We evaluate the teacher-forced bootstrap and each subsequent DAgger iteration on this window, selecting the readout with the lowest NMSE or the longest valid prediction time VV, depending on the task. The regression coefficients are fitted using training-window samples, while the retained iteration is determined by the evaluation-window targets. We report the selected readout’s performance on that same window, without an additional independent test window. The reported scores may therefore be optimistic relative to the performance on independent data. Every configuration goes through the same selection protocol, so comparisons between coherence values and between reservoirs share this optimism, although not necessarily to the same degree.

Why teacher forcing is not sufficient

The standard method fits Eq. (12) on teacher-forced states, which are states visited when the true signal is fed in at every step. Although it minimizes the one-step error during the training phase, it does not say anything about what happens when the loop is closed. In general, the trained map is a deterministic dynamical system and the target orbit is not an attractor of it. As a consequence, the closed loop either contracts onto a fixed point or departs from the target trajectory. In order to confirm that this is not a matter of insufficient regularization, we swept the ridge parameter over αr∈{10−6,10−4,10−2,1}\alpha_r \in\{10^{-6}, 10^{-4}, 10^{-2}, 1\}. On the chaotic target, it left the ordering of every coherence comparison unchanged over all 12 combinations of αr\alpha_r and training seed. The selection did not keep the teacher-forced bootstrap for any value of αr\alpha_r and the aggregated fit was selected in all 24 runs.

This failure mode is the exposure bias of autoregressive models [46, 47], and is not specific to quantum reservoirs. The model is trained only on states drawn from the data distribution and then asked to operate on states drawn from its own. As a result, it drifts away from the moment a small error accumulates.

Algorithm 1 In-loop readout training. The reservoir map T\mathcal{T} and the set of polynomial monomials are fixed, while W~\widetilde{\mathbf{W}} is learned.

Require: target y1:Ly_{1:L}, windows [k0,k1)[k_0,k_1) for training and [e0,e1)[e_0,e_1) for evaluation, number of iterations QQ, final feedback probability pendp_{\mathrm{end}}, ridge parameter αr\alpha_r, feature noise scale ς\varsigma

  1. D←∅\mathbb{D} \leftarrow\varnothing; W~←∅\widetilde{\mathbf{W}} \leftarrow\varnothing

  2. for q=0,1,…,Qq=0,1,\ldots,Q do

  3. pq←0p_q \leftarrow0 if q=0q=0 else pendq/Qp_{\mathrm{end}}q/Q ▷ scheduled sampling ramp

  4. ρ←ξ\rho\leftarrow\xi; yˉ←∅\bar{y} \leftarrow\varnothing

  5. for k=0,…,k1−1k=0,\ldots,k_1-1 do

  6. uk←max⁡(0,min⁡(yˉ,1))u_k \leftarrow\max(0,\min(\bar{y},1)) with probability pqp_q if k≥k0k\geq k_0 and W~≠∅\widetilde{\mathbf{W}}\ne\varnothing, else uk←yku_k\leftarrow y_k ▷ drive from own output or from truth

  7. ρ←T[ρ;uk]\rho\leftarrow\mathcal{T}[\rho;u_k] by Eq. (4); x←\mathbf{x}\leftarrow observables of ρ\rho

  8. if k≥k0k\geq k_0 then

  9. D←D∪{(x,yk+1)}\mathbb{D}\leftarrow\mathbb{D}\cup\{(\mathbf{x},y_{k+1})\} ▷ label with the TRUE next value

  10. end if

  11. yˉ←W~T(Φd(x)1)\bar{y}\leftarrow\widetilde{\mathbf{W}}^{\mathsf{T}}\begin{pmatrix}\boldsymbol{\Phi}_d(\mathbf{x})\\1\end{pmatrix} if W~≠∅\widetilde{\mathbf{W}}\ne\varnothing

  12. end for

  13. recompute the min–max constants over all D\mathbb{D} ▷ normalization tracks the visited distribution

  14. W~(q)←argmin⁡W~ L(W~)\widetilde{\mathbf{W}}^{(q)}\leftarrow\underset{\widetilde{\mathbf{W}}}{\operatorname{argmin}}\ \mathcal{L}(\widetilde{\mathbf{W}}) on D\mathbb{D}, features perturbed by N(0,ς2)\mathcal{N}(0,\varsigma^2)

  15. Mq←\mathcal{M}_q\leftarrow autonomous score of W~(q)\widetilde{\mathbf{W}}^{(q)} on [e0,e1)[e_0,e_1) ▷ closed loop, not one-step

  16. W~←W~(q)\widetilde{\mathbf{W}}\leftarrow\widetilde{\mathbf{W}}^{(q)}

  17. end for

  18. return W~(q⋆)\widetilde{\mathbf{W}}^{(q^\star)} with q⋆={argmin⁡q Mq,if Mq=NMSEargmax⁡q Mq,if Mq=Vq^\star=\begin{cases}\underset{q}{\operatorname{argmin}}\ \mathcal{M}_q,&\text{if }\mathcal{M}_q=\mathrm{NMSE}\\\underset{q}{\operatorname{argmax}}\ \mathcal{M}_q,&\text{if }\mathcal{M}_q=\mathcal{V}\end{cases}

In-loop training with dataset aggregation

In order to avoid the problem mentioned above we therefore have to train the readout on the states it will actually visit. We show the pipeline in Algorithm 1. We start from the teacher-forced bootstrap fit W~(0)\widetilde{\mathbf{W}}^{(0)}. At each iteration qq we run the reservoir over the training window feeding back the model’s own prediction with probability pqp_q and the true value with probability 1−pq1-p_q. This routine is called scheduled sampling [48]. We label every state on that run, including off-attractor ones, with the true next value. We append those pairs to the aggregated dataset Dq=Dq−1∪{(x,y)}\mathbb{D}_q=\mathbb{D}_{q-1}\cup\{(\mathbf{x},y)\} and refit W~(q)\widetilde{\mathbf{W}}^{(q)} on the whole. This is called dataset aggregation (DAgger) [49] and we borrow it from imitation learning. The aggregated dataset Dq\mathbb{D}_q teaches the readout both what next step value is on the attractor and also how to get back to the attractor from the neighborhood it drifts into.

Each fitted readout is stored with its corresponding min–max normalization constants, which are retained with the selected iteration. We ramp the feedback probability linearly as pq=pendq/Qp_q = p_{\mathrm{end}}q/Q with Q=15Q = 15 iterations and pend=0.8p_{\mathrm{end}} = 0.8, both of which are hyperparameters. We also add a small Gaussian perturbation of standard deviation ς=10−3\varsigma= 10^{-3} to the normalized training features at every refit. The performance of the closed-loop map is not monotonic in qq and the procedure usually passes through a rough transient. The best readout is some 0≤q⋆≤Q0 \leq q^\star\leq Q, so we evaluate the autonomous performance after each iteration and keep the best snapshot:

q⋆={argmin⁡0≤q≤QMq,if Mq=NMSE,argmax⁡0≤q≤QMq,if Mq=V,W~⋆=W~(q⋆)(14)\begin{aligned} q^\star&= \begin{cases} \underset{0 \leq q \leq Q}{\operatorname{argmin}}\mathcal{M}_q, & \text{if }\mathcal{M}_q = \mathrm{NMSE},\\ \underset{0 \leq q \leq Q}{\operatorname{argmax}}\mathcal{M}_q, & \text{if }\mathcal{M}_q = \mathcal{V}, \end{cases}\\ \widetilde{\mathbf{W}}^\star&= \widetilde{\mathbf{W}}^{(q^\star)} \tag*{(14)} \end{aligned}

where M\mathcal{M} is the selection metric defined in § 4.4.4. Given that the bootstrap fit W~(0)\widetilde{\mathbf{W}}^{(0)} is included in the comparison, the procedure can in principle fall back on pure teacher forcing (i.e., baseline) whenever the in-loop iterations fail to improve on it.

We note that across all 70 non-chaotic runs of the coherence sweep the bootstrap fit was never selected. The median choice was iteration 11 out of 15, while 53 out of 70 runs picked an interior iteration 0<q⋆<Q0 < q^\star< Q and the remaining 17 selected q⋆=Qq^\star= Q. The aggregated fit therefore scored better than the bootstrap on the selection window in every run. Finally, in about 1/41/4 of our runs it kept improving until QQ, meaning it could potentially offer improvements if we extended the iteration range.

Metrics

For evaluation, predictions are indexed by their target time: y^k=y‾k\widehat{y}_k = \overline{y}_k for NARMA and y^k=y‾k−1\widehat{y}_k = \overline{y}_{k-1} for MG and SNA. The metrics below compare y^k\widehat{y}_k with yky_k.

For the NARMA, non-chaotic MG and SNA tasks, we score the evaluation window with the normalized mean-square error (NMSE), which is also the selection metric for the two autonomous ones:

NMSE=∑k∈E(y^k−yk)2∑k∈E(yk−μy)2,μy=1∣E∣∑k∈Eyk(15)\mathrm{NMSE} = \frac{\sum_{k \in\mathcal{E}}(\widehat{y}_k-y_k)^2}{\sum_{k \in\mathcal{E}}(y_k-\mu_y)^2}, \qquad\mu_y = \frac{1}{|\mathcal{E}|}\sum_{k \in\mathcal{E}} y_k \tag*{(15)}

with E\mathcal{E} the evaluation index set, so that NMSE =1= 1 is the performance of predicting the mean and values above one are worse than the constant-mean baseline.

For the chaotic target we select and report by the valid prediction time V\mathcal{V}, the standard horizon metric in the chaotic RC literature [57, 58]. Two trajectories that start arbitrarily close separate exponentially, so the NMSE over a long autonomous window saturates regardless of the model’s quality, and we quote it only as a secondary measure. It is defined as:

V=min⁡({ℓ∈{0,…,Keval−1}:∣y^ka+ℓ−yka+ℓ∣>θσy}∪{Keval})(16)\mathcal{V} = \min\left(\left\{\ell\in\{0,\ldots,K_{\mathrm{eval}}-1\}: \left|\widehat{y}_{k_a+\ell}-y_{k_a+\ell}\right| > \theta\sigma_y\right\} \cup\{K_{\mathrm{eval}}\}\right) \tag*{(16)}

where kak_a is the target-time index of the first scored autonomous prediction, σy\sigma_y is the standard deviation of the aligned target values over the scored autonomous window, and θ\theta a threshold which we set to θ=1\theta= 1 here. We see that V\mathcal{V} effectively calculates the time before the error exceeds one standard deviation of the signal. Here, KevalK_{\mathrm{eval}} is the number of autonomous predictions for which aligned target values are available, and ℓ\ell counts steps from the first scored autonomous prediction. If the error never exceeds the threshold within this window, we set V=Keval\mathcal{V} = K_{\mathrm{eval}}. Thus, 0≤V≤Keval0 \le\mathcal{V} \le K_{\mathrm{eval}}, with Keval=1699K_{\mathrm{eval}} = 1699 in our MG evaluations. Reaching this ceiling means that the prediction remains within the chosen error threshold throughout the scored autonomous window, and its validity beyond that window is not assessed. Unlike ref. [58], we do not normalize V\mathcal{V} by the maximum Lyapunov exponent since we compare a single system against itself across coherence values instead of different attractors. This makes the raw step count a more directly interpretable quantity.

Use of AI tools

We used a large language model to refine the language of the manuscript, which included rewording and condensing the authors’ text, and to produce the TikZ source for the schematic in Figure 1 to the authors’ design. The authors directed the work, reviewed all text, and take full responsibility for the content of this article.

Code availability

The simulation and analysis code used in this study is available from the corresponding author on reasonable request subject to approval by Equal1 Labs.

Acknowledgements

We would like to thank our colleagues Andrii Semenov for insightful discussions and Niall Murphy for providing assistance with the code. Part of this work appeared in the PhD thesis of N.P. [59].

Author Contributions

Nikolaos Petropoulos: Conceptualization, Methodology, Software, Writing – original draft. Elena Blokhina: Conceptualization, Writing – review & editing, Supervision, Project administration, Funding acquisition.

Data availability

The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request.

Competing Interests

Both authors are affiliated with Equal1 Labs, a company developing silicon spin-qubit quantum computers. The authors declare no other competing interests.

References

  1. [1]Tanaka, G., Yamane, T., Héroux, J.B., Nakane, R., Kanazawa, N., Takeda, S., Numata, H., Nakano, D., Hirose, A.: Recent advances in physical reservoir computing: A review. Neural Networks 115, 100–123 (2019) https://doi.org/10.1016/j.neunet.2019.03.005
  2. [2]Nakajima, K.: Physical reservoir computing—an introductory perspective. Japanese Journal of Applied Physics 59(6), 060501 (2020) https://doi.org/10.35848/1347-4065/ab8d4f
  3. [3]Yan, M., Huang, C., Bienstman, P., Tino, P., Lin, W., Sun, J.: Emerging opportunities and challenges for the future of reservoir computing. Nature Communications 15, 2056 (2024) https://doi.org/10.1038/s41467-024-45187-1
  4. [4]Roy, K., Jaiswal, A., Panda, P.: Towards spike-based machine intelligence with neuromorphic computing. Nature 575(7784), 607–617 (2019) https://doi.org/10.1038/s41586-019-1677-2
  5. [5]Sangwan, V.K., Hersam, M.C.: Neuromorphic nanoelectronic materials. Nature Nanotechnology 15(7), 517–528 (2020) https://doi.org/10.1038/s41565-020-0647-z
  6. [6]Lv, Z., Wang, Y., Chen, J., Wang, J., Zhou, Y., Han, S.-T.: Semiconductor quantum dots for memories and neuromorphic computing systems. Chemical Reviews 120(9), 3941–4006 (2020) https://doi.org/10.1021/acs.chemrev.9b00730
  7. [7]Fujii, K., Nakajima, K.: Harnessing Disordered-Ensemble Quantum Dynamics for Machine Learning. Phys. Rev. Applied 8(2), 024030 (2017) https://doi.org/10.1103/PhysRevApplied.8.024030 . Accessed 2024-01-25
  8. [8]Ghosh, S., Opala, A., Matuszewski, M., Paterek, T., Liew, T.C.H.: Quantum reservoir processing. npj Quantum Inf 5(1), 35 (2019) https://doi.org/10.1038/s41534-019-0149-8 . Accessed 2023-08-03
  9. [9]Mujal, P., Martínez-Peña, R., Nokkala, J., García-Beni, J., Giorgi, G.L., Soriano, M.C., Zambrini, R.: Opportunities in quantum reservoir computing and extreme learning machines. Advanced Quantum Technologies 4(8), 2100027 (2021) https://doi.org/10.1002/qute.202100027 https://advanced.onlinelibrary.wiley.com/doi/pdf/10.1002/qute.202100027
  10. [10]Fujii, K., Nakajima, K.: Quantum Reservoir Computing: A Reservoir Approach Toward Quantum Machine Learning on Near-Term Quantum Devices, pp. 423–450. Springer, ??? (2021). http://dx.doi.org/10.1007/978-981-13-1687-6_18
  11. [11]Marković, D., Grollier, J.: Quantum neuromorphic computing. Applied Physics Letters 117(15), 150501 (2020) https://doi.org/10.1063/5.0002014 . Accessed 2023-08-03
  12. [12]Carles, B., Dudas, J., Balembois, L., Grollier, J., Markovic, D.: Experimental quantum reservoir computing with a circuit-quantum-electrodynamics system. Phys. Rev. Appl. 25, 054005 (2026) https://doi.org/10.1103/fmm7-3qm9
  13. [13]Suzuki, Y., Gao, Q., Pradel, K.C., Yasuoka, K., Yamamoto, N.: Natural quantum reservoir computing for temporal information processing. Sci Rep 12(1), 1353 (2022) https://doi.org/10.1038/s41598-022-05061-w . Accessed 2023-08-03
  14. [14]García-Beni, J., Giorgi, G.L., Soriano, M.C., Zambrini, R.: Scalable photonic platform for real-time quantum reservoir computing. Phys. Rev. Appl. 20, 014051 (2023) https://doi.org/10.1103/PhysRevApplied.20.014051
  15. [15]Zia, D., Innocenti, L., Minati, G., Lorenzo, S., Suprano, A., Di Bartolo, R., Spagnolo, N., Giordani, T., Cimini, V., Palma, G.M., Ferraro, A., Sciarrino, F., Paternostro, M.: Quantum reservoir computing for photonic entanglement witnessing. Science Advances 11(50) (2025) https://doi.org/10.1126/sciadv.ady7987
  16. [16]Paparelle, I., Henaff, J., García-Beni, J., Gillet, É., Montesinos, D., Giorgi, G.L., Soriano, M.C., Zambrini, R., Parigi, V.: Experimental memory control in continuous-variable optical quantum reservoir computing. Nature Photonics 20(4), 413–420 (2026) https://doi.org/10.1038/s41566-026-01880-9
  17. [17]Negoro, M., Mitarai, K., Fujii, K., Nakajima, K., Kitagawa, M.: Machine learning with controllable quantum dynamics of a nuclear spin ensemble in a solid. arXiv (2018). https://doi.org/10.48550/ARXIV.1806.10910 . https://arxiv.org/abs/1806.10910
  18. [18]Govia, L.C.G., Ribeill, G.J., Rowlands, G.E., Krovi, H.K., Ohki, T.A.: Quantum reservoir computing with a single nonlinear oscillator. Phys. Rev. Research 3(1), 013077 (2021) https://doi.org/10.1103/PhysRevResearch.3.013077 . arXiv:2004.14965 [cond-mat, physics:quant-ph]. Accessed 2023-08-03
  19. [19]Kornjača, M., Hu, H.-Y., Zhao, C., Wurtz, J., Weinberg, P., Hamdan, M., Zhdanov, A., Cantu, S.H., Zhou, H., Araiza Bravo, R.: Large-scale quantum reservoir learning with an analog quantum computer. PREPRINT as of 2026-08-26 (2024)arxiv.org/abs/2407.02553
  20. [20]Liu, D.-S., Jie, Q.-X., Zou, C.-L., Ren, X.-F., Guo, G.-C.: Practical quantum reservoir computing in rydberg atom arrays. Phys. Rev. A 113, 042401 (2026) https://doi.org/10.1103/pkhd-pl3w
  21. [21]Nakajima, K., Fujii, K., Negoro, M., Mitarai, K., Kitagawa, M.: Boosting Computational Power through Spatial Multiplexing in Quantum Reservoir Computing. Phys. Rev. Applied 11(3), 034021 (2019) https://doi.org/10.1103/PhysRevApplied.11.034021 . Accessed 2023-08-03
  22. [22]Hou, Y., Hua, J., Wu, Z., Xia, W., Chen, Y., Li, X., Li, Z., Peng, X., Du, J.: High-accuracy temporal prediction via experimental quantum reservoir computing in correlated spins. Phys. Rev. Lett. 136, 120602 (2026) https://doi.org/10.1103/r8ww-qw7j
  23. [23]Angelatos, G., Khan, S.A., Türeci, H.E.: Reservoir computing approach to quantum state measurement. Phys. Rev. X 11, 041062 (2021) https://doi.org/10.1103/PhysRevX.11.041062
  24. [24]Ghosh, S., Opala, A., Matuszewski, M., Paterek, T., Liew, T.C.H.: Reconstructing Quantum States With Quantum Reservoir Networks. IEEE Trans. Neural Netw. Learning Syst. 32(7), 3148–3155 (2021) https://doi.org/10.1109/TNNLS.2020.3009716 . Accessed 2023-08-03
  25. [25]Ghosh, S., Krisnanda, T., Paterek, T., Liew, T.C.H.: Realising and compressing quantum circuits with quantum reservoir computing. Commun Phys 4(1), 105 (2021) https://doi.org/10.1038/s42005-021-00606-3 . Accessed 2023-08-03
  26. [26]Gyurik, C., Wudarski, F., Philip, E., Sannia, A., Sadeghi, H., Kyriienko, O., Venturelli, D., Gentile, A.A.: From quantum feature maps to quantum reservoir computing: an applicative perspective. Philosophical Transactions of the Royal Society A 384(2315), 20250085 (2026) https://doi.org/10.1098/rsta.2025.0085
  27. [27]Tariq, S., Talha, M., Ali, A., Divan, M., Chatzinotas, S.: Quantum Reservoir Computing: Recent Advances and Future Directions. PREPRINT as of 2026-08-26 (2026)arxiv.org/abs/2607.18552
  28. [28]Kalfus, W.D., Ribeill, G.J., Rowlands, G.E., Krovi, H.K., Ohki, T.A., Govia, L.C.G.: Hilbert space as a computational resource in reservoir computing. Phys. Rev. Research 4(3), 033007 (2022) https://doi.org/10.1103/PhysRevResearch.4.033007 . Accessed 2023-08-03
  29. [29]Ghosh, S., Krisnanda, T., Paterek, T., Liew, T.C.H.: Realising and compressing quantum circuits with quantum reservoir computing. Communications Physics 4(1) (2021) https://doi.org/10.1038/s42005-021-00606-3
  30. [30]Martínez-Peña, R., Nokkala, J., Giorgi, G.L., Zambrini, R., Soriano, M.C.: Information Processing Capacity of Spin-Based Quantum Reservoir Computing Systems. Cognitive Computation (2020) https://doi.org/10.1007/s12559-020-09772-y . arXiv: 2010.06369
  31. [31]Sannia, A., Martínez-Peña, R., Soriano, M.C., Giorgi, G.L., Zambrini, R.: Dissipation as a resource for quantum reservoir computing. Quantum 8, 1291 (2024) https://doi.org/10.22331/q-2024-03-20-1291
  32. [32]Monzani, F., Ricci, E., Nigro, L., Prati, E.: Non-unital noise in a superconducting quantum computer as a computational resource for reservoir computing. PREPRINT as of 2026-08-26 (2024)arxiv.org/abs/2409.07886
  33. [33]Mifune, S., Kanao, T., Tanamoto, T.: Effects of dissipation in reservoir computing using spin qubit array. Japanese Journal of Applied Physics 64(4), 04–08 (2025) https://doi.org/10.35848/1347-4065/adbf9e
  34. [34]Ricci, E., Monzani, F., Nigro, L., Prati, E.: Quantum reservoir computing induced by controllable damping. npj Quantum Information 12, 107 (2026) https://doi.org/10.1038/s41534-026-01229-8
  35. [35]Palacios, A., Martínez-Peña, R., Soriano, M.C., Giorgi, G.L., Zambrini, R.: Role of coherence in many-body quantum reservoir computing. Communications Physics 7, 369 (2024) https://doi.org/10.1038/s42005-024-01859-4
  36. [36]Vandersypen, L.M.K., Eriksson, M.A.: Quantum computing with semiconductor spins. Physics Today 72(8), 38–45 (2019) https://doi.org/10.1063/pt.3.4270
  37. [37]Philips, S.G.J., Madzik, M.T., Amitonov, S.V., Snoo, S.L., Russ, M., Kalhor, N., Volk, C., Lawrie, W.I.L., Brousse, D., Tryputen, L., Wuetz, B.P., Sammak, A., Veldhorst, M., Scappucci, G., Vandersypen, L.M.K.: Universal control of a six-qubit quantum processor in silicon. Nature 609(7929), 919–924 (2022) https://doi.org/10.1038/s41586-022-05117-x
  38. [38]Fuentes, I., Raymenants, E., Undseth, B., Pietx-Casas, O., Philips, S., Madzik, M., Snoo, S.L., Amitonov, S.V., Tryputen, L., Schmitz, A.T., Matsuura, A.Y., Scappucci, G., Vandersypen, L.M.K.: Running a six-qubit quantum circuit on a silicon spin-qubit array. PRX Quantum 7(1), 010308 (2026) https://doi.org/10.1103/p285-l2v5
  39. [39]Amitonov, S., Apra, A., Asker, M., Bals, R., Barry, B., Bashir, I., Blokhina, E., Giounanlis, P., Harkin, M., Hanos-Puskai, P., Kriekouki, I., Leipold, D., Moras, M., Murphy, N., Petropoulos, N., Power, C., Sammak, A., Samkharadze, N., Semenov, A., Sokolov, A., Redmond, D., Rohrbacher, C., Wu, X.: Persistence of entangled states and high-fidelity quantum gate operations in Si/SiGe spin qubits at high temperature. Applied Physics Letters 129(7), 073503 (2026) https://doi.org/10.1063/5.0312195
  40. [40]Borjans, F., Zajac, D.M., Hazard, T.M., Petta, J.R.: Single-spin relaxation in a synthetic spin-orbit field. Physical Review Applied 11(4), 044063 (2019) https://doi.org/10.1103/PhysRevApplied.11.044063
  41. [41]Wang, N., Wang, S.-M., Zhang, R.-Z., Kang, J.-M., Lu, W.-L., Li, H.-O., Cao, G., Wang, B.-C., Guo, G.-P.: Pursuing high-fidelity control of spin qubits in natural Si/SiGe quantum dot. Applied Physics Letters 125(20), 204002 (2024) https://doi.org/10.1063/5.0230605
  42. [42]Chen, J., Nurdin, H.I., Yamamoto, N.: Temporal Information Processing on Noisy Quantum Computers. Phys. Rev. Applied 14(2), 024065 (2020) https://doi.org/10.1103/PhysRevApplied.14.024065 . Accessed 2023-08-03
  43. [43]Mackey, M.C., Glass, L.: Oscillation and chaos in physiological control systems. Science 197(4300), 287–289 (1977) https://doi.org/10.1126/science.267326
  44. [44]Grebogi, C., Ott, E., Pelikan, S., Yorke, J.A.: Strange attractors that are not chaotic. Physica D: Nonlinear Phenomena 13(1–2), 261–268 (1984) https://doi.org/10.1016/0167-2789(84)90282-3
  45. [45]Jin-Tu, S., Yan, Z., Ying-Hai, W.: Strange Nonchaotic Attractors in a Time-Delay System. Chinese Phys. Lett. 25(7), 2389–2391 (2008) https://doi.org/10.1088/0256-307X/25/7/016 . Accessed 2024-03-19
  46. [46]Ranzato, M., Chopra, S., Auli, M., Zaremba, W.: Sequence level training with recurrent neural networks. In: International Conference on Learning Representations (ICLR) (2016)
  47. [47]He, T., Zhang, J., Zhou, Z., Glass, J.: Exposure bias versus self-recovery: Are distortions really incremental for autoregressive text generation? In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 5087–5102. Association for Computational Linguistics, ??? (2021). https://doi.org/10.18653/v1/2021.emnlp-main.415
  48. [48]Bengio, S., Vinyals, O., Jaitly, N., Shazeer, N.: Scheduled sampling for sequence prediction with recurrent neural networks. In: Advances in Neural Information Processing Systems, vol. 28 (2015)arxiv.org/abs/1506.03099
  49. [49]Ross, S., Gordon, G.J., Bagnell, J.A.: A reduction of imitation learning and structured prediction to no-regret online learning. In: Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics (AISTATS). Proceedings of Machine Learning Research, vol. 15, pp. 627–635. PMLR, ??? (2011)arxiv.org/abs/1011.0686
  50. [50]Petropoulos, N., Wu, X., Sokolov, A., Giounanlis, P., Bashir, I., Mitchell, A.K., Asker, M., Leipold, D., Staszewski, R.B., Blokhina, E.: Nanoscale single-electron box with a floating lead for quantum sensing: Modeling and device characterization. Applied Physics Letters 124(17) (2024) https://doi.org/10.1063/5.0203421
  51. [51]Xia, W., Zou, J., Qiu, X., Li, X.: The Reservoir Learning Power across Quantum Many-Body Localisation Transition. Front. Phys. 17(3), 33506 (2022) https://doi.org/10.1007/s11467-022-1158-1 . arXiv:2104.02727 [cond-mat, physics:quant-ph]. Accessed 2023-08-03
  52. [52]Petropoulos, N., Bogdan Staszewski, R., Leipold, D., Blokhina, E.: Topological order detection and qubit encoding in su–schrieffer–heeger type quantum dot arrays. Journal of Applied Physics 131(7) (2022) https://doi.org/10.1063/5.0082214
  53. [53]Christandl, M., Datta, N., Ekert, A., Landahl, A.J.: Perfect state transfer in quantum spin networks. Physical Review Letters 92(18) (2004) https://doi.org/10.1103/physrevlett.92.187902
  54. [54]Gauthier, D.J., Bollt, E., Griffith, A., Barbosa, W.A.S.: Next generation reservoir computing. Nature Communications 12, 5564 (2021) https://doi.org/10.1038/s41467-021-25801-2
  55. [55]Bollt, E.: On explaining the surprising success of reservoir computing forecaster of chaos? the universal machine learning dynamical system with contrast to VAR and DMD. Chaos: An Interdisciplinary Journal of Nonlinear Science 31(1), 013108 (2021) https://doi.org/10.1063/5.0024890
  56. [56]Petropoulos, N., Blokhina, E.: Impact of quantum information encoding and metallic leads on dynamical multipartite correlation formation in semiconductor quantum dot arrays. Journal of Physics: Condensed Matter 37(5), 055303 (2024) https://doi.org/10.1088/1361-648x/ad906e
  57. [57]Pathak, J., Hunt, B., Girvan, M., Lu, Z., Ott, E.: Model-free prediction of large spatiotemporally chaotic systems from data: A reservoir computing approach. Physical Review Letters 120(2), 024102 (2018) https://doi.org/10.1103/PhysRevLett.120.024102
  58. [58]Vlachas, P.R., Pathak, J., Hunt, B.R., Sapsis, T.P., Girvan, M., Ott, E., Koumoutsakos, P.: Backpropagation algorithms and Reservoir Computing in Recurrent Neural Networks for the forecasting of complex spatiotemporal dynamics. Elsevier BV (2020). https://doi.org/10.1016/j.neunet.2020.02.016 · http://dx.doi.org/10.1016/j.neunet.2020.02.016
  59. [59]Petropoulos, N.: Elements of Quantum Neuromorphic Learning and Information Processing on Semiconductor Quantum Dot Arrays. PhD thesis, University College Dublin (2024). http://hdl.handle.net/10197/30502

Paper details

vlnc:W_6xc6qx83
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

a large language model

View statement
Looking for related papers…
Contents