{"id":"7c685e97-7c5f-44fd-b1ae-18efa8cfa743","arxiv_id":"2412.07910","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A quantum echo-state network with data re-uploading and measure-and-reset readout estimates the y and z components of the Lorenz system from x alone in simulation, claiming lower error than an equal-size classical echo-state network.","lead":"This paper proposes a quantum echo-state network circuit that feeds a chaotic time series through random, sparse quantum gates and trains a classical linear model on the measurement distribution. On a simulated Lorenz benchmark it reports lower error than a small classical echo-state network, but the comparison is weakened by best-seed selection and test-set tuning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The QESN-versus-ESN comparison is not apples-to-apples: the QESN uses 2^(nq/2) distribution features while the classical ESN is limited to nq/2 reservoir nodes, so the reported advantage may simply reflect feature count, not quantum reservoir dynamics.","rationale":"The reader's weakest_assumption correctly identifies evaluation bias from best-seed selection and test-error tuning, and also notes the classical ESN is deliberately small. My stress-test focuses on a sharper, more load-bearing problem: the comparison fixes reservoir nodes but not feature dimension, so the QESN's probability-distribution readout has exponentially more linear readout features than the classical ESN has state variables. This asymmetry alone can explain the reported performance gap, independent of any quantum advantage. The paper's own results support this reading: in Table 1, the distribution-based recovery consistently outperforms expectation-value recovery, and the expectation-value variant uses only nq/2 features, matching the classical ESN's state dimension. That internal comparison suggests the gain is driven by feature dimensionality rather than by the quantum reservoir's temporal memory. This does not invalidate the architecture as a reservoir-computing proposal, but it does invalidate the headline comparative claim as currently evidenced. I therefore keep the REJECT verdict, with the caveat that a capacity-matched comparison could rehabilitate the claim.","tokens_in":6215,"tokens_out":2179,"duration_ms":23620,"concrete_test":"Re-run the Figure 4 comparison with a classical ESN whose reservoir dimension is set to the number of QESN distribution features: for nq/2 readout qubits, use a classical ESN with 2^(nq/2) reservoir nodes (or, alternatively, a classical nonlinear random feature map with the same output dimension), trained under identical validation-based hyperparameter selection and averaged over seeds. If the classical model achieves RMSE at or below the QESN for each nq, the central comparative claim is refuted in its current form.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is stated in 'Comparison with Classical Techniques': 'we consistently observed a large performance gain on the noiseless QESN simulation in comparison to the classical ESN when the number of classical ESN reservoir nodes is fixed to the number of readout qubits in the QESN.' This comparison is not equating model capacity. For nq/2 readout qubits, the QESN's 'entire probability distribution' recovery produces 2^(nq/2) features (Table 1 and the text: 'the entire probability distributions over the computational basis, which give 2^(nq/2) features'), whereas the classical ESN has only nq/2 reservoir state variables. Even with a linear readout, the QESN distribution method therefore fits a readout weight matrix of size 2^(nq/2) x 2, while the classical ESN fits roughly (nq/2) x 2 weights. At nq=16, that is 256 features versus 8 reservoir nodes. A classical reservoir with 256 nonlinear features—or any classical nonlinear random feature map of the same dimension—could plausibly match or beat the QESN on the same task without invoking quantum memory. The paper's attribution of the gain to 'the inherent expressivity of qubits over bits' is therefore unsupported by the controlled comparison. The additional evaluation weaknesses noted by the reader (best-of-seeds selection and elastic-net tuning on test error, as stated in 'Data Handling and Training' and the Table 1 caption) compound this: even the QESN's reported numbers are optimistically biased, so the actual gap against a capacity-matched classical baseline is unknown and possibly negative.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a quantum echo-state network (QESN) architecture based on a measure-and-reset readout, context-window data embedding, data re-uploading, and sparse random entanglement. The authors simulate the QESN on a Lorenz system prediction task (x(t) as input, y(t) and z(t) as targets), comparing expectation-value and full-probability-distribution readouts under noiseless and noisy models, and benchmarking against a classical ESN and linear regression. The central empirical claim is that the QESN achieves lower test RMSE than a classical ESN when the number of readout qubits equals the number of reservoir nodes, which the authors attribute to the expressivity of qubits over bits.","tokens_in":6510,"tokens_out":5091,"duration_ms":52487,"significance":"If the central empirical claim were validated, the architecture would be a practical blueprint for recurrent quantum reservoir computing on near-term hardware, particularly because the circuit runs indefinitely without re-initialization. The manuscript has concrete strengths: a clear pseudocode specification (Algorithm 1), a physically motivated measure-and-reset mechanism building on Hu et al. (2024), and the inclusion of a realistic IBM Fez noise model. However, the main claim is not currently supported: the QESN comparison is not capacity-matched, and the evaluation protocol is optimistically biased. I do not see a circularity problem, since the readout is fitted on training data and evaluated on test data; the issues are the uncontrolled feature dimension and the test-set-based tuning, both of which are openly described in the manuscript.","major_comments":[{"comment":"The comparison between the QESN and the classical ESN is not apples-to-apples because the feature dimensions differ by an exponential factor. As stated in 'Results and Analysis', the QESN distribution readout uses 2^(nq/2) features, while the classical ESN uses only nq/2 reservoir nodes. For nq = 16, this is 256 features versus 8 nodes. The performance gap in Figure 4 could therefore be explained by the number of nonlinear features rather than by quantum reservoir dynamics. To support the claim that qubits provide 'inherent expressivity over bits', the authors must compare against a classical reservoir with a comparable number of nonlinear features, for example a classical ESN with 2^(nq/2) nodes or a random nonlinear feature map of the same dimension. This is load-bearing because the stated advantage over the classical ESN is the paper's primary empirical result.","section":"Comparison with Classical Techniques / Figure 4"},{"comment":"The evaluation protocol is not a valid estimate of generalization. The paper states that elastic net parameters are tuned 'with the goal being to minimize test error rate', and the Table 1 caption reports that 'the best run from each category was used'. This selection of the best seed and direct optimization on test error injects an optimistic bias into every reported RMSE value in Table 1 and Figure 4. The authors should split the data into training, validation, and test sets, select hyperparameters on the validation set, and report the mean and standard deviation (or confidence intervals) of the RMSE over multiple random seeds. Without this, the reported performance gains are not statistically grounded.","section":"Data Handling and Training / Table 1 caption"}],"minor_comments":[{"comment":"The definitions of the pairing sets P and M are ambiguous; the pseudocode refers to P1, P2 and M1, M2 but the mapping from these indices to physical qubit pairs is not specified. Please clarify with explicit indexing.","section":"System Architecture / Algorithm 1"},{"comment":"The sentence 'Randomly initialized weights gates play an important role' contains a typo; it should read 'Randomly initialized weight gates'.","section":"System Architecture"},{"comment":"The caption writes '2^nq/2 Hilbert space'; this should be formatted as 2^(nq/2) to avoid ambiguity.","section":"Figure 1 caption"},{"comment":"The number of random seeds tested is not specified; the phrase 'several seeds' is too vague, and without this information the reader cannot assess the variability of the results.","section":"Data Handling and Training"},{"comment":"The classical ESN baseline is underspecified: the text does not report the reservoir size (beyond the node count), spectral radius, leakage rate, sparsity, or how the elastic net regularization was tuned. This information is necessary for reproducibility.","section":"Comparison with Classical Techniques"},{"comment":"The 'Linear Regression' baseline is not described; it is unclear whether it uses the same context-window features as the QESN. Please specify the feature set used for this baseline.","section":"Figure 4"}],"recommendation":"reject","confidential_remarks":"This manuscript appears to be a preliminary workshop submission. The central empirical claim is undermined by two documented methodological choices: tuning hyperparameters directly on test error and reporting only the best seed per configuration. Additionally, the QESN-versus-ESN comparison does not control for the exponential difference in feature count. These are not local presentation issues; they require a redesign of the experiments and a reinterpretation of the main conclusion. I would not support publication until the authors provide a capacity-matched baseline and a statistically valid evaluation protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: this is a workshop-style preprint with a coherent QESN architecture, but the central performance claim is not supported by the evidence as presented. The specific circuit—context-window Euler-angle embedding, sparse random two-qubit gates, data re-uploading, and measure-and-reset readout—is a reasonable incremental combination of known ideas, and the claim that the circuit can run indefinitely without reinitialization is the most interesting part. The Lorenz results show the reservoir is doing something sensible, and including a realistic Fez-like noise model is a plus.\n\nThe soft spot is the comparison in Figure 4. The classical ESN is capped at nq/2 reservoir nodes, while the QESN's distribution recovery produces 2^(nq/2) features. At 16 qubits that's 256 features versus 8 nodes. A classical random feature map with 256 nonlinear features might match or beat the QESN on the same task, so the attributed 'expressivity of qubits over bits' is not established. The reported numbers are also optimistic: the authors select the best of several seeds and tune elastic-net regularization directly on test error (stated in Data Handling and Training and the Table 1 caption), and there are no confidence intervals or ablations of c, r, and kappa. That makes the actual gap against any proper baseline unknown.\n\nThe 'predicting chaotic systems' framing overstates the task: it is same-time estimation of y(t) and z(t) from x(t), not forecasting. That is still a legitimate reservoir-computing task, but the title oversells it.\n\nNone of this is fatal to the architecture idea. The measure-and-reset fading memory from Hu et al. is applied sensibly, and the self-citation there is appropriate. The paper just needs a re-evaluation: validation-based hyperparameter selection, seed-averaged metrics, and a capacity-matched classical baseline—ideally a random feature map with the same dimension as the QESN distribution. If the QESN advantage survives that, it's a useful result.\n\nWho's it for: people working in quantum reservoir computing and NISQ time-series, who will read it for the architecture. I would not cite it in its current form, but it deserves a serious referee, and a major revision could make it a solid paper.","headline":"A coherent QESN architecture with a useful no-reset mechanism, but the central performance comparison is undermined by a feature-count mismatch and optimistic evaluation.","tokens_in":7132,"tokens_out":2136,"would_cite":false,"duration_ms":19775,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A quantum echo-state network with only a few readout qubits predicts the chaotic Lorenz system with lower error than an equal-sized classical echo-state network, and the circuit can run continuously without re-initialization.","keywords":["quantum echo-state network","reservoir computing","time-series prediction","Lorenz system","chaos","measure-and-reset","NISQ","quantum machine learning"],"falsifier":"Rerun the Lorenz prediction task with a classical echo-state network whose reservoir has $2^{{number of readout qubits}}$ nodes (the dimension of the quantum feature space), selecting hyperparameters on a held-out validation set and averaging over many seeds; if the classical network matches or beats the QESN's test RMSE, the reported quantum advantage collapses.","tokens_in":5942,"feed_emoji":"⚛️","tokens_out":6706,"duration_ms":59309,"temperature":0.7,"pith_summary":"This paper proposes a quantum circuit that acts as an echo-state network, a type of recurrent neural network whose random reservoir learns a high-dimensional embedding. The circuit is designed to run indefinitely by measuring and resetting only the readout qubits, preserving a fading memory over long time horizons. In simulations of the chaotic Lorenz system, the authors report that the quantum reservoir predicts the y and z variables from the single x signal with lower root-mean-square error than a classical echo-state network with the same number of reservoir nodes. They also find that using the full measurement probability distribution as regression features outperforms using Pauli expectation values. The work aims to show that small noisy quantum devices are capable of competitive time-series forecasting.","feed_headline":"Quantum reservoir outdoes equal classical net on chaotic forecasting","feed_subtitle":"A few readout qubits beat an equal-sized classical reservoir on the chaotic Lorenz system.","key_machinery":"The QESN circuit is the central object: a parameterized quantum circuit with randomly initialized, sparsely connected weights, split into memory and readout registers. The input is embedded through a sliding context window that maps to three Euler angles per qubit via a fully connected classical layer; repeated circuit blocks of single-qubit rotations, CNOT gates, and controlled RY, RX, RZ rotations create nonlinear combinations of current and past inputs. Mid-circuit measurement and deterministic reset of the readout qubits supplies the echo-state property, the fading-memory feature that makes reservoir states depend on recent inputs. The regression features are the full probability distribution over the readout computational basis, combined with a classical elastic-net fit.","core_discovery":"The paper's own claim, stated in its comparison section, is that a noiseless simulated quantum echo-state network (QESN) consistently yields a large performance gain over a classical echo-state network when the classical reservoir's node count is fixed to the number of readout qubits in the QESN. On the chaotic Lorenz system the authors show that with only the x(t) signal as input, the QESN predicts the y(t) and z(t) components with lower test RMSE than the equal-sized classical reservoir. They further claim that the circuit, which interleaves data-reuploaded rotations with entangling gates and randomly sparse weights, can be run indefinitely by measuring and resetting only the readout register, preserving a fading memory over long horizons. The paper also reports that using the full measured probability distribution as regression features outperforms using Pauli expectation values, and that results collected under a realistic IBM Fez noise model are degraded but still promising for near-term hardware.","pith_inferences":["The reported advantage may stem from the quantum feature space's dimension (2^{readout qubits}) rather than from intrinsic quantum effects; a classical reservoir with an equal number of random nonlinear features could be a stronger baseline than the one used in the paper.","The indefinite-run design is not specific to the Lorenz system; extending it to other chaotic systems or PDE forecasting is a direct next step, but the paper demonstrates only this one benchmark.","Because the QESN trains only the classical readout, the architecture should be robust to barren plateaus in the trained stage; the main tunable quantum resource is the random initialization and sparsity of the reservoir, which deserves separate study."],"forward_implications":["If the result holds, a quantum reservoir of only a few qubits can substitute for a much larger classical reservoir in forecasting chaotic signals, lowering the hardware budget for reservoir computing.","The measure-and-reset design means the circuit need not be stopped or re-initialized, so long-horizon streaming prediction on NISQ devices is in principle possible.","Using full probability distributions as features is a low-cost way to extract more information from each quantum measurement, and should be preferred over expectation values in similar QESN setups.","The noise-model results suggest that current NISQ hardware can still sustain useful predictions, though with degraded accuracy, supporting further attempts on real devices."],"supporting_citations":[{"why":"Defines the classical echo-state network and the echo-state (fading memory) property that the QESN is designed to emulate; also supplies the baseline architecture.","marker":"(Jaeger 2001)"},{"why":"Introduces the measure-and-reset paradigm that lets a quantum circuit run indefinitely while maintaining a fading memory, the design the QESN directly builds upon.","marker":"(Hu et al. 2024)"},{"why":"Provides the data re-uploading technique used to insert nonlinearity into the circuit and to re-encode each context window.","marker":"(Pérez-Salinas et al. 2020)"},{"why":"Argues that nonlinear input transformations are standard in quantum reservoir computing, supporting the use of re-uploaded rotations and entangling gates to generate nonlinear features.","marker":"(Govia et al. 2022)"},{"why":"Makes the case that Hilbert space size acts as a computational resource in reservoir computing, motivating the use of the full probability distribution as a high-dimensional feature set.","marker":"(Kalfus et al. 2022)"},{"why":"Supports the use of sparse, randomly structured reservoir connections, which the QESN mimics through its sparsity parameter and random weight matrices.","marker":"(Miao, Narayanan, and Li 2022)"}],"fun_headline_variants":["Quantum ESN beats classical reservoir on chaos","Small quantum net outpredicts equal classical one","Noiseless QESN slashes error on Lorenz prediction","Quantum echo-state net wins on chaotic system","QESN outperforms classical net on chaos forecast"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that picking the best seed and tuning the elastic-net regularization directly on the test set gives a fair estimate of predictive performance, and that a classical echo-state network capped at 2–8 reservoir nodes is the right baseline to compare against.","fun_headline_variants_meta":{"raw":{"variants":["Quantum ESN beats classical reservoir on chaos","Small quantum net outpredicts equal classical one","Noiseless QESN slashes error on Lorenz prediction","Quantum echo-state net wins on chaotic system","QESN outperforms classical net on chaos forecast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1598,"prompt_tokens":950,"completion_tokens":648,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":575}},"tokens_in":566,"tokens_out":648,"duration_ms":6329,"temperature":1.0,"reasoning_tokens":575,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:25:52.808628+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the Lorenz prediction task with a classical echo-state network whose reservoir has $2^{{number of readout qubits}}$ nodes (the dimension of the quantum feature space), selecting hyperparameters on a held-out validation set and averaging over many seeds; if the classical network matches or beats the QESN's test RMSE, the reported quantum advantage collapses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the classical echo-state network and the echo-state (fading memory) property that the QESN is designed to emulate; also supplies the baseline architecture."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the data re-uploading technique used to insert nonlinearity into the circuit and to re-encode each context window."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Argues that nonlinear input transformations are standard in quantum reservoir computing, supporting the use of re-uploaded rotations and entangling gates to generate nonlinear features."},{"cited_title":"D.; Ribeill, G","cited_arxiv_id":null,"evidence_quote":"Makes the case that Hilbert space size acts as a computational resource in reservoir computing, motivating the use of the full probability distribution as a high-dimensional feature set."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the use of sparse, randomly structured reservoir connections, which the QESN mimics through its sparsity parameter and random weight matrices."}],"review_version":1}