{"id":"666f9494-5462-4859-bd99-5efaa3255a5d","arxiv_id":"2504.14002","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A support vector regressor using a projected quantum kernel from a 4-qubit Rydberg reservoir predicts 1D fermionic densities and, at suitable measurement times, beats a linear kernel and matches an RBF kernel.","lead":"The authors use a machine-learning kernel built from the measurements of a small Rydberg-atom quantum reservoir to predict electron densities in 1D model molecules and wells. A scientist or engineer might read this to see whether small quantum processors can already add value to density-functional-theory calculations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported errors are measured against truncated basis densities, not true KS densities; basis adequacy is verified only for selected H2 samples and never for the triple-well model, so the central claim may overstate prediction accuracy where the three-region well basis is poor.","rationale":"The reader's weakest assumption identifies exactly this issue: the training targets are the coefficients of the density in the single-particle well basis, and the paper verifies the truncation only on selected samples. This is the most load-bearing structural premise. If the basis is faithful, the reported errors accurately reflect density prediction; if not, the central claim is overstated. The paper's own text acknowledges the basis choice is motivated by the three-region structure of both potentials, and Appendix C defines the basis from fixed well geometries. The concern is not that the basis is wrong for all samples, but that its adequacy is unverified statistically, especially for the triple-well model and for parameter ranges where the wells are shallow or merge. The proposed concrete test directly measures truncation error on all hidden samples and recomputes the error against the true KS density, which would settle whether the reported E is a valid proxy. I found no other concern more fundamental: the kernel construction is not circular, the comparison protocols are described sufficiently for reproduction, and the measurement-time scanning is a performance-curve report rather than a single cherry-picked test result. Since the reader already assigned CONDITIONAL based substantially on this same assumption, my verdict remains unchanged; the conditionality is the appropriate status until the truncation check is carried out.","tokens_in":21518,"tokens_out":4422,"duration_ms":43247,"concrete_test":"Compute for every hidden sample in both models the L1 truncation error T_k = ∫|n_k(x) - n_approx(x)| dx and the prediction error E_true = (1/N_hidden) Σ_k ∫|n_ML,k(x) - n_k(x)| dx, using the true KS density n_k rather than n_approx. Report the mean and maximum T_k over all hidden samples, and compare E_true with the reported E from Eq. (9). Also repeat the truncation check for the wider parameter range of Appendix E (heights in [0.4, 2] a.u., widths in [0.1, 0.7] a.u.). If max T_k is comparable to the typical integrated density or if E_true exceeds E by more than the reported error bars, the basis truncation is the limiting factor and the central claim must be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the PQK-SVR predicts fermionic ground-state densities. However, the error measure in Eq. (9) compares predictions to n_approx, the truncated basis expansion defined in Eq. (4), not to the KS density n_k. The basis is constructed once from fixed single-particle wells in three predefined spatial regions (Appendix C, Eqs. C1-C4), independent of the sample. For configurations where the external potential does not produce three well-separated regions, this fixed basis may represent the true density poorly. In the triple-well model, barrier heights are drawn from [0.8, 1.6] a.u. and widths from [0.2, 0.5] a.u.; low barriers can make the wells merge, and the density can delocalize across regions. With only Ntrunc = 18 basis states (6 per region), a delocalized density may not be well captured. The paper's evidence that n_k ≈ n_approx is a single H2 sample shown in Fig. 2(c); the text states 'there is good agreement between n_k and n_approx, proving that Ntrunc = 30 is sufficient,' but this is not a statistical check and no analogous check is shown for the triple-well model. If the truncation error is large for a non-negligible fraction of hidden samples, the reported E is not a faithful estimate of true density prediction error, and the conclusion that the method 'predicts the density structure' is unsupported for those configurations. This is a load-bearing assumption because the entire training target is defined through this basis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a support vector regression (SVR) scheme based on a projected quantum kernel (PQK) to predict the ground-state density of 1D two-fermion systems, specifically a one-dimensional H2 molecule and a triple-well potential. The kernel is constructed from the measured magnetizations and correlation functions of a four-qubit Rydberg reservoir, with potential features encoded into the detunings of the reservoir Hamiltonian. The SVR is trained on the coefficients of the density in a fixed three-region single-particle basis (Appendix C), and the prediction error is defined in Eq. (9) as the L1 difference between the predicted density and the truncated basis expansion n^(approx) of the Kohn-Sham density. The authors compare PQK-SVR with linear and RBF kernel SVRs and report that, for sufficiently large measurement times, the PQK method outperforms the linear kernel and is competitive with the RBF kernel. They also analyze the dependence of the error on reservoir parameters such as interaction strength, Rabi frequency, and detuning, and observe a common plateau–drop–stabilization behavior of the error as a function of measurement time.","tokens_in":21829,"tokens_out":7164,"duration_ms":64377,"significance":"If the central claim holds, the paper provides a concrete and platform-relevant demonstration that a very small (four-qubit) quantum reservoir can act as a feature map for predicting densities in model fermionic systems, with numerical protocols that are described in detail (Pulser simulations, scikit-learn training, and appendices with formulas). The strength of the paper is its careful numerical setup: 20 random data sets, explicit hyperparameter grids, and a systematic scan over reservoir parameters. The observation that the reservoir needs a minimum interaction time before the error drops, and that the error then stabilizes at a value competitive with classical kernels, is a potentially useful design principle for reservoir-based quantum kernels. However, the significance of the result as a route to 'bypassing Kohn-Sham equations' is currently limited by two methodological issues: the error is measured against a truncated-basis proxy rather than the full KS density, and the headline comparison is made at a measurement time chosen after inspecting the hidden-test error curves. These issues narrow the scope of the claim as presently stated.","major_comments":[{"comment":"The error measure in Eq. (9) compares the predicted density to n^(approx)_k, the truncated basis expansion defined in Eq. (4), not to the full Kohn-Sham density n_k. The only evidence that n_k ≈ n^(approx) is a single H2 sample shown in Fig. 2(c), with the text stating 'there is good agreement between n_k and n^(approx)_k, proving that Ntrunc = 30 is sufficient.' No statistical check over the sample distribution is provided, and no analogous check is shown for the triple-well model, where Ntrunc = 18 and barrier heights can be as low as 0.8 a.u. (Appendix A). Since the basis is built once from fixed well regions (Appendix C), configurations with low barriers that merge the wells may be poorly represented, and for those configurations E is not a faithful estimate of the true density prediction error. The authors should either report the distribution of ||n_k − n^(approx)|| over all samples for both models, or explicitly restate the claim as predicting the truncated-basis density rather than the KS density. As written, the assertion that the method 'predicts the density structure' exceeds what Eq. (9) actually measures.","section":"Sec. II, Eq. (9); Sec. III B; Appendix C"},{"comment":"The headline comparison ('overall, the best performance of the PQK method is obtained in the interval π ≲ t*Ωmax ≲ 3π/2' and the conclusion that PQK 'often outperforms the linear kernel results and can be competitive with the RBF kernel') is drawn from error curves computed on the hidden test set. The paper does not specify a protocol for selecting t* from training data alone, so the reported advantage over linear and competitiveness with RBF is subject to selection bias. Please define a deterministic rule for choosing t* (for example, based on a validation split or on the magnetization dynamics discussed in Sec. III B) and report errors at that fixed t* for all methods, or report the error averaged over a time window that is chosen without reference to the hidden labels.","section":"Sec. III A, Figs. 2(a), 3, 4; Sec. IV"},{"comment":"The definition of the piecewise function is internally inconsistent: P(z,y) = θ(z−x) − θ(y−x) with z < y equals −1 on the interval [z,y], not +1 as stated in the text ('The function P(z,y) is one in the interval [z,y] and zero elsewhere'). This sign error propagates through Eqs. (C1)–(C4), which define the single-particle potentials used to construct the basis for the density expansion. Because the basis is load-bearing for the target coefficients, this inconsistency must be corrected (either the definition or the description) and the basis construction re-verified so that the numerical protocol is reproducible and self-consistent.","section":"Appendix C, Eq. (C1)"},{"comment":"The grid-search procedure for the SVR hyperparameters C and ϵ (and γ for the RBF kernel) is described, but the manuscript does not state whether the grid search is performed using the training set only (e.g., by cross-validation) or using the hidden test set. If the hidden set is used for hyperparameter selection, the reported errors are optimistic and the comparison is not a fair assessment of prediction performance. Please clarify the protocol and, if necessary, modify the evaluation so that all hyperparameters, including t*, are selected without access to the test labels.","section":"Sec. II, Fig. 2 caption; Appendix D"}],"minor_comments":[{"comment":"The scalar product in Eq. (C3) should read ⟨Ψ(ℓ′), Ψ̃(ℓ)⟩ = ∫ dx Ψ*_ℓ′(x) Ψ̃_ℓ(x); the current expression contains Ψ̃_ℓ′(x) on the right-hand side, which appears to be a typo.","section":"Appendix C, Eq. (C3)"},{"comment":"In the last line of Eq. (D9), the second term of the kernel should involve ρ_t*(v_l) σ^i σ^j with i < j, consistent with the definition in Eq. (6); the current expression Tr[ρ_t*(v_l) σ^i σ^i] is dimensionally consistent only for i = j and appears to be a typographical error.","section":"Appendix D, Eq. (D9)"},{"comment":"The caption states 'In any case the global detuning term is set to ∆glob = 5 rad/µs' but the description of the panels is somewhat ambiguous; please clarify which parameters are common to panels (a), (b), and (c) and which vary.","section":"Sec. III B, Fig. 4 caption"},{"comment":"The bottom panels of Fig. 2(c) appear to show signed differences between the approximated KS density and the ML densities, whereas Eq. (11) defines δn^ML as an absolute value; please clarify the plotting convention.","section":"Fig. 2(c) and Eq. (11)"},{"comment":"The abstract says the method 'outperforms the classical linear kernel method and can be competitive with the radial basis function method' at large measurement times, while the conclusion says 'often outperforms the linear kernel results and can be competitive with the RBF kernel.' These statements should be aligned and, if the post hoc time selection is corrected, the wording should reflect the protocol under which the claim holds.","section":"Abstract and Sec. IV"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is interesting but the current evidence does not yet support the full strength of the claim. The main concern is that the reported errors are computed against a truncated-basis proxy with only a single-sample check of the proxy quality, and the comparison advantage is obtained at a measurement time selected from the hidden-test error curves. These are fixable by adding statistical checks of the truncation error for both models and by adopting a pre-specified t* selection rule. The sign inconsistency in the definition of P(z,y) in Appendix C also needs correction. If these issues are addressed, the paper could be a solid contribution to quantum reservoir-based kernel methods for model fermionic systems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nRead Perciavalle et al. on predicting fermionic densities with a projected quantum kernel. The short version: it's a careful numerical benchmark, well within the capability of a competent group, and the central conclusion—PQK competitive with RBF and better than linear at suitable measurement times—is plausible from the data. But the result is narrower than the abstract suggests, and there are two methodological soft spots worth taking seriously.\n\nWhat's new: the combination of a Rydberg-based PQK with SVR for density prediction, and the systematic map of error versus measurement time, including the plateau-drop-stabilization shape and its tie to reservoir interactions. The reservoir is only four qubits, which makes the demonstration clean and easily reproducible in principle. The internal consistency checks—20 random data sets, grid searches, parameter sweeps—are more than many QML papers bother with.\n\nThe first soft spot: the error in Eq. (9) is measured against the truncated basis density n_approx, not the full KS density n_k. The basis is built from fixed single-particle wells in three spatial regions. The paper shows one H2 sample where n_k ≈ n_approx, but that's an anecdote, not a distributional check. For the triple-well model, no such check is shown at all. If low barriers cause densities to delocalize across regions, the basis may miss real structure, and then the reported E understates the true prediction error. The learned targets are coefficients in this basis, so this is load-bearing.\n\nSecond: the measurement-time interval where the PQK 'wins' is selected after looking at the hidden test error curves. That's post hoc. The paper honestly reports full curves, so a reader can see the dependence, but any headline claim of outperformance should be tied to a pre-specified or validation-selected t*. Without that, the comparison is optimistic.\n\nI don't think either issue kills the paper. The method is what it is: a kernel on reservoir observables, and the benchmarks are toy 1D models. The advantage over RBF is parity, and over linear is expected for a nonlinear kernel. Still, the empirical characterization of the error's time-dependence is genuinely useful for people building quantum reservoir kernels.\n\nWho's this for: QML researchers benchmarking kernel methods, and anyone interested in ML surrogates for DFT on small model systems. The citation pattern is ordinary—prior PQK and reservoir papers are cited correctly, and self-citations appear only as background. It's not a breakthrough, but it's a serious piece of numerical work.\n\nRecommendation: send it to review, but ask for two things before acceptance: a statistical check of truncation error across the full sample distribution (at minimum for the triple-well model), and a selection protocol for t* that doesn't use the hidden test set. Also, releasing code would help.\n\nBest,\n[Your name]","headline":"A careful but modest numerical study of a projected quantum kernel for 1D density prediction; the main novelty is the error-vs-time characterization, and the central claim is plausible but rests on a truncated-basis target that is only spot-checked.","tokens_in":22421,"tokens_out":5083,"would_cite":false,"duration_ms":40755,"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 projected quantum kernel built from the measured observables of a four-qubit Rydberg reservoir predicts the ground-state densities of 1D fermionic systems, outperforming a classical linear kernel and matching an RBF kernel once the…","keywords":["projected quantum kernel","support vector regression","Rydberg reservoir","fermionic density","Kohn-Sham equations","density functional theory","quantum machine learning","1D quantum chemistry"],"falsifier":"Take a triple-well configuration with barrier heights near zero so the three wells merge into one broad well, compute the exact KS density $n_k$ and its truncated expansion $n_{\\mathrm{approx}}$, and evaluate the L1 distance; if the distance is large and the PQK prediction tracks $n_{\\mathrm{approx}}$ rather than $n_k$, the method's accuracy is limited by the basis, not by the quantum kernel. Also, a statistical check across the full test distribution (not just selected samples in Fig. 2(c)) comparing $n_k$ and $n_{\\mathrm{approx}}$ would settle whether the truncation error is negligible on average.","tokens_in":21271,"feed_emoji":"⚛️","tokens_out":7499,"duration_ms":60959,"temperature":0.7,"pith_summary":"This paper claims that a support vector regressor armed with a projected quantum kernel can predict the ground-state density of two-fermion 1D systems from the external potential alone, without solving the Kohn-Sham equations. The kernel is built from single-spin and two-spin measurements on a four-qubit Rydberg reservoir whose detunings encode the potential features. Across two model systems (an H2-like molecule and a triple-well potential), the prediction error shows the same shape in measurement time: a short plateau, a sharp drop, then stabilization. At stabilized times the quantum kernel beats a linear kernel and performs on par with an RBF kernel, and the authors read this as evidence that a minimal quantum reservoir can serve as a feature map for density prediction.","feed_headline":"A quantum kernel from four Rydberg atoms predicts fermionic densities","feed_subtitle":"With enough measurement time, it beats the linear kernel and rivals the RBF kernel on model molecules.","key_machinery":"The central object is the projected quantum kernel (PQK). It is $K(v_k,v_l) = m_k \\cdot m_l$, where $m_k$ is the vector of measured expectation values $\\langle \\sigma_j^z \\rangle$ and $\\langle \\sigma_i^z \\sigma_j^z \\rangle$ at time $t^*$, evolved under a four-qubit Rydberg Hamiltonian with the potential features encoded in local detunings. This mapping enlarges the feature space from $N_f$ inputs to ten observables, and the time-dependent dynamics make the kernel a tunable, nonlinear feature map; the reservoir interaction is what lifts degeneracies among the observables. The regression targets are the coefficients $\\{u^{(\\ell)}\\}$ of the density in a truncated basis formed from single-particle eigenstates of left, center, and right well potentials, orthonormalized via Gram-Schmidt.","core_discovery":"The paper's central claim is that the map from potential parameters to fermionic density can be learned by an SVR whose kernel is the inner product of observable vectors $m_k = \\{\\langle \\sigma_j^z \\rangle\\} \\cup \\{\\langle \\sigma_i^z \\sigma_j^z \\rangle\\}$ taken after a fixed evolution time $t^*$. With four interacting Rydberg qubits and the right choice of $t^*$, the PQK error on held-out density profiles is systematically lower than a linear kernel in the two test problems, and sits in the same range as an RBF kernel. The authors also identify a robust qualitative law: the error is flat at short times, drops rapidly once the reservoir dynamics have had time to spread information, and then stabilizes; the drop time is set by the interaction strength, and the non-addressed qubits become distinguishable only when interactions break the encoding symmetry. The method's best operating window, for the H2 problem, is around $\\pi \\leq t^*\\Omega_{\\max} \\leq 3\\pi/2$, where errors are competitive with RBF.","pith_inferences":["Since the training target is itself a truncated basis expansion, the method's ceiling is the representational power of the three-well basis; away from the sampled potential ranges (e.g., very low barriers that merge the wells), the reported advantage over linear and RBF kernels may not transfer even if the kernel is perfect.","A direct test of the reservoir's contribution would be to replace the PQK with a random but fixed nonlinear feature map of the same dimension and compare errors; if the quantum dynamics matter beyond dimension, the PQK should win across measurement times.","The plateau-drop structure suggests the kernel's effective rank grows with time; measuring the observable-vector covariance as a function of $t^*$ could quantify when the feature space saturates, and whether the optimal window coincides with maximal rank."],"forward_implications":["If correct, KS-free density prediction for these model systems needs only four qubits, making the method testable on current Rydberg hardware.","The universal plateau-drop-stabilize error curve provides a practical rule for choosing measurement time: aim for the stabilized window instead of tuning the pulse naively.","The finding that reservoir interactions improve performance suggests interaction strength is a useful control knob for encoding quality, not just a source of noise.","Extending the pipeline with a second SVR that maps predicted densities to energies would give a complete bypass of the KS self-consistency loop."],"supporting_citations":[{"why":"Establishes the task this paper follows: learning the potential-to-density map to bypass Kohn-Sham equations.","marker":"[7]"},{"why":"Provides the LDA exchange-correlation potential used to generate the training densities.","marker":"[46]"},{"why":"Supplies the local-encoding quantum reservoir method and the projected-quantum-kernel construction adapted here.","marker":"[29]"},{"why":"Defines projected quantum kernels, the kernel class at the heart of the method.","marker":"[23]"},{"why":"Provides the one-dimensional molecular model and reference electronic structure framework for the H2 test case.","marker":"[30]"},{"why":"Gives the support vector regression implementation used to train the models and tune hyperparameters.","marker":"[3]"},{"why":"Provides the simulator used to evolve the Rydberg reservoir and compute the measured observables.","marker":"[52]"}],"fun_headline_variants":["Four Rydberg atoms power a kernel that predicts fermionic densities","Rydberg quantum kernel rivals RBF and beats linear on fermionic densities","Quantum kernel with four Rydberg atoms beats linear density models","Rydberg-based quantum kernel predicts fermionic densities accurately"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training targets are coefficients of the density in a truncated basis of single-particle well states for the three potential regions; if a test potential yields a density that this basis cannot capture, the SVR target stops being a faithful proxy for the true density.","fun_headline_variants_meta":{"raw":{"variants":["Four Rydberg atoms power a kernel that predicts fermionic densities","Rydberg quantum kernel rivals RBF and beats linear on fermionic densities","Quantum kernel with four Rydberg atoms beats linear density models","Rydberg-based quantum kernel predicts fermionic densities accurately"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000826,"raw_usage":{"total_tokens":3573,"prompt_tokens":871,"completion_tokens":2702,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":2628}},"tokens_in":487,"tokens_out":2702,"duration_ms":17851,"temperature":1.0,"reasoning_tokens":2628,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:58:35.169468+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a triple-well configuration with barrier heights near zero so the three wells merge into one broad well, compute the exact KS density $n_k$ and its truncated expansion $n_{\\mathrm{approx}}$, and evaluate the L1 distance; if the distance is large and the PQK prediction tracks $n_{\\mathrm{approx}}$ rather than $n_k$, the method's accuracy is limited by the basis, not by the quantum kernel. Also, a statistical check across the full test distribution (not just selected samples in Fig. 2(c)) comparing $n_k$ and $n_{\\mathrm{approx}}$ would settle whether the truncation error is negligible on average.","supporting_citations":[{"cited_title":"Lo Monaco, M","cited_arxiv_id":null,"evidence_quote":"Defines projected quantum kernels, the kernel class at the heart of the method."},{"cited_title":"Jirauschek and T","cited_arxiv_id":null,"evidence_quote":"Provides the LDA exchange-correlation potential used to generate the training densities."},{"cited_title":"D’Amore, L","cited_arxiv_id":null,"evidence_quote":"Supplies the local-encoding quantum reservoir method and the projected-quantum-kernel construction adapted here."}],"review_version":1}