{"id":"8932f956-811c-410b-8563-cb029fa3755b","arxiv_id":"2602.14736","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Coupled photonic quantum memristors with crossed feedback, driven by a room-temperature SiV− single-photon source, show enhanced, self-intersecting hysteresis in inter-memristor input-output curves.","lead":"Two coupled photonic quantum memristors — adaptive light switches whose setting depends on past photon counts — were built on one silicon nitride chip and fed by a room-temperature nanodiamond single-photon source. Coupled in this way they show richer, self-intersecting memory behavior than single devices, a step toward photonic neuromorphic computing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Polarization drift during a run can corrupt the calibrated MZI response, biasing every reconstructed photon flux and hence the hysteresis loops that constitute the central demonstration.","rationale":"The central claim requires reliable reconstruction of ⟨N_in⟩ and ⟨N_out⟩; if those are corrupted, every hysteresis curve and the crossed-feedback demonstration is compromised. The paper itself concedes that polarization control is infeasible and that realignment changes the polarization, so the sensitivity is not a speculative external worry but an acknowledged limitation. The single-memristor data reproduce the earlier independent experiment [34], giving useful external grounding, and the coupled-device model is parameter-light in its core structure (M, Φ). However, the per-dataset fitted visibilities and phase offsets in Appendix D, the absence of error bars on the hysteresis traces, and the unsupported NARMA sentence in the abstract all reinforce the need for a direct polarization-stability check. None of these secondary issues by itself overturns the coupled-PQM demonstration, so the reader's CONDITIONAL verdict remains appropriate; the specific reference-monitor test above would settle whether the central measurement premise holds.","tokens_in":21245,"tokens_out":13481,"duration_ms":135414,"concrete_test":"Repeat the Fig. 4(d) coupled-memristor run while inserting, in every time bin, a short fixed-phase calibration interval on MZI-M1 (e.g., set θ to the 50% transmission point and record counts). If the inferred reference transmission drifts by more than the Poisson shot-noise level (≈1/√N, with N the counts) over the run, or if the reconstructed R changes by more than ~0.02, polarization/calibration drift is present. Independently, recompute the hysteresis loop from even vs. odd time bins only; if the self-intersection and form factor change qualitatively between the two subsets, the reported loop geometry is not robust to slow calibration drift.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 states that because of the 20 dB insertion loss and the 170 kcps source rate, in-fiber polarization control is not feasible, and that the polarization can change after each realignment, requiring recalibration of the polarization-sensitive MZIs (designed for TE0). The measurement premise of Appendix C (Eqs. 8-10) is that detector-click numbers from one output port, taken at phase settings θ and θ+π/2, can be converted into ⟨N_in⟩ and ⟨N_out⟩ using the calibrated MZI response R = ½[1 − V cos(φ_real)] (Appendix D, Eq. 15). If the input polarization drifts or the calibration is slightly wrong during a 20 s bin, Eq. 15 no longer maps the commanded phase to the actual reflectance R that enters the feedback law (Eq. 7). Because R is updated from these inferred fluxes, every later memristor state is biased. The headline experimental features — larger form factors and self-intersecting inter-relations — are precisely loop-shape properties that a slow time-dependent calibration error could mimic or destroy. Tables 2–3 fit a single static visibility and phase offset per dataset, so they cannot distinguish a slowly drifting polarization from genuine memristive memory dynamics. The paper itself flags the polarization sensitivity; this is therefore not an ad hoc objection but a load-bearing measurement risk.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports the experimental implementation and characterization of a photonic integrated circuit containing two parallel Mach–Zehnder-based photonic quantum memristors (PQMs) with crossed feedback, driven by a room-temperature SiV− single-photon source. The central claim is that the crossed feedback law (Eq. 7) produces non-Markovian inter-memristor dynamics that a single PQM or independent pair cannot produce, evidenced by hysteresis curves with larger form factors and self-intersecting loops. The paper also reports single-PQM results that reproduce Ref. [34], and it presents numerical simulations mapping the form factor and self-intersection behavior over the (Φ, T/Tosc) plane. The abstract further claims an experimental NARMA test, but no NARMA experiment appears in the main text or appendices.","tokens_in":21434,"tokens_out":5107,"duration_ms":56298,"significance":"If the measured hysteresis features are genuine, the work would be a meaningful step beyond isolated PQMs: it demonstrates that crossed feedback between integrated photonic memristors generates richer input–output relations and could serve as a building block for quantum neuromorphic and reservoir computing. The use of a portable, room-temperature SiV− single-photon source coupled to a silicon nitride PIC is also a practical advance. The single-memristor results matching the independent prior experiment [34] provide a useful sanity check. However, the quantitative edge of the paper—larger form factors and self-intersecting loops—rests partly on simulations whose parameters are fitted per dataset (Appendix D, Tables 2–3), and the experimental hysteresis curves are shown without error bars or a direct experimental form-factor measurement. The manuscript itself flags that in-fiber polarization control is infeasible and that the polarization-sensitive MZIs require recalibration after realignment, which is a load-bearing measurement risk for the coupled-device claims.","major_comments":[{"comment":"The central observable—the hysteresis loop—is reconstructed from single-output-port detector clicks using the calibrated MZI law R = ½[1−V cos(φ_real)] and the π/2 output-phase-shift relation. The paper states that in-fiber polarization control is not feasible and that the input polarization can change after each realignment, while the MZIs are designed for TE0 and are polarization-sensitive. Under these conditions, any polarization drift during a 20 s time bin or over a run changes V and the effective phase, biasing every reconstructed ⟨N_in⟩ and ⟨N_out⟩ and hence the feedback update in Eq. (7). The headline features—larger form factors and self-intersections—are loop-shape properties that a slow time-dependent calibration error could mimic or destroy. Since Tables 2–3 fit a single static V and δφ_sta per dataset, they cannot distinguish a slowly drifting polarization from genuine memri","section":"Section 4 and Appendix C, Eqs. (8)–(10); Appendix D, Eq. (15)"},{"comment":"The curves labeled 'simulation' in Figs. 3–4 are not parameter-free predictions: the visibilities and static phase errors in Tables 2–3 are obtained from an optimization routine per experiment, bounded only by calibration values, and stochastic noise is set to zero. Thus the experiment–simulation agreement is partly a fitting exercise. Moreover, the quantitative headline numbers—inter-relation form factors up to ≈0.95 and the self-intersection map in Figs. 5–6—are simulation outputs at the red-marked parameter choices; the paper does not report experimentally measured form factors or uncertainties for the data in Fig. 4. This overstates the experimental support for the central claim. Please either generate simulations using calibration data alone with full uncertainty propagation, or explicitly report experimental form factors with error bars and separate simulation-only statements from","section":"Section 4 and Appendix D, Eqs. (14)–(16); Tables 2–3; Figs. 5–6"},{"comment":"The abstract states: 'We experimentally test the performance of our system in the NARMA task.' I could not find any NARMA experiment, data, or analysis in the main text, Methods, or appendices; Section 5 only discusses possible reservoir-computing applications. This claim is unsupported and should be removed unless the corresponding experiment and results are added.","section":"Abstract"}],"minor_comments":[{"comment":"Typo: 'photonic quantum meristors' should be 'memristors'; also 'briigthness' should be 'brightness'.","section":"Section 2"},{"comment":"'non-unitary visibilities' appears to be a typo for 'non-ideal visibilities' or 'non-unit visibilities'.","section":"Section 4"},{"comment":"The figures show experimental points and simulation curves but no error bars or confidence bands on the experimental data. Adding uncertainty estimates would greatly help the reader assess the significance of loop shapes, especially for the claimed self-intersection.","section":"Fig. 3 and Fig. 4 captions"},{"comment":"The stochastic term ξ_err is introduced but then set to zero in the simulations. Please specify its assumed distribution/amplitude and whether any residual noise was included in the fits.","section":"Appendix D, Eq. (16)"},{"comment":"Typo: 'self-intesecting' should be 'self-intersecting'.","section":"Fig. 6 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is timely and the single-PQM validation is useful, but the coupled-device claims currently lean on fitted simulations and a polarization-sensitive measurement premise that the paper itself concedes is uncontrolled. I do not think this requires rejection, but the authors should add a polarization-stability/calibration-drift analysis, report experimental form factors with uncertainties, and remove or substantiate the unsupported NARMA claim. The three configurations in Fig. 4 are a reasonable starting set, but they are too few to establish the full (Φ, T/Tosc) phenomenology claimed in Figs. 5–6."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine first — two parallel photonic quantum memristors with crossed feedback on an integrated silicon nitride chip, driven by a room-temperature SiV- source. The single-memristor part reproduces Spagnolo et al. [34], which is good external grounding, and the coupled-device phenomenology (non-pinched inter-relation loops, form factors up to ~0.95 in simulation, one self-intersecting inter-relation at a time) follows directly from the crossed-feedback law in Eq. 7. I think the central physics is likely right.\n\nWhere it gets soft. First, the abstract says 'We experimentally test the performance of our system in the NARMA task' but there is no NARMA data or method anywhere in the paper. That claim should either be removed or supported; as written it is misleading.\n\nSecond, the simulation curves in Figs. 3-4 are generated with visibilities and static phase errors fitted per dataset (Appendix D, Tables 2-3). The visibilities are bounded by calibration values, but they are still free parameters per run. So the agreement between experiment and simulation is partly a fit, not an independent prediction. The headline features — F≈0.95 maxima and the self-intersection boundaries in Figs. 5-6 — are simulation outputs, not directly measured. That doesn't kill the paper, but the claims should be softened accordingly.\n\nThird, the hysteresis traces have no error bars, and I don't see any raw data or code deposited. That makes it hard to judge whether the non-pinched loops are robust.\n\nFourth, and most load-bearing: the polarization issue the authors themselves flag in Section 4. The MZIs are polarization-sensitive, designed for TE0, and in-fiber polarization control is not feasible with this source and loss. The input/output fluxes that drive the feedback law are reconstructed from clicks on one output port using the calibrated reflectance law R = ½[1 − V cos(φ_real)]. If the polarization drifts during a 20 s bin, that law no longer maps the commanded phase to the actual reflectance. Since the feedback state R is updated from these inferred fluxes, every hysteresis loop is potentially biased. Tables 2-3 fit a single static visibility per dataset, so they can't distinguish a slow polarization drift from genuine memory dynamics. This is not an ad hoc objection — the paper itself admits the risk. I'd want to see either a polarization stability check over the measurement window or a different way to infer the fluxes.\n\nNet: this is a solid experimental step, and the authors have been honest about some limitations. With the NARMA claim removed or substantiated, error bars added, and a polarization-drift test reported, it would be a good contribution. As it stands, I'd send it to a knowledgeable referee — the core idea deserves referee time — but I'd expect significant revision before acceptance.","headline":"First coupled crossed-feedback photonic quantum memristors: a credible experiment, but an unsupported NARMA claim and polarization-drift risk need attention before I trust the headline numbers.","tokens_in":22094,"tokens_out":5772,"would_cite":true,"duration_ms":49677,"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":"Two coupled photonic quantum memristors with crossed feedback produce inter-memristor hysteresis loops that are larger and self-intersect—behaviour a single device cannot show, pointing to scalable quantum neuromorphic building blocks.","keywords":["photonic quantum memristor","crossed feedback","quantum neuromorphic computing","reservoir computing","silicon nitride photonic integrated circuit","silicon-vacancy color center","single-photon source","hysteresis loop"],"falsifier":"Run the coupled-memristor protocol with four detectors so both outputs of each MZI are monitored simultaneously, and add active polarization stabilization at the input. If the inter-memristor self-intersection or the form factor above the single-memristor value (≈0.95 vs ≈0.58) fails to appear under stabilized conditions, the claimed topology is an artifact of the polarization-sensitive single-port reconstruction rather than of crossed feedback.","tokens_in":21028,"feed_emoji":"🌀","tokens_out":9063,"duration_ms":88027,"temperature":0.7,"pith_summary":"This paper sets out to show that a pair of photonic quantum memristors—integrated Mach–Zehnder interferometers whose splitting ratio is updated by past photon-count measurements—can be coupled so that each memristor's history is driven by the other's input flux (crossed feedback). Working on a silicon nitride chip fed by a room-temperature silicon-vacancy color-center single-photon source, the authors report inter-memristor input-output hysteresis curves that are not pinched at the origin, reach form factors up to about 0.95 (versus about 0.58 for a single memristor), and self-intersect for one inter-relation at a time. Simulations show these features emerge from the interplay of memory depth T/Tosc and the relative phase between the two sinusoidal inputs. If correct, the result would establish coupled photonic quantum memristors as scalable nonlinear elements for quantum neuromorphic and reservoir computing, using only measurement-induced dynamics and standard integrated photonics.","feed_headline":"Crossed-feedback quantum memristors create self-crossing loops","feed_subtitle":"Two interferometer memristors wired to each other's history show richer hysteresis for quantum neuromorphic chips.","key_machinery":"The central object is the Mach–Zehnder-interferometer photonic quantum memristor in dual-rail encoding: a single photon is split between a reference path and an interferometer whose reflectance R(t) acts as the memristive state. R(t) is updated from the history of measured average input photon number over a sliding buffer of M time bins (Eq. 6), which is what makes the dynamics measurement-induced and non-Markovian. The novel element is the crossed-feedback law of Eq. (7), where the reflectance of each memristor is updated from the other memristor's input history, interleaved with a π/2 phase-shift trick that lets one detector per memristor sample both interferometer outputs. The coupled dyn","core_discovery":"The central claim, stated in the abstract and Section 4, is that two parallel photonic quantum memristors governed by the crossed feedback laws R^(1)(t_k) = 1/2 + (1/M) Σ (⟨N_in^(2)(t_j)⟩ − 1/2) and symmetrically for R^(2) realize non-Markovian input-output dynamics that a single PQM or two uncoupled parallel memristors cannot produce. The experimental signatures are inter-memristor hysteresis loops that are large—simulated form factor up to ≈0.95 versus ≈0.58 for a single memristor—and self-intersect, with a self-intersection occurring for exactly one inter-relation at a time. The same chip and source also reproduce the known single-memristor hysteresis behaviour, and the abstract reports a","pith_inferences":["Inference: because the feedback law uses averaged photon numbers, the same chip driven by a classical coherent state at comparable flux should reproduce much of the hysteresis; comparing the two settings would isolate the genuinely single-photon contribution.","Inference: replacing the single-port, two-phase reconstruction with simultaneous monitoring of both outputs and active polarization stabilization is the natural robustness check; if the self-intersection disappears under stabilized conditions, the topology is an artifact of the measurement chain rather than of crossed feedback.","Inference: the observation that only one inter-relation self-intersects at a time suggests a structural constraint in the crossed-feedback equations; this could serve as a diagnostic for non-Markovian cross-correlations in larger memristor networks.","Inference: the two-memristor crossed feedback is the minimal case of cyclic coupling; extending to three or more memristors with staggered phase lags could produce programmable hysteresis geometries for reservoir computing, though this goes beyond the paper's data."],"forward_implications":["If the central claim holds, two crossed-feedback PQMs already produce dynamics—larger-area, non-pinched, self-intersecting hysteresis—that a single PQM cannot, so network connectivity itself becomes a resource for nonlinearity and memory.","The crossed-feedback correlations arise without any direct coupling between the photonic modes; they are mediated entirely by the measurement feedback, so the scheme can be extended to many memristors on a single photonic chip with few detectors.","Room-temperature operation of both the source and the memristors removes a cryogenic bottleneck that limits other single-photon-source PQM demonstrations, easing integration into compact quantum neuromorphic hardware.","The parameter maps for form factor and self-intersection provide a practical recipe for choosing buffer length and input phase to engineer a desired hysteresis topology.","Using one detector per memristor with the π/2 output phase relation halves hardware requirements at the cost of doubled experiment time; future on-chip detection could remove that cost."],"fun_headline_variants":["Coupled quantum memristors self-intersect in memory loops","Two quantum memristors cross their own hysteresis paths","Coupled quantum memristors loop with larger memory hysteresis","Quantum memristor duo self-intersect in hysteresis"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the measured photon counts faithfully encode the memristor input and output fluxes—but the counts are reconstructed from one output port using a polarization-sensitive calibration of each interferometer, and Section 4 concedes that in-fiber polarization cannot be actively controlled and can drift after realignment, so any polarization or calibration error during a 20-second bin corrupts every inferred flux and therefore the reported hysteresis","fun_headline_variants_meta":{"raw":{"variants":["Coupled quantum memristors self-intersect in memory loops","Two quantum memristors cross their own hysteresis paths","Coupled quantum memristors loop with larger memory hysteresis","Quantum memristor duo self-intersect in hysteresis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000699,"raw_usage":{"total_tokens":3028,"prompt_tokens":812,"completion_tokens":2216,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":2148}},"tokens_in":556,"tokens_out":2216,"duration_ms":16291,"temperature":1.0,"reasoning_tokens":2148,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T23:05:09.927187+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the coupled-memristor protocol with four detectors so both outputs of each MZI are monitored simultaneously, and add active polarization stabilization at the input. If the inter-memristor self-intersection or the form factor above the single-memristor value (≈0.95 vs ≈0.58) fails to appear under stabilized conditions, the claimed topology is an artifact of the polarization-sensitive single-port reconstruction rather than of crossed feedback.","supporting_citations":[],"review_version":1}