{"id":"fd06b1a6-bef2-4f4f-8359-e08d5f221d1b","arxiv_id":"2506.15399","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A Raman memory in rubidium vapor stores squeezed vacuum pulses with up to 24 MHz bandwidth, 80% memory efficiency, and 1.0 dB retrieved squeezing from a 1.6 dB input.","lead":"This experiment stores and retrieves squeezed light pulses in a hot rubidium vapor using a far-detuned Raman memory, reaching 80% memory efficiency at up to 24 MHz bandwidth. It provides a path to broadband quantum memories for high-speed continuous-variable quantum networks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported η=0.642 and δ=0.025 SNU cannot yield the claimed 1.0 dB retrieved squeezing from 1.6 dB input under Eq. (1); Eq. (D2) predicts 0.82 dB.","rationale":"The reader's weakest assumption identified exactly the same arithmetic inconsistency: with η = 0.642 and δ = 0.025, a 1.6 dB input yields 0.82 dB output squeezing, not 1.0 dB. My independent calculation confirms this. The concern is concrete, quantitative, and touches the central performance claims of the abstract. It is not an objection to the existence of a broadband Raman memory of squeezed light, nor to the qualitative demonstration; it is a demand for internal consistency among the quoted efficiency, noise, and squeezing values. Because the reader already assigned CONDITIONAL and the same concern is confirmed, no verdict adjustment is needed. The paper should be accepted only after the authors either provide raw variance data showing which number is wrong or clarify the exact definition and scaling of η and δ. The absence of error bars amplifies the concern, since even a modest uncertainty in the 1.6 dB input or 1.0 dB output cannot reconcile a required δ of −0.008 SNU with the reported +0.025 SNU.","tokens_in":10162,"tokens_out":6671,"duration_ms":68194,"concrete_test":"Request the raw phase-resolved variance traces for the B = 4.4 MHz row of Table I. For those traces, compute V_in and V_out directly, apply the stated detector-efficiency corrections (HD1 98.5%, HD2 95%) explicitly, and re-evaluate Eq. (D2) with η = 0.642. If the corrected V_out is 10^(−0.1) = 0.794 SNU and V_in is 10^(−0.16) = 0.692 SNU, the derived δ is −0.008 SNU, which contradicts the reported δ = 0.025. If instead the authors used a different η or a different definition of δ (e.g., input-referred noise, or a separate retrieval efficiency), that definition must be stated and the headline numbers recomputed under it. This one re-analysis of the raw B = 4.4 data settles whether the inconsistency is real or a definitional artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is internally inconsistent. The model is Eq. (1): a_out = sqrt(η) a_in + sqrt(1−η) v + b_th, with excess noise δ = Var(b_th) entered in Eq. (D2) as δ = [V_out − η V_in − (1−η)]/V_vac. For the headline row (Table I, B = 4.4 MHz), the paper reports η = 0.642, δ = 0.025, input squeezing 1.6 dB, and output squeezing 1.0 dB. Converting: V_in = 10^(−1.6/10) = 0.692 SNU; V_out = 10^(−1.0/10) = 0.794 SNU. Inserting the quoted η and δ into Eq. (1) gives V_out = 0.642 × 0.692 + (1 − 0.642) + 0.025 = 0.827 SNU, corresponding to 0.82 dB of squeezing, not 1.0 dB. Even ignoring δ entirely, loss alone predicts V_out = 0.802 SNU, i.e. 0.96 dB, so the claimed 1.0 dB leaves no room for positive excess noise. To match the reported 1.0 dB with η = 0.642, δ would have to be −0.008 SNU, which is unphysical. Thus at least one of the quoted input squeezing, end-to-end efficiency, excess noise, or output squeezing is misreported or defined on a different scale. Since these four numbers are the headline performance claims, the inconsistency is load-bearing: the claim '80% memory efficiency, 0.025 SNU excess noise, and 1.0 dB retrieved squeezing' cannot be simultaneously true under the paper's own model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript reports an experimental far-off-resonant Raman quantum memory for squeezed vacuum states in warm rubidium vapor. The paper claims storage of squeezed light over bandwidths from 4.4 to 24 MHz, with a memory efficiency of 80%, an end-to-end efficiency of 64.2%, excess noise as low as 0.025 SNU, retrieved squeezing of 1.0 dB from a 1.6 dB input, and unconditional fidelity above 92% across the full bandwidth range. The technical approach combines an OPA-based squeezer, write-pulse waveform optimization by a differential evolution algorithm, a backward retrieval strategy, and a dual-pulse time-domain homodyne tomography protocol. The central quantitative claims are summarized in Figs. 2 and 3 and Table I, and are connected through the noisy-channel model of Eq. (1) and the excess-noise estimator Eq. (D2).","tokens_in":10447,"tokens_out":7351,"duration_ms":73158,"significance":"If the quantitative claims survive scrutiny, this would be a substantial experimental advance: the factor-of-12 bandwidth increase over previous narrowband EIT- or QND-based squeezed-light memories, combined with high end-to-end efficiency and sub-shot-noise retrieval, would make the system a practical interface for broadband continuous-variable quantum processing. The dual-pulse phase-tracking tomography and the backward-retrieval noise reduction are concrete technical contributions. The claims are explicit and falsifiable, which is a strength, but they must first be made mutually consistent under the paper's own model.","major_comments":[{"comment":"The four headline numbers are internally inconsistent under the paper's own model. With η=0.642, δ=0.025 SNU, and input squeezing of 1.6 dB (V_in=10^(-1.6/10)=0.692 SNU), Eq. (1) predicts V_out = η V_in + (1−η) + δ = 0.642×0.692 + 0.358 + 0.025 = 0.827 SNU, which is 0.82 dB of squeezing, not the quoted 1.0 dB (0.794 SNU). To reproduce the reported 1.0 dB with η=0.642, one would need δ=−0.008 SNU, which is unphysical. The same contradiction follows directly from Eq. (D2) if the phase-resolved variances are substituted consistently. At least one of the quoted input squeezing, end-to-end efficiency, excess noise, or output squeezing is defined on a different reference plane or is misreported. The manuscript must resolve this by reporting the raw variances at all measured phases and stating explicitly which detection efficiencies enter V_in, V_out, and η. I note that using η=0.80 (the memory efficiency) instead of η=0.642 in the same formula would give V_out≈0.779 SNU, i.e. 1.09 dB, much closer to the quoted 1.0 dB; the authors should clarify which efficiency is actually used in Eq. (D2).","section":"Memory Performance, Eq. (1), Table I, Appendix D (Eq. D2)"},{"comment":"No statistical or systematic uncertainties are given for the squeezing values, efficiencies, excess noise, or fidelities. Without error bars, the reader cannot tell whether the 0.18 dB discrepancy between the quoted output squeezing and the model prediction is a typo, a normalization issue, or a genuine contradiction. The paper should provide standard errors from repeated measurements and a systematic error budget for the homodyne interference efficiencies (98.5% versus 95%) and the 92% photodiode quantum efficiency reported in Appendix A. This is load-bearing because the central claims are numerical and the consistency check depends on these values.","section":"Figs. 2(b), 3, and Table I"},{"comment":"The two homodyne detection paths have different interference efficiencies (98.5% for the input channel and 95% for the output channel), but the main text does not state whether the quoted dB squeezing values are corrected for these differences. Since Eq. (1) relates variances in a common reference plane, the manuscript must specify whether η includes detection losses and whether V_in and V_out are corrected before substitution into Eq. (D2). This point is closely related to the inconsistency above and needs to be settled before the performance claims can be evaluated.","section":"Appendix A and Eq. (D2)"}],"minor_comments":[{"comment":"The figure references are inconsistent: the experimental setup is shown in Fig. 1, but the text in the 'Experimental setup' section refers to 'Fig. 4 (a)', 'Fig. 4 (b)', and 'Fig. 4 (c)', while Fig. 4 is also the detailed setup diagram. Please correct the cross-references.","section":"Experimental setup section"},{"comment":"There are several typos: '87Rbatomic transition' should read '^87Rb atomic transition'; the Fig. 1 caption says 'filp mirror' instead of 'flip mirror'; and the input FWHM is given as 227.2 ns in one place and 227.5 ns in another.","section":"Throughout"},{"comment":"The reconstructed Wigner functions are shown without error bars, contour levels, or a statement of the number of quadrature samples and Hilbert-space truncation used in the maximum-likelihood reconstruction; this information is needed to assess the fidelity claim.","section":"Fig. 2(c)"},{"comment":"The Raman memory equations use g_s, g_a, and Δk without defining all symbols or specifying the boundary conditions; please add a sentence stating the assumptions under which these equations are solved.","section":"Appendix C, Eqs. (C1)-(C3)"},{"comment":"Reference [28] is incomplete: 'See Supplemental Material at for additional details' contains no URL or identifier.","section":"Reference list"}],"recommendation":"major_revision","confidential_remarks":"This is a potentially important experimental result, and the paper is well matched to the journal's scope. The central issue is not circularity: the excess-noise estimate is a standard parameter extraction from measured variances. The risk is that the authors may have used the memory efficiency (0.80) in place of the end-to-end efficiency (0.642) in the noisy-channel model, which would explain the discrepancy. I recommend major revision rather than rejection because the inconsistency appears fixable by a careful resubmission with raw variances, error bars, and explicit normalization."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper reports the first Raman memory of squeezed vacuum states, and the 24 MHz bandwidth is a real step beyond MHz-level EIT/QND systems. But the headline numbers do not survive arithmetic: under the paper's own Eq. (1) and (D2), eta=0.642, delta=0.025, and 1.6 dB input squeezing predict 0.82 dB output squeezing, not the claimed 1.0 dB. So the central quantitative claim is internally inconsistent.\n\nWhat is genuinely new: extending far-off-resonant Raman memory from coherent and single-photon pulses to squeezed vacuum states. The backward retrieval strategy and the dual-pulse time-domain homodyne tomography are practical engineering contributions others will likely adopt. The observed bandwidth scaling and >92% fidelity across 4.4-24 MHz are plausible and, if reproduced, important.\n\nWhere it is soft: the inconsistency above is load-bearing. The quoted memory efficiency, end-to-end efficiency, excess noise, and retrieved squeezing cannot all be true simultaneously. Even ignoring excess noise, loss alone predicts 0.96 dB from 1.6 dB input, so the claimed 1.0 dB leaves no room for positive excess noise; delta would have to be negative. One or more numbers are misreported, or the loss accounting across the two homodyne paths (98.5% vs 95% interference efficiency) is off. There are no error bars anywhere, and the excess noise is inferred from the same variance data via the model, so it is not an independent check. The 43.1 MHz output bandwidth claim also deserves scrutiny: it is derived from a 4.4 MHz input and relies on the read pulse, but the retrieved spectrum is not shown directly.\n\nThe qualitative demonstration is likely real; the numbers need a careful re-analysis with raw data. This is exactly what a serious referee should demand.\n\nWho it is for: CV quantum information experimentalists and the quantum memory community. It deserves peer review, but as a major-revision candidate, not as-is. I would not cite the quantitative claims until they are fixed.\n\nRecommendation: send to peer review, but flag the internal inconsistency and request loss-corrected data with uncertainties.","headline":"Raman memory of squeezed vacuum is a real first, but the headline numbers contradict the paper's own noise model and need major correction before quantitative claims can be trusted.","tokens_in":11063,"tokens_out":4634,"would_cite":false,"duration_ms":39826,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A far-off-resonant Raman memory stores squeezed vacuum light at up to 24 MHz bandwidth with 80 percent memory efficiency, preserving 1.0 dB of squeezing from a 1.6 dB input.","keywords":["quantum memory","squeezed light","Raman memory","continuous-variable quantum information","homodyne tomography","broadband quantum memory","atomic ensemble","excess noise"],"falsifier":"Reconstruct the input and output states with a single calibrated homodyne detector, or apply the paper's own model: with $\\eta = 0.642$ and $\\delta = 0.025$ SNU, a 1.6 dB input should come out with about 0.8 dB of squeezing rather than the quoted 1.0 dB, so a direct measurement of the output squeezing under one common calibration decides whether the quoted numbers are mutually consistent.","tokens_in":1876,"feed_emoji":"⚛️","tokens_out":2531,"duration_ms":70073,"temperature":0.7,"pith_summary":"The paper reports a quantum memory for squeezed vacuum light based on a far-off-resonant Raman process in a hot rubidium vapor. The central claim is that this memory can store squeezed states whose bandwidth covers the full squeezing spectrum of the source, up to 24 MHz, while keeping the stored state's quantum character: an input squeezed by 1.6 dB is retrieved with 1.0 dB of squeezing, an end-to-end efficiency of 64.2 percent, and unconditional fidelity above 92 percent. The authors attribute the performance to a backward retrieval strategy that reaches high write-read efficiency at low read power, lowering excess noise to 0.025 shot-noise units. The demonstration matters because previous squeezed-light memories were narrowband, so high-frequency quadrature information was lost; this memory keeps the full bandwidth available for high-speed continuous-variable protocols.","feed_headline":"Raman memory stores squeezed light at 24 MHz with 80% efficiency","feed_subtitle":"Output keeps 1.0 dB of squeezing from a 1.6 dB input, with fidelity above 92 percent.","key_machinery":"The load-bearing mechanism is a backward-retrieval Raman memory in a hot $^{87}\\mathrm{Rb}$ vapor, in which the write pulse is temporally mode-matched to the input squeezed pulse and the read pulse counter-propagates to extract the spin wave. Backward retrieval is the piece that makes the claims cohere: it delivers high memory efficiency at relatively low read power, and since excess noise grows with read power, reaching 80 percent efficiency without adding much noise is what lets the retrieved state keep its squeezing. The paper's quantitative claims are connected through a noisy-channel model, $\\hat{a}_{\\mathrm{out}} = \\sqrt{\\eta}\\,\\hat{a}_{\\mathrm{in}} + \\sqrt{1-\\eta}\\,\\hat{v} + \\hat{b}_{\\mathrm{th}}$, where $\\eta$ is the end-to-end efficiency and $\\hat{b}_{\\mathrm{th}}$ adds phase-independent excess noise $\\delta$; the authors use this model to extract $\\delta = 0.025$ SNU from measured quadrature variances. The dual-pulse time-domain homodyne tomography supplies the state characterization: a squeezed vacuum pulse gives the quadrature values and a delayed bright coherent pulse with the same waveform tracks the phase, cancelling Stark-shift and phase-drift effects.","core_discovery":"The central discovery is that a far-off-resonant Raman memory in an atomic vapor can act as a broadband, low-noise, high-efficiency memory for squeezed vacuum states. The authors demonstrate write-in and retrieval of a squeezed vacuum with an 80 percent memory efficiency and 80.3 percent optical transmission, giving a 64.2 percent end-to-end efficiency; the output state retains 1.0 dB of squeezing from a 1.6 dB input, with excess noise estimated at 0.025 SNU by a noisy-channel model. They further store squeezed states with bandwidths from 4.4 to 24 MHz, and every retrieved state remains below the shot-noise level with fidelity above 92 percent, including 0.55 dB output squeezing at 24 MHz from a 0.9 dB input. In the authors' picture, the memory protects the full squeezing bandwidth rather than only its low-frequency part, and the backward-retrieval strategy is what makes high efficiency and low noise compatible.","pith_inferences":["The bandwidth-enhancement observation (4.4 MHz input to 43.1 MHz output) suggests the same Raman memory could act as a quantum state converter between sources and processors with different spectral widths, though the paper does not test that function directly.","If the noisy-channel model is the right description, then the squeezing of the retrieved state should be predictable from $\\eta$ and $\\delta$ alone; checking that prediction at more than one input squeezing level would be a direct test of whether the quoted noise is phase-independent.","The excess noise of order 0.02 SNU limits the achievable output squeezing at higher input squeezing; pushing input squeezing beyond 1.6 dB would likely require further suppression of the four-wave-mixing noise, so the present performance may set a practical ceiling for this architecture.","The method of using a delayed bright coherent pulse for phase tracking could be applied to other pulsed squeezed-state measurements where the photon number per pulse is too low for conventional time-domain homodyne tomography."],"forward_implications":["A squeezed state of up to 24 MHz bandwidth can be stored and retrieved with the retrieved noise still below the shot-noise level, so the memory covers the full bandwidth of the squeezed source rather than a narrow slice.","Retrieved squeezing of 1.0 dB from an input of 1.6 dB, with memory efficiency 80 percent and end-to-end efficiency 64.2 percent, implies the memory can serve as a loss- and noise-tolerant interface for continuous-variable protocols that require non-classical states.","A 4.4 MHz input state can be retrieved with 43.1 MHz bandwidth, since the output bandwidth is set by the read pulse, offering bandwidth conversion within a memory.","Fidelity above 92 percent across the 4.4 to 24 MHz range means the memory process preserves the quantum state, not just its photon statistics.","A backward retrieval strategy, rather than forward retrieval, is the practical route to simultaneously high efficiency and low excess noise in Raman memories."],"supporting_citations":[{"why":"Prior electromagnetically induced transparency memory of squeezed light with narrow bandwidth; defines the baseline the paper's 24 MHz bandwidth exceeds.","marker":"[5]"},{"why":"Prior narrowband squeezed-light memory using time-domain homodyne tomography; provides the baseline and the measurement approach the paper extends.","marker":"[6]"},{"why":"Demonstration of broadband Raman memory for weak coherent pulses and single photons; establishes the GHz-bandwidth Raman platform applied here to squeezed light.","marker":"[19]"},{"why":"Broadband Raman memory demonstration; supports the claim that Raman memory can store full-bandwidth states.","marker":"[20]"},{"why":"Optimization of Raman memory efficiency via write-pulse shaping; supplies the mode-matching and retrieval-strategy optimization used in the experiment.","marker":"[21]"},{"why":"Noisy Raman memory equations in the co-moving frame; underlies the backward-retrieval modeling in Appendix C.","marker":"[25]"},{"why":"Optical parametric amplifier squeezer under de-amplification; supplies the source of the 1.6 dB input squeezed states.","marker":"[29]"},{"why":"Maximum-likelihood quantum state tomography; supplies the reconstruction method behind the Wigner functions and fidelity values.","marker":"[30]"},{"why":"Noisy channel model used in Eq. (1) and Eq. (D2); connects measured quadrature variances to the estimated excess noise.","marker":"[31]"},{"why":"Differential evolution algorithm for pulse optimization; provides the write-pulse optimization procedure.","marker":"[32]"}],"fun_headline_variants":["Squeezed light memory stores full bandwidth at 24 MHz","80% efficient memory for squeezed light with 1 dB output","Low-noise Raman memory stores squeezed light up to 24 MHz","Backward retrieval yields 80% efficiency for squeezed light memory"],"cache_read_input_tokens":13056,"weakest_assumption_plain":"The load-bearing premise is that the measured input and output quadrature variances are placed on the same loss-normalized scale, so that one efficiency $\\eta$ and one excess noise $\\delta$ describe the whole memory channel; if the two homodyne paths are calibrated differently, the quoted 1.0 dB output squeezing no longer follows from the 1.6 dB input.","fun_headline_variants_meta":{"raw":{"variants":["Squeezed light memory stores full bandwidth at 24 MHz","80% efficient memory for squeezed light with 1 dB output","Low-noise Raman memory stores squeezed light up to 24 MHz","Backward retrieval yields 80% efficiency for squeezed light memory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000319,"raw_usage":{"total_tokens":1787,"prompt_tokens":916,"completion_tokens":871,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":800}},"tokens_in":532,"tokens_out":871,"duration_ms":7637,"temperature":1.0,"reasoning_tokens":800,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:35:11.888227+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reconstruct the input and output states with a single calibrated homodyne detector, or apply the paper's own model: with $\\eta = 0.642$ and $\\delta = 0.025$ SNU, a 1.6 dB input should come out with about 0.8 dB of squeezing rather than the quoted 1.0 dB, so a direct measurement of the output squeezing under one common calibration decides whether the quoted numbers are mutually consistent.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior electromagnetically induced transparency memory of squeezed light with narrow bandwidth; defines the baseline the paper's 24 MHz bandwidth exceeds."},{"cited_title":"Honda, D","cited_arxiv_id":null,"evidence_quote":"Prior narrowband squeezed-light memory using time-domain homodyne tomography; provides the baseline and the measurement approach the paper extends."},{"cited_title":"Arikawa, K","cited_arxiv_id":null,"evidence_quote":"Demonstration of broadband Raman memory for weak coherent pulses and single photons; establishes the GHz-bandwidth Raman platform applied here to squeezed light."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Broadband Raman memory demonstration; supports the claim that Raman memory can store full-bandwidth states."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Optimization of Raman memory efficiency via write-pulse shaping; supplies the mode-matching and retrieval-strategy optimization used in the experiment."},{"cited_title":"Zhang, J","cited_arxiv_id":null,"evidence_quote":"Noisy Raman memory equations in the co-moving frame; underlies the backward-retrieval modeling in Appendix C."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Optical parametric amplifier squeezer under de-amplification; supplies the source of the 1.6 dB input squeezed states."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Maximum-likelihood quantum state tomography; supplies the reconstruction method behind the Wigner functions and fidelity values."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Noisy channel model used in Eq. (1) and Eq. (D2); connects measured quadrature variances to the estimated excess noise."},{"cited_title":"Zhang, Y","cited_arxiv_id":null,"evidence_quote":"Differential evolution algorithm for pulse optimization; provides the write-pulse optimization procedure."}],"review_version":2}