{"id":"ad44a9d3-2226-402c-8a22-d8211a736564","arxiv_id":"2512.00753","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An OPA-based nonlinear interferometer is proposed that keeps GBS entanglement linear in the number of modes under realistic photon loss, which the authors argue prevents efficient classical simulation.","lead":"This paper proposes a photonic sampling machine in which the passive interferometer of standard Gaussian Boson Sampling is replaced by a network of optical parametric amplifiers, and shows numerically that the light's entanglement keeps growing with system size even when photons are lost. If the result holds, it could make photonic quantum advantage experiments much more tolerant to loss.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's hardness-under-loss claim rests on an unproven inference from linear E_N scaling to exponential MPS/MPO cost; the cited simulability theorems only prove the converse.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: linear E_N scaling is taken as sufficient to conclude exponential MPS/MPO cost, even though the cited simulability results only give the converse implication. My stress test confirms this is the most serious gap in the paper's central argument. The paper's Gaussian-state derivations (Appendix C on loss commutation) are sound, and the numerical observation of linear E_N under loss is interesting and potentially useful. However, the main advertised claim—that computational hardness persists under realistic loss—is not established by the presented evidence. The exact Hafnian hardness of output probabilities does not by itself settle the hardness of approximate sampling from the lossy distribution, and the entanglement-scaling argument conflates E_N with the entanglement measure that controls tensor-network simulation cost. A direct computation of the required MPO bond dimension for the proposed architecture would settle whether the linear E_N actually translates to exponential simulation cost. Since the paper makes a concrete, testable proposal and the deficiencies are fillable, the appropriate verdict remains CONDITIONAL, unchanged from the reader's disposition.","tokens_in":16769,"tokens_out":4211,"duration_ms":45661,"concrete_test":"Implement the MPO simulation algorithm of Oh et al. (Ref. [19]) for the OPA network at t=0.8, d=8, r=0.8, with n=8,10,12,...,30, and measure the minimal bond dimension required to approximate the photon-number distribution to a fixed total variation distance (e.g., ε=10^-3). Compare with the equivalent lossy SU(2) GBS (same squeezing and loss) and with the lossless OPA network. If the bond dimension for the lossy OPA network grows exponentially in n (or in total photon number), the linear-E_N claim supports hardness; if it grows polynomially, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that OPA-boosted GBS remains classically intractable under realistic loss rests on the bridge from the numerical linear scaling of logarithmic negativity E_N with mode number n (Fig. 3e) to the conclusion that tensor-network simulations require exponential resources. That bridge is not established. Refs. [18-20] prove only that sublinear entanglement entropy suffices for efficient MPS/MPO simulation of lossy GBS; they do not prove that linear E_N forces exponential bond dimension. E_N is a mixed-state entanglement monotone, not the von Neumann or operator entanglement entropy that controls MPO bond dimension for Gaussian states; no known theorem maps E_N ~ n to an exponential simulation cost. Indeed, a product of n independent two-mode squeezed states with local loss can have E_N ~ n while being exactly simulable with O(1) bond dimension. The OPA network may generate genuine multipartite entanglement, but the paper never computes the bond dimension actually required by MPS/MPO algorithms for this architecture. Furthermore, exact Hafnian #P-hardness (Eq. 3) does not automatically imply that approximate sampling from the lossy distribution is classically hard; the paper provides no approximate-sampling complexity framework or noise-robustness argument, even though the loss model is central to the claim. These gaps are addressable, which is why the reader's CONDITIONAL verdict is appropriate rather than REJECT.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes replacing the passive SU(2) interferometer in Gaussian Boson Sampling (GBS) with an active network of optical parametric amplifiers (OPAs) forming an SU(1,1) interferometer. The authors derive the covariance-matrix transformation for this network, including a model of photon loss via interleaved beam-splitter layers, and note that the output click probabilities are still given by a Hafnian (Eq. 3). They numerically compute the logarithmic negativity E_N for up to n=10 modes, d=20 layers, and transmittances t in [0.4,1]. They report linear scaling of E_N with squeezing r and depth d in the lossless case, and approximately linear scaling of E_N with mode number n under realistic loss. From this they conclude that the OPA-boosted scheme remains classically hard to simulate under loss, in contrast to standard passive GBS. The manuscript includes appendices with the symplectic derivation, the lossy-channel composition, a Bloch-Messiah argument, and a comparison table of boson-sampling variants.","tokens_in":17027,"tokens_out":5859,"duration_ms":59811,"significance":"The idea of injecting nonclassicality inside the interferometer to counteract loss is timely and, if rigorously supported, would be an important contribution to the photonic quantum-advantage program. The paper uses standard Gaussian-state tools, and the numerical simulations are transparent and reproducible with open-source software; no parameters are fitted to a target result. However, the two load-bearing steps—(i) the bridge from E_N scaling to tensor-network simulation cost, and (ii) the step from exact Hafnian hardness to approximate sampling hardness under loss—are not established. The paper is therefore more a proposal with encouraging numerics than a definitive proof of loss-robust hardness; the claims can be strengthened by direct bond-dimension or approximate-sampling analysis.","major_comments":[{"comment":"The paper's central assertion that the scheme 'sustains computational hardness' under loss (Discussion) rests on the numerical observation that 'EN continues to scale linearly with the number of modes even under lossy conditions' (Fig. 3(e), d=8, r=0.8). This inference is not justified. Refs. [18-20] prove only that sublinear entanglement entropy suffices for efficient MPS/MPO simulation; they do not establish that linear logarithmic negativity forces exponential bond dimension. E_N is a mixed-state entanglement monotone, not the von Neumann or operator entanglement entropy that controls MPS/MPO bond dimension for Gaussian states. A product of n independent two-mode squeezed states with local loss has E_N ~ n yet is exactly simulable with O(1) bond dimension. To support Table I's 'Cannot be classically simulated using MPS/MPO [19,20]', the authors must compute the relevant bond dimension","section":"Entanglement section; Fig. 3(e); Discussion; Appendix D Table I"},{"comment":"The argument from Eq. (3) to sampling hardness is incomplete. Eq. (3) gives an exact Hafnian expression for a single output probability. Exact Hafnian #P-hardness does not imply that approximate sampling from the lossy distribution is classically hard, particularly in the presence of noise. The standard GBS hardness proof requires average-case hardness of the Hafnian, anticoncentration, and a noise-robustness argument. The manuscript does not provide these elements, nor does it explain how the cited works [41,42] apply to the OPA network. Without this framework, the statement in the text that 'the same complexity restrictions that apply to linear GBS also extend to our scheme' is too strong.","section":"Computational complexity section, Eq. (3)"},{"comment":"The asymptotic claim that E_N 'scales linearly with the number of modes even under lossy conditions' is supported by only five data points (n=2,4,6,8,10) at d=8, r=0.8, t>=0.6. No error bars or fits are reported, and the smallest transmittance in this panel is t=0.6, while the text and Fig. 3(c) refer to t=0.4. Given the central role of this scaling, the numerical evidence should be extended to larger n and a wider loss range, and the slope and fit quality should be reported to justify the asymptotic extrapolation.","section":"Fig. 3(e), Discussion"}],"minor_comments":[{"comment":"The phrase 'a more effective implementations' is ungrammatical. In addition, the abstract's statement that the scaling results 'suggest that classical simulation in lossy scenarios remains computationally intractable' overstates what the numerics can show without the missing complexity analysis.","section":"Abstract"},{"comment":"The displayed formula for p({n_k}) is typeset incorrectly; it should read p({n_k}) = haf(W̃({n_k})) / (sqrt(det(I+G)) ∏_k n_k!). As written, the square root and product are misplaced, which obscures the meaning.","section":"Eq. (3)"},{"comment":"Refs. [22], [41], and [42] are cited only in Appendix D but are directly relevant to the main complexity discussion; consider moving them into the main text. Also, in the Computational Complexity section, 'in the complexity class #P' should be '#P-complete' for permanent and Hafnian functions.","section":"References"},{"comment":"Please state that the logarithm in F(x) is natural and that E_N is the Gaussian-state logarithmic negativity obtained from the symplectic eigenvalues of the partially transposed covariance matrix. This is standard but should be made explicit for the non-specialist reader.","section":"Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"This is an interesting proposal with encouraging numerics, but the main claim currently outruns the evidence. The authors should be encouraged to either add direct bond-dimension simulations or frame the sampling-hardness claim properly within an approximate-sampling framework. I do not see a circularity problem: the model parameters are scanned inputs, and the Hafnian form follows from the Gaussian-state formalism. The manuscript could become acceptable with the suggested revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: there's a real idea here — replacing the passive SU(2) interferometer in GBS with an active SU(1,1) OPA network so squeezing is distributed through the circuit. The loss-noncommutation observation in Appendix C is solid and does distinguish this from simply lumping loss at the input. The numerics are clean, and the paper is honest that the lossless limit is Bloch-Messiah equivalent to standard GBS.\n\nThe soft spot is exactly where the reader and the stress test put it. The central claim that the scheme \"preserves computational hardness under realistic loss\" rests on E_N scaling linearly with mode number. That's a necessary-condition argument, not a sufficient one. The cited MPS/MPO simulability theorems (refs 18-20) prove sublinear entanglement entropy is enough for efficient simulation; they don't prove linear logarithmic negativity forces exponential bond dimension. E_N is a mixed-state entanglement monotone; it can be extensive (linear in n) for a product of independent two-mode squeezed states with local loss, a distribution you can sample from with O(1) bond dimension. So the bridge from Fig. 3(e) to exponential classical cost is exactly the missing argument.\n\nThe complexity section also skips from exact Hafnian #P-hardness to sampling hardness. For a noisy device, approximate sampling is the relevant task, and that needs average-case conjectures or a noise-robustness proof. The paper doesn't provide it. These gaps are addressable: compute the actual MPO bond dimension for the OPA network, and either cite a proper approximate-sampling hardness result or soften the claim to \"no known efficient simulation.\"\n\nWhat's good: the architecture is concrete, the covariance/Hafnian formalism is correct, and the loss-noncommutation calculation is a useful observation on its own. The uniform-parameter numerics are limited (n up to 10, t >= 0.4) but that's not the main issue. The paper's own abstract hedges with \"suggest,\" but the Discussion and conclusion overstate the case.\n\nBottom line: this deserves a serious referee, but as written the load-bearing claim isn't established. I'd send it to review with a request to fix or downgrade the hardness claim. If I were writing on loss-robust GBS variants, I'd cite this for the architecture and the noncommutation result.","headline":"A genuinely new architecture for loss-robust GBS, but the hardness-under-loss claim is not proven because linear E_N does not imply exponential MPS/MPO cost.","tokens_in":17589,"tokens_out":2767,"would_cite":true,"duration_ms":29001,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Replacing passive interferometers with optical parametric amplifier networks keeps Gaussian boson sampling classically hard even when photon loss is present, because the output entanglement continues to scale linearly with mode number.","keywords":["Gaussian boson sampling","optical parametric amplifier","SU(1,1) interferometer","logarithmic negativity","photon loss","computational hardness","Hafnian","quantum advantage"],"falsifier":"A direct calculation of logarithmic negativity for more than 10 modes (say n=20 or 40) at transmittance values around 0.4–0.6 that shows sublinear growth with mode number would refute the paper's central scaling claim; alternatively, demonstrating an MPO/MPS classical algorithm that approximates the OPA-network output distribution to constant error in polynomial time for these parameters would refute the hardness conclusion even if the scaling holds.","tokens_in":16590,"feed_emoji":"⚛️","tokens_out":3673,"duration_ms":34006,"temperature":0.7,"pith_summary":"The paper argues that inserting optical parametric amplifiers (OPAs) into the interferometer of Gaussian boson sampling creates an active network that replenishes entanglement lost to photon loss. The authors show numerically that the output state's logarithmic negativity grows linearly with the number of modes even at realistic transmittance, whereas in a passive interferometer loss makes entanglement grow sublinearly and enables classical simulation. Since linear entanglement growth is their criterion for exponential cost of tensor-network classical simulation, they conclude this 'OPA-boosted GBS' preserves #P-hard sampling complexity in lossy conditions. If correct, this would give a practical route to quantum computational advantage that tolerates optical loss.","feed_headline":"OPA-boosted boson sampling stays hard under photon loss","feed_subtitle":"Adding optical parametric amplifiers makes output entanglement scale linearly with mode count even in lossy interferometers, blocking classi","key_machinery":"The central object is the SU(1,1) network: a multilayer array of optical parametric amplifiers described by two-mode squeezing operations, interleaved with beam-splitter layers that model photon loss. The symplectic covariance-matrix formalism tracks the Gaussian state through the network; the output photon-number distribution is given by the Hafnian master theorem, keeping the #P-hard complexity of GBS. The entanglement measure is logarithmic negativity EN, computed from symplectic eigenvalues of the partially transposed covariance matrix. The argument's load-bearing step is the numerical finding that EN grows linearly with mode number n under loss, which the paper links to the known result","core_discovery":"In standard GBS a lossy linear interferometer is equivalent to a lossless one with weaker squeezing, so loss reduces entanglement and can make the output classically simulable. The paper proposes replacing beam splitters with two-mode squeezers (OPAs) arranged in layers, so the network itself injects nonclassicality. The output probabilities still have Hafnian form, hence the sampling problem remains #P-hard. Using logarithmic negativity computed from the output covariance matrix, the paper reports that with vacuum inputs the entanglement grows linearly with both squeezing parameter and network depth in the lossless case, and—crucially—that under photon loss with transmittance down to 0.6 (a","pith_inferences":["The paper's complexity argument assumes that linear scaling of logarithmic negativity (a mixed-state entanglement measure) with mode number forces exponential bond dimension for MPS/MPO simulation. This is an extrapolation from the converse result about sublinear entanglement; a direct simulation algorithm for this specific state would settle it.","The numerical evidence covers n up to 10 and transmittance down to 0.4; experimental realizations will have inhomogeneous loss and phase errors, which may alter the scaling. Testing the scheme with larger n or with local loss imbalance would be a natural next step.","If the hardness claim holds, the same OPA-network idea could be applied to other Gaussian sampling tasks, e.g., threshold detection or scattershot settings, to restore loss-robustness.","The paper does not provide a full approximate-sampling hardness proof (e.g., under noise models with constant error), so the practical complexity gap remains open; a rigorous reduction would strengthen the claim."],"forward_implications":["If correct, the scheme offers a loss-tolerant photonic platform for quantum computational advantage, avoiding the need for extremely low-loss interferometers.","The equivalence between lossy passive GBS and weaker-squeezing lossless GBS is broken; loss cannot be lumped into a single transmissivity, so input squeezing cannot be tuned to mimic the effect.","The OPA network can be pumped synchronously and is experimentally feasible with current pulsed-laser and frequency-decoupling techniques.","Since the input need not be squeezed vacuum, the noncompact SU(1,1) structure may allow even harder sampling problems than standard GBS.","The linear entanglement scaling indicates that tensor-network algorithms that efficiently simulate current GBS experiments under loss would fail for this scheme."],"fun_headline_variants":["OPA network makes boson sampling loss-proof","Amplified boson sampling stays hard despite photon loss","Parametric amplifiers rescue boson sampling from loss","OPA-boosted GBS keeps quantum advantage under loss"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that the scheme is classically intractable under loss rests on the numerical observation—for at most 10 modes and transmittance at least 0.4—that logarithmic negativity grows linearly with mode number, and on the assumption that this linear growth of a mixed-state entanglement measure forces exponential cost for tensor-network classical simulation.","fun_headline_variants_meta":{"raw":{"variants":["OPA network makes boson sampling loss-proof","Amplified boson sampling stays hard despite photon loss","Parametric amplifiers rescue boson sampling from loss","OPA-boosted GBS keeps quantum advantage under loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000438,"raw_usage":{"total_tokens":2029,"prompt_tokens":675,"completion_tokens":1354,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":419,"completion_tokens_details":{"reasoning_tokens":1290}},"tokens_in":419,"tokens_out":1354,"duration_ms":10497,"temperature":1.0,"reasoning_tokens":1290,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:21:39.486004+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct calculation of logarithmic negativity for more than 10 modes (say n=20 or 40) at transmittance values around 0.4–0.6 that shows sublinear growth with mode number would refute the paper's central scaling claim; alternatively, demonstrating an MPO/MPS classical algorithm that approximates the OPA-network output distribution to constant error in polynomial time for these parameters would refute the hardness conclusion even if the scaling holds.","supporting_citations":[],"review_version":1}