{"id":"89ec5cfc-4ff0-41c3-8462-9d5521201f01","arxiv_id":"2507.08570","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"In a benchmark of variational quantum process tomography on 2-qubit unitaries, a classical one-hot optical processor and a quantum photonic processor both outperformed superconducting processors, reaching process fidelity about 0.8.","lead":"The authors benchmarked a variational quantum process tomography algorithm on an optical processor and compared it with IBM, QuTech, and Quandela quantum hardware. They report that optical processors, including a classical-light device, reach higher process fidelities than superconducting processors at circuit depths 3 and 6.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Process fidelity for the local optical processor is computed from intensity-only data with no specified phase recovery; the central cross-platform comparison may rest on invalid process matrices.","rationale":"The reader identified the phase-reconstruction issue as the weakest assumption, and the manuscript text supports that concern: Sec. VIE explicitly describes amplitude extraction only 'up to the sign,' and Sec. VIF describes passing 'probabilities' to the optimizer as the entries of the implemented matrix. For the cost function in Eq.1 and the process fidelity in Eq.3, complex phases are indispensable; if they are dropped or guessed, the reported numbers do not describe the actual physical process. This concern is more load-bearing than the acknowledged known-U assumption, because even granting that U is known, the fidelity metric used to compare platforms must be correctly computed. The paper could still be correct if a hidden Hadamard-test structure makes the ancilla probability directly yield the real part of the overlap, but no such circuit-level explanation is provided, and the 'up to the sign' language suggests amplitudes are used without a proper phase reference. The manuscript also gives credit for disclosing the joint-unitary limitation and the small statistics, but those are secondary to the phase issue. The appropriate verdict remains CONDITIONAL: release the data-processing details and raw data, and show that the reconstructed χ_actual matches a directly measurable process fidelity. I therefore do not change the reader's verdict, but I emphasize that the requested clarification is not optional for the central claim.","tokens_in":13109,"tokens_out":9301,"duration_ms":120048,"concrete_test":"Ask the authors to release the raw, noise-subtracted, normalized intensity vectors for the three Haar-random unitaries (both depths) together with the exact code that converts intensities to χ_actual. Independently recompute F_process = Tr(χ_actual χ_ideal) from these intensities, and also compute the true fidelity F_true of the implemented composite unitary (including a random diagonal D to model the inactive output phase shifters, as described in Sec. VIA). If |F_process − F_true| exceeds 0.05 on average, the intensity-only reconstruction is invalid and the reported outperformance is an artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that optical processors reach process fidelities of 0.8 and outperform superconducting processors by 0.10–0.22—rests on the fidelity values computed from Eq.3. For the local optical processor, Sec. VIE states that the measurement yields only a normalized intensity distribution, and that the square root of intensity corresponds to the probability amplitude 'up to the sign.' No procedure is given for determining the complex phases needed to construct χ_actual in Eq.3. If the optimizer receives the 'entries of the implemented matrix' as probabilities (Sec. VIF), then the cost function of Eq.1, which requires Re⟨Ψtr|Ψpr⟩, cannot be evaluated correctly, and the trace in Eq.3 is not the actual process fidelity. The paper neither specifies a Hadamard-test/ancilla phase-extraction step nor explains how off-diagonal phase information is recovered. Without this, the reported 0.71–0.80 fidelities could be artifacts of a phase-insensitive reconstruction (e.g., assuming all amplitudes positive), making the comparison to IBM Sherbrooke and QuTech Tuna-5 unsupported. Because the abstract's headline result is quantified entirely by these fidelity numbers, this is the most load-bearing assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript benchmarks a variational quantum process tomography algorithm on a local 8-mode classical optical processor using one-hot encoding with coherent light, and compares it against Quandela's Ascella photonic processor, IBM Sherbrooke, and QuTech Tuna-5 at circuit depths d=3 and d=6. The central quantitative claim is that the optical processors achieve process fidelities up to 0.8 and outperform the superconducting processors by 0.10–0.22. The paper also attributes the dominant noise in the local processor to thermal phase-shifter noise. The experimental setup is described in detail, but the fidelity extraction from intensity-only data is not specified.","tokens_in":13316,"tokens_out":9688,"duration_ms":119256,"significance":"If the reported fidelities were valid, this would provide a useful cross-platform benchmark for variational algorithms on photonic hardware and would support the practical value of one-hot encoding for small-scale variational circuits. The paper's strengths are its detailed description of the optical hardware, the one-hot encoding rationale, and the attempt to compare multiple publicly accessible platforms. However, the validity of the principal result hinges entirely on the process-fidelity extraction, which is not established; as it stands, the main quantitative conclusion is not supported.","major_comments":[{"comment":"The process fidelity values for the local optical processor cannot be computed from the data described. The measurement yields a normalized intensity distribution, i.e., the squared moduli |U_ij|^2, and Sec. VIE states that the square root of the intensity corresponds to the probability amplitude only \"up to the sign.\" Eq. (3) requires the complex process matrix chi_actual, whose off-diagonal phase information is not available from these measurements. Moreover, Sec. VI.A notes that the output phase row (the diagonal matrix D in the Clements decomposition) cannot be set, so the implemented unitary is defined only up to an unknown diagonal phase matrix; that D cancels in intensities but enters directly into Tr(chi_actual chi_ideal). No Hadamard-test, ancilla, or interferometric phase-extraction step is described. The optimizer is fed \"probabilities\" (Sec. VIF) rather than complex amplitudes, so the cost function of Eq. (1), which contains Re<Psi_tr|Psi_pr>, is not the true overlap. Consequently, the reported fidelities of 0.71–0.80 and the 0.10–0.22 gap against superconducting devices are unsupported by the described experimental procedure.","section":"Sec. VIE/VIF, Eq. (3)"},{"comment":"The experimental protocol assumes the target unitary U is known and uses it to construct the composite unitary. This makes the experiment a variational compilation or benchmarking task with a known target, not a process tomography experiment, in which U must be unknown. The title and abstract's claim of \"variational quantum process tomography\" is therefore overstated. The assumption is disclosed in Sec. VIF and in the Conclusions, but it changes the interpretation of the benchmark: the results do not demonstrate tomography of an unknown process, and the term \"tomography\" should be either justified or replaced with a more accurate description such as \"variational unitary compilation.\"","section":"Sec. VIF and Conclusions"},{"comment":"The statement that \"thermal noise in the phase-shifters dominates over other optical imperfections, such as mode mismatch and dark counts from single-photon sources\" is not supported by any measurement or model presented in the paper. The comparison between the local optical processor and Ascella does not isolate thermal phase-shifter noise from other classical or quantum error sources; no thermal noise characterization, such as phase-drift or heater-stability measurements, is reported. This attribution should be removed or substantiated with explicit experimental evidence.","section":"Sec. VII and abstract"},{"comment":"The cross-platform comparison is based on only three Haar-random unitaries for the fidelity and cost-function uncertainties and only five optimization iterations on the quantum hardware (ten on the local processor and the simulation). With three samples, the reported 0.10–0.22 fidelity gap may be within statistical uncertainty, and the abstract's claim about convergence behavior for superconducting devices is inferred from a truncated run. The paper should provide explicit error bars and justify that five iterations are sufficient for the convergence comparison, or restrict the convergence claims to the devices that were run for the full ten iterations.","section":"Sec. VII"}],"minor_comments":[{"comment":"Decimal commas (e.g., \"0,71\" and \"0,10\") are used inconsistently with the English text; decimal points should be used for consistency.","section":"Throughout"},{"comment":"The phrase \"up to relative phase differences on the modes\" should be stated as an explicit limitation on the implementable unitary set; this is directly relevant to the phase-recovery issue raised in the major comments.","section":"Sec. VI.A"},{"comment":"The phrase \"up to the sign\" is ambiguous for complex probability amplitudes; the missing information is a phase, not merely a sign, and the wording should be corrected accordingly.","section":"Sec. VI.E"},{"comment":"The combination of \"quantum-analogous\" with \"classical one-hot encoding\" is confusing; the local device operates with coherent light and is a classical stand-in, and this should be stated directly in the abstract.","section":"Abstract"},{"comment":"The phrase \"Haar measurement\" should be \"Haar measure.\"","section":"Sec. IV"},{"comment":"The sentence \"the local optical processor and the 5-qubit Tuna processor, which exhibit significantly shorter and longer processing times of up to 400 s and 4400 s, respectively\" is grammatically unclear and should be rewritten for readability.","section":"Sec. VII"},{"comment":"The caption should specify what the shaded regions represent (e.g., standard deviation or confidence interval) and how many samples are used for each data point.","section":"Fig. 3 caption"}],"recommendation":"reject","confidential_remarks":"The central quantitative claim rests on a phase-recovery gap that cannot be fixed by textual revision alone. To support the fidelity-based conclusions, the authors would need to either perform phase-sensitive measurements on the local optical processor, or explicitly reframe the paper as a benchmark of intensity-only variational unitary compilation and remove the process-fidelity comparisons. The hardware description and one-hot encoding analysis are valuable, but the headline result as stated is not supported by the reported experimental procedure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the useful part: this is the first benchmark that runs the same variational process tomography protocol on an actual classical-optical one-hot processor, a single-photon optical processor, and two superconducting platforms, with a clear metric set. The hardware description in Sec VI is careful, and the attribution of the local processor's noise to thermal phase-shifter noise rather than mode mismatch or dark counts is a genuine experimental finding.\n\nThe soft spot is the one the stress-test flags, and it is load-bearing. Sec VIE says the square root of normalized intensity gives the probability amplitude 'up to the sign.' That explains magnitudes, not phases. The process fidelity in Eq.3 needs the full complex process matrix; Sec VIF says the resulting probabilities are passed to the optimizer 'as the entries of the implemented matrix,' which doesn't reconcile with the cost function in Eq.1 requiring Re⟨Ψtr|Ψpr⟩. If phases are never recovered, the reported fidelities of 0.71–0.80 for the optical devices could be optimistic by an unspecified amount. Since the abstract's headline is exactly those numbers versus 0.42–0.30 for superconducting, this gap may be an artifact. The authors need to state explicitly how χ_actual in Eq.3 is reconstructed from intensity-only data, or the central comparison is unsupported.\n\nOther issues are minor by comparison: three Haar-random unitaries, five hardware iterations, no code/data release, and the local processor is classical so the phrase 'photonic processors are strong contenders' overreaches to the quantum case. Those are fixable or at least disclosed.\n\nBottom line: the paper is worth refereeing because the benchmark design and data have value, but it will need major revisions on the fidelity extraction before the results support the claimed conclusions.","headline":"Useful first cross-platform benchmark of variational quantum process tomography, but the optical fidelity numbers rest on an unexplained phase-recovery step that is load-bearing.","tokens_in":13834,"tokens_out":4503,"would_cite":false,"duration_ms":58900,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","81P50"],"pacs":["03.67.Lx","42.50.Ex"],"model":"deepseek-v4-flash","headline":"A cross-platform benchmark shows photonic processors reaching 0.8 process fidelity for variational quantum process tomography while superconducting processors plateau near 0.42.","keywords":["variational quantum process tomography","photonic processor","one-hot encoding","process fidelity","benchmarking","superconducting qubits","unitary decomposition","NISQ hardware"],"falsifier":"Recalculate the optical processors' process fidelity using phase-resolving measurements—for instance, interfering each output mode with a reference beam to recover complex amplitudes—for the same Haar-random unitaries and iteration counts. If the phase-sensitive fidelity for the local optical processor falls below the superconducting platforms' values (about 0.42 or lower at depth 6), the paper's central claim is refuted.","tokens_in":12872,"feed_emoji":"⚛️","tokens_out":6750,"duration_ms":67041,"temperature":0.7,"pith_summary":"The paper claims that for two-qubit variational quantum process tomography, an 8-mode integrated optical processor running with classical one-hot encoding achieves process fidelities up to 0.8 after nine iterations, matching a commercial 12-mode quantum photonic processor and clearly exceeding two superconducting processors tested on the same task. It presents this as the first direct benchmark of variational quantum process tomography across different hardware platforms, using cost function, process fidelity, and time per iteration at circuit depths 3 and 6. The authors argue that the classical-laser experiment is a valid stand-in for a quantum photonic experiment because, for one photon in one spatial mode, the classical scattering matrix of a linear interferometer coincides with its quantum transformation. If the benchmark is correct, it shifts practical near-term expectation toward photonic processors for hybrid variational algorithms.","feed_headline":"Photonic processors beat superconducting ones in 2-qubit tomography","feed_subtitle":"Benchmark reaches 0.8 process fidelity for photonic platforms; superconducting chips plateau near 0.42.","key_machinery":"The load-bearing objects are the one-hot encoding and the unitary decomposition that lets a single laser-powered mesh realize the whole variational circuit. In one-hot encoding, each logical $n$-qubit basis state is mapped to a single occupied mode among $2^n$ optical modes; for a single photon, the linear-optical scattering matrix equals the quantum transformation, so a coherent-state experiment can stand in for a single-photon one. The Clements decomposition factorizes the joint unitary into tunable beam-splitter and phase-shifter angles, and the optimizer uses a four-term shift rule on the cost function $C(\\hat\\theta) = \\frac{1}{2^{n-1}}\\sum_i \\left(1 - \\Re \\langle \\Psi^{tr}_i | \\Psi^{pr}_i \\rangle\\right)$ (Eq. 1). Process fidelity is then evaluated as $F_{\\text{process}} = \\operatorname{Tr}(\\chi_E^{\\text{actual}} \\chi_E^{\\text{ideal}})$ (Eq. 3) between the implemented and ideal process matrices.","core_discovery":"The central discovery reported is that variational quantum process tomography—fitting an unknown two-qubit process with a parametrized circuit—runs more accurately and with better convergence on photonic processors than on superconducting ones. Using a joint unitary that combines the target Haar-random process with the variational ansatz, the local optical processor compiles the circuit onto tunable beam splitters and phase shifters, measures only output intensities, and updates parameters via a four-parameter-shift rule. The reported process fidelity reaches 0.8 at depth 6 after nine iterations for both the local classical-light processor and the 12-mode quantum photonic processor, whereas the superconducting platforms start around 0.42 and do not improve beyond four iterations at depth 3. The paper also reports that increasing circuit depth helps the optical platforms but not the superconducting ones, and that thermal noise in the phase shifters, rather than single-photon effects such as dark counts or mode mismatch, dominates the optical platform's error budget.","pith_inferences":["An editor's inference: the comparison does not equalize calibration drift, transpilation, or error-mitigation choices, so the reported favorability to photonics could be platform-specific rather than a general photonics-vs-superconducting law.","Since the optical fidelity is derived from intensity-only data, a phase-sensitive variant of the same benchmark would be the natural stress test; if phases matter, the 0.8 fidelity may be an upper bound rather than the true process fidelity.","The protocol currently assumes the target unitary is known when constructing the joint unitary, so it is not yet blind tomography; implementing the target and the ansatz on two separate processors, as the paper suggests, would make the result a genuine unknown-process benchmark.","One could extend the benchmark to an ion or neutral-atom processor, which the paper notes were unavailable, to test whether the optical advantage holds against a third hardware family."],"forward_implications":["If the benchmark is correct, photonic processors—both classical one-hot and quantum—are currently the better NISQ platform for variational circuits that repeatedly execute shallow two-qubit unitaries.","The one-hot classical encoding reproduces the quantum photonic results within uncertainty, meaning laser-based setups can prototype photonic algorithms before single-photon sources are deployed.","Increasing VQC depth from 3 to 6 helps optical platforms but not superconducting ones, so the fidelity gap likely grows with circuit width and depth.","The same benchmarking framework and VQC loop is claimed to extend beyond process tomography to algorithms such as QAOA and VQE.","The local optical processor's processing time per iteration (up to about 400 seconds) is an order of magnitude shorter than the noisiest superconducting platform's (up to 4400 seconds), because intensity measurements directly yield amplitudes without sampling."],"supporting_citations":[{"why":"Supplies the variational quantum process tomography algorithm, the circuit architecture, and the definition of optimal depth d=6.","marker":"[19]"},{"why":"Provides the universal multiport interferometer decomposition used to compile the joint unitary onto the optical processor.","marker":"[26]"},{"why":"Establishes that linear optics can implement every quantum operation for one photon, justifying the one-hot classical stand-in.","marker":"[27]"},{"why":"Gives the criterion and constructive recipe for which quantum operations linear optics can realize, backing the one-hot encoding equivalence.","marker":"[28]"},{"why":"Documents the 12-mode quantum photonic processor used as the quantum-optical comparison platform.","marker":"[21]"},{"why":"Describes the 5-qubit superconducting processor used as one of the two superconducting comparison platforms.","marker":"[20]"},{"why":"Introduced variational quantum process tomography on an optical processor, the approach the benchmark extends.","marker":"[16]"},{"why":"Supplies the QR-decomposition method used to generate Haar-random target unitaries.","marker":"[23]"}],"fun_headline_variants":["Photonic processors win in variational tomography benchmark","Light-based chips hit 0.8 fidelity in process tomography","Optical quantum processors beat IBM, QuTech in tomography","Photonic platforms outperform superconducting in tomography","Variational tomography: photonic beats superconducting at depth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The optical-processor fidelities are computed from intensity-only measurements, so the central comparison assumes that the missing phase information either does not affect the fidelity metric or can be recovered from the known unitary design; if that fails, the reported photonic advantage is not measuring the true implemented process.","fun_headline_variants_meta":{"raw":{"variants":["Photonic processors win in variational tomography benchmark","Light-based chips hit 0.8 fidelity in process tomography","Optical quantum processors beat IBM, QuTech in tomography","Photonic platforms outperform superconducting in tomography","Variational tomography: photonic beats superconducting at depth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1406,"prompt_tokens":1026,"completion_tokens":380,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":306}},"tokens_in":642,"tokens_out":380,"duration_ms":4337,"temperature":1.0,"reasoning_tokens":306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:15:56.597088+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recalculate the optical processors' process fidelity using phase-resolving measurements—for instance, interfering each output mode with a reference beam to recover complex amplitudes—for the same Haar-random unitaries and iteration counts. If the phase-sensitive fidelity for the local optical processor falls below the superconducting platforms' values (about 0.42 or lower at depth 6), the paper's central claim is refuted.","supporting_citations":[{"cited_title":"That only means it implements the correct inten- sity distribution on the output when fed with light at the inputs","cited_arxiv_id":null,"evidence_quote":"Supplies the variational quantum process tomography algorithm, the circuit architecture, and the definition of optimal depth d=6."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the universal multiport interferometer decomposition used to compile the joint unitary onto the optical processor."},{"cited_title":"QR and LQ Decomposition Matrix Backpropagation Algorithms for Square, Wide, and Deep -- Real or Complex -- Matrices and Their Software Implementation","cited_arxiv_id":"2009.10071","evidence_quote":"Establishes that linear optics can implement every quantum operation for one photon, justifying the one-hot classical stand-in."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the criterion and constructive recipe for which quantum operations linear optics can realize, backing the one-hot encoding equivalence."},{"cited_title":"Carolan, M","cited_arxiv_id":null,"evidence_quote":"Documents the 12-mode quantum photonic processor used as the quantum-optical comparison platform."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced variational quantum process tomography on an optical processor, the approach the benchmark extends."},{"cited_title":"Galetsky, P","cited_arxiv_id":null,"evidence_quote":"Supplies the QR-decomposition method used to generate Haar-random target unitaries."}],"review_version":1}