{"id":"1ec1fd51-9b10-4305-9b44-164616a9fe37","arxiv_id":"2611.16791","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A photonic chip implements quantum tunneling search for non-convex optimization on 1D and 2D landscapes, claiming speedups over classical SGD.","lead":"This paper presents a photonic chip that performs non-convex optimization by encoding the objective function into the refractive index of waveguides and using quantum tunneling to explore the landscape. The authors claim that their quantum tunneling search (QTS) algorithm finds local and global minima on small 1D and 2D test functions and demonstrates speedups over classical stochastic gradient descent.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The claim of quantum speedups over SGD is not supported by the data presented, as the authors themselves acknowledge in the supplementary material.","rationale":"The reader's verdict is CONDITIONAL, but the stress-test reveals a more fundamental issue: the paper's own supplementary material explicitly states that the data do not allow an end-to-end speedup estimate, and the comparison to SGD is not a fair resource comparison. The authors acknowledge that the horizontal axes are fundamentally different, the plotting factor of 60 is a display convention, and the SGD baseline is weak. The theoretical speedup assumes access to the local ground state, which is not rigorously achieved. The paper also acknowledges that the varying Hamiltonian approach does not cause relaxation to the ground state. Therefore, the central claim of quantum speedups is not supported by the evidence, and the paper should be rejected unless the authors provide a fair comparison and rigorous characterization of the input state. The reader correctly identified the weakest assumption, but the verdict should be stronger given the paper's own admissions.","tokens_in":18779,"tokens_out":566,"duration_ms":5132,"concrete_test":"Re-analyze the data without the rescaling factor of 60. Compare the number of objective function evaluations (or gradient calls) required by QTS (including fabrication, calibration, and measurement overhead) to achieve a given success probability on a specific landscape, against a state-of-the-art classical optimizer (e.g., CMA-ES, basin-hopping, or a well-tuned SGD with momentum) under a matched budget. If the classical optimizer matches or exceeds QTS performance, the speedup claim is unsupported.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim of quantum speedups over SGD (sentences b5, b23) is undermined by the paper's own statements. The supplementary material explicitly states that the horizontal axes are fundamentally different (propagation length vs. iterations), the plotting factor of 60 is a display convention and not an independently established comparison of resources, and the data presented do not allow an end-to-end speedup estimate (sentences b180, b181, b184). The SGD baseline is a simple gradient descent with noise, not a state-of-the-art classical optimizer. The theoretical speedup proven in Ref. [5] assumes access to the local ground state of the initial well, which the paper attempts to approximate via local annealing, but the quality of this approximation is not rigorously characterized. The paper also acknowledges that the varying Hamiltonian approach does not by itself cause relaxation to the instantaneous ground state (sentence b268). Therefore, the experimental results do not demonstrate quantum speedups, and the claim is not supported by the evidence provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper presents an experimental implementation of quantum tunneling search (QTS) on a photonic chip for non-convex optimization. The authors map objective functions to the potential term of a waveguide Hamiltonian, use local annealing to prepare input states, and employ a varying Hamiltonian approach to enhance search. They report experimental results on 1D and 2D landscapes, claiming that QTS successfully finds local and global minima and demonstrates quantum speedups over classical stochastic gradient descent (SGD).","tokens_in":19041,"tokens_out":1694,"duration_ms":11762,"significance":"If the claims were fully supported, this would be a significant experimental demonstration of a quantum algorithm for non-convex optimization, a classically hard problem. The photonic platform is interesting and the experimental techniques are non-trivial. However, the central claim of quantum speedups is not supported by the data presented, as the authors themselves acknowledge in the supplementary material. The comparison to SGD is not a fair resource comparison, the SGD baseline is not state-of-the-art, and the theoretical speedup assumptions are not fully met. The experimental results are consistent with the engineered Hamiltonian dynamics but do not isolate the contribution of quantum tunneling or demonstrate a practical advantage. The paper is well-structured and the experimental work is careful, but the claims are overstated relative to the evidence.","major_comments":[{"comment":"The claim of quantum speedups over SGD (sentences b5, b23) is not supported by the data. The supplementary material explicitly states that the horizontal axes are fundamentally different (propagation length vs. iterations), the plotting factor of 60 is a display convention and not an independently established comparison of resources, and the data presented do not allow an end-to-end speedup estimate (sentences b180, b181, b184). The SGD baseline is a simple gradient descent with noise, not a state-of-the-art classical optimizer. The theoretical speedup proven in Ref. [5] assumes access to the local ground state of the initial well, which the paper attempts to approximate via local annealing, but the quality of this approximation is not rigorously characterized. The paper also acknowledges that the varying Hamiltonian approach does not by itself cause relaxation to the instantaneous ground state (sentence b268). Therefore, the experimental results do not demonstrate quantum speedups, and the claim is not supported by the evidence provided.","section":"Abstract, §II (Multi-well random landscape), Supplementary Note 1 (Benchmark accounting and scope)","authors_acknowledge":"The data presented here do not allow such an end-to-end speedup estimate.","authors_acknowledge_p":0.91},{"comment":"The comparison to SGD is not a fair resource comparison. The SGD baseline uses a simple gradient descent with noise, not a state-of-the-art classical optimizer. The number of SGD steps is rescaled by a factor of 60 to bring its spatial spread close to that of the photonic search, but this factor is arbitrary and not based on any resource equivalence. The paper acknowledges that the data do not allow an end-to-end speedup estimate (sentence b184). Without a meaningful comparison, the claim of quantum speedups is unsubstantiated.","section":"§II (Multi-well random landscape), Supplementary Note 1 (Benchmark accounting and scope)","authors_acknowledge":"The plotting factor of 60 is a display convention and not an independently established comparison of resources.","authors_acknowledge_p":0.94},{"comment":"The experimental results are consistent with the engineered Hamiltonian dynamics, but the paper does not provide controls to isolate the contribution of quantum tunneling. The authors acknowledge that output intensity at a distant minimum is consistent with the engineered Hamiltonian dynamics but does not isolate the contribution of a programmed potential barrier (sentence b232). No additional barrier-control data are reported (sentence b234). Without such controls, the claim that the observed behavior is due to quantum tunneling rather than coherent coupling or other classical effects is not fully supported.","section":"§II (Global minimum search), Supplementary Note 2 (Evidence and controls for tunneling-inspired transport)","authors_acknowledge":"Consequently, output intensity at a distant minimum is consistent with the engineered Hamiltonian dynamics but, on its own, does not isolate the contribution of a programmed potential barrier.","authors_acknowledge_p":0.95},{"comment":"The theoretical speedup in Ref. [5] assumes access to the local ground state of the initial well. The paper uses an incomplete local annealing process to approximate this state, but the quality of this approximation is not rigorously characterized. The energy of the input state is reported as E=0.162, compared to the ground state energy E0=0.034 (Figure 3a), but the impact of this deviation on the theoretical speedup is not analyzed. The paper acknowledges that the input state is not the exact ground state (sentence b64), but does not quantify how this affects the query complexity or mixing time.","section":"§II (Local annealing), Supplementary Note 1 (Initial state significance)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The varying Hamiltonian approach is introduced to accelerate search and identify the global minimum, but the paper acknowledges that increasing the ratio in a lossless, coherently evolving array does not by itself cause relaxation to the instantaneous ground state (sentence b268). The mechanism by which the varying Hamiltonian leads to global minimum identification is not clearly explained or theoretically justified. The experimental results show that the probability at the global minimum increases, but without a clear theoretical understanding, it is difficult to assess whether this is due to the varying Hamiltonian or other factors.","section":"§II (Varying Hamiltonian), Supplementary Note 2 (Varying the ratio)","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The paper uses the term 'quantum speedups' but the experimental demonstration does not provide a meaningful comparison of resources. The authors should either remove this claim or provide a more careful discussion of what would be required to demonstrate a speedup.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The caption of Figure 4 states that the SGD curve shows the empirical distribution of positions after 1,600 iterations, but the text says the number of SGD steps is rescaled by a factor of 60. This is confusing and should be clarified.","section":"§II (Multi-well random landscape), Figure 4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The supplementary material provides a detailed and honest discussion of the limitations of the comparison to SGD. This is commendable, but it directly contradicts the claims made in the abstract and main text. The authors should reconcile these statements.","section":"Supplementary Note 1 (Benchmark accounting and scope)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper states that z=55 mm denotes the hitting 'time', but the assignment of a first detectable arrival length would require a specified intensity threshold and is limited by the spacing between sampled lengths (sentence b199). This should be acknowledged in the main text.","section":"§II (Double-well landscape)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper suggests a discriminating control experiment but does not perform it. This is a reasonable suggestion for future work, but the lack of such controls should be acknowledged as a limitation in the main text.","section":"Supplementary Note 2 (Evidence and controls for tunneling-inspired transport)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper states that the maximum available waveguide length is insufficient for convergence (sentence b73). This is an important limitation that should be discussed more prominently.","section":"§II (Multi-well landscapes)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper provides a realistic assessment of the fabrication overhead and scalability challenges. This is good, but it further undermines the claim of practical quantum speedups.","section":"Supplementary Note 2 (Fabrication overhead and scope of scalability)","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":null,"pith_extraction":{"msc":["81V80","68Q12","78A60"],"pacs":["42.82.Et","03.67.Ac","42.50.Ex"],"model":"deepseek/deepseek-v4-flash","headline":"The paper reports an experimental demonstration of quantum tunneling search (QTS) on a photonic chip, claiming it can find local and global minima in non-convex landscapes and demonstrate quantum speedups over classical stochastic…","keywords":["quantum tunneling search","non-convex optimization","photonic chip","waveguide array","quantum annealing","quantum speedup","analog quantum simulation","local minima"],"falsifier":"A head-to-head comparison on a suite of standard non-convex benchmark functions (e.g., Ackley, Rastrigin, Rosenbrock) between QTS on a photonic chip and a state-of-the-art classical optimizer (e.g., CMA-ES or Adam), with matched budgets of objective-function evaluations and wall-clock time, would either confirm or refute the claimed speedup. The paper's own data do not provide such a comparison.","tokens_in":21757,"feed_emoji":"🔦","tokens_out":1580,"duration_ms":19566,"temperature":0.7,"pith_summary":"The paper reports an experimental demonstration of a quantum optimization algorithm called quantum tunneling search (QTS) on a photonic chip. The authors encode a non-convex objective function into the refractive index profile of a waveguide array, then launch an optical state prepared by local annealing into the array. As the light propagates, quantum tunneling allows probability amplitude to pass through energy barriers that would trap classical gradient-descent methods. By averaging the output intensity over multiple propagation lengths, the system highlights local minima, and by gradually increasing the potential depth (the varying Hamiltonian technique) it can concentrate probability on the global minimum. The authors claim that this approach demonstrates quantum speedups over classical stochastic gradient descent, though they acknowledge that the comparison is limited because the horizontal axes (propagation length vs. iterations) are fundamentally different and the data do not permit an end-to-end speedup estimate. The work is presented as a proof-of-principle for near-term analog quantum simulators in non-convex optimization.","feed_headline":"Photonic chip finds minima by quantum tunneling","feed_subtitle":"A waveguide array encodes a non-convex landscape; light tunnels through barriers to reveal local and global minima.","key_machinery":"The central object is the quantum tunneling walk (QTW) implemented in a photonic waveguide array. The Hamiltonian is $H = \\sum_i \\beta_i a_i^\\dagger a_i + \\sum_{i \\neq j} C_{i,j} a_i^\\dagger a_j$, where the on-site propagation constants $\\beta_i$ are set proportional to the objective function values $f(x_i)$ at grid points $x_i$, and $C_{i,j}$ gives the nearest-neighbor coupling. Light injected into the array evolves according to $|\\Psi(z)\\rangle = e^{-iHz}|\\Psi(0)\\rangle$, and the probability $P_j(z)=|c_j(z)|^2$ at output length $z$ identifies candidate minima. The key identity is that quantum tunneling allows the walker to pass through energy barriers, unlike classical gradient descent. The local annealing step prepares a low-energy input state by slowly reducing the potential in one well, and the varying Hamiltonian technique increases the potential depth along $z$ to sharpen localization on the global minimum.","core_discovery":"The central claim is that a photonic chip implementing quantum tunneling walk (QTW) can solve non-convex optimization problems by using the quantum tunneling effect to traverse barriers that trap classical methods. The authors experimentally verify that the average probability distribution of light exiting a waveguide array, whose refractive index encodes the objective function, highlights local minima in one- and two-dimensional landscapes. By adding a varying Hamiltonian (gradually increasing the potential depth along the propagation direction), they further show that the output intensity can concentrate on the global minimum in a landscape with uneven local minima. The paper thus asserts that QTS is a successful and efficient non-convex optimization solver that demonstrates quantum speedups over classical stochastic gradient descent.","pith_inferences":["The paper's claim of quantum speedups rests on a comparison to a simple stochastic gradient descent baseline, not to state-of-the-art classical optimizers like Adam or quasi-Newton methods. A stronger test would compare QTS to classical algorithms that also leverage non-local exploration, such as simulated annealing or evolutionary strategies, on the same landscape.","The experimental setup uses classical laser light, not single photons or entangled states, so the observed dynamics are classical wave optics obeying the same Schrödinger-like equation. True quantum advantage would require demonstrating that multi-photon interference (e.g., indistinguishable photon pairs) yields a computational benefit beyond what classical coherent light can achieve.","The scaling of the waveguide approach is severely limited: an $n$-dimensional grid with $k$ points per dimension requires $k^n$ waveguides. Without a more compact encoding, the method cannot address high-dimensional problems typical of machine learning. The authors acknowledge this but do not propose a concrete path forward beyond mentioning additive structure or low effective dimension as possibl","A critical open question is whether the local annealing step can be performed efficiently for an arbitrary landscape without prior knowledge of which well contains the low-energy state. The paper assumes that a rough local minimum is known, which may not be available in general non-convex problems."],"forward_implications":["If QTS works as claimed, photonic analog simulators could provide a complementary route to classical digital methods for non-convex optimization, especially for problems with high, thin barriers where tunneling is exponentially faster than thermal activation.","The demonstration that a fixed photonic chip can solve a specific landscape suggests that reconfigurable photonic platforms could, in principle, be programmed to tackle a family of optimization problems without repeated fabrication.","The local annealing preparation method offers a practical way to approximate the ground state of a potential well, which is required for the theoretical quantum speedup proven in Ref. [5].","The varying Hamiltonian technique gives a concrete mechanism to improve search accuracy and identify the global minimum, potentially extending the reach of waveguide-based optimization beyond simple local-minima detection."],"supporting_citations":[{"why":"Provides the theoretical framework for quantum tunneling walk and proves exponential speedup over classical local-query algorithms for double-well landscapes, assuming access to the local ground state.","marker":"[5]"},{"why":"Introduces quantum Hamiltonian descent, which motivates the varying Hamiltonian approach used in the paper to accelerate search and identify the global minimum.","marker":"[7]"},{"why":"Supplies the coupling coefficient calibration method and the basic waveguide fabrication and characterization techniques used in the experiments.","marker":"[10]"}],"fun_headline_variants":["Quantum Tunneling Search Solves Non-Convex Problems on a Photonic Chip","Light-Based Chip Uses Tunneling to Find Global Minima","Photonic Tunneling Algorithm Outperforms Classical Optimization","Chip-Scale Quantum Tunneling Finds Hidden Minima Faster","Photonic Quantum Tunneling Search Beats Gradient Descent"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The claim of quantum speedups over classical SGD assumes that the comparison of propagation length to SGD iterations is meaningful and fair, but the paper itself states that the axes are fundamentally different, the factor of 60 is only a display convention, and the data do not allow an end-to-end speedup estimate.","fun_headline_variants_meta":{"raw":{"variants":["Quantum Tunneling Search Solves Non-Convex Problems on a Photonic Chip","Light-Based Chip Uses Tunneling to Find Global Minima","Photonic Tunneling Algorithm Outperforms Classical Optimization","Chip-Scale Quantum Tunneling Finds Hidden Minima Faster","Photonic Quantum Tunneling Search Beats Gradient Descent"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":8.6e-05,"raw_usage":{"cost":8.55582e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":8.55582e-05,"upstream_inference_prompt_cost":7.09982e-05,"upstream_inference_completions_cost":1.456e-05},"total_tokens":1065,"prompt_tokens":985,"completion_tokens":80,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":0}},"tokens_in":985,"tokens_out":80,"cost_source":"openrouter_usage","duration_ms":1750,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":0,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T12:44:26.217614+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"A head-to-head comparison on a suite of standard non-convex benchmark functions (e.g., Ackley, Rastrigin, Rosenbrock) between QTS on a photonic chip and a state-of-the-art classical optimizer (e.g., CMA-ES or Adam), with matched budgets of objective-function evaluations and wall-clock time, would either confirm or refute the claimed speedup. The paper's own data do not provide such a comparison.","supporting_citations":[{"cited_title":"On quantum speedups for nonconvex optimization via quantum tunneling walks, Quantum7, 1030","cited_arxiv_id":null,"evidence_quote":"Provides the theoretical framework for quantum tunneling walk and proves exponential speedup over classical local-query algorithms for double-well landscapes, assuming access to the local ground state."},{"cited_title":"Quantum Hamiltonian Descent","cited_arxiv_id":"2303.01471","evidence_quote":"Introduces quantum Hamiltonian descent, which motivates the varying Hamiltonian approach used in the paper to accelerate search and identify the global minimum."},{"cited_title":"Y.-H. Lu, J. Gao, Y.-J. Chang, R.-J. Ren, Z.- Q. Jiao, Z.-Y. Zhang, and X.-M. Jin, Topologically pro- tected polarization quantum entanglement on a photonic chip, Chip1, 100003","cited_arxiv_id":null,"evidence_quote":"Supplies the coupling coefficient calibration method and the basic waveguide fabrication and characterization techniques used in the experiments."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}