{"id":"eca01ae6-0307-478d-b211-365882509735","arxiv_id":"2505.05158","paper_version":1,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":1.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"This survey summarizes the potential of qudit-based quantum machine learning for expressivity and resource efficiency, along with current hardware demonstrations and remaining challenges.","lead":"This paper reviews how multi-level quantum systems (qudits) are being used in quantum machine learning, covering variational algorithms, quantum neural networks, and software tools. It argues that qudits can make quantum machine learning more compact and expressive, but notes significant hardware and software challenges.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Qudit QML advantage rests on the net-decoherence premise that the review cites but never re-derives for QML workloads.","rationale":"The reader's verdict is UNVERDICTED and the weakest assumption is exactly the net-decoherence claim. My independent reading reaches the same load-bearing concern: the review's central trade-off claim, that qudits reduce resources while maintaining competitive fidelity, depends on the unexamined premise that shorter qudit circuits compensate for noisier gates. That premise is cited rather than analyzed, and the supporting references are not QML-specific. Moreover, the paper itself flags in Section VI that no general advantage is established, so the strong conclusion in the abstract and conclusion is not fully supported by the evidence presented. I considered whether this should change the verdict to CONDITIONAL, but because the review is explicitly a survey and the reader already flagged the same weakness, the existing UNVERDICTED verdict with moderate confidence is appropriate. The concrete test I propose would settle the question by computing whether the quoted resource reductions actually translate into better end-to-end fidelity under realistic error rates for the very applications the review cites as its best evidence.","tokens_in":10617,"tokens_out":4098,"duration_ms":47321,"concrete_test":"Take the gate counts and architecture from the qutrit QAOA comparison [25] for graph 3-coloring: record the number of layers L and the number of two-body gates per layer for the qutrit encoding and for the optimized qubit encoding. Using the measured per-gate error rates reported for the relevant hardware (e.g., transmon qutrit versus qubit in [24]), compute the total success probability for both encodings with a simple error model, such as P_success proportional to exp(-N_gates * epsilon_gate). If the qutrit circuit's total success probability is not higher than or comparable to the qubit circuit's for those realistic error rates, the statement that 'net decoherence can decrease despite noisier gates' fails for the paper's own headline application.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of the review is that qudit architectures can reduce circuit depth and parameter counts while maintaining competitive fidelity. The load-bearing premise for the fidelity part appears in Section II: 'qudit circuits can be much shorter, the net decoherence over an algorithm can decrease despite noisier gates', supported only by citations [16]–[20]. This premise is not re-derived in the QML context, and the cited papers are mostly about gate-efficiency conditions, quantum steering, QEC, or magic-state distillation, not about QML workloads. The comparison is also not apples-to-apples: a single d-level qudit is contrasted with multiple qubits by logical Hilbert-space dimension, but the qudit has different physical error channels, leakage, control complexity, and calibration overhead. A reduction in depth or gate count does not by itself imply lower end-to-end error if the qudit per-gate error rate is higher. The one quantitative hardware example given, the Toffoli decomposition [24], uses 4 versus 8 two-transmon operations, but this is a single gate, not a QML circuit, and the reported fidelity gain is modest. The review's own Section VI admits that 'we still lack a comprehensive understanding of when and why a qudit-based QML model outperforms a qubit one' and that 'current evidence is often case-by-case'. That admission directly undercuts the stronger conclusion that qudits are 'a powerful tool for NISQ-era quantum machine learning'. The concern is not that the cited experiments are wrong; it is that the review extrapolates from resource counts to net fidelity without a quantitative error model, and then presents the extrapolation as a settled trade-off.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a short review of qudit-based quantum machine learning (QML). It argues that encoding information in d-level systems rather than qubits can benefit variational quantum algorithms and quantum neural networks by enlarging the local Hilbert space, enabling richer SU(d) feature maps, reducing circuit depth and parameter counts, and mapping multi-valued problems more naturally. The review surveys representative applications (transmon qudit VQE, qutrit QAOA for graph coloring, qudit classifiers, data re-uploading), discusses the software ecosystem (Cirq, MQT Qudits, QuDiet, QuForge, QuTiP), and closes with hardware challenges, open theoretical questions, and a cautiously optimistic outlook.","tokens_in":10902,"tokens_out":4656,"duration_ms":50906,"significance":"The review is timely and potentially useful as a compact entry point to qudit QML. It collects a reasonable set of experimental and numerical demonstrations and gives a clear taxonomy of the proposed advantages and current limitations. Its main technical claim, that qudits offer a favorable trade-off for NISQ-era QML, is plausible but is not established quantitatively in the manuscript. The review's own Section VI honestly admits that comparisons are case-by-case and that no comprehensive understanding exists, which is a strength; however, the introduction and conclusion occasionally state the advantage in stronger terms than the assembled evidence supports. The strengths are the breadth of the survey, the concrete examples, and the explicit identification of open problems, while the main weakness is the lack of a quantitative or at least carefully scoped analysis of the net-decoherence premise that underpins the practical benefit.","major_comments":[{"comment":"The load-bearing premise that 'because qudit circuits can be much shorter, the net decoherence over an algorithm can decrease despite noisier gates [16]-[20]' is cited but never re-derived or quantitatively justified in a QML context. References [16]-[20] include work on gate-efficiency conditions, quantum steering, qudit error-correcting codes, and magic-state distillation, which are not demonstrations for QML workloads. A shorter circuit reduces the number of operations but the per-gate error rate of qudit operations is typically higher, so the net end-to-end fidelity depends on the ratio of depth reduction to per-gate error increase. I recommend either (a) deriving an explicit condition, e.g., total error scaling as 1 - (1 - epsilon_q)^(D_q) vs. 1 - (1 - epsilon_b)^(D_b), applied to the cited experimental parameters, or (b) explicitly reframing this claim as an open assumption rather than an established advantage.","section":"Section II"},{"comment":"The conclusion states that qudit-based classifiers 'even achieve tasks unattainable by equivalent qubit networks.' This is not supported by any specific evidence cited in the review, and it conflicts with Section VI, which admits 'we still lack a comprehensive understanding of when and why a qudit-based QML model outperforms a qubit one' and that 'current evidence is often case-by-case.' Unless the authors can point to a specific result with a proven separation (not just an empirical demonstration on a particular dataset), this sentence should be removed or replaced with a statement about empirical advantages in specific case studies. The same applies to the stronger phrasing in the abstract and introduction that qudits are 'a powerful tool for NISQ-era quantum machine learning'.","section":"Section VII"},{"comment":"The quantitative hardware example used to illustrate resource savings, the qudit Toffoli decomposition from [24], is a single gate and not a QML circuit. The reported fidelity improvement is described as 'modest,' and the comparison is between one qutrit and multiple qubits with different error channels and calibration overhead. This example does not by itself validate the central claim that qudit QML circuits reduce net decoherence. I recommend adding a quantitative comparison from one of the QML demonstrations discussed in the paper, or clearly labeling the Toffoli example as an illustration of gate-count reduction rather than as evidence for a QML fidelity advantage.","section":"Sections II and III"}],"minor_comments":[{"comment":"The sentence 'loped an 8-dimensional Gell-Mann feature map for a qutrit' appears garbled or truncated; it should be completed (e.g., 'developed') and the appropriate reference should be cited explicitly.","section":"Section IV"},{"comment":"References [7] and [44] are the same paper (Mandilara et al., 'Classification of data with a qudit, a geometric approach,' Quantum Machine Intelligence, vol. 6, no. 1, p. 17, 2024). Please merge the duplicate citation.","section":"References"},{"comment":"The first bullet is labeled 'Circ' but the text and reference [47] refer to 'Cirq'; please fix the typo.","section":"Section V"},{"comment":"The QuForge description in Section V and the barren-plateau discussion in Section VI ('As demonstrated in [54]') concern the authors' own work. Neither is circular, but for transparency the manuscript should explicitly identify these as self-citations, for instance by saying 'our recent work' or adding a footnote.","section":"Section V and Section VI"},{"comment":"The phrase 'high-level superconducting transmons' should likely be 'higher-level superconducting transmons' to refer to multi-level transmon states.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The review contains two self-citations (QuForge in Section V and the barren plateau result in Section VI). Neither is circular, but the QuForge paragraph is noticeably longer than the other software entries and the barren-plateau sentence does not identify it as the authors' own work; editors may wish to ask for explicit disclosure. The more serious issue is the mismatch between the cautious Section VI and the stronger claims in the introduction and conclusion; this is the main reason for my recommendation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis is a competent short review of qudit quantum machine learning, worth reading if you want a map of the area, but don't expect new analysis. The central claim—that qudits reduce depth and parameter counts while keeping fidelity—is presented more confidently than the evidence supports.\n\nWhat's actually new: nothing, as expected for a review. The value is in collecting recent experimental and software developments—transmon VQE, qutrit QAOA, single-qudit classifiers, MQT Qudits, QuDiet, QuForge—and organizing them by theme. The survey of the tool ecosystem is current and useful. The challenges section honestly covers control complexity, leakage, software immaturity, and the barren plateau problem. The authors also cite competently, including their own QuForge and BP papers, which is appropriate in context.\n\nThe soft spot is the net-decoherence premise in Section II. The sentence 'qudit circuits can be much shorter, the net decoherence over an algorithm can decrease despite noisier gates' is cited to five papers, none of which test QML workloads. The review does not provide an error model or a cross-platform comparison, so the fidelity advantage is a borrowed assumption, not a demonstrated result. The Toffoli example is single-gate, not a QML circuit. Worse, the conclusion says qudits can 'achieve tasks unattainable by equivalent qubit networks,' which the body's own statement—that evidence is case-by-case—undercuts. These are review overreach, but they're fixable. Minor typos, like 'loped' in Section IV, suggest a light proofread.\n\nThis is a review aimed at newcomers or researchers needing a quick reference on software and experiments. It deserves a serious referee: it's a competent survey that will be citable. The referee should ask the authors to bring the caveats from Section VI into the abstract and conclusion, and to either supply a concrete error model for the depth-vs-fidelity claim or moderate that claim.\n\nRecommendation: accept after minor revision, with the overreach removed. Don't desk reject.","headline":"A useful, competent survey of qudit QML that overstates its evidence; worth refereeing but the authors should temper the fidelity claims.","tokens_in":11396,"tokens_out":2530,"would_cite":true,"duration_ms":25123,"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":"Qudits—d-level quantum systems—can make quantum machine learning circuits shorter and more expressive than qubit circuits.","keywords":["qudits","quantum machine learning","variational quantum algorithms","quantum neural networks","QAOA","Gell-Mann feature map","qutrit","NISQ devices"],"falsifier":"Run one fixed benchmark, such as 3-class classification or graph 3-coloring, on calibrated hardware with an optimized qutrit circuit and an optimized qubit circuit; tally end-to-end success probability and accumulated error at matched solution quality. If the qubit version matches the qutrit version's accuracy at equal or lower end-to-end error in the regime where the circuits have comparable depth, the claim that qudit compression reduces net decoherence fails in that setting.","tokens_in":10432,"feed_emoji":"⚛️","tokens_out":11104,"duration_ms":104566,"temperature":0.7,"pith_summary":"This review makes the case that the qubit is not the only—or always the best—unit for quantum machine learning. It argues that qudits, quantum systems with d > 2 basis states, store more information per physical system and support a richer rotation algebra, so variational quantum algorithms and quantum neural networks built from them can be shallower, use fewer parameters, and still match or beat qubit models on classification and optimization tasks. The review supports the case with recent hardware demonstrations on superconducting transmons and with comparative numerical studies, and it maps the remaining control, noise, and software bottlenecks. The practical stake is concrete: if the compressed-circuit claim holds, noisy near-term quantum devices could run learning tasks that their qubit-only versions cannot reach at the same fidelity.","feed_headline":"Qudits can shrink quantum-ML circuits while keeping accuracy","feed_subtitle":"More states per quantum unit mean shallower circuits and fewer trainable parameters, a practical edge on today's noisy devices.","key_machinery":"The load-bearing object is the qudit register and its $\\mathrm{SU}(d)$ algebra. A d-level system holds $\\log_2 d$ bits and has a $(d^2-1)$-dimensional generalized Bloch sphere; the Gell-Mann feature map is a named qutrit encoding that places classical inputs on that eight-dimensional sphere. These structures supply the argument's mechanism: wider rotation space per physical unit grants expressivity without extra entangling gates, natural d-ary encodings remove binary overhead, and shorter circuits reduce accumulated decoherence even when individual gates are noisier.","core_discovery":"In the paper's own framing, the central claim is that qudits are a resource-efficient alternative to qubits for QML, not a niche exoticism. An n-qudit register spans $d^n$ states, a qutrit carries $\\log_2 3 \\approx 1.53$ bits, and $\\mathrm{SU}(d)$ rotations give $d^2-1$ independent directions, so a single qudit can draw nonlinear decision boundaries that would otherwise need entangling several qubits. The survey points to a four-level transmon emulating two qubits in a VQE within chemical accuracy, a qutrit QAOA for graph 3-coloring that uses half the two-qubit gates per layer, a ternary Toffoli decomposition using four transmon operations rather than eight, and qutrit quantum neural networks that reach target accuracy with fewer parameters. Its conclusion is that qudit architectures reduce circuit depth and parameter counts while keeping competitive fidelity, which is the property that matters on noisy devices.","pith_inferences":["The compression argument extends beyond learning: algorithms dominated by multi-controlled gates, such as arithmetic or search oracles, should inherit the same qudit savings, since the Toffoli example is not QML-specific.","A quantitative crossover threshold—how much noisier each qudit gate may be before the shorter-circuit benefit is wiped out—would turn the review's trade-off into a design rule; the review does not derive that number.","The single-qudit nonlinear classifier hints that high-dimensional local rotations may substitute for entanglement in some learning tasks, a testable architectural hypothesis that would reshape how QML models are compared.","As compilers mature, automatic qubit-qudit partitioning—compiling some subcircuits to d-level systems and leaving others binary—could be the most direct route to practical adoption, extending the review's note that hybrid approaches are coming."],"forward_implications":["On noisy near-term hardware, quantum machine learning tasks that are currently depth-limited could absorb qudits as a resource reduction: the same Hilbert-space work with fewer physical units and fewer entangling gates.","Problems with naturally multi-valued variables—three-class classification, graph 3-coloring, spin-1 simulation—can be mapped directly, eliminating binary encodings, penalty terms, and wasted basis states.","Training machinery must be adapted: parameter-shift gradient rules become more expensive for higher-dimensional gates, and barren plateaus are known to worsen with qudit dimension, so existing mitigation ideas need reworking.","Software infrastructure is moving from qubit-only to qubit-qudit simulation and compilation; differentiable qudit simulators make gradient-based training of these circuits practical.","The relevant hardware comparison becomes end-to-end fidelity rather than per-gate error rate, so today's per-gate fidelity gap between qudits and qubits does not, by itself, decide the question."],"supporting_citations":[{"why":"Comparative study of qubit versus qutrit variational quantum neural networks, reported to reach equal or higher accuracy with fewer parameters for angle encoding.","marker":"[5]"},{"why":"Introduces the Gell-Mann feature map, the qutrit encoding that puts classical inputs into an eight-dimensional rotation space and underpins the expressivity argument.","marker":"[6]"},{"why":"Shows a single qudit can solve nonlinear classification tasks without entangling gates, grounding the claim that high-dimensional local operations are expressive.","marker":"[7]"},{"why":"Provides the condition under which noisy qudit gates still improve fidelity because circuits shorten, the core of the decoherence trade-off.","marker":"[16]"},{"why":"Hardware demonstration that a ternary Toffoli decomposition on transmon qutrits uses four operations instead of eight, exemplifying circuit compression.","marker":"[24]"},{"why":"Qutrit QAOA for graph 3-coloring with half the two-qubit gates per layer and higher sampling probability than the qubit encoding.","marker":"[25]"},{"why":"Single four-level transmon emulating two qubits in a VQE within chemical accuracy, showing qudits can stand in for qubit registers on hardware.","marker":"[36]"},{"why":"Single-qudit data re-uploading scheme, showing repeated re-encoding can be done efficiently, a QML technique central to compact qudit classifiers.","marker":"[45]"}],"fun_headline_variants":["Qudits shrink quantum ML circuits while preserving accuracy","Fewer parameters, shallower circuits: qudit quantum ML","Qudit architectures reduce QML depth and parameter counts","Multi-level qudits offer efficient path for quantum ML"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The case rests on the assumption, adopted from earlier studies and not re-derived here, that a qudit circuit can be shortened so much that its total accumulated error is smaller than a qubit circuit's even though each qudit gate is individually noisier.","fun_headline_variants_meta":{"raw":{"variants":["Qudits shrink quantum ML circuits while preserving accuracy","Fewer parameters, shallower circuits: qudit quantum ML","Qudit architectures reduce QML depth and parameter counts","Multi-level qudits offer efficient path for quantum ML"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000416,"raw_usage":{"total_tokens":2119,"prompt_tokens":890,"completion_tokens":1229,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":1174}},"tokens_in":506,"tokens_out":1229,"duration_ms":9684,"temperature":1.0,"reasoning_tokens":1174,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:10:33.110436+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run one fixed benchmark, such as 3-class classification or graph 3-coloring, on calibrated hardware with an optimized qutrit circuit and an optimized qubit circuit; tally end-to-end success probability and accumulated error at matched solution quality. If the qubit version matches the qutrit version's accuracy at equal or lower end-to-end error in the regime where the circuits have comparable depth, the claim that qudit compression reduces net decoherence fails in that setting.","supporting_citations":[{"cited_title":"Unlocking the high dimensional’ potential: Comparative analysis of qubits and qutrits in variational quantum neural networks,","cited_arxiv_id":null,"evidence_quote":"Comparative study of qubit versus qutrit variational quantum neural networks, reported to reach equal or higher accuracy with fewer parameters for angle encoding."},{"cited_title":"The Gell-Mann feature map of qutrits and its applications in classification tasks","cited_arxiv_id":"2312.11150","evidence_quote":"Introduces the Gell-Mann feature map, the qutrit encoding that puts classical inputs into an eight-dimensional rotation space and underpins the expressivity argument."},{"cited_title":"Noisy qudit vs multiple qubits: conditions on gate efficiency for enhancing fidelity,","cited_arxiv_id":null,"evidence_quote":"Provides the condition under which noisy qudit gates still improve fidelity because circuits shorten, the core of the decoherence trade-off."},{"cited_title":"Implementing a Ternary Decomposition of the Toffoli Gate on Fixed-FrequencyTransmon Qutrits","cited_arxiv_id":"2109.00558","evidence_quote":"Hardware demonstration that a ternary Toffoli decomposition on transmon qutrits uses four operations instead of eight, exemplifying circuit compression."},{"cited_title":"Exploring the potential of qutrits for quantum optimization of graph coloring,","cited_arxiv_id":null,"evidence_quote":"Qutrit QAOA for graph 3-coloring with half the two-qubit gates per layer and higher sampling probability than the qubit encoding."},{"cited_title":"Emulating two qubits with a four-level transmon qudit for variational quantum algorithms,","cited_arxiv_id":null,"evidence_quote":"Single four-level transmon emulating two qubits in a VQE within chemical accuracy, showing qudits can stand in for qubit registers on hardware."},{"cited_title":"Data re-uploading with a single qudit,","cited_arxiv_id":null,"evidence_quote":"Single-qudit data re-uploading scheme, showing repeated re-encoding can be done efficiently, a QML technique central to compact qudit classifiers."}],"review_version":1}