{"id":"b13bd433-fb67-4322-be39-f16b5e0af8b8","arxiv_id":"2411.11660","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TT-cross tensor-network state preparation encodes log-normal finance distributions into quantum circuits with linear depth scaling, but the supporting data and code are not public.","lead":"The authors show that a classical tensor-network algorithm (TT-cross) can encode financial probability distributions into quantum circuits with circuit depth that grows linearly with qubit count. They benchmark the approach on log-normal distributions used by Itaú Unibanco and report improved scalability over Qiskit's built-in state preparation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TT-cross accuracy degrades beyond ~15 qubits, so the claimed high accuracy and scalability for QMC does not hold at large qubit counts.","rationale":"The reader's weakest_assumption identified low-rankness of the target distributions as load-bearing. My concern is sharper and grounded in the paper's own Figure 4: even when low-rankness holds (small bond dimensions suffice), the KS error stops decreasing and then increases with qubit count. This is an acknowledged limitation that directly affects the 'accuracy' component of the central claim. It is more specific than a generic low-rankness worry, so I partially agree with the reader. I do not see this as requiring a verdict change because the paper is presented as an evaluation with conditional acceptance; the reader already called for addressing limitations and reproducibility. However, the condition should explicitly require the authors to either demonstrate that relevant accuracy metrics (e.g., L2 or total variation) converge with increasing qubits, or clearly state the optimal qubit range and temper the scalability claim. I also note a secondary internal inconsistency: Algorithm 1 contains suspicious lines ('R <- (ΣV∗) ⊗ I2' and 'W.insert(G[i])' instead of inserting U). This is not the main attack because the surrounding text describes the correct procedure, but it should be fixed for reproducibility.","tokens_in":10957,"tokens_out":8563,"duration_ms":84572,"concrete_test":"Reproduce Figure 4 for the univariate log-normal with the same parameters (bond dimensions 2, 4, 8, 16) and extend the horizontal axis to 30 qubits. Compute both the KS distance and the L2 norm error between the TT approximation and the target discrete distribution. Determine whether the KS distance continues to increase beyond 16 qubits while the L2 error decreases (indicating a metric artifact) or whether the actual approximation error (L2) also increases (indicating a true non-convergence). If the true error increases, the central accuracy claim fails; if the L2 error decreases, the KS metric alone is misleading and the authors should report L2 error as the primary accuracy metric.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of the paper is that the TT-cross method provides an accurate and scalable solution for loading probability distributions in QMC. The paper's own Figure 4 (left) shows that for a univariate log-normal distribution, the KS distance between the TT approximation and the target reaches a minimum of 7.2e-5 at 16 qubits and then increases for larger qubit numbers, even when the bond dimension is increased beyond 8. The text explicitly attributes this to accumulation of small errors across the domain. This means the approximation error is not a monotonically decreasing function of the discretization resolution; instead, there is an optimal qubit count beyond which the loaded state becomes less faithful. In a QMC pipeline, the state preparation error contributes directly to the bias of the estimated expectation value. If a 25-qubit discretization is needed to meet the QMC discretization error tolerance, the TT-cross state preparation error could exceed the target. Thus the method's 'scalability in qubits' does not translate into scalable accuracy, which is a load-bearing condition for the claimed advantage over Grover-Rudolph. The authors acknowledge the effect but do not propose a mitigation, leaving the central claim of high accuracy at scale unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to solve the probability-loading problem for Quantum Monte Carlo by first approximating a target distribution in tensor-train (TT) form with the TT-cross algorithm, then mapping the resulting TT cores to a quantum circuit using an SVD-based construction. The method is benchmarked on log-normal distributions calibrated to financial data from Itaú Unibanco, for univariate and multivariate cases up to five dimensions. The reported results compare the TT-cross circuit depth and training time with Qiskit's built-in Grover-Rudolph implementation, and include a 5-qubit hardware demonstration on IBM Eagle. The main claims are that the TT-cross approach achieves logarithmic training complexity in the number of discretization points, linear circuit-depth scaling in the number of qubits, and high accuracy for financial distributions.","tokens_in":11143,"tokens_out":6359,"duration_ms":67198,"significance":"If the scaling claims hold, the proposed encoding is a useful state-preparation subroutine for QMC, since it provides a classical, systematically improvable route from a probability distribution to a shallow quantum circuit. The core SVD-based mapping in Section 3 is mathematically standard and appears correct, and the comparison against Qiskit on realistic financial distributions is a practical strength. The use of real market-calibrated data and the hardware experiment are also positive features. However, the accuracy of the encoding saturates and then degrades with increasing qubit count, and the empirical evidence currently lacks statistical error bars and reproducibility details; as a result, the \"scalable accuracy\" part of the central claim is not yet established.","major_comments":[{"comment":"The paper's own results show that the KS distance for the univariate log-normal distribution reaches a minimum of 7.2e-5 at 16 qubits and then increases for larger qubit numbers, and that bond dimensions larger than 8 lead to convergence issues beyond about 15 qubits. This is load-bearing for the QMC motivation: in QMC, state-preparation error contributes directly to the bias of the estimated expectation value, so if a QMC discretization error target requires more than roughly 16 qubits, the TT-cross encoding error can grow instead of shrink. The manuscript acknowledges the effect but proposes no mitigation and no error bound, so the claim of high accuracy at scale is unsupported. Please provide an analysis of when the optimal-qubit regime is sufficient for QMC tolerances, or an error-aware rank/truncation strategy that prevents this degradation.","section":"Section 4, Figures 4, 8, 9, 10"},{"comment":"The KS and KL plots and the training-time plots report \"average\" values without error bars, confidence intervals, number of independent runs, or random seeds. Since TT-cross relies on heuristic index selection and the plots are the main evidence for the logarithmic and linear scaling claims, this missing statistical information makes it impossible to assess run-to-run variability or to determine whether the flat and then increasing KS region is robust. Please include error bars, the number of repetitions, and the exact experimental settings for each curve.","section":"Section 4, Figures 4, 8, 9, 10"},{"comment":"The numerical results are not reproducible as reported. The paper does not give the parameters (mu, sigma) of the log-normal distributions, the domain range, the discretization grid, the TT-cross termination criteria, the truncation tolerance, or the maximum number of sweeps; nor is code or detailed pseudocode for the TT-cross implementation provided. These details are needed to verify the central accuracy and scaling claims, because the observed behavior depends sensitively on these choices. Please add a table of distribution parameters and algorithmic settings, and make the implementation available or provide a complete algorithmic specification.","section":"Section 4"},{"comment":"The quantization ordering (sequential, mirroring, interleaving) is selected after observing the results on the same benchmark distributions: the text says mirroring \"led to better outcomes\" and interleaving \"yielded the best outcome.\" Because the same data are used to choose the ordering and to report the accuracy, the reported KL values are subject to selection bias, and the statement that interleaving is the most effective method is not validated on independent data. Please provide a principled criterion for choosing the ordering, or evaluate the method on held-out distribution families, to support the general claim.","section":"Section 2.2 and Section 4"},{"comment":"Although the abstract and conclusion frame the contribution as a pathway for Quantum Monte Carlo, no end-to-end QMC experiment is performed. The paper measures encoding fidelity and circuit depth, but does not estimate a financial expectation value with quantum amplitude estimation or another QMC scheme, so the impact of the encoding bias and circuit depth on the final QMC error is not quantified. A concrete QMC example (for example, an expected payoff under the encoded log-normal distribution) would substantiate the claim that the method provides an efficient pathway for QMC.","section":"Section 4 and Section 5"}],"minor_comments":[{"comment":"The encoding formulas for the multivariate cases write f(x_j, y_l) as amplitudes, whereas Eq. (2.2) uses sqrt(p(x_i)). Please clarify whether f denotes a probability value or an amplitude and adjust the notation consistently.","section":"Section 2.2, Eqs. (2.4), (2.6), (2.8)"},{"comment":"The pseudocode calls \"truncate(Sigma)\" without specifying the truncation criterion or tolerance. Please define the truncation rule, since it directly controls the bond dimension and the approximation error.","section":"Section 3, Algorithm 1"},{"comment":"The notation \"2 ^ n1\" is nonstandard and hard to parse; please write it as min(2, n1) or use explicit powers of two throughout.","section":"Section 3, Eq. (3.4) and surrounding text"},{"comment":"The hardware experiment is presented only visually. Please report quantitative fidelity measures (for example, KS or KL distance between the measured histogram and the target) for both optimization levels, rather than relying on a qualitative comparison.","section":"Section 4, Figures 6 and 7"},{"comment":"The conclusion says \"circuit and operation complexity scaling logarithmically with system size,\" but the results show circuit depth scaling linearly with qubit count and training time scaling logarithmically with the number of discretization points. Please clarify which quantity is meant by \"system size\" to avoid conflating qubits with grid points.","section":"Section 5"},{"comment":"There are several typographical errors, including \"dimesnion\" in the Figure 5 caption and \"Simrnov\" in the KS definition. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a useful incremental paper on a recognized bottleneck: loading probability distributions for quantum Monte Carlo. The genuinely new piece is the interleaving quantization ordering for multivariate distributions, and the paper does a fair job benchmarking it against sequential and mirroring encodings on Itaú-calibrated log-normal data. The SVD-based TT-to-circuit mapping is standard, and the reported scalings—linear circuit depth in qubits, logarithmic training time in discretization points—are consistent with what tensor-train methods should give. I also appreciate that the hardware section is honest about noise rather than overselling a 5-qubit demo.\n\nThe main weakness is load-bearing and the paper admits it without fixing it. Figure 4 shows KS error decreasing to 7.2e-5 at 16 qubits and then increasing for larger qubit counts, even with bond dimension 16. That means the encoding accuracy does not improve with finer discretization beyond a point; it gets worse. In a QMC pipeline, that state-preparation error contributes directly to bias. So the claim that the method gives accurate and scalable loading at large qubit counts is not supported. The authors mention error accumulation and convergence issues for bond dimensions above 8 beyond 15 qubits, but they provide no mitigation. I would want them to either bound this degradation or narrow the claim.\n\nThe other issues are more ordinary: no error bars on the KS/KL curves, no released code or data, and the choice of quantization ordering and bond-dimension cap appears to be selected post hoc on the same benchmark data. Those are addressable in revision. The citation pattern is reasonable; the relation to Sakaue et al. is stated fairly.\n\nWho gets value from this? Researchers working on quantum finance state preparation who want a concrete comparison point and a new ordering heuristic. It is not a breakthrough, but it is a serious benchmark. It deserves a real referee; the derivation is sound and the comparison is useful. I would send it out, with a request for reproducibility material and a direct response to the 16-qubit accuracy ceiling.","headline":"A solid incremental benchmark for TT-cross loading of financial distributions, with a genuinely new interleaving ordering, but the accuracy-at-scale claim breaks down around 16 qubits and the paper does not mitigate it.","tokens_in":11723,"tokens_out":2813,"would_cite":true,"duration_ms":27933,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68"],"pacs":["03.67.Ac"],"model":"deepseek-v4-flash","headline":"This paper claims that probability distributions can be loaded into quantum circuits with linear circuit depth using tensor-train cross approximation, avoiding the exponential scaling of standard state preparation.","keywords":["tensor networks","tensor-train cross approximation","quantum Monte Carlo","probability loading","state preparation","quantum finance","log-normal distribution","circuit depth"],"falsifier":"For a distribution with a known high tensor-train rank, such as a sharply multimodal or oscillatory density, measure the minimal bond dimension needed to reach a fixed Kolmogorov-Smirnov error as the number of qubits grows; if that bond dimension grows exponentially, the linear circuit-depth scaling and logarithmic training cost cease to hold for that input class.","tokens_in":10736,"feed_emoji":"📈","tokens_out":7876,"duration_ms":67246,"temperature":0.7,"pith_summary":"The paper tackles the probability-loading bottleneck of Quantum Monte Carlo: converting a classical distribution into a quantum state without a classically intractable precomputation. It argues that the tensor-train cross approximation (TT-cross) can do this efficiently for financial log-normal distributions, mapping the resulting tensor train into a quantum circuit with guaranteed convergence and scalable gate count. The central quantitative claim is that circuit depth grows linearly with the number of qubits and training time grows logarithmically with the number of discretization points, while the built-in Grover-Rudolph method grows exponentially. If correct, this gives near-term quantum hardware a practical path for QMC-based financial pricing and risk analysis.","feed_headline":"Tensor-train encoding makes quantum circuit depth scale linearly","feed_subtitle":"Loads financial distributions into qubits at linear depth, beating Grover-Rudolph.","key_machinery":"The tensor-train cross approximation (TT-cross) generalizes matrix cross approximation to high-order tensors by sampling nested index sets and building TT cores with small bond dimensions. A quantization scheme—sequential, mirroring, or interleaving—maps the discretized distribution onto a binary-indexed tensor, and each TT core is then SVD-factorized into a unitary acting on a growing set of qubits, giving a circuit whose depth can be further reduced by merging gates. The interleaving scheme is reported as the most effective for multivariate distributions.","core_discovery":"Using a tensor-train (matrix product state) representation of the discretized probability vector, the TT-cross algorithm finds a low-rank approximation whose cores are converted one by one into unitary gates via singular value decomposition, with a subsequent gate-merging step reducing depth. On univariate log-normal distributions calibrated to real market data the method reaches a Kolmogorov-Smirnov distance of $7.2\\times10^{-5}$ at 16 qubits, and on multivariate distributions up to five dimensions the Kullback-Leibler divergence stays near $10^{-2}$ while using up to $2^{50}$ discretization points. The paper's key numerical finding is the scaling gap: circuit depth is linear in qubits versus exponential for the built-in comparison method, and training complexity is logarithmic in the number of points versus linear.","pith_inferences":["The method's advantage depends on the target distribution having small TT-rank; distributions with long-range correlations or sharp oscillations would require larger bond dimensions, which the paper already observes leads to convergence issues beyond 15 qubits.","The same TT-cross pipeline could be applied to other QMC loading tasks—stochastic volatility, credit risk, or physics sampling—provided the quantized tensor is approximately low-rank, a condition the paper tests only for log-normal data.","Because the choice of quantization ordering is distribution-dependent, a data-driven heuristic for selecting sequential, mirroring, or interleaving would make the method robust beyond the reported cases.","An analytical error bound for the TT-cross truncation and the SVD mapping would let users predict the observed degradation past 15 qubits instead of discovering it empirically."],"forward_implications":["Quantum Monte Carlo on near-term hardware gains a data-loading routine for log-normal and similarly low-rank financial distributions with linear-depth circuits.","Larger discretization grids become reachable: up to $2^{50}$ points for five-dimensional distributions were encoded in simulation at training times that scale logarithmically.","The interleaving quantization order keeps relative entropy around $10^{-2}$ for up to five dimensions, indicating practical accuracy for multi-asset pricing.","A 5-qubit hardware run on a superconducting processor preserves the overall distribution shape, suggesting the encoded states are not immediately destroyed by device noise."],"supporting_citations":[{"why":"Supplies the TT-cross algorithm that constructs the low-rank tensor-train approximation of the distribution.","marker":"[12]"},{"why":"Provides the base method for encoding matrix product states into quantum circuits of one- and two-qubit gates, which the paper adapts.","marker":"[15]"},{"why":"Defines the Grover-Rudolph state-preparation baseline against which circuit depth and training complexity are compared.","marker":"[7]"},{"why":"Names the quantum processor used in the 5-qubit hardware test of the encoded distribution.","marker":"[9]"},{"why":"Independent related work on learning tensor trains from noisy functions, cited to position the paper's specific focus on distribution loading.","marker":"[17]"},{"why":"Gives the tensor-network formalism and notation the paper builds on.","marker":"[11]"}],"fun_headline_variants":["Linear-depth quantum Monte Carlo via tensor trains","TT-cross cuts circuit depth from exponential to linear","Tensor-train cross beats Grover-Rudolph depth scaling","Quantum finance loading at linear qubit depth","Tensor networks crack the probability loading bottleneck"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The target distribution must be well approximated by a tensor train with small bond dimensions (roughly 2 to 8); if it is not low-rank, the TT-cross method's accuracy and its linear-depth advantage collapse.","fun_headline_variants_meta":{"raw":{"variants":["Linear-depth quantum Monte Carlo via tensor trains","TT-cross cuts circuit depth from exponential to linear","Tensor-train cross beats Grover-Rudolph depth scaling","Quantum finance loading at linear qubit depth","Tensor networks crack the probability loading bottleneck"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000322,"raw_usage":{"total_tokens":1753,"prompt_tokens":833,"completion_tokens":920,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":449,"completion_tokens_details":{"reasoning_tokens":850}},"tokens_in":449,"tokens_out":920,"duration_ms":7751,"temperature":1.0,"reasoning_tokens":850,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:15:51.802354+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a distribution with a known high tensor-train rank, such as a sharply multimodal or oscillatory density, measure the minimal bond dimension needed to reach a fixed Kolmogorov-Smirnov error as the number of qubits grows; if that bond dimension grows exponentially, the linear circuit-depth scaling and logarithmic training cost cease to hold for that input class.","supporting_citations":[{"cited_title":"Ibm unveils breakthrough 127-qubit quantum processor, Novem- ber 2021","cited_arxiv_id":null,"evidence_quote":"Names the quantum processor used in the 5-qubit hardware test of the encoded distribution."},{"cited_title":"Oseledets and E","cited_arxiv_id":null,"evidence_quote":"Supplies the TT-cross algorithm that constructs the low-rank tensor-train approximation of the distribution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the base method for encoding matrix product states into quantum circuits of one- and two-qubit gates, which the paper adapts."},{"cited_title":"Grover and T","cited_arxiv_id":null,"evidence_quote":"Defines the Grover-Rudolph state-preparation baseline against which circuit depth and training complexity are compared."},{"cited_title":"Sakaue, H","cited_arxiv_id":null,"evidence_quote":"Independent related work on learning tensor trains from noisy functions, cited to position the paper's specific focus on distribution loading."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the tensor-network formalism and notation the paper builds on."}],"review_version":1}