{"id":"6af24b9a-b301-4382-8d24-e9324e025deb","arxiv_id":"2607.28435","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SFFT tensor-train Carr–Madan pricing matches FFT accuracy on Black–Scholes while using far less memory at large grids, and is compared directly to QFT on simulators and IBM hardware.","lead":"The paper prices European calls with Carr–Madan by running the Fourier step as a compressed tensor-train operator (SFFT) instead of a dense FFT, cutting memory at fine grids. It also times the same transform on IBM hardware and simulators so classical tensor networks and QFT can be compared side by side.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"Rank saturation (hence memory/runtime gains) is shown only for 1D Black–Scholes; the asymptotic claim rests on an unproven transfer to the jump/SV models the paper motivates.","rationale":"The reader’s strongest claim is correctly scoped to the BS benchmark numbers (memory plateau, core crossover ~16, full ~23, NRMSE floors). Those numbers are internally consistent with the reported figures and do not appear to contain an arithmetic or algorithmic contradiction for 1D BS. The single point on which the stronger asymptotic / “large-scale finance” language stands or falls is precisely the reader’s weakest_assumption: bounded TT ranks under grid refinement. That condition is empirically supported only for BS; it is not demonstrated for the models the paper uses to justify relevance. Keeping CONDITIONAL is therefore right—accept the BS methods result, condition any broader efficiency claim on rank behavior in non-BS and multi-asset regimes. No tighter verdict is warranted: the BS evidence is real, code/hardware gaps are secondary, and there is no internal inconsistency in the 1D construction. Concrete VG/Heston rank curves would settle the issue cleanly.","tokens_in":18689,"tokens_out":662,"duration_ms":68930,"concrete_test":"Re-run the exact SFFT pipeline (same TT-cross / Zip-Up tolerances, α=2.5, moneyness window, n=12..24) on Variance-Gamma and Heston European calls with their closed CFs. Record max TT rank of ψ and of the post-SFFT train, peak core memory, and NRMSE vs a reference. If max rank grows materially with n or exceeds a few hundred by n=24, or if the core-time crossover vs FFT vanishes, the load-bearing rank-saturation assumption fails outside BS.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency claim—that SFFT Carr–Madan keeps high accuracy while memory plateaus and core cost scales as O(n r_SFFT² r²) rather than O(n 2^n)—requires the binary-tensorized Carr–Madan quantities (especially ψ from the characteristic function, and the TT after SFFT application) to have small, non-growing max TT rank as n increases. Figs. 2–4 establish this only for European calls under Black–Scholes (rank saturates ~120 after n≈12; NRMSE ~3.1×10^{-13}). The introduction and abstract explicitly motivate jumps, stochastic volatility, and high-dimensional/multi-asset settings (Heston, Merton, Variance Gamma; cf. Kastoryano–Pancotti), where the CF is less entire/analytic and the tensorized integrand need not stay low-rank under the same binary encoding. No rank plot, crossover, or NRMSE is given for any non-BS model. If ranks grow with n (or with model roughness) the memory plateau and the core-time crossover near n≈16 disappear, and the “subexponential / large-scale / high-dimensional” framing does not hold. The BS evidence is solid for what it is; it does not underwrite the broader claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper reformulates the Carr–Madan Fourier pricing method in Tensor Train format by applying the Superfast Fourier Transform (SFFT), a low-rank TT operator realization of the reduced Quantum Fourier Transform, so that the damped characteristic-function integrand, rank-one damping factors, and inverse DFT are handled without materializing the full 2^n grid. It places this classical TN algorithm alongside a QFT-based Carr–Madan pipeline (following Ewen) and reports memory, runtime, and NRMSE for European calls under Black–Scholes, plus inverse-QFT execution statistics on IBM hardware and accuracy on statevector simulation. For the BS benchmark the TT ranks saturate (max rank ~120 after n≈12), memory plateaus relative to dense FFT, core SFFT runtime crosses FFT near n≈16 (full workflow near n≈23), and NRMSE reaches ~3.1×10^{-13} versus FFT ~4.4×10^{-15}; both SFFT and QFT avoid explicit exponential storage of the Fourier operator.","tokens_in":19060,"tokens_out":1531,"duration_ms":31762,"significance":"If the efficiency story transfers beyond the reported setting, the work would usefully unify classical FFT pricing, tensor-network compression of the Fourier operator itself (rather than only of pricing surfaces), and QFT-based quantum pricing under one Carr–Madan discretization. Strengths that should be credited: a clear, reproducible pipeline (TT-cross for ψ, analytic rank-one dampings, Zip-Up SFFT-TTO); an external analytical BS benchmark rather than self-consistency checks; concrete hardware circuit-depth/CZ/timing data; and an explicit core-vs-preprocessing runtime split that makes the asymptotic claim falsifiable. The main significance is therefore methodological and comparative for large 1D Fourier grids; claims of a scalable route to jump/SV or multi-asset pricing remain aspirational until supported by evidence.","major_comments":[{"comment":"Abstract, Introduction, and Conclusion frame the method as addressing large-scale / high-dimensional pricing and motivate jumps, stochastic volatility, and multi-asset settings (Heston, Merton, Variance Gamma; cf. Kastoryano–Pancotti). All numerical evidence (Figs. 2–4, NRMSE, rank saturation ~120) is for non-dividend 1D Black–Scholes only. The load-bearing efficiency claim—memory plateau and O(n r_SFFT² r²) core cost—requires that the binary-tensorized ψ and post-SFFT TT keep small, non-growing max rank as n grows. That is demonstrated for BS but not for any CF the introduction cites as motivation. Either add at least one non-BS CF experiment (e.g. Heston or VG) with rank/NRMSE/runtime plots, or substantially narrow the abstract/conclusion language to what is shown (fine 1D Fourier grids under BS).","section":"Abstract; Introduction; Experimental Evaluation; Conclusion"},{"comment":"The QFT vs SFFT/FFT comparison in “Comparison between Classical and Quantum Methods” and Fig. 9 measures QPU time as post-transpiled inverse-QFT execution only (state preparation, transpilation, queueing excluded), while full SFFT includes construction and FFT is end-to-end. The paper notes this, but the speed-up narrative and abstract claim of a “direct comparison” still risk overstating quantum advantage relative to classical kernels. Please report, side-by-side, (i) inverse-QFT-only vs SFFT-core and (ii) a full pricing workflow estimate including a concrete state-preparation cost model (or explicitly label Fig. 9 as kernel-only and remove workflow-level superiority language).","section":"QFT-based Option Pricing; Comparison between Classical and Quantum Methods; Fig. 9"},{"comment":"Eqs. (16)–(20): price recovery from computational-basis shots uses e^{iπl} ỹ_l ≈ |ỹ_l|=√p_l. This is stated as accurate when the discretization is good, but residual phase/imaginary error is a second error source beyond sampling. Classical FFT/SFFT reach ~10^{-15}–10^{-13} NRMSE while the statevector floor is higher (~10^{-12}–10^{-9}, Fig. 8). Quantify the phase-approximation error (e.g. max |Im(e^{iπl} ỹ_l)| or price error with exact complex amplitudes vs √p_l) on the same grids so readers can separate discretization, reconstruction, and shot noise—especially before any hardware pricing claims.","section":"QFT-based Option Pricing, Eqs. (16)–(20); Fig. 8"}],"minor_comments":[{"comment":"Several figure cross-references are broken or placeholder: “see Appendix??”, “Figure X”, and incomplete author/email fields (“tbf”). Fix before production.","section":"Experimental Evaluation; title page"},{"comment":"Damping is described as “varied over set to a value of α=2.5” (typo/grammar). State the fixed α used for classical runs and how the five α values in the simulator median NRMSE were chosen.","section":"Experimental Evaluation"},{"comment":"Grid definitions: Δk=N^{-1/2} and b=log(S0)-NΔk/2 should be written with clear math formatting; confirm consistency with Δv Δk=π/N from the discretization section.","section":"European Option Pricing…; Experimental Evaluation"},{"comment":"SFFT construction via Zip-Up is deferred; a brief complexity/rank bound or pointer to Hauck et al. / Dolgov et al. with the truncation tolerance used in experiments would aid reproducibility.","section":"Superfast Fourier Transform"},{"comment":"Fig. 2–4 axis labels appear as Unicode entity strings in the manuscript source; ensure vectorized figures render cleanly.","section":"Figures 2–4"},{"comment":"Related TN Fourier pricing (Glau et al.; Sakurai et al.; Kastoryano–Pancotti) is cited; a short explicit contrast—compressing the DFT operator vs compressing the pricing function/surface—would sharpen the novelty claim.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"Fit is reasonable for a quant-ph / computational-finance venue interested in TN–QC bridges, but the current draft’s “high-dimensional financial computations” framing exceeds the BS-only evidence. I would not reject on novelty grounds: the SFFT-Carr–Madan pipeline and the side-by-side FFT/SFFT/QFT study are publishable after scope correction and the comparison/phase-error clarifications. No integrity concerns. Moderate confidence: central 1D BS results look solid; transferability is the open risk."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real package here is an end-to-end Carr–Madan pipeline that never materializes the 2^n grid: TT-cross for ψ, analytic rank-1 dampings, Zip-Up SFFT-TTO, then the usual reconstruction, plus a side-by-side FFT vs SFFT vs inverse-QFT (statevector + IBM Boston) comparison. That joint benchmark is what is new; SFFT itself, QFT pricing, and TN finance are already on the shelf.\n\nWhat they do well is concrete and reproducible from the text. Memory plateaus after n≈12 while FFT keeps growing; max TT rank saturates near 120; core contraction crosses FFT around n≈16 and the full workflow around n≈23; NRMSE bottoms near 3×10^{-13} against closed-form BS (FFT ~4×10^{-15}). The unitary DFT/QFT conventions and the rank-1 exponential cores are clean. Hardware section is careful: they time the transpiled inverse-QFT only, report depth/CZ growth, and do not pretend noisy full prices. For anyone who already cares about fine 1D Fourier grids or classical–quantum Fourier comparisons, the figures are useful.\n\nSoft spots are real but bounded. All scaling and accuracy claims are Black–Scholes only. The introduction and abstract lean on jumps, Heston, Variance Gamma, and multi-asset motivation; none of those appear in the rank or timing plots. If the binary-tensorized CF does not stay low-rank, the O(n r^2) story and the “high-dimensional” framing weaken. That is the main caveat, not a hidden contradiction in the BS math. Minor: no code/data release, α and moneyness window are free parameters, and QPU numbers exclude prep and queue. None of that sinks the 1D result.\n\nWho it is for: computational-finance numerics people and quantum-inspired TN folks who want a worked Carr–Madan example with scaling curves. Not a pricing-theory paper and not yet a multi-asset method. I would send it to referees; the BS evidence is solid enough to deserve a proper read and a request for at least one non-BS CF experiment. Worth engaging if you work in this lane; cite the scaling plots, not the broader claim.","headline":"Solid end-to-end SFFT Carr–Madan with honest FFT/SFFT/QFT scaling numbers on 1D BS; the high-dimensional framing is ahead of the evidence.","tokens_in":19719,"tokens_out":569,"would_cite":true,"duration_ms":11873,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A compressed tensor-train form of the quantum Fourier transform prices European calls on huge grids without building the full FFT vectors.","keywords":["Option pricing","Carr–Madan method","Fourier methods","Tensor Train","Superfast Fourier Transform","Quantum Fourier Transform","Tensor networks","European call"],"falsifier":"Re-run the SFFT Carr–Madan pipeline on the same Black–Scholes parameters while forcing the TT rank of ψ or of the SFFT operator to grow with n (or switch to a model whose characteristic function is known to be high-rank under binary encoding); if memory and core runtime then track classical FFT instead of saturating, the claimed subexponential advantage fails.","tokens_in":19534,"feed_emoji":"📈","tokens_out":1028,"duration_ms":21461,"temperature":0.7,"pith_summary":"Classical Carr–Madan Fourier pricing needs a discrete Fourier transform over a strike–frequency grid. As that grid is refined to 2^n points, ordinary FFT memory and work grow exponentially and become the bottleneck. This paper rewrites the same pricing pipeline so every intermediate object lives in Tensor Train form and the transform itself is applied as the Superfast Fourier Transform—a low-rank Tensor Train Operator built from the reduced quantum Fourier circuit. The algorithm never materializes the huge vectors or the dense Fourier matrix. On a Black–Scholes European-call benchmark the method keeps pricing error near machine precision while memory saturates instead of exploding, and the core contraction eventually beats classical FFT. The same structure is run as a true inverse QFT on simulators and IBM hardware, giving a side-by-side classical-versus-quantum comparison of two ways to avoid exponential Fourier storage.","feed_headline":"Tensor-train QFT prices options without huge FFT grids","feed_subtitle":"Carr–Madan calls keep near-machine accuracy while memory saturates instead of exploding with grid size.","key_machinery":"The Superfast Fourier Transform (SFFT): a low-rank Tensor Train Operator for the reduced QFT circuit (without bit-reversal), applied by TT contraction of cost O(n r_SFFT² r²) to the tensorized Carr–Madan input, together with exact rank-one TT factors for the exponential damping and phase terms.","core_discovery":"Carr–Madan European call pricing can be executed entirely inside Tensor Trains by applying the Superfast Fourier Transform (a compressed Tensor Train Operator of the reduced QFT) to a TT representation of the damped characteristic-function data, so option prices on grids of size 2^n are obtained without ever storing O(2^n) vectors or operators, while accuracy stays comparable to classical FFT and both the TT method and hardware QFT avoid the exponential scaling of dense Fourier pricing.","pith_inferences":["If TT ranks stay modest for characteristic functions of Lévy or stochastic-volatility models, the same pipeline would undercut FFT memory for the settings where closed-form densities are unavailable and Fourier methods are most needed.","Precomputing a library of SFFT operators at fixed accuracy would remove the main offline bottleneck the authors report, making the core contraction the only online cost.","The phase-reconstruction step required by quantum measurement (taking square roots of probabilities) is a structural accuracy ceiling classical SFFT does not face; hybrid workflows may therefore keep the transform classical even when state preparation is quantum.","Binary tensorization of the frequency grid is doing the real compression work; other Fourier option methods (COS, PROJ) could be tensorized the same way if their kernels separate under the same encoding."],"forward_implications":["Large strike–frequency grids that are memory-infeasible for dense FFT become routine once the pricing tensors admit modest TT ranks.","The same SFFT operator can be built once offline and reused across many strikes, maturities, or parameter sweeps.","Classical tensor-network Fourier pricing and QFT-based quantum pricing become directly comparable inside one Carr–Madan discretization.","Hardware QFT wall-clock for the transform alone scales only polynomially in qubits, while classical FFT scales with the full grid size.","Future multi-asset or path-dependent Fourier pricers can target the same compressed-transform pattern rather than low-rank structure only in payoffs or Greeks."],"fun_headline_variants":["TT Superfast FFT prices Carr–Madan calls without O(2^n) grids","Tensor-train QFT keeps Carr–Madan accuracy as memory saturates","SFFT in Tensor Trains: option prices at 2^n without dense FFT storage","Carr–Madan via compressed QFT: TT and hardware both skip exp scaling","Classical TT vs QFT: large-grid option pricing without huge Fourier arrays"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method stays cheap only if the characteristic-function data and the intermediate pricing tensors keep small, saturating Tensor Train ranks as the grid is refined—shown here for one-dimensional Black–Scholes, not yet for the jump, stochastic-volatility, or multi-asset cases that motivate the work.","fun_headline_variants_meta":{"raw":{"variants":["TT Superfast FFT prices Carr–Madan calls without O(2^n) grids","Tensor-train QFT keeps Carr–Madan accuracy as memory saturates","SFFT in Tensor Trains: option prices at 2^n without dense FFT storage","Carr–Madan via compressed QFT: TT and hardware both skip exp scaling","Classical TT vs QFT: large-grid option pricing without huge Fourier arrays"]},"model":"grok-4.5","effort":"low","cost_usd":0.001931,"raw_usage":{"total_tokens":953,"prompt_tokens":838,"num_sources_used":0,"completion_tokens":95,"cost_in_usd_ticks":19308000,"prompt_tokens_details":{"text_tokens":838,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":20,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":838,"tokens_out":95,"duration_ms":3297,"temperature":1.0,"reasoning_tokens":20,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T07:24:49.775321+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the SFFT Carr–Madan pipeline on the same Black–Scholes parameters while forcing the TT rank of ψ or of the SFFT operator to grow with n (or switch to a model whose characteristic function is known to be high-rank under binary encoding); if memory and core runtime then track classical FFT instead of saturating, the claimed subexponential advantage fails.","supporting_citations":[],"review_version":1}