{"id":"94ff7be0-989b-4677-a168-6559cd8e438f","arxiv_id":"2501.16323","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The stitching method reduces a high-dimensional real-time path integral to low-dimensional integrals and fast Fourier transforms, demonstrated on quantum mechanical barrier potentials.","lead":"This paper introduces a numerical method that evaluates real-time Feynman path integrals by splitting them into small pieces and gluing them back together with fast Fourier transforms. The method is fast, parallelizable, and works for potentials that behave like a quadratic at large distances, which could make real-time quantum dynamics and cosmology calculations much cheaper.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The residual-decay assumption fails for the stated 'quadratic-at-infinity' class: for a quadratic potential, the saddle point shifts by O(a^2 grad V/m), so delta I_n does not decay and the FFT stitch has uncontrolled truncation error.","rationale":"The reader's weakest assumption is the decay of the residuals delta I_n. I agree that this is the load-bearing point, but I want to sharpen it: the issue is not merely that decay estimates are missing; for a genuinely quadratic potential the ansatz \\bar J_n is not the leading asymptotic at all, because the saddle point of the combined kinetic-plus-potential exponent shifts by a relative amount a^2 H V / m. Since the phase difference grows like q_n^2, delta I_n is non-decaying, and the FFT step is invalid on any finite lattice. The paper's condition H V << 2 (Eq. 23) omits the factor a^2 / m and is therefore not the correct small parameter. This matters because the abstract and introduction advertise the method for potentials dominated by a quadratic at infinity, a class that includes V = (1/2) omega^2 q^2. The worked examples are subquadratic (Rosen-Morse, smooth step) or truncated to be subquadratic, and for those the residual may well decay; the Rosen-Morse comparison with the exact propagator is a genuine piece of independent support. But it does not test the stated class. A conditional verdict is appropriate: the core stitching identity is sound, the examples are encouraging, but the scope and error control need to be corrected. I would not reject outright, because the method is likely repairable by restricting the class to potentials with decaying Hessian and adding decay estimates, or by replacing \\bar J_n with the quadratic saddle-point form. The reader's CONDITIONAL verdict already captures this; my concern does not move it, so I mark UNCHANGED. Agreement is partial: we identify the same assumption, but I add a concrete failure mode and a missing dimensionless parameter.","tokens_in":16529,"tokens_out":13254,"duration_ms":127939,"concrete_test":"Run the stitching recursion (39) for V(q) = (1/2) omega^2 q^2 with m = 1, hbar = 0.25, T = 10, x0 = 0, N = 20, and omega = 1 on a uniform lattice of width A and fixed spacing h. Compare G_N(x1, 0, T) with the exact lattice-regularized propagator, which can be computed in closed form by Gaussian integration of the quadratic lattice action. Also plot |delta I_2(q2)| = |I_2(q2) - \\bar J_2(q2)| for |q2| out to several hundred. If |delta I_2| stays of order 1 while I_2 and \\bar J_2 are also of order 1, and the G_N error does not decrease as A grows, the FFT truncation is uncontrolled and the central claim fails for quadratic potentials.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The stitching identity (39) is algebraically exact for any choice of \\bar J_n, but the numerical step replaces the convolution of delta I_n by an FFT on a finite padded lattice. This is only controlled if delta I_n decays in |q_n|. Section IIIB asserts that this follows from H V << 2 at infinity (Eq. 23), but the condition is insufficient. For the natural member of the stated class V(q) = (1/2) omega^2 q^2, the saddle equation (20) gives q1 = (x0 + q2) / (2 - a^2 omega^2 / m), whereas \\bar J_2 (Eq. 24) is built around q1 = (x0 + q2) / 2. The phase difference between I_2 and \\bar J_2 is of order (a^2 omega^2 / m) q2^2, so delta I_2 does not decay; it is an oscillatory function of order one. Convolving it with the Gaussian kernel on a finite lattice therefore introduces an O(1) truncation error. The relevant dimensionless parameter is a^2 H V / m at the saddle, not H V << 2. For the Rosen-Morse and smooth-step examples, V' and V'' decay at infinity, so the residual may decay, but the abstract and introduction claim the broader class of potentials dominated by a quadratic at infinity, which includes the quadratic failure case. The exactness and robustness claim is therefore not established for the stated class.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 'stitching method' for evaluating lattice-regularized real-time Feynman path integrals. Starting from the iterative composition law I_{n+1}(q_{n+1}) = c ∫ I_n(q_n) exp(im(q_n-q_{n+1})^2/(2aℏ) - iaV(q_n)/ℏ) dq_n, the authors split each intermediate integral I_n into a smooth asymptotic part \\bar J_n and a residual δI_n. The asymptotic parts are evaluated as low-dimensional oscillatory integrals (numerically or with an eikonal approximation), while the residuals are propagated by FFT-based convolutions. The claim is that this converts a d(N-1)-dimensional integral into N-1 d-dimensional integrals plus 2(N-1) FFTs, with computational cost linear in N, for potentials dominated by a quadratic at infinity. The method is demonstrated on the Rosen-Morse barrier, a smooth step, and a truncated double-well potential, and compared with the exact spectral propagator for the Rosen-Morse case.","tokens_in":16923,"tokens_out":5246,"duration_ms":50346,"significance":"If the central residual-decay assumption is established, the method would be a genuinely useful tool: it replaces a high-dimensional oscillatory integral by a sequence of low-dimensional integrals and FFTs, it is trivially parallelizable, and the numerical experiments are benchmarked against an independent exact spectral propagator without fitted parameters. The algebraic decomposition leading to Eq. (39) is exact, and the examples correctly reproduce caustic structures and approach the continuum propagator as N increases. However, the advertised scope ('potentials dominated by a quadratic at infinity') is broader than what is actually proven or benchmarked, and the missing decay estimates are load-bearing for the claimed exactness.","major_comments":[{"comment":"The paper states that for potentials with Hessian H V ≪ 2 at infinity, I_2 approaches \\bar J_2 for large |q_2|, but no decay statement is proven. For a harmonic-oscillator potential V(q) = ½κq^2, which is clearly 'dominated by a quadratic at infinity' and has H V = κ < 2 for κ < 2, the saddle equation (20) gives q_1 = (x_0 + q_2)/(2 - a^2κ/m), whereas \\bar J_2 in Eq. (24) is centered at q_1 = (x_0 + q_2)/2. The phase difference between I_2 and \\bar J_2 is of order (a^2κ/m) q_2^2, so δI_2 = I_2 - \\bar J_2 is an oscillatory function that does not decay as |q_2| → ∞. The condition H V ≪ 2 is therefore insufficient; the relevant control parameter involves a^2 H V / m, and the residual is generically not small at infinity for a natural member of the stated class.","section":"III.B, Eqs. (20)-(24)"},{"comment":"The stitching identity (39) is algebraically exact for any choice of \\bar J_n, but the numerical implementation replaces the convolution of δI_n with an FFT on a finite zero-padded lattice. This is controlled only if δI_n decays sufficiently fast in |q_n|. No decay estimate or error bound is supplied anywhere in the paper; Section VI.A, step 3 merely says the required padding 'depends on the physical problem in question.' For the harmonic-oscillator example in the previous comment, the residual is of order one over the entire lattice, so the FFT truncation error is uncontrolled. Consequently, the advertised 'exact, robust' recovery of the lattice path integral is not established for the stated class of potentials.","section":"III.C, Eq. (39); VI.A"},{"comment":"The numerical evidence covers the Rosen-Morse potential, the smooth step, and a truncated double well. In the first two cases V' and V'' decay at infinity, so the residual may decay; in the third the potential is modified to be bounded. These examples do not test the claimed class of potentials 'dominated by a quadratic at infinity.' A benchmark against the exactly known real-time propagator of the harmonic oscillator would directly probe the residual-decay assumption and the finite-lattice FFT error. Without such a test, or a proof of decay for the stated class, the central claim of the abstract and introduction is supported only for a narrower class than is claimed.","section":"IV, VI.B"}],"minor_comments":[{"comment":"The term 'word-line quantization' should be 'world-line quantization'; the same typo appears in Section V.","section":"Abstract and Section I"},{"comment":"The object called the 'effective reduced Planck constant' does not have units of action: from Eq. (42), ℏ_eff = Tℏ/(mN) · (n-1)/n has units of length squared in d=1. Please clarify the intended scaling or rename this quantity to avoid dimensional confusion.","section":"III.D, Eq. (47)"},{"comment":"The introduction claims the method 'runs significantly faster than the generalized Lefschetz thimble method,' but no runtime comparison with that method is reported. Either provide a benchmark or temper the claim to the observed linear-in-N scaling.","section":"VI.C"},{"comment":"The choice θ = ½θ_max is stated without explanation; a sentence justifying why this angle balances suppression of oscillations with avoidance of the singularity hull would improve reproducibility.","section":"IV.A, Eq. (58)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the algebraic decomposition is a useful contribution. The main concern is the mismatch between the abstract's claim for 'potentials dominated by a quadratic at infinity' and the unproven residual-decay assumption that controls the FFT stitching. This is fixable by either proving decay estimates under the stated assumptions, narrowing the claimed class to potentials with derivatives decaying at infinity, or adding a harmonic-oscillator benchmark that shows the finite-lattice error is controlled. I see no circularity issue: the Rosen-Morse spectral comparison is independent, and the cited numerical tools are not fitted to the target propagator."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The stitching decomposition is the real thing: Eq. (39) is an exact identity, and replacing a d(N-1)-dimensional integral by N-1 low-dimensional integrals plus FFTs is a genuinely new approach to real-time lattice path integrals. The Rosen-Morse comparison against the exact spectral propagator is the right test, and it is persuasive, with no parameter fitted to the target. The self-citations are to independent numerical tools, not a circular web. On those points the paper earns its place.\n\nThe soft spots are not cosmetic. The residual-decay assumption is asserted in Section IIIB, not proved, and the stress-test counterexample lands. For V(q) = (1/2) m omega^2 q^2, the saddle for I_2 is at q_1 = (x0+q_2)/(2-a^2 omega^2), while \\bar J_2 is centered on the midpoint. For fixed a^2 omega^2, the difference in the q2^2 phase coefficient makes delta I_2 a non-decaying oscillatory function of order one. Convolving it on a finite padded lattice then has uncontrolled truncation error. So the 'exact, robust' claim is not established for the stated class of potentials 'dominated by a quadratic at infinity.' The class that actually works is the one the examples use: potentials whose first and second derivatives vanish at infinity, so the kinetic quadratic dominates at large distance. That should be stated precisely.\n\nThe other issues are smaller but real. The speed advantage over generalized Lefschetz thimbles is asserted with no benchmark. The J_n integrals are evaluated approximately (Gauss-Hermite or eikonal), so even with decaying residuals the pipeline is not exact in the strict sense, and no error bounds are given. No code or data are released, so the numerical claims are not independently checkable from the paper alone.\n\nWho is this for? People computing real-time propagators in 1D quantum mechanics and world-line quantization, and method developers in the sign-problem community. They will get a useful tool if the class statement is corrected and the decay/error analysis is added. As it stands, the central algorithm is sound in spirit, but the paper overclaims. It deserves a serious referee. I would send it to peer review with the request that the authors prove or numerically demonstrate residual decay for the claimed class, remove or qualify the exactness and speed claims, and release code.","headline":"A genuinely new stitching trick for real-time lattice path integrals, undermined by an unproven residual-decay assumption that fails for the literal quadratic-at-infinity class.","tokens_in":17385,"tokens_out":7524,"would_cite":true,"duration_ms":73247,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81S40","65T50","65D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The lattice-regularized real-time Feynman path integral can be rewritten as N-1 low-dimensional integrals stitched by fast Fourier transforms, making evaluation cost scale linearly with the number of time steps.","keywords":["real-time path integral","lattice regularization","fast Fourier transform","Picard-Lefschetz theory","eikonal approximation","oscillatory integrals","quantum propagator","world-line quantization"],"falsifier":"Take the Rosen-Morse barrier with its exactly known spectral propagator, compute the stitched discretized propagator at fixed $N$ while extending the lattice bounds and zero-padding, and compare to the spectral result; also directly measure $|\\delta I_n(q_n)|$ at large $|q_n|$. If the residual decays slowly enough that increasing the padding changes the result beyond roundoff, the claimed exactness fails for that potential.","tokens_in":16327,"feed_emoji":"⚛️","tokens_out":8083,"duration_ms":69701,"temperature":0.7,"pith_summary":"The paper claims that for potentials whose second-derivative matrix decays at infinity (potentials \"dominated by a quadratic at infinity\"), the lattice-regularized real-time Feynman path integral—normally a $d \\times (N-1)$-dimensional, highly oscillatory integral—can be rewritten exactly as a chain of $N-1$ low-dimensional integrals stitched together by $2(N-1)$ fast Fourier transforms. The reduction is exact in the sense that the only remaining numerical approximation is the evaluation of the low-dimensional integrals $J_n$; the stitching itself exactly reproduces the lattice path integral. This matters because real-time path integrals are notoriously expensive and numerically delicate, and the proposed scheme reduces the cost to linear in $N$ and parallelizes trivially. The paper demonstrates the method on exactly solvable and model potentials and ties the resulting interference patterns to the caustics of the corresponding classical theory.","feed_headline":"Real-time path integrals run in linear time via Fourier stitching","feed_subtitle":"This splits the high-dimensional oscillatory sum into N-1 small integrals joined by FFTs.","key_machinery":"The load-bearing objects are the asymptotic family $\\bar J_n$ (Eq. 35) and the stitching identity (Eq. 39). At each step the exact integral $I_n$ is split as $I_n = \\bar J_n + \\delta I_n$; the residual $\\delta I_n$ decays at infinity by construction, and the next integral is $I_{n+1} = J_{n+1} + \\mathcal{F}^{-1}[\\, \\mathcal{F}(\\delta I_n e^{-iaV/\\hbar})\\, e^{-ia\\hbar k^2/(2m)}\\,]$. The free kinetic kernel is a Gaussian whose Fourier transform is again a Gaussian, turning the convolution into two $d$-dimensional FFTs. The $J_n$ integrals carry all the potential dependence and are evaluated by contour rotation with Gauss-Hermite quadrature or by the eikonal approximation, whose accuracy improves as $N$ grows because the effective reduced Planck constant $\\hbar_{\\rm eff} = T\\hbar/(mN)\\,(n-1)/n$ shrinks.","core_discovery":"On the paper's own terms, the central claim is that the discretized propagator $G_N$ is recovered exactly by an iterative decomposition: starting from $I_1$, each step writes $I_{n+1}(q_{n+1})$ as a low-dimensional integral $J_{n+1}(q_{n+1})$ plus a convolution of the residual $\\delta I_n$ with the free kinetic Gaussian kernel. The low-dimensional integral is built from the asymptotic $\\bar J_n(q_n) = n^{-d/2} \\exp\\left(\\frac{i m}{2 a \\hbar} \\frac{(x_0-q_n)^2}{n} - \\frac{i a}{\\hbar} \\sum_{k=1}^{n-1} V(\\bar q_k^n)\\right)$, where $\\bar q_k^n$ is the linear interpolation between $x_0$ and $q_n$. Because the potential is dominated by a quadratic at infinity, $\\delta I_n$ decays for large $\\|q_n\\|$, so the convolution can be evaluated with two fast Fourier transforms on a padded lattice. The result is that the $d(N-1)$-dimensional path-integral evaluation reduces to $N-1$ independent $d$-dimensional integrals $J_2,\\dots,J_N$—evaluated by Picard-Lefschetz methods or by the eikonal approximation—plus $2(N-1)$ FFTs.","pith_inferences":["A strengthening of the paper's asymptotic assumption—for example, explicit decay rates for $\\delta I_n$—would turn the method's heuristic exactness into a rigorous numerical quadrature bound; without it, the required FFT padding length is an uncontrolled parameter.","The same stitching idea could be applied with a different leading-order ansatz, such as a WKB prefactor instead of the quadratic-phase Gaussian, for potentials that are not asymptotically quadratic, although the paper leaves that extension open.","Because the $J_n$ integrals are independent, the method's parallel structure suggests it could be embedded in larger studies that scan many initial positions, potentials, or propagation times at marginal extra cost.","For time-dependent potentials or nonlocal interactions the convolution may no longer be a pure Gaussian, but a generalized stitching based on the composition law might still reduce the effective dimension."],"forward_implications":["For a fixed initial position, one run of the stitching method produces the propagator $G_n(x_1,x_0,an)$ on a lattice of final positions and all intermediate times $n=1,\\dots,N$ at once.","The computational cost grows linearly with the number of time steps $N$ and is trivially parallelizable, so much finer discretizations become feasible than with a direct high-dimensional evaluation.","The method applies to potentials that are asymptotically quadratic, including Rosen-Morse barriers, smooth steps, and truncated double wells; the double-well propagator is handled by truncating or approximating the potential by Gaussians.","For the world-line quantization of relativistic particles, the stitching method applies to the inner path integral over Schwinger time, with the lapse integral treated by complex deformation.","Observables and unequal-time correlation functions follow immediately from the composition law once the propagator is known as a function of time and endpoints."],"supporting_citations":[{"why":"Supplies the Picard-Lefschetz numerical integration method used to evaluate the low-dimensional $J_n$ integrals.","marker":"[15]"},{"why":"Defines the generalized Lefschetz thimble method that the paper compares against as the main alternative for real-time path integrals.","marker":"[22]"},{"why":"The multi-plane lens integration technique that inspired the stitching construction and its FFT-based assembly.","marker":"[17]"},{"why":"Justifies the smooth regularization of conditionally convergent real-time path integrals on which the method's exactness relies.","marker":"[32]"},{"why":"Establishes the composition-law structure of the Feynman propagator that the stitching scheme mirrors.","marker":"[2]"},{"why":"Provides the spectral numerical evaluation of the Rosen-Morse propagator used as the exact reference for convergence tests.","marker":"[12]"}],"fun_headline_variants":["O(N) time real-time path integrals via FFT stitching","Low-dim integrals + FFTs: fast real-time path integrals","Picard-Lefschetz meets FFT for real-time path integrals","Exact real-time path integrals from low-dim pieces and FFTs","Efficient path integrals: split and stitch with FFTs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the asymptotic form $\\bar J_n$ captures the leading behavior of $I_n$ for large $|q_n|$, so each residual $\\delta I_n$ decays fast enough for the finite zero-padded Fourier transform to resolve the convolution; the paper states this for potentials with Hessian $H V \\ll 2$ at infinity but proves no decay rate or error bound.","fun_headline_variants_meta":{"raw":{"variants":["O(N) time real-time path integrals via FFT stitching","Low-dim integrals + FFTs: fast real-time path integrals","Picard-Lefschetz meets FFT for real-time path integrals","Exact real-time path integrals from low-dim pieces and FFTs","Efficient path integrals: split and stitch with FFTs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000456,"raw_usage":{"total_tokens":2307,"prompt_tokens":979,"completion_tokens":1328,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":1237}},"tokens_in":595,"tokens_out":1328,"duration_ms":11153,"temperature":1.0,"reasoning_tokens":1237,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:32:50.961576+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the Rosen-Morse barrier with its exactly known spectral propagator, compute the stitched discretized propagator at fixed $N$ while extending the lattice bounds and zero-padding, and compare to the spectral result; also directly measure $|\\delta I_n(q_n)|$ at large $|q_n|$. If the residual decays slowly enough that increasing the padding changes the result beyond roundoff, the claimed exactness fails for that potential.","supporting_citations":[{"cited_title":"Kleinert, Path integrals in quantum mechanics, statistics, polymer physics, and financial markets; 3rd ed","cited_arxiv_id":null,"evidence_quote":"The multi-plane lens integration technique that inspired the stitching construction and its FFT-based assembly."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the composition-law structure of the Feynman propagator that the stitching scheme mirrors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the spectral numerical evaluation of the Rosen-Morse propagator used as the exact reference for convergence tests."}],"review_version":1}