{"id":"dd97c67e-3d22-4750-86b2-c6b9d42e8938","arxiv_id":"2501.12643","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A tensor cross interpolation solver evaluates weak-coupling expansion integrals for quantum impurity problems up to 40th order and reproduces DMFT Mott physics.","lead":"This paper applies the tensor cross interpolation algorithm to the weak-coupling expansion of quantum impurity problems, allowing high-order terms up to 40th order to be integrated efficiently. The solver matches exact solutions and quantum Monte Carlo in benchmarks, and inside dynamical mean-field theory it reproduces the metal-to-Mott-insulator transition.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The weak-coupling integrand's low-rank structure is inferred from numerical plateaus that are absent at the highest order tested (nmax=40, Fig. 2(d)); the resulting scalability limit is the main unproven pillar of the method.","rationale":"The paper's core algorithm is credible: the exactly solvable benchmark reaches ~1e-5 accuracy, the DMFT results at beta=16/t agree with CT-QMC to ~1e-4, and the limitation section honestly reports the beta=30/t, U=5t failure. These are genuine independent checks. The remaining load-bearing assumption is the low tensor-train rank of the transformed weak-coupling integrands. That assumption is supported only by observed convergence plateaus in a limited parameter window; the absence of a plateau at the highest order tested (nmax=40, Fig. 2(d)) and the reported breakdown at beta=30/t indicate the rank requirement is parameter- and order-dependent. Because the paper does not claim a formal guarantee, this is not an internal inconsistency, but it is a correctness/scalability risk for the central claim that the weak-coupling TCI solver is efficient up to 40th order and beyond. The reader's conditional verdict already captures this risk; my stress test does not change it, though it sharpens the specific test needed. The abstract's 'free from the sign problem' phrasing is also somewhat strong given the cancellation issue in Sec. III.C, but that is secondary to the rank assumption.","tokens_in":22439,"tokens_out":8320,"duration_ms":91262,"concrete_test":"For the exactly solvable model, compute the minimal bond dimension chi_min needed to reach a fixed relative error (e.g. 1e-5) in the partition-function integral at beta=20/t, U=5t for n=30, 40, 56, using the same TCI code and quadrature. Then repeat at beta=30/t, U=5t for n=40 and n=56, increasing chi up to 1000. If chi_min grows worse than linearly with n, or fails to saturate at beta=30/t, the low-rank hypothesis is not strong enough to support the claimed efficiency outside the benchmarked window. A complementary check is to rerun the beta=30/t, U=5t DMFT loop with an environment-optimized TCI; if the unphysical positive Gloc disappears, the bottleneck is the element-wise TCI objective rather than intrinsic rank.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency claim rests on the assertion in Sec. III.A that, after the simplex-to-hypercube map, the integrands P(h(v)) and Q_k(v;tau) have small tensor-train rank. The evidence is empirical, and at the highest order tested it is not conclusive: Fig. 2(d) shows that for nmax=40 the error at tau=beta/2 decreases monotonically with chi and has no plateau up to chi=200, so the rank actually needed to represent the n=40 integrand is not shown to be small. Since the cost scales as O(chi^2 d n^6) (Sec. III.A), a rank that grows with order undermines the claim of efficient evaluation of high-order terms. Sec. III.C documents the practical failure mode: at beta=30/t, U=5t, where the estimate Eq. (51) requires nmax ~ 56, chi=200 gives slightly positive Gloc(tau), violating the exact inequality Gloc(tau)<0. This shows the low-rank assumption can break in the low-temperature/strong-coupling regime that is the method's advertised target. No rank bound or scaling analysis is provided, so the method's broader applicability remains conditional on this empirical property.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a deterministic impurity solver based on tensor cross interpolation (TCI) applied to the weak-coupling expansion. The high-dimensional integrals over the simplex are mapped to hypercubes, discontinuities are handled by splitting the time-ordered domain into k sectors, and a tensor-train representation of each integrand is constructed by TCI. The solver is benchmarked on an exactly solvable impurity model and then used as a DMFT impurity solver for the Hubbard model on the Bethe lattice, where it reproduces the metal-insulator crossover, the first-order Mott transition with hysteresis, the doublon number, and the lattice free energy. The central numerical claims are supported by external comparisons with exact solutions and CT-HYB/CT-AUX results.","tokens_in":22675,"tokens_out":5500,"duration_ms":61583,"significance":"If the numerical results are correct, the method offers a deterministic, sign-problem-free (in principle) alternative to CT-QMC for impurity models, with the additional capability of computing the free energy directly. The paper includes strong external benchmarks: errors near 1e-5 for the exactly solvable model, agreement with CT-HYB and CT-AUX at the 1e-4 level in DMFT, and a correctly located first-order Mott transition. The method has no fitted parameters in the physical results, and the authors are candid about known failure modes. The main open question is the scaling of the required tensor-train rank with perturbation order, which is the pillar on which the efficiency claim rests.","major_comments":[{"comment":"The statement in Sec. III.A that 'the integrand in the weak-coupling expansion has a low-rank structure' is not established at the highest order tested. For nmax=40, the error at tau=beta/2 decreases monotonically with chi and shows no plateau up to chi=200, so the rank needed to represent the n=40 integrand is not demonstrated to be small. Since the cost estimate in Sec. III.A scales as O(chi^2 d n^6), a rank that grows with n would limit the method to lower orders than the abstract implies. Please provide a scaling analysis (e.g., the minimal chi required to reach a target error as a function of n for n=10,...,40, or the bond-dimension profile of the tensor train at n=40) and qualify the abstract's 'naturally have a low-rank structure' accordingly.","section":"III.A, Fig. 2(d)"},{"comment":"The rough estimate (51) is used both to choose nmax and to draw the 'works stably' boundary in Fig. 4. This estimate controls only the truncation error of the perturbative series, not the convergence of the TCI rank or the cancellation among k-sectors. The observed failure at beta=30/t, U=5t (slightly positive Gloc despite chi=200) illustrates that accuracy is also limited by rank convergence, which Eq. (51) does not capture. I recommend stating explicitly in Sec. III.C that Eq. (51) is a necessary but not sufficient condition for accuracy, and, if feasible, reporting the bond-dimension requirement at the failure point.","section":"III.C and Eq. (51)"}],"minor_comments":[{"comment":"The summation in Eq. (38) starts at n=1, omitting the n=0 term (which equals unity); either start at n=0 or add a sentence explaining that the n=0 contribution has been separated.","section":"Eq. (38)"},{"comment":"In the comparison of the three summation methods, the phrase 'approximately n+1 times smaller' should be clarified, e.g., as 'by a factor of about n+1' or 'roughly 1/(n+1) of the cost'; the current wording is ambiguous.","section":"II.E.3"},{"comment":"Reference [39] is typeset as 'N´u˜nez-Fern´andezet al.' with a missing space before 'et al.'; please fix the formatting.","section":"References"},{"comment":"The abstract's claim of being 'free from the sign problem' is later qualified in Sec. III.C by the cancellation between positive and negative contributions in the k-sum, which can require larger bond dimensions. Adding a sentence to the abstract or Introduction with this qualification would prevent overreading of the claim.","section":"Abstract and Sec. III.C"},{"comment":"In Fig. 2(a) and (b), the TCI and exact curves are nearly indistinguishable; adding explicit error insets or distinct markers at the plotted points would improve readability.","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"This is a solid methods paper with strong external benchmarks and an unusually candid limitations section. The main uncertainty is whether the low-rank behavior of the weak-coupling integrand persists beyond n=30; requiring the authors to document the rank-versus-order scaling is a reasonable and fixable request. The paper fits the scope of cond-mat.str-el, and I see no issues with novelty or citation practice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely new: it applies tensor cross interpolation to the equilibrium weak-coupling expansion of quantum impurity problems, and takes it all the way to DMFT with a correctly placed first-order Mott transition. The benchmarks are the strong part. Against an exactly solvable model the error reaches ~1e-5, and in DMFT the Green's functions agree with CT-HYB and CT-AUX to about 1e-4. The free energy calculation is a real plus, since CT-QMC struggles with that. The limitations section is unusually candid: it admits the breakdown at beta=30/t, U=5t where the Green's function turns slightly positive, and it distinguishes the cancellation problem from QMC sign problems. That honesty earns credit.\n\nThe main soft spot is the central efficiency claim. The paper asserts that the weak-coupling integrands have low tensor-train rank after the simplex-to-hypercube mapping. The evidence is numerical, and at the highest order tested (nmax=40) the error in Fig. 2(d) has not plateaued by chi=200. So the rank needed for 40th-order terms is not actually shown to be small; it may grow with order. That does not sink the paper, because the achieved accuracy at chi=200 is still excellent, but it means the broad applicability of the method in the low-temperature/strong-coupling regime rests on an empirical property that the authors themselves show can fail. The abstract's \"free from the sign problem\" is also a bit strong; the paper's own Sec. III.C documents cancellation issues that are analogous in effect, if not in origin. Missing code and data artifacts make independent verification harder, though the comparisons to exact and QMC results mitigate that.\n\nThese are real but proportionate concerns. The method is not oversold in the body, the numerics are careful, and the DMFT results are believable. The low-rank assumption is heuristic, but that is standard for TCI methods and the paper is upfront about it. Who is this for? Anyone working on impurity solvers, especially people interested in deterministic alternatives to QMC or in computing free energies. It deserves a serious referee; the main requests should be for a code release and a more careful discussion of rank scaling with perturbation order.","headline":"A credible, well-benchmarked first application of TCI to equilibrium weak-coupling impurity problems, with the low-rank assumption empirical rather than proven and the sign-problem claim a bit overstated, but clearly worth refereeing.","tokens_in":23222,"tokens_out":1173,"would_cite":true,"duration_ms":13702,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["71.27.+a","71.30.+h","02.70.-c"],"model":"deepseek-v4-flash","headline":"Tensor cross interpolation makes quantum impurity weak-coupling expansions tractable to 40th order.","keywords":["tensor cross interpolation","weak-coupling expansion","quantum impurity problem","dynamical mean-field theory","Mott transition","tensor train","free energy","Green's function"],"falsifier":"A direct test would be to run the TCI solver at $\\beta=30/t$ and $U=5t$ (where the paper reports unphysical positive $G_{\\rm loc}(\\tau)$ even with $\\chi=200$) and measure the bond dimension needed to drive $G_{\\rm loc}(\\tau)$ below zero and to match a CT-QMC reference to $10^{-4}$; if no plateau in the error-versus-$\\chi$ curve appears within feasible bond dimensions, the low-rank premise fails in that regime.","tokens_in":22213,"feed_emoji":"🧮","tokens_out":7163,"duration_ms":69303,"temperature":0.7,"pith_summary":"This paper aims to show that tensor cross interpolation (TCI) can make the weak-coupling expansion of quantum impurity problems practical, evaluating the perturbative integrals up to 40th order instead of being limited to low orders. The core claim is that after a variable transformation from the time-ordered simplex to a hypercube, the integrand—a product of two determinants—has a low-rank tensor-train structure that the TCI algorithm can discover from a small number of samples. If this is true, impurity solvers no longer need Monte Carlo sampling: they are sign-problem-free and can compute the partition function, and hence the free energy, directly. Benchmarks on an exactly solvable impurity model and on DMFT for the Hubbard model support the claim, reproducing the metal-to-Mott-insulator crossover and the first-order Mott transition with accuracy comparable to continuous-time quantum Monte Carlo.","feed_headline":"Weak-coupling series to 40th order via tensor cross interpolation","feed_subtitle":"Sign-problem-free impurity solver that also yields free energies and reproduces the Mott transition in DMFT.","key_machinery":"The machinery is the tensor cross interpolation (TCI) algorithm, an active-learning scheme that approximates a high-dimensional tensor by a tensor train (matrix product state) built from a few selected pivot slices instead of the full tensor. The paper combines TCI with Gauss-Kronrod quadrature: the simplex domain of the imaginary-time integrals is mapped to a hypercube by the bijection $h^{a,b}_n$, whose Jacobian is separable, and the Green's function's $\\tau$-dependence is handled by adding $\\tau$ as an extra tensor leg and splitting the simplex at $\\tau$ into $n+1$ regions so the integrand is continuous. The load-bearing step is that TCI finds a low-rank tensor-train representation of $\\tilde P$ and $\\tilde Q^k_\\sigma$ from a polynomial number of determinant evaluations, reducing an $n$-fold integral to a product of one-dimensional sums with $\\mathrm{O}(nd\\chi^2)$ cost.","core_discovery":"The central discovery is that the integrands in the weak-coupling expansion, $\\tilde P(v_1,\\dots,v_n)=P(h^{0,\\beta}_n(v))J_{h^{0,\\beta}_n}(v)$ for the partition function and the analogous $\\tilde Q^k_\\sigma(v;\\tau)$ for Green's functions, are low-rank when viewed as tensors, and that the rank stays small enough for TCI to capture the integral to high precision. The paper demonstrates this by the fast convergence of the computed Green's function with respect to the tensor-train bond dimension $\\chi$, and by agreement with exact solutions at order $n_{\\max}=40$, where errors of order $10^{-5}$ are reached with $\\chi=200$. In DMFT, the solver with $\\chi\\simeq200$ reproduces CT-HYB and CT-AUX results at the level of $10^{-4}$ and captures both the crossover at $\\beta=16/t$ and the first-order transition with hysteresis at $\\beta=20/t$.","pith_inferences":["Editorial inference: if the low-rank property survives when site or orbital indices are included as tensor legs, the same TCI machinery could solve multi-site cluster and multi-orbital problems in regimes where CT-QMC has a sign problem; the paper only offers this as a future possibility.","Editorial inference: the observed failure at $\\beta=30/t$, $U=5t$ suggests the required bond dimension grows with $\\beta U$, so an environment-aware TCI or a different simplex-to-hypercube map (both mentioned as future work) may be needed before the solver reaches the deep-Mott regime.","Editorial inference: TCI's deterministic quadrature errors could make it useful as a reference for validating CT-QMC codes on benchmark impurity models, since it provides independent free-energy estimates."],"forward_implications":["The weak-coupling TCI solver evaluates impurity Green's functions and partition functions without stochastic sampling, so it is free of the sign problem that limits CT-QMC in some multi-orbital, cluster, spin-orbit, and nonequilibrium setups.","In DMFT on the Bethe lattice it reproduces both the metal-to-insulator crossover at $\\beta=16/t$ and the first-order Mott transition with coexisting solutions and hysteresis at $\\beta=20/t$, matching CT-HYB and CT-AUX to roughly $10^{-4}$.","Because it computes $Z/Z_0$ directly, observables such as the doublon number and the lattice free energy are obtained almost without extra cost, quantities QMC accesses only with difficulty.","The computational cost grows polynomially, roughly $\\mathrm{O}(\\chi^2 d\\, n_{\\max}^6)$, so reaching order 40 is feasible with bond dimension $\\chi\\approx200$ and small-scale parallelization.","The same approach is expected to extend to multi-site cluster impurity problems, where strong-coupling TCI struggles with exponentially large local Hilbert spaces."],"supporting_citations":[{"why":"Supplies the TCI algorithm and library used to construct the tensor-train approximations.","marker":"[39]"},{"why":"Introduces the TCI treatment of impurity problems and the variable-transformation technique that this weak-coupling work adapts.","marker":"[40]"},{"why":"Provides the weak-coupling expansion formulas and determinant structure on which the solver is built.","marker":"[31]"},{"why":"Gives the exact Falicov-Kimball solution used as the benchmark.","marker":"[48]"},{"why":"CT-HYB solver whose DMFT results are compared with the TCI results.","marker":"[28]"},{"why":"CT-AUX solver whose DMFT results are compared with the TCI results.","marker":"[30]"},{"why":"Sets out the DMFT mapping from lattice to impurity problem used throughout.","marker":"[12]"}],"fun_headline_variants":["Sign-problem-free TCI solver hits 40th order","TCI solves quantum impurity problems at order 40","TCI solver captures Mott transition in DMFT","Weak-coupling expansion to 40th order without sign problem","TCI exploits low-rank tensors in weak-coupling expansion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption, supported only by numerical evidence, that the transformed integrands are well approximated by a tensor train with small bond dimension; this assumption is already seen to fail in the paper at $\\beta=30/t$, $U=5t$, where the computed Green's function becomes slightly positive.","fun_headline_variants_meta":{"raw":{"variants":["Sign-problem-free TCI solver hits 40th order","TCI solves quantum impurity problems at order 40","TCI solver captures Mott transition in DMFT","Weak-coupling expansion to 40th order without sign problem","TCI exploits low-rank tensors in weak-coupling expansion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001723,"raw_usage":{"total_tokens":6806,"prompt_tokens":926,"completion_tokens":5880,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":5799}},"tokens_in":542,"tokens_out":5880,"duration_ms":32290,"temperature":1.0,"reasoning_tokens":5799,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:57:09.110335+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to run the TCI solver at $\\beta=30/t$ and $U=5t$ (where the paper reports unphysical positive $G_{\\rm loc}(\\tau)$ even with $\\chi=200$) and measure the bond dimension needed to drive $G_{\\rm loc}(\\tau)$ below zero and to match a CT-QMC reference to $10^{-4}$; if no plateau in the error-versus-$\\chi$ curve appears within feasible bond dimensions, the low-rank premise fails in that regime.","supporting_citations":[{"cited_title":"Werner, A","cited_arxiv_id":null,"evidence_quote":"Introduces the TCI treatment of impurity problems and the variable-transformation technique that this weak-coupling work adapts."},{"cited_title":"Bl¨ umer, Metal-Insulator Transition and Optical Con- ductivity in High Dimensions (Shaker Verlag, Aachen, 2003)","cited_arxiv_id":null,"evidence_quote":"Provides the weak-coupling expansion formulas and determinant structure on which the solver is built."},{"cited_title":"Bulla, Zero Temperature Metal-Insulator Transition in the Infinite-Dimensional Hubbard Model, Phys","cited_arxiv_id":null,"evidence_quote":"CT-AUX solver whose DMFT results are compared with the TCI results."},{"cited_title":"(A9) The Jacobian of this map is given by Jha,b n (z) = Jf a,b n (gn(z))Jgn (z) = (b − a)n(1 − z1)n−1 · · ·(1 − zn−1)","cited_arxiv_id":null,"evidence_quote":"Sets out the DMFT mapping from lattice to impurity problem used throughout."}],"review_version":1}