{"id":"a54e7cb8-a707-434b-872b-c7026800695e","arxiv_id":"2607.27916","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Quarter-filled attractive Hubbard cylinders crossover from itinerant fermions to tightly bound singlet pairs; PCA and UMAP on correlation matrices track the same crossover region.","lead":"The paper maps how pairing grows in the quarter-filled attractive Hubbard model on narrow cylinders using exact diagonalization, DMRG, and unsupervised machine-learning dimensionality reduction. It reports a smooth BCS-to-BEC-like crossover near U≈-4 to -6 that it claims survives finite-size extrapolation and is visible to PCA/UMAP without assuming an order parameter.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Thermodynamic-limit extrapolation is only to a 4-leg cylinder (Lx fixed=4); a+b/Ly fits cannot establish a 2D BCS-BEC crossover or rule out finite-size artifacts.","rationale":"The reader identified the finite-size scaling as the weakest assumption, focusing on the linear 1/Ly form and the small number of widths. My concern is closely related but more structural: the extrapolation is not only based on five widths, it is to the wrong limit, because Lx is fixed at 4. This directly threatens the abstract's claim that the crossover is robust in the thermodynamic limit and not a finite-size artifact. That said, the qualitative physics of a smooth BCS-BEC-like crossover in the attractive Hubbard model is supported by prior literature and by the paper's internal ED/DMRG consistency, so I would not reject the paper. The recommendation remains CONDITIONAL, requiring the proposed larger-Lx check or an explicit reframing of the extrapolation as a 4-leg cylinder limit. Since the reader's verdict was already CONDITIONAL and this concern reinforces it without changing the verdict, I mark UNCHANGED.","tokens_in":19805,"tokens_out":5950,"duration_ms":91485,"concrete_test":"Repeat the DMRG/ED analysis at quarter filling on cylinders with Lx=6 and Lx=8 (e.g., 6x6, 8x6, 8x8) over the same U grid, and re-fit SP(0,0;Ly) and lambda_1(Ly) using both the current 1/Ly form and a two-variable form such as a + b/Ly + c/Lx. If the extrapolated crossover location shifts by more than about 1 in U, or the extrapolated SP(0,0) at U=-5 changes by more than ~20% relative to the Lx=4 result, then the fixed-circumference extrapolation is not a reliable 2D thermodynamic limit and the claims of robustness need to be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the crossover survives in the thermodynamic limit is extrapolated from cylinders with fixed circumference Lx=4, as stated in Sec. II. The finite-size scaling in Sec. III C (Eqs. 18-19) fits SP(0,0) and the PCA variance ratio lambda_1 linearly in 1/Ly with Ly=2..6. The extrapolated a values are therefore the thermodynamic limit of an infinite 4-leg cylinder, not of the 2D square lattice. In the true 2D limit both transverse width and longitudinal length must grow; with Lx fixed at 4, finite-circumference corrections are never sampled. The paper's own caveat that the extrapolations are 'estimates of the thermodynamic trend rather than precise thermodynamic-limit determinations' (Sec. III C) undercuts the abstract's unqualified claim that the signatures are 'robust with increasing system size.' The visible deviations from linearity in Fig. 10 and the sign change of the fitted b coefficient for lambda_1 at U=-5 (Appendix B, Table II) are based on only five widths with no error bars; they provide weak evidence for a qualitative change in the thermodynamic limit. If the leading 2D finite-size correction is not 1/Ly, the extrapolated a values are not valid thermodynamic-limit estimates and the 'not a finite-size artifact' conclusion is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the quarter-filled attractive Hubbard model on finite-width cylindrical lattices, fixing Lx=4 and varying Ly from 2 to 6. Exact diagonalization is used for 2x4, 3x4 and 4x4 clusters, and DMRG extends the study to 5x4 and 6x4. The authors report a monotonic lowering of the ground-state energy, a broad maximum of the excitation gap at intermediate coupling, increasing double occupancy and pairing correlations, decreasing spin correlations, negative two-hole binding energy across the attractive regime, and size-dependent three-hole binding. They then apply PCA and UMAP to charge-charge and pair-pair correlation matrices and claim that these unsupervised methods identify the same BCS-BEC-like crossover, centered near U≈-4 to -6, as conventional observables. Finite-size scaling of the ground-state energy, pairing structure factor and leading PCA variance ratio with 1/Ly is used to argue that the crossover signatures persist in the thermodynamic limit.","tokens_in":20159,"tokens_out":6331,"duration_ms":85706,"significance":"If the claims were fully established, the paper would provide a useful multi-method benchmark for the interaction-driven pairing crossover in finite-width attractive Hubbard systems and would demonstrate unsupervised dimensionality reduction as a complement to conventional observables. The strengths of the manuscript include the ED/DMRG cross-validation, the direct hole-binding-energy evidence, the explicit finite-size tables, and the consistency of several standard observables. However, the central thermodynamic-limit claim is only demonstrated for infinitely long four-leg cylinders, not for the two-dimensional square lattice, and the 'unbiased' machine-learning claim is weakened by the use of pairing-sensitive input features. The quantitative results also lack uncertainty estimates. The paper is therefore a sound finite-size study whose broader conclusions need substantial qualification.","major_comments":[{"comment":"The statement that the pairing and PCA signatures are 'robust in the thermodynamic limit' and 'not finite-size artifacts' is stronger than the evidence. Because Lx is fixed at 4 throughout (Sec. II), the extrapolated intercepts a in Eqs. (18) and (19) are thermodynamic limits of an infinitely long 4-leg cylinder, not of the 2D square lattice. The fits use only Ly=2,...,6, the paper admits 'some deviations from strict linearity are visible' (Sec. III C, Fig. 10), and Table II shows a sign change in b with no uncertainties. The authors themselves call the extrapolations 'estimates of the thermodynamic trend rather than precise thermodynamic-limit determinations.' The abstract and conclusion should be qualified to finite-width cylinder extrapolations unless additional scaling in Lx or a justification that 1/Ly captures the leading 2D correction is provided.","section":"Sec. III C, Eqs. (18)-(19); Abstract"},{"comment":"The claim that PCA/UMAP identify the crossover 'without prior assumptions regarding the relevant order parameters' is overstated. The feature vectors are not the full many-body state but the independent elements of charge-charge and pair-pair correlation matrices (Eq. (2)); the pair-pair matrix is explicitly a pairing-sensitive correlation. Dimensionality reduction of these matrices therefore assumes that pairing correlations are the relevant degrees of freedom. The agreement with conventional observables is reassuring, but it is not an independent or unbiased confirmation. Please replace 'independently' and 'order-parameter-independent' with statements about complementary analysis of a chosen correlation set.","section":"Sec. III B 1 and Fig. 6-7; Abstract"},{"comment":"No uncertainties are reported for the DMRG data or for the fitted parameters a and b in Tables I-II. With five data points and visible deviations from linearity, the fitted values—for example the sign change of b for λ1 at U=-5—cannot be assessed. The authors should report fit errors, goodness-of-fit measures, and DMRG truncation error estimates. This is necessary to support the quantitative location of the crossover at U≈-4 to -6.","section":"Sec. III C, Tables I-II"},{"comment":"UMAP hyperparameters (n_neighbors, min_dist, metric, initialization) and the number of samples per U are not given. The two-branch separation in pair-pair UMAP is a central machine-learning signature; without these settings the result is not reproducible, and it is unknown whether the branching is robust or an artifact of the embedding parameters. Please provide the full UMAP configuration and, ideally, a stability check over hyperparameter choices.","section":"Sec. III B 2, Fig. 7"}],"minor_comments":[{"comment":"The inset of Fig. 1 gives a broad gap maximum at U≈-6 to -7, while the Abstract and Conclusion place the crossover near U≈-4 to -6. These are not inconsistent for a broad crossover, but the text should explicitly acknowledge both estimates to avoid confusion.","section":"Sec. III A 1, Fig. 1"},{"comment":"Feature preprocessing is described inconsistently: Sec. II says the data are standardized by subtracting the mean and dividing by the standard deviation, while Sec. III B 1 says only centering is performed before constructing Σ. Please clarify which preprocessing was actually used.","section":"Sec. II and Sec. III B 1"},{"comment":"The horizontal axis R is not defined. If it is a site-distance label or a lattice distance, spell it out and specify how distances are measured under the periodic x and open y boundary conditions.","section":"Fig. 4"},{"comment":"The statement that DMRG energies 'are monitored throughout the sweeps and found to converge to the reported precision' is too vague. Please report the largest discarded weight or energy variance per system size and interaction strength so that the reliability of the extrapolations can be judged.","section":"Sec. II, DMRG"}],"recommendation":"major_revision","confidential_remarks":"The core finite-size ED/DMRG analysis appears sound and the conventional observables support a smooth pairing crossover on the cylinders studied. The main weaknesses are in the framing: the thermodynamic-limit claim is an extrapolation of a fixed-circumference cylinder, not of the 2D square lattice, and the machine-learning 'independence' claim is overstated. I would not reject the paper, but it should be revised to state precisely what is extrapolated, to add uncertainty estimates, and to temper the abstract and conclusion. The high number of self-citations to closely related preprints by the same group should also be checked for completeness and appropriateness."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nPunchline: this is a solid numerical study of the quarter-filled attractive Hubbard model on Ly×4 cylinders, and the ED/DMRG benchmarks are worth having. But the two headline claims—that the crossover is not a finite-size artifact and that the ML analysis independently confirms it—are softer than the abstract suggests. I would send it to a referee, expecting revision.\n\nWhat is actually new: the systematic Ly×4 dataset at quarter filling, the two-/three-/four-hole binding-energy trends, and the PCA/UMAP treatment of full correlation matrices. The ED/DMRG agreement for 2×4, 3×4 and 4×4 clusters is a genuinely useful cross-check, and the binding-energy results are clean: EB2 is negative through the attractive regime, EB3 changes sign with a pronounced finite-size shift, and EB4 stays positive. Those are concrete quantitative outputs someone else can build on.\n\nWhere it is soft: the finite-size scaling is the main issue. SP(0,0) and λ̃1 are extrapolated linearly in 1/Ly with Ly=2–6 and Lx=4 fixed. That gives, at best, the thermodynamic limit of an infinite 4-leg cylinder, not of the 2D square lattice. The paper concedes as much in Sec. III C, but the abstract and conclusion still say the signatures 'remain robust with increasing system size.' The extrapolations have no error bars, and visible nonlinearities appear near the crossover; the sign change of b in Table II rests on five widths. I do not think this kills the physics, but it means 'not a finite-size artifact' is not established by the data.\n\nSecond soft spot: the ML narrative. Calling PCA/UMAP 'order-parameter-independent' is misleading when the feature vectors are built from charge-charge and pair-pair correlation matrices—those choices already encode physical assumptions. PCA on pair correlations finding a crossover in pairing is closer to a consistency check than an independent confirmation. UMAP hyperparameters are not reported, and no code or data are released, which matters for a methods-heavy paper. In fairness, the physical conclusion does not depend on the ML; the conventional observables and binding energies carry it.\n\nBottom line: the crossover at quarter filling is known physics, so the paper's value is the finite-width benchmark plus a cautionary case study in applying unsupervised learning. The numerical core looks competent and the qualitative conclusions align with prior work. I would take it with revision: temper the abstract, add error bars or at least discuss the Lx=4 limitation honestly, and release data or hyperparameters.\n\nRecommendation: yes to peer review, but the referee should push on the scaling extrapolation and the 'independent' ML language.","headline":"Useful finite-width ED/DMRG benchmarks for the quarter-filled attractive Hubbard model, but the ML 'independent confirmation' is oversold and the thermodynamic-limit claim rests on fits to a single circumference.","tokens_in":20626,"tokens_out":2439,"would_cite":true,"duration_ms":38217,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["71.10.Fd","71.27.+a","74.20.-z"],"model":"deepseek-v4-flash","headline":"The quarter-filled attractive Hubbard model exhibits a continuous BCS-BEC-like pairing crossover near U ≈ -4 to -6 that persists after finite-size extrapolation.","keywords":["attractive Hubbard model","BCS-BEC crossover","pairing correlations","exact diagonalization","DMRG","principal component analysis","UMAP","finite-size scaling"],"falsifier":"A reader could compute the same observables on wider cylinders (e.g., Ly=8 or 10) or with periodic boundary conditions in both directions, and check whether the pairing structure factor and leading PCA variance ratio still extrapolate to nonzero values that increase monotonically with |U|, and whether the crossover remains near U ≈ -4 to -6; alternatively, a Quantum Monte Carlo calculation in the 2D thermodynamic limit could test whether the crossover persists.","tokens_in":19681,"feed_emoji":"⚛️","tokens_out":5488,"duration_ms":69347,"temperature":0.7,"pith_summary":"The paper argues that as the on-site attraction in the quarter-filled Hubbard model strengthens, the ground state evolves smoothly from itinerant, weakly correlated fermions into tightly bound local singlet pairs, with the crossover centered near U ≈ -4 to -6. It supports this with exact diagonalization and DMRG on cylindrical clusters, showing that excitation gaps, double occupancy, pairing structure factors, and hole-binding energies all change continuously across this range. It further shows that unsupervised PCA and UMAP applied to the full correlation matrices recover the same crossover without assuming any order parameter, and that linear finite-size scaling of the pairing structure factor and leading PCA variance ratio indicates these signatures persist toward the thermodynamic limit. A sympathetic reader would care because it suggests that pairing crossovers in this model are genuine many-body phenomena rather than finite-size artifacts, and that unsupervised machine learning can serve as an order-parameter-independent diagnostic for such crossovers.","feed_headline":"Machine learning spots Hubbard pairing crossover","feed_subtitle":"Unsupervised PCA and UMAP detect the same crossover as conventional observables, with no order parameters assumed.","key_machinery":"The key machinery is the competition between kinetic-energy-driven itinerancy and interaction-driven on-site pair formation, encoded in the Hubbard Hamiltonian with attractive U. The analysis is carried by real-space correlation matrices (charge-charge, spin-spin, pair-pair) whose structure is probed by PCA and UMAP, and by hole-binding energies as energetic criteria for bound states. Finite-size scaling uses leading-order 1/Ly extrapolations to connect finite cylinders to the thermodynamic limit.","core_discovery":"The central claim is that the quarter-filled attractive Hubbard model on finite-width cylinders hosts a continuous, interaction-driven crossover from a weakly correlated Fermi regime to a regime of tightly bound singlet pairs, analogous to the BCS-BEC crossover, and that this crossover is located around U ≈ -4 to -6. The authors establish this through multiple independent probes: the excitation gap shows a broad maximum at intermediate attraction; double occupancy rises while local moments fall; charge and pairing structure factors grow while spin correlations are suppressed; and the two-hole binding energy is negative throughout the attractive region while three-hole binding only appears at","pith_inferences":["The claim that the crossover is 'not a finite-size artifact' rests on a linear 1/Ly extrapolation from only five widths with fixed Lx=4; if higher-order corrections or the narrow circumference matter, the thermodynamic-limit location of the crossover could shift, though the qualitative picture likely survives.","A natural extension would be to apply the same PCA/UMAP pipeline to other fillings or to models with longer-range interactions to test whether the branch separation in pair-pair correlations is a generic signature of pairing crossovers.","The identified crossover scale near U ≈ -4 to -6 is close to the bandwidth scale for these cylinders; it would be instructive to test whether the crossover tracks the bandwidth or the density as lattice geometry changes."],"forward_implications":["If the crossover is robust, experiments with ultracold fermions in optical lattices at quarter filling should observe a smooth evolution of pairing correlations and a broad maximum in the excitation gap near the predicted interaction range.","Unsupervised machine learning on correlation matrices can serve as an order-parameter-independent tool for locating interaction-driven crossovers in other strongly correlated models.","The negative two-hole binding energy throughout the attractive regime and positive four-hole binding energy indicate that the dominant instability is pair formation rather than phase separation, which can guide quantum-gas-microscope searches.","Finite-size scaling suggests that the pairing structure factor enhancement persists in the thermodynamic limit, so pairing correlations are not merely a small-cluster effect."],"fun_headline_variants":["Machine learning finds Hubbard pairing crossover","AI reveals Hubbard crossover without order parameters","Unsupervised ML pinpoints Hubbard pairing crossover","PCA and UMAP spot crossover in Hubbard model","ML detects Hubbard pairing crossover without order parameters"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The extrapolation to the thermodynamic limit assumes that the leading finite-size corrections scale as 1/Ly, but only five cylinder widths (Ly=2 to 6) are used and the paper itself notes some visible deviations from strict linearity in the intermediate-coupling region.","fun_headline_variants_meta":{"raw":{"variants":["Machine learning finds Hubbard pairing crossover","AI reveals Hubbard crossover without order parameters","Unsupervised ML pinpoints Hubbard pairing crossover","PCA and UMAP spot crossover in Hubbard model","ML detects Hubbard pairing crossover without order parameters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001351,"raw_usage":{"total_tokens":5329,"prompt_tokens":759,"completion_tokens":4570,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":4505}},"tokens_in":503,"tokens_out":4570,"duration_ms":44888,"temperature":1.0,"reasoning_tokens":4505,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T03:14:55.657303+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could compute the same observables on wider cylinders (e.g., Ly=8 or 10) or with periodic boundary conditions in both directions, and check whether the pairing structure factor and leading PCA variance ratio still extrapolate to nonzero values that increase monotonically with |U|, and whether the crossover remains near U ≈ -4 to -6; alternatively, a Quantum Monte Carlo calculation in the 2D thermodynamic limit could test whether the crossover persists.","supporting_citations":[],"review_version":2}