{"id":"cba8e112-338b-4d1a-88f3-eb49dc27b916","arxiv_id":"2501.18069","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Minimizing a new local L1 adaptive loss via basis rotation reduces the negative sign problem in QMC for frustrated spin systems.","lead":"Quantum Monte Carlo simulations of frustrated magnets can suffer from an exponential slowdown called the negative sign problem. This paper proposes a systematic local-basis rotation, tuned by a new L1 adaptive loss, to reduce that slowdown and tests it on several frustrated spin models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 drops a factor of |B|: the per-bond L1 adaptive loss bounds the extensive gap negativity only as L_L1 >= Delta/|B|, not L_L1 >= Delta; Eq. (8)'s common-psi equality is also generally false.","rationale":"I read the paper as proposing a practical local-basis optimization method, with Theorem 1 as the theoretical certificate connecting the optimized L1 loss to the physically relevant gap negativity. The numerical sections are substantial: random frustration-free 1D/2D ensembles, reconstruction of known sign-free transformations (J0-J1-J2-J3, BLBQ, Shastry-Sutherland), and kagome results. These empirical results are credible support for the heuristic value of the method, and the conclusion appropriately limits claims outside frustration-free systems. However, the central theorem is not supported by its proof. The derivation in Eqs. (12)-(13) produces the bound Delta <= |B| L_L1, not Delta <= L_L1. This is not merely a typo: L_L1 is an average over bonds, while Delta is extensive, so the inequality as printed cannot hold asymptotically. The authors' own Fig. 3 uses y=6x for L=6, confirming that the observed proportionality is Delta vs |B| L_L1. The common-psi equality in Eq. (8) is a second independent gap between definition and implementation; it may explain the 'adaptive' behavior but it is not justified by Perron-Frobenius alone. Because the abstract's claim that the loss 'effectively approximates negativity' rests on Theorem 1, the paper's central claim is invalid as written. The correct response is revision, fixing the factor, stating the weaker bound, and clarifying the loss actually optimized, rather than acceptance of the theorem as stated. I therefore agree with the reader's REJECT and, since the magnitude of the problem justifies that verdict, recommend no change to the verdict.","tokens_in":14528,"tokens_out":8077,"duration_ms":85078,"concrete_test":"Run exact diagonalization on the L=6 1D random frustration-free parent Hamiltonians used in Sec. V A: for a fixed optimized basis u, compute L_L1 = (1/|B|) sum_b [lambda_+((u tensor u) g_b (u dagger tensor u dagger)) - lambda(g_b)] and Delta = lambda_+(U G U dagger) - lambda(G). Check whether any instance satisfies Delta > L_L1; if so, Theorem 1 is false as stated. The corrected bound predicts Delta <= |B| L_L1 = 6 L_L1, so fit Delta versus L_L1 to see whether the slope is approximately |B|. In the same run, evaluate max_psi (1/|B|) sum_b psi^T (g_b')_+ psi separately from (1/|B|) sum_b lambda_+(g_b'); if they differ, Eq. (8)'s equality is invalid for the test ensemble.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 1 yields only a factor-|B| weaker bound. From Eq. (12), lambda_+(G) <= sum_b lambda_+(g_b). If the transformed Hamiltonian is frustration-free, lambda(G) = sum_b lambda(g_b), so subtracting gives lambda_+(G) - lambda(G) <= sum_b [lambda_+(g_b) - lambda(g_b)] = |B| L_L1(G,u). Thus the stated inequality L_L1 >= lambda_+(G) - lambda(G) is not implied; the proof gives L_L1 >= Delta/|B|. Since L_L1 is an intensive per-bond average while Delta is extensive, the missing factor is not a harmless constant: for larger systems the stated bound is off by a system-size-dependent factor. The authors' own Fig. 3 plots improvement in Delta against improvement in L_L1 for L=6 with the line y=6x, explicitly noting |B|=6; most points lie on or below that line, i.e. Delta_improvement <= 6 L_L1_improvement, not Delta_improvement <= L_L1_improvement. Separately, Eq. (8) equates (1/|B|) sum_b lambda_+(g_b') with max_psi (1/|B|) sum_b psi^T (g_b')_+ psi. The right-hand side is generally smaller than the left unless a single state is simultaneously the Perron vector of every (g_b')_+; frustration-freeness gives a common ground state of the g_b', not of the (g_b')_+. So the loss actually minimized by the adaptive algorithm is not provably the quantity used in Theorem 1. The numerical demonstrations are extensive and reproduce known sign-free regions, but they do not repair the stated certificate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a systematic local-basis-transformation approach to mitigating the negative sign problem in quantum Monte Carlo simulations of frustrated spin systems. It defines a negativity measure, proves that the stoquastic Hamiltonian G+ is the optimal virtual Hamiltonian in a reweighting scheme (Proposition 1), and introduces an L1 adaptive loss function for local basis optimization. The central theoretical claim is Theorem 1, which states that this loss upper-bounds the gap negativity λ_+(G)-λ(G) for frustration-free systems. The numerical sections apply the method to random 1D and 2D frustration-free parent Hamiltonians, the J0-J1-J2-J3 model, the bilinear-biquadratic chain, the Shastry-Sutherland model, and the kagome Heisenberg model, and compare orthogonal versus unitary local transformations. The paper reports substantial reductions of the gap negativity and recovery of several known sign-free basis choices.","tokens_in":14912,"tokens_out":7012,"duration_ms":84829,"significance":"If Theorem 1 were correct, the paper would be a significant contribution: it would provide a local, system-size-independent, computationally tractable loss function whose minimization provably reduces the sign problem, including in 2D systems. The numerical work is extensive and, as far as the manuscript shows, honestly executed: the gap negativity is evaluated by independent exact diagonalization, no fitted parameters are introduced, and the tests reproduce analytically known sign-free regions. Proposition 1 is a useful and correct formal observation. However, the main formal guarantee is not correct as stated. Theorem 1 drops a factor of |B|, Eq. (13) has the wrong sign/direction, and the equality in Eq. (8) between the sum of bond-wise Perron eigenvalues and the adaptive max-over-|ψ⟩ loss is unjustified. These are not presentation issues: they affect the paper's central claim that minimizing the proposed loss certifiably reduces the extensive gap negativity. The numerical demonstrations are promising, but they do not repair the broken certificate.","major_comments":[{"comment":"The stated inequality L_L1(G,u) ≥ λ_+(G)-λ(G) does not follow from the proof. From Eq. (12) and frustration-freeness one obtains λ_+(G)-λ(G) ≤ Σ_b [λ_+(g_b)-λ(g_b)] = |B| L_L1(G,u), i.e., L_L1(G,u) ≥ (λ_+(G)-λ(G))/|B|. The proof therefore establishes only a factor-|B|-weaker bound. Moreover, Eq. (13) has the wrong direction and sign: because G+ is entrywise larger than G, the Weyl monotonicity theorem gives λ_+(G) ≥ λ(G), so λ(G)-λ_+(G) ≤ 0, while the proof requires λ(G)-λ_+(G) ≥ 0. This is load-bearing because L_L1 is an intensive per-bond average while λ_+-λ is extensive; the missing factor depends on system size. The central claim that minimizing L_L1 certifiably reduces the gap negativity is therefore unsupported.","section":"IV.C, Theorem 1 and Eqs. (11)-(13)"},{"comment":"The equality (1/|B|) Σ_b λ_+((u⊗u)g_b(u†⊗u†)) = max_ψ (1/|B|) Σ_b Σ_ij ψ_i ψ_j^* |((u⊗u)g_b(u†⊗u†))_ij| would require a single normalized state |ψ⟩ that simultaneously saturates the Perron-Frobenius variational bound for every bond. Frustration-freeness (Definition 1) supplies a common eigenstate of the original operators g_b, but not of their absolute-value matrices (g_b)_+. The Perron vectors of the different (g_b)_+ will generally differ from bond to bond. Consequently, the quantity actually optimized by the adaptive max-over-|ψ⟩ procedure is not provably equal to the L_L1 used in Theorem 1; it is, in fact, pointwise no larger than L_L1, so it can be smaller than any bound derived for L_L1.","section":"IV.C, Eq. (8)"},{"comment":"The numerical data in Fig. 3 are consistent with the factor-|B| bound and not with Theorem 1 as stated. The authors draw the line y=6x, explicitly noting |B|=6, and state that most points lie below this line. This is exactly the relation Δ_improvement ≤ |B| L_L1_improvement, not Δ_improvement ≤ L_L1_improvement. The text's claim that this linear relationship is a consequence of Theorem 1 is therefore misleading: the slope 6 is precisely the factor missing from the theorem. This is not a minor plotting detail but a direct numerical confirmation that the stated upper bound is false.","section":"V.A, Fig. 3"}],"minor_comments":[{"comment":"The cross-reference to Eq. (10) should be to Eq. (9), and \"servers\" should be \"serves\" in the sentence \"it servers as a local Hamiltonian.\"","section":"IV.C, Remark 1"},{"comment":"The sentence beginning \"The reason for the apparent persistence of the negative sign in this region is that, although log(η)/β reaches the minimum value after optimization, the degeneracy of the ground state increases compared to the original Hamiltonian G\" is duplicated verbatim.","section":"V.B.1"},{"comment":"The model name is misspelled as \"Shustry-Sutherland\" in the text; it should be \"Shastry-Sutherland.\"","section":"V.B.3"},{"comment":"The phrase \"in certain parameter region Fig. 11\" should read \"in the parameter region shown in Fig. 11\" (or similar) for clarity.","section":"V.C"}],"recommendation":"reject","confidential_remarks":"The failure of Theorem 1 is load-bearing and is effectively acknowledged by the paper's own Fig. 3, where the line y=6x is drawn. The numerical results are promising and could support a revised manuscript that presents the L1 adaptive loss as a heuristic with strong empirical backing, but the current central formal claim about the upper bound is false. I see no evidence of circularity or parameter fitting; the problem is a genuine derivation gap in the main theoretical result."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know: the paper's L1 adaptive loss—weighting local off-diagonal elements by the Perron vector and updating during optimization—is a genuinely new algorithmic idea, and the numerical work is broad and honestly presented. It reproduces known sign-free regions on the J0-J1-J2-J3 chain, the BLBQ chain, and the Shastry-Sutherland model, shows clear mitigation on random 1D and 2D frustration-free models, and extends to the kagome Heisenberg model. The unitary-versus-orthogonal comparison is a nice addition.\n\nThe soft spot is the central theory. Theorem 1 as stated (L_L1 ≥ λ_+ − λ) is false for systems with more than one bond. The proof actually yields λ_+ − λ ≤ |B| L_L1, so the bound is missing a factor |B|. That is not a cosmetic slip: L_L1 is intensive, λ_+ − λ is extensive, so the claimed inequality cannot hold at large system size. The authors' own Fig. 3 uses the line y = 6x for L = 6—exactly the |B|-weighted relation—and most points lie below it, consistent with the weaker bound. Inequality (13) also has the wrong sign on the right: λ_+(G) ≥ λ(G) is the general direction, so λ(G) − λ_+(G) ≤ 0.\n\nEq. (8) is a second real problem. It equates the bond-average of local Perron roots with a single-state max of the bond-averaged L1 norm. In general, max of a sum ≤ sum of maxes; equality needs one state to be the Perron vector of every local (g_b)_+. Frustration-freeness gives a common eigenstate of the g_b, not of their absolute values. So the loss actually minimized is not provably the loss used in Theorem 1.\n\nCredit where due: the numerical results are extensive, the authors flag the frustration-free limitation and the weaker 2D performance, and the conclusion is careful not to overclaim. The paper does not fit parameters or hide negative results. But the theorem is load-bearing: the abstract's central claim—that a local quantity gives an upper bound on gap negativity—is unsupported as written.\n\nAnyone working on QMC sign problems and basis optimization will get value from the numerical methodology; the L1 adaptive loss is a plausible practical tool. But the proof of the central theorem must be fixed or explicitly downgraded to a heuristic. I'd send this to a serious referee, with a clear request for major revision: correct Theorem 1 to the |B|-weighted bound, fix the direction in Eq. (13), replace Eq. (8) with an inequality and say what the algorithm actually minimizes, and re-tie the numerical claims to the corrected statement. If that is done, the empirical story is strong enough to publish.","headline":"The L1 adaptive loss and the numerical work are worth attention, but Theorem 1 as stated is wrong by a factor |B| and Eq. (8) is unjustified; the paper needs a major revision before its central claim can be trusted.","tokens_in":15431,"tokens_out":9343,"would_cite":false,"duration_ms":100525,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that the negative sign problem in quantum Monte Carlo can be systematically mitigated by minimizing a locally computable L1 adaptive loss through local spin-basis rotations, with a proven upper bound for frustration-free…","keywords":["negative sign problem","quantum Monte Carlo","basis rotation","frustration-free Hamiltonian","stoquastic Hamiltonian","L1 adaptive loss","unitary transformation","frustrated quantum spin systems"],"falsifier":"Take a small translation-invariant frustration-free chain (e.g., an L=6 random parent Hamiltonian of the kind used in Fig. 2), minimize the L1 adaptive loss over all allowed local unitaries, and then compute the true gap negativity $\\lambda_+(G)-\\lambda(G)$ by exact diagonalization of the rotated Hamiltonian; finding any case with $\\lambda_+(G)-\\lambda(G) > L_{L1}(G,u)$ would refute Theorem 1. For the method's usefulness outside the frustration-free regime, run the kagome anisotropic Heisenberg model at $J_z=1, h_x=0$ with $\\beta=0.5$ and compare the average sign before and after loss minimization: if the sign worsens while the loss drops, the surrogate has detached from the true sign severity.","tokens_in":14327,"feed_emoji":"🎲","tokens_out":8646,"duration_ms":97768,"temperature":0.7,"pith_summary":"The paper's thesis is that the negative sign problem of quantum Monte Carlo is not an irreducible obstruction: it can be turned into an optimization objective over the representation basis. The authors introduce a metric (the negativity, and its temperature-independent form, the gap negativity) and prove that the usual fallback of simulating the absolutized stoquastic Hamiltonian is always the optimal virtual Hamiltonian, so only the basis remains to be chosen. Under the frustration-free assumption, they define the L1 adaptive loss, a sum of local bond terms that is independent of system size when the lattice is translationally invariant, and prove it is an upper bound on the gap negativity. Their numerical experiments show that optimizing local orthogonal or unitary rotations against this loss removes or substantially reduces the sign problem in random 1D and 2D frustration-free models, reproduces known sign-free dimer bases in the J0-J1-J2-J3 chain, bilinear-biquadratic chain, and Shastry-Sutherland model, and partially helps even in the strongly frustrated kagome Heisenberg model.","feed_headline":"Local spin rotations shrink the quantum Monte Carlo sign problem","feed_subtitle":"A single local loss both bounds and reduces the exponential sampling slowdown of frustrated quantum magnets.","key_machinery":"The central object is the L1 adaptive loss, $L_{L1}(G,u) = |B|^{-1}\\sum_{b\\in B}[\\lambda_+((u\\otimes u)g_b(u^\\dagger\\otimes u^\\dagger)) - \\lambda(g_b)]$, with $G=\\sum_b g_b=-H$ and $\\lambda_+$ the largest eigenvalue after taking absolute values elementwise. Rewriting the sum as a maximum over normalized non-negative states $|\\psi\\rangle$ of $\\sum_b \\sum_{i,j} \\psi_i\\psi_j^* |((u\\otimes u)g_b(u^\\dagger\\otimes u^\\dagger))_{ij}|$ turns basis optimization into a min-max problem. Perron-Frobenius lets the maximizing state be taken non-negative; the frustration-free property (one common eigenstate for all $g_b$) is what makes the inequality $L_{L1}\\ge \\lambda_+(G)-\\lambda(G)$ hold, so the loss is a computable, system-size-independent surrogate for the gap negativity. Riemannian gradient descent with Adam optimizes $u$ under the constraint $U=\\otimes_i u_i$.","core_discovery":"The central claim is that gap negativity $\\eta' = \\lambda_+(G)-\\lambda(G)$, the exponential rate at which the average sign decays, can be bounded from above by the L1 adaptive loss $L_{L1}(G,u)$ whenever $G=\\sum_b g_b$ is frustration-free. The theorem $L_{L1}(G,u)\\ge \\lambda_+(G)-\\lambda(G)$ means that a quantity computable from local, system-size-independent data controls the global severity of the sign problem, so reducing the loss through local basis rotations is a safe strategy. The numerical results support that this is more than an inequality: in frustration-free benchmarks the optimized rotations either reach $\\eta'=0$ (complete sign removal) or approach it, and in non-frustration-free cases such as the kagome Heisenberg model the method still lowers the sign problem in some parameter regimes. The paper also finds that allowing unitary rather than only orthogonal local rotations can restore symmetries and further reduce the loss, even though the original Hamiltonian is real symmetric.","pith_inferences":["Beyond the paper's claims, the upper-bound theorem suggests frustration-free-like structure is the regime where local losses can be reliable; for strongly frustrated systems, a hierarchical or adaptive cell partition that shrinks the cell-boundary terms noted in Remark 1 might be a natural extension.","Editorial extension: the min-max form of the loss resembles a local ground-state problem, so one testable next step is to vary the cell size and measure whether the optimized gap negativity improves monotonically, cleanly separating the approximation error of the loss from the optimization error.","Editorial extension: because unitary transformations help by breaking real-symmetry constraints, a concrete test would be to compare optimized complex phases against loop frustration patterns in the kagome model, to see whether they cancel sign-inducing amplitudes rather than just reducing their magnitude."],"forward_implications":["For any frustration-free Hamiltonian, the L1 adaptive loss is a safe objective: because it upper-bounds the gap negativity, lowering it cannot make the true sign problem worse than the loss itself.","The optimal-virtual-Hamiltonian proposition reduces sign-problem mitigation to basis selection alone, so future methods need not spend effort designing reweighting Hamiltonians.","The loss is system-size independent under translational invariance, which means the same optimized local rotation can be applied to arbitrarily large lattices without re-solving the optimization.","Numerically, the method reproduces analytically constructed sign-free bases for the J0-J1-J2-J3 chain, the bilinear-biquadratic chain, and the Shastry-Sutherland model, showing a common mechanism behind previously ad hoc transformations.","Allowing unitary local rotations can go beyond orthogonal ones, restoring lattice symmetries and further reducing the sign problem in models like the kagome Heisenberg antiferromagnet."],"supporting_citations":[{"why":"Formulates local basis optimization against a loss and supplies the supplementary remark that $G^+$ is not the sum of local $g_b^+$; the paper builds its local loss from this line of work.","marker":"[12]"},{"why":"Gives the related quasiprobability/optimal-stoquastic argument that underlies Proposition 1 on the optimal virtual Hamiltonian.","marker":"[13]"},{"why":"Provides the analytically known dimer-basis rotation for the J0-J1-J2-J3 chain that the optimized L1 loss reproduces.","marker":"[17]"},{"why":"Supplies the Shastry-Sutherland QMC benchmark whose known sign-free dimer basis is recovered in the frustration-free regime.","marker":"[19]"},{"why":"Defines the SO(N) bilinear-biquadratic chain basis transformation whose sign-free region the orthogonal optimization matches.","marker":"[20]"},{"why":"Establishes the exact dimer ground state of the Shastry-Sutherland lattice, the origin of the dimer-product sign-free basis.","marker":"[24]"},{"why":"Source of the frustration-free lemma used in Theorem 1: the ground state of the sum is a ground state of every local term.","marker":"[32]"},{"why":"Perron-Frobenius theorem, which justifies taking the maximizing state in the L1 adaptive loss to be non-negative.","marker":"[33]"}],"fun_headline_variants":["Local loss bounds negative sign in quantum Monte Carlo","Local spin rotations shrink the sign problem","Unitary rotations beat orthogonal for sign problem","Frustration-free bound: local loss controls sign issue","Optimize local basis to curb QMC sign problem"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Hamiltonian is frustration-free: a single state is a simultaneous eigenstate of every bond term, and this same non-negative state realizes the Perron-Frobenius maxima for all bonds at once; if that shared-state property fails, the L1 adaptive loss is only a heuristic and the upper-bound guarantee collapses.","fun_headline_variants_meta":{"raw":{"variants":["Local loss bounds negative sign in quantum Monte Carlo","Local spin rotations shrink the sign problem","Unitary rotations beat orthogonal for sign problem","Frustration-free bound: local loss controls sign issue","Optimize local basis to curb QMC sign problem"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000297,"raw_usage":{"total_tokens":1702,"prompt_tokens":906,"completion_tokens":796,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":725}},"tokens_in":522,"tokens_out":796,"duration_ms":9199,"temperature":1.0,"reasoning_tokens":725,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T00:48:10.525782+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small translation-invariant frustration-free chain (e.g., an L=6 random parent Hamiltonian of the kind used in Fig. 2), minimize the L1 adaptive loss over all allowed local unitaries, and then compute the true gap negativity $\\lambda_+(G)-\\lambda(G)$ by exact diagonalization of the rotated Hamiltonian; finding any case with $\\lambda_+(G)-\\lambda(G) > L_{L1}(G,u)$ would refute Theorem 1. For the method's usefulness outside the frustration-free regime, run the kagome anisotropic Heisenberg model at $J_z=1, h_x=0$ with $\\beta=0.5$ and compare the average sign before and after loss minimization: if the sign worsens while the loss drops, the surrogate has detached from the true sign severity.","supporting_citations":[{"cited_title":"Quantum Monte Carlo simulation method for spin systems","cited_arxiv_id":null,"evidence_quote":"Formulates local basis optimization against a loss and supplies the supplementary remark that $G^+$ is not the sum of local $g_b^+$; the paper builds its local loss from this line of work."},{"cited_title":"Reweighting method for quantum Monte Carlo simulations with the negative-sign problem","cited_arxiv_id":null,"evidence_quote":"Gives the related quasiprobability/optimal-stoquastic argument that underlies Proposition 1 on the optimal virtual Hamiltonian."},{"cited_title":"Wallman, and Stephen D","cited_arxiv_id":null,"evidence_quote":"Provides the analytically known dimer-basis rotation for the J0-J1-J2-J3 chain that the optimized L1 loss reproduces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Shastry-Sutherland QMC benchmark whose known sign-free dimer basis is recovered in the frustration-free regime."},{"cited_title":"Haldane and dimer gap in general double spin-chain models","cited_arxiv_id":null,"evidence_quote":"Defines the SO(N) bilinear-biquadratic chain basis transformation whose sign-free region the orthogonal optimization matches."},{"cited_title":"Symmetry- protected topological order and negative-sign problem for SO( N ) bilinear-biquadratic chains","cited_arxiv_id":null,"evidence_quote":"Establishes the exact dimer ground state of the Shastry-Sutherland lattice, the origin of the dimer-product sign-free basis."},{"cited_title":"Calculation of spin correlations in two-dimensional Ising systems from one-dimensional kinetic models","cited_arxiv_id":null,"evidence_quote":"Source of the frustration-free lemma used in Theorem 1: the ground state of the sum is a ground state of every local term."},{"cited_title":"Yu Kitaev","cited_arxiv_id":null,"evidence_quote":"Perron-Frobenius theorem, which justifies taking the maximizing state in the L1 adaptive loss to be non-negative."}],"review_version":1}