{"id":"9c8f9fe0-c94c-4b1a-a8f2-e045baa04182","arxiv_id":"2412.19904","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An exact hybrid Monte Carlo framework is built by embedding SU(N) into complex matrix space, enabling non-separable Hamiltonians like Riemannian manifold HMC in lattice gauge theory without gauge fixing.","lead":"This paper presents a new way to run the hybrid Monte Carlo algorithm for lattice gauge theory, embedding SU(N) gauge fields into the space of general complex matrices so the Monte Carlo update runs in an unconstrained space. The framework makes it possible to use non-separable Hamiltonians, such as Riemannian manifold HMC, exactly, which could eventually speed up lattice QCD simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that non-separable kernels are handled exactly hinges on the implicit leapfrog map being single-valued and reversible; the paper neither proves uniqueness for general K(w) nor exercises a non-trivial K, and the p-update (Eq. 27) is a quadratic equation that can have multiple roots.","rationale":"The paper makes a clean theoretical construction: the factorization of the path integral in Sec. II.C is well motivated, the measure decomposition is explicit, and the numerical test with K=1 confirms that the embedding itself preserves SU(N) expectation values. These parts are solid. However, the headline contribution is the exact treatment of non-separable Hamiltonians, and that claim depends entirely on the implicit leapfrog integrator being a well-defined reversible map. The paper neither proves existence/uniqueness of solutions to the implicit equations for a general K(w) nor provides a numerical test in the non-separable regime. The reader identified exactly this as the weakest assumption, and I agree. The concern is load-bearing because the implicit p-update, Eq. (27), is a quadratic equation in p_{1/2} whenever the kinetic term depends on w (as in Eq. 25), and quadratic equations can have multiple roots. A multi-valued root selection breaks single-valuedness and reversibility, which are prerequisites for HMC exactness. The concrete test proposed here would settle the issue directly: if the implicit equations have a unique solution and reversibility holds for a non-trivial kernel, the central claim would be supported; if not, the claim fails in the intended regime of use. Since the reader's CONDITIONAL verdict already reflects this open question, my assessment does not change the verdict, hence UNCHANGED.","tokens_in":13298,"tokens_out":15410,"duration_ms":145752,"concrete_test":"Run the existing 32×32 2D SU(3) code with a non-trivial kernel, e.g., K(w)=1+ε Φ(w) or the RMHMC kernel G(U)=1+c(-D^2), and check: (i) unique solvability of Eqs. (27) and (29) for a sample of (w,p) by solving the fixed-point equations with several different initial guesses (e.g., p and w) via Newton iteration; record whether a unique root is found and whether the chosen branch is continuous in τ; (ii) reversibility, by integrating one MD trajectory forward from (w,p) to (w',p') and then backward from (w',-p') with the same solver, verifying return to (w,-p) to machine precision; (iii) exactness, by comparing the plaquette expectation values to the exact character-expansion values as in Fig. 2. If more than one root, a discontinuous branch, or a reversibility violation appears for any τ with acceptance above 0.5, the claim of exact non-separable HMC is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that embedding SU(N) into M_N(C) makes non-separable HMC Hamiltonians exact, because MD runs on unconstrained W and the physical SU(N) variables are recovered by polar decomposition. Exactness of the HMC step requires the discrete MD map (Eqs. 27-30) to be a single-valued, reversible, volume-preserving diffeomorphism. App. A proves symplecticity of the first pair only under the implicit-function assumption that Eq. (27) determines p_{1/2} uniquely for the given (w,p); the second pair (Eq. 29) similarly requires a unique w'. For the Hamiltonian of Eq. (25), ∂_w H contains a term quadratic in p, so Eq. (27) is a vector quadratic fixed-point equation. Even in a scalar toy with k(w)=1+ε w^2, the scalar equation p_{1/2} + (ετ/4)p_{1/2}^2 = p - (τ/2)S'(w) has two real roots for large enough ετ|p|. In the matrix case of Eq. (35), with K^{-1}(w)=J^T diag(G^{-1},m)J, no Lipschitz or small-τ condition is stated that would ensure a unique root. If the root is selected by an iteration's initial guess, the resulting map can be discontinuous or non-reversible, breaking detailed balance. This is not a cosmetic issue: the paper's numerical test (Sec. III) deliberately sets K=1, for which both implicit equations become explicit and the quadratic term vanishes, so the claimed non-separable regime is never exercised. The discussion mentions the cost of implicit iterations but does not address existence, uniqueness, or branch selection.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an extended HMC framework for lattice gauge theory by embedding SU(N) into the space M_N(C) of complex matrices. The physical SU(N) variables are recovered from the extended variable W via a reversible polar decomposition, and an auxiliary action S_0(Φ) is introduced to control the non-compact directions. The authors show that the integration measure factorizes into a physical Haar measure and an auxiliary Φ-dependent factor, so expectation values of SU(N) observables are unchanged. Molecular dynamics is performed on the unconstrained matrix elements of W, which permits, in principle, the use of non-separable Hamiltonians such as the Riemannian manifold HMC Hamiltonian. The paper presents a numerical test for 2D SU(3) pure gauge theory with a trivial kinetic kernel K=1, comparing plaquette expectation values to exact character-expansion values. The test shows agreement within 3σ, validating the embedding and the explicit leapfrog integration for that special case, but the non-separable regime, which is the stated motivation, is not numerically exercised.","tokens_in":13657,"tokens_out":4356,"duration_ms":47280,"significance":"If the framework indeed enables exact HMC with non-separable Hamiltonians without gauge fixing, it would be an important step toward Fourier-accelerated lattice QCD and a remedy for the finite-step-size inexactness of earlier Riemannian manifold HMC implementations. The path-integral factorization in Sec. II.C is derived carefully, the reversibility construction in Sec. II.B is thoughtful, and the numerical benchmark is a clean check of the embedding itself; no parameters are fitted to produce the expectation values. The main weakness is that the central claimed capability for non-separable kernels rests on an unproven well-posedness assumption for the implicit leapfrog integrator, and the provided numerical test deliberately avoids that regime. This gap is load-bearing for the paper's central claim.","major_comments":[{"comment":"The exactness claim for non-separable Hamiltonians depends on the implicit leapfrog map being a single-valued, reversible, volume-preserving diffeomorphism. The proof in App. A that the first pair preserves the symplectic form assumes that Eq. (27) uniquely determines p_{1/2} for given (w,p), and the second pair is justified only by time-reversal symmetry, inheriting the same assumption. For a general non-separable kernel, for example the Hamiltonian in Eq. (35) with K^{-1}(w)=J^T diag(G^{-1},m)J, the force ∂_w H contains terms quadratic in p, so Eq. (27) is a vector quadratic fixed-point equation. The paper states no Lipschitz condition, small-τ bound, or branch-selection rule that would guarantee a unique root. A scalar toy with k(w)=1+εw^2 already exhibits two real roots for the p-half-step when ετ|p| is large. If the root is selected by the initial guess of an iterative solver, the resulting map can be discontinuous or non-reversible, breaking detailed balance. This must be addressed for the claim of exact non-separable HMC to hold.","section":"II.D, Eqs. (27)–(30); App. A"},{"comment":"The numerical test uses the trivial kernel K(w)=1_{2N^2}, as stated in the second paragraph of Sec. III. With this kernel, both implicit equations (27) and (29) become explicit, the quadratic-in-p terms vanish, and the integrator reduces to an ordinary explicit leapfrog. The benchmark therefore exercises the embedding and the polar-decomposition step, but it does not exercise the non-separable regime that motivates the 'exact non-separable Hamiltonians' claim in the abstract and introduction. A numerical test with a nontrivial, gauge-covariant kernel (e.g., one based on a covariant Laplacian) is needed to demonstrate that the implicit update can be solved in practice, that a consistent branch choice preserves reversibility, and that the acceptance remains acceptable. Without such a test, the paper's central new capability remains unvalidated.","section":"III, Sec. II.D"}],"minor_comments":[{"comment":"The y-axis label of Fig. 3, '(deviation)/(stat. error)', has a stray parenthesis; it should read '(deviation)/(stat. error)' or more clearly 'difference in units of statistical error'.","section":"III, Fig. 3"},{"comment":"The notation \\lfloor N\\delta\\omega\\rfloor for the projection of an angle onto the range [-π,π) is confusing because the same symbol is standard for the integer floor function. A different notation, such as \\mathrm{proj}_{[-\\pi,\\pi)}(N\\delta\\omega), would improve clarity.","section":"App. C, Eq. (C2)"},{"comment":"The phrase 'a prior' should be 'a priori' in the sentence 'Since we do not know the typical values of the derivatives a prior'.","section":"App. A, after Eq. (A8)"},{"comment":"Step 5 says 'We add to the ensemble the physical configuration U, calculated from the accepted W', but the Markov-chain state is described earlier in the same section as the triplet (U,θ,Φ). It would be clearer to specify that U is taken from the stored triplet after acceptance, to avoid implying that W alone is the state variable.","section":"II.D, step 5"},{"comment":"The text says 'We provide an implementation of the algorithm with the decomposition (D1) in the Grid Python Toolkit (GPT) [29]', but the manuscript gives no repository link, version, or instructions to access the code. If the implementation is meant to be a deliverable, an availability statement would be helpful.","section":"App. D, final paragraph"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern raised in the review is valid and falls squarely on the central claim. The paper's factorization and embedding are sound and the 2D test is a nice check, but the non-separable case—the entire motivation—neither is proven well-posed nor numerically exercised. This is a fixable gap: the authors could add a uniqueness theorem under stated conditions, or replace the implicit leapfrog with a formulation that is manifestly single-valued, and should certainly run a test with a non-trivial kernel. With those additions the paper would be a strong contribution to hep-lat."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper does something genuinely new: it embeds SU(N) into M_N(C), runs HMC on the unconstrained matrix elements, and recovers the physical gauge field by a carefully reversible polar decomposition. The path-integral measure factorizes so the auxiliary directions cancel, which makes the embedding exact. Second, the numerical evidence for this construction uses only the trivial kinetic kernel K=1. The non-separable Hamiltonian that motivates the whole paper is never exercised, and the implicit integrator required for such a kernel is assumed to have a unique solution without proof.\n\nThe construction is solid in the regime it tests. The measure factorization in Sec. II.C is derived carefully, including the U(1)/center ambiguity in the U(N) decomposition, and the treatment of the measure-zero edge case with skew detailed balance is genuinely careful. The 2D SU(3) pure-gauge test matches exact plaquette values within 3σ, which is real evidence that the embedding plus HMC is exact when the kinetic term is trivial. The authors also explicitly acknowledge that their earlier RMHMC work was inexact; that is honest.\n\nThe gap is the non-separable case. The implicit leapfrog in Eqs. (27)-(30) requires solving p_{1/2}=p-(τ/2)∂_w H(w,p_{1/2}). For a general kernel K(w), ∂_w H contains a term quadratic in p, so this is a vector quadratic equation. Multiple roots are possible; a scalar toy k(w)=1+εw^2 already has two roots for large ετ|p|. If the root is selected by an iterative solver's initial guess, the map can be discontinuous or fail to be reversible, breaking detailed balance. App. A proves symplecticity only under an implicit-function assumption on uniqueness. The paper neither proves uniqueness for small τ nor specifies a branch-selection rule. This is central, not cosmetic, because exact HMC requires a single-valued reversible volume-preserving map on all of phase space. The discussion of computational cost skips over existence and convergence.\n\nMinor issues: the test is pure gauge without fermions, which is acceptable for a first demonstration, and the auxiliary action S0 is admittedly hand-tuned, with a four-dimensional quartic variant noted only in passing.\n\nWho this is for: lattice practitioners working on HMC acceleration, specifically RMHMC and Fourier acceleration. The paper deserves a serious referee. The referee should ask for a non-trivial kernel test, e.g., a simple gauge-field-dependent K(w), and for a clear statement of conditions under which Eqs. (27)-(30) have a unique solution, or a proof for small enough τ, or a reversible branch rule. If that can be supplied, the framework would be a significant advance. As it stands, the exactness claim for non-separable Hamiltonians is not yet established. I would send it to review, not desk-reject it.","headline":"Clever exact embedding for HMC; the headline non-separable capability is untested and rests on an unproven uniqueness assumption.","tokens_in":14191,"tokens_out":5269,"would_cite":true,"duration_ms":49912,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["11.15.Ha"],"model":"deepseek-v4-flash","headline":"Embedding SU(N) in complex matrix space lets lattice gauge HMC use non-separable Hamiltonians exactly.","keywords":["hybrid Monte Carlo","lattice gauge theory","non-separable Hamiltonian","Riemannian manifold HMC","polar decomposition","symplectic integrator","SU(N) embedding","Fourier acceleration"],"falsifier":"Run the extended HMC on a small lattice, for example two-dimensional $\\mathrm{SU}(3)$, with a nontrivial non-separable kernel such as a covariant Laplacian, and compare forward and reverse trajectories while shrinking the step size $\\tau$. A visible fraction of trajectories whose backward integration lands on a different configuration, or an acceptance rate that fails to approach one as $\\tau\\to 0$ in cases where the trivial-kernel run succeeds, would show the unique-solution assumption fails in the regime the paper targets.","tokens_in":13119,"feed_emoji":"🎲","tokens_out":10244,"duration_ms":91060,"temperature":0.7,"pith_summary":"This paper proposes a way to run hybrid Monte Carlo in lattice gauge theory with non-separable Hamiltonians, the kind needed for Riemannian-manifold HMC and Fourier acceleration, without gauge fixing and without the finite-step-size inexactness that has plagued earlier attempts. The move is to let the dynamical variable drift from $\\mathrm{SU}(N)$ into all complex $N\\times N$ matrices, run the molecular-dynamics update on unconstrained matrix elements, and recover the physical gauge field at each step by a reversible polar decomposition. The extra directions are given a confining auxiliary action, and the integration measure factorizes so they cancel out of expectation values. The authors verify exactness numerically on a $32\\times 32$ two-dimensional $\\mathrm{SU}(3)$ pure gauge theory with a trivial kinetic kernel $K=1$, matching known plaquette values to sub-percent precision. The claim that non-separable kernels work rests on the implicit leapfrog having a unique solution at each half-step, which the paper proves only for that trivial kernel.","feed_headline":"Embedding SU(N) in complex space makes non-separable gauge HMC exact","feed_subtitle":"Riemannian-manifold HMC can run without gauge fixing or finite-step errors.","key_machinery":"The load-bearing object is the polar decomposition $W=e^{i\\theta}\\Phi U$ with a continuity-preserving, reversible phase convention: after each molecular-dynamics step the $U(1)$ factor is updated by $\\delta\\theta=\\frac{1}{N}\\arg\\det(e^{-i\\theta}\\Omega')$, keeping the $\\mathrm{SU}(N)$ part smooth along the trajectory. On top of it, the algorithm uses a flat-space implicit leapfrog integrator on the $2N^2$ real variables $(x_{jk},y_{jk})$, with forces obtained by pulling gradients from the $(U,\\theta,\\Phi)$ basis back through the Jacobian $J(U,\\theta,\\Phi)$. The measure factorization $(dW)=\\sqrt{\\det g(\\Phi)}\\,d\\theta\\,(dU)\\,(d\\Phi)$ is what makes the auxiliary directions harmless: the $\\Phi$ integral is finite and factorizes out of every expectation value. An alternative decomposition using $T_0=\\mathrm{diag}(1,0,\\ldots,0)$ removes the $Z_N$ center ambiguity of $\\mathrm{SU}(N)$ entirely.","core_discovery":"The central claim is that the obstruction to non-separable HMC in gauge theory is not fundamental: it comes from treating $\\mathrm{SU}(N)$-valued links as constrained variables, so updating them requires exponentiating the force and generically breaks the symplectic two-form at finite step size. By embedding $\\mathrm{SU}(N)$ in $M_N(\\mathbb{C})$ and using the unconstrained matrix elements of $W$ as the molecular-dynamics variables, the leapfrog update becomes an ordinary symplectic integrator on a flat phase space. The physical $\\mathrm{SU}(N)$ configuration is recovered by the polar decomposition $W=e^{i\\theta}\\Phi U$ with a reversible rule for fixing the $U(1)$ angle, and the $\\Phi$, $\\theta$ directions contribute a factorized, well-defined path-integral factor that does not alter expectation values of $\\mathrm{SU}(N)$ observables. In the continuous-time limit the reduced equations on the physical variables coincide with the Riemannian-manifold HMC equations, so the framework lifts RMHMC from an approximate method with finite-step errors to an exact one, provided the implicit update equations have a unique solution.","pith_inferences":["A natural stress test the paper does not include: run the same exactness check with a nontrivial kernel, for example $K(w)=1-c\\nabla^2$ or a covariant Laplacian, and confirm that acceptances and plaquette values stay within errors as $\\tau$ shrinks.","If the implicit leapfrog equations develop multiple solutions for strong non-separable kernels, the discrete map becomes multi-valued and detailed balance could fail; monitoring fixed-point iteration convergence and comparing forward and backward trajectories would expose where that happens.","The auxiliary $\\Phi$ action is a tunable lever: the authors report that a quartic form worked better than the quadratic form in exploratory runs, suggesting systematic tuning of $S_0(\\Phi)$ could reduce rejections while keeping the algorithm away from $\\det W=0$.","A floating-point implementation may want to prefer the alternative $T_0$ decomposition to avoid the measure-zero but numerically awkward phase discontinuities of the $Z_N$ convention."],"forward_implications":["Riemannian-manifold HMC in lattice QCD can in principle be run without gauge fixing and without the finite-step symplecticity error of earlier implementations.","A covariant-Laplacian kinetic kernel, previously known to break symplecticity, becomes an allowed non-separable Hamiltonian inside the extended framework.","Expectation values of $\\mathrm{SU}(N)$-valued observables are unchanged by the embedding, so no reweighting or gauge-fixing correction is needed.","The transfer-matrix viewpoint yields higher-order symplectic integrators for the non-separable case, not just the basic leapfrog.","The extra cost of implicit iterations can be kept additive rather than multiplicative when fermion forces dominate, because the kernel-force terms can be pushed into a lower multilevel integrator."],"supporting_citations":[{"why":"Defines the hybrid Monte Carlo algorithm that the paper extends to non-separable Hamiltonians.","marker":"[1]"},{"why":"Warns that a nontrivial kernel $G(U)$ generically violates symplecticity for group-valued links, the problem this paper removes.","marker":"[6]"},{"why":"Introduces Riemannian-manifold HMC, the immediate application the new framework makes exact.","marker":"[7]"},{"why":"Describes the earlier lattice-QCD RMHMC implementation that the paper says was inexact at finite step size.","marker":"[8]"},{"why":"Uses gauge fixing for Fourier acceleration, an approach the embedding framework avoids.","marker":"[12]"},{"why":"Gives the group-manifold symplectic structure whose nonlinearities cause the failure documented in Appendix B.","marker":"[27]"},{"why":"Provides the multilevel integration scheme used to argue the implicit-iteration cost stays mild.","marker":"[20]"}],"fun_headline_variants":["Exact Riemannian HMC via embedding gauge links in complex space","Complex embedding makes gauge HMC symplectic and exact","Non-separable gauge HMC becomes exact with complex embedding","Embedding SU(N) in complex space yields exact RMHMC","Gauge HMC: complex embedding removes finite-step errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The discrete molecular-dynamics map is single-valued and reversible even when the kinetic kernel $K(w)$ is non-separable, because the implicit leapfrog equations are assumed to have a unique solution at every half-step; the paper proves and tests this only for the trivial kernel $K=1$.","fun_headline_variants_meta":{"raw":{"variants":["Exact Riemannian HMC via embedding gauge links in complex space","Complex embedding makes gauge HMC symplectic and exact","Non-separable gauge HMC becomes exact with complex embedding","Embedding SU(N) in complex space yields exact RMHMC","Gauge HMC: complex embedding removes finite-step errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0007,"raw_usage":{"total_tokens":3133,"prompt_tokens":890,"completion_tokens":2243,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":2158}},"tokens_in":506,"tokens_out":2243,"duration_ms":15535,"temperature":1.0,"reasoning_tokens":2158,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:47:53.360801+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the extended HMC on a small lattice, for example two-dimensional $\\mathrm{SU}(3)$, with a nontrivial non-separable kernel such as a covariant Laplacian, and compare forward and reverse trajectories while shrinking the step size $\\tau$. A visible fraction of trajectories whose backward integration lands on a different configuration, or an acceptance rate that fails to approach one as $\\tau\\to 0$ in cases where the trivial-kernel run succeeds, would show the unique-solution assumption fails in the regime the paper targets.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the hybrid Monte Carlo algorithm that the paper extends to non-separable Hamiltonians."},{"cited_title":"Parisi, Prolegomena to any future computer evalua- tion of the qcd mass spectrum, inProgress in Gauge Field Theory, edited by G","cited_arxiv_id":null,"evidence_quote":"Introduces Riemannian-manifold HMC, the immediate application the new framework makes exact."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the earlier lattice-QCD RMHMC implementation that the paper says was inexact at finite step size."},{"cited_title":"Duane and B","cited_arxiv_id":null,"evidence_quote":"Uses gauge fixing for Fourier acceleration, an approach the embedding framework avoids."},{"cited_title":"Cohen, M","cited_arxiv_id":null,"evidence_quote":"Provides the multilevel integration scheme used to argue the implicit-iteration cost stays mild."}],"review_version":1}