{"id":"2124cd5d-5fac-448f-8b41-6a52c0b2b3a8","arxiv_id":"1908.09096","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A regularized density-based convex formulation with an accelerated projected gradient method computes MGPE ground states efficiently, including the strong higher-order interaction regime.","lead":"The authors propose a faster way to compute the ground state of a Bose-Einstein condensate with higher-order interactions by minimizing a convex energy written in terms of the particle density rather than the wave function. The algorithm, called rDF-APG, is designed to stay efficient when interactions are strong, a regime where existing methods slow down or fail.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The O(1/k^2) APG guarantee is not established: the absolute-value extension (4.6) is not C^1 across density zeros, so the FISTA smoothness hypothesis fails.","rationale":"The paper's genuinely novel contribution is the density reformulation, and the regularization and discretization results (Theorem 2.6, Theorem 3.3, Theorem 3.5) are plausible and supported by the numerical tables. The most load-bearing spot is the optimization step: without a rigorous convergence guarantee for Algorithm 4.1, the abstract's efficiency claim has no theoretical anchor. The reader identified this same assumption, and the present read sharpens it from 'not proven' to 'actually violated': the absolute-value extension used to move off the feasible set is not C^1 across the hyperplanes rho_j = 0, so the standard FISTA theorem invoked in Section 4 does not apply. This does not invalidate the numerical evidence: the experiments in Section 5 show consistent second-order spatial accuracy and practical convergence, and the method may well be efficient. But the current manuscript claims a quadratic energy convergence rate without supplying the required smoothness or an alternative nonsmooth analysis. Since the concern points to a missing or incorrect proof rather than to a demonstrated failure of the method, the appropriate outcome remains conditional acceptance rather than rejection. I therefore keep the reader's verdict unchanged while strengthening the stated reason.","tokens_in":24049,"tokens_out":10434,"duration_ms":110737,"concrete_test":"Compute the one-sided partial derivatives of the kinetic term in (4.6) with respect to rho_j at rho_j = 0, fixing rho_{j+1} = b > 0, h = 1, and eps > 0. With T(a) = (a-b)^2/(|a|+|b|+2eps), the right derivative at a=0 is -b(3b+4eps)/(b+2eps)^2 and the left derivative is -b(b+4eps)/(b+2eps)^2. These are unequal for every b>0 and eps>0. If this derivative jump is confirmed, tilde_E_eps^h is not differentiable at such points, so the FISTA C^{1,1} hypothesis in [12] is violated and the O(1/k^2) proof does not apply. A further check would be to replace |rho_j| in the kinetic denominator by sqrt(rho_j^2+eps^2), run Algorithm 4.1 on Case II of Example 5.1, and compare iteration counts and line-search behavior with the current implementation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency claim rests on applying FISTA/APG to the extended objective tilde_E_eps^h in Eq. (4.6). The paper invokes the Beck-Teboulle O(1/k^2) rate, which requires f in (4.1) to be convex and C^{1,1} on all of R^{N-1}. But tilde_E_eps^h is not continuously differentiable there. For a single kinetic term T(a,b) = (a-b)^2/(|a|+|b|+2eps), at a=0 with b>0, the right and left partial derivatives with respect to a differ: for eps=0, b=1, the right derivative is -3 and the left derivative is -1. Consequently the gradient formula (4.8), which contains the discontinuous factor s_j = sign(rho_j), is not the gradient of a C^1 function, and the majorization Q_L in (4.11) cannot be justified by the standard Lipschitz-gradient argument. The same non-smooth absolute-value kinetic term appears in the modified functional E_hat_eps^h used for all numerical experiments in Section 5 via (A.1); Appendix A asserts the convergence results carry over but does not prove the hypotheses needed for the APG rate. Thus the rigorous basis for the headline 'much more efficient' and for the claimed O(1/k^2) convergence is missing. The numerical experiments provide practical evidence, but they do not substitute for a proof of the optimization-theoretic premise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a numerical method, rDF-APG, for computing ground states of the modified Gross-Pitaevskii equation with higher-order interaction. The nonconvex minimization over wave functions is recast as a convex minimization over densities ρ=|φ|^2, the kinetic term is regularized with a small parameter ε, convergence of the regularized ground states as ε→0 is stated (Theorem 2.6), a finite-difference discretization is analyzed (Theorem 3.3), and an accelerated projected gradient method is adapted. Numerical experiments in 1D, 2D, and 3D compare the method with a regularized Newton method and show strong advantages in the large-δ regime.","tokens_in":24372,"tokens_out":6579,"duration_ms":70348,"significance":"If the stated results all hold, this would be a useful contribution: the density-function formulation turns a difficult nonconvex constrained problem into a convex one, and the numerical evidence indicates favorable scaling with interaction strength. The paper provides no fitted parameters, and the main mathematical claims are clearly formulated. The numerical tables support the predicted spatial orders and the ε→0 behavior. However, several load-bearing proof ingredients are either omitted or only asserted, most importantly the smoothness/convexity hypothesis needed for the claimed FISTA O(1/k^2) rate, and the fact that all numerical experiments use a modified functional whose convergence theory is deferred to an appendix without proof.","major_comments":[{"comment":"The O(1/k^2) convergence of FISTA is invoked under the requirement that f in (4.1) be convex with C^{1,1} gradient on R^{N-1}. The proposed extension tilde_E_eps^h in (4.6) uses |ρ_j|, so it is not differentiable at points with ρ_j=0; the factor s_j=sign(ρ_j) in (4.8) is discontinuous, and direct computation of a single kinetic term (a-b)^2/(|a|+|b|+2ε) shows different one-sided derivatives at a=0. No restricted-domain version of the Beck–Teboulle proof with Lipschitz gradient only on the simplex W_h is provided, and convexity of the extension on all of R^{N-1} is not proved. Consequently the majorization Q_L in (4.11), the backtracking line search, and the claimed O(1/k^2) energy convergence are not justified by the cited theorem. The numerical experiments suggest practical convergence, but they do not substitute for the missing hypothesis.","section":"Section 4, Eqs. (4.5)–(4.8) and Algorithm 4.1"},{"comment":"All reported numerical experiments are formulated with the modified functional hat_E_eps^h in (A.1), not with E_eps^h whose convergence theory is proved in Sections 2–3. The appendix states that Theorems 3.2, 3.3, and 3.5 'still hold true' for hat_E_eps^h and then omits the proofs. Thus the numerical verification of the O(h) H^1-seminorm error and of the ε→0 convergence is performed for a functional whose regularity and convergence properties are asserted rather than established. The paper should either provide the missing proofs for hat_E_eps^h or perform the numerical checks on the functional actually covered by the theorems.","section":"Section 5 preamble and Appendix A, Eq. (A.1)"},{"comment":"The proof of Theorem 2.6 says the conclusion follows from Γ-convergence, but the text only shows monotonicity (Lemma 2.5) and lower semicontinuity of E_ε for a fixed ε. The Γ-convergence liminf inequality must treat sequences with ε_n→0 and ρ_n→ρ simultaneously, and Remark 2.7 concedes that the ε=0 case requires a more complicated argument. As written, the claimed H^1 convergence of ρ_ε^g to ρ_g is not fully demonstrated. The authors should supply the full Γ-convergence argument or give a precise reference that covers exactly this family of regularized functionals.","section":"Section 2.3, Theorem 2.6"},{"comment":"The proof of the spatial error estimate relies on estimates (3.31) and (3.32) with constants C2 and C3, but the bound for C3 depends on a uniform bound of ‖δ+ ρ_ε_g,h‖ whose derivation is omitted, and the claimed second-order consistency of the discretized singular kinetic term is only sketched. Since Theorem 3.3 is the central spatial-accuracy result used to justify the method, these omitted steps need to be supplied in a revised version.","section":"Section 3.2, Theorem 3.3"}],"minor_comments":[{"comment":"There are typographical errors that should be corrected, e.g. 'methd' on page 21 and 'tolerence' in Algorithm 4.1; also 'Gross-Pitaveskii' in the introduction should be 'Gross-Pitaevskii'.","section":"Throughout"},{"comment":"The notation tilde_E_eps^h and hat_E_eps^h is easy to confuse, especially because Section 5 uses hat_E_eps^h while referring to the results of Sections 2–3. A unified notation or an explicit table of the three functionals would improve readability.","section":"Eqs. (4.6), (4.14), (A.1)"},{"comment":"The abstract claims the method is 'much more efficient than the existing methods in the literature', but the numerical comparison in Table 5.5 is against a single method (regularized Newton). The conclusion should be restricted to the compared method, or additional comparisons should be reported.","section":"Abstract and Section 5.3"},{"comment":"The derivation of the gradient formula (3.11) is not shown; although the formula appears correct, a short derivation or reference would help readers verify the signs of the finite-difference terms.","section":"Section 3.1, Eq. (3.11)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the main idea is attractive. I do not see circularity or fabricated results; the concern is rigor in the optimization-theoretic premise and the gap between the analyzed functional and the functional used in the experiments. These gaps are fixable with additional proofs, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi —\n\nThis one deserves a serious referee, despite a real gap in the optimization theory. The new piece is the regularized density-function formulation for MGPE ground states: rewrite the energy in rho = |phi|^2, smooth the sqrt(rho) kinetic term with eps, prove Gamma-convergence as eps -> 0, discretize with finite differences, and solve the convex problem. That combination is not in the earlier literature, and it is an elegant, transferable idea for strong HOI. The spatial accuracy results match the numerics (O(h) H^1-seminorm, better in L2), and the convergence rho_eps^g -> rho_g is cleanly established. The paper also avoids fitted parameters; the only free constant is the regularization eps, whose effect is analyzed.\n\nNow the soft spot. The stress-test is right: the APG/FISTA rate in Section 4 is not proven. The extended objective (4.6) replaces rho_j by |rho_j| in the kinetic denominator, and that function is not C^1 on R^{N-1}. At a point with rho_j=0 and unequal neighbors, the partial derivative jumps. Beck-Teboulle's O(1/k^2) theorem requires convexity plus a Lipschitz-continuous gradient; that hypothesis is not met, so the majorization Q_L cannot be justified by the standard argument. The same issue carries over to the modified functional in Appendix A that all the numerics actually use. This does not mean the method fails — the experiments show practical convergence — but the headline 'much more efficient' is currently supported by numerics plus an unproved rate, not by a theorem.\n\nLesser issues: several proof steps are omitted (Lemma 2.1, parts of Theorem 2.4, Lemma 2.9, constant in Theorem 3.3). Most are routine, but Lemma 2.1 is load-bearing for convexity and is only cited. The efficiency comparison uses one baseline, regularized Newton, so the abstract's 'much more efficient than existing methods' is broader than the evidence. The self-citations are fine; they point to the existence and Thomas-Fermi results this work builds on.\n\nBottom line: the core mathematical strategy is sound and useful. If the authors either prove the needed smoothness (or a modified rate) for the absolute-value extension, or switch to a genuinely smooth extension and prove it, this becomes a solid paper. Send it to peer review; just make the APG premise a required revision.","headline":"A genuinely useful convex density formulation for MGPE ground states, with a real gap in the claimed FISTA O(1/k^2) rate because the extended objective is not C^1.","tokens_in":24874,"tokens_out":4571,"would_cite":true,"duration_ms":43823,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35Q55","65N06","65N25","90C30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that ground states of the modified Gross-Pitaevskii equation can be computed by minimizing a regularized convex functional of the density, and that the resulting method is convergent and much faster than existing…","keywords":["Bose-Einstein condensate","modified Gross-Pitaevskii equation","higher order interaction","ground state","density function formulation","accelerated projected gradient method","convex optimization","regularization"],"falsifier":"Take a small 1D grid, fix $\\varepsilon>0$, and test convexity of the extended energy $\\tilde{E}^\\varepsilon_h$ along the line segment from a feasible density to its negative reflection; a negative second directional derivative at any point with a negative coordinate would disprove the global convexity premise of the APG analysis. A second check is to record the energy decrease per iteration of Algorithm 4.1 for a very small $\\varepsilon$ (e.g., $10^{-8}$) on a coarse grid and see whether the claimed $O(1/k^2)$ bound actually holds in that regime.","tokens_in":23832,"feed_emoji":"⚛️","tokens_out":6800,"duration_ms":66257,"temperature":0.7,"pith_summary":"Computing the ground state of a Bose-Einstein condensate with higher-order interactions is normally a non-convex minimization problem over wave functions, but this paper shows it can be reformulated as a convex minimization over the density $\\rho=|\\phi|^2$ when the interaction parameters are nonnegative. Because the kinetic term $|\\nabla\\sqrt{\\rho}|^2$ becomes singular as $\\rho\\to 0$, the authors introduce a small regularization $\\varepsilon>0$ and prove that the regularized ground-state density converges to the true one in $H^1$ as $\\varepsilon\\to 0$. They discretize the regularized convex problem with second-order finite differences, prove first-order spatial accuracy in the $H^1$-seminorm, and solve the resulting finite-dimensional problem with the accelerated projected gradient method. The payoff is a method that works for all nonnegative interaction strengths and that the numerical experiments show to be dramatically faster than a regularized Newton method when the higher-order interaction is strong.","feed_headline":"A density rewrite makes BEC ground-state search convex and fast","feed_subtitle":"New regularized method converges in theory and beats existing solvers when interactions are strong.","key_machinery":"The central object is the regularized density functional $E_\\varepsilon(\\rho)=\\int\\big[\\tfrac{|\\nabla\\rho|^2}{8(\\rho+\\varepsilon)}+V\\rho+\\tfrac{\\beta}{2}\\rho^2+\\tfrac{\\delta}{2}|\\nabla\\rho|^2\\big]dx$ over the convex simplex $\\{\\rho\\ge 0,\\int\\rho=1\\}$. This functional replaces the non-convex wave-function energy with a convex problem in the density; the parameter $\\varepsilon$ removes the singularity of $|\\nabla\\sqrt{\\rho}|$ near $\\rho=0$ and also keeps the gradient of the energy bounded. The discrete version uses the denominator $\\rho_j+\\rho_{j+1}+2\\varepsilon$ in the finite-difference kinetic term, which preserves convexity on the simplex, and the optimization is carried out by the accelerated projected gradient method with projection onto the simplex, using an absolute-value extension of the discrete energy so that gradients are defined everywhere in $\\mathbb{R}^{N-1}$.","core_discovery":"For $\\beta\\ge 0$ and $\\delta\\ge 0$, the ground state of the modified Gross-Pitaevskii equation can be characterized as the minimizer $\\rho_g$ of the convex energy functional $E(\\rho)=\\int_{\\mathbb{R}^d}\\big[\\tfrac12|\\nabla\\sqrt{\\rho}|^2+V(x)\\rho+\\tfrac{\\beta}{2}\\rho^2+\\tfrac{\\delta}{2}|\\nabla\\rho|^2\\big]dx$ over densities with $\\int\\rho=1$ and $\\rho\\ge 0$. To make this tractable, the paper introduces the regularized functional $E_\\varepsilon(\\rho)=\\int_{\\mathbb{R}^d}\\big[\\tfrac{|\\nabla\\rho|^2}{8(\\rho+\\varepsilon)}+V\\rho+\\tfrac{\\beta}{2}\\rho^2+\\tfrac{\\delta}{2}|\\nabla\\rho|^2\\big]dx$, proves that its minimizer $\\rho^\\varepsilon_g$ converges to $\\rho_g$ in $H^1$ as $\\varepsilon\\to 0^+$, and derives an energy-difference bound that controls the $L^2$ and $H^1$ density errors. After a second-order finite-difference discretization on a bounded domain with homogeneous Dirichlet boundary conditions, the discrete minimizer converges to $\\rho^\\varepsilon_g$ with $O(h)$ error in the $H^1$-seminorm, and numerical results show second-order convergence in energy and $L^2$ density. The discrete problem is solved by the accelerated projected gradient method applied to an absolute-value extension of the discrete energy over all of $\\mathbb{R}^{N-1}$, with the paper claiming an $O(1/k^2)$ energy convergence rate and demonstrating in experiments that the method outperforms the regularized Newton method, especially for large $\\delta$.","pith_inferences":["The convexity of the density energy suggests a connection to Wasserstein gradient flows of $E_\\varepsilon$, which could yield unconditionally stable time-splitting schemes for computing ground states or dynamics, a direction the paper does not explore.","The observed nearly linear convergence in $\\varepsilon$ hints that Richardson extrapolation across two small values of $\\varepsilon$ could produce higher-order estimates of $\\rho_g$ at negligible extra cost, since the error appears systematic in the tested harmonic-trap cases.","The absolute-value extension of the discrete energy is the fragile part of the analysis; a natural repair would be to prove a restricted Lipschitz-gradient property on the simplex itself, or to replace the extension by a barrier/projection strategy that stays inside the feasible set.","The density formulation could be adapted to compute excited states or rotating condensates by adding a phase or angular-momentum constraint, but such extensions would require a separate convexity analysis."],"forward_implications":["Ground-state densities of the modified Gross-Pitaevskii equation for $\\beta,\\delta\\ge 0$ can be obtained by convex minimization, so global convergence guarantees from convex optimization apply directly.","The regularization error can be made arbitrarily small by driving $\\varepsilon\\to 0$, and the numerical evidence indicates nearly linear convergence of the density in $\\varepsilon$ for harmonic traps.","The second-order finite-difference discretization gives first-order accuracy in the $H^1$-seminorm and second-order accuracy in energy and $L^2$ density, so the main numerical cost is solving a well-conditioned convex problem rather than handling non-convexity.","In the strong-interaction regime, especially $\\delta\\gg 1$, the method's cost decreases as the interaction strength grows, unlike the regularized Newton method, making the density formulation attractive for strongly interacting condensates.","The method extends naturally to two and three dimensions on tensor grids, since the gradient evaluation and the simplex projection are per-node operations with no dimension-specific coupling."],"supporting_citations":[{"why":"Defines the ground-state problem for the MGPE and proves existence and uniqueness for $\\beta,\\delta\\ge 0$, giving the theoretical basis for the density reformulation.","marker":"[7]"},{"why":"Supplies the accelerated proximal gradient (FISTA) framework and the $O(1/k^2)$ energy convergence rate that Algorithm 4.1 relies on.","marker":"[12]"},{"why":"Provides the Gamma-convergence background used in the proof of Theorem 2.6 for convergence of the regularized ground states as $\\varepsilon\\to 0$.","marker":"[16]"},{"why":"Shows that the normalized gradient flow method can fail for the MGPE, motivating the need for a new approach such as the density formulation.","marker":"[40]"},{"why":"Derives Thomas-Fermi approximations for strong interactions, which the paper uses as initial guesses in the rDF-APG method.","marker":"[42]"},{"why":"The regularized Newton method is the baseline algorithm for the efficiency comparisons in the numerical experiments.","marker":"[44]"},{"why":"Derives the higher-order interaction term that leads to the modified Gross-Pitaevskii equation, defining the physical problem being solved.","marker":"[23]"}],"fun_headline_variants":["Convex density trick speeds BEC ground-state search","Regularized density method tames non-convex BEC problem","Faster BEC ground states via convex density reformulation","Accelerated gradient on regularized density solves BEC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The rigorous $O(1/k^2)$ rate and the line-search guarantee require the absolute-value extension of the discrete energy to be convex with a Lipschitz continuous gradient over all of $\\mathbb{R}^{N-1}$, a property the paper invokes but does not prove.","fun_headline_variants_meta":{"raw":{"variants":["Convex density trick speeds BEC ground-state search","Regularized density method tames non-convex BEC problem","Faster BEC ground states via convex density reformulation","Accelerated gradient on regularized density solves BEC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000201,"raw_usage":{"total_tokens":1477,"prompt_tokens":1143,"completion_tokens":334,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":759,"completion_tokens_details":{"reasoning_tokens":267}},"tokens_in":759,"tokens_out":334,"duration_ms":3942,"temperature":1.0,"reasoning_tokens":267,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:22:42.351522+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small 1D grid, fix $\\varepsilon>0$, and test convexity of the extended energy $\\tilde{E}^\\varepsilon_h$ along the line segment from a feasible density to its negative reflection; a negative second directional derivative at any point with a negative coordinate would disprove the global convexity premise of the APG analysis. A second check is to record the energy decrease per iteration of Algorithm 4.1 for a very small $\\varepsilon$ (e.g., $10^{-8}$) on a coarse grid and see whether the claimed $O(1/k^2)$ bound actually holds in that regime.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the ground-state problem for the MGPE and proves existence and uniqueness for $\\beta,\\delta\\ge 0$, giving the theoretical basis for the density reformulation."},{"cited_title":"Beck and M","cited_arxiv_id":null,"evidence_quote":"Supplies the accelerated proximal gradient (FISTA) framework and the $O(1/k^2)$ energy convergence rate that Algorithm 4.1 relies on."},{"cited_title":"Braides, A handbook of Γ-convegence, in M","cited_arxiv_id":null,"evidence_quote":"Provides the Gamma-convergence background used in the proof of Theorem 2.6 for convergence of the regularized ground states as $\\varepsilon\\to 0$."},{"cited_title":"Ruan, A normalized gradient flow method with attractive-repulsive splitting for comput- ing ground states of Bose-Einstein condensates with higher-order interaction , J","cited_arxiv_id":null,"evidence_quote":"Shows that the normalized gradient flow method can fail for the MGPE, motivating the need for a new approach such as the density formulation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Derives Thomas-Fermi approximations for strong interactions, which the paper uses as initial guesses in the rDF-APG method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The regularized Newton method is the baseline algorithm for the efficiency comparisons in the numerical experiments."},{"cited_title":"Collin, P","cited_arxiv_id":null,"evidence_quote":"Derives the higher-order interaction term that leads to the modified Gross-Pitaevskii equation, defining the physical problem being solved."}],"review_version":1}