{"id":"7aa8654f-0cd1-4f59-a6ac-d1cc579ece84","arxiv_id":"2509.05279","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A block-diagonal Hessian plus two Newton-Raphson steps makes subspace harmonic relaxation reproduce (H2O)10 thermodynamics nearly exactly at roughly comparable cost.","lead":"Researchers made a coarse-grained molecular model cheaper to evaluate by using just two Newton-Raphson relaxation steps and a block-diagonal curvature matrix. On a water cluster, the simplified model matches the full model's heat capacity and structure at comparable total cost.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'overall computational cost becomes comparable' claim is unsupported: only MC step counts are compared, but per-step CG-SHR cost is higher and no wall-clock benchmark is reported.","rationale":"The reader's weakest_assumption concerns block-diagonal Hessian truncation, which is a legitimate generalization risk, especially for proteins. However, for the water-cluster demonstration, the direct comparison to AA already tests whether the block-diagonal approximation preserves thermodynamic consistency; the reported 1 k_B agreement suggests it does for (H2O)10. The more load-bearing gap is the cost-equivalence claim: the abstract and conclusions explicitly promise 'overall computational cost becomes comparable,' yet the manuscript supplies only MC step counts. Since CG-SHR steps include multiple gradient and Hessian evaluations, step-count reduction alone cannot establish wall-clock competitiveness. This is not an internal mathematical inconsistency, but it is a missing piece of evidence for the paper's primary practical motivation. The reader also flagged the missing wall-clock baseline in the rationale, but not as the weakest assumption, so my agreement is partial. The verdict should remain CONDITIONAL: the accuracy claim is credible for water clusters, but the computational-efficiency claim needs a direct benchmark before acceptance as stated.","tokens_in":7844,"tokens_out":5802,"duration_ms":68524,"concrete_test":"Run the AA and CG-SHR replica-exchange simulations from §4 on identical hardware and record total wall-clock time to reach the same statistical error in C_v(T) and the O-O RDF. If the CG-SHR wall-clock time is not within a small factor (say ≤2–3×) of the AA time, the 'overall computational cost becomes comparable' claim is falsified. Report per-MC-step timing for energy, gradient, Hessian construction, and Newton–Raphson iterations to document the actual per-step cost ratio.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central efficiency claim (abstract and Conclusions) is that CG-SHR needs roughly an order of magnitude fewer MC steps, so the overall computational cost becomes comparable to AA. This does not follow from the data presented. Each CG-SHR step is more expensive than an AA step: it requires P=2 Newton-Raphson iterations, each needing a gradient evaluation, plus two block-diagonal Hessian evaluations (one at the initial guess r^(0), and one at the final point r^(P) for the harmonic free-energy correction). The block Hessian K_i is not cheap for large systems: even with the block-diagonal truncation, the second-derivative blocks depend on all pairwise interactions involving atoms in group i, so the scaling is still ~N^2 with a larger prefactor than a single energy evaluation. The manuscript reports no wall-clock times, no per-step cost breakdown, and no code/artifact. Therefore the 'comparable cost' half of the central claim is untested. A factor-of-10 reduction in MC steps could easily be erased if the CG-SHR step is more than ~10x slower, which is plausible for (H2O)10 and even more so for larger systems. The accuracy claim on (H2O)10 is credible, but the efficiency claim—central to the title 'computationally efficient'—is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper revisits the subspace harmonic relaxation (SHR) method for coarse-graining molecular clusters, originally introduced by the same authors. It proposes two algorithmic improvements: (i) replacing the conjugate-gradient subspace minimization with a fixed small number P of Newton-Raphson iterations and (ii) replacing the full Hessian by its block-diagonal approximation, so that the pseudoinverse and normal-mode calculation factorize over the rigid groups. The resulting CG potential is tested on the (H2O)10 cluster with q-TIP4P/F, comparing against all-atom replica-exchange Monte Carlo. The authors report that CG-SHR reproduces the heat capacity within 1 k_B, matches O-O RDFs essentially exactly, and recovers OH-bond and HOH-angle distributions via cheap harmonic backmapping, while requiring about 20x fewer MC steps than AA. On this basis they conclude that the overall computational cost is comparable to AA simulations and that the approach is applicable to larger systems such as proteins.","tokens_in":8210,"tokens_out":4321,"duration_ms":47374,"significance":"If the efficiency claim is valid, the paper offers a practical route to nearly exact coarse-graining for systems with well-separated fast and slow degrees of freedom, and the block-diagonal Hessian idea could remove the main bottleneck of the original SHR method. The accuracy results on (H2O)10 are credible and substantial: no target observable is fitted, the CG results are benchmarked against independent AA replica-exchange simulations, and the agreement is demonstrated for several complementary observables, including backmapped atomistic distributions. The algorithmic contribution is a natural extension of Paper I and is clearly described. However, the headline computational-efficiency claim is not supported by the data presented, and the extension to large covalently bonded systems is asserted without evidence. These issues are central to the paper's stated purpose and must be addressed.","major_comments":[{"comment":"The abstract and Conclusions claim that the CG simulation's 'overall computational cost becomes comparable' to AA because it requires roughly an order of magnitude fewer MC steps. The manuscript reports only MC step counts (4.5e9 vs 2.2e8) and no wall-clock times, per-step cost decomposition, or complexity comparison including Hessian construction/diagonalization. Each CG-SHR step is more expensive: it requires P=2 Newton-Raphson gradient evaluations plus construction and eigendecomposition of the block Hessian at r^(0) and again at r^(P). Even with block-diagonal truncation, each block Hessian K_i depends on all pairwise interactions involving group i, so the step cost has a larger prefactor than a single AA energy evaluation. A 20x reduction in MC steps can be erased if the CG step is more than 20x slower; this is plausible and untested. Please add actual CPU timings or careful flop co","section":"Numerical example; Conclusions"},{"comment":"The block-diagonal approximation K = diag{K_1,...,K_I} discards all inter-group second-derivative couplings. The paper presents this as enabling application to large proteins, but the numerical demonstration is limited to (H2O)10, where groups are independent monomers. For a covalently bonded protein, the inter-block Hessian elements between connected groups are not negligible, so the MEM condition (6) and the Newton step (11) are not obviously accurate. No test or error estimate is given for such systems. The broad applicability claim should either be supported by a test on a bonded multi-group system or explicitly narrowed.","section":"Eqs. (2)-(6); Introduction"},{"comment":"The algorithm computes the pseudoinverse tilde-K(r^(0)) once and reuses it for all P Newton-Raphson iterations. The text says 'we confirmed that this is the case' but provides no quantitative data on how the projected gradient or the resulting frequencies change if tilde-K is updated. This is a key approximation for large systems. Please report a comparison for (H2O)10 between fixed and updated pseudoinverse, or provide a convergence criterion.","section":"Algorithm, step 2"}],"minor_comments":[{"comment":"The statement that 'P=2 will maintain the accuracy on the order of 1 cm^-1 for the CG free energy of (H2O)10' is not supported by a figure, table, or error analysis. Please provide evidence or remove the quantitative claim.","section":"Preliminary Setup"},{"comment":"The y-axis label 'Cv (kb)' should be typeset as C_v/k_B, with units and notation defined in the caption. Currently the label is ambiguous.","section":"Fig. 1"},{"comment":"The statement that agreement is 'within the thickness of the lines' is visual. For the RDFs and backmapped distributions, no statistical errors or quantitative deviation (e.g., RMS difference) are reported. Adding a measure of uncertainty or deviation would strengthen the accuracy claim.","section":"Figs. 2-4"},{"comment":"The statement says 'The data that supports the findings of this study are available within the article.' However, the numerical data underlying the figures is not provided. Please deposit raw simulation data or specify a repository.","section":"Data Availability Statement"}],"recommendation":"major_revision","confidential_remarks":"The self-citation burden is high (the core formalism is from Paper I), but the incremental contribution is clear. The manuscript would be acceptable if the efficiency claim is backed by wall-clock timings and if the generalizability claim is tempered or tested. The main concern for the editor is that the central performance claim, appearing in the abstract and title, is currently unsubstantiated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The accuracy story is credible: on (H2O)10, the CG-SHR heat capacity sits within 1 k_B of the AA replica-exchange reference, and the O-O, O-H, and HOH distributions agree essentially to line thickness. The algorithmic changes—block-diagonal Hessian, fixed P=1 or 2 Newton–Raphson steps, and not updating the pseudoinverse inside the iteration—are genuine simplifications over Paper I and are described clearly. What is not established is the headline 'overall computational cost becomes comparable' claim: the paper only compares MC step counts (2.2×10^8 vs 4.5×10^9) and never measures wall-clock. Since each CG-SHR step costs at least one gradient and one block-Hessian evaluation (two of each in the implementation), a plausible per-step overhead of ~10x would erase the step-count advantage. The stress-test note is fair on that point.\n\nThe block-diagonal Hessian is the central assumption, validated on a single stiff cluster of essentially rigid monomers. For a protein, where groups are covalently connected and inter-block couplings are strong, the approximation is untested. The authors admit in the conclusions that direct evaluation may still be too expensive for large systems; that honesty is welcome, but it sits awkwardly with the abstract's promise of applicability to large proteins.\n\nThis is an incremental extension of the authors' own Paper I, not a new framework, and no code or data artifact is provided. Still, the numerics look careful, the backmapping via Eq. (12) is a useful low-cost feature, and even if the efficiency claim collapses, the near-exact CG potential is worth having for mapping studies. I would not cite it in my own work this year, but I would send it to a referee: someone should ask for wall-clock timings and at least one additional system before the title's 'computationally efficient' is justified.","headline":"Solid incremental algorithm with credible accuracy on (H2O)10, but the 'comparable cost' claim needs a wall-clock benchmark and a second system.","tokens_in":8644,"tokens_out":4103,"would_cite":false,"duration_ms":41908,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Two Newton–Raphson steps keep a coarse-grained water cluster thermodynamically consistent with its all-atom reference to within 1 k_B.","keywords":["coarse-graining","subspace harmonic relaxation","rigid-body model","block-diagonal Hessian","Newton-Raphson iteration","minimum energy manifold","thermodynamic consistency","water cluster"],"falsifier":"Run the same CG-SHR protocol with $P=2$ and the block-diagonal Hessian on a small peptide or protein fragment, and compare the heat capacity and pair distributions against an all-atom replica-exchange simulation. If the SHR results deviate by more than the statistical error or the heat-capacity peak shifts by several kelvin, the block-truncation assumption is falsified. A direct check would compare the MEM point and free-energy correction obtained with the block-diagonal pseudoinverse versus the full-Hessian pseudoinverse: a meaningful difference in the gradient projection or in the harmonic fre","tokens_in":7768,"feed_emoji":"💧","tokens_out":6281,"duration_ms":56560,"temperature":0.7,"texified_at":"2026-08-05T20:25:24.981536+00:00","pith_summary":"The paper claims that coarse-graining a molecular system by relaxing fast intramolecular degrees of freedom—rather than freezing them—can reproduce all-atom thermodynamics almost exactly, at a cost comparable to an all-atom simulation. It turns the Subspace Harmonic Relaxation (SHR) method into a practical algorithm: two Newton-Raphson steps using a block-diagonal Hessian pseudoinverse bring the system to the minimum energy manifold, and a harmonic free-energy correction defines the coarse-grained potential. On the (H2O)10 cluster, the resulting heat capacity and structural distributions agree with the all-atom reference within about 1 k_B, while needing roughly ten times fewer Monte Carlo samples. The approach also recovers all-atom bond and angle distributions from the coarse-grained simulation for negligible extra cost.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":2821,"prompt_tokens":811,"completion_tokens":2010,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":811,"completion_tokens_details":{"reasoning_tokens":1269}},"feed_headline":"Coarse-grained water matches all-atom thermodynamics within 1 k_B","feed_subtitle":"Block-diagonal Hessian cuts coarse-graining cost to all-atom levels while keeping heat capacity and structure nearly exact.","key_machinery":"The minimum energy manifold $R^*$, defined by $\\tilde{K}(r^*) M^{-1/2} \\nabla V(r^*) = 0$, where $\\tilde{K}$ is the pseudoinverse of the mass-scaled Hessian restricted to the fast degrees of freedom. The algorithm approximates $R^*$ with the iteration $r \\leftarrow r - M^{-1/2} \\tilde{K} M^{-1/2} \\nabla V$, with $\\tilde{K}$ assembled from the $L$ largest eigenpairs of each block-diagonal Hessian block $K_i$. The block-diagonal form $K = \\mathrm{diag}\\{K_1,\\ldots,K_I\\}$ is the object that cuts the diagonalization cost from $(3N)^3$ to the sum of per-group cubes, and the final Hessian evaluation at the relaxed point supplies the harmonic frequencies for the free-energy correction.","core_discovery":"The central claim is that near-exact thermodynamic consistency survives two aggressive approximations: truncating the Hessian to its per-monomer blocks and limiting subspace relaxation to $P=2$ Newton-Raphson iterations. For each rigid-body configuration, the algorithm relaxes the fast coordinates along the pseudoinverse of the block-diagonal Hessian, landing on a point near the minimum energy manifold, then evaluates the coarse-grained free energy as the all-atom potential plus a harmonic correction from the selected fast-mode frequencies. Demonstrated with the q-TIP4P/F model of (H2O)10, this cheap recipe reproduces the all-atom heat capacity and O–O, O–H, and H–O–H distributions within stat","pith_inferences":["We infer that the number of Newton-Raphson iterations P could be tuned per system: P=1 is exact for a harmonic system, and anharmonicity or strong inter-mode coupling would show up as a growing need for P>2, measurable by comparing CG-SHR with AA across temperatures.","We infer that the block-diagonal truncation's success on water may not carry over to covalently linked groups (e.g., protein backbone fragments), where inter-block Hessian couplings could be strong; a cheap diagnostic would compare the block-diagonal pseudoinverse with the full-Hessian pseudoinverse on a small peptide.","The harmonic reconstruction of AA configurations (Eq. 12) could serve as a fast surrogate for backmapping; one could test whether the reconstructed ensemble reproduces not just static distributions but also correlated two-body properties, which the paper does not report."],"forward_implications":["For (H2O)10, the CG-SHR heat capacity stays within about 1 k_B of the all-atom result at all temperatures, while the fixed-shape CG model shifts the heat-capacity peak by roughly 10 K.","Because the CG simulation needs about an order of magnitude fewer Monte Carlo steps, the overall cost of the SHR simulation is comparable to the all-atom simulation despite two Hessian evaluations per step.","All-atom properties such as OH bond-length and HOH-angle distributions follow from the CG simulation with almost no extra cost, by adding harmonic Gaussian fluctuations along the fast modes.","The block-diagonal Hessian makes the SHR evaluation scale per group rather than with the full system, which is the step that opens the method to larger molecules such as proteins."],"supporting_citations":[{"why":"introduces the original SHR (Rigid-HA) method, including the bijection to CG/fast coordinates, the MEM, and the harmonic free-energy expressions that this paper accelerates","marker":"[9]"},{"why":"supplies the q-TIP4P/F all-atom water potential used for both the AA reference and the constructed CG potential","marker":"[13]"},{"why":"provides the block-diagonal Hessian approximation for low-frequency mode analysis that motivates replacing the full Hessian","marker":"[12]"},{"why":"introduces replica exchange Monte Carlo used to sample the CG and AA ensembles","marker":"[20]"},{"why":"adds the replica exchange Monte Carlo protocol that the simulations follow","marker":"[21]"}],"fun_headline_variants":["Block-Hessian coarse-graining hits near-exact thermodynamics","Cheap CG water: block-diagonal Hessian keeps thermodynamics exact","Subspace harmonic relaxation: CG water at AA cost, near-exact","Fast coarse-graining with block-diagonal Hessian matches AA thermodynamics","Near-exact CG water from truncated Hessian and 2-iteration relaxation"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method assumes the off-diagonal blocks of the Hessian—the couplings between different rigid groups—can be dropped without shifting the relaxed fast coordinates or the selected fast-mode frequencies enough to change thermodynamic averages; this is validated only on a single small water cluster and not tested on covalently bonded groups such as a protein backbone.","fun_headline_variants_meta":{"raw":{"variants":["Block-Hessian coarse-graining hits near-exact thermodynamics","Cheap CG water: block-diagonal Hessian keeps thermodynamics exact","Subspace harmonic relaxation: CG water at AA cost, near-exact","Fast coarse-graining with block-diagonal Hessian matches AA thermodynamics","Near-exact CG water from truncated Hessian and 2-iteration relaxation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000482,"raw_usage":{"total_tokens":2225,"prompt_tokens":758,"completion_tokens":1467,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":1374}},"tokens_in":502,"tokens_out":1467,"duration_ms":8619,"temperature":1.0,"reasoning_tokens":1374,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:24:42.721755+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same CG-SHR protocol with $P=2$ and the block-diagonal Hessian on a small peptide or protein fragment, and compare the heat capacity and pair distributions against an all-atom replica-exchange simulation. If the SHR results deviate by more than the statistical error or the heat-capacity peak shifts by several kelvin, the block-truncation assumption is falsified. A direct check would compare the MEM point and free-energy correction obtained with the block-diagonal pseudoinverse versus the full-Hessian pseudoinverse: a meaningful difference in the gradient projection or in the harmonic fre","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces the original SHR (Rigid-HA) method, including the bijection to CG/fast coordinates, the MEM, and the harmonic free-energy expressions that this paper accelerates"},{"cited_title":"Tama , author F","cited_arxiv_id":null,"evidence_quote":"provides the block-diagonal Hessian approximation for low-frequency mode analysis that motivates replacing the full Hessian"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces replica exchange Monte Carlo used to sample the CG and AA ensembles"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"adds the replica exchange Monte Carlo protocol that the simulations follow"}],"review_version":1}