{"id":"cc715b26-30e9-46a3-af32-3e532732c612","arxiv_id":"2506.21133","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A rank-reduced relativistic CCSD method is proposed that compresses complex double-excitation amplitudes with SVD; benchmark tests indicate 1 kJ/mol accuracy with only a few percent of amplitudes retained for gold and ytterbium systems.","lead":"This paper studies a way to compress the large tables of numbers inside relativistic coupled cluster calculations for heavy-element molecules, using a matrix factorization that keeps only the most important contributions. Benchmarks on gold chloride chains, gold clusters, and an ytterbium chloride crystal model suggest that about 3 percent of the compressed amplitudes may be enough to stay near 1 kJ/mol accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The feasibility claim rests on post-hoc truncation of converged CCSD amplitudes, not on solving the rank-reduced equations; MP3-projector quality and nonlinear error propagation remain untested.","rationale":"The paper does several things well: the SVD/Takagi discussion recognizes the complex symmetric non-Hermitian issue, the Goldstone versus Brandow choice is substantive, Eqs. (13)-(16) are derived cleanly, and the MP3-over-MP2 projector observation is useful. The benchmarks are also clearly described. However, the load-bearing numerical evidence is the truncation simulation in Section IV. The abstract's central claim of 1 kJ/mol accuracy with only a few percent of amplitudes is about feasibility of solving the rank-reduced equations, and that requires demonstrating that MP3 projectors plus the compressed equations deliver such energies. The current evidence is essentially a necessary condition: exact CCSD amplitudes have low effective rank. It is not sufficient because the actual RR-RCCSD solution is constrained to the frozen MP3-generated subspace and solves projected residual equations, so truncation error can propagate and amplify differently. The reader identified exactly this weakness, and I agree. The paper remains a worthwhile feasibility study, but the central claim should be framed as 'a posteriori compressibility is promising' until a self-consistent RR-RCCSD benchmark is provided. Since the reader's CONDITIONAL verdict already reflects this, no verdict change is needed.","tokens_in":18625,"tokens_out":6517,"duration_ms":86762,"concrete_test":"Run the actual rank-reduced RCCSD iteration for at least one benchmark, e.g. YbCl7@CTEP or (AuCl)3, with the same basis sets and pseudopotentials: (1) compute MP3 amplitudes and their SVD projectors U at ε=1e-4; (2) solve the compressed equations (13)/(16) to convergence for T_XY using the reported residual factorizations; (3) compare the resulting correlation energy and a reaction energy (e.g. Au_n cohesion) with the exact CCSD values. If |ΔE| ≤1 kJ/mol and the retained T_XY fraction is close to ~3%, the feasibility claim survives. If the error exceeds 1 kJ/mol or the iterations do not converge, the a posteriori truncation study is not a faithful proxy and the central claim must be qualified or rejected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim is established by a posteriori truncation, not by the actual rank-reduced algorithm. In Section IV the authors state: 'converged cluster amplitudes were decomposed, singular values below the given threshold ε were discarded, and then the full tensor was reconstructed to simulate the rank-reduced CC approach.' This simulation is not equivalent to solving Eq. (13) with projectors fixed from MP3 amplitudes. The SVD truncation of the exact amplitude tensor is an optimal rank-r approximation to the exact solution in the full space; the rank-reduced RCCSD iterates instead in the frozen MP3-generated subspace and projects the residual equations. The two solutions coincide only if the MP3 subspace exactly contains the exact CCSD amplitudes, which is not shown and is generally false. Nonlinear feedback in the CCD terms (e.g. Appendix A, Eq. A3) can amplify small projector errors, and the transformed residual R_XY discards couplings to the orthogonal complement. Thus Fig. 4 gives an optimistic error estimate, not a bound on the self-consistent method. The conclusion 'rank reduction ... improving its computational scaling is feasible' therefore rests on an unverified assumption. The paper also ships no converged RR-CCSD energies or code output; the stated ongoing implementation supports that this test has not yet been performed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a relativistic generalization of rank-reduced coupled cluster theory with single and double excitations (RR-CCSD), based on a Tucker/SVD decomposition of the complex symmetric doubles amplitude tensor. The authors derive compressed working equations (Eqs. (13)–(16)), argue for the Goldstone rather than the antisymmetrized Brandow diagrammatic formalism because the decomposition does not respect antisymmetry, and recommend SVD over Takagi factorization for obtaining projectors. Benchmarks are reported for (AuCl)_n chains, Au_n clusters, and a YbCl_7@CTEP cluster model of solid YbCl_2. In these benchmarks, converged exact RCCSD amplitudes are truncated by discarding singular values below a threshold ε, and correlation energies are reconstructed from the truncated tensor. The authors find that ε ≈ 10^-4 gives errors near 1 kJ/mol, with only a few percent of double-amplitude components retained for the largest system, and conclude that rank reduction for relativistic CCSD with sub-O(N^6) scaling is feasible.","tokens_in":18818,"tokens_out":7468,"duration_ms":93880,"significance":"If the central numerical claims were established for the actual rank-reduced algorithm, this would be a practically useful step toward relativistic coupled-cluster calculations on embedded clusters and medium-sized heavy-element systems. The theoretical part is self-contained and the working equations appear algebraically coherent. The choice of an absolute 1 kJ/mol accuracy target is sensible and avoids the extensivity problem of relative-error criteria. The Goldstone-versus-Brandow discussion is a useful practical contribution for implementers. However, the benchmark protocol does not solve the rank-reduced equations at all; it only truncates converged exact CCSD amplitudes and reconstructs energies. The paper itself states in Section V that the implementation of the relativistic RR-CCSD method is ongoing and no self-consistent RR-CCSD results are reported. The headline accuracy and compression claims therefore rest on an optimistic proxy rather than on a test of the proposed method, which is the main weakness of the manuscript.","major_comments":[{"comment":"The benchmark protocol described in Section IV ('converged cluster amplitudes were decomposed, singular values below the given threshold ε were discarded, and then the full tensor was reconstructed to simulate the rank-reduced CC approach') does not simulate the rank-reduced RCCSD method defined by Eq. (13). The actual method fixes projectors from MP3 amplitudes and iteratively solves the compressed amplitude equations, while the reported test performs an optimal low-rank truncation of the converged exact CCSD amplitude tensor. These two procedures coincide only if the MP3-generated subspace exactly contains the exact CCSD amplitudes, which is not demonstrated and is generally false. The projected residual equations discard couplings to the orthogonal complement, and nonlinear terms such as the ladder-type contribution in Eq. (A3) can amplify small projector errors. Consequently, the error curves in Figs. 4 and 6 and the claimed universal threshold ε ≈ 10^-4 are optimistic estimates, not error bounds for the proposed algorithm. Since no self-consistent RR-CCSD energies are provided, the central feasibility claim is not yet supported by the numerical evidence.","section":"Section IV and Eq. (13)"},{"comment":"The statement in the Conclusions that MP3 projectors are 'substantially superior' to MP2 projectors is inferred only from the singular value spectra shown in Fig. 3. No benchmark is reported that actually uses MP2- or MP3-derived projectors to project the exact CCSD residual or to solve Eq. (13). Since the projector choice is the key approximation that determines the accuracy of the rank-reduced iteration, this inference is not established. A minimally sufficient test would be to compare reconstructed energies using MP2 versus MP3 projectors on the converged amplitudes, or to report the norm of the exact residual projected onto the MP3 subspace.","section":"Section IV, Fig. 3"}],"minor_comments":[{"comment":"There is a typo in the abstract: 'amplitudesvia' should read 'amplitudes via'.","section":"Abstract"},{"comment":"The definition of the transformed residual is misprinted: the sentence beginning 'where R_{X'Y'} = sum_{X'Y'} r_ab^ij ...' uses summation indices that duplicate the free indices. It should read R_{X'Y'} = sum_{ia,jb} r_{ab}^{ij} U_{X'}^{ia*} U_{Y'}^{jb*}.","section":"Section II C"},{"comment":"The comparison 'cf. Fig. 4a' is ambiguous because Fig. 4 panels are not labeled (a), (b), (c) in the caption; the relevant comparison for Au_n cohesion energies should refer to the middle panel of Fig. 4 or to a specifically labeled panel.","section":"Section IV, Fig. 6 discussion"},{"comment":"For a complex symmetric matrix, the SVD in Eq. (9) has right singular vectors V related to U by V = U^* up to column phases (the Takagi form). The text should state this explicitly and explain how the phases are handled when the same projector U is used on both indices in Eq. (10), since a general-purpose SVD routine returns U and V independently.","section":"Section II B, Eqs. (9)-(10)"},{"comment":"The percentage 'only 2.6% of doubles amplitudes are significant' refers to the fraction N_SVD^2/(N_occ N_virt)^2, but this is not directly the storage or operation-count reduction of the full RR-CCSD method, which also stores the projectors, the additional intermediate decompositions of Appendix A, and the density-fitting integrals. The text should avoid implying that this percentage equals the overall memory saving.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The central derivation is sound and the paper is potentially useful, but the numerical section does not test the method it proposes. The authors should either add self-consistent RR-CCSD results with MP3 projectors (even for a subset of the benchmark systems) or substantially soften the feasibility claim and reframe the paper as an a posteriori compression analysis. Given that the implementation is described as ongoing, the missing test appears to be the critical gap between the current manuscript and a publishable claim about the method's accuracy and scalability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper gives the first relativistic generalization of rank-reduced coupled cluster, and that part is genuinely useful. The authors work through the complex symmetric amplitude tensor correctly: SVD instead of eigendecomposition, a clear note that Takagi has the same singular values but no high-performance implementation, and a good argument for why Goldstone diagrams are necessary rather than Brandow. The benchmark suite is also well chosen: (AuCl)n chains, Aun clusters, and a YbCl7@CTEP embedded cluster model, spanning open-shell odd gold clusters and a three-dimensional extended object. The singularity spectra and the error-vs-threshold curves in Figs. 3–6 are clearly presented, and the 1 kJ/mol target is a sensible and honest accuracy metric.\n\nWhat the paper does not do is solve the rank-reduced RCCSD equations. The error estimates in Fig. 4 come from truncating the singular values of already-converged exact CCSD amplitudes and reconstructing the full tensor. That is an optimal low-rank approximation in the full space; it tells you how compressible the exact solution is. It does not test the actual iterative algorithm of Eq. (13), where the projector subspace is fixed from MP3 amplitudes and the residual is projected onto that subspace. The two can differ substantially if the MP3 subspace does not contain the exact CCSD amplitudes, and the nonlinear terms in Appendix A can amplify small projector errors. The paper acknowledges the implementation is ongoing, so this is a known limitation rather than a hidden one, but the abstract's claim that rank reduction is 'feasible' is supported only by the post-hoc proxy. The rank-scaling evidence is also thin: N_SVD ~ N^1.5 for systems up to a few monomers, not the linear scaling that would justify the O(N^5) claim for extended systems. That extrapolation is plausible but not demonstrated.\n\nThe math that is presented is coherent. The compressed working equations follow from the unitarity of the projectors, the factorized residual terms in Appendix A look right, and the complexity estimate O(N_occ N_virt N_DF N_SVD^2) is consistent with the non-relativistic literature. No code or data are shipped, which limits reproducibility, though the supplementary material apparently contains geometries and basis sets.\n\nWho should read this: anyone working on relativistic CC implementations or rank-reduced methods. It is a useful feasibility study and a reasonable basis for an implementation project. I would send it to a serious referee, but the referee should insist that the self-consistent rank-reduced calculations be either included or explicitly deferred to future work, and that the 'feasible' claim be softened accordingly.\n\nYes, worth a peer-review slot.\n\nBest.","headline":"Solid feasibility study for relativistic rank-reduced CCSD; the compression benchmarks are informative, but the central claim rests on post-hoc truncation of converged amplitudes, not on solving the rank-reduced equations.","tokens_in":19372,"tokens_out":1965,"would_cite":true,"duration_ms":21681,"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":"The paper claims that relativistic CCSD double-excitation amplitudes are so compressible that keeping singular values above about $10^{-4}$ reproduces 1 kJ/mol-accurate correlation and reaction energies while storing only a few percent of…","keywords":["rank-reduced coupled cluster","relativistic CCSD","Tucker decomposition","singular value decomposition","Goldstone diagrams","heavy-element molecules","amplitude compression","ytterbium chloride cluster"],"falsifier":"Implement the full iterative rank-reduced RCCSD equations for one of the benchmark systems, such as Au3 or the YbCl7@CTEP cluster, with $\\varepsilon=10^{-4}$, and compare the correlation energy to the exact CCSD value; if the discrepancy exceeds 1 kJ/mol, or if the MP3-derived projector subspace misses the dominant doubly-excited configurations, the central feasibility claim would be refuted.","tokens_in":18407,"feed_emoji":"⚛️","tokens_out":7031,"duration_ms":69152,"temperature":0.7,"pith_summary":"This paper claims that the double-excitation amplitudes of relativistic coupled cluster theory, which form complex symmetric tensors, can be compressed with a Tucker decomposition built from singular vectors of MP3 amplitudes. In benchmark calculations on gold chloride chains, gold clusters, and an embedded ytterbium chloride cluster, discarding singular values below about $10^{-4}$ reproduces CCSD correlation energies and reaction energies within 1 kJ/mol while retaining only a few percent of the amplitude data. The authors argue this makes a sub-$O(N^6)$ relativistic CCSD feasible and that the method should be formulated with Goldstone (symmetric) amplitudes rather than antisymmetrized Brandow amplitudes. If correct, the result would let high-precision relativistic calculations treat larger heavy-element molecules and crystal-embedded clusters.","feed_headline":"Relativistic CCSD needs only 3% of its doubles amplitudes","feed_subtitle":"SVD-based compression with a 1e-4 cutoff keeps heavy-element correlation energies within 1 kJ/mol.","key_machinery":"The load-bearing object is the Tucker/SVD decomposition of the doubles amplitude supermatrix $t_{ia,jb}\\approx\\sum_{XY} T_{XY} U^X_{ia} U^Y_{jb}$, where the unitary projectors $U$ come from the SVD of MP3 amplitudes. Because the relativistic amplitude matrix is complex symmetric but not Hermitian, the paper uses SVD rather than eigendecomposition; the compressed working equations $ET+TE=-R$ (with $E$ the transformed energy denominator matrix) are derived by projecting onto the $U$ subspace. To keep the quadratic ladder-type terms from restoring $O(N^6)$ cost, the paper adopts the device of decomposing the intermediate tensors $O_{ki,lj}$ and $Z_{kj,bc}$ into additional Tucker factors, which drives the formal cost toward $O(N^5)$.","core_discovery":"The central discovery is that the tensor $t^{ab}_{ij}$ of double-excitation amplitudes, viewed as a complex symmetric matrix $t_{ia,jb}$, has a low effective rank: $t_{ia,jb} \\approx \\sum_{XY} T_{XY} U^X_{ia} U^Y_{jb}$, with unitary projectors $U$ obtained by singular value decomposition of MP3 amplitudes. For three families of relativistic benchmarks, singular values below $\\varepsilon \\sim 10^{-4}$ can be discarded with correlation-energy errors below 1 kJ/mol, and for the largest embedded cluster only about 3% of the compressed doubles amplitudes remain significant. The paper also establishes that MP3 projectors are substantially superior to MP2 projectors and that compression must be applied to symmetric Goldstone amplitudes, because antisymmetrized Brandow amplitudes cannot be reformulated efficiently in reduced rank.","pith_inferences":["The benchmarks test compression by truncating already-converged exact amplitudes and reconstructing the tensor, not by iteratively solving the compressed equations (13); the true rank-reduced solver could show different errors, and the first test should compare the two on the smaller systems.","Because SVD applies naturally to non-square matrices, the same compression route should extend to Fock-space multireference relativistic coupled cluster, where amplitude tensors are rectangular; this is a direct next step the paper does not implement.","The generality of the $\\varepsilon\\sim 10^{-4}$ threshold could be probed by applying it to actinide-containing molecules or to property calculations, where the amplitude subspaces may need to remain field-independent.","If the effective rank keeps growing only as a low power of system size for three-dimensional embedded clusters, the method could make two-component CCSD with large basis sets routine for impurity centers in crystals, where basis-set incompleteness errors already exceed the truncation error."],"forward_implications":["Relativistic CCSD correlation energies of moderate-size heavy-element systems can be computed within about 1 kJ/mol while storing only a few percent of doubles amplitudes, dramatically reducing memory and disk requirements.","A singular-value threshold near $\\varepsilon\\sim 10^{-4}$ appears to control accuracy across the benchmarked systems, making the truncation a tunable one-parameter approximation.","Relative energies such as cohesion energies benefit from error compensation, allowing a looser threshold of about $3\\times10^{-4}$ for the same 1 kJ/mol target.","The effective rank grows slower than the full rank $N_{occ}N_{virt}$, roughly as $N_{occ}^{1.5}$ for the compact systems studied, so a sub-$N^6$ relativistic rank-reduced CCSD is plausible for spatially extended objects.","The framework is positioned for extension to CCSD(T) and CCSDT, although the number of Goldstone diagrams grows and becomes a practical challenge."],"supporting_citations":[{"why":"Introduces the Tucker decomposition of doubles amplitudes and the MP2/MP3 projector construction that the relativistic version adapts.","marker":"[48]"},{"why":"Shows how additional Tucker decompositions of intermediate tensors reduce the ladder terms to $O(N^5)$ scaling, the key route to sub-$N^6$ cost.","marker":"[51]"},{"why":"Provides the tensor-hypercontraction variant and error-compensation observations used to justify looser thresholds for relative energies.","marker":"[52]"},{"why":"Supplies the Kramers-unrestricted relativistic CCSD framework that the compressed equations generalize.","marker":"[5]"},{"why":"Documents the complex-valued, low-symmetry structure of relativistic integrals and amplitudes that forces the SVD-based approach.","marker":"[21]"},{"why":"Gives the antisymmetrized tensor factorization for relativistic integrals that the paper considers and rejects for amplitudes.","marker":"[33]"},{"why":"Defines the Goldstone and Brandow forms of the double-excitation operator, the formal distinction underlying the method choice.","marker":"[62]"},{"why":"Provides the compound-tunable embedding potential model of YbCl2 used as the three-dimensional benchmark system.","marker":"[72]"}],"fun_headline_variants":["SVD compression cuts relativistic CCSD amplitudes to 3%","3% of doubles amplitudes yield 1 kJ/mol CCSD accuracy","Rank-reduced CCSD: 3% of amplitudes keeps heavy-atom errors low","Relativistic CCSD: only 3% of amplitudes are significant","Tucker decomposition makes relativistic CCSD feasible"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy claims are tested by cutting small pieces out of already-finished exact calculations, not by running the compressed method itself; if the compressed equations behave differently, the 1 kJ/mol estimate could change.","fun_headline_variants_meta":{"raw":{"variants":["SVD compression cuts relativistic CCSD amplitudes to 3%","3% of doubles amplitudes yield 1 kJ/mol CCSD accuracy","Rank-reduced CCSD: 3% of amplitudes keeps heavy-atom errors low","Relativistic CCSD: only 3% of amplitudes are significant","Tucker decomposition makes relativistic CCSD feasible"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001575,"raw_usage":{"total_tokens":6272,"prompt_tokens":921,"completion_tokens":5351,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":5261}},"tokens_in":537,"tokens_out":5351,"duration_ms":40255,"temperature":1.0,"reasoning_tokens":5261,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:33:42.069749+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement the full iterative rank-reduced RCCSD equations for one of the benchmark systems, such as Au3 or the YbCl7@CTEP cluster, with $\\varepsilon=10^{-4}$, and compare the correlation energy to the exact CCSD value; if the discrepancy exceeds 1 kJ/mol, or if the MP3-derived projector subspace misses the dominant doubly-excited configurations, the central feasibility claim would be refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Tucker decomposition of doubles amplitudes and the MP2/MP3 projector construction that the relativistic version adapts."},{"cited_title":"Lesiuk, Quintic-scaling rank-reduced coupled clus- ter theory with single and double excitations, J","cited_arxiv_id":null,"evidence_quote":"Shows how additional Tucker decompositions of intermediate tensors reduce the ladder terms to $O(N^5)$ scaling, the key route to sub-$N^6$ cost."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the tensor-hypercontraction variant and error-compensation observations used to justify looser thresholds for relative energies."},{"cited_title":"Frozen natural spinors for Cholesky decomposition based two-component relativistic coupled cluster method","cited_arxiv_id":"2412.18395","evidence_quote":"Gives the antisymmetrized tensor factorization for relativistic integrals that the paper considers and rejects for amplitudes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the compound-tunable embedding potential model of YbCl2 used as the three-dimensional benchmark system."}],"review_version":1}