{"id":"609c7b16-d09e-46e8-a32f-8a86e9f58f49","arxiv_id":"2608.10658","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SVD-xTC-CCSD(T) compresses the virtual space of a large-basis transcorrelated calculation via singular value decomposition and reaches chemical accuracy on the G2-1 atomization energy benchmark with triple-zeta quality bases.","lead":"Researchers introduce SVD-xTC, a way to run transcorrelated quantum chemistry calculations in a large basis while doing the expensive correlation step in a much smaller virtual space. It reaches about 0.5 kcal/mol average error on atomization energies for 55 benchmark molecules, matching near-exact references at a fraction of the cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SVD compression's safety rests on the empirical sharpness of the singular-value cliff, which is only demonstrated for one molecule; if the discarded tail is not negligible for some system, the compressed virtual space is not the full small-basis virtual space and the central accuracy claim…","rationale":"The reader's weakest assumption identifies the SVD subspace retention as load-bearing, and that is the right target. I agree with that identification and sharpen it: the truncation is exact only under ideal nesting of basis sets, and for realistic non-nested basis sets the discarded singular values are finite. The paper's only displayed spectrum shows a clear cliff, but the assertion that this holds throughout the G2-1 set is not backed by per-molecule data in the main text. Since the MAE benchmarks could hide a small systematic loss arising from the discarded tail, a direct spectral and ablation check is the decisive test. The paper has real independent support: the construction is mathematically clean, the benchmark is extensive, the timings are reported, and the per-molecule FCIQMC checks for outliers are a good-faith diagnostic. The concern here is addressable and does not by itself overturn the conditional verdict; it strengthens the conditions under which the method should be accepted. Hence I recommend no change to the reader's CONDITIONAL verdict, with the concrete spectral test as a requested revision.","tokens_in":16914,"tokens_out":15377,"duration_ms":169709,"concrete_test":"For all 55 G2-1 molecules, compute from Eq. (9) the largest discarded singular value lambda_max = max_{k>n_sel} sigma_k and report its distribution. Then take the molecule with the largest lambda_max and rerun SVD-xTC-CCSD(T) at A VTZ retaining n_sel+1, n_sel+2, and n_sel+5 singular vectors; if any atomization energy shifts by more than 0.1 kcal/mol, or if any molecule has lambda_max > 0.05, the compression is not uniformly safe and the central claim must be restricted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the n_sel-dimensional SVD subspace from Eqs. (8)-(12) preserves the accuracy of the large-basis calculation while reducing bottleneck costs. The retention argument is exact only if the large-basis occupied space is exactly representable in the small-basis AO space, so that the last n_occ singular values are true zeros. With non-nested basis sets such as aug-cc-pVxZ versus aug-cc-pV5Z, the occupied orbitals have components outside the small AO span, and the discarded singular values are small but nonzero (e.g., 0.023 for Si2H6 in Fig. 1). The paper states that the sharp cliff is found throughout the G2-1 set, but it displays only this one spectrum. The benchmark MAEs could conceal a small systematic loss of correlation, especially if the cliff sharpness degrades for molecules with diffuse or highly polar bonding. If for any molecule the largest discarded singular value is not tiny, the retained space is no longer the full small-basis virtual space, and the claimed one-step equivalence between SVD-xTC and a large-reference/small-virtual calculation is broken.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SVD-xTC, a scheme that constructs a small virtual space out of a large-basis transcorrelated calculation by projecting large-basis virtual orbitals onto a small-basis AO space and truncating via SVD. The retained virtual space has dimension n_S - n_occ, matching the small basis, while the reference and occupied orbitals come from the large-basis SCF solution. The method is applied to atomization energies of the G2-1 set using xTC-CCSD(T) and xTC-PP-CCSD(T), with comparisons to SHCI+PBE+CV and experimental values. The authors report MAEs of 0.86 kcal/mol at AVTZ and 0.47 kcal/mol at AVQZ for all-electron SVD-xTC-CCSD(T), chemical-accuracy-level MAEs against experiment, faster basis-set convergence with ECPs, and wall-clock timings over the full G2-1 set.","tokens_in":17172,"tokens_out":11663,"duration_ms":124697,"significance":"If the central assumption holds, SVD-xTC is a valuable non-composite alternative to reference-corrected transcorrelated methods: it preserves a large-basis reference and compresses the post-HF calculation into a smaller virtual space, potentially reducing the bottleneck cost of TC-CCSD(T) while retaining near-basis-set-limit accuracy. The paper's strengths are its extensive 55-molecule benchmark, the absence of benchmark-fitted parameters (the SVD truncation is set by the singular-value structure and the Jastrow parameters are variance-optimized), the external comparison against SHCI+PBE+CV and experiment, and the practical timing data. The main risk is that the reliability of the method rests on an empirical claim about the singular-value spectrum that is only illustrated for one molecule, and the headline accuracy is an MAE statement rather than a per-molecule guarantee.","major_comments":[{"comment":"The correctness of the method rests on the assertion that the last n_occ singular values in Eq. (9) are negligible, so that the n_sel-dimensional space retained in Eq. (10) preserves essentially all correlation energy of the large-basis virtual space. For non-nested basis sets this is not an identity: the last singular values are small but nonzero (0.023 for Si2H6 in Fig. 1), and the paper states that the sharp cliff is found throughout G2-1 but shows only this single spectrum. Please report the distribution of the largest discarded singular value (or the ratio sigma_{n_sel+1}/sigma_{n_sel}) over all 55 molecules and basis pairs, and show per-molecule convergence of atomization energies with respect to retaining one or more extra SVD vectors. Without such evidence, the central claim that SVD-xTC reproduces the large-basis accuracy in all systems is not established.","section":"Section 2, Eqs. (8)-(12), Fig. 1"},{"comment":"The headline \"chemical accuracy already with triple-zeta basis sets\" is an MAE statement and should be qualified as such. At A VTZ the MAE is 0.86 kcal/mol but the MaxE is 4.48 kcal/mol; at A VQZ the MAE is 0.47 kcal/mol but the MaxE is 2.21 kcal/mol, with the outliers CN, O2, F2, Si2, P2, and CH3Cl discussed in Fig. 4. Even the per-molecule \"best estimate\" retains a MaxE of 3.81 kcal/mol in Table 1. Please report the fraction of molecules within 1 kcal/mol of the reference and either rephrase the abstract or justify why an MAE-level statement is the appropriate reading of \"chemical accuracy\".","section":"Section 4, Table 1 and Fig. 3"},{"comment":"The efficiency claim that \"TC integral calculation ... incur[s] the cost of only a small virtual space calculation\" is not directly supported. The compressed orbitals C_final in Eq. (12) are n_L x n_sel coefficient matrices expanded in the large-basis AO set, so the numerical evaluation of the xTC corrections still involves the large AO space; it is not obvious that the integration cost reduces to that of an n_S-basis calculation. Figure 8 reports only aggregate node-hours over the whole pipeline, without separating VMC, xTC integral evaluation, and CCSD(T). Please provide a complexity estimate for the xTC integral step as a function of n_L and n_sel, and stage-resolved timings, to substantiate the claimed bottleneck reduction.","section":"Section 3, Fig. 8 and Eq. (12)"}],"minor_comments":[{"comment":"The caption reads \"standard, RC- and CSV-xTC-CCSD(T)\"; \"CSV\" should be \"SVD\".","section":"Fig. 4 caption"},{"comment":"Notation for basis sets is inconsistent: the table and text use both \"avQZ\" and \"A VQZ\"; please standardize.","section":"Table 1 and throughout"},{"comment":"The outlier label \"CLF\" should be \"ClF\" for consistency with the molecular formulas used elsewhere.","section":"Fig. 6"},{"comment":"The \"best estimate\" row is a per-molecule, post-hoc selection of AVQZ, AV5Z, and FCIQMC results; it should not be presented as the prediction of a single method without an explicit caveat, although the paper does label it.","section":"Section 4, Fig. 4 and Table 1"},{"comment":"The SVD notation with V of dimension n_virt_L x n_S is a thin SVD; writing M = U Sigma V^T with these dimensions is fine, but a sentence clarifying that the full V is truncated to n_S columns would help readers.","section":"Section 2, Eq. (9)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a strong methodological benchmark with a clean construction and extensive numerical evidence. My main concern is that the empirical basis for the SVD retention assumption is currently limited to one illustrative spectrum; the requested distribution of discarded singular values and per-molecule convergence tests are essential before the central equivalence claim can be accepted. The stage-resolved timing question is also important because the paper's efficiency claim depends on the actual cost of the xTC integral step for orbitals expanded in the large-basis AO set. The paper is within scope for the journal and the work is likely publishable after these points are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about arXiv:2608.10658: the SVD-xTC construction is a real one-step replacement for the composite RC-xTC scheme, and the benchmark on the G2-1 set is extensive and honestly reported. The math is clean, the comparison against SHCI+PBE+CV and experiment is external, and nothing is fitted to the benchmark. I came away thinking this is a solid methods paper.\n\nWhat is actually new: Eqs. (8)-(12) define a compression of the large-basis virtual space by projecting onto the small-basis AO space via SVD and dropping the near-zero principal angles. That is distinct from the reference-correction composite in Ref. 32. The paper also makes a useful empirical point: SVD-xTC-CCSD(T) reaches MAE 0.86 kcal/mol at AVTZ and 0.47 at AVQZ against the SHCI reference, and the per-molecule plot in Fig. 4 does not hide systematic failures. The FCIQMC spot-checks on the negative outliers are a good-faith effort to separate basis-set error from correlation error. The ECP analysis with ccECPs, including the SR/CV bookkeeping, is careful and yields a concrete message: ECP-TC converges by AVTZ.\n\nSoft spots, in order of severity. First, the abstract's \"chemical accuracy with triple-zeta\" is an MAE statement, not a per-molecule guarantee. At AVTZ the MaxE is 4.48 kcal/mol, and even at AVQZ there are per-molecule errors above 2. The results are good, but the headline overstates what MAE alone means. The language should be tightened. Second, the robustness of the SVD compression rests on the sharp singular-value cliff, but the paper shows only one spectrum (Si2H6). The text says the cliff is found throughout the G2-1 set, but no figure or table supports that. A referee should ask for the singular value gaps for all 55 molecules, or at least for the worst cases. This is a completeness issue, not a demonstrated flaw: the tiny MAEs over a diverse set are themselves empirical evidence that the compression is working, and if the cliff were not sharp somewhere, the per-molecule errors would likely show it. Third, no code is released. The data are online as interactive figures, and the in-house codes are named, but PyTCHInt is not available. That limits independent replay. Again, this is standard practice in this subfield, but it is a real reproducibility gap.\n\nThe circularity burden is nil: the Jastrow parameters are variance-optimized in VMC, the SVD truncation is set by the number of orbitals, and the references are external. The stress-test worry about non-nested bases and the last n_occ singular values being nonzero is worth a sentence in the paper, but the benchmark already partially answers it. The paper holds up.\n\nWho gets value: quantum chemists using transcorrelated methods, method developers interested in basis-set reduction, and anyone benchmarking CCSD(T) on the G2 set. I would take it to reading group and cite it for the method. It deserves peer review; with the clarifications above it should be publishable in a strong topical journal.\n\nRecommendation: send it to a serious referee. The central claim is supported by the data; the requested changes are presentation and completeness, not a change in conclusion.","headline":"SVD-xTC is a genuine one-step alternative to RC-xTC, the G2-1 benchmark is extensive and honest, and the paper deserves a serious referee with minor requests for tighter accuracy language and more singular-value spectra.","tokens_in":17699,"tokens_out":4950,"would_cite":true,"duration_ms":51249,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By compressing the virtual orbital space with a singular value decomposition, transcorrelated CCSD(T) reaches chemical accuracy for the G2-1 atomization energies at triple-zeta basis cost.","keywords":["transcorrelation","singular value decomposition","virtual orbital compression","CCSD(T)","atomization energies","G2 benchmark set","basis-set convergence","pseudopotentials"],"falsifier":"Run SVD-xTC-CCSD(T) and the uncompressed xTC-CCSD(T) in the same large basis (for instance AVQZ) over the G2-1 set and examine their atomization-energy differences; any molecule where the difference exceeds about 1 kcal/mol would show that the compression discarded correlation energy that the large-basis calculation captures.","tokens_in":16746,"feed_emoji":"⚛️","tokens_out":7709,"duration_ms":70979,"temperature":0.7,"pith_summary":"Transcorrelated methods accelerate basis-set convergence by folding electron–electron cusp correlation into a Jastrow similarity transformation, but their reference energy still converges slowly with basis size. This paper proposes to keep the accurate large-basis reference orbitals while compressing the virtual space to the size of a small basis through a singular value decomposition (SVD) of the overlap between the two orbital spaces. The resulting workflow, SVD-xTC-CCSD(T), is a one-step, non-composite method that reproduces the accuracy of a large-basis calculation at the cost of a small virtual space. On the 55-molecule G2-1 benchmark, it reaches chemical accuracy in atomization energies already with triple-zeta basis sets, and it gives the lowest mean absolute error of the compared methods at quadruple-zeta level.","feed_headline":"SVD compression brings transcorrelated CCSD(T) to chemical accuracy","feed_subtitle":"Projecting large-basis virtuals onto a small-basis space keeps reference accuracy at a fraction of the cost.","key_machinery":"The machinery is the SVD projection matrix $M = S_{SS}^{-1/2} S_{SL} C_{\\text{virt}}^{L}$, where $S_{SS}$ is the small-basis overlap matrix, $S_{SL}$ is the cross-overlap between small- and large-basis atomic orbitals, and $C_{\\text{virt}}^{L}$ are the large-basis canonical virtual coefficients. Its singular values are cosines of the principal angles between the large-basis virtual space and the small-basis AO space; the smallest $n_{\\text{occ}}$ of them are near zero because large-basis virtuals are orthogonal to the large-basis occupied space, and this near-zero block is cleanly separated by a sharp cliff. Truncating the corresponding right singular vectors and rotating the remaining ones gives a compressed virtual space of dimension $n_S - n_{\\text{occ}}$ that is orthonormal in the large-basis metric, and a final pseudocanonicalization diagonalizes the projected Fock matrix. This construction yields a one-step TC workflow: it preserves the large-basis reference and occupied space, compresses only the virtuals, and thereby keeps the expensive TC integral and CCSD(T) steps in a small virtual space.","core_discovery":"The central discovery is that the large-basis virtual space can be replaced, without loss of accuracy, by an SVD-compressed subspace of dimension equal to the small-basis virtual space, provided the occupied reference orbitals are taken from the large-basis SCF solution. The compression is constructed by projecting the large-basis canonical virtuals onto the small-basis orthonormal atomic orbitals and discarding the right singular vectors whose singular values are near zero; these near-zero values correspond to the occupied-block deficiency of the small-basis AO space and are cleanly separated by a sharp singular-value cliff. Because the transcorrelated correlation energy converges faster than the reference energy, the small compressed virtual space carries essentially all the correlation while the large-basis reference supplies the slowly-converging part. On the G2-1 set, SVD-xTC-CCSD(T) achieves a mean absolute error of 0.86 kcal/mol at AVTZ and 0.47 kcal/mol at AVQZ against near-exact SHCI+PBE+CV reference atomization energies, outperforming standard and reference-corrected xTC-CCSD(T), and it maintains chemical accuracy when compared with experiment.","pith_inferences":["The sharp singular-value gap could be used as a diagnostic: if a molecule's spectrum lacks a clear cliff, the compression would be unreliable and the method should fall back to the full virtual space.","The same SVD construction could be applied to other post-Hartree–Fock solvers operating on transcorrelated Hamiltonians, such as DMRG or selected CI, whenever their cost scales steeply with the number of virtuals.","The method's reliance on occupied-virtual orthogonality suggests that for strongly multi-configurational references the active space must be separated before compression, an extension the paper notes but does not benchmark.","The faster basis-set convergence observed with pseudopotentials hints that a frozen-core all-electron TC variant might combine the all-electron accuracy with the ECP workflow's cheaper Jastrow optimization."],"forward_implications":["SVD-xTC-CCSD(T) reaches chemical accuracy (MAE 0.86 kcal/mol) for G2-1 atomization energies already at triple-zeta level, and the lowest MAE (0.47 kcal/mol) among tested methods at quadruple-zeta.","The one-step workflow eliminates the composite reference-correction step, requiring only a single large-basis SCF calculation while keeping the correlation solver in a small virtual space.","With pseudopotentials, SVD-xTC-PP-CCSD(T) converges by triple-zeta (MAE about 0.63–0.65 kcal/mol) and runs at only 10.5 node-hours for the full 55-molecule set, indicating that ECP-based TC is both accurate and inexpensive.","The SVD compression also removes the systematic positive outliers seen with standard and reference-corrected xTC for silicon-containing second-row molecules.","The results imply that CCSD(T), when combined with transcorrelation and SVD compression, is sufficient for chemically accurate atomization energies of first- and second-row molecules."],"supporting_citations":[{"why":"Establishes that the TC reference energy converges more slowly than the correlation energy, the motivation for the SVD compression.","marker":"32"},{"why":"Defines the xTC approximation used to reduce the three-body transcorrelated terms to two-body form.","marker":"8"},{"why":"Provides the transcorrelated pseudopotential treatment used in the ECP benchmark.","marker":"11"},{"why":"Supplies the near-exact SHCI+PBE+CV atomization energies used as the primary reference.","marker":"34"},{"why":"Defines the G2-1 (G2/55) benchmark set of molecules used for all comparisons.","marker":"33"},{"why":"Gives the variational Monte Carlo Jastrow-factor optimization procedure used in the workflow.","marker":"9"}],"fun_headline_variants":["SVD-shrunk virtuals give chemical accuracy in transcorrelated CCSD(T)","Small basis, large accuracy: SVD-compressed virtuals for TC-CCSD(T)","SVD compression achieves chemical accuracy in transcorrelated CCSD(T)","SVD cuts virtual space, transcorrelated CCSD(T) hits chemical accuracy","Compress virtuals via SVD: transcorrelated accuracy without the cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the SVD-truncated virtual space retains essentially all of the large-basis correlation energy; this is justified empirically by the sharp singular-value cliff and the benchmark results, but not proven analytically.","fun_headline_variants_meta":{"raw":{"variants":["SVD-shrunk virtuals give chemical accuracy in transcorrelated CCSD(T)","Small basis, large accuracy: SVD-compressed virtuals for TC-CCSD(T)","SVD compression achieves chemical accuracy in transcorrelated CCSD(T)","SVD cuts virtual space, transcorrelated CCSD(T) hits chemical accuracy","Compress virtuals via SVD: transcorrelated accuracy without the cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00098,"raw_usage":{"total_tokens":4247,"prompt_tokens":1115,"completion_tokens":3132,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":731,"completion_tokens_details":{"reasoning_tokens":3027}},"tokens_in":731,"tokens_out":3132,"duration_ms":27029,"temperature":1.0,"reasoning_tokens":3027,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:59:08.360992+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SVD-xTC-CCSD(T) and the uncompressed xTC-CCSD(T) in the same large basis (for instance AVQZ) over the G2-1 set and examine their atomization-energy differences; any molecule where the difference exceeds about 1 kcal/mol would show that the compression discarded correlation energy that the large-basis calculation captures.","supporting_citations":[{"cited_title":"The Journal of Chemical Physics , volume =","cited_arxiv_id":null,"evidence_quote":"Provides the transcorrelated pseudopotential treatment used in the ECP benchmark."},{"cited_title":", title =","cited_arxiv_id":null,"evidence_quote":"Gives the variational Monte Carlo Jastrow-factor optimization procedure used in the workflow."}],"review_version":1}