{"id":"685230d3-40dc-479f-abe9-7be2ce5178ec","arxiv_id":"1908.06043","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SISLICE combines density-of-states shift selection with k-means shift migration to make parallel shift-invert spectrum slicing communication-light and load-balanced for self-consistent field eigenvalue problems.","lead":"This paper introduces SISLICE, a parallel algorithm for computing many eigenvalues of the large symmetric matrices that appear in quantum chemistry simulations, by splitting the eigenvalue range into slices handled simultaneously on different processor groups. It picks slice boundaries from an estimate of the eigenvalue density and then moves them between iterations with k-means clustering, which keeps communication low.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation relies on counting negative diagonal entries of an LDL^T factor; with Bunch-Kaufman 2x2 pivots this does not give the inertia needed for n_exact, so the slice-validation count in Sec. 2.2 can be wrong.","rationale":"The reader correctly identified the midpoint heuristic as an unproven empirical premise. I agree that premise is worth testing, but the method also has a failure-detection mechanism for missing eigenvalues that partially mitigates it. The inertia-count issue is more load-bearing because it attacks the exactness of the count used by every validation branch. The paper explicitly says Bunch-Kaufman and explicitly counts negative diagonal elements; this is a formal mismatch, not a matter of tuning or additional evidence. I still do not move the verdict: the scheme is plausible and the bug is localized and fixable, so conditional acceptance with a required correction, plus a test showing no 2x2-pivot miscounts in the reported runs, is appropriate. The numerical scaling comparisons, the use of pregenerated matrix sequences, and the PFEAST comparison give the paper partial independent support, but they do not exercise the flawed counting rule in a way that would expose it.","tokens_in":29973,"tokens_out":10950,"duration_ms":110280,"concrete_test":"Construct a small symmetric indefinite matrix whose Bunch-Kaufman factorization has a 2x2 pivot, e.g. A=[[0,1,0],[1,0,0],[0,0,1]] with σ=0, and run the same LDL^T routine used in SISLICE. Compare the number of negative diagonal entries of D with the inertia from a full eigendecomposition; if they differ, n_exact as defined in Sec. 2.2 is wrong for this pivot. Then, for all probe shifts in the Silane, Graphene, and Ga10As10 experiments, recompute n_exact by summing the inertia of each 1x1/2x2 block of D and compare with the validation decisions reported in Sec. 5; any change in a slice count would show that the published validation is unreliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The robustness claim in Sec. 5 and the conclusion depends on exact slice validation, and the paper's stated validation count is not valid for the factorization it specifies. In Sec. 2.2, n_exact for a slice is computed as the difference of 'the number of negative diagonal elements of D_{j+1} and D_j', where D is the factor from an LDL^T factorization. Sec. 2.1 identifies this factorization as Bunch-Kaufman. Bunch-Kaufman LDL^T represents D as a block-diagonal matrix with 1x1 and 2x2 pivots. The inertia of a 2x2 pivot block is not determined by the signs of the two diagonal entries; e.g. [[0,1],[1,0]] has two zero diagonal entries but inertia (1,1), one positive and one negative eigenvalue. A shifted pencil A-σB is indefinite and can produce such pivots, so the reported n_exact can be wrong. Since the validation logic either accepts candidates (n_exact=n_cand), triggers shift insertion (n_exact>n_cand), or removes candidates as spurious (n_exact<n_cand), an incorrect n_exact directly undermines the claimed robustness. No proof or experiment is given that all pivots were 1x1 in the reported runs; PARDISO and LAPACK commonly use 2x2 pivots for indefinite systems. Correcting the count to sum the inertias of the 1x1 and 2x2 blocks would likely fix this, so the algorithm is probably salvageable, but as written the central validation step is not exact.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SISLICE, a parallel shift-invert spectrum slicing eigensolver for symmetric generalized eigenproblems arising in self-consistent field (SCF) calculations. The method selects all spectral shifts simultaneously using a Lanczos-based density-of-states (DOS) estimate, refines and migrates shifts between SCF iterations via k-means clustering of validated Ritz values, validates spectral slices by comparing candidate counts with an inertia-based eigenvalue count from LDL^T factorizations, and reuses previously computed eigenvectors as initial guesses for the shift-invert subspace iteration. Numerical experiments on Silane (N=1109), Graphene (N=9360), and the Ga10As10 matrix (N=113,081) demonstrate convergence of the SCF sequences and strong scaling of the method, with a performance comparison to PFEAST.","tokens_in":30278,"tokens_out":7054,"duration_ms":68164,"significance":"If the central claims hold, SISLICE is a useful contribution: it attacks a real bottleneck in SCF eigensolvers, avoids communication-intensive orthogonality checks between slices, and shows good strong scaling on a large sparse problem. The paper's strengths include clearly specified algorithms and pseudocode, experiments on two qualitatively different spectra (clustered and uniform) plus a large sparse matrix, a sensible comparison with PFEAST, and an honest accounting of limitations such as the cost of rare probe insertions and the lack of integration into production SCF codes. However, the correctness of the slice-validation step is load-bearing and is not established as written, and several performance claims go beyond what the experiments actually compare against.","major_comments":[{"comment":"The exact slice count n_exact is defined as the difference of the numbers of negative diagonal entries of the LDL^T factors D_j and D_{j+1}, and this count drives all three validation outcomes. However, Sec. 2.1 specifies the Bunch-Kaufman LDL^T factorization, whose D factor contains 2x2 blocks, and for such blocks the signs of the two diagonal entries do not determine the inertia. For example, the block [[0,1],[1,0]] has two zero diagonal entries but inertia (1,1). Since A - sigma B is indefinite, 2x2 pivots can occur in the reported experiments, and both LAPACK and PARDISO routinely use such pivots for indefinite systems. An incorrect n_exact can therefore accept spurious candidates, fail to detect missing eigenvalues, or discard genuine eigenpairs, so the central robustness claim is not supported as written. The fix is to sum the inertias of the 1x1 and 2x2 diagonal blocks (or to use a factorization with only 1x1 pivots) and to verify in the experiments which pivot structure actually occurred.","section":"Sec. 2.2 (validation count; n_exact)"},{"comment":"The abstract and conclusion claim that SISLICE 'drastically reduces' the communication requirement of current parallel shift-invert eigenvalue algorithms, but the only quantitative comparison in the paper is to PFEAST. The closest prior shift-invert spectrum slicing methods, SIPs and SIESTA-SIPs, are described in Sec. 1 but are never benchmarked. Without a comparison to these methods, the communication and scalability advantage over the specifically targeted class of algorithms is not established. Please add such a comparison or revise the claims to refer only to the PFEAST comparison.","section":"Sec. 5.5 (Figs. 16a and 16c)"},{"comment":"The validation logic also assumes that Ritz values closer to a shift converge faster, so the midpoint rule assigns each eigenpair to the correct probe. The paper states that this is demonstrated numerically, but the experiments in Sec. 5 do not provide a direct quantitative check, such as a plot of residual norm as a function of distance to the nearest shift across many slices. If this heuristic fails for some eigenvalue distributions, n_cand will be wrong and the method may miss or duplicate eigenpairs. Since the robustness claim depends on exact validation, please provide a quantitative study of this assumption or explicitly state its failure modes as a limitation.","section":"Sec. 2.2 (candidate selection; Sec. 5 robustness)"}],"minor_comments":[{"comment":"The timing data appear to be single runs without repeated measurements or error bars; adding a few repetitions would make the strong-scaling claim more robust.","section":"Sec. 5.5, Fig. 16"},{"comment":"Equation (4) contains a corrupted font token ('/afii10069.italj') in place of the first Lanczos vector component; the notation should be cleaned up.","section":"Eq. (4)"},{"comment":"The symbol for the exact count is typeset inconsistently as 'nex act j'; it should be a single symbol such as n_exact^{(j)}.","section":"Sec. 2.2"},{"comment":"The x-axis label of Fig. 16c is ambiguous: it should state clearly whether the horizontal axis is the number of shifts per MPI rank or the total processor count.","section":"Fig. 16c"},{"comment":"The text says that a 10-fold increase in basis dimension has a negligible effect on timing, but Fig. 15a shows a visible increase; please quantify the percentage increase and discuss its relevance.","section":"Sec. 5.4, Fig. 15a"},{"comment":"The claim that PFEAST 'requires considerable tuning' is only partially supported, since the experiments use the default elliptic contour and the same DOS-based intervals; consider tempering this statement.","section":"Sec. 1 and Sec. 5.5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for ACM TOMS and the algorithmic ideas are clearly described. The main obstacle is the inertia-counting flaw in the validation step, which is load-bearing for the robustness claims but appears fixable. The absence of a comparison against SIPs/SIESTA-SIPs should also be addressed in revision, as the abstract's central communication claim is currently benchmarked only against PFEAST."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, SISLICE is a sensible integration of known pieces—DOS-based shift selection, k-means shift migration, and communication-light validation—and the authors are honest about what is heuristic. Second, the paper as written has a technical error in its validation step: it counts negative diagonal entries of the LDL^T factor from a Bunch-Kaufman factorization, but with 2x2 pivot blocks that does not give the inertia. The stress-test note is right. So the \"exact\" count that drives accept/insert/remove logic is not exact as specified.\n\nWhat is genuinely useful here is the specific integration. The simultaneous allocation of all shifts from a DOS estimate, the cluster-to-probe mapping that lets shifts migrate while reusing Ritz vectors from the previous SCF iteration, and the avoidance of cross-slice vector communication by relying on inertia counts and residual norms are all coherent design choices. The experiments cover two qualitatively different spectra (clustered Silane, uniform Graphene) and a large sparse matrix, with strong scaling out to one shift per MPI rank and a fair comparison against PFEAST. The paper is clearly written, and the authors explicitly show the cost of probe insertion rather than hiding it.\n\nThe soft spots are real but mostly proportionate. The inertia bug is load-bearing and needs to be fixed; the fix is straightforward, by summing the inertia of the 1x1 and 2x2 blocks in D or by using an explicit inertia routine, so the algorithm is likely salvageable. The midpoint-rule partitioning of validation candidates rests on the empirical claim that Ritz values closer to a shift converge faster; it is demonstrated only on the test cases and not proven. There is no shipped code or data, and no direct comparison to SIPs or SIESTA-SIPs, so the stronger communication-reduction claims rely on the PFEAST comparison alone. Minor, but worth noting: no repeated runs or error bars.\n\nWho is this for? People working on parallel eigenvalue algorithms for SCF and on spectrum slicing more generally. It deserves a serious referee and a request for revision rather than a desk rejection: the core idea is sound, the presentation is honest, and the main flaw is a correctable technical detail in the validation counting.","headline":"A useful algorithmic integration for parallel shift-invert slicing, but the validation step as written miscounts inertia under Bunch-Kaufman factorizations and needs correction before the robustness claim holds.","tokens_in":30874,"tokens_out":2230,"would_cite":true,"duration_ms":24161,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F15","65Y05","65F50"],"pacs":[],"model":"deepseek-v4-flash","headline":"The SISLICE method claims that symmetric self-consistent eigenvalue problems can be solved by shift-invert spectrum slicing with all spectral shifts chosen at once from a density-of-states estimate, shift positions tracked by k-means…","keywords":["shift-invert spectrum slicing","self-consistent field eigenvalue problem","density of states estimation","k-means shift migration","symmetric generalized eigenvalue problem","parallel eigensolver","Sylvester inertia theorem","subspace iteration"],"falsifier":"Run SISLICE on a symmetric pencil whose spectrum contains a tight pair of eigenvalues straddling the midpoint between two shifts, with the pair separated by less than the Ritz error of the probes; if the validated set for the slice does not match the exact inertia count, or if the smaller-residual candidate is the eigenvalue that is actually farther from its probe's shift, the midpoint partition has failed.","tokens_in":29695,"feed_emoji":"🧮","tokens_out":8987,"duration_ms":83671,"temperature":0.7,"pith_summary":"This paper introduces SISLICE, a parallel algorithm for computing a large contiguous block of eigenpairs of a symmetric generalized eigenvalue problem that must be solved repeatedly as the matrix changes, as in self-consistent field (SCF) electronic structure calculations. The central claim is that the two bottlenecks of earlier shift-invert spectrum slicing, sequential shift placement and the need to communicate eigenvectors to validate slices, can both be removed: all shifts are chosen simultaneously from a Lanczos-based density-of-states estimate, and slice validation is done with exact inertia counts plus residual norms rather than cross-node orthogonality checks. Shift locations are then migrated between SCF iterations by k-means clustering of validated Ritz values, letting the solver track a drifting spectrum without recomputing the density of states. On two representative density-functional test cases the paper reports convergence in a few SCF iterations and linear strong scaling down to one shift per MPI rank, with measured scaling limited only by factorization cost. If correct, this makes spectrum slicing a practical black-box route to interior eigenpairs of large sparse symmetric pencils on thousands of processors.","feed_headline":"Spectrum-slicing eigensolver scales linearly on many cores","feed_subtitle":"One density-of-states pass fixes all slices up front; k-means tracks eigenvalue drift across SCF iterations.","key_machinery":"The load-bearing objects are the spectral probe and the validated spectral slice. A spectral probe SP($\\sigma$) runs shift-invert subspace iteration at shift $\\sigma$ and returns approximate eigenpairs (Ritz pairs); the slice between adjacent shifts is validated by taking, from each bounding probe, only Ritz values on that probe's half of the slice, split at the midpoint, an assignment justified by the assumption that values closer to a shift converge faster. The exact eigenvalue count for the slice comes from the difference in negative diagonal entries of the LDL^T factors at the two shifts (Sylvester's inertia theorem), and when candidates outnumber the exact count the excess is discarded by residual norm. The machinery also includes the Lanczos-Gaussian density-of-states estimate, with cluster refinement by recursive local-maximum search, and the k-means/k-means++ clustering that computes new shift positions and drives probe insertion and deletion while keeping the number of probes fixed. These pieces together let the method localize all communication to scalar Ritz values and residuals, never eigenvectors.","core_discovery":"The paper's central discovery is a shift-selection and validation protocol that turns shift-invert spectrum slicing into a communication-light, load-balanced parallel method. For the first SCF iteration, a short Lanczos run produces an approximate density of states; the spectral region is partitioned into slices either by solving cumulative-density root equations for equal-count intervals or by locating local maxima to isolate tight clusters, placing one shift per slice. In later iterations the validated Ritz values from the previous iteration are clustered with k-means (seeded by k-means++), and each cluster's centroid becomes the new shift of the probe with which it overlaps most; probes left without a cluster are deleted and new probes inserted into the largest clusters to keep the probe count fixed. Missing eigenvalues are caught because the LDL^T factorization at each shift yields an exact eigenvalue count by Sylvester's inertia theorem, and residual norms, not inter-node vector exchange, select the validated eigenpairs. Numerical experiments on clustered (silane) and uniform (graphene) spectra show convergence of validated residuals below machine tolerance within 2-6 SCF iterations, with the k-means migration reducing iteration counts in irregular regions and never degrading uniform ones; strong-scaling tests to 1280 eigenvalues of a 113,081-dimensional matrix show linear scaling to one shift per MPI rank and a two-fold speed advantage over a contour-integral solver because SISLICE works in real arithmetic.","pith_inferences":["Beyond the paper, the same two-stage protocol, one-shot density-of-states partitioning followed by cluster-centroid tracking, could be dropped into other spectrum-slicing eigensolvers, such as contour-integral or polynomial-filtering methods, whenever those methods' inner iterations would benefit from better shift or interval placement.","The midpoint validation heuristic is only as safe as the convergence-rate assumption; a natural testable extension is an adaptive validation boundary that tightens near dense clusters or uses estimated local eigenvalue density to set a confidence interval around the midpoint.","The trace-difference monitor for detecting spectral change could be generalized to other cheap observables, such as subspace angles or partial traces of B^{-1}A, to trigger density-of-states recomputation only when the spectrum actually moves, reducing the risk of missing eigenvalues in early SCF iterations.","Since the method never exchanges Ritz vectors across nodes, its communication pattern is well matched to accelerator-heavy machines where vector movement is the dominant cost, suggesting the reported linear scaling could persist on GPU clusters with replicated matrices."],"forward_implications":["Large sparse symmetric generalized eigenproblems with dimension well above 100,000 can be partially diagonalized by factorization plus subspace iteration, avoiding the Rayleigh-Ritz bottleneck of Krylov methods and the dense diagonalization bottleneck of full eigensolvers.","In an SCF loop, eigenvectors from the previous iteration can be reused as initial guesses, so the cost of each successive partial diagonalization drops as the sequence converges.","The method's communication is limited to Ritz values and residual norms, so scaling is bounded by the cost of sparse LDL^T factorizations per shift, not by data movement, down to one shift per MPI rank.","The k-means shift migration can correct initial density-of-states misplacements, including deleting probes in empty spectral regions and splitting overloaded clusters, without ever exceeding a fixed probe count.","Because all shifts are chosen before the first subspace iteration, the initial slicing can be load-balanced across processor groups in a single pass."],"supporting_citations":[{"why":"Supplies the Lanczos-based Gaussian density-of-states approximation used to select the initial set of spectral shifts.","marker":"Lin et al. 2016"},{"why":"Provides the generalized-eigenvalue spectral density estimation technique that the density-of-states shift selection builds on for the pencil (A, B).","marker":"Xi et al. 2018"},{"why":"Defines the contour-integral solver used as the performance baseline for the numerical comparisons.","marker":"Kestyn et al. 2016"},{"why":"Introduces the parallel shift-invert spectrum-slicing framework whose sequential shift selection SISLICE replaces.","marker":"Zhang et al. 2007"},{"why":"Describes the closest prior spectrum-tracking variant with k-means-based migration, which SISLICE modifies by keeping probe count fixed and reusing eigenvector information.","marker":"Keçeli et al. 2018"},{"why":"Supplies the inertia theorem that gives the exact eigenvalue count in every slice from LDL^T factor diagonal signs.","marker":"Sylvester 1852"},{"why":"Provides the k-means++ seeding used to avoid poor-quality cluster centroids in the shift migration step.","marker":"Arthur and Vassilvitskii 2007"},{"why":"Documents the sequential dynamic shift-selection strategy for shift-invert Lanczos that motivates the all-at-once density-of-states shift selection.","marker":"Grimes et al. 1994"}],"fun_headline_variants":["SISLICE: one pass to set shifts, then parallel slice","SISLICE: communication-light spectrum slicing for parallel eigensolvers","Shift selection via density of states and k-means for fast eigensolvers","SISLICE: real-arithmetic spectrum slicing that scales to one shift per rank","One Lanczos run sets all shifts: SISLICE cuts eigensolver communication"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a Ritz value computed from the shift closer to the eigenvalue it approximates is always the more accurate one, so that the midpoint rule assigns every eigenvalue to exactly one slice and residual-based selection keeps the right eigenpair.","fun_headline_variants_meta":{"raw":{"variants":["SISLICE: one pass to set shifts, then parallel slice","SISLICE: communication-light spectrum slicing for parallel eigensolvers","Shift selection via density of states and k-means for fast eigensolvers","SISLICE: real-arithmetic spectrum slicing that scales to one shift per rank","One Lanczos run sets all shifts: SISLICE cuts eigensolver communication"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001541,"raw_usage":{"total_tokens":6214,"prompt_tokens":1047,"completion_tokens":5167,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":5061}},"tokens_in":663,"tokens_out":5167,"duration_ms":31640,"temperature":1.0,"reasoning_tokens":5061,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:57:39.801206+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SISLICE on a symmetric pencil whose spectrum contains a tight pair of eigenvalues straddling the midpoint between two shifts, with the pair separated by less than the Ritz error of the probes; if the validated set for the slice does not match the exact inertia count, or if the smaller-residual candidate is the eigenvalue that is actually farther from its probe's shift, the midpoint partition has failed.","supporting_citations":[],"review_version":1}