{"id":"a78ee091-c00e-4c4c-90dd-bfc4f09c5104","arxiv_id":"2603.25233","paper_version":3,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Rank-adaptive sweep-based SI-DSA with mild space augmentation matches full-rank accuracy and outer iterations while cutting memory and runtime for steady-state RTE even at 30–45% effective rank.","lead":"A new low-rank iterative solver for the radiative transfer equation adapts the solution rank by adding only a few basis vectors at a time, not a multiple of the current rank. This keeps memory and runtime savings even when the solution needs 30–45% of the full rank, a regime where older low-rank methods often lose their edge.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The reader correctly isolates the residual-based greedy sampler and dual stopping test as the algorithmic hinge that makes mild (rank-independent) augmentation work. The manuscript already supplies the natural stress tests: identical outer iteration counts across all four homogeneous regimes, lattice, variable-scattering (p,q) matrix, and eight random pin-cell seeds; oversampling ratios stay well below 15% even at 42% of full rank; and scalar-flux errors remain an order of magnitude below the outer tolerance. Because those checks already address the weakest assumption and no contradictory evidence appears, the ACCEPT verdict stands. The suggested concrete test is a modest further probe of stopping-criterion robustness rather than a requirement for acceptance.","tokens_in":26393,"tokens_out":434,"duration_ms":4255,"concrete_test":"Re-run the pin-cell configuration (Sec. 4.4) with the same mesh/quadrature but force the inner residual tolerance ε_res to 10^{-5} (one order looser than the default 10^{-7}) while keeping ε_diff=10^{-7}; if outer iteration count rises above the full-rank baseline of 21 or ||φ_LR−φ_FR||_2 exceeds 10^{-6}, the dual stopping criterion is more fragile than claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that mild residual-based greedy angular subsampling (p=1 from q=8 candidates) plus the dual residual/scalar-flux inner stopping test yields outer SI-DSA iteration counts and accuracy matching full-rank SI-DSA while still delivering memory/runtime savings at effective ranks of 30–45%—is supported by the reported experiments. The paper’s own ablations (lattice MGS2 vs MGS/QR, variable-scattering p/q sweeps, pin-cell multi-seed runs) already probe the reader’s weakest assumption and show no premature termination or outer stagnation. No internal inconsistency or untested regime that would overturn the claim is evident from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a rank-adaptive sweep-based SI–DSA solver for the first-order steady-state RTE. Its core is an inner-loop low-rank source iteration that enriches the spatial basis by a small, rank-independent number p of angular snapshots per iteration (default p=1), selected by residual-based greedy random subsampling from q candidates, with a single SVD truncation only after the inner loop converges. Incremental MGS2-RO basis/operator updates and a dual residual/scalar-flux stopping test support non-intrusive reuse of existing transport sweeps; DSA is applied in the outer loop. Systematic 2D2V SN–DG experiments (homogeneous regimes, lattice, multiscale variable scattering, pin-cell) report outer iteration counts matching full-rank SI–DSA, scalar-flux differences O(10^{-7}–10^{-8}) below ε_SI-SA=10^{-6}, compression ratios ~8–45%, and speedups from ~1.5× up to >12×, including cases whose effective rank is 30–45% of full rank.","tokens_in":26638,"tokens_out":1494,"duration_ms":32457,"significance":"The work addresses a genuine practical bottleneck: rank-proportional space augmentation (e.g., 2r or 4r) can erase low-rank gains once the effective rank reaches tens of percent of the full rank, a regime that appears in standard multiscale RTE benchmarks. Delivering the first rank-adaptive sweep-based iterative solver for the first-order formulation, while preserving classical transport sweeps and DSA and requiring only mild augmentation, is a clear and useful contribution. Strengths include non-intrusive design, explicit ablations (MGS2 vs MGS vs full QR; p/q sweeps; multi-seed pin-cell robustness), and direct comparison to the same full-rank SI–DSA baseline on identical discretizations. If the empirical picture holds under broader use, the method is a practical path to genuine memory and runtime savings for moderately high-rank RTE problems where aggressive rank adaptation fails.","major_comments":[{"comment":"Sec. 3.3 (eqs. 3.7–3.9) and Sec. 3.5: when the candidate residual test passes, the stopping procedure solves Galerkin systems for all remaining unsampled angles at cost O((N_Ω−q−r_k)r_k^3) before checking the scalar-flux difference. The text asserts this occurs only in the final few inner iterations, but Sec. 4 reports no counts of how often the full-angle projection is invoked, nor a wall-clock breakdown separating sweeps, incremental updates, candidate residuals, and full projections. In the motivating 30–45% rank regime this term is not a priori negligible; a short quantification (or timing table) is needed to substantiate that mild augmentation plus the dual stopping test still yields the claimed net speedups rather than shifting cost into the convergence monitor.","section":"Sec. 3.3, 3.5; Sec. 4 tables"},{"comment":"Sec. 3.2 and Remark 3.1: residual norms on a small random candidate set (default q=8, p=1) are treated as sufficiently informative for greedy enrichment. The p/q ablation (Tab. 4.3) and eight-seed pin-cell study (Tab. 4.4) support robustness on the reported tests, but the manuscript does not discuss or test a failure mode in which important angular structure is systematically under-sampled by the random candidates (e.g., highly anisotropic beams or very sparse angular features). A brief limitation statement and, if feasible, one more extreme angular configuration would make the weakest algorithmic assumption falsifiable rather than only empirically successful on the current suite.","section":"Sec. 3.2; Sec. 4.3–4.4"}],"minor_comments":[{"comment":"Title page and running headers show spaced-out words (e.g., “HIGHL Y”, “RADIA TIVE”, “AUGMENT A TION”), likely PDF extraction artifacts; ensure the production PDF has clean titles and section headings.","section":"Title / headers"},{"comment":"Alg. 3.1 line 3 uses “while k ≤ NIter” but the outer counter is n; align the loop variable with the surrounding text (n).","section":"Alg. 3.1"},{"comment":"Fig. 4.2 caption refers to “Sec. 4.3” for the lattice problem, which is Sec. 4.2; Fig. 4.6 caption similarly points to Sec. 4.3 for the pin-cell problem (Sec. 4.4). Correct cross-references.","section":"Fig. 4.2, Fig. 4.6 captions"},{"comment":"Notation: r_k is used both as the current spatial rank and, in places, as if it equals p k; state explicitly that r_k = r_0 + k p (or the actual recurrence) when reorthogonalization does not drop columns.","section":"Sec. 3.1–3.2"},{"comment":"Default hyperparameters (p=1, q=8, ε_res=ε_diff=10^{-7}, ε_SVD=10^{-8}, ε_MGS) are stated in Sec. 4 but not collected in one place; a short “default parameters” paragraph or table would help reproducibility.","section":"Sec. 4 intro"},{"comment":"The relation to concurrent collocation-based DLR [6] is discussed well in Sec. 3.6; a one-sentence clarification that no numerical head-to-head is attempted (different problem class / fixed vs adaptive rank) would prevent readers from expecting such a comparison in Sec. 4.","section":"Sec. 3.6 / Sec. 4"},{"comment":"Appx. B: the definitions of D^±_x, D^±_y are dense; a brief remark that they are the standard upwind DG streaming matrices would help readers who skip the appendix.","section":"Appx. B"}],"recommendation":"minor_revision","confidential_remarks":"The central empirical claim is well supported; the two major points are documentation/quantification gaps rather than contradictions. Concurrent arXiv:2601.18705 is fixed-rank and time-dependent, so the novelty claim for rank-adaptive sweep-based SI on the first-order steady problem is fair. Fit for a numerical-analysis / computational transport venue is good. I would not block acceptance over theory: this is a methods paper whose contribution is algorithmic and empirical."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is practical: Guo and Peng give a rank-adaptive, sweep-based SI-DSA for the first-order steady-state RTE that keeps outer iteration counts and accuracy essentially identical to full-rank SI-DSA while still cutting memory and runtime when the effective rank is 30–45% of full. That is the regime where rank-proportional enrichment (2r BUG-style, 4r in their earlier second-order work) usually kills the speedup.\n\nWhat is new is the inner-loop design. They add a fixed small number of angles (default p=1 from q=8 residual-greedy candidates) without truncation, update the spatial basis and projected operators incrementally with MGS2-RO, and truncate only once after the inner loop. Residual-based greedy angular subsampling plus a dual residual/scalar-flux stop (eqs. 3.7–3.9) replaces full angular residual evaluation. Transport sweeps and DSA stay untouched, so existing codes can be reused. This is distinct from the fixed-rank collocation DLR in the concurrent Haut et al. work and from their own second-order SI-DSA.\n\nThe numerics are the right ones: homogeneous regimes, lattice, multiscale variable scattering, pin-cell. Outer iterations match full-rank; scalar-flux differences sit at 1e-7–1e-8 below the 1e-6 outer tolerance; compression and speedups are reported; ablations cover MGS2 vs MGS vs full QR, p/q sweeps, and multi-seed randomness. Mild oversampling (often <10–15%) is verified. No formal convergence theory and no code, but the empirical claim is cleanly supported against the correct baseline.\n\nSoft spots are minor and already partly probed: hyperparameters are hand-chosen (though sensitivity is shown), and everything is 2D2V Cartesian with isotropic scattering. That does not undercut the central result.\n\nThis is for people who write or use production-style discrete-ordinates solvers and care about low-rank methods that still win at moderate ranks. It deserves a serious referee. I would engage with it.","headline":"Solid methods paper: first rank-adaptive sweep SI-DSA for first-order RTE that stays faster than full-rank even at 30–45% ranks via mild (constant-p) augmentation.","tokens_in":27236,"tokens_out":534,"would_cite":true,"duration_ms":7330,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M60","65F10","85A25"],"pacs":[],"model":"grok-4.5","headline":"A mild-augmentation rank-adaptive SI-DSA matches full-rank accuracy and outer iterations for the radiative transfer equation while cutting memory and runtime even when ranks hit 30–45% of full size.","keywords":["radiative transfer equation","source iteration","diffusion synthetic acceleration","low-rank approximation","rank adaptivity","transport sweep","mild space augmentation","discrete ordinates"],"falsifier":"A multiscale benchmark (e.g., the pin-cell or lattice problem) on which the low-rank outer residual stagnates or the scalar-flux error exceeds the full-rank SI-DSA error by more than an order of magnitude once the true rank exceeds roughly 40 percent of full size, while the same residual-greedy schedule and stopping tolerances are used.","tokens_in":27279,"feed_emoji":"☁️","tokens_out":939,"duration_ms":8375,"temperature":0.7,"pith_summary":"Discrete-ordinates solvers for the steady radiative transfer equation store and update an angular flux over a high-dimensional phase space, which becomes expensive. Low-rank approximations can cut that cost, but standard rank-adaptive schemes enlarge the basis by a multiple of the current rank before truncating; when the true rank is already tens of percent of full size, that overhead cancels the savings. This paper builds a source-iteration DSA solver that works directly on the first-order form and reuses ordinary transport sweeps. Inside each source-iteration step an inner loop adds only a fixed handful of angular snapshots chosen by residual-based greedy sampling, updates the spatial basis and projected operators incrementally, and truncates only once the inner loop has converged. Outer DSA acceleration is left unchanged. On homogeneous, lattice, variable-scattering and pin-cell problems the method recovers the same outer iteration counts and scalar-flux accuracy as full-rank SI-DSA, while the memory and runtime drop by factors of roughly two (or more than twelve in the pure-diffusion regime). The practical claim is that genuine speed-ups remain available even for multiscale media whose solution rank is no longer tiny.","feed_headline":"Mild rank adaptivity cuts RTE cost even at 45% rank","feed_subtitle":"Sweep-based SI-DSA matches full-rank outer iterations while still saving memory and time on multiscale media","key_machinery":"Mild space augmentation via residual-based greedy angular subsampling: each inner iteration solves transport sweeps for a fixed small number p of newly sampled angles, incrementally enlarges the spatial basis by those p vectors without truncation, and selects the next p angles from a random candidate pool by residual size; truncation occurs only after the inner residual and scalar-flux tests both pass.","core_discovery":"A rank-adaptive, sweep-based SI-DSA for the first-order steady RTE can be realized by mild (rank-independent) space augmentation inside an inner residual-driven loop; the resulting solver matches full-rank outer iteration counts and accuracy while still reducing memory and runtime when the effective rank reaches 30–45 percent of the full discrete rank.","pith_inferences":["The residual-greedy sampler may transfer to other kinetic equations that already possess fast directional sweeps (e.g., linear Boltzmann or neutron transport).","If the candidate-pool size q can be reduced further without loss of outer convergence, the method becomes still cheaper for very high angular resolution.","Coupling the same mild-augmentation idea with quasi-diffusion or data-driven synthetic accelerators could push the high-rank regime beyond 45 percent while preserving speed-up."],"forward_implications":["Existing transport-sweep and DSA codes can be reused with only non-intrusive wrappers that supply angular snapshots and projected operators.","Memory and runtime savings remain available for practical multiscale media whose solution ranks are 30–45 percent of the full discrete rank.","The same mild-augmentation SI can be substituted into implicit time-stepping of time-dependent RTE without redesigning the angular sampler.","Aggressive O(r) basis enrichment is unnecessary for rank-adaptive first-order transport solvers.","Compression ratios of roughly 40 percent already yield roughly 2\times wall-clock speed-ups under standard outer tolerances."],"fun_headline_variants":["Mild space augmentation keeps low-rank RTE gains at 45% rank","Rank-adaptive SI-DSA matches full-rank accuracy with less memory","Sweep-based mild rank adaptivity cuts RTE cost up to 45% rank","Inner residual loop enables efficient rank-adaptive RTE SI-DSA","Mild rank-independent augmentation trims RTE runtime at 30-45% rank"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That sampling only a handful of residual-largest angles from a small random candidate set, together with the dual residual-plus-scalar-flux stopping test, is enough to adapt rank correctly without premature inner termination or outer-loop stagnation.","fun_headline_variants_meta":{"raw":{"variants":["Mild space augmentation keeps low-rank RTE gains at 45% rank","Rank-adaptive SI-DSA matches full-rank accuracy with less memory","Sweep-based mild rank adaptivity cuts RTE cost up to 45% rank","Inner residual loop enables efficient rank-adaptive RTE SI-DSA","Mild rank-independent augmentation trims RTE runtime at 30-45% rank"]},"model":"grok-4.5","effort":"low","cost_usd":0.003848,"raw_usage":{"total_tokens":1174,"prompt_tokens":799,"num_sources_used":0,"completion_tokens":100,"cost_in_usd_ticks":38480000,"prompt_tokens_details":{"text_tokens":799,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":275,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":799,"tokens_out":100,"duration_ms":2797,"temperature":1.0,"reasoning_tokens":275,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T11:47:25.428853+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"A multiscale benchmark (e.g., the pin-cell or lattice problem) on which the low-rank outer residual stagnates or the scalar-flux error exceeds the full-rank SI-DSA error by more than an order of magnitude once the true rank exceeds roughly 40 percent of full size, while the same residual-greedy schedule and stopping tolerances are used.","supporting_citations":[],"review_version":2}