{"id":"560445a1-1b0d-451e-91e4-252c6817d02a","arxiv_id":"2509.07179","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A single auxiliary space framework yields exact error-propagation and condition-number identities for iterative methods on semidefinite systems, unifying and generalizing prior SPD and semidefinite analyses.","lead":"This paper builds a unified mathematical framework, called auxiliary space theory, for analyzing iterative solvers of linear systems whose matrices may be semidefinite. It gives exact formulas for the error reduction and condition numbers of such solvers, and uses the framework to re-derive convergence results for several classic preconditioning and multigrid methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sharp identities require the factorization B = Π B_e Π^t, but the paper gives no converse or completeness criterion; the advertised 'unified framework' is narrower than the abstract implies.","rationale":"The paper's central theorems are mathematically sound: I checked the reductions in Theorems 3.5–3.8, 4.4–4.7, and 5.1–5.4, including the delicate compression step in Theorem 3.7 where W_f ⊃ R(A_e) ensures W_f^⊥ ⊂ N(Π), and found no gap. Small examples with singular A confirm the semidefinite identities. The applications (PSC/SSC, Hiptmair–Xu, auxiliary grid) fit the framework and reproduce known results. The reader's weakest assumption names the factorization B = Π B_e Π^t as the load-bearing premise; I agree. The absence of any converse or completeness statement means the auxiliary-space sharp identities are conditional on a structural condition that is not characterized. This is a legitimate scope concern about the central claim, though it does not rise to a correctness error. The secondary issue—the abstract's 'refined analyses' claim—is supported only by proof-technique novelty, not by quantitative improvements or numerical comparisons, so it should be tempered. Since the reader's CONDITIONAL verdict already asks for such clarification, I recommend keeping the verdict unchanged.","tokens_in":21693,"tokens_out":48379,"duration_ms":374847,"concrete_test":"Take a standard iterative method not built as a subspace correction—for example, the symmetric Gauss–Seidel (SSOR) preconditioner on the semidefinite model system A = diag(1,0), B = (D+L)(D+U)-type operator—and attempt to exhibit a finite-dimensional space V_e, a surjective Π, and an SPD B_e satisfying B = Π B_e Π^t (Eq. 3.2b). If no such factorization can be constructed or characterized, add an explicit statement that Theorems 5.1–5.4 apply only to methods admitting this auxiliary-space structure, and soften the 'unified framework' phrasing in the abstract. Alternatively, for the Hiptmair–Xu and auxiliary-grid applications, recompute the constants in Theorems 7.1 and 8.1 and compare them with the sharp constants from [15] and [36]; if the bounds are identical up to the same O(1) factors, the word 'refined' should be replaced by 'unified' to avoid overclaiming.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorems 5.1–5.4 all assume that the iteration operator B can be written as B = Π B_e Π^t with surjective Π and with B_e (or its symmetrization) SPD on a space containing R(A_e). This factorization, introduced in Eq. (3.2b), is the genuine load-bearing premise of the auxiliary-space part of the paper. The theorems are proven correctly and the identities are exact when the premise holds, but the paper never characterizes which methods admit such a factorization or proves a converse. The general semidefinite results in Section 4 (Theorems 4.4–4.7) do not need the factorization and therefore cover arbitrary B, but the auxiliary-space sharp identities—which are the paper's main advertised contribution—silently exclude any method whose B cannot be represented in this way. The three applications are all constructed so that the factorization holds by design, so they do not test the generality of the framework. The abstract's 'unified framework' and 'generalize and refine' claims are consequently stronger than what is established: the theory is a powerful sufficient framework, not a complete analysis of all iterative methods of the form (1.2). This is a scope limitation rather than an internal inconsistency, and it does not invalidate any theorem, but it should be stated explicitly for the paper's central claim to be accurately framed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops an auxiliary-space framework for analyzing iterative methods of the form u_{m+1}=u_m+B(f-Au_m) for symmetric positive semidefinite linear systems. The central object is a factorization B=Π B_e Π^t through an auxiliary space V_e; under SPD assumptions on B_e or its symmetrized version, the authors derive exact identities for the error-propagation seminorm |I-BA|_A and for the extremal eigenvalues of BA, with the semidefinite case involving an additional infimum over N(A). Sections 3–5 build the abstract theory as a sequence of corollaries of the SPD auxiliary-space theorem; Section 4 recovers the semidefinite theory for general B by a canonical choice of auxiliary space; Sections 6–8 apply the framework to parallel and successive subspace correction methods, Hiptmair–Xu preconditioners, and auxiliary grid methods, yielding the Xu–Zikatanov identity and uniform condition-number bounds as special cases.","tokens_in":21945,"tokens_out":20718,"duration_ms":166098,"significance":"If the results hold—and the derivations appear sound—the paper provides a clean, elementary unification of several sharp convergence analyses that previously required separate arguments. Its main strengths are the exact identities rather than estimates with undetermined constants, the careful reduction of the semidefinite theory to the SPD case, and the uniform derivation of known results such as the Xu–Zikatanov identity and Hiptmair–Xu optimality. The treatment in Section 6 avoids pseudoinverses, which is a genuine pedagogical advantage. The formal contribution is a useful organizing theorem rather than a fundamentally new class of methods, but the unified viewpoint and the extension to semidefinite problems justify publication after minor revision.","major_comments":[],"minor_comments":[{"comment":"The statement that the framework is 'unified' should be qualified with one sentence noting that every B admits the trivial factorization V_e=V, Π=I, B_e=B, while the sharp auxiliary-space identities are most useful when a factorization with a simple, SPD B_e on a larger space is available; this would address the otherwise unstated completeness question for the factorization assumption.","section":"Abstract and §5 (Eq. (3.2b))"},{"comment":"The proof of Theorem 5.2 is omitted with the sentence 'the proof is omitted, as it follow from an argument analogous to that of Theorem 5.1'; please supply a brief proof or expand the remark to say explicitly that the same substitution used in Theorem 5.1 applies with B_e in place of \\bar{B}_e and with Theorem 3.6 replacing Theorem 3.5.","section":"§5, Theorem 5.2"},{"comment":"The sentence 'We also present counterparts of Theorems 3.7 and 3.8 for the semi-SPD case, given in Theorems 4.7 and 5.3, respectively' misstates the pairing: Theorem 4.6 and Theorem 5.3 correspond to Theorem 3.7, while Theorem 4.7 and Theorem 5.4 correspond to Theorem 3.8.","section":"§5, paragraph before Theorem 5.3"},{"comment":"The proof of Theorem 7.3 is omitted with the explanation that it is 'nearly identical' to that of Theorem 7.1; since H(div) is one of the three advertised applications, please include at least a proof sketch indicating how the discrete regular decomposition (7.5) and Theorem 5.2 are used in both the ϵ>0 and ϵ=0 cases.","section":"§7.3, Theorem 7.3"},{"comment":"'the conditioner number κ(BA)' should read 'the condition number κ(BA)'.","section":"§8, after Eq. (8.3)"},{"comment":"In the verification of (6.16c), the phrase 'where the last inequality is due to (6.15)' should say 'where the last equality is due to (6.15)'.","section":"§6, proof of Theorem 6.7"}],"recommendation":"minor_revision","confidential_remarks":"The paper is mathematically sound and well organized. The only substantive reservations are about framing: the factorization assumption is stated but not discussed at the level of completeness, and a few proofs are omitted. Both are easily fixed by revision. The novelty is moderate—much of the content reorganizes and extends known results—but the unified semidefinite treatment and the clean derivation of the Xu–Zikatanov identity are valuable enough for a standard numerical-analysis journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid theory paper and the central identities survive a close read. The genuinely new piece is the semi-SPD auxiliary space analysis with exact identities (Theorems 5.1–5.4), derived by reducing to the existing SPD theorem in [38] instead of dragging in pseudoinverses. The infimum over N(A) is a clean way to handle the singularity, and the proof of Theorem 5.1 in particular is tidy. The unification claim is legitimate: PSC, SSC, Hiptmair–Xu, and auxiliary grid methods all appear as instances of the same framework, and the paper is honest about what it inherits from [26,34].\n\nThe soft spots are in the framing rather than the mathematics. The sharp auxiliary-space theorems all assume B = Π B_e Π^t with surjective Π; without it they do not apply, and the paper gives no converse or completeness criterion. That makes the theory a powerful sufficient framework, not a general analysis of all methods of the form (1.2). The abstract should say so. Likewise, 'refined analyses' is only true in the sense of 'more elementary and unified'; the three applications recover known estimates rather than producing sharper constants or numerical comparisons. If that is all 'refined' means, fine, but say it. The omitted proofs of Theorem 5.2 and 7.3 are harmless; they are genuinely analogous. The citation pattern is appropriate, and the prior semidefinite work gets proper credit.\n\nBottom line: the math is sound, the scope is narrower than the abstract implies, and the fix is a revision, not a rejection. I'd send it to a serious referee.","headline":"Solid, honest theory paper: the new semi-SPD auxiliary space identities are correct and unifying, but the 'unified framework' claim overreaches the factorization assumption and the applications mostly repackage known results.","tokens_in":22501,"tokens_out":2625,"would_cite":true,"duration_ms":24997,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F08","65F10","65J05","65N55"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that for semidefinite linear systems, every iterative method of the form $u_{m+1}=u_m+B(f-Au_m)$ whose operator $B$ factors through an auxiliary space satisfies an exact, constant-free formula for its convergence rate and…","keywords":["auxiliary space theory","semidefinite linear systems","iterative methods","sharp convergence estimates","error propagation operator","condition number","subspace correction methods","Hiptmair–Xu preconditioners"],"falsifier":"Take $A=\\mathrm{diag}(1,0)$ on $\\mathbb{R}^2$, let $\\Pi:\\mathbb{R}\\to\\mathbb{R}^2$ be any nonzero column, and let $B_e$ be a positive scalar. Form $B=\\Pi B_e\\Pi^t$; compute $|I-BA|_A^2$ directly by maximizing $|(I-BA)v|_A^2$ over $v\\in R(A)$ with $|v|_A=1$, and compare it with the right-hand side of Theorem 5.1. Any mismatch would refute the identity, and repeating the test over random $\\Pi$ and $B_e$ would probe the factorization's sufficiency as well.","tokens_in":21469,"feed_emoji":"🧮","tokens_out":15257,"duration_ms":110371,"temperature":0.7,"pith_summary":"The paper sets out to prove that the convergence behavior of a broad class of iterative solvers for symmetric positive semidefinite linear systems is captured exactly by an auxiliary-space formula: any method of the form $u_{m+1}=u_m+B(f-Au_m)$ whose operator $B$ can be factored as $B=\\Pi B_e\\Pi^t$ through a surjective map $\\Pi$ has its error-propagation seminorm and, when $B$ is symmetric, its condition number given by explicitly computed quantities in the auxiliary space. The theory applies to singular systems without special-case tools, treating the semidefinite problem by viewing it as the auxiliary system of its own restriction to the range of $A$. A sympathetic reader would care because this replaces the usual sequence of problem-specific estimates with one uniform algebraic argument and yields sharp, constant-free identities for subspace correction methods, Hiptmair–Xu preconditioners, and auxiliary grid methods.","feed_headline":"Exact convergence formula for semidefinite linear-system solvers","feed_subtitle":"An auxiliary-space identity pins down the error rate with no undetermined constants, even for singular system matrices.","key_machinery":"The load-bearing object is the factorization identity $B=\\Pi B_e\\Pi^t$ together with the symmetrized auxiliary operator $\\bar B_e=B_e+B_e^t-B_e^tA_eB_e$, where $A_e=\\Pi^t A\\Pi$ and $\\Pi:V_e\\to V$ is surjective. The auxiliary space lemma, $(B^{-1}v,v)=\\inf_{\\Pi v_e=v}(B_e^{-1}v_e,v_e)$, converts inversion in the original space into an infimum over preimages in the auxiliary space; applying this to $\\bar B$ or to $B$ in the abstract symmetric-positive-definite identities of Theorem 2.3 yields the sharp formulas. The semidefinite case enters by choosing $V_e=V$ and $\\Pi=Q$, the orthogonal projection onto $R(A)$, making (1.1) the auxiliary system of the restricted SPD problem (4.1); the null-space infimum in the final formulas records the freedom in the kernel of $A$.","core_discovery":"The central discovery is an exact identity for the error propagation of the iteration. Let $A$ be symmetric positive semidefinite on a finite-dimensional space $V$, and suppose the iteration operator satisfies $B=\\Pi B_e\\Pi^t$ for a surjective $\\Pi:V_e\\to V$ and an auxiliary operator $B_e$ whose symmetrized form $\\bar B_e=B_e+B_e^t-B_e^tA_eB_e$, with $A_e=\\Pi^t A\\Pi$, is positive definite. Then the iteration (1.2) converges in the $A$-seminorm and $$|I-BA|$_A^{2}$=1-\\left(\\sup_{v\\in R(A),|v|_A=1}\\inf_{\\phi\\in N(A)}\\inf_{v_e\\in V_e,\\Pi v_e=v+\\phi}(\\bar $B_e^{{-1}}$v_e,v_e)\\right)^{-1}.$$ The same auxiliary-space mechanism supplies matching formulas for the smallest nonzero and largest eigenvalues of $BA$ when $B_e$ is positive definite, and a relaxed version in which positive definiteness is required only on a subspace containing $R(A_e)$. The semidefinite case arises by regarding the original singular system as the auxiliary system of its own restriction to $R(A)$; the extra infimum over $\\phi\\in N(A)$ is the trace of the singular kernel.","pith_inferences":["Not in the paper: the identity could be turned into a design principle — choose $V_e$, $\\Pi$, and $B_e$ to minimize the sup-inf expression, giving a variational route to optimize preconditioners rather than merely analyze them.","Not in the paper: because the formulas are algebraic and finite-dimensional, numerical evaluation of the right-hand side on small problems could certify a proposed factorization's convergence rate, possibly offering an a posteriori check for multigrid and domain decomposition methods.","Not in the paper: one might test whether the theory extends to infinite-dimensional Hilbert spaces when $B$ is a bounded operator and $B_e$ is coercive on $R(A_e)$; the finite-dimensional proof does not by itself cover that case."],"forward_implications":["For any iterative method of the form (1.2) whose $B$ admits the factorization, the convergence rate in the $A$-seminorm is known exactly from Theorem 5.1, with no undetermined constants.","The condition-number formulas of Theorems 5.2 and 5.4 make the analysis of preconditioned conjugate gradient methods on singular systems a matter of evaluating one auxiliary-space infimum.","The successive subspace correction method converges on semidefinite systems whenever each restricted local solver is convergent, and the sharp Xu–Zikatanov identity (6.16) holds in this setting.","The Hiptmair–Xu preconditioners for $H(\\mathrm{curl})$ and $H(\\mathrm{div})$ problems have uniformly bounded condition number by the same argument in both the definite case $\\epsilon>0$ and the semidefinite case $\\epsilon=0$.","An auxiliary grid preconditioner is condition-number equivalent to the preconditioner on the structured auxiliary grid, so optimal structured-grid solvers transfer to unstructured grids."],"supporting_citations":[{"why":"Supplies the abstract theory of iterative methods for SPD systems, including Theorem 2.3 and the symmetrized operator $\\bar B$, which the paper extends to semidefinite problems.","marker":"[38]"},{"why":"Introduces the auxiliary space method and auxiliary grid preconditioners that the paper formalizes into a general framework.","marker":"[36]"},{"why":"States the Xu–Zikatanov identity for the successive subspace correction method that the paper re-derives sharply in the semidefinite setting.","marker":"[37]"},{"why":"Shows how the auxiliary space method derives the Xu–Zikatanov identity for SPD problems; the SSC proof in Section 6 follows this template.","marker":"[7]"},{"why":"Defines the Hiptmair–Xu preconditioners for $H(\\mathrm{curl})$ and $H(\\mathrm{div})$ and supplies the discrete regular decompositions used in the condition-number bounds.","marker":"[15]"},{"why":"Gives the prior sharp convergence estimate for subspace corrections for singular systems that the paper generalizes and reproduces without pseudoinverses.","marker":"[26]"},{"why":"Provides the semidefinite convergence results, including Lemma 2.2 invoked in Theorem 4.6, that the auxiliary-space viewpoint unifies.","marker":"[34]"},{"why":"Develops a nearly optimal multigrid method on unstructured grids via an auxiliary grid, the application revisited in Section 8.","marker":"[12]"}],"fun_headline_variants":["Exact convergence identity for semidefinite solvers","Auxiliary space formula predicts iterative error","Sharp error formula for singular matrix iterations","Unified theory gives exact convergence rates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that the iteration operator $B$ can be factored through a surjective auxiliary map as $B=\\Pi B_e\\Pi^t$, with the symmetrized auxiliary operator $\\bar B_e$ positive definite on $V_e$ (or on a subspace $W_f$ containing $R(A_e)$); if $B$ does not have such a factorization, the paper's exact formulas do not apply.","fun_headline_variants_meta":{"raw":{"variants":["Exact convergence identity for semidefinite solvers","Auxiliary space formula predicts iterative error","Sharp error formula for singular matrix iterations","Unified theory gives exact convergence rates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000646,"raw_usage":{"total_tokens":2974,"prompt_tokens":957,"completion_tokens":2017,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":1962}},"tokens_in":573,"tokens_out":2017,"duration_ms":13250,"temperature":1.0,"reasoning_tokens":1962,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:12:57.589193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $A=\\mathrm{diag}(1,0)$ on $\\mathbb{R}^2$, let $\\Pi:\\mathbb{R}\\to\\mathbb{R}^2$ be any nonzero column, and let $B_e$ be a positive scalar. Form $B=\\Pi B_e\\Pi^t$; compute $|I-BA|_A^2$ directly by maximizing $|(I-BA)v|_A^2$ over $v\\in R(A)$ with $|v|_A=1$, and compare it with the right-hand side of Theorem 5.1. Any mismatch would refute the identity, and repeating the test over random $\\Pi$ and $B_e$ would probe the factorization's sufficiency as well.","supporting_citations":[{"cited_title":"Xu and L","cited_arxiv_id":null,"evidence_quote":"Supplies the abstract theory of iterative methods for SPD systems, including Theorem 2.3 and the symmetrized operator $\\bar B$, which the paper extends to semidefinite problems."},{"cited_title":"Xu , The auxiliary space method and optimal multigrid preconditioning techniques for unstructured grids , Computing, 56 (1996), pp","cited_arxiv_id":null,"evidence_quote":"Introduces the auxiliary space method and auxiliary grid preconditioners that the paper formalizes into a general framework."},{"cited_title":"Xu and L","cited_arxiv_id":null,"evidence_quote":"States the Xu–Zikatanov identity for the successive subspace correction method that the paper re-derives sharply in the semidefinite setting."},{"cited_title":"Chen , Deriving the X -- Z identity from auxiliary space method , in Domain Decomposition Methods in Science and Engineering XIX , vol","cited_arxiv_id":null,"evidence_quote":"Shows how the auxiliary space method derives the Xu–Zikatanov identity for SPD problems; the SSC proof in Section 6 follows this template."},{"cited_title":"Hiptmair and J","cited_arxiv_id":null,"evidence_quote":"Defines the Hiptmair–Xu preconditioners for $H(\\mathrm{curl})$ and $H(\\mathrm{div})$ and supplies the discrete regular decompositions used in the condition-number bounds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the prior sharp convergence estimate for subspace corrections for singular systems that the paper generalizes and reproduces without pseudoinverses."},{"cited_title":"Wu, Y.-J","cited_arxiv_id":null,"evidence_quote":"Provides the semidefinite convergence results, including Lemma 2.2 invoked in Theorem 4.6, that the auxiliary-space viewpoint unifies."},{"cited_title":"Grasedyck, L","cited_arxiv_id":null,"evidence_quote":"Develops a nearly optimal multigrid method on unstructured grids via an auxiliary grid, the application revisited in Section 8."}],"review_version":2}