{"id":"6ccb1019-e9cb-44c7-a8f7-1f2b84ee2f7c","arxiv_id":"2607.29270","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Newton–Schulz iterations, initialized by a scaled transpose or an LS–Gram polynomial, converge to the i-conjugate Moore–Penrose inverse of any rectangular split-quaternion matrix, and C†AR† is the optimal CUR middle factor.","lead":"The paper builds iterative methods for computing Moore–Penrose inverses of split-quaternion matrices by embedding them as real 2x2-block matrices and running Newton–Schulz iterations. It adds a least-squares polynomial start, proves convergence conditions, and shows the resulting CUR/cross approximations behave like the standard real case.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"The reader identified the i-conjugate transpose as the weakest assumption. I judge this not to be a load-bearing concern: the real representation definition (19) is self-consistent, and the H-adjoint is the unique adjoint for which the real transpose identity holds. A full-conjugate MP inverse would fail the Penrose symmetry equations for zero divisors, so H is not an arbitrary convention. The spectral-bound premise is also addressable by taking ρ=∥M∥_F², which is always a certified upper bound. The only genuine mathematical defect I found is a likely typographical error in Proposition 3.1's first branch, but it does not propagate to the central iterative results. The reader's CONDITIONAL verdict rests on reproducibility gaps (missing code, single-realization timings), which are outside the scope of this stress-test. Since I find no load-bearing correctness concern, the reader's verdict should remain unchanged.","tokens_in":21772,"tokens_out":32943,"duration_ms":299768,"concrete_test":"Verify Proposition 3.1 for scalars q=1+i and q=1+j: compute (q_σ)† and map back via σ^{-1}. For N(q)≠0 the result should equal \\bar{q}/N(q); if the printed formula q/N(q) is used, it differs by the full conjugate, confirming the typo. This test does not affect the main theorems but would settle the correctness of the scalar statement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—Theorem 4.1's exact Newton–Schulz convergence for arbitrary-rank real matrices, its transfer to split-quaternion matrices via the real representation, and the CUR optimality result—is mathematically sound. The i-conjugate transpose H is not an arbitrary choice: it is the adjoint for which (A^H)_σ = (A_σ)^T, and it is necessary for the Penrose symmetry equations to be satisfiable for zero divisors such as q=1+j. The only defect I find is a likely typo in Proposition 3.1's first branch: for N(q)≠0, the scalar Moore–Penrose inverse should be \\bar{q}/N(q), not q/N(q). The proof itself derives Q^{-1}=(\\bar{q})_σ/N(q), so this is a statement/proof mismatch. This does not affect the main iterative or CUR theorems, which rely on the representation-based Definition 3.3 rather than the scalar formula. The practical concern about a certified upper bound ρ is also muted because ρ=∥M∥_F² always provides a valid conservative choice, albeit at the cost of slower convergence.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops iterative Newton–Schulz algorithms for the Moore–Penrose inverse of split-quaternion matrices via a 2×2 real representation and the i-conjugate transpose H, so that (A^H)_σ=(A_σ)^T. It defines A† by σ^{-1}((Aσ)†), proves scalar formulas including zero-divisor cases, and establishes existence and uniqueness of the H-based Moore–Penrose inverse. The main analytic result, Theorem 4.1, gives exact thin-SVD expressions for the iterates and residuals of the Newton–Schulz iteration for arbitrary-rank real rectangular matrices. These results are transferred to embedded and native split-quaternion iterations in Theorems 4.6 and 4.7. A Souriau-inspired low-degree polynomial initialization with a spectral acceptance test and safe fallback is proposed in Section 4.6. The paper then proves that U⋆=C†AR† is a Frobenius-optimal CUR middle factor and gives a rank condition under which the cross factor W† is exact. Numerical examples in MATLAB and Python illustrate the behavior, including an ablation of the polynomial initialization.","tokens_in":22055,"tokens_out":26710,"duration_ms":249515,"significance":"If the results stand, this is the first iteration-count-certified Newton–Schulz analysis for split-quaternion Moore–Penrose inverses without full-rank restrictions. The thin-SVD argument in Theorem 4.1 is clean and yields exact formulas for iterates and residuals rather than asymptotic rates, and the transfer through the real representation is elegant. The CUR/cross optimality results are useful and are proved without restrictive rank assumptions. I agree with the stress-test assessment that the main derivations are sound and non-circular: the Moore–Penrose inverse is defined through the real representation and then shown to satisfy the Penrose equations, and the convergence theorems are derived from a thin SVD without fitted constants. The paper is also careful to label the timing examples as single-realization implementation checks and to disclose that the polynomial-initialization experiments use exactly known spectra and ρ. The principal caveat is that the object computed is the H-adjoint Moore–Penrose inverse, not the full-conjugate inverse; the paper is explicit about this, but it should be kept prominent.","major_comments":[{"comment":"The first branch of the scalar formula is stated as q†=q/N(q) when N(q)≠0. This is false: for q=i it gives q†=i, yet i·i·i=-i≠i. The proof in the same proposition derives Q^{-1}=(\\bar{q})_σ/N(q), which corresponds to q^{-1}=\\bar{q}/N(q); the displayed formula is missing the full conjugate. This is a must-fix. It does not propagate to Theorem 4.1, 4.7, 5.1, or 5.3, which use Definition 3.3, but the scalar formula is part of the paper's stated framework and should be corrected.","section":"§3.1, Eq. (17)"}],"minor_comments":[{"comment":"The text says the complete Python implementation is available in a Google Colab notebook, but no URL or repository is given. Please provide a link or attach the code as supplementary material.","section":"§6.4"},{"comment":"The LS–Gram initialization is used only when full row/column rank is certified, and the iteration requires a certified upper bound ρ≥||M||_2^2. No certification procedure is specified. Please state how rank certification and the spectral-norm bound are obtained in practice (e.g., rank-revealing QR, power iteration/Lanczos with a safety factor) and how an overestimate of ρ affects the polynomial fit.","section":"§4.6, Algorithm 4"},{"comment":"The Moore–Penrose inverse is defined with respect to the i-conjugate transpose H, not the full conjugate A^*. Although Section 2.2 is explicit, the abstract and title could mislead readers familiar with other split-quaternion conventions. Consider stating this qualifier in the abstract.","section":"§2.2 / Abstract"},{"comment":"The timing comparisons are based on one matrix realization each. The paper labels them as implementation checks, which is appropriate; if they are meant to support performance claims, multiple seeds and interquartile ranges should be reported as in Example 6.4.","section":"§6.1–6.3"},{"comment":"Minor typo: 'Sigma' appears instead of '\\Sigma' in the thin-SVD expression for the Frobenius norm of αM^T−M†. Also, check the hyphenation of 'split-quaternion' for consistency.","section":"§4.5, Prop. 4.5 proof"}],"recommendation":"minor_revision","confidential_remarks":"The paper's main theorems are sound and the presentation is generally careful. The scalar formula error in Proposition 3.1 is isolated and does not affect the central iterative or CUR results, but it must be corrected. The missing Colab URL is a reproducibility issue that should be fixed. I see no need for another round of technical review after these local revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Main theorems are correct. The thin-SVD proof of Newton–Schulz for rectangular matrices, Theorem 4.1, is clean and gives exact iteration and residual formulas. The transfer to split quaternions via the real representation is legitimate, and the equivalence of embedded and native iterations is a nice touch. The CUR optimal middle factor is standard but correctly adapted. Credit where due: the paper is honest about the extrapolation heuristic and the LS–Gram initialization's failure mode, and includes a safe fallback.\n\nSoft spots, in order. First, Proposition 3.1 has a typo: the first branch says q† = q/N(q), but the proof derives \\bar{q}/N(q). Since q\\bar{q}=N(q), the statement is wrong; it should be corrected. It does not affect the matrix-level theorems, which use Definition 3.3, but a reader using the scalar formula will get a wrong inverse. Second, reproducibility: a Colab notebook is claimed but no URL is given, and the MATLAB experiments are single-realization timings. The paper labels them as implementation checks, fair enough, but the key practical step—estimating ρ or the spectral norm—is side-stepped in the experiments because ρ=1 by construction. A test with an estimated ρ and the fallback path would be more convincing. Third, the choice of the i-conjugate transpose is explicit and reasonable; it is a scope limitation, not a flaw. If someone needs the Moore–Penrose inverse with respect to the full conjugate, they are computing a different object.\n\nNone of these are load-bearing. The central argument holds. The paper is a solid extension of known iterative techniques to a niche algebra, with a useful convergence analysis and an honest experimental section. It deserves a serious referee; with the typo fixed and code linked, I would be comfortable accepting.","headline":"Main theorems are sound: the thin-SVD Newton–Schulz analysis is clean and the split-quaternion transfer works, but there is a typo in the scalar Moore–Penrose formula and the experimental reproducibility is not as claimed.","tokens_in":22506,"tokens_out":2870,"would_cite":true,"duration_ms":27027,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A09","15A66","65F10","65F35"],"pacs":[],"model":"deepseek-v4-flash","headline":"A Newton–Schulz iteration, initialized by a scaled i-conjugate transpose, converges to the Moore–Penrose inverse of any rectangular or rank-deficient split-quaternion matrix, with exact residual formulas certifying the convergence rate.","keywords":["split-quaternion matrix","Moore–Penrose inverse","Newton–Schulz iteration","polynomial initialization","real representation","CUR approximation","zero divisors","i-conjugate transpose"],"falsifier":"Take the 2×2 real matrix M=diag(1,0) and run X_{k+1}=2X_k−X_kMX_k with X_0=αM^T. For α=1.5 the theorem predicts X_1=diag(0.75,0), X_2=diag(0.9375,0), converging to diag(1,0); for α=2.0 it predicts non-convergence (the (1,1) entry alternates between 0 and 2). If either prediction fails, the thin-SVD convergence claim is wrong. For the split-quaternion side, computing q† for the zero divisor q=1+j via the native iteration and comparing with τ(q)/(2∥q∥_E^2) would test the transfer.","tokens_in":21691,"feed_emoji":"🔁","tokens_out":7536,"duration_ms":63658,"temperature":0.7,"pith_summary":"The paper establishes that the Moore–Penrose inverse of a split-quaternion matrix can be computed by a Newton–Schulz iteration, with convergence guaranteed even for rectangular, rank-deficient, and zero-divisor cases. Using a 2×2 real representation and the i-conjugate transpose, it derives exact formulas for the iterates and projector residuals, and shows that an embedded real implementation and a native split-quaternion implementation are exactly equivalent. It also proves that for fixed sampled rows and columns, the CUR middle factor C†AR† is Frobenius-optimal, and gives a rank condition under which the cheaper cross factor W† yields exact reconstruction. This matters because split quaternions contain zero divisors and have an indefinite norm, so generalized inverses are non-obvious; certified iterative algorithms with explicit iteration counts make such computations practical for large or low-rank problems.","feed_headline":"Newton–Schulz iteration provably finds split-quaternion pseudoinverse","feed_subtitle":"Certified convergence for rectangular, rank-deficient, and zero-divisor split-quaternion matrices.","key_machinery":"The machinery is the block 2×2 real representation σ (from split quaternions to real 2×2 matrices) and its matrix extension, together with the i-conjugate transpose H satisfying (A^H)_σ=(A_σ)^T. The real representation converts split-quaternion arithmetic into real matrix arithmetic in a product-preserving way (Frobenius norms scale by √2), so the thin SVD of A_σ reduces the iteration to independent scalar recurrences per singular direction. The H-adjoint is exactly what makes the real transpose correspond to the split-quaternion adjoint, so the four Penrose equations carry over unchanged. Each singular-direction coefficient d_{i,k} follows the scalar Newton map d←d(2−σ_i d), whose error squ","core_discovery":"Starting from a real matrix M of rank r with positive singular values σ1≥…≥σr, the paper proves that the Newton–Schulz iteration X_{k+1}=2X_k−X_kMX_k with X_0=αM^T converges to M† whenever 0<α<2/σ1^2. The proof uses a thin SVD to obtain exact formulas: X_k=V_r diag((1−θ_i^{2^k})/σ_i)U_r^T and the projector residual PL−MX_k=U_r diag(θ_i^{2^k})U_r^T, with θ_i=1−ασ_i^2. These formulas are transferred to split-quaternion matrices through the real representation σ and the i-conjugate transpose H, for which (A^H)_σ=(A_σ)^T. Both the embedded real iteration and the native split-quaternion iteration are shown to converge to A†, and the CUR middle factor C†AR† is shown to be Frobenius-optimal for fix","pith_inferences":["Because the Moore–Penrose inverse here is defined with respect to the i-conjugate transpose, the same iteration applied with the full conjugate transpose would compute a different object; applications that use the full conjugate (e.g., some Lorentzian formulations) should translate the scaling and residual formulas before adopting the algorithm.","The exact residual–error identity suggests a direct a posteriori error bound that could be used for adaptive stopping or as a certificate in verified numerical libraries, a use the paper does not develop.","The same real-representation transfer should apply to other generalized inverses (e.g., the Drazin inverse) and to other hypercomplex algebras with a fixed anti-involution, since the paper's proof only needs (A^H)_σ=(A_σ)^T and the real Newton–Schulz theorem.","The polynomial initialization's benefits are shown for dense matrices; a testable extension is to use it in matrix-free or sparse settings where Gram-matrix powers are approximated, since the setup cost currently dominates for small matrices."],"forward_implications":["For every split-quaternion matrix A, the native iteration X_{k+1}=2X_k−X_k A X_k with X_0=αA^H converges to A† whenever 0<α<2/∥Aσ∥_2^2, and it is exactly equivalent to the embedded real iteration in exact arithmetic.","The normalized residual ∥A X_k A−A∥_F / ∥A∥_F is a certified stopping criterion: Theorem 4.4 gives ∥X_k−A†∥_F ≤ ∥A†∥_2^2 ∥A X_k A−A∥_F, so users can stop on a computable quantity without knowing the exact inverse.","For fixed sampled column and row sets, C†AR† minimizes ∥A−CUR∥_F over all k×k split-quaternion middle factors U, and when rankσ(A(R,C))=rankσ(A), the cross factor W† equals C†AR† and reconstructs A exactly.","The LS–Gram polynomial initialization, supplemented by a spectral acceptance test and a safe scaled-transpose fallback, never sacrifices the convergence guarantee and reduces the number of Newton–Schulz iterations in the reported tests (about 20% fewer than a scaled transpose at κ=10^3)."],"fun_headline_variants":["Split-quaternion pseudoinverse via Newton–Schulz: proven convergence","Newton–Schulz converges to split-quaternion pseudoinverse—conditions shown","Certified Newton–Schulz iteration for split-quaternion pseudoinverse","How Newton–Schulz nails the split-quaternion pseudoinverse","Pseudoinverse of split-quaternion matrices: Newton–Schulz with proof"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the Moore–Penrose inverse is defined with respect to the i-conjugate transpose, because only then does the real transpose (A_σ)^T correspond to the adjoint used in the Penrose equations; if the full conjugate is the intended adjoint, every algorithm in the paper converges to a different object.","fun_headline_variants_meta":{"raw":{"variants":["Split-quaternion pseudoinverse via Newton–Schulz: proven convergence","Newton–Schulz converges to split-quaternion pseudoinverse—conditions shown","Certified Newton–Schulz iteration for split-quaternion pseudoinverse","How Newton–Schulz nails the split-quaternion pseudoinverse","Pseudoinverse of split-quaternion matrices: Newton–Schulz with proof"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000303,"raw_usage":{"total_tokens":1628,"prompt_tokens":838,"completion_tokens":790,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":688}},"tokens_in":582,"tokens_out":790,"duration_ms":7188,"temperature":1.0,"reasoning_tokens":688,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:18:54.881398+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the 2×2 real matrix M=diag(1,0) and run X_{k+1}=2X_k−X_kMX_k with X_0=αM^T. For α=1.5 the theorem predicts X_1=diag(0.75,0), X_2=diag(0.9375,0), converging to diag(1,0); for α=2.0 it predicts non-convergence (the (1,1) entry alternates between 0 and 2). If either prediction fails, the thin-SVD convergence claim is wrong. For the split-quaternion side, computing q† for the zero divisor q=1+j via the native iteration and comparing with τ(q)/(2∥q∥_E^2) would test the transfer.","supporting_citations":[],"review_version":1}