{"id":"a08b9334-0833-4a0a-bc91-2d700580af16","arxiv_id":"2508.06378","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper develops a duality-based framework and an iterative algorithm (m-d-Lawson) for discrete minimax rational approximation of matrix-valued functions with a common denominator.","lead":"This paper builds a rigorous minimax approximation theory for matrix-valued functions by extending the scalar duality framework and proposes the m-d-Lawson algorithm, which computes rational approximants that minimize the worst-case Frobenius-norm error over sample points. The method targets model reduction, microwave duplexer design, and nonlinear eigenvalue problems, where existing algorithms usually minimize a root-mean-square error without an optimality guarantee.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The rational-case convergence proof in Theorem 7.2 does not cover the implemented β=1 setting and gives only per-iteration local monotonicity, leaving the central claim that m-d-Lawson solves (2.19) without a uniform guarantee.","rationale":"The reader identified the β=1 gap and the a posteriori strong-duality condition as the weakest assumptions. I agree that the β=1 gap is significant, and I add that even the stated small-β result is only per-iteration and iteration-dependent, so it does not yield a uniform convergence guarantee for any fixed β. The a posteriori strong-duality condition is a limitation rather than a flaw, since Theorem 4.1 itself is valid (indeed it follows immediately from weak duality). The most load-bearing issue is therefore the mismatch between the proved convergence (local, small-β, and non-uniform) and the actual algorithm (β=1), which undermines the claim that m-d-Lawson provably solves the dual problem. This reinforces the reader's CONDITIONAL verdict without changing it, because the contribution is still substantive and the gap is addressable by either proving a uniform result or explicitly labeling the β=1 experiments as heuristic.","tokens_in":38136,"tokens_out":25520,"duration_ms":290868,"concrete_test":"Run Algorithm 1 with β = 1 on a large set of random matrix-valued rational minimax instances (e.g., 1000 problems with s=t=2, n_ij=d=3, m=20, random nodes and random F(x) values) and record d(w^(k)) at every iteration. If any instance exhibits d(w^(k+1)) < d(w^(k)) for some k, the implementation contradicts the only proven ascent property and the concern lands. If no violation is found, the practical concern is mitigated, but the theoretical gap (non-uniform β0, no convergence to the dual maximum) remains and should be acknowledged in the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central computational claim is that m-d-Lawson solves the dual problem (2.19) for matrix-valued rational minimax approximation. The only convergence result for the rational case, Theorem 7.2, asserts that at each iteration k there exists β0 > 0, possibly depending on k, such that for all β ∈ (0, β0), d(w^(k+1)) ≥ d(w^(k)). It does not establish a uniform β0 working for all iterations, nor does it prove that the sequence {d(w^(k))} converges to max_{w∈S} d(w). The algorithm is implemented and evaluated with β = 1 (Section 6, Step 5), for which the theorem provides no guarantee even on a single step. Moreover, the proof of Theorem 7.2 borrows the argument of [43, Theorem 5.2] and assumes, among other things, that q(β) is continuously differentiable near β=0 and that d(w^(k)) is a simple eigenvalue; these hypotheses are not verified in the numerical experiments. If β0(k) can shrink as k grows, then no fixed β is certified, so the theoretical support for the default algorithm is absent. The reported monotone decrease of the duality gap in Example 8.2 is empirical evidence, not a consequence of the stated theorems. Thus the central claim that m-d-Lawson solves (2.19) rests on an unproven convergence property for the parameter regime actually used.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a duality-based framework for discrete rational minimax approximation of matrix-valued functions with a common scalar denominator, minimizing the worst-case Frobenius-norm error over sampled nodes. It reformulates the original problem (1.2) as a linearized optimization (2.6), derives the Lagrange dual problem (2.19), and proves weak duality (Theorem 2.3). Strong duality is established conditionally in Theorem 4.1 via the checkable equality d(w*) = e(R*), and unconditionally for the polynomial case d=0 in Theorem 4.2. The proposed algorithm, m-d-Lawson (Algorithm 1), generalizes Lawson's iteration to the matrix-valued setting using the dual function. Section 7 gives convergence results: a monotonicity statement for the polynomial case for beta in [0,2], and for the rational case only a local monotonicity result for sufficiently small Lawson exponent beta (Theorem 7.2). Numerical experiments compare m-d-Lawson with v-AAA-Lawson, block-AAA, RKFIT, Loewner and VF on three test problems.","tokens_in":38500,"tokens_out":6118,"duration_ms":59913,"significance":"If the framework is fully validated, the paper would provide a principled optimization formulation and a practical algorithm for a problem class that is usually handled by heuristics; this is a useful extension of the authors' scalar d-Lawson work to matrix-valued functions with a common denominator. The block treatment in Sections 2 and 4 is algebraically coherent, and the sufficient-condition/checkable-gap perspective in Theorem 4.1 is genuinely useful. The paper also ships reproducible MATLAB code and gives extensive comparisons, which strengthens its practical value. However, the convergence guarantee for the rational case, which is the central computational claim, is considerably weaker than the narrative suggests: Theorem 7.2 does not cover the implemented beta=1 regime and does not prove convergence to a dual maximizer.","major_comments":[{"comment":"The only convergence result for the rational case d >= 1 is local: for each fixed k, there exists beta0(k) > 0 such that d(w^(k+1)) >= d(w^(k)) for all beta in (0, beta0(k)). It establishes neither a uniform beta0, nor convergence of {d(w^(k))} to max_{w in S} d(w). Yet Algorithm 1 is implemented and evaluated with beta = 1 (Section 6, Step 5), a regime to which the theorem provides no guarantee even for a single step. Therefore the abstract and Section 1 claim that 'convergence analysis of m-d-Lawson is established' overstates the proved statement; the monotone decrease of the duality gap in Example 8.2 is empirical evidence, not a consequence of the stated theorem.","section":"Section 7.3, Theorem 7.2 and Section 6, Algorithm 1"},{"comment":"Theorem 7.2 assumes d(w^(k)) is a simple eigenvalue of the pencil and, in the proof, that q(beta) is continuously differentiable near beta = 0. It also assumes (A1) and (A2). None of these conditions is verified or monitored in the numerical experiments of Section 8, and no experiment reports the behavior for beta other than 1. Thus the numerical protocol is not covered by the theorem's hypotheses either. The authors should either prove a theorem valid for beta = 1 under verifiable assumptions, or state explicitly that beta = 1 is a heuristic choice for rational problems and the convergence theory applies only to sufficiently small Lawson exponents.","section":"Section 7.3, Theorem 7.2 assumptions"},{"comment":"For rational d >= 1, the optimality of a computed R* rests on the a posteriori condition d(w*) = e(R*) in (4.1). The paper gives no result showing that m-d-Lawson, with either beta = 1 or the small beta of Theorem 7.2, drives the duality gap (4.5) to zero or that a sequence of iterates reaches a dual maximizer. Hence the central claim that m-d-Lawson solves the dual problem (2.19) and hence the original problem (1.2) is conditional in two layers: strong duality must hold for the instance, and the iteration must converge to a maximizer. This limitation should be stated prominently as a theorem-level restriction, not only as a numerical check.","section":"Section 4, Theorem 4.1 and stopping criterion (6.1)"}],"minor_comments":[{"comment":"The title spacing 'matrix-valued function s' and the abstract 'minima x approximation' should be corrected. Proofreading is needed.","section":"Title and Abstract"},{"comment":"The text says '401 (i.e., m = 400) equidistant sampled frequencies' but then lists m = 400; this is inconsistent. Also Example 8.1's 'interval [1,100]i' should be written as {i omega : 1 <= omega <= 100} for clarity.","section":"Example 8.3, Section 8"},{"comment":"Figure 8.1 caption says 'the up-right represents' and 'demonstrates' where 'upper-right' and 'shows' are intended. Please also check that Figure 8.5's subfigure labels are legible in print.","section":"Figure captions"},{"comment":"Several proofs are stated as 'same as [45]' or 'adapted from [43]' with details omitted. This is acceptable for a generalization, but the paper would benefit from a short summary of which algebraic steps in the block setting are new and which are direct transcriptions of the scalar arguments.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is heavily built on the authors' own prior work [43,45], and the genuinely new content is the block/common-denominator extension. The gap between Theorem 7.2 and the beta=1 experiments is the main obstacle; I would not accept the paper in its current form without either a proof covering the implemented parameter regime or a substantial rewriting of the convergence claims. The numerical study is useful and the conditional strong-duality framework is sound, so major revision rather than rejection seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this is a genuine block/matrix-valued generalization of the d-Lawson duality framework, and the matrix algebra is largely sound. The catch is that the rational-case convergence theorem only gives per-iteration monotonic improvement for a β small enough that may depend on the iteration, while the implemented algorithm runs at β=1. So the theory does not cover the default experiment regime.\n\nWhat is new: the paper formulates the discrete rational minimax problem for matrix-valued functions with a common denominator, derives a Lagrangian dual over the simplex, proves weak/strong duality under a numerically checkable condition (Theorem 4.1), gives an extremal-point lower bound (Theorem 5.1), and proposes m-d-Lawson, a block extension of Lawson's iteration. For the polynomial case (d=0), strong duality is unconditional (Theorem 4.2) and the convergence analysis covers β∈[0,2], so that part is solid. The paper also ships MATLAB code, which helps reproducibility.\n\nThe main soft spot is exactly what the stress-test flags. Theorem 7.2 assumes d(w^k) is a simple eigenvalue and the denominator coefficients are continuously differentiable in β near 0; but even then it only gives a β0(k) for monotonicity of one step. There is no uniform β0 and no proof that the dual iterates converge to d*. Since all rational experiments use β=1, the reported monotone decay of the duality gap is empirical, not a consequence of the stated theorems. The paper is somewhat transparent about this—Section 6 explicitly says \"any sufficiently small β\" for the rational case—but the abstract's blanket claim that \"convergence analysis of m-d-Lawson is established\" overstates what is proven. The a posteriori strong-duality condition (4.1) for rational approximation is also a real limitation, though it is standard in this line of work and the paper does report the duality gap.\n\nThe self-citation pattern is heavy—many proofs are \"same as [45]\" or \"adapted from [43].\" That is not a flaw per se, because the block extension is non-trivial and those results are the natural base. But a referee should ask the authors to make the dependency clearer and to either close the β=1 gap or label β=1 as heuristic.\n\nBottom line: this deserves a serious referee. The polynomial case is strong and the rational framework is a useful contribution even if the convergence guarantee is weaker than advertised. I'd send it to review with a request for revision rather than desk reject.\n\nBest,","headline":"A genuine and mostly correct block extension of the d-Lawson framework, but the rational-case convergence theory does not cover the β=1 used in experiments.","tokens_in":38975,"tokens_out":2629,"would_cite":true,"duration_ms":26706,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["41A50","41A20","65D15","90C46"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that matrix-valued rational minimax approximation can be solved by maximizing a dual problem over the probability simplex, and that a checkable strong-duality condition certifies when the recovered rational function is the","keywords":["matrix-valued rational approximation","minimax approximation","Lawson iteration","Lagrangian duality","strong duality","Frobenius norm error","discrete rational approximation","dual problem"],"falsifier":"Take a discrete rational approximation of type $(n,d)$ with $d\\ge1$ and a known exact minimax error (for instance a scalar rational function with a known equioscillation point distribution). Run m-d-Lawson to convergence and compute the duality gap (4.5), $|e(R)-d(w)|/e(R)$. If the gap stabilizes at a strictly positive value, the sufficient condition (4.1) is violated for that instance, so no global optimality certificate is obtained; collecting such instances would reveal when the paper's central guarantee actually applies.","tokens_in":37990,"feed_emoji":"🎯","tokens_out":8757,"duration_ms":85540,"temperature":0.7,"pith_summary":"This paper tackles a discrete worst-case approximation problem: given samples of a matrix-valued function $F:\\mathbb{C}\\to\\mathbb{C}^{s\\times t}$, find a matrix rational function $R(x)=P(x)/q(x)$ with a common scalar denominator that minimizes the largest squared Frobenius error over the sample nodes. The authors show that this min-max problem can be linearized into a standard constrained minimization, and that its Lagrange dual is a maximization over the probability simplex. They prove weak duality always holds, and give a sufficient condition under which strong duality holds: if at the computed dual maximizer the recovered rational function has maximal squared error exactly equal to the dual value, then that rational function solves the original minimax problem. In the polynomial case (denominator degree zero), the condition is unnecessary and strong duality holds unconditionally. The numerical method m-d-Lawson solves the dual by iterated weighted least-squares-type updates, with convergence established for both polynomial and rational settings.","feed_headline":"Certify the best rational fit to matrix samples with a dual iteration","feed_subtitle":"When the duality gap closes, the computed rational function is provably the worst-case-optimal approximant.","key_machinery":"The key object is the dual function $d(w)$, defined for each weight vector $w$ on the simplex as the minimum, subject to $\\sum_\\ell w_\\ell |q(x_\\ell)|^2=1$, of the weighted sum of squared Frobenius deviations. For fixed $w$, this minimum is the smallest eigenvalue of a Hermitian positive semidefinite pencil $(A_w,B_w)$, and after Vandermonde-with-Arnoldi orthogonalization its square root is the smallest singular value of a structured tall matrix. The Lawson update $w_\\ell \\leftarrow w_\\ell \\|F(x_\\ell)-R(x_\\ell)\\|_F^\\beta / \\sum_i w_i \\|F(x_i)-R(x_i)\\|_F^\\beta$ is a projected gradient ascent step for $d$, so each iteration only requires one thin SVD and one closed-form weight update. The pape","core_discovery":"The central claim is that the matrix-valued rational minimax problem (1.2) can be solved through the dual maximization (2.19) over the simplex, and that m-d-Lawson is a reliable solver for that dual. The linearized primal (2.6) admits a Lagrange dual function $d(w)$ defined by a constrained minimum that, for fixed weights, reduces to the smallest eigenvalue of a Hermitian positive semidefinite pencil—equivalently, to the smallest singular value of a tall matrix built from weighted Vandermonde bases. Theorem 4.1 is the paper's hinge: if $w^*$ maximizes the dual and the rational function $R^*$ recovered from the minimizer of $d(w^*)$ satisfies $e(R^*) = \\max_{x_\\ell \\in X} \\|F(x_\\ell)-R^*(x_\\e","pith_inferences":["Beyond the paper: if strong duality holds generically for matrix-valued rational problems, m-d-Lawson would provide a certified alternative to RMSE-based fitting in multiport microwave modeling and nonlinear eigenvalue problems, where worst-case error is the operative metric.","Beyond the paper: the duality gap (4.5) can be used as an adaptive certification switch—run Lawson until the gap is small, then declare the fit minimax; a persistent gap would signal a type or node-set change rather than more iterations.","Beyond the paper: the same dual formulation likely extends to other matrix norms, such as the spectral norm, by replacing the Frobenius trace terms with the dominant singular value of the error matrix and reweighting by the corresponding singular vectors.","Beyond the paper: because Example 2.1 shows the infimum in (1.2) can be unattainable, any practical solver must distinguish the certified regime (gap zero) from the merely approximate regime—a distinction the paper's framework makes explicit."],"forward_implications":["If the sufficient condition (4.1) holds, the computed $R^*$ is a provable global solution of (1.2), giving users a numerical certificate of optimality via the duality gap (4.5).","For matrix-valued polynomial minimax approximation ($d=0$), strong duality is unconditional, so m-d-Lawson with $\\beta\\in(0,2)$ monotonically increases $d(w)$ and converges to the global minimax polynomial.","Complementary slackness implies the optimal weight vector is supported only on extreme points, so non-extreme nodes can be discarded to reduce cost without changing the solution.","At least $\\min_{i,j}(n_{ij}+2)$ extreme points must exist at the optimum, a matrix analogue of the equioscillation property that can be used as a sanity check.","In the rational case ($d\\ge1$), the framework inherits a Ruttan-type strong-duality behavior: the condition is frequently satisfied in practice, so the scalar success of d-Lawson extends to block/vector-valued functions."],"supporting_citations":[{"why":"Supplies the scalar convex dual problem and d-Lawson iteration that this paper generalizes to matrix-valued functions.","marker":"[45]"},{"why":"Provides the convergence analysis of Lawson's iteration for polynomial and rational minimax approximation that the proofs adapt.","marker":"[43]"},{"why":"Introduces the Lq-weighted dual programming viewpoint for linear Chebyshev approximation that motivates the duality approach.","marker":"[42]"},{"why":"Supplies the Vandermonde-with-Arnoldi process used to evaluate the dual function stably at given weights.","marker":"[6]"},{"why":"Provides benchmark algorithms and test problems for matrix-valued rational approximation against which m-d-Lawson is compared.","marker":"[14]"},{"why":"Supplies the adaptive barycentric rational minimax approximation context and the AAA-Lawson comparison baseline.","marker":"[12]"},{"why":"Introduces Lawson's original iteration, the method that m-d-Lawson extends to the dual of the matrix-valued problem.","marker":"[24]"},{"why":"Establishes Ruttan's sufficient condition for best complex rational approximants, whose matrix analogue is Theorem 4.1.","marker":"[32]"},{"why":"Supplies optimality and uniqueness conditions in complex rational Chebyshev approximation used to interpret strong duality.","marker":"[35]"},{"why":"Supplies the convex duality and Slater condition facts used to prove unconditional strong duality for the polynomial case.","marker":"[5]"}],"fun_headline_variants":["Matrix rational minimax: dual iteration gives provably optimal fit","Solve matrix rational fits with a certified minimax dual iteration","Matrix minimax via dual Lawson - proven worst-case fit","When duality closes, rational matrix fit is provably best","m-d-Lawson: provably optimal matrix rational minimax"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that, for rational approximations with denominator degree $d\\ge1$, the computed dual maximizer produces a rational function whose largest squared Frobenius error exactly equals the dual value; this equality is not guaranteed in advance and is only verified after the computation through the duality gap (4.5).","fun_headline_variants_meta":{"raw":{"variants":["Matrix rational minimax: dual iteration gives provably optimal fit","Solve matrix rational fits with a certified minimax dual iteration","Matrix minimax via dual Lawson - proven worst-case fit","When duality closes, rational matrix fit is provably best","m-d-Lawson: provably optimal matrix rational minimax"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00078,"raw_usage":{"total_tokens":3347,"prompt_tokens":869,"completion_tokens":2478,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":2394}},"tokens_in":613,"tokens_out":2478,"duration_ms":15646,"temperature":1.0,"reasoning_tokens":2394,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:45:19.179215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a discrete rational approximation of type $(n,d)$ with $d\\ge1$ and a known exact minimax error (for instance a scalar rational function with a known equioscillation point distribution). Run m-d-Lawson to convergence and compute the duality gap (4.5), $|e(R)-d(w)|/e(R)$. If the gap stabilizes at a strictly positive value, the sufficient condition (4.1) is violated for that instance, so no global optimality certificate is obtained; collecting such instances would reveal when the paper's central guarantee actually applies.","supporting_citations":[{"cited_title":"Yang, L.-H","cited_arxiv_id":null,"evidence_quote":"Introduces the Lq-weighted dual programming viewpoint for linear Chebyshev approximation that motivates the duality approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Vandermonde-with-Arnoldi process used to evaluate the dual function stably at given weights."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides benchmark algorithms and test problems for matrix-valued rational approximation against which m-d-Lawson is compared."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces Lawson's original iteration, the method that m-d-Lawson extends to the dual of the matrix-valued problem."},{"cited_title":"Boyd and L","cited_arxiv_id":null,"evidence_quote":"Supplies the convex duality and Slater condition facts used to prove unconditional strong duality for the polynomial case."}],"review_version":1}