{"id":"36d7a181-1b6c-45b0-8c12-aaae5cfec125","arxiv_id":"2505.06806","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper claims the Laplacian kernel enables better kernel DMD reconstruction from sparse irregular data and proves Koopman closability, but the proofs have invalid steps and circular reasoning.","lead":"This paper claims the Laplacian kernel makes kernel dynamic mode decomposition rebuild sparse, irregular spatiotemporal data better than the Gaussian kernel, and it attempts an operator-theoretic proof. The experimental pictures are suggestive, but the central theorems contain invalid steps and circular reasoning, so the justification does not hold.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.4 proves only square-integrability; the Laplacian kernel sections are not holomorphic, so the dictionary used by Lap-KeDMD is not in H_L and the closability guarantee never reaches the algorithm.","rationale":"The reader's weakest assumption is precisely the membership of the Laplacian kernel in H_L, and the stress-test confirms that this is the most load-bearing defect. The claim that the algorithm's dictionary is contained in the RKHS where closability is proved is not merely unproved; it is false, because the kernel sections are not holomorphic. This invalidates the claimed operator-theoretic justification for the empirical method, independent of whether the closability theorem itself could be repaired. The secondary algebraic error in Theorem 6.5 reinforces the rejection but is not the primary concern, since even a corrected boundedness proof would not place the Laplacian kernel sections in H_L. The empirical demonstrations remain suggestive but do not compensate for the collapse of the theoretical bridge. Therefore the reader's REJECT verdict is unchanged.","tokens_in":23554,"tokens_out":4429,"duration_ms":45771,"concrete_test":"Fix D=1 and x=0. Compute the Wirtinger derivative ∂/∂z̄ of f(z)=exp(-|z|/σ) for z=a+ib; show it equals -(z/(2σ|z|)) f(z), which is nonzero for z≠0. Since a nonzero antiholomorphic derivative implies f is not holomorphic on C\\{0} and f is not differentiable at 0, f cannot belong to H_L defined in (4.8). Additionally, in Theorem 6.5, substitute u=Az+b into the integral ∫|g_m(Az+b)|^2 exp(-||z||^2/σ)dV_D(z), and check whether the resulting expression equals exp(-||Au||^2/σ)dμ_L(u) for a non-orthogonal A; this equality fails for generic A, so the displayed boundedness chain is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires the dictionary elements used by Lap-KeDMD, namely the Laplacian kernel sections K^{1,σ}_exp(·,x), to lie in the RKHS H_L defined in (4.8) as a space of holomorphic functions. Theorem 4.4 claims this membership, but its proof bounds only the integral of |K^{1,σ}_exp(z,x)|^2 against dμ_L; it never establishes holomorphy, and the function is not holomorphic in z on C^D. For D=1, writing z=a+ib and x=c+id, K^{1,σ}_exp(z,x)=exp(-sqrt((a-c)^2+(b-d)^2)/σ), which satisfies the Cauchy-Riemann equations only at the single point z=x, where it is not differentiable. Since H_L is contained in the holomorphic square-integrable space by (4.8), K^{1,σ}_exp(·,x) is not in H_L. Therefore the closable Koopman operators K_φ:H_L→H_L analyzed in Theorems 5.3 and 6.5 are never applied to the actual kernel sections used in Algorithm 3.1; the operator-theoretic justification for the empirical Lap-KeDMD reconstructions is severed at its first step. A separate, independent defect is visible in Theorem 6.5: the substitution u=Az+b leads to exp(-||A^{-1}(u-b)||^2/σ)|det A|^{-1}dV_D(u), yet the proof replaces this with exp(-||Au||^2/σ)dμ_L(u), an equality that is false for general invertible A. Both issues point in the same direction: the paper does not supply a valid bridge from its RKHS closability theory to the Laplacian-kernel algorithm it advocates.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a kernel-based extended dynamic mode decomposition (Lap-KeDMD) for sparse spatiotemporal reconstruction, together with an operator-theoretic analysis. The authors construct an RKHS H_L of holomorphic functions square-integrable with respect to a \"Laplacian measure\" (4.6), claim that the Laplacian kernel sections belong to H_L (Theorem 4.4), establish boundedness and compactness of Koopman operators with affine symbols on H_L (Theorems 5.3 and 6.5), contrast this with an alleged failure of closability for the GRBF kernel (Theorem 6.7), and derive approximation and faithfulness results for Koopman spectral measures and mode decompositions (Theorems 7.1 and 7.2). The paper also reports qualitative comparisons of Lap-KeDMD versus GRBF-KeDMD on seven datasets, including chaotic and PDE examples.","tokens_in":24024,"tokens_out":6339,"duration_ms":62750,"significance":"If correct, the paper would supply a principled operator-theoretic justification for using the Laplacian kernel in KeDMD, addressing a recognized open problem about closability of Koopman operators on RKHSs. The empirical scope (seven datasets, including Burgers equation, flow past a cylinder, Lorenz and Rössler attractors, traffic data, and NOAA SST anomalies) is a genuine strength, and the problem of connecting kernel choice to closable Koopman operators is timely and relevant. However, the central theoretical results contain multiple load-bearing errors: the Laplacian kernel sections are not established to be (and in general are not) holomorphic, so they need not lie in H_L; the closability proof in Theorem 6.5 uses an invalid change of variables and does not verify the norm-convergence condition required for closability; Theorem 7.1 uses a false identity about spectral-measure integrals; and Theorem 7.2 essentially restates the conclusion already built into Definition 2.4. Because these steps are load-bearing for the advertised operator-theoretic guarantee, the paper's central claim is not supported.","major_comments":[{"comment":"The proof of Theorem 4.4 bounds only the L^2(μ_L) integral of |K^{1,σ}_exp(z,x)|^2 and never establishes that the function is holomorphic. For D=1, K^{1,σ}_exp(z,x)=exp(-|z-x|/σ) is not holomorphic in z on C, and the same failure occurs in higher dimensions, so the Laplacian kernel sections are not members of H_L as defined in (4.8). Consequently, the dictionary used in Algorithm 3.1 is not contained in the Hilbert space on which Theorems 5.3 and 6.5 apply, severing the proposed bridge from the RKHS theory to the Lap-KeDMD algorithm.","section":"§4.3, Theorem 4.4"},{"comment":"The proof's change of variables u=Az+b yields the factor exp(-||A^{-1}(u-b)||^2/σ) dV_D(u)/|det A|, but the next inequality replaces this with exp(-||Au||^2/σ) and then identifies the resulting integral with ||g_m||^2_{H_L}; this identification is false for a general invertible A because dμ_L(u) contains exp(-||u||/σ), not exp(-||Au||/σ). In addition, the proof does not establish closability: Lemma 6.4 requires that g_m→0 in H_L implies K_φ g_m→0 in H_L in norm, whereas (6.3) only shows pointwise convergence through the reproducing kernel. The theorem's conclusion is therefore not derived.","section":"§6.2, Theorem 6.5"},{"comment":"The identity [∫_D λ dV(λ) g](x) - [∫_D λ dV(λ) g](Ax) = [∫_D λ dV(λ) g](x-Ax) is false: the spectral measure integral is a linear operator applied to the function g, not a function of the evaluation point in such a way that argument differences can be pulled through. The subsequent steps also mix pointwise values with norms and treat |V(ρ(K_Ft))| as a scalar measure despite V being a projection-valued measure. The claimed bound (7.4)–(7.5) is therefore not proven.","section":"§7.2, Theorem 7.1, Eq. (7.8)"},{"comment":"The conclusion 𭟋Δ^(0,m)_{H_L} ≡ 0 is already asserted in Definition 2.4, Eq. (2.10), so Theorem 7.2 is a restatement rather than a proof. Moreover, the proof's claim that an infinite series summing to zero has partial sums equal to zero is incorrect; partial sums of a convergent series need not vanish, and the individual terms c_n[K^m_{F_t}ζ_n(x_0)-ζ_n(A^m x_0 + b∑ A^i)] need not be zero. The faithfulness result is thus true by construction, not by the argument given.","section":"§7.2, Theorem 7.2"},{"comment":"The proof concludes ∥K_φ g_m∥_{H_σ} = ∞ from an unbounded upper bound on the integral in (6.4). An upper bound that diverges does not imply that the actual norm diverges, so the assertion that K_φ g_m ∉ H_σ is not established. Furthermore, even if some sequence produced an infinite norm, closability failure would require exhibiting a sequence g_m→0 with K_φ g_m converging to a nonzero limit; no such sequence is given. The claimed contrast with the GRBF kernel is therefore unsupported.","section":"§6.2.1, Theorem 6.7"}],"minor_comments":[{"comment":"The definition states X=∅, which contradicts the intended use of a nonempty input space throughout the paper; this should be X≠∅.","section":"Definition 2.5"},{"comment":"The entry for W reads \"Field of real, complex and whole numbers respectively\"; the phrase \"whole numbers\" is unclear and should be replaced by the intended set (for example, natural numbers or integers).","section":"Table 3"},{"comment":"The experimental comparison is based on visual inspection of heat maps and error plots; no quantitative summary statistic (such as RMSE, mean absolute error, or correlation) is reported, which makes the claimed superiority of the Laplacian kernel difficult to assess.","section":"§3.3"},{"comment":"The figure combines unexplained symbols (for example, the script F and the notation Δ^(0,m)_H) and is not self-contained; the caption should define all quantities shown.","section":"Figure 1"}],"recommendation":"reject","confidential_remarks":"The empirical study and the motivation for studying closable Koopman operators are of interest, but the theory sections contain invalid derivations in the central claims. In my view the errors in Theorem 4.4, Theorem 6.5, Theorem 7.1, and Theorem 7.2 are load-bearing and cannot be repaired by local edits; a substantially rebuilt theoretical framework would be needed. The paper may merit resubmission after such a rework, but in its present form the advertised operator-theoretic guarantee for Lap-KeDMD is not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's central claim — that the Laplacian kernel yields closable Koopman operators on a natural RKHS — is not supported. The first step fails: the Laplacian kernel sections used in the algorithm are not holomorphic, so they do not lie in the space H_L where closability is proved. Theorem 4.4 only establishes square-integrability; it never shows holomorphy, and the function exp(-|z-x|_2/σ) is not holomorphic in z. This severs the theoretical bridge from the RKHS results to the Lap-KeDMD algorithm.\n\nThere is something worth taking seriously here. The question of how kernel choice interacts with closability of Koopman operators is legitimate and under-explored, and the idea of embedding a kernel as an L2 measure to build an operator-friendly space is a reasonable route. The seven-dataset empirical comparison also shows that Laplacian kernels can beat GRBF for sparse irregular reconstruction, at least in the figures shown. The specific combination of the Laplacian kernel with KeDMD appears new.\n\nBut the theory has load-bearing flaws beyond the holomorphy issue. Theorem 6.5's change of variables is wrong: substituting u = Az + b does not produce the bound claimed, and the equality used later is false for general invertible A. Theorem 7.1 assumes g(x) - g(Ax) = g(x - Ax), which is not true for a fixed function. Theorem 7.2 derives a conclusion already baked into Definition 2.4, where the faithful difference is defined to be identically zero. Theorem 6.7 is a non-sequitur: an unbounded upper bound does not imply the norm itself is infinite. On the empirical side, there is no code, no seeds, no aggregated error numbers, and no baseline beyond the GRBF kernel, so the performance claims are not yet reproducible or properly quantified.\n\nThe paper is not a serious threat to the literature; it is a well-intentioned but mathematically unsound attempt at an important problem. I would not cite it as it stands, and I would not send it to peer review. The underlying question deserves better, but this manuscript needs a complete rework of its theory and a much more careful empirical study before it can be evaluated seriously.","headline":"The Laplacian-kernel closability claim collapses because the kernel sections are not holomorphic and several key proofs rely on invalid steps; the empirical comparison is suggestive but not enough to carry the paper.","tokens_in":24514,"tokens_out":3577,"would_cite":false,"duration_ms":37951,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["86A08","47N60","47N70","46E22","47B32","46E20","70G60","76F20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Koopman operators are closable over the Laplacian kernel's function space, giving sparse reconstruction a formal guarantee.","keywords":["Dynamic Mode Decomposition","Koopman operator","closable operator","Laplacian kernel","reproducing kernel Hilbert space","sparse spatiotemporal reconstruction","Koopman mode decomposition","kernel extended DMD"],"falsifier":"For $D=1$, take $f(z)=\\exp(-|z-x|/\\sigma)$ and evaluate its Wirtinger derivative $\\partial f/\\partial\\bar z$; it is nonzero almost everywhere, so $f$ is not holomorphic on $\\mathbb{C}$. Since $H_L$ was defined to contain only holomorphic functions, this single calculation would disprove the membership asserted in Theorem 4.4 and thereby remove the dictionary from the space where closability is proven.","tokens_in":23303,"feed_emoji":"🌀","tokens_out":13278,"duration_ms":106402,"temperature":0.7,"pith_summary":"This paper aims to settle a standing question in kernel-based dynamic mode decomposition: when does the infinite-dimensional Koopman operator, lifted from a nonlinear system onto a reproducing kernel Hilbert space, admit a well-behaved (closable) extension? The authors construct a function space from the Laplacian kernel treated as an $L^2$ density, prove that Koopman operators with affine, contractive symbols are closable and compact on that space, and prove that the Gaussian radial basis function kernel fails the same property. They then run kernel extended DMD with the Laplacian kernel on seven datasets, ranging from Burgers' equation and cylinder flow to Lorenz, Rössler, traffic, and sea-surface temperature, and report that it reconstructs dominant spatial-temporal modes from sparse, irregularly sampled snapshots better than the GRBF kernel. The payoff is a data-driven reconstruction method that carries an operator-theoretic guarantee rather than a purely empirical one.","feed_headline":"Laplacian kernel yields closable Koopman operators","feed_subtitle":"A new function-space construction gives irregular-data reconstruction a proven mathematical guarantee.","key_machinery":"The load-bearing object is the RKHS $H_L$ built from the Laplacian measure, with reproducing kernel $K^\\sigma(z,w) = \\sinh\\!\\big(\\sqrt{\\langle z,w\\rangle/\\sigma^2}\\big)/\\sqrt{\\langle z,w\\rangle/\\sigma^2}$; the space is assembled by applying the Moore–Aronszajn theorem to an orthonormal basis of holomorphic functions. On this space the Koopman operator is taken with affine holomorphic symbol $\\varphi(z)=Az+b$, with $A$ invertible and $0<\\|A\\|_{\\mathrm{Frob}}<1$, so the composition operator is compact; compactness is the mechanism that yields closability, since every compact operator has a closed graph. The argument's second engine is the 0th-snapshot mth-Koopman-mode difference $\\Delta^{(0,m)}_{H_L}$, a weighted sum over Koopman modes comparing $K_{F_t}^m\\zeta_n(x_0)$ with $\\zeta_n(A^m x_0 + b\\sum_{i=0}^{m-1} A^i)$; Theorem 7.2 shows this sum converges to zero, which is the paper's definition of faithful reconstruction.","core_discovery":"On the paper's own terms, the central discovery is that the RKHS $H_L$ generated by the Laplacian kernel, defined as the space of holomorphic functions square-integrable against $d\\mu_L(z) = (2\\pi\\sigma^2)^{-D}\\exp(-\\|z\\|_2/\\sigma)\\,dV_D(z)$, supports closable Koopman operators: for every affine holomorphic symbol $\\varphi(z) = Az + b$ with invertible $A$ and $0 < \\|A\\|_{\\mathrm{Frob}} < 1$, the composition operator $K_\\varphi$ is compact and therefore closable on $H_L$ (Theorem 6.5). The same argument fails for the GRBF kernel, whose RKHS norm forces $\\|K_\\varphi g_m\\|_{H_\\sigma} = \\infty$ for null sequences $g_m$, so the Koopman operator is not closable there (Theorem 6.7). The paper then defines a 'faithful' Koopman mode decomposition difference $\\Delta^{(0,m)}_{H_L}$ that measures the gap between the true dynamical evolution and the RKHS evolution of each Koopman eigenfunction, and proves this difference vanishes identically under the closable affine construction (Theorem 7.2). That vanishing is the theoretical justification offered for using the Laplacian kernel in kernel extended DMD for sparse spatiotemporal reconstruction.","pith_inferences":["The paper's own conclusion notes that the practical side of the faithfulness identity (7.10) is yet to be demonstrated; a numerical study of how fast the partial sums $P_N\\Delta^{(0,m)}_{H_L}$ approach zero would supply that missing check.","The membership question in Theorem 4.4 is testable directly: since $\\exp(-\\|z-x\\|_2/\\sigma)$ is not holomorphic in $z$ on $\\mathbb{C}^D$, a Cauchy–Riemann check would settle whether the algorithm's dictionary actually lies inside $H_L$.","Because the Laplacian kernel is the $\\nu=1/2$ Matern kernel, neighboring low-smoothness Matern kernels may inherit closability, whereas the GRBF failure corresponds to the $\\nu\\to\\infty$ limit.","The faithfulness criterion could be reused to audit other kernel choices, such as polynomial or compactly supported kernels, for Koopman reconstruction, not just Laplacian versus GRBF."],"forward_implications":["Lap-KeDMD, the Laplacian-kernel version of kernel extended DMD, is guaranteed in the idealized setting to reconstruct dominant spatial-temporal modes from sparse and irregularly timed snapshots, because the reconstruction error defined by the Koopman mode difference vanishes.","The GRBF kernel, the standard benchmark, is theoretically disfavored for this task: its RKHS does not support closable Koopman operators, so the same faithfulness guarantee is unavailable.","Bounded, compact, closable Koopman operators admit finite-rank approximations, which is the formal justification for approximating the infinite-dimensional operator by a finite matrix in extended DMD.","For large state-space dimension $D$, the operator norm gap between the true Koopman evolution and the RKHS evolution can be made smaller than any chosen $\\epsilon > (D+1)/(D+1-2\\,\\mathrm{Tr}(A))$, so the spectral measures of the two operators are close (Theorem 7.1)."],"supporting_citations":[{"why":"Supplies the Moore–Aronszajn theorem used to build the holomorphic RKHS from an orthonormal basis.","marker":"[5]"},{"why":"Provides Bochner's theorem and the spectral densities of the squared-exponential and Matern covariance functions that ground the Laplacian measure embedding.","marker":"[55]"},{"why":"Defines the extended DMD / kernel trick algorithm that the paper adapts into Lap-KeDMD with the Laplacian kernel.","marker":"[76]"},{"why":"Gives the explicit RKHS norm for the Gaussian RBF kernel, which the paper uses to prove Koopman operators are not closable on that space.","marker":"[71]"},{"why":"Contributes the boundedness result for Koopman operators on the Gaussian RKHS invoked in Theorem 6.7 to show failure of closability.","marker":"[29]"},{"why":"Frames the requirement that a kernel's RKHS make the Koopman operator closable, which is the operator-theoretic challenge the paper aims to settle.","marker":"[14,15]"}],"fun_headline_variants":["Laplacian kernel makes Koopman operators closable","Closable Koopman operators via Laplacian kernel","Laplacian kernel proves Koopman closability in RKHS","Sparse reconstruction backed by closable Koopman theory","Kernel DMD with Laplacian: closable Koopman operators"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole guarantee rests on the Laplacian kernel actually belonging to the space of holomorphic functions the paper builds; the proof only estimates an integral and does not show the kernel is holomorphic in $z$, so if that membership fails the closability result does not cover the dictionary used in practice.","fun_headline_variants_meta":{"raw":{"variants":["Laplacian kernel makes Koopman operators closable","Closable Koopman operators via Laplacian kernel","Laplacian kernel proves Koopman closability in RKHS","Sparse reconstruction backed by closable Koopman theory","Kernel DMD with Laplacian: closable Koopman operators"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000298,"raw_usage":{"total_tokens":1790,"prompt_tokens":1072,"completion_tokens":718,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":688,"completion_tokens_details":{"reasoning_tokens":630}},"tokens_in":688,"tokens_out":718,"duration_ms":5974,"temperature":1.0,"reasoning_tokens":630,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:33:00.386842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For $D=1$, take $f(z)=\\exp(-|z-x|/\\sigma)$ and evaluate its Wirtinger derivative $\\partial f/\\partial\\bar z$; it is nonzero almost everywhere, so $f$ is not holomorphic on $\\mathbb{C}$. Since $H_L$ was defined to contain only holomorphic functions, this single calculation would disprove the membership asserted in Theorem 4.4 and thereby remove the dictionary from the space where closability is proven.","supporting_citations":[{"cited_title":"Aronszajn, Theory of reproducing kernels , Transactions of the American Mathematical Society, 68 (1950), pp","cited_arxiv_id":null,"evidence_quote":"Supplies the Moore–Aronszajn theorem used to build the holomorphic RKHS from an orthonormal basis."},{"cited_title":"1, Springer, 2006","cited_arxiv_id":null,"evidence_quote":"Provides Bochner's theorem and the spectral densities of the squared-exponential and Matern covariance functions that ground the Laplacian measure embedding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the extended DMD / kernel trick algorithm that the paper adapts into Lap-KeDMD with the Laplacian kernel."},{"cited_title":"Steinwart, D","cited_arxiv_id":null,"evidence_quote":"Gives the explicit RKHS norm for the Gaussian RBF kernel, which the paper uses to prove Koopman operators are not closable on that space."},{"cited_title":"Gonzalez, M","cited_arxiv_id":null,"evidence_quote":"Contributes the boundedness result for Koopman operators on the Gaussian RKHS invoked in Theorem 6.7 to show failure of closability."}],"review_version":1}