{"id":"a3b637e4-1b69-4b08-b9da-bd4ae8d3b1c0","arxiv_id":"2505.11435","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Kernel interpolation converges with rate ε^θ for targets in interpolation spaces between the RKHS and the image of the adjoint of an embedding, continuously interpolating between classical and doubled rates.","lead":"This paper proves general superconvergence bounds for kernel-based interpolation: target functions lying in the image of certain adjoint operators are approximated at faster rates, and the speed-up can be tuned continuously via interpolation spaces. It unifies prior doubling results for smoother targets, covers Sobolev spaces with boundary conditions, and gives a framework for predicting when smoother target functions actually converge faster.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 7.4 misstates the Sobolev smoothness of the random periodic functions: for fα in (48), the correct threshold is s < α − 1/2, not s < α + 1/2, which invalidates the claimed saturation predictions and explains the reported 'unexplained' discrepancies.","rationale":"In good faith, the central theorems appear correct: Theorem 6 is proven cleanly, and Corollary 10 is valid up to a harmless factor of 2 in the constant (the proof gives 2ε^θ∥v∥, not ε^θ∥v∥). The reader's identified weakness—checkability of v ∈ A*(V')—is a genuine limitation and is honestly acknowledged in Sections 1 and 6, but it is not an internal error. The concrete error I find is in Section 7.4: the smoothness threshold for the random trigonometric series is misstated by one full derivative. This is easily verified and is more load-bearing than the reader's limitation because it directly invalidates the experimental calibration for the periodic setting, which is one of the few settings where the power spaces admit exact characterizations (Example 18). The reported saturation discrepancies, which the paper leaves unexplained, are resolved by the correct threshold: r = 1's saturation at α = 3/2 matches τ < α − 1/2, and r = 2's anomaly is at least shifted in the right direction. Thus the practical recommendation is to keep the verdict CONDITIONAL: the main theorems stand, but the periodic experiment and its statements need correction before the numerical support can be taken at face value.","tokens_in":28119,"tokens_out":23753,"duration_ms":216262,"concrete_test":"Recompute the Sobolev membership of fα in (48) analytically: E∥fα∥²_{W^τ_{2,per}} = (Eξ₁²/4) Σ_{j=1}^∞ j^{2τ−2α}. For α = 1 and τ = 1.2 the series is Σ j^{0.4} = ∞, contradicting the paper's claim that fα ∈ W^{1.2} almost surely because 1.2 < 1 + 1/2. A direct check is to evaluate the series and the implied threshold τ < α − 1/2; if the threshold is α − 1/2, then the probabilities and saturation values in Section 7.4 must be corrected, and the 'unexplained' numerical discrepancies disappear for r = 1 and are repositioned for r = 2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest point is not the interpolation argument but the calibration of the periodic numerical experiment in Section 7.4. After defining fα(x) = 1 + Σ_{j≥1} j^{-α} ξ_j cos(2πjx), the paper asserts that fα ∈ W^τ_{2,per} with probability one iff τ < α + 1/2. This is false. The Fourier coefficients of fα decay like |j|^{-α}, so E∥fα∥²_{W^τ} = (Eξ₁²/4) Σ_{j=1}^∞ j^{2τ−2α}, which is finite exactly when τ < α − 1/2; by the three-series theorem the same threshold holds almost surely. The stated condition is off by one full derivative. Consequently the predicted membership thresholds in the discussion of Figure 6 are wrong: for r = 1 the paper's formula would predict saturation for α > 1/2, whereas the observed saturation at α = 3/2 matches the correct threshold α > 3/2; for r = 2 the reported anomaly (saturation at α = 4 instead of α = 7/2) is not a genuine unexplained phenomenon but a consequence of the miscalibrated smoothness. This error does not affect Theorem 6 or Corollary 16, but it undermines the claim that the periodic experiments verify the theory and it directly explains the discrepancies the authors say they lack an explanation for.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a general operator-theoretic framework for superconvergence in kernel-based approximation. Starting from a general Hilbert-space setting, Theorem 6 shows that if a bounded operator A satisfies a global error bound ∥A(f−Pf)∥_V ≤ ε∥f∥_H for an orthogonal projection P, then every v = A*g in the range of the adjoint satisfies ∥v−Pv∥_H ≤ ∥g∥_{V'} ε. Corollary 10 extends this to real interpolation spaces (H, A*(V'))_{θ,∞}, yielding intermediate rates ε^θ. These results are specialized to Mercer power spaces (Section 4), to general kernel integral operators mapping into L_p (Section 5), and to Sobolev spaces, where the dependence on hidden boundary conditions is discussed (Section 6). Numerical experiments in Section 7 illustrate the theory.","tokens_in":28421,"tokens_out":21588,"duration_ms":184678,"significance":"If the central results hold, the paper provides a clean unification and generalization of earlier superconvergence results by Schaback, Sloan–Kaarnioja, and others. The operator-range characterization in Theorem 6 is elegant, and the interpolation-scale extension in Corollary 10 is novel and yields a continuous family of rates between the classical and doubled rates. The application to Sobolev spaces, with the explicit role of boundary conditions, is a useful contribution. The derived rates are parameter-free in the sense that no fitted constants are used; the numerical experiments are illustrative. However, the numerical verification in Section 7.4 contains a clear mathematical error about the Sobolev smoothness of the random periodic functions, which undermines the claimed experimental support and the reported 'unexplained' phenomena. The core theoretical results (Theorem 6, Corollaries 15–16) appear sound.","major_comments":[{"comment":"The statement that f_α(x)=1+Σ_{j≥1} j^{-α} ξ_j cos(2πjx) lies in W^τ_{2,per}(Ω) with probability one iff τ < α+1/2 is false. The Fourier coefficients decay as |c_j| ≍ j^{-α}, so E∥f_α∥^2_{W^τ_{2,per}} ≍ Σ_{j≥1} j^{2τ-2α}, which converges exactly when τ < α−1/2; by the three-series theorem the same threshold holds almost surely. Consequently, all membership predictions for H^θ(k_r,Ω) in the discussion of Figure 6 are shifted by one full derivative. In particular, the reported anomaly for r=2 (saturation of the L1 and L2 rates at α=4 instead of the predicted α=7/2) is not a genuine unexplained phenomenon: with the corrected smoothness, the prediction for the h^4-rate saturation moves to α>9/2, which is consistent with the observation that saturation occurs later than 7/2. The claimed 'agreement with theory' for r=1 is likewise not established as stated. Section 7.4 and the related comments in Section 8 must be re-evaluated with the corrected threshold.","section":"Section 7.4, Eq. (48)"},{"comment":"The displayed bound ∥v−Pv∥_H ≤ ∥v∥_{(H,A*(V'))_{θ,∞}} ε^θ omits a factor 2 that appears in the proof. The proof yields ∥v−Pv∥_H ≤ 2K(ε,v) ≤ 2∥v∥_{(H,A*(V'))_{θ,∞}} ε^θ. The statement should include this factor, and the same constant should be tracked in the subsequent corollaries (or the text should note explicitly that constants are ignored). As written, the corollary states a stronger inequality than the proof establishes.","section":"Corollary 10, Eq. (23)"}],"minor_comments":[{"comment":"The bound is stated with ∥v∥_{L2(Ω)} on the right-hand side, but the standard argument via identity (13) gives ∥g∥_{L2(Ω)} for v=Tg, i.e., the T(L2)-norm of v. Please verify the exact statement in [37] and in Theorem 11.23 of [42]; if the L2 norm is intended, explain how it is obtained, and otherwise correct Theorem 4 and Corollary 5 accordingly.","section":"Section 2.4, Theorem 4"},{"comment":"There is a typo: 'In thi section' should read 'In this section'.","section":"Section 5, first paragraph"},{"comment":"The label 'TL 1(Ω)' is ambiguous; it should be typeset as T(L^1(Ω)) to avoid confusion with a product or a new space name.","section":"Figure 1"},{"comment":"For integer α, the function x^α is a polynomial and therefore belongs to W^σ_2(Ω) for every σ; the statement 'except for integer values of α, the Sobolev smoothness of f_α scales according to α' is imprecise and should be qualified.","section":"Section 7.1"},{"comment":"The text reports saturation thresholds using the parameter α, while Figure 6 labels the horizontal axis as 'Smoothness α+1/2'. Please state explicitly whether statements such as 'after α = 3/2' refer to the exponent α or to the smoothness value α+1/2, to avoid an apparent mismatch with the figure.","section":"Section 7.4"},{"comment":"The sentence referring to 'a saturation of these rates to values that are larger than predicted by a 1/2 term' is vague and depends on the miscalibrated Section 7.4; it should be rewritten after correcting the periodic experiment.","section":"Section 8"}],"recommendation":"major_revision","confidential_remarks":"The mathematical error in Section 7.4 is a clear false statement that must be fixed; it does not affect the main theorems, but it weakens the paper's numerical-validation claim. The central results are sound and the interpolation-scale extension is a genuine contribution. The paper is likely acceptable after a major revision that corrects the periodic smoothness calculation and the constant in Corollary 10, and updates the discussion of the numerical observations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the core of the paper is in good shape. Theorem 6 is the right abstraction: it replaces the embedding-specific superconvergence of Schaback and Sloan–Kaarnioja with a general bounded operator A and its adjoint, and the proof is short and correct. Corollary 10, extending the gain to the whole interpolation scale between H and A*(V'), is genuinely new and gives the intermediate rates ε^θ. The Mercer power-space characterization (Prop. 13, Cor. 15–16) and the L_p results (Prop. 19, Thm. 25) are solid extensions that go beyond the cited literature. The paper also deserves credit for being upfront about the practical difficulty of checking v ∈ A*(V') and about hidden boundary conditions.\n\nThe soft spots are real but not load-bearing. Corollary 10's statement omits a factor 2 that appears in its own proof; the rate is unchanged, so this is a minor constant error. Bigger is Section 7.4. The paper claims that the random periodic function f_α(x)=1+Σ j^{-α} ξ_j cos(2πjx) is in W^τ_{2,per} with probability one iff τ < α+1/2. That's wrong. The Fourier coefficients decay like j^{-α}, so E||f_α||^2_{W^τ} ~ Σ j^{2τ−2α}, finite iff τ < α−1/2. The threshold is off by a full derivative. The stress-test note suggested this explains the 'unexplained' saturation anomalies in Figure 6, but on reading the paper that doesn't hold: for r=1, the observed saturation at α=3/2 matches the authors' incorrect threshold (since that gives f_α ∈ W^2), while the correct threshold would predict saturation only near α=5/2. So the miscalibration makes the experiment look better than it should, not worse, and the discrepancies the authors flag remain unexplained. They should correct the threshold statement and re-run the experiment.\n\nThe paper still deserves peer review. The main theorems are correct, the framework is useful, and the errors are in a numerical illustration and a constant. A referee should ask for a corrected Corollary 10, a fixed Section 7.4, and ideally public code. If you work in kernel approximation, this is worth your time; the operator-adjoint view is a genuine step forward.","headline":"Solid generalization of superconvergence theory with a clean operator-adjoint framework, but a genuine smoothness-threshold error in the periodic experiment and a missing factor 2 keep it from being fully clean.","tokens_in":28957,"tokens_out":13628,"would_cite":true,"duration_ms":115058,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["41A05","41A25","46E22","46B70","65D05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Kernel interpolation's superconvergence is governed by adjoint-operator ranges, yielding rates anywhere between the classical and doubled extremes.","keywords":["kernel interpolation","superconvergence","reproducing kernel Hilbert space","Mercer operator","Mercer power spaces","real interpolation","Sobolev spaces","boundary conditions"],"falsifier":"Take a kernel whose Green-function/PDE form is known exactly, such as the reproducing kernel of $W^1_2(0,1)$ with the standard inner product, for which $T(L_2)$ consists of functions $u\\in W^2_2(0,1)$ with $u'(0)=u'(1)=0$. Construct $v=Tf$ for a known $f\\in L_2(0,1)$, interpolate $v$ at equally spaced interior points, and measure the $L_2$ and $W^1_2$ errors as the fill distance tends to zero. If the rate fails to improve from $h^{\\tau}$ to $h^{2\\tau}$ (or, for an intermediate member of the interpolation scale, fails to match $\\varepsilon^\\theta$), the central bound is false; conversely, if a function of the same Sobolev smoothness that violates the boundary conditions achieves the doubled rate, the boundary-condition characterization is incomplete.","tokens_in":27935,"feed_emoji":"📈","tokens_out":8563,"duration_ms":79810,"temperature":0.7,"pith_summary":"The paper claims that superconvergence in kernel-based approximation is not a special feature of particular kernels or spaces but a general phenomenon controlled by a single operator-theoretic mechanism: if a target function lies in the range of the adjoint of a bounded linear operator $A$, then the Hilbert-space projection error inherits the factor $\\varepsilon$ that bounds the $A$-error, which doubles the convergence rate. It then interpolates between this special range and the full Hilbert space, showing that every intermediate rate $\\varepsilon^\\theta$, $0\\le\\theta\\le1$, is attained by functions in the corresponding real interpolation space. In the reproducing-kernel setting these intermediate spaces are the Mercer power spaces $H^{1+\\theta}$, so for Sobolev kernels the paper obtains explicit error bounds $h^{(1+\\theta)\\tau-m-d(1/2-1/q)_+}$ in $W_q^m$. A sympathetic reader would care because this unifies earlier doubling results, gives a continuous family of rates instead of only the classical and doubled extremes, and exposes the hidden boundary conditions that decide whether smoother targets actually converge faster.","feed_headline":"Adjoint ranges explain when kernel interpolation speeds up","feed_subtitle":"A general theorem ties improved convergence to operator ranges and interpolation spaces, with hidden boundary conditions as the catch.","key_machinery":"The load-bearing object is the adjoint-range condition $v\\in A^*(V')$ together with the real $K$-functional of interpolation theory. The proof of Theorem 6 uses the identity $\\|v-Pv\\|_H=\\sup_{\\|f\\|_H\\le1}|\\langle v,f-Pf\\rangle_H|$, valid for orthogonal projections, and the duality $\\langle A^*g,f\\rangle_H=g(Af)$; this hands the $\\varepsilon$ from the $A$-error to the $H$-error. Corollary 10 then applies the $K$-functional $K(t,v)=\\inf_{v_0\\in A^*(V')}(\\|v-v_0\\|_H+t\\|v_0\\|_{A^*(V')})$ to extract the fractional rate $\\varepsilon^\\theta$, without any extra structure. In the kernel case, $A$ is the embedding $H(\\Omega)\\hookrightarrow L_2(\\Omega)$, $A^*$ is the Mercer integral operator $T$, and the interpolation spaces are the power spaces $H^{1+\\theta}(\\Omega)$; for Sobolev kernels these are norm-equivalent to Sobolev spaces of fractional smoothness $(1+\\theta)\\tau$.","core_discovery":"The central discovery is Theorem 6: for any Hilbert space $H$, Banach space $V$, bounded linear operator $A:H\\to V$ with adjoint $A^*:V'\\to H$, and orthogonal projection $P$, the error bound $\\|A(f-Pf)\\|_V\\le\\varepsilon\\|f\\|_H$ for all $f\\in H$ implies $\\|v-Pv\\|_H\\le\\|g\\|_{V'}\\,\\varepsilon$ for every $v=A^*g$. Corollary 10 extends this to $v$ in the real interpolation space $(H,A^*(V'))_{\\theta,\\infty}$, with the rate $\\varepsilon^\\theta$. In the Mercer setting the interpolation spaces coincide with the power spaces $H^{1+\\theta}(\\Omega)$, and for Sobolev kernels this yields the Sobolev-norm bound of Corollary 16, interpolating between the standard rate and the doubled rate. The paper also shows that the adjoint of a general operator into $L_p$ is a kernel integral operator, connects the images of such adjoints to Mercer power spaces (Theorem 25), and demonstrates through explicit one-dimensional kernels that norm-equivalent Sobolev RKHSs can demand different boundary conditions for superconvergence.","pith_inferences":["A testable extension is to turn the boundary-condition characterization into a diagnostic: fit the observed interpolation rate and check whether it saturates at the value predicted for $A^*(V')$; a saturation below the doubled rate would indicate the target has left the adjoint range, and this could be used to estimate which boundary conditions a kernel secretly enforces.","The same adjoint-range mechanism suggests that adaptive or greedy approximation schemes, where Theorem 8 applies to a single function, could exhibit superconvergence without the full uniform bound (15); this is a natural place to look for practical rate improvements.","The observed extra $1/2$ in the numerical rates beyond the theory's saturation point hints that the true interpolation spaces between $H(\\Omega)$ and $T(L_2(\\Omega))$ may be larger than the power spaces, or that boundary-condition effects persist past $\\theta=2$; a sharper interpolation characterization would either explain or refute that.","If membership in $A^*(V')$ could be checked algorithmically from kernel derivatives at the boundary, kernel interpolation would become a reliable high-order method for solving elliptic PDEs by collocation, because the improved rates would be guaranteed only for the correct solution space."],"forward_implications":["For Sobolev kernels, functions in the power space $H^{1+\\theta}(\\Omega)$ are approximated in $W_q^m$ at rate $h^{(1+\\theta)\\tau-m-d(1/2-1/q)_+}$, so practitioners can choose any rate between the classical and doubled extremes by controlling how much smoothness and structure they assume.","Functions in $T(L_2(\\Omega))$ get doubled rates in every $W_q^m$ norm covered by Theorem 3, extending the classical $L_2$ doubling to higher-order norms.","General operators $A$ beyond embeddings, including differential operators, produce superconvergence for functions in $A^*(L_p)$, with the adjoint realized as an explicit kernel integral operator.","Two kernels with norm-equivalent Sobolev RKHSs can have different superconvergence subspaces, so switching kernels in an application changes which functions are approximated at improved rates.","On domains without boundary (e.g., the sphere) or for periodic kernels, the power spaces coincide with Sobolev spaces of order $(1+\\theta)\\tau$, so improved rates follow from smoothness alone."],"supporting_citations":[{"why":"Defines the superconvergence setting for kernel interpolation and stresses the role of hidden boundary conditions; the present paper generalizes its operator-based argument.","marker":"[33]"},{"why":"Proves the doubling trick for orthogonal projections in general Hilbert spaces, which is the special case of Theorem 6 that motivates the general result.","marker":"[37]"},{"why":"Supplies the fundamental Sobolev interpolation error bounds (Theorem 3) and the identity used in the classical doubling proof.","marker":"[42]"},{"why":"Establishes the Mercer expansion and the power spaces $H^\\theta(\\Omega)$ that Proposition 13 and Lemma 14 build upon.","marker":"[40]"},{"why":"The original result that interpolation into the range of the Mercer operator doubles the order of convergence.","marker":"[32]"},{"why":"Characterizes adjoints of Sobolev embedding operators as solutions of boundary-value problems, used in Proposition 26 for the boundary-condition analysis.","marker":"[11]"},{"why":"Gives the explicit reproducing kernels of $W^m_2(a,b)$ as Green kernels, providing the one-dimensional examples of boundary conditions.","marker":"[28]"},{"why":"Proves the Hilbert-space interpolation characterization used in Lemma 14 to identify power spaces as real interpolation spaces.","marker":"[4]"},{"why":"Supplies the $K$-functional and real interpolation machinery used in Corollary 10.","marker":"[20]"}],"fun_headline_variants":["Operator ranges explain kernel superconvergence","Boundary conditions gate kernel interpolation speed","Adjoint ranges set kernel convergence rates","Superconvergence hinges on interpolation spaces","Kernel speedup depends on adjoint ranges"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction rests on the target function actually lying in the special subspace $A^*(V')$ (or one of its interpolation spaces); in Sobolev settings that membership is controlled by boundary conditions that the kernel usually keeps hidden.","fun_headline_variants_meta":{"raw":{"variants":["Operator ranges explain kernel superconvergence","Boundary conditions gate kernel interpolation speed","Adjoint ranges set kernel convergence rates","Superconvergence hinges on interpolation spaces","Kernel speedup depends on adjoint ranges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1538,"prompt_tokens":989,"completion_tokens":549,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":484}},"tokens_in":605,"tokens_out":549,"duration_ms":5708,"temperature":1.0,"reasoning_tokens":484,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:54:13.891205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a kernel whose Green-function/PDE form is known exactly, such as the reproducing kernel of $W^1_2(0,1)$ with the standard inner product, for which $T(L_2)$ consists of functions $u\\in W^2_2(0,1)$ with $u'(0)=u'(1)=0$. Construct $v=Tf$ for a known $f\\in L_2(0,1)$, interpolate $v$ at equally spaced interior points, and measure the $L_2$ and $W^1_2$ errors as the fill distance tends to zero. If the rate fails to improve from $h^{\\tau}$ to $h^{2\\tau}$ (or, for an intermediate member of the interpolation scale, fails to match $\\varepsilon^\\theta$), the central bound is false; conversely, if a function of the same Sobolev smoothness that violates the boundary conditions achieves the doubled rate, the boundary-condition characterization is incomplete.","supporting_citations":[{"cited_title":"Schaback","cited_arxiv_id":null,"evidence_quote":"Defines the superconvergence setting for kernel interpolation and stresses the role of hidden boundary conditions; the present paper generalizes its operator-based argument."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves the doubling trick for orthogonal projections in general Hilbert spaces, which is the special case of Theorem 6 that motivates the general result."},{"cited_title":"Wendland","cited_arxiv_id":null,"evidence_quote":"Supplies the fundamental Sobolev interpolation error bounds (Theorem 3) and the identity used in the classical doubling proof."},{"cited_title":"Steinwart and C","cited_arxiv_id":null,"evidence_quote":"Establishes the Mercer expansion and the power spaces $H^\\theta(\\Omega)$ that Proposition 13 and Lemma 14 build upon."},{"cited_title":"Schaback","cited_arxiv_id":null,"evidence_quote":"The original result that interpolation into the range of the Mercer operator doubles the order of convergence."},{"cited_title":"Hubmer, E","cited_arxiv_id":null,"evidence_quote":"Characterizes adjoints of Sobolev embedding operators as solutions of boundary-value problems, used in Proposition 26 for the boundary-condition analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the explicit reproducing kernels of $W^m_2(a,b)$ as Green kernels, providing the one-dimensional examples of boundary conditions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves the Hilbert-space interpolation characterization used in Lemma 14 to identify power spaces as real interpolation spaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the $K$-functional and real interpolation machinery used in Corollary 10."}],"review_version":1}