REVIEW 2 major objections 4 minor 1 cited by
Polynomial Approximation in $ L^2 $ of the Double Exponential via Complex Analysis
T0 review · 2 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash
Pith's one-line read This paper proves a sharp, coefficient-level logarithmic bound for polynomial approximation in $L^2$ of the double-exponential measure, and shows the logarithmic weight cannot be improved.
desk verdict The coefficient inequality is a genuine advance over Lubinsky, but the paper as written contains a load-bearing gap: the claimed reduction from the double-exponential measure μ1 to the hyperbolic-secant measure ν is invalid, because the two densities are not comparable. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The carrying object is the generating function $G_\ell(x,s)=e^{x\arctan s}/(1+s^2)^{\ell/2}$, whose Taylor coefficients are the Meixner-Pollaczek polynomials $P_k^{(\ell)}$, orthogonal for the convolution measures $\nu_\ell$ with densities $|\Gamma((\ell+ix)/2)|^2/(2\pi)$. For $\ell=1$, $\nu$ has density $1/(2\cosh(\pi x/2))$, the smoothed proxy for $\mu_1$. The central mechanism is Lemma 10: for a positive weight $\varphi$, the weighted coefficient sum $\sum_k \Gamma_\varphi(k)\langle f,P_k\rangle^2$ is sandwiched between two integrals over the Fourier transform of $K_u(x)\,(f(x+i)-f(x-i))$, with $K_u(x)=x e^{ux}/\sinh(\pi x/2)$; choosing $\varphi(\varepsilon)=\log^2(\varepsilon)$ makes $\Gamma_\varphi(k)\asymp \log^2(e+k)$. The proof obtains this by mapping the unit disk through $\arctan$ onto the strip $|\operatorname{Re} z|<\pi/4$, using a Paley-Wiener identity and Parseval's formula, and bounding the resulting integrals with elementary hyperbolic-function inequalities.
What would settle it
Evaluate the ratio $e^{-|x|}\big/(1/(2\cosh(\pi x/2)))$ as $|x|\to\infty$; it behaves like $2 e^{(\pi/2-1)|x|}$, so no universal constants make the two densities comparable, directly settling the validity of the reduction step and showing an intermediate comparison is required.
Extended reading notes
Core claim
On the paper's own terms, the central discovery is that the logarithmic decay of polynomial approximation under $\mu_1$ is not merely an error-rate phenomenon but a coefficient phenomenon. Theorem 1 states that $\sum_{k\ge 1}\log^2(e+k)\langle f,P_k\rangle^2 \lesssim \int \log^2(e+|x|) f^2\,d\mu_1 + \int (f')^2\,d\mu_1$, together with a variant where the derivative is weighted by $\log^2(e+|x|)$; when the right-hand side is bounded this gives a logarithmic rate of approximation, recovering Lubinsky's result, while the left-hand side records the full weighted $\ell^1$ summability of the coefficients. The sharpness clause asserts that the weight $\log^2(e+k)$ is optimal: for any positive increasing $a_k\to\infty$ there is a function with bounded right-hand side whose $a_k\log^2(e+k)$-weighted coefficient sum diverges, via a Banach-Steinhaus argument applied to scaled Gaussians. A tensorization theorem transfers the one-dimensional control to $\mu_1^{\otimes d}$, with explicit rates for Lipschitz functions. The authors also conjecture the matching reverse inequality.
Load-bearing premise
The load-bearing premise is that the double-exponential measure and the hyperbolic-secant measure are mutually bounded up to constants, a comparison that fails as written and whose repair is needed for the theorem to transfer.
Editorial extensions
If this is right
- For any absolutely continuous $f$ with the weighted derivative norm finite, the coefficients satisfy $\sum_{k\ge 1}\log^2(e+k)\langle f,P_k\rangle^2<\infty$, a stronger $\ell^1$-type statement than the earlier weak-$\ell^1$ error estimates.
- When the right-hand side is bounded, Theorem 1 yields $E_n(f,\mu_1)\lesssim \log^{-2}(n)$, recovering and refining Lubinsky's logarithmic Jackson-type rate.
- The tensorization theorem gives, for $\mu_1^{\otimes d}$ and $1$-Lipschitz $f$, approximation rates of order $\log\log d/\log n$, showing the rate degrades only logarithmically in dimension.
- The optimality clause rules out any improvement of the logarithm: no sequence $a_k\to\infty$ can multiply $\log^2(e+k)$ while preserving the inequality for all admissible $f$.
- The comparison with Laguerre polynomials on the half line shows the two-sided exponential is the critical case: the one-sided exponential gives a linear rate, so the logarithmic rate is genuinely a two-sided phenomenon.
Reading between the lines
- Extending the same coefficient machinery to the general Meixner-Pollaczek family ($\ell>1$) would yield sharp weighted-sum inequalities for the convolution measures $\nu_\ell$; the paper's proof already contains the $\ell$-dependent generating function.
- A testable repair of the reduction step is to insert the intermediate weight $1/(2\cosh x)$ between $\mu_1$ and $\nu$ and then dilate; the density ratio argument currently needs this extra comparison to be rigorous.
- The tensorization theorem should apply to any product of measures whose orthogonal-polynomial coefficients satisfy weighted Poincaré-type inequalities, so anisotropic weights with distinct $\varphi_i$ and $w_i$ would give dimension-dependent rates beyond the isotropic case.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies L^2 polynomial approximation for the two-sided exponential measure μ1(dx)=e^{-|x|}/2 dx. Its main theorem, Theorem 1, asserts a logarithmic coefficient-weight estimate: for every absolutely continuous f, the weighted sum Σ_{k≥1} log^2(e+k)⟨f,P_k⟩^2 is controlled by ∫ log^2(e+|x|) f^2 dμ1 plus ∫(f')^2 dμ1, with an analogous variant in which the logarithmic weight is moved onto (f')^2. The theorem also claims sharpness in the strong sense that log^2(e+k) cannot be multiplied by a sequence tending to infinity. The proof passes to the Meixner-Pollaczek weight ν(dx)=1/(2 cosh(πx/2)) dx, uses the explicit generating function G_ℓ(x,s)=e^{x arctan s}/(1+s^2)^{ℓ/2}, derives identities expressing weighted coefficient sums as integrals of holomorphic differences, and then proves the desired inequality by Fourier/Laplace estimates and a Poincaré inequality. A tensorization argument yields a d-dimensional corollary with a logarithmic rate for Lipschitz functions.
Significance. The analytic core of the paper—the generating-function identity, the weighted Parseval identities in Lemmas 9 and 10, and the sharpness construction via scaled Gaussians—is explicit, parameter-free, and of independent interest. If the measure-reduction gap is repaired, the result would significantly sharpen Lubinsky's logarithmic error estimate to a coefficient-wise ℓ^1 statement with a sharp weight, and it would give the first tensorizable higher-dimensional rate for the double-exponential measure. The paper contains no fitted parameters and the main coefficient identities are stated in machine-checkable form. However, as written, the proof of Theorem 1 for μ1 is conditional on an invalid comparison between μ1 and ν.
major comments (2)
- [1. Preliminaries] The reduction from μ1 to ν is invalid as written. The paragraph beginning 'Remark that 1/(2 cosh(x)) ≤ exp(−|x|) ≤ 1/(2 cosh(x))' cannot be correct as printed, since the two inequalities are mutually incompatible. More importantly, the measure comparison needed for Lemma 5 and Eq. (8) fails: for ν(dx)=1/(2 cosh(πx/2))dx, the ratio e^{-|x|}/(1/(2 cosh(πx/2))) = 2e^{-|x|}cosh(πx/2) ∼ e^{(π/2−1)|x|} as |x|→∞, so no constants c,C satisfy cν≤μ1≤Cν. Consequently the invocation of Lemma 5 and the chain of inequalities in Eq. (8) do not transfer the theorem from ν to μ1, and the proof of Theorem 1 for μ1 is unsupported. The gap is repairable: first compare μ1 with η(dx)=1/(2 cosh x)dx, where 1/(2 cosh x)≤e^{-|x|}≤2/(2 cosh x), then use the dilation y=πx/2 to pass from η to ν, keeping track of the universal constants in the log(e+|x|) factors and in the orthonormal expansion. This intermediate step must be written explicitly before Lemma 5 is applied.
- [2, Lemma 14] The statement and proof of Lemma 14 are corrupted as printed. The displayed lower bound 'φ(1/k)/(32 + k/2 Φ(1/2k)) ≤ Γφ(k)' does not follow from the proof that follows it, and the proof's interval '[1/2k, 3k/4], which has length k/4' is inconsistent: for k>1 the interval is not contained in [0,1] and its length is not k/4. The intended argument appears to be a lower bound of the form Γφ(k) ≥ φ(1/k)/32 + (k/2)Φ(1/(2k)), with the second interval [1/(2k), 3/(4k)] of length 1/(4k). Since Lemma 14 is the source of the crucial equivalence Γφ(k) ≍ log^2(e+k) used in both the upper and lower bounds, the statement and proof must be corrected.
minor comments (4)
- [1, Eq. (8)] The constants c² and C² in Eq. (8) do not match the comparison cν≤μ≤Cν; using c and C would be consistent, although the argument is unaffected up to renaming constants.
- [2, Lemma 5] In the statement of Lemma 5, 'If Σ bkφk ≤ ∞' should read 'If Σ bkφk < ∞'.
- [3.2, final paragraph] The appeal to the Banach-Steinhaus theorem should specify the Hilbert space (for instance H^1(ν) with norm ∥f∥²=∫ f² dν + ∫(f')² dν) and the bounded linear operators T_N(f)=(√(a_k) log(e+k)⟨f,P_k⟩)_{k≤N}; with that specification the argument is valid, but as written the space and operators are not identified.
- [3.1] The sentence 'We proved (3) and it remains to establish (4)' is placed after the derivation of (3) from Lemma 10; since the following translation argument proves (4), the paragraph would be clearer if the two steps were labeled separately.
Circularity Check
No circularity: the coefficient inequalities are derived from explicit generating functions, with no fitted parameters and no self-citation chain; the known comparability gap is a correctness issue, not a circular one.
full rationale
The derivation chain is self-contained: Theorem 1 for ν is obtained from Lemma 10, which is proved from the explicit generating function G_1(x,s)=exp(x arctan s)/sqrt(1+s^2) and standard Fourier/Laplace identities; no estimate in the proof is assumed rather than derived. The passage from ν back to μ1 uses Lemma 5 and Eq. (8), which would be a legitimate transfer under measure comparability; the paper's assertion that 1/(2 cosh x)≤exp(-|x|)≤1/(2 cosh x) is false as stated, since exp(-|x|)/sech(πx/2) is unbounded as |x|→∞. However, this is an ordinary mathematical error in an auxiliary comparison, not a circular step: the target inequality (3) is not built into the hypotheses, no fitted parameter is renamed as a prediction, and no load-bearing self-citation is used. The cited MP-polynomial background (Szegő and Araaya) supplies the standard integral and generating-function identities; the uniqueness of the polynomials is not used to forbid alternatives. Tightness is established by an explicit family F_λ and Banach-Steinhaus, not by assuming the result. Hence the circularity score is 0, with the correctness gap flagged separately.
Assumptions & free parameters
assumptions (5)
- standard math Summation by parts for positive sequences, used in Lemma 5.
- ad hoc to paper The density comparison c nu <= mu1 <= C nu for nu with density 1/(2 cosh(pi x/2)).
- standard math The standard moment integral integral e^{alpha x} d nu_l(x) = 1/cos^l(alpha) for |alpha| < pi/2, cited from Ara04.
- domain assumption Density of polynomials in L^2(nu_tilde), where nu_tilde has density log^2(e+|x|)/cosh(pi x/2).
- standard math Banach-Steinhaus theorem in the tightness proof.
Cite this review
Pith. "Pith review of Polynomial Approximation in $ L^2 $ of the Double Exponential via Complex Analysis." pith.science (2026). https://pith.science/paper/YERGJSI5
@misc{pith2026250207448,
author = {Pith},
title = {Pith review of: Polynomial Approximation in $ L^2 $ of the Double Exponential via Complex Analysis},
year = {2026},
howpublished = {\url{https://pith.science/paper/YERGJSI5}},
note = {Machine review of arXiv:2502.07448}
}
abstract
We study the polynomial approximation problem in $L^2(\mu_1)$ where $\mu_1(dx) = e^{-|x|}/2 dx$. We show that for any absolutely continuous function $f$, $$ \sum_{k=1}^{\infty} \log^2(e+k) \langle f, P_k \rangle^2 \ \leq C \left( \int_{\mathbb{R}} \log^2(e+\lvert x \rvert) f^2 \, d\mu_1 \ + \ \int_{\mathbb{R}} (f')^2 \, d\mu_1 \right) $$ for some universal constant $C>0$, where $(P_k)_{k \in N}$ are the orthonormal polynomials associated with $\mu_1$. This inequality is tight in the sense that $\log^2(e +k)$ on the left hand-side cannot be replaced by $a_k \log^2(e +k)$ with a sequence $a_k \longrightarrow \infty$. When the right hand-side is bounded this inequality implies a logarithmic rate of approximation for $f$, which was previously obtained by Lubinsky. We also obtain some rates of approximation for the product measure $\mu_1^{\otimes d}$ in $\mathbb{R}^d$ via a tensorization argument. Our proof relies on an explicit formula for the generating function of orthonormal polynomials associated with the weight $\frac{1}{2\cosh(\pi x/2)}$ and some complex analysis.
Figures
Forward citations
Cited by 1 Pith paper
-
Entropy and Learning of Lipschitz Functions under Log-Concave Measures
For regression of 1-Lipschitz functions under log-concave measures with Gaussian-like polynomial approximation, low-degree polynomial estimators achieve the minimax L2 risk of order log d / log n when n is subexponent...
Reference graph
Works this paper leans on
-
[1]
The M eixner- P ollaczek polynomials and a system of orthogonal polynomials in a strip
Tsehaye K Araaya. The M eixner- P ollaczek polynomials and a system of orthogonal polynomials in a strip. Journal of computational and applied mathematics , 170(2):241--254, 2004
work page 2004
-
[2]
S Bernstein. La probl \`e me de l'approximation des fonctions continues sur tout l'axe r \'e el et l'une de ses applications. Bulletin de la Soci \'e t \'e Math \'e matique de France , 52:399--410, 1924
work page 1924
-
[3]
Analysis and geometry of M arkov diffusion operators , volume 103
Dominique Bakry, Ivan Gentil, and Michel Ledoux. Analysis and geometry of M arkov diffusion operators , volume 103. Springer, 2014
work page 2014
- [4]
-
[5]
G. Freud. On M arkov- B ernstein- T ype I nequalities and T heir A pplications. Journal of Approximation Theory , 19(1):22--37, 1977
work page 1977
-
[6]
A.L. Levin and Doron S. Lubinsky. Canonical products and the weights exp( - |x|^ ), > 1 , with applications. Journal of Approximation Theory , 49(2):149--169, 1987
work page 1987
- [7]
-
[8]
Orthogonal polynomials , volume 23
Gabor Szeg. Orthogonal polynomials , volume 23. American Mathematical Soc., 1939
work page 1939
Reviewed August 8, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.