betaFeedback

WorksMathematics

Polynomial Spaces Are Locally Paley-Wiener: Bulk Universality and Christoffel Asymptotics on Symmetric Convex Bodies

Abstract

We prove a bulk universality limit for total-degree polynomial spaces on symmetric convex bodies. For a Bernstein-Markov measure with a positive continuous density near an interior point, the rescaled reproducing kernel converges to the Fourier integral over the dual unit ball of the Baran metric, which we call the local spectral body. After diagonal normalization, the limit is the Fourier transform of the uniform probability measure on this body. The result holds uniformly on compact subsets where the local density assumption holds and yields the corresponding Christoffel asymptotic. Its main constructive ingredient is recovery, with the exact rescaled L2L^2 norm, of Paley-Wiener functions at arbitrary interior centres by Chebyshev ridge polynomials. This extends the classical polynomial-entire function limit at the symmetry centre to a position-dependent local spectrum. Ellipsoidal spectral bodies recover the known Bessel profile on the ball; polytopal spectral bodies give other Fourier profiles. For rectangles, the two-dimensional profile is a product of two sinc functions. On a regular hexagon, the local spectral body changes from a hexagon to a parallelogram across the inscribed circle; explicit Fourier profiles are available in both regimes. The proof uses the Siciak extremal function and does not require a differential operator or an explicit orthogonal-polynomial kernel.

Introduction

Bulk universality describes the microscopic behaviour of reproducing kernels of polynomial spaces. In one variable its basic profile is the sine kernel, sin⁡π(a−b)/(π(a−b))\sin\pi(a-b)/(\pi(a-b)). Riemann–Hilbert methods provide precise asymptotics for varying analytic exponential weights [6], while Lubinsky’s comparison method establishes bulk universality for broad classes of fixed regular measures under local density assumptions [17]. Christoffel asymptotics relate the diagonal kernel to the equilibrium density; see Nevai [18] and Totik [21]. Levin and Lubinsky [16] formulate one-dimensional Christoffel and universality problems in Paley–Wiener spaces. For a recent treatment through tangent measures and chains of de Branges spaces, see Eichinger, Lukić, and Woracek [8].

Kroó and Lubinsky [15] prove multivariate Christoffel and kernel comparison theorems and derive explicit bulk universality limits on the ball and simplex. Their reference kernels use Xu’s formulas [24, 23]; Jacobi Mehler–Heine asymptotics [20] then give normalized Bessel functions. Earlier Christoffel asymptotics for the classical ball and simplex weights were obtained by Xu [22, 25]. Their Theorems 1.1 and 1.5 isolate the transfer from a reference measure to other regular measures, subject to a degree-stability hypothesis. The explicit shape of a reference limit and its transfer between weights are therefore distinct issues. General background on multivariate Christoffel–Darboux kernels is given in Dunkl and Xu [7 Section 3.6].

We identify the established Baran co-ball as the local spectral region and determine the full local kernel for every symmetric convex body. Here symmetric means symmetric about a point. Translation preserves the total-degree filtration, so we normalize the symmetry centre to the origin. Let E=−E⊂RdE=-E\subset\mathbb{R}^{d} be such a body and let

E∗={y∈Rd:t⋅y≤1 for every t∈E}E^{*}=\{y\in\mathbb{R}^{d}:t\cdot y\le1\text{ for every }t\in E\}

be its polar. At x∈int⁡Ex\in\operatorname{int}E, the Baran metric and its dual unit ball are

FE(x;v)=sup⁡y∈∂E∗∣y⋅v∣1−(x⋅y)2,Ax=conv⁡{y1−(x⋅y)2:y∈∂E∗}.hAx(v)=FE(x;v).(1)\begin{aligned} F_{E}(x;v)&=\sup_{y\in\partial E^{*}}\frac{\lvert y\cdot v\rvert}{\sqrt{1-(x\cdot y)^{2}}},\qquad A_{x}=\operatorname{conv}\left\{\frac{y}{\sqrt{1-(x\cdot y)^{2}}}:y\in\partial E^{*}\right\}.\\ h_{A_{x}}(v)&=F_{E}(x;v). \tag*{(1)} \end{aligned}

These geometric formulas are established results. They follow from the Lundin–Baran representation of the extremal function, in the form recorded by Bos, Calvi, and Levenberg [3 Proposition 1.2]; see also Bos, Levenberg, and Waldron [4]. Kroó [14] surveys the broader connection between polynomial inequalities and convex geometry. We call AxA_{x} the local spectral body at xx; its geometry will be referred to as the local spectral shape.

The polynomial–Paley–Wiener connection itself is classical. Ganzburg’s expanding-domain approximation theorem [9 Theorem 5.2] already identifies the polar body E∗E^{*} as the limiting spectrum at the symmetry centre. As a consequence of that theorem at p=2p=2, convergence of best-approximation distances gives strong convergence of the corresponding orthogonal projections; standard local polynomial estimates then yield the full unweighted centre kernel limit. His later sharp-constant theorem [10 Theorem 1.2] gives the centre Christoffel constant explicitly. We do not claim these centre limits as new. The compactness-and-recovery passage to extremal limits likewise belongs to an established method; see, for example, [11].

There are also direct geometric and spectral antecedents. Bedford and Taylor [1] express the equilibrium density through the volume of this dual unit ball; Burns, Levenberg, Ma’u, and Révész [5] extend that formula to general convex bodies. Piazzon [19] identifies polynomial eigenspaces of the Baran Laplacian on the ball and simplex. Berman and Ortega-Cerdà [2] relate real polynomial sampling to Paley–Wiener sampling and pluripotential equilibrium measures. In Section 1.7 of that paper, they also distinguish weak Bergman-measure convergence from microscopic universality and explicitly raise the higher-dimensional local-kernel problem beyond the known examples.

The result proved here concerns every interior centre of an arbitrary centrally symmetric convex body, including bodies with nonsmooth boundary and non-ellipsoidal Baran unit balls. Under global Bernstein–Markov control and a positive continuous local density ww, we prove

n−dKnμ(x+u/n,x+v/n)⟶1(2π)dw(x)∫Axei(u−v)⋅ξ dξn^{-d}K_{n}^{\mu}(x+u/n,x+v/n)\longrightarrow\frac{1}{(2\pi)^{d}w(x)}\int_{A_{x}}e^{i(u-v)\cdot\xi}\,d\xi

uniformly for xx in compact interior sets and u,vu,v in bounded sets. The substantive step is an exact-norm recovery construction at arbitrary xx, where the allowed spectrum varies with position. Translating a centre theorem does not supply this step: translating the base point to zero generally destroys symmetry about zero.

The proof has two parts. The local expansion of the Siciak extremal function bounds the Fourier support of every bounded-energy polynomial limit. Conversely, products of finitely many Chebyshev ridge polynomials, with a smooth amplitude on a degree-allocation simplex, recover a dense class in PW⁡(Ax)\operatorname{PW}(A_x) with the exact rescaled norm. Convergence of finite evaluation-functional norms then identifies the kernel. The diagonal gives the Christoffel limit. The familiar Bessel profile is thereby recovered as the Fourier transform of an ellipsoid; this Fourier identity and the previously known ball and simplex limits are examples of the result’s interpretation, rather than separate novelty claims. Rectangles give a simple non-ellipsoidal profile, while the regular hexagon in section 8.3 shows a more striking phenomenon: the combinatorial type of the local spectral body changes with the interior base point.

Setting and main result

We write ∥⋅∥\lVert\cdot\rVert for the Euclidean norm on Rd\mathbb{R}^d; absolute-value bars denote scalar absolute values. Throughout, E⊂RdE \subset\mathbb{R}^d is a centrally symmetric convex body:

E=−E,0∈E∘,int⁡E≠∅.E=-E,\qquad0\in E^\circ,\qquad\operatorname{int}E\ne\varnothing.

Its polar is E∗={y:x⋅y≤1 for all x∈E}E^*=\{y:x\cdot y\le1\text{ for all }x\in E\}. We write

Πnd={p∈C[x1,…,xd]:deg⁡p≤n},Nn=dim⁡Πnd=(n+dd).\Pi_n^d=\{p\in\mathbb{C}[x_1,\ldots,x_d]:\deg p\le n\},\qquad N_n=\dim\Pi_n^d=\binom{n+d}{d}.

Let μ\mu be a finite positive measure with supp⁡μ=E\operatorname{supp}\mu=E. We assume the following L2L^2 Bernstein–Markov condition: there are constants Mn≥1M_n\ge1 such that

∥p∥L∞(E)≤Mn∥p∥L2(μ),p∈Πnd,log⁡Mn=o(n).(2)\lVert p\rVert_{L^\infty(E)}\le M_n\lVert p\rVert_{L^2(\mu)},\qquad p\in\Pi_n^d,\qquad\log M_n=o(n). \tag*{(2)}

We also assume that on an open set U⊂int⁡EU\subset\operatorname{int}E,

dμ(t)=w(t) dt,w∈C(U),w(t)>0(t∈U).(3)d\mu(t)=w(t)\,dt,\qquad w\in C(U),\qquad w(t)>0\quad(t\in U). \tag*{(3)}

The equality in (3) is an equality of restricted measures, so there is no singular part in UU. The measure may have an arbitrary finite part outside UU, subject to (2).

For x∈int⁡Ex\in\operatorname{int}E, define

gx(y)=y1−(x⋅y)2,y∈E∗.(4)g_x(y)=\frac{y}{\sqrt{1-(x\cdot y)^2}},\qquad y\in E^*. \tag*{(4)}

and let AxA_x be the convex hull defined in the introduction, with support function (1). Since xx is an interior point, ∣x⋅y∣<1|x\cdot y|<1 uniformly for y∈E∗y\in E^*, so AxA_x is a compact, full-dimensional, centrally symmetric convex body.

We use the Fourier transform convention, initially for f∈L1∩L2f\in L^1\cap L^2 and then extended to L2L^2 by Plancherel,

f^(ξ)=∫Rdf(u)e−iu⋅ξ du.\widehat{f}(\xi)=\int_{\mathbb{R}^d}f(u)e^{-iu\cdot\xi}\,du.

For complex arguments, choose a real-valued orthonormal basis {pj,n}j=1Nn\{p_{j,n}\}_{j=1}^{N_n} of Πnd\Pi_n^d in L2(μ)L^2(\mu) and use the holomorphic continuation

Knμ(z,w)=∑j=1Nnpj,n(z)pj,n(w),z,w∈Cd.K_n^\mu(z,w)=\sum_{j=1}^{N_n}p_{j,n}(z)p_{j,n}(w),\qquad z,w\in\mathbb{C}^d.

This agrees with the usual reproducing kernel for real z,wz,w; the Hermitian reproducing kernel on complex arguments is Knμ(z,w‾)K_n^\mu(z,\overline{w}). All complexified kernel statements below refer to the displayed holomorphic continuation. The Christoffel function is

λnμ(x)=1Knμ(x,x)=inf⁡{∫E∣p∣2 dμ:p∈Πnd, p(x)=1}.\lambda_n^\mu(x)=\frac{1}{K_n^\mu(x,x)}=\inf\left\{\int_E |p|^2\,d\mu:p\in\Pi_n^d,\ p(x)=1\right\}.

For a compact convex body A⊂RdA\subset\mathbb{R}^d, the Paley–Wiener space is

PW⁡(A)={f∈L2(Rd):supp⁡f^⊆A},\operatorname{PW}(A)=\{f\in L^2(\mathbb{R}^d):\operatorname{supp}\widehat{f}\subseteq A\},

where each member is identified with its canonical entire extension. It is equipped below with the squared norm w(x)∥f∥22w(x)\lVert f\rVert_2^2.

Theorem 2.1 (Local Paley–Wiener limit). Assume (2)–(3). Let D⋐UD\Subset U be compact and let R>0R>0. Then, uniformly for x∈Dx\in D and u,v∈Rdu,v\in\mathbb{R}^d with ∥u∥,∥v∥≤R\lVert u\rVert,\lVert v\rVert\le R,

1ndKnμ(x+un,x+vn)⟶1(2π)dw(x)∫Axei(u−v)⋅ξ dξ.(5)\frac{1}{n^d}K_n^\mu\left(x+\frac{u}{n},x+\frac{v}{n}\right)\longrightarrow\frac{1}{(2\pi)^d w(x)}\int_{A_x}e^{i(u-v)\cdot\xi}\,d\xi. \tag*{(5)}

The convergence also holds locally uniformly for the complexifications of uu and vv.

Corollary 2.2 (Bulk universality). Under the hypotheses of theorem 2.1, uniformly for xx in compact subsets of UU and for u,vu,v in bounded subsets of Rd\mathbb{R}^d,

Knμ(x+u/n,x+v/n)Knμ(x,x)⟶1vol⁡(Ax)∫Axei(u−v)⋅ξ dξ.(6)\frac{K_n^\mu(x+u/n,x+v/n)}{K_n^\mu(x,x)} \longrightarrow\frac{1}{\operatorname{vol}(A_x)}\int_{A_x}e^{i(u-v)\cdot\xi}\,d\xi. \tag*{(6)}

Proof. Divide eq. (5) by its diagonal specialization u=v=0u=v=0. The limiting diagonal is strictly positive because AxA_x has nonempty interior.

Corollary 2.3 (Christoffel asymptotics). Under the hypotheses of theorem 2.1, uniformly for xx in compact subsets of UU,

Nnλnμ(x)⟶(2π)dd!w(x)vol⁡(Ax).(7)N_n\lambda_n^\mu(x)\longrightarrow\frac{(2\pi)^d}{d!}\frac{w(x)}{\operatorname{vol}(A_x)}. \tag*{(7)}

Proof. The diagonal specialization of the main kernel limit gives the diagonal-kernel asymptotic. The Christoffel limit then follows from the dimension asymptotic and the defining relation between the Christoffel function and the diagonal kernel.

Remark 2.4 (Known phase-volume formula). For the probability-normalized pluripotential equilibrium measure, the established formula of Bedford–Taylor [1], also given in [5 Theorem 1.1], is

ρE(x)=d!(2π)dvol⁡(Ax).(8)\rho_E(x)=\frac{d!}{(2\pi)^d}\operatorname{vol}(A_x). \tag*{(8)}

Thus theorem 2.3 reads Nnλnμ(x)→w(x)/ρE(x)N_n\lambda_n^\mu(x)\to w(x)/\rho_E(x). The volume identity is an established geometric result used here for interpretation. Neither it nor weak convergence of Bergman measures alone identifies the two-point limit in theorem 2.1.

Remark 2.5 (Scope of the theorem). The theorem is an interior result. No smoothness of ∂E\partial E is assumed, and no boundary asymptotics are asserted. The density ww need not be symmetric. The central symmetry assumption is imposed on the domain, where it is needed for the global extremal-function representation.

The local extremal geometry

We first identify the local exponential type of degree-nn polynomials. Let VEV_E denote the Siciak extremal function of EE.

Lemma 3.1 (BCL representation and local blow-up). For a centrally symmetric convex body EE,

VE(z)=sup⁡y∈∂E∗V[−1,1](y⋅z),z∈Cd.(9)V_E(z)=\sup_{y\in\partial E^*}V_{[-1,1]}(y\cdot z),\qquad z\in\mathbb{C}^d. \tag*{(9)}

For every compact D⋐int⁡ED\Subset\operatorname{int}E and every compact C⊂CdC\subset\mathbb{C}^d,

nVE(x+ζn)=hAx(Im⁡ζ)+OD,C(n−1),x∈D, ζ∈C.(10)nV_E\left(x+\frac{\zeta}{n}\right)=h_{A_x}(\operatorname{Im}\zeta)+O_{D,C}(n^{-1}),\qquad x\in D,\ \zeta\in C. \tag*{(10)}

Proof. The identity (9) is the consequence of Lundin’s and Baran’s results stated in Proposition 1.2 of Bos–Calvi–Levenberg [3], after normalizing each nonzero real linear functional y⋅xy\cdot x by max⁡t∈E∣y⋅t∣=1\max_{t\in E}|y\cdot t|=1. This normalization is exactly y∈∂E∗y\in\partial E^*.

For ss in a compact subinterval of (−1,1)(-1,1) and h∈Ch\in\mathbb{C} sufficiently small, the interval extremal function satisfies

V[−1,1](s+h)=∣Im⁡h∣1−s2+O(∣h∣2),(11)V_{[-1,1]}(s+h)=\frac{|\operatorname{Im}h|}{\sqrt{1-s^2}}+O(|h|^2), \tag*{(11)}

uniformly in ss and hh. Indeed, locally V[−1,1](z)=∣Im⁡arccos⁡z∣V_{[-1,1]}(z)=|\operatorname{Im}\arccos z|, and Taylor expansion of the analytic branch of arccos⁡\arccos gives (11).

If x∈Dx\in D, compactness of DD and ∂E∗\partial E^* gives ∣x⋅y∣≤1−δD|x\cdot y|\leq1-\delta_D for some δD>0\delta_D>0. Apply (11) with s=x⋅ys=x\cdot y and h=(y⋅ζ)/nh=(y\cdot\zeta)/n. Uniformly in x∈Dx\in D, y∈∂E∗y\in\partial E^*, and ζ∈C\zeta\in C,

nV[−1,1](x⋅y+y⋅ζn)=∣y⋅Im⁡ζ∣1−(x⋅y)2+O(n−1).nV_{[-1,1]}\left(x\cdot y+\frac{y\cdot\zeta}{n}\right)=\frac{|y\cdot\operatorname{Im}\zeta|}{\sqrt{1-(x\cdot y)^2}}+O(n^{-1}).

Taking the supremum over yy and using (4) yields (10).

Lemma 3.2 (Interior ridge directions generate the co-ball). For every x∈int⁡Ex\in\operatorname{int}E,

Ax=conv⁡{gx(y):y∈int⁡E∗}.A_x=\operatorname{conv}\{g_x(y):y\in\operatorname{int}E^*\}.

Consequently, if C⋐int⁡AxC\Subset\operatorname{int}A_x, there exist finitely many points y1,…,yr∈int⁡E∗y_1,\ldots,y_r\in\operatorname{int}E^*, including dd linearly independent directions and their negatives, such that

C⊂int⁡P⊂P⊂Ax,P=conv⁡{0,gx(y1),…,gx(yr)}.C\subset\operatorname{int}P\subset P\subset A_x,\qquad P=\operatorname{conv}\{0,g_x(y_1),\ldots,g_x(y_r)\}.

Proof. Fix y0∈∂E∗y_0\in\partial E^* and 0<t<10<t<1. Then ty0∈int⁡E∗ty_0\in\operatorname{int}E^*, and

gx(ty0)=t1−(x⋅y0)21−t2(x⋅y0)2gx(y0).g_x(ty_0)=\frac{t\sqrt{1-(x\cdot y_0)^2}}{\sqrt{1-t^2(x\cdot y_0)^2}}g_x(y_0).

The scalar factor belongs to (0,1)(0,1) and tends to 11 as t↑1t\uparrow1. Since 0∈Ax0\in A_x and AxA_x is convex, gx(ty0)∈Axg_x(ty_0)\in A_x and converges to gx(y0)g_x(y_0). This proves the closure identity.

For the finite assertion, choose 0<ρ<ρ′<10<\rho<\rho'<1 such that C⊂ρAxC\subset\rho A_x. Since AxA_x contains a ball about zero, uniform approximation of its support function by finite subsets of the generating set gives a polytope PP with hP≥ρ′hAxh_P\geq\rho'h_{A_x} and P⊂AxP\subset A_x. Therefore C⊂ρAx⊂int⁡(ρ′Ax)⊂int⁡PC\subset\rho A_x\subset\operatorname{int}(\rho'A_x)\subset\operatorname{int}P. Add independent interior directions and their negatives, ordering the first dd to be independent. This does not shrink PP.

Recovery: Chebyshev ridge wave packets

This section proves that the entire Paley–Wiener space is locally recoverable by total-degree polynomials. The construction is explicit and does not use an orthogonal basis adapted to EE.

A smooth lifting lemma

Let y1,…,yry_1,\ldots,y_r be chosen as in theorem 3.2, and abbreviate gj=gx(yj)g_j=g_x(y_j). Define

B:Rr→Rd,Bλ=∑j=1rλjgj.B:\mathbb{R}^r\to\mathbb{R}^d,\qquad B\lambda=\sum_{j=1}^{r}\lambda_jg_j.

and the open frequency-allocation simplex

Δr={λ∈Rr:λj>0, λ1+⋯+λr<1}.\Delta_r=\{\lambda\in\mathbb{R}^r:\lambda_j>0,\ \lambda_1+\cdots+\lambda_r<1\}.

Since BB has rank dd, B(Δr)=int⁡PB(\Delta_r)=\operatorname{int}P.

Lemma 4.1 (Smooth amplitude lifting). For every b∈Cc∞(int⁡P)b\in C^\infty_c(\operatorname{int}P) there exists a∈Cc∞(Δr)a\in C^\infty_c(\Delta_r) such that

B∗(a(λ) dλ)=b(ξ) dξ.(12)B_*\bigl(a(\lambda)\,d\lambda\bigr)=b(\xi)\,d\xi. \tag*{(12)}

Equivalently, for every continuous φ\varphi,

∫Δra(λ)φ(Bλ) dλ=∫Rdb(ξ)φ(ξ) dξ.\int_{\Delta_r}a(\lambda)\varphi(B\lambda)\,d\lambda=\int_{\mathbb{R}^d}b(\xi)\varphi(\xi)\,d\xi.

Proof. Choose a linear right inverse S:Rd→RrS:\mathbb{R}^d\to\mathbb{R}^r of BB and a linear isomorphism R:Rr−d→ker⁡BR:\mathbb{R}^{r-d}\to\ker B. The map

T(ξ,η)=Sξ+RηT(\xi,\eta)=S\xi+R\eta

is an invertible linear map from Rd×Rr−d\mathbb{R}^d\times\mathbb{R}^{r-d} to Rr\mathbb{R}^r, and BT(ξ,η)=ξBT(\xi,\eta)=\xi. Write dλ=J dξ dηd\lambda=J\,d\xi\,d\eta, where J=∣det⁡T∣J=\lvert\det T\rvert.

For every ξ0∈supp⁡b⋐int⁡P\xi_0\in\operatorname{supp}b\Subset\operatorname{int}P, the fiber B−1(ξ0)B^{-1}(\xi_0) meets Δr\Delta_r. Hence there are neighbourhoods Uξ0U_{\xi_0} of ξ0\xi_0 and Vξ0V_{\xi_0} in Rr−d\mathbb{R}^{r-d} such that T(Uξ0×Vξ0)⋐ΔrT(U_{\xi_0}\times V_{\xi_0})\Subset\Delta_r. By compactness, choose finitely many such product neighbourhoods Uℓ×VℓU_\ell\times V_\ell covering supp⁡b\operatorname{supp}b. Take a smooth partition of unity {ψℓ}\{\psi_\ell\} near supp⁡b\operatorname{supp}b, with supp⁡ψℓ⋐Uℓ\operatorname{supp}\psi_\ell\Subset U_\ell, and choose χℓ∈Cc∞(Vℓ)\chi_\ell\in C^\infty_c(V_\ell) with ∫χℓ=1\int\chi_\ell=1.

Define, in the (ξ,η)(\xi,\eta) coordinates,

a(T(ξ,η))=J−1b(ξ)∑ℓψℓ(ξ)χℓ(η).a(T(\xi,\eta))=J^{-1}b(\xi)\sum_\ell\psi_\ell(\xi)\chi_\ell(\eta).

and extend by zero. The support is compactly contained in Δr\Delta_r, so a∈Cc∞(Δr)a\in C^\infty_c(\Delta_r). Integrating first in η\eta gives

∫a(λ)φ(Bλ) dλ=∫b(ξ)φ(ξ)∑ℓψℓ(ξ) dξ=∫b(ξ)φ(ξ) dξ.\int a(\lambda)\varphi(B\lambda)\,d\lambda =\int b(\xi)\varphi(\xi)\sum_\ell\psi_\ell(\xi)\,d\xi =\int b(\xi)\varphi(\xi)\,d\xi.

The wave-packet construction

For yj∈int⁡E∗y_j \in\operatorname{int} E^*, define on the real body EE

θj(t)=arccos⁡(yj⋅t).\theta_j(t) = \arccos(y_j \cdot t).

There is εj>0\varepsilon_j > 0 such that

εj≤θj(t)≤π−εj,t∈E.(13)\varepsilon_j \le\theta_j(t) \le\pi- \varepsilon_j,\qquad t \in E. \tag*{(13)}

Moreover,

∇θj(x)=−gj,θj(x+un)=θj(x)−gj⋅un+O(n−2),(14)\nabla\theta_j(x) = -g_j,\qquad\theta_j\left(x+\frac{u}{n}\right) = \theta_j(x) - \frac{g_j \cdot u}{n} + O(n^{-2}), \tag*{(14)}

uniformly for uu in bounded sets.

Let b∈Cc∞(int⁡Ax)b \in C^\infty_c(\operatorname{int} A_x). Choose the finite directions so that supp⁡b⋐int⁡P\operatorname{supp} b \Subset\operatorname{int} P, and let aa be given by theorem 4.1. Set Θ(x)=(θ1(x),…,θr(x))\Theta(x) = (\theta_1(x),\ldots,\theta_r(x)) and define

Pn(t)=2rn−r∑k∈Zra(k/n)≠0a(k/n)eik⋅Θ(x)∏j=1rTkj(yj⋅t).(15)P_n(t) = 2^r n^{-r} \sum_{\substack{k\in\mathbb{Z}^r\\a(k/n)\ne0}} a(k/n)e^{ik\cdot\Theta(x)}\prod_{j=1}^{r}T_{k_j}(y_j\cdot t). \tag*{(15)}

Since aa is compactly supported in Δr\Delta_r, every nonzero summand has kj≥1k_j \ge1 and ∣k∣1≤(1−η)n|k|_1 \le(1-\eta)n for a fixed η>0\eta> 0. Thus no negative-index Chebyshev polynomial occurs. Thus

Pn∈Πnd.(16)P_n \in\Pi_n^d. \tag*{(16)}

Lemma 4.2 (Poisson estimate). For a∈Cc∞(Rr)a \in C^\infty_c(\mathbb{R}^r), define

Sn(t)=n−r∑k∈Zra(k/n)eik⋅t.S_n(t) = n^{-r}\sum_{k\in\mathbb{Z}^r}a(k/n)e^{ik\cdot t}.

For every M>0M > 0 there is CMC_M such that

∣Sn(t)∣≤CM(1+ndist⁡(t,2πZr))−M,t∈Rr.(17)|S_n(t)| \le C_M(1+n\operatorname{dist}(t,2\pi\mathbb{Z}^r))^{-M},\qquad t\in\mathbb{R}^r. \tag*{(17)}

Proof. Poisson summation gives

Sn(t)=∑ℓ∈Zra^(n(2πℓ−t)),S_n(t) = \sum_{\ell\in\mathbb{Z}^r}\widehat{a}\bigl(n(2\pi\ell-t)\bigr),

with a^(η)=∫a(λ)e−iλ⋅η dλ\widehat{a}(\eta)=\int a(\lambda)e^{-i\lambda\cdot\eta}\,d\lambda. Since aa is smooth and compactly supported, a^\widehat{a} is rapidly decreasing. Summing the resulting lattice estimates yields (17).

Lemma 4.3 (Global localization of the wave packet). For every M>0M > 0 there is CMC_M, independent of nn and t∈Et \in E, such that

∣Pn(t)∣≤CM(1+n∥t−x∥)−M+CMn−M.(18)|P_n(t)| \le C_M(1+n\lVert t-x\rVert)^{-M}+C_Mn^{-M}. \tag*{(18)}

Proof. Use

2Tk(s)=eikarccos⁡s+e−ikarccos⁡s.2T_k(s)=e^{ik\arccos s}+e^{-ik\arccos s}.

Expanding the product in (15) gives 2r2^r phase branches indexed by σ∈{−1,1}r\sigma\in\{-1,1\}^r, with phase vector

tjσ(t)=σjθj(t)+θj(x).t_j^\sigma(t)=\sigma_j\theta_j(t)+\theta_j(x).

The branch σ=(−1,…,−1)\sigma= (-1,\ldots,-1) is the main branch. Its phase coordinates are θj(x)−θj(t)\theta_j(x)-\theta_j(t). By (13), these differences lie strictly between −π-\pi and π\pi, so their nearest lattice point in 2πZ2\pi\mathbb{Z} is zero. For the first dd directions, which are linearly independent,

∣θj(t)−θj(x)∣≥∣yj⋅(t−x)∣.|\theta_j(t)-\theta_j(x)| \ge|y_j\cdot(t-x)|.

Indeed, ∣ddsarccos⁡s∣≥1|\frac{d}{ds}\arccos s|\ge1 on (−1,1)(-1,1). If YY is the matrix with rows y1T,…,ydTy_1^T,\ldots,y_d^T, then ∥Y(t−x)∥≥c∥t−x∥\|Y(t-x)\|\ge c\|t-x\|. Hence

dist⁡(t−σ(t),2πZr)≥c∥t−x∥.\operatorname{dist}(t^{-\sigma}(t),2\pi\mathbb{Z}^r)\ge c\|t-x\|.

Lemma 4.2 bounds the main branch by CM(1+n∥t−x∥)−MC_M(1+n\|t-x\|)^{-M}.

For every other branch, at least one coordinate has σj=1\sigma_j=1, so

tjσ(t)=θj(t)+θj(x)∈[2εj,2π−2εj].t_j^\sigma(t)=\theta_j(t)+\theta_j(x)\in[2\varepsilon_j,2\pi-2\varepsilon_j].

That coordinate stays a fixed positive distance from 2πZ2\pi\mathbb{Z}. Lemma 4.2 therefore bounds every non-main branch by CMn−MC_Mn^{-M}. Summing the finitely many branches proves (18).

Theorem 4.4 (Ridge–Chebyshev recovery). Let b∈Cc∞(int⁡Ax)b\in C^\infty_{\mathrm{c}}(\operatorname{int} A_x) and define

fb(u)=∫Axb(ξ)eiu⋅ξ dξ.f_b(u)=\int_{A_x}b(\xi)e^{iu\cdot\xi}\,d\xi.

There exist Pn∈ΠndP_n\in\Pi_n^d such that

Pn(x+un)⟶fb(u)(19)P_n\left(x+\frac{u}{n}\right)\longrightarrow f_b(u) \tag*{(19)}

locally uniformly for u∈Rdu\in\mathbb{R}^d, and

nd∫E∣Pn(t)∣2 dμ(t)⟶(2π)dw(x)∫Ax∣b(ξ)∣2 dξ.(20)n^d\int_E|P_n(t)|^2\,d\mu(t)\longrightarrow(2\pi)^d w(x)\int_{A_x}|b(\xi)|^2\,d\xi. \tag*{(20)}

Proof. The degree bound is (16). Expanding the Chebyshev factors, the main branch is

Sn(Θ(x)−Θ(t)),S_n(\Theta(x)-\Theta(t)),

where SnS_n is as in theorem 4.2. By (14), for t=x+u/nt=x+u/n,

Θ(x)−Θ(x+un)=1n(g1⋅u,…,gr⋅u)+O(n−2).\Theta(x)-\Theta\left(x+\frac{u}{n}\right)=\frac{1}{n}(g_1\cdot u,\ldots,g_r\cdot u)+O(n^{-2}).

Consequently, the main branch converges, by a Riemann-sum argument, to

∫Ara(λ)ei(Bλ)⋅u dλ=∫Axb(ξ)eiu⋅ξ dξ,\int_{A_r}a(\lambda)e^{i(B\lambda)\cdot u}\,d\lambda=\int_{A_x}b(\xi)e^{iu\cdot\xi}\,d\xi,

where the last equality is (12). The other branches tend to zero locally uniformly by the proof of theorem 4.3. This proves (19).

For the norm, choose a neighbourhood Ux⋐UU_x\Subset U on which dμ=w(t) dtd\mu=w(t)\,dt. Split the integral into UxU_x and E∖UxE\setminus U_x. On the complement, the distance from xx is positive, and (18) gives a contribution O(nd−2M)O(n^{d-2M}), which tends to zero for 2M>d2M>d. On UxU_x, set u=n(t−x)u=n(t-x). Then

nd∫Ux∣Pn(t)∣2w(t) dt=∫n(Ux−x)∣Pn(x+un)∣2w(x+un) du.n^d\int_{U_x}|P_n(t)|^2w(t)\,dt =\int_{n(U_x-x)}\left|P_n\left(x+\frac{u}{n}\right)\right|^2w\left(x+\frac{u}{n}\right)\,du.

The localization estimate gives an integrable majorant after choosing 2M>d2M>d; the residual n−Mn^{-M} term contributes O(nd−2M)O(n^{d-2M}). Dominated convergence and (19) yield

w(x)∫Rd∣fb(u)∣2 du.w(x)\int_{\mathbb{R}^{d}} |f_b(u)|^2\,du.

Finally, Plancherel’s theorem gives

∫Rd∣fb(u)∣2 du=(2π)d∫Ax∣b(ξ)∣2 dξ,\int_{\mathbb{R}^{d}} |f_b(u)|^2\,du=(2\pi)^d\int_{A_x}|b(\xi)|^2\,d\xi,

which proves (20).

Remark 4.5 (Density of the recovery class). The functions fbf_b, b∈Cc∞(int⁡Ax)b\in C^\infty_c(\operatorname{int} A_x), form a dense subspace of PW⁡(Ax)\operatorname{PW}(A_x) in L2(Rd)L^2(\mathbb{R}^d). Indeed, ∂Ax\partial A_x has Lebesgue measure zero and Cc∞(int⁡Ax)C^\infty_c(\operatorname{int} A_x) is dense in L2(Ax)L^2(A_x).

Compactness of rescaled polynomial spaces

We now prove the reverse inclusion: every bounded-energy local limit belongs to PW⁡(Ax)\operatorname{PW}(A_x).

Proposition 5.1 (Local compactness). Let x∈Ux\in U, and let pn∈Πndp_n\in\Pi_n^d satisfy

sup⁡nnd∥pn∥L2(μ)2<∞.(21)\sup_n n^d\|p_n\|_{L^2(\mu)}^2<\infty. \tag*{(21)}

Then the rescaled functions qn(u)=pn(x+u/n)q_n(u)=p_n(x+u/n) have a subsequence converging locally uniformly on Cd\mathbb{C}^d to an entire function ff. Every such limit satisfies

w(x)∥f∥L2(Rd)2≤lim inf⁡j→∞njd∥pnj∥L2(μ)2,(22)w(x)\|f\|_{L^2(\mathbb{R}^d)}^2\leq\liminf_{j\to\infty}n_j^d\|p_{n_j}\|_{L^2(\mu)}^2, \tag*{(22)}

where qnj→fq_{n_j}\to f is the chosen convergent subsequence, and

supp⁡f^⊆Ax.(23)\operatorname{supp}\widehat f\subseteq A_x. \tag*{(23)}

The same assertion holds when nn is restricted at the outset to any increasing subsequence.

Proof. The proof is divided into three steps.

Step 1: local normality. Choose nested cubes Q1⋐Q0⋐UQ_1\Subset Q_0\Subset U around xx. Since ww is positive and continuous, w≥c>0w\geq c>0 on Q0Q_0. By (21),

∫Q0∣pn(t)∣2 dt≤Cn−d.\int_{Q_0}|p_n(t)|^2\,dt\leq Cn^{-d}.

Here is a direct interior evaluation estimate. On Id=[−1,1]dI^d=[-1,1]^d, set R(t)=p(t)∏j(1−tj2)R(t)=p(t)\prod_j(1-t_j^2) and use the probability measure dν=π−d∏j(1−tj2)−1/2 dtd\nu=\pi^{-d}\prod_j(1-t_j^2)^{-1/2}\,dt. Then

∥R∥L2(ν)2≤π−d∥p∥L2(Id)2.\|R\|_{L^2(\nu)}^2\leq\pi^{-d}\|p\|_{L^2(I^d)}^2.

The univariate orthonormal system is 1,2T1,2T2,…1,\sqrt{2}T_1,\sqrt{2}T_2,\ldots, whose diagonal kernel through degree n+2n+2 is at most 2n+52n+5. Tensor-product Cauchy–Schwarz gives

∣p(s)∣2≤(2n+5)dπd∏j(1−sj2)2∥p∥L2(Id)2,s∈(−1,1)d.|p(s)|^2\leq\frac{(2n+5)^d}{\pi^d\prod_j(1-s_j^2)^{2}}\|p\|_{L^2(I^d)}^2,\qquad s\in(-1,1)^d.

Affine rescaling to the nested cubes, followed by the preceding integral bound, gives

∥pn∥L∞(Q1)≤C.(24)\lVert p_n\rVert_{L^\infty(Q_1)} \le C. \tag*{(24)}

Applying the one-variable Bernstein–Walsh inequality successively in the coordinates of Q1Q_1 bounds ∣pn(z)∣|p_n(z)| by Cexp⁡(n∑jVIj(zj))C\exp\left(n\sum_j V_{I_j}(z_j)\right), where Q1=∏jIjQ_1=\prod_j I_j. The interval expansion (11) therefore implies that

sup⁡ζ∈C∣pn(x+ζn)∣≤CC\sup_{\zeta\in C}\left|p_n\left(x+\frac{\zeta}{n}\right)\right|\le CC

for every compact C⊂CdC\subset\mathbb{C}^d. Thus qnq_n is a normal family. Montel’s theorem gives a locally uniformly convergent subsequence with entire limit ff.

Step 2: the L2L^2 bound. For every fixed R>0R>0 and all sufficiently large nn, the set x+BR/nx+B_R/n lies in UU. Changing variables gives

nd∫x+BR/n∣pn(t)∣2 dμ(t)=∫BR∣qn(u)∣2w(x+un) du.n^d\int_{x+B_R/n}|p_n(t)|^2\,d\mu(t)=\int_{B_R}|q_n(u)|^2w\left(x+\frac{u}{n}\right)\,du.

Along the chosen convergent subsequence, local uniform convergence and continuity of ww imply

w(x)∫BR∣f(u)∣2 du≤lim inf⁡j→∞njd∥pnj∥L2(μ)2.w(x)\int_{B_R}|f(u)|^2\,du\le\liminf_{j\to\infty}n_j^d\lVert p_{n_j}\rVert_{L^2(\mu)}^2.

Letting R→∞R\to\infty proves (22), and in particular f∣Rd∈L2(Rd)f|_{\mathbb{R}^d}\in L^2(\mathbb{R}^d).

Step 3: Fourier support. The local bound (24) is not a global bound on EE. By (2) and (21),

∥pn∥L∞(E)≤CMnn−d/2.\lVert p_n\rVert_{L^\infty(E)}\le CM_nn^{-d/2}.

Choose

mn=⌈n(1+log⁡Mn)⌉.m_n=\left\lceil\sqrt{n(1+\log M_n)}\right\rceil.

Then mn=o(n)m_n=o(n) and log⁡Mn/mn→0\log M_n/m_n\to0. Choose ε>0\varepsilon>0 so small that

0<1−ε∥t−x∥2≤1,t∈E,0<1-\varepsilon\lVert t-x\rVert^2\le1,\qquad t\in E,

and define the polynomial localizer

rn(z)=(1−ε∑j=1d(zj−xj)2)mn.r_n(z)=\left(1-\varepsilon\sum_{j=1}^d(z_j-x_j)^2\right)^{m_n}.

On the real body,

0<rn(t)≤e−εmn∥t−x∥2.0<r_n(t)\le e^{-\varepsilon m_n\lVert t-x\rVert^2}.

Near xx, (24) bounds pnp_n uniformly. Away from a fixed neighbourhood of xx, the decay of rnr_n dominates MnM_n, because log⁡Mn=o(mn)\log M_n=o(m_n). Hence

∥pnrn∥L∞(E)≤C.(25)\lVert p_nr_n\rVert_{L^\infty(E)}\le C. \tag*{(25)}

Also, deg⁡(pnrn)≤n+2mn=n(1+o(1))\deg(p_nr_n)\le n+2m_n=n(1+o(1)), and

rn(x+ζn)=(1−εn2∑j=1dζj2)mn⟶1r_n\left(x+\frac{\zeta}{n}\right)=\left(1-\frac{\varepsilon}{n^2}\sum_{j=1}^d\zeta_j^2\right)^{m_n}\longrightarrow1

locally uniformly on Cd\mathbb{C}^d. Thus pnrnp_nr_n has the same local limit ff as pnp_n. The Bernstein–Walsh inequality now gives

∣pnrn(x+ζn)∣≤Cexp⁡((n+2mn)VE(x+ζn)).\left|p_nr_n\left(x+\frac{\zeta}{n}\right)\right| \le C\exp\left((n+2m_n)V_E\left(x+\frac{\zeta}{n}\right)\right).

By theorem 3.1,

∣f(ζ)∣≤Cexp⁡(hAx(Im⁡ζ)),ζ∈Cd.|f(\zeta)| \le C\exp\left(h_{A_x}(\operatorname{Im}\zeta)\right), \qquad\zeta\in\mathbb{C}^d.

The constant CC is the same for all ζ\zeta: it comes from (25), not from a compact-dependent normal-family bound. The L2L^2 Paley–Wiener theorem for compact convex spectral sets (see, e.g., [13]) therefore implies (23). This completes the proof.

Remark 5.2 (Why the localizer is sufficient). The multiplier rnr_n is a simple quadratic polynomial raised to a sublinear power. Its degree is o(n)o(n), so it does not change the limiting exponential type, while its real restriction decays like e−cmn∥t−x∥2e^{-cm_n\lVert t-x\rVert^2}. The only global growth to suppress is the subexponential Bernstein–Markov factor. No boundary regularity or needle-polynomial construction is needed.

Convergence of the local reproducing kernels

We now use the standard two-sided extremal argument associated with recovery and compactness; compare the general framework in [11]. The additional polarization step identifies all finite kernel matrices. For fixed x∈Ux\in U, define

Hn,x={u↦p(x+u/n):p∈Πnd}.\mathcal{H}_{n,x}=\{u\mapsto p(x+u/n):p\in\Pi_n^d\}.

with norm

∥q∥n,x2=nd∫E∣p(t)∣2 dμ(t),q(u)=p(x+u/n).(26)\|q\|_{n,x}^2=n^d\int_E|p(t)|^2\,d\mu(t),\qquad q(u)=p(x+u/n). \tag*{(26)}

Its reproducing kernel is

kn,x(u,v)=n−dKnμ(x+u/n,x+v/n).(27)k_{n,x}(u,v)=n^{-d}K_n^\mu(x+u/n,x+v/n). \tag*{(27)}

The candidate limit is Hx=PW⁡(Ax)\mathcal{H}_x=\operatorname{PW}(A_x) with norm

∥f∥x2=w(x)∫Rd∣f(u)∣2 du.\|f\|_x^2=w(x)\int_{\mathbb{R}^d}|f(u)|^2\,du.

Using the Fourier representation f(u)=∫Axb(ξ)eiu⋅ξ dξf(u)=\int_{A_x}b(\xi)e^{iu\cdot\xi}\,d\xi, Plancherel gives

∥f∥x2=(2π)dw(x)∫Ax∣b(ξ)∣2 dξ.\|f\|_x^2=(2\pi)^dw(x)\int_{A_x}|b(\xi)|^2\,d\xi.

Its reproducing kernel is

kx(u,v)=1(2π)dw(x)∫Axei(u−v)⋅ξ dξ.(28)k_x(u,v)=\frac{1}{(2\pi)^dw(x)}\int_{A_x}e^{i(u-v)\cdot\xi}\,d\xi. \tag*{(28)}

Lemma 6.1 (Convergence of evaluation functionals). Let u1,…,um∈Rdu_1,\ldots,u_m\in\mathbb{R}^d and c1,…,cm∈Cc_1,\ldots,c_m\in\mathbb{C}. Define L(f)=∑j=1mcjf(uj)L(f)=\sum_{j=1}^m c_jf(u_j). Then

∥L∥Hn,x∗⟶∥L∥Hx∗.\|L\|_{\mathcal{H}_{n,x}^*}\longrightarrow\|L\|_{\mathcal{H}_x^*}.

Proof. The local evaluation bound from Step 1 of theorem 5.1 makes the dual norms uniformly bounded. For the upper bound, first select a subsequence along which these norms tend to their limsup; relabel it by nn. Choose near-extremizers qn∈Hn,xq_n \in\mathcal{H}_{n,x} with ∥qn∥n,x≤1\lVert q_n\rVert_{n,x} \le1 and ∣L(qn)∣≥∥L∥Hn,x∗−o(1)\lvert L(q_n)\rvert\ge\lVert L\rVert_{\mathcal{H}_{n,x}^{*}} - o(1). Proposition 5.1 gives a subsequence converging locally uniformly to q∈Hxq \in\mathcal{H}_x with ∥q∥x≤1\lVert q\rVert_x \le1. Hence

lim sup⁡n→∞∥L∥Hn,x∗≤∥L∥Hx∗.\limsup_{n\to\infty}\lVert L\rVert_{\mathcal{H}_{n,x}^{*}} \le\lVert L\rVert_{\mathcal{H}_x^{*}}.

For the lower bound, take any nonzero ff in the dense recovery class from theorem 4.4. There are qn∈Hn,xq_n \in\mathcal{H}_{n,x} with qn→fq_n \to f locally uniformly and ∥qn∥n,x→∥f∥x\lVert q_n\rVert_{n,x} \to\lVert f\rVert_x. Thus

lim inf⁡n→∞∥L∥Hn,x∗≥∣L(f)∣∥f∥x.\liminf_{n\to\infty}\lVert L\rVert_{\mathcal{H}_{n,x}^{*}} \ge\frac{\lvert L(f)\rvert}{\lVert f\rVert_x}.

Taking the supremum over the dense recovery class gives the reverse inequality.

Proof of theorem 2.1. For any finite set of real points u1,…,umu_1,\ldots,u_m, the squared norm of a linear combination of evaluations is the corresponding quadratic form in the kernel matrix (kn,x(uj,uk))j,k(k_{n,x}(u_j,u_k))_{j,k}. By theorem 6.1, these quadratic forms converge to those of (kx(uj,uk))j,k(k_x(u_j,u_k))_{j,k}. Polarization therefore gives

kn,x(u,v)⟶kx(u,v),u,v∈Rd.k_{n,x}(u,v) \longrightarrow k_x(u,v), \qquad u,v \in\mathbb{R}^d.

We upgrade this to local uniform convergence. If q∈Hn,xq \in\mathcal{H}_{n,x} has norm at most one, then its representing polynomial has L2(μ)L^2(\mu) norm at most n−d/2n^{-d/2}. The local Nikolskii estimate and the Bernstein–Walsh inequality on a fixed interior cube, as in Step 1 of theorem 5.1, show that evaluation at any complex point in a fixed compact set has uniformly bounded norm on Hn,x\mathcal{H}_{n,x}. Cauchy–Schwarz therefore gives local uniform bounds for the complexified kernels. They form a normal family. Every subsequential limit agrees with kxk_x on Rd×Rd\mathbb{R}^d \times\mathbb{R}^d, and the identity theorem in each variable identifies the limit uniquely with (28). Thus convergence is locally uniform in the complex variables, in particular uniformly for real u,vu,v in compact sets.

For uniformity in xx on compact subsets of UU, argue by contradiction. Suppose that there are nj→∞n_j \to\infty, xj∈Dx_j \in D, and bounded uj,vju_j,v_j for which the asserted convergence fails. After passing to a subsequence, assume xj→x∞∈Dx_j \to x_\infty\in D and uj→uu_j \to u, vj→vv_j \to v.

The compactness argument is uniform for these moving centres. Since D⋐UD \Subset U, choose a fixed r>0r>0 such that the cubes Q0,j=xj+[−2r,2r]dQ_{0,j}=x_j+[-2r,2r]^d and Q1,j=xj+[−r,r]dQ_{1,j}=x_j+[-r,r]^d lie in UU for every jj. The density ww has a uniform positive lower bound on their union. The elementary cube estimate, the local Bernstein–Walsh bound, and the polynomial localizer centred at xjx_j therefore have constants independent of jj. The uniform extremal expansion in theorem 3.1 then gives the Fourier-support constraint with limiting body Ax∞A_{x_\infty}.

For recovery, take b∈Cc∞(int⁡Ax∞)b \in C^\infty_c(\operatorname{int} A_{x_\infty}) and choose the finite directions y1,…,yry_1,\ldots,y_r and lifting amplitude aa at x∞x_\infty. For each jj, use the same directions and amplitude, but centre the wave packet at xjx_j, with phases θℓ(xj)\theta_\ell(x_j). Since gxj(yℓ)→gx∞(yℓ)g_{x_j}(y_\ell) \to g_{x_\infty}(y_\ell), the main Riemann sums converge locally uniformly to

∫Δra(λ)ei(Bx∞λ)⋅u dλ=∫Ax∞b(ξ)eiξ⋅u dξ.\int_{\Delta_r} a(\lambda)e^{i(B_{x_\infty}\lambda)\cdot u}\,d\lambda= \int_{A_{x_\infty}} b(\xi)e^{i\xi\cdot u}\,d\xi.

The global localization constants are uniform in jj: the directions are fixed in int⁡Ex∞∗\operatorname{int} E_{x_\infty}^{*}, and the first dd remain linearly independent. The same dominated-convergence argument gives convergence of the rescaled norms, with w(xj)→w(x∞)w(x_j) \to w(x_\infty). Thus recovery also holds along the moving centres.

The compactness and recovery arguments now give convergence of all finite evaluation Gram matrices to those of the limiting Paley–Wiener space at x∞x_\infty. Finally, x↦Axx \mapsto A_x is continuous in the Hausdorff metric, because gx(y)g_x(y) is jointly continuous on D×∂E∗D \times\partial E^*. Convex bodies with nonempty interior have boundary of Lebesgue measure zero; Hausdorff convergence here therefore gives 1Axj→1Ax∞\mathbf{1}_{A_{x_j}} \to\mathbf{1}_{A_{x_\infty}} in L1L^1 on a common bounded set. Consequently their Fourier integrals converge locally uniformly, contradicting the assumed discrepancy. This proves the stated compact-uniform convergence.

Classical bulk profiles as Fourier transforms of local spectral bodies

The main theorem gives the limiting normalized kernel in the form

Ux(z)=1vol⁡(Ax)∫Axeiz⋅ξ dξ,z=u−v.U_x(z) = \frac{1}{\operatorname{vol}(\mathcal{A}_x)} \int_{\mathcal{A}_x} e^{iz \cdot\xi}\,d\xi,\qquad z = u-v.

Thus the universal profile is the characteristic function (Fourier transform) of the uniform probability measure on the local spectral body. This section explains how the familiar sine and Bessel profiles arise from this formula.

Ellipsoids and normalized Bessel functions

For ν>−1\nu> -1, write

Jν∗(r)=r−νJν(r),Jν∗(0)=12νΓ(ν+1),J_\nu^*(r) = r^{-\nu}J_\nu(r),\qquad J_\nu^*(0) = \frac{1}{2^\nu\Gamma(\nu+1)},

where the formula for positive real rr is extended by the even entire power series

Jν∗(r)=2−ν∑k=0∞(−r2/4)kk! Γ(k+ν+1).J_\nu^*(r) = 2^{-\nu}\sum_{k=0}^{\infty}\frac{(-r^2/4)^k}{k!\,\Gamma(k+\nu+1)}.

This convention removes any branch ambiguity.

Proposition 7.1 (Fourier transform of an ellipsoid). Let MM be a real symmetric positive-definite d×dd \times d matrix and let

EM={ξ∈Rd:ξTMξ≤1}.\mathcal{E}_M = \{\xi\in\mathbb{R}^d : \xi^{\mathsf{T}}M\xi\le1\}.

Then

vol⁡(EM)=ωddet⁡M,ωd=πd/2Γ(d/2+1),(29)\operatorname{vol}(\mathcal{E}_M) = \frac{\omega_d}{\sqrt{\det M}},\qquad\omega_d = \frac{\pi^{d/2}}{\Gamma(d/2+1)}, \tag*{(29)}

and, for z∈Rdz \in\mathbb{R}^d,

1vol⁡(EM)∫EMeiz⋅ξ dξ=Jd/2∗(zTM−1z)Jd/2∗(0).(30)\frac{1}{\operatorname{vol}(\mathcal{E}_M)}\int_{\mathcal{E}_M}e^{iz\cdot\xi}\,d\xi= \frac{J_{d/2}^*\left(\sqrt{z^{\mathsf{T}}M^{-1}z}\right)}{J_{d/2}^*(0)}. \tag*{(30)}

The identity extends to complex zz by analytic continuation, with the right-hand side interpreted through the entire power series in zTM−1zz^{\mathsf{T}}M^{-1}z.

Proof. Use the change of variables ξ=M−1/2η\xi=M^{-1/2}\eta. The ellipsoid is mapped to the unit ball BdB_d, and the Jacobian (det⁡M)−1/2(\det M)^{-1/2} cancels after normalization. Rotational invariance reduces the remaining integral to

1ωd∫Bdei(M−1/2z)⋅η dη.\frac{1}{\omega_d}\int_{B_d}e^{i(M^{-1/2}z)\cdot\eta}\,d\eta.

The standard Fourier transform of the indicator of the unit ball is

∫Bdeiζ⋅η dη=(2π)d/2Jd/2(∥ζ∥)∥ζ∥d/2.\int_{B_d} e^{i\zeta\cdot\eta}\,d\eta=(2\pi)^{d/2}\frac{J_{d/2}(\|\zeta\|)}{\|\zeta\|^{d/2}}.

Dividing by ωd\omega_d and using Jd/2∗(0)=1/(2d/2Γ(d/2+1))J^*_{d/2}(0)=1/(2^{d/2}\Gamma(d/2+1)) gives (30).

The proposition shows that the normalized Bessel function is not tied to a particular orthogonal-polynomial basis. It is the Fourier transform of uniform volume measure on a Euclidean ball, transported by a linear map. The dimension fixes the Bessel order d/2d/2, while the ellipsoid’s shape enters only through the quadratic form zTM−1zz^{\mathsf T}M^{-1}z.

The same Fourier form has a classical spectral antecedent. Piazzon [19 Theorems 1 and 2] identifies the total-degree polynomial spaces on the ball and simplex, for their equilibrium weights, with spectral subspaces of the Baran Laplacians. Combining this identification with Hörmander’s local spectral-function theorem [12 Theorem 5.1] gives the interior Fourier integral over the principal-symbol unit co-ball, and hence the Bessel profile. Here the degree-nn spectral threshold is n2+O(n)n^2+O(n); the local Weyl remainder is of lower order on compact interior sets. This consequence of the two results is distinct from the direct recovery proof for general, possibly non-ellipsoidal, co-balls given above.

The interval and the sine kernel

For E=[−1,1]E=[-1,1], the Baran co-ball at x∈(−1,1)x\in(-1,1) is

Ax=[−11−x2,11−x2].\mathcal{A}_x=\left[-\frac{1}{\sqrt{1-x^2}},\frac{1}{\sqrt{1-x^2}}\right].

Consequently,

Ux(z)=1vol⁡(Ax)∫Axeizξ dξ=sin⁡(z/1−x2)z/1−x2.(31)\mathcal{U}_x(z)=\frac{1}{\operatorname{vol}(\mathcal{A}_x)}\int_{\mathcal{A}_x}e^{iz\xi}\,d\xi=\frac{\sin\left(z/\sqrt{1-x^2}\right)}{z/\sqrt{1-x^2}}. \tag*{(31)}

For d=1d=1, the Christoffel consequence of the main theorem gives

w(x)Knμ(x,x)∼nπ1−x2.w(x)K_n^\mu(x,x)\sim\frac{n}{\pi\sqrt{1-x^2}}.

Thus the microscopic displacement a/(w(x)Knμ(x,x))a/(w(x)K_n^\mu(x,x)) is asymptotic to πa1−x2/n\pi a\sqrt{1-x^2}/n. Substituting this scale into (31) yields

sin⁡π(a−b)π(a−b).\frac{\sin\pi(a-b)}{\pi(a-b)}.

the classical sine-kernel profile in Lubinsky’s one-dimensional bulk universality theorem [17]. In this sense, the sine kernel is the one-dimensional instance of the same local Fourier-geometric principle.

The ball: the Bessel profile of Kroó–Lubinsky

For the unit ball BdB^d, the Baran co-ball is the ellipsoid

AxB={ξ:ξTMB(x)ξ≤1},MB(x)=I−xxT.\mathcal{A}_x^B=\{\xi:\xi^{\mathsf T}M_B(x)\xi\le1\},\qquad M_B(x)=I-xx^{\mathsf T}.

The matrix determinant lemma and the Sherman–Morrison formula give

det⁡MB(x)=1−∥x∥2,MB(x)−1=I+xxT1−∥x∥2.\det M_B(x)=1-\|x\|^2,\qquad M_B(x)^{-1}=I+\frac{xx^{\mathsf T}}{1-\|x\|^2}.

Hence, for z=u−vz=u-v,

zTMB(x)−1z=∥u−v∥2+(x⋅(u−v))21−∥x∥2=:G(x,u,v).z^{\mathsf T}M_{B}(x)^{-1}z=\lVert u-v\rVert^{2}+\frac{(x\cdot(u-v))^{2}}{1-\lVert x\rVert^{2}}=:G(x,u,v).

Proposition 7.1 therefore yields

uxB(u−v)=Jd/2∗(G(x,u,v‾))Jd/2∗(0).(32)u_{x}^{B}(u-v)=\frac{J^{*}_{d/2}\left(\sqrt{G(x,u,\overline{v})}\right)}{J^{*}_{d/2}(0)}. \tag*{(32)}

This is exactly the normalized Bessel profile in Theorem 1.7 of Kroó and Lubinsky [15]. Their theorem establishes the limit for regular measures on the ball. The reference-kernel identity is based on Xu’s formula for Jacobi-weight Fourier series on the ball [24]; Kroó and Lubinsky combine it with the Mehler–Heine formula and then transfer the limit to general measures. Formula (32) supplies a geometric explanation for the resulting special function: the local spectral body is an ellipsoid, and its normalized Fourier transform is Bessel.

The simplex: a parallel ellipsoidal example

The standard simplex is not centrally symmetric, so it is not covered by the main theorem. It nevertheless provides an instructive parallel example. Write

Σd={xj>0, ∑j=1dxj<1},xd+1=1−∑j=1dxj.\Sigma^{d}=\left\{x_j>0,\ \sum_{j=1}^{d}x_j<1\right\},\qquad x_{d+1}=1-\sum_{j=1}^{d}x_j.

The Baran metric of the simplex is Riemannian; its inverse metric matrix, equivalently the quadratic form defining the local co-ball, is

MT(x)=diag⁡(x1,…,xd)−xxT.M_{T}(x)=\operatorname{diag}(x_1,\ldots,x_d)-xx^{\mathsf T}.

The Sherman–Morrison formula gives

MT(x)−1=diag⁡(x1−1,…,xd−1)+11Txd+1.M_{T}(x)^{-1}=\operatorname{diag}(x_1^{-1},\ldots,x_d^{-1})+\frac{\mathbf{1}\mathbf{1}^{\mathsf T}}{x_{d+1}}.

Thus, with zj=uj−vjz_j=u_j-v_j for 1≤j≤d1\leq j\leq d and zd+1=−∑j=1dzjz_{d+1}=-\sum_{j=1}^{d}z_j,

zTMT(x)−1z=∑j=1d+1zj2xj=:H(x,u,v).z^{\mathsf T}M_{T}(x)^{-1}z=\sum_{j=1}^{d+1}\frac{z_j^2}{x_j}=:H(x,u,v).

The corresponding ellipsoid Fourier transform is therefore

Jd/2∗(H(x,u,v‾))Jd/2∗(0),\frac{J^{*}_{d/2}\left(\sqrt{H(x,u,\overline{v})}\right)}{J^{*}_{d/2}(0)},

which is precisely the profile in Theorem 1.8 of Kroó and Lubinsky [15]. Their reference-kernel representation uses Xu’s Jacobi-weight Fourier-series formula for the simplex [23]. The Riemannian Baran geometry of the simplex and its relation to polynomial eigenfunctions are established by Piazzon [19]. We include the simplex as a comparison, not as an application of the centrally symmetric theorem.

Beyond ellipsoids

The Bessel function is therefore not the universal bulk kernel for multivariate polynomials. The universal object is the normalized Fourier transform of the local spectral body Ax\mathcal{A}_x. Bessel functions occur when Ax\mathcal{A}_x is an ellipsoid. For the cube, Ax\mathcal{A}_x is a weighted cross-polytope; for other centrally symmetric polytopes it can be a position-dependent polytope. Their limiting kernels are the corresponding polyhedral Fourier profiles, not in general Bessel functions.

This distinction is important: the Christoffel function sees only the volume of Ax\mathcal{A}_x, whereas the full near-diagonal kernel sees its shape. The phase-space description therefore predicts not just a local density, but the entire directional correlation profile of the polynomial space.

Non-ellipsoidal examples

Throughout this section,

sinc⁡(t)={sin⁡(t)/t,t≠0,1,t=0.\operatorname{sinc}(t) = \begin{cases} \sin(t)/t, & t \ne0, \\ 1, & t = 0. \end{cases}

For the local spectral body Ax\mathcal{A}_x, write

Φx(z)=1vol⁡(Ax)∫Axeiz⋅ξ dξ.\Phi_x(z) = \frac{1}{\operatorname{vol}(\mathcal{A}_x)} \int_{\mathcal{A}_x} e^{i z \cdot\xi}\,d\xi.

For a centrally symmetric polytope with E∗=conv⁡{±y1,…,±ym}E^*=\operatorname{conv}\{\pm y_1,\ldots,\pm y_m\}, the co-ball reduces to

Ax=conv⁡{±gx(y1),…,±gx(ym)}.(33)\mathcal{A}_x=\operatorname{conv}\{\pm g_x(y_1),\ldots,\pm g_x(y_m)\}. \tag*{(33)}

Indeed, fix vv and put M=max⁡j∣v⋅gx(yj)∣M=\max_j |v\cdot g_x(y_j)|. The convex quadratic function (v⋅y)2+M2(x⋅y)2(v\cdot y)^2+M^2(x\cdot y)^2 is at most M2M^2 at every vertex, and therefore throughout E∗E^*. Since ∣x⋅y∣<1|x\cdot y|<1 there, ∣v⋅gx(y)∣≤M|v\cdot g_x(y)|\le M. Equality of support functions proves (33).

The cube

For Qd=[−1,1]dQ_d=[-1,1]^d, the polar is the cross-polytope {y:∑j∣yj∣≤1}\{y:\sum_j |y_j|\le1\}. The Baran metric is

FQ(x;v)=max⁡1≤j≤d∣vj∣1−xj2.F_{Q}(x;v)=\max_{1\le j\le d}\frac{|v_j|}{\sqrt{1-x_j^2}}.

and its dual unit ball is

Ax={ξ:∑j=1d1−xj2∣ξj∣≤1}.\mathcal{A}_x=\left\{\xi:\sum_{j=1}^{d}\sqrt{1-x_j^2}|\xi_j|\le1\right\}.

This is a weighted cross-polytope, not an ellipsoid. Its volume is

vol⁡(Ax)=2dd!∏j=1d1−xj2.\operatorname{vol}(\mathcal{A}_x)=\frac{2^d}{d!\prod_{j=1}^{d}\sqrt{1-x_j^2}}.

The main theorem yields

Nnλnμ(x)⟶πdw(x)∏j=1d1−xj2.N_n\lambda_n^\mu(x)\longrightarrow\pi^d w(x)\prod_{j=1}^{d}\sqrt{1-x_j^2}.

For the product Chebyshev measure, the near-diagonal kernel can also be computed directly by an elementary Riemann-sum calculation from its tensor-product Chebyshev basis with a total-degree cutoff. We include this example as a consistency check, without a novelty claim for the cube. The resulting Fourier region is exactly the cross-polytope above. Its Fourier transform is generally not a Bessel function, illustrating that the local spectral geometry, rather than the name of the orthogonal-polynomial family, determines the universal profile.

Rectangles and boxes

Let

B=∏j=1d[αj,βj],x∈int⁡B,B = \prod_{j=1}^{d}[\alpha_j,\beta_j], \qquad x \in\operatorname{int} B,

and put

ℓj(x)=(xj−αj)(βj−xj).\ell_j(x) = \sqrt{(x_j-\alpha_j)(\beta_j-x_j)}.

The product property of the Siciak extremal function, together with the one-dimensional interval formula, gives

FB(x;v)=max⁡1≤j≤d∣vj∣ℓj(x).F_B(x;v) = \max_{1\le j\le d} \frac{|v_j|}{\ell_j(x)}.

Consequently the local spectral body is the weighted cross-polytope

Ax={ξ∈Rd:∑j=1dℓj(x)∣ξj∣≤1},vol⁡(Ax)=2dd!∏j=1dℓj(x).(34)A_x = \left\{\xi\in\mathbb{R}^d : \sum_{j=1}^{d}\ell_j(x)|\xi_j| \le1\right\}, \qquad\operatorname{vol}(A_x) = \frac{2^d}{d!\prod_{j=1}^{d}\ell_j(x)}. \tag*{(34)}

The normalized Fourier profile can be written in every dimension as

Φx(z)=d!∫Δd∏j=1dcos⁡(zjtjℓj(x)) dt,Δd={tj≥0:t1+⋯+td≤1}.(35)\Phi_x(z) = d! \int_{\Delta_d} \prod_{j=1}^{d}\cos\left(\frac{z_jt_j}{\ell_j(x)}\right)\,dt, \qquad\Delta_d = \{t_j \ge0 : t_1+\cdots+t_d \le1\}. \tag*{(35)}

Thus the main theorem yields

Nnλnμ(x)⟶πdw(x)∏j=1dℓj(x).(36)N_n\lambda_n^\mu(x) \longrightarrow\pi^d w(x)\prod_{j=1}^{d}\ell_j(x). \tag*{(36)}

In dimension two, write qj=zj/ℓj(x)q_j=z_j/\ell_j(x). The diamond {∣η1∣+∣η2∣≤1}\{|\eta_1|+|\eta_2|\le1\} becomes the square [−1,1]2[-1,1]^2 under s=η1+η2s=\eta_1+\eta_2, t=η1−η2t=\eta_1-\eta_2, with dη1dη2=12 ds dtd\eta_1d\eta_2=\frac{1}{2}\,ds\,dt. Hence the Fourier integral separates:

Φx(z)=sinc⁡(q1+q22)sinc⁡(q1−q22).(37)\Phi_x(z) = \operatorname{sinc}\left(\frac{q_1+q_2}{2}\right)\operatorname{sinc}\left(\frac{q_1-q_2}{2}\right). \tag*{(37)}

In particular, for a rectangle [α1,β1]×[α2,β2][\alpha_1,\beta_1]\times[\alpha_2,\beta_2], the normalized bulk kernel has the explicit limit

Knμ(x+u/n,x+v/n)Knμ(x,x)⟶sinc⁡[12(u1−v1ℓ1(x)+u2−v2ℓ2(x))]sinc⁡[12(u1−v1ℓ1(x)−u2−v2ℓ2(x))].\frac{K_n^\mu(x+u/n,x+v/n)}{K_n^\mu(x,x)} \longrightarrow \operatorname{sinc}\left[\frac{1}{2}\left(\frac{u_1-v_1}{\ell_1(x)}+\frac{u_2-v_2}{\ell_2(x)}\right)\right] \operatorname{sinc}\left[\frac{1}{2}\left(\frac{u_1-v_1}{\ell_1(x)}-\frac{u_2-v_2}{\ell_2(x)}\right)\right].

A rectangle in physical space therefore gives a diamond, not a rectangle, in local spectral space. This reflects the total-degree constraint. By contrast, a tensor-product degree cutoff would produce a box-shaped spectral region and a product of coordinate sinc functions. More generally, an affine image c+L[−1,1]dc + L[-1,1]^d has local spectral body

L−T{η:∑j=1d1−sj2 ∣ηj∣≤1},s=L−1(x−c),L^{-\mathsf{T}}\left\{\eta:\sum_{j=1}^{d}\sqrt{1-s_j^2}\,|\eta_j|\le1\right\},\qquad s=L^{-1}(x-c),

so the same calculation covers rotated rectangles and parallelepipeds.

The regular hexagon: a change of local spectral shape

Let

H=conv⁡{(cos⁡(kπ/3),sin⁡(kπ/3)):0≤k<6}.H=\operatorname{conv}\{(\cos(k\pi/3),\sin(k\pi/3)):0\le k<6\}.

Its polar has vertices ±y1,±y2,±y3\pm y_1,\pm y_2,\pm y_3, where

y1=(1,1/3),y2=(0,2/3),y3=(−1,1/3),y2=y1+y3.y_1=(1,1/\sqrt{3}),\qquad y_2=(0,2/\sqrt{3}),\qquad y_3=(-1,1/\sqrt{3}),\qquad y_2=y_1+y_3.

For an arbitrary interior base point x=(x1,x2)x=(x_1,x_2), set

qj=x⋅yj,hj=1−qj2,bj=yjhj,S=h1+h2+h3.q_j=x\cdot y_j,\qquad h_j=\sqrt{1-q_j^2},\qquad b_j=\frac{y_j}{h_j},\qquad S=h_1+h_2+h_3.

The finite-polar formula (33) gives

Ax=conv⁡{±b1,±b2,±b3}.(38)A_x=\operatorname{conv}\{\pm b_1,\pm b_2,\pm b_3\}. \tag*{(38)}

The vectors satisfy

y1=y2−y3,y2=y1+y3,y3=y2−y1.y_1=y_2-y_3,\qquad y_2=y_1+y_3,\qquad y_3=y_2-y_1.

It follows that the pair ±bj\pm b_j is redundant precisely when

hj≥hk+hℓ,{j,k,ℓ}={1,2,3}.h_j\ge h_k+h_\ell,\qquad\{j,k,\ell\}=\{1,2,3\}.

For example, if h2≥h1+h3h_2\ge h_1+h_3, then

b2=h1h2b1+h3h2b3b_2=\frac{h_1}{h_2}b_1+\frac{h_3}{h_2}b_3

lies in conv⁡{±b1,±b3}\operatorname{conv}\{\pm b_1,\pm b_3\}; the other cases follow cyclically. The location of the transition has a particularly simple description. The identity

S(−h1+h2+h3)(h1−h2+h3)(h1+h2−h3)=3−4∥x∥2(39)S(-h_1+h_2+h_3)(h_1-h_2+h_3)(h_1+h_2-h_3)=3-4\lVert x\rVert^2 \tag*{(39)}

follows by substituting q1=x1+x2/3q_1=x_1+x_2/\sqrt{3}, q2=2x2/3q_2=2x_2/\sqrt{3}, q3=−x1+x2/3q_3=-x_1+x_2/\sqrt{3} and using q2=q1+q3q_2=q_1+q_3 and q12+q22+q32=2∥x∥2q_1^2+q_2^2+q_3^2=2\lVert x\rVert^2. Since all hjh_j are positive in the interior, (39) shows that

{∥x∥<3/2:all three strict triangle inequalities hold,∥x∥=3/2:exactly one triangle equality holds,∥x∥>3/2:exactly one of the three triangle inequalities fails.\begin{cases} \lVert x\rVert<\sqrt{3}/2 &: \text{all three strict triangle inequalities hold},\\ \lVert x\rVert=\sqrt{3}/2 &: \text{exactly one triangle equality holds},\\ \lVert x\rVert>\sqrt{3}/2 &: \text{exactly one of the three triangle inequalities fails}. \end{cases}

The circle ∥x∥=3/2\lVert x\rVert=\sqrt{3}/2 is the incircle of HH; its six tangency points lie on ∂H\partial H and are excluded from the interior statement. Thus the central disk corresponds to a hexagonal local spectral body, whereas the six corner regions outside the incircle correspond to parallelograms.

The incircle region: hexagonal local spectral bodies. If ∥x∥<3/2\lVert x\rVert< \sqrt{3}/2, then all six points in (38) are vertices. The area is

vol⁡(Ax)=2S3h1h2h3.(40)\operatorname{vol}(\mathcal{A}_x)=\frac{2S}{\sqrt{3}h_1h_2h_3}. \tag*{(40)}

For an explicit Fourier profile, define

g1=b2−b12,g2=b3−b22,g3=−b1+b32.g_1=\frac{b_2-b_1}{2},\qquad g_2=\frac{b_3-b_2}{2},\qquad g_3=-\frac{b_1+b_3}{2}.

The identities

b1=−g1−g2−g3,b2=g1−g2−g3,b3=g1+g2−g3b_1=-g_1-g_2-g_3,\qquad b_2=g_1-g_2-g_3,\qquad b_3=g_1+g_2-g_3

show that Ax⊆Z\mathcal{A}_x\subseteq Z, where

Z=[−g1,g1]+[−g2,g2]+[−g3,g3].Z=[-g_1,g_1]+[-g_2,g_2]+[-g_3,g_3].

The determinant identities below imply

vol⁡(Z)=4∑i<jdet⁡(gi,gj)=2S3h1h2h3.\operatorname{vol}(Z)=4\sum_{i<j}\det(g_i,g_j)=\frac{2S}{\sqrt{3}h_1h_2h_3}.

This equals the shoelace area of Ax\mathcal{A}_x, so Ax=Z\mathcal{A}_x=Z. We have

det⁡(g1,g2)=h1+h3−h223h1h2h3,\det(g_1,g_2)=\frac{h_1+h_3-h_2}{2\sqrt{3}h_1h_2h_3},
det⁡(g1,g3)=h2+h3−h123h1h2h3,\det(g_1,g_3)=\frac{h_2+h_3-h_1}{2\sqrt{3}h_1h_2h_3},
det⁡(g2,g3)=h1+h2−h323h1h2h3.\det(g_2,g_3)=\frac{h_1+h_2-h_3}{2\sqrt{3}h_1h_2h_3}.

All three determinants are positive in this regime. Put

ω12=h1+h3−h2S,ω13=h2+h3−h1S,ω23=h1+h2−h3S.\omega_{12}=\frac{h_1+h_3-h_2}{S},\qquad\omega_{13}=\frac{h_2+h_3-h_1}{S},\qquad\omega_{23}=\frac{h_1+h_2-h_3}{S}.

They are positive and sum to one. For a planar three-generator zonotope Z=[−g1,g1]+[−g2,g2]+[−g3,g3]Z=[-g_1,g_1]+[-g_2,g_2]+[-g_3,g_3], the six parallelograms

Pij±=±gk+[−gi,gi]+[−gj,gj],{i,j,k}={1,2,3},P_{ij}^{\pm}=\pm g_k+[-g_i,g_i]+[-g_j,g_j],\qquad\{i,j,k\}=\{1,2,3\},

cover ZZ exactly twice, up to boundaries. Indeed, a generic fiber of the linear map

[−1,1]3∋(t1,t2,t3)⟼t1g1+t2g2+t3g3[-1,1]^3\ni(t_1,t_2,t_3)\longmapsto t_1g_1+t_2g_2+t_3g_3

is a line segment whose two endpoints lie on two faces of the cube. The images of those faces are precisely the six parallelograms above. Since

∫Pij±eiz⋅ξ dξ=4det⁡(gi,gj)e±iz⋅gksinc⁡(z⋅gi)sinc⁡(z⋅gj),\int_{P_{ij}^{\pm}}e^{iz\cdot\xi}\,d\xi=4\det(g_i,g_j)e^{\pm iz\cdot g_k}\operatorname{sinc}(z\cdot g_i)\operatorname{sinc}(z\cdot g_j),

the two-fold covering gives the normalized Fourier formula below. The weights ωij\omega_{ij} are exactly

ωij=det⁡(gi,gj)det⁡(g1,g2)+det⁡(g1,g3)+det⁡(g2,g3).\omega_{ij}=\frac{\det(g_i,g_j)}{\det(g_1,g_2)+\det(g_1,g_3)+\det(g_2,g_3)}.

The normalized Fourier transform of a planar zonotope generated by three segments therefore has the elementary formula

Φx(z)=ω12cos⁡(z⋅g3)sinc⁡(z⋅g1)sinc⁡(z⋅g2)+ω13cos⁡(z⋅g2)sinc⁡(z⋅g1)sinc⁡(z⋅g3)+ω23cos⁡(z⋅g1)sinc⁡(z⋅g2)sinc⁡(z⋅g3).(41)\begin{aligned} \Phi_x(z)={}&\omega_{12}\cos(z\cdot g_3)\operatorname{sinc}(z\cdot g_1)\operatorname{sinc}(z\cdot g_2)\\ &+\omega_{13}\cos(z\cdot g_2)\operatorname{sinc}(z\cdot g_1)\operatorname{sinc}(z\cdot g_3)\\ &+\omega_{23}\cos(z\cdot g_1)\operatorname{sinc}(z\cdot g_2)\operatorname{sinc}(z\cdot g_3). \tag*{(41)} \end{aligned}

The outer corner regions: parallelogram local spectral bodies. Suppose ∥x∥>3/2\lVert x\rVert> \sqrt{3}/2, and let jj be the unique index for which hj>hk+hℓh_j > h_k + h_\ell, with {j,k,ℓ}={1,2,3}\{j,k,\ell\} = \{1,2,3\}. Then ±bj\pm b_j lie inside the parallelogram generated by the other two pairs, and

Ax=conv⁡{±bk,±bℓ}.\mathcal{A}_x = \operatorname{conv}\{\pm b_k,\pm b_\ell\}.

Put

a=bk+bℓ2,c=bk−bℓ2.a = \frac{b_k+b_\ell}{2}, \qquad c = \frac{b_k-b_\ell}{2}.

Since

Ax=[−a,a]+[−c,c],\mathcal{A}_x = [-a,a] + [-c,c],

its area and normalized Fourier transform are

vol⁡(Ax)=43hkhℓ,Φx(z)=sinc⁡(z⋅a)sinc⁡(z⋅c).(42)\operatorname{vol}(\mathcal{A}_x) = \frac{4}{\sqrt{3}h_kh_\ell}, \qquad\Phi_x(z) = \operatorname{sinc}(z\cdot a)\operatorname{sinc}(z\cdot c). \tag*{(42)}

At ∥x∥=3/2\lVert x\rVert= \sqrt{3}/2 in the interior, one determinant above vanishes and one pair of generators lies on an edge; the body is already a parallelogram. The two area formulas agree, and the central Fourier formula reduces continuously to the parallelogram formula.

Under the hypotheses of theorem 2.1, the Christoffel limit is, in both regimes,

Nnλnμ(x)⟶2π2w(x)vol⁡(Ax),Nn=(n+22).(43)N_n\lambda_n^\mu(x) \longrightarrow\frac{2\pi^2w(x)}{\operatorname{vol}(\mathcal{A}_x)}, \qquad N_n = \binom{n+2}{2}. \tag*{(43)}

The normalized local kernel converges to Φx(u−v)\Phi_x(u-v), with Φx\Phi_x given by (41) or (42) according to the position of xx.

Along the symmetry axis x=(r,0)x=(r,0), one has h1=h3=1−r2h_1=h_3=\sqrt{1-r^2} and h2=1h_2=1. Writing t=(1−r2)−1/2t=(1-r^2)^{-1/2}, the hexagonal regime is 0≤r<3/20\le r<\sqrt{3}/2 and the parallelogram is the axis-aligned rectangle

[−t,t]×[−t/3,t/3][-t,t]\times[-t/\sqrt{3},t/\sqrt{3}]

when 3/2≤r<1\sqrt{3}/2\le r<1. In this case (42) reduces to

Φ(r,0)(z)=sinc⁡(tz1)sinc⁡(tz2/3),vol⁡(A(r,0))={2(t2+2t)/3,1≤t≤2,4t2/3,t≥2.\Phi_{(r,0)}(z)=\operatorname{sinc}(tz_1)\operatorname{sinc}(tz_2/\sqrt{3}), \qquad\operatorname{vol}(\mathcal{A}_{(r,0)}) = \begin{cases} 2(t^2+2t)/\sqrt{3}, & 1\le t\le2,\\ 4t^2/\sqrt{3}, & t\ge2. \end{cases}

Thus the earlier axial calculation is a special case of the full off-axis description.

Figure 1 illustrates the central and corner regimes at both axial and off-axis points.

Figure 1. The physical regular hexagon, its incircle, and local spectral bodies at representative interior points. The central disk gives hexagonal spectral bodies; outside it, the spectral body is a parallelogram. The axis-aligned rectangle is a special case of the parallelogram regime.

The example shows that the local spectral shape can change combinatorial type within the bulk, even though the domain and the total-degree filtration remain fixed. The Fourier profiles change accordingly, from the three-term zonotope formula to a product of two sinc functions. No special-function kernel for orthogonal polynomials on the hexagon is needed.

Scope and further questions

The theorem is a local limit for the entire total-degree filtration, not a construction of a second-order differential operator on an arbitrary convex body. The Baran co-ball and its equilibrium-volume formula are established geometric objects; the proof realizes that body as the full local L2L^2 spectrum, or local spectral body, at every interior point.

Central symmetry is used to match the extremal growth with the frequencies generated by affine Chebyshev ridges. For general convex bodies, the Baran derivative and equilibrium-volume formula are already available [5]. The unresolved step for the present method is a recovery construction attaining the entire corresponding co-ball with the correct global energy. The global ridge representation (9) need not hold without symmetry [3]; this failure alone does not disprove a general local-kernel conjecture. Boundary scaling is a separate problem and is not addressed here.

References

  1. [1]E. Bedford and B. A. Taylor, The complex equilibrium measure of a symmetric convex set in Rⁿ, Transactions of the American Mathematical Society 294 (1986), 705–717. doi:10.1090/S0002-9947-1986-0825731-8.DOI
  2. [2]R. J. Berman and J. Ortega-Cerdà, Sampling of real multivariate polynomials and pluripotential theory, American Journal of Mathematics 140 (2018), no. 3, 789–820. doi:10.1353/ajm.2018.0019; arXiv:1509.00956.DOI
  3. [3]L. Bos, J.-P. Calvi, and N. Levenberg, On the Siciak extremal function for real compact convex sets, Arkiv för Matematik 39 (2001), no. 2, 245–262. doi:10.1007/BF02384556.DOI
  4. [4]L. Bos, N. Levenberg, and S. Waldron, Pseudometrics, distances and multivariate polynomial inequalities, Journal of Approximation Theory 153 (2008), no. 1, 80–96. doi:10.1016/j.jat.2008.02.002.DOI
  5. [5]D. Burns, N. Levenberg, S. Ma’u, and Sz. Révész, Monge–Ampère measures for convex bodies and Bernstein–Markov type inequalities, Transactions of the American Mathematical Society 362 (2010), 6325–6340. arXiv:0705.1095.arxiv.org/abs/0705.1095
  6. [6]P. Deift, T. Kriecherbauer, K. T.-R. McLaughlin, S. Venakides, and X. Zhou, Uniform asymptotics for polynomials orthogonal with respect to varying exponential weights and applications to universality questions in random matrix theory, Communications on Pure and Applied Mathematics 52 (1999), no. 11, 1335–1425.doi.org/10.1002/(sici)1097-0312(199911)52:11<1335::aid-cpa1>3.0.co;2-1
  7. [7]C. F. Dunkl and Y. Xu, Orthogonal Polynomials of Several Variables, 2nd ed., Encyclopedia of Mathematics and its Applications, vol. 155, Cambridge University Press, 2014.
  8. [8]B. Eichinger, M. Lukić, and H. Woracek, Necessary and sufficient conditions for universality limits, Inventiones Mathematicae (2026), published online 17 November 2026. doi:10.1007/s00222-026-01460-2; arXiv:2409.18045.arxiv.org/abs/2409.18045
  9. [9]M. I. Ganzburg, Multidimensional limit theorems of the theory of best polynomial approximations, Sibirsk. Mat. Zh. 23 (1982), no. 3, 30–47 (Russian); English translation: Siberian Mathematical Journal 23, 316–331. doi:10.1007/BF00973488.DOI
  10. [10]M. I. Ganzburg, Sharp constants of approximation theory. III. Certain polynomial inequalities of different metrics on convex sets, Journal of Approximation Theory 252 (2020), article 105351. doi:10.1016/j.jat.2019.105351; arXiv:1902.10215.DOI
  11. [11]M. I. Ganzburg, Sharp constants of approximation theory. IV. Asymptotic relations in general settings, Analysis Mathematica 49 (2023), 79–136. doi:10.1007/s10476-022-0185-z; arXiv:2002.10512.DOI
  12. [12]L. Hörmander, The spectral function of an elliptic operator, Acta Mathematica 121 (1968), 193–218. doi:10.1007/BF02391913.DOI
  13. [13]L. Hörmander, The Analysis of Linear Partial Differential Operators I: Distribution Theory and Fourier Analysis, Springer-Verlag, Berlin, 1983.DOI
  14. [14]A. Kroó, Classical polynomial inequalities in several variables, in Constructive Theory of Functions, Varna 2002 (B. Bojanov, ed.), DARBA, Sofia, 2003, pp. 19–32.
  15. [15]A. Kroó and D. S. Lubinsky, Christoffel functions and universality in the bulk for multivariate orthogonal polynomials, Canadian Journal of Mathematics 65 (2013), no. 3, 600–620. doi:10.4153/CJM-2012-016-x.DOI
  16. [16]E. Levin and D. S. Lubinsky, L^p Christoffel functions, L^p universality, and Paley–Wiener spaces, Journal d’Analyse Mathématique 125 (2015), no. 1, 243–283. doi:10.1007/s11854-015-0008-2.DOI
  17. [17]D. S. Lubinsky, A new approach to universality limits involving orthogonal polynomials, Annals of Mathematics 170 (2009), no. 2, 915–939. doi:10.4007/annals.2009.170.915; arXiv:math/0701307.DOI
  18. [18]P. Nevai, Géza Freud, orthogonal polynomials and Christoffel functions. A case study, Journal of Approximation Theory 48 (1986), 3–167. doi:10.1016/0021-9045(86)90016-X.DOI
  19. [19]F. Piazzon, Laplace Beltrami operator in the Baran metric and pluripotential equilibrium measure: the ball, the simplex, and the sphere, Computational Methods and Function Theory 19 (2019), no. 4, 547–582. doi:10.1007/s40315-019-00286-9.DOI
  20. [20]G. Szegő, Orthogonal Polynomials, 4th ed., American Mathematical Society Colloquium Publications, vol. 23, Providence, RI, 1975.
  21. [21]V. Totik, Asymptotics for Christoffel functions for general measures on the real line, Journal d’Analyse Mathématique 81 (2000), 283–303. doi:10.1007/BF02788993.DOI
  22. [22]Y. Xu, Asymptotics for orthogonal polynomials and Christoffel functions on a ball, Methods and Applications of Analysis 3 (1996), 257–272.DOI
  23. [23]Y. Xu, Summability of Fourier orthogonal series for Jacobi weight functions on the simplex in R^d, Proceedings of the American Mathematical Society 126 (1998), no. 10, 3027–3036. doi:10.1090/S0002-9939-98-04415-3.DOI
  24. [24]Y. Xu, Summability of Fourier orthogonal series for Jacobi weight on a ball in R^d, Transactions of the American Mathematical Society 351 (1999), no. 6, 2439–2458. doi:10.1090/S0002-9947-99-02225-4.DOI
  25. [25]Y. Xu, Asymptotics of the Christoffel functions on a simplex in R^d, Journal of Approximation Theory 99 (1999), no. 1, 122–133. doi:10.1006/jath.1998.3312.DOI

Paper details

vlnc:W_7ggww47e
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents