betaFeedback

WorksMathematics

Solutions to Bianchi's covariogram problem and Kobayashi's Fourier zero-set problem for C+2C^2_+ convex bodies

Abstract

In each dimension n≥2n\ge2, the covariogram of a C+2C^2_+ convex body determines it among all convex bodies, up to translation and point reflection. For a general convex body, it suffices that the densities of the absolutely continuous part of its surface area measure at uu and −u-u be positive and unequal on a set of positive spherical measure. When the two densities are positive on a set of positive spherical measure, the complex zero set of the Fourier–Laplace transform, given without multiplicities, determines the body up to translation within that class. An averaged stationary phase formula locates simple zeros that account for the leading term in Jensen's bound, and a growth theorem controls the factors supported on the remaining zeros. In dimension three, width and the ordered antipodal curvature ratio also determine a body up to translation in the class with C1,1C^{1,1} support functions and uniformly positive curvature-radius matrices.

Introduction

For a convex body X⊂RnX \subset\mathbb{R}^{n}, n≥2n \ge2, define

gX(x)=vol⁡n(X∩(X+x)),FX(ζ)=∫Xeiζ⋅x dx,ζ∈Cn.g_X(x) = \operatorname{vol}_{n}(X \cap(X + x)), \qquad F_X(\zeta) = \int_X e^{i\zeta\cdot x}\,dx,\qquad\zeta\in\mathbb{C}^{n}.

A convex body is a compact convex set with nonempty interior. The dot product in the transform is complex bilinear. The convolution identity gX=1X∗1−Xg_X = \mathbf{1}_X * \mathbf{1}_{-X} gives, by Fubini on complex frequencies, the identity below whenever gK=gLg_K = g_L.

FK(ζ)FK(−ζ)=FL(ζ)FL(−ζ)(ζ∈Cn).(1)F_K(\zeta)F_K(-\zeta) = F_L(\zeta)F_L(-\zeta)\qquad(\zeta\in\mathbb{C}^{n}). \tag*{(1)}

On real frequencies the product is ∣FX∣2|F_X|^2. Both translations and point reflections preserve gXg_X.

Matheron’s planar conjecture [22] was proved for C+2C_{+}^{2} bodies by Bianchi, Segala and Volčič [8] and in full by Averkov and Bianchi [1]. Here C+2C_{+}^{2} means that the boundary is C2C^2 and its Gauss curvature, the product of its principal curvatures, is positive everywhere. Bianchi proved uniqueness for three-dimensional convex polytopes and constructed direct-product counterexamples in dimensions at least four [3, 5]. He asked whether uniqueness holds for three-dimensional C+2C_{+}^{2} bodies [4 Problem 5]. His Fourier–Laplace method gives a uniqueness theorem for pairs of positively curved bodies with a dimension-dependent smoothness assumption, which is C9C^{9} in dimension three [6 Theorem 1.1]. Averkov and Bianchi [2 Theorem 4(i)] proved generic uniqueness in the Baire category sense in each dimension, using covariogram data near the boundary of the difference body.

Let σ=Hn−1∣Sn−1\sigma= \mathcal{H}^{n-1}|_{\mathbb{S}^{n-1}} be spherical area measure. Write SXS_X for the pushforward of boundary area by the almost everywhere defined outer unit normal. In the Lebesgue decomposition SX=sXσ+(SX)sS_X = s_X \sigma+ (S_X)_{\mathrm{s}}, choose a finite nonnegative Borel representative of sXs_X. Set

TX={u∈Sn−1:sX(u)>0, sX(−u)>0},T_X = \{u \in\mathbb{S}^{n-1} : s_X(u) > 0,\ s_X(-u) > 0\},
UX={u∈TX:sX(u)≠sX(−u)}.U_X = \{u \in T_X : s_X(u) \ne s_X(-u)\}.

The sets are independent of this choice up to spherical null sets.

Theorem 1.1. Let K,L⊂RnK, L \subset\mathbb{R}^{n} be convex bodies with gK=gLg_K = g_L. If σ(UK)>0\sigma(U_K) > 0, then LL is a translate of KK or of −K-K.

Corollary 1.2. If SKS_K is absolutely continuous with strictly positive density almost everywhere, its covariogram determines KK among all convex bodies up to translation and point reflection. In particular this holds for every C+2C_{+}^{2} convex body.

Corollary 1.2 answers Bianchi’s three-dimensional C+2C_{+}^{2} question and allows an arbitrary convex competitor. The direct-product examples have singular surface area measure. For a product X=X1×X2X = X_1 \times X_2, the points with both coordinates in the respective factor boundaries form a set of boundary area zero. At almost every other boundary point the normal lies in a factor subspace, and the two subspheres of factor normals are spherical null sets. Under an invertible linear map BB, the normals change by the sphere diffeomorphism u↦B−Tu/∣B−Tu∣u \mapsto B^{-\mathsf{T}}u/\lvert B^{-\mathsf{T}}u\rvert. Thus singularity persists under the linear images used in these examples.

Kobayashi asked whether the complex zero set of FXF_X determines XX up to translation [18, 19]. Bianchi states the convex-body problem in [7 Problem 3.24]. Kobayashi proved injectivity in the smooth positively curved planar class [19 Corollaries 1.12 and 2.3.11].

Theorem 1.3. Let K,L⊂RnK, L \subset\mathbb{R}^{n} be convex bodies. Suppose TKT_K and TLT_L each have positive spherical measure. If

{ζ∈Cn:FK(ζ)=0}={ζ∈Cn:FL(ζ)=0}\{\zeta\in\mathbb{C}^{n} : F_K(\zeta) = 0\} = \{\zeta\in\mathbb{C}^{n} : F_L(\zeta) = 0\}

as sets, then LL is a translate of KK.

Theorem 1.3 applies in particular to any two C+2C^2_+ bodies. Kobayashi also asked about determination by width and the ordered antipodal Gauss-curvature ratio [19 Problem 1.13]. Bianchi restated this question in [7 Problem 3.27]. Write hXh_X for the support function and wX(u)=hX(u)+hX(−u)w_X(u)=h_X(u)+h_X(-u) for the width. The curvature-radius matrix is AX(u)=∇Sn−12hX(u)+hX(u)IA_X(u)=\nabla^2_{S^{n-1}}h_X(u)+h_X(u)I almost everywhere for a C1,1C^{1,1} support function. For a smooth positively curved boundary, let κX(u)\kappa_X(u) be the Gauss curvature at the point with outer normal uu. Then sX(u)/sX(−u)=κX(−u)/κX(u)s_X(u)/s_X(-u)=\kappa_X(-u)/\kappa_X(u).

Theorem 1.4. Let K,L⊂R3K,L\subset\mathbb{R}^3 be convex bodies whose support functions are C1,1C^{1,1} and whose curvature-radius matrices are bounded below almost everywhere by a positive multiple of the identity on the tangent plane. If

wK=wL,sK(u)sK(−u)=sL(u)sL(−u)for almost every u∈S2,w_K=w_L,\qquad\frac{s_K(u)}{s_K(-u)}=\frac{s_L(u)}{s_L(-u)}\quad\text{for almost every }u\in S^2,

then KK and LL are translates. In particular, the ordered data determine any C+2C^2_+ body in R3\mathbb{R}^3 within that class.

The large radial zeros for smooth bodies were studied by Kobayashi [19 Theorem 2.3.6] and Bianchi [6 Theorems 4.1 and 4.2]. In Sections 2 and 3, we identify the endpoint coefficients from the density sXs_X and prove an averaged formula in a complex strip. The estimate permits an angular exceptional set of arbitrarily small measure, which accommodates concentration at singular parts of the surface area measure. It gives simple radial zeros whose angular count exhausts the leading term in Jensen’s bound (§ 4). Thus any part of the divisor that misses these simple zeros has sublinear average radial count on the chosen directions. Section 5 uses a Remez inequality on the Taylor coefficients of the even product of a factor and its reflection. This product has minimal exponential type, and the factor contributes a linear functional to the growth along purely imaginary rays. Since that growth recovers the support function of each body, the divisor comparisons in Section 6 yield translation rigidity. Section 7 proves the width and ratio result by a separate elliptic argument in dimension three.

Support functions and section asymptotics

Write HX(x)=max⁡p∈Xp⋅xH_X(x)=\max_{p\in X}p\cdot x for the homogeneous support function and set

aX(u)=−hX(−u),bX(u)=hX(u),d=n−1,β=d/2.a_X(u)=-h_X(-u),\qquad b_X(u)=h_X(u),\qquad d=n-1,\qquad\beta=d/2.

Thus wX=bX−aXw_X=b_X-a_X. An inscribed ball gives a positive lower bound on wXw_X, and the diameter of XX gives an upper bound. Alexandrov’s theorem [15 Section 6.4] gives a second-order expansion of HXH_X at Lebesgue-almost every point of a Euclidean annulus. Since HXH_X is positively homogeneous of degree one, the existence of this expansion is invariant along each positive ray. It therefore holds at σ\sigma-almost every u∈Sn−1u\in S^{n-1}. At such a direction, pX(u)=DHX(u)p_X(u)=DH_X(u) is the unique supporting point, and

AX(u)=D2HX(u)∣u⊥=∇Sn−12hX(u)+hX(u)IA_X(u)=D^2H_X(u)|_{u^\perp}=\nabla^2_{S^{n-1}}h_X(u)+h_X(u)I

is positive semidefinite. The last equality follows by expanding HX(u+v)=∣u+v∣hX((u+v)/∣u+v∣)H_X(u+v)=|u+v|h_X((u+v)/|u+v|) for v∈u⊥v\in u^\perp. The surface area density satisfies

sX(u)=det⁡AX(u)(2)s_X(u)=\det A_X(u) \tag*{(2)}

for almost every uu.

For the normalization in (2), let F1(HX,⋅)F_1(H_X,\cdot) denote the Hessian measure with index 1 in Colesanti and Hug’s local Steiner formula. In that formula the coefficient of ρn−1\rho^{n-1} is nF1(HX,⋅)nF_1(H_X,\cdot). The homogeneity ∂HX(ru)=∂HX(u)\partial H_X(ru)=\partial H_X(u) for r>0r>0 implies F1(HX,bG)=bF1(HX,G)F_1(H_X,bG)=bF_1(H_X,G) for b>0b>0. Indeed, the parallel set {x+ρp:x∈bG, p∈∂HX(x)}\{x+\rho p:x\in bG,\ p\in\partial H_X(x)\} is bb times the corresponding parallel set of GG at scale ρ/b\rho/b. Comparing the coefficients of ρn−1\rho^{n-1} gives the scaling. In particular, F1(HX,⋅)F_1(H_X,\cdot) gives zero mass to the origin. For any Borel E⊂Sn−1E\subset S^{n-1}, [12 Corollary 5.10] now gives

nF1(HX,{ru:0<r≤1, u∈E})=SX(E).nF_1(H_X,\{ru:0<r\leq1,\ u\in E\})=S_X(E).

Scaling and subtracting give, for 0<a<b0<a<b,

F1(HX,{ru:a<r≤b, u∈E})=b−anSX(E).F_1(H_X,\{ru:a<r\leq b,\ u\in E\})=\frac{b-a}{n}S_X(E).

The annular sectors determine the measure off the origin. In polar coordinates the preceding identity gives n−1(dr)⊗SXn^{-1}(dr)\otimes S_X. The singular angular part is carried by a Lebesgue-null cone, and Euclidean volume is rn−1 dr dσ(u)r^{n-1}\,dr\,d\sigma(u). The Lebesgue density at ruru is thus r1−nsX(u)/nr^{1-n}s_X(u)/n. By [11 Theorem 2], the same density is σn−1(D2HX(ru))/n\sigma_{n-1}(D^2H_X(ru))/n, where σn−1\sigma_{n-1} is the elementary symmetric function of degree n−1n-1. The identity HX((1+t)u+v)=(1+t)HX(u+v/(1+t))H_X((1+t)u+v)=(1+t)H_X(u+v/(1+t)) for v⊥uv\perp u and ∣t∣<1|t|<1 shows that the radial row and column of the Alexandrov Hessian at uu are zero. Scaling the expansion to ruru gives the tangent block r−1AX(u)r^{-1}A_X(u). Equating the two expressions gives (2).

Lemma 2.1. At almost every uu, the section at depth ss below the upper supporting hyperplane has dd-dimensional volume

Hd(X∩{x:u⋅x=bX(u)−s})=ωd(2s)βsX(u)+o(sβ)(s↓0),(3)\mathcal{H}^{d}\bigl(X\cap\{x:u\cdot x=b_X(u)-s\}\bigr)=\omega_d(2s)^\beta\sqrt{s_X(u)}+o(s^\beta)\qquad(s\downarrow0), \tag*{(3)}

where ωd\omega_d is the volume of the unit ball in Rd\mathbb{R}^d. At a direction with an Alexandrov expansion of HXH_X, the matrix AX(u)A_X(u) is positive definite if and only if the boundary near the supporting point is a graph over u⊥u^\perp and its depth has a positive-definite quadratic expansion there. In that case the depth Hessian is AX(u)−1A_X(u)^{-1}.

Proof. Translate the supporting point to 00, use u⊥u^\perp as the horizontal plane, and let A=AX(u)A=A_X(u). In these coordinates

ϕ(v):=HX(u+v)=12vTAv+o(∣v∣2),v∈u⊥.\phi(v):=H_X(u+v)=\frac{1}{2}v^{\mathsf T}Av+o(|v|^2),\qquad v\in u^\perp.

If y−su∈Xy-su\in X, then y⋅v≤s+ϕ(v)y\cdot v\le s+\phi(v). The vectors yy in these sections tend uniformly to 00 as s↓0s\downarrow0, by compactness and uniqueness of the supporting point. For fixed δ>0\delta>0 insert v=(A+δI)−1yv=(A+\delta I)^{-1}y and use ϕ(v)≤12vT(A+δI)v\phi(v)\le\frac{1}{2}v^{\mathsf T}(A+\delta I)v near 00. The section is contained in

yT(A+δI)−1y≤2s.y^{\mathsf T}(A+\delta I)^{-1}y\le2s.

If AA is singular, the ellipsoid gives a bound for the limsup of the section volume divided by sβs^\beta. That bound tends to zero as δ↓0\delta\downarrow0.

Suppose A>0A>0. A second unit normal at 00 would have a nonzero tangential component, since opposite normals at the same point force XX into a hyperplane. Adding a small multiple of the second normal to uu and dividing by the uu-component gives a support normal u+vu+v for arbitrarily small nonzero v∈u⊥v\in u^\perp. This gives ϕ(v)=0\phi(v)=0 in contradiction to A>0A>0. Thus uu is the unique normal at 00. It follows that 00 belongs to the interior of the projection of XX onto u⊥u^\perp, because a supporting hyperplane of the projection at 00 would give a support normal at 00 tangent to u⊥u^\perp. The projection of the interior of XX equals the interior of the projection. Indeed, any point in the latter set lies on an open segment from the projection of a fixed interior point to another projected point, and the corresponding lifted segment contains interior points of XX. It follows that −s0u-s_0u is interior to XX for some s0>0s_0>0.

Let f(y)f(y) be the depth of the upper boundary, extended by +∞+\infty outside the projection. It is closed and convex, with f(0)=0f(0)=0, and ϕ=f∗\phi=f^* by the definition of support. The unique normal at 00 implies ∂f(0)={0}\partial f(0)=\{0\}. Subgradients of ff at yy near zero are uniformly bounded by local Lipschitz continuity and converge to zero as y→0y\to0, by closedness of the subgradient graph. For 0<δ<λmin⁡(A)0<\delta<\lambda_{\min}(A), the inequalities 12vT(A−δI)v≤ϕ(v)≤12vT(A+δI)v\frac{1}{2}v^{\mathsf T}(A-\delta I)v\le\phi(v)\le\frac{1}{2}v^{\mathsf T}(A+\delta I)v near zero yield

12yT(A+δI)−1y≤f(y)≤12yT(A−δI)−1y.\frac{1}{2}y^{\mathsf T}(A+\delta I)^{-1}y\le f(y)\le\frac{1}{2}y^{\mathsf T}(A-\delta I)^{-1}y.

For the lower bound, evaluate the conjugate at v=(A+δI)−1yv=(A+\delta I)^{-1}y. For the upper bound, take v∈∂f(y)v\in\partial f(y), so that f(y)=y⋅v−ϕ(v)f(y)=y\cdot v-\phi(v), and maximize y⋅v−12vT(A−δI)vy\cdot v-\frac{1}{2}v^{\mathsf T}(A-\delta I)v over vv. The two quadratic forms converge to the form with matrix A−1A^{-1} as δ↓0\delta\downarrow0. This proves the second-order expansion. The interior point −s0u-s_0u supplies a point of depth s0s_0 in each nearby vertical fibre. For s<s0s<s_0 in these fibres, y−su∈Xy-su\in X is equivalent to f(y)≤sf(y)\le s. Since the whole section lies in this neighborhood for small ss, the quadratic bounds sandwich it between ellipsoids of volumes ωd(2s)βdet⁡(A−δI)\omega_d(2s)^\beta\sqrt{\det(A-\delta I)} and ωd(2s)βdet⁡(A+δI)\omega_d(2s)^\beta\sqrt{\det(A+\delta I)}. Letting δ↓0\delta\downarrow0 gives (3).

Conversely, a positive-definite tangent depth expansion implies uniqueness of the supporting point, since a second supporting point would give a segment starting there with zero depth. Supporting points for u+vu+v approach that point as v→0v\to0, by compactness. Thus the maximizers in the conjugate formula for f∗(v)f^*(v) approach zero. If Q>0Q>0 is the depth Hessian, use the lower quadratic bound 12yT(Q−δI)y\frac{1}{2}y^{\mathsf T}(Q-\delta I)y at the maximizers for an upper bound on f∗(v)f^{*}(v). Evaluating at y=(Q+δI)−1vy=(Q+\delta I)^{-1}v and using the upper quadratic bound gives the reverse estimate. This yields the support Hessian Q−1Q^{-1}. □

Corollary 2.2. For almost every uu,

lim⁡t→∞tβ+1e−tbX(u)FX(−itu)=(2π)βsX(u).(4)\lim_{t\to\infty}t^{\beta+1}e^{-tb_X(u)}F_X(-itu)=(2\pi)^\beta\sqrt{s_X(u)}. \tag*{(4)}

If gK=gLg_K=g_L, then wK=wLw_K=w_L and

sK(u)sK(−u)=sL(u)sL(−u)for almost every u.(5)s_K(u)s_K(-u)=s_L(u)s_L(-u)\qquad\text{for almost every }u. \tag*{(5)}

Proof. Let V(s)V(s) be the section volume in (3). By Fubini,

e−tbX(u)FX(−itu)=∫0wX(u)e−tsV(s) ds.e^{-tb_X(u)}F_X(-itu)=\int_0^{w_X(u)}e^{-ts}V(s)\,ds.

On a fixed small interval at zero one has V(s)≤CsβV(s)\le Cs^\beta. There the change of variable v=tsv=ts, followed by dominated convergence, gives the limit. The integral away from zero is exponentially small. The constant is 2βωdΓ(1+β)=(2π)β2^\beta\omega_d\Gamma(1+\beta)=(2\pi)^\beta.

The support of gXg_X equals X−XX-X, since differences of interior points have gX>0g_X>0 and are dense in X−XX-X. Its support function is wXw_X, which gives the width identity. Evaluate (1) at ζ=−itu\zeta=-itu and multiply each side by t2β+2e−twX(u)t^{2\beta+2}e^{-tw_X(u)}, with X=KX=K or LL respectively. By (4) the limit on the XX-side is (2π)2βsX(u)sX(−u)(2\pi)^{2\beta}\sqrt{s_X(u)s_X(-u)} for almost every $u. Squaring the equality of the limits gives (5). For C+2C^2_+ bodies this recovers the product identity in [6 Proposition 3.1(I)]. □

Averaged stationary phase

For Re⁡z>0\operatorname{Re}z>0, use the holomorphic power of zz that is positive on the positive real axis. Define

DX(z,u)=izβ+1FX(zu),(6)D_X(z,u)=iz^{\beta+1}F_X(zu), \tag*{(6)}
MX(z,u)=(2π)β(e−iπβ/2sX(u)eizbX(u)−eiπβ/2sX(−u)eizaX(u)).(7)M_X(z,u)=(2\pi)^\beta\left(e^{-i\pi\beta/2}\sqrt{s_X(u)}e^{izb_X(u)}-e^{i\pi\beta/2}\sqrt{s_X(-u)}e^{iza_X(u)}\right). \tag*{(7)}

For a measurable V⊂Sn−1V\subset S^{n-1}, use the normalized norm

∥G∥R,V2=1R∫V∫R2R∣G(r,u)∣2 dr dσ(u).\|G\|_{R,V}^2=\frac{1}{R}\int_V\int_R^{2R}|G(r,u)|^2\,dr\,d\sigma(u).

Proposition 3.1. Let XX be any convex body. For every B,ε,η>0B,\varepsilon,\eta>0 there is a measurable V⊂Sn−1V\subset S^{n-1}, independent of RR and vv, with σ(Sn−1∖V)<ε\sigma(S^{n-1}\setminus V)<\varepsilon and

lim sup⁡R→∞sup⁡∣v∣≤B∥DX(r+iv,u)−MX(r+iv,u)∥R,V≤η.(8)\limsup_{R\to\infty}\sup_{|v|\le B}\|D_X(r+iv,u)-M_X(r+iv,u)\|_{R,V}\le\eta. \tag*{(8)}

Brandolini, Hofmann and Iosevich [9] proved an O(r−(n+1))O(r^{-(n+1)}) bound for the spherical mean square of the Fourier transform of a convex body. In (8), the angular exceptional set allows for concentration at facet normals. For example, for X=[−1,1]nX=[-1,1]^n one has MX=0M_X=0 almost everywhere and

∥DX(r,u)∥R,Sn−12=1R∫R≤∣ξ∣≤2R∣ξ∣2∏j=1n∣2sin⁡ξjξj∣2 dξ.\|D_X(r,u)\|_{R,S^{n-1}}^2=\frac{1}{R}\int_{R\leq|\xi|\leq2R}|\xi|^2\prod_{j=1}^{n}\left|\frac{2\sin\xi_j}{\xi_j}\right|^2\,d\xi.

On the region R+1≤ξ1≤2R−1R+1\leq\xi_1\leq2R-1, ∣ξj∣≤1/2|\xi_j|\leq1/2 for j≥2j\geq2, the integrand is bounded below by a positive multiple of sin⁡2ξ1\sin^2 \xi_1. This region lies in the annulus for large RR, so the full-sphere norm has positive liminf.

Definite Hessians

Use orthonormal coordinates xf(y)=(y,f(y))x_f(y)=(y,f(y)) on a graph over an open ball in Rd\mathbb{R}^d. For a∈L2a\in L^2 compactly supported in the ball, write

If,a(η,λ)=∫a(y)ei(η⋅y+λf(y)) dy.I_{f,a}(\eta,\lambda)=\int a(y)e^{i(\eta\cdot y+\lambda f(y))}\,dy.

At fixed real λ\lambda, Plancherel gives

∫Rd∣If,a(η,λ)∣2 dη=(2π)d∥a∥22.(9)\int_{\mathbb{R}^d}|I_{f,a}(\eta,\lambda)|^2\,d\eta=(2\pi)^d\|a\|_2^2. \tag*{(9)}

The polar coordinate Jacobian for (η,λ)=ru(\eta,\lambda)=ru is rdr^d. Using 2β=d2\beta=d and integrating (9) over ∣λ∣≤2R|\lambda|\leq2R gives

∫Sn−1∫R2R∣rβIf,a(ru)∣2 dr dσ(u)≤4(2π)dR∥a∥22.(10)\int_{S^{n-1}}\int_R^{2R}|r^\beta I_{f,a}(ru)|^2\,dr\,d\sigma(u)\leq4(2\pi)^dR\|a\|_2^2. \tag*{(10)}

A fixed translation of the graph multiplies the real-frequency integral by a phase of modulus one.

Use a Borel representative for aa and extend it by zero outside its support. For u=(u′,un)u=(u',u_n), suppose that yuy_u solves u′+unDf(yu)=0u'+u_nDf(y_u)=0 at a point with definite Hessian. Here un≠0u_n\ne0. The stationary term at this point is

(Tf,ra)(u)=(2π)βa(yu)eiru⋅xf(yu)+iπsig⁡(unD2f(yu))/4∣un∣β∣det⁡D2f(yu)∣.(11)(T_{f,r}a)(u)=\frac{(2\pi)^\beta a(y_u)e^{iru\cdot x_f(y_u)+i\pi\operatorname{sig}(u_nD^2f(y_u))/4}}{|u_n|^\beta\sqrt{|\det D^2f(y_u)|}}. \tag*{(11)}

Here sig⁡\operatorname{sig} is the number of positive eigenvalues minus the number of negative eigenvalues. For a convex or concave graph there is at most one such point in a given direction. Two points with the same gradient force affinity along the segment between them, contradicting a definite quadratic expansion at either point. We use Alexandrov Hessians for nonsmooth graphs and set the stationary term to zero when there is no definite point. The area formula below makes the term independent, almost everywhere in uu, of the representative chosen for aa.

Lemma 3.2. Suppose a∈L2a\in L^2 has compact support in a ball on which ff is C2C^2 and its Hessian is uniformly positive definite or uniformly negative

definite. Compute Tf,rT_{f,r} on that ball. Then

1R∫Sn−1∫R2R∣rβIf,a(ru)−(Tf,ra)(u)∣2 dr dσ(u)⟶0.\frac{1}{R}\int_{\mathbb{S}^{n-1}}\int_R^{2R}\left|r^\beta I_{f,a}(ru)-(T_{f,r}a)(u)\right|^2\,dr\,d\sigma(u)\longrightarrow0.

Proof. First take aa smooth. Choose an open ball WW containing supp⁡a\operatorname{supp} a whose closure lies in the ball of uniform definiteness. For a real, smooth, compactly supported even kernel φ\varphi of integral one, put

ρ(y)=43φ(y)−132−dφ(y/2).\rho(y)=\frac{4}{3}\varphi(y)-\frac{1}{3}2^{-d}\varphi(y/2).

This kernel has integral one and vanishing first and second moments. In particular convolution by ρh(y)=h−dρ(y/h)\rho_h(y)=h^{-d}\rho(y/h) fixes polynomials of degree at most two. Take h=R−1/2h=R^{-1/2} and fh=f∗ρhf_h=f*\rho_h on a fixed neighborhood of W‾\overline{W}. Uniformly there,

∥fh−f∥∞=o(h2),∥Dfh−Df∥∞=O(h),∥Dkfh∥∞=Ok(h2−k)(k≥2).(12)\begin{aligned} \|f_h-f\|_\infty&=o(h^2),\qquad\|Df_h-Df\|_\infty=O(h),\\ \|D^k f_h\|_\infty&=O_k(h^{2-k})\qquad(k\ge2). \tag*{(12)} \end{aligned}

The first estimate follows by integrating the quadratic Taylor expansion and using uniform continuity of the Hessian. The second uses the Lipschitz bound for DfDf, and the last follows by differentiating the convolution of D2fD^2f. For ∣λ∣≤2R|\lambda|\le2R one has ∣λ∣∥fh−f∥∞=o(1)|\lambda|\|f_h-f\|_\infty=o(1), so ∥a(eiλfh−eiλf)∥2=o(1)\|a(e^{i\lambda f_h}-e^{i\lambda f})\|_2=o(1) uniformly. Plancherel on each λ\lambda-slice, followed by integration over the annulus as in (10), shows that replacing ff by fhf_h costs o(1)o(1) in the norm ∥⋅∥R,Sn−1\|\cdot\|_{R,\mathbb{S}^{n-1}}.

The gradient map of ff is injective on WW with open image, and ∣Df(y)−Df(y′)∣≥c∣y−y′∣|Df(y)-Df(y')|\ge c|y-y'| there. This follows from the definite Hessian bounds by integrating along the line segment. The inverse function theorem gives openness. When a stationary point yuy_u lies in WW, one has ∣un∣=(1+∣Df(yu)∣2)−1/2≥c′>0|u_n|=(1+|Df(y_u)|^2)^{-1/2}\ge c'>0. Use y=yu+hqy=y_u+hq and write ϕ(q)=r(u′⋅(yu+hq)+unfh(yu+hq))\phi(q)=r(u'\cdot(y_u+hq)+u_nf_h(y_u+hq)). The rescaled gradient is

∇qϕ(q)=rh un(Dfh(yu+hq)−Df(yu)).\nabla_q\phi(q)=rh\,u_n\bigl(Df_h(y_u+hq)-Df(y_u)\bigr).

By (12) and the gradient lower bound,

∣∇qϕ(q)∣≥c1∣q∣−c2,∣∂qαϕ(q)∣≤Cα(∣α∣≥2),|\nabla_q\phi(q)|\ge c_1|q|-c_2,\qquad|\partial_q^\alpha\phi(q)|\le C_\alpha\quad(|\alpha|\ge2),

for R≤r≤2RR\le r\le2R and yu+hqy_u+hq in the amplitude support. The same bounds hold on a neighborhood of that support within WW. The prefactor rβhd=(r/R)βr^\beta h^d=(r/R)^\beta is bounded. To estimate the tail, multiply the rescaled amplitude by a smooth cutoff that vanishes for ∣q∣≤A|q|\le A and equals one for ∣q∣≥2A|q|\ge2A. For large AA, the vector field

Vϕ=∇qϕi∣∇qϕ∣2V_\phi=\frac{\nabla_q\phi}{i|\nabla_q\phi|^2}

and all its derivatives of fixed order are O(∣q∣−1)O(|q|^{-1}) on this region where the amplitude is supported. This follows by the chain rule, since the order-kk derivatives of v↦v/∣v∣2v\mapsto v/|v|^2 are homogeneous of degree −1−k-1-k and the phase derivatives of order at least two are bounded. The operator Vϕ⋅∇qV_\phi\cdot\nabla_q fixes eiϕe^{i\phi}. Applying its formal transpose NN times to the amplitude with cutoff yields terms each containing NN vector-field factors or their derivatives. The remaining derivatives of the amplitude and cutoff are bounded independently of h≤1h \le1 and A≥1A \ge1. For N>dN>d the tail is bounded by the integral of CN∣q∣−NC_N|q|^{-N} over ∣q∣≥A|q|\ge A, or ON(Ad−N)O_N(A^{d-N}).

On bounded qq-sets, the phase relative to ru⋅xf(yu)ru\cdot x_f(y_u) is

12(r/R)unTD2f(yu)q+o(1).\frac{1}{2}(r/R)u_n^{\mathsf T}D^2f(y_u)q+o(1).

This follows from (12), stationarity, and the quadratic Taylor expansion of ff, uniformly in uu with yu∈Wy_u\in W and R≤r≤2RR\le r\le2R. The amplitude converges uniformly there to a(yu)a(y_u) on each bounded qq-set. The tail bound just proved also holds for the quadratic phase with a factor e−ϵ∣q∣2e^{-\epsilon|q|^2}, 0<ϵ≤10<\epsilon\le1, since its derivatives of fixed order are uniformly bounded. Therefore the quadratic oscillatory integral defined by smooth cutoffs has the same value as the limit of the Gaussian-regularized integrals. For each eigenvalue μ≠0\mu\ne0 of unD2f(yu)u_nD^2f(y_u), the one-dimensional Gaussian factor in the limit is

(2π(r/R)∣μ∣)1/2eiπsgn⁡(μ)/4.\left(\frac{2\pi}{(r/R)|\mu|}\right)^{1/2}e^{i\pi\operatorname{sgn}(\mu)/4}.

Multiplying these factors and cancelling (r/R)d/2(r/R)^{d/2} gives (11). Taking R→∞R\to\infty with AA fixed and then A→∞A\to\infty is justified by the uniform tail bounds. The estimate includes stationary points outside the amplitude support, where a(yu)=0a(y_u)=0.

For directions with no stationary point in WW, the size of u′+unDf(y)u'+u_nDf(y) has a uniformly positive lower bound on the support. Otherwise a convergent subsequence of directions and support points would give a stationary point in WW, and openness of the gradient image would give stationary points in WW for the nearby directions. In coordinates y=hqy=hq, the phase gradient for fhf_h is now bounded below by ch−1ch^{-1}. The vector field and its derivatives are O(h)O(h) on the scaled support. The same integrations by parts give O(hN−d)O(h^{N-d}) for the normalized integral, since the scaled support has volume O(h−d)O(h^{-d}). Thus the normalized smoothed integral converges to the model with an error uniform in uu and R≤r≤2RR\le r\le2R. The Plancherel comparison proves the assertion for smooth aa.

For the L2L^2 extension, the two maps

y⟼±(−Df(y),1)1+∣Df(y)∣2y\longmapsto\pm\frac{(-Df(y),1)}{\sqrt{1+|Df(y)|^2}}

have spherical area Jacobian ∣det⁡D2f(y)∣(1+∣Df(y)∣2)−(d+1)/2|\det D^2f(y)|(1+|Df(y)|^2)^{-(d+1)/2}. The exponent follows by differentiating the map p↦(−p,1)/1+∣p∣2p\mapsto(-p,1)/\sqrt{1+|p|^2}. Its squared length factors are (1+∣p∣2)−1(1+|p|^2)^{-1} in the d−1d-1 directions perpendicular to pp and (1+∣p∣2)−2(1+|p|^2)^{-2} in its direction. The area formula [15 Section 3.3] applies to each normal map. Multiplication by the Jacobian cancels the factor ∣det⁡D2f∣−1|\det D^2f|^{-1} in the squared modulus of the stationary term and changes ∣un∣−2β|u_n|^{-2\beta} to (1+∣Df∣2)−1/2(1+|Df|^2)^{-1/2}. Summing over the two signs gives

∥Tf,ra∥L2(σ)2=2(2π)d∫∣a(y)∣2(1+∣Df(y)∣2)−1/2 dy.(13)\|T_{f,r}a\|_{L^2(\sigma)}^2=2(2\pi)^d\int|a(y)|^2(1+|Df(y)|^2)^{-1/2}\,dy. \tag*{(13)}

The resulting bound depends only on dd and ∥a∥2\lVert a\rVert_2. Approximation by smooth amplitudes in L2L^2, together with (10), completes the proof. □

Localization for a Lipschitz graph

Lemma 3.3. If ff is Lipschitz on the graph domain and a∈L2a \in L^2 is compactly supported there, then for ψ∈Cc∞(Rd)\psi\in C^\infty_c(\mathbb{R}^d)

lim⁡∣λ∣→∞∫Rdψ(η/λ)∣If,a(η,λ)∣2 dη=(2π)d∫∣a(y)∣2ψ(−Df(y)) dy.(14)\lim_{|\lambda|\to\infty}\int_{\mathbb{R}^d}\psi(\eta/\lambda)|I_{f,a}(\eta,\lambda)|^2\,d\eta=(2\pi)^d\int|a(y)|^2\psi(-Df(y))\,dy. \tag*{(14)}

Proof. Extend ff as a Lipschitz function outside the compact region in use and aa by zero. Expand the square and, for λ≠0\lambda\ne0, set η=λt\eta=\lambda t and p=λ(y−y′)p=\lambda(y-y'). The Jacobians ∣λ∣d|\lambda|^d and ∣λ∣−d|\lambda|^{-d} cancel. With Ψ(p)=∫ψ(t)eit⋅p dt\Psi(p)=\int\psi(t)e^{it\cdot p}\,dt, the result is the integral against Ψ(p)\Psi(p) of

∫a(y)a(y−p/λ)eiλ(f(y)−f(y−p/λ)) dy.\int a(y)a(y-p/\lambda)e^{i\lambda(f(y)-f(y-p/\lambda))}\,dy.

For fixed pp, the phase converges to Df(y)⋅pDf(y)\cdot p for almost every yy. Translation continuity in L2L^2 replaces a(y−p/λ)a(y-p/\lambda) by a(y)a(y) at a cost tending to zero. Dominated convergence first with respect to ∣a(y)∣2 dy|a(y)|^2\,dy and then against the integrable function ∣Ψ∣|\Psi| is justified by the bound ∥a∥22\lVert a\rVert_2^2 for the inner integral. Fourier inversion gives (14), with the displayed sign. □

Convex graph approximation

Lemma 3.4. Let ff be a Lipschitz convex or concave function on the graph ball and a∈L2a\in L^2 compactly supported there. Let Ωdef\Omega_{\mathrm{def}} be the set of points with definite Alexandrov Hessians on this ball. Then

∥Tf,ra∥L2(σ)2=2(2π)d∫Ωdef∣a(y)∣21+∣Df(y)∣2 dy.(15)\lVert T_{f,r}a\rVert_{L^2(\sigma)}^2=2(2\pi)^d\int_{\Omega_{\mathrm{def}}}\frac{|a(y)|^2}{\sqrt{1+|Df(y)|^2}}\,dy. \tag*{(15)}

For every ε,γ>0\varepsilon,\gamma>0 there is a measurable V⊂Sn−1V\subset\mathbb{S}^{n-1}, with σ(Sn−1∖V)<ε\sigma(\mathbb{S}^{n-1}\setminus V)<\varepsilon, such that

lim sup⁡R→∞∥rβIf,a(ru)−(Tf,ra)(u)∥R,V≤γ.\limsup_{R\to\infty}\lVert r^\beta I_{f,a}(ru)-(T_{f,r}a)(u)\rVert_{R,V}\le\gamma.

Proof. On a compact subregion of a Lipschitz convex graph, the points with Alexandrov expansions are a countable union of sets on each of which the gradient is Lipschitz. To see this, fix C,δ>0C,\delta>0, with δ\delta smaller than the distance of the subregion to the graph domain boundary, and consider points xx with

0≤f(x+h)−f(x)−Df(x)⋅h≤C∣h∣2(∣h∣<δ).0\le f(x+h)-f(x)-Df(x)\cdot h\le C|h|^2\qquad(|h|<\delta).

Every point of quadratic expansion belongs to one of these sets for suitable C,δC,\delta drawn from countably many choices. If r=∣x−y∣<δ/3r=|x-y|<\delta/3, the supporting affine functions at xx and yy give, for every unit ee,

(Df(y)−Df(x))⋅e≤f(y+re)−f(x)−Df(x)⋅(y+re−x)r≤4Cr.(Df(y)-Df(x))\cdot e\le\frac{f(y+re)-f(x)-Df(x)\cdot(y+re-x)}{r}\le4Cr.

The Lipschitz bound on ff handles pairs separated by at least δ/3\delta/3. Apply the argument to −f-f for a concave graph. Consequently the normal maps take parameter null sets of quadratically differentiable points to spherical null sets.

By [14 Proposition A.1], for every ϵ>0\epsilon> 0 there is a C2C^2 function agreeing with ff, DfDf, and the Alexandrov Hessian of ff outside a set of Lebesgue measure less than ϵ\epsilon. On the compact subregion containing supp⁡a\operatorname{supp} a, take approximations with ϵ→0\epsilon\to0. Inner regularity gives countably many compact pieces of agreement covering the part of the definite set in use up to a null set. On each piece the normal maps agree with the C1C^1 maps of the approximant, whose Jacobian is the one used in (13). Make the pieces disjoint by assigning each point to its first piece. Each sign of the normal map is injective on the definite set, including across the pieces, and has image in a distinct hemisphere. The omitted points have spherical null image by the gradient estimate above. The area formula on the pieces therefore gives (15).

Choose one C2C^2 approximation with sufficiently small exceptional measure. Within its agreement set choose disjoint compact sets PP and NN of points with Alexandrov expansions, with definite Hessian on PP and singular Hessian on NN. The exceptional measure and these sets can be chosen so that ∥a1(P∪N)c∥2\lVert a1_{(P\cup N)^c}\rVert_2 is arbitrarily small, by absolute continuity and inner regularity of ∣a∣2 dy|a|^2\,dy. Cover PP by finitely many balls of uniform definiteness of the agreeing function. Partition PP into measurable pieces each contained in a compact subball of one of those balls. For the amplitude on each piece, equality of the values of the graph functions gives equality of the integrals. Equality of their first and second derivatives gives equality of the models, since the amplitudes vanish off the piece. Apply Lemma 3.2 to each piece in its ball of uniform definiteness and add the finitely many estimates. This gives a vanishing normalized error for a1Pa1_P. By (10) and (15), the normalized error off P∪NP\cup N is at most Cd∥a1(P∪N)c∥2C_d\lVert a1_{(P\cup N)^c}\rVert_2 for every RR and every angular set.

The two normal images of NN are compact spherical null sets. Indeed, the normal maps of the agreeing C2C^2 function are continuous and have Jacobian zero on NN. The area formula gives spherical measure zero for their images. Remove an open neighborhood of these images of spherical measure less than ϵ\epsilon, and let VV be its closed complement. The stationary model for a1Na1_N vanishes. For the graph integral with a1Na1_N, the part with ∣λ∣<δR|\lambda| < \delta R contributes at most 2(2π)dδ∥a1N∥222(2\pi)^d\delta\lVert a1_N\rVert_2^2 to the squared norm, by (9). For ∣λ∣≥δR|\lambda| \ge\delta R on the annulus and u∈Vu \in V, write ∣λ∣≥δ/2|\lambda| \ge\delta/2. The slopes λu′/un\lambda u'/u_n range over the compact image of V∩{∣un∣≥δ/2}V\cap\{|u_n|\ge\delta/2\} under the slope map. This image is disjoint from −Df(N)-Df(N), since each slope in −Df(N)-Df(N) comes from the two removed normal directions above it. Choose a nonnegative smooth compactly supported function equal to one on the slope set and zero near −Df(N)-Df(N). Lemma 3.3, with the amplitude a1Na1_N, bounds the weighted η\eta-integral by o(1)o(1) uniformly for δR≤∣λ∣≤2R\delta R\le|\lambda|\le2R. The total length of the λ\lambda-range is at most 4R4R. Letting R→∞R\to\infty and then δ↓0\delta\downarrow0 shows that the contribution from NN tends to zero in ∥⋅∥R,V\|\cdot\|_{R,V}. The estimates on the three amplitudes prove the lemma. □\square

Boundary integral in the strip

Proof of Proposition 3.1. The divergence theorem on the Lipschitz boundary gives

izFX(zu)=∫∂Xeizu⋅xu⋅ν(x) dHd(x).(16)izF_X(zu)=\int_{\partial X}e^{izu\cdot x}u\cdot\nu(x)\,d\mathcal{H}^{d}(x). \tag*{(16)}

Here ν\nu is the outer unit normal, defined for Hd\mathcal{H}^{d}-almost every boundary point. To obtain a graph chart at p∈∂Xp\in\partial X, take the vertical direction from an interior point p0p_0 toward pp. Their common projection lies in the interior of the projected body. The vertical fibre contains the interior point p0p_0 and has pp as its upper endpoint. On a smaller ball of the projection the two boundary graphs are separated, and the upper graph is Lipschitz and concave. Reversing the vertical axis gives a convex graph. Compactness gives finitely many charts with a partition of unity compactly supported in them. On each graph the area element times the normal is the vector ±(−Df,1) dy\pm(-Df,1)\,dy in chart coordinates. Thus (u⋅ν) dHd(u\cdot\nu)\,d\mathcal{H}^{d} times the partition function is a sum of terms ujaj(y) dyu_j a_j(y)\,dy in fixed ambient coordinates, with bounded compactly supported amplitudes aja_j.

For z=r+ivz=r+iv, the additional weight e−vu⋅xf(y)e^{-vu\cdot x_f(y)} in (16) has the absolutely convergent expansion

∑α∈Nn(−v)∣α∣uαα!xf(y)α.\sum_{\alpha\in\mathbb{N}^{n}}\frac{(-v)^{|\alpha|}u^\alpha}{\alpha!}x_f(y)^\alpha.

In this expansion xf(y)x_f(y) is the boundary point in ambient coordinates, including the rotation and translation of the chart. If its coordinates on the amplitude support are bounded by MM, then ∥ajxfα∥2≤M∣α∣∥aj∥2\|a_jx_f^\alpha\|_2\le M^{|\alpha|}\|a_j\|_2. In the graph estimates, rotate the real frequency variables with the chart and use rotational invariance of σ\sigma. By (10) and (15), the sum of the normalized norms of the terms, for the integrals and their models, is bounded uniformly in RR and ∣v∣≤B|v|\le B by a constant times ∑α(BM)∣α∣/α!=enBM\sum_{\alpha}(BM)^{|\alpha|}/\alpha!=e^{nBM}. Choose a finite truncation with tail error less than η/2\eta/2 in normalized norm, on the full sphere. Apply Lemma 3.4 to each retained amplitude ajxfαa_jx_f^\alpha. Choose the omitted angular measures to sum to less than ε\varepsilon and the individual errors, multiplied by the uniform bounds for their coefficients, to sum to less than η/2\eta/2. The intersection of the resulting sets works uniformly for ∣v∣≤B|v|\le B. By (10) the norms of the weighted integrals are bounded on the full sphere uniformly in RR and vv. Hence replacing rβr^\beta there by zβz^\beta changes the result by o(1)o(1) in normalized norm, since zβ/rβ=1+OB(R−1)z^\beta/r^\beta=1+O_B(R^{-1}) on the band.

We now compute the two coefficients of the model. At a differentiable point in a graph chart there is a unique outer normal. Indeed, the base point is interior to the projection, the vertical fibre contains interior points, and the slope of the graph is unique. Projection to the tangent plane also gives a graph near this point, by the interior projection argument in the proof of Lemma 2.1. Let HH be the graph Hessian at a stationary point yuy_u. In orthonormal tangent coordinates the graph’s tangent map LL satisfies LTL=I+Df(Df)TL^{\mathsf{T}}L=I+Df(Df)^{\mathsf{T}}, where DfDf is written as a column. Its projection to the tangent plane has the expansion t=L(y−yu)+O(∣y−yu∣2)t=L(y-y_u)+O(|y-y_u|^2), so y−yu=L−1t+O(∣t∣2)y-y_u=L^{-1}t+O(|t|^2). The quadratic part of the height in direction uu is 12(u⋅en)(y−yu)TH(y−yu)\frac{1}{2}(u\cdot e_n)(y-y_u)^{\mathsf{T}}H(y-y_u), where ene_n is the chart’s vertical vector. Thus the Hessian in tangent coordinates is (u⋅en)L−THL−1(u\cdot e_n)L^{-\mathsf{T}}HL^{-1}. The same invertible change of coordinates, applied to the tangent graph, gives the converse when the second-order expansion is given in tangent coordinates. In either case the absolute determinant of the tangent height Hessian is

∣det⁡H∣(1+∣Df∣2)(d+2)/2.\frac{|\det H|}{(1+|Df|^2)^{(d+2)/2}}.

since ∣u⋅en∣=(1+∣Df∣2)−1/2|u\cdot e_n|=(1+|Df|^2)^{-1/2}. The factor from (11) and the surface element is

1+∣Df∣2∣u⋅en∣β∣det⁡H∣=(1+∣Df∣2)(d+2)/4∣det⁡H∣,\frac{\sqrt{1+|Df|^2}}{|u\cdot e_n|^\beta\sqrt{|\det H|}}=\frac{(1+|Df|^2)^{(d+2)/4}}{\sqrt{|\det H|}},

the reciprocal square root of the displayed determinant. For almost every uu the supporting points in directions uu and −u-u are both unique. By Lemma 2.1 and (2), their reciprocal square roots of curvature are sX(u)\sqrt{s_X(u)} and sX(−u)\sqrt{s_X(-u)} at definite jets. If AX(u)A_X(u) or AX(−u)A_X(-u) is singular, there is no definite jet at that endpoint. The phase Hessian is negative definite at the upper supporting point and positive definite at the lower one. The values of u⋅νu\cdot\nu are +1+1 and −1-1, respectively. The phase factors in (11) are e−iπd/4e^{-i\pi d/4} at the upper point and eiπd/4e^{i\pi d/4} at the lower point. Reinstating the complex weight and summing the partition functions therefore gives precisely (7). The estimates on the graph integrals give (8). □

Simple zeros and radial counting

For u∈TXu\in T_X, the zeros of the model (7) are simple. Solving eiwX(u)z=eiπβsX(−u)/sX(u)e^{iw_X(u)z}=e^{i\pi\beta}\sqrt{s_X(-u)/s_X(u)} gives

qjX(u)=2πj+πβwX(u)+iYX(u),YX(u)=12wX(u)log⁡sX(u)sX(−u),j∈Z.(17)q_j^X(u)=\frac{2\pi j+\pi\beta}{w_X(u)}+iY_X(u),\qquad Y_X(u)=\frac{1}{2w_X(u)}\log\frac{s_X(u)}{s_X(-u)},\qquad j\in\mathbb{Z}. \tag*{(17)}

The reflected body −X-X has the same width and the height −YX(u)-Y_X(u). On a direction set E⊂TXE\subset T_X where both endpoint densities are bounded above and bounded away from zero, the heights are bounded. The real spacing of the model zeros is bounded above and below by positive constants. On any bounded horizontal strip the model has a uniform positive lower bound in modulus at any prescribed positive distance from its zeros. In fact, for every jj,

MX(z,u)=(2π)βeiπβ/2sX(−u) eizaX(u)(eiwX(u)(z−qjX(u))−1).M_X(z,u)=(2\pi)^\beta e^{i\pi\beta/2}\sqrt{s_X(-u)}\,e^{iz a_X(u)}\left(e^{iw_X(u)(z-q_j^X(u))}-1\right).

The modulus of the prefactor is bounded away from zero in the strip. Choose jj with the closest real part, so that Re⁡(wXX(u)(z−qjX(u)))∈[−π,π]\operatorname{Re}(wX^{X}(u)(z-q_{j}^{X}(u)))\in[-\pi,\pi]. The values of wXX(u)(z−qjX(u))wX^{X}(u)(z-q_{j}^{X}(u)) under consideration lie in a compact rectangle and stay a positive distance from the zeros of eiz−1e^{iz}-1.

Lemma 4.1. Let E⊂TXE\subset T_{X} be measurable and satisfy the preceding endpoint density bounds. Fix a finite family F\mathcal{F} of convex bodies containing XX, a radius ρ>0\rho>0 less than one quarter of the infimum of the model spacing on EE, and a threshold τ>0\tau>0. For u∈Eu\in E, consider closed disks Bj(u)B_{j}(u) of radius ρ\rho centered at the qjX(u)q_{j}^{X}(u) with R≤Re⁡qjX(u)≤2RR\leq\operatorname{Re}q_{j}^{X}(u)\leq2R. Call a disk bad if

sup⁡z∈Bj(u)∣DW(z,u)−MW(z,u)∣>τfor at least one W∈F.\sup_{z\in B_{j}(u)}\lvert D_{W}(z,u)-M_{W}(z,u)\rvert>\tau\quad\text{for at least one }W\in\mathcal{F}.

The integral over EE of the number of bad disks is o(R)o(R).

Proof. The disks with the same centers and twice the radius are disjoint for each direction by the choice of ρ\rho. They lie in a fixed horizontal strip ∣Im⁡z∣≤B\lvert\operatorname{Im}z\rvert\leq B and, for large RR, in R/2≤Re⁡z≤4RR/2\leq\operatorname{Re}z\leq4R. For each point of the original disk, the disk of radius ρ\rho centered at that point lies in the enlarged disk. The submean inequality for the squared modulus of the holomorphic error gives a bound for the supremum of its square by 1/(πρ2)1/(\pi\rho^{2}) times the integral of the square over the enlarged disk. If bbad(R,u)b_{\mathrm{bad}}(R,u) is the bad count, summing over the disjoint enlarged disks gives

πρ2τ2bbad(R,u)≤∑W∈F∫−BB∫R/24R∣DW(r+iv,u)−MW(r+iv,u)∣2 dr dv.\pi\rho^{2}\tau^{2}b_{\mathrm{bad}}(R,u)\leq\sum_{W\in\mathcal{F}}\int_{-B}^{B}\int_{R/2}^{4R}\lvert D_{W}(r+iv,u)-M_{W}(r+iv,u)\rvert^{2}\,dr\,dv.

Let V∗V_{*} be the intersection of the sets from Proposition 3.1 for the comparison family, with strip bound BB and tolerances ε,η\varepsilon,\eta. Integrate this bound over E∩V∗E\cap V_{*}. Apply (8) to the intervals [R/2,R][R/2,R], [R,2R][R,2R] and [2R,4R][2R,4R], and integrate in vv. The resulting limsup of the integrated bad count divided by RR is OB,ρ,F(η2/τ2)O_{B,\rho,\mathcal{F}}(\eta^{2}/\tau^{2}). On E∖V∗E\setminus V_{*} the total count is O(R)O(R) per direction and the angular measure is at most ∣F∣ε\lvert\mathcal{F}\rvert\varepsilon. Sending ε,η\varepsilon,\eta to zero proves the bound. The counts are measurable, since each disk supremum can be taken on a fixed countable dense subset translated by its measurable center. □

Take τ\tau smaller than the uniform lower bound for ∣MX∣\lvert M_{X}\rvert on the disk boundaries. Call the disks meeting the error threshold for the whole comparison family good. Rouché’s theorem gives one zero of DX(z,u)D_{X}(z,u), counted with multiplicity, in each good disk. The normalizing factor in (6) is nonzero there, so this is a simple zero of z↦FX(zu)z\mapsto F_{X}(zu). The symmetry

FX(−ζˉ)=FX(ζ)‾(18)F_{X}(-\bar{\zeta})=\overline{F_{X}(\zeta)} \tag*{(18)}

gives the corresponding simple zero in the reflected disk across the imaginary axis. Fix a positive lower bound for the real parts of the original centers, greater than $2\rho$, so that the original disks and the reflected disks are disjoint. Let b(t,u)b(t,u) count the good disks and their reflections that lie entirely in ∣z∣≤t\lvert z\rvert\leq t. A disk with center qq lies inside this circle exactly when $\lvert q\rvert+\rho\leq t. Since the heights are bounded, the upper bound for the real part of qq is t−ρ+O(t−1)t-\rho+O(t^{-1}), uniformly in uu. The real spacing 2π/wX(u)2\pi/w_X(u) gives wX(u)t/π+O(1)w_X(u)t/\pi+O(1) disks counting both signs before subtracting the bad ones. The loss is o(t)o(t) in integral over EE by summing Lemma 4.1 over dyadic intervals. Indeed for every δ>0\delta> 0, the loss on [R,2R][R,2R] is at most δR\delta R in integral for all sufficiently large RR. The bounded range contributes O(1)O(1), and the other ranges contribute O(δt)O(\delta t). It follows that

∫Eb(t,u) dσ(u)=tπ∫EwX(u) dσ(u)+o(t).(19)\int_E b(t,u)\,d\sigma(u)=\frac{t}{\pi}\int_E w_X(u)\,d\sigma(u)+o(t). \tag*{(19)}

Let nX(t,u)n_X(t,u) count the zeros of z↦FX(zu)z\mapsto F_X(zu) in ∣z∣≤t|z|\le t with multiplicity. Bound the modulus by FX(0)exp⁡(−aX(u)Im⁡z)F_X(0)\exp(-a_X(u)\operatorname{Im}z) in the upper half-plane and by FX(0)exp⁡(−bX(u)Im⁡z)F_X(0)\exp(-b_X(u)\operatorname{Im}z) in the lower one. On the circle z=teiθz=te^{i\theta}, the upper and lower semicircles contribute −taX(u)/π-ta_X(u)/\pi and tbX(u)/πtb_X(u)/\pi, respectively, to the mean of the logarithmic bound after subtracting log⁡FX(0)\log F_X(0). Jensen’s formula therefore gives

∫0tnX(s,u) dss≤wX(u)tπ.(20)\int_0^t n_X(s,u)\,\frac{ds}{s}\le\frac{w_X(u)t}{\pi}. \tag*{(20)}

In particular nX(t,u)≤CXtn_X(t,u)\le C_Xt uniformly in uu, by using (20) at 2t2t and monotonicity.

Effective subdivisors

An effective subdivisor D≤div⁡FXD\le\operatorname{div}F_X assigns to each irreducible hypersurface of div⁡FX\operatorname{div}F_X a nonnegative integer at most equal to the multiplicity of that hypersurface. We use the finite-order divisor theorem of Lelong and Stoll [20, 24] in the form recorded in § 1 of [13]. An effective divisor of finite order in Cn\mathbb{C}^n is the zero divisor of an entire function of order at most that of the divisor. In particular, an integrated count O(R)O(R) gives a representative of order at most one.

Since FX(0)>0F_X(0)>0, every subdivisor avoids a neighborhood of zero. Its integrated count, up to a dimensional normalization, is

ND(R)=∫0Rvol⁡2n−2(D∩BCn(0,t))t2n−1 dt.N_D(R)=\int_0^R\frac{\operatorname{vol}_{2n-2}(D\cap B_{\mathbb{C}^n}(0,t))}{t^{2n-1}}\,dt.

where the hypersurface volume includes multiplicities. For div⁡FX\operatorname{div}F_X, Jensen’s formula in several variables [13 Section 1(c)–(d)] identifies this count, with the appropriate normalization, with the spherical mean of log⁡∣FX∣\log|F_X| at radius RR minus log⁡FX(0)\log F_X(0). It is O(R)O(R) since

∣FX(ζ)∣≤FX(0)exp⁡(RX∣ζ∣),RX=max⁡x∈X∣x∣.|F_X(\zeta)|\le F_X(0)\exp(R_X|\zeta|),\qquad R_X=\max_{x\in X}|x|.

The bound D≤div⁡FXD\le\operatorname{div}F_X gives ND(R)=O(R)N_D(R)=O(R). The divisor theorem gives an entire AA of order at most one with div⁡A=D\operatorname{div}A=D and A(0)≠0A(0)\ne0. If D=0D=0, take A=1A=1.

Write nD(t,u)n_D(t,u) for the number of zeros of A(zu)A(zu) in ∣z∣≤t|z|\le t, counted with multiplicity. At any point of a complex line through the origin, local defining functions for the same divisor differ by a holomorphic unit, whose restriction stays a unit. Also FXF_X is nonzero at the origin of every such line. Factoring FXF_X locally as AA times a defining function of the complementary divisor, up to a unit, and restricting this equality to the line gives additivity of orders even at tangent intersections. Thus the count is intrinsic to DD, with nD(t,u)≤nX(t,u)n_D(t,u) \le n_X(t,u). The counts are upper semicontinuous in (t,u)(t,u). At a given (t,u)(t,u), choose a slightly larger circle enclosing the same zeros and avoiding zeros on its boundary. Its count persists for nearby directions by the argument principle and bounds the count in nearby smaller closed disks. In particular the counts are measurable.

Lemma 4.2. Suppose EE has positive measure and all the simple zeros counted by b(t,u)b(t,u) in (19) lie outside DD on their radial lines. Then

∫EnD(t,u) dσ(u)=o(t).(21)\int_E n_D(t,u)\,d\sigma(u) = o(t). \tag*{(21)}

Proof. The count nX(s,u)n_X(s,u) dominates nD(s,u)+b(s,u)n_D(s,u)+b(s,u). Integrate (20) over EE and use (19). In the latter formula, integrating the o(s)o(s) error against ds/sds/s gives o(t)o(t): for any ϵ>0\epsilon>0 it is bounded in modulus by ϵs\epsilon s beyond a fixed cutoff, and the integral below the cutoff is finite. Consequently

∫0t∫Eb(s,u) dσ(u) dss=tπ∫EwX dσ+o(t).\int_0^t \int_E b(s,u)\,d\sigma(u)\,\frac{ds}{s} = \frac{t}{\pi}\int_E w_X\,d\sigma+ o(t).

The leading terms cancel and give

0≤∫0t∫EnD(s,u) dσ(u) dss=o(t).0 \le\int_0^t \int_E n_D(s,u)\,d\sigma(u)\,\frac{ds}{s} = o(t).

The estimate at 2t2t bounds (log⁡2)∫EnD(t,u) dσ(u)(\log2)\int_E n_D(t,u)\,d\sigma(u) by o(t)o(t).

Growth of factors with sparse radial zeros

Proposition 5.1. Let D≤div⁡FXD \le\operatorname{div} F_X be effective and invariant, including multiplicities, under ζ↦−ζ\zeta\mapsto-\zeta. Suppose (21) holds on a set EE of positive spherical measure. For any entire representative AA of order at most one there is a real linear functional ℓA\ell_A on Rn\mathbb{R}^n such that

lim⁡r→∞log⁡∣A(iru)∣r=ℓA(u)(u∈Sn−1).(22)\lim_{r\to\infty}\frac{\log|A(iru)|}{r} = \ell_A(u) \qquad(u \in S^{n-1}). \tag*{(22)}

Propagation by a polynomial inequality

Lemma 5.2. Suppose D≤div⁡FXD \le\operatorname{div} F_X satisfies (21). Let AA represent DD, with order at most one and A(0)=1A(0)=1. Put P(ζ)=A(ζ)A(−ζ)P(\zeta)=A(\zeta)A(-\zeta). Then

log⁡max⁡∣ζ∣≤R∣P(ζ)∣=o(R)(23)\log\max_{|\zeta|\le R}|P(\zeta)| = o(R) \tag*{(23)}

and

sup⁡u∈Sn−1nD(R,u)=o(R).(24)\sup_{u\in S^{n-1}} n_D(R,u) = o(R). \tag*{(24)}

Proof. For each real unit uu, P(zu)P(zu) is an even entire function of order at most one, with zero count nP(t,u)=2nD(t,u)≤Ctn_P(t,u)=2n_D(t,u)\le Ct. In the one-variable Hadamard factorization [21 Chapter I], pair the roots λ,−λ\lambda,-\lambda with multiplicities. The genus-one exponential factors cancel in each pair. The remaining exponential is constant by evenness, and the normalization at zero gives

P(zu)=∏{λ,−λ}(1−z2λ2).P(zu)=\prod_{\{\lambda,-\lambda\}}\left(1-\frac{z^2}{\lambda^2}\right).

The zero count O(t)O(t) gives ∑λ∣λ∣−2<∞\sum_\lambda|\lambda|^{-2}<\infty, so the product converges locally uniformly.

Taking moduli and integrating by parts in the zero-counting measure gives

log⁡max⁡∣z∣≤R∣P(zu)∣≤12∫0∞log⁡(1+R2/t2) dnP(t,u)=∫0∞R2nP(t,u)t(t2+R2) dt.(25)\log\max_{|z|\le R}|P(zu)|\le\frac{1}{2}\int_0^\infty\log(1+R^2/t^2)\,dn_P(t,u)=\int_0^\infty\frac{R^2n_P(t,u)}{t(t^2+R^2)}\,dt. \tag*{(25)}

There is no boundary contribution at zero, and the bound nP(t,u)=O(t)n_P(t,u)=O(t) gives the vanishing boundary term at infinity. Average over EE, divide by RR, and substitute t=Rst=Rs. The resulting integrand is

11+s2∫EnP(Rs,u) dσ(u)Rs.\frac{1}{1+s^2}\frac{\int_E n_P(Rs,u)\,d\sigma(u)}{Rs}.

It tends to zero for each s>0s>0 by (21) and is dominated by Cσ(E)/(1+s2)C\sigma(E)/(1+s^2). Consequently the average of the log maximum on the left of (25) is o(R)o(R).

The log maximum is nonnegative since P(0)=1P(0)=1. For each δ>0\delta>0 and all large RR, a subset ER⊂EE_R\subset E of measure at least σ(E)/2\sigma(E)/2 satisfies max⁡∣z∣≤R∣P(zu)∣≤eδR\max_{|z|\le R}|P(zu)|\le e^{\delta R}. Cauchy’s estimate for the degree-mm homogeneous Taylor polynomial PmP_m gives ∣Pm(u)∣≤eδRR−m|P_m(u)|\le e^{\delta R}R^{-m} on ERE_R. By homogeneity this holds on the cone over ERE_R inside the real unit ball, of volume at least σ(E)/(2n)\sigma(E)/(2n). The multivariate Remez inequality of Brudnyi and Ganzburg [10 Theorem 2 and Eq. (8)] gives the bound CmeδRR−mC^m e^{\delta R}R^{-m} on the whole real unit ball. The constant CC depends only on nn and the lower bound on the cone volume. For a complex polynomial the inequality applies to the real part of a constant phase multiple of that polynomial, with the phase chosen at the evaluation point.

For a complex-valued homogeneous polynomial QmQ_m of degree m≥1m\ge1, write BmB_m for its symmetric multilinear form. For real unit vectors x1,…,xmx_1,\ldots,x_m, polarization gives

Bm(x1,…,xm)=12mm!∑ϵj=±1(∏j=1mϵj)Qm(∑j=1mϵjxj).B_m(x_1,\ldots,x_m)=\frac{1}{2^m m!}\sum_{\epsilon_j=\pm1}\left(\prod_{j=1}^m\epsilon_j\right)Q_m\left(\sum_{j=1}^m\epsilon_jx_j\right).

The summed vector has norm at most mm. Homogeneity bounds ∣Bm∣|B_m| on real unit vectors by mm/m!m^m/m! times the supremum of ∣Qm∣|Q_m| on the real unit ball. Expand each of the mm arguments of Bm(ζ,…,ζ)B_m(\zeta,\ldots,\zeta) into real and imaginary parts of norm at most one for ∣ζ∣≤1|\zeta|\le1. Since mm/m!≤emm^m/m!\le e^m, the resulting complex supremum is at most (2e)m(2e)^m times the real supremum. Absorbing this factor into CC, the estimates on PmP_m give

∣Pm(ζ)∣≤eδR(C∣ζ∣/R)m.|P_m(\zeta)| \le e^{\delta R}(C|\zeta|/R)^m.

Summing the Taylor series on ∣ζ∣≤R/(2C)|\zeta| \le R/(2C) gives ∣P(ζ)∣≤2eδR|P(\zeta)| \le2e^{\delta R}. Rescaling the radius and letting δ↓0\delta\downarrow0 proves (23). Apply Jensen’s formula to P(zu)P(zu) at 2R2R. Then log⁡max⁡∣ζ∣≤2R∣P(ζ)∣\log\max_{|\zeta|\le2R}|P(\zeta)| bounds (log⁡2)nP(R,u)(\log2)n_P(R,u) uniformly in uu. By (23) this is o(R)o(R), which proves (24). □

Two bounds for the radial zeros

Lemma 5.3. The radial zeros λ\lambda of FX(zu)F_X(zu), for u∈Sn−1u \in S^{n-1}, lie outside a fixed disk at zero and satisfy

∣Re⁡λ∣≥cn∣λ∣(26)|\operatorname{Re}\lambda| \ge c_n|\lambda| \tag*{(26)}

for a constant cn>0c_n > 0. Counting with multiplicity,

sup⁡u∈Sn−1∑∣λ∣>s∣Im⁡λ∣∣λ∣2≤cX1+log⁡(1+s)s(s≥1).(27)\sup_{u\in S^{n-1}} \sum_{|\lambda|>s}\frac{|\operatorname{Im}\lambda|}{|\lambda|^2} \le c_X\frac{1+\log(1+s)}{s}\qquad(s\ge1). \tag*{(27)}

Proof. The disk at zero follows from

∣FX(zu)/FX(0)−1∣≤exp⁡(RX∣z∣)−1.\left|F_X(zu)/F_X(0)-1\right| \le\exp(R_X|z|)-1.

For the cone use the depth t=u⋅p−aX(u)≥0t=u\cdot p-a_X(u)\ge0, p∈Xp\in X, from the lower supporting plane, and let C(t)C(t) be the cap volume up to depth tt. A homothety of ratio 1/k1/k about a point on this plane in XX sends the cap of depth ktkt into the cap of depth tt for k≥1k\ge1. Convexity puts the image inside XX, and the volume scales by k−nk^{-n}. Thus C(kt)≤knC(t)C(kt)\le k^nC(t). At z=x+iyz=x+iy with y>0y>0, remove the exponential of the lower supporting height and normalize by the positive integral of e−yt dC(t)e^{-yt}\,dC(t). The resulting value is the expectation of eixte^{ixt} under this normalized measure. The denominator is at least e−1C(1/y)>0e^{-1}C(1/y)>0. Split ytyt into the ranges k≤yt<k+1k\le yt<k+1 for nonnegative integers kk. The numerator for the expectation of ytyt is at most

∑k=0∞(k+1)e−kC((k+1)/y)≤C(1/y)∑k=0∞(k+1)n+1e−k.\sum_{k=0}^{\infty}(k+1)e^{-k}C((k+1)/y) \le C(1/y)\sum_{k=0}^{\infty}(k+1)^{n+1}e^{-k}.

Therefore the expectation of ytyt is bounded by a constant depending only on nn. Since ∣eixt−1∣≤∣x∣t|e^{ixt}-1|\le|x|t, the normalized value stays within distance 1/21/2 of one when ∣x∣/y|x|/y is sufficiently small. Apply the same argument to the other endpoint in the lower half-plane. The two zero-free cones give (26).

For (27), the function

G(z)=FX(zu)e−iaX(u)z/FX(0)G(z)=F_X(zu)e^{-ia_X(u)z}/F_X(0)

has modulus at most one in the upper half-plane. A homothetic copy of XX with ratio min⁡(1,(swX(u))−1)\min(1,(sw_X(u))^{-1}), centered at a point in the lower support, lies in the cap of depth 1/s1/s. Integration over that cap gives

G(is)≥e−1min⁡{1,(swX(u))−n}≥cX(1+s)−n>0.G(is)\ge e^{-1}\min\{1,(sw_X(u))^{-n}\}\ge c_X(1+s)^{-n}>0.

Map the upper half-plane to the disk by z↦(z−is)/(z+is)z \mapsto(z-is)/(z+is) and apply Jensen’s formula at its center, letting the disk radius increase to one. The zeros in the upper half-plane satisfy

∑Im⁡λ>0log⁡∣λ+isλ−is∣≤−log⁡G(is)≤CX+nlog⁡(1+s).\sum_{\operatorname{Im}\lambda>0}\log\left|\frac{\lambda+is}{\lambda-is}\right|\leq-\log G(is)\leq C_X+n\log(1+s).

The inequality −log⁡r≥(1−r2)/2-\log r\geq(1-r^2)/2 for 0<r≤10<r\leq1 bounds each logarithm from below by 2sIm⁡λ/∣λ+is∣22s\operatorname{Im}\lambda/|\lambda+is|^2. For ∣λ∣>s|\lambda|>s the denominator is at most 4∣λ∣24|\lambda|^2. Apply the upper half-plane estimate in direction −u-u for the lower half-plane zeros. This proves (27). □

Linearity of the imaginary growth

Proof of Proposition 5.1. Normalize A(0)=1A(0)=1 and DA(0)=0DA(0)=0 by dividing by its value at zero and multiplying by a linear exponential. This changes the growth in (22) by a real linear functional of uu. By Lemma 5.2, (23) and (24) hold. For each uu pair the radial roots as λ\lambda, −λ‾-\overline{\lambda}, including multiplicity. The symmetry is the one in the hypothesis, and there are no fixed roots by (26). Define

Qu(z)=∏{λ,−λ‾}(1−zλ)(1+zλ‾),S(u)=∑{λ,−λ‾}(1λ−1λ‾).Q_u(z)=\prod_{\{\lambda,-\overline{\lambda}\}}\left(1-\frac{z}{\lambda}\right)\left(1+\frac{z}{\overline{\lambda}}\right),\qquad S(u)=\sum_{\{\lambda,-\overline{\lambda}\}}\left(\frac{1}{\lambda}-\frac{1}{\overline{\lambda}}\right).

In each pair the linear coefficient has absolute value 2∣Im⁡λ∣/∣λ∣22|\operatorname{Im}\lambda|/|\lambda|^2, and the quadratic coefficient has absolute value ∣λ∣−2|\lambda|^{-2}. The sums converge by the tail bound (27), the uniform count nD(t,u)≤CXtn_D(t,u)\leq C_Xt, and the root-free disk. These bounds are uniform in uu and give absolute, locally uniform convergence of the paired product in zz. In the genus-one Hadamard product each factor has value one and derivative zero at zero, so the value and derivative normalizations remove the prefactor. The two genus-one factors of a pair multiply to (1−z/λ)(1+z/λ‾)exp⁡(z(1/λ−1/λ‾))(1-z/\lambda)(1+z/\overline{\lambda})\exp(z(1/\lambda-1/\overline{\lambda})). This gives

A(zu)=exp⁡(zS(u))Qu(z).(28)A(zu)=\exp(zS(u))Q_u(z). \tag*{(28)}

The numbers S(u)S(u) are purely imaginary and uniformly bounded by (27) and the count of roots in a fixed disk.

For ∣z∣≤R|z|\leq R, bound log⁡∣Qu(z)∣\log|Q_u(z)| above by

∑{λ,−λ‾}log⁡(1+2R∣Im⁡(1/λ)∣+R2∣λ∣2).\sum_{\{\lambda,-\overline{\lambda}\}}\log\left(1+2R|\operatorname{Im}(1/\lambda)|+\frac{R^2}{|\lambda|^2}\right).

Use log⁡(1+a+b)≤log⁡(1+a)+log⁡(1+b)\log(1+a+b)\leq\log(1+a)+\log(1+b) for a,b≥0a,b\geq0. The sum of the quadratic logarithms is o(R)o(R) uniformly by the counting integral in (25), with nDn_D in place of nPn_P, and (24). For the linear logarithms, the pairs with ∣λ∣>s|\lambda|>s contribute at most 2R∑∣λ∣>s∣Im⁡(1/λ)∣2R\sum_{|\lambda|>s}|\operatorname{Im}(1/\lambda)|. The remaining pairs cost Os(log⁡(1+R))O_s(\log(1+R)) by the zero count and the root-free disk. Letting s→∞s\to\infty in (27) proves the uniform upper bound log⁡∣Qu(z)∣≤o(R)\log|Q_u(z)|\leq o(R) for ∣z∣≤R|z|\leq R. For z=±iRz=\pm iR, each individual factor satisfies

∣1−zλ∣=∣λ−z∣∣λ∣≥∣Re⁡λ∣∣λ∣≥cn.\left|1-\frac{z}{\lambda}\right|=\frac{|\lambda-z|}{|\lambda|}\ge\frac{|\operatorname{Re}\lambda|}{|\lambda|}\ge c_n.

The pairs with ∣λ∣≤2R|\lambda|\le2R therefore contribute at least −o(R)-o(R) to log⁡∣Qu(z)∣\log|Q_u(z)|. For ∣λ∣>2R|\lambda|>2R, apply log⁡∣1−w∣=−Re⁡w+O(∣w∣2)\log|1-w|=-\operatorname{Re}w+O(|w|^2) when ∣w∣≤1/2|w|\le1/2. The absolute sum of the paired linear terms is at most 2R∑∣λ∣>2R∣Im⁡(1/λ)∣=o(R)2R\sum_{|\lambda|>2R}|\operatorname{Im}(1/\lambda)|=o(R). The remainders are bounded by a constant times

R2∑∣λ∣>2R1∣λ∣2≤2R2∫2R∞nD(t,u)t3 dt=o(R)R^2\sum_{|\lambda|>2R}\frac{1}{|\lambda|^2}\le2R^2\int_{2R}^{\infty}\frac{n_D(t,u)}{t^3}\,dt=o(R)

uniformly, by (24). Thus

log⁡∣Qu(±iR)∣=o(R)uniformly for u∈Sn−1.(29)\log|Q_u(\pm iR)|=o(R)\qquad\text{uniformly for }u\in S^{n-1}. \tag*{(29)}

In particular the two imaginary limits in each direction exist and are opposite, with the positive one Re⁡(iS(u))\operatorname{Re}(iS(u)). The upper bound for QuQ_u and boundedness of S(u)S(u) give max⁡∣z∣≤R∣A(zu)∣≤exp⁡(CR)\max_{|z|\le R}|A(zu)|\le\exp(CR) for all large RR, uniformly in uu. If AmA_m is the degree-mm Taylor term of AA, Cauchy’s estimate on the real radial lines gives sup⁡∣u∣≤1, u∈Rn∣Am(u)∣≤eCRR−m\sup_{|u|\le1,\ u\in\mathbb{R}^n}|A_m(u)|\le e^{CR}R^{-m}. The polarization bound from the proof of Lemma 5.2 gives ∣Am(ζ)∣≤eCR(2e∣ζ∣/R)m|A_m(\zeta)|\le e^{CR}(2e|\zeta|/R)^m. Summation on ∣ζ∣≤R/(4e)|\zeta|\le R/(4e) gives ∣A(ζ)∣≤2eCR|A(\zeta)|\le2e^{CR} there, establishing finite exponential type on Cn\mathbb{C}^n.

Set uR(ζ)=R−1log⁡∣A(Rζ)∣u_R(\zeta)=R^{-1}\log|A(R\zeta)|. These functions are subharmonic on R2n\mathbb{R}^{2n}, locally uniformly bounded above, and uR(0)=0u_R(0)=0. Subharmonic compactness [17 Theorem 4.1.9] gives a sequence Rj→∞R_j\to\infty for which uRju_{R_j} converges in Lloc1L^1_{\mathrm{loc}} to a subharmonic function vv. The value at zero excludes local uniform convergence to −∞-\infty. For every fixed T>0T>0, (23) gives

sup⁡∣ζ∣≤T(uR(ζ)+uR(−ζ))≤R−1log⁡max⁡∣ξ∣≤RT∣P(ξ)∣⟶0.\sup_{|\zeta|\le T}\bigl(u_R(\zeta)+u_R(-\zeta)\bigr)\le R^{-1}\log\max_{|\xi|\le RT}|P(\xi)|\longrightarrow0.

Passing to the Lloc1L^1_{\mathrm{loc}} limit shows that v(ζ)+v(−ζ)≤0v(\zeta)+v(-\zeta)\le0 almost everywhere. The submean inequality for the sum yields the same bound everywhere. At the origin, uR(0)=0u_R(0)=0 and the submean inequality give

0≤1∣B(0,t)∣∫B(0,t)v(ζ) dζ(t>0)0\le\frac{1}{|B(0,t)|}\int_{B(0,t)}v(\zeta)\,d\zeta\qquad(t>0)

on passing to the limit. Upper semicontinuity then gives v(0)≥0v(0)\ge0. The sum has maximum zero at the origin, so it vanishes everywhere. Consequently −v=v(−⋅)-v=v(-\cdot) is subharmonic, and vv is harmonic. The exponential type bound on AA implies v(ζ)≤C∣ζ∣v(\zeta)\le C|\zeta| almost everywhere and hence everywhere by continuity. Oddness gives the corresponding lower bound. On balls of radius tending to infinity centered at any fixed point, the interior second-derivative estimate for vv tends to zero, by this linear growth bound. Thus vv is real linear.

At each ζ\zeta, take the submean bound for uRju_{R_j} on a ball centered at ζ\zeta, pass to the limit of the averages, and use the mean value property of vv. This gives lim sup⁡j→∞uRj(ζ)≤v(ζ)\limsup_{j\to\infty} u_{R_j}(\zeta) \le v(\zeta). For ζ=±iu\zeta=\pm iu, the limits of uR(±iu)u_R(\pm iu) as R→∞R\to\infty are g(u)=Re⁡(iS(u))g(u)=\operatorname{Re}(iS(u)) and −g(u)-g(u) by (28) and (29). Thus g(u)≤v(iu)g(u)\le v(iu) and −g(u)≤v(−iu)=−v(iu)-g(u)\le v(-iu)=-v(iu), which gives equality. This proves (22) with ℓA(u)=v(iu)\ell_A(u)=v(iu) under the chosen normalization. □

Determination of the bodies

For convex bodies K,LK,L, set

Δ=div⁡FK−div⁡FL,DK=Δ+,DL=(−Δ)+.\Delta=\operatorname{div} F_K-\operatorname{div} F_L,\qquad D_K=\Delta_+,\qquad D_L=(-\Delta)_+.

where positive parts are taken coefficientwise. By (18), both effective divisors are invariant under ζ↦−ζ\zeta\mapsto-\zeta.

Lemma 6.1. If each of DK,DLD_K,D_L satisfies (21) on some positive-measure set of real directions, then KK and LL are translates.

Proof. Let AK,ALA_K,A_L be their entire representatives of order at most one, with AK(0)=AL(0)=1A_K(0)=A_L(0)=1. The identity DK−DL=ΔD_K-D_L=\Delta makes the divisors of FKALF_KA_L and FLAKF_LA_K equal. On each complex line through zero the restrictions have the same zeros with multiplicity, by the restriction rule in § 4. Both restricted functions have order at most one. Their quotient on the line extends through the zeros and is the exponential of an affine function by Hadamard factorization. In a neighborhood of zero in Cn\mathbb{C}^n, the quotient is holomorphic and nonvanishing. Its value cc and its logarithmic differential Λ\Lambda at zero fix the constant factor and the linear term on each line. Varying the line gives

FK(ζ)AL(ζ)=cexp⁡(Λ⋅ζ)FL(ζ)AK(ζ)(ζ∈Cn)(30)F_K(\zeta)A_L(\zeta)=c\exp(\Lambda\cdot\zeta)F_L(\zeta)A_K(\zeta)\qquad(\zeta\in\mathbb{C}^n) \tag*{(30)}

for c≠0c\ne0 and Λ∈Cn\Lambda\in\mathbb{C}^n.

For every real unit uu,

lim⁡r→∞r−1log⁡FX(−iru)=hX(u).(31)\lim_{r\to\infty}r^{-1}\log F_X(-iru)=h_X(u). \tag*{(31)}

The integral is positive and is at most FX(0)erhX(u)F_X(0)e^{rh_X(u)}. A fixed cap of depth δ>0\delta>0 has positive volume and gives a lower bound of that volume times er(hX(u)−δ)e^{r(h_X(u)-\delta)}. Letting δ↓0\delta\downarrow0 proves (31). The functions FX(±iru)F_X(\pm iru) are positive, so the two subdivisors also avoid these rays. Evaluate (30) at −iru-iru, take logarithms of moduli, and divide by rr. Proposition 5.1 and (31) give

hK(u)−hL(u)=Re⁡(−iΛ⋅u)+ℓAK(−u)−ℓAL(−u).h_K(u)-h_L(u)=\operatorname{Re}(-i\Lambda\cdot u)+\ell_{A_K}(-u)-\ell_{A_L}(-u).

The right-hand side is a real linear functional of uu. Support functions determine convex bodies, and adding a linear functional to a support function translates the body. □

Proof of Theorem 1.1. The width and density product identities of Corollary 2.2 hold. The two endpoint densities of LL are positive almost everywhere on UKU_K. By (17), YKY_K is nonzero there. Choose E0⊂UKE_0\subset U_K of positive measure on which all four densities are bounded above and away from zero and ∣YK∣≥δ>0|Y_K|\ge\delta>0. Indeed the sets on which the densities lie in [m−1,m][m^{-1},m] and ∣YK∣≥m−1|Y_K| \ge m^{-1}, for positive integers mm, cover UKU_K up to a null set. For each u∈E0u \in E_0,

max⁡{∣YK(u)−YL(u)∣, ∣YK(u)+YL(u)∣}≥∣YK(u)∣.\max\{|Y_K(u)-Y_L(u)|,\ |Y_K(u)+Y_L(u)|\}\ge|Y_K(u)|.

At least one of the two separations is at least δ\delta on a positive-measure subset EE of E0E_0. Reflect LL if necessary to obtain ∣YK(u)+YL(u)∣≥δ|Y_K(u)+Y_L(u)|\ge\delta on EE.

Apply Lemma 4.1 with the centers for KK and with the comparison family {K,−L}\{K,-L\}. Choose ρ<δ/2\rho<\delta/2 and small enough for the enlarged disks to be disjoint. The height separation and the endpoint density bounds give a uniform positive lower bound for ∣M−L∣|M-L| on each whole disk. Choose τ\tau below both this bound and the lower bound for ∣MK∣|M_K| on the disk boundaries. Each good disk then contains a simple zero of FK(zu)F_K(zu), and F−L(zu)=FL(−zu)F_{-L}(zu)=F_L(-zu) is nonvanishing throughout the disk. By symmetry these statements also hold in the disks reflected across the imaginary axis.

At each of these points ζ0=zu\zeta_0=zu, FL(−ζ)F_L(-\zeta) is nonvanishing in a neighborhood in Cn\mathbb{C}^n. In that neighborhood (1) gives FL(ζ)=FK(ζ)FK(−ζ)/FL(−ζ)F_L(\zeta)=F_K(\zeta)F_K(-\zeta)/F_L(-\zeta), so FKF_K divides FLF_L there. The simple zeros counted in (19) therefore avoid the support of DKD_K. Lemma 4.2 gives (21) for DKD_K. For j(ζ)=−ζj(\zeta)=-\zeta, equality of the two product divisors in (1) is Δ+j∗Δ=0\Delta+j_*\Delta=0. Taking positive and negative parts gives DL=j∗DKD_L=j_*D_K. On a radial line, pullback by jj negates the zeros and preserves their multiplicities, so the two counts agree for each centered disk. Lemma 6.1 now gives the translation after the possible reflection of LL. □

Proof of Corollary 1.2. If σ(UK)>0\sigma(U_K)>0, apply Theorem 1.1. Otherwise positivity almost everywhere gives sK(u)=sK(−u)s_K(u)=s_K(-u) almost everywhere. Absolute continuity then gives SK=S−KS_K=S_{-K}, and Minkowski uniqueness [23 Theorem 7.2.1] makes KK centrally symmetric. Centrally symmetric convex bodies are determined by their covariograms, as in [7 Theorem 5.3]. For the convex competitors here, equality of covariograms gives K−K=L−LK-K=L-L and vol⁡n(K)=vol⁡n(L)\operatorname{vol}_n(K)=\operatorname{vol}_n(L). Hence

vol⁡n(L−L)=vol⁡n(K−K)=2nvol⁡n(K)=2nvol⁡n(L).\operatorname{vol}_n(L-L)=\operatorname{vol}_n(K-K)=2^n\operatorname{vol}_n(K)=2^n\operatorname{vol}_n(L).

This is equality in Brunn–Minkowski for LL and −L-L [23 Section 7.1]. The bodies are homothetic, and equal volume makes them translates, so LL is centrally symmetric. For each centrally symmetric body XX with center cXc_X, one has X−X=2(X−cX)X-X=2(X-c_X), which identifies LL and KK up to translation.

For C+C^+ bodies every normal has a unique supporting point, since a boundary segment would contradict positive curvature. The Gauss map is a C1C^1 diffeomorphism onto the sphere, its tangential derivative being the positive-definite curvature operator. Changing variables gives the continuous positive density sXs_X as the reciprocal of the Gauss curvature.

Proof of Theorem 1.3. Choose a positive-measure subset of TKT_K with uniform endpoint density bounds and apply Lemma 4.1 to {K}\{K\}. At any resulting simple radial zero ζ0=zu\zeta_0=zu, ∂uFK(ζ0)≠0\partial_uF_K(\zeta_0)\ne0. The holomorphic inverse function theorem supplies local coordinates ww in which w1=FKw_1=F_K, and the local divisor of FKF_K is w1=0w_1=0 with multiplicity one. Equality of the zero sets gives FL(0,w2,…,wn)=0F_L(0,w_2,\ldots,w_n)=0 in these coordinates, so the power series in w1w_1 shows that w1w_1 divides FLF_L. The zeros counted in (19), including the reflected ones, avoid the support of DKD_K. Lemma 4.2 applies to that divisor. Exchanging KK and LL proves (21) for DLD_L on a positive-measure subset of TLT_L. Lemma 6.1 uses the two sets of directions independently and yields the claimed translation.

Width and ordered curvature ratio in dimension three

Proof of Theorem 1.4. Let p=hK−hLp=h_K-h_L and P(x)=∣x∣p(x/∣x∣)P(x)=|x|p(x/|x|) for x≠0x\ne0, with P(0)=0P(0)=0. The width identity holds everywhere by continuity, so pp and PP are odd. Identify the tangent planes at uu and −u-u with the same Euclidean plane u⊥u^\perp, and set Ap(u)=AK(u)−AL(u)A_p(u)=A_K(u)-A_L(u). The homogeneous extension is Cloc1,1C_{\mathrm{loc}}^{1,1} off the origin. Its Hessian has radial row and column zero and tangent block ∣x∣−1Ap(x/∣x∣)|x|^{-1}A_p(x/|x|) almost everywhere, as in Section 2. Apply the chain rule twice to P(−x)=−P(x)P(-x)=-P(x) to obtain D2P(−x)=−D2P(x)D^2P(-x)=-D^2P(x) at points of twice differentiability in both directions. Restriction to the tangent plane yields Ap(−u)=−Ap(u)A_p(-u)=-A_p(u) almost everywhere. For 0≤t≤10\le t\le1 let Rt(u)=(1−t)AL(u)+tAK(u)R_t(u)=(1-t)A_L(u)+tA_K(u). Use (2) at uu and −u-u to write the logarithm of the ratio identity as

[log⁡det⁡Rt(u)−log⁡det⁡Rt(−u)]t=0t=1=0.[\log\det R_t(u)-\log\det R_t(-u)]_{t=0}^{t=1}=0.

Jacobi’s determinant formula and integration in tt give

0=tr⁡[(∫01(Rt(u)−1+Rt(−u)−1) dt)Ap(u)].0=\operatorname{tr}\left[\left(\int_0^1\left(R_t(u)^{-1}+R_t(-u)^{-1}\right)\,dt\right)A_p(u)\right].

The plus sign comes from Ap(−u)=−Ap(u)A_p(-u)=-A_p(u). The C1,1C^{1,1} bound and positivity give mI≤Rt(±u)≤MImI\le R_t(\pm u)\le MI uniformly almost everywhere, for some 0<m≤M<∞0<m\le M<\infty. The matrix in parentheses is bounded between 2M−1I2M^{-1}I and 2m−1I2m^{-1}I. For u=x/∣x∣u=x/|x|, let B(x)B(x) have this matrix on u⊥u^\perp and eigenvalue one on the radial direction. Set B=IB=I on the exceptional null set, including at zero. This gives a measurable uniformly elliptic coefficient field and tr⁡(B(x)D2P(x))=0\operatorname{tr}(B(x)D^2P(x))=0 almost everywhere.

The bounds ∣P(x)∣≤C∣x∣|P(x)|\le C|x|, ∣DP(x)∣≤C|DP(x)|\le C, and ∣D2P(x)∣≤C/∣x∣|D^2P(x)|\le C/|x| hold off the origin, with the Hessian bound holding almost everywhere. In dimension three ∫∣x∣≤1∣x∣−2 dx<∞\int_{|x|\le1}|x|^{-2}\,dx<\infty. To see that these derivatives are also weak derivatives across zero, integrate by parts outside the ball of radius ε\varepsilon. The first-derivative boundary term is O(ε3)O(\varepsilon^3) by the bound on PP, and the boundary term for the derivative of DPDP is O(ε2)O(\varepsilon^2). Both vanish in the limit, so P∈Wloc2,2(R3)P\in W_{\mathrm{loc}}^{2,2}(\mathbb{R}^3) and is an order-one homogeneous strong solution of the uniformly elliptic equation. Theorem 1.1 of Han, Nadirashvili and Yuan [16] gives that PP is linear, which amounts to translating the body.

Finally, for a C2C^2 boundary the inverse Gauss map x(u)=∇Sn−1hX(u)+hX(u)ux(u)=\nabla_{\mathbb{S}^{n-1}}h_X(u)+h_X(u)u is C1C^1. This identity gives hX∈C2h_X\in C^2, and its matrix AXA_X is the inverse of the Gauss map differential. Compactness therefore gives the uniform positive lower bound in the hypotheses. □\square

AI assistance

OpenAI Codex assisted with literature searches, proof development and checking, and preparation of the manuscript.

Chongqing Bashu Science City Secondary School, Chongqing, China

References

  1. [1]G. Averkov and G. Bianchi, Confirmation of Matheron’s conjecture on the covariogram of a planar convex body, J. Eur. Math. Soc. 11 (2009), 1187–1202. doi:10.4171/JEMS/179.DOI
  2. [2]G. Averkov and G. Bianchi, Retrieving convex bodies from restricted covariogram functions, Adv. Appl. Probab. 39 (2007), 613–629. doi:10.1239/aap/1189518630.arxiv.org/abs/math/0702892
  3. [3]G. Bianchi, Matheron’s conjecture for the covariogram problem, J. London Math. Soc. (2) 71 (2005), 203–220. doi:10.1112/S0024610704006039.DOI
  4. [4]G. Bianchi, Some open problems regarding the determination of a set from its covariogram, Le Matematiche 60 (2005), no. 2, 247–257. Journal record.
  5. [5]G. Bianchi, The covariogram determines three-dimensional convex polytopes, Adv. Math. 220 (2009), 1771–1808. doi:10.1016/j.aim.2008.11.011.DOI
  6. [6]G. Bianchi, The covariogram and Fourier–Laplace transform in Cⁿ, Proc. Lond. Math. Soc. (3) 113 (2016), 1–23. doi:10.1112/plms/pdw020.DOI
  7. [7]G. Bianchi, The covariogram problem, in Harmonic Analysis and Convexity, A. Koldobsky and A. Volberg (eds.), Adv. Anal. Geom., vol. 9, De Gruyter, Berlin, 2023, pp. 37–82. doi:10.1515/978311075389-002.DOI
  8. [8]G. Bianchi, F. F. Segala and A. Volčič, The solution of the covariogram problem for plane C²₊ convex bodies, J. Differential Geom. 60 (2002), 177–198. doi:10.4310/jdg/1090351101.DOI
  9. [9]L. Brandolini, S. Hofmann and A. Iosevich, Sharp rate of average decay of the Fourier transform of a bounded set, Geom. Funct. Anal. 13 (2003), 671–680. doi:10.1007/s00039-003-0426-7.DOI
  10. [10]Yu. A. Brudnyi and M. I. Ganzburg, On an extremal problem for polynomials in n variables, Math. USSR-Izv. 7 (1973), 345–356. doi:10.1070/IM1973v007n02ABEH001941.DOI
  11. [11]A. Colesanti and D. Hug, Hessian measures of convex functions and applications to area measures, J. London Math. Soc. (2) 71 (2005), 221–235. doi:10.1112/S0024610704005915.DOI
  12. [12]A. Colesanti and D. Hug, Hessian measures of semi-convex functions and applications to support measures of convex bodies, Manuscripta Math. 101 (2000), 209–238. doi:10.1007/s002290050015.DOI
  13. [13]M. Cornalba and P. A. Griffiths, Analytic cycles and vector bundles on non-compact algebraic varieties, Invent. Math. 28 (1975), 1–106. doi:10.1007/BF01389905.DOI
  14. [14]L. C. Evans and W. Gangbo, Differential equations methods for the Monge–Kantorovich mass transfer problem, Mem. Amer. Math. Soc. 137 (1999), no. 653. doi:10.1090/memo/0653.DOI
  15. [15]L. C. Evans and R. F. Gariepy, Measure Theory and Fine Properties of Functions, CRC Press, Boca Raton, FL, 1992.
  16. [16]Q. Han, N. Nadirashvili and Y. Yuan, Linearity of homogeneous order-one solutions to elliptic equations in dimension three, Comm. Pure Appl. Math. 56 (2003), 425–432. doi:10.1002/cpa.10064.DOI
  17. [17]L. Hörmander, The Analysis of Linear Partial Differential Operators I. Distribution Theory and Fourier Analysis, 2nd ed., Springer-Verlag, Berlin, 1990.
  18. [18]T. Kobayashi, The null variety of the Fourier–Laplace transform of the characteristic function of a bounded domain, Seminar Reports of Unitary Representation 6 (1986), 1–18 (in Japanese).
  19. [19]T. Kobayashi, Asymptotic behaviour of the null variety for a convex domain in a non-positively curved space form, J. Fac. Sci. Univ. Tokyo Sect. IA Math. 36 (1989), no. 3, 389–478. Author copy.
  20. [20]P. Lelong, Fonctions entières (n variables) et fonctions plurisousharmoniques d’ordre fini dans Cⁿ, J. Analyse Math. 12 (1964), 365–407. doi:10.1007/BF02807441.DOI
  21. [21]B. Ya. Levin, Distribution of Zeros of Entire Functions, revised ed., Translations of Mathematical Monographs, vol. 5, American Mathematical Society, Providence, RI, 1980. doi:10.1090/mmono/005.
  22. [22]G. Matheron, Le covariogramme géométrique des compacts convexes de R², Technical Report N-2/86/G, Centre de Géostatistique, École Nationale Supérieure des Mines de Paris, 1986.
  23. [23]R. Schneider, Convex Bodies: The Brunn–Minkowski Theory, 2nd expanded ed., Encyclopedia Math. Appl., vol. 151, Cambridge University Press, Cambridge, 2014. doi:10.1017/CBO9781139003858.
  24. [24]W. Stoll, Holomorphic Functions of Finite Order in Several Complex Variables, CBMS Regional Conf. Ser. in Math., vol. 21, American Mathematical Society, Providence, RI, 1974. doi:10.1090/cbms/021.

Paper details

vlnc:W_qufy8tqn
Version
v1 · latest
Published
Fresh — published within the last 30 days
License
CC BY 4.0
Usage

views · 1

downloads · 0

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

  • Mathematics

OpenAI ChatGPT GPT 6.2 Sol

View statement
Looking for related papers…
Contents